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 /* Interrupts must be disabled (for the fallback code to work right) */
358 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
359 void *freelist_old, unsigned long counters_old,
360 void *freelist_new, unsigned long counters_new,
363 VM_BUG_ON(!irqs_disabled());
364 #ifdef CONFIG_CMPXCHG_DOUBLE
365 if (s->flags & __CMPXCHG_DOUBLE) {
366 if (cmpxchg_double(&page->freelist,
367 freelist_old, counters_old,
368 freelist_new, counters_new))
374 if (page->freelist == freelist_old && page->counters == counters_old) {
375 page->freelist = freelist_new;
376 page->counters = counters_new;
384 stat(s, CMPXCHG_DOUBLE_FAIL);
386 #ifdef SLUB_DEBUG_CMPXCHG
387 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
393 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
394 void *freelist_old, unsigned long counters_old,
395 void *freelist_new, unsigned long counters_new,
398 #ifdef CONFIG_CMPXCHG_DOUBLE
399 if (s->flags & __CMPXCHG_DOUBLE) {
400 if (cmpxchg_double(&page->freelist,
401 freelist_old, counters_old,
402 freelist_new, counters_new))
409 local_irq_save(flags);
411 if (page->freelist == freelist_old && page->counters == counters_old) {
412 page->freelist = freelist_new;
413 page->counters = counters_new;
415 local_irq_restore(flags);
419 local_irq_restore(flags);
423 stat(s, CMPXCHG_DOUBLE_FAIL);
425 #ifdef SLUB_DEBUG_CMPXCHG
426 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432 #ifdef CONFIG_SLUB_DEBUG
434 * Determine a map of object in use on a page.
436 * Node listlock must be held to guarantee that the page does
437 * not vanish from under us.
439 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
442 void *addr = page_address(page);
444 for (p = page->freelist; p; p = get_freepointer(s, p))
445 set_bit(slab_index(p, s, addr), map);
451 #ifdef CONFIG_SLUB_DEBUG_ON
452 static int slub_debug = DEBUG_DEFAULT_FLAGS;
454 static int slub_debug;
457 static char *slub_debug_slabs;
458 static int disable_higher_order_debug;
463 static void print_section(char *text, u8 *addr, unsigned int length)
471 for (i = 0; i < length; i++) {
473 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
476 printk(KERN_CONT " %02x", addr[i]);
478 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
480 printk(KERN_CONT " %s\n", ascii);
487 printk(KERN_CONT " ");
491 printk(KERN_CONT " %s\n", ascii);
495 static struct track *get_track(struct kmem_cache *s, void *object,
496 enum track_item alloc)
501 p = object + s->offset + sizeof(void *);
503 p = object + s->inuse;
508 static void set_track(struct kmem_cache *s, void *object,
509 enum track_item alloc, unsigned long addr)
511 struct track *p = get_track(s, object, alloc);
515 p->cpu = smp_processor_id();
516 p->pid = current->pid;
519 memset(p, 0, sizeof(struct track));
522 static void init_tracking(struct kmem_cache *s, void *object)
524 if (!(s->flags & SLAB_STORE_USER))
527 set_track(s, object, TRACK_FREE, 0UL);
528 set_track(s, object, TRACK_ALLOC, 0UL);
531 static void print_track(const char *s, struct track *t)
536 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
537 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
540 static void print_tracking(struct kmem_cache *s, void *object)
542 if (!(s->flags & SLAB_STORE_USER))
545 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
546 print_track("Freed", get_track(s, object, TRACK_FREE));
549 static void print_page_info(struct page *page)
551 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
552 page, page->objects, page->inuse, page->freelist, page->flags);
556 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
562 vsnprintf(buf, sizeof(buf), fmt, args);
564 printk(KERN_ERR "========================================"
565 "=====================================\n");
566 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
567 printk(KERN_ERR "----------------------------------------"
568 "-------------------------------------\n\n");
571 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
577 vsnprintf(buf, sizeof(buf), fmt, args);
579 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
582 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
584 unsigned int off; /* Offset of last byte */
585 u8 *addr = page_address(page);
587 print_tracking(s, p);
589 print_page_info(page);
591 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
592 p, p - addr, get_freepointer(s, p));
595 print_section("Bytes b4", p - 16, 16);
597 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
599 if (s->flags & SLAB_RED_ZONE)
600 print_section("Redzone", p + s->objsize,
601 s->inuse - s->objsize);
604 off = s->offset + sizeof(void *);
608 if (s->flags & SLAB_STORE_USER)
609 off += 2 * sizeof(struct track);
612 /* Beginning of the filler is the free pointer */
613 print_section("Padding", p + off, s->size - off);
618 static void object_err(struct kmem_cache *s, struct page *page,
619 u8 *object, char *reason)
621 slab_bug(s, "%s", reason);
622 print_trailer(s, page, object);
625 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
631 vsnprintf(buf, sizeof(buf), fmt, args);
633 slab_bug(s, "%s", buf);
634 print_page_info(page);
638 static void init_object(struct kmem_cache *s, void *object, u8 val)
642 if (s->flags & __OBJECT_POISON) {
643 memset(p, POISON_FREE, s->objsize - 1);
644 p[s->objsize - 1] = POISON_END;
647 if (s->flags & SLAB_RED_ZONE)
648 memset(p + s->objsize, val, s->inuse - s->objsize);
651 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
654 if (*start != (u8)value)
662 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
663 void *from, void *to)
665 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
666 memset(from, data, to - from);
669 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
670 u8 *object, char *what,
671 u8 *start, unsigned int value, unsigned int bytes)
676 fault = check_bytes(start, value, bytes);
681 while (end > fault && end[-1] == value)
684 slab_bug(s, "%s overwritten", what);
685 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
686 fault, end - 1, fault[0], value);
687 print_trailer(s, page, object);
689 restore_bytes(s, what, value, fault, end);
697 * Bytes of the object to be managed.
698 * If the freepointer may overlay the object then the free
699 * pointer is the first word of the object.
701 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
704 * object + s->objsize
705 * Padding to reach word boundary. This is also used for Redzoning.
706 * Padding is extended by another word if Redzoning is enabled and
709 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
710 * 0xcc (RED_ACTIVE) for objects in use.
713 * Meta data starts here.
715 * A. Free pointer (if we cannot overwrite object on free)
716 * B. Tracking data for SLAB_STORE_USER
717 * C. Padding to reach required alignment boundary or at mininum
718 * one word if debugging is on to be able to detect writes
719 * before the word boundary.
721 * Padding is done using 0x5a (POISON_INUSE)
724 * Nothing is used beyond s->size.
726 * If slabcaches are merged then the objsize and inuse boundaries are mostly
727 * ignored. And therefore no slab options that rely on these boundaries
728 * may be used with merged slabcaches.
731 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
733 unsigned long off = s->inuse; /* The end of info */
736 /* Freepointer is placed after the object. */
737 off += sizeof(void *);
739 if (s->flags & SLAB_STORE_USER)
740 /* We also have user information there */
741 off += 2 * sizeof(struct track);
746 return check_bytes_and_report(s, page, p, "Object padding",
747 p + off, POISON_INUSE, s->size - off);
750 /* Check the pad bytes at the end of a slab page */
751 static int slab_pad_check(struct kmem_cache *s, struct page *page)
759 if (!(s->flags & SLAB_POISON))
762 start = page_address(page);
763 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
764 end = start + length;
765 remainder = length % s->size;
769 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
772 while (end > fault && end[-1] == POISON_INUSE)
775 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
776 print_section("Padding", end - remainder, remainder);
778 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
782 static int check_object(struct kmem_cache *s, struct page *page,
783 void *object, u8 val)
786 u8 *endobject = object + s->objsize;
788 if (s->flags & SLAB_RED_ZONE) {
789 if (!check_bytes_and_report(s, page, object, "Redzone",
790 endobject, val, s->inuse - s->objsize))
793 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
794 check_bytes_and_report(s, page, p, "Alignment padding",
795 endobject, POISON_INUSE, s->inuse - s->objsize);
799 if (s->flags & SLAB_POISON) {
800 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
801 (!check_bytes_and_report(s, page, p, "Poison", p,
802 POISON_FREE, s->objsize - 1) ||
803 !check_bytes_and_report(s, page, p, "Poison",
804 p + s->objsize - 1, POISON_END, 1)))
807 * check_pad_bytes cleans up on its own.
809 check_pad_bytes(s, page, p);
812 if (!s->offset && val == SLUB_RED_ACTIVE)
814 * Object and freepointer overlap. Cannot check
815 * freepointer while object is allocated.
819 /* Check free pointer validity */
820 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
821 object_err(s, page, p, "Freepointer corrupt");
823 * No choice but to zap it and thus lose the remainder
824 * of the free objects in this slab. May cause
825 * another error because the object count is now wrong.
827 set_freepointer(s, p, NULL);
833 static int check_slab(struct kmem_cache *s, struct page *page)
837 VM_BUG_ON(!irqs_disabled());
839 if (!PageSlab(page)) {
840 slab_err(s, page, "Not a valid slab page");
844 maxobj = order_objects(compound_order(page), s->size, s->reserved);
845 if (page->objects > maxobj) {
846 slab_err(s, page, "objects %u > max %u",
847 s->name, page->objects, maxobj);
850 if (page->inuse > page->objects) {
851 slab_err(s, page, "inuse %u > max %u",
852 s->name, page->inuse, page->objects);
855 /* Slab_pad_check fixes things up after itself */
856 slab_pad_check(s, page);
861 * Determine if a certain object on a page is on the freelist. Must hold the
862 * slab lock to guarantee that the chains are in a consistent state.
864 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
869 unsigned long max_objects;
872 while (fp && nr <= page->objects) {
875 if (!check_valid_pointer(s, page, fp)) {
877 object_err(s, page, object,
878 "Freechain corrupt");
879 set_freepointer(s, object, NULL);
882 slab_err(s, page, "Freepointer corrupt");
883 page->freelist = NULL;
884 page->inuse = page->objects;
885 slab_fix(s, "Freelist cleared");
891 fp = get_freepointer(s, object);
895 max_objects = order_objects(compound_order(page), s->size, s->reserved);
896 if (max_objects > MAX_OBJS_PER_PAGE)
897 max_objects = MAX_OBJS_PER_PAGE;
899 if (page->objects != max_objects) {
900 slab_err(s, page, "Wrong number of objects. Found %d but "
901 "should be %d", page->objects, max_objects);
902 page->objects = max_objects;
903 slab_fix(s, "Number of objects adjusted.");
905 if (page->inuse != page->objects - nr) {
906 slab_err(s, page, "Wrong object count. Counter is %d but "
907 "counted were %d", page->inuse, page->objects - nr);
908 page->inuse = page->objects - nr;
909 slab_fix(s, "Object count adjusted.");
911 return search == NULL;
914 static void trace(struct kmem_cache *s, struct page *page, void *object,
917 if (s->flags & SLAB_TRACE) {
918 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
920 alloc ? "alloc" : "free",
925 print_section("Object", (void *)object, s->objsize);
932 * Hooks for other subsystems that check memory allocations. In a typical
933 * production configuration these hooks all should produce no code at all.
935 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
937 flags &= gfp_allowed_mask;
938 lockdep_trace_alloc(flags);
939 might_sleep_if(flags & __GFP_WAIT);
941 return should_failslab(s->objsize, flags, s->flags);
944 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
946 flags &= gfp_allowed_mask;
947 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
948 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
951 static inline void slab_free_hook(struct kmem_cache *s, void *x)
953 kmemleak_free_recursive(x, s->flags);
956 * Trouble is that we may no longer disable interupts in the fast path
957 * So in order to make the debug calls that expect irqs to be
958 * disabled we need to disable interrupts temporarily.
960 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
964 local_irq_save(flags);
965 kmemcheck_slab_free(s, x, s->objsize);
966 debug_check_no_locks_freed(x, s->objsize);
967 local_irq_restore(flags);
970 if (!(s->flags & SLAB_DEBUG_OBJECTS))
971 debug_check_no_obj_freed(x, s->objsize);
975 * Tracking of fully allocated slabs for debugging purposes.
977 * list_lock must be held.
979 static void add_full(struct kmem_cache *s,
980 struct kmem_cache_node *n, struct page *page)
982 if (!(s->flags & SLAB_STORE_USER))
985 list_add(&page->lru, &n->full);
989 * list_lock must be held.
991 static void remove_full(struct kmem_cache *s, struct page *page)
993 if (!(s->flags & SLAB_STORE_USER))
996 list_del(&page->lru);
999 /* Tracking of the number of slabs for debugging purposes */
1000 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1002 struct kmem_cache_node *n = get_node(s, node);
1004 return atomic_long_read(&n->nr_slabs);
1007 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1009 return atomic_long_read(&n->nr_slabs);
1012 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1014 struct kmem_cache_node *n = get_node(s, node);
1017 * May be called early in order to allocate a slab for the
1018 * kmem_cache_node structure. Solve the chicken-egg
1019 * dilemma by deferring the increment of the count during
1020 * bootstrap (see early_kmem_cache_node_alloc).
1023 atomic_long_inc(&n->nr_slabs);
1024 atomic_long_add(objects, &n->total_objects);
1027 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1029 struct kmem_cache_node *n = get_node(s, node);
1031 atomic_long_dec(&n->nr_slabs);
1032 atomic_long_sub(objects, &n->total_objects);
1035 /* Object debug checks for alloc/free paths */
1036 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1039 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1042 init_object(s, object, SLUB_RED_INACTIVE);
1043 init_tracking(s, object);
1046 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1047 void *object, unsigned long addr)
1049 if (!check_slab(s, page))
1052 if (!check_valid_pointer(s, page, object)) {
1053 object_err(s, page, object, "Freelist Pointer check fails");
1057 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1060 /* Success perform special debug activities for allocs */
1061 if (s->flags & SLAB_STORE_USER)
1062 set_track(s, object, TRACK_ALLOC, addr);
1063 trace(s, page, object, 1);
1064 init_object(s, object, SLUB_RED_ACTIVE);
1068 if (PageSlab(page)) {
1070 * If this is a slab page then lets do the best we can
1071 * to avoid issues in the future. Marking all objects
1072 * as used avoids touching the remaining objects.
1074 slab_fix(s, "Marking all objects used");
1075 page->inuse = page->objects;
1076 page->freelist = NULL;
1081 static noinline int free_debug_processing(struct kmem_cache *s,
1082 struct page *page, void *object, unsigned long addr)
1084 unsigned long flags;
1087 local_irq_save(flags);
1090 if (!check_slab(s, page))
1093 if (!check_valid_pointer(s, page, object)) {
1094 slab_err(s, page, "Invalid object pointer 0x%p", object);
1098 if (on_freelist(s, page, object)) {
1099 object_err(s, page, object, "Object already free");
1103 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1106 if (unlikely(s != page->slab)) {
1107 if (!PageSlab(page)) {
1108 slab_err(s, page, "Attempt to free object(0x%p) "
1109 "outside of slab", object);
1110 } else if (!page->slab) {
1112 "SLUB <none>: no slab for object 0x%p.\n",
1116 object_err(s, page, object,
1117 "page slab pointer corrupt.");
1121 if (s->flags & SLAB_STORE_USER)
1122 set_track(s, object, TRACK_FREE, addr);
1123 trace(s, page, object, 0);
1124 init_object(s, object, SLUB_RED_INACTIVE);
1128 local_irq_restore(flags);
1132 slab_fix(s, "Object at 0x%p not freed", object);
1136 static int __init setup_slub_debug(char *str)
1138 slub_debug = DEBUG_DEFAULT_FLAGS;
1139 if (*str++ != '=' || !*str)
1141 * No options specified. Switch on full debugging.
1147 * No options but restriction on slabs. This means full
1148 * debugging for slabs matching a pattern.
1152 if (tolower(*str) == 'o') {
1154 * Avoid enabling debugging on caches if its minimum order
1155 * would increase as a result.
1157 disable_higher_order_debug = 1;
1164 * Switch off all debugging measures.
1169 * Determine which debug features should be switched on
1171 for (; *str && *str != ','; str++) {
1172 switch (tolower(*str)) {
1174 slub_debug |= SLAB_DEBUG_FREE;
1177 slub_debug |= SLAB_RED_ZONE;
1180 slub_debug |= SLAB_POISON;
1183 slub_debug |= SLAB_STORE_USER;
1186 slub_debug |= SLAB_TRACE;
1189 slub_debug |= SLAB_FAILSLAB;
1192 printk(KERN_ERR "slub_debug option '%c' "
1193 "unknown. skipped\n", *str);
1199 slub_debug_slabs = str + 1;
1204 __setup("slub_debug", setup_slub_debug);
1206 static unsigned long kmem_cache_flags(unsigned long objsize,
1207 unsigned long flags, const char *name,
1208 void (*ctor)(void *))
1211 * Enable debugging if selected on the kernel commandline.
1213 if (slub_debug && (!slub_debug_slabs ||
1214 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1215 flags |= slub_debug;
1220 static inline void setup_object_debug(struct kmem_cache *s,
1221 struct page *page, void *object) {}
1223 static inline int alloc_debug_processing(struct kmem_cache *s,
1224 struct page *page, void *object, unsigned long addr) { return 0; }
1226 static inline int free_debug_processing(struct kmem_cache *s,
1227 struct page *page, void *object, unsigned long addr) { return 0; }
1229 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1231 static inline int check_object(struct kmem_cache *s, struct page *page,
1232 void *object, u8 val) { return 1; }
1233 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1234 struct page *page) {}
1235 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1236 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1237 unsigned long flags, const char *name,
1238 void (*ctor)(void *))
1242 #define slub_debug 0
1244 #define disable_higher_order_debug 0
1246 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1248 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1250 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1252 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1255 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1258 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1261 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1263 #endif /* CONFIG_SLUB_DEBUG */
1266 * Slab allocation and freeing
1268 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1269 struct kmem_cache_order_objects oo)
1271 int order = oo_order(oo);
1273 flags |= __GFP_NOTRACK;
1275 if (node == NUMA_NO_NODE)
1276 return alloc_pages(flags, order);
1278 return alloc_pages_exact_node(node, flags, order);
1281 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1284 struct kmem_cache_order_objects oo = s->oo;
1287 flags &= gfp_allowed_mask;
1289 if (flags & __GFP_WAIT)
1292 flags |= s->allocflags;
1295 * Let the initial higher-order allocation fail under memory pressure
1296 * so we fall-back to the minimum order allocation.
1298 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1300 page = alloc_slab_page(alloc_gfp, node, oo);
1301 if (unlikely(!page)) {
1304 * Allocation may have failed due to fragmentation.
1305 * Try a lower order alloc if possible
1307 page = alloc_slab_page(flags, node, oo);
1310 stat(s, ORDER_FALLBACK);
1313 if (flags & __GFP_WAIT)
1314 local_irq_disable();
1319 if (kmemcheck_enabled
1320 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1321 int pages = 1 << oo_order(oo);
1323 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1326 * Objects from caches that have a constructor don't get
1327 * cleared when they're allocated, so we need to do it here.
1330 kmemcheck_mark_uninitialized_pages(page, pages);
1332 kmemcheck_mark_unallocated_pages(page, pages);
1335 page->objects = oo_objects(oo);
1336 mod_zone_page_state(page_zone(page),
1337 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1338 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1344 static void setup_object(struct kmem_cache *s, struct page *page,
1347 setup_object_debug(s, page, object);
1348 if (unlikely(s->ctor))
1352 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1359 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1361 page = allocate_slab(s,
1362 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1366 inc_slabs_node(s, page_to_nid(page), page->objects);
1368 page->flags |= 1 << PG_slab;
1370 start = page_address(page);
1372 if (unlikely(s->flags & SLAB_POISON))
1373 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1376 for_each_object(p, s, start, page->objects) {
1377 setup_object(s, page, last);
1378 set_freepointer(s, last, p);
1381 setup_object(s, page, last);
1382 set_freepointer(s, last, NULL);
1384 page->freelist = start;
1391 static void __free_slab(struct kmem_cache *s, struct page *page)
1393 int order = compound_order(page);
1394 int pages = 1 << order;
1396 if (kmem_cache_debug(s)) {
1399 slab_pad_check(s, page);
1400 for_each_object(p, s, page_address(page),
1402 check_object(s, page, p, SLUB_RED_INACTIVE);
1405 kmemcheck_free_shadow(page, compound_order(page));
1407 mod_zone_page_state(page_zone(page),
1408 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1409 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1412 __ClearPageSlab(page);
1413 reset_page_mapcount(page);
1414 if (current->reclaim_state)
1415 current->reclaim_state->reclaimed_slab += pages;
1416 __free_pages(page, order);
1419 #define need_reserve_slab_rcu \
1420 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1422 static void rcu_free_slab(struct rcu_head *h)
1426 if (need_reserve_slab_rcu)
1427 page = virt_to_head_page(h);
1429 page = container_of((struct list_head *)h, struct page, lru);
1431 __free_slab(page->slab, page);
1434 static void free_slab(struct kmem_cache *s, struct page *page)
1436 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1437 struct rcu_head *head;
1439 if (need_reserve_slab_rcu) {
1440 int order = compound_order(page);
1441 int offset = (PAGE_SIZE << order) - s->reserved;
1443 VM_BUG_ON(s->reserved != sizeof(*head));
1444 head = page_address(page) + offset;
1447 * RCU free overloads the RCU head over the LRU
1449 head = (void *)&page->lru;
1452 call_rcu(head, rcu_free_slab);
1454 __free_slab(s, page);
1457 static void discard_slab(struct kmem_cache *s, struct page *page)
1459 dec_slabs_node(s, page_to_nid(page), page->objects);
1464 * Management of partially allocated slabs.
1466 * list_lock must be held.
1468 static inline void add_partial(struct kmem_cache_node *n,
1469 struct page *page, int tail)
1473 list_add_tail(&page->lru, &n->partial);
1475 list_add(&page->lru, &n->partial);
1479 * list_lock must be held.
1481 static inline void remove_partial(struct kmem_cache_node *n,
1484 list_del(&page->lru);
1489 * Lock slab, remove from the partial list and put the object into the
1492 * Must hold list_lock.
1494 static inline int acquire_slab(struct kmem_cache *s,
1495 struct kmem_cache_node *n, struct page *page)
1498 unsigned long counters;
1502 * Zap the freelist and set the frozen bit.
1503 * The old freelist is the list of objects for the
1504 * per cpu allocation list.
1507 freelist = page->freelist;
1508 counters = page->counters;
1509 new.counters = counters;
1510 new.inuse = page->objects;
1512 VM_BUG_ON(new.frozen);
1515 } while (!__cmpxchg_double_slab(s, page,
1518 "lock and freeze"));
1520 remove_partial(n, page);
1523 /* Populate the per cpu freelist */
1524 this_cpu_write(s->cpu_slab->freelist, freelist);
1525 this_cpu_write(s->cpu_slab->page, page);
1526 this_cpu_write(s->cpu_slab->node, page_to_nid(page));
1530 * Slab page came from the wrong list. No object to allocate
1531 * from. Put it onto the correct list and continue partial
1534 printk(KERN_ERR "SLUB: %s : Page without available objects on"
1535 " partial list\n", s->name);
1541 * Try to allocate a partial slab from a specific node.
1543 static struct page *get_partial_node(struct kmem_cache *s,
1544 struct kmem_cache_node *n)
1549 * Racy check. If we mistakenly see no partial slabs then we
1550 * just allocate an empty slab. If we mistakenly try to get a
1551 * partial slab and there is none available then get_partials()
1554 if (!n || !n->nr_partial)
1557 spin_lock(&n->list_lock);
1558 list_for_each_entry(page, &n->partial, lru)
1559 if (acquire_slab(s, n, page))
1563 spin_unlock(&n->list_lock);
1568 * Get a page from somewhere. Search in increasing NUMA distances.
1570 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1573 struct zonelist *zonelist;
1576 enum zone_type high_zoneidx = gfp_zone(flags);
1580 * The defrag ratio allows a configuration of the tradeoffs between
1581 * inter node defragmentation and node local allocations. A lower
1582 * defrag_ratio increases the tendency to do local allocations
1583 * instead of attempting to obtain partial slabs from other nodes.
1585 * If the defrag_ratio is set to 0 then kmalloc() always
1586 * returns node local objects. If the ratio is higher then kmalloc()
1587 * may return off node objects because partial slabs are obtained
1588 * from other nodes and filled up.
1590 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1591 * defrag_ratio = 1000) then every (well almost) allocation will
1592 * first attempt to defrag slab caches on other nodes. This means
1593 * scanning over all nodes to look for partial slabs which may be
1594 * expensive if we do it every time we are trying to find a slab
1595 * with available objects.
1597 if (!s->remote_node_defrag_ratio ||
1598 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1602 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1603 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1604 struct kmem_cache_node *n;
1606 n = get_node(s, zone_to_nid(zone));
1608 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1609 n->nr_partial > s->min_partial) {
1610 page = get_partial_node(s, n);
1623 * Get a partial page, lock it and return it.
1625 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1628 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1630 page = get_partial_node(s, get_node(s, searchnode));
1631 if (page || node != NUMA_NO_NODE)
1634 return get_any_partial(s, flags);
1637 #ifdef CONFIG_PREEMPT
1639 * Calculate the next globally unique transaction for disambiguiation
1640 * during cmpxchg. The transactions start with the cpu number and are then
1641 * incremented by CONFIG_NR_CPUS.
1643 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1646 * No preemption supported therefore also no need to check for
1652 static inline unsigned long next_tid(unsigned long tid)
1654 return tid + TID_STEP;
1657 static inline unsigned int tid_to_cpu(unsigned long tid)
1659 return tid % TID_STEP;
1662 static inline unsigned long tid_to_event(unsigned long tid)
1664 return tid / TID_STEP;
1667 static inline unsigned int init_tid(int cpu)
1672 static inline void note_cmpxchg_failure(const char *n,
1673 const struct kmem_cache *s, unsigned long tid)
1675 #ifdef SLUB_DEBUG_CMPXCHG
1676 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1678 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1680 #ifdef CONFIG_PREEMPT
1681 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1682 printk("due to cpu change %d -> %d\n",
1683 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1686 if (tid_to_event(tid) != tid_to_event(actual_tid))
1687 printk("due to cpu running other code. Event %ld->%ld\n",
1688 tid_to_event(tid), tid_to_event(actual_tid));
1690 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1691 actual_tid, tid, next_tid(tid));
1693 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1696 void init_kmem_cache_cpus(struct kmem_cache *s)
1700 for_each_possible_cpu(cpu)
1701 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1704 * Remove the cpu slab
1708 * Remove the cpu slab
1710 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1712 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1713 struct page *page = c->page;
1714 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1716 enum slab_modes l = M_NONE, m = M_NONE;
1723 if (page->freelist) {
1724 stat(s, DEACTIVATE_REMOTE_FREES);
1728 c->tid = next_tid(c->tid);
1730 freelist = c->freelist;
1734 * Stage one: Free all available per cpu objects back
1735 * to the page freelist while it is still frozen. Leave the
1738 * There is no need to take the list->lock because the page
1741 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1743 unsigned long counters;
1746 prior = page->freelist;
1747 counters = page->counters;
1748 set_freepointer(s, freelist, prior);
1749 new.counters = counters;
1751 VM_BUG_ON(!new.frozen);
1753 } while (!__cmpxchg_double_slab(s, page,
1755 freelist, new.counters,
1756 "drain percpu freelist"));
1758 freelist = nextfree;
1762 * Stage two: Ensure that the page is unfrozen while the
1763 * list presence reflects the actual number of objects
1766 * We setup the list membership and then perform a cmpxchg
1767 * with the count. If there is a mismatch then the page
1768 * is not unfrozen but the page is on the wrong list.
1770 * Then we restart the process which may have to remove
1771 * the page from the list that we just put it on again
1772 * because the number of objects in the slab may have
1777 old.freelist = page->freelist;
1778 old.counters = page->counters;
1779 VM_BUG_ON(!old.frozen);
1781 /* Determine target state of the slab */
1782 new.counters = old.counters;
1785 set_freepointer(s, freelist, old.freelist);
1786 new.freelist = freelist;
1788 new.freelist = old.freelist;
1792 if (!new.inuse && n->nr_partial < s->min_partial)
1794 else if (new.freelist) {
1799 * Taking the spinlock removes the possiblity
1800 * that acquire_slab() will see a slab page that
1803 spin_lock(&n->list_lock);
1807 if (kmem_cache_debug(s) && !lock) {
1810 * This also ensures that the scanning of full
1811 * slabs from diagnostic functions will not see
1814 spin_lock(&n->list_lock);
1822 remove_partial(n, page);
1824 else if (l == M_FULL)
1826 remove_full(s, page);
1828 if (m == M_PARTIAL) {
1830 add_partial(n, page, tail);
1831 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1833 } else if (m == M_FULL) {
1835 stat(s, DEACTIVATE_FULL);
1836 add_full(s, n, page);
1842 if (!__cmpxchg_double_slab(s, page,
1843 old.freelist, old.counters,
1844 new.freelist, new.counters,
1849 spin_unlock(&n->list_lock);
1852 stat(s, DEACTIVATE_EMPTY);
1853 discard_slab(s, page);
1858 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1860 stat(s, CPUSLAB_FLUSH);
1861 deactivate_slab(s, c);
1867 * Called from IPI handler with interrupts disabled.
1869 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1871 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1873 if (likely(c && c->page))
1877 static void flush_cpu_slab(void *d)
1879 struct kmem_cache *s = d;
1881 __flush_cpu_slab(s, smp_processor_id());
1884 static void flush_all(struct kmem_cache *s)
1886 on_each_cpu(flush_cpu_slab, s, 1);
1890 * Check if the objects in a per cpu structure fit numa
1891 * locality expectations.
1893 static inline int node_match(struct kmem_cache_cpu *c, int node)
1896 if (node != NUMA_NO_NODE && c->node != node)
1902 static int count_free(struct page *page)
1904 return page->objects - page->inuse;
1907 static unsigned long count_partial(struct kmem_cache_node *n,
1908 int (*get_count)(struct page *))
1910 unsigned long flags;
1911 unsigned long x = 0;
1914 spin_lock_irqsave(&n->list_lock, flags);
1915 list_for_each_entry(page, &n->partial, lru)
1916 x += get_count(page);
1917 spin_unlock_irqrestore(&n->list_lock, flags);
1921 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1923 #ifdef CONFIG_SLUB_DEBUG
1924 return atomic_long_read(&n->total_objects);
1930 static noinline void
1931 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1936 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1938 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1939 "default order: %d, min order: %d\n", s->name, s->objsize,
1940 s->size, oo_order(s->oo), oo_order(s->min));
1942 if (oo_order(s->min) > get_order(s->objsize))
1943 printk(KERN_WARNING " %s debugging increased min order, use "
1944 "slub_debug=O to disable.\n", s->name);
1946 for_each_online_node(node) {
1947 struct kmem_cache_node *n = get_node(s, node);
1948 unsigned long nr_slabs;
1949 unsigned long nr_objs;
1950 unsigned long nr_free;
1955 nr_free = count_partial(n, count_free);
1956 nr_slabs = node_nr_slabs(n);
1957 nr_objs = node_nr_objs(n);
1960 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1961 node, nr_slabs, nr_objs, nr_free);
1966 * Slow path. The lockless freelist is empty or we need to perform
1969 * Interrupts are disabled.
1971 * Processing is still very fast if new objects have been freed to the
1972 * regular freelist. In that case we simply take over the regular freelist
1973 * as the lockless freelist and zap the regular freelist.
1975 * If that is not working then we fall back to the partial lists. We take the
1976 * first element of the freelist as the object to allocate now and move the
1977 * rest of the freelist to the lockless freelist.
1979 * And if we were unable to get a new slab from the partial slab lists then
1980 * we need to allocate a new slab. This is the slowest path since it involves
1981 * a call to the page allocator and the setup of a new slab.
1983 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1984 unsigned long addr, struct kmem_cache_cpu *c)
1988 unsigned long flags;
1990 unsigned long counters;
1992 local_irq_save(flags);
1993 #ifdef CONFIG_PREEMPT
1995 * We may have been preempted and rescheduled on a different
1996 * cpu before disabling interrupts. Need to reload cpu area
1999 c = this_cpu_ptr(s->cpu_slab);
2002 /* We handle __GFP_ZERO in the caller */
2003 gfpflags &= ~__GFP_ZERO;
2009 if (unlikely(!node_match(c, node))) {
2010 stat(s, ALLOC_NODE_MISMATCH);
2011 deactivate_slab(s, c);
2015 stat(s, ALLOC_SLOWPATH);
2018 object = page->freelist;
2019 counters = page->counters;
2020 new.counters = counters;
2021 VM_BUG_ON(!new.frozen);
2024 * If there is no object left then we use this loop to
2025 * deactivate the slab which is simple since no objects
2026 * are left in the slab and therefore we do not need to
2027 * put the page back onto the partial list.
2029 * If there are objects left then we retrieve them
2030 * and use them to refill the per cpu queue.
2033 new.inuse = page->objects;
2034 new.frozen = object != NULL;
2036 } while (!__cmpxchg_double_slab(s, page,
2041 if (unlikely(!object)) {
2043 stat(s, DEACTIVATE_BYPASS);
2047 stat(s, ALLOC_REFILL);
2050 VM_BUG_ON(!page->frozen);
2051 c->freelist = get_freepointer(s, object);
2052 c->tid = next_tid(c->tid);
2053 local_irq_restore(flags);
2057 page = get_partial(s, gfpflags, node);
2059 stat(s, ALLOC_FROM_PARTIAL);
2060 object = c->freelist;
2062 if (kmem_cache_debug(s))
2067 page = new_slab(s, gfpflags, node);
2070 c = __this_cpu_ptr(s->cpu_slab);
2075 * No other reference to the page yet so we can
2076 * muck around with it freely without cmpxchg
2078 object = page->freelist;
2079 page->freelist = NULL;
2080 page->inuse = page->objects;
2082 stat(s, ALLOC_SLAB);
2083 c->node = page_to_nid(page);
2086 if (kmem_cache_debug(s))
2090 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2091 slab_out_of_memory(s, gfpflags, node);
2092 local_irq_restore(flags);
2096 if (!object || !alloc_debug_processing(s, page, object, addr))
2099 c->freelist = get_freepointer(s, object);
2100 deactivate_slab(s, c);
2102 c->node = NUMA_NO_NODE;
2103 local_irq_restore(flags);
2108 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2109 * have the fastpath folded into their functions. So no function call
2110 * overhead for requests that can be satisfied on the fastpath.
2112 * The fastpath works by first checking if the lockless freelist can be used.
2113 * If not then __slab_alloc is called for slow processing.
2115 * Otherwise we can simply pick the next object from the lockless free list.
2117 static __always_inline void *slab_alloc(struct kmem_cache *s,
2118 gfp_t gfpflags, int node, unsigned long addr)
2121 struct kmem_cache_cpu *c;
2124 if (slab_pre_alloc_hook(s, gfpflags))
2130 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2131 * enabled. We may switch back and forth between cpus while
2132 * reading from one cpu area. That does not matter as long
2133 * as we end up on the original cpu again when doing the cmpxchg.
2135 c = __this_cpu_ptr(s->cpu_slab);
2138 * The transaction ids are globally unique per cpu and per operation on
2139 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2140 * occurs on the right processor and that there was no operation on the
2141 * linked list in between.
2146 object = c->freelist;
2147 if (unlikely(!object || !node_match(c, node)))
2149 object = __slab_alloc(s, gfpflags, node, addr, c);
2153 * The cmpxchg will only match if there was no additional
2154 * operation and if we are on the right processor.
2156 * The cmpxchg does the following atomically (without lock semantics!)
2157 * 1. Relocate first pointer to the current per cpu area.
2158 * 2. Verify that tid and freelist have not been changed
2159 * 3. If they were not changed replace tid and freelist
2161 * Since this is without lock semantics the protection is only against
2162 * code executing on this cpu *not* from access by other cpus.
2164 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2165 s->cpu_slab->freelist, s->cpu_slab->tid,
2167 get_freepointer_safe(s, object), next_tid(tid)))) {
2169 note_cmpxchg_failure("slab_alloc", s, tid);
2172 stat(s, ALLOC_FASTPATH);
2175 if (unlikely(gfpflags & __GFP_ZERO) && object)
2176 memset(object, 0, s->objsize);
2178 slab_post_alloc_hook(s, gfpflags, object);
2183 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2185 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2187 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2191 EXPORT_SYMBOL(kmem_cache_alloc);
2193 #ifdef CONFIG_TRACING
2194 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2196 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2197 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2200 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2202 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2204 void *ret = kmalloc_order(size, flags, order);
2205 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2208 EXPORT_SYMBOL(kmalloc_order_trace);
2212 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2214 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2216 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2217 s->objsize, s->size, gfpflags, node);
2221 EXPORT_SYMBOL(kmem_cache_alloc_node);
2223 #ifdef CONFIG_TRACING
2224 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2226 int node, size_t size)
2228 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2230 trace_kmalloc_node(_RET_IP_, ret,
2231 size, s->size, gfpflags, node);
2234 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2239 * Slow patch handling. This may still be called frequently since objects
2240 * have a longer lifetime than the cpu slabs in most processing loads.
2242 * So we still attempt to reduce cache line usage. Just take the slab
2243 * lock and free the item. If there is no additional partial page
2244 * handling required then we can return immediately.
2246 static void __slab_free(struct kmem_cache *s, struct page *page,
2247 void *x, unsigned long addr)
2250 void **object = (void *)x;
2254 unsigned long counters;
2255 struct kmem_cache_node *n = NULL;
2256 unsigned long uninitialized_var(flags);
2258 stat(s, FREE_SLOWPATH);
2260 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2264 prior = page->freelist;
2265 counters = page->counters;
2266 set_freepointer(s, object, prior);
2267 new.counters = counters;
2268 was_frozen = new.frozen;
2270 if ((!new.inuse || !prior) && !was_frozen && !n) {
2271 n = get_node(s, page_to_nid(page));
2273 * Speculatively acquire the list_lock.
2274 * If the cmpxchg does not succeed then we may
2275 * drop the list_lock without any processing.
2277 * Otherwise the list_lock will synchronize with
2278 * other processors updating the list of slabs.
2280 spin_lock_irqsave(&n->list_lock, flags);
2284 } while (!cmpxchg_double_slab(s, page,
2286 object, new.counters,
2291 * The list lock was not taken therefore no list
2292 * activity can be necessary.
2295 stat(s, FREE_FROZEN);
2300 * was_frozen may have been set after we acquired the list_lock in
2301 * an earlier loop. So we need to check it here again.
2304 stat(s, FREE_FROZEN);
2306 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2310 * Objects left in the slab. If it was not on the partial list before
2313 if (unlikely(!prior)) {
2314 remove_full(s, page);
2315 add_partial(n, page, 0);
2316 stat(s, FREE_ADD_PARTIAL);
2319 spin_unlock_irqrestore(&n->list_lock, flags);
2325 * Slab still on the partial list.
2327 remove_partial(n, page);
2328 stat(s, FREE_REMOVE_PARTIAL);
2331 spin_unlock_irqrestore(&n->list_lock, flags);
2333 discard_slab(s, page);
2337 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2338 * can perform fastpath freeing without additional function calls.
2340 * The fastpath is only possible if we are freeing to the current cpu slab
2341 * of this processor. This typically the case if we have just allocated
2344 * If fastpath is not possible then fall back to __slab_free where we deal
2345 * with all sorts of special processing.
2347 static __always_inline void slab_free(struct kmem_cache *s,
2348 struct page *page, void *x, unsigned long addr)
2350 void **object = (void *)x;
2351 struct kmem_cache_cpu *c;
2354 slab_free_hook(s, x);
2359 * Determine the currently cpus per cpu slab.
2360 * The cpu may change afterward. However that does not matter since
2361 * data is retrieved via this pointer. If we are on the same cpu
2362 * during the cmpxchg then the free will succedd.
2364 c = __this_cpu_ptr(s->cpu_slab);
2369 if (likely(page == c->page)) {
2370 set_freepointer(s, object, c->freelist);
2372 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2373 s->cpu_slab->freelist, s->cpu_slab->tid,
2375 object, next_tid(tid)))) {
2377 note_cmpxchg_failure("slab_free", s, tid);
2380 stat(s, FREE_FASTPATH);
2382 __slab_free(s, page, x, addr);
2386 void kmem_cache_free(struct kmem_cache *s, void *x)
2390 page = virt_to_head_page(x);
2392 slab_free(s, page, x, _RET_IP_);
2394 trace_kmem_cache_free(_RET_IP_, x);
2396 EXPORT_SYMBOL(kmem_cache_free);
2399 * Object placement in a slab is made very easy because we always start at
2400 * offset 0. If we tune the size of the object to the alignment then we can
2401 * get the required alignment by putting one properly sized object after
2404 * Notice that the allocation order determines the sizes of the per cpu
2405 * caches. Each processor has always one slab available for allocations.
2406 * Increasing the allocation order reduces the number of times that slabs
2407 * must be moved on and off the partial lists and is therefore a factor in
2412 * Mininum / Maximum order of slab pages. This influences locking overhead
2413 * and slab fragmentation. A higher order reduces the number of partial slabs
2414 * and increases the number of allocations possible without having to
2415 * take the list_lock.
2417 static int slub_min_order;
2418 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2419 static int slub_min_objects;
2422 * Merge control. If this is set then no merging of slab caches will occur.
2423 * (Could be removed. This was introduced to pacify the merge skeptics.)
2425 static int slub_nomerge;
2428 * Calculate the order of allocation given an slab object size.
2430 * The order of allocation has significant impact on performance and other
2431 * system components. Generally order 0 allocations should be preferred since
2432 * order 0 does not cause fragmentation in the page allocator. Larger objects
2433 * be problematic to put into order 0 slabs because there may be too much
2434 * unused space left. We go to a higher order if more than 1/16th of the slab
2437 * In order to reach satisfactory performance we must ensure that a minimum
2438 * number of objects is in one slab. Otherwise we may generate too much
2439 * activity on the partial lists which requires taking the list_lock. This is
2440 * less a concern for large slabs though which are rarely used.
2442 * slub_max_order specifies the order where we begin to stop considering the
2443 * number of objects in a slab as critical. If we reach slub_max_order then
2444 * we try to keep the page order as low as possible. So we accept more waste
2445 * of space in favor of a small page order.
2447 * Higher order allocations also allow the placement of more objects in a
2448 * slab and thereby reduce object handling overhead. If the user has
2449 * requested a higher mininum order then we start with that one instead of
2450 * the smallest order which will fit the object.
2452 static inline int slab_order(int size, int min_objects,
2453 int max_order, int fract_leftover, int reserved)
2457 int min_order = slub_min_order;
2459 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2460 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2462 for (order = max(min_order,
2463 fls(min_objects * size - 1) - PAGE_SHIFT);
2464 order <= max_order; order++) {
2466 unsigned long slab_size = PAGE_SIZE << order;
2468 if (slab_size < min_objects * size + reserved)
2471 rem = (slab_size - reserved) % size;
2473 if (rem <= slab_size / fract_leftover)
2481 static inline int calculate_order(int size, int reserved)
2489 * Attempt to find best configuration for a slab. This
2490 * works by first attempting to generate a layout with
2491 * the best configuration and backing off gradually.
2493 * First we reduce the acceptable waste in a slab. Then
2494 * we reduce the minimum objects required in a slab.
2496 min_objects = slub_min_objects;
2498 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2499 max_objects = order_objects(slub_max_order, size, reserved);
2500 min_objects = min(min_objects, max_objects);
2502 while (min_objects > 1) {
2504 while (fraction >= 4) {
2505 order = slab_order(size, min_objects,
2506 slub_max_order, fraction, reserved);
2507 if (order <= slub_max_order)
2515 * We were unable to place multiple objects in a slab. Now
2516 * lets see if we can place a single object there.
2518 order = slab_order(size, 1, slub_max_order, 1, reserved);
2519 if (order <= slub_max_order)
2523 * Doh this slab cannot be placed using slub_max_order.
2525 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2526 if (order < MAX_ORDER)
2532 * Figure out what the alignment of the objects will be.
2534 static unsigned long calculate_alignment(unsigned long flags,
2535 unsigned long align, unsigned long size)
2538 * If the user wants hardware cache aligned objects then follow that
2539 * suggestion if the object is sufficiently large.
2541 * The hardware cache alignment cannot override the specified
2542 * alignment though. If that is greater then use it.
2544 if (flags & SLAB_HWCACHE_ALIGN) {
2545 unsigned long ralign = cache_line_size();
2546 while (size <= ralign / 2)
2548 align = max(align, ralign);
2551 if (align < ARCH_SLAB_MINALIGN)
2552 align = ARCH_SLAB_MINALIGN;
2554 return ALIGN(align, sizeof(void *));
2558 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2561 spin_lock_init(&n->list_lock);
2562 INIT_LIST_HEAD(&n->partial);
2563 #ifdef CONFIG_SLUB_DEBUG
2564 atomic_long_set(&n->nr_slabs, 0);
2565 atomic_long_set(&n->total_objects, 0);
2566 INIT_LIST_HEAD(&n->full);
2570 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2572 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2573 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2576 * Must align to double word boundary for the double cmpxchg
2577 * instructions to work; see __pcpu_double_call_return_bool().
2579 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2580 2 * sizeof(void *));
2585 init_kmem_cache_cpus(s);
2590 static struct kmem_cache *kmem_cache_node;
2593 * No kmalloc_node yet so do it by hand. We know that this is the first
2594 * slab on the node for this slabcache. There are no concurrent accesses
2597 * Note that this function only works on the kmalloc_node_cache
2598 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2599 * memory on a fresh node that has no slab structures yet.
2601 static void early_kmem_cache_node_alloc(int node)
2604 struct kmem_cache_node *n;
2606 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2608 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2611 if (page_to_nid(page) != node) {
2612 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2614 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2615 "in order to be able to continue\n");
2620 page->freelist = get_freepointer(kmem_cache_node, n);
2623 kmem_cache_node->node[node] = n;
2624 #ifdef CONFIG_SLUB_DEBUG
2625 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2626 init_tracking(kmem_cache_node, n);
2628 init_kmem_cache_node(n, kmem_cache_node);
2629 inc_slabs_node(kmem_cache_node, node, page->objects);
2631 add_partial(n, page, 0);
2634 static void free_kmem_cache_nodes(struct kmem_cache *s)
2638 for_each_node_state(node, N_NORMAL_MEMORY) {
2639 struct kmem_cache_node *n = s->node[node];
2642 kmem_cache_free(kmem_cache_node, n);
2644 s->node[node] = NULL;
2648 static int init_kmem_cache_nodes(struct kmem_cache *s)
2652 for_each_node_state(node, N_NORMAL_MEMORY) {
2653 struct kmem_cache_node *n;
2655 if (slab_state == DOWN) {
2656 early_kmem_cache_node_alloc(node);
2659 n = kmem_cache_alloc_node(kmem_cache_node,
2663 free_kmem_cache_nodes(s);
2668 init_kmem_cache_node(n, s);
2673 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2675 if (min < MIN_PARTIAL)
2677 else if (min > MAX_PARTIAL)
2679 s->min_partial = min;
2683 * calculate_sizes() determines the order and the distribution of data within
2686 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2688 unsigned long flags = s->flags;
2689 unsigned long size = s->objsize;
2690 unsigned long align = s->align;
2694 * Round up object size to the next word boundary. We can only
2695 * place the free pointer at word boundaries and this determines
2696 * the possible location of the free pointer.
2698 size = ALIGN(size, sizeof(void *));
2700 #ifdef CONFIG_SLUB_DEBUG
2702 * Determine if we can poison the object itself. If the user of
2703 * the slab may touch the object after free or before allocation
2704 * then we should never poison the object itself.
2706 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2708 s->flags |= __OBJECT_POISON;
2710 s->flags &= ~__OBJECT_POISON;
2714 * If we are Redzoning then check if there is some space between the
2715 * end of the object and the free pointer. If not then add an
2716 * additional word to have some bytes to store Redzone information.
2718 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2719 size += sizeof(void *);
2723 * With that we have determined the number of bytes in actual use
2724 * by the object. This is the potential offset to the free pointer.
2728 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2731 * Relocate free pointer after the object if it is not
2732 * permitted to overwrite the first word of the object on
2735 * This is the case if we do RCU, have a constructor or
2736 * destructor or are poisoning the objects.
2739 size += sizeof(void *);
2742 #ifdef CONFIG_SLUB_DEBUG
2743 if (flags & SLAB_STORE_USER)
2745 * Need to store information about allocs and frees after
2748 size += 2 * sizeof(struct track);
2750 if (flags & SLAB_RED_ZONE)
2752 * Add some empty padding so that we can catch
2753 * overwrites from earlier objects rather than let
2754 * tracking information or the free pointer be
2755 * corrupted if a user writes before the start
2758 size += sizeof(void *);
2762 * Determine the alignment based on various parameters that the
2763 * user specified and the dynamic determination of cache line size
2766 align = calculate_alignment(flags, align, s->objsize);
2770 * SLUB stores one object immediately after another beginning from
2771 * offset 0. In order to align the objects we have to simply size
2772 * each object to conform to the alignment.
2774 size = ALIGN(size, align);
2776 if (forced_order >= 0)
2777 order = forced_order;
2779 order = calculate_order(size, s->reserved);
2786 s->allocflags |= __GFP_COMP;
2788 if (s->flags & SLAB_CACHE_DMA)
2789 s->allocflags |= SLUB_DMA;
2791 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2792 s->allocflags |= __GFP_RECLAIMABLE;
2795 * Determine the number of objects per slab
2797 s->oo = oo_make(order, size, s->reserved);
2798 s->min = oo_make(get_order(size), size, s->reserved);
2799 if (oo_objects(s->oo) > oo_objects(s->max))
2802 return !!oo_objects(s->oo);
2806 static int kmem_cache_open(struct kmem_cache *s,
2807 const char *name, size_t size,
2808 size_t align, unsigned long flags,
2809 void (*ctor)(void *))
2811 memset(s, 0, kmem_size);
2816 s->flags = kmem_cache_flags(size, flags, name, ctor);
2819 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2820 s->reserved = sizeof(struct rcu_head);
2822 if (!calculate_sizes(s, -1))
2824 if (disable_higher_order_debug) {
2826 * Disable debugging flags that store metadata if the min slab
2829 if (get_order(s->size) > get_order(s->objsize)) {
2830 s->flags &= ~DEBUG_METADATA_FLAGS;
2832 if (!calculate_sizes(s, -1))
2837 #ifdef CONFIG_CMPXCHG_DOUBLE
2838 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
2839 /* Enable fast mode */
2840 s->flags |= __CMPXCHG_DOUBLE;
2844 * The larger the object size is, the more pages we want on the partial
2845 * list to avoid pounding the page allocator excessively.
2847 set_min_partial(s, ilog2(s->size));
2850 s->remote_node_defrag_ratio = 1000;
2852 if (!init_kmem_cache_nodes(s))
2855 if (alloc_kmem_cache_cpus(s))
2858 free_kmem_cache_nodes(s);
2860 if (flags & SLAB_PANIC)
2861 panic("Cannot create slab %s size=%lu realsize=%u "
2862 "order=%u offset=%u flags=%lx\n",
2863 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2869 * Determine the size of a slab object
2871 unsigned int kmem_cache_size(struct kmem_cache *s)
2875 EXPORT_SYMBOL(kmem_cache_size);
2877 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2880 #ifdef CONFIG_SLUB_DEBUG
2881 void *addr = page_address(page);
2883 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2884 sizeof(long), GFP_ATOMIC);
2887 slab_err(s, page, "%s", text);
2890 get_map(s, page, map);
2891 for_each_object(p, s, addr, page->objects) {
2893 if (!test_bit(slab_index(p, s, addr), map)) {
2894 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2896 print_tracking(s, p);
2905 * Attempt to free all partial slabs on a node.
2907 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2909 unsigned long flags;
2910 struct page *page, *h;
2912 spin_lock_irqsave(&n->list_lock, flags);
2913 list_for_each_entry_safe(page, h, &n->partial, lru) {
2915 remove_partial(n, page);
2916 discard_slab(s, page);
2918 list_slab_objects(s, page,
2919 "Objects remaining on kmem_cache_close()");
2922 spin_unlock_irqrestore(&n->list_lock, flags);
2926 * Release all resources used by a slab cache.
2928 static inline int kmem_cache_close(struct kmem_cache *s)
2933 free_percpu(s->cpu_slab);
2934 /* Attempt to free all objects */
2935 for_each_node_state(node, N_NORMAL_MEMORY) {
2936 struct kmem_cache_node *n = get_node(s, node);
2939 if (n->nr_partial || slabs_node(s, node))
2942 free_kmem_cache_nodes(s);
2947 * Close a cache and release the kmem_cache structure
2948 * (must be used for caches created using kmem_cache_create)
2950 void kmem_cache_destroy(struct kmem_cache *s)
2952 down_write(&slub_lock);
2956 if (kmem_cache_close(s)) {
2957 printk(KERN_ERR "SLUB %s: %s called for cache that "
2958 "still has objects.\n", s->name, __func__);
2961 if (s->flags & SLAB_DESTROY_BY_RCU)
2963 sysfs_slab_remove(s);
2965 up_write(&slub_lock);
2967 EXPORT_SYMBOL(kmem_cache_destroy);
2969 /********************************************************************
2971 *******************************************************************/
2973 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2974 EXPORT_SYMBOL(kmalloc_caches);
2976 static struct kmem_cache *kmem_cache;
2978 #ifdef CONFIG_ZONE_DMA
2979 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2982 static int __init setup_slub_min_order(char *str)
2984 get_option(&str, &slub_min_order);
2989 __setup("slub_min_order=", setup_slub_min_order);
2991 static int __init setup_slub_max_order(char *str)
2993 get_option(&str, &slub_max_order);
2994 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2999 __setup("slub_max_order=", setup_slub_max_order);
3001 static int __init setup_slub_min_objects(char *str)
3003 get_option(&str, &slub_min_objects);
3008 __setup("slub_min_objects=", setup_slub_min_objects);
3010 static int __init setup_slub_nomerge(char *str)
3016 __setup("slub_nomerge", setup_slub_nomerge);
3018 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3019 int size, unsigned int flags)
3021 struct kmem_cache *s;
3023 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3026 * This function is called with IRQs disabled during early-boot on
3027 * single CPU so there's no need to take slub_lock here.
3029 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3033 list_add(&s->list, &slab_caches);
3037 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3042 * Conversion table for small slabs sizes / 8 to the index in the
3043 * kmalloc array. This is necessary for slabs < 192 since we have non power
3044 * of two cache sizes there. The size of larger slabs can be determined using
3047 static s8 size_index[24] = {
3074 static inline int size_index_elem(size_t bytes)
3076 return (bytes - 1) / 8;
3079 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3085 return ZERO_SIZE_PTR;
3087 index = size_index[size_index_elem(size)];
3089 index = fls(size - 1);
3091 #ifdef CONFIG_ZONE_DMA
3092 if (unlikely((flags & SLUB_DMA)))
3093 return kmalloc_dma_caches[index];
3096 return kmalloc_caches[index];
3099 void *__kmalloc(size_t size, gfp_t flags)
3101 struct kmem_cache *s;
3104 if (unlikely(size > SLUB_MAX_SIZE))
3105 return kmalloc_large(size, flags);
3107 s = get_slab(size, flags);
3109 if (unlikely(ZERO_OR_NULL_PTR(s)))
3112 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3114 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3118 EXPORT_SYMBOL(__kmalloc);
3121 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3126 flags |= __GFP_COMP | __GFP_NOTRACK;
3127 page = alloc_pages_node(node, flags, get_order(size));
3129 ptr = page_address(page);
3131 kmemleak_alloc(ptr, size, 1, flags);
3135 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3137 struct kmem_cache *s;
3140 if (unlikely(size > SLUB_MAX_SIZE)) {
3141 ret = kmalloc_large_node(size, flags, node);
3143 trace_kmalloc_node(_RET_IP_, ret,
3144 size, PAGE_SIZE << get_order(size),
3150 s = get_slab(size, flags);
3152 if (unlikely(ZERO_OR_NULL_PTR(s)))
3155 ret = slab_alloc(s, flags, node, _RET_IP_);
3157 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3161 EXPORT_SYMBOL(__kmalloc_node);
3164 size_t ksize(const void *object)
3168 if (unlikely(object == ZERO_SIZE_PTR))
3171 page = virt_to_head_page(object);
3173 if (unlikely(!PageSlab(page))) {
3174 WARN_ON(!PageCompound(page));
3175 return PAGE_SIZE << compound_order(page);
3178 return slab_ksize(page->slab);
3180 EXPORT_SYMBOL(ksize);
3182 void kfree(const void *x)
3185 void *object = (void *)x;
3187 trace_kfree(_RET_IP_, x);
3189 if (unlikely(ZERO_OR_NULL_PTR(x)))
3192 page = virt_to_head_page(x);
3193 if (unlikely(!PageSlab(page))) {
3194 BUG_ON(!PageCompound(page));
3199 slab_free(page->slab, page, object, _RET_IP_);
3201 EXPORT_SYMBOL(kfree);
3204 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3205 * the remaining slabs by the number of items in use. The slabs with the
3206 * most items in use come first. New allocations will then fill those up
3207 * and thus they can be removed from the partial lists.
3209 * The slabs with the least items are placed last. This results in them
3210 * being allocated from last increasing the chance that the last objects
3211 * are freed in them.
3213 int kmem_cache_shrink(struct kmem_cache *s)
3217 struct kmem_cache_node *n;
3220 int objects = oo_objects(s->max);
3221 struct list_head *slabs_by_inuse =
3222 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3223 unsigned long flags;
3225 if (!slabs_by_inuse)
3229 for_each_node_state(node, N_NORMAL_MEMORY) {
3230 n = get_node(s, node);
3235 for (i = 0; i < objects; i++)
3236 INIT_LIST_HEAD(slabs_by_inuse + i);
3238 spin_lock_irqsave(&n->list_lock, flags);
3241 * Build lists indexed by the items in use in each slab.
3243 * Note that concurrent frees may occur while we hold the
3244 * list_lock. page->inuse here is the upper limit.
3246 list_for_each_entry_safe(page, t, &n->partial, lru) {
3248 remove_partial(n, page);
3249 discard_slab(s, page);
3251 list_move(&page->lru,
3252 slabs_by_inuse + page->inuse);
3257 * Rebuild the partial list with the slabs filled up most
3258 * first and the least used slabs at the end.
3260 for (i = objects - 1; i >= 0; i--)
3261 list_splice(slabs_by_inuse + i, n->partial.prev);
3263 spin_unlock_irqrestore(&n->list_lock, flags);
3266 kfree(slabs_by_inuse);
3269 EXPORT_SYMBOL(kmem_cache_shrink);
3271 #if defined(CONFIG_MEMORY_HOTPLUG)
3272 static int slab_mem_going_offline_callback(void *arg)
3274 struct kmem_cache *s;
3276 down_read(&slub_lock);
3277 list_for_each_entry(s, &slab_caches, list)
3278 kmem_cache_shrink(s);
3279 up_read(&slub_lock);
3284 static void slab_mem_offline_callback(void *arg)
3286 struct kmem_cache_node *n;
3287 struct kmem_cache *s;
3288 struct memory_notify *marg = arg;
3291 offline_node = marg->status_change_nid;
3294 * If the node still has available memory. we need kmem_cache_node
3297 if (offline_node < 0)
3300 down_read(&slub_lock);
3301 list_for_each_entry(s, &slab_caches, list) {
3302 n = get_node(s, offline_node);
3305 * if n->nr_slabs > 0, slabs still exist on the node
3306 * that is going down. We were unable to free them,
3307 * and offline_pages() function shouldn't call this
3308 * callback. So, we must fail.
3310 BUG_ON(slabs_node(s, offline_node));
3312 s->node[offline_node] = NULL;
3313 kmem_cache_free(kmem_cache_node, n);
3316 up_read(&slub_lock);
3319 static int slab_mem_going_online_callback(void *arg)
3321 struct kmem_cache_node *n;
3322 struct kmem_cache *s;
3323 struct memory_notify *marg = arg;
3324 int nid = marg->status_change_nid;
3328 * If the node's memory is already available, then kmem_cache_node is
3329 * already created. Nothing to do.
3335 * We are bringing a node online. No memory is available yet. We must
3336 * allocate a kmem_cache_node structure in order to bring the node
3339 down_read(&slub_lock);
3340 list_for_each_entry(s, &slab_caches, list) {
3342 * XXX: kmem_cache_alloc_node will fallback to other nodes
3343 * since memory is not yet available from the node that
3346 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3351 init_kmem_cache_node(n, s);
3355 up_read(&slub_lock);
3359 static int slab_memory_callback(struct notifier_block *self,
3360 unsigned long action, void *arg)
3365 case MEM_GOING_ONLINE:
3366 ret = slab_mem_going_online_callback(arg);
3368 case MEM_GOING_OFFLINE:
3369 ret = slab_mem_going_offline_callback(arg);
3372 case MEM_CANCEL_ONLINE:
3373 slab_mem_offline_callback(arg);
3376 case MEM_CANCEL_OFFLINE:
3380 ret = notifier_from_errno(ret);
3386 #endif /* CONFIG_MEMORY_HOTPLUG */
3388 /********************************************************************
3389 * Basic setup of slabs
3390 *******************************************************************/
3393 * Used for early kmem_cache structures that were allocated using
3394 * the page allocator
3397 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3401 list_add(&s->list, &slab_caches);
3404 for_each_node_state(node, N_NORMAL_MEMORY) {
3405 struct kmem_cache_node *n = get_node(s, node);
3409 list_for_each_entry(p, &n->partial, lru)
3412 #ifdef CONFIG_SLUB_DEBUG
3413 list_for_each_entry(p, &n->full, lru)
3420 void __init kmem_cache_init(void)
3424 struct kmem_cache *temp_kmem_cache;
3426 struct kmem_cache *temp_kmem_cache_node;
3427 unsigned long kmalloc_size;
3429 kmem_size = offsetof(struct kmem_cache, node) +
3430 nr_node_ids * sizeof(struct kmem_cache_node *);
3432 /* Allocate two kmem_caches from the page allocator */
3433 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3434 order = get_order(2 * kmalloc_size);
3435 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3438 * Must first have the slab cache available for the allocations of the
3439 * struct kmem_cache_node's. There is special bootstrap code in
3440 * kmem_cache_open for slab_state == DOWN.
3442 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3444 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3445 sizeof(struct kmem_cache_node),
3446 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3448 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3450 /* Able to allocate the per node structures */
3451 slab_state = PARTIAL;
3453 temp_kmem_cache = kmem_cache;
3454 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3455 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3456 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3457 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3460 * Allocate kmem_cache_node properly from the kmem_cache slab.
3461 * kmem_cache_node is separately allocated so no need to
3462 * update any list pointers.
3464 temp_kmem_cache_node = kmem_cache_node;
3466 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3467 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3469 kmem_cache_bootstrap_fixup(kmem_cache_node);
3472 kmem_cache_bootstrap_fixup(kmem_cache);
3474 /* Free temporary boot structure */
3475 free_pages((unsigned long)temp_kmem_cache, order);
3477 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3480 * Patch up the size_index table if we have strange large alignment
3481 * requirements for the kmalloc array. This is only the case for
3482 * MIPS it seems. The standard arches will not generate any code here.
3484 * Largest permitted alignment is 256 bytes due to the way we
3485 * handle the index determination for the smaller caches.
3487 * Make sure that nothing crazy happens if someone starts tinkering
3488 * around with ARCH_KMALLOC_MINALIGN
3490 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3491 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3493 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3494 int elem = size_index_elem(i);
3495 if (elem >= ARRAY_SIZE(size_index))
3497 size_index[elem] = KMALLOC_SHIFT_LOW;
3500 if (KMALLOC_MIN_SIZE == 64) {
3502 * The 96 byte size cache is not used if the alignment
3505 for (i = 64 + 8; i <= 96; i += 8)
3506 size_index[size_index_elem(i)] = 7;
3507 } else if (KMALLOC_MIN_SIZE == 128) {
3509 * The 192 byte sized cache is not used if the alignment
3510 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3513 for (i = 128 + 8; i <= 192; i += 8)
3514 size_index[size_index_elem(i)] = 8;
3517 /* Caches that are not of the two-to-the-power-of size */
3518 if (KMALLOC_MIN_SIZE <= 32) {
3519 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3523 if (KMALLOC_MIN_SIZE <= 64) {
3524 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3528 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3529 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3535 /* Provide the correct kmalloc names now that the caches are up */
3536 if (KMALLOC_MIN_SIZE <= 32) {
3537 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3538 BUG_ON(!kmalloc_caches[1]->name);
3541 if (KMALLOC_MIN_SIZE <= 64) {
3542 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3543 BUG_ON(!kmalloc_caches[2]->name);
3546 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3547 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3550 kmalloc_caches[i]->name = s;
3554 register_cpu_notifier(&slab_notifier);
3557 #ifdef CONFIG_ZONE_DMA
3558 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3559 struct kmem_cache *s = kmalloc_caches[i];
3562 char *name = kasprintf(GFP_NOWAIT,
3563 "dma-kmalloc-%d", s->objsize);
3566 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3567 s->objsize, SLAB_CACHE_DMA);
3572 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3573 " CPUs=%d, Nodes=%d\n",
3574 caches, cache_line_size(),
3575 slub_min_order, slub_max_order, slub_min_objects,
3576 nr_cpu_ids, nr_node_ids);
3579 void __init kmem_cache_init_late(void)
3584 * Find a mergeable slab cache
3586 static int slab_unmergeable(struct kmem_cache *s)
3588 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3595 * We may have set a slab to be unmergeable during bootstrap.
3597 if (s->refcount < 0)
3603 static struct kmem_cache *find_mergeable(size_t size,
3604 size_t align, unsigned long flags, const char *name,
3605 void (*ctor)(void *))
3607 struct kmem_cache *s;
3609 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3615 size = ALIGN(size, sizeof(void *));
3616 align = calculate_alignment(flags, align, size);
3617 size = ALIGN(size, align);
3618 flags = kmem_cache_flags(size, flags, name, NULL);
3620 list_for_each_entry(s, &slab_caches, list) {
3621 if (slab_unmergeable(s))
3627 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3630 * Check if alignment is compatible.
3631 * Courtesy of Adrian Drzewiecki
3633 if ((s->size & ~(align - 1)) != s->size)
3636 if (s->size - size >= sizeof(void *))
3644 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3645 size_t align, unsigned long flags, void (*ctor)(void *))
3647 struct kmem_cache *s;
3653 down_write(&slub_lock);
3654 s = find_mergeable(size, align, flags, name, ctor);
3658 * Adjust the object sizes so that we clear
3659 * the complete object on kzalloc.
3661 s->objsize = max(s->objsize, (int)size);
3662 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3664 if (sysfs_slab_alias(s, name)) {
3668 up_write(&slub_lock);
3672 n = kstrdup(name, GFP_KERNEL);
3676 s = kmalloc(kmem_size, GFP_KERNEL);
3678 if (kmem_cache_open(s, n,
3679 size, align, flags, ctor)) {
3680 list_add(&s->list, &slab_caches);
3681 if (sysfs_slab_add(s)) {
3687 up_write(&slub_lock);
3694 up_write(&slub_lock);
3696 if (flags & SLAB_PANIC)
3697 panic("Cannot create slabcache %s\n", name);
3702 EXPORT_SYMBOL(kmem_cache_create);
3706 * Use the cpu notifier to insure that the cpu slabs are flushed when
3709 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3710 unsigned long action, void *hcpu)
3712 long cpu = (long)hcpu;
3713 struct kmem_cache *s;
3714 unsigned long flags;
3717 case CPU_UP_CANCELED:
3718 case CPU_UP_CANCELED_FROZEN:
3720 case CPU_DEAD_FROZEN:
3721 down_read(&slub_lock);
3722 list_for_each_entry(s, &slab_caches, list) {
3723 local_irq_save(flags);
3724 __flush_cpu_slab(s, cpu);
3725 local_irq_restore(flags);
3727 up_read(&slub_lock);
3735 static struct notifier_block __cpuinitdata slab_notifier = {
3736 .notifier_call = slab_cpuup_callback
3741 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3743 struct kmem_cache *s;
3746 if (unlikely(size > SLUB_MAX_SIZE))
3747 return kmalloc_large(size, gfpflags);
3749 s = get_slab(size, gfpflags);
3751 if (unlikely(ZERO_OR_NULL_PTR(s)))
3754 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3756 /* Honor the call site pointer we received. */
3757 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3763 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3764 int node, unsigned long caller)
3766 struct kmem_cache *s;
3769 if (unlikely(size > SLUB_MAX_SIZE)) {
3770 ret = kmalloc_large_node(size, gfpflags, node);
3772 trace_kmalloc_node(caller, ret,
3773 size, PAGE_SIZE << get_order(size),
3779 s = get_slab(size, gfpflags);
3781 if (unlikely(ZERO_OR_NULL_PTR(s)))
3784 ret = slab_alloc(s, gfpflags, node, caller);
3786 /* Honor the call site pointer we received. */
3787 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3794 static int count_inuse(struct page *page)
3799 static int count_total(struct page *page)
3801 return page->objects;
3805 #ifdef CONFIG_SLUB_DEBUG
3806 static int validate_slab(struct kmem_cache *s, struct page *page,
3810 void *addr = page_address(page);
3812 if (!check_slab(s, page) ||
3813 !on_freelist(s, page, NULL))
3816 /* Now we know that a valid freelist exists */
3817 bitmap_zero(map, page->objects);
3819 get_map(s, page, map);
3820 for_each_object(p, s, addr, page->objects) {
3821 if (test_bit(slab_index(p, s, addr), map))
3822 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3826 for_each_object(p, s, addr, page->objects)
3827 if (!test_bit(slab_index(p, s, addr), map))
3828 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3833 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3837 validate_slab(s, page, map);
3841 static int validate_slab_node(struct kmem_cache *s,
3842 struct kmem_cache_node *n, unsigned long *map)
3844 unsigned long count = 0;
3846 unsigned long flags;
3848 spin_lock_irqsave(&n->list_lock, flags);
3850 list_for_each_entry(page, &n->partial, lru) {
3851 validate_slab_slab(s, page, map);
3854 if (count != n->nr_partial)
3855 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3856 "counter=%ld\n", s->name, count, n->nr_partial);
3858 if (!(s->flags & SLAB_STORE_USER))
3861 list_for_each_entry(page, &n->full, lru) {
3862 validate_slab_slab(s, page, map);
3865 if (count != atomic_long_read(&n->nr_slabs))
3866 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3867 "counter=%ld\n", s->name, count,
3868 atomic_long_read(&n->nr_slabs));
3871 spin_unlock_irqrestore(&n->list_lock, flags);
3875 static long validate_slab_cache(struct kmem_cache *s)
3878 unsigned long count = 0;
3879 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3880 sizeof(unsigned long), GFP_KERNEL);
3886 for_each_node_state(node, N_NORMAL_MEMORY) {
3887 struct kmem_cache_node *n = get_node(s, node);
3889 count += validate_slab_node(s, n, map);
3895 * Generate lists of code addresses where slabcache objects are allocated
3900 unsigned long count;
3907 DECLARE_BITMAP(cpus, NR_CPUS);
3913 unsigned long count;
3914 struct location *loc;
3917 static void free_loc_track(struct loc_track *t)
3920 free_pages((unsigned long)t->loc,
3921 get_order(sizeof(struct location) * t->max));
3924 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3929 order = get_order(sizeof(struct location) * max);
3931 l = (void *)__get_free_pages(flags, order);
3936 memcpy(l, t->loc, sizeof(struct location) * t->count);
3944 static int add_location(struct loc_track *t, struct kmem_cache *s,
3945 const struct track *track)
3947 long start, end, pos;
3949 unsigned long caddr;
3950 unsigned long age = jiffies - track->when;
3956 pos = start + (end - start + 1) / 2;
3959 * There is nothing at "end". If we end up there
3960 * we need to add something to before end.
3965 caddr = t->loc[pos].addr;
3966 if (track->addr == caddr) {
3972 if (age < l->min_time)
3974 if (age > l->max_time)
3977 if (track->pid < l->min_pid)
3978 l->min_pid = track->pid;
3979 if (track->pid > l->max_pid)
3980 l->max_pid = track->pid;
3982 cpumask_set_cpu(track->cpu,
3983 to_cpumask(l->cpus));
3985 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3989 if (track->addr < caddr)
3996 * Not found. Insert new tracking element.
3998 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4004 (t->count - pos) * sizeof(struct location));
4007 l->addr = track->addr;
4011 l->min_pid = track->pid;
4012 l->max_pid = track->pid;
4013 cpumask_clear(to_cpumask(l->cpus));
4014 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4015 nodes_clear(l->nodes);
4016 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4020 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4021 struct page *page, enum track_item alloc,
4024 void *addr = page_address(page);
4027 bitmap_zero(map, page->objects);
4028 get_map(s, page, map);
4030 for_each_object(p, s, addr, page->objects)
4031 if (!test_bit(slab_index(p, s, addr), map))
4032 add_location(t, s, get_track(s, p, alloc));
4035 static int list_locations(struct kmem_cache *s, char *buf,
4036 enum track_item alloc)
4040 struct loc_track t = { 0, 0, NULL };
4042 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4043 sizeof(unsigned long), GFP_KERNEL);
4045 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4048 return sprintf(buf, "Out of memory\n");
4050 /* Push back cpu slabs */
4053 for_each_node_state(node, N_NORMAL_MEMORY) {
4054 struct kmem_cache_node *n = get_node(s, node);
4055 unsigned long flags;
4058 if (!atomic_long_read(&n->nr_slabs))
4061 spin_lock_irqsave(&n->list_lock, flags);
4062 list_for_each_entry(page, &n->partial, lru)
4063 process_slab(&t, s, page, alloc, map);
4064 list_for_each_entry(page, &n->full, lru)
4065 process_slab(&t, s, page, alloc, map);
4066 spin_unlock_irqrestore(&n->list_lock, flags);
4069 for (i = 0; i < t.count; i++) {
4070 struct location *l = &t.loc[i];
4072 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4074 len += sprintf(buf + len, "%7ld ", l->count);
4077 len += sprintf(buf + len, "%pS", (void *)l->addr);
4079 len += sprintf(buf + len, "<not-available>");
4081 if (l->sum_time != l->min_time) {
4082 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4084 (long)div_u64(l->sum_time, l->count),
4087 len += sprintf(buf + len, " age=%ld",
4090 if (l->min_pid != l->max_pid)
4091 len += sprintf(buf + len, " pid=%ld-%ld",
4092 l->min_pid, l->max_pid);
4094 len += sprintf(buf + len, " pid=%ld",
4097 if (num_online_cpus() > 1 &&
4098 !cpumask_empty(to_cpumask(l->cpus)) &&
4099 len < PAGE_SIZE - 60) {
4100 len += sprintf(buf + len, " cpus=");
4101 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4102 to_cpumask(l->cpus));
4105 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4106 len < PAGE_SIZE - 60) {
4107 len += sprintf(buf + len, " nodes=");
4108 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4112 len += sprintf(buf + len, "\n");
4118 len += sprintf(buf, "No data\n");
4123 #ifdef SLUB_RESILIENCY_TEST
4124 static void resiliency_test(void)
4128 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4130 printk(KERN_ERR "SLUB resiliency testing\n");
4131 printk(KERN_ERR "-----------------------\n");
4132 printk(KERN_ERR "A. Corruption after allocation\n");
4134 p = kzalloc(16, GFP_KERNEL);
4136 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4137 " 0x12->0x%p\n\n", p + 16);
4139 validate_slab_cache(kmalloc_caches[4]);
4141 /* Hmmm... The next two are dangerous */
4142 p = kzalloc(32, GFP_KERNEL);
4143 p[32 + sizeof(void *)] = 0x34;
4144 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4145 " 0x34 -> -0x%p\n", p);
4147 "If allocated object is overwritten then not detectable\n\n");
4149 validate_slab_cache(kmalloc_caches[5]);
4150 p = kzalloc(64, GFP_KERNEL);
4151 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4153 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4156 "If allocated object is overwritten then not detectable\n\n");
4157 validate_slab_cache(kmalloc_caches[6]);
4159 printk(KERN_ERR "\nB. Corruption after free\n");
4160 p = kzalloc(128, GFP_KERNEL);
4163 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4164 validate_slab_cache(kmalloc_caches[7]);
4166 p = kzalloc(256, GFP_KERNEL);
4169 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4171 validate_slab_cache(kmalloc_caches[8]);
4173 p = kzalloc(512, GFP_KERNEL);
4176 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4177 validate_slab_cache(kmalloc_caches[9]);
4181 static void resiliency_test(void) {};
4186 enum slab_stat_type {
4187 SL_ALL, /* All slabs */
4188 SL_PARTIAL, /* Only partially allocated slabs */
4189 SL_CPU, /* Only slabs used for cpu caches */
4190 SL_OBJECTS, /* Determine allocated objects not slabs */
4191 SL_TOTAL /* Determine object capacity not slabs */
4194 #define SO_ALL (1 << SL_ALL)
4195 #define SO_PARTIAL (1 << SL_PARTIAL)
4196 #define SO_CPU (1 << SL_CPU)
4197 #define SO_OBJECTS (1 << SL_OBJECTS)
4198 #define SO_TOTAL (1 << SL_TOTAL)
4200 static ssize_t show_slab_objects(struct kmem_cache *s,
4201 char *buf, unsigned long flags)
4203 unsigned long total = 0;
4206 unsigned long *nodes;
4207 unsigned long *per_cpu;
4209 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4212 per_cpu = nodes + nr_node_ids;
4214 if (flags & SO_CPU) {
4217 for_each_possible_cpu(cpu) {
4218 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4220 if (!c || c->node < 0)
4224 if (flags & SO_TOTAL)
4225 x = c->page->objects;
4226 else if (flags & SO_OBJECTS)
4232 nodes[c->node] += x;
4238 lock_memory_hotplug();
4239 #ifdef CONFIG_SLUB_DEBUG
4240 if (flags & SO_ALL) {
4241 for_each_node_state(node, N_NORMAL_MEMORY) {
4242 struct kmem_cache_node *n = get_node(s, node);
4244 if (flags & SO_TOTAL)
4245 x = atomic_long_read(&n->total_objects);
4246 else if (flags & SO_OBJECTS)
4247 x = atomic_long_read(&n->total_objects) -
4248 count_partial(n, count_free);
4251 x = atomic_long_read(&n->nr_slabs);
4258 if (flags & SO_PARTIAL) {
4259 for_each_node_state(node, N_NORMAL_MEMORY) {
4260 struct kmem_cache_node *n = get_node(s, node);
4262 if (flags & SO_TOTAL)
4263 x = count_partial(n, count_total);
4264 else if (flags & SO_OBJECTS)
4265 x = count_partial(n, count_inuse);
4272 x = sprintf(buf, "%lu", total);
4274 for_each_node_state(node, N_NORMAL_MEMORY)
4276 x += sprintf(buf + x, " N%d=%lu",
4279 unlock_memory_hotplug();
4281 return x + sprintf(buf + x, "\n");
4284 #ifdef CONFIG_SLUB_DEBUG
4285 static int any_slab_objects(struct kmem_cache *s)
4289 for_each_online_node(node) {
4290 struct kmem_cache_node *n = get_node(s, node);
4295 if (atomic_long_read(&n->total_objects))
4302 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4303 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4305 struct slab_attribute {
4306 struct attribute attr;
4307 ssize_t (*show)(struct kmem_cache *s, char *buf);
4308 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4311 #define SLAB_ATTR_RO(_name) \
4312 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4314 #define SLAB_ATTR(_name) \
4315 static struct slab_attribute _name##_attr = \
4316 __ATTR(_name, 0644, _name##_show, _name##_store)
4318 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4320 return sprintf(buf, "%d\n", s->size);
4322 SLAB_ATTR_RO(slab_size);
4324 static ssize_t align_show(struct kmem_cache *s, char *buf)
4326 return sprintf(buf, "%d\n", s->align);
4328 SLAB_ATTR_RO(align);
4330 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4332 return sprintf(buf, "%d\n", s->objsize);
4334 SLAB_ATTR_RO(object_size);
4336 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4338 return sprintf(buf, "%d\n", oo_objects(s->oo));
4340 SLAB_ATTR_RO(objs_per_slab);
4342 static ssize_t order_store(struct kmem_cache *s,
4343 const char *buf, size_t length)
4345 unsigned long order;
4348 err = strict_strtoul(buf, 10, &order);
4352 if (order > slub_max_order || order < slub_min_order)
4355 calculate_sizes(s, order);
4359 static ssize_t order_show(struct kmem_cache *s, char *buf)
4361 return sprintf(buf, "%d\n", oo_order(s->oo));
4365 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4367 return sprintf(buf, "%lu\n", s->min_partial);
4370 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4376 err = strict_strtoul(buf, 10, &min);
4380 set_min_partial(s, min);
4383 SLAB_ATTR(min_partial);
4385 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4389 return sprintf(buf, "%pS\n", s->ctor);
4393 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4395 return sprintf(buf, "%d\n", s->refcount - 1);
4397 SLAB_ATTR_RO(aliases);
4399 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4401 return show_slab_objects(s, buf, SO_PARTIAL);
4403 SLAB_ATTR_RO(partial);
4405 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4407 return show_slab_objects(s, buf, SO_CPU);
4409 SLAB_ATTR_RO(cpu_slabs);
4411 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4413 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4415 SLAB_ATTR_RO(objects);
4417 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4419 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4421 SLAB_ATTR_RO(objects_partial);
4423 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4425 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4428 static ssize_t reclaim_account_store(struct kmem_cache *s,
4429 const char *buf, size_t length)
4431 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4433 s->flags |= SLAB_RECLAIM_ACCOUNT;
4436 SLAB_ATTR(reclaim_account);
4438 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4440 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4442 SLAB_ATTR_RO(hwcache_align);
4444 #ifdef CONFIG_ZONE_DMA
4445 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4447 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4449 SLAB_ATTR_RO(cache_dma);
4452 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4454 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4456 SLAB_ATTR_RO(destroy_by_rcu);
4458 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4460 return sprintf(buf, "%d\n", s->reserved);
4462 SLAB_ATTR_RO(reserved);
4464 #ifdef CONFIG_SLUB_DEBUG
4465 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4467 return show_slab_objects(s, buf, SO_ALL);
4469 SLAB_ATTR_RO(slabs);
4471 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4473 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4475 SLAB_ATTR_RO(total_objects);
4477 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4479 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4482 static ssize_t sanity_checks_store(struct kmem_cache *s,
4483 const char *buf, size_t length)
4485 s->flags &= ~SLAB_DEBUG_FREE;
4486 if (buf[0] == '1') {
4487 s->flags &= ~__CMPXCHG_DOUBLE;
4488 s->flags |= SLAB_DEBUG_FREE;
4492 SLAB_ATTR(sanity_checks);
4494 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4496 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4499 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4502 s->flags &= ~SLAB_TRACE;
4503 if (buf[0] == '1') {
4504 s->flags &= ~__CMPXCHG_DOUBLE;
4505 s->flags |= SLAB_TRACE;
4511 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4513 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4516 static ssize_t red_zone_store(struct kmem_cache *s,
4517 const char *buf, size_t length)
4519 if (any_slab_objects(s))
4522 s->flags &= ~SLAB_RED_ZONE;
4523 if (buf[0] == '1') {
4524 s->flags &= ~__CMPXCHG_DOUBLE;
4525 s->flags |= SLAB_RED_ZONE;
4527 calculate_sizes(s, -1);
4530 SLAB_ATTR(red_zone);
4532 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4534 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4537 static ssize_t poison_store(struct kmem_cache *s,
4538 const char *buf, size_t length)
4540 if (any_slab_objects(s))
4543 s->flags &= ~SLAB_POISON;
4544 if (buf[0] == '1') {
4545 s->flags &= ~__CMPXCHG_DOUBLE;
4546 s->flags |= SLAB_POISON;
4548 calculate_sizes(s, -1);
4553 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4555 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4558 static ssize_t store_user_store(struct kmem_cache *s,
4559 const char *buf, size_t length)
4561 if (any_slab_objects(s))
4564 s->flags &= ~SLAB_STORE_USER;
4565 if (buf[0] == '1') {
4566 s->flags &= ~__CMPXCHG_DOUBLE;
4567 s->flags |= SLAB_STORE_USER;
4569 calculate_sizes(s, -1);
4572 SLAB_ATTR(store_user);
4574 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4579 static ssize_t validate_store(struct kmem_cache *s,
4580 const char *buf, size_t length)
4584 if (buf[0] == '1') {
4585 ret = validate_slab_cache(s);
4591 SLAB_ATTR(validate);
4593 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4595 if (!(s->flags & SLAB_STORE_USER))
4597 return list_locations(s, buf, TRACK_ALLOC);
4599 SLAB_ATTR_RO(alloc_calls);
4601 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4603 if (!(s->flags & SLAB_STORE_USER))
4605 return list_locations(s, buf, TRACK_FREE);
4607 SLAB_ATTR_RO(free_calls);
4608 #endif /* CONFIG_SLUB_DEBUG */
4610 #ifdef CONFIG_FAILSLAB
4611 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4613 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4616 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4619 s->flags &= ~SLAB_FAILSLAB;
4621 s->flags |= SLAB_FAILSLAB;
4624 SLAB_ATTR(failslab);
4627 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4632 static ssize_t shrink_store(struct kmem_cache *s,
4633 const char *buf, size_t length)
4635 if (buf[0] == '1') {
4636 int rc = kmem_cache_shrink(s);
4647 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4649 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4652 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4653 const char *buf, size_t length)
4655 unsigned long ratio;
4658 err = strict_strtoul(buf, 10, &ratio);
4663 s->remote_node_defrag_ratio = ratio * 10;
4667 SLAB_ATTR(remote_node_defrag_ratio);
4670 #ifdef CONFIG_SLUB_STATS
4671 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4673 unsigned long sum = 0;
4676 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4681 for_each_online_cpu(cpu) {
4682 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4688 len = sprintf(buf, "%lu", sum);
4691 for_each_online_cpu(cpu) {
4692 if (data[cpu] && len < PAGE_SIZE - 20)
4693 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4697 return len + sprintf(buf + len, "\n");
4700 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4704 for_each_online_cpu(cpu)
4705 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4708 #define STAT_ATTR(si, text) \
4709 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4711 return show_stat(s, buf, si); \
4713 static ssize_t text##_store(struct kmem_cache *s, \
4714 const char *buf, size_t length) \
4716 if (buf[0] != '0') \
4718 clear_stat(s, si); \
4723 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4724 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4725 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4726 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4727 STAT_ATTR(FREE_FROZEN, free_frozen);
4728 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4729 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4730 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4731 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4732 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4733 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4734 STAT_ATTR(FREE_SLAB, free_slab);
4735 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4736 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4737 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4738 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4739 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4740 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4741 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4742 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4743 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4744 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4747 static struct attribute *slab_attrs[] = {
4748 &slab_size_attr.attr,
4749 &object_size_attr.attr,
4750 &objs_per_slab_attr.attr,
4752 &min_partial_attr.attr,
4754 &objects_partial_attr.attr,
4756 &cpu_slabs_attr.attr,
4760 &hwcache_align_attr.attr,
4761 &reclaim_account_attr.attr,
4762 &destroy_by_rcu_attr.attr,
4764 &reserved_attr.attr,
4765 #ifdef CONFIG_SLUB_DEBUG
4766 &total_objects_attr.attr,
4768 &sanity_checks_attr.attr,
4770 &red_zone_attr.attr,
4772 &store_user_attr.attr,
4773 &validate_attr.attr,
4774 &alloc_calls_attr.attr,
4775 &free_calls_attr.attr,
4777 #ifdef CONFIG_ZONE_DMA
4778 &cache_dma_attr.attr,
4781 &remote_node_defrag_ratio_attr.attr,
4783 #ifdef CONFIG_SLUB_STATS
4784 &alloc_fastpath_attr.attr,
4785 &alloc_slowpath_attr.attr,
4786 &free_fastpath_attr.attr,
4787 &free_slowpath_attr.attr,
4788 &free_frozen_attr.attr,
4789 &free_add_partial_attr.attr,
4790 &free_remove_partial_attr.attr,
4791 &alloc_from_partial_attr.attr,
4792 &alloc_slab_attr.attr,
4793 &alloc_refill_attr.attr,
4794 &alloc_node_mismatch_attr.attr,
4795 &free_slab_attr.attr,
4796 &cpuslab_flush_attr.attr,
4797 &deactivate_full_attr.attr,
4798 &deactivate_empty_attr.attr,
4799 &deactivate_to_head_attr.attr,
4800 &deactivate_to_tail_attr.attr,
4801 &deactivate_remote_frees_attr.attr,
4802 &deactivate_bypass_attr.attr,
4803 &order_fallback_attr.attr,
4804 &cmpxchg_double_fail_attr.attr,
4805 &cmpxchg_double_cpu_fail_attr.attr,
4807 #ifdef CONFIG_FAILSLAB
4808 &failslab_attr.attr,
4814 static struct attribute_group slab_attr_group = {
4815 .attrs = slab_attrs,
4818 static ssize_t slab_attr_show(struct kobject *kobj,
4819 struct attribute *attr,
4822 struct slab_attribute *attribute;
4823 struct kmem_cache *s;
4826 attribute = to_slab_attr(attr);
4829 if (!attribute->show)
4832 err = attribute->show(s, buf);
4837 static ssize_t slab_attr_store(struct kobject *kobj,
4838 struct attribute *attr,
4839 const char *buf, size_t len)
4841 struct slab_attribute *attribute;
4842 struct kmem_cache *s;
4845 attribute = to_slab_attr(attr);
4848 if (!attribute->store)
4851 err = attribute->store(s, buf, len);
4856 static void kmem_cache_release(struct kobject *kobj)
4858 struct kmem_cache *s = to_slab(kobj);
4864 static const struct sysfs_ops slab_sysfs_ops = {
4865 .show = slab_attr_show,
4866 .store = slab_attr_store,
4869 static struct kobj_type slab_ktype = {
4870 .sysfs_ops = &slab_sysfs_ops,
4871 .release = kmem_cache_release
4874 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4876 struct kobj_type *ktype = get_ktype(kobj);
4878 if (ktype == &slab_ktype)
4883 static const struct kset_uevent_ops slab_uevent_ops = {
4884 .filter = uevent_filter,
4887 static struct kset *slab_kset;
4889 #define ID_STR_LENGTH 64
4891 /* Create a unique string id for a slab cache:
4893 * Format :[flags-]size
4895 static char *create_unique_id(struct kmem_cache *s)
4897 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4904 * First flags affecting slabcache operations. We will only
4905 * get here for aliasable slabs so we do not need to support
4906 * too many flags. The flags here must cover all flags that
4907 * are matched during merging to guarantee that the id is
4910 if (s->flags & SLAB_CACHE_DMA)
4912 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4914 if (s->flags & SLAB_DEBUG_FREE)
4916 if (!(s->flags & SLAB_NOTRACK))
4920 p += sprintf(p, "%07d", s->size);
4921 BUG_ON(p > name + ID_STR_LENGTH - 1);
4925 static int sysfs_slab_add(struct kmem_cache *s)
4931 if (slab_state < SYSFS)
4932 /* Defer until later */
4935 unmergeable = slab_unmergeable(s);
4938 * Slabcache can never be merged so we can use the name proper.
4939 * This is typically the case for debug situations. In that
4940 * case we can catch duplicate names easily.
4942 sysfs_remove_link(&slab_kset->kobj, s->name);
4946 * Create a unique name for the slab as a target
4949 name = create_unique_id(s);
4952 s->kobj.kset = slab_kset;
4953 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4955 kobject_put(&s->kobj);
4959 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4961 kobject_del(&s->kobj);
4962 kobject_put(&s->kobj);
4965 kobject_uevent(&s->kobj, KOBJ_ADD);
4967 /* Setup first alias */
4968 sysfs_slab_alias(s, s->name);
4974 static void sysfs_slab_remove(struct kmem_cache *s)
4976 if (slab_state < SYSFS)
4978 * Sysfs has not been setup yet so no need to remove the
4983 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4984 kobject_del(&s->kobj);
4985 kobject_put(&s->kobj);
4989 * Need to buffer aliases during bootup until sysfs becomes
4990 * available lest we lose that information.
4992 struct saved_alias {
4993 struct kmem_cache *s;
4995 struct saved_alias *next;
4998 static struct saved_alias *alias_list;
5000 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5002 struct saved_alias *al;
5004 if (slab_state == SYSFS) {
5006 * If we have a leftover link then remove it.
5008 sysfs_remove_link(&slab_kset->kobj, name);
5009 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5012 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5018 al->next = alias_list;
5023 static int __init slab_sysfs_init(void)
5025 struct kmem_cache *s;
5028 down_write(&slub_lock);
5030 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5032 up_write(&slub_lock);
5033 printk(KERN_ERR "Cannot register slab subsystem.\n");
5039 list_for_each_entry(s, &slab_caches, list) {
5040 err = sysfs_slab_add(s);
5042 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5043 " to sysfs\n", s->name);
5046 while (alias_list) {
5047 struct saved_alias *al = alias_list;
5049 alias_list = alias_list->next;
5050 err = sysfs_slab_alias(al->s, al->name);
5052 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5053 " %s to sysfs\n", s->name);
5057 up_write(&slub_lock);
5062 __initcall(slab_sysfs_init);
5063 #endif /* CONFIG_SYSFS */
5066 * The /proc/slabinfo ABI
5068 #ifdef CONFIG_SLABINFO
5069 static void print_slabinfo_header(struct seq_file *m)
5071 seq_puts(m, "slabinfo - version: 2.1\n");
5072 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5073 "<objperslab> <pagesperslab>");
5074 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5075 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5079 static void *s_start(struct seq_file *m, loff_t *pos)
5083 down_read(&slub_lock);
5085 print_slabinfo_header(m);
5087 return seq_list_start(&slab_caches, *pos);
5090 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5092 return seq_list_next(p, &slab_caches, pos);
5095 static void s_stop(struct seq_file *m, void *p)
5097 up_read(&slub_lock);
5100 static int s_show(struct seq_file *m, void *p)
5102 unsigned long nr_partials = 0;
5103 unsigned long nr_slabs = 0;
5104 unsigned long nr_inuse = 0;
5105 unsigned long nr_objs = 0;
5106 unsigned long nr_free = 0;
5107 struct kmem_cache *s;
5110 s = list_entry(p, struct kmem_cache, list);
5112 for_each_online_node(node) {
5113 struct kmem_cache_node *n = get_node(s, node);
5118 nr_partials += n->nr_partial;
5119 nr_slabs += atomic_long_read(&n->nr_slabs);
5120 nr_objs += atomic_long_read(&n->total_objects);
5121 nr_free += count_partial(n, count_free);
5124 nr_inuse = nr_objs - nr_free;
5126 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5127 nr_objs, s->size, oo_objects(s->oo),
5128 (1 << oo_order(s->oo)));
5129 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5130 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5136 static const struct seq_operations slabinfo_op = {
5143 static int slabinfo_open(struct inode *inode, struct file *file)
5145 return seq_open(file, &slabinfo_op);
5148 static const struct file_operations proc_slabinfo_operations = {
5149 .open = slabinfo_open,
5151 .llseek = seq_lseek,
5152 .release = seq_release,
5155 static int __init slab_proc_init(void)
5157 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5160 module_init(slab_proc_init);
5161 #endif /* CONFIG_SLABINFO */