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>
31 #include <linux/stacktrace.h>
32 #include <linux/prefetch.h>
34 #include <trace/events/kmem.h>
38 * 1. slub_lock (Global Semaphore)
40 * 3. slab_lock(page) (Only on some arches and for debugging)
44 * The role of the slub_lock is to protect the list of all the slabs
45 * and to synchronize major metadata changes to slab cache structures.
47 * The slab_lock is only used for debugging and on arches that do not
48 * have the ability to do a cmpxchg_double. It only protects the second
49 * double word in the page struct. Meaning
50 * A. page->freelist -> List of object free in a page
51 * B. page->counters -> Counters of objects
52 * C. page->frozen -> frozen state
54 * If a slab is frozen then it is exempt from list management. It is not
55 * on any list. The processor that froze the slab is the one who can
56 * perform list operations on the page. Other processors may put objects
57 * onto the freelist but the processor that froze the slab is the only
58 * one that can retrieve the objects from the page's freelist.
60 * The list_lock protects the partial and full list on each node and
61 * the partial slab counter. If taken then no new slabs may be added or
62 * removed from the lists nor make the number of partial slabs be modified.
63 * (Note that the total number of slabs is an atomic value that may be
64 * modified without taking the list lock).
66 * The list_lock is a centralized lock and thus we avoid taking it as
67 * much as possible. As long as SLUB does not have to handle partial
68 * slabs, operations can continue without any centralized lock. F.e.
69 * allocating a long series of objects that fill up slabs does not require
71 * Interrupts are disabled during allocation and deallocation in order to
72 * make the slab allocator safe to use in the context of an irq. In addition
73 * interrupts are disabled to ensure that the processor does not change
74 * while handling per_cpu slabs, due to kernel preemption.
76 * SLUB assigns one slab for allocation to each processor.
77 * Allocations only occur from these slabs called cpu slabs.
79 * Slabs with free elements are kept on a partial list and during regular
80 * operations no list for full slabs is used. If an object in a full slab is
81 * freed then the slab will show up again on the partial lists.
82 * We track full slabs for debugging purposes though because otherwise we
83 * cannot scan all objects.
85 * Slabs are freed when they become empty. Teardown and setup is
86 * minimal so we rely on the page allocators per cpu caches for
87 * fast frees and allocs.
89 * Overloading of page flags that are otherwise used for LRU management.
91 * PageActive The slab is frozen and exempt from list processing.
92 * This means that the slab is dedicated to a purpose
93 * such as satisfying allocations for a specific
94 * processor. Objects may be freed in the slab while
95 * it is frozen but slab_free will then skip the usual
96 * list operations. It is up to the processor holding
97 * the slab to integrate the slab into the slab lists
98 * when the slab is no longer needed.
100 * One use of this flag is to mark slabs that are
101 * used for allocations. Then such a slab becomes a cpu
102 * slab. The cpu slab may be equipped with an additional
103 * freelist that allows lockless access to
104 * free objects in addition to the regular freelist
105 * that requires the slab lock.
107 * PageError Slab requires special handling due to debug
108 * options set. This moves slab handling out of
109 * the fast path and disables lockless freelists.
112 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
113 SLAB_TRACE | SLAB_DEBUG_FREE)
115 static inline int kmem_cache_debug(struct kmem_cache *s)
117 #ifdef CONFIG_SLUB_DEBUG
118 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
125 * Issues still to be resolved:
127 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
129 * - Variable sizing of the per node arrays
132 /* Enable to test recovery from slab corruption on boot */
133 #undef SLUB_RESILIENCY_TEST
135 /* Enable to log cmpxchg failures */
136 #undef SLUB_DEBUG_CMPXCHG
139 * Mininum number of partial slabs. These will be left on the partial
140 * lists even if they are empty. kmem_cache_shrink may reclaim them.
142 #define MIN_PARTIAL 5
145 * Maximum number of desirable partial slabs.
146 * The existence of more partial slabs makes kmem_cache_shrink
147 * sort the partial list by the number of objects in the.
149 #define MAX_PARTIAL 10
151 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
152 SLAB_POISON | SLAB_STORE_USER)
155 * Debugging flags that require metadata to be stored in the slab. These get
156 * disabled when slub_debug=O is used and a cache's min order increases with
159 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
162 * Set of flags that will prevent slab merging
164 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
165 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
168 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
169 SLAB_CACHE_DMA | SLAB_NOTRACK)
172 #define OO_MASK ((1 << OO_SHIFT) - 1)
173 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
175 /* Internal SLUB flags */
176 #define __OBJECT_POISON 0x80000000UL /* Poison object */
177 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
179 static int kmem_size = sizeof(struct kmem_cache);
182 static struct notifier_block slab_notifier;
186 DOWN, /* No slab functionality available */
187 PARTIAL, /* Kmem_cache_node works */
188 UP, /* Everything works but does not show up in sysfs */
192 /* A list of all slab caches on the system */
193 static DECLARE_RWSEM(slub_lock);
194 static LIST_HEAD(slab_caches);
197 * Tracking user of a slab.
199 #define TRACK_ADDRS_COUNT 16
201 unsigned long addr; /* Called from address */
202 #ifdef CONFIG_STACKTRACE
203 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
205 int cpu; /* Was running on cpu */
206 int pid; /* Pid context */
207 unsigned long when; /* When did the operation occur */
210 enum track_item { TRACK_ALLOC, TRACK_FREE };
213 static int sysfs_slab_add(struct kmem_cache *);
214 static int sysfs_slab_alias(struct kmem_cache *, const char *);
215 static void sysfs_slab_remove(struct kmem_cache *);
218 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
219 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
221 static inline void sysfs_slab_remove(struct kmem_cache *s)
229 static inline void stat(const struct kmem_cache *s, enum stat_item si)
231 #ifdef CONFIG_SLUB_STATS
232 __this_cpu_inc(s->cpu_slab->stat[si]);
236 /********************************************************************
237 * Core slab cache functions
238 *******************************************************************/
240 int slab_is_available(void)
242 return slab_state >= UP;
245 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
247 return s->node[node];
250 /* Verify that a pointer has an address that is valid within a slab page */
251 static inline int check_valid_pointer(struct kmem_cache *s,
252 struct page *page, const void *object)
259 base = page_address(page);
260 if (object < base || object >= base + page->objects * s->size ||
261 (object - base) % s->size) {
268 static inline void *get_freepointer(struct kmem_cache *s, void *object)
270 return *(void **)(object + s->offset);
273 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
275 prefetch(object + s->offset);
278 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
282 #ifdef CONFIG_DEBUG_PAGEALLOC
283 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
285 p = get_freepointer(s, object);
290 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
292 *(void **)(object + s->offset) = fp;
295 /* Loop over all objects in a slab */
296 #define for_each_object(__p, __s, __addr, __objects) \
297 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
300 /* Determine object index from a given position */
301 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
303 return (p - addr) / s->size;
306 static inline size_t slab_ksize(const struct kmem_cache *s)
308 #ifdef CONFIG_SLUB_DEBUG
310 * Debugging requires use of the padding between object
311 * and whatever may come after it.
313 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
318 * If we have the need to store the freelist pointer
319 * back there or track user information then we can
320 * only use the space before that information.
322 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
325 * Else we can use all the padding etc for the allocation
330 static inline int order_objects(int order, unsigned long size, int reserved)
332 return ((PAGE_SIZE << order) - reserved) / size;
335 static inline struct kmem_cache_order_objects oo_make(int order,
336 unsigned long size, int reserved)
338 struct kmem_cache_order_objects x = {
339 (order << OO_SHIFT) + order_objects(order, size, reserved)
345 static inline int oo_order(struct kmem_cache_order_objects x)
347 return x.x >> OO_SHIFT;
350 static inline int oo_objects(struct kmem_cache_order_objects x)
352 return x.x & OO_MASK;
356 * Per slab locking using the pagelock
358 static __always_inline void slab_lock(struct page *page)
360 bit_spin_lock(PG_locked, &page->flags);
363 static __always_inline void slab_unlock(struct page *page)
365 __bit_spin_unlock(PG_locked, &page->flags);
368 /* Interrupts must be disabled (for the fallback code to work right) */
369 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
370 void *freelist_old, unsigned long counters_old,
371 void *freelist_new, unsigned long counters_new,
374 VM_BUG_ON(!irqs_disabled());
375 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
376 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
377 if (s->flags & __CMPXCHG_DOUBLE) {
378 if (cmpxchg_double(&page->freelist, &page->counters,
379 freelist_old, counters_old,
380 freelist_new, counters_new))
386 if (page->freelist == freelist_old && page->counters == counters_old) {
387 page->freelist = freelist_new;
388 page->counters = counters_new;
396 stat(s, CMPXCHG_DOUBLE_FAIL);
398 #ifdef SLUB_DEBUG_CMPXCHG
399 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
405 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
406 void *freelist_old, unsigned long counters_old,
407 void *freelist_new, unsigned long counters_new,
410 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
411 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
412 if (s->flags & __CMPXCHG_DOUBLE) {
413 if (cmpxchg_double(&page->freelist, &page->counters,
414 freelist_old, counters_old,
415 freelist_new, counters_new))
422 local_irq_save(flags);
424 if (page->freelist == freelist_old && page->counters == counters_old) {
425 page->freelist = freelist_new;
426 page->counters = counters_new;
428 local_irq_restore(flags);
432 local_irq_restore(flags);
436 stat(s, CMPXCHG_DOUBLE_FAIL);
438 #ifdef SLUB_DEBUG_CMPXCHG
439 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
445 #ifdef CONFIG_SLUB_DEBUG
447 * Determine a map of object in use on a page.
449 * Node listlock must be held to guarantee that the page does
450 * not vanish from under us.
452 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
455 void *addr = page_address(page);
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(slab_index(p, s, addr), map);
464 #ifdef CONFIG_SLUB_DEBUG_ON
465 static int slub_debug = DEBUG_DEFAULT_FLAGS;
467 static int slub_debug;
470 static char *slub_debug_slabs;
471 static int disable_higher_order_debug;
476 static void print_section(char *text, u8 *addr, unsigned int length)
478 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
482 static struct track *get_track(struct kmem_cache *s, void *object,
483 enum track_item alloc)
488 p = object + s->offset + sizeof(void *);
490 p = object + s->inuse;
495 static void set_track(struct kmem_cache *s, void *object,
496 enum track_item alloc, unsigned long addr)
498 struct track *p = get_track(s, object, alloc);
501 #ifdef CONFIG_STACKTRACE
502 struct stack_trace trace;
505 trace.nr_entries = 0;
506 trace.max_entries = TRACK_ADDRS_COUNT;
507 trace.entries = p->addrs;
509 save_stack_trace(&trace);
511 /* See rant in lockdep.c */
512 if (trace.nr_entries != 0 &&
513 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
516 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
520 p->cpu = smp_processor_id();
521 p->pid = current->pid;
524 memset(p, 0, sizeof(struct track));
527 static void init_tracking(struct kmem_cache *s, void *object)
529 if (!(s->flags & SLAB_STORE_USER))
532 set_track(s, object, TRACK_FREE, 0UL);
533 set_track(s, object, TRACK_ALLOC, 0UL);
536 static void print_track(const char *s, struct track *t)
541 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
542 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
543 #ifdef CONFIG_STACKTRACE
546 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
548 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
555 static void print_tracking(struct kmem_cache *s, void *object)
557 if (!(s->flags & SLAB_STORE_USER))
560 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
561 print_track("Freed", get_track(s, object, TRACK_FREE));
564 static void print_page_info(struct page *page)
566 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
567 page, page->objects, page->inuse, page->freelist, page->flags);
571 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
577 vsnprintf(buf, sizeof(buf), fmt, args);
579 printk(KERN_ERR "========================================"
580 "=====================================\n");
581 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
582 printk(KERN_ERR "----------------------------------------"
583 "-------------------------------------\n\n");
586 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
592 vsnprintf(buf, sizeof(buf), fmt, args);
594 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
597 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
599 unsigned int off; /* Offset of last byte */
600 u8 *addr = page_address(page);
602 print_tracking(s, p);
604 print_page_info(page);
606 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607 p, p - addr, get_freepointer(s, p));
610 print_section("Bytes b4 ", p - 16, 16);
612 print_section("Object ", p, min_t(unsigned long, s->objsize,
614 if (s->flags & SLAB_RED_ZONE)
615 print_section("Redzone ", p + s->objsize,
616 s->inuse - s->objsize);
619 off = s->offset + sizeof(void *);
623 if (s->flags & SLAB_STORE_USER)
624 off += 2 * sizeof(struct track);
627 /* Beginning of the filler is the free pointer */
628 print_section("Padding ", p + off, s->size - off);
633 static void object_err(struct kmem_cache *s, struct page *page,
634 u8 *object, char *reason)
636 slab_bug(s, "%s", reason);
637 print_trailer(s, page, object);
640 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
646 vsnprintf(buf, sizeof(buf), fmt, args);
648 slab_bug(s, "%s", buf);
649 print_page_info(page);
653 static void init_object(struct kmem_cache *s, void *object, u8 val)
657 if (s->flags & __OBJECT_POISON) {
658 memset(p, POISON_FREE, s->objsize - 1);
659 p[s->objsize - 1] = POISON_END;
662 if (s->flags & SLAB_RED_ZONE)
663 memset(p + s->objsize, val, s->inuse - s->objsize);
666 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667 void *from, void *to)
669 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670 memset(from, data, to - from);
673 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674 u8 *object, char *what,
675 u8 *start, unsigned int value, unsigned int bytes)
680 fault = memchr_inv(start, value, bytes);
685 while (end > fault && end[-1] == value)
688 slab_bug(s, "%s overwritten", what);
689 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690 fault, end - 1, fault[0], value);
691 print_trailer(s, page, object);
693 restore_bytes(s, what, value, fault, end);
701 * Bytes of the object to be managed.
702 * If the freepointer may overlay the object then the free
703 * pointer is the first word of the object.
705 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
708 * object + s->objsize
709 * Padding to reach word boundary. This is also used for Redzoning.
710 * Padding is extended by another word if Redzoning is enabled and
713 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714 * 0xcc (RED_ACTIVE) for objects in use.
717 * Meta data starts here.
719 * A. Free pointer (if we cannot overwrite object on free)
720 * B. Tracking data for SLAB_STORE_USER
721 * C. Padding to reach required alignment boundary or at mininum
722 * one word if debugging is on to be able to detect writes
723 * before the word boundary.
725 * Padding is done using 0x5a (POISON_INUSE)
728 * Nothing is used beyond s->size.
730 * If slabcaches are merged then the objsize and inuse boundaries are mostly
731 * ignored. And therefore no slab options that rely on these boundaries
732 * may be used with merged slabcaches.
735 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
737 unsigned long off = s->inuse; /* The end of info */
740 /* Freepointer is placed after the object. */
741 off += sizeof(void *);
743 if (s->flags & SLAB_STORE_USER)
744 /* We also have user information there */
745 off += 2 * sizeof(struct track);
750 return check_bytes_and_report(s, page, p, "Object padding",
751 p + off, POISON_INUSE, s->size - off);
754 /* Check the pad bytes at the end of a slab page */
755 static int slab_pad_check(struct kmem_cache *s, struct page *page)
763 if (!(s->flags & SLAB_POISON))
766 start = page_address(page);
767 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768 end = start + length;
769 remainder = length % s->size;
773 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
776 while (end > fault && end[-1] == POISON_INUSE)
779 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780 print_section("Padding ", end - remainder, remainder);
782 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
786 static int check_object(struct kmem_cache *s, struct page *page,
787 void *object, u8 val)
790 u8 *endobject = object + s->objsize;
792 if (s->flags & SLAB_RED_ZONE) {
793 if (!check_bytes_and_report(s, page, object, "Redzone",
794 endobject, val, s->inuse - s->objsize))
797 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
798 check_bytes_and_report(s, page, p, "Alignment padding",
799 endobject, POISON_INUSE, s->inuse - s->objsize);
803 if (s->flags & SLAB_POISON) {
804 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
805 (!check_bytes_and_report(s, page, p, "Poison", p,
806 POISON_FREE, s->objsize - 1) ||
807 !check_bytes_and_report(s, page, p, "Poison",
808 p + s->objsize - 1, POISON_END, 1)))
811 * check_pad_bytes cleans up on its own.
813 check_pad_bytes(s, page, p);
816 if (!s->offset && val == SLUB_RED_ACTIVE)
818 * Object and freepointer overlap. Cannot check
819 * freepointer while object is allocated.
823 /* Check free pointer validity */
824 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
825 object_err(s, page, p, "Freepointer corrupt");
827 * No choice but to zap it and thus lose the remainder
828 * of the free objects in this slab. May cause
829 * another error because the object count is now wrong.
831 set_freepointer(s, p, NULL);
837 static int check_slab(struct kmem_cache *s, struct page *page)
841 VM_BUG_ON(!irqs_disabled());
843 if (!PageSlab(page)) {
844 slab_err(s, page, "Not a valid slab page");
848 maxobj = order_objects(compound_order(page), s->size, s->reserved);
849 if (page->objects > maxobj) {
850 slab_err(s, page, "objects %u > max %u",
851 s->name, page->objects, maxobj);
854 if (page->inuse > page->objects) {
855 slab_err(s, page, "inuse %u > max %u",
856 s->name, page->inuse, page->objects);
859 /* Slab_pad_check fixes things up after itself */
860 slab_pad_check(s, page);
865 * Determine if a certain object on a page is on the freelist. Must hold the
866 * slab lock to guarantee that the chains are in a consistent state.
868 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
873 unsigned long max_objects;
876 while (fp && nr <= page->objects) {
879 if (!check_valid_pointer(s, page, fp)) {
881 object_err(s, page, object,
882 "Freechain corrupt");
883 set_freepointer(s, object, NULL);
886 slab_err(s, page, "Freepointer corrupt");
887 page->freelist = NULL;
888 page->inuse = page->objects;
889 slab_fix(s, "Freelist cleared");
895 fp = get_freepointer(s, object);
899 max_objects = order_objects(compound_order(page), s->size, s->reserved);
900 if (max_objects > MAX_OBJS_PER_PAGE)
901 max_objects = MAX_OBJS_PER_PAGE;
903 if (page->objects != max_objects) {
904 slab_err(s, page, "Wrong number of objects. Found %d but "
905 "should be %d", page->objects, max_objects);
906 page->objects = max_objects;
907 slab_fix(s, "Number of objects adjusted.");
909 if (page->inuse != page->objects - nr) {
910 slab_err(s, page, "Wrong object count. Counter is %d but "
911 "counted were %d", page->inuse, page->objects - nr);
912 page->inuse = page->objects - nr;
913 slab_fix(s, "Object count adjusted.");
915 return search == NULL;
918 static void trace(struct kmem_cache *s, struct page *page, void *object,
921 if (s->flags & SLAB_TRACE) {
922 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
924 alloc ? "alloc" : "free",
929 print_section("Object ", (void *)object, s->objsize);
936 * Hooks for other subsystems that check memory allocations. In a typical
937 * production configuration these hooks all should produce no code at all.
939 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
941 flags &= gfp_allowed_mask;
942 lockdep_trace_alloc(flags);
943 might_sleep_if(flags & __GFP_WAIT);
945 return should_failslab(s->objsize, flags, s->flags);
948 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
950 flags &= gfp_allowed_mask;
951 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
952 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
955 static inline void slab_free_hook(struct kmem_cache *s, void *x)
957 kmemleak_free_recursive(x, s->flags);
960 * Trouble is that we may no longer disable interupts in the fast path
961 * So in order to make the debug calls that expect irqs to be
962 * disabled we need to disable interrupts temporarily.
964 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
968 local_irq_save(flags);
969 kmemcheck_slab_free(s, x, s->objsize);
970 debug_check_no_locks_freed(x, s->objsize);
971 local_irq_restore(flags);
974 if (!(s->flags & SLAB_DEBUG_OBJECTS))
975 debug_check_no_obj_freed(x, s->objsize);
979 * Tracking of fully allocated slabs for debugging purposes.
981 * list_lock must be held.
983 static void add_full(struct kmem_cache *s,
984 struct kmem_cache_node *n, struct page *page)
986 if (!(s->flags & SLAB_STORE_USER))
989 list_add(&page->lru, &n->full);
993 * list_lock must be held.
995 static void remove_full(struct kmem_cache *s, struct page *page)
997 if (!(s->flags & SLAB_STORE_USER))
1000 list_del(&page->lru);
1003 /* Tracking of the number of slabs for debugging purposes */
1004 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1006 struct kmem_cache_node *n = get_node(s, node);
1008 return atomic_long_read(&n->nr_slabs);
1011 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1013 return atomic_long_read(&n->nr_slabs);
1016 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1018 struct kmem_cache_node *n = get_node(s, node);
1021 * May be called early in order to allocate a slab for the
1022 * kmem_cache_node structure. Solve the chicken-egg
1023 * dilemma by deferring the increment of the count during
1024 * bootstrap (see early_kmem_cache_node_alloc).
1027 atomic_long_inc(&n->nr_slabs);
1028 atomic_long_add(objects, &n->total_objects);
1031 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1033 struct kmem_cache_node *n = get_node(s, node);
1035 atomic_long_dec(&n->nr_slabs);
1036 atomic_long_sub(objects, &n->total_objects);
1039 /* Object debug checks for alloc/free paths */
1040 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1043 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1046 init_object(s, object, SLUB_RED_INACTIVE);
1047 init_tracking(s, object);
1050 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1051 void *object, unsigned long addr)
1053 if (!check_slab(s, page))
1056 if (!check_valid_pointer(s, page, object)) {
1057 object_err(s, page, object, "Freelist Pointer check fails");
1061 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1064 /* Success perform special debug activities for allocs */
1065 if (s->flags & SLAB_STORE_USER)
1066 set_track(s, object, TRACK_ALLOC, addr);
1067 trace(s, page, object, 1);
1068 init_object(s, object, SLUB_RED_ACTIVE);
1072 if (PageSlab(page)) {
1074 * If this is a slab page then lets do the best we can
1075 * to avoid issues in the future. Marking all objects
1076 * as used avoids touching the remaining objects.
1078 slab_fix(s, "Marking all objects used");
1079 page->inuse = page->objects;
1080 page->freelist = NULL;
1085 static noinline int free_debug_processing(struct kmem_cache *s,
1086 struct page *page, void *object, unsigned long addr)
1088 unsigned long flags;
1091 local_irq_save(flags);
1094 if (!check_slab(s, page))
1097 if (!check_valid_pointer(s, page, object)) {
1098 slab_err(s, page, "Invalid object pointer 0x%p", object);
1102 if (on_freelist(s, page, object)) {
1103 object_err(s, page, object, "Object already free");
1107 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1110 if (unlikely(s != page->slab)) {
1111 if (!PageSlab(page)) {
1112 slab_err(s, page, "Attempt to free object(0x%p) "
1113 "outside of slab", object);
1114 } else if (!page->slab) {
1116 "SLUB <none>: no slab for object 0x%p.\n",
1120 object_err(s, page, object,
1121 "page slab pointer corrupt.");
1125 if (s->flags & SLAB_STORE_USER)
1126 set_track(s, object, TRACK_FREE, addr);
1127 trace(s, page, object, 0);
1128 init_object(s, object, SLUB_RED_INACTIVE);
1132 local_irq_restore(flags);
1136 slab_fix(s, "Object at 0x%p not freed", object);
1140 static int __init setup_slub_debug(char *str)
1142 slub_debug = DEBUG_DEFAULT_FLAGS;
1143 if (*str++ != '=' || !*str)
1145 * No options specified. Switch on full debugging.
1151 * No options but restriction on slabs. This means full
1152 * debugging for slabs matching a pattern.
1156 if (tolower(*str) == 'o') {
1158 * Avoid enabling debugging on caches if its minimum order
1159 * would increase as a result.
1161 disable_higher_order_debug = 1;
1168 * Switch off all debugging measures.
1173 * Determine which debug features should be switched on
1175 for (; *str && *str != ','; str++) {
1176 switch (tolower(*str)) {
1178 slub_debug |= SLAB_DEBUG_FREE;
1181 slub_debug |= SLAB_RED_ZONE;
1184 slub_debug |= SLAB_POISON;
1187 slub_debug |= SLAB_STORE_USER;
1190 slub_debug |= SLAB_TRACE;
1193 slub_debug |= SLAB_FAILSLAB;
1196 printk(KERN_ERR "slub_debug option '%c' "
1197 "unknown. skipped\n", *str);
1203 slub_debug_slabs = str + 1;
1208 __setup("slub_debug", setup_slub_debug);
1210 static unsigned long kmem_cache_flags(unsigned long objsize,
1211 unsigned long flags, const char *name,
1212 void (*ctor)(void *))
1215 * Enable debugging if selected on the kernel commandline.
1217 if (slub_debug && (!slub_debug_slabs ||
1218 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1219 flags |= slub_debug;
1224 static inline void setup_object_debug(struct kmem_cache *s,
1225 struct page *page, void *object) {}
1227 static inline int alloc_debug_processing(struct kmem_cache *s,
1228 struct page *page, void *object, unsigned long addr) { return 0; }
1230 static inline int free_debug_processing(struct kmem_cache *s,
1231 struct page *page, void *object, unsigned long addr) { return 0; }
1233 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1235 static inline int check_object(struct kmem_cache *s, struct page *page,
1236 void *object, u8 val) { return 1; }
1237 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1238 struct page *page) {}
1239 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1240 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1241 unsigned long flags, const char *name,
1242 void (*ctor)(void *))
1246 #define slub_debug 0
1248 #define disable_higher_order_debug 0
1250 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1252 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1254 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1256 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1259 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1262 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1265 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1267 #endif /* CONFIG_SLUB_DEBUG */
1270 * Slab allocation and freeing
1272 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1273 struct kmem_cache_order_objects oo)
1275 int order = oo_order(oo);
1277 flags |= __GFP_NOTRACK;
1279 if (node == NUMA_NO_NODE)
1280 return alloc_pages(flags, order);
1282 return alloc_pages_exact_node(node, flags, order);
1285 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1288 struct kmem_cache_order_objects oo = s->oo;
1291 flags &= gfp_allowed_mask;
1293 if (flags & __GFP_WAIT)
1296 flags |= s->allocflags;
1299 * Let the initial higher-order allocation fail under memory pressure
1300 * so we fall-back to the minimum order allocation.
1302 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1304 page = alloc_slab_page(alloc_gfp, node, oo);
1305 if (unlikely(!page)) {
1308 * Allocation may have failed due to fragmentation.
1309 * Try a lower order alloc if possible
1311 page = alloc_slab_page(flags, node, oo);
1314 stat(s, ORDER_FALLBACK);
1317 if (flags & __GFP_WAIT)
1318 local_irq_disable();
1323 if (kmemcheck_enabled
1324 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1325 int pages = 1 << oo_order(oo);
1327 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1330 * Objects from caches that have a constructor don't get
1331 * cleared when they're allocated, so we need to do it here.
1334 kmemcheck_mark_uninitialized_pages(page, pages);
1336 kmemcheck_mark_unallocated_pages(page, pages);
1339 page->objects = oo_objects(oo);
1340 mod_zone_page_state(page_zone(page),
1341 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1342 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1348 static void setup_object(struct kmem_cache *s, struct page *page,
1351 setup_object_debug(s, page, object);
1352 if (unlikely(s->ctor))
1356 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1363 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1365 page = allocate_slab(s,
1366 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1370 inc_slabs_node(s, page_to_nid(page), page->objects);
1372 page->flags |= 1 << PG_slab;
1374 start = page_address(page);
1376 if (unlikely(s->flags & SLAB_POISON))
1377 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1380 for_each_object(p, s, start, page->objects) {
1381 setup_object(s, page, last);
1382 set_freepointer(s, last, p);
1385 setup_object(s, page, last);
1386 set_freepointer(s, last, NULL);
1388 page->freelist = start;
1389 page->inuse = page->objects;
1395 static void __free_slab(struct kmem_cache *s, struct page *page)
1397 int order = compound_order(page);
1398 int pages = 1 << order;
1400 if (kmem_cache_debug(s)) {
1403 slab_pad_check(s, page);
1404 for_each_object(p, s, page_address(page),
1406 check_object(s, page, p, SLUB_RED_INACTIVE);
1409 kmemcheck_free_shadow(page, compound_order(page));
1411 mod_zone_page_state(page_zone(page),
1412 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1413 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1416 __ClearPageSlab(page);
1417 reset_page_mapcount(page);
1418 if (current->reclaim_state)
1419 current->reclaim_state->reclaimed_slab += pages;
1420 __free_pages(page, order);
1423 #define need_reserve_slab_rcu \
1424 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1426 static void rcu_free_slab(struct rcu_head *h)
1430 if (need_reserve_slab_rcu)
1431 page = virt_to_head_page(h);
1433 page = container_of((struct list_head *)h, struct page, lru);
1435 __free_slab(page->slab, page);
1438 static void free_slab(struct kmem_cache *s, struct page *page)
1440 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1441 struct rcu_head *head;
1443 if (need_reserve_slab_rcu) {
1444 int order = compound_order(page);
1445 int offset = (PAGE_SIZE << order) - s->reserved;
1447 VM_BUG_ON(s->reserved != sizeof(*head));
1448 head = page_address(page) + offset;
1451 * RCU free overloads the RCU head over the LRU
1453 head = (void *)&page->lru;
1456 call_rcu(head, rcu_free_slab);
1458 __free_slab(s, page);
1461 static void discard_slab(struct kmem_cache *s, struct page *page)
1463 dec_slabs_node(s, page_to_nid(page), page->objects);
1468 * Management of partially allocated slabs.
1470 * list_lock must be held.
1472 static inline void add_partial(struct kmem_cache_node *n,
1473 struct page *page, int tail)
1476 if (tail == DEACTIVATE_TO_TAIL)
1477 list_add_tail(&page->lru, &n->partial);
1479 list_add(&page->lru, &n->partial);
1483 * list_lock must be held.
1485 static inline void remove_partial(struct kmem_cache_node *n,
1488 list_del(&page->lru);
1493 * Remove slab from the partial list, freeze it and
1494 * return the pointer to the freelist.
1496 * Returns a list of objects or NULL if it fails.
1498 * Must hold list_lock since we modify the partial list.
1500 static inline void *acquire_slab(struct kmem_cache *s,
1501 struct kmem_cache_node *n, struct page *page,
1505 unsigned long counters;
1509 * Zap the freelist and set the frozen bit.
1510 * The old freelist is the list of objects for the
1511 * per cpu allocation list.
1513 freelist = page->freelist;
1514 counters = page->counters;
1515 new.counters = counters;
1517 new.inuse = page->objects;
1519 VM_BUG_ON(new.frozen);
1522 if (!__cmpxchg_double_slab(s, page,
1529 remove_partial(n, page);
1534 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1537 * Try to allocate a partial slab from a specific node.
1539 static void *get_partial_node(struct kmem_cache *s,
1540 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1542 struct page *page, *page2;
1543 void *object = NULL;
1546 * Racy check. If we mistakenly see no partial slabs then we
1547 * just allocate an empty slab. If we mistakenly try to get a
1548 * partial slab and there is none available then get_partials()
1551 if (!n || !n->nr_partial)
1554 spin_lock(&n->list_lock);
1555 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1556 void *t = acquire_slab(s, n, page, object == NULL);
1564 stat(s, ALLOC_FROM_PARTIAL);
1566 available = page->objects - page->inuse;
1569 available = put_cpu_partial(s, page, 0);
1570 stat(s, CPU_PARTIAL_NODE);
1572 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1576 spin_unlock(&n->list_lock);
1581 * Get a page from somewhere. Search in increasing NUMA distances.
1583 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1584 struct kmem_cache_cpu *c)
1587 struct zonelist *zonelist;
1590 enum zone_type high_zoneidx = gfp_zone(flags);
1592 unsigned int cpuset_mems_cookie;
1595 * The defrag ratio allows a configuration of the tradeoffs between
1596 * inter node defragmentation and node local allocations. A lower
1597 * defrag_ratio increases the tendency to do local allocations
1598 * instead of attempting to obtain partial slabs from other nodes.
1600 * If the defrag_ratio is set to 0 then kmalloc() always
1601 * returns node local objects. If the ratio is higher then kmalloc()
1602 * may return off node objects because partial slabs are obtained
1603 * from other nodes and filled up.
1605 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1606 * defrag_ratio = 1000) then every (well almost) allocation will
1607 * first attempt to defrag slab caches on other nodes. This means
1608 * scanning over all nodes to look for partial slabs which may be
1609 * expensive if we do it every time we are trying to find a slab
1610 * with available objects.
1612 if (!s->remote_node_defrag_ratio ||
1613 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1617 cpuset_mems_cookie = get_mems_allowed();
1618 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1619 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1620 struct kmem_cache_node *n;
1622 n = get_node(s, zone_to_nid(zone));
1624 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1625 n->nr_partial > s->min_partial) {
1626 object = get_partial_node(s, n, c);
1629 * Return the object even if
1630 * put_mems_allowed indicated that
1631 * the cpuset mems_allowed was
1632 * updated in parallel. It's a
1633 * harmless race between the alloc
1634 * and the cpuset update.
1636 put_mems_allowed(cpuset_mems_cookie);
1641 } while (!put_mems_allowed(cpuset_mems_cookie));
1647 * Get a partial page, lock it and return it.
1649 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1650 struct kmem_cache_cpu *c)
1653 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1655 object = get_partial_node(s, get_node(s, searchnode), c);
1656 if (object || node != NUMA_NO_NODE)
1659 return get_any_partial(s, flags, c);
1662 #ifdef CONFIG_PREEMPT
1664 * Calculate the next globally unique transaction for disambiguiation
1665 * during cmpxchg. The transactions start with the cpu number and are then
1666 * incremented by CONFIG_NR_CPUS.
1668 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1671 * No preemption supported therefore also no need to check for
1677 static inline unsigned long next_tid(unsigned long tid)
1679 return tid + TID_STEP;
1682 static inline unsigned int tid_to_cpu(unsigned long tid)
1684 return tid % TID_STEP;
1687 static inline unsigned long tid_to_event(unsigned long tid)
1689 return tid / TID_STEP;
1692 static inline unsigned int init_tid(int cpu)
1697 static inline void note_cmpxchg_failure(const char *n,
1698 const struct kmem_cache *s, unsigned long tid)
1700 #ifdef SLUB_DEBUG_CMPXCHG
1701 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1703 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1705 #ifdef CONFIG_PREEMPT
1706 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1707 printk("due to cpu change %d -> %d\n",
1708 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1711 if (tid_to_event(tid) != tid_to_event(actual_tid))
1712 printk("due to cpu running other code. Event %ld->%ld\n",
1713 tid_to_event(tid), tid_to_event(actual_tid));
1715 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1716 actual_tid, tid, next_tid(tid));
1718 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1721 void init_kmem_cache_cpus(struct kmem_cache *s)
1725 for_each_possible_cpu(cpu)
1726 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1730 * Remove the cpu slab
1732 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1734 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1735 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1737 enum slab_modes l = M_NONE, m = M_NONE;
1739 int tail = DEACTIVATE_TO_HEAD;
1743 if (page->freelist) {
1744 stat(s, DEACTIVATE_REMOTE_FREES);
1745 tail = DEACTIVATE_TO_TAIL;
1749 * Stage one: Free all available per cpu objects back
1750 * to the page freelist while it is still frozen. Leave the
1753 * There is no need to take the list->lock because the page
1756 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1758 unsigned long counters;
1761 prior = page->freelist;
1762 counters = page->counters;
1763 set_freepointer(s, freelist, prior);
1764 new.counters = counters;
1766 VM_BUG_ON(!new.frozen);
1768 } while (!__cmpxchg_double_slab(s, page,
1770 freelist, new.counters,
1771 "drain percpu freelist"));
1773 freelist = nextfree;
1777 * Stage two: Ensure that the page is unfrozen while the
1778 * list presence reflects the actual number of objects
1781 * We setup the list membership and then perform a cmpxchg
1782 * with the count. If there is a mismatch then the page
1783 * is not unfrozen but the page is on the wrong list.
1785 * Then we restart the process which may have to remove
1786 * the page from the list that we just put it on again
1787 * because the number of objects in the slab may have
1792 old.freelist = page->freelist;
1793 old.counters = page->counters;
1794 VM_BUG_ON(!old.frozen);
1796 /* Determine target state of the slab */
1797 new.counters = old.counters;
1800 set_freepointer(s, freelist, old.freelist);
1801 new.freelist = freelist;
1803 new.freelist = old.freelist;
1807 if (!new.inuse && n->nr_partial > s->min_partial)
1809 else if (new.freelist) {
1814 * Taking the spinlock removes the possiblity
1815 * that acquire_slab() will see a slab page that
1818 spin_lock(&n->list_lock);
1822 if (kmem_cache_debug(s) && !lock) {
1825 * This also ensures that the scanning of full
1826 * slabs from diagnostic functions will not see
1829 spin_lock(&n->list_lock);
1837 remove_partial(n, page);
1839 else if (l == M_FULL)
1841 remove_full(s, page);
1843 if (m == M_PARTIAL) {
1845 add_partial(n, page, tail);
1848 } else if (m == M_FULL) {
1850 stat(s, DEACTIVATE_FULL);
1851 add_full(s, n, page);
1857 if (!__cmpxchg_double_slab(s, page,
1858 old.freelist, old.counters,
1859 new.freelist, new.counters,
1864 spin_unlock(&n->list_lock);
1867 stat(s, DEACTIVATE_EMPTY);
1868 discard_slab(s, page);
1873 /* Unfreeze all the cpu partial slabs */
1874 static void unfreeze_partials(struct kmem_cache *s)
1876 struct kmem_cache_node *n = NULL;
1877 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1878 struct page *page, *discard_page = NULL;
1880 while ((page = c->partial)) {
1881 enum slab_modes { M_PARTIAL, M_FREE };
1882 enum slab_modes l, m;
1886 c->partial = page->next;
1891 old.freelist = page->freelist;
1892 old.counters = page->counters;
1893 VM_BUG_ON(!old.frozen);
1895 new.counters = old.counters;
1896 new.freelist = old.freelist;
1900 if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1903 struct kmem_cache_node *n2 = get_node(s,
1909 spin_unlock(&n->list_lock);
1912 spin_lock(&n->list_lock);
1917 if (l == M_PARTIAL) {
1918 remove_partial(n, page);
1919 stat(s, FREE_REMOVE_PARTIAL);
1921 add_partial(n, page,
1922 DEACTIVATE_TO_TAIL);
1923 stat(s, FREE_ADD_PARTIAL);
1929 } while (!cmpxchg_double_slab(s, page,
1930 old.freelist, old.counters,
1931 new.freelist, new.counters,
1932 "unfreezing slab"));
1935 page->next = discard_page;
1936 discard_page = page;
1941 spin_unlock(&n->list_lock);
1943 while (discard_page) {
1944 page = discard_page;
1945 discard_page = discard_page->next;
1947 stat(s, DEACTIVATE_EMPTY);
1948 discard_slab(s, page);
1954 * Put a page that was just frozen (in __slab_free) into a partial page
1955 * slot if available. This is done without interrupts disabled and without
1956 * preemption disabled. The cmpxchg is racy and may put the partial page
1957 * onto a random cpus partial slot.
1959 * If we did not find a slot then simply move all the partials to the
1960 * per node partial list.
1962 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1964 struct page *oldpage;
1971 oldpage = this_cpu_read(s->cpu_slab->partial);
1974 pobjects = oldpage->pobjects;
1975 pages = oldpage->pages;
1976 if (drain && pobjects > s->cpu_partial) {
1977 unsigned long flags;
1979 * partial array is full. Move the existing
1980 * set to the per node partial list.
1982 local_irq_save(flags);
1983 unfreeze_partials(s);
1984 local_irq_restore(flags);
1987 stat(s, CPU_PARTIAL_DRAIN);
1992 pobjects += page->objects - page->inuse;
1994 page->pages = pages;
1995 page->pobjects = pobjects;
1996 page->next = oldpage;
1998 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2002 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2004 stat(s, CPUSLAB_FLUSH);
2005 deactivate_slab(s, c->page, c->freelist);
2007 c->tid = next_tid(c->tid);
2015 * Called from IPI handler with interrupts disabled.
2017 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2019 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2025 unfreeze_partials(s);
2029 static void flush_cpu_slab(void *d)
2031 struct kmem_cache *s = d;
2033 __flush_cpu_slab(s, smp_processor_id());
2036 static bool has_cpu_slab(int cpu, void *info)
2038 struct kmem_cache *s = info;
2039 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2041 return c->page || c->partial;
2044 static void flush_all(struct kmem_cache *s)
2046 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2050 * Check if the objects in a per cpu structure fit numa
2051 * locality expectations.
2053 static inline int node_match(struct page *page, int node)
2056 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2062 static int count_free(struct page *page)
2064 return page->objects - page->inuse;
2067 static unsigned long count_partial(struct kmem_cache_node *n,
2068 int (*get_count)(struct page *))
2070 unsigned long flags;
2071 unsigned long x = 0;
2074 spin_lock_irqsave(&n->list_lock, flags);
2075 list_for_each_entry(page, &n->partial, lru)
2076 x += get_count(page);
2077 spin_unlock_irqrestore(&n->list_lock, flags);
2081 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2083 #ifdef CONFIG_SLUB_DEBUG
2084 return atomic_long_read(&n->total_objects);
2090 static noinline void
2091 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2096 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2098 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2099 "default order: %d, min order: %d\n", s->name, s->objsize,
2100 s->size, oo_order(s->oo), oo_order(s->min));
2102 if (oo_order(s->min) > get_order(s->objsize))
2103 printk(KERN_WARNING " %s debugging increased min order, use "
2104 "slub_debug=O to disable.\n", s->name);
2106 for_each_online_node(node) {
2107 struct kmem_cache_node *n = get_node(s, node);
2108 unsigned long nr_slabs;
2109 unsigned long nr_objs;
2110 unsigned long nr_free;
2115 nr_free = count_partial(n, count_free);
2116 nr_slabs = node_nr_slabs(n);
2117 nr_objs = node_nr_objs(n);
2120 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2121 node, nr_slabs, nr_objs, nr_free);
2125 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2126 int node, struct kmem_cache_cpu **pc)
2129 struct kmem_cache_cpu *c = *pc;
2132 freelist = get_partial(s, flags, node, c);
2137 page = new_slab(s, flags, node);
2139 c = __this_cpu_ptr(s->cpu_slab);
2144 * No other reference to the page yet so we can
2145 * muck around with it freely without cmpxchg
2147 freelist = page->freelist;
2148 page->freelist = NULL;
2150 stat(s, ALLOC_SLAB);
2160 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2161 * or deactivate the page.
2163 * The page is still frozen if the return value is not NULL.
2165 * If this function returns NULL then the page has been unfrozen.
2167 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2170 unsigned long counters;
2174 freelist = page->freelist;
2175 counters = page->counters;
2177 new.counters = counters;
2178 VM_BUG_ON(!new.frozen);
2180 new.inuse = page->objects;
2181 new.frozen = freelist != NULL;
2183 } while (!cmpxchg_double_slab(s, page,
2192 * Slow path. The lockless freelist is empty or we need to perform
2195 * Processing is still very fast if new objects have been freed to the
2196 * regular freelist. In that case we simply take over the regular freelist
2197 * as the lockless freelist and zap the regular freelist.
2199 * If that is not working then we fall back to the partial lists. We take the
2200 * first element of the freelist as the object to allocate now and move the
2201 * rest of the freelist to the lockless freelist.
2203 * And if we were unable to get a new slab from the partial slab lists then
2204 * we need to allocate a new slab. This is the slowest path since it involves
2205 * a call to the page allocator and the setup of a new slab.
2207 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2208 unsigned long addr, struct kmem_cache_cpu *c)
2212 unsigned long flags;
2214 local_irq_save(flags);
2215 #ifdef CONFIG_PREEMPT
2217 * We may have been preempted and rescheduled on a different
2218 * cpu before disabling interrupts. Need to reload cpu area
2221 c = this_cpu_ptr(s->cpu_slab);
2229 if (unlikely(!node_match(page, node))) {
2230 stat(s, ALLOC_NODE_MISMATCH);
2231 deactivate_slab(s, page, c->freelist);
2237 /* must check again c->freelist in case of cpu migration or IRQ */
2238 freelist = c->freelist;
2242 stat(s, ALLOC_SLOWPATH);
2244 freelist = get_freelist(s, page);
2248 stat(s, DEACTIVATE_BYPASS);
2252 stat(s, ALLOC_REFILL);
2256 * freelist is pointing to the list of objects to be used.
2257 * page is pointing to the page from which the objects are obtained.
2258 * That page must be frozen for per cpu allocations to work.
2260 VM_BUG_ON(!c->page->frozen);
2261 c->freelist = get_freepointer(s, freelist);
2262 c->tid = next_tid(c->tid);
2263 local_irq_restore(flags);
2269 page = c->page = c->partial;
2270 c->partial = page->next;
2271 stat(s, CPU_PARTIAL_ALLOC);
2276 freelist = new_slab_objects(s, gfpflags, node, &c);
2278 if (unlikely(!freelist)) {
2279 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2280 slab_out_of_memory(s, gfpflags, node);
2282 local_irq_restore(flags);
2287 if (likely(!kmem_cache_debug(s)))
2290 /* Only entered in the debug case */
2291 if (!alloc_debug_processing(s, page, freelist, addr))
2292 goto new_slab; /* Slab failed checks. Next slab needed */
2294 deactivate_slab(s, page, get_freepointer(s, freelist));
2297 local_irq_restore(flags);
2302 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2303 * have the fastpath folded into their functions. So no function call
2304 * overhead for requests that can be satisfied on the fastpath.
2306 * The fastpath works by first checking if the lockless freelist can be used.
2307 * If not then __slab_alloc is called for slow processing.
2309 * Otherwise we can simply pick the next object from the lockless free list.
2311 static __always_inline void *slab_alloc(struct kmem_cache *s,
2312 gfp_t gfpflags, int node, unsigned long addr)
2315 struct kmem_cache_cpu *c;
2319 if (slab_pre_alloc_hook(s, gfpflags))
2325 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2326 * enabled. We may switch back and forth between cpus while
2327 * reading from one cpu area. That does not matter as long
2328 * as we end up on the original cpu again when doing the cmpxchg.
2330 c = __this_cpu_ptr(s->cpu_slab);
2333 * The transaction ids are globally unique per cpu and per operation on
2334 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2335 * occurs on the right processor and that there was no operation on the
2336 * linked list in between.
2341 object = c->freelist;
2343 if (unlikely(!object || !node_match(page, node)))
2345 object = __slab_alloc(s, gfpflags, node, addr, c);
2348 void *next_object = get_freepointer_safe(s, object);
2351 * The cmpxchg will only match if there was no additional
2352 * operation and if we are on the right processor.
2354 * The cmpxchg does the following atomically (without lock semantics!)
2355 * 1. Relocate first pointer to the current per cpu area.
2356 * 2. Verify that tid and freelist have not been changed
2357 * 3. If they were not changed replace tid and freelist
2359 * Since this is without lock semantics the protection is only against
2360 * code executing on this cpu *not* from access by other cpus.
2362 if (unlikely(!this_cpu_cmpxchg_double(
2363 s->cpu_slab->freelist, s->cpu_slab->tid,
2365 next_object, next_tid(tid)))) {
2367 note_cmpxchg_failure("slab_alloc", s, tid);
2370 prefetch_freepointer(s, next_object);
2371 stat(s, ALLOC_FASTPATH);
2374 if (unlikely(gfpflags & __GFP_ZERO) && object)
2375 memset(object, 0, s->objsize);
2377 slab_post_alloc_hook(s, gfpflags, object);
2382 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2384 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2386 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2390 EXPORT_SYMBOL(kmem_cache_alloc);
2392 #ifdef CONFIG_TRACING
2393 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2395 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2396 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2399 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2401 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2403 void *ret = kmalloc_order(size, flags, order);
2404 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2407 EXPORT_SYMBOL(kmalloc_order_trace);
2411 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2413 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2415 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2416 s->objsize, s->size, gfpflags, node);
2420 EXPORT_SYMBOL(kmem_cache_alloc_node);
2422 #ifdef CONFIG_TRACING
2423 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2425 int node, size_t size)
2427 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2429 trace_kmalloc_node(_RET_IP_, ret,
2430 size, s->size, gfpflags, node);
2433 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2438 * Slow patch handling. This may still be called frequently since objects
2439 * have a longer lifetime than the cpu slabs in most processing loads.
2441 * So we still attempt to reduce cache line usage. Just take the slab
2442 * lock and free the item. If there is no additional partial page
2443 * handling required then we can return immediately.
2445 static void __slab_free(struct kmem_cache *s, struct page *page,
2446 void *x, unsigned long addr)
2449 void **object = (void *)x;
2453 unsigned long counters;
2454 struct kmem_cache_node *n = NULL;
2455 unsigned long uninitialized_var(flags);
2457 stat(s, FREE_SLOWPATH);
2459 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2463 prior = page->freelist;
2464 counters = page->counters;
2465 set_freepointer(s, object, prior);
2466 new.counters = counters;
2467 was_frozen = new.frozen;
2469 if ((!new.inuse || !prior) && !was_frozen && !n) {
2471 if (!kmem_cache_debug(s) && !prior)
2474 * Slab was on no list before and will be partially empty
2475 * We can defer the list move and instead freeze it.
2479 else { /* Needs to be taken off a list */
2481 n = get_node(s, page_to_nid(page));
2483 * Speculatively acquire the list_lock.
2484 * If the cmpxchg does not succeed then we may
2485 * drop the list_lock without any processing.
2487 * Otherwise the list_lock will synchronize with
2488 * other processors updating the list of slabs.
2490 spin_lock_irqsave(&n->list_lock, flags);
2496 } while (!cmpxchg_double_slab(s, page,
2498 object, new.counters,
2504 * If we just froze the page then put it onto the
2505 * per cpu partial list.
2507 if (new.frozen && !was_frozen) {
2508 put_cpu_partial(s, page, 1);
2509 stat(s, CPU_PARTIAL_FREE);
2512 * The list lock was not taken therefore no list
2513 * activity can be necessary.
2516 stat(s, FREE_FROZEN);
2521 * was_frozen may have been set after we acquired the list_lock in
2522 * an earlier loop. So we need to check it here again.
2525 stat(s, FREE_FROZEN);
2527 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2531 * Objects left in the slab. If it was not on the partial list before
2534 if (unlikely(!prior)) {
2535 remove_full(s, page);
2536 add_partial(n, page, DEACTIVATE_TO_TAIL);
2537 stat(s, FREE_ADD_PARTIAL);
2540 spin_unlock_irqrestore(&n->list_lock, flags);
2546 * Slab on the partial list.
2548 remove_partial(n, page);
2549 stat(s, FREE_REMOVE_PARTIAL);
2551 /* Slab must be on the full list */
2552 remove_full(s, page);
2554 spin_unlock_irqrestore(&n->list_lock, flags);
2556 discard_slab(s, page);
2560 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2561 * can perform fastpath freeing without additional function calls.
2563 * The fastpath is only possible if we are freeing to the current cpu slab
2564 * of this processor. This typically the case if we have just allocated
2567 * If fastpath is not possible then fall back to __slab_free where we deal
2568 * with all sorts of special processing.
2570 static __always_inline void slab_free(struct kmem_cache *s,
2571 struct page *page, void *x, unsigned long addr)
2573 void **object = (void *)x;
2574 struct kmem_cache_cpu *c;
2577 slab_free_hook(s, x);
2581 * Determine the currently cpus per cpu slab.
2582 * The cpu may change afterward. However that does not matter since
2583 * data is retrieved via this pointer. If we are on the same cpu
2584 * during the cmpxchg then the free will succedd.
2586 c = __this_cpu_ptr(s->cpu_slab);
2591 if (likely(page == c->page)) {
2592 set_freepointer(s, object, c->freelist);
2594 if (unlikely(!this_cpu_cmpxchg_double(
2595 s->cpu_slab->freelist, s->cpu_slab->tid,
2597 object, next_tid(tid)))) {
2599 note_cmpxchg_failure("slab_free", s, tid);
2602 stat(s, FREE_FASTPATH);
2604 __slab_free(s, page, x, addr);
2608 void kmem_cache_free(struct kmem_cache *s, void *x)
2612 page = virt_to_head_page(x);
2614 slab_free(s, page, x, _RET_IP_);
2616 trace_kmem_cache_free(_RET_IP_, x);
2618 EXPORT_SYMBOL(kmem_cache_free);
2621 * Object placement in a slab is made very easy because we always start at
2622 * offset 0. If we tune the size of the object to the alignment then we can
2623 * get the required alignment by putting one properly sized object after
2626 * Notice that the allocation order determines the sizes of the per cpu
2627 * caches. Each processor has always one slab available for allocations.
2628 * Increasing the allocation order reduces the number of times that slabs
2629 * must be moved on and off the partial lists and is therefore a factor in
2634 * Mininum / Maximum order of slab pages. This influences locking overhead
2635 * and slab fragmentation. A higher order reduces the number of partial slabs
2636 * and increases the number of allocations possible without having to
2637 * take the list_lock.
2639 static int slub_min_order;
2640 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2641 static int slub_min_objects;
2644 * Merge control. If this is set then no merging of slab caches will occur.
2645 * (Could be removed. This was introduced to pacify the merge skeptics.)
2647 static int slub_nomerge;
2650 * Calculate the order of allocation given an slab object size.
2652 * The order of allocation has significant impact on performance and other
2653 * system components. Generally order 0 allocations should be preferred since
2654 * order 0 does not cause fragmentation in the page allocator. Larger objects
2655 * be problematic to put into order 0 slabs because there may be too much
2656 * unused space left. We go to a higher order if more than 1/16th of the slab
2659 * In order to reach satisfactory performance we must ensure that a minimum
2660 * number of objects is in one slab. Otherwise we may generate too much
2661 * activity on the partial lists which requires taking the list_lock. This is
2662 * less a concern for large slabs though which are rarely used.
2664 * slub_max_order specifies the order where we begin to stop considering the
2665 * number of objects in a slab as critical. If we reach slub_max_order then
2666 * we try to keep the page order as low as possible. So we accept more waste
2667 * of space in favor of a small page order.
2669 * Higher order allocations also allow the placement of more objects in a
2670 * slab and thereby reduce object handling overhead. If the user has
2671 * requested a higher mininum order then we start with that one instead of
2672 * the smallest order which will fit the object.
2674 static inline int slab_order(int size, int min_objects,
2675 int max_order, int fract_leftover, int reserved)
2679 int min_order = slub_min_order;
2681 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2682 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2684 for (order = max(min_order,
2685 fls(min_objects * size - 1) - PAGE_SHIFT);
2686 order <= max_order; order++) {
2688 unsigned long slab_size = PAGE_SIZE << order;
2690 if (slab_size < min_objects * size + reserved)
2693 rem = (slab_size - reserved) % size;
2695 if (rem <= slab_size / fract_leftover)
2703 static inline int calculate_order(int size, int reserved)
2711 * Attempt to find best configuration for a slab. This
2712 * works by first attempting to generate a layout with
2713 * the best configuration and backing off gradually.
2715 * First we reduce the acceptable waste in a slab. Then
2716 * we reduce the minimum objects required in a slab.
2718 min_objects = slub_min_objects;
2720 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2721 max_objects = order_objects(slub_max_order, size, reserved);
2722 min_objects = min(min_objects, max_objects);
2724 while (min_objects > 1) {
2726 while (fraction >= 4) {
2727 order = slab_order(size, min_objects,
2728 slub_max_order, fraction, reserved);
2729 if (order <= slub_max_order)
2737 * We were unable to place multiple objects in a slab. Now
2738 * lets see if we can place a single object there.
2740 order = slab_order(size, 1, slub_max_order, 1, reserved);
2741 if (order <= slub_max_order)
2745 * Doh this slab cannot be placed using slub_max_order.
2747 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2748 if (order < MAX_ORDER)
2754 * Figure out what the alignment of the objects will be.
2756 static unsigned long calculate_alignment(unsigned long flags,
2757 unsigned long align, unsigned long size)
2760 * If the user wants hardware cache aligned objects then follow that
2761 * suggestion if the object is sufficiently large.
2763 * The hardware cache alignment cannot override the specified
2764 * alignment though. If that is greater then use it.
2766 if (flags & SLAB_HWCACHE_ALIGN) {
2767 unsigned long ralign = cache_line_size();
2768 while (size <= ralign / 2)
2770 align = max(align, ralign);
2773 if (align < ARCH_SLAB_MINALIGN)
2774 align = ARCH_SLAB_MINALIGN;
2776 return ALIGN(align, sizeof(void *));
2780 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2783 spin_lock_init(&n->list_lock);
2784 INIT_LIST_HEAD(&n->partial);
2785 #ifdef CONFIG_SLUB_DEBUG
2786 atomic_long_set(&n->nr_slabs, 0);
2787 atomic_long_set(&n->total_objects, 0);
2788 INIT_LIST_HEAD(&n->full);
2792 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2794 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2795 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2798 * Must align to double word boundary for the double cmpxchg
2799 * instructions to work; see __pcpu_double_call_return_bool().
2801 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2802 2 * sizeof(void *));
2807 init_kmem_cache_cpus(s);
2812 static struct kmem_cache *kmem_cache_node;
2815 * No kmalloc_node yet so do it by hand. We know that this is the first
2816 * slab on the node for this slabcache. There are no concurrent accesses
2819 * Note that this function only works on the kmalloc_node_cache
2820 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2821 * memory on a fresh node that has no slab structures yet.
2823 static void early_kmem_cache_node_alloc(int node)
2826 struct kmem_cache_node *n;
2828 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2830 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2833 if (page_to_nid(page) != node) {
2834 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2836 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2837 "in order to be able to continue\n");
2842 page->freelist = get_freepointer(kmem_cache_node, n);
2845 kmem_cache_node->node[node] = n;
2846 #ifdef CONFIG_SLUB_DEBUG
2847 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2848 init_tracking(kmem_cache_node, n);
2850 init_kmem_cache_node(n, kmem_cache_node);
2851 inc_slabs_node(kmem_cache_node, node, page->objects);
2853 add_partial(n, page, DEACTIVATE_TO_HEAD);
2856 static void free_kmem_cache_nodes(struct kmem_cache *s)
2860 for_each_node_state(node, N_NORMAL_MEMORY) {
2861 struct kmem_cache_node *n = s->node[node];
2864 kmem_cache_free(kmem_cache_node, n);
2866 s->node[node] = NULL;
2870 static int init_kmem_cache_nodes(struct kmem_cache *s)
2874 for_each_node_state(node, N_NORMAL_MEMORY) {
2875 struct kmem_cache_node *n;
2877 if (slab_state == DOWN) {
2878 early_kmem_cache_node_alloc(node);
2881 n = kmem_cache_alloc_node(kmem_cache_node,
2885 free_kmem_cache_nodes(s);
2890 init_kmem_cache_node(n, s);
2895 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2897 if (min < MIN_PARTIAL)
2899 else if (min > MAX_PARTIAL)
2901 s->min_partial = min;
2905 * calculate_sizes() determines the order and the distribution of data within
2908 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2910 unsigned long flags = s->flags;
2911 unsigned long size = s->objsize;
2912 unsigned long align = s->align;
2916 * Round up object size to the next word boundary. We can only
2917 * place the free pointer at word boundaries and this determines
2918 * the possible location of the free pointer.
2920 size = ALIGN(size, sizeof(void *));
2922 #ifdef CONFIG_SLUB_DEBUG
2924 * Determine if we can poison the object itself. If the user of
2925 * the slab may touch the object after free or before allocation
2926 * then we should never poison the object itself.
2928 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2930 s->flags |= __OBJECT_POISON;
2932 s->flags &= ~__OBJECT_POISON;
2936 * If we are Redzoning then check if there is some space between the
2937 * end of the object and the free pointer. If not then add an
2938 * additional word to have some bytes to store Redzone information.
2940 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2941 size += sizeof(void *);
2945 * With that we have determined the number of bytes in actual use
2946 * by the object. This is the potential offset to the free pointer.
2950 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2953 * Relocate free pointer after the object if it is not
2954 * permitted to overwrite the first word of the object on
2957 * This is the case if we do RCU, have a constructor or
2958 * destructor or are poisoning the objects.
2961 size += sizeof(void *);
2964 #ifdef CONFIG_SLUB_DEBUG
2965 if (flags & SLAB_STORE_USER)
2967 * Need to store information about allocs and frees after
2970 size += 2 * sizeof(struct track);
2972 if (flags & SLAB_RED_ZONE)
2974 * Add some empty padding so that we can catch
2975 * overwrites from earlier objects rather than let
2976 * tracking information or the free pointer be
2977 * corrupted if a user writes before the start
2980 size += sizeof(void *);
2984 * Determine the alignment based on various parameters that the
2985 * user specified and the dynamic determination of cache line size
2988 align = calculate_alignment(flags, align, s->objsize);
2992 * SLUB stores one object immediately after another beginning from
2993 * offset 0. In order to align the objects we have to simply size
2994 * each object to conform to the alignment.
2996 size = ALIGN(size, align);
2998 if (forced_order >= 0)
2999 order = forced_order;
3001 order = calculate_order(size, s->reserved);
3008 s->allocflags |= __GFP_COMP;
3010 if (s->flags & SLAB_CACHE_DMA)
3011 s->allocflags |= SLUB_DMA;
3013 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3014 s->allocflags |= __GFP_RECLAIMABLE;
3017 * Determine the number of objects per slab
3019 s->oo = oo_make(order, size, s->reserved);
3020 s->min = oo_make(get_order(size), size, s->reserved);
3021 if (oo_objects(s->oo) > oo_objects(s->max))
3024 return !!oo_objects(s->oo);
3028 static int kmem_cache_open(struct kmem_cache *s,
3029 const char *name, size_t size,
3030 size_t align, unsigned long flags,
3031 void (*ctor)(void *))
3033 memset(s, 0, kmem_size);
3038 s->flags = kmem_cache_flags(size, flags, name, ctor);
3041 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3042 s->reserved = sizeof(struct rcu_head);
3044 if (!calculate_sizes(s, -1))
3046 if (disable_higher_order_debug) {
3048 * Disable debugging flags that store metadata if the min slab
3051 if (get_order(s->size) > get_order(s->objsize)) {
3052 s->flags &= ~DEBUG_METADATA_FLAGS;
3054 if (!calculate_sizes(s, -1))
3059 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3060 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3061 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3062 /* Enable fast mode */
3063 s->flags |= __CMPXCHG_DOUBLE;
3067 * The larger the object size is, the more pages we want on the partial
3068 * list to avoid pounding the page allocator excessively.
3070 set_min_partial(s, ilog2(s->size) / 2);
3073 * cpu_partial determined the maximum number of objects kept in the
3074 * per cpu partial lists of a processor.
3076 * Per cpu partial lists mainly contain slabs that just have one
3077 * object freed. If they are used for allocation then they can be
3078 * filled up again with minimal effort. The slab will never hit the
3079 * per node partial lists and therefore no locking will be required.
3081 * This setting also determines
3083 * A) The number of objects from per cpu partial slabs dumped to the
3084 * per node list when we reach the limit.
3085 * B) The number of objects in cpu partial slabs to extract from the
3086 * per node list when we run out of per cpu objects. We only fetch 50%
3087 * to keep some capacity around for frees.
3089 if (kmem_cache_debug(s))
3091 else if (s->size >= PAGE_SIZE)
3093 else if (s->size >= 1024)
3095 else if (s->size >= 256)
3096 s->cpu_partial = 13;
3098 s->cpu_partial = 30;
3102 s->remote_node_defrag_ratio = 1000;
3104 if (!init_kmem_cache_nodes(s))
3107 if (alloc_kmem_cache_cpus(s))
3110 free_kmem_cache_nodes(s);
3112 if (flags & SLAB_PANIC)
3113 panic("Cannot create slab %s size=%lu realsize=%u "
3114 "order=%u offset=%u flags=%lx\n",
3115 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3121 * Determine the size of a slab object
3123 unsigned int kmem_cache_size(struct kmem_cache *s)
3127 EXPORT_SYMBOL(kmem_cache_size);
3129 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3132 #ifdef CONFIG_SLUB_DEBUG
3133 void *addr = page_address(page);
3135 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3136 sizeof(long), GFP_ATOMIC);
3139 slab_err(s, page, "%s", text);
3142 get_map(s, page, map);
3143 for_each_object(p, s, addr, page->objects) {
3145 if (!test_bit(slab_index(p, s, addr), map)) {
3146 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3148 print_tracking(s, p);
3157 * Attempt to free all partial slabs on a node.
3158 * This is called from kmem_cache_close(). We must be the last thread
3159 * using the cache and therefore we do not need to lock anymore.
3161 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3163 struct page *page, *h;
3165 list_for_each_entry_safe(page, h, &n->partial, lru) {
3167 remove_partial(n, page);
3168 discard_slab(s, page);
3170 list_slab_objects(s, page,
3171 "Objects remaining on kmem_cache_close()");
3177 * Release all resources used by a slab cache.
3179 static inline int kmem_cache_close(struct kmem_cache *s)
3184 free_percpu(s->cpu_slab);
3185 /* Attempt to free all objects */
3186 for_each_node_state(node, N_NORMAL_MEMORY) {
3187 struct kmem_cache_node *n = get_node(s, node);
3190 if (n->nr_partial || slabs_node(s, node))
3193 free_kmem_cache_nodes(s);
3198 * Close a cache and release the kmem_cache structure
3199 * (must be used for caches created using kmem_cache_create)
3201 void kmem_cache_destroy(struct kmem_cache *s)
3203 down_write(&slub_lock);
3207 up_write(&slub_lock);
3208 if (kmem_cache_close(s)) {
3209 printk(KERN_ERR "SLUB %s: %s called for cache that "
3210 "still has objects.\n", s->name, __func__);
3213 if (s->flags & SLAB_DESTROY_BY_RCU)
3215 sysfs_slab_remove(s);
3217 up_write(&slub_lock);
3219 EXPORT_SYMBOL(kmem_cache_destroy);
3221 /********************************************************************
3223 *******************************************************************/
3225 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3226 EXPORT_SYMBOL(kmalloc_caches);
3228 static struct kmem_cache *kmem_cache;
3230 #ifdef CONFIG_ZONE_DMA
3231 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3234 static int __init setup_slub_min_order(char *str)
3236 get_option(&str, &slub_min_order);
3241 __setup("slub_min_order=", setup_slub_min_order);
3243 static int __init setup_slub_max_order(char *str)
3245 get_option(&str, &slub_max_order);
3246 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3251 __setup("slub_max_order=", setup_slub_max_order);
3253 static int __init setup_slub_min_objects(char *str)
3255 get_option(&str, &slub_min_objects);
3260 __setup("slub_min_objects=", setup_slub_min_objects);
3262 static int __init setup_slub_nomerge(char *str)
3268 __setup("slub_nomerge", setup_slub_nomerge);
3270 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3271 int size, unsigned int flags)
3273 struct kmem_cache *s;
3275 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3278 * This function is called with IRQs disabled during early-boot on
3279 * single CPU so there's no need to take slub_lock here.
3281 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3285 list_add(&s->list, &slab_caches);
3289 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3294 * Conversion table for small slabs sizes / 8 to the index in the
3295 * kmalloc array. This is necessary for slabs < 192 since we have non power
3296 * of two cache sizes there. The size of larger slabs can be determined using
3299 static s8 size_index[24] = {
3326 static inline int size_index_elem(size_t bytes)
3328 return (bytes - 1) / 8;
3331 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3337 return ZERO_SIZE_PTR;
3339 index = size_index[size_index_elem(size)];
3341 index = fls(size - 1);
3343 #ifdef CONFIG_ZONE_DMA
3344 if (unlikely((flags & SLUB_DMA)))
3345 return kmalloc_dma_caches[index];
3348 return kmalloc_caches[index];
3351 void *__kmalloc(size_t size, gfp_t flags)
3353 struct kmem_cache *s;
3356 if (unlikely(size > SLUB_MAX_SIZE))
3357 return kmalloc_large(size, flags);
3359 s = get_slab(size, flags);
3361 if (unlikely(ZERO_OR_NULL_PTR(s)))
3364 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3366 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3370 EXPORT_SYMBOL(__kmalloc);
3373 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3378 flags |= __GFP_COMP | __GFP_NOTRACK;
3379 page = alloc_pages_node(node, flags, get_order(size));
3381 ptr = page_address(page);
3383 kmemleak_alloc(ptr, size, 1, flags);
3387 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3389 struct kmem_cache *s;
3392 if (unlikely(size > SLUB_MAX_SIZE)) {
3393 ret = kmalloc_large_node(size, flags, node);
3395 trace_kmalloc_node(_RET_IP_, ret,
3396 size, PAGE_SIZE << get_order(size),
3402 s = get_slab(size, flags);
3404 if (unlikely(ZERO_OR_NULL_PTR(s)))
3407 ret = slab_alloc(s, flags, node, _RET_IP_);
3409 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3413 EXPORT_SYMBOL(__kmalloc_node);
3416 size_t ksize(const void *object)
3420 if (unlikely(object == ZERO_SIZE_PTR))
3423 page = virt_to_head_page(object);
3425 if (unlikely(!PageSlab(page))) {
3426 WARN_ON(!PageCompound(page));
3427 return PAGE_SIZE << compound_order(page);
3430 return slab_ksize(page->slab);
3432 EXPORT_SYMBOL(ksize);
3434 #ifdef CONFIG_SLUB_DEBUG
3435 bool verify_mem_not_deleted(const void *x)
3438 void *object = (void *)x;
3439 unsigned long flags;
3442 if (unlikely(ZERO_OR_NULL_PTR(x)))
3445 local_irq_save(flags);
3447 page = virt_to_head_page(x);
3448 if (unlikely(!PageSlab(page))) {
3449 /* maybe it was from stack? */
3455 if (on_freelist(page->slab, page, object)) {
3456 object_err(page->slab, page, object, "Object is on free-list");
3464 local_irq_restore(flags);
3467 EXPORT_SYMBOL(verify_mem_not_deleted);
3470 void kfree(const void *x)
3473 void *object = (void *)x;
3475 trace_kfree(_RET_IP_, x);
3477 if (unlikely(ZERO_OR_NULL_PTR(x)))
3480 page = virt_to_head_page(x);
3481 if (unlikely(!PageSlab(page))) {
3482 BUG_ON(!PageCompound(page));
3487 slab_free(page->slab, page, object, _RET_IP_);
3489 EXPORT_SYMBOL(kfree);
3492 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3493 * the remaining slabs by the number of items in use. The slabs with the
3494 * most items in use come first. New allocations will then fill those up
3495 * and thus they can be removed from the partial lists.
3497 * The slabs with the least items are placed last. This results in them
3498 * being allocated from last increasing the chance that the last objects
3499 * are freed in them.
3501 int kmem_cache_shrink(struct kmem_cache *s)
3505 struct kmem_cache_node *n;
3508 int objects = oo_objects(s->max);
3509 struct list_head *slabs_by_inuse =
3510 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3511 unsigned long flags;
3513 if (!slabs_by_inuse)
3517 for_each_node_state(node, N_NORMAL_MEMORY) {
3518 n = get_node(s, node);
3523 for (i = 0; i < objects; i++)
3524 INIT_LIST_HEAD(slabs_by_inuse + i);
3526 spin_lock_irqsave(&n->list_lock, flags);
3529 * Build lists indexed by the items in use in each slab.
3531 * Note that concurrent frees may occur while we hold the
3532 * list_lock. page->inuse here is the upper limit.
3534 list_for_each_entry_safe(page, t, &n->partial, lru) {
3535 list_move(&page->lru, slabs_by_inuse + page->inuse);
3541 * Rebuild the partial list with the slabs filled up most
3542 * first and the least used slabs at the end.
3544 for (i = objects - 1; i > 0; i--)
3545 list_splice(slabs_by_inuse + i, n->partial.prev);
3547 spin_unlock_irqrestore(&n->list_lock, flags);
3549 /* Release empty slabs */
3550 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3551 discard_slab(s, page);
3554 kfree(slabs_by_inuse);
3557 EXPORT_SYMBOL(kmem_cache_shrink);
3559 #if defined(CONFIG_MEMORY_HOTPLUG)
3560 static int slab_mem_going_offline_callback(void *arg)
3562 struct kmem_cache *s;
3564 down_read(&slub_lock);
3565 list_for_each_entry(s, &slab_caches, list)
3566 kmem_cache_shrink(s);
3567 up_read(&slub_lock);
3572 static void slab_mem_offline_callback(void *arg)
3574 struct kmem_cache_node *n;
3575 struct kmem_cache *s;
3576 struct memory_notify *marg = arg;
3579 offline_node = marg->status_change_nid;
3582 * If the node still has available memory. we need kmem_cache_node
3585 if (offline_node < 0)
3588 down_read(&slub_lock);
3589 list_for_each_entry(s, &slab_caches, list) {
3590 n = get_node(s, offline_node);
3593 * if n->nr_slabs > 0, slabs still exist on the node
3594 * that is going down. We were unable to free them,
3595 * and offline_pages() function shouldn't call this
3596 * callback. So, we must fail.
3598 BUG_ON(slabs_node(s, offline_node));
3600 s->node[offline_node] = NULL;
3601 kmem_cache_free(kmem_cache_node, n);
3604 up_read(&slub_lock);
3607 static int slab_mem_going_online_callback(void *arg)
3609 struct kmem_cache_node *n;
3610 struct kmem_cache *s;
3611 struct memory_notify *marg = arg;
3612 int nid = marg->status_change_nid;
3616 * If the node's memory is already available, then kmem_cache_node is
3617 * already created. Nothing to do.
3623 * We are bringing a node online. No memory is available yet. We must
3624 * allocate a kmem_cache_node structure in order to bring the node
3627 down_read(&slub_lock);
3628 list_for_each_entry(s, &slab_caches, list) {
3630 * XXX: kmem_cache_alloc_node will fallback to other nodes
3631 * since memory is not yet available from the node that
3634 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3639 init_kmem_cache_node(n, s);
3643 up_read(&slub_lock);
3647 static int slab_memory_callback(struct notifier_block *self,
3648 unsigned long action, void *arg)
3653 case MEM_GOING_ONLINE:
3654 ret = slab_mem_going_online_callback(arg);
3656 case MEM_GOING_OFFLINE:
3657 ret = slab_mem_going_offline_callback(arg);
3660 case MEM_CANCEL_ONLINE:
3661 slab_mem_offline_callback(arg);
3664 case MEM_CANCEL_OFFLINE:
3668 ret = notifier_from_errno(ret);
3674 #endif /* CONFIG_MEMORY_HOTPLUG */
3676 /********************************************************************
3677 * Basic setup of slabs
3678 *******************************************************************/
3681 * Used for early kmem_cache structures that were allocated using
3682 * the page allocator
3685 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3689 list_add(&s->list, &slab_caches);
3692 for_each_node_state(node, N_NORMAL_MEMORY) {
3693 struct kmem_cache_node *n = get_node(s, node);
3697 list_for_each_entry(p, &n->partial, lru)
3700 #ifdef CONFIG_SLUB_DEBUG
3701 list_for_each_entry(p, &n->full, lru)
3708 void __init kmem_cache_init(void)
3712 struct kmem_cache *temp_kmem_cache;
3714 struct kmem_cache *temp_kmem_cache_node;
3715 unsigned long kmalloc_size;
3717 if (debug_guardpage_minorder())
3720 kmem_size = offsetof(struct kmem_cache, node) +
3721 nr_node_ids * sizeof(struct kmem_cache_node *);
3723 /* Allocate two kmem_caches from the page allocator */
3724 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3725 order = get_order(2 * kmalloc_size);
3726 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3729 * Must first have the slab cache available for the allocations of the
3730 * struct kmem_cache_node's. There is special bootstrap code in
3731 * kmem_cache_open for slab_state == DOWN.
3733 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3735 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3736 sizeof(struct kmem_cache_node),
3737 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3739 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3741 /* Able to allocate the per node structures */
3742 slab_state = PARTIAL;
3744 temp_kmem_cache = kmem_cache;
3745 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3746 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3747 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3748 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3751 * Allocate kmem_cache_node properly from the kmem_cache slab.
3752 * kmem_cache_node is separately allocated so no need to
3753 * update any list pointers.
3755 temp_kmem_cache_node = kmem_cache_node;
3757 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3758 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3760 kmem_cache_bootstrap_fixup(kmem_cache_node);
3763 kmem_cache_bootstrap_fixup(kmem_cache);
3765 /* Free temporary boot structure */
3766 free_pages((unsigned long)temp_kmem_cache, order);
3768 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3771 * Patch up the size_index table if we have strange large alignment
3772 * requirements for the kmalloc array. This is only the case for
3773 * MIPS it seems. The standard arches will not generate any code here.
3775 * Largest permitted alignment is 256 bytes due to the way we
3776 * handle the index determination for the smaller caches.
3778 * Make sure that nothing crazy happens if someone starts tinkering
3779 * around with ARCH_KMALLOC_MINALIGN
3781 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3782 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3784 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3785 int elem = size_index_elem(i);
3786 if (elem >= ARRAY_SIZE(size_index))
3788 size_index[elem] = KMALLOC_SHIFT_LOW;
3791 if (KMALLOC_MIN_SIZE == 64) {
3793 * The 96 byte size cache is not used if the alignment
3796 for (i = 64 + 8; i <= 96; i += 8)
3797 size_index[size_index_elem(i)] = 7;
3798 } else if (KMALLOC_MIN_SIZE == 128) {
3800 * The 192 byte sized cache is not used if the alignment
3801 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3804 for (i = 128 + 8; i <= 192; i += 8)
3805 size_index[size_index_elem(i)] = 8;
3808 /* Caches that are not of the two-to-the-power-of size */
3809 if (KMALLOC_MIN_SIZE <= 32) {
3810 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3814 if (KMALLOC_MIN_SIZE <= 64) {
3815 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3819 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3820 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3826 /* Provide the correct kmalloc names now that the caches are up */
3827 if (KMALLOC_MIN_SIZE <= 32) {
3828 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3829 BUG_ON(!kmalloc_caches[1]->name);
3832 if (KMALLOC_MIN_SIZE <= 64) {
3833 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3834 BUG_ON(!kmalloc_caches[2]->name);
3837 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3838 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3841 kmalloc_caches[i]->name = s;
3845 register_cpu_notifier(&slab_notifier);
3848 #ifdef CONFIG_ZONE_DMA
3849 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3850 struct kmem_cache *s = kmalloc_caches[i];
3853 char *name = kasprintf(GFP_NOWAIT,
3854 "dma-kmalloc-%d", s->objsize);
3857 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3858 s->objsize, SLAB_CACHE_DMA);
3863 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3864 " CPUs=%d, Nodes=%d\n",
3865 caches, cache_line_size(),
3866 slub_min_order, slub_max_order, slub_min_objects,
3867 nr_cpu_ids, nr_node_ids);
3870 void __init kmem_cache_init_late(void)
3875 * Find a mergeable slab cache
3877 static int slab_unmergeable(struct kmem_cache *s)
3879 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3886 * We may have set a slab to be unmergeable during bootstrap.
3888 if (s->refcount < 0)
3894 static struct kmem_cache *find_mergeable(size_t size,
3895 size_t align, unsigned long flags, const char *name,
3896 void (*ctor)(void *))
3898 struct kmem_cache *s;
3900 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3906 size = ALIGN(size, sizeof(void *));
3907 align = calculate_alignment(flags, align, size);
3908 size = ALIGN(size, align);
3909 flags = kmem_cache_flags(size, flags, name, NULL);
3911 list_for_each_entry(s, &slab_caches, list) {
3912 if (slab_unmergeable(s))
3918 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3921 * Check if alignment is compatible.
3922 * Courtesy of Adrian Drzewiecki
3924 if ((s->size & ~(align - 1)) != s->size)
3927 if (s->size - size >= sizeof(void *))
3935 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3936 size_t align, unsigned long flags, void (*ctor)(void *))
3938 struct kmem_cache *s;
3944 down_write(&slub_lock);
3945 s = find_mergeable(size, align, flags, name, ctor);
3949 * Adjust the object sizes so that we clear
3950 * the complete object on kzalloc.
3952 s->objsize = max(s->objsize, (int)size);
3953 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3955 if (sysfs_slab_alias(s, name)) {
3959 up_write(&slub_lock);
3963 n = kstrdup(name, GFP_KERNEL);
3967 s = kmalloc(kmem_size, GFP_KERNEL);
3969 if (kmem_cache_open(s, n,
3970 size, align, flags, ctor)) {
3971 list_add(&s->list, &slab_caches);
3972 up_write(&slub_lock);
3973 if (sysfs_slab_add(s)) {
3974 down_write(&slub_lock);
3986 up_write(&slub_lock);
3988 if (flags & SLAB_PANIC)
3989 panic("Cannot create slabcache %s\n", name);
3994 EXPORT_SYMBOL(kmem_cache_create);
3998 * Use the cpu notifier to insure that the cpu slabs are flushed when
4001 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
4002 unsigned long action, void *hcpu)
4004 long cpu = (long)hcpu;
4005 struct kmem_cache *s;
4006 unsigned long flags;
4009 case CPU_UP_CANCELED:
4010 case CPU_UP_CANCELED_FROZEN:
4012 case CPU_DEAD_FROZEN:
4013 down_read(&slub_lock);
4014 list_for_each_entry(s, &slab_caches, list) {
4015 local_irq_save(flags);
4016 __flush_cpu_slab(s, cpu);
4017 local_irq_restore(flags);
4019 up_read(&slub_lock);
4027 static struct notifier_block __cpuinitdata slab_notifier = {
4028 .notifier_call = slab_cpuup_callback
4033 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4035 struct kmem_cache *s;
4038 if (unlikely(size > SLUB_MAX_SIZE))
4039 return kmalloc_large(size, gfpflags);
4041 s = get_slab(size, gfpflags);
4043 if (unlikely(ZERO_OR_NULL_PTR(s)))
4046 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4048 /* Honor the call site pointer we received. */
4049 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4055 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4056 int node, unsigned long caller)
4058 struct kmem_cache *s;
4061 if (unlikely(size > SLUB_MAX_SIZE)) {
4062 ret = kmalloc_large_node(size, gfpflags, node);
4064 trace_kmalloc_node(caller, ret,
4065 size, PAGE_SIZE << get_order(size),
4071 s = get_slab(size, gfpflags);
4073 if (unlikely(ZERO_OR_NULL_PTR(s)))
4076 ret = slab_alloc(s, gfpflags, node, caller);
4078 /* Honor the call site pointer we received. */
4079 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4086 static int count_inuse(struct page *page)
4091 static int count_total(struct page *page)
4093 return page->objects;
4097 #ifdef CONFIG_SLUB_DEBUG
4098 static int validate_slab(struct kmem_cache *s, struct page *page,
4102 void *addr = page_address(page);
4104 if (!check_slab(s, page) ||
4105 !on_freelist(s, page, NULL))
4108 /* Now we know that a valid freelist exists */
4109 bitmap_zero(map, page->objects);
4111 get_map(s, page, map);
4112 for_each_object(p, s, addr, page->objects) {
4113 if (test_bit(slab_index(p, s, addr), map))
4114 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4118 for_each_object(p, s, addr, page->objects)
4119 if (!test_bit(slab_index(p, s, addr), map))
4120 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4125 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4129 validate_slab(s, page, map);
4133 static int validate_slab_node(struct kmem_cache *s,
4134 struct kmem_cache_node *n, unsigned long *map)
4136 unsigned long count = 0;
4138 unsigned long flags;
4140 spin_lock_irqsave(&n->list_lock, flags);
4142 list_for_each_entry(page, &n->partial, lru) {
4143 validate_slab_slab(s, page, map);
4146 if (count != n->nr_partial)
4147 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4148 "counter=%ld\n", s->name, count, n->nr_partial);
4150 if (!(s->flags & SLAB_STORE_USER))
4153 list_for_each_entry(page, &n->full, lru) {
4154 validate_slab_slab(s, page, map);
4157 if (count != atomic_long_read(&n->nr_slabs))
4158 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4159 "counter=%ld\n", s->name, count,
4160 atomic_long_read(&n->nr_slabs));
4163 spin_unlock_irqrestore(&n->list_lock, flags);
4167 static long validate_slab_cache(struct kmem_cache *s)
4170 unsigned long count = 0;
4171 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4172 sizeof(unsigned long), GFP_KERNEL);
4178 for_each_node_state(node, N_NORMAL_MEMORY) {
4179 struct kmem_cache_node *n = get_node(s, node);
4181 count += validate_slab_node(s, n, map);
4187 * Generate lists of code addresses where slabcache objects are allocated
4192 unsigned long count;
4199 DECLARE_BITMAP(cpus, NR_CPUS);
4205 unsigned long count;
4206 struct location *loc;
4209 static void free_loc_track(struct loc_track *t)
4212 free_pages((unsigned long)t->loc,
4213 get_order(sizeof(struct location) * t->max));
4216 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4221 order = get_order(sizeof(struct location) * max);
4223 l = (void *)__get_free_pages(flags, order);
4228 memcpy(l, t->loc, sizeof(struct location) * t->count);
4236 static int add_location(struct loc_track *t, struct kmem_cache *s,
4237 const struct track *track)
4239 long start, end, pos;
4241 unsigned long caddr;
4242 unsigned long age = jiffies - track->when;
4248 pos = start + (end - start + 1) / 2;
4251 * There is nothing at "end". If we end up there
4252 * we need to add something to before end.
4257 caddr = t->loc[pos].addr;
4258 if (track->addr == caddr) {
4264 if (age < l->min_time)
4266 if (age > l->max_time)
4269 if (track->pid < l->min_pid)
4270 l->min_pid = track->pid;
4271 if (track->pid > l->max_pid)
4272 l->max_pid = track->pid;
4274 cpumask_set_cpu(track->cpu,
4275 to_cpumask(l->cpus));
4277 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4281 if (track->addr < caddr)
4288 * Not found. Insert new tracking element.
4290 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4296 (t->count - pos) * sizeof(struct location));
4299 l->addr = track->addr;
4303 l->min_pid = track->pid;
4304 l->max_pid = track->pid;
4305 cpumask_clear(to_cpumask(l->cpus));
4306 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4307 nodes_clear(l->nodes);
4308 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4312 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4313 struct page *page, enum track_item alloc,
4316 void *addr = page_address(page);
4319 bitmap_zero(map, page->objects);
4320 get_map(s, page, map);
4322 for_each_object(p, s, addr, page->objects)
4323 if (!test_bit(slab_index(p, s, addr), map))
4324 add_location(t, s, get_track(s, p, alloc));
4327 static int list_locations(struct kmem_cache *s, char *buf,
4328 enum track_item alloc)
4332 struct loc_track t = { 0, 0, NULL };
4334 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4335 sizeof(unsigned long), GFP_KERNEL);
4337 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4340 return sprintf(buf, "Out of memory\n");
4342 /* Push back cpu slabs */
4345 for_each_node_state(node, N_NORMAL_MEMORY) {
4346 struct kmem_cache_node *n = get_node(s, node);
4347 unsigned long flags;
4350 if (!atomic_long_read(&n->nr_slabs))
4353 spin_lock_irqsave(&n->list_lock, flags);
4354 list_for_each_entry(page, &n->partial, lru)
4355 process_slab(&t, s, page, alloc, map);
4356 list_for_each_entry(page, &n->full, lru)
4357 process_slab(&t, s, page, alloc, map);
4358 spin_unlock_irqrestore(&n->list_lock, flags);
4361 for (i = 0; i < t.count; i++) {
4362 struct location *l = &t.loc[i];
4364 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4366 len += sprintf(buf + len, "%7ld ", l->count);
4369 len += sprintf(buf + len, "%pS", (void *)l->addr);
4371 len += sprintf(buf + len, "<not-available>");
4373 if (l->sum_time != l->min_time) {
4374 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4376 (long)div_u64(l->sum_time, l->count),
4379 len += sprintf(buf + len, " age=%ld",
4382 if (l->min_pid != l->max_pid)
4383 len += sprintf(buf + len, " pid=%ld-%ld",
4384 l->min_pid, l->max_pid);
4386 len += sprintf(buf + len, " pid=%ld",
4389 if (num_online_cpus() > 1 &&
4390 !cpumask_empty(to_cpumask(l->cpus)) &&
4391 len < PAGE_SIZE - 60) {
4392 len += sprintf(buf + len, " cpus=");
4393 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4394 to_cpumask(l->cpus));
4397 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4398 len < PAGE_SIZE - 60) {
4399 len += sprintf(buf + len, " nodes=");
4400 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4404 len += sprintf(buf + len, "\n");
4410 len += sprintf(buf, "No data\n");
4415 #ifdef SLUB_RESILIENCY_TEST
4416 static void resiliency_test(void)
4420 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4422 printk(KERN_ERR "SLUB resiliency testing\n");
4423 printk(KERN_ERR "-----------------------\n");
4424 printk(KERN_ERR "A. Corruption after allocation\n");
4426 p = kzalloc(16, GFP_KERNEL);
4428 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4429 " 0x12->0x%p\n\n", p + 16);
4431 validate_slab_cache(kmalloc_caches[4]);
4433 /* Hmmm... The next two are dangerous */
4434 p = kzalloc(32, GFP_KERNEL);
4435 p[32 + sizeof(void *)] = 0x34;
4436 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4437 " 0x34 -> -0x%p\n", p);
4439 "If allocated object is overwritten then not detectable\n\n");
4441 validate_slab_cache(kmalloc_caches[5]);
4442 p = kzalloc(64, GFP_KERNEL);
4443 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4445 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4448 "If allocated object is overwritten then not detectable\n\n");
4449 validate_slab_cache(kmalloc_caches[6]);
4451 printk(KERN_ERR "\nB. Corruption after free\n");
4452 p = kzalloc(128, GFP_KERNEL);
4455 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4456 validate_slab_cache(kmalloc_caches[7]);
4458 p = kzalloc(256, GFP_KERNEL);
4461 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4463 validate_slab_cache(kmalloc_caches[8]);
4465 p = kzalloc(512, GFP_KERNEL);
4468 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4469 validate_slab_cache(kmalloc_caches[9]);
4473 static void resiliency_test(void) {};
4478 enum slab_stat_type {
4479 SL_ALL, /* All slabs */
4480 SL_PARTIAL, /* Only partially allocated slabs */
4481 SL_CPU, /* Only slabs used for cpu caches */
4482 SL_OBJECTS, /* Determine allocated objects not slabs */
4483 SL_TOTAL /* Determine object capacity not slabs */
4486 #define SO_ALL (1 << SL_ALL)
4487 #define SO_PARTIAL (1 << SL_PARTIAL)
4488 #define SO_CPU (1 << SL_CPU)
4489 #define SO_OBJECTS (1 << SL_OBJECTS)
4490 #define SO_TOTAL (1 << SL_TOTAL)
4492 static ssize_t show_slab_objects(struct kmem_cache *s,
4493 char *buf, unsigned long flags)
4495 unsigned long total = 0;
4498 unsigned long *nodes;
4499 unsigned long *per_cpu;
4501 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4504 per_cpu = nodes + nr_node_ids;
4506 if (flags & SO_CPU) {
4509 for_each_possible_cpu(cpu) {
4510 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4514 page = ACCESS_ONCE(c->page);
4518 node = page_to_nid(page);
4519 if (flags & SO_TOTAL)
4521 else if (flags & SO_OBJECTS)
4529 page = ACCESS_ONCE(c->partial);
4540 lock_memory_hotplug();
4541 #ifdef CONFIG_SLUB_DEBUG
4542 if (flags & SO_ALL) {
4543 for_each_node_state(node, N_NORMAL_MEMORY) {
4544 struct kmem_cache_node *n = get_node(s, node);
4546 if (flags & SO_TOTAL)
4547 x = atomic_long_read(&n->total_objects);
4548 else if (flags & SO_OBJECTS)
4549 x = atomic_long_read(&n->total_objects) -
4550 count_partial(n, count_free);
4553 x = atomic_long_read(&n->nr_slabs);
4560 if (flags & SO_PARTIAL) {
4561 for_each_node_state(node, N_NORMAL_MEMORY) {
4562 struct kmem_cache_node *n = get_node(s, node);
4564 if (flags & SO_TOTAL)
4565 x = count_partial(n, count_total);
4566 else if (flags & SO_OBJECTS)
4567 x = count_partial(n, count_inuse);
4574 x = sprintf(buf, "%lu", total);
4576 for_each_node_state(node, N_NORMAL_MEMORY)
4578 x += sprintf(buf + x, " N%d=%lu",
4581 unlock_memory_hotplug();
4583 return x + sprintf(buf + x, "\n");
4586 #ifdef CONFIG_SLUB_DEBUG
4587 static int any_slab_objects(struct kmem_cache *s)
4591 for_each_online_node(node) {
4592 struct kmem_cache_node *n = get_node(s, node);
4597 if (atomic_long_read(&n->total_objects))
4604 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4605 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4607 struct slab_attribute {
4608 struct attribute attr;
4609 ssize_t (*show)(struct kmem_cache *s, char *buf);
4610 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4613 #define SLAB_ATTR_RO(_name) \
4614 static struct slab_attribute _name##_attr = \
4615 __ATTR(_name, 0400, _name##_show, NULL)
4617 #define SLAB_ATTR(_name) \
4618 static struct slab_attribute _name##_attr = \
4619 __ATTR(_name, 0600, _name##_show, _name##_store)
4621 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4623 return sprintf(buf, "%d\n", s->size);
4625 SLAB_ATTR_RO(slab_size);
4627 static ssize_t align_show(struct kmem_cache *s, char *buf)
4629 return sprintf(buf, "%d\n", s->align);
4631 SLAB_ATTR_RO(align);
4633 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4635 return sprintf(buf, "%d\n", s->objsize);
4637 SLAB_ATTR_RO(object_size);
4639 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4641 return sprintf(buf, "%d\n", oo_objects(s->oo));
4643 SLAB_ATTR_RO(objs_per_slab);
4645 static ssize_t order_store(struct kmem_cache *s,
4646 const char *buf, size_t length)
4648 unsigned long order;
4651 err = strict_strtoul(buf, 10, &order);
4655 if (order > slub_max_order || order < slub_min_order)
4658 calculate_sizes(s, order);
4662 static ssize_t order_show(struct kmem_cache *s, char *buf)
4664 return sprintf(buf, "%d\n", oo_order(s->oo));
4668 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4670 return sprintf(buf, "%lu\n", s->min_partial);
4673 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4679 err = strict_strtoul(buf, 10, &min);
4683 set_min_partial(s, min);
4686 SLAB_ATTR(min_partial);
4688 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4690 return sprintf(buf, "%u\n", s->cpu_partial);
4693 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4696 unsigned long objects;
4699 err = strict_strtoul(buf, 10, &objects);
4702 if (objects && kmem_cache_debug(s))
4705 s->cpu_partial = objects;
4709 SLAB_ATTR(cpu_partial);
4711 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4715 return sprintf(buf, "%pS\n", s->ctor);
4719 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4721 return sprintf(buf, "%d\n", s->refcount - 1);
4723 SLAB_ATTR_RO(aliases);
4725 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4727 return show_slab_objects(s, buf, SO_PARTIAL);
4729 SLAB_ATTR_RO(partial);
4731 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4733 return show_slab_objects(s, buf, SO_CPU);
4735 SLAB_ATTR_RO(cpu_slabs);
4737 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4739 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4741 SLAB_ATTR_RO(objects);
4743 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4745 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4747 SLAB_ATTR_RO(objects_partial);
4749 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4756 for_each_online_cpu(cpu) {
4757 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4760 pages += page->pages;
4761 objects += page->pobjects;
4765 len = sprintf(buf, "%d(%d)", objects, pages);
4768 for_each_online_cpu(cpu) {
4769 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4771 if (page && len < PAGE_SIZE - 20)
4772 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4773 page->pobjects, page->pages);
4776 return len + sprintf(buf + len, "\n");
4778 SLAB_ATTR_RO(slabs_cpu_partial);
4780 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4782 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4785 static ssize_t reclaim_account_store(struct kmem_cache *s,
4786 const char *buf, size_t length)
4788 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4790 s->flags |= SLAB_RECLAIM_ACCOUNT;
4793 SLAB_ATTR(reclaim_account);
4795 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4797 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4799 SLAB_ATTR_RO(hwcache_align);
4801 #ifdef CONFIG_ZONE_DMA
4802 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4804 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4806 SLAB_ATTR_RO(cache_dma);
4809 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4811 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4813 SLAB_ATTR_RO(destroy_by_rcu);
4815 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4817 return sprintf(buf, "%d\n", s->reserved);
4819 SLAB_ATTR_RO(reserved);
4821 #ifdef CONFIG_SLUB_DEBUG
4822 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4824 return show_slab_objects(s, buf, SO_ALL);
4826 SLAB_ATTR_RO(slabs);
4828 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4830 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4832 SLAB_ATTR_RO(total_objects);
4834 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4836 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4839 static ssize_t sanity_checks_store(struct kmem_cache *s,
4840 const char *buf, size_t length)
4842 s->flags &= ~SLAB_DEBUG_FREE;
4843 if (buf[0] == '1') {
4844 s->flags &= ~__CMPXCHG_DOUBLE;
4845 s->flags |= SLAB_DEBUG_FREE;
4849 SLAB_ATTR(sanity_checks);
4851 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4853 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4856 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4859 s->flags &= ~SLAB_TRACE;
4860 if (buf[0] == '1') {
4861 s->flags &= ~__CMPXCHG_DOUBLE;
4862 s->flags |= SLAB_TRACE;
4868 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4870 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4873 static ssize_t red_zone_store(struct kmem_cache *s,
4874 const char *buf, size_t length)
4876 if (any_slab_objects(s))
4879 s->flags &= ~SLAB_RED_ZONE;
4880 if (buf[0] == '1') {
4881 s->flags &= ~__CMPXCHG_DOUBLE;
4882 s->flags |= SLAB_RED_ZONE;
4884 calculate_sizes(s, -1);
4887 SLAB_ATTR(red_zone);
4889 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4891 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4894 static ssize_t poison_store(struct kmem_cache *s,
4895 const char *buf, size_t length)
4897 if (any_slab_objects(s))
4900 s->flags &= ~SLAB_POISON;
4901 if (buf[0] == '1') {
4902 s->flags &= ~__CMPXCHG_DOUBLE;
4903 s->flags |= SLAB_POISON;
4905 calculate_sizes(s, -1);
4910 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4912 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4915 static ssize_t store_user_store(struct kmem_cache *s,
4916 const char *buf, size_t length)
4918 if (any_slab_objects(s))
4921 s->flags &= ~SLAB_STORE_USER;
4922 if (buf[0] == '1') {
4923 s->flags &= ~__CMPXCHG_DOUBLE;
4924 s->flags |= SLAB_STORE_USER;
4926 calculate_sizes(s, -1);
4929 SLAB_ATTR(store_user);
4931 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4936 static ssize_t validate_store(struct kmem_cache *s,
4937 const char *buf, size_t length)
4941 if (buf[0] == '1') {
4942 ret = validate_slab_cache(s);
4948 SLAB_ATTR(validate);
4950 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4952 if (!(s->flags & SLAB_STORE_USER))
4954 return list_locations(s, buf, TRACK_ALLOC);
4956 SLAB_ATTR_RO(alloc_calls);
4958 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4960 if (!(s->flags & SLAB_STORE_USER))
4962 return list_locations(s, buf, TRACK_FREE);
4964 SLAB_ATTR_RO(free_calls);
4965 #endif /* CONFIG_SLUB_DEBUG */
4967 #ifdef CONFIG_FAILSLAB
4968 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4970 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4973 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4976 s->flags &= ~SLAB_FAILSLAB;
4978 s->flags |= SLAB_FAILSLAB;
4981 SLAB_ATTR(failslab);
4984 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4989 static ssize_t shrink_store(struct kmem_cache *s,
4990 const char *buf, size_t length)
4992 if (buf[0] == '1') {
4993 int rc = kmem_cache_shrink(s);
5004 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5006 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5009 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5010 const char *buf, size_t length)
5012 unsigned long ratio;
5015 err = strict_strtoul(buf, 10, &ratio);
5020 s->remote_node_defrag_ratio = ratio * 10;
5024 SLAB_ATTR(remote_node_defrag_ratio);
5027 #ifdef CONFIG_SLUB_STATS
5028 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5030 unsigned long sum = 0;
5033 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5038 for_each_online_cpu(cpu) {
5039 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5045 len = sprintf(buf, "%lu", sum);
5048 for_each_online_cpu(cpu) {
5049 if (data[cpu] && len < PAGE_SIZE - 20)
5050 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5054 return len + sprintf(buf + len, "\n");
5057 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5061 for_each_online_cpu(cpu)
5062 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5065 #define STAT_ATTR(si, text) \
5066 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5068 return show_stat(s, buf, si); \
5070 static ssize_t text##_store(struct kmem_cache *s, \
5071 const char *buf, size_t length) \
5073 if (buf[0] != '0') \
5075 clear_stat(s, si); \
5080 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5081 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5082 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5083 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5084 STAT_ATTR(FREE_FROZEN, free_frozen);
5085 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5086 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5087 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5088 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5089 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5090 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5091 STAT_ATTR(FREE_SLAB, free_slab);
5092 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5093 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5094 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5095 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5096 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5097 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5098 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5099 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5100 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5101 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5102 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5103 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5104 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5105 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5108 static struct attribute *slab_attrs[] = {
5109 &slab_size_attr.attr,
5110 &object_size_attr.attr,
5111 &objs_per_slab_attr.attr,
5113 &min_partial_attr.attr,
5114 &cpu_partial_attr.attr,
5116 &objects_partial_attr.attr,
5118 &cpu_slabs_attr.attr,
5122 &hwcache_align_attr.attr,
5123 &reclaim_account_attr.attr,
5124 &destroy_by_rcu_attr.attr,
5126 &reserved_attr.attr,
5127 &slabs_cpu_partial_attr.attr,
5128 #ifdef CONFIG_SLUB_DEBUG
5129 &total_objects_attr.attr,
5131 &sanity_checks_attr.attr,
5133 &red_zone_attr.attr,
5135 &store_user_attr.attr,
5136 &validate_attr.attr,
5137 &alloc_calls_attr.attr,
5138 &free_calls_attr.attr,
5140 #ifdef CONFIG_ZONE_DMA
5141 &cache_dma_attr.attr,
5144 &remote_node_defrag_ratio_attr.attr,
5146 #ifdef CONFIG_SLUB_STATS
5147 &alloc_fastpath_attr.attr,
5148 &alloc_slowpath_attr.attr,
5149 &free_fastpath_attr.attr,
5150 &free_slowpath_attr.attr,
5151 &free_frozen_attr.attr,
5152 &free_add_partial_attr.attr,
5153 &free_remove_partial_attr.attr,
5154 &alloc_from_partial_attr.attr,
5155 &alloc_slab_attr.attr,
5156 &alloc_refill_attr.attr,
5157 &alloc_node_mismatch_attr.attr,
5158 &free_slab_attr.attr,
5159 &cpuslab_flush_attr.attr,
5160 &deactivate_full_attr.attr,
5161 &deactivate_empty_attr.attr,
5162 &deactivate_to_head_attr.attr,
5163 &deactivate_to_tail_attr.attr,
5164 &deactivate_remote_frees_attr.attr,
5165 &deactivate_bypass_attr.attr,
5166 &order_fallback_attr.attr,
5167 &cmpxchg_double_fail_attr.attr,
5168 &cmpxchg_double_cpu_fail_attr.attr,
5169 &cpu_partial_alloc_attr.attr,
5170 &cpu_partial_free_attr.attr,
5171 &cpu_partial_node_attr.attr,
5172 &cpu_partial_drain_attr.attr,
5174 #ifdef CONFIG_FAILSLAB
5175 &failslab_attr.attr,
5181 static struct attribute_group slab_attr_group = {
5182 .attrs = slab_attrs,
5185 static ssize_t slab_attr_show(struct kobject *kobj,
5186 struct attribute *attr,
5189 struct slab_attribute *attribute;
5190 struct kmem_cache *s;
5193 attribute = to_slab_attr(attr);
5196 if (!attribute->show)
5199 err = attribute->show(s, buf);
5204 static ssize_t slab_attr_store(struct kobject *kobj,
5205 struct attribute *attr,
5206 const char *buf, size_t len)
5208 struct slab_attribute *attribute;
5209 struct kmem_cache *s;
5212 attribute = to_slab_attr(attr);
5215 if (!attribute->store)
5218 err = attribute->store(s, buf, len);
5223 static void kmem_cache_release(struct kobject *kobj)
5225 struct kmem_cache *s = to_slab(kobj);
5231 static const struct sysfs_ops slab_sysfs_ops = {
5232 .show = slab_attr_show,
5233 .store = slab_attr_store,
5236 static struct kobj_type slab_ktype = {
5237 .sysfs_ops = &slab_sysfs_ops,
5238 .release = kmem_cache_release
5241 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5243 struct kobj_type *ktype = get_ktype(kobj);
5245 if (ktype == &slab_ktype)
5250 static const struct kset_uevent_ops slab_uevent_ops = {
5251 .filter = uevent_filter,
5254 static struct kset *slab_kset;
5256 #define ID_STR_LENGTH 64
5258 /* Create a unique string id for a slab cache:
5260 * Format :[flags-]size
5262 static char *create_unique_id(struct kmem_cache *s)
5264 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5271 * First flags affecting slabcache operations. We will only
5272 * get here for aliasable slabs so we do not need to support
5273 * too many flags. The flags here must cover all flags that
5274 * are matched during merging to guarantee that the id is
5277 if (s->flags & SLAB_CACHE_DMA)
5279 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5281 if (s->flags & SLAB_DEBUG_FREE)
5283 if (!(s->flags & SLAB_NOTRACK))
5287 p += sprintf(p, "%07d", s->size);
5288 BUG_ON(p > name + ID_STR_LENGTH - 1);
5292 static int sysfs_slab_add(struct kmem_cache *s)
5298 if (slab_state < SYSFS)
5299 /* Defer until later */
5302 unmergeable = slab_unmergeable(s);
5305 * Slabcache can never be merged so we can use the name proper.
5306 * This is typically the case for debug situations. In that
5307 * case we can catch duplicate names easily.
5309 sysfs_remove_link(&slab_kset->kobj, s->name);
5313 * Create a unique name for the slab as a target
5316 name = create_unique_id(s);
5319 s->kobj.kset = slab_kset;
5320 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5322 kobject_put(&s->kobj);
5326 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5328 kobject_del(&s->kobj);
5329 kobject_put(&s->kobj);
5332 kobject_uevent(&s->kobj, KOBJ_ADD);
5334 /* Setup first alias */
5335 sysfs_slab_alias(s, s->name);
5341 static void sysfs_slab_remove(struct kmem_cache *s)
5343 if (slab_state < SYSFS)
5345 * Sysfs has not been setup yet so no need to remove the
5350 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5351 kobject_del(&s->kobj);
5352 kobject_put(&s->kobj);
5356 * Need to buffer aliases during bootup until sysfs becomes
5357 * available lest we lose that information.
5359 struct saved_alias {
5360 struct kmem_cache *s;
5362 struct saved_alias *next;
5365 static struct saved_alias *alias_list;
5367 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5369 struct saved_alias *al;
5371 if (slab_state == SYSFS) {
5373 * If we have a leftover link then remove it.
5375 sysfs_remove_link(&slab_kset->kobj, name);
5376 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5379 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5385 al->next = alias_list;
5390 static int __init slab_sysfs_init(void)
5392 struct kmem_cache *s;
5395 down_write(&slub_lock);
5397 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5399 up_write(&slub_lock);
5400 printk(KERN_ERR "Cannot register slab subsystem.\n");
5406 list_for_each_entry(s, &slab_caches, list) {
5407 err = sysfs_slab_add(s);
5409 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5410 " to sysfs\n", s->name);
5413 while (alias_list) {
5414 struct saved_alias *al = alias_list;
5416 alias_list = alias_list->next;
5417 err = sysfs_slab_alias(al->s, al->name);
5419 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5420 " %s to sysfs\n", s->name);
5424 up_write(&slub_lock);
5429 __initcall(slab_sysfs_init);
5430 #endif /* CONFIG_SYSFS */
5433 * The /proc/slabinfo ABI
5435 #ifdef CONFIG_SLABINFO
5436 static void print_slabinfo_header(struct seq_file *m)
5438 seq_puts(m, "slabinfo - version: 2.1\n");
5439 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5440 "<objperslab> <pagesperslab>");
5441 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5442 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5446 static void *s_start(struct seq_file *m, loff_t *pos)
5450 down_read(&slub_lock);
5452 print_slabinfo_header(m);
5454 return seq_list_start(&slab_caches, *pos);
5457 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5459 return seq_list_next(p, &slab_caches, pos);
5462 static void s_stop(struct seq_file *m, void *p)
5464 up_read(&slub_lock);
5467 static int s_show(struct seq_file *m, void *p)
5469 unsigned long nr_partials = 0;
5470 unsigned long nr_slabs = 0;
5471 unsigned long nr_inuse = 0;
5472 unsigned long nr_objs = 0;
5473 unsigned long nr_free = 0;
5474 struct kmem_cache *s;
5477 s = list_entry(p, struct kmem_cache, list);
5479 for_each_online_node(node) {
5480 struct kmem_cache_node *n = get_node(s, node);
5485 nr_partials += n->nr_partial;
5486 nr_slabs += atomic_long_read(&n->nr_slabs);
5487 nr_objs += atomic_long_read(&n->total_objects);
5488 nr_free += count_partial(n, count_free);
5491 nr_inuse = nr_objs - nr_free;
5493 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5494 nr_objs, s->size, oo_objects(s->oo),
5495 (1 << oo_order(s->oo)));
5496 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5497 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5503 static const struct seq_operations slabinfo_op = {
5510 static int slabinfo_open(struct inode *inode, struct file *file)
5512 return seq_open(file, &slabinfo_op);
5515 static const struct file_operations proc_slabinfo_operations = {
5516 .open = slabinfo_open,
5518 .llseek = seq_lseek,
5519 .release = seq_release,
5522 static int __init slab_proc_init(void)
5524 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5527 module_init(slab_proc_init);
5528 #endif /* CONFIG_SLABINFO */