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>
31 #include <trace/events/kmem.h>
38 * The slab_lock protects operations on the object of a particular
39 * slab and its metadata in the page struct. If the slab lock
40 * has been taken then no allocations nor frees can be performed
41 * on the objects in the slab nor can the slab be added or removed
42 * from the partial or full lists since this would mean modifying
43 * the page_struct of the slab.
45 * The list_lock protects the partial and full list on each node and
46 * the partial slab counter. If taken then no new slabs may be added or
47 * removed from the lists nor make the number of partial slabs be modified.
48 * (Note that the total number of slabs is an atomic value that may be
49 * modified without taking the list lock).
51 * The list_lock is a centralized lock and thus we avoid taking it as
52 * much as possible. As long as SLUB does not have to handle partial
53 * slabs, operations can continue without any centralized lock. F.e.
54 * allocating a long series of objects that fill up slabs does not require
57 * The lock order is sometimes inverted when we are trying to get a slab
58 * off a list. We take the list_lock and then look for a page on the list
59 * to use. While we do that objects in the slabs may be freed. We can
60 * only operate on the slab if we have also taken the slab_lock. So we use
61 * a slab_trylock() on the slab. If trylock was successful then no frees
62 * can occur anymore and we can use the slab for allocations etc. If the
63 * slab_trylock() does not succeed then frees are in progress in the slab and
64 * we must stay away from it for a while since we may cause a bouncing
65 * cacheline if we try to acquire the lock. So go onto the next slab.
66 * If all pages are busy then we may allocate a new slab instead of reusing
67 * a partial slab. A new slab has no one operating on it and thus there is
68 * no danger of cacheline contention.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache *s)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
135 * Mininum number of partial slabs. These will be left on the partial
136 * lists even if they are empty. kmem_cache_shrink may reclaim them.
138 #define MIN_PARTIAL 5
141 * Maximum number of desirable partial slabs.
142 * The existence of more partial slabs makes kmem_cache_shrink
143 * sort the partial list by the number of objects in the.
145 #define MAX_PARTIAL 10
147 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
148 SLAB_POISON | SLAB_STORE_USER)
151 * Debugging flags that require metadata to be stored in the slab. These get
152 * disabled when slub_debug=O is used and a cache's min order increases with
155 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
158 * Set of flags that will prevent slab merging
160 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
161 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
164 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
165 SLAB_CACHE_DMA | SLAB_NOTRACK)
168 #define OO_MASK ((1 << OO_SHIFT) - 1)
169 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
171 /* Internal SLUB flags */
172 #define __OBJECT_POISON 0x80000000UL /* Poison object */
174 static int kmem_size = sizeof(struct kmem_cache);
177 static struct notifier_block slab_notifier;
181 DOWN, /* No slab functionality available */
182 PARTIAL, /* Kmem_cache_node works */
183 UP, /* Everything works but does not show up in sysfs */
187 /* A list of all slab caches on the system */
188 static DECLARE_RWSEM(slub_lock);
189 static LIST_HEAD(slab_caches);
192 * Tracking user of a slab.
194 #define TRACK_ADDRS_COUNT 16
196 unsigned long addr; /* Called from address */
197 #ifdef CONFIG_STACKTRACE
198 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
200 int cpu; /* Was running on cpu */
201 int pid; /* Pid context */
202 unsigned long when; /* When did the operation occur */
205 enum track_item { TRACK_ALLOC, TRACK_FREE };
208 static int sysfs_slab_add(struct kmem_cache *);
209 static int sysfs_slab_alias(struct kmem_cache *, const char *);
210 static void sysfs_slab_remove(struct kmem_cache *);
213 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
214 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
216 static inline void sysfs_slab_remove(struct kmem_cache *s)
224 static inline void stat(const struct kmem_cache *s, enum stat_item si)
226 #ifdef CONFIG_SLUB_STATS
227 __this_cpu_inc(s->cpu_slab->stat[si]);
231 /********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
235 int slab_is_available(void)
237 return slab_state >= UP;
240 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
242 return s->node[node];
245 /* Verify that a pointer has an address that is valid within a slab page */
246 static inline int check_valid_pointer(struct kmem_cache *s,
247 struct page *page, const void *object)
254 base = page_address(page);
255 if (object < base || object >= base + page->objects * s->size ||
256 (object - base) % s->size) {
263 static inline void *get_freepointer(struct kmem_cache *s, void *object)
265 return *(void **)(object + s->offset);
268 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
272 #ifdef CONFIG_DEBUG_PAGEALLOC
273 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
275 p = get_freepointer(s, object);
280 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
282 *(void **)(object + s->offset) = fp;
285 /* Loop over all objects in a slab */
286 #define for_each_object(__p, __s, __addr, __objects) \
287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
290 /* Determine object index from a given position */
291 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293 return (p - addr) / s->size;
296 static inline size_t slab_ksize(const struct kmem_cache *s)
298 #ifdef CONFIG_SLUB_DEBUG
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
315 * Else we can use all the padding etc for the allocation
320 static inline int order_objects(int order, unsigned long size, int reserved)
322 return ((PAGE_SIZE << order) - reserved) / size;
325 static inline struct kmem_cache_order_objects oo_make(int order,
326 unsigned long size, int reserved)
328 struct kmem_cache_order_objects x = {
329 (order << OO_SHIFT) + order_objects(order, size, reserved)
335 static inline int oo_order(struct kmem_cache_order_objects x)
337 return x.x >> OO_SHIFT;
340 static inline int oo_objects(struct kmem_cache_order_objects x)
342 return x.x & OO_MASK;
345 #ifdef CONFIG_SLUB_DEBUG
347 * Determine a map of object in use on a page.
349 * Slab lock or node listlock must be held to guarantee that the page does
350 * not vanish from under us.
352 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
355 void *addr = page_address(page);
357 for (p = page->freelist; p; p = get_freepointer(s, p))
358 set_bit(slab_index(p, s, addr), map);
364 #ifdef CONFIG_SLUB_DEBUG_ON
365 static int slub_debug = DEBUG_DEFAULT_FLAGS;
367 static int slub_debug;
370 static char *slub_debug_slabs;
371 static int disable_higher_order_debug;
376 static void print_section(char *text, u8 *addr, unsigned int length)
384 for (i = 0; i < length; i++) {
386 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
389 printk(KERN_CONT " %02x", addr[i]);
391 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
393 printk(KERN_CONT " %s\n", ascii);
400 printk(KERN_CONT " ");
404 printk(KERN_CONT " %s\n", ascii);
408 static struct track *get_track(struct kmem_cache *s, void *object,
409 enum track_item alloc)
414 p = object + s->offset + sizeof(void *);
416 p = object + s->inuse;
421 static void set_track(struct kmem_cache *s, void *object,
422 enum track_item alloc, unsigned long addr)
424 struct track *p = get_track(s, object, alloc);
427 #ifdef CONFIG_STACKTRACE
428 struct stack_trace trace;
431 trace.nr_entries = 0;
432 trace.max_entries = TRACK_ADDRS_COUNT;
433 trace.entries = p->addrs;
435 save_stack_trace(&trace);
437 /* See rant in lockdep.c */
438 if (trace.nr_entries != 0 &&
439 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
442 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
446 p->cpu = smp_processor_id();
447 p->pid = current->pid;
450 memset(p, 0, sizeof(struct track));
453 static void init_tracking(struct kmem_cache *s, void *object)
455 if (!(s->flags & SLAB_STORE_USER))
458 set_track(s, object, TRACK_FREE, 0UL);
459 set_track(s, object, TRACK_ALLOC, 0UL);
462 static void print_track(const char *s, struct track *t)
467 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
468 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
469 #ifdef CONFIG_STACKTRACE
472 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
474 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
481 static void print_tracking(struct kmem_cache *s, void *object)
483 if (!(s->flags & SLAB_STORE_USER))
486 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
487 print_track("Freed", get_track(s, object, TRACK_FREE));
490 static void print_page_info(struct page *page)
492 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
493 page, page->objects, page->inuse, page->freelist, page->flags);
497 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
503 vsnprintf(buf, sizeof(buf), fmt, args);
505 printk(KERN_ERR "========================================"
506 "=====================================\n");
507 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
508 printk(KERN_ERR "----------------------------------------"
509 "-------------------------------------\n\n");
512 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
518 vsnprintf(buf, sizeof(buf), fmt, args);
520 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
523 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
525 unsigned int off; /* Offset of last byte */
526 u8 *addr = page_address(page);
528 print_tracking(s, p);
530 print_page_info(page);
532 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
533 p, p - addr, get_freepointer(s, p));
536 print_section("Bytes b4", p - 16, 16);
538 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
540 if (s->flags & SLAB_RED_ZONE)
541 print_section("Redzone", p + s->objsize,
542 s->inuse - s->objsize);
545 off = s->offset + sizeof(void *);
549 if (s->flags & SLAB_STORE_USER)
550 off += 2 * sizeof(struct track);
553 /* Beginning of the filler is the free pointer */
554 print_section("Padding", p + off, s->size - off);
559 static void object_err(struct kmem_cache *s, struct page *page,
560 u8 *object, char *reason)
562 slab_bug(s, "%s", reason);
563 print_trailer(s, page, object);
566 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
572 vsnprintf(buf, sizeof(buf), fmt, args);
574 slab_bug(s, "%s", buf);
575 print_page_info(page);
579 static void init_object(struct kmem_cache *s, void *object, u8 val)
583 if (s->flags & __OBJECT_POISON) {
584 memset(p, POISON_FREE, s->objsize - 1);
585 p[s->objsize - 1] = POISON_END;
588 if (s->flags & SLAB_RED_ZONE)
589 memset(p + s->objsize, val, s->inuse - s->objsize);
592 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
595 if (*start != (u8)value)
603 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
604 void *from, void *to)
606 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
607 memset(from, data, to - from);
610 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
611 u8 *object, char *what,
612 u8 *start, unsigned int value, unsigned int bytes)
617 fault = check_bytes(start, value, bytes);
622 while (end > fault && end[-1] == value)
625 slab_bug(s, "%s overwritten", what);
626 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
627 fault, end - 1, fault[0], value);
628 print_trailer(s, page, object);
630 restore_bytes(s, what, value, fault, end);
638 * Bytes of the object to be managed.
639 * If the freepointer may overlay the object then the free
640 * pointer is the first word of the object.
642 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
645 * object + s->objsize
646 * Padding to reach word boundary. This is also used for Redzoning.
647 * Padding is extended by another word if Redzoning is enabled and
650 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
651 * 0xcc (RED_ACTIVE) for objects in use.
654 * Meta data starts here.
656 * A. Free pointer (if we cannot overwrite object on free)
657 * B. Tracking data for SLAB_STORE_USER
658 * C. Padding to reach required alignment boundary or at mininum
659 * one word if debugging is on to be able to detect writes
660 * before the word boundary.
662 * Padding is done using 0x5a (POISON_INUSE)
665 * Nothing is used beyond s->size.
667 * If slabcaches are merged then the objsize and inuse boundaries are mostly
668 * ignored. And therefore no slab options that rely on these boundaries
669 * may be used with merged slabcaches.
672 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
674 unsigned long off = s->inuse; /* The end of info */
677 /* Freepointer is placed after the object. */
678 off += sizeof(void *);
680 if (s->flags & SLAB_STORE_USER)
681 /* We also have user information there */
682 off += 2 * sizeof(struct track);
687 return check_bytes_and_report(s, page, p, "Object padding",
688 p + off, POISON_INUSE, s->size - off);
691 /* Check the pad bytes at the end of a slab page */
692 static int slab_pad_check(struct kmem_cache *s, struct page *page)
700 if (!(s->flags & SLAB_POISON))
703 start = page_address(page);
704 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
705 end = start + length;
706 remainder = length % s->size;
710 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
713 while (end > fault && end[-1] == POISON_INUSE)
716 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
717 print_section("Padding", end - remainder, remainder);
719 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
723 static int check_object(struct kmem_cache *s, struct page *page,
724 void *object, u8 val)
727 u8 *endobject = object + s->objsize;
729 if (s->flags & SLAB_RED_ZONE) {
730 if (!check_bytes_and_report(s, page, object, "Redzone",
731 endobject, val, s->inuse - s->objsize))
734 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
735 check_bytes_and_report(s, page, p, "Alignment padding",
736 endobject, POISON_INUSE, s->inuse - s->objsize);
740 if (s->flags & SLAB_POISON) {
741 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
742 (!check_bytes_and_report(s, page, p, "Poison", p,
743 POISON_FREE, s->objsize - 1) ||
744 !check_bytes_and_report(s, page, p, "Poison",
745 p + s->objsize - 1, POISON_END, 1)))
748 * check_pad_bytes cleans up on its own.
750 check_pad_bytes(s, page, p);
753 if (!s->offset && val == SLUB_RED_ACTIVE)
755 * Object and freepointer overlap. Cannot check
756 * freepointer while object is allocated.
760 /* Check free pointer validity */
761 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
762 object_err(s, page, p, "Freepointer corrupt");
764 * No choice but to zap it and thus lose the remainder
765 * of the free objects in this slab. May cause
766 * another error because the object count is now wrong.
768 set_freepointer(s, p, NULL);
774 static int check_slab(struct kmem_cache *s, struct page *page)
778 VM_BUG_ON(!irqs_disabled());
780 if (!PageSlab(page)) {
781 slab_err(s, page, "Not a valid slab page");
785 maxobj = order_objects(compound_order(page), s->size, s->reserved);
786 if (page->objects > maxobj) {
787 slab_err(s, page, "objects %u > max %u",
788 s->name, page->objects, maxobj);
791 if (page->inuse > page->objects) {
792 slab_err(s, page, "inuse %u > max %u",
793 s->name, page->inuse, page->objects);
796 /* Slab_pad_check fixes things up after itself */
797 slab_pad_check(s, page);
802 * Determine if a certain object on a page is on the freelist. Must hold the
803 * slab lock to guarantee that the chains are in a consistent state.
805 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
808 void *fp = page->freelist;
810 unsigned long max_objects;
812 while (fp && nr <= page->objects) {
815 if (!check_valid_pointer(s, page, fp)) {
817 object_err(s, page, object,
818 "Freechain corrupt");
819 set_freepointer(s, object, NULL);
822 slab_err(s, page, "Freepointer corrupt");
823 page->freelist = NULL;
824 page->inuse = page->objects;
825 slab_fix(s, "Freelist cleared");
831 fp = get_freepointer(s, object);
835 max_objects = order_objects(compound_order(page), s->size, s->reserved);
836 if (max_objects > MAX_OBJS_PER_PAGE)
837 max_objects = MAX_OBJS_PER_PAGE;
839 if (page->objects != max_objects) {
840 slab_err(s, page, "Wrong number of objects. Found %d but "
841 "should be %d", page->objects, max_objects);
842 page->objects = max_objects;
843 slab_fix(s, "Number of objects adjusted.");
845 if (page->inuse != page->objects - nr) {
846 slab_err(s, page, "Wrong object count. Counter is %d but "
847 "counted were %d", page->inuse, page->objects - nr);
848 page->inuse = page->objects - nr;
849 slab_fix(s, "Object count adjusted.");
851 return search == NULL;
854 static void trace(struct kmem_cache *s, struct page *page, void *object,
857 if (s->flags & SLAB_TRACE) {
858 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
860 alloc ? "alloc" : "free",
865 print_section("Object", (void *)object, s->objsize);
872 * Hooks for other subsystems that check memory allocations. In a typical
873 * production configuration these hooks all should produce no code at all.
875 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
877 flags &= gfp_allowed_mask;
878 lockdep_trace_alloc(flags);
879 might_sleep_if(flags & __GFP_WAIT);
881 return should_failslab(s->objsize, flags, s->flags);
884 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
886 flags &= gfp_allowed_mask;
887 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
888 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
891 static inline void slab_free_hook(struct kmem_cache *s, void *x)
893 kmemleak_free_recursive(x, s->flags);
896 * Trouble is that we may no longer disable interupts in the fast path
897 * So in order to make the debug calls that expect irqs to be
898 * disabled we need to disable interrupts temporarily.
900 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
904 local_irq_save(flags);
905 kmemcheck_slab_free(s, x, s->objsize);
906 debug_check_no_locks_freed(x, s->objsize);
907 local_irq_restore(flags);
910 if (!(s->flags & SLAB_DEBUG_OBJECTS))
911 debug_check_no_obj_freed(x, s->objsize);
915 * Tracking of fully allocated slabs for debugging purposes.
917 static void add_full(struct kmem_cache_node *n, struct page *page)
919 spin_lock(&n->list_lock);
920 list_add(&page->lru, &n->full);
921 spin_unlock(&n->list_lock);
924 static void remove_full(struct kmem_cache *s, struct page *page)
926 struct kmem_cache_node *n;
928 if (!(s->flags & SLAB_STORE_USER))
931 n = get_node(s, page_to_nid(page));
933 spin_lock(&n->list_lock);
934 list_del(&page->lru);
935 spin_unlock(&n->list_lock);
938 /* Tracking of the number of slabs for debugging purposes */
939 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
941 struct kmem_cache_node *n = get_node(s, node);
943 return atomic_long_read(&n->nr_slabs);
946 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
948 return atomic_long_read(&n->nr_slabs);
951 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
953 struct kmem_cache_node *n = get_node(s, node);
956 * May be called early in order to allocate a slab for the
957 * kmem_cache_node structure. Solve the chicken-egg
958 * dilemma by deferring the increment of the count during
959 * bootstrap (see early_kmem_cache_node_alloc).
962 atomic_long_inc(&n->nr_slabs);
963 atomic_long_add(objects, &n->total_objects);
966 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
968 struct kmem_cache_node *n = get_node(s, node);
970 atomic_long_dec(&n->nr_slabs);
971 atomic_long_sub(objects, &n->total_objects);
974 /* Object debug checks for alloc/free paths */
975 static void setup_object_debug(struct kmem_cache *s, struct page *page,
978 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
981 init_object(s, object, SLUB_RED_INACTIVE);
982 init_tracking(s, object);
985 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
986 void *object, unsigned long addr)
988 if (!check_slab(s, page))
991 if (!on_freelist(s, page, object)) {
992 object_err(s, page, object, "Object already allocated");
996 if (!check_valid_pointer(s, page, object)) {
997 object_err(s, page, object, "Freelist Pointer check fails");
1001 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1004 /* Success perform special debug activities for allocs */
1005 if (s->flags & SLAB_STORE_USER)
1006 set_track(s, object, TRACK_ALLOC, addr);
1007 trace(s, page, object, 1);
1008 init_object(s, object, SLUB_RED_ACTIVE);
1012 if (PageSlab(page)) {
1014 * If this is a slab page then lets do the best we can
1015 * to avoid issues in the future. Marking all objects
1016 * as used avoids touching the remaining objects.
1018 slab_fix(s, "Marking all objects used");
1019 page->inuse = page->objects;
1020 page->freelist = NULL;
1025 static noinline int free_debug_processing(struct kmem_cache *s,
1026 struct page *page, void *object, unsigned long addr)
1028 if (!check_slab(s, page))
1031 if (!check_valid_pointer(s, page, object)) {
1032 slab_err(s, page, "Invalid object pointer 0x%p", object);
1036 if (on_freelist(s, page, object)) {
1037 object_err(s, page, object, "Object already free");
1041 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1044 if (unlikely(s != page->slab)) {
1045 if (!PageSlab(page)) {
1046 slab_err(s, page, "Attempt to free object(0x%p) "
1047 "outside of slab", object);
1048 } else if (!page->slab) {
1050 "SLUB <none>: no slab for object 0x%p.\n",
1054 object_err(s, page, object,
1055 "page slab pointer corrupt.");
1059 /* Special debug activities for freeing objects */
1060 if (!PageSlubFrozen(page) && !page->freelist)
1061 remove_full(s, page);
1062 if (s->flags & SLAB_STORE_USER)
1063 set_track(s, object, TRACK_FREE, addr);
1064 trace(s, page, object, 0);
1065 init_object(s, object, SLUB_RED_INACTIVE);
1069 slab_fix(s, "Object at 0x%p not freed", object);
1073 static int __init setup_slub_debug(char *str)
1075 slub_debug = DEBUG_DEFAULT_FLAGS;
1076 if (*str++ != '=' || !*str)
1078 * No options specified. Switch on full debugging.
1084 * No options but restriction on slabs. This means full
1085 * debugging for slabs matching a pattern.
1089 if (tolower(*str) == 'o') {
1091 * Avoid enabling debugging on caches if its minimum order
1092 * would increase as a result.
1094 disable_higher_order_debug = 1;
1101 * Switch off all debugging measures.
1106 * Determine which debug features should be switched on
1108 for (; *str && *str != ','; str++) {
1109 switch (tolower(*str)) {
1111 slub_debug |= SLAB_DEBUG_FREE;
1114 slub_debug |= SLAB_RED_ZONE;
1117 slub_debug |= SLAB_POISON;
1120 slub_debug |= SLAB_STORE_USER;
1123 slub_debug |= SLAB_TRACE;
1126 slub_debug |= SLAB_FAILSLAB;
1129 printk(KERN_ERR "slub_debug option '%c' "
1130 "unknown. skipped\n", *str);
1136 slub_debug_slabs = str + 1;
1141 __setup("slub_debug", setup_slub_debug);
1143 static unsigned long kmem_cache_flags(unsigned long objsize,
1144 unsigned long flags, const char *name,
1145 void (*ctor)(void *))
1148 * Enable debugging if selected on the kernel commandline.
1150 if (slub_debug && (!slub_debug_slabs ||
1151 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1152 flags |= slub_debug;
1157 static inline void setup_object_debug(struct kmem_cache *s,
1158 struct page *page, void *object) {}
1160 static inline int alloc_debug_processing(struct kmem_cache *s,
1161 struct page *page, void *object, unsigned long addr) { return 0; }
1163 static inline int free_debug_processing(struct kmem_cache *s,
1164 struct page *page, void *object, unsigned long addr) { return 0; }
1166 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1168 static inline int check_object(struct kmem_cache *s, struct page *page,
1169 void *object, u8 val) { return 1; }
1170 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1171 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1172 unsigned long flags, const char *name,
1173 void (*ctor)(void *))
1177 #define slub_debug 0
1179 #define disable_higher_order_debug 0
1181 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1183 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1185 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1187 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1190 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1193 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1196 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1198 #endif /* CONFIG_SLUB_DEBUG */
1201 * Slab allocation and freeing
1203 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1204 struct kmem_cache_order_objects oo)
1206 int order = oo_order(oo);
1208 flags |= __GFP_NOTRACK;
1210 if (node == NUMA_NO_NODE)
1211 return alloc_pages(flags, order);
1213 return alloc_pages_exact_node(node, flags, order);
1216 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1219 struct kmem_cache_order_objects oo = s->oo;
1222 flags |= s->allocflags;
1225 * Let the initial higher-order allocation fail under memory pressure
1226 * so we fall-back to the minimum order allocation.
1228 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1230 page = alloc_slab_page(alloc_gfp, node, oo);
1231 if (unlikely(!page)) {
1234 * Allocation may have failed due to fragmentation.
1235 * Try a lower order alloc if possible
1237 page = alloc_slab_page(flags, node, oo);
1241 stat(s, ORDER_FALLBACK);
1244 if (kmemcheck_enabled
1245 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1246 int pages = 1 << oo_order(oo);
1248 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1251 * Objects from caches that have a constructor don't get
1252 * cleared when they're allocated, so we need to do it here.
1255 kmemcheck_mark_uninitialized_pages(page, pages);
1257 kmemcheck_mark_unallocated_pages(page, pages);
1260 page->objects = oo_objects(oo);
1261 mod_zone_page_state(page_zone(page),
1262 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1263 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1269 static void setup_object(struct kmem_cache *s, struct page *page,
1272 setup_object_debug(s, page, object);
1273 if (unlikely(s->ctor))
1277 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1284 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1286 page = allocate_slab(s,
1287 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1291 inc_slabs_node(s, page_to_nid(page), page->objects);
1293 page->flags |= 1 << PG_slab;
1295 start = page_address(page);
1297 if (unlikely(s->flags & SLAB_POISON))
1298 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1301 for_each_object(p, s, start, page->objects) {
1302 setup_object(s, page, last);
1303 set_freepointer(s, last, p);
1306 setup_object(s, page, last);
1307 set_freepointer(s, last, NULL);
1309 page->freelist = start;
1315 static void __free_slab(struct kmem_cache *s, struct page *page)
1317 int order = compound_order(page);
1318 int pages = 1 << order;
1320 if (kmem_cache_debug(s)) {
1323 slab_pad_check(s, page);
1324 for_each_object(p, s, page_address(page),
1326 check_object(s, page, p, SLUB_RED_INACTIVE);
1329 kmemcheck_free_shadow(page, compound_order(page));
1331 mod_zone_page_state(page_zone(page),
1332 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1333 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1336 __ClearPageSlab(page);
1337 reset_page_mapcount(page);
1338 if (current->reclaim_state)
1339 current->reclaim_state->reclaimed_slab += pages;
1340 __free_pages(page, order);
1343 #define need_reserve_slab_rcu \
1344 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1346 static void rcu_free_slab(struct rcu_head *h)
1350 if (need_reserve_slab_rcu)
1351 page = virt_to_head_page(h);
1353 page = container_of((struct list_head *)h, struct page, lru);
1355 __free_slab(page->slab, page);
1358 static void free_slab(struct kmem_cache *s, struct page *page)
1360 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1361 struct rcu_head *head;
1363 if (need_reserve_slab_rcu) {
1364 int order = compound_order(page);
1365 int offset = (PAGE_SIZE << order) - s->reserved;
1367 VM_BUG_ON(s->reserved != sizeof(*head));
1368 head = page_address(page) + offset;
1371 * RCU free overloads the RCU head over the LRU
1373 head = (void *)&page->lru;
1376 call_rcu(head, rcu_free_slab);
1378 __free_slab(s, page);
1381 static void discard_slab(struct kmem_cache *s, struct page *page)
1383 dec_slabs_node(s, page_to_nid(page), page->objects);
1388 * Per slab locking using the pagelock
1390 static __always_inline void slab_lock(struct page *page)
1392 bit_spin_lock(PG_locked, &page->flags);
1395 static __always_inline void slab_unlock(struct page *page)
1397 __bit_spin_unlock(PG_locked, &page->flags);
1400 static __always_inline int slab_trylock(struct page *page)
1404 rc = bit_spin_trylock(PG_locked, &page->flags);
1409 * Management of partially allocated slabs
1411 static void add_partial(struct kmem_cache_node *n,
1412 struct page *page, int tail)
1414 spin_lock(&n->list_lock);
1417 list_add_tail(&page->lru, &n->partial);
1419 list_add(&page->lru, &n->partial);
1420 spin_unlock(&n->list_lock);
1423 static inline void __remove_partial(struct kmem_cache_node *n,
1426 list_del(&page->lru);
1430 static void remove_partial(struct kmem_cache *s, struct page *page)
1432 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1434 spin_lock(&n->list_lock);
1435 __remove_partial(n, page);
1436 spin_unlock(&n->list_lock);
1440 * Lock slab and remove from the partial list.
1442 * Must hold list_lock.
1444 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1447 if (slab_trylock(page)) {
1448 __remove_partial(n, page);
1449 __SetPageSlubFrozen(page);
1456 * Try to allocate a partial slab from a specific node.
1458 static struct page *get_partial_node(struct kmem_cache_node *n)
1463 * Racy check. If we mistakenly see no partial slabs then we
1464 * just allocate an empty slab. If we mistakenly try to get a
1465 * partial slab and there is none available then get_partials()
1468 if (!n || !n->nr_partial)
1471 spin_lock(&n->list_lock);
1472 list_for_each_entry(page, &n->partial, lru)
1473 if (lock_and_freeze_slab(n, page))
1477 spin_unlock(&n->list_lock);
1482 * Get a page from somewhere. Search in increasing NUMA distances.
1484 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1487 struct zonelist *zonelist;
1490 enum zone_type high_zoneidx = gfp_zone(flags);
1494 * The defrag ratio allows a configuration of the tradeoffs between
1495 * inter node defragmentation and node local allocations. A lower
1496 * defrag_ratio increases the tendency to do local allocations
1497 * instead of attempting to obtain partial slabs from other nodes.
1499 * If the defrag_ratio is set to 0 then kmalloc() always
1500 * returns node local objects. If the ratio is higher then kmalloc()
1501 * may return off node objects because partial slabs are obtained
1502 * from other nodes and filled up.
1504 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1505 * defrag_ratio = 1000) then every (well almost) allocation will
1506 * first attempt to defrag slab caches on other nodes. This means
1507 * scanning over all nodes to look for partial slabs which may be
1508 * expensive if we do it every time we are trying to find a slab
1509 * with available objects.
1511 if (!s->remote_node_defrag_ratio ||
1512 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1516 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1517 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1518 struct kmem_cache_node *n;
1520 n = get_node(s, zone_to_nid(zone));
1522 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1523 n->nr_partial > s->min_partial) {
1524 page = get_partial_node(n);
1537 * Get a partial page, lock it and return it.
1539 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1542 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1544 page = get_partial_node(get_node(s, searchnode));
1545 if (page || node != NUMA_NO_NODE)
1548 return get_any_partial(s, flags);
1552 * Move a page back to the lists.
1554 * Must be called with the slab lock held.
1556 * On exit the slab lock will have been dropped.
1558 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1561 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1563 __ClearPageSlubFrozen(page);
1566 if (page->freelist) {
1567 add_partial(n, page, tail);
1568 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1570 stat(s, DEACTIVATE_FULL);
1571 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1576 stat(s, DEACTIVATE_EMPTY);
1577 if (n->nr_partial < s->min_partial) {
1579 * Adding an empty slab to the partial slabs in order
1580 * to avoid page allocator overhead. This slab needs
1581 * to come after the other slabs with objects in
1582 * so that the others get filled first. That way the
1583 * size of the partial list stays small.
1585 * kmem_cache_shrink can reclaim any empty slabs from
1588 add_partial(n, page, 1);
1593 discard_slab(s, page);
1598 #ifdef CONFIG_PREEMPT
1600 * Calculate the next globally unique transaction for disambiguiation
1601 * during cmpxchg. The transactions start with the cpu number and are then
1602 * incremented by CONFIG_NR_CPUS.
1604 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1607 * No preemption supported therefore also no need to check for
1613 static inline unsigned long next_tid(unsigned long tid)
1615 return tid + TID_STEP;
1618 static inline unsigned int tid_to_cpu(unsigned long tid)
1620 return tid % TID_STEP;
1623 static inline unsigned long tid_to_event(unsigned long tid)
1625 return tid / TID_STEP;
1628 static inline unsigned int init_tid(int cpu)
1633 static inline void note_cmpxchg_failure(const char *n,
1634 const struct kmem_cache *s, unsigned long tid)
1636 #ifdef SLUB_DEBUG_CMPXCHG
1637 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1639 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1641 #ifdef CONFIG_PREEMPT
1642 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1643 printk("due to cpu change %d -> %d\n",
1644 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1647 if (tid_to_event(tid) != tid_to_event(actual_tid))
1648 printk("due to cpu running other code. Event %ld->%ld\n",
1649 tid_to_event(tid), tid_to_event(actual_tid));
1651 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1652 actual_tid, tid, next_tid(tid));
1654 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1657 void init_kmem_cache_cpus(struct kmem_cache *s)
1661 for_each_possible_cpu(cpu)
1662 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1665 * Remove the cpu slab
1667 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1670 struct page *page = c->page;
1674 stat(s, DEACTIVATE_REMOTE_FREES);
1676 * Merge cpu freelist into slab freelist. Typically we get here
1677 * because both freelists are empty. So this is unlikely
1680 while (unlikely(c->freelist)) {
1683 tail = 0; /* Hot objects. Put the slab first */
1685 /* Retrieve object from cpu_freelist */
1686 object = c->freelist;
1687 c->freelist = get_freepointer(s, c->freelist);
1689 /* And put onto the regular freelist */
1690 set_freepointer(s, object, page->freelist);
1691 page->freelist = object;
1695 c->tid = next_tid(c->tid);
1696 unfreeze_slab(s, page, tail);
1699 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1701 stat(s, CPUSLAB_FLUSH);
1703 deactivate_slab(s, c);
1709 * Called from IPI handler with interrupts disabled.
1711 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1713 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1715 if (likely(c && c->page))
1719 static void flush_cpu_slab(void *d)
1721 struct kmem_cache *s = d;
1723 __flush_cpu_slab(s, smp_processor_id());
1726 static void flush_all(struct kmem_cache *s)
1728 on_each_cpu(flush_cpu_slab, s, 1);
1732 * Check if the objects in a per cpu structure fit numa
1733 * locality expectations.
1735 static inline int node_match(struct kmem_cache_cpu *c, int node)
1738 if (node != NUMA_NO_NODE && c->node != node)
1744 static int count_free(struct page *page)
1746 return page->objects - page->inuse;
1749 static unsigned long count_partial(struct kmem_cache_node *n,
1750 int (*get_count)(struct page *))
1752 unsigned long flags;
1753 unsigned long x = 0;
1756 spin_lock_irqsave(&n->list_lock, flags);
1757 list_for_each_entry(page, &n->partial, lru)
1758 x += get_count(page);
1759 spin_unlock_irqrestore(&n->list_lock, flags);
1763 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1765 #ifdef CONFIG_SLUB_DEBUG
1766 return atomic_long_read(&n->total_objects);
1772 static noinline void
1773 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1778 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1780 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1781 "default order: %d, min order: %d\n", s->name, s->objsize,
1782 s->size, oo_order(s->oo), oo_order(s->min));
1784 if (oo_order(s->min) > get_order(s->objsize))
1785 printk(KERN_WARNING " %s debugging increased min order, use "
1786 "slub_debug=O to disable.\n", s->name);
1788 for_each_online_node(node) {
1789 struct kmem_cache_node *n = get_node(s, node);
1790 unsigned long nr_slabs;
1791 unsigned long nr_objs;
1792 unsigned long nr_free;
1797 nr_free = count_partial(n, count_free);
1798 nr_slabs = node_nr_slabs(n);
1799 nr_objs = node_nr_objs(n);
1802 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1803 node, nr_slabs, nr_objs, nr_free);
1808 * Slow path. The lockless freelist is empty or we need to perform
1811 * Interrupts are disabled.
1813 * Processing is still very fast if new objects have been freed to the
1814 * regular freelist. In that case we simply take over the regular freelist
1815 * as the lockless freelist and zap the regular freelist.
1817 * If that is not working then we fall back to the partial lists. We take the
1818 * first element of the freelist as the object to allocate now and move the
1819 * rest of the freelist to the lockless freelist.
1821 * And if we were unable to get a new slab from the partial slab lists then
1822 * we need to allocate a new slab. This is the slowest path since it involves
1823 * a call to the page allocator and the setup of a new slab.
1825 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1826 unsigned long addr, struct kmem_cache_cpu *c)
1830 unsigned long flags;
1832 local_irq_save(flags);
1833 #ifdef CONFIG_PREEMPT
1835 * We may have been preempted and rescheduled on a different
1836 * cpu before disabling interrupts. Need to reload cpu area
1839 c = this_cpu_ptr(s->cpu_slab);
1842 /* We handle __GFP_ZERO in the caller */
1843 gfpflags &= ~__GFP_ZERO;
1850 if (unlikely(!node_match(c, node)))
1853 stat(s, ALLOC_REFILL);
1856 object = page->freelist;
1857 if (unlikely(!object))
1859 if (kmem_cache_debug(s))
1862 c->freelist = get_freepointer(s, object);
1863 page->inuse = page->objects;
1864 page->freelist = NULL;
1867 c->tid = next_tid(c->tid);
1868 local_irq_restore(flags);
1869 stat(s, ALLOC_SLOWPATH);
1873 deactivate_slab(s, c);
1876 page = get_partial(s, gfpflags, node);
1878 stat(s, ALLOC_FROM_PARTIAL);
1879 c->node = page_to_nid(page);
1884 gfpflags &= gfp_allowed_mask;
1885 if (gfpflags & __GFP_WAIT)
1888 page = new_slab(s, gfpflags, node);
1890 if (gfpflags & __GFP_WAIT)
1891 local_irq_disable();
1894 c = __this_cpu_ptr(s->cpu_slab);
1895 stat(s, ALLOC_SLAB);
1900 __SetPageSlubFrozen(page);
1901 c->node = page_to_nid(page);
1905 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1906 slab_out_of_memory(s, gfpflags, node);
1907 local_irq_restore(flags);
1910 if (!alloc_debug_processing(s, page, object, addr))
1914 page->freelist = get_freepointer(s, object);
1915 deactivate_slab(s, c);
1917 c->node = NUMA_NO_NODE;
1918 local_irq_restore(flags);
1923 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1924 * have the fastpath folded into their functions. So no function call
1925 * overhead for requests that can be satisfied on the fastpath.
1927 * The fastpath works by first checking if the lockless freelist can be used.
1928 * If not then __slab_alloc is called for slow processing.
1930 * Otherwise we can simply pick the next object from the lockless free list.
1932 static __always_inline void *slab_alloc(struct kmem_cache *s,
1933 gfp_t gfpflags, int node, unsigned long addr)
1936 struct kmem_cache_cpu *c;
1939 if (slab_pre_alloc_hook(s, gfpflags))
1945 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
1946 * enabled. We may switch back and forth between cpus while
1947 * reading from one cpu area. That does not matter as long
1948 * as we end up on the original cpu again when doing the cmpxchg.
1950 c = __this_cpu_ptr(s->cpu_slab);
1953 * The transaction ids are globally unique per cpu and per operation on
1954 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
1955 * occurs on the right processor and that there was no operation on the
1956 * linked list in between.
1961 object = c->freelist;
1962 if (unlikely(!object || !node_match(c, node)))
1964 object = __slab_alloc(s, gfpflags, node, addr, c);
1968 * The cmpxchg will only match if there was no additional
1969 * operation and if we are on the right processor.
1971 * The cmpxchg does the following atomically (without lock semantics!)
1972 * 1. Relocate first pointer to the current per cpu area.
1973 * 2. Verify that tid and freelist have not been changed
1974 * 3. If they were not changed replace tid and freelist
1976 * Since this is without lock semantics the protection is only against
1977 * code executing on this cpu *not* from access by other cpus.
1979 if (unlikely(!irqsafe_cpu_cmpxchg_double(
1980 s->cpu_slab->freelist, s->cpu_slab->tid,
1982 get_freepointer_safe(s, object), next_tid(tid)))) {
1984 note_cmpxchg_failure("slab_alloc", s, tid);
1987 stat(s, ALLOC_FASTPATH);
1990 if (unlikely(gfpflags & __GFP_ZERO) && object)
1991 memset(object, 0, s->objsize);
1993 slab_post_alloc_hook(s, gfpflags, object);
1998 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2000 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2002 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2006 EXPORT_SYMBOL(kmem_cache_alloc);
2008 #ifdef CONFIG_TRACING
2009 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2011 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2012 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2015 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2017 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2019 void *ret = kmalloc_order(size, flags, order);
2020 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2023 EXPORT_SYMBOL(kmalloc_order_trace);
2027 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2029 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2031 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2032 s->objsize, s->size, gfpflags, node);
2036 EXPORT_SYMBOL(kmem_cache_alloc_node);
2038 #ifdef CONFIG_TRACING
2039 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2041 int node, size_t size)
2043 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2045 trace_kmalloc_node(_RET_IP_, ret,
2046 size, s->size, gfpflags, node);
2049 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2054 * Slow patch handling. This may still be called frequently since objects
2055 * have a longer lifetime than the cpu slabs in most processing loads.
2057 * So we still attempt to reduce cache line usage. Just take the slab
2058 * lock and free the item. If there is no additional partial page
2059 * handling required then we can return immediately.
2061 static void __slab_free(struct kmem_cache *s, struct page *page,
2062 void *x, unsigned long addr)
2065 void **object = (void *)x;
2066 unsigned long flags;
2068 local_irq_save(flags);
2070 stat(s, FREE_SLOWPATH);
2072 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2075 prior = page->freelist;
2076 set_freepointer(s, object, prior);
2077 page->freelist = object;
2080 if (unlikely(PageSlubFrozen(page))) {
2081 stat(s, FREE_FROZEN);
2085 if (unlikely(!page->inuse))
2089 * Objects left in the slab. If it was not on the partial list before
2092 if (unlikely(!prior)) {
2093 add_partial(get_node(s, page_to_nid(page)), page, 1);
2094 stat(s, FREE_ADD_PARTIAL);
2099 local_irq_restore(flags);
2105 * Slab still on the partial list.
2107 remove_partial(s, page);
2108 stat(s, FREE_REMOVE_PARTIAL);
2111 local_irq_restore(flags);
2113 discard_slab(s, page);
2117 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2118 * can perform fastpath freeing without additional function calls.
2120 * The fastpath is only possible if we are freeing to the current cpu slab
2121 * of this processor. This typically the case if we have just allocated
2124 * If fastpath is not possible then fall back to __slab_free where we deal
2125 * with all sorts of special processing.
2127 static __always_inline void slab_free(struct kmem_cache *s,
2128 struct page *page, void *x, unsigned long addr)
2130 void **object = (void *)x;
2131 struct kmem_cache_cpu *c;
2134 slab_free_hook(s, x);
2139 * Determine the currently cpus per cpu slab.
2140 * The cpu may change afterward. However that does not matter since
2141 * data is retrieved via this pointer. If we are on the same cpu
2142 * during the cmpxchg then the free will succedd.
2144 c = __this_cpu_ptr(s->cpu_slab);
2149 if (likely(page == c->page)) {
2150 set_freepointer(s, object, c->freelist);
2152 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2153 s->cpu_slab->freelist, s->cpu_slab->tid,
2155 object, next_tid(tid)))) {
2157 note_cmpxchg_failure("slab_free", s, tid);
2160 stat(s, FREE_FASTPATH);
2162 __slab_free(s, page, x, addr);
2166 void kmem_cache_free(struct kmem_cache *s, void *x)
2170 page = virt_to_head_page(x);
2172 slab_free(s, page, x, _RET_IP_);
2174 trace_kmem_cache_free(_RET_IP_, x);
2176 EXPORT_SYMBOL(kmem_cache_free);
2179 * Object placement in a slab is made very easy because we always start at
2180 * offset 0. If we tune the size of the object to the alignment then we can
2181 * get the required alignment by putting one properly sized object after
2184 * Notice that the allocation order determines the sizes of the per cpu
2185 * caches. Each processor has always one slab available for allocations.
2186 * Increasing the allocation order reduces the number of times that slabs
2187 * must be moved on and off the partial lists and is therefore a factor in
2192 * Mininum / Maximum order of slab pages. This influences locking overhead
2193 * and slab fragmentation. A higher order reduces the number of partial slabs
2194 * and increases the number of allocations possible without having to
2195 * take the list_lock.
2197 static int slub_min_order;
2198 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2199 static int slub_min_objects;
2202 * Merge control. If this is set then no merging of slab caches will occur.
2203 * (Could be removed. This was introduced to pacify the merge skeptics.)
2205 static int slub_nomerge;
2208 * Calculate the order of allocation given an slab object size.
2210 * The order of allocation has significant impact on performance and other
2211 * system components. Generally order 0 allocations should be preferred since
2212 * order 0 does not cause fragmentation in the page allocator. Larger objects
2213 * be problematic to put into order 0 slabs because there may be too much
2214 * unused space left. We go to a higher order if more than 1/16th of the slab
2217 * In order to reach satisfactory performance we must ensure that a minimum
2218 * number of objects is in one slab. Otherwise we may generate too much
2219 * activity on the partial lists which requires taking the list_lock. This is
2220 * less a concern for large slabs though which are rarely used.
2222 * slub_max_order specifies the order where we begin to stop considering the
2223 * number of objects in a slab as critical. If we reach slub_max_order then
2224 * we try to keep the page order as low as possible. So we accept more waste
2225 * of space in favor of a small page order.
2227 * Higher order allocations also allow the placement of more objects in a
2228 * slab and thereby reduce object handling overhead. If the user has
2229 * requested a higher mininum order then we start with that one instead of
2230 * the smallest order which will fit the object.
2232 static inline int slab_order(int size, int min_objects,
2233 int max_order, int fract_leftover, int reserved)
2237 int min_order = slub_min_order;
2239 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2240 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2242 for (order = max(min_order,
2243 fls(min_objects * size - 1) - PAGE_SHIFT);
2244 order <= max_order; order++) {
2246 unsigned long slab_size = PAGE_SIZE << order;
2248 if (slab_size < min_objects * size + reserved)
2251 rem = (slab_size - reserved) % size;
2253 if (rem <= slab_size / fract_leftover)
2261 static inline int calculate_order(int size, int reserved)
2269 * Attempt to find best configuration for a slab. This
2270 * works by first attempting to generate a layout with
2271 * the best configuration and backing off gradually.
2273 * First we reduce the acceptable waste in a slab. Then
2274 * we reduce the minimum objects required in a slab.
2276 min_objects = slub_min_objects;
2278 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2279 max_objects = order_objects(slub_max_order, size, reserved);
2280 min_objects = min(min_objects, max_objects);
2282 while (min_objects > 1) {
2284 while (fraction >= 4) {
2285 order = slab_order(size, min_objects,
2286 slub_max_order, fraction, reserved);
2287 if (order <= slub_max_order)
2295 * We were unable to place multiple objects in a slab. Now
2296 * lets see if we can place a single object there.
2298 order = slab_order(size, 1, slub_max_order, 1, reserved);
2299 if (order <= slub_max_order)
2303 * Doh this slab cannot be placed using slub_max_order.
2305 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2306 if (order < MAX_ORDER)
2312 * Figure out what the alignment of the objects will be.
2314 static unsigned long calculate_alignment(unsigned long flags,
2315 unsigned long align, unsigned long size)
2318 * If the user wants hardware cache aligned objects then follow that
2319 * suggestion if the object is sufficiently large.
2321 * The hardware cache alignment cannot override the specified
2322 * alignment though. If that is greater then use it.
2324 if (flags & SLAB_HWCACHE_ALIGN) {
2325 unsigned long ralign = cache_line_size();
2326 while (size <= ralign / 2)
2328 align = max(align, ralign);
2331 if (align < ARCH_SLAB_MINALIGN)
2332 align = ARCH_SLAB_MINALIGN;
2334 return ALIGN(align, sizeof(void *));
2338 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2341 spin_lock_init(&n->list_lock);
2342 INIT_LIST_HEAD(&n->partial);
2343 #ifdef CONFIG_SLUB_DEBUG
2344 atomic_long_set(&n->nr_slabs, 0);
2345 atomic_long_set(&n->total_objects, 0);
2346 INIT_LIST_HEAD(&n->full);
2350 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2352 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2353 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2355 #ifdef CONFIG_CMPXCHG_LOCAL
2357 * Must align to double word boundary for the double cmpxchg instructions
2360 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *));
2362 /* Regular alignment is sufficient */
2363 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2369 init_kmem_cache_cpus(s);
2374 static struct kmem_cache *kmem_cache_node;
2377 * No kmalloc_node yet so do it by hand. We know that this is the first
2378 * slab on the node for this slabcache. There are no concurrent accesses
2381 * Note that this function only works on the kmalloc_node_cache
2382 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2383 * memory on a fresh node that has no slab structures yet.
2385 static void early_kmem_cache_node_alloc(int node)
2388 struct kmem_cache_node *n;
2389 unsigned long flags;
2391 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2393 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2396 if (page_to_nid(page) != node) {
2397 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2399 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2400 "in order to be able to continue\n");
2405 page->freelist = get_freepointer(kmem_cache_node, n);
2407 kmem_cache_node->node[node] = n;
2408 #ifdef CONFIG_SLUB_DEBUG
2409 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2410 init_tracking(kmem_cache_node, n);
2412 init_kmem_cache_node(n, kmem_cache_node);
2413 inc_slabs_node(kmem_cache_node, node, page->objects);
2416 * lockdep requires consistent irq usage for each lock
2417 * so even though there cannot be a race this early in
2418 * the boot sequence, we still disable irqs.
2420 local_irq_save(flags);
2421 add_partial(n, page, 0);
2422 local_irq_restore(flags);
2425 static void free_kmem_cache_nodes(struct kmem_cache *s)
2429 for_each_node_state(node, N_NORMAL_MEMORY) {
2430 struct kmem_cache_node *n = s->node[node];
2433 kmem_cache_free(kmem_cache_node, n);
2435 s->node[node] = NULL;
2439 static int init_kmem_cache_nodes(struct kmem_cache *s)
2443 for_each_node_state(node, N_NORMAL_MEMORY) {
2444 struct kmem_cache_node *n;
2446 if (slab_state == DOWN) {
2447 early_kmem_cache_node_alloc(node);
2450 n = kmem_cache_alloc_node(kmem_cache_node,
2454 free_kmem_cache_nodes(s);
2459 init_kmem_cache_node(n, s);
2464 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2466 if (min < MIN_PARTIAL)
2468 else if (min > MAX_PARTIAL)
2470 s->min_partial = min;
2474 * calculate_sizes() determines the order and the distribution of data within
2477 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2479 unsigned long flags = s->flags;
2480 unsigned long size = s->objsize;
2481 unsigned long align = s->align;
2485 * Round up object size to the next word boundary. We can only
2486 * place the free pointer at word boundaries and this determines
2487 * the possible location of the free pointer.
2489 size = ALIGN(size, sizeof(void *));
2491 #ifdef CONFIG_SLUB_DEBUG
2493 * Determine if we can poison the object itself. If the user of
2494 * the slab may touch the object after free or before allocation
2495 * then we should never poison the object itself.
2497 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2499 s->flags |= __OBJECT_POISON;
2501 s->flags &= ~__OBJECT_POISON;
2505 * If we are Redzoning then check if there is some space between the
2506 * end of the object and the free pointer. If not then add an
2507 * additional word to have some bytes to store Redzone information.
2509 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2510 size += sizeof(void *);
2514 * With that we have determined the number of bytes in actual use
2515 * by the object. This is the potential offset to the free pointer.
2519 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2522 * Relocate free pointer after the object if it is not
2523 * permitted to overwrite the first word of the object on
2526 * This is the case if we do RCU, have a constructor or
2527 * destructor or are poisoning the objects.
2530 size += sizeof(void *);
2533 #ifdef CONFIG_SLUB_DEBUG
2534 if (flags & SLAB_STORE_USER)
2536 * Need to store information about allocs and frees after
2539 size += 2 * sizeof(struct track);
2541 if (flags & SLAB_RED_ZONE)
2543 * Add some empty padding so that we can catch
2544 * overwrites from earlier objects rather than let
2545 * tracking information or the free pointer be
2546 * corrupted if a user writes before the start
2549 size += sizeof(void *);
2553 * Determine the alignment based on various parameters that the
2554 * user specified and the dynamic determination of cache line size
2557 align = calculate_alignment(flags, align, s->objsize);
2561 * SLUB stores one object immediately after another beginning from
2562 * offset 0. In order to align the objects we have to simply size
2563 * each object to conform to the alignment.
2565 size = ALIGN(size, align);
2567 if (forced_order >= 0)
2568 order = forced_order;
2570 order = calculate_order(size, s->reserved);
2577 s->allocflags |= __GFP_COMP;
2579 if (s->flags & SLAB_CACHE_DMA)
2580 s->allocflags |= SLUB_DMA;
2582 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2583 s->allocflags |= __GFP_RECLAIMABLE;
2586 * Determine the number of objects per slab
2588 s->oo = oo_make(order, size, s->reserved);
2589 s->min = oo_make(get_order(size), size, s->reserved);
2590 if (oo_objects(s->oo) > oo_objects(s->max))
2593 return !!oo_objects(s->oo);
2597 static int kmem_cache_open(struct kmem_cache *s,
2598 const char *name, size_t size,
2599 size_t align, unsigned long flags,
2600 void (*ctor)(void *))
2602 memset(s, 0, kmem_size);
2607 s->flags = kmem_cache_flags(size, flags, name, ctor);
2610 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2611 s->reserved = sizeof(struct rcu_head);
2613 if (!calculate_sizes(s, -1))
2615 if (disable_higher_order_debug) {
2617 * Disable debugging flags that store metadata if the min slab
2620 if (get_order(s->size) > get_order(s->objsize)) {
2621 s->flags &= ~DEBUG_METADATA_FLAGS;
2623 if (!calculate_sizes(s, -1))
2629 * The larger the object size is, the more pages we want on the partial
2630 * list to avoid pounding the page allocator excessively.
2632 set_min_partial(s, ilog2(s->size));
2635 s->remote_node_defrag_ratio = 1000;
2637 if (!init_kmem_cache_nodes(s))
2640 if (alloc_kmem_cache_cpus(s))
2643 free_kmem_cache_nodes(s);
2645 if (flags & SLAB_PANIC)
2646 panic("Cannot create slab %s size=%lu realsize=%u "
2647 "order=%u offset=%u flags=%lx\n",
2648 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2654 * Determine the size of a slab object
2656 unsigned int kmem_cache_size(struct kmem_cache *s)
2660 EXPORT_SYMBOL(kmem_cache_size);
2662 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2665 #ifdef CONFIG_SLUB_DEBUG
2666 void *addr = page_address(page);
2668 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2669 sizeof(long), GFP_ATOMIC);
2672 slab_err(s, page, "%s", text);
2675 get_map(s, page, map);
2676 for_each_object(p, s, addr, page->objects) {
2678 if (!test_bit(slab_index(p, s, addr), map)) {
2679 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2681 print_tracking(s, p);
2690 * Attempt to free all partial slabs on a node.
2692 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2694 unsigned long flags;
2695 struct page *page, *h;
2697 spin_lock_irqsave(&n->list_lock, flags);
2698 list_for_each_entry_safe(page, h, &n->partial, lru) {
2700 __remove_partial(n, page);
2701 discard_slab(s, page);
2703 list_slab_objects(s, page,
2704 "Objects remaining on kmem_cache_close()");
2707 spin_unlock_irqrestore(&n->list_lock, flags);
2711 * Release all resources used by a slab cache.
2713 static inline int kmem_cache_close(struct kmem_cache *s)
2718 free_percpu(s->cpu_slab);
2719 /* Attempt to free all objects */
2720 for_each_node_state(node, N_NORMAL_MEMORY) {
2721 struct kmem_cache_node *n = get_node(s, node);
2724 if (n->nr_partial || slabs_node(s, node))
2727 free_kmem_cache_nodes(s);
2732 * Close a cache and release the kmem_cache structure
2733 * (must be used for caches created using kmem_cache_create)
2735 void kmem_cache_destroy(struct kmem_cache *s)
2737 down_write(&slub_lock);
2741 if (kmem_cache_close(s)) {
2742 printk(KERN_ERR "SLUB %s: %s called for cache that "
2743 "still has objects.\n", s->name, __func__);
2746 if (s->flags & SLAB_DESTROY_BY_RCU)
2748 sysfs_slab_remove(s);
2750 up_write(&slub_lock);
2752 EXPORT_SYMBOL(kmem_cache_destroy);
2754 /********************************************************************
2756 *******************************************************************/
2758 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2759 EXPORT_SYMBOL(kmalloc_caches);
2761 static struct kmem_cache *kmem_cache;
2763 #ifdef CONFIG_ZONE_DMA
2764 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2767 static int __init setup_slub_min_order(char *str)
2769 get_option(&str, &slub_min_order);
2774 __setup("slub_min_order=", setup_slub_min_order);
2776 static int __init setup_slub_max_order(char *str)
2778 get_option(&str, &slub_max_order);
2779 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2784 __setup("slub_max_order=", setup_slub_max_order);
2786 static int __init setup_slub_min_objects(char *str)
2788 get_option(&str, &slub_min_objects);
2793 __setup("slub_min_objects=", setup_slub_min_objects);
2795 static int __init setup_slub_nomerge(char *str)
2801 __setup("slub_nomerge", setup_slub_nomerge);
2803 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2804 int size, unsigned int flags)
2806 struct kmem_cache *s;
2808 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2811 * This function is called with IRQs disabled during early-boot on
2812 * single CPU so there's no need to take slub_lock here.
2814 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2818 list_add(&s->list, &slab_caches);
2822 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2827 * Conversion table for small slabs sizes / 8 to the index in the
2828 * kmalloc array. This is necessary for slabs < 192 since we have non power
2829 * of two cache sizes there. The size of larger slabs can be determined using
2832 static s8 size_index[24] = {
2859 static inline int size_index_elem(size_t bytes)
2861 return (bytes - 1) / 8;
2864 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2870 return ZERO_SIZE_PTR;
2872 index = size_index[size_index_elem(size)];
2874 index = fls(size - 1);
2876 #ifdef CONFIG_ZONE_DMA
2877 if (unlikely((flags & SLUB_DMA)))
2878 return kmalloc_dma_caches[index];
2881 return kmalloc_caches[index];
2884 void *__kmalloc(size_t size, gfp_t flags)
2886 struct kmem_cache *s;
2889 if (unlikely(size > SLUB_MAX_SIZE))
2890 return kmalloc_large(size, flags);
2892 s = get_slab(size, flags);
2894 if (unlikely(ZERO_OR_NULL_PTR(s)))
2897 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2899 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2903 EXPORT_SYMBOL(__kmalloc);
2906 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2911 flags |= __GFP_COMP | __GFP_NOTRACK;
2912 page = alloc_pages_node(node, flags, get_order(size));
2914 ptr = page_address(page);
2916 kmemleak_alloc(ptr, size, 1, flags);
2920 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2922 struct kmem_cache *s;
2925 if (unlikely(size > SLUB_MAX_SIZE)) {
2926 ret = kmalloc_large_node(size, flags, node);
2928 trace_kmalloc_node(_RET_IP_, ret,
2929 size, PAGE_SIZE << get_order(size),
2935 s = get_slab(size, flags);
2937 if (unlikely(ZERO_OR_NULL_PTR(s)))
2940 ret = slab_alloc(s, flags, node, _RET_IP_);
2942 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2946 EXPORT_SYMBOL(__kmalloc_node);
2949 size_t ksize(const void *object)
2953 if (unlikely(object == ZERO_SIZE_PTR))
2956 page = virt_to_head_page(object);
2958 if (unlikely(!PageSlab(page))) {
2959 WARN_ON(!PageCompound(page));
2960 return PAGE_SIZE << compound_order(page);
2963 return slab_ksize(page->slab);
2965 EXPORT_SYMBOL(ksize);
2967 void kfree(const void *x)
2970 void *object = (void *)x;
2972 trace_kfree(_RET_IP_, x);
2974 if (unlikely(ZERO_OR_NULL_PTR(x)))
2977 page = virt_to_head_page(x);
2978 if (unlikely(!PageSlab(page))) {
2979 BUG_ON(!PageCompound(page));
2984 slab_free(page->slab, page, object, _RET_IP_);
2986 EXPORT_SYMBOL(kfree);
2989 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2990 * the remaining slabs by the number of items in use. The slabs with the
2991 * most items in use come first. New allocations will then fill those up
2992 * and thus they can be removed from the partial lists.
2994 * The slabs with the least items are placed last. This results in them
2995 * being allocated from last increasing the chance that the last objects
2996 * are freed in them.
2998 int kmem_cache_shrink(struct kmem_cache *s)
3002 struct kmem_cache_node *n;
3005 int objects = oo_objects(s->max);
3006 struct list_head *slabs_by_inuse =
3007 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3008 unsigned long flags;
3010 if (!slabs_by_inuse)
3014 for_each_node_state(node, N_NORMAL_MEMORY) {
3015 n = get_node(s, node);
3020 for (i = 0; i < objects; i++)
3021 INIT_LIST_HEAD(slabs_by_inuse + i);
3023 spin_lock_irqsave(&n->list_lock, flags);
3026 * Build lists indexed by the items in use in each slab.
3028 * Note that concurrent frees may occur while we hold the
3029 * list_lock. page->inuse here is the upper limit.
3031 list_for_each_entry_safe(page, t, &n->partial, lru) {
3032 if (!page->inuse && slab_trylock(page)) {
3034 * Must hold slab lock here because slab_free
3035 * may have freed the last object and be
3036 * waiting to release the slab.
3038 __remove_partial(n, page);
3040 discard_slab(s, page);
3042 list_move(&page->lru,
3043 slabs_by_inuse + page->inuse);
3048 * Rebuild the partial list with the slabs filled up most
3049 * first and the least used slabs at the end.
3051 for (i = objects - 1; i >= 0; i--)
3052 list_splice(slabs_by_inuse + i, n->partial.prev);
3054 spin_unlock_irqrestore(&n->list_lock, flags);
3057 kfree(slabs_by_inuse);
3060 EXPORT_SYMBOL(kmem_cache_shrink);
3062 #if defined(CONFIG_MEMORY_HOTPLUG)
3063 static int slab_mem_going_offline_callback(void *arg)
3065 struct kmem_cache *s;
3067 down_read(&slub_lock);
3068 list_for_each_entry(s, &slab_caches, list)
3069 kmem_cache_shrink(s);
3070 up_read(&slub_lock);
3075 static void slab_mem_offline_callback(void *arg)
3077 struct kmem_cache_node *n;
3078 struct kmem_cache *s;
3079 struct memory_notify *marg = arg;
3082 offline_node = marg->status_change_nid;
3085 * If the node still has available memory. we need kmem_cache_node
3088 if (offline_node < 0)
3091 down_read(&slub_lock);
3092 list_for_each_entry(s, &slab_caches, list) {
3093 n = get_node(s, offline_node);
3096 * if n->nr_slabs > 0, slabs still exist on the node
3097 * that is going down. We were unable to free them,
3098 * and offline_pages() function shouldn't call this
3099 * callback. So, we must fail.
3101 BUG_ON(slabs_node(s, offline_node));
3103 s->node[offline_node] = NULL;
3104 kmem_cache_free(kmem_cache_node, n);
3107 up_read(&slub_lock);
3110 static int slab_mem_going_online_callback(void *arg)
3112 struct kmem_cache_node *n;
3113 struct kmem_cache *s;
3114 struct memory_notify *marg = arg;
3115 int nid = marg->status_change_nid;
3119 * If the node's memory is already available, then kmem_cache_node is
3120 * already created. Nothing to do.
3126 * We are bringing a node online. No memory is available yet. We must
3127 * allocate a kmem_cache_node structure in order to bring the node
3130 down_read(&slub_lock);
3131 list_for_each_entry(s, &slab_caches, list) {
3133 * XXX: kmem_cache_alloc_node will fallback to other nodes
3134 * since memory is not yet available from the node that
3137 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3142 init_kmem_cache_node(n, s);
3146 up_read(&slub_lock);
3150 static int slab_memory_callback(struct notifier_block *self,
3151 unsigned long action, void *arg)
3156 case MEM_GOING_ONLINE:
3157 ret = slab_mem_going_online_callback(arg);
3159 case MEM_GOING_OFFLINE:
3160 ret = slab_mem_going_offline_callback(arg);
3163 case MEM_CANCEL_ONLINE:
3164 slab_mem_offline_callback(arg);
3167 case MEM_CANCEL_OFFLINE:
3171 ret = notifier_from_errno(ret);
3177 #endif /* CONFIG_MEMORY_HOTPLUG */
3179 /********************************************************************
3180 * Basic setup of slabs
3181 *******************************************************************/
3184 * Used for early kmem_cache structures that were allocated using
3185 * the page allocator
3188 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3192 list_add(&s->list, &slab_caches);
3195 for_each_node_state(node, N_NORMAL_MEMORY) {
3196 struct kmem_cache_node *n = get_node(s, node);
3200 list_for_each_entry(p, &n->partial, lru)
3203 #ifdef CONFIG_SLUB_DEBUG
3204 list_for_each_entry(p, &n->full, lru)
3211 void __init kmem_cache_init(void)
3215 struct kmem_cache *temp_kmem_cache;
3217 struct kmem_cache *temp_kmem_cache_node;
3218 unsigned long kmalloc_size;
3220 kmem_size = offsetof(struct kmem_cache, node) +
3221 nr_node_ids * sizeof(struct kmem_cache_node *);
3223 /* Allocate two kmem_caches from the page allocator */
3224 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3225 order = get_order(2 * kmalloc_size);
3226 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3229 * Must first have the slab cache available for the allocations of the
3230 * struct kmem_cache_node's. There is special bootstrap code in
3231 * kmem_cache_open for slab_state == DOWN.
3233 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3235 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3236 sizeof(struct kmem_cache_node),
3237 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3239 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3241 /* Able to allocate the per node structures */
3242 slab_state = PARTIAL;
3244 temp_kmem_cache = kmem_cache;
3245 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3246 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3247 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3248 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3251 * Allocate kmem_cache_node properly from the kmem_cache slab.
3252 * kmem_cache_node is separately allocated so no need to
3253 * update any list pointers.
3255 temp_kmem_cache_node = kmem_cache_node;
3257 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3258 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3260 kmem_cache_bootstrap_fixup(kmem_cache_node);
3263 kmem_cache_bootstrap_fixup(kmem_cache);
3265 /* Free temporary boot structure */
3266 free_pages((unsigned long)temp_kmem_cache, order);
3268 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3271 * Patch up the size_index table if we have strange large alignment
3272 * requirements for the kmalloc array. This is only the case for
3273 * MIPS it seems. The standard arches will not generate any code here.
3275 * Largest permitted alignment is 256 bytes due to the way we
3276 * handle the index determination for the smaller caches.
3278 * Make sure that nothing crazy happens if someone starts tinkering
3279 * around with ARCH_KMALLOC_MINALIGN
3281 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3282 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3284 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3285 int elem = size_index_elem(i);
3286 if (elem >= ARRAY_SIZE(size_index))
3288 size_index[elem] = KMALLOC_SHIFT_LOW;
3291 if (KMALLOC_MIN_SIZE == 64) {
3293 * The 96 byte size cache is not used if the alignment
3296 for (i = 64 + 8; i <= 96; i += 8)
3297 size_index[size_index_elem(i)] = 7;
3298 } else if (KMALLOC_MIN_SIZE == 128) {
3300 * The 192 byte sized cache is not used if the alignment
3301 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3304 for (i = 128 + 8; i <= 192; i += 8)
3305 size_index[size_index_elem(i)] = 8;
3308 /* Caches that are not of the two-to-the-power-of size */
3309 if (KMALLOC_MIN_SIZE <= 32) {
3310 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3314 if (KMALLOC_MIN_SIZE <= 64) {
3315 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3319 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3320 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3326 /* Provide the correct kmalloc names now that the caches are up */
3327 if (KMALLOC_MIN_SIZE <= 32) {
3328 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3329 BUG_ON(!kmalloc_caches[1]->name);
3332 if (KMALLOC_MIN_SIZE <= 64) {
3333 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3334 BUG_ON(!kmalloc_caches[2]->name);
3337 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3338 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3341 kmalloc_caches[i]->name = s;
3345 register_cpu_notifier(&slab_notifier);
3348 #ifdef CONFIG_ZONE_DMA
3349 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3350 struct kmem_cache *s = kmalloc_caches[i];
3353 char *name = kasprintf(GFP_NOWAIT,
3354 "dma-kmalloc-%d", s->objsize);
3357 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3358 s->objsize, SLAB_CACHE_DMA);
3363 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3364 " CPUs=%d, Nodes=%d\n",
3365 caches, cache_line_size(),
3366 slub_min_order, slub_max_order, slub_min_objects,
3367 nr_cpu_ids, nr_node_ids);
3370 void __init kmem_cache_init_late(void)
3375 * Find a mergeable slab cache
3377 static int slab_unmergeable(struct kmem_cache *s)
3379 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3386 * We may have set a slab to be unmergeable during bootstrap.
3388 if (s->refcount < 0)
3394 static struct kmem_cache *find_mergeable(size_t size,
3395 size_t align, unsigned long flags, const char *name,
3396 void (*ctor)(void *))
3398 struct kmem_cache *s;
3400 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3406 size = ALIGN(size, sizeof(void *));
3407 align = calculate_alignment(flags, align, size);
3408 size = ALIGN(size, align);
3409 flags = kmem_cache_flags(size, flags, name, NULL);
3411 list_for_each_entry(s, &slab_caches, list) {
3412 if (slab_unmergeable(s))
3418 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3421 * Check if alignment is compatible.
3422 * Courtesy of Adrian Drzewiecki
3424 if ((s->size & ~(align - 1)) != s->size)
3427 if (s->size - size >= sizeof(void *))
3435 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3436 size_t align, unsigned long flags, void (*ctor)(void *))
3438 struct kmem_cache *s;
3444 down_write(&slub_lock);
3445 s = find_mergeable(size, align, flags, name, ctor);
3449 * Adjust the object sizes so that we clear
3450 * the complete object on kzalloc.
3452 s->objsize = max(s->objsize, (int)size);
3453 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3455 if (sysfs_slab_alias(s, name)) {
3459 up_write(&slub_lock);
3463 n = kstrdup(name, GFP_KERNEL);
3467 s = kmalloc(kmem_size, GFP_KERNEL);
3469 if (kmem_cache_open(s, n,
3470 size, align, flags, ctor)) {
3471 list_add(&s->list, &slab_caches);
3472 if (sysfs_slab_add(s)) {
3478 up_write(&slub_lock);
3485 up_write(&slub_lock);
3487 if (flags & SLAB_PANIC)
3488 panic("Cannot create slabcache %s\n", name);
3493 EXPORT_SYMBOL(kmem_cache_create);
3497 * Use the cpu notifier to insure that the cpu slabs are flushed when
3500 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3501 unsigned long action, void *hcpu)
3503 long cpu = (long)hcpu;
3504 struct kmem_cache *s;
3505 unsigned long flags;
3508 case CPU_UP_CANCELED:
3509 case CPU_UP_CANCELED_FROZEN:
3511 case CPU_DEAD_FROZEN:
3512 down_read(&slub_lock);
3513 list_for_each_entry(s, &slab_caches, list) {
3514 local_irq_save(flags);
3515 __flush_cpu_slab(s, cpu);
3516 local_irq_restore(flags);
3518 up_read(&slub_lock);
3526 static struct notifier_block __cpuinitdata slab_notifier = {
3527 .notifier_call = slab_cpuup_callback
3532 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3534 struct kmem_cache *s;
3537 if (unlikely(size > SLUB_MAX_SIZE))
3538 return kmalloc_large(size, gfpflags);
3540 s = get_slab(size, gfpflags);
3542 if (unlikely(ZERO_OR_NULL_PTR(s)))
3545 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3547 /* Honor the call site pointer we received. */
3548 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3554 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3555 int node, unsigned long caller)
3557 struct kmem_cache *s;
3560 if (unlikely(size > SLUB_MAX_SIZE)) {
3561 ret = kmalloc_large_node(size, gfpflags, node);
3563 trace_kmalloc_node(caller, ret,
3564 size, PAGE_SIZE << get_order(size),
3570 s = get_slab(size, gfpflags);
3572 if (unlikely(ZERO_OR_NULL_PTR(s)))
3575 ret = slab_alloc(s, gfpflags, node, caller);
3577 /* Honor the call site pointer we received. */
3578 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3585 static int count_inuse(struct page *page)
3590 static int count_total(struct page *page)
3592 return page->objects;
3596 #ifdef CONFIG_SLUB_DEBUG
3597 static int validate_slab(struct kmem_cache *s, struct page *page,
3601 void *addr = page_address(page);
3603 if (!check_slab(s, page) ||
3604 !on_freelist(s, page, NULL))
3607 /* Now we know that a valid freelist exists */
3608 bitmap_zero(map, page->objects);
3610 get_map(s, page, map);
3611 for_each_object(p, s, addr, page->objects) {
3612 if (test_bit(slab_index(p, s, addr), map))
3613 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3617 for_each_object(p, s, addr, page->objects)
3618 if (!test_bit(slab_index(p, s, addr), map))
3619 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3624 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3627 if (slab_trylock(page)) {
3628 validate_slab(s, page, map);
3631 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3635 static int validate_slab_node(struct kmem_cache *s,
3636 struct kmem_cache_node *n, unsigned long *map)
3638 unsigned long count = 0;
3640 unsigned long flags;
3642 spin_lock_irqsave(&n->list_lock, flags);
3644 list_for_each_entry(page, &n->partial, lru) {
3645 validate_slab_slab(s, page, map);
3648 if (count != n->nr_partial)
3649 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3650 "counter=%ld\n", s->name, count, n->nr_partial);
3652 if (!(s->flags & SLAB_STORE_USER))
3655 list_for_each_entry(page, &n->full, lru) {
3656 validate_slab_slab(s, page, map);
3659 if (count != atomic_long_read(&n->nr_slabs))
3660 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3661 "counter=%ld\n", s->name, count,
3662 atomic_long_read(&n->nr_slabs));
3665 spin_unlock_irqrestore(&n->list_lock, flags);
3669 static long validate_slab_cache(struct kmem_cache *s)
3672 unsigned long count = 0;
3673 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3674 sizeof(unsigned long), GFP_KERNEL);
3680 for_each_node_state(node, N_NORMAL_MEMORY) {
3681 struct kmem_cache_node *n = get_node(s, node);
3683 count += validate_slab_node(s, n, map);
3689 * Generate lists of code addresses where slabcache objects are allocated
3694 unsigned long count;
3701 DECLARE_BITMAP(cpus, NR_CPUS);
3707 unsigned long count;
3708 struct location *loc;
3711 static void free_loc_track(struct loc_track *t)
3714 free_pages((unsigned long)t->loc,
3715 get_order(sizeof(struct location) * t->max));
3718 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3723 order = get_order(sizeof(struct location) * max);
3725 l = (void *)__get_free_pages(flags, order);
3730 memcpy(l, t->loc, sizeof(struct location) * t->count);
3738 static int add_location(struct loc_track *t, struct kmem_cache *s,
3739 const struct track *track)
3741 long start, end, pos;
3743 unsigned long caddr;
3744 unsigned long age = jiffies - track->when;
3750 pos = start + (end - start + 1) / 2;
3753 * There is nothing at "end". If we end up there
3754 * we need to add something to before end.
3759 caddr = t->loc[pos].addr;
3760 if (track->addr == caddr) {
3766 if (age < l->min_time)
3768 if (age > l->max_time)
3771 if (track->pid < l->min_pid)
3772 l->min_pid = track->pid;
3773 if (track->pid > l->max_pid)
3774 l->max_pid = track->pid;
3776 cpumask_set_cpu(track->cpu,
3777 to_cpumask(l->cpus));
3779 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3783 if (track->addr < caddr)
3790 * Not found. Insert new tracking element.
3792 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3798 (t->count - pos) * sizeof(struct location));
3801 l->addr = track->addr;
3805 l->min_pid = track->pid;
3806 l->max_pid = track->pid;
3807 cpumask_clear(to_cpumask(l->cpus));
3808 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3809 nodes_clear(l->nodes);
3810 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3814 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3815 struct page *page, enum track_item alloc,
3818 void *addr = page_address(page);
3821 bitmap_zero(map, page->objects);
3822 get_map(s, page, map);
3824 for_each_object(p, s, addr, page->objects)
3825 if (!test_bit(slab_index(p, s, addr), map))
3826 add_location(t, s, get_track(s, p, alloc));
3829 static int list_locations(struct kmem_cache *s, char *buf,
3830 enum track_item alloc)
3834 struct loc_track t = { 0, 0, NULL };
3836 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3837 sizeof(unsigned long), GFP_KERNEL);
3839 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3842 return sprintf(buf, "Out of memory\n");
3844 /* Push back cpu slabs */
3847 for_each_node_state(node, N_NORMAL_MEMORY) {
3848 struct kmem_cache_node *n = get_node(s, node);
3849 unsigned long flags;
3852 if (!atomic_long_read(&n->nr_slabs))
3855 spin_lock_irqsave(&n->list_lock, flags);
3856 list_for_each_entry(page, &n->partial, lru)
3857 process_slab(&t, s, page, alloc, map);
3858 list_for_each_entry(page, &n->full, lru)
3859 process_slab(&t, s, page, alloc, map);
3860 spin_unlock_irqrestore(&n->list_lock, flags);
3863 for (i = 0; i < t.count; i++) {
3864 struct location *l = &t.loc[i];
3866 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3868 len += sprintf(buf + len, "%7ld ", l->count);
3871 len += sprintf(buf + len, "%pS", (void *)l->addr);
3873 len += sprintf(buf + len, "<not-available>");
3875 if (l->sum_time != l->min_time) {
3876 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3878 (long)div_u64(l->sum_time, l->count),
3881 len += sprintf(buf + len, " age=%ld",
3884 if (l->min_pid != l->max_pid)
3885 len += sprintf(buf + len, " pid=%ld-%ld",
3886 l->min_pid, l->max_pid);
3888 len += sprintf(buf + len, " pid=%ld",
3891 if (num_online_cpus() > 1 &&
3892 !cpumask_empty(to_cpumask(l->cpus)) &&
3893 len < PAGE_SIZE - 60) {
3894 len += sprintf(buf + len, " cpus=");
3895 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3896 to_cpumask(l->cpus));
3899 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3900 len < PAGE_SIZE - 60) {
3901 len += sprintf(buf + len, " nodes=");
3902 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3906 len += sprintf(buf + len, "\n");
3912 len += sprintf(buf, "No data\n");
3917 #ifdef SLUB_RESILIENCY_TEST
3918 static void resiliency_test(void)
3922 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3924 printk(KERN_ERR "SLUB resiliency testing\n");
3925 printk(KERN_ERR "-----------------------\n");
3926 printk(KERN_ERR "A. Corruption after allocation\n");
3928 p = kzalloc(16, GFP_KERNEL);
3930 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3931 " 0x12->0x%p\n\n", p + 16);
3933 validate_slab_cache(kmalloc_caches[4]);
3935 /* Hmmm... The next two are dangerous */
3936 p = kzalloc(32, GFP_KERNEL);
3937 p[32 + sizeof(void *)] = 0x34;
3938 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3939 " 0x34 -> -0x%p\n", p);
3941 "If allocated object is overwritten then not detectable\n\n");
3943 validate_slab_cache(kmalloc_caches[5]);
3944 p = kzalloc(64, GFP_KERNEL);
3945 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3947 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3950 "If allocated object is overwritten then not detectable\n\n");
3951 validate_slab_cache(kmalloc_caches[6]);
3953 printk(KERN_ERR "\nB. Corruption after free\n");
3954 p = kzalloc(128, GFP_KERNEL);
3957 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3958 validate_slab_cache(kmalloc_caches[7]);
3960 p = kzalloc(256, GFP_KERNEL);
3963 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3965 validate_slab_cache(kmalloc_caches[8]);
3967 p = kzalloc(512, GFP_KERNEL);
3970 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3971 validate_slab_cache(kmalloc_caches[9]);
3975 static void resiliency_test(void) {};
3980 enum slab_stat_type {
3981 SL_ALL, /* All slabs */
3982 SL_PARTIAL, /* Only partially allocated slabs */
3983 SL_CPU, /* Only slabs used for cpu caches */
3984 SL_OBJECTS, /* Determine allocated objects not slabs */
3985 SL_TOTAL /* Determine object capacity not slabs */
3988 #define SO_ALL (1 << SL_ALL)
3989 #define SO_PARTIAL (1 << SL_PARTIAL)
3990 #define SO_CPU (1 << SL_CPU)
3991 #define SO_OBJECTS (1 << SL_OBJECTS)
3992 #define SO_TOTAL (1 << SL_TOTAL)
3994 static ssize_t show_slab_objects(struct kmem_cache *s,
3995 char *buf, unsigned long flags)
3997 unsigned long total = 0;
4000 unsigned long *nodes;
4001 unsigned long *per_cpu;
4003 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4006 per_cpu = nodes + nr_node_ids;
4008 if (flags & SO_CPU) {
4011 for_each_possible_cpu(cpu) {
4012 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4014 if (!c || c->node < 0)
4018 if (flags & SO_TOTAL)
4019 x = c->page->objects;
4020 else if (flags & SO_OBJECTS)
4026 nodes[c->node] += x;
4032 lock_memory_hotplug();
4033 #ifdef CONFIG_SLUB_DEBUG
4034 if (flags & SO_ALL) {
4035 for_each_node_state(node, N_NORMAL_MEMORY) {
4036 struct kmem_cache_node *n = get_node(s, node);
4038 if (flags & SO_TOTAL)
4039 x = atomic_long_read(&n->total_objects);
4040 else if (flags & SO_OBJECTS)
4041 x = atomic_long_read(&n->total_objects) -
4042 count_partial(n, count_free);
4045 x = atomic_long_read(&n->nr_slabs);
4052 if (flags & SO_PARTIAL) {
4053 for_each_node_state(node, N_NORMAL_MEMORY) {
4054 struct kmem_cache_node *n = get_node(s, node);
4056 if (flags & SO_TOTAL)
4057 x = count_partial(n, count_total);
4058 else if (flags & SO_OBJECTS)
4059 x = count_partial(n, count_inuse);
4066 x = sprintf(buf, "%lu", total);
4068 for_each_node_state(node, N_NORMAL_MEMORY)
4070 x += sprintf(buf + x, " N%d=%lu",
4073 unlock_memory_hotplug();
4075 return x + sprintf(buf + x, "\n");
4078 #ifdef CONFIG_SLUB_DEBUG
4079 static int any_slab_objects(struct kmem_cache *s)
4083 for_each_online_node(node) {
4084 struct kmem_cache_node *n = get_node(s, node);
4089 if (atomic_long_read(&n->total_objects))
4096 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4097 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4099 struct slab_attribute {
4100 struct attribute attr;
4101 ssize_t (*show)(struct kmem_cache *s, char *buf);
4102 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4105 #define SLAB_ATTR_RO(_name) \
4106 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4108 #define SLAB_ATTR(_name) \
4109 static struct slab_attribute _name##_attr = \
4110 __ATTR(_name, 0644, _name##_show, _name##_store)
4112 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4114 return sprintf(buf, "%d\n", s->size);
4116 SLAB_ATTR_RO(slab_size);
4118 static ssize_t align_show(struct kmem_cache *s, char *buf)
4120 return sprintf(buf, "%d\n", s->align);
4122 SLAB_ATTR_RO(align);
4124 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4126 return sprintf(buf, "%d\n", s->objsize);
4128 SLAB_ATTR_RO(object_size);
4130 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4132 return sprintf(buf, "%d\n", oo_objects(s->oo));
4134 SLAB_ATTR_RO(objs_per_slab);
4136 static ssize_t order_store(struct kmem_cache *s,
4137 const char *buf, size_t length)
4139 unsigned long order;
4142 err = strict_strtoul(buf, 10, &order);
4146 if (order > slub_max_order || order < slub_min_order)
4149 calculate_sizes(s, order);
4153 static ssize_t order_show(struct kmem_cache *s, char *buf)
4155 return sprintf(buf, "%d\n", oo_order(s->oo));
4159 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4161 return sprintf(buf, "%lu\n", s->min_partial);
4164 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4170 err = strict_strtoul(buf, 10, &min);
4174 set_min_partial(s, min);
4177 SLAB_ATTR(min_partial);
4179 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4183 return sprintf(buf, "%pS\n", s->ctor);
4187 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4189 return sprintf(buf, "%d\n", s->refcount - 1);
4191 SLAB_ATTR_RO(aliases);
4193 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4195 return show_slab_objects(s, buf, SO_PARTIAL);
4197 SLAB_ATTR_RO(partial);
4199 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4201 return show_slab_objects(s, buf, SO_CPU);
4203 SLAB_ATTR_RO(cpu_slabs);
4205 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4207 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4209 SLAB_ATTR_RO(objects);
4211 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4213 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4215 SLAB_ATTR_RO(objects_partial);
4217 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4219 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4222 static ssize_t reclaim_account_store(struct kmem_cache *s,
4223 const char *buf, size_t length)
4225 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4227 s->flags |= SLAB_RECLAIM_ACCOUNT;
4230 SLAB_ATTR(reclaim_account);
4232 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4234 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4236 SLAB_ATTR_RO(hwcache_align);
4238 #ifdef CONFIG_ZONE_DMA
4239 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4241 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4243 SLAB_ATTR_RO(cache_dma);
4246 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4248 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4250 SLAB_ATTR_RO(destroy_by_rcu);
4252 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4254 return sprintf(buf, "%d\n", s->reserved);
4256 SLAB_ATTR_RO(reserved);
4258 #ifdef CONFIG_SLUB_DEBUG
4259 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4261 return show_slab_objects(s, buf, SO_ALL);
4263 SLAB_ATTR_RO(slabs);
4265 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4267 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4269 SLAB_ATTR_RO(total_objects);
4271 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4273 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4276 static ssize_t sanity_checks_store(struct kmem_cache *s,
4277 const char *buf, size_t length)
4279 s->flags &= ~SLAB_DEBUG_FREE;
4281 s->flags |= SLAB_DEBUG_FREE;
4284 SLAB_ATTR(sanity_checks);
4286 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4288 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4291 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4294 s->flags &= ~SLAB_TRACE;
4296 s->flags |= SLAB_TRACE;
4301 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4303 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4306 static ssize_t red_zone_store(struct kmem_cache *s,
4307 const char *buf, size_t length)
4309 if (any_slab_objects(s))
4312 s->flags &= ~SLAB_RED_ZONE;
4314 s->flags |= SLAB_RED_ZONE;
4315 calculate_sizes(s, -1);
4318 SLAB_ATTR(red_zone);
4320 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4322 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4325 static ssize_t poison_store(struct kmem_cache *s,
4326 const char *buf, size_t length)
4328 if (any_slab_objects(s))
4331 s->flags &= ~SLAB_POISON;
4333 s->flags |= SLAB_POISON;
4334 calculate_sizes(s, -1);
4339 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4341 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4344 static ssize_t store_user_store(struct kmem_cache *s,
4345 const char *buf, size_t length)
4347 if (any_slab_objects(s))
4350 s->flags &= ~SLAB_STORE_USER;
4352 s->flags |= SLAB_STORE_USER;
4353 calculate_sizes(s, -1);
4356 SLAB_ATTR(store_user);
4358 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4363 static ssize_t validate_store(struct kmem_cache *s,
4364 const char *buf, size_t length)
4368 if (buf[0] == '1') {
4369 ret = validate_slab_cache(s);
4375 SLAB_ATTR(validate);
4377 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4379 if (!(s->flags & SLAB_STORE_USER))
4381 return list_locations(s, buf, TRACK_ALLOC);
4383 SLAB_ATTR_RO(alloc_calls);
4385 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4387 if (!(s->flags & SLAB_STORE_USER))
4389 return list_locations(s, buf, TRACK_FREE);
4391 SLAB_ATTR_RO(free_calls);
4392 #endif /* CONFIG_SLUB_DEBUG */
4394 #ifdef CONFIG_FAILSLAB
4395 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4397 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4400 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4403 s->flags &= ~SLAB_FAILSLAB;
4405 s->flags |= SLAB_FAILSLAB;
4408 SLAB_ATTR(failslab);
4411 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4416 static ssize_t shrink_store(struct kmem_cache *s,
4417 const char *buf, size_t length)
4419 if (buf[0] == '1') {
4420 int rc = kmem_cache_shrink(s);
4431 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4433 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4436 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4437 const char *buf, size_t length)
4439 unsigned long ratio;
4442 err = strict_strtoul(buf, 10, &ratio);
4447 s->remote_node_defrag_ratio = ratio * 10;
4451 SLAB_ATTR(remote_node_defrag_ratio);
4454 #ifdef CONFIG_SLUB_STATS
4455 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4457 unsigned long sum = 0;
4460 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4465 for_each_online_cpu(cpu) {
4466 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4472 len = sprintf(buf, "%lu", sum);
4475 for_each_online_cpu(cpu) {
4476 if (data[cpu] && len < PAGE_SIZE - 20)
4477 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4481 return len + sprintf(buf + len, "\n");
4484 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4488 for_each_online_cpu(cpu)
4489 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4492 #define STAT_ATTR(si, text) \
4493 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4495 return show_stat(s, buf, si); \
4497 static ssize_t text##_store(struct kmem_cache *s, \
4498 const char *buf, size_t length) \
4500 if (buf[0] != '0') \
4502 clear_stat(s, si); \
4507 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4508 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4509 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4510 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4511 STAT_ATTR(FREE_FROZEN, free_frozen);
4512 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4513 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4514 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4515 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4516 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4517 STAT_ATTR(FREE_SLAB, free_slab);
4518 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4519 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4520 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4521 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4522 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4523 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4524 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4527 static struct attribute *slab_attrs[] = {
4528 &slab_size_attr.attr,
4529 &object_size_attr.attr,
4530 &objs_per_slab_attr.attr,
4532 &min_partial_attr.attr,
4534 &objects_partial_attr.attr,
4536 &cpu_slabs_attr.attr,
4540 &hwcache_align_attr.attr,
4541 &reclaim_account_attr.attr,
4542 &destroy_by_rcu_attr.attr,
4544 &reserved_attr.attr,
4545 #ifdef CONFIG_SLUB_DEBUG
4546 &total_objects_attr.attr,
4548 &sanity_checks_attr.attr,
4550 &red_zone_attr.attr,
4552 &store_user_attr.attr,
4553 &validate_attr.attr,
4554 &alloc_calls_attr.attr,
4555 &free_calls_attr.attr,
4557 #ifdef CONFIG_ZONE_DMA
4558 &cache_dma_attr.attr,
4561 &remote_node_defrag_ratio_attr.attr,
4563 #ifdef CONFIG_SLUB_STATS
4564 &alloc_fastpath_attr.attr,
4565 &alloc_slowpath_attr.attr,
4566 &free_fastpath_attr.attr,
4567 &free_slowpath_attr.attr,
4568 &free_frozen_attr.attr,
4569 &free_add_partial_attr.attr,
4570 &free_remove_partial_attr.attr,
4571 &alloc_from_partial_attr.attr,
4572 &alloc_slab_attr.attr,
4573 &alloc_refill_attr.attr,
4574 &free_slab_attr.attr,
4575 &cpuslab_flush_attr.attr,
4576 &deactivate_full_attr.attr,
4577 &deactivate_empty_attr.attr,
4578 &deactivate_to_head_attr.attr,
4579 &deactivate_to_tail_attr.attr,
4580 &deactivate_remote_frees_attr.attr,
4581 &order_fallback_attr.attr,
4583 #ifdef CONFIG_FAILSLAB
4584 &failslab_attr.attr,
4590 static struct attribute_group slab_attr_group = {
4591 .attrs = slab_attrs,
4594 static ssize_t slab_attr_show(struct kobject *kobj,
4595 struct attribute *attr,
4598 struct slab_attribute *attribute;
4599 struct kmem_cache *s;
4602 attribute = to_slab_attr(attr);
4605 if (!attribute->show)
4608 err = attribute->show(s, buf);
4613 static ssize_t slab_attr_store(struct kobject *kobj,
4614 struct attribute *attr,
4615 const char *buf, size_t len)
4617 struct slab_attribute *attribute;
4618 struct kmem_cache *s;
4621 attribute = to_slab_attr(attr);
4624 if (!attribute->store)
4627 err = attribute->store(s, buf, len);
4632 static void kmem_cache_release(struct kobject *kobj)
4634 struct kmem_cache *s = to_slab(kobj);
4640 static const struct sysfs_ops slab_sysfs_ops = {
4641 .show = slab_attr_show,
4642 .store = slab_attr_store,
4645 static struct kobj_type slab_ktype = {
4646 .sysfs_ops = &slab_sysfs_ops,
4647 .release = kmem_cache_release
4650 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4652 struct kobj_type *ktype = get_ktype(kobj);
4654 if (ktype == &slab_ktype)
4659 static const struct kset_uevent_ops slab_uevent_ops = {
4660 .filter = uevent_filter,
4663 static struct kset *slab_kset;
4665 #define ID_STR_LENGTH 64
4667 /* Create a unique string id for a slab cache:
4669 * Format :[flags-]size
4671 static char *create_unique_id(struct kmem_cache *s)
4673 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4680 * First flags affecting slabcache operations. We will only
4681 * get here for aliasable slabs so we do not need to support
4682 * too many flags. The flags here must cover all flags that
4683 * are matched during merging to guarantee that the id is
4686 if (s->flags & SLAB_CACHE_DMA)
4688 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4690 if (s->flags & SLAB_DEBUG_FREE)
4692 if (!(s->flags & SLAB_NOTRACK))
4696 p += sprintf(p, "%07d", s->size);
4697 BUG_ON(p > name + ID_STR_LENGTH - 1);
4701 static int sysfs_slab_add(struct kmem_cache *s)
4707 if (slab_state < SYSFS)
4708 /* Defer until later */
4711 unmergeable = slab_unmergeable(s);
4714 * Slabcache can never be merged so we can use the name proper.
4715 * This is typically the case for debug situations. In that
4716 * case we can catch duplicate names easily.
4718 sysfs_remove_link(&slab_kset->kobj, s->name);
4722 * Create a unique name for the slab as a target
4725 name = create_unique_id(s);
4728 s->kobj.kset = slab_kset;
4729 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4731 kobject_put(&s->kobj);
4735 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4737 kobject_del(&s->kobj);
4738 kobject_put(&s->kobj);
4741 kobject_uevent(&s->kobj, KOBJ_ADD);
4743 /* Setup first alias */
4744 sysfs_slab_alias(s, s->name);
4750 static void sysfs_slab_remove(struct kmem_cache *s)
4752 if (slab_state < SYSFS)
4754 * Sysfs has not been setup yet so no need to remove the
4759 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4760 kobject_del(&s->kobj);
4761 kobject_put(&s->kobj);
4765 * Need to buffer aliases during bootup until sysfs becomes
4766 * available lest we lose that information.
4768 struct saved_alias {
4769 struct kmem_cache *s;
4771 struct saved_alias *next;
4774 static struct saved_alias *alias_list;
4776 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4778 struct saved_alias *al;
4780 if (slab_state == SYSFS) {
4782 * If we have a leftover link then remove it.
4784 sysfs_remove_link(&slab_kset->kobj, name);
4785 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4788 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4794 al->next = alias_list;
4799 static int __init slab_sysfs_init(void)
4801 struct kmem_cache *s;
4804 down_write(&slub_lock);
4806 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4808 up_write(&slub_lock);
4809 printk(KERN_ERR "Cannot register slab subsystem.\n");
4815 list_for_each_entry(s, &slab_caches, list) {
4816 err = sysfs_slab_add(s);
4818 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4819 " to sysfs\n", s->name);
4822 while (alias_list) {
4823 struct saved_alias *al = alias_list;
4825 alias_list = alias_list->next;
4826 err = sysfs_slab_alias(al->s, al->name);
4828 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4829 " %s to sysfs\n", s->name);
4833 up_write(&slub_lock);
4838 __initcall(slab_sysfs_init);
4839 #endif /* CONFIG_SYSFS */
4842 * The /proc/slabinfo ABI
4844 #ifdef CONFIG_SLABINFO
4845 static void print_slabinfo_header(struct seq_file *m)
4847 seq_puts(m, "slabinfo - version: 2.1\n");
4848 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4849 "<objperslab> <pagesperslab>");
4850 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4851 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4855 static void *s_start(struct seq_file *m, loff_t *pos)
4859 down_read(&slub_lock);
4861 print_slabinfo_header(m);
4863 return seq_list_start(&slab_caches, *pos);
4866 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4868 return seq_list_next(p, &slab_caches, pos);
4871 static void s_stop(struct seq_file *m, void *p)
4873 up_read(&slub_lock);
4876 static int s_show(struct seq_file *m, void *p)
4878 unsigned long nr_partials = 0;
4879 unsigned long nr_slabs = 0;
4880 unsigned long nr_inuse = 0;
4881 unsigned long nr_objs = 0;
4882 unsigned long nr_free = 0;
4883 struct kmem_cache *s;
4886 s = list_entry(p, struct kmem_cache, list);
4888 for_each_online_node(node) {
4889 struct kmem_cache_node *n = get_node(s, node);
4894 nr_partials += n->nr_partial;
4895 nr_slabs += atomic_long_read(&n->nr_slabs);
4896 nr_objs += atomic_long_read(&n->total_objects);
4897 nr_free += count_partial(n, count_free);
4900 nr_inuse = nr_objs - nr_free;
4902 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4903 nr_objs, s->size, oo_objects(s->oo),
4904 (1 << oo_order(s->oo)));
4905 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4906 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4912 static const struct seq_operations slabinfo_op = {
4919 static int slabinfo_open(struct inode *inode, struct file *file)
4921 return seq_open(file, &slabinfo_op);
4924 static const struct file_operations proc_slabinfo_operations = {
4925 .open = slabinfo_open,
4927 .llseek = seq_lseek,
4928 .release = seq_release,
4931 static int __init slab_proc_init(void)
4933 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4936 module_init(slab_proc_init);
4937 #endif /* CONFIG_SLABINFO */