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 or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
32 #include <trace/events/kmem.h>
36 * 1. slub_lock (Global Semaphore)
38 * 3. slab_lock(page) (Only on some arches and for debugging)
42 * The role of the slub_lock is to protect the list of all the slabs
43 * and to synchronize major metadata changes to slab cache structures.
45 * The slab_lock is only used for debugging and on arches that do not
46 * have the ability to do a cmpxchg_double. It only protects the second
47 * double word in the page struct. Meaning
48 * A. page->freelist -> List of object free in a page
49 * B. page->counters -> Counters of objects
50 * C. page->frozen -> frozen state
52 * If a slab is frozen then it is exempt from list management. It is not
53 * on any list. The processor that froze the slab is the one who can
54 * perform list operations on the page. Other processors may put objects
55 * onto the freelist but the processor that froze the slab is the only
56 * one that can retrieve the objects from the page's freelist.
58 * The list_lock protects the partial and full list on each node and
59 * the partial slab counter. If taken then no new slabs may be added or
60 * removed from the lists nor make the number of partial slabs be modified.
61 * (Note that the total number of slabs is an atomic value that may be
62 * modified without taking the list lock).
64 * The list_lock is a centralized lock and thus we avoid taking it as
65 * much as possible. As long as SLUB does not have to handle partial
66 * slabs, operations can continue without any centralized lock. F.e.
67 * allocating a long series of objects that fill up slabs does not require
69 * Interrupts are disabled during allocation and deallocation in order to
70 * make the slab allocator safe to use in the context of an irq. In addition
71 * interrupts are disabled to ensure that the processor does not change
72 * while handling per_cpu slabs, due to kernel preemption.
74 * SLUB assigns one slab for allocation to each processor.
75 * Allocations only occur from these slabs called cpu slabs.
77 * Slabs with free elements are kept on a partial list and during regular
78 * operations no list for full slabs is used. If an object in a full slab is
79 * freed then the slab will show up again on the partial lists.
80 * We track full slabs for debugging purposes though because otherwise we
81 * cannot scan all objects.
83 * Slabs are freed when they become empty. Teardown and setup is
84 * minimal so we rely on the page allocators per cpu caches for
85 * fast frees and allocs.
87 * Overloading of page flags that are otherwise used for LRU management.
89 * PageActive The slab is frozen and exempt from list processing.
90 * This means that the slab is dedicated to a purpose
91 * such as satisfying allocations for a specific
92 * processor. Objects may be freed in the slab while
93 * it is frozen but slab_free will then skip the usual
94 * list operations. It is up to the processor holding
95 * the slab to integrate the slab into the slab lists
96 * when the slab is no longer needed.
98 * One use of this flag is to mark slabs that are
99 * used for allocations. Then such a slab becomes a cpu
100 * slab. The cpu slab may be equipped with an additional
101 * freelist that allows lockless access to
102 * free objects in addition to the regular freelist
103 * that requires the slab lock.
105 * PageError Slab requires special handling due to debug
106 * options set. This moves slab handling out of
107 * the fast path and disables lockless freelists.
110 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
111 SLAB_TRACE | SLAB_DEBUG_FREE)
113 static inline int kmem_cache_debug(struct kmem_cache *s)
115 #ifdef CONFIG_SLUB_DEBUG
116 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
123 * Issues still to be resolved:
125 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
127 * - Variable sizing of the per node arrays
130 /* Enable to test recovery from slab corruption on boot */
131 #undef SLUB_RESILIENCY_TEST
133 /* Enable to log cmpxchg failures */
134 #undef SLUB_DEBUG_CMPXCHG
137 * Mininum number of partial slabs. These will be left on the partial
138 * lists even if they are empty. kmem_cache_shrink may reclaim them.
140 #define MIN_PARTIAL 5
143 * Maximum number of desirable partial slabs.
144 * The existence of more partial slabs makes kmem_cache_shrink
145 * sort the partial list by the number of objects in the.
147 #define MAX_PARTIAL 10
149 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
150 SLAB_POISON | SLAB_STORE_USER)
153 * Debugging flags that require metadata to be stored in the slab. These get
154 * disabled when slub_debug=O is used and a cache's min order increases with
157 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
160 * Set of flags that will prevent slab merging
162 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
163 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
166 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
167 SLAB_CACHE_DMA | SLAB_NOTRACK)
170 #define OO_MASK ((1 << OO_SHIFT) - 1)
171 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
173 /* Internal SLUB flags */
174 #define __OBJECT_POISON 0x80000000UL /* Poison object */
175 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
177 static int kmem_size = sizeof(struct kmem_cache);
180 static struct notifier_block slab_notifier;
184 DOWN, /* No slab functionality available */
185 PARTIAL, /* Kmem_cache_node works */
186 UP, /* Everything works but does not show up in sysfs */
190 /* A list of all slab caches on the system */
191 static DECLARE_RWSEM(slub_lock);
192 static LIST_HEAD(slab_caches);
195 * Tracking user of a slab.
198 unsigned long addr; /* Called from address */
199 int cpu; /* Was running on cpu */
200 int pid; /* Pid context */
201 unsigned long when; /* When did the operation occur */
204 enum track_item { TRACK_ALLOC, TRACK_FREE };
207 static int sysfs_slab_add(struct kmem_cache *);
208 static int sysfs_slab_alias(struct kmem_cache *, const char *);
209 static void sysfs_slab_remove(struct kmem_cache *);
212 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
213 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
215 static inline void sysfs_slab_remove(struct kmem_cache *s)
223 static inline void stat(const struct kmem_cache *s, enum stat_item si)
225 #ifdef CONFIG_SLUB_STATS
226 __this_cpu_inc(s->cpu_slab->stat[si]);
230 /********************************************************************
231 * Core slab cache functions
232 *******************************************************************/
234 int slab_is_available(void)
236 return slab_state >= UP;
239 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
241 return s->node[node];
244 /* Verify that a pointer has an address that is valid within a slab page */
245 static inline int check_valid_pointer(struct kmem_cache *s,
246 struct page *page, const void *object)
253 base = page_address(page);
254 if (object < base || object >= base + page->objects * s->size ||
255 (object - base) % s->size) {
262 static inline void *get_freepointer(struct kmem_cache *s, void *object)
264 return *(void **)(object + s->offset);
267 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
271 #ifdef CONFIG_DEBUG_PAGEALLOC
272 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
274 p = get_freepointer(s, object);
279 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
281 *(void **)(object + s->offset) = fp;
284 /* Loop over all objects in a slab */
285 #define for_each_object(__p, __s, __addr, __objects) \
286 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
289 /* Determine object index from a given position */
290 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292 return (p - addr) / s->size;
295 static inline size_t slab_ksize(const struct kmem_cache *s)
297 #ifdef CONFIG_SLUB_DEBUG
299 * Debugging requires use of the padding between object
300 * and whatever may come after it.
302 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
307 * If we have the need to store the freelist pointer
308 * back there or track user information then we can
309 * only use the space before that information.
311 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
314 * Else we can use all the padding etc for the allocation
319 static inline int order_objects(int order, unsigned long size, int reserved)
321 return ((PAGE_SIZE << order) - reserved) / size;
324 static inline struct kmem_cache_order_objects oo_make(int order,
325 unsigned long size, int reserved)
327 struct kmem_cache_order_objects x = {
328 (order << OO_SHIFT) + order_objects(order, size, reserved)
334 static inline int oo_order(struct kmem_cache_order_objects x)
336 return x.x >> OO_SHIFT;
339 static inline int oo_objects(struct kmem_cache_order_objects x)
341 return x.x & OO_MASK;
345 * Per slab locking using the pagelock
347 static __always_inline void slab_lock(struct page *page)
349 bit_spin_lock(PG_locked, &page->flags);
352 static __always_inline void slab_unlock(struct page *page)
354 __bit_spin_unlock(PG_locked, &page->flags);
357 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
358 void *freelist_old, unsigned long counters_old,
359 void *freelist_new, unsigned long counters_new,
362 #ifdef CONFIG_CMPXCHG_DOUBLE
363 if (s->flags & __CMPXCHG_DOUBLE) {
364 if (cmpxchg_double(&page->freelist,
365 freelist_old, counters_old,
366 freelist_new, counters_new))
372 if (page->freelist == freelist_old && page->counters == counters_old) {
373 page->freelist = freelist_new;
374 page->counters = counters_new;
382 stat(s, CMPXCHG_DOUBLE_FAIL);
384 #ifdef SLUB_DEBUG_CMPXCHG
385 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
391 #ifdef CONFIG_SLUB_DEBUG
393 * Determine a map of object in use on a page.
395 * Node listlock must be held to guarantee that the page does
396 * not vanish from under us.
398 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
401 void *addr = page_address(page);
403 for (p = page->freelist; p; p = get_freepointer(s, p))
404 set_bit(slab_index(p, s, addr), map);
410 #ifdef CONFIG_SLUB_DEBUG_ON
411 static int slub_debug = DEBUG_DEFAULT_FLAGS;
413 static int slub_debug;
416 static char *slub_debug_slabs;
417 static int disable_higher_order_debug;
422 static void print_section(char *text, u8 *addr, unsigned int length)
430 for (i = 0; i < length; i++) {
432 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
435 printk(KERN_CONT " %02x", addr[i]);
437 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
439 printk(KERN_CONT " %s\n", ascii);
446 printk(KERN_CONT " ");
450 printk(KERN_CONT " %s\n", ascii);
454 static struct track *get_track(struct kmem_cache *s, void *object,
455 enum track_item alloc)
460 p = object + s->offset + sizeof(void *);
462 p = object + s->inuse;
467 static void set_track(struct kmem_cache *s, void *object,
468 enum track_item alloc, unsigned long addr)
470 struct track *p = get_track(s, object, alloc);
474 p->cpu = smp_processor_id();
475 p->pid = current->pid;
478 memset(p, 0, sizeof(struct track));
481 static void init_tracking(struct kmem_cache *s, void *object)
483 if (!(s->flags & SLAB_STORE_USER))
486 set_track(s, object, TRACK_FREE, 0UL);
487 set_track(s, object, TRACK_ALLOC, 0UL);
490 static void print_track(const char *s, struct track *t)
495 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
496 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
499 static void print_tracking(struct kmem_cache *s, void *object)
501 if (!(s->flags & SLAB_STORE_USER))
504 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
505 print_track("Freed", get_track(s, object, TRACK_FREE));
508 static void print_page_info(struct page *page)
510 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
511 page, page->objects, page->inuse, page->freelist, page->flags);
515 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
521 vsnprintf(buf, sizeof(buf), fmt, args);
523 printk(KERN_ERR "========================================"
524 "=====================================\n");
525 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
526 printk(KERN_ERR "----------------------------------------"
527 "-------------------------------------\n\n");
530 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
536 vsnprintf(buf, sizeof(buf), fmt, args);
538 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
541 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
543 unsigned int off; /* Offset of last byte */
544 u8 *addr = page_address(page);
546 print_tracking(s, p);
548 print_page_info(page);
550 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
551 p, p - addr, get_freepointer(s, p));
554 print_section("Bytes b4", p - 16, 16);
556 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
558 if (s->flags & SLAB_RED_ZONE)
559 print_section("Redzone", p + s->objsize,
560 s->inuse - s->objsize);
563 off = s->offset + sizeof(void *);
567 if (s->flags & SLAB_STORE_USER)
568 off += 2 * sizeof(struct track);
571 /* Beginning of the filler is the free pointer */
572 print_section("Padding", p + off, s->size - off);
577 static void object_err(struct kmem_cache *s, struct page *page,
578 u8 *object, char *reason)
580 slab_bug(s, "%s", reason);
581 print_trailer(s, page, object);
584 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
590 vsnprintf(buf, sizeof(buf), fmt, args);
592 slab_bug(s, "%s", buf);
593 print_page_info(page);
597 static void init_object(struct kmem_cache *s, void *object, u8 val)
601 if (s->flags & __OBJECT_POISON) {
602 memset(p, POISON_FREE, s->objsize - 1);
603 p[s->objsize - 1] = POISON_END;
606 if (s->flags & SLAB_RED_ZONE)
607 memset(p + s->objsize, val, s->inuse - s->objsize);
610 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
613 if (*start != (u8)value)
621 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
622 void *from, void *to)
624 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
625 memset(from, data, to - from);
628 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
629 u8 *object, char *what,
630 u8 *start, unsigned int value, unsigned int bytes)
635 fault = check_bytes(start, value, bytes);
640 while (end > fault && end[-1] == value)
643 slab_bug(s, "%s overwritten", what);
644 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
645 fault, end - 1, fault[0], value);
646 print_trailer(s, page, object);
648 restore_bytes(s, what, value, fault, end);
656 * Bytes of the object to be managed.
657 * If the freepointer may overlay the object then the free
658 * pointer is the first word of the object.
660 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
663 * object + s->objsize
664 * Padding to reach word boundary. This is also used for Redzoning.
665 * Padding is extended by another word if Redzoning is enabled and
668 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
669 * 0xcc (RED_ACTIVE) for objects in use.
672 * Meta data starts here.
674 * A. Free pointer (if we cannot overwrite object on free)
675 * B. Tracking data for SLAB_STORE_USER
676 * C. Padding to reach required alignment boundary or at mininum
677 * one word if debugging is on to be able to detect writes
678 * before the word boundary.
680 * Padding is done using 0x5a (POISON_INUSE)
683 * Nothing is used beyond s->size.
685 * If slabcaches are merged then the objsize and inuse boundaries are mostly
686 * ignored. And therefore no slab options that rely on these boundaries
687 * may be used with merged slabcaches.
690 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
692 unsigned long off = s->inuse; /* The end of info */
695 /* Freepointer is placed after the object. */
696 off += sizeof(void *);
698 if (s->flags & SLAB_STORE_USER)
699 /* We also have user information there */
700 off += 2 * sizeof(struct track);
705 return check_bytes_and_report(s, page, p, "Object padding",
706 p + off, POISON_INUSE, s->size - off);
709 /* Check the pad bytes at the end of a slab page */
710 static int slab_pad_check(struct kmem_cache *s, struct page *page)
718 if (!(s->flags & SLAB_POISON))
721 start = page_address(page);
722 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
723 end = start + length;
724 remainder = length % s->size;
728 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
731 while (end > fault && end[-1] == POISON_INUSE)
734 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
735 print_section("Padding", end - remainder, remainder);
737 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
741 static int check_object(struct kmem_cache *s, struct page *page,
742 void *object, u8 val)
745 u8 *endobject = object + s->objsize;
747 if (s->flags & SLAB_RED_ZONE) {
748 if (!check_bytes_and_report(s, page, object, "Redzone",
749 endobject, val, s->inuse - s->objsize))
752 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
753 check_bytes_and_report(s, page, p, "Alignment padding",
754 endobject, POISON_INUSE, s->inuse - s->objsize);
758 if (s->flags & SLAB_POISON) {
759 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
760 (!check_bytes_and_report(s, page, p, "Poison", p,
761 POISON_FREE, s->objsize - 1) ||
762 !check_bytes_and_report(s, page, p, "Poison",
763 p + s->objsize - 1, POISON_END, 1)))
766 * check_pad_bytes cleans up on its own.
768 check_pad_bytes(s, page, p);
771 if (!s->offset && val == SLUB_RED_ACTIVE)
773 * Object and freepointer overlap. Cannot check
774 * freepointer while object is allocated.
778 /* Check free pointer validity */
779 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
780 object_err(s, page, p, "Freepointer corrupt");
782 * No choice but to zap it and thus lose the remainder
783 * of the free objects in this slab. May cause
784 * another error because the object count is now wrong.
786 set_freepointer(s, p, NULL);
792 static int check_slab(struct kmem_cache *s, struct page *page)
796 VM_BUG_ON(!irqs_disabled());
798 if (!PageSlab(page)) {
799 slab_err(s, page, "Not a valid slab page");
803 maxobj = order_objects(compound_order(page), s->size, s->reserved);
804 if (page->objects > maxobj) {
805 slab_err(s, page, "objects %u > max %u",
806 s->name, page->objects, maxobj);
809 if (page->inuse > page->objects) {
810 slab_err(s, page, "inuse %u > max %u",
811 s->name, page->inuse, page->objects);
814 /* Slab_pad_check fixes things up after itself */
815 slab_pad_check(s, page);
820 * Determine if a certain object on a page is on the freelist. Must hold the
821 * slab lock to guarantee that the chains are in a consistent state.
823 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
828 unsigned long max_objects;
831 while (fp && nr <= page->objects) {
834 if (!check_valid_pointer(s, page, fp)) {
836 object_err(s, page, object,
837 "Freechain corrupt");
838 set_freepointer(s, object, NULL);
841 slab_err(s, page, "Freepointer corrupt");
842 page->freelist = NULL;
843 page->inuse = page->objects;
844 slab_fix(s, "Freelist cleared");
850 fp = get_freepointer(s, object);
854 max_objects = order_objects(compound_order(page), s->size, s->reserved);
855 if (max_objects > MAX_OBJS_PER_PAGE)
856 max_objects = MAX_OBJS_PER_PAGE;
858 if (page->objects != max_objects) {
859 slab_err(s, page, "Wrong number of objects. Found %d but "
860 "should be %d", page->objects, max_objects);
861 page->objects = max_objects;
862 slab_fix(s, "Number of objects adjusted.");
864 if (page->inuse != page->objects - nr) {
865 slab_err(s, page, "Wrong object count. Counter is %d but "
866 "counted were %d", page->inuse, page->objects - nr);
867 page->inuse = page->objects - nr;
868 slab_fix(s, "Object count adjusted.");
870 return search == NULL;
873 static void trace(struct kmem_cache *s, struct page *page, void *object,
876 if (s->flags & SLAB_TRACE) {
877 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
879 alloc ? "alloc" : "free",
884 print_section("Object", (void *)object, s->objsize);
891 * Hooks for other subsystems that check memory allocations. In a typical
892 * production configuration these hooks all should produce no code at all.
894 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
896 flags &= gfp_allowed_mask;
897 lockdep_trace_alloc(flags);
898 might_sleep_if(flags & __GFP_WAIT);
900 return should_failslab(s->objsize, flags, s->flags);
903 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
905 flags &= gfp_allowed_mask;
906 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
907 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
910 static inline void slab_free_hook(struct kmem_cache *s, void *x)
912 kmemleak_free_recursive(x, s->flags);
915 * Trouble is that we may no longer disable interupts in the fast path
916 * So in order to make the debug calls that expect irqs to be
917 * disabled we need to disable interrupts temporarily.
919 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
923 local_irq_save(flags);
924 kmemcheck_slab_free(s, x, s->objsize);
925 debug_check_no_locks_freed(x, s->objsize);
926 local_irq_restore(flags);
929 if (!(s->flags & SLAB_DEBUG_OBJECTS))
930 debug_check_no_obj_freed(x, s->objsize);
934 * Tracking of fully allocated slabs for debugging purposes.
936 * list_lock must be held.
938 static void add_full(struct kmem_cache *s,
939 struct kmem_cache_node *n, struct page *page)
941 if (!(s->flags & SLAB_STORE_USER))
944 list_add(&page->lru, &n->full);
948 * list_lock must be held.
950 static void remove_full(struct kmem_cache *s, struct page *page)
952 if (!(s->flags & SLAB_STORE_USER))
955 list_del(&page->lru);
958 /* Tracking of the number of slabs for debugging purposes */
959 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
961 struct kmem_cache_node *n = get_node(s, node);
963 return atomic_long_read(&n->nr_slabs);
966 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
968 return atomic_long_read(&n->nr_slabs);
971 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
973 struct kmem_cache_node *n = get_node(s, node);
976 * May be called early in order to allocate a slab for the
977 * kmem_cache_node structure. Solve the chicken-egg
978 * dilemma by deferring the increment of the count during
979 * bootstrap (see early_kmem_cache_node_alloc).
982 atomic_long_inc(&n->nr_slabs);
983 atomic_long_add(objects, &n->total_objects);
986 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
988 struct kmem_cache_node *n = get_node(s, node);
990 atomic_long_dec(&n->nr_slabs);
991 atomic_long_sub(objects, &n->total_objects);
994 /* Object debug checks for alloc/free paths */
995 static void setup_object_debug(struct kmem_cache *s, struct page *page,
998 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1001 init_object(s, object, SLUB_RED_INACTIVE);
1002 init_tracking(s, object);
1005 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1006 void *object, unsigned long addr)
1008 if (!check_slab(s, page))
1011 if (!check_valid_pointer(s, page, object)) {
1012 object_err(s, page, object, "Freelist Pointer check fails");
1016 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1019 /* Success perform special debug activities for allocs */
1020 if (s->flags & SLAB_STORE_USER)
1021 set_track(s, object, TRACK_ALLOC, addr);
1022 trace(s, page, object, 1);
1023 init_object(s, object, SLUB_RED_ACTIVE);
1027 if (PageSlab(page)) {
1029 * If this is a slab page then lets do the best we can
1030 * to avoid issues in the future. Marking all objects
1031 * as used avoids touching the remaining objects.
1033 slab_fix(s, "Marking all objects used");
1034 page->inuse = page->objects;
1035 page->freelist = NULL;
1040 static noinline int free_debug_processing(struct kmem_cache *s,
1041 struct page *page, void *object, unsigned long addr)
1043 unsigned long flags;
1046 local_irq_save(flags);
1049 if (!check_slab(s, page))
1052 if (!check_valid_pointer(s, page, object)) {
1053 slab_err(s, page, "Invalid object pointer 0x%p", object);
1057 if (on_freelist(s, page, object)) {
1058 object_err(s, page, object, "Object already free");
1062 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1065 if (unlikely(s != page->slab)) {
1066 if (!PageSlab(page)) {
1067 slab_err(s, page, "Attempt to free object(0x%p) "
1068 "outside of slab", object);
1069 } else if (!page->slab) {
1071 "SLUB <none>: no slab for object 0x%p.\n",
1075 object_err(s, page, object,
1076 "page slab pointer corrupt.");
1080 if (s->flags & SLAB_STORE_USER)
1081 set_track(s, object, TRACK_FREE, addr);
1082 trace(s, page, object, 0);
1083 init_object(s, object, SLUB_RED_INACTIVE);
1087 local_irq_restore(flags);
1091 slab_fix(s, "Object at 0x%p not freed", object);
1095 static int __init setup_slub_debug(char *str)
1097 slub_debug = DEBUG_DEFAULT_FLAGS;
1098 if (*str++ != '=' || !*str)
1100 * No options specified. Switch on full debugging.
1106 * No options but restriction on slabs. This means full
1107 * debugging for slabs matching a pattern.
1111 if (tolower(*str) == 'o') {
1113 * Avoid enabling debugging on caches if its minimum order
1114 * would increase as a result.
1116 disable_higher_order_debug = 1;
1123 * Switch off all debugging measures.
1128 * Determine which debug features should be switched on
1130 for (; *str && *str != ','; str++) {
1131 switch (tolower(*str)) {
1133 slub_debug |= SLAB_DEBUG_FREE;
1136 slub_debug |= SLAB_RED_ZONE;
1139 slub_debug |= SLAB_POISON;
1142 slub_debug |= SLAB_STORE_USER;
1145 slub_debug |= SLAB_TRACE;
1148 slub_debug |= SLAB_FAILSLAB;
1151 printk(KERN_ERR "slub_debug option '%c' "
1152 "unknown. skipped\n", *str);
1158 slub_debug_slabs = str + 1;
1163 __setup("slub_debug", setup_slub_debug);
1165 static unsigned long kmem_cache_flags(unsigned long objsize,
1166 unsigned long flags, const char *name,
1167 void (*ctor)(void *))
1170 * Enable debugging if selected on the kernel commandline.
1172 if (slub_debug && (!slub_debug_slabs ||
1173 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1174 flags |= slub_debug;
1179 static inline void setup_object_debug(struct kmem_cache *s,
1180 struct page *page, void *object) {}
1182 static inline int alloc_debug_processing(struct kmem_cache *s,
1183 struct page *page, void *object, unsigned long addr) { return 0; }
1185 static inline int free_debug_processing(struct kmem_cache *s,
1186 struct page *page, void *object, unsigned long addr) { return 0; }
1188 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1190 static inline int check_object(struct kmem_cache *s, struct page *page,
1191 void *object, u8 val) { return 1; }
1192 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1193 struct page *page) {}
1194 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1195 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1196 unsigned long flags, const char *name,
1197 void (*ctor)(void *))
1201 #define slub_debug 0
1203 #define disable_higher_order_debug 0
1205 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1207 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1209 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1211 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1214 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1217 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1220 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1222 #endif /* CONFIG_SLUB_DEBUG */
1225 * Slab allocation and freeing
1227 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1228 struct kmem_cache_order_objects oo)
1230 int order = oo_order(oo);
1232 flags |= __GFP_NOTRACK;
1234 if (node == NUMA_NO_NODE)
1235 return alloc_pages(flags, order);
1237 return alloc_pages_exact_node(node, flags, order);
1240 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1243 struct kmem_cache_order_objects oo = s->oo;
1246 flags &= gfp_allowed_mask;
1248 if (flags & __GFP_WAIT)
1251 flags |= s->allocflags;
1254 * Let the initial higher-order allocation fail under memory pressure
1255 * so we fall-back to the minimum order allocation.
1257 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1259 page = alloc_slab_page(alloc_gfp, node, oo);
1260 if (unlikely(!page)) {
1263 * Allocation may have failed due to fragmentation.
1264 * Try a lower order alloc if possible
1266 page = alloc_slab_page(flags, node, oo);
1269 stat(s, ORDER_FALLBACK);
1272 if (flags & __GFP_WAIT)
1273 local_irq_disable();
1278 if (kmemcheck_enabled
1279 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1280 int pages = 1 << oo_order(oo);
1282 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1285 * Objects from caches that have a constructor don't get
1286 * cleared when they're allocated, so we need to do it here.
1289 kmemcheck_mark_uninitialized_pages(page, pages);
1291 kmemcheck_mark_unallocated_pages(page, pages);
1294 page->objects = oo_objects(oo);
1295 mod_zone_page_state(page_zone(page),
1296 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1297 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1303 static void setup_object(struct kmem_cache *s, struct page *page,
1306 setup_object_debug(s, page, object);
1307 if (unlikely(s->ctor))
1311 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1318 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1320 page = allocate_slab(s,
1321 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1325 inc_slabs_node(s, page_to_nid(page), page->objects);
1327 page->flags |= 1 << PG_slab;
1329 start = page_address(page);
1331 if (unlikely(s->flags & SLAB_POISON))
1332 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1335 for_each_object(p, s, start, page->objects) {
1336 setup_object(s, page, last);
1337 set_freepointer(s, last, p);
1340 setup_object(s, page, last);
1341 set_freepointer(s, last, NULL);
1343 page->freelist = start;
1350 static void __free_slab(struct kmem_cache *s, struct page *page)
1352 int order = compound_order(page);
1353 int pages = 1 << order;
1355 if (kmem_cache_debug(s)) {
1358 slab_pad_check(s, page);
1359 for_each_object(p, s, page_address(page),
1361 check_object(s, page, p, SLUB_RED_INACTIVE);
1364 kmemcheck_free_shadow(page, compound_order(page));
1366 mod_zone_page_state(page_zone(page),
1367 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1368 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1371 __ClearPageSlab(page);
1372 reset_page_mapcount(page);
1373 if (current->reclaim_state)
1374 current->reclaim_state->reclaimed_slab += pages;
1375 __free_pages(page, order);
1378 #define need_reserve_slab_rcu \
1379 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1381 static void rcu_free_slab(struct rcu_head *h)
1385 if (need_reserve_slab_rcu)
1386 page = virt_to_head_page(h);
1388 page = container_of((struct list_head *)h, struct page, lru);
1390 __free_slab(page->slab, page);
1393 static void free_slab(struct kmem_cache *s, struct page *page)
1395 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1396 struct rcu_head *head;
1398 if (need_reserve_slab_rcu) {
1399 int order = compound_order(page);
1400 int offset = (PAGE_SIZE << order) - s->reserved;
1402 VM_BUG_ON(s->reserved != sizeof(*head));
1403 head = page_address(page) + offset;
1406 * RCU free overloads the RCU head over the LRU
1408 head = (void *)&page->lru;
1411 call_rcu(head, rcu_free_slab);
1413 __free_slab(s, page);
1416 static void discard_slab(struct kmem_cache *s, struct page *page)
1418 dec_slabs_node(s, page_to_nid(page), page->objects);
1423 * Management of partially allocated slabs.
1425 * list_lock must be held.
1427 static inline void add_partial(struct kmem_cache_node *n,
1428 struct page *page, int tail)
1432 list_add_tail(&page->lru, &n->partial);
1434 list_add(&page->lru, &n->partial);
1438 * list_lock must be held.
1440 static inline void remove_partial(struct kmem_cache_node *n,
1443 list_del(&page->lru);
1448 * Lock slab, remove from the partial list and put the object into the
1451 * Must hold list_lock.
1453 static inline int acquire_slab(struct kmem_cache *s,
1454 struct kmem_cache_node *n, struct page *page)
1457 unsigned long counters;
1461 * Zap the freelist and set the frozen bit.
1462 * The old freelist is the list of objects for the
1463 * per cpu allocation list.
1466 freelist = page->freelist;
1467 counters = page->counters;
1468 new.counters = counters;
1469 new.inuse = page->objects;
1471 VM_BUG_ON(new.frozen);
1474 } while (!cmpxchg_double_slab(s, page,
1477 "lock and freeze"));
1479 remove_partial(n, page);
1482 /* Populate the per cpu freelist */
1483 this_cpu_write(s->cpu_slab->freelist, freelist);
1484 this_cpu_write(s->cpu_slab->page, page);
1485 this_cpu_write(s->cpu_slab->node, page_to_nid(page));
1489 * Slab page came from the wrong list. No object to allocate
1490 * from. Put it onto the correct list and continue partial
1493 printk(KERN_ERR "SLUB: %s : Page without available objects on"
1494 " partial list\n", s->name);
1500 * Try to allocate a partial slab from a specific node.
1502 static struct page *get_partial_node(struct kmem_cache *s,
1503 struct kmem_cache_node *n)
1508 * Racy check. If we mistakenly see no partial slabs then we
1509 * just allocate an empty slab. If we mistakenly try to get a
1510 * partial slab and there is none available then get_partials()
1513 if (!n || !n->nr_partial)
1516 spin_lock(&n->list_lock);
1517 list_for_each_entry(page, &n->partial, lru)
1518 if (acquire_slab(s, n, page))
1522 spin_unlock(&n->list_lock);
1527 * Get a page from somewhere. Search in increasing NUMA distances.
1529 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1532 struct zonelist *zonelist;
1535 enum zone_type high_zoneidx = gfp_zone(flags);
1539 * The defrag ratio allows a configuration of the tradeoffs between
1540 * inter node defragmentation and node local allocations. A lower
1541 * defrag_ratio increases the tendency to do local allocations
1542 * instead of attempting to obtain partial slabs from other nodes.
1544 * If the defrag_ratio is set to 0 then kmalloc() always
1545 * returns node local objects. If the ratio is higher then kmalloc()
1546 * may return off node objects because partial slabs are obtained
1547 * from other nodes and filled up.
1549 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1550 * defrag_ratio = 1000) then every (well almost) allocation will
1551 * first attempt to defrag slab caches on other nodes. This means
1552 * scanning over all nodes to look for partial slabs which may be
1553 * expensive if we do it every time we are trying to find a slab
1554 * with available objects.
1556 if (!s->remote_node_defrag_ratio ||
1557 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1561 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1562 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1563 struct kmem_cache_node *n;
1565 n = get_node(s, zone_to_nid(zone));
1567 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1568 n->nr_partial > s->min_partial) {
1569 page = get_partial_node(s, n);
1582 * Get a partial page, lock it and return it.
1584 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1587 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1589 page = get_partial_node(s, get_node(s, searchnode));
1590 if (page || node != NUMA_NO_NODE)
1593 return get_any_partial(s, flags);
1596 #ifdef CONFIG_PREEMPT
1598 * Calculate the next globally unique transaction for disambiguiation
1599 * during cmpxchg. The transactions start with the cpu number and are then
1600 * incremented by CONFIG_NR_CPUS.
1602 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1605 * No preemption supported therefore also no need to check for
1611 static inline unsigned long next_tid(unsigned long tid)
1613 return tid + TID_STEP;
1616 static inline unsigned int tid_to_cpu(unsigned long tid)
1618 return tid % TID_STEP;
1621 static inline unsigned long tid_to_event(unsigned long tid)
1623 return tid / TID_STEP;
1626 static inline unsigned int init_tid(int cpu)
1631 static inline void note_cmpxchg_failure(const char *n,
1632 const struct kmem_cache *s, unsigned long tid)
1634 #ifdef SLUB_DEBUG_CMPXCHG
1635 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1637 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1639 #ifdef CONFIG_PREEMPT
1640 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1641 printk("due to cpu change %d -> %d\n",
1642 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1645 if (tid_to_event(tid) != tid_to_event(actual_tid))
1646 printk("due to cpu running other code. Event %ld->%ld\n",
1647 tid_to_event(tid), tid_to_event(actual_tid));
1649 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1650 actual_tid, tid, next_tid(tid));
1652 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1655 void init_kmem_cache_cpus(struct kmem_cache *s)
1659 for_each_possible_cpu(cpu)
1660 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1663 * Remove the cpu slab
1667 * Remove the cpu slab
1669 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1671 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1672 struct page *page = c->page;
1673 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1675 enum slab_modes l = M_NONE, m = M_NONE;
1682 if (page->freelist) {
1683 stat(s, DEACTIVATE_REMOTE_FREES);
1687 c->tid = next_tid(c->tid);
1689 freelist = c->freelist;
1693 * Stage one: Free all available per cpu objects back
1694 * to the page freelist while it is still frozen. Leave the
1697 * There is no need to take the list->lock because the page
1700 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1702 unsigned long counters;
1705 prior = page->freelist;
1706 counters = page->counters;
1707 set_freepointer(s, freelist, prior);
1708 new.counters = counters;
1710 VM_BUG_ON(!new.frozen);
1712 } while (!cmpxchg_double_slab(s, page,
1714 freelist, new.counters,
1715 "drain percpu freelist"));
1717 freelist = nextfree;
1721 * Stage two: Ensure that the page is unfrozen while the
1722 * list presence reflects the actual number of objects
1725 * We setup the list membership and then perform a cmpxchg
1726 * with the count. If there is a mismatch then the page
1727 * is not unfrozen but the page is on the wrong list.
1729 * Then we restart the process which may have to remove
1730 * the page from the list that we just put it on again
1731 * because the number of objects in the slab may have
1736 old.freelist = page->freelist;
1737 old.counters = page->counters;
1738 VM_BUG_ON(!old.frozen);
1740 /* Determine target state of the slab */
1741 new.counters = old.counters;
1744 set_freepointer(s, freelist, old.freelist);
1745 new.freelist = freelist;
1747 new.freelist = old.freelist;
1751 if (!new.inuse && n->nr_partial < s->min_partial)
1753 else if (new.freelist) {
1758 * Taking the spinlock removes the possiblity
1759 * that acquire_slab() will see a slab page that
1762 spin_lock(&n->list_lock);
1766 if (kmem_cache_debug(s) && !lock) {
1769 * This also ensures that the scanning of full
1770 * slabs from diagnostic functions will not see
1773 spin_lock(&n->list_lock);
1781 remove_partial(n, page);
1783 else if (l == M_FULL)
1785 remove_full(s, page);
1787 if (m == M_PARTIAL) {
1789 add_partial(n, page, tail);
1790 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1792 } else if (m == M_FULL) {
1794 stat(s, DEACTIVATE_FULL);
1795 add_full(s, n, page);
1801 if (!cmpxchg_double_slab(s, page,
1802 old.freelist, old.counters,
1803 new.freelist, new.counters,
1808 spin_unlock(&n->list_lock);
1811 stat(s, DEACTIVATE_EMPTY);
1812 discard_slab(s, page);
1817 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1819 stat(s, CPUSLAB_FLUSH);
1820 deactivate_slab(s, c);
1826 * Called from IPI handler with interrupts disabled.
1828 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1830 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1832 if (likely(c && c->page))
1836 static void flush_cpu_slab(void *d)
1838 struct kmem_cache *s = d;
1840 __flush_cpu_slab(s, smp_processor_id());
1843 static void flush_all(struct kmem_cache *s)
1845 on_each_cpu(flush_cpu_slab, s, 1);
1849 * Check if the objects in a per cpu structure fit numa
1850 * locality expectations.
1852 static inline int node_match(struct kmem_cache_cpu *c, int node)
1855 if (node != NUMA_NO_NODE && c->node != node)
1861 static int count_free(struct page *page)
1863 return page->objects - page->inuse;
1866 static unsigned long count_partial(struct kmem_cache_node *n,
1867 int (*get_count)(struct page *))
1869 unsigned long flags;
1870 unsigned long x = 0;
1873 spin_lock_irqsave(&n->list_lock, flags);
1874 list_for_each_entry(page, &n->partial, lru)
1875 x += get_count(page);
1876 spin_unlock_irqrestore(&n->list_lock, flags);
1880 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1882 #ifdef CONFIG_SLUB_DEBUG
1883 return atomic_long_read(&n->total_objects);
1889 static noinline void
1890 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1895 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1897 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1898 "default order: %d, min order: %d\n", s->name, s->objsize,
1899 s->size, oo_order(s->oo), oo_order(s->min));
1901 if (oo_order(s->min) > get_order(s->objsize))
1902 printk(KERN_WARNING " %s debugging increased min order, use "
1903 "slub_debug=O to disable.\n", s->name);
1905 for_each_online_node(node) {
1906 struct kmem_cache_node *n = get_node(s, node);
1907 unsigned long nr_slabs;
1908 unsigned long nr_objs;
1909 unsigned long nr_free;
1914 nr_free = count_partial(n, count_free);
1915 nr_slabs = node_nr_slabs(n);
1916 nr_objs = node_nr_objs(n);
1919 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1920 node, nr_slabs, nr_objs, nr_free);
1925 * Slow path. The lockless freelist is empty or we need to perform
1928 * Interrupts are disabled.
1930 * Processing is still very fast if new objects have been freed to the
1931 * regular freelist. In that case we simply take over the regular freelist
1932 * as the lockless freelist and zap the regular freelist.
1934 * If that is not working then we fall back to the partial lists. We take the
1935 * first element of the freelist as the object to allocate now and move the
1936 * rest of the freelist to the lockless freelist.
1938 * And if we were unable to get a new slab from the partial slab lists then
1939 * we need to allocate a new slab. This is the slowest path since it involves
1940 * a call to the page allocator and the setup of a new slab.
1942 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1943 unsigned long addr, struct kmem_cache_cpu *c)
1947 unsigned long flags;
1949 unsigned long counters;
1951 local_irq_save(flags);
1952 #ifdef CONFIG_PREEMPT
1954 * We may have been preempted and rescheduled on a different
1955 * cpu before disabling interrupts. Need to reload cpu area
1958 c = this_cpu_ptr(s->cpu_slab);
1961 /* We handle __GFP_ZERO in the caller */
1962 gfpflags &= ~__GFP_ZERO;
1968 if (unlikely(!node_match(c, node)))
1971 stat(s, ALLOC_SLOWPATH);
1974 object = page->freelist;
1975 counters = page->counters;
1976 new.counters = counters;
1977 new.inuse = page->objects;
1978 VM_BUG_ON(!new.frozen);
1980 } while (!cmpxchg_double_slab(s, page,
1986 VM_BUG_ON(!page->frozen);
1988 if (unlikely(!object))
1991 stat(s, ALLOC_REFILL);
1993 c->freelist = get_freepointer(s, object);
1994 c->tid = next_tid(c->tid);
1995 local_irq_restore(flags);
1999 deactivate_slab(s, c);
2002 page = get_partial(s, gfpflags, node);
2004 stat(s, ALLOC_FROM_PARTIAL);
2005 object = c->freelist;
2007 if (kmem_cache_debug(s))
2012 page = new_slab(s, gfpflags, node);
2015 c = __this_cpu_ptr(s->cpu_slab);
2020 * No other reference to the page yet so we can
2021 * muck around with it freely without cmpxchg
2023 object = page->freelist;
2024 page->freelist = NULL;
2025 page->inuse = page->objects;
2027 stat(s, ALLOC_SLAB);
2028 c->node = page_to_nid(page);
2032 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2033 slab_out_of_memory(s, gfpflags, node);
2034 local_irq_restore(flags);
2038 if (!object || !alloc_debug_processing(s, page, object, addr))
2041 c->freelist = get_freepointer(s, object);
2042 deactivate_slab(s, c);
2044 c->node = NUMA_NO_NODE;
2045 local_irq_restore(flags);
2050 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2051 * have the fastpath folded into their functions. So no function call
2052 * overhead for requests that can be satisfied on the fastpath.
2054 * The fastpath works by first checking if the lockless freelist can be used.
2055 * If not then __slab_alloc is called for slow processing.
2057 * Otherwise we can simply pick the next object from the lockless free list.
2059 static __always_inline void *slab_alloc(struct kmem_cache *s,
2060 gfp_t gfpflags, int node, unsigned long addr)
2063 struct kmem_cache_cpu *c;
2066 if (slab_pre_alloc_hook(s, gfpflags))
2072 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2073 * enabled. We may switch back and forth between cpus while
2074 * reading from one cpu area. That does not matter as long
2075 * as we end up on the original cpu again when doing the cmpxchg.
2077 c = __this_cpu_ptr(s->cpu_slab);
2080 * The transaction ids are globally unique per cpu and per operation on
2081 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2082 * occurs on the right processor and that there was no operation on the
2083 * linked list in between.
2088 object = c->freelist;
2089 if (unlikely(!object || !node_match(c, node)))
2091 object = __slab_alloc(s, gfpflags, node, addr, c);
2095 * The cmpxchg will only match if there was no additional
2096 * operation and if we are on the right processor.
2098 * The cmpxchg does the following atomically (without lock semantics!)
2099 * 1. Relocate first pointer to the current per cpu area.
2100 * 2. Verify that tid and freelist have not been changed
2101 * 3. If they were not changed replace tid and freelist
2103 * Since this is without lock semantics the protection is only against
2104 * code executing on this cpu *not* from access by other cpus.
2106 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2107 s->cpu_slab->freelist, s->cpu_slab->tid,
2109 get_freepointer_safe(s, object), next_tid(tid)))) {
2111 note_cmpxchg_failure("slab_alloc", s, tid);
2114 stat(s, ALLOC_FASTPATH);
2117 if (unlikely(gfpflags & __GFP_ZERO) && object)
2118 memset(object, 0, s->objsize);
2120 slab_post_alloc_hook(s, gfpflags, object);
2125 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2127 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2129 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2133 EXPORT_SYMBOL(kmem_cache_alloc);
2135 #ifdef CONFIG_TRACING
2136 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2138 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2139 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2142 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2144 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2146 void *ret = kmalloc_order(size, flags, order);
2147 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2150 EXPORT_SYMBOL(kmalloc_order_trace);
2154 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2156 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2158 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2159 s->objsize, s->size, gfpflags, node);
2163 EXPORT_SYMBOL(kmem_cache_alloc_node);
2165 #ifdef CONFIG_TRACING
2166 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2168 int node, size_t size)
2170 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2172 trace_kmalloc_node(_RET_IP_, ret,
2173 size, s->size, gfpflags, node);
2176 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2181 * Slow patch handling. This may still be called frequently since objects
2182 * have a longer lifetime than the cpu slabs in most processing loads.
2184 * So we still attempt to reduce cache line usage. Just take the slab
2185 * lock and free the item. If there is no additional partial page
2186 * handling required then we can return immediately.
2188 static void __slab_free(struct kmem_cache *s, struct page *page,
2189 void *x, unsigned long addr)
2192 void **object = (void *)x;
2196 unsigned long counters;
2197 struct kmem_cache_node *n = NULL;
2198 unsigned long uninitialized_var(flags);
2200 stat(s, FREE_SLOWPATH);
2202 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2206 prior = page->freelist;
2207 counters = page->counters;
2208 set_freepointer(s, object, prior);
2209 new.counters = counters;
2210 was_frozen = new.frozen;
2212 if ((!new.inuse || !prior) && !was_frozen && !n) {
2213 n = get_node(s, page_to_nid(page));
2215 * Speculatively acquire the list_lock.
2216 * If the cmpxchg does not succeed then we may
2217 * drop the list_lock without any processing.
2219 * Otherwise the list_lock will synchronize with
2220 * other processors updating the list of slabs.
2222 spin_lock_irqsave(&n->list_lock, flags);
2226 } while (!cmpxchg_double_slab(s, page,
2228 object, new.counters,
2233 * The list lock was not taken therefore no list
2234 * activity can be necessary.
2237 stat(s, FREE_FROZEN);
2242 * was_frozen may have been set after we acquired the list_lock in
2243 * an earlier loop. So we need to check it here again.
2246 stat(s, FREE_FROZEN);
2248 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2252 * Objects left in the slab. If it was not on the partial list before
2255 if (unlikely(!prior)) {
2256 remove_full(s, page);
2257 add_partial(n, page, 0);
2258 stat(s, FREE_ADD_PARTIAL);
2261 spin_unlock_irqrestore(&n->list_lock, flags);
2267 * Slab still on the partial list.
2269 remove_partial(n, page);
2270 stat(s, FREE_REMOVE_PARTIAL);
2273 spin_unlock_irqrestore(&n->list_lock, flags);
2275 discard_slab(s, page);
2279 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2280 * can perform fastpath freeing without additional function calls.
2282 * The fastpath is only possible if we are freeing to the current cpu slab
2283 * of this processor. This typically the case if we have just allocated
2286 * If fastpath is not possible then fall back to __slab_free where we deal
2287 * with all sorts of special processing.
2289 static __always_inline void slab_free(struct kmem_cache *s,
2290 struct page *page, void *x, unsigned long addr)
2292 void **object = (void *)x;
2293 struct kmem_cache_cpu *c;
2296 slab_free_hook(s, x);
2301 * Determine the currently cpus per cpu slab.
2302 * The cpu may change afterward. However that does not matter since
2303 * data is retrieved via this pointer. If we are on the same cpu
2304 * during the cmpxchg then the free will succedd.
2306 c = __this_cpu_ptr(s->cpu_slab);
2311 if (likely(page == c->page)) {
2312 set_freepointer(s, object, c->freelist);
2314 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2315 s->cpu_slab->freelist, s->cpu_slab->tid,
2317 object, next_tid(tid)))) {
2319 note_cmpxchg_failure("slab_free", s, tid);
2322 stat(s, FREE_FASTPATH);
2324 __slab_free(s, page, x, addr);
2328 void kmem_cache_free(struct kmem_cache *s, void *x)
2332 page = virt_to_head_page(x);
2334 slab_free(s, page, x, _RET_IP_);
2336 trace_kmem_cache_free(_RET_IP_, x);
2338 EXPORT_SYMBOL(kmem_cache_free);
2341 * Object placement in a slab is made very easy because we always start at
2342 * offset 0. If we tune the size of the object to the alignment then we can
2343 * get the required alignment by putting one properly sized object after
2346 * Notice that the allocation order determines the sizes of the per cpu
2347 * caches. Each processor has always one slab available for allocations.
2348 * Increasing the allocation order reduces the number of times that slabs
2349 * must be moved on and off the partial lists and is therefore a factor in
2354 * Mininum / Maximum order of slab pages. This influences locking overhead
2355 * and slab fragmentation. A higher order reduces the number of partial slabs
2356 * and increases the number of allocations possible without having to
2357 * take the list_lock.
2359 static int slub_min_order;
2360 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2361 static int slub_min_objects;
2364 * Merge control. If this is set then no merging of slab caches will occur.
2365 * (Could be removed. This was introduced to pacify the merge skeptics.)
2367 static int slub_nomerge;
2370 * Calculate the order of allocation given an slab object size.
2372 * The order of allocation has significant impact on performance and other
2373 * system components. Generally order 0 allocations should be preferred since
2374 * order 0 does not cause fragmentation in the page allocator. Larger objects
2375 * be problematic to put into order 0 slabs because there may be too much
2376 * unused space left. We go to a higher order if more than 1/16th of the slab
2379 * In order to reach satisfactory performance we must ensure that a minimum
2380 * number of objects is in one slab. Otherwise we may generate too much
2381 * activity on the partial lists which requires taking the list_lock. This is
2382 * less a concern for large slabs though which are rarely used.
2384 * slub_max_order specifies the order where we begin to stop considering the
2385 * number of objects in a slab as critical. If we reach slub_max_order then
2386 * we try to keep the page order as low as possible. So we accept more waste
2387 * of space in favor of a small page order.
2389 * Higher order allocations also allow the placement of more objects in a
2390 * slab and thereby reduce object handling overhead. If the user has
2391 * requested a higher mininum order then we start with that one instead of
2392 * the smallest order which will fit the object.
2394 static inline int slab_order(int size, int min_objects,
2395 int max_order, int fract_leftover, int reserved)
2399 int min_order = slub_min_order;
2401 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2402 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2404 for (order = max(min_order,
2405 fls(min_objects * size - 1) - PAGE_SHIFT);
2406 order <= max_order; order++) {
2408 unsigned long slab_size = PAGE_SIZE << order;
2410 if (slab_size < min_objects * size + reserved)
2413 rem = (slab_size - reserved) % size;
2415 if (rem <= slab_size / fract_leftover)
2423 static inline int calculate_order(int size, int reserved)
2431 * Attempt to find best configuration for a slab. This
2432 * works by first attempting to generate a layout with
2433 * the best configuration and backing off gradually.
2435 * First we reduce the acceptable waste in a slab. Then
2436 * we reduce the minimum objects required in a slab.
2438 min_objects = slub_min_objects;
2440 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2441 max_objects = order_objects(slub_max_order, size, reserved);
2442 min_objects = min(min_objects, max_objects);
2444 while (min_objects > 1) {
2446 while (fraction >= 4) {
2447 order = slab_order(size, min_objects,
2448 slub_max_order, fraction, reserved);
2449 if (order <= slub_max_order)
2457 * We were unable to place multiple objects in a slab. Now
2458 * lets see if we can place a single object there.
2460 order = slab_order(size, 1, slub_max_order, 1, reserved);
2461 if (order <= slub_max_order)
2465 * Doh this slab cannot be placed using slub_max_order.
2467 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2468 if (order < MAX_ORDER)
2474 * Figure out what the alignment of the objects will be.
2476 static unsigned long calculate_alignment(unsigned long flags,
2477 unsigned long align, unsigned long size)
2480 * If the user wants hardware cache aligned objects then follow that
2481 * suggestion if the object is sufficiently large.
2483 * The hardware cache alignment cannot override the specified
2484 * alignment though. If that is greater then use it.
2486 if (flags & SLAB_HWCACHE_ALIGN) {
2487 unsigned long ralign = cache_line_size();
2488 while (size <= ralign / 2)
2490 align = max(align, ralign);
2493 if (align < ARCH_SLAB_MINALIGN)
2494 align = ARCH_SLAB_MINALIGN;
2496 return ALIGN(align, sizeof(void *));
2500 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2503 spin_lock_init(&n->list_lock);
2504 INIT_LIST_HEAD(&n->partial);
2505 #ifdef CONFIG_SLUB_DEBUG
2506 atomic_long_set(&n->nr_slabs, 0);
2507 atomic_long_set(&n->total_objects, 0);
2508 INIT_LIST_HEAD(&n->full);
2512 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2514 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2515 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2518 * Must align to double word boundary for the double cmpxchg
2519 * instructions to work; see __pcpu_double_call_return_bool().
2521 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2522 2 * sizeof(void *));
2527 init_kmem_cache_cpus(s);
2532 static struct kmem_cache *kmem_cache_node;
2535 * No kmalloc_node yet so do it by hand. We know that this is the first
2536 * slab on the node for this slabcache. There are no concurrent accesses
2539 * Note that this function only works on the kmalloc_node_cache
2540 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2541 * memory on a fresh node that has no slab structures yet.
2543 static void early_kmem_cache_node_alloc(int node)
2546 struct kmem_cache_node *n;
2548 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2550 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2553 if (page_to_nid(page) != node) {
2554 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2556 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2557 "in order to be able to continue\n");
2562 page->freelist = get_freepointer(kmem_cache_node, n);
2565 kmem_cache_node->node[node] = n;
2566 #ifdef CONFIG_SLUB_DEBUG
2567 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2568 init_tracking(kmem_cache_node, n);
2570 init_kmem_cache_node(n, kmem_cache_node);
2571 inc_slabs_node(kmem_cache_node, node, page->objects);
2573 add_partial(n, page, 0);
2576 static void free_kmem_cache_nodes(struct kmem_cache *s)
2580 for_each_node_state(node, N_NORMAL_MEMORY) {
2581 struct kmem_cache_node *n = s->node[node];
2584 kmem_cache_free(kmem_cache_node, n);
2586 s->node[node] = NULL;
2590 static int init_kmem_cache_nodes(struct kmem_cache *s)
2594 for_each_node_state(node, N_NORMAL_MEMORY) {
2595 struct kmem_cache_node *n;
2597 if (slab_state == DOWN) {
2598 early_kmem_cache_node_alloc(node);
2601 n = kmem_cache_alloc_node(kmem_cache_node,
2605 free_kmem_cache_nodes(s);
2610 init_kmem_cache_node(n, s);
2615 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2617 if (min < MIN_PARTIAL)
2619 else if (min > MAX_PARTIAL)
2621 s->min_partial = min;
2625 * calculate_sizes() determines the order and the distribution of data within
2628 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2630 unsigned long flags = s->flags;
2631 unsigned long size = s->objsize;
2632 unsigned long align = s->align;
2636 * Round up object size to the next word boundary. We can only
2637 * place the free pointer at word boundaries and this determines
2638 * the possible location of the free pointer.
2640 size = ALIGN(size, sizeof(void *));
2642 #ifdef CONFIG_SLUB_DEBUG
2644 * Determine if we can poison the object itself. If the user of
2645 * the slab may touch the object after free or before allocation
2646 * then we should never poison the object itself.
2648 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2650 s->flags |= __OBJECT_POISON;
2652 s->flags &= ~__OBJECT_POISON;
2656 * If we are Redzoning then check if there is some space between the
2657 * end of the object and the free pointer. If not then add an
2658 * additional word to have some bytes to store Redzone information.
2660 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2661 size += sizeof(void *);
2665 * With that we have determined the number of bytes in actual use
2666 * by the object. This is the potential offset to the free pointer.
2670 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2673 * Relocate free pointer after the object if it is not
2674 * permitted to overwrite the first word of the object on
2677 * This is the case if we do RCU, have a constructor or
2678 * destructor or are poisoning the objects.
2681 size += sizeof(void *);
2684 #ifdef CONFIG_SLUB_DEBUG
2685 if (flags & SLAB_STORE_USER)
2687 * Need to store information about allocs and frees after
2690 size += 2 * sizeof(struct track);
2692 if (flags & SLAB_RED_ZONE)
2694 * Add some empty padding so that we can catch
2695 * overwrites from earlier objects rather than let
2696 * tracking information or the free pointer be
2697 * corrupted if a user writes before the start
2700 size += sizeof(void *);
2704 * Determine the alignment based on various parameters that the
2705 * user specified and the dynamic determination of cache line size
2708 align = calculate_alignment(flags, align, s->objsize);
2712 * SLUB stores one object immediately after another beginning from
2713 * offset 0. In order to align the objects we have to simply size
2714 * each object to conform to the alignment.
2716 size = ALIGN(size, align);
2718 if (forced_order >= 0)
2719 order = forced_order;
2721 order = calculate_order(size, s->reserved);
2728 s->allocflags |= __GFP_COMP;
2730 if (s->flags & SLAB_CACHE_DMA)
2731 s->allocflags |= SLUB_DMA;
2733 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2734 s->allocflags |= __GFP_RECLAIMABLE;
2737 * Determine the number of objects per slab
2739 s->oo = oo_make(order, size, s->reserved);
2740 s->min = oo_make(get_order(size), size, s->reserved);
2741 if (oo_objects(s->oo) > oo_objects(s->max))
2744 return !!oo_objects(s->oo);
2748 static int kmem_cache_open(struct kmem_cache *s,
2749 const char *name, size_t size,
2750 size_t align, unsigned long flags,
2751 void (*ctor)(void *))
2753 memset(s, 0, kmem_size);
2758 s->flags = kmem_cache_flags(size, flags, name, ctor);
2761 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2762 s->reserved = sizeof(struct rcu_head);
2764 if (!calculate_sizes(s, -1))
2766 if (disable_higher_order_debug) {
2768 * Disable debugging flags that store metadata if the min slab
2771 if (get_order(s->size) > get_order(s->objsize)) {
2772 s->flags &= ~DEBUG_METADATA_FLAGS;
2774 if (!calculate_sizes(s, -1))
2779 #ifdef CONFIG_CMPXCHG_DOUBLE
2780 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
2781 /* Enable fast mode */
2782 s->flags |= __CMPXCHG_DOUBLE;
2786 * The larger the object size is, the more pages we want on the partial
2787 * list to avoid pounding the page allocator excessively.
2789 set_min_partial(s, ilog2(s->size));
2792 s->remote_node_defrag_ratio = 1000;
2794 if (!init_kmem_cache_nodes(s))
2797 if (alloc_kmem_cache_cpus(s))
2800 free_kmem_cache_nodes(s);
2802 if (flags & SLAB_PANIC)
2803 panic("Cannot create slab %s size=%lu realsize=%u "
2804 "order=%u offset=%u flags=%lx\n",
2805 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2811 * Determine the size of a slab object
2813 unsigned int kmem_cache_size(struct kmem_cache *s)
2817 EXPORT_SYMBOL(kmem_cache_size);
2819 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2822 #ifdef CONFIG_SLUB_DEBUG
2823 void *addr = page_address(page);
2825 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2826 sizeof(long), GFP_ATOMIC);
2829 slab_err(s, page, "%s", text);
2832 get_map(s, page, map);
2833 for_each_object(p, s, addr, page->objects) {
2835 if (!test_bit(slab_index(p, s, addr), map)) {
2836 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2838 print_tracking(s, p);
2847 * Attempt to free all partial slabs on a node.
2849 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2851 unsigned long flags;
2852 struct page *page, *h;
2854 spin_lock_irqsave(&n->list_lock, flags);
2855 list_for_each_entry_safe(page, h, &n->partial, lru) {
2857 remove_partial(n, page);
2858 discard_slab(s, page);
2860 list_slab_objects(s, page,
2861 "Objects remaining on kmem_cache_close()");
2864 spin_unlock_irqrestore(&n->list_lock, flags);
2868 * Release all resources used by a slab cache.
2870 static inline int kmem_cache_close(struct kmem_cache *s)
2875 free_percpu(s->cpu_slab);
2876 /* Attempt to free all objects */
2877 for_each_node_state(node, N_NORMAL_MEMORY) {
2878 struct kmem_cache_node *n = get_node(s, node);
2881 if (n->nr_partial || slabs_node(s, node))
2884 free_kmem_cache_nodes(s);
2889 * Close a cache and release the kmem_cache structure
2890 * (must be used for caches created using kmem_cache_create)
2892 void kmem_cache_destroy(struct kmem_cache *s)
2894 down_write(&slub_lock);
2898 if (kmem_cache_close(s)) {
2899 printk(KERN_ERR "SLUB %s: %s called for cache that "
2900 "still has objects.\n", s->name, __func__);
2903 if (s->flags & SLAB_DESTROY_BY_RCU)
2905 sysfs_slab_remove(s);
2907 up_write(&slub_lock);
2909 EXPORT_SYMBOL(kmem_cache_destroy);
2911 /********************************************************************
2913 *******************************************************************/
2915 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2916 EXPORT_SYMBOL(kmalloc_caches);
2918 static struct kmem_cache *kmem_cache;
2920 #ifdef CONFIG_ZONE_DMA
2921 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2924 static int __init setup_slub_min_order(char *str)
2926 get_option(&str, &slub_min_order);
2931 __setup("slub_min_order=", setup_slub_min_order);
2933 static int __init setup_slub_max_order(char *str)
2935 get_option(&str, &slub_max_order);
2936 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2941 __setup("slub_max_order=", setup_slub_max_order);
2943 static int __init setup_slub_min_objects(char *str)
2945 get_option(&str, &slub_min_objects);
2950 __setup("slub_min_objects=", setup_slub_min_objects);
2952 static int __init setup_slub_nomerge(char *str)
2958 __setup("slub_nomerge", setup_slub_nomerge);
2960 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2961 int size, unsigned int flags)
2963 struct kmem_cache *s;
2965 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2968 * This function is called with IRQs disabled during early-boot on
2969 * single CPU so there's no need to take slub_lock here.
2971 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2975 list_add(&s->list, &slab_caches);
2979 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2984 * Conversion table for small slabs sizes / 8 to the index in the
2985 * kmalloc array. This is necessary for slabs < 192 since we have non power
2986 * of two cache sizes there. The size of larger slabs can be determined using
2989 static s8 size_index[24] = {
3016 static inline int size_index_elem(size_t bytes)
3018 return (bytes - 1) / 8;
3021 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3027 return ZERO_SIZE_PTR;
3029 index = size_index[size_index_elem(size)];
3031 index = fls(size - 1);
3033 #ifdef CONFIG_ZONE_DMA
3034 if (unlikely((flags & SLUB_DMA)))
3035 return kmalloc_dma_caches[index];
3038 return kmalloc_caches[index];
3041 void *__kmalloc(size_t size, gfp_t flags)
3043 struct kmem_cache *s;
3046 if (unlikely(size > SLUB_MAX_SIZE))
3047 return kmalloc_large(size, flags);
3049 s = get_slab(size, flags);
3051 if (unlikely(ZERO_OR_NULL_PTR(s)))
3054 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3056 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3060 EXPORT_SYMBOL(__kmalloc);
3063 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3068 flags |= __GFP_COMP | __GFP_NOTRACK;
3069 page = alloc_pages_node(node, flags, get_order(size));
3071 ptr = page_address(page);
3073 kmemleak_alloc(ptr, size, 1, flags);
3077 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3079 struct kmem_cache *s;
3082 if (unlikely(size > SLUB_MAX_SIZE)) {
3083 ret = kmalloc_large_node(size, flags, node);
3085 trace_kmalloc_node(_RET_IP_, ret,
3086 size, PAGE_SIZE << get_order(size),
3092 s = get_slab(size, flags);
3094 if (unlikely(ZERO_OR_NULL_PTR(s)))
3097 ret = slab_alloc(s, flags, node, _RET_IP_);
3099 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3103 EXPORT_SYMBOL(__kmalloc_node);
3106 size_t ksize(const void *object)
3110 if (unlikely(object == ZERO_SIZE_PTR))
3113 page = virt_to_head_page(object);
3115 if (unlikely(!PageSlab(page))) {
3116 WARN_ON(!PageCompound(page));
3117 return PAGE_SIZE << compound_order(page);
3120 return slab_ksize(page->slab);
3122 EXPORT_SYMBOL(ksize);
3124 void kfree(const void *x)
3127 void *object = (void *)x;
3129 trace_kfree(_RET_IP_, x);
3131 if (unlikely(ZERO_OR_NULL_PTR(x)))
3134 page = virt_to_head_page(x);
3135 if (unlikely(!PageSlab(page))) {
3136 BUG_ON(!PageCompound(page));
3141 slab_free(page->slab, page, object, _RET_IP_);
3143 EXPORT_SYMBOL(kfree);
3146 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3147 * the remaining slabs by the number of items in use. The slabs with the
3148 * most items in use come first. New allocations will then fill those up
3149 * and thus they can be removed from the partial lists.
3151 * The slabs with the least items are placed last. This results in them
3152 * being allocated from last increasing the chance that the last objects
3153 * are freed in them.
3155 int kmem_cache_shrink(struct kmem_cache *s)
3159 struct kmem_cache_node *n;
3162 int objects = oo_objects(s->max);
3163 struct list_head *slabs_by_inuse =
3164 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3165 unsigned long flags;
3167 if (!slabs_by_inuse)
3171 for_each_node_state(node, N_NORMAL_MEMORY) {
3172 n = get_node(s, node);
3177 for (i = 0; i < objects; i++)
3178 INIT_LIST_HEAD(slabs_by_inuse + i);
3180 spin_lock_irqsave(&n->list_lock, flags);
3183 * Build lists indexed by the items in use in each slab.
3185 * Note that concurrent frees may occur while we hold the
3186 * list_lock. page->inuse here is the upper limit.
3188 list_for_each_entry_safe(page, t, &n->partial, lru) {
3190 remove_partial(n, page);
3191 discard_slab(s, page);
3193 list_move(&page->lru,
3194 slabs_by_inuse + page->inuse);
3199 * Rebuild the partial list with the slabs filled up most
3200 * first and the least used slabs at the end.
3202 for (i = objects - 1; i >= 0; i--)
3203 list_splice(slabs_by_inuse + i, n->partial.prev);
3205 spin_unlock_irqrestore(&n->list_lock, flags);
3208 kfree(slabs_by_inuse);
3211 EXPORT_SYMBOL(kmem_cache_shrink);
3213 #if defined(CONFIG_MEMORY_HOTPLUG)
3214 static int slab_mem_going_offline_callback(void *arg)
3216 struct kmem_cache *s;
3218 down_read(&slub_lock);
3219 list_for_each_entry(s, &slab_caches, list)
3220 kmem_cache_shrink(s);
3221 up_read(&slub_lock);
3226 static void slab_mem_offline_callback(void *arg)
3228 struct kmem_cache_node *n;
3229 struct kmem_cache *s;
3230 struct memory_notify *marg = arg;
3233 offline_node = marg->status_change_nid;
3236 * If the node still has available memory. we need kmem_cache_node
3239 if (offline_node < 0)
3242 down_read(&slub_lock);
3243 list_for_each_entry(s, &slab_caches, list) {
3244 n = get_node(s, offline_node);
3247 * if n->nr_slabs > 0, slabs still exist on the node
3248 * that is going down. We were unable to free them,
3249 * and offline_pages() function shouldn't call this
3250 * callback. So, we must fail.
3252 BUG_ON(slabs_node(s, offline_node));
3254 s->node[offline_node] = NULL;
3255 kmem_cache_free(kmem_cache_node, n);
3258 up_read(&slub_lock);
3261 static int slab_mem_going_online_callback(void *arg)
3263 struct kmem_cache_node *n;
3264 struct kmem_cache *s;
3265 struct memory_notify *marg = arg;
3266 int nid = marg->status_change_nid;
3270 * If the node's memory is already available, then kmem_cache_node is
3271 * already created. Nothing to do.
3277 * We are bringing a node online. No memory is available yet. We must
3278 * allocate a kmem_cache_node structure in order to bring the node
3281 down_read(&slub_lock);
3282 list_for_each_entry(s, &slab_caches, list) {
3284 * XXX: kmem_cache_alloc_node will fallback to other nodes
3285 * since memory is not yet available from the node that
3288 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3293 init_kmem_cache_node(n, s);
3297 up_read(&slub_lock);
3301 static int slab_memory_callback(struct notifier_block *self,
3302 unsigned long action, void *arg)
3307 case MEM_GOING_ONLINE:
3308 ret = slab_mem_going_online_callback(arg);
3310 case MEM_GOING_OFFLINE:
3311 ret = slab_mem_going_offline_callback(arg);
3314 case MEM_CANCEL_ONLINE:
3315 slab_mem_offline_callback(arg);
3318 case MEM_CANCEL_OFFLINE:
3322 ret = notifier_from_errno(ret);
3328 #endif /* CONFIG_MEMORY_HOTPLUG */
3330 /********************************************************************
3331 * Basic setup of slabs
3332 *******************************************************************/
3335 * Used for early kmem_cache structures that were allocated using
3336 * the page allocator
3339 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3343 list_add(&s->list, &slab_caches);
3346 for_each_node_state(node, N_NORMAL_MEMORY) {
3347 struct kmem_cache_node *n = get_node(s, node);
3351 list_for_each_entry(p, &n->partial, lru)
3354 #ifdef CONFIG_SLUB_DEBUG
3355 list_for_each_entry(p, &n->full, lru)
3362 void __init kmem_cache_init(void)
3366 struct kmem_cache *temp_kmem_cache;
3368 struct kmem_cache *temp_kmem_cache_node;
3369 unsigned long kmalloc_size;
3371 kmem_size = offsetof(struct kmem_cache, node) +
3372 nr_node_ids * sizeof(struct kmem_cache_node *);
3374 /* Allocate two kmem_caches from the page allocator */
3375 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3376 order = get_order(2 * kmalloc_size);
3377 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3380 * Must first have the slab cache available for the allocations of the
3381 * struct kmem_cache_node's. There is special bootstrap code in
3382 * kmem_cache_open for slab_state == DOWN.
3384 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3386 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3387 sizeof(struct kmem_cache_node),
3388 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3390 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3392 /* Able to allocate the per node structures */
3393 slab_state = PARTIAL;
3395 temp_kmem_cache = kmem_cache;
3396 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3397 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3398 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3399 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3402 * Allocate kmem_cache_node properly from the kmem_cache slab.
3403 * kmem_cache_node is separately allocated so no need to
3404 * update any list pointers.
3406 temp_kmem_cache_node = kmem_cache_node;
3408 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3409 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3411 kmem_cache_bootstrap_fixup(kmem_cache_node);
3414 kmem_cache_bootstrap_fixup(kmem_cache);
3416 /* Free temporary boot structure */
3417 free_pages((unsigned long)temp_kmem_cache, order);
3419 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3422 * Patch up the size_index table if we have strange large alignment
3423 * requirements for the kmalloc array. This is only the case for
3424 * MIPS it seems. The standard arches will not generate any code here.
3426 * Largest permitted alignment is 256 bytes due to the way we
3427 * handle the index determination for the smaller caches.
3429 * Make sure that nothing crazy happens if someone starts tinkering
3430 * around with ARCH_KMALLOC_MINALIGN
3432 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3433 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3435 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3436 int elem = size_index_elem(i);
3437 if (elem >= ARRAY_SIZE(size_index))
3439 size_index[elem] = KMALLOC_SHIFT_LOW;
3442 if (KMALLOC_MIN_SIZE == 64) {
3444 * The 96 byte size cache is not used if the alignment
3447 for (i = 64 + 8; i <= 96; i += 8)
3448 size_index[size_index_elem(i)] = 7;
3449 } else if (KMALLOC_MIN_SIZE == 128) {
3451 * The 192 byte sized cache is not used if the alignment
3452 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3455 for (i = 128 + 8; i <= 192; i += 8)
3456 size_index[size_index_elem(i)] = 8;
3459 /* Caches that are not of the two-to-the-power-of size */
3460 if (KMALLOC_MIN_SIZE <= 32) {
3461 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3465 if (KMALLOC_MIN_SIZE <= 64) {
3466 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3470 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3471 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3477 /* Provide the correct kmalloc names now that the caches are up */
3478 if (KMALLOC_MIN_SIZE <= 32) {
3479 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3480 BUG_ON(!kmalloc_caches[1]->name);
3483 if (KMALLOC_MIN_SIZE <= 64) {
3484 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3485 BUG_ON(!kmalloc_caches[2]->name);
3488 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3489 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3492 kmalloc_caches[i]->name = s;
3496 register_cpu_notifier(&slab_notifier);
3499 #ifdef CONFIG_ZONE_DMA
3500 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3501 struct kmem_cache *s = kmalloc_caches[i];
3504 char *name = kasprintf(GFP_NOWAIT,
3505 "dma-kmalloc-%d", s->objsize);
3508 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3509 s->objsize, SLAB_CACHE_DMA);
3514 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3515 " CPUs=%d, Nodes=%d\n",
3516 caches, cache_line_size(),
3517 slub_min_order, slub_max_order, slub_min_objects,
3518 nr_cpu_ids, nr_node_ids);
3521 void __init kmem_cache_init_late(void)
3526 * Find a mergeable slab cache
3528 static int slab_unmergeable(struct kmem_cache *s)
3530 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3537 * We may have set a slab to be unmergeable during bootstrap.
3539 if (s->refcount < 0)
3545 static struct kmem_cache *find_mergeable(size_t size,
3546 size_t align, unsigned long flags, const char *name,
3547 void (*ctor)(void *))
3549 struct kmem_cache *s;
3551 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3557 size = ALIGN(size, sizeof(void *));
3558 align = calculate_alignment(flags, align, size);
3559 size = ALIGN(size, align);
3560 flags = kmem_cache_flags(size, flags, name, NULL);
3562 list_for_each_entry(s, &slab_caches, list) {
3563 if (slab_unmergeable(s))
3569 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3572 * Check if alignment is compatible.
3573 * Courtesy of Adrian Drzewiecki
3575 if ((s->size & ~(align - 1)) != s->size)
3578 if (s->size - size >= sizeof(void *))
3586 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3587 size_t align, unsigned long flags, void (*ctor)(void *))
3589 struct kmem_cache *s;
3595 down_write(&slub_lock);
3596 s = find_mergeable(size, align, flags, name, ctor);
3600 * Adjust the object sizes so that we clear
3601 * the complete object on kzalloc.
3603 s->objsize = max(s->objsize, (int)size);
3604 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3606 if (sysfs_slab_alias(s, name)) {
3610 up_write(&slub_lock);
3614 n = kstrdup(name, GFP_KERNEL);
3618 s = kmalloc(kmem_size, GFP_KERNEL);
3620 if (kmem_cache_open(s, n,
3621 size, align, flags, ctor)) {
3622 list_add(&s->list, &slab_caches);
3623 if (sysfs_slab_add(s)) {
3629 up_write(&slub_lock);
3636 up_write(&slub_lock);
3638 if (flags & SLAB_PANIC)
3639 panic("Cannot create slabcache %s\n", name);
3644 EXPORT_SYMBOL(kmem_cache_create);
3648 * Use the cpu notifier to insure that the cpu slabs are flushed when
3651 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3652 unsigned long action, void *hcpu)
3654 long cpu = (long)hcpu;
3655 struct kmem_cache *s;
3656 unsigned long flags;
3659 case CPU_UP_CANCELED:
3660 case CPU_UP_CANCELED_FROZEN:
3662 case CPU_DEAD_FROZEN:
3663 down_read(&slub_lock);
3664 list_for_each_entry(s, &slab_caches, list) {
3665 local_irq_save(flags);
3666 __flush_cpu_slab(s, cpu);
3667 local_irq_restore(flags);
3669 up_read(&slub_lock);
3677 static struct notifier_block __cpuinitdata slab_notifier = {
3678 .notifier_call = slab_cpuup_callback
3683 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3685 struct kmem_cache *s;
3688 if (unlikely(size > SLUB_MAX_SIZE))
3689 return kmalloc_large(size, gfpflags);
3691 s = get_slab(size, gfpflags);
3693 if (unlikely(ZERO_OR_NULL_PTR(s)))
3696 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3698 /* Honor the call site pointer we received. */
3699 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3705 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3706 int node, unsigned long caller)
3708 struct kmem_cache *s;
3711 if (unlikely(size > SLUB_MAX_SIZE)) {
3712 ret = kmalloc_large_node(size, gfpflags, node);
3714 trace_kmalloc_node(caller, ret,
3715 size, PAGE_SIZE << get_order(size),
3721 s = get_slab(size, gfpflags);
3723 if (unlikely(ZERO_OR_NULL_PTR(s)))
3726 ret = slab_alloc(s, gfpflags, node, caller);
3728 /* Honor the call site pointer we received. */
3729 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3736 static int count_inuse(struct page *page)
3741 static int count_total(struct page *page)
3743 return page->objects;
3747 #ifdef CONFIG_SLUB_DEBUG
3748 static int validate_slab(struct kmem_cache *s, struct page *page,
3752 void *addr = page_address(page);
3754 if (!check_slab(s, page) ||
3755 !on_freelist(s, page, NULL))
3758 /* Now we know that a valid freelist exists */
3759 bitmap_zero(map, page->objects);
3761 get_map(s, page, map);
3762 for_each_object(p, s, addr, page->objects) {
3763 if (test_bit(slab_index(p, s, addr), map))
3764 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3768 for_each_object(p, s, addr, page->objects)
3769 if (!test_bit(slab_index(p, s, addr), map))
3770 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3775 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3779 validate_slab(s, page, map);
3783 static int validate_slab_node(struct kmem_cache *s,
3784 struct kmem_cache_node *n, unsigned long *map)
3786 unsigned long count = 0;
3788 unsigned long flags;
3790 spin_lock_irqsave(&n->list_lock, flags);
3792 list_for_each_entry(page, &n->partial, lru) {
3793 validate_slab_slab(s, page, map);
3796 if (count != n->nr_partial)
3797 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3798 "counter=%ld\n", s->name, count, n->nr_partial);
3800 if (!(s->flags & SLAB_STORE_USER))
3803 list_for_each_entry(page, &n->full, lru) {
3804 validate_slab_slab(s, page, map);
3807 if (count != atomic_long_read(&n->nr_slabs))
3808 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3809 "counter=%ld\n", s->name, count,
3810 atomic_long_read(&n->nr_slabs));
3813 spin_unlock_irqrestore(&n->list_lock, flags);
3817 static long validate_slab_cache(struct kmem_cache *s)
3820 unsigned long count = 0;
3821 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3822 sizeof(unsigned long), GFP_KERNEL);
3828 for_each_node_state(node, N_NORMAL_MEMORY) {
3829 struct kmem_cache_node *n = get_node(s, node);
3831 count += validate_slab_node(s, n, map);
3837 * Generate lists of code addresses where slabcache objects are allocated
3842 unsigned long count;
3849 DECLARE_BITMAP(cpus, NR_CPUS);
3855 unsigned long count;
3856 struct location *loc;
3859 static void free_loc_track(struct loc_track *t)
3862 free_pages((unsigned long)t->loc,
3863 get_order(sizeof(struct location) * t->max));
3866 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3871 order = get_order(sizeof(struct location) * max);
3873 l = (void *)__get_free_pages(flags, order);
3878 memcpy(l, t->loc, sizeof(struct location) * t->count);
3886 static int add_location(struct loc_track *t, struct kmem_cache *s,
3887 const struct track *track)
3889 long start, end, pos;
3891 unsigned long caddr;
3892 unsigned long age = jiffies - track->when;
3898 pos = start + (end - start + 1) / 2;
3901 * There is nothing at "end". If we end up there
3902 * we need to add something to before end.
3907 caddr = t->loc[pos].addr;
3908 if (track->addr == caddr) {
3914 if (age < l->min_time)
3916 if (age > l->max_time)
3919 if (track->pid < l->min_pid)
3920 l->min_pid = track->pid;
3921 if (track->pid > l->max_pid)
3922 l->max_pid = track->pid;
3924 cpumask_set_cpu(track->cpu,
3925 to_cpumask(l->cpus));
3927 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3931 if (track->addr < caddr)
3938 * Not found. Insert new tracking element.
3940 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3946 (t->count - pos) * sizeof(struct location));
3949 l->addr = track->addr;
3953 l->min_pid = track->pid;
3954 l->max_pid = track->pid;
3955 cpumask_clear(to_cpumask(l->cpus));
3956 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3957 nodes_clear(l->nodes);
3958 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3962 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3963 struct page *page, enum track_item alloc,
3966 void *addr = page_address(page);
3969 bitmap_zero(map, page->objects);
3970 get_map(s, page, map);
3972 for_each_object(p, s, addr, page->objects)
3973 if (!test_bit(slab_index(p, s, addr), map))
3974 add_location(t, s, get_track(s, p, alloc));
3977 static int list_locations(struct kmem_cache *s, char *buf,
3978 enum track_item alloc)
3982 struct loc_track t = { 0, 0, NULL };
3984 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3985 sizeof(unsigned long), GFP_KERNEL);
3987 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3990 return sprintf(buf, "Out of memory\n");
3992 /* Push back cpu slabs */
3995 for_each_node_state(node, N_NORMAL_MEMORY) {
3996 struct kmem_cache_node *n = get_node(s, node);
3997 unsigned long flags;
4000 if (!atomic_long_read(&n->nr_slabs))
4003 spin_lock_irqsave(&n->list_lock, flags);
4004 list_for_each_entry(page, &n->partial, lru)
4005 process_slab(&t, s, page, alloc, map);
4006 list_for_each_entry(page, &n->full, lru)
4007 process_slab(&t, s, page, alloc, map);
4008 spin_unlock_irqrestore(&n->list_lock, flags);
4011 for (i = 0; i < t.count; i++) {
4012 struct location *l = &t.loc[i];
4014 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4016 len += sprintf(buf + len, "%7ld ", l->count);
4019 len += sprintf(buf + len, "%pS", (void *)l->addr);
4021 len += sprintf(buf + len, "<not-available>");
4023 if (l->sum_time != l->min_time) {
4024 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4026 (long)div_u64(l->sum_time, l->count),
4029 len += sprintf(buf + len, " age=%ld",
4032 if (l->min_pid != l->max_pid)
4033 len += sprintf(buf + len, " pid=%ld-%ld",
4034 l->min_pid, l->max_pid);
4036 len += sprintf(buf + len, " pid=%ld",
4039 if (num_online_cpus() > 1 &&
4040 !cpumask_empty(to_cpumask(l->cpus)) &&
4041 len < PAGE_SIZE - 60) {
4042 len += sprintf(buf + len, " cpus=");
4043 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4044 to_cpumask(l->cpus));
4047 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4048 len < PAGE_SIZE - 60) {
4049 len += sprintf(buf + len, " nodes=");
4050 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4054 len += sprintf(buf + len, "\n");
4060 len += sprintf(buf, "No data\n");
4065 #ifdef SLUB_RESILIENCY_TEST
4066 static void resiliency_test(void)
4070 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4072 printk(KERN_ERR "SLUB resiliency testing\n");
4073 printk(KERN_ERR "-----------------------\n");
4074 printk(KERN_ERR "A. Corruption after allocation\n");
4076 p = kzalloc(16, GFP_KERNEL);
4078 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4079 " 0x12->0x%p\n\n", p + 16);
4081 validate_slab_cache(kmalloc_caches[4]);
4083 /* Hmmm... The next two are dangerous */
4084 p = kzalloc(32, GFP_KERNEL);
4085 p[32 + sizeof(void *)] = 0x34;
4086 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4087 " 0x34 -> -0x%p\n", p);
4089 "If allocated object is overwritten then not detectable\n\n");
4091 validate_slab_cache(kmalloc_caches[5]);
4092 p = kzalloc(64, GFP_KERNEL);
4093 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4095 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4098 "If allocated object is overwritten then not detectable\n\n");
4099 validate_slab_cache(kmalloc_caches[6]);
4101 printk(KERN_ERR "\nB. Corruption after free\n");
4102 p = kzalloc(128, GFP_KERNEL);
4105 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4106 validate_slab_cache(kmalloc_caches[7]);
4108 p = kzalloc(256, GFP_KERNEL);
4111 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4113 validate_slab_cache(kmalloc_caches[8]);
4115 p = kzalloc(512, GFP_KERNEL);
4118 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4119 validate_slab_cache(kmalloc_caches[9]);
4123 static void resiliency_test(void) {};
4128 enum slab_stat_type {
4129 SL_ALL, /* All slabs */
4130 SL_PARTIAL, /* Only partially allocated slabs */
4131 SL_CPU, /* Only slabs used for cpu caches */
4132 SL_OBJECTS, /* Determine allocated objects not slabs */
4133 SL_TOTAL /* Determine object capacity not slabs */
4136 #define SO_ALL (1 << SL_ALL)
4137 #define SO_PARTIAL (1 << SL_PARTIAL)
4138 #define SO_CPU (1 << SL_CPU)
4139 #define SO_OBJECTS (1 << SL_OBJECTS)
4140 #define SO_TOTAL (1 << SL_TOTAL)
4142 static ssize_t show_slab_objects(struct kmem_cache *s,
4143 char *buf, unsigned long flags)
4145 unsigned long total = 0;
4148 unsigned long *nodes;
4149 unsigned long *per_cpu;
4151 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4154 per_cpu = nodes + nr_node_ids;
4156 if (flags & SO_CPU) {
4159 for_each_possible_cpu(cpu) {
4160 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4162 if (!c || c->node < 0)
4166 if (flags & SO_TOTAL)
4167 x = c->page->objects;
4168 else if (flags & SO_OBJECTS)
4174 nodes[c->node] += x;
4180 lock_memory_hotplug();
4181 #ifdef CONFIG_SLUB_DEBUG
4182 if (flags & SO_ALL) {
4183 for_each_node_state(node, N_NORMAL_MEMORY) {
4184 struct kmem_cache_node *n = get_node(s, node);
4186 if (flags & SO_TOTAL)
4187 x = atomic_long_read(&n->total_objects);
4188 else if (flags & SO_OBJECTS)
4189 x = atomic_long_read(&n->total_objects) -
4190 count_partial(n, count_free);
4193 x = atomic_long_read(&n->nr_slabs);
4200 if (flags & SO_PARTIAL) {
4201 for_each_node_state(node, N_NORMAL_MEMORY) {
4202 struct kmem_cache_node *n = get_node(s, node);
4204 if (flags & SO_TOTAL)
4205 x = count_partial(n, count_total);
4206 else if (flags & SO_OBJECTS)
4207 x = count_partial(n, count_inuse);
4214 x = sprintf(buf, "%lu", total);
4216 for_each_node_state(node, N_NORMAL_MEMORY)
4218 x += sprintf(buf + x, " N%d=%lu",
4221 unlock_memory_hotplug();
4223 return x + sprintf(buf + x, "\n");
4226 #ifdef CONFIG_SLUB_DEBUG
4227 static int any_slab_objects(struct kmem_cache *s)
4231 for_each_online_node(node) {
4232 struct kmem_cache_node *n = get_node(s, node);
4237 if (atomic_long_read(&n->total_objects))
4244 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4245 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4247 struct slab_attribute {
4248 struct attribute attr;
4249 ssize_t (*show)(struct kmem_cache *s, char *buf);
4250 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4253 #define SLAB_ATTR_RO(_name) \
4254 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4256 #define SLAB_ATTR(_name) \
4257 static struct slab_attribute _name##_attr = \
4258 __ATTR(_name, 0644, _name##_show, _name##_store)
4260 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4262 return sprintf(buf, "%d\n", s->size);
4264 SLAB_ATTR_RO(slab_size);
4266 static ssize_t align_show(struct kmem_cache *s, char *buf)
4268 return sprintf(buf, "%d\n", s->align);
4270 SLAB_ATTR_RO(align);
4272 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4274 return sprintf(buf, "%d\n", s->objsize);
4276 SLAB_ATTR_RO(object_size);
4278 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4280 return sprintf(buf, "%d\n", oo_objects(s->oo));
4282 SLAB_ATTR_RO(objs_per_slab);
4284 static ssize_t order_store(struct kmem_cache *s,
4285 const char *buf, size_t length)
4287 unsigned long order;
4290 err = strict_strtoul(buf, 10, &order);
4294 if (order > slub_max_order || order < slub_min_order)
4297 calculate_sizes(s, order);
4301 static ssize_t order_show(struct kmem_cache *s, char *buf)
4303 return sprintf(buf, "%d\n", oo_order(s->oo));
4307 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4309 return sprintf(buf, "%lu\n", s->min_partial);
4312 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4318 err = strict_strtoul(buf, 10, &min);
4322 set_min_partial(s, min);
4325 SLAB_ATTR(min_partial);
4327 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4331 return sprintf(buf, "%pS\n", s->ctor);
4335 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4337 return sprintf(buf, "%d\n", s->refcount - 1);
4339 SLAB_ATTR_RO(aliases);
4341 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4343 return show_slab_objects(s, buf, SO_PARTIAL);
4345 SLAB_ATTR_RO(partial);
4347 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4349 return show_slab_objects(s, buf, SO_CPU);
4351 SLAB_ATTR_RO(cpu_slabs);
4353 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4355 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4357 SLAB_ATTR_RO(objects);
4359 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4361 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4363 SLAB_ATTR_RO(objects_partial);
4365 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4367 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4370 static ssize_t reclaim_account_store(struct kmem_cache *s,
4371 const char *buf, size_t length)
4373 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4375 s->flags |= SLAB_RECLAIM_ACCOUNT;
4378 SLAB_ATTR(reclaim_account);
4380 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4382 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4384 SLAB_ATTR_RO(hwcache_align);
4386 #ifdef CONFIG_ZONE_DMA
4387 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4389 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4391 SLAB_ATTR_RO(cache_dma);
4394 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4396 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4398 SLAB_ATTR_RO(destroy_by_rcu);
4400 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4402 return sprintf(buf, "%d\n", s->reserved);
4404 SLAB_ATTR_RO(reserved);
4406 #ifdef CONFIG_SLUB_DEBUG
4407 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4409 return show_slab_objects(s, buf, SO_ALL);
4411 SLAB_ATTR_RO(slabs);
4413 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4415 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4417 SLAB_ATTR_RO(total_objects);
4419 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4421 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4424 static ssize_t sanity_checks_store(struct kmem_cache *s,
4425 const char *buf, size_t length)
4427 s->flags &= ~SLAB_DEBUG_FREE;
4428 if (buf[0] == '1') {
4429 s->flags &= ~__CMPXCHG_DOUBLE;
4430 s->flags |= SLAB_DEBUG_FREE;
4434 SLAB_ATTR(sanity_checks);
4436 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4438 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4441 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4444 s->flags &= ~SLAB_TRACE;
4445 if (buf[0] == '1') {
4446 s->flags &= ~__CMPXCHG_DOUBLE;
4447 s->flags |= SLAB_TRACE;
4453 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4455 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4458 static ssize_t red_zone_store(struct kmem_cache *s,
4459 const char *buf, size_t length)
4461 if (any_slab_objects(s))
4464 s->flags &= ~SLAB_RED_ZONE;
4465 if (buf[0] == '1') {
4466 s->flags &= ~__CMPXCHG_DOUBLE;
4467 s->flags |= SLAB_RED_ZONE;
4469 calculate_sizes(s, -1);
4472 SLAB_ATTR(red_zone);
4474 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4476 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4479 static ssize_t poison_store(struct kmem_cache *s,
4480 const char *buf, size_t length)
4482 if (any_slab_objects(s))
4485 s->flags &= ~SLAB_POISON;
4486 if (buf[0] == '1') {
4487 s->flags &= ~__CMPXCHG_DOUBLE;
4488 s->flags |= SLAB_POISON;
4490 calculate_sizes(s, -1);
4495 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4497 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4500 static ssize_t store_user_store(struct kmem_cache *s,
4501 const char *buf, size_t length)
4503 if (any_slab_objects(s))
4506 s->flags &= ~SLAB_STORE_USER;
4507 if (buf[0] == '1') {
4508 s->flags &= ~__CMPXCHG_DOUBLE;
4509 s->flags |= SLAB_STORE_USER;
4511 calculate_sizes(s, -1);
4514 SLAB_ATTR(store_user);
4516 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4521 static ssize_t validate_store(struct kmem_cache *s,
4522 const char *buf, size_t length)
4526 if (buf[0] == '1') {
4527 ret = validate_slab_cache(s);
4533 SLAB_ATTR(validate);
4535 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4537 if (!(s->flags & SLAB_STORE_USER))
4539 return list_locations(s, buf, TRACK_ALLOC);
4541 SLAB_ATTR_RO(alloc_calls);
4543 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4545 if (!(s->flags & SLAB_STORE_USER))
4547 return list_locations(s, buf, TRACK_FREE);
4549 SLAB_ATTR_RO(free_calls);
4550 #endif /* CONFIG_SLUB_DEBUG */
4552 #ifdef CONFIG_FAILSLAB
4553 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4555 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4558 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4561 s->flags &= ~SLAB_FAILSLAB;
4563 s->flags |= SLAB_FAILSLAB;
4566 SLAB_ATTR(failslab);
4569 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4574 static ssize_t shrink_store(struct kmem_cache *s,
4575 const char *buf, size_t length)
4577 if (buf[0] == '1') {
4578 int rc = kmem_cache_shrink(s);
4589 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4591 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4594 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4595 const char *buf, size_t length)
4597 unsigned long ratio;
4600 err = strict_strtoul(buf, 10, &ratio);
4605 s->remote_node_defrag_ratio = ratio * 10;
4609 SLAB_ATTR(remote_node_defrag_ratio);
4612 #ifdef CONFIG_SLUB_STATS
4613 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4615 unsigned long sum = 0;
4618 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4623 for_each_online_cpu(cpu) {
4624 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4630 len = sprintf(buf, "%lu", sum);
4633 for_each_online_cpu(cpu) {
4634 if (data[cpu] && len < PAGE_SIZE - 20)
4635 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4639 return len + sprintf(buf + len, "\n");
4642 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4646 for_each_online_cpu(cpu)
4647 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4650 #define STAT_ATTR(si, text) \
4651 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4653 return show_stat(s, buf, si); \
4655 static ssize_t text##_store(struct kmem_cache *s, \
4656 const char *buf, size_t length) \
4658 if (buf[0] != '0') \
4660 clear_stat(s, si); \
4665 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4666 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4667 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4668 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4669 STAT_ATTR(FREE_FROZEN, free_frozen);
4670 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4671 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4672 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4673 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4674 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4675 STAT_ATTR(FREE_SLAB, free_slab);
4676 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4677 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4678 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4679 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4680 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4681 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4682 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4683 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4684 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4687 static struct attribute *slab_attrs[] = {
4688 &slab_size_attr.attr,
4689 &object_size_attr.attr,
4690 &objs_per_slab_attr.attr,
4692 &min_partial_attr.attr,
4694 &objects_partial_attr.attr,
4696 &cpu_slabs_attr.attr,
4700 &hwcache_align_attr.attr,
4701 &reclaim_account_attr.attr,
4702 &destroy_by_rcu_attr.attr,
4704 &reserved_attr.attr,
4705 #ifdef CONFIG_SLUB_DEBUG
4706 &total_objects_attr.attr,
4708 &sanity_checks_attr.attr,
4710 &red_zone_attr.attr,
4712 &store_user_attr.attr,
4713 &validate_attr.attr,
4714 &alloc_calls_attr.attr,
4715 &free_calls_attr.attr,
4717 #ifdef CONFIG_ZONE_DMA
4718 &cache_dma_attr.attr,
4721 &remote_node_defrag_ratio_attr.attr,
4723 #ifdef CONFIG_SLUB_STATS
4724 &alloc_fastpath_attr.attr,
4725 &alloc_slowpath_attr.attr,
4726 &free_fastpath_attr.attr,
4727 &free_slowpath_attr.attr,
4728 &free_frozen_attr.attr,
4729 &free_add_partial_attr.attr,
4730 &free_remove_partial_attr.attr,
4731 &alloc_from_partial_attr.attr,
4732 &alloc_slab_attr.attr,
4733 &alloc_refill_attr.attr,
4734 &free_slab_attr.attr,
4735 &cpuslab_flush_attr.attr,
4736 &deactivate_full_attr.attr,
4737 &deactivate_empty_attr.attr,
4738 &deactivate_to_head_attr.attr,
4739 &deactivate_to_tail_attr.attr,
4740 &deactivate_remote_frees_attr.attr,
4741 &order_fallback_attr.attr,
4742 &cmpxchg_double_fail_attr.attr,
4743 &cmpxchg_double_cpu_fail_attr.attr,
4745 #ifdef CONFIG_FAILSLAB
4746 &failslab_attr.attr,
4752 static struct attribute_group slab_attr_group = {
4753 .attrs = slab_attrs,
4756 static ssize_t slab_attr_show(struct kobject *kobj,
4757 struct attribute *attr,
4760 struct slab_attribute *attribute;
4761 struct kmem_cache *s;
4764 attribute = to_slab_attr(attr);
4767 if (!attribute->show)
4770 err = attribute->show(s, buf);
4775 static ssize_t slab_attr_store(struct kobject *kobj,
4776 struct attribute *attr,
4777 const char *buf, size_t len)
4779 struct slab_attribute *attribute;
4780 struct kmem_cache *s;
4783 attribute = to_slab_attr(attr);
4786 if (!attribute->store)
4789 err = attribute->store(s, buf, len);
4794 static void kmem_cache_release(struct kobject *kobj)
4796 struct kmem_cache *s = to_slab(kobj);
4802 static const struct sysfs_ops slab_sysfs_ops = {
4803 .show = slab_attr_show,
4804 .store = slab_attr_store,
4807 static struct kobj_type slab_ktype = {
4808 .sysfs_ops = &slab_sysfs_ops,
4809 .release = kmem_cache_release
4812 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4814 struct kobj_type *ktype = get_ktype(kobj);
4816 if (ktype == &slab_ktype)
4821 static const struct kset_uevent_ops slab_uevent_ops = {
4822 .filter = uevent_filter,
4825 static struct kset *slab_kset;
4827 #define ID_STR_LENGTH 64
4829 /* Create a unique string id for a slab cache:
4831 * Format :[flags-]size
4833 static char *create_unique_id(struct kmem_cache *s)
4835 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4842 * First flags affecting slabcache operations. We will only
4843 * get here for aliasable slabs so we do not need to support
4844 * too many flags. The flags here must cover all flags that
4845 * are matched during merging to guarantee that the id is
4848 if (s->flags & SLAB_CACHE_DMA)
4850 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4852 if (s->flags & SLAB_DEBUG_FREE)
4854 if (!(s->flags & SLAB_NOTRACK))
4858 p += sprintf(p, "%07d", s->size);
4859 BUG_ON(p > name + ID_STR_LENGTH - 1);
4863 static int sysfs_slab_add(struct kmem_cache *s)
4869 if (slab_state < SYSFS)
4870 /* Defer until later */
4873 unmergeable = slab_unmergeable(s);
4876 * Slabcache can never be merged so we can use the name proper.
4877 * This is typically the case for debug situations. In that
4878 * case we can catch duplicate names easily.
4880 sysfs_remove_link(&slab_kset->kobj, s->name);
4884 * Create a unique name for the slab as a target
4887 name = create_unique_id(s);
4890 s->kobj.kset = slab_kset;
4891 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4893 kobject_put(&s->kobj);
4897 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4899 kobject_del(&s->kobj);
4900 kobject_put(&s->kobj);
4903 kobject_uevent(&s->kobj, KOBJ_ADD);
4905 /* Setup first alias */
4906 sysfs_slab_alias(s, s->name);
4912 static void sysfs_slab_remove(struct kmem_cache *s)
4914 if (slab_state < SYSFS)
4916 * Sysfs has not been setup yet so no need to remove the
4921 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4922 kobject_del(&s->kobj);
4923 kobject_put(&s->kobj);
4927 * Need to buffer aliases during bootup until sysfs becomes
4928 * available lest we lose that information.
4930 struct saved_alias {
4931 struct kmem_cache *s;
4933 struct saved_alias *next;
4936 static struct saved_alias *alias_list;
4938 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4940 struct saved_alias *al;
4942 if (slab_state == SYSFS) {
4944 * If we have a leftover link then remove it.
4946 sysfs_remove_link(&slab_kset->kobj, name);
4947 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4950 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4956 al->next = alias_list;
4961 static int __init slab_sysfs_init(void)
4963 struct kmem_cache *s;
4966 down_write(&slub_lock);
4968 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4970 up_write(&slub_lock);
4971 printk(KERN_ERR "Cannot register slab subsystem.\n");
4977 list_for_each_entry(s, &slab_caches, list) {
4978 err = sysfs_slab_add(s);
4980 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4981 " to sysfs\n", s->name);
4984 while (alias_list) {
4985 struct saved_alias *al = alias_list;
4987 alias_list = alias_list->next;
4988 err = sysfs_slab_alias(al->s, al->name);
4990 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4991 " %s to sysfs\n", s->name);
4995 up_write(&slub_lock);
5000 __initcall(slab_sysfs_init);
5001 #endif /* CONFIG_SYSFS */
5004 * The /proc/slabinfo ABI
5006 #ifdef CONFIG_SLABINFO
5007 static void print_slabinfo_header(struct seq_file *m)
5009 seq_puts(m, "slabinfo - version: 2.1\n");
5010 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5011 "<objperslab> <pagesperslab>");
5012 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5013 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5017 static void *s_start(struct seq_file *m, loff_t *pos)
5021 down_read(&slub_lock);
5023 print_slabinfo_header(m);
5025 return seq_list_start(&slab_caches, *pos);
5028 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5030 return seq_list_next(p, &slab_caches, pos);
5033 static void s_stop(struct seq_file *m, void *p)
5035 up_read(&slub_lock);
5038 static int s_show(struct seq_file *m, void *p)
5040 unsigned long nr_partials = 0;
5041 unsigned long nr_slabs = 0;
5042 unsigned long nr_inuse = 0;
5043 unsigned long nr_objs = 0;
5044 unsigned long nr_free = 0;
5045 struct kmem_cache *s;
5048 s = list_entry(p, struct kmem_cache, list);
5050 for_each_online_node(node) {
5051 struct kmem_cache_node *n = get_node(s, node);
5056 nr_partials += n->nr_partial;
5057 nr_slabs += atomic_long_read(&n->nr_slabs);
5058 nr_objs += atomic_long_read(&n->total_objects);
5059 nr_free += count_partial(n, count_free);
5062 nr_inuse = nr_objs - nr_free;
5064 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5065 nr_objs, s->size, oo_objects(s->oo),
5066 (1 << oo_order(s->oo)));
5067 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5068 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5074 static const struct seq_operations slabinfo_op = {
5081 static int slabinfo_open(struct inode *inode, struct file *file)
5083 return seq_open(file, &slabinfo_op);
5086 static const struct file_operations proc_slabinfo_operations = {
5087 .open = slabinfo_open,
5089 .llseek = seq_lseek,
5090 .release = seq_release,
5093 static int __init slab_proc_init(void)
5095 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5098 module_init(slab_proc_init);
5099 #endif /* CONFIG_SLABINFO */