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
134 /* Enable to log cmpxchg failures */
135 #undef SLUB_DEBUG_CMPXCHG
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
141 #define MIN_PARTIAL 5
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
148 #define MAX_PARTIAL 10
150 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
158 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
161 * Set of flags that will prevent slab merging
163 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
167 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
171 #define OO_MASK ((1 << OO_SHIFT) - 1)
172 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
174 /* Internal SLUB flags */
175 #define __OBJECT_POISON 0x80000000UL /* Poison object */
176 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
178 static int kmem_size = sizeof(struct kmem_cache);
181 static struct notifier_block slab_notifier;
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
191 /* A list of all slab caches on the system */
192 static DECLARE_RWSEM(slub_lock);
193 static LIST_HEAD(slab_caches);
196 * Tracking user of a slab.
199 unsigned long addr; /* 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 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
346 void *freelist_old, unsigned long counters_old,
347 void *freelist_new, unsigned long counters_new,
350 #ifdef CONFIG_CMPXCHG_DOUBLE
351 if (s->flags & __CMPXCHG_DOUBLE) {
352 if (cmpxchg_double(&page->freelist,
353 freelist_old, counters_old,
354 freelist_new, counters_new))
359 if (page->freelist == freelist_old && page->counters == counters_old) {
360 page->freelist = freelist_new;
361 page->counters = counters_new;
367 stat(s, CMPXCHG_DOUBLE_FAIL);
369 #ifdef SLUB_DEBUG_CMPXCHG
370 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
376 #ifdef CONFIG_SLUB_DEBUG
378 * Determine a map of object in use on a page.
380 * Slab lock or node listlock must be held to guarantee that the page does
381 * not vanish from under us.
383 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
386 void *addr = page_address(page);
388 for (p = page->freelist; p; p = get_freepointer(s, p))
389 set_bit(slab_index(p, s, addr), map);
395 #ifdef CONFIG_SLUB_DEBUG_ON
396 static int slub_debug = DEBUG_DEFAULT_FLAGS;
398 static int slub_debug;
401 static char *slub_debug_slabs;
402 static int disable_higher_order_debug;
407 static void print_section(char *text, u8 *addr, unsigned int length)
415 for (i = 0; i < length; i++) {
417 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
420 printk(KERN_CONT " %02x", addr[i]);
422 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
424 printk(KERN_CONT " %s\n", ascii);
431 printk(KERN_CONT " ");
435 printk(KERN_CONT " %s\n", ascii);
439 static struct track *get_track(struct kmem_cache *s, void *object,
440 enum track_item alloc)
445 p = object + s->offset + sizeof(void *);
447 p = object + s->inuse;
452 static void set_track(struct kmem_cache *s, void *object,
453 enum track_item alloc, unsigned long addr)
455 struct track *p = get_track(s, object, alloc);
459 p->cpu = smp_processor_id();
460 p->pid = current->pid;
463 memset(p, 0, sizeof(struct track));
466 static void init_tracking(struct kmem_cache *s, void *object)
468 if (!(s->flags & SLAB_STORE_USER))
471 set_track(s, object, TRACK_FREE, 0UL);
472 set_track(s, object, TRACK_ALLOC, 0UL);
475 static void print_track(const char *s, struct track *t)
480 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
481 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
484 static void print_tracking(struct kmem_cache *s, void *object)
486 if (!(s->flags & SLAB_STORE_USER))
489 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
490 print_track("Freed", get_track(s, object, TRACK_FREE));
493 static void print_page_info(struct page *page)
495 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
496 page, page->objects, page->inuse, page->freelist, page->flags);
500 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
506 vsnprintf(buf, sizeof(buf), fmt, args);
508 printk(KERN_ERR "========================================"
509 "=====================================\n");
510 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
511 printk(KERN_ERR "----------------------------------------"
512 "-------------------------------------\n\n");
515 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
521 vsnprintf(buf, sizeof(buf), fmt, args);
523 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
526 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
528 unsigned int off; /* Offset of last byte */
529 u8 *addr = page_address(page);
531 print_tracking(s, p);
533 print_page_info(page);
535 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
536 p, p - addr, get_freepointer(s, p));
539 print_section("Bytes b4", p - 16, 16);
541 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
543 if (s->flags & SLAB_RED_ZONE)
544 print_section("Redzone", p + s->objsize,
545 s->inuse - s->objsize);
548 off = s->offset + sizeof(void *);
552 if (s->flags & SLAB_STORE_USER)
553 off += 2 * sizeof(struct track);
556 /* Beginning of the filler is the free pointer */
557 print_section("Padding", p + off, s->size - off);
562 static void object_err(struct kmem_cache *s, struct page *page,
563 u8 *object, char *reason)
565 slab_bug(s, "%s", reason);
566 print_trailer(s, page, object);
569 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
575 vsnprintf(buf, sizeof(buf), fmt, args);
577 slab_bug(s, "%s", buf);
578 print_page_info(page);
582 static void init_object(struct kmem_cache *s, void *object, u8 val)
586 if (s->flags & __OBJECT_POISON) {
587 memset(p, POISON_FREE, s->objsize - 1);
588 p[s->objsize - 1] = POISON_END;
591 if (s->flags & SLAB_RED_ZONE)
592 memset(p + s->objsize, val, s->inuse - s->objsize);
595 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
598 if (*start != (u8)value)
606 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
607 void *from, void *to)
609 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
610 memset(from, data, to - from);
613 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
614 u8 *object, char *what,
615 u8 *start, unsigned int value, unsigned int bytes)
620 fault = check_bytes(start, value, bytes);
625 while (end > fault && end[-1] == value)
628 slab_bug(s, "%s overwritten", what);
629 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
630 fault, end - 1, fault[0], value);
631 print_trailer(s, page, object);
633 restore_bytes(s, what, value, fault, end);
641 * Bytes of the object to be managed.
642 * If the freepointer may overlay the object then the free
643 * pointer is the first word of the object.
645 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
648 * object + s->objsize
649 * Padding to reach word boundary. This is also used for Redzoning.
650 * Padding is extended by another word if Redzoning is enabled and
653 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
654 * 0xcc (RED_ACTIVE) for objects in use.
657 * Meta data starts here.
659 * A. Free pointer (if we cannot overwrite object on free)
660 * B. Tracking data for SLAB_STORE_USER
661 * C. Padding to reach required alignment boundary or at mininum
662 * one word if debugging is on to be able to detect writes
663 * before the word boundary.
665 * Padding is done using 0x5a (POISON_INUSE)
668 * Nothing is used beyond s->size.
670 * If slabcaches are merged then the objsize and inuse boundaries are mostly
671 * ignored. And therefore no slab options that rely on these boundaries
672 * may be used with merged slabcaches.
675 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
677 unsigned long off = s->inuse; /* The end of info */
680 /* Freepointer is placed after the object. */
681 off += sizeof(void *);
683 if (s->flags & SLAB_STORE_USER)
684 /* We also have user information there */
685 off += 2 * sizeof(struct track);
690 return check_bytes_and_report(s, page, p, "Object padding",
691 p + off, POISON_INUSE, s->size - off);
694 /* Check the pad bytes at the end of a slab page */
695 static int slab_pad_check(struct kmem_cache *s, struct page *page)
703 if (!(s->flags & SLAB_POISON))
706 start = page_address(page);
707 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
708 end = start + length;
709 remainder = length % s->size;
713 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
716 while (end > fault && end[-1] == POISON_INUSE)
719 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
720 print_section("Padding", end - remainder, remainder);
722 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
726 static int check_object(struct kmem_cache *s, struct page *page,
727 void *object, u8 val)
730 u8 *endobject = object + s->objsize;
732 if (s->flags & SLAB_RED_ZONE) {
733 if (!check_bytes_and_report(s, page, object, "Redzone",
734 endobject, val, s->inuse - s->objsize))
737 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
738 check_bytes_and_report(s, page, p, "Alignment padding",
739 endobject, POISON_INUSE, s->inuse - s->objsize);
743 if (s->flags & SLAB_POISON) {
744 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
745 (!check_bytes_and_report(s, page, p, "Poison", p,
746 POISON_FREE, s->objsize - 1) ||
747 !check_bytes_and_report(s, page, p, "Poison",
748 p + s->objsize - 1, POISON_END, 1)))
751 * check_pad_bytes cleans up on its own.
753 check_pad_bytes(s, page, p);
756 if (!s->offset && val == SLUB_RED_ACTIVE)
758 * Object and freepointer overlap. Cannot check
759 * freepointer while object is allocated.
763 /* Check free pointer validity */
764 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
765 object_err(s, page, p, "Freepointer corrupt");
767 * No choice but to zap it and thus lose the remainder
768 * of the free objects in this slab. May cause
769 * another error because the object count is now wrong.
771 set_freepointer(s, p, NULL);
777 static int check_slab(struct kmem_cache *s, struct page *page)
781 VM_BUG_ON(!irqs_disabled());
783 if (!PageSlab(page)) {
784 slab_err(s, page, "Not a valid slab page");
788 maxobj = order_objects(compound_order(page), s->size, s->reserved);
789 if (page->objects > maxobj) {
790 slab_err(s, page, "objects %u > max %u",
791 s->name, page->objects, maxobj);
794 if (page->inuse > page->objects) {
795 slab_err(s, page, "inuse %u > max %u",
796 s->name, page->inuse, page->objects);
799 /* Slab_pad_check fixes things up after itself */
800 slab_pad_check(s, page);
805 * Determine if a certain object on a page is on the freelist. Must hold the
806 * slab lock to guarantee that the chains are in a consistent state.
808 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
811 void *fp = page->freelist;
813 unsigned long max_objects;
815 while (fp && nr <= page->objects) {
818 if (!check_valid_pointer(s, page, fp)) {
820 object_err(s, page, object,
821 "Freechain corrupt");
822 set_freepointer(s, object, NULL);
825 slab_err(s, page, "Freepointer corrupt");
826 page->freelist = NULL;
827 page->inuse = page->objects;
828 slab_fix(s, "Freelist cleared");
834 fp = get_freepointer(s, object);
838 max_objects = order_objects(compound_order(page), s->size, s->reserved);
839 if (max_objects > MAX_OBJS_PER_PAGE)
840 max_objects = MAX_OBJS_PER_PAGE;
842 if (page->objects != max_objects) {
843 slab_err(s, page, "Wrong number of objects. Found %d but "
844 "should be %d", page->objects, max_objects);
845 page->objects = max_objects;
846 slab_fix(s, "Number of objects adjusted.");
848 if (page->inuse != page->objects - nr) {
849 slab_err(s, page, "Wrong object count. Counter is %d but "
850 "counted were %d", page->inuse, page->objects - nr);
851 page->inuse = page->objects - nr;
852 slab_fix(s, "Object count adjusted.");
854 return search == NULL;
857 static void trace(struct kmem_cache *s, struct page *page, void *object,
860 if (s->flags & SLAB_TRACE) {
861 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
863 alloc ? "alloc" : "free",
868 print_section("Object", (void *)object, s->objsize);
875 * Hooks for other subsystems that check memory allocations. In a typical
876 * production configuration these hooks all should produce no code at all.
878 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
880 flags &= gfp_allowed_mask;
881 lockdep_trace_alloc(flags);
882 might_sleep_if(flags & __GFP_WAIT);
884 return should_failslab(s->objsize, flags, s->flags);
887 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
889 flags &= gfp_allowed_mask;
890 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
891 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
894 static inline void slab_free_hook(struct kmem_cache *s, void *x)
896 kmemleak_free_recursive(x, s->flags);
899 * Trouble is that we may no longer disable interupts in the fast path
900 * So in order to make the debug calls that expect irqs to be
901 * disabled we need to disable interrupts temporarily.
903 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
907 local_irq_save(flags);
908 kmemcheck_slab_free(s, x, s->objsize);
909 debug_check_no_locks_freed(x, s->objsize);
910 local_irq_restore(flags);
913 if (!(s->flags & SLAB_DEBUG_OBJECTS))
914 debug_check_no_obj_freed(x, s->objsize);
918 * Tracking of fully allocated slabs for debugging purposes.
920 * list_lock must be held.
922 static void add_full(struct kmem_cache *s,
923 struct kmem_cache_node *n, struct page *page)
925 if (!(s->flags & SLAB_STORE_USER))
928 list_add(&page->lru, &n->full);
932 * list_lock must be held.
934 static void remove_full(struct kmem_cache *s, struct page *page)
936 if (!(s->flags & SLAB_STORE_USER))
939 list_del(&page->lru);
942 /* Tracking of the number of slabs for debugging purposes */
943 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
945 struct kmem_cache_node *n = get_node(s, node);
947 return atomic_long_read(&n->nr_slabs);
950 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
952 return atomic_long_read(&n->nr_slabs);
955 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
957 struct kmem_cache_node *n = get_node(s, node);
960 * May be called early in order to allocate a slab for the
961 * kmem_cache_node structure. Solve the chicken-egg
962 * dilemma by deferring the increment of the count during
963 * bootstrap (see early_kmem_cache_node_alloc).
966 atomic_long_inc(&n->nr_slabs);
967 atomic_long_add(objects, &n->total_objects);
970 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
972 struct kmem_cache_node *n = get_node(s, node);
974 atomic_long_dec(&n->nr_slabs);
975 atomic_long_sub(objects, &n->total_objects);
978 /* Object debug checks for alloc/free paths */
979 static void setup_object_debug(struct kmem_cache *s, struct page *page,
982 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
985 init_object(s, object, SLUB_RED_INACTIVE);
986 init_tracking(s, object);
989 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
990 void *object, unsigned long addr)
992 if (!check_slab(s, page))
995 if (!on_freelist(s, page, object)) {
996 object_err(s, page, object, "Object already allocated");
1000 if (!check_valid_pointer(s, page, object)) {
1001 object_err(s, page, object, "Freelist Pointer check fails");
1005 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1008 /* Success perform special debug activities for allocs */
1009 if (s->flags & SLAB_STORE_USER)
1010 set_track(s, object, TRACK_ALLOC, addr);
1011 trace(s, page, object, 1);
1012 init_object(s, object, SLUB_RED_ACTIVE);
1016 if (PageSlab(page)) {
1018 * If this is a slab page then lets do the best we can
1019 * to avoid issues in the future. Marking all objects
1020 * as used avoids touching the remaining objects.
1022 slab_fix(s, "Marking all objects used");
1023 page->inuse = page->objects;
1024 page->freelist = NULL;
1029 static noinline int free_debug_processing(struct kmem_cache *s,
1030 struct page *page, void *object, unsigned long addr)
1032 if (!check_slab(s, page))
1035 if (!check_valid_pointer(s, page, object)) {
1036 slab_err(s, page, "Invalid object pointer 0x%p", object);
1040 if (on_freelist(s, page, object)) {
1041 object_err(s, page, object, "Object already free");
1045 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1048 if (unlikely(s != page->slab)) {
1049 if (!PageSlab(page)) {
1050 slab_err(s, page, "Attempt to free object(0x%p) "
1051 "outside of slab", object);
1052 } else if (!page->slab) {
1054 "SLUB <none>: no slab for object 0x%p.\n",
1058 object_err(s, page, object,
1059 "page slab pointer corrupt.");
1063 /* Special debug activities for freeing objects */
1064 if (!page->frozen && !page->freelist) {
1065 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1067 spin_lock(&n->list_lock);
1068 remove_full(s, page);
1069 spin_unlock(&n->list_lock);
1071 if (s->flags & SLAB_STORE_USER)
1072 set_track(s, object, TRACK_FREE, addr);
1073 trace(s, page, object, 0);
1074 init_object(s, object, SLUB_RED_INACTIVE);
1078 slab_fix(s, "Object at 0x%p not freed", object);
1082 static int __init setup_slub_debug(char *str)
1084 slub_debug = DEBUG_DEFAULT_FLAGS;
1085 if (*str++ != '=' || !*str)
1087 * No options specified. Switch on full debugging.
1093 * No options but restriction on slabs. This means full
1094 * debugging for slabs matching a pattern.
1098 if (tolower(*str) == 'o') {
1100 * Avoid enabling debugging on caches if its minimum order
1101 * would increase as a result.
1103 disable_higher_order_debug = 1;
1110 * Switch off all debugging measures.
1115 * Determine which debug features should be switched on
1117 for (; *str && *str != ','; str++) {
1118 switch (tolower(*str)) {
1120 slub_debug |= SLAB_DEBUG_FREE;
1123 slub_debug |= SLAB_RED_ZONE;
1126 slub_debug |= SLAB_POISON;
1129 slub_debug |= SLAB_STORE_USER;
1132 slub_debug |= SLAB_TRACE;
1135 slub_debug |= SLAB_FAILSLAB;
1138 printk(KERN_ERR "slub_debug option '%c' "
1139 "unknown. skipped\n", *str);
1145 slub_debug_slabs = str + 1;
1150 __setup("slub_debug", setup_slub_debug);
1152 static unsigned long kmem_cache_flags(unsigned long objsize,
1153 unsigned long flags, const char *name,
1154 void (*ctor)(void *))
1157 * Enable debugging if selected on the kernel commandline.
1159 if (slub_debug && (!slub_debug_slabs ||
1160 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1161 flags |= slub_debug;
1166 static inline void setup_object_debug(struct kmem_cache *s,
1167 struct page *page, void *object) {}
1169 static inline int alloc_debug_processing(struct kmem_cache *s,
1170 struct page *page, void *object, unsigned long addr) { return 0; }
1172 static inline int free_debug_processing(struct kmem_cache *s,
1173 struct page *page, void *object, unsigned long addr) { return 0; }
1175 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1177 static inline int check_object(struct kmem_cache *s, struct page *page,
1178 void *object, u8 val) { return 1; }
1179 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1180 struct page *page) {}
1181 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1182 unsigned long flags, const char *name,
1183 void (*ctor)(void *))
1187 #define slub_debug 0
1189 #define disable_higher_order_debug 0
1191 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1193 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1195 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1197 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1200 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1203 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1206 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1208 #endif /* CONFIG_SLUB_DEBUG */
1211 * Slab allocation and freeing
1213 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1214 struct kmem_cache_order_objects oo)
1216 int order = oo_order(oo);
1218 flags |= __GFP_NOTRACK;
1220 if (node == NUMA_NO_NODE)
1221 return alloc_pages(flags, order);
1223 return alloc_pages_exact_node(node, flags, order);
1226 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1229 struct kmem_cache_order_objects oo = s->oo;
1232 flags &= gfp_allowed_mask;
1234 if (flags & __GFP_WAIT)
1237 flags |= s->allocflags;
1240 * Let the initial higher-order allocation fail under memory pressure
1241 * so we fall-back to the minimum order allocation.
1243 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1245 page = alloc_slab_page(alloc_gfp, node, oo);
1246 if (unlikely(!page)) {
1249 * Allocation may have failed due to fragmentation.
1250 * Try a lower order alloc if possible
1252 page = alloc_slab_page(flags, node, oo);
1255 stat(s, ORDER_FALLBACK);
1258 if (flags & __GFP_WAIT)
1259 local_irq_disable();
1264 if (kmemcheck_enabled
1265 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1266 int pages = 1 << oo_order(oo);
1268 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1271 * Objects from caches that have a constructor don't get
1272 * cleared when they're allocated, so we need to do it here.
1275 kmemcheck_mark_uninitialized_pages(page, pages);
1277 kmemcheck_mark_unallocated_pages(page, pages);
1280 page->objects = oo_objects(oo);
1281 mod_zone_page_state(page_zone(page),
1282 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1283 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1289 static void setup_object(struct kmem_cache *s, struct page *page,
1292 setup_object_debug(s, page, object);
1293 if (unlikely(s->ctor))
1297 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1304 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1306 page = allocate_slab(s,
1307 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1311 inc_slabs_node(s, page_to_nid(page), page->objects);
1313 page->flags |= 1 << PG_slab;
1315 start = page_address(page);
1317 if (unlikely(s->flags & SLAB_POISON))
1318 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1321 for_each_object(p, s, start, page->objects) {
1322 setup_object(s, page, last);
1323 set_freepointer(s, last, p);
1326 setup_object(s, page, last);
1327 set_freepointer(s, last, NULL);
1329 page->freelist = start;
1336 static void __free_slab(struct kmem_cache *s, struct page *page)
1338 int order = compound_order(page);
1339 int pages = 1 << order;
1341 if (kmem_cache_debug(s)) {
1344 slab_pad_check(s, page);
1345 for_each_object(p, s, page_address(page),
1347 check_object(s, page, p, SLUB_RED_INACTIVE);
1350 kmemcheck_free_shadow(page, compound_order(page));
1352 mod_zone_page_state(page_zone(page),
1353 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1354 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1357 __ClearPageSlab(page);
1358 reset_page_mapcount(page);
1359 if (current->reclaim_state)
1360 current->reclaim_state->reclaimed_slab += pages;
1361 __free_pages(page, order);
1364 #define need_reserve_slab_rcu \
1365 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1367 static void rcu_free_slab(struct rcu_head *h)
1371 if (need_reserve_slab_rcu)
1372 page = virt_to_head_page(h);
1374 page = container_of((struct list_head *)h, struct page, lru);
1376 __free_slab(page->slab, page);
1379 static void free_slab(struct kmem_cache *s, struct page *page)
1381 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1382 struct rcu_head *head;
1384 if (need_reserve_slab_rcu) {
1385 int order = compound_order(page);
1386 int offset = (PAGE_SIZE << order) - s->reserved;
1388 VM_BUG_ON(s->reserved != sizeof(*head));
1389 head = page_address(page) + offset;
1392 * RCU free overloads the RCU head over the LRU
1394 head = (void *)&page->lru;
1397 call_rcu(head, rcu_free_slab);
1399 __free_slab(s, page);
1402 static void discard_slab(struct kmem_cache *s, struct page *page)
1404 dec_slabs_node(s, page_to_nid(page), page->objects);
1409 * Per slab locking using the pagelock
1411 static __always_inline void slab_lock(struct page *page)
1413 bit_spin_lock(PG_locked, &page->flags);
1416 static __always_inline void slab_unlock(struct page *page)
1418 __bit_spin_unlock(PG_locked, &page->flags);
1421 static __always_inline int slab_trylock(struct page *page)
1425 rc = bit_spin_trylock(PG_locked, &page->flags);
1430 * Management of partially allocated slabs.
1432 * list_lock must be held.
1434 static inline void add_partial(struct kmem_cache_node *n,
1435 struct page *page, int tail)
1439 list_add_tail(&page->lru, &n->partial);
1441 list_add(&page->lru, &n->partial);
1445 * list_lock must be held.
1447 static inline void remove_partial(struct kmem_cache_node *n,
1450 list_del(&page->lru);
1455 * Lock slab, remove from the partial list and put the object into the
1458 * Must hold list_lock.
1460 static inline int lock_and_freeze_slab(struct kmem_cache *s,
1461 struct kmem_cache_node *n, struct page *page)
1463 if (slab_trylock(page)) {
1464 remove_partial(n, page);
1471 * Try to allocate a partial slab from a specific node.
1473 static struct page *get_partial_node(struct kmem_cache *s,
1474 struct kmem_cache_node *n)
1479 * Racy check. If we mistakenly see no partial slabs then we
1480 * just allocate an empty slab. If we mistakenly try to get a
1481 * partial slab and there is none available then get_partials()
1484 if (!n || !n->nr_partial)
1487 spin_lock(&n->list_lock);
1488 list_for_each_entry(page, &n->partial, lru)
1489 if (lock_and_freeze_slab(s, n, page))
1493 spin_unlock(&n->list_lock);
1498 * Get a page from somewhere. Search in increasing NUMA distances.
1500 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1503 struct zonelist *zonelist;
1506 enum zone_type high_zoneidx = gfp_zone(flags);
1510 * The defrag ratio allows a configuration of the tradeoffs between
1511 * inter node defragmentation and node local allocations. A lower
1512 * defrag_ratio increases the tendency to do local allocations
1513 * instead of attempting to obtain partial slabs from other nodes.
1515 * If the defrag_ratio is set to 0 then kmalloc() always
1516 * returns node local objects. If the ratio is higher then kmalloc()
1517 * may return off node objects because partial slabs are obtained
1518 * from other nodes and filled up.
1520 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1521 * defrag_ratio = 1000) then every (well almost) allocation will
1522 * first attempt to defrag slab caches on other nodes. This means
1523 * scanning over all nodes to look for partial slabs which may be
1524 * expensive if we do it every time we are trying to find a slab
1525 * with available objects.
1527 if (!s->remote_node_defrag_ratio ||
1528 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1532 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1533 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1534 struct kmem_cache_node *n;
1536 n = get_node(s, zone_to_nid(zone));
1538 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1539 n->nr_partial > s->min_partial) {
1540 page = get_partial_node(s, n);
1553 * Get a partial page, lock it and return it.
1555 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1558 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1560 page = get_partial_node(s, get_node(s, searchnode));
1561 if (page || node != NUMA_NO_NODE)
1564 return get_any_partial(s, flags);
1568 * Move a page back to the lists.
1570 * Must be called with the slab lock held.
1572 * On exit the slab lock will have been dropped.
1574 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1577 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1581 if (page->freelist) {
1582 spin_lock(&n->list_lock);
1583 add_partial(n, page, tail);
1584 spin_unlock(&n->list_lock);
1585 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1587 stat(s, DEACTIVATE_FULL);
1588 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER)) {
1589 spin_lock(&n->list_lock);
1590 add_full(s, n, page);
1591 spin_unlock(&n->list_lock);
1596 stat(s, DEACTIVATE_EMPTY);
1597 if (n->nr_partial < s->min_partial) {
1599 * Adding an empty slab to the partial slabs in order
1600 * to avoid page allocator overhead. This slab needs
1601 * to come after the other slabs with objects in
1602 * so that the others get filled first. That way the
1603 * size of the partial list stays small.
1605 * kmem_cache_shrink can reclaim any empty slabs from
1608 spin_lock(&n->list_lock);
1609 add_partial(n, page, 1);
1610 spin_unlock(&n->list_lock);
1615 discard_slab(s, page);
1620 #ifdef CONFIG_PREEMPT
1622 * Calculate the next globally unique transaction for disambiguiation
1623 * during cmpxchg. The transactions start with the cpu number and are then
1624 * incremented by CONFIG_NR_CPUS.
1626 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1629 * No preemption supported therefore also no need to check for
1635 static inline unsigned long next_tid(unsigned long tid)
1637 return tid + TID_STEP;
1640 static inline unsigned int tid_to_cpu(unsigned long tid)
1642 return tid % TID_STEP;
1645 static inline unsigned long tid_to_event(unsigned long tid)
1647 return tid / TID_STEP;
1650 static inline unsigned int init_tid(int cpu)
1655 static inline void note_cmpxchg_failure(const char *n,
1656 const struct kmem_cache *s, unsigned long tid)
1658 #ifdef SLUB_DEBUG_CMPXCHG
1659 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1661 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1663 #ifdef CONFIG_PREEMPT
1664 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1665 printk("due to cpu change %d -> %d\n",
1666 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1669 if (tid_to_event(tid) != tid_to_event(actual_tid))
1670 printk("due to cpu running other code. Event %ld->%ld\n",
1671 tid_to_event(tid), tid_to_event(actual_tid));
1673 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1674 actual_tid, tid, next_tid(tid));
1676 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1679 void init_kmem_cache_cpus(struct kmem_cache *s)
1683 for_each_possible_cpu(cpu)
1684 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1687 * Remove the cpu slab
1689 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1692 struct page *page = c->page;
1696 stat(s, DEACTIVATE_REMOTE_FREES);
1698 * Merge cpu freelist into slab freelist. Typically we get here
1699 * because both freelists are empty. So this is unlikely
1702 while (unlikely(c->freelist)) {
1705 tail = 0; /* Hot objects. Put the slab first */
1707 /* Retrieve object from cpu_freelist */
1708 object = c->freelist;
1709 c->freelist = get_freepointer(s, c->freelist);
1711 /* And put onto the regular freelist */
1712 set_freepointer(s, object, page->freelist);
1713 page->freelist = object;
1717 c->tid = next_tid(c->tid);
1719 unfreeze_slab(s, page, tail);
1722 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1724 stat(s, CPUSLAB_FLUSH);
1726 deactivate_slab(s, c);
1732 * Called from IPI handler with interrupts disabled.
1734 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1736 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1738 if (likely(c && c->page))
1742 static void flush_cpu_slab(void *d)
1744 struct kmem_cache *s = d;
1746 __flush_cpu_slab(s, smp_processor_id());
1749 static void flush_all(struct kmem_cache *s)
1751 on_each_cpu(flush_cpu_slab, s, 1);
1755 * Check if the objects in a per cpu structure fit numa
1756 * locality expectations.
1758 static inline int node_match(struct kmem_cache_cpu *c, int node)
1761 if (node != NUMA_NO_NODE && c->node != node)
1767 static int count_free(struct page *page)
1769 return page->objects - page->inuse;
1772 static unsigned long count_partial(struct kmem_cache_node *n,
1773 int (*get_count)(struct page *))
1775 unsigned long flags;
1776 unsigned long x = 0;
1779 spin_lock_irqsave(&n->list_lock, flags);
1780 list_for_each_entry(page, &n->partial, lru)
1781 x += get_count(page);
1782 spin_unlock_irqrestore(&n->list_lock, flags);
1786 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1788 #ifdef CONFIG_SLUB_DEBUG
1789 return atomic_long_read(&n->total_objects);
1795 static noinline void
1796 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1801 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1803 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1804 "default order: %d, min order: %d\n", s->name, s->objsize,
1805 s->size, oo_order(s->oo), oo_order(s->min));
1807 if (oo_order(s->min) > get_order(s->objsize))
1808 printk(KERN_WARNING " %s debugging increased min order, use "
1809 "slub_debug=O to disable.\n", s->name);
1811 for_each_online_node(node) {
1812 struct kmem_cache_node *n = get_node(s, node);
1813 unsigned long nr_slabs;
1814 unsigned long nr_objs;
1815 unsigned long nr_free;
1820 nr_free = count_partial(n, count_free);
1821 nr_slabs = node_nr_slabs(n);
1822 nr_objs = node_nr_objs(n);
1825 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1826 node, nr_slabs, nr_objs, nr_free);
1831 * Slow path. The lockless freelist is empty or we need to perform
1834 * Interrupts are disabled.
1836 * Processing is still very fast if new objects have been freed to the
1837 * regular freelist. In that case we simply take over the regular freelist
1838 * as the lockless freelist and zap the regular freelist.
1840 * If that is not working then we fall back to the partial lists. We take the
1841 * first element of the freelist as the object to allocate now and move the
1842 * rest of the freelist to the lockless freelist.
1844 * And if we were unable to get a new slab from the partial slab lists then
1845 * we need to allocate a new slab. This is the slowest path since it involves
1846 * a call to the page allocator and the setup of a new slab.
1848 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1849 unsigned long addr, struct kmem_cache_cpu *c)
1853 unsigned long flags;
1855 local_irq_save(flags);
1856 #ifdef CONFIG_PREEMPT
1858 * We may have been preempted and rescheduled on a different
1859 * cpu before disabling interrupts. Need to reload cpu area
1862 c = this_cpu_ptr(s->cpu_slab);
1865 /* We handle __GFP_ZERO in the caller */
1866 gfpflags &= ~__GFP_ZERO;
1873 if (unlikely(!node_match(c, node)))
1876 stat(s, ALLOC_REFILL);
1879 VM_BUG_ON(!page->frozen);
1881 object = page->freelist;
1882 if (unlikely(!object))
1884 if (kmem_cache_debug(s))
1887 c->freelist = get_freepointer(s, object);
1888 page->inuse = page->objects;
1889 page->freelist = NULL;
1892 c->tid = next_tid(c->tid);
1893 local_irq_restore(flags);
1894 stat(s, ALLOC_SLOWPATH);
1898 deactivate_slab(s, c);
1901 page = get_partial(s, gfpflags, node);
1903 stat(s, ALLOC_FROM_PARTIAL);
1905 c->node = page_to_nid(page);
1910 page = new_slab(s, gfpflags, node);
1913 c = __this_cpu_ptr(s->cpu_slab);
1914 stat(s, ALLOC_SLAB);
1920 c->node = page_to_nid(page);
1924 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1925 slab_out_of_memory(s, gfpflags, node);
1926 local_irq_restore(flags);
1929 if (!alloc_debug_processing(s, page, object, addr))
1933 page->freelist = get_freepointer(s, object);
1934 deactivate_slab(s, c);
1936 c->node = NUMA_NO_NODE;
1937 local_irq_restore(flags);
1942 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1943 * have the fastpath folded into their functions. So no function call
1944 * overhead for requests that can be satisfied on the fastpath.
1946 * The fastpath works by first checking if the lockless freelist can be used.
1947 * If not then __slab_alloc is called for slow processing.
1949 * Otherwise we can simply pick the next object from the lockless free list.
1951 static __always_inline void *slab_alloc(struct kmem_cache *s,
1952 gfp_t gfpflags, int node, unsigned long addr)
1955 struct kmem_cache_cpu *c;
1958 if (slab_pre_alloc_hook(s, gfpflags))
1964 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
1965 * enabled. We may switch back and forth between cpus while
1966 * reading from one cpu area. That does not matter as long
1967 * as we end up on the original cpu again when doing the cmpxchg.
1969 c = __this_cpu_ptr(s->cpu_slab);
1972 * The transaction ids are globally unique per cpu and per operation on
1973 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
1974 * occurs on the right processor and that there was no operation on the
1975 * linked list in between.
1980 object = c->freelist;
1981 if (unlikely(!object || !node_match(c, node)))
1983 object = __slab_alloc(s, gfpflags, node, addr, c);
1987 * The cmpxchg will only match if there was no additional
1988 * operation and if we are on the right processor.
1990 * The cmpxchg does the following atomically (without lock semantics!)
1991 * 1. Relocate first pointer to the current per cpu area.
1992 * 2. Verify that tid and freelist have not been changed
1993 * 3. If they were not changed replace tid and freelist
1995 * Since this is without lock semantics the protection is only against
1996 * code executing on this cpu *not* from access by other cpus.
1998 if (unlikely(!irqsafe_cpu_cmpxchg_double(
1999 s->cpu_slab->freelist, s->cpu_slab->tid,
2001 get_freepointer_safe(s, object), next_tid(tid)))) {
2003 note_cmpxchg_failure("slab_alloc", s, tid);
2006 stat(s, ALLOC_FASTPATH);
2009 if (unlikely(gfpflags & __GFP_ZERO) && object)
2010 memset(object, 0, s->objsize);
2012 slab_post_alloc_hook(s, gfpflags, object);
2017 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2019 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2021 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2025 EXPORT_SYMBOL(kmem_cache_alloc);
2027 #ifdef CONFIG_TRACING
2028 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2030 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2031 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2034 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2036 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2038 void *ret = kmalloc_order(size, flags, order);
2039 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2042 EXPORT_SYMBOL(kmalloc_order_trace);
2046 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2048 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2050 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2051 s->objsize, s->size, gfpflags, node);
2055 EXPORT_SYMBOL(kmem_cache_alloc_node);
2057 #ifdef CONFIG_TRACING
2058 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2060 int node, size_t size)
2062 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2064 trace_kmalloc_node(_RET_IP_, ret,
2065 size, s->size, gfpflags, node);
2068 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2073 * Slow patch handling. This may still be called frequently since objects
2074 * have a longer lifetime than the cpu slabs in most processing loads.
2076 * So we still attempt to reduce cache line usage. Just take the slab
2077 * lock and free the item. If there is no additional partial page
2078 * handling required then we can return immediately.
2080 static void __slab_free(struct kmem_cache *s, struct page *page,
2081 void *x, unsigned long addr)
2084 void **object = (void *)x;
2085 unsigned long uninitialized_var(flags);
2087 local_irq_save(flags);
2089 stat(s, FREE_SLOWPATH);
2091 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2094 prior = page->freelist;
2095 set_freepointer(s, object, prior);
2096 page->freelist = object;
2099 if (unlikely(page->frozen)) {
2100 stat(s, FREE_FROZEN);
2104 if (unlikely(!page->inuse))
2108 * Objects left in the slab. If it was not on the partial list before
2111 if (unlikely(!prior)) {
2112 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2114 spin_lock(&n->list_lock);
2115 add_partial(get_node(s, page_to_nid(page)), page, 1);
2116 spin_unlock(&n->list_lock);
2117 stat(s, FREE_ADD_PARTIAL);
2122 local_irq_restore(flags);
2128 * Slab still on the partial list.
2130 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2132 spin_lock(&n->list_lock);
2133 remove_partial(n, page);
2134 spin_unlock(&n->list_lock);
2135 stat(s, FREE_REMOVE_PARTIAL);
2138 local_irq_restore(flags);
2140 discard_slab(s, page);
2144 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2145 * can perform fastpath freeing without additional function calls.
2147 * The fastpath is only possible if we are freeing to the current cpu slab
2148 * of this processor. This typically the case if we have just allocated
2151 * If fastpath is not possible then fall back to __slab_free where we deal
2152 * with all sorts of special processing.
2154 static __always_inline void slab_free(struct kmem_cache *s,
2155 struct page *page, void *x, unsigned long addr)
2157 void **object = (void *)x;
2158 struct kmem_cache_cpu *c;
2161 slab_free_hook(s, x);
2166 * Determine the currently cpus per cpu slab.
2167 * The cpu may change afterward. However that does not matter since
2168 * data is retrieved via this pointer. If we are on the same cpu
2169 * during the cmpxchg then the free will succedd.
2171 c = __this_cpu_ptr(s->cpu_slab);
2176 if (likely(page == c->page)) {
2177 set_freepointer(s, object, c->freelist);
2179 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2180 s->cpu_slab->freelist, s->cpu_slab->tid,
2182 object, next_tid(tid)))) {
2184 note_cmpxchg_failure("slab_free", s, tid);
2187 stat(s, FREE_FASTPATH);
2189 __slab_free(s, page, x, addr);
2193 void kmem_cache_free(struct kmem_cache *s, void *x)
2197 page = virt_to_head_page(x);
2199 slab_free(s, page, x, _RET_IP_);
2201 trace_kmem_cache_free(_RET_IP_, x);
2203 EXPORT_SYMBOL(kmem_cache_free);
2206 * Object placement in a slab is made very easy because we always start at
2207 * offset 0. If we tune the size of the object to the alignment then we can
2208 * get the required alignment by putting one properly sized object after
2211 * Notice that the allocation order determines the sizes of the per cpu
2212 * caches. Each processor has always one slab available for allocations.
2213 * Increasing the allocation order reduces the number of times that slabs
2214 * must be moved on and off the partial lists and is therefore a factor in
2219 * Mininum / Maximum order of slab pages. This influences locking overhead
2220 * and slab fragmentation. A higher order reduces the number of partial slabs
2221 * and increases the number of allocations possible without having to
2222 * take the list_lock.
2224 static int slub_min_order;
2225 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2226 static int slub_min_objects;
2229 * Merge control. If this is set then no merging of slab caches will occur.
2230 * (Could be removed. This was introduced to pacify the merge skeptics.)
2232 static int slub_nomerge;
2235 * Calculate the order of allocation given an slab object size.
2237 * The order of allocation has significant impact on performance and other
2238 * system components. Generally order 0 allocations should be preferred since
2239 * order 0 does not cause fragmentation in the page allocator. Larger objects
2240 * be problematic to put into order 0 slabs because there may be too much
2241 * unused space left. We go to a higher order if more than 1/16th of the slab
2244 * In order to reach satisfactory performance we must ensure that a minimum
2245 * number of objects is in one slab. Otherwise we may generate too much
2246 * activity on the partial lists which requires taking the list_lock. This is
2247 * less a concern for large slabs though which are rarely used.
2249 * slub_max_order specifies the order where we begin to stop considering the
2250 * number of objects in a slab as critical. If we reach slub_max_order then
2251 * we try to keep the page order as low as possible. So we accept more waste
2252 * of space in favor of a small page order.
2254 * Higher order allocations also allow the placement of more objects in a
2255 * slab and thereby reduce object handling overhead. If the user has
2256 * requested a higher mininum order then we start with that one instead of
2257 * the smallest order which will fit the object.
2259 static inline int slab_order(int size, int min_objects,
2260 int max_order, int fract_leftover, int reserved)
2264 int min_order = slub_min_order;
2266 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2267 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2269 for (order = max(min_order,
2270 fls(min_objects * size - 1) - PAGE_SHIFT);
2271 order <= max_order; order++) {
2273 unsigned long slab_size = PAGE_SIZE << order;
2275 if (slab_size < min_objects * size + reserved)
2278 rem = (slab_size - reserved) % size;
2280 if (rem <= slab_size / fract_leftover)
2288 static inline int calculate_order(int size, int reserved)
2296 * Attempt to find best configuration for a slab. This
2297 * works by first attempting to generate a layout with
2298 * the best configuration and backing off gradually.
2300 * First we reduce the acceptable waste in a slab. Then
2301 * we reduce the minimum objects required in a slab.
2303 min_objects = slub_min_objects;
2305 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2306 max_objects = order_objects(slub_max_order, size, reserved);
2307 min_objects = min(min_objects, max_objects);
2309 while (min_objects > 1) {
2311 while (fraction >= 4) {
2312 order = slab_order(size, min_objects,
2313 slub_max_order, fraction, reserved);
2314 if (order <= slub_max_order)
2322 * We were unable to place multiple objects in a slab. Now
2323 * lets see if we can place a single object there.
2325 order = slab_order(size, 1, slub_max_order, 1, reserved);
2326 if (order <= slub_max_order)
2330 * Doh this slab cannot be placed using slub_max_order.
2332 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2333 if (order < MAX_ORDER)
2339 * Figure out what the alignment of the objects will be.
2341 static unsigned long calculate_alignment(unsigned long flags,
2342 unsigned long align, unsigned long size)
2345 * If the user wants hardware cache aligned objects then follow that
2346 * suggestion if the object is sufficiently large.
2348 * The hardware cache alignment cannot override the specified
2349 * alignment though. If that is greater then use it.
2351 if (flags & SLAB_HWCACHE_ALIGN) {
2352 unsigned long ralign = cache_line_size();
2353 while (size <= ralign / 2)
2355 align = max(align, ralign);
2358 if (align < ARCH_SLAB_MINALIGN)
2359 align = ARCH_SLAB_MINALIGN;
2361 return ALIGN(align, sizeof(void *));
2365 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2368 spin_lock_init(&n->list_lock);
2369 INIT_LIST_HEAD(&n->partial);
2370 #ifdef CONFIG_SLUB_DEBUG
2371 atomic_long_set(&n->nr_slabs, 0);
2372 atomic_long_set(&n->total_objects, 0);
2373 INIT_LIST_HEAD(&n->full);
2377 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2379 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2380 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2383 * Must align to double word boundary for the double cmpxchg
2384 * instructions to work; see __pcpu_double_call_return_bool().
2386 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2387 2 * sizeof(void *));
2392 init_kmem_cache_cpus(s);
2397 static struct kmem_cache *kmem_cache_node;
2400 * No kmalloc_node yet so do it by hand. We know that this is the first
2401 * slab on the node for this slabcache. There are no concurrent accesses
2404 * Note that this function only works on the kmalloc_node_cache
2405 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2406 * memory on a fresh node that has no slab structures yet.
2408 static void early_kmem_cache_node_alloc(int node)
2411 struct kmem_cache_node *n;
2413 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2415 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2418 if (page_to_nid(page) != node) {
2419 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2421 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2422 "in order to be able to continue\n");
2427 page->freelist = get_freepointer(kmem_cache_node, n);
2430 kmem_cache_node->node[node] = n;
2431 #ifdef CONFIG_SLUB_DEBUG
2432 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2433 init_tracking(kmem_cache_node, n);
2435 init_kmem_cache_node(n, kmem_cache_node);
2436 inc_slabs_node(kmem_cache_node, node, page->objects);
2438 add_partial(n, page, 0);
2441 static void free_kmem_cache_nodes(struct kmem_cache *s)
2445 for_each_node_state(node, N_NORMAL_MEMORY) {
2446 struct kmem_cache_node *n = s->node[node];
2449 kmem_cache_free(kmem_cache_node, n);
2451 s->node[node] = NULL;
2455 static int init_kmem_cache_nodes(struct kmem_cache *s)
2459 for_each_node_state(node, N_NORMAL_MEMORY) {
2460 struct kmem_cache_node *n;
2462 if (slab_state == DOWN) {
2463 early_kmem_cache_node_alloc(node);
2466 n = kmem_cache_alloc_node(kmem_cache_node,
2470 free_kmem_cache_nodes(s);
2475 init_kmem_cache_node(n, s);
2480 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2482 if (min < MIN_PARTIAL)
2484 else if (min > MAX_PARTIAL)
2486 s->min_partial = min;
2490 * calculate_sizes() determines the order and the distribution of data within
2493 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2495 unsigned long flags = s->flags;
2496 unsigned long size = s->objsize;
2497 unsigned long align = s->align;
2501 * Round up object size to the next word boundary. We can only
2502 * place the free pointer at word boundaries and this determines
2503 * the possible location of the free pointer.
2505 size = ALIGN(size, sizeof(void *));
2507 #ifdef CONFIG_SLUB_DEBUG
2509 * Determine if we can poison the object itself. If the user of
2510 * the slab may touch the object after free or before allocation
2511 * then we should never poison the object itself.
2513 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2515 s->flags |= __OBJECT_POISON;
2517 s->flags &= ~__OBJECT_POISON;
2521 * If we are Redzoning then check if there is some space between the
2522 * end of the object and the free pointer. If not then add an
2523 * additional word to have some bytes to store Redzone information.
2525 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2526 size += sizeof(void *);
2530 * With that we have determined the number of bytes in actual use
2531 * by the object. This is the potential offset to the free pointer.
2535 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2538 * Relocate free pointer after the object if it is not
2539 * permitted to overwrite the first word of the object on
2542 * This is the case if we do RCU, have a constructor or
2543 * destructor or are poisoning the objects.
2546 size += sizeof(void *);
2549 #ifdef CONFIG_SLUB_DEBUG
2550 if (flags & SLAB_STORE_USER)
2552 * Need to store information about allocs and frees after
2555 size += 2 * sizeof(struct track);
2557 if (flags & SLAB_RED_ZONE)
2559 * Add some empty padding so that we can catch
2560 * overwrites from earlier objects rather than let
2561 * tracking information or the free pointer be
2562 * corrupted if a user writes before the start
2565 size += sizeof(void *);
2569 * Determine the alignment based on various parameters that the
2570 * user specified and the dynamic determination of cache line size
2573 align = calculate_alignment(flags, align, s->objsize);
2577 * SLUB stores one object immediately after another beginning from
2578 * offset 0. In order to align the objects we have to simply size
2579 * each object to conform to the alignment.
2581 size = ALIGN(size, align);
2583 if (forced_order >= 0)
2584 order = forced_order;
2586 order = calculate_order(size, s->reserved);
2593 s->allocflags |= __GFP_COMP;
2595 if (s->flags & SLAB_CACHE_DMA)
2596 s->allocflags |= SLUB_DMA;
2598 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2599 s->allocflags |= __GFP_RECLAIMABLE;
2602 * Determine the number of objects per slab
2604 s->oo = oo_make(order, size, s->reserved);
2605 s->min = oo_make(get_order(size), size, s->reserved);
2606 if (oo_objects(s->oo) > oo_objects(s->max))
2609 return !!oo_objects(s->oo);
2613 static int kmem_cache_open(struct kmem_cache *s,
2614 const char *name, size_t size,
2615 size_t align, unsigned long flags,
2616 void (*ctor)(void *))
2618 memset(s, 0, kmem_size);
2623 s->flags = kmem_cache_flags(size, flags, name, ctor);
2626 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2627 s->reserved = sizeof(struct rcu_head);
2629 if (!calculate_sizes(s, -1))
2631 if (disable_higher_order_debug) {
2633 * Disable debugging flags that store metadata if the min slab
2636 if (get_order(s->size) > get_order(s->objsize)) {
2637 s->flags &= ~DEBUG_METADATA_FLAGS;
2639 if (!calculate_sizes(s, -1))
2644 #ifdef CONFIG_CMPXCHG_DOUBLE
2645 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
2646 /* Enable fast mode */
2647 s->flags |= __CMPXCHG_DOUBLE;
2651 * The larger the object size is, the more pages we want on the partial
2652 * list to avoid pounding the page allocator excessively.
2654 set_min_partial(s, ilog2(s->size));
2657 s->remote_node_defrag_ratio = 1000;
2659 if (!init_kmem_cache_nodes(s))
2662 if (alloc_kmem_cache_cpus(s))
2665 free_kmem_cache_nodes(s);
2667 if (flags & SLAB_PANIC)
2668 panic("Cannot create slab %s size=%lu realsize=%u "
2669 "order=%u offset=%u flags=%lx\n",
2670 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2676 * Determine the size of a slab object
2678 unsigned int kmem_cache_size(struct kmem_cache *s)
2682 EXPORT_SYMBOL(kmem_cache_size);
2684 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2687 #ifdef CONFIG_SLUB_DEBUG
2688 void *addr = page_address(page);
2690 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2691 sizeof(long), GFP_ATOMIC);
2694 slab_err(s, page, "%s", text);
2697 get_map(s, page, map);
2698 for_each_object(p, s, addr, page->objects) {
2700 if (!test_bit(slab_index(p, s, addr), map)) {
2701 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2703 print_tracking(s, p);
2712 * Attempt to free all partial slabs on a node.
2714 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2716 unsigned long flags;
2717 struct page *page, *h;
2719 spin_lock_irqsave(&n->list_lock, flags);
2720 list_for_each_entry_safe(page, h, &n->partial, lru) {
2722 remove_partial(n, page);
2723 discard_slab(s, page);
2725 list_slab_objects(s, page,
2726 "Objects remaining on kmem_cache_close()");
2729 spin_unlock_irqrestore(&n->list_lock, flags);
2733 * Release all resources used by a slab cache.
2735 static inline int kmem_cache_close(struct kmem_cache *s)
2740 free_percpu(s->cpu_slab);
2741 /* Attempt to free all objects */
2742 for_each_node_state(node, N_NORMAL_MEMORY) {
2743 struct kmem_cache_node *n = get_node(s, node);
2746 if (n->nr_partial || slabs_node(s, node))
2749 free_kmem_cache_nodes(s);
2754 * Close a cache and release the kmem_cache structure
2755 * (must be used for caches created using kmem_cache_create)
2757 void kmem_cache_destroy(struct kmem_cache *s)
2759 down_write(&slub_lock);
2763 if (kmem_cache_close(s)) {
2764 printk(KERN_ERR "SLUB %s: %s called for cache that "
2765 "still has objects.\n", s->name, __func__);
2768 if (s->flags & SLAB_DESTROY_BY_RCU)
2770 sysfs_slab_remove(s);
2772 up_write(&slub_lock);
2774 EXPORT_SYMBOL(kmem_cache_destroy);
2776 /********************************************************************
2778 *******************************************************************/
2780 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2781 EXPORT_SYMBOL(kmalloc_caches);
2783 static struct kmem_cache *kmem_cache;
2785 #ifdef CONFIG_ZONE_DMA
2786 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2789 static int __init setup_slub_min_order(char *str)
2791 get_option(&str, &slub_min_order);
2796 __setup("slub_min_order=", setup_slub_min_order);
2798 static int __init setup_slub_max_order(char *str)
2800 get_option(&str, &slub_max_order);
2801 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2806 __setup("slub_max_order=", setup_slub_max_order);
2808 static int __init setup_slub_min_objects(char *str)
2810 get_option(&str, &slub_min_objects);
2815 __setup("slub_min_objects=", setup_slub_min_objects);
2817 static int __init setup_slub_nomerge(char *str)
2823 __setup("slub_nomerge", setup_slub_nomerge);
2825 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2826 int size, unsigned int flags)
2828 struct kmem_cache *s;
2830 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2833 * This function is called with IRQs disabled during early-boot on
2834 * single CPU so there's no need to take slub_lock here.
2836 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2840 list_add(&s->list, &slab_caches);
2844 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2849 * Conversion table for small slabs sizes / 8 to the index in the
2850 * kmalloc array. This is necessary for slabs < 192 since we have non power
2851 * of two cache sizes there. The size of larger slabs can be determined using
2854 static s8 size_index[24] = {
2881 static inline int size_index_elem(size_t bytes)
2883 return (bytes - 1) / 8;
2886 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2892 return ZERO_SIZE_PTR;
2894 index = size_index[size_index_elem(size)];
2896 index = fls(size - 1);
2898 #ifdef CONFIG_ZONE_DMA
2899 if (unlikely((flags & SLUB_DMA)))
2900 return kmalloc_dma_caches[index];
2903 return kmalloc_caches[index];
2906 void *__kmalloc(size_t size, gfp_t flags)
2908 struct kmem_cache *s;
2911 if (unlikely(size > SLUB_MAX_SIZE))
2912 return kmalloc_large(size, flags);
2914 s = get_slab(size, flags);
2916 if (unlikely(ZERO_OR_NULL_PTR(s)))
2919 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2921 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2925 EXPORT_SYMBOL(__kmalloc);
2928 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2933 flags |= __GFP_COMP | __GFP_NOTRACK;
2934 page = alloc_pages_node(node, flags, get_order(size));
2936 ptr = page_address(page);
2938 kmemleak_alloc(ptr, size, 1, flags);
2942 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2944 struct kmem_cache *s;
2947 if (unlikely(size > SLUB_MAX_SIZE)) {
2948 ret = kmalloc_large_node(size, flags, node);
2950 trace_kmalloc_node(_RET_IP_, ret,
2951 size, PAGE_SIZE << get_order(size),
2957 s = get_slab(size, flags);
2959 if (unlikely(ZERO_OR_NULL_PTR(s)))
2962 ret = slab_alloc(s, flags, node, _RET_IP_);
2964 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2968 EXPORT_SYMBOL(__kmalloc_node);
2971 size_t ksize(const void *object)
2975 if (unlikely(object == ZERO_SIZE_PTR))
2978 page = virt_to_head_page(object);
2980 if (unlikely(!PageSlab(page))) {
2981 WARN_ON(!PageCompound(page));
2982 return PAGE_SIZE << compound_order(page);
2985 return slab_ksize(page->slab);
2987 EXPORT_SYMBOL(ksize);
2989 void kfree(const void *x)
2992 void *object = (void *)x;
2994 trace_kfree(_RET_IP_, x);
2996 if (unlikely(ZERO_OR_NULL_PTR(x)))
2999 page = virt_to_head_page(x);
3000 if (unlikely(!PageSlab(page))) {
3001 BUG_ON(!PageCompound(page));
3006 slab_free(page->slab, page, object, _RET_IP_);
3008 EXPORT_SYMBOL(kfree);
3011 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3012 * the remaining slabs by the number of items in use. The slabs with the
3013 * most items in use come first. New allocations will then fill those up
3014 * and thus they can be removed from the partial lists.
3016 * The slabs with the least items are placed last. This results in them
3017 * being allocated from last increasing the chance that the last objects
3018 * are freed in them.
3020 int kmem_cache_shrink(struct kmem_cache *s)
3024 struct kmem_cache_node *n;
3027 int objects = oo_objects(s->max);
3028 struct list_head *slabs_by_inuse =
3029 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3030 unsigned long flags;
3032 if (!slabs_by_inuse)
3036 for_each_node_state(node, N_NORMAL_MEMORY) {
3037 n = get_node(s, node);
3042 for (i = 0; i < objects; i++)
3043 INIT_LIST_HEAD(slabs_by_inuse + i);
3045 spin_lock_irqsave(&n->list_lock, flags);
3048 * Build lists indexed by the items in use in each slab.
3050 * Note that concurrent frees may occur while we hold the
3051 * list_lock. page->inuse here is the upper limit.
3053 list_for_each_entry_safe(page, t, &n->partial, lru) {
3054 if (!page->inuse && slab_trylock(page)) {
3056 * Must hold slab lock here because slab_free
3057 * may have freed the last object and be
3058 * waiting to release the slab.
3060 remove_partial(n, page);
3062 discard_slab(s, page);
3064 list_move(&page->lru,
3065 slabs_by_inuse + page->inuse);
3070 * Rebuild the partial list with the slabs filled up most
3071 * first and the least used slabs at the end.
3073 for (i = objects - 1; i >= 0; i--)
3074 list_splice(slabs_by_inuse + i, n->partial.prev);
3076 spin_unlock_irqrestore(&n->list_lock, flags);
3079 kfree(slabs_by_inuse);
3082 EXPORT_SYMBOL(kmem_cache_shrink);
3084 #if defined(CONFIG_MEMORY_HOTPLUG)
3085 static int slab_mem_going_offline_callback(void *arg)
3087 struct kmem_cache *s;
3089 down_read(&slub_lock);
3090 list_for_each_entry(s, &slab_caches, list)
3091 kmem_cache_shrink(s);
3092 up_read(&slub_lock);
3097 static void slab_mem_offline_callback(void *arg)
3099 struct kmem_cache_node *n;
3100 struct kmem_cache *s;
3101 struct memory_notify *marg = arg;
3104 offline_node = marg->status_change_nid;
3107 * If the node still has available memory. we need kmem_cache_node
3110 if (offline_node < 0)
3113 down_read(&slub_lock);
3114 list_for_each_entry(s, &slab_caches, list) {
3115 n = get_node(s, offline_node);
3118 * if n->nr_slabs > 0, slabs still exist on the node
3119 * that is going down. We were unable to free them,
3120 * and offline_pages() function shouldn't call this
3121 * callback. So, we must fail.
3123 BUG_ON(slabs_node(s, offline_node));
3125 s->node[offline_node] = NULL;
3126 kmem_cache_free(kmem_cache_node, n);
3129 up_read(&slub_lock);
3132 static int slab_mem_going_online_callback(void *arg)
3134 struct kmem_cache_node *n;
3135 struct kmem_cache *s;
3136 struct memory_notify *marg = arg;
3137 int nid = marg->status_change_nid;
3141 * If the node's memory is already available, then kmem_cache_node is
3142 * already created. Nothing to do.
3148 * We are bringing a node online. No memory is available yet. We must
3149 * allocate a kmem_cache_node structure in order to bring the node
3152 down_read(&slub_lock);
3153 list_for_each_entry(s, &slab_caches, list) {
3155 * XXX: kmem_cache_alloc_node will fallback to other nodes
3156 * since memory is not yet available from the node that
3159 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3164 init_kmem_cache_node(n, s);
3168 up_read(&slub_lock);
3172 static int slab_memory_callback(struct notifier_block *self,
3173 unsigned long action, void *arg)
3178 case MEM_GOING_ONLINE:
3179 ret = slab_mem_going_online_callback(arg);
3181 case MEM_GOING_OFFLINE:
3182 ret = slab_mem_going_offline_callback(arg);
3185 case MEM_CANCEL_ONLINE:
3186 slab_mem_offline_callback(arg);
3189 case MEM_CANCEL_OFFLINE:
3193 ret = notifier_from_errno(ret);
3199 #endif /* CONFIG_MEMORY_HOTPLUG */
3201 /********************************************************************
3202 * Basic setup of slabs
3203 *******************************************************************/
3206 * Used for early kmem_cache structures that were allocated using
3207 * the page allocator
3210 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3214 list_add(&s->list, &slab_caches);
3217 for_each_node_state(node, N_NORMAL_MEMORY) {
3218 struct kmem_cache_node *n = get_node(s, node);
3222 list_for_each_entry(p, &n->partial, lru)
3225 #ifdef CONFIG_SLUB_DEBUG
3226 list_for_each_entry(p, &n->full, lru)
3233 void __init kmem_cache_init(void)
3237 struct kmem_cache *temp_kmem_cache;
3239 struct kmem_cache *temp_kmem_cache_node;
3240 unsigned long kmalloc_size;
3242 kmem_size = offsetof(struct kmem_cache, node) +
3243 nr_node_ids * sizeof(struct kmem_cache_node *);
3245 /* Allocate two kmem_caches from the page allocator */
3246 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3247 order = get_order(2 * kmalloc_size);
3248 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3251 * Must first have the slab cache available for the allocations of the
3252 * struct kmem_cache_node's. There is special bootstrap code in
3253 * kmem_cache_open for slab_state == DOWN.
3255 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3257 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3258 sizeof(struct kmem_cache_node),
3259 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3261 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3263 /* Able to allocate the per node structures */
3264 slab_state = PARTIAL;
3266 temp_kmem_cache = kmem_cache;
3267 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3268 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3269 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3270 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3273 * Allocate kmem_cache_node properly from the kmem_cache slab.
3274 * kmem_cache_node is separately allocated so no need to
3275 * update any list pointers.
3277 temp_kmem_cache_node = kmem_cache_node;
3279 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3280 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3282 kmem_cache_bootstrap_fixup(kmem_cache_node);
3285 kmem_cache_bootstrap_fixup(kmem_cache);
3287 /* Free temporary boot structure */
3288 free_pages((unsigned long)temp_kmem_cache, order);
3290 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3293 * Patch up the size_index table if we have strange large alignment
3294 * requirements for the kmalloc array. This is only the case for
3295 * MIPS it seems. The standard arches will not generate any code here.
3297 * Largest permitted alignment is 256 bytes due to the way we
3298 * handle the index determination for the smaller caches.
3300 * Make sure that nothing crazy happens if someone starts tinkering
3301 * around with ARCH_KMALLOC_MINALIGN
3303 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3304 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3306 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3307 int elem = size_index_elem(i);
3308 if (elem >= ARRAY_SIZE(size_index))
3310 size_index[elem] = KMALLOC_SHIFT_LOW;
3313 if (KMALLOC_MIN_SIZE == 64) {
3315 * The 96 byte size cache is not used if the alignment
3318 for (i = 64 + 8; i <= 96; i += 8)
3319 size_index[size_index_elem(i)] = 7;
3320 } else if (KMALLOC_MIN_SIZE == 128) {
3322 * The 192 byte sized cache is not used if the alignment
3323 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3326 for (i = 128 + 8; i <= 192; i += 8)
3327 size_index[size_index_elem(i)] = 8;
3330 /* Caches that are not of the two-to-the-power-of size */
3331 if (KMALLOC_MIN_SIZE <= 32) {
3332 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3336 if (KMALLOC_MIN_SIZE <= 64) {
3337 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3341 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3342 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3348 /* Provide the correct kmalloc names now that the caches are up */
3349 if (KMALLOC_MIN_SIZE <= 32) {
3350 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3351 BUG_ON(!kmalloc_caches[1]->name);
3354 if (KMALLOC_MIN_SIZE <= 64) {
3355 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3356 BUG_ON(!kmalloc_caches[2]->name);
3359 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3360 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3363 kmalloc_caches[i]->name = s;
3367 register_cpu_notifier(&slab_notifier);
3370 #ifdef CONFIG_ZONE_DMA
3371 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3372 struct kmem_cache *s = kmalloc_caches[i];
3375 char *name = kasprintf(GFP_NOWAIT,
3376 "dma-kmalloc-%d", s->objsize);
3379 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3380 s->objsize, SLAB_CACHE_DMA);
3385 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3386 " CPUs=%d, Nodes=%d\n",
3387 caches, cache_line_size(),
3388 slub_min_order, slub_max_order, slub_min_objects,
3389 nr_cpu_ids, nr_node_ids);
3392 void __init kmem_cache_init_late(void)
3397 * Find a mergeable slab cache
3399 static int slab_unmergeable(struct kmem_cache *s)
3401 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3408 * We may have set a slab to be unmergeable during bootstrap.
3410 if (s->refcount < 0)
3416 static struct kmem_cache *find_mergeable(size_t size,
3417 size_t align, unsigned long flags, const char *name,
3418 void (*ctor)(void *))
3420 struct kmem_cache *s;
3422 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3428 size = ALIGN(size, sizeof(void *));
3429 align = calculate_alignment(flags, align, size);
3430 size = ALIGN(size, align);
3431 flags = kmem_cache_flags(size, flags, name, NULL);
3433 list_for_each_entry(s, &slab_caches, list) {
3434 if (slab_unmergeable(s))
3440 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3443 * Check if alignment is compatible.
3444 * Courtesy of Adrian Drzewiecki
3446 if ((s->size & ~(align - 1)) != s->size)
3449 if (s->size - size >= sizeof(void *))
3457 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3458 size_t align, unsigned long flags, void (*ctor)(void *))
3460 struct kmem_cache *s;
3466 down_write(&slub_lock);
3467 s = find_mergeable(size, align, flags, name, ctor);
3471 * Adjust the object sizes so that we clear
3472 * the complete object on kzalloc.
3474 s->objsize = max(s->objsize, (int)size);
3475 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3477 if (sysfs_slab_alias(s, name)) {
3481 up_write(&slub_lock);
3485 n = kstrdup(name, GFP_KERNEL);
3489 s = kmalloc(kmem_size, GFP_KERNEL);
3491 if (kmem_cache_open(s, n,
3492 size, align, flags, ctor)) {
3493 list_add(&s->list, &slab_caches);
3494 if (sysfs_slab_add(s)) {
3500 up_write(&slub_lock);
3507 up_write(&slub_lock);
3509 if (flags & SLAB_PANIC)
3510 panic("Cannot create slabcache %s\n", name);
3515 EXPORT_SYMBOL(kmem_cache_create);
3519 * Use the cpu notifier to insure that the cpu slabs are flushed when
3522 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3523 unsigned long action, void *hcpu)
3525 long cpu = (long)hcpu;
3526 struct kmem_cache *s;
3527 unsigned long flags;
3530 case CPU_UP_CANCELED:
3531 case CPU_UP_CANCELED_FROZEN:
3533 case CPU_DEAD_FROZEN:
3534 down_read(&slub_lock);
3535 list_for_each_entry(s, &slab_caches, list) {
3536 local_irq_save(flags);
3537 __flush_cpu_slab(s, cpu);
3538 local_irq_restore(flags);
3540 up_read(&slub_lock);
3548 static struct notifier_block __cpuinitdata slab_notifier = {
3549 .notifier_call = slab_cpuup_callback
3554 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3556 struct kmem_cache *s;
3559 if (unlikely(size > SLUB_MAX_SIZE))
3560 return kmalloc_large(size, gfpflags);
3562 s = get_slab(size, gfpflags);
3564 if (unlikely(ZERO_OR_NULL_PTR(s)))
3567 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3569 /* Honor the call site pointer we received. */
3570 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3576 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3577 int node, unsigned long caller)
3579 struct kmem_cache *s;
3582 if (unlikely(size > SLUB_MAX_SIZE)) {
3583 ret = kmalloc_large_node(size, gfpflags, node);
3585 trace_kmalloc_node(caller, ret,
3586 size, PAGE_SIZE << get_order(size),
3592 s = get_slab(size, gfpflags);
3594 if (unlikely(ZERO_OR_NULL_PTR(s)))
3597 ret = slab_alloc(s, gfpflags, node, caller);
3599 /* Honor the call site pointer we received. */
3600 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3607 static int count_inuse(struct page *page)
3612 static int count_total(struct page *page)
3614 return page->objects;
3618 #ifdef CONFIG_SLUB_DEBUG
3619 static int validate_slab(struct kmem_cache *s, struct page *page,
3623 void *addr = page_address(page);
3625 if (!check_slab(s, page) ||
3626 !on_freelist(s, page, NULL))
3629 /* Now we know that a valid freelist exists */
3630 bitmap_zero(map, page->objects);
3632 get_map(s, page, map);
3633 for_each_object(p, s, addr, page->objects) {
3634 if (test_bit(slab_index(p, s, addr), map))
3635 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3639 for_each_object(p, s, addr, page->objects)
3640 if (!test_bit(slab_index(p, s, addr), map))
3641 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3646 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3649 if (slab_trylock(page)) {
3650 validate_slab(s, page, map);
3653 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3657 static int validate_slab_node(struct kmem_cache *s,
3658 struct kmem_cache_node *n, unsigned long *map)
3660 unsigned long count = 0;
3662 unsigned long flags;
3664 spin_lock_irqsave(&n->list_lock, flags);
3666 list_for_each_entry(page, &n->partial, lru) {
3667 validate_slab_slab(s, page, map);
3670 if (count != n->nr_partial)
3671 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3672 "counter=%ld\n", s->name, count, n->nr_partial);
3674 if (!(s->flags & SLAB_STORE_USER))
3677 list_for_each_entry(page, &n->full, lru) {
3678 validate_slab_slab(s, page, map);
3681 if (count != atomic_long_read(&n->nr_slabs))
3682 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3683 "counter=%ld\n", s->name, count,
3684 atomic_long_read(&n->nr_slabs));
3687 spin_unlock_irqrestore(&n->list_lock, flags);
3691 static long validate_slab_cache(struct kmem_cache *s)
3694 unsigned long count = 0;
3695 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3696 sizeof(unsigned long), GFP_KERNEL);
3702 for_each_node_state(node, N_NORMAL_MEMORY) {
3703 struct kmem_cache_node *n = get_node(s, node);
3705 count += validate_slab_node(s, n, map);
3711 * Generate lists of code addresses where slabcache objects are allocated
3716 unsigned long count;
3723 DECLARE_BITMAP(cpus, NR_CPUS);
3729 unsigned long count;
3730 struct location *loc;
3733 static void free_loc_track(struct loc_track *t)
3736 free_pages((unsigned long)t->loc,
3737 get_order(sizeof(struct location) * t->max));
3740 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3745 order = get_order(sizeof(struct location) * max);
3747 l = (void *)__get_free_pages(flags, order);
3752 memcpy(l, t->loc, sizeof(struct location) * t->count);
3760 static int add_location(struct loc_track *t, struct kmem_cache *s,
3761 const struct track *track)
3763 long start, end, pos;
3765 unsigned long caddr;
3766 unsigned long age = jiffies - track->when;
3772 pos = start + (end - start + 1) / 2;
3775 * There is nothing at "end". If we end up there
3776 * we need to add something to before end.
3781 caddr = t->loc[pos].addr;
3782 if (track->addr == caddr) {
3788 if (age < l->min_time)
3790 if (age > l->max_time)
3793 if (track->pid < l->min_pid)
3794 l->min_pid = track->pid;
3795 if (track->pid > l->max_pid)
3796 l->max_pid = track->pid;
3798 cpumask_set_cpu(track->cpu,
3799 to_cpumask(l->cpus));
3801 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3805 if (track->addr < caddr)
3812 * Not found. Insert new tracking element.
3814 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3820 (t->count - pos) * sizeof(struct location));
3823 l->addr = track->addr;
3827 l->min_pid = track->pid;
3828 l->max_pid = track->pid;
3829 cpumask_clear(to_cpumask(l->cpus));
3830 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3831 nodes_clear(l->nodes);
3832 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3836 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3837 struct page *page, enum track_item alloc,
3840 void *addr = page_address(page);
3843 bitmap_zero(map, page->objects);
3844 get_map(s, page, map);
3846 for_each_object(p, s, addr, page->objects)
3847 if (!test_bit(slab_index(p, s, addr), map))
3848 add_location(t, s, get_track(s, p, alloc));
3851 static int list_locations(struct kmem_cache *s, char *buf,
3852 enum track_item alloc)
3856 struct loc_track t = { 0, 0, NULL };
3858 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3859 sizeof(unsigned long), GFP_KERNEL);
3861 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3864 return sprintf(buf, "Out of memory\n");
3866 /* Push back cpu slabs */
3869 for_each_node_state(node, N_NORMAL_MEMORY) {
3870 struct kmem_cache_node *n = get_node(s, node);
3871 unsigned long flags;
3874 if (!atomic_long_read(&n->nr_slabs))
3877 spin_lock_irqsave(&n->list_lock, flags);
3878 list_for_each_entry(page, &n->partial, lru)
3879 process_slab(&t, s, page, alloc, map);
3880 list_for_each_entry(page, &n->full, lru)
3881 process_slab(&t, s, page, alloc, map);
3882 spin_unlock_irqrestore(&n->list_lock, flags);
3885 for (i = 0; i < t.count; i++) {
3886 struct location *l = &t.loc[i];
3888 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3890 len += sprintf(buf + len, "%7ld ", l->count);
3893 len += sprintf(buf + len, "%pS", (void *)l->addr);
3895 len += sprintf(buf + len, "<not-available>");
3897 if (l->sum_time != l->min_time) {
3898 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3900 (long)div_u64(l->sum_time, l->count),
3903 len += sprintf(buf + len, " age=%ld",
3906 if (l->min_pid != l->max_pid)
3907 len += sprintf(buf + len, " pid=%ld-%ld",
3908 l->min_pid, l->max_pid);
3910 len += sprintf(buf + len, " pid=%ld",
3913 if (num_online_cpus() > 1 &&
3914 !cpumask_empty(to_cpumask(l->cpus)) &&
3915 len < PAGE_SIZE - 60) {
3916 len += sprintf(buf + len, " cpus=");
3917 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3918 to_cpumask(l->cpus));
3921 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3922 len < PAGE_SIZE - 60) {
3923 len += sprintf(buf + len, " nodes=");
3924 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3928 len += sprintf(buf + len, "\n");
3934 len += sprintf(buf, "No data\n");
3939 #ifdef SLUB_RESILIENCY_TEST
3940 static void resiliency_test(void)
3944 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3946 printk(KERN_ERR "SLUB resiliency testing\n");
3947 printk(KERN_ERR "-----------------------\n");
3948 printk(KERN_ERR "A. Corruption after allocation\n");
3950 p = kzalloc(16, GFP_KERNEL);
3952 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3953 " 0x12->0x%p\n\n", p + 16);
3955 validate_slab_cache(kmalloc_caches[4]);
3957 /* Hmmm... The next two are dangerous */
3958 p = kzalloc(32, GFP_KERNEL);
3959 p[32 + sizeof(void *)] = 0x34;
3960 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3961 " 0x34 -> -0x%p\n", p);
3963 "If allocated object is overwritten then not detectable\n\n");
3965 validate_slab_cache(kmalloc_caches[5]);
3966 p = kzalloc(64, GFP_KERNEL);
3967 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3969 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3972 "If allocated object is overwritten then not detectable\n\n");
3973 validate_slab_cache(kmalloc_caches[6]);
3975 printk(KERN_ERR "\nB. Corruption after free\n");
3976 p = kzalloc(128, GFP_KERNEL);
3979 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3980 validate_slab_cache(kmalloc_caches[7]);
3982 p = kzalloc(256, GFP_KERNEL);
3985 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3987 validate_slab_cache(kmalloc_caches[8]);
3989 p = kzalloc(512, GFP_KERNEL);
3992 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3993 validate_slab_cache(kmalloc_caches[9]);
3997 static void resiliency_test(void) {};
4002 enum slab_stat_type {
4003 SL_ALL, /* All slabs */
4004 SL_PARTIAL, /* Only partially allocated slabs */
4005 SL_CPU, /* Only slabs used for cpu caches */
4006 SL_OBJECTS, /* Determine allocated objects not slabs */
4007 SL_TOTAL /* Determine object capacity not slabs */
4010 #define SO_ALL (1 << SL_ALL)
4011 #define SO_PARTIAL (1 << SL_PARTIAL)
4012 #define SO_CPU (1 << SL_CPU)
4013 #define SO_OBJECTS (1 << SL_OBJECTS)
4014 #define SO_TOTAL (1 << SL_TOTAL)
4016 static ssize_t show_slab_objects(struct kmem_cache *s,
4017 char *buf, unsigned long flags)
4019 unsigned long total = 0;
4022 unsigned long *nodes;
4023 unsigned long *per_cpu;
4025 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4028 per_cpu = nodes + nr_node_ids;
4030 if (flags & SO_CPU) {
4033 for_each_possible_cpu(cpu) {
4034 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4036 if (!c || c->node < 0)
4040 if (flags & SO_TOTAL)
4041 x = c->page->objects;
4042 else if (flags & SO_OBJECTS)
4048 nodes[c->node] += x;
4054 lock_memory_hotplug();
4055 #ifdef CONFIG_SLUB_DEBUG
4056 if (flags & SO_ALL) {
4057 for_each_node_state(node, N_NORMAL_MEMORY) {
4058 struct kmem_cache_node *n = get_node(s, node);
4060 if (flags & SO_TOTAL)
4061 x = atomic_long_read(&n->total_objects);
4062 else if (flags & SO_OBJECTS)
4063 x = atomic_long_read(&n->total_objects) -
4064 count_partial(n, count_free);
4067 x = atomic_long_read(&n->nr_slabs);
4074 if (flags & SO_PARTIAL) {
4075 for_each_node_state(node, N_NORMAL_MEMORY) {
4076 struct kmem_cache_node *n = get_node(s, node);
4078 if (flags & SO_TOTAL)
4079 x = count_partial(n, count_total);
4080 else if (flags & SO_OBJECTS)
4081 x = count_partial(n, count_inuse);
4088 x = sprintf(buf, "%lu", total);
4090 for_each_node_state(node, N_NORMAL_MEMORY)
4092 x += sprintf(buf + x, " N%d=%lu",
4095 unlock_memory_hotplug();
4097 return x + sprintf(buf + x, "\n");
4100 #ifdef CONFIG_SLUB_DEBUG
4101 static int any_slab_objects(struct kmem_cache *s)
4105 for_each_online_node(node) {
4106 struct kmem_cache_node *n = get_node(s, node);
4111 if (atomic_long_read(&n->total_objects))
4118 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4119 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4121 struct slab_attribute {
4122 struct attribute attr;
4123 ssize_t (*show)(struct kmem_cache *s, char *buf);
4124 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4127 #define SLAB_ATTR_RO(_name) \
4128 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4130 #define SLAB_ATTR(_name) \
4131 static struct slab_attribute _name##_attr = \
4132 __ATTR(_name, 0644, _name##_show, _name##_store)
4134 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4136 return sprintf(buf, "%d\n", s->size);
4138 SLAB_ATTR_RO(slab_size);
4140 static ssize_t align_show(struct kmem_cache *s, char *buf)
4142 return sprintf(buf, "%d\n", s->align);
4144 SLAB_ATTR_RO(align);
4146 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4148 return sprintf(buf, "%d\n", s->objsize);
4150 SLAB_ATTR_RO(object_size);
4152 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4154 return sprintf(buf, "%d\n", oo_objects(s->oo));
4156 SLAB_ATTR_RO(objs_per_slab);
4158 static ssize_t order_store(struct kmem_cache *s,
4159 const char *buf, size_t length)
4161 unsigned long order;
4164 err = strict_strtoul(buf, 10, &order);
4168 if (order > slub_max_order || order < slub_min_order)
4171 calculate_sizes(s, order);
4175 static ssize_t order_show(struct kmem_cache *s, char *buf)
4177 return sprintf(buf, "%d\n", oo_order(s->oo));
4181 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4183 return sprintf(buf, "%lu\n", s->min_partial);
4186 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4192 err = strict_strtoul(buf, 10, &min);
4196 set_min_partial(s, min);
4199 SLAB_ATTR(min_partial);
4201 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4205 return sprintf(buf, "%pS\n", s->ctor);
4209 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4211 return sprintf(buf, "%d\n", s->refcount - 1);
4213 SLAB_ATTR_RO(aliases);
4215 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4217 return show_slab_objects(s, buf, SO_PARTIAL);
4219 SLAB_ATTR_RO(partial);
4221 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4223 return show_slab_objects(s, buf, SO_CPU);
4225 SLAB_ATTR_RO(cpu_slabs);
4227 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4229 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4231 SLAB_ATTR_RO(objects);
4233 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4235 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4237 SLAB_ATTR_RO(objects_partial);
4239 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4241 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4244 static ssize_t reclaim_account_store(struct kmem_cache *s,
4245 const char *buf, size_t length)
4247 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4249 s->flags |= SLAB_RECLAIM_ACCOUNT;
4252 SLAB_ATTR(reclaim_account);
4254 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4256 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4258 SLAB_ATTR_RO(hwcache_align);
4260 #ifdef CONFIG_ZONE_DMA
4261 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4263 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4265 SLAB_ATTR_RO(cache_dma);
4268 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4270 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4272 SLAB_ATTR_RO(destroy_by_rcu);
4274 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4276 return sprintf(buf, "%d\n", s->reserved);
4278 SLAB_ATTR_RO(reserved);
4280 #ifdef CONFIG_SLUB_DEBUG
4281 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4283 return show_slab_objects(s, buf, SO_ALL);
4285 SLAB_ATTR_RO(slabs);
4287 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4289 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4291 SLAB_ATTR_RO(total_objects);
4293 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4295 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4298 static ssize_t sanity_checks_store(struct kmem_cache *s,
4299 const char *buf, size_t length)
4301 s->flags &= ~SLAB_DEBUG_FREE;
4302 if (buf[0] == '1') {
4303 s->flags &= ~__CMPXCHG_DOUBLE;
4304 s->flags |= SLAB_DEBUG_FREE;
4308 SLAB_ATTR(sanity_checks);
4310 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4312 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4315 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4318 s->flags &= ~SLAB_TRACE;
4319 if (buf[0] == '1') {
4320 s->flags &= ~__CMPXCHG_DOUBLE;
4321 s->flags |= SLAB_TRACE;
4327 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4329 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4332 static ssize_t red_zone_store(struct kmem_cache *s,
4333 const char *buf, size_t length)
4335 if (any_slab_objects(s))
4338 s->flags &= ~SLAB_RED_ZONE;
4339 if (buf[0] == '1') {
4340 s->flags &= ~__CMPXCHG_DOUBLE;
4341 s->flags |= SLAB_RED_ZONE;
4343 calculate_sizes(s, -1);
4346 SLAB_ATTR(red_zone);
4348 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4350 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4353 static ssize_t poison_store(struct kmem_cache *s,
4354 const char *buf, size_t length)
4356 if (any_slab_objects(s))
4359 s->flags &= ~SLAB_POISON;
4360 if (buf[0] == '1') {
4361 s->flags &= ~__CMPXCHG_DOUBLE;
4362 s->flags |= SLAB_POISON;
4364 calculate_sizes(s, -1);
4369 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4371 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4374 static ssize_t store_user_store(struct kmem_cache *s,
4375 const char *buf, size_t length)
4377 if (any_slab_objects(s))
4380 s->flags &= ~SLAB_STORE_USER;
4381 if (buf[0] == '1') {
4382 s->flags &= ~__CMPXCHG_DOUBLE;
4383 s->flags |= SLAB_STORE_USER;
4385 calculate_sizes(s, -1);
4388 SLAB_ATTR(store_user);
4390 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4395 static ssize_t validate_store(struct kmem_cache *s,
4396 const char *buf, size_t length)
4400 if (buf[0] == '1') {
4401 ret = validate_slab_cache(s);
4407 SLAB_ATTR(validate);
4409 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4411 if (!(s->flags & SLAB_STORE_USER))
4413 return list_locations(s, buf, TRACK_ALLOC);
4415 SLAB_ATTR_RO(alloc_calls);
4417 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4419 if (!(s->flags & SLAB_STORE_USER))
4421 return list_locations(s, buf, TRACK_FREE);
4423 SLAB_ATTR_RO(free_calls);
4424 #endif /* CONFIG_SLUB_DEBUG */
4426 #ifdef CONFIG_FAILSLAB
4427 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4429 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4432 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4435 s->flags &= ~SLAB_FAILSLAB;
4437 s->flags |= SLAB_FAILSLAB;
4440 SLAB_ATTR(failslab);
4443 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4448 static ssize_t shrink_store(struct kmem_cache *s,
4449 const char *buf, size_t length)
4451 if (buf[0] == '1') {
4452 int rc = kmem_cache_shrink(s);
4463 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4465 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4468 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4469 const char *buf, size_t length)
4471 unsigned long ratio;
4474 err = strict_strtoul(buf, 10, &ratio);
4479 s->remote_node_defrag_ratio = ratio * 10;
4483 SLAB_ATTR(remote_node_defrag_ratio);
4486 #ifdef CONFIG_SLUB_STATS
4487 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4489 unsigned long sum = 0;
4492 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4497 for_each_online_cpu(cpu) {
4498 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4504 len = sprintf(buf, "%lu", sum);
4507 for_each_online_cpu(cpu) {
4508 if (data[cpu] && len < PAGE_SIZE - 20)
4509 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4513 return len + sprintf(buf + len, "\n");
4516 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4520 for_each_online_cpu(cpu)
4521 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4524 #define STAT_ATTR(si, text) \
4525 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4527 return show_stat(s, buf, si); \
4529 static ssize_t text##_store(struct kmem_cache *s, \
4530 const char *buf, size_t length) \
4532 if (buf[0] != '0') \
4534 clear_stat(s, si); \
4539 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4540 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4541 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4542 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4543 STAT_ATTR(FREE_FROZEN, free_frozen);
4544 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4545 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4546 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4547 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4548 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4549 STAT_ATTR(FREE_SLAB, free_slab);
4550 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4551 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4552 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4553 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4554 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4555 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4556 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4557 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4558 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4561 static struct attribute *slab_attrs[] = {
4562 &slab_size_attr.attr,
4563 &object_size_attr.attr,
4564 &objs_per_slab_attr.attr,
4566 &min_partial_attr.attr,
4568 &objects_partial_attr.attr,
4570 &cpu_slabs_attr.attr,
4574 &hwcache_align_attr.attr,
4575 &reclaim_account_attr.attr,
4576 &destroy_by_rcu_attr.attr,
4578 &reserved_attr.attr,
4579 #ifdef CONFIG_SLUB_DEBUG
4580 &total_objects_attr.attr,
4582 &sanity_checks_attr.attr,
4584 &red_zone_attr.attr,
4586 &store_user_attr.attr,
4587 &validate_attr.attr,
4588 &alloc_calls_attr.attr,
4589 &free_calls_attr.attr,
4591 #ifdef CONFIG_ZONE_DMA
4592 &cache_dma_attr.attr,
4595 &remote_node_defrag_ratio_attr.attr,
4597 #ifdef CONFIG_SLUB_STATS
4598 &alloc_fastpath_attr.attr,
4599 &alloc_slowpath_attr.attr,
4600 &free_fastpath_attr.attr,
4601 &free_slowpath_attr.attr,
4602 &free_frozen_attr.attr,
4603 &free_add_partial_attr.attr,
4604 &free_remove_partial_attr.attr,
4605 &alloc_from_partial_attr.attr,
4606 &alloc_slab_attr.attr,
4607 &alloc_refill_attr.attr,
4608 &free_slab_attr.attr,
4609 &cpuslab_flush_attr.attr,
4610 &deactivate_full_attr.attr,
4611 &deactivate_empty_attr.attr,
4612 &deactivate_to_head_attr.attr,
4613 &deactivate_to_tail_attr.attr,
4614 &deactivate_remote_frees_attr.attr,
4615 &order_fallback_attr.attr,
4616 &cmpxchg_double_fail_attr.attr,
4617 &cmpxchg_double_cpu_fail_attr.attr,
4619 #ifdef CONFIG_FAILSLAB
4620 &failslab_attr.attr,
4626 static struct attribute_group slab_attr_group = {
4627 .attrs = slab_attrs,
4630 static ssize_t slab_attr_show(struct kobject *kobj,
4631 struct attribute *attr,
4634 struct slab_attribute *attribute;
4635 struct kmem_cache *s;
4638 attribute = to_slab_attr(attr);
4641 if (!attribute->show)
4644 err = attribute->show(s, buf);
4649 static ssize_t slab_attr_store(struct kobject *kobj,
4650 struct attribute *attr,
4651 const char *buf, size_t len)
4653 struct slab_attribute *attribute;
4654 struct kmem_cache *s;
4657 attribute = to_slab_attr(attr);
4660 if (!attribute->store)
4663 err = attribute->store(s, buf, len);
4668 static void kmem_cache_release(struct kobject *kobj)
4670 struct kmem_cache *s = to_slab(kobj);
4676 static const struct sysfs_ops slab_sysfs_ops = {
4677 .show = slab_attr_show,
4678 .store = slab_attr_store,
4681 static struct kobj_type slab_ktype = {
4682 .sysfs_ops = &slab_sysfs_ops,
4683 .release = kmem_cache_release
4686 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4688 struct kobj_type *ktype = get_ktype(kobj);
4690 if (ktype == &slab_ktype)
4695 static const struct kset_uevent_ops slab_uevent_ops = {
4696 .filter = uevent_filter,
4699 static struct kset *slab_kset;
4701 #define ID_STR_LENGTH 64
4703 /* Create a unique string id for a slab cache:
4705 * Format :[flags-]size
4707 static char *create_unique_id(struct kmem_cache *s)
4709 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4716 * First flags affecting slabcache operations. We will only
4717 * get here for aliasable slabs so we do not need to support
4718 * too many flags. The flags here must cover all flags that
4719 * are matched during merging to guarantee that the id is
4722 if (s->flags & SLAB_CACHE_DMA)
4724 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4726 if (s->flags & SLAB_DEBUG_FREE)
4728 if (!(s->flags & SLAB_NOTRACK))
4732 p += sprintf(p, "%07d", s->size);
4733 BUG_ON(p > name + ID_STR_LENGTH - 1);
4737 static int sysfs_slab_add(struct kmem_cache *s)
4743 if (slab_state < SYSFS)
4744 /* Defer until later */
4747 unmergeable = slab_unmergeable(s);
4750 * Slabcache can never be merged so we can use the name proper.
4751 * This is typically the case for debug situations. In that
4752 * case we can catch duplicate names easily.
4754 sysfs_remove_link(&slab_kset->kobj, s->name);
4758 * Create a unique name for the slab as a target
4761 name = create_unique_id(s);
4764 s->kobj.kset = slab_kset;
4765 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4767 kobject_put(&s->kobj);
4771 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4773 kobject_del(&s->kobj);
4774 kobject_put(&s->kobj);
4777 kobject_uevent(&s->kobj, KOBJ_ADD);
4779 /* Setup first alias */
4780 sysfs_slab_alias(s, s->name);
4786 static void sysfs_slab_remove(struct kmem_cache *s)
4788 if (slab_state < SYSFS)
4790 * Sysfs has not been setup yet so no need to remove the
4795 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4796 kobject_del(&s->kobj);
4797 kobject_put(&s->kobj);
4801 * Need to buffer aliases during bootup until sysfs becomes
4802 * available lest we lose that information.
4804 struct saved_alias {
4805 struct kmem_cache *s;
4807 struct saved_alias *next;
4810 static struct saved_alias *alias_list;
4812 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4814 struct saved_alias *al;
4816 if (slab_state == SYSFS) {
4818 * If we have a leftover link then remove it.
4820 sysfs_remove_link(&slab_kset->kobj, name);
4821 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4824 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4830 al->next = alias_list;
4835 static int __init slab_sysfs_init(void)
4837 struct kmem_cache *s;
4840 down_write(&slub_lock);
4842 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4844 up_write(&slub_lock);
4845 printk(KERN_ERR "Cannot register slab subsystem.\n");
4851 list_for_each_entry(s, &slab_caches, list) {
4852 err = sysfs_slab_add(s);
4854 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4855 " to sysfs\n", s->name);
4858 while (alias_list) {
4859 struct saved_alias *al = alias_list;
4861 alias_list = alias_list->next;
4862 err = sysfs_slab_alias(al->s, al->name);
4864 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4865 " %s to sysfs\n", s->name);
4869 up_write(&slub_lock);
4874 __initcall(slab_sysfs_init);
4875 #endif /* CONFIG_SYSFS */
4878 * The /proc/slabinfo ABI
4880 #ifdef CONFIG_SLABINFO
4881 static void print_slabinfo_header(struct seq_file *m)
4883 seq_puts(m, "slabinfo - version: 2.1\n");
4884 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4885 "<objperslab> <pagesperslab>");
4886 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4887 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4891 static void *s_start(struct seq_file *m, loff_t *pos)
4895 down_read(&slub_lock);
4897 print_slabinfo_header(m);
4899 return seq_list_start(&slab_caches, *pos);
4902 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4904 return seq_list_next(p, &slab_caches, pos);
4907 static void s_stop(struct seq_file *m, void *p)
4909 up_read(&slub_lock);
4912 static int s_show(struct seq_file *m, void *p)
4914 unsigned long nr_partials = 0;
4915 unsigned long nr_slabs = 0;
4916 unsigned long nr_inuse = 0;
4917 unsigned long nr_objs = 0;
4918 unsigned long nr_free = 0;
4919 struct kmem_cache *s;
4922 s = list_entry(p, struct kmem_cache, list);
4924 for_each_online_node(node) {
4925 struct kmem_cache_node *n = get_node(s, node);
4930 nr_partials += n->nr_partial;
4931 nr_slabs += atomic_long_read(&n->nr_slabs);
4932 nr_objs += atomic_long_read(&n->total_objects);
4933 nr_free += count_partial(n, count_free);
4936 nr_inuse = nr_objs - nr_free;
4938 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4939 nr_objs, s->size, oo_objects(s->oo),
4940 (1 << oo_order(s->oo)));
4941 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4942 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4948 static const struct seq_operations slabinfo_op = {
4955 static int slabinfo_open(struct inode *inode, struct file *file)
4957 return seq_open(file, &slabinfo_op);
4960 static const struct file_operations proc_slabinfo_operations = {
4961 .open = slabinfo_open,
4963 .llseek = seq_lseek,
4964 .release = seq_release,
4967 static int __init slab_proc_init(void)
4969 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4972 module_init(slab_proc_init);
4973 #endif /* CONFIG_SLABINFO */