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>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
35 #include <trace/events/kmem.h>
41 * 1. slab_mutex (Global Mutex)
43 * 3. slab_lock(page) (Only on some arches and for debugging)
47 * The role of the slab_mutex is to protect the list of all the slabs
48 * and to synchronize major metadata changes to slab cache structures.
50 * The slab_lock is only used for debugging and on arches that do not
51 * have the ability to do a cmpxchg_double. It only protects the second
52 * double word in the page struct. Meaning
53 * A. page->freelist -> List of object free in a page
54 * B. page->counters -> Counters of objects
55 * C. page->frozen -> frozen state
57 * If a slab is frozen then it is exempt from list management. It is not
58 * on any list. The processor that froze the slab is the one who can
59 * perform list operations on the page. Other processors may put objects
60 * onto the freelist but the processor that froze the slab is the only
61 * one that can retrieve the objects from the page's freelist.
63 * The list_lock protects the partial and full list on each node and
64 * the partial slab counter. If taken then no new slabs may be added or
65 * removed from the lists nor make the number of partial slabs be modified.
66 * (Note that the total number of slabs is an atomic value that may be
67 * modified without taking the list lock).
69 * The list_lock is a centralized lock and thus we avoid taking it as
70 * much as possible. As long as SLUB does not have to handle partial
71 * slabs, operations can continue without any centralized lock. F.e.
72 * allocating a long series of objects that fill up slabs does not require
74 * Interrupts are disabled during allocation and deallocation in order to
75 * make the slab allocator safe to use in the context of an irq. In addition
76 * interrupts are disabled to ensure that the processor does not change
77 * while handling per_cpu slabs, due to kernel preemption.
79 * SLUB assigns one slab for allocation to each processor.
80 * Allocations only occur from these slabs called cpu slabs.
82 * Slabs with free elements are kept on a partial list and during regular
83 * operations no list for full slabs is used. If an object in a full slab is
84 * freed then the slab will show up again on the partial lists.
85 * We track full slabs for debugging purposes though because otherwise we
86 * cannot scan all objects.
88 * Slabs are freed when they become empty. Teardown and setup is
89 * minimal so we rely on the page allocators per cpu caches for
90 * fast frees and allocs.
92 * Overloading of page flags that are otherwise used for LRU management.
94 * PageActive The slab is frozen and exempt from list processing.
95 * This means that the slab is dedicated to a purpose
96 * such as satisfying allocations for a specific
97 * processor. Objects may be freed in the slab while
98 * it is frozen but slab_free will then skip the usual
99 * list operations. It is up to the processor holding
100 * the slab to integrate the slab into the slab lists
101 * when the slab is no longer needed.
103 * One use of this flag is to mark slabs that are
104 * used for allocations. Then such a slab becomes a cpu
105 * slab. The cpu slab may be equipped with an additional
106 * freelist that allows lockless access to
107 * free objects in addition to the regular freelist
108 * that requires the slab lock.
110 * PageError Slab requires special handling due to debug
111 * options set. This moves slab handling out of
112 * the fast path and disables lockless freelists.
115 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
116 SLAB_TRACE | SLAB_DEBUG_FREE)
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
128 * Issues still to be resolved:
130 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
132 * - Variable sizing of the per node arrays
135 /* Enable to test recovery from slab corruption on boot */
136 #undef SLUB_RESILIENCY_TEST
138 /* Enable to log cmpxchg failures */
139 #undef SLUB_DEBUG_CMPXCHG
142 * Mininum number of partial slabs. These will be left on the partial
143 * lists even if they are empty. kmem_cache_shrink may reclaim them.
145 #define MIN_PARTIAL 5
148 * Maximum number of desirable partial slabs.
149 * The existence of more partial slabs makes kmem_cache_shrink
150 * sort the partial list by the number of objects in the.
152 #define MAX_PARTIAL 10
154 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_STORE_USER)
158 * Debugging flags that require metadata to be stored in the slab. These get
159 * disabled when slub_debug=O is used and a cache's min order increases with
162 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
165 * Set of flags that will prevent slab merging
167 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
168 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
171 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
172 SLAB_CACHE_DMA | SLAB_NOTRACK)
175 #define OO_MASK ((1 << OO_SHIFT) - 1)
176 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
178 /* Internal SLUB flags */
179 #define __OBJECT_POISON 0x80000000UL /* Poison object */
180 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static int kmem_size = sizeof(struct kmem_cache);
185 static struct notifier_block slab_notifier;
189 * Tracking user of a slab.
191 #define TRACK_ADDRS_COUNT 16
193 unsigned long addr; /* Called from address */
194 #ifdef CONFIG_STACKTRACE
195 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
197 int cpu; /* Was running on cpu */
198 int pid; /* Pid context */
199 unsigned long when; /* When did the operation occur */
202 enum track_item { TRACK_ALLOC, TRACK_FREE };
205 static int sysfs_slab_add(struct kmem_cache *);
206 static int sysfs_slab_alias(struct kmem_cache *, const char *);
207 static void sysfs_slab_remove(struct kmem_cache *);
210 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
211 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
213 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
217 static inline void stat(const struct kmem_cache *s, enum stat_item si)
219 #ifdef CONFIG_SLUB_STATS
220 __this_cpu_inc(s->cpu_slab->stat[si]);
224 /********************************************************************
225 * Core slab cache functions
226 *******************************************************************/
228 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
230 return s->node[node];
233 /* Verify that a pointer has an address that is valid within a slab page */
234 static inline int check_valid_pointer(struct kmem_cache *s,
235 struct page *page, const void *object)
242 base = page_address(page);
243 if (object < base || object >= base + page->objects * s->size ||
244 (object - base) % s->size) {
251 static inline void *get_freepointer(struct kmem_cache *s, void *object)
253 return *(void **)(object + s->offset);
256 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
258 prefetch(object + s->offset);
261 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
265 #ifdef CONFIG_DEBUG_PAGEALLOC
266 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
268 p = get_freepointer(s, object);
273 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
275 *(void **)(object + s->offset) = fp;
278 /* Loop over all objects in a slab */
279 #define for_each_object(__p, __s, __addr, __objects) \
280 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
283 /* Determine object index from a given position */
284 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
286 return (p - addr) / s->size;
289 static inline size_t slab_ksize(const struct kmem_cache *s)
291 #ifdef CONFIG_SLUB_DEBUG
293 * Debugging requires use of the padding between object
294 * and whatever may come after it.
296 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
297 return s->object_size;
301 * If we have the need to store the freelist pointer
302 * back there or track user information then we can
303 * only use the space before that information.
305 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
308 * Else we can use all the padding etc for the allocation
313 static inline int order_objects(int order, unsigned long size, int reserved)
315 return ((PAGE_SIZE << order) - reserved) / size;
318 static inline struct kmem_cache_order_objects oo_make(int order,
319 unsigned long size, int reserved)
321 struct kmem_cache_order_objects x = {
322 (order << OO_SHIFT) + order_objects(order, size, reserved)
328 static inline int oo_order(struct kmem_cache_order_objects x)
330 return x.x >> OO_SHIFT;
333 static inline int oo_objects(struct kmem_cache_order_objects x)
335 return x.x & OO_MASK;
339 * Per slab locking using the pagelock
341 static __always_inline void slab_lock(struct page *page)
343 bit_spin_lock(PG_locked, &page->flags);
346 static __always_inline void slab_unlock(struct page *page)
348 __bit_spin_unlock(PG_locked, &page->flags);
351 /* Interrupts must be disabled (for the fallback code to work right) */
352 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
353 void *freelist_old, unsigned long counters_old,
354 void *freelist_new, unsigned long counters_new,
357 VM_BUG_ON(!irqs_disabled());
358 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
359 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
360 if (s->flags & __CMPXCHG_DOUBLE) {
361 if (cmpxchg_double(&page->freelist, &page->counters,
362 freelist_old, counters_old,
363 freelist_new, counters_new))
369 if (page->freelist == freelist_old && page->counters == counters_old) {
370 page->freelist = freelist_new;
371 page->counters = counters_new;
379 stat(s, CMPXCHG_DOUBLE_FAIL);
381 #ifdef SLUB_DEBUG_CMPXCHG
382 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
388 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
389 void *freelist_old, unsigned long counters_old,
390 void *freelist_new, unsigned long counters_new,
393 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
394 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
395 if (s->flags & __CMPXCHG_DOUBLE) {
396 if (cmpxchg_double(&page->freelist, &page->counters,
397 freelist_old, counters_old,
398 freelist_new, counters_new))
405 local_irq_save(flags);
407 if (page->freelist == freelist_old && page->counters == counters_old) {
408 page->freelist = freelist_new;
409 page->counters = counters_new;
411 local_irq_restore(flags);
415 local_irq_restore(flags);
419 stat(s, CMPXCHG_DOUBLE_FAIL);
421 #ifdef SLUB_DEBUG_CMPXCHG
422 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
428 #ifdef CONFIG_SLUB_DEBUG
430 * Determine a map of object in use on a page.
432 * Node listlock must be held to guarantee that the page does
433 * not vanish from under us.
435 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
438 void *addr = page_address(page);
440 for (p = page->freelist; p; p = get_freepointer(s, p))
441 set_bit(slab_index(p, s, addr), map);
447 #ifdef CONFIG_SLUB_DEBUG_ON
448 static int slub_debug = DEBUG_DEFAULT_FLAGS;
450 static int slub_debug;
453 static char *slub_debug_slabs;
454 static int disable_higher_order_debug;
459 static void print_section(char *text, u8 *addr, unsigned int length)
461 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
465 static struct track *get_track(struct kmem_cache *s, void *object,
466 enum track_item alloc)
471 p = object + s->offset + sizeof(void *);
473 p = object + s->inuse;
478 static void set_track(struct kmem_cache *s, void *object,
479 enum track_item alloc, unsigned long addr)
481 struct track *p = get_track(s, object, alloc);
484 #ifdef CONFIG_STACKTRACE
485 struct stack_trace trace;
488 trace.nr_entries = 0;
489 trace.max_entries = TRACK_ADDRS_COUNT;
490 trace.entries = p->addrs;
492 save_stack_trace(&trace);
494 /* See rant in lockdep.c */
495 if (trace.nr_entries != 0 &&
496 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
499 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
503 p->cpu = smp_processor_id();
504 p->pid = current->pid;
507 memset(p, 0, sizeof(struct track));
510 static void init_tracking(struct kmem_cache *s, void *object)
512 if (!(s->flags & SLAB_STORE_USER))
515 set_track(s, object, TRACK_FREE, 0UL);
516 set_track(s, object, TRACK_ALLOC, 0UL);
519 static void print_track(const char *s, struct track *t)
524 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
525 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
526 #ifdef CONFIG_STACKTRACE
529 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
531 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
538 static void print_tracking(struct kmem_cache *s, void *object)
540 if (!(s->flags & SLAB_STORE_USER))
543 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
544 print_track("Freed", get_track(s, object, TRACK_FREE));
547 static void print_page_info(struct page *page)
549 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
550 page, page->objects, page->inuse, page->freelist, page->flags);
554 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
560 vsnprintf(buf, sizeof(buf), fmt, args);
562 printk(KERN_ERR "========================================"
563 "=====================================\n");
564 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
565 printk(KERN_ERR "----------------------------------------"
566 "-------------------------------------\n\n");
569 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
575 vsnprintf(buf, sizeof(buf), fmt, args);
577 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
580 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
582 unsigned int off; /* Offset of last byte */
583 u8 *addr = page_address(page);
585 print_tracking(s, p);
587 print_page_info(page);
589 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
590 p, p - addr, get_freepointer(s, p));
593 print_section("Bytes b4 ", p - 16, 16);
595 print_section("Object ", p, min_t(unsigned long, s->object_size,
597 if (s->flags & SLAB_RED_ZONE)
598 print_section("Redzone ", p + s->object_size,
599 s->inuse - s->object_size);
602 off = s->offset + sizeof(void *);
606 if (s->flags & SLAB_STORE_USER)
607 off += 2 * sizeof(struct track);
610 /* Beginning of the filler is the free pointer */
611 print_section("Padding ", p + off, s->size - off);
616 static void object_err(struct kmem_cache *s, struct page *page,
617 u8 *object, char *reason)
619 slab_bug(s, "%s", reason);
620 print_trailer(s, page, object);
623 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
629 vsnprintf(buf, sizeof(buf), fmt, args);
631 slab_bug(s, "%s", buf);
632 print_page_info(page);
636 static void init_object(struct kmem_cache *s, void *object, u8 val)
640 if (s->flags & __OBJECT_POISON) {
641 memset(p, POISON_FREE, s->object_size - 1);
642 p[s->object_size - 1] = POISON_END;
645 if (s->flags & SLAB_RED_ZONE)
646 memset(p + s->object_size, val, s->inuse - s->object_size);
649 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
650 void *from, void *to)
652 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
653 memset(from, data, to - from);
656 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
657 u8 *object, char *what,
658 u8 *start, unsigned int value, unsigned int bytes)
663 fault = memchr_inv(start, value, bytes);
668 while (end > fault && end[-1] == value)
671 slab_bug(s, "%s overwritten", what);
672 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
673 fault, end - 1, fault[0], value);
674 print_trailer(s, page, object);
676 restore_bytes(s, what, value, fault, end);
684 * Bytes of the object to be managed.
685 * If the freepointer may overlay the object then the free
686 * pointer is the first word of the object.
688 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
691 * object + s->object_size
692 * Padding to reach word boundary. This is also used for Redzoning.
693 * Padding is extended by another word if Redzoning is enabled and
694 * object_size == inuse.
696 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
697 * 0xcc (RED_ACTIVE) for objects in use.
700 * Meta data starts here.
702 * A. Free pointer (if we cannot overwrite object on free)
703 * B. Tracking data for SLAB_STORE_USER
704 * C. Padding to reach required alignment boundary or at mininum
705 * one word if debugging is on to be able to detect writes
706 * before the word boundary.
708 * Padding is done using 0x5a (POISON_INUSE)
711 * Nothing is used beyond s->size.
713 * If slabcaches are merged then the object_size and inuse boundaries are mostly
714 * ignored. And therefore no slab options that rely on these boundaries
715 * may be used with merged slabcaches.
718 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
720 unsigned long off = s->inuse; /* The end of info */
723 /* Freepointer is placed after the object. */
724 off += sizeof(void *);
726 if (s->flags & SLAB_STORE_USER)
727 /* We also have user information there */
728 off += 2 * sizeof(struct track);
733 return check_bytes_and_report(s, page, p, "Object padding",
734 p + off, POISON_INUSE, s->size - off);
737 /* Check the pad bytes at the end of a slab page */
738 static int slab_pad_check(struct kmem_cache *s, struct page *page)
746 if (!(s->flags & SLAB_POISON))
749 start = page_address(page);
750 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
751 end = start + length;
752 remainder = length % s->size;
756 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
759 while (end > fault && end[-1] == POISON_INUSE)
762 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
763 print_section("Padding ", end - remainder, remainder);
765 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
769 static int check_object(struct kmem_cache *s, struct page *page,
770 void *object, u8 val)
773 u8 *endobject = object + s->object_size;
775 if (s->flags & SLAB_RED_ZONE) {
776 if (!check_bytes_and_report(s, page, object, "Redzone",
777 endobject, val, s->inuse - s->object_size))
780 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
781 check_bytes_and_report(s, page, p, "Alignment padding",
782 endobject, POISON_INUSE, s->inuse - s->object_size);
786 if (s->flags & SLAB_POISON) {
787 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
788 (!check_bytes_and_report(s, page, p, "Poison", p,
789 POISON_FREE, s->object_size - 1) ||
790 !check_bytes_and_report(s, page, p, "Poison",
791 p + s->object_size - 1, POISON_END, 1)))
794 * check_pad_bytes cleans up on its own.
796 check_pad_bytes(s, page, p);
799 if (!s->offset && val == SLUB_RED_ACTIVE)
801 * Object and freepointer overlap. Cannot check
802 * freepointer while object is allocated.
806 /* Check free pointer validity */
807 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
808 object_err(s, page, p, "Freepointer corrupt");
810 * No choice but to zap it and thus lose the remainder
811 * of the free objects in this slab. May cause
812 * another error because the object count is now wrong.
814 set_freepointer(s, p, NULL);
820 static int check_slab(struct kmem_cache *s, struct page *page)
824 VM_BUG_ON(!irqs_disabled());
826 if (!PageSlab(page)) {
827 slab_err(s, page, "Not a valid slab page");
831 maxobj = order_objects(compound_order(page), s->size, s->reserved);
832 if (page->objects > maxobj) {
833 slab_err(s, page, "objects %u > max %u",
834 s->name, page->objects, maxobj);
837 if (page->inuse > page->objects) {
838 slab_err(s, page, "inuse %u > max %u",
839 s->name, page->inuse, page->objects);
842 /* Slab_pad_check fixes things up after itself */
843 slab_pad_check(s, page);
848 * Determine if a certain object on a page is on the freelist. Must hold the
849 * slab lock to guarantee that the chains are in a consistent state.
851 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
856 unsigned long max_objects;
859 while (fp && nr <= page->objects) {
862 if (!check_valid_pointer(s, page, fp)) {
864 object_err(s, page, object,
865 "Freechain corrupt");
866 set_freepointer(s, object, NULL);
869 slab_err(s, page, "Freepointer corrupt");
870 page->freelist = NULL;
871 page->inuse = page->objects;
872 slab_fix(s, "Freelist cleared");
878 fp = get_freepointer(s, object);
882 max_objects = order_objects(compound_order(page), s->size, s->reserved);
883 if (max_objects > MAX_OBJS_PER_PAGE)
884 max_objects = MAX_OBJS_PER_PAGE;
886 if (page->objects != max_objects) {
887 slab_err(s, page, "Wrong number of objects. Found %d but "
888 "should be %d", page->objects, max_objects);
889 page->objects = max_objects;
890 slab_fix(s, "Number of objects adjusted.");
892 if (page->inuse != page->objects - nr) {
893 slab_err(s, page, "Wrong object count. Counter is %d but "
894 "counted were %d", page->inuse, page->objects - nr);
895 page->inuse = page->objects - nr;
896 slab_fix(s, "Object count adjusted.");
898 return search == NULL;
901 static void trace(struct kmem_cache *s, struct page *page, void *object,
904 if (s->flags & SLAB_TRACE) {
905 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
907 alloc ? "alloc" : "free",
912 print_section("Object ", (void *)object, s->object_size);
919 * Hooks for other subsystems that check memory allocations. In a typical
920 * production configuration these hooks all should produce no code at all.
922 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
924 flags &= gfp_allowed_mask;
925 lockdep_trace_alloc(flags);
926 might_sleep_if(flags & __GFP_WAIT);
928 return should_failslab(s->object_size, flags, s->flags);
931 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
933 flags &= gfp_allowed_mask;
934 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
935 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
938 static inline void slab_free_hook(struct kmem_cache *s, void *x)
940 kmemleak_free_recursive(x, s->flags);
943 * Trouble is that we may no longer disable interupts in the fast path
944 * So in order to make the debug calls that expect irqs to be
945 * disabled we need to disable interrupts temporarily.
947 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
951 local_irq_save(flags);
952 kmemcheck_slab_free(s, x, s->object_size);
953 debug_check_no_locks_freed(x, s->object_size);
954 local_irq_restore(flags);
957 if (!(s->flags & SLAB_DEBUG_OBJECTS))
958 debug_check_no_obj_freed(x, s->object_size);
962 * Tracking of fully allocated slabs for debugging purposes.
964 * list_lock must be held.
966 static void add_full(struct kmem_cache *s,
967 struct kmem_cache_node *n, struct page *page)
969 if (!(s->flags & SLAB_STORE_USER))
972 list_add(&page->lru, &n->full);
976 * list_lock must be held.
978 static void remove_full(struct kmem_cache *s, struct page *page)
980 if (!(s->flags & SLAB_STORE_USER))
983 list_del(&page->lru);
986 /* Tracking of the number of slabs for debugging purposes */
987 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
989 struct kmem_cache_node *n = get_node(s, node);
991 return atomic_long_read(&n->nr_slabs);
994 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
996 return atomic_long_read(&n->nr_slabs);
999 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1001 struct kmem_cache_node *n = get_node(s, node);
1004 * May be called early in order to allocate a slab for the
1005 * kmem_cache_node structure. Solve the chicken-egg
1006 * dilemma by deferring the increment of the count during
1007 * bootstrap (see early_kmem_cache_node_alloc).
1010 atomic_long_inc(&n->nr_slabs);
1011 atomic_long_add(objects, &n->total_objects);
1014 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1016 struct kmem_cache_node *n = get_node(s, node);
1018 atomic_long_dec(&n->nr_slabs);
1019 atomic_long_sub(objects, &n->total_objects);
1022 /* Object debug checks for alloc/free paths */
1023 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1026 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1029 init_object(s, object, SLUB_RED_INACTIVE);
1030 init_tracking(s, object);
1033 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1034 void *object, unsigned long addr)
1036 if (!check_slab(s, page))
1039 if (!check_valid_pointer(s, page, object)) {
1040 object_err(s, page, object, "Freelist Pointer check fails");
1044 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1047 /* Success perform special debug activities for allocs */
1048 if (s->flags & SLAB_STORE_USER)
1049 set_track(s, object, TRACK_ALLOC, addr);
1050 trace(s, page, object, 1);
1051 init_object(s, object, SLUB_RED_ACTIVE);
1055 if (PageSlab(page)) {
1057 * If this is a slab page then lets do the best we can
1058 * to avoid issues in the future. Marking all objects
1059 * as used avoids touching the remaining objects.
1061 slab_fix(s, "Marking all objects used");
1062 page->inuse = page->objects;
1063 page->freelist = NULL;
1068 static noinline struct kmem_cache_node *free_debug_processing(
1069 struct kmem_cache *s, struct page *page, void *object,
1070 unsigned long addr, unsigned long *flags)
1072 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1074 spin_lock_irqsave(&n->list_lock, *flags);
1077 if (!check_slab(s, page))
1080 if (!check_valid_pointer(s, page, object)) {
1081 slab_err(s, page, "Invalid object pointer 0x%p", object);
1085 if (on_freelist(s, page, object)) {
1086 object_err(s, page, object, "Object already free");
1090 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1093 if (unlikely(s != page->slab)) {
1094 if (!PageSlab(page)) {
1095 slab_err(s, page, "Attempt to free object(0x%p) "
1096 "outside of slab", object);
1097 } else if (!page->slab) {
1099 "SLUB <none>: no slab for object 0x%p.\n",
1103 object_err(s, page, object,
1104 "page slab pointer corrupt.");
1108 if (s->flags & SLAB_STORE_USER)
1109 set_track(s, object, TRACK_FREE, addr);
1110 trace(s, page, object, 0);
1111 init_object(s, object, SLUB_RED_INACTIVE);
1115 * Keep node_lock to preserve integrity
1116 * until the object is actually freed
1122 spin_unlock_irqrestore(&n->list_lock, *flags);
1123 slab_fix(s, "Object at 0x%p not freed", object);
1127 static int __init setup_slub_debug(char *str)
1129 slub_debug = DEBUG_DEFAULT_FLAGS;
1130 if (*str++ != '=' || !*str)
1132 * No options specified. Switch on full debugging.
1138 * No options but restriction on slabs. This means full
1139 * debugging for slabs matching a pattern.
1143 if (tolower(*str) == 'o') {
1145 * Avoid enabling debugging on caches if its minimum order
1146 * would increase as a result.
1148 disable_higher_order_debug = 1;
1155 * Switch off all debugging measures.
1160 * Determine which debug features should be switched on
1162 for (; *str && *str != ','; str++) {
1163 switch (tolower(*str)) {
1165 slub_debug |= SLAB_DEBUG_FREE;
1168 slub_debug |= SLAB_RED_ZONE;
1171 slub_debug |= SLAB_POISON;
1174 slub_debug |= SLAB_STORE_USER;
1177 slub_debug |= SLAB_TRACE;
1180 slub_debug |= SLAB_FAILSLAB;
1183 printk(KERN_ERR "slub_debug option '%c' "
1184 "unknown. skipped\n", *str);
1190 slub_debug_slabs = str + 1;
1195 __setup("slub_debug", setup_slub_debug);
1197 static unsigned long kmem_cache_flags(unsigned long object_size,
1198 unsigned long flags, const char *name,
1199 void (*ctor)(void *))
1202 * Enable debugging if selected on the kernel commandline.
1204 if (slub_debug && (!slub_debug_slabs ||
1205 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1206 flags |= slub_debug;
1211 static inline void setup_object_debug(struct kmem_cache *s,
1212 struct page *page, void *object) {}
1214 static inline int alloc_debug_processing(struct kmem_cache *s,
1215 struct page *page, void *object, unsigned long addr) { return 0; }
1217 static inline struct kmem_cache_node *free_debug_processing(
1218 struct kmem_cache *s, struct page *page, void *object,
1219 unsigned long addr, unsigned long *flags) { return NULL; }
1221 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1223 static inline int check_object(struct kmem_cache *s, struct page *page,
1224 void *object, u8 val) { return 1; }
1225 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1226 struct page *page) {}
1227 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1228 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1229 unsigned long flags, const char *name,
1230 void (*ctor)(void *))
1234 #define slub_debug 0
1236 #define disable_higher_order_debug 0
1238 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1240 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1242 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1244 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1247 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1250 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1253 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1255 #endif /* CONFIG_SLUB_DEBUG */
1258 * Slab allocation and freeing
1260 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1261 struct kmem_cache_order_objects oo)
1263 int order = oo_order(oo);
1265 flags |= __GFP_NOTRACK;
1267 if (node == NUMA_NO_NODE)
1268 return alloc_pages(flags, order);
1270 return alloc_pages_exact_node(node, flags, order);
1273 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1276 struct kmem_cache_order_objects oo = s->oo;
1279 flags &= gfp_allowed_mask;
1281 if (flags & __GFP_WAIT)
1284 flags |= s->allocflags;
1287 * Let the initial higher-order allocation fail under memory pressure
1288 * so we fall-back to the minimum order allocation.
1290 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1292 page = alloc_slab_page(alloc_gfp, node, oo);
1293 if (unlikely(!page)) {
1296 * Allocation may have failed due to fragmentation.
1297 * Try a lower order alloc if possible
1299 page = alloc_slab_page(flags, node, oo);
1302 stat(s, ORDER_FALLBACK);
1305 if (kmemcheck_enabled && page
1306 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1307 int pages = 1 << oo_order(oo);
1309 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1312 * Objects from caches that have a constructor don't get
1313 * cleared when they're allocated, so we need to do it here.
1316 kmemcheck_mark_uninitialized_pages(page, pages);
1318 kmemcheck_mark_unallocated_pages(page, pages);
1321 if (flags & __GFP_WAIT)
1322 local_irq_disable();
1326 page->objects = oo_objects(oo);
1327 mod_zone_page_state(page_zone(page),
1328 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1329 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1335 static void setup_object(struct kmem_cache *s, struct page *page,
1338 setup_object_debug(s, page, object);
1339 if (unlikely(s->ctor))
1343 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1350 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1352 page = allocate_slab(s,
1353 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1357 inc_slabs_node(s, page_to_nid(page), page->objects);
1359 __SetPageSlab(page);
1360 if (page->pfmemalloc)
1361 SetPageSlabPfmemalloc(page);
1363 start = page_address(page);
1365 if (unlikely(s->flags & SLAB_POISON))
1366 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1369 for_each_object(p, s, start, page->objects) {
1370 setup_object(s, page, last);
1371 set_freepointer(s, last, p);
1374 setup_object(s, page, last);
1375 set_freepointer(s, last, NULL);
1377 page->freelist = start;
1378 page->inuse = page->objects;
1384 static void __free_slab(struct kmem_cache *s, struct page *page)
1386 int order = compound_order(page);
1387 int pages = 1 << order;
1389 if (kmem_cache_debug(s)) {
1392 slab_pad_check(s, page);
1393 for_each_object(p, s, page_address(page),
1395 check_object(s, page, p, SLUB_RED_INACTIVE);
1398 kmemcheck_free_shadow(page, compound_order(page));
1400 mod_zone_page_state(page_zone(page),
1401 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1402 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1405 __ClearPageSlabPfmemalloc(page);
1406 __ClearPageSlab(page);
1407 reset_page_mapcount(page);
1408 if (current->reclaim_state)
1409 current->reclaim_state->reclaimed_slab += pages;
1410 __free_pages(page, order);
1413 #define need_reserve_slab_rcu \
1414 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1416 static void rcu_free_slab(struct rcu_head *h)
1420 if (need_reserve_slab_rcu)
1421 page = virt_to_head_page(h);
1423 page = container_of((struct list_head *)h, struct page, lru);
1425 __free_slab(page->slab, page);
1428 static void free_slab(struct kmem_cache *s, struct page *page)
1430 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1431 struct rcu_head *head;
1433 if (need_reserve_slab_rcu) {
1434 int order = compound_order(page);
1435 int offset = (PAGE_SIZE << order) - s->reserved;
1437 VM_BUG_ON(s->reserved != sizeof(*head));
1438 head = page_address(page) + offset;
1441 * RCU free overloads the RCU head over the LRU
1443 head = (void *)&page->lru;
1446 call_rcu(head, rcu_free_slab);
1448 __free_slab(s, page);
1451 static void discard_slab(struct kmem_cache *s, struct page *page)
1453 dec_slabs_node(s, page_to_nid(page), page->objects);
1458 * Management of partially allocated slabs.
1460 * list_lock must be held.
1462 static inline void add_partial(struct kmem_cache_node *n,
1463 struct page *page, int tail)
1466 if (tail == DEACTIVATE_TO_TAIL)
1467 list_add_tail(&page->lru, &n->partial);
1469 list_add(&page->lru, &n->partial);
1473 * list_lock must be held.
1475 static inline void remove_partial(struct kmem_cache_node *n,
1478 list_del(&page->lru);
1483 * Remove slab from the partial list, freeze it and
1484 * return the pointer to the freelist.
1486 * Returns a list of objects or NULL if it fails.
1488 * Must hold list_lock since we modify the partial list.
1490 static inline void *acquire_slab(struct kmem_cache *s,
1491 struct kmem_cache_node *n, struct page *page,
1495 unsigned long counters;
1499 * Zap the freelist and set the frozen bit.
1500 * The old freelist is the list of objects for the
1501 * per cpu allocation list.
1503 freelist = page->freelist;
1504 counters = page->counters;
1505 new.counters = counters;
1507 new.inuse = page->objects;
1508 new.freelist = NULL;
1510 new.freelist = freelist;
1513 VM_BUG_ON(new.frozen);
1516 if (!__cmpxchg_double_slab(s, page,
1518 new.freelist, new.counters,
1522 remove_partial(n, page);
1527 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1530 * Try to allocate a partial slab from a specific node.
1532 static void *get_partial_node(struct kmem_cache *s,
1533 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1535 struct page *page, *page2;
1536 void *object = NULL;
1539 * Racy check. If we mistakenly see no partial slabs then we
1540 * just allocate an empty slab. If we mistakenly try to get a
1541 * partial slab and there is none available then get_partials()
1544 if (!n || !n->nr_partial)
1547 spin_lock(&n->list_lock);
1548 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1549 void *t = acquire_slab(s, n, page, object == NULL);
1557 stat(s, ALLOC_FROM_PARTIAL);
1559 available = page->objects - page->inuse;
1561 available = put_cpu_partial(s, page, 0);
1562 stat(s, CPU_PARTIAL_NODE);
1564 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1568 spin_unlock(&n->list_lock);
1573 * Get a page from somewhere. Search in increasing NUMA distances.
1575 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1576 struct kmem_cache_cpu *c)
1579 struct zonelist *zonelist;
1582 enum zone_type high_zoneidx = gfp_zone(flags);
1584 unsigned int cpuset_mems_cookie;
1587 * The defrag ratio allows a configuration of the tradeoffs between
1588 * inter node defragmentation and node local allocations. A lower
1589 * defrag_ratio increases the tendency to do local allocations
1590 * instead of attempting to obtain partial slabs from other nodes.
1592 * If the defrag_ratio is set to 0 then kmalloc() always
1593 * returns node local objects. If the ratio is higher then kmalloc()
1594 * may return off node objects because partial slabs are obtained
1595 * from other nodes and filled up.
1597 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1598 * defrag_ratio = 1000) then every (well almost) allocation will
1599 * first attempt to defrag slab caches on other nodes. This means
1600 * scanning over all nodes to look for partial slabs which may be
1601 * expensive if we do it every time we are trying to find a slab
1602 * with available objects.
1604 if (!s->remote_node_defrag_ratio ||
1605 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1609 cpuset_mems_cookie = get_mems_allowed();
1610 zonelist = node_zonelist(slab_node(), flags);
1611 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1612 struct kmem_cache_node *n;
1614 n = get_node(s, zone_to_nid(zone));
1616 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1617 n->nr_partial > s->min_partial) {
1618 object = get_partial_node(s, n, c);
1621 * Return the object even if
1622 * put_mems_allowed indicated that
1623 * the cpuset mems_allowed was
1624 * updated in parallel. It's a
1625 * harmless race between the alloc
1626 * and the cpuset update.
1628 put_mems_allowed(cpuset_mems_cookie);
1633 } while (!put_mems_allowed(cpuset_mems_cookie));
1639 * Get a partial page, lock it and return it.
1641 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1642 struct kmem_cache_cpu *c)
1645 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1647 object = get_partial_node(s, get_node(s, searchnode), c);
1648 if (object || node != NUMA_NO_NODE)
1651 return get_any_partial(s, flags, c);
1654 #ifdef CONFIG_PREEMPT
1656 * Calculate the next globally unique transaction for disambiguiation
1657 * during cmpxchg. The transactions start with the cpu number and are then
1658 * incremented by CONFIG_NR_CPUS.
1660 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1663 * No preemption supported therefore also no need to check for
1669 static inline unsigned long next_tid(unsigned long tid)
1671 return tid + TID_STEP;
1674 static inline unsigned int tid_to_cpu(unsigned long tid)
1676 return tid % TID_STEP;
1679 static inline unsigned long tid_to_event(unsigned long tid)
1681 return tid / TID_STEP;
1684 static inline unsigned int init_tid(int cpu)
1689 static inline void note_cmpxchg_failure(const char *n,
1690 const struct kmem_cache *s, unsigned long tid)
1692 #ifdef SLUB_DEBUG_CMPXCHG
1693 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1695 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1697 #ifdef CONFIG_PREEMPT
1698 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1699 printk("due to cpu change %d -> %d\n",
1700 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1703 if (tid_to_event(tid) != tid_to_event(actual_tid))
1704 printk("due to cpu running other code. Event %ld->%ld\n",
1705 tid_to_event(tid), tid_to_event(actual_tid));
1707 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1708 actual_tid, tid, next_tid(tid));
1710 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1713 void init_kmem_cache_cpus(struct kmem_cache *s)
1717 for_each_possible_cpu(cpu)
1718 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1722 * Remove the cpu slab
1724 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1726 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1727 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1729 enum slab_modes l = M_NONE, m = M_NONE;
1731 int tail = DEACTIVATE_TO_HEAD;
1735 if (page->freelist) {
1736 stat(s, DEACTIVATE_REMOTE_FREES);
1737 tail = DEACTIVATE_TO_TAIL;
1741 * Stage one: Free all available per cpu objects back
1742 * to the page freelist while it is still frozen. Leave the
1745 * There is no need to take the list->lock because the page
1748 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1750 unsigned long counters;
1753 prior = page->freelist;
1754 counters = page->counters;
1755 set_freepointer(s, freelist, prior);
1756 new.counters = counters;
1758 VM_BUG_ON(!new.frozen);
1760 } while (!__cmpxchg_double_slab(s, page,
1762 freelist, new.counters,
1763 "drain percpu freelist"));
1765 freelist = nextfree;
1769 * Stage two: Ensure that the page is unfrozen while the
1770 * list presence reflects the actual number of objects
1773 * We setup the list membership and then perform a cmpxchg
1774 * with the count. If there is a mismatch then the page
1775 * is not unfrozen but the page is on the wrong list.
1777 * Then we restart the process which may have to remove
1778 * the page from the list that we just put it on again
1779 * because the number of objects in the slab may have
1784 old.freelist = page->freelist;
1785 old.counters = page->counters;
1786 VM_BUG_ON(!old.frozen);
1788 /* Determine target state of the slab */
1789 new.counters = old.counters;
1792 set_freepointer(s, freelist, old.freelist);
1793 new.freelist = freelist;
1795 new.freelist = old.freelist;
1799 if (!new.inuse && n->nr_partial > s->min_partial)
1801 else if (new.freelist) {
1806 * Taking the spinlock removes the possiblity
1807 * that acquire_slab() will see a slab page that
1810 spin_lock(&n->list_lock);
1814 if (kmem_cache_debug(s) && !lock) {
1817 * This also ensures that the scanning of full
1818 * slabs from diagnostic functions will not see
1821 spin_lock(&n->list_lock);
1829 remove_partial(n, page);
1831 else if (l == M_FULL)
1833 remove_full(s, page);
1835 if (m == M_PARTIAL) {
1837 add_partial(n, page, tail);
1840 } else if (m == M_FULL) {
1842 stat(s, DEACTIVATE_FULL);
1843 add_full(s, n, page);
1849 if (!__cmpxchg_double_slab(s, page,
1850 old.freelist, old.counters,
1851 new.freelist, new.counters,
1856 spin_unlock(&n->list_lock);
1859 stat(s, DEACTIVATE_EMPTY);
1860 discard_slab(s, page);
1866 * Unfreeze all the cpu partial slabs.
1868 * This function must be called with interrupt disabled.
1870 static void unfreeze_partials(struct kmem_cache *s)
1872 struct kmem_cache_node *n = NULL, *n2 = NULL;
1873 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1874 struct page *page, *discard_page = NULL;
1876 while ((page = c->partial)) {
1880 c->partial = page->next;
1882 n2 = get_node(s, page_to_nid(page));
1885 spin_unlock(&n->list_lock);
1888 spin_lock(&n->list_lock);
1893 old.freelist = page->freelist;
1894 old.counters = page->counters;
1895 VM_BUG_ON(!old.frozen);
1897 new.counters = old.counters;
1898 new.freelist = old.freelist;
1902 } while (!__cmpxchg_double_slab(s, page,
1903 old.freelist, old.counters,
1904 new.freelist, new.counters,
1905 "unfreezing slab"));
1907 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1908 page->next = discard_page;
1909 discard_page = page;
1911 add_partial(n, page, DEACTIVATE_TO_TAIL);
1912 stat(s, FREE_ADD_PARTIAL);
1917 spin_unlock(&n->list_lock);
1919 while (discard_page) {
1920 page = discard_page;
1921 discard_page = discard_page->next;
1923 stat(s, DEACTIVATE_EMPTY);
1924 discard_slab(s, page);
1930 * Put a page that was just frozen (in __slab_free) into a partial page
1931 * slot if available. This is done without interrupts disabled and without
1932 * preemption disabled. The cmpxchg is racy and may put the partial page
1933 * onto a random cpus partial slot.
1935 * If we did not find a slot then simply move all the partials to the
1936 * per node partial list.
1938 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1940 struct page *oldpage;
1947 oldpage = this_cpu_read(s->cpu_slab->partial);
1950 pobjects = oldpage->pobjects;
1951 pages = oldpage->pages;
1952 if (drain && pobjects > s->cpu_partial) {
1953 unsigned long flags;
1955 * partial array is full. Move the existing
1956 * set to the per node partial list.
1958 local_irq_save(flags);
1959 unfreeze_partials(s);
1960 local_irq_restore(flags);
1964 stat(s, CPU_PARTIAL_DRAIN);
1969 pobjects += page->objects - page->inuse;
1971 page->pages = pages;
1972 page->pobjects = pobjects;
1973 page->next = oldpage;
1975 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1979 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1981 stat(s, CPUSLAB_FLUSH);
1982 deactivate_slab(s, c->page, c->freelist);
1984 c->tid = next_tid(c->tid);
1992 * Called from IPI handler with interrupts disabled.
1994 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1996 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2002 unfreeze_partials(s);
2006 static void flush_cpu_slab(void *d)
2008 struct kmem_cache *s = d;
2010 __flush_cpu_slab(s, smp_processor_id());
2013 static bool has_cpu_slab(int cpu, void *info)
2015 struct kmem_cache *s = info;
2016 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2018 return c->page || c->partial;
2021 static void flush_all(struct kmem_cache *s)
2023 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2027 * Check if the objects in a per cpu structure fit numa
2028 * locality expectations.
2030 static inline int node_match(struct page *page, int node)
2033 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2039 static int count_free(struct page *page)
2041 return page->objects - page->inuse;
2044 static unsigned long count_partial(struct kmem_cache_node *n,
2045 int (*get_count)(struct page *))
2047 unsigned long flags;
2048 unsigned long x = 0;
2051 spin_lock_irqsave(&n->list_lock, flags);
2052 list_for_each_entry(page, &n->partial, lru)
2053 x += get_count(page);
2054 spin_unlock_irqrestore(&n->list_lock, flags);
2058 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2060 #ifdef CONFIG_SLUB_DEBUG
2061 return atomic_long_read(&n->total_objects);
2067 static noinline void
2068 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2073 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2075 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2076 "default order: %d, min order: %d\n", s->name, s->object_size,
2077 s->size, oo_order(s->oo), oo_order(s->min));
2079 if (oo_order(s->min) > get_order(s->object_size))
2080 printk(KERN_WARNING " %s debugging increased min order, use "
2081 "slub_debug=O to disable.\n", s->name);
2083 for_each_online_node(node) {
2084 struct kmem_cache_node *n = get_node(s, node);
2085 unsigned long nr_slabs;
2086 unsigned long nr_objs;
2087 unsigned long nr_free;
2092 nr_free = count_partial(n, count_free);
2093 nr_slabs = node_nr_slabs(n);
2094 nr_objs = node_nr_objs(n);
2097 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2098 node, nr_slabs, nr_objs, nr_free);
2102 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2103 int node, struct kmem_cache_cpu **pc)
2106 struct kmem_cache_cpu *c = *pc;
2109 freelist = get_partial(s, flags, node, c);
2114 page = new_slab(s, flags, node);
2116 c = __this_cpu_ptr(s->cpu_slab);
2121 * No other reference to the page yet so we can
2122 * muck around with it freely without cmpxchg
2124 freelist = page->freelist;
2125 page->freelist = NULL;
2127 stat(s, ALLOC_SLAB);
2136 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2138 if (unlikely(PageSlabPfmemalloc(page)))
2139 return gfp_pfmemalloc_allowed(gfpflags);
2145 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2146 * or deactivate the page.
2148 * The page is still frozen if the return value is not NULL.
2150 * If this function returns NULL then the page has been unfrozen.
2152 * This function must be called with interrupt disabled.
2154 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2157 unsigned long counters;
2161 freelist = page->freelist;
2162 counters = page->counters;
2164 new.counters = counters;
2165 VM_BUG_ON(!new.frozen);
2167 new.inuse = page->objects;
2168 new.frozen = freelist != NULL;
2170 } while (!__cmpxchg_double_slab(s, page,
2179 * Slow path. The lockless freelist is empty or we need to perform
2182 * Processing is still very fast if new objects have been freed to the
2183 * regular freelist. In that case we simply take over the regular freelist
2184 * as the lockless freelist and zap the regular freelist.
2186 * If that is not working then we fall back to the partial lists. We take the
2187 * first element of the freelist as the object to allocate now and move the
2188 * rest of the freelist to the lockless freelist.
2190 * And if we were unable to get a new slab from the partial slab lists then
2191 * we need to allocate a new slab. This is the slowest path since it involves
2192 * a call to the page allocator and the setup of a new slab.
2194 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2195 unsigned long addr, struct kmem_cache_cpu *c)
2199 unsigned long flags;
2201 local_irq_save(flags);
2202 #ifdef CONFIG_PREEMPT
2204 * We may have been preempted and rescheduled on a different
2205 * cpu before disabling interrupts. Need to reload cpu area
2208 c = this_cpu_ptr(s->cpu_slab);
2216 if (unlikely(!node_match(page, node))) {
2217 stat(s, ALLOC_NODE_MISMATCH);
2218 deactivate_slab(s, page, c->freelist);
2225 * By rights, we should be searching for a slab page that was
2226 * PFMEMALLOC but right now, we are losing the pfmemalloc
2227 * information when the page leaves the per-cpu allocator
2229 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2230 deactivate_slab(s, page, c->freelist);
2236 /* must check again c->freelist in case of cpu migration or IRQ */
2237 freelist = c->freelist;
2241 stat(s, ALLOC_SLOWPATH);
2243 freelist = get_freelist(s, page);
2247 stat(s, DEACTIVATE_BYPASS);
2251 stat(s, ALLOC_REFILL);
2255 * freelist is pointing to the list of objects to be used.
2256 * page is pointing to the page from which the objects are obtained.
2257 * That page must be frozen for per cpu allocations to work.
2259 VM_BUG_ON(!c->page->frozen);
2260 c->freelist = get_freepointer(s, freelist);
2261 c->tid = next_tid(c->tid);
2262 local_irq_restore(flags);
2268 page = c->page = c->partial;
2269 c->partial = page->next;
2270 stat(s, CPU_PARTIAL_ALLOC);
2275 freelist = new_slab_objects(s, gfpflags, node, &c);
2277 if (unlikely(!freelist)) {
2278 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2279 slab_out_of_memory(s, gfpflags, node);
2281 local_irq_restore(flags);
2286 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2289 /* Only entered in the debug case */
2290 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2291 goto new_slab; /* Slab failed checks. Next slab needed */
2293 deactivate_slab(s, page, get_freepointer(s, freelist));
2296 local_irq_restore(flags);
2301 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2302 * have the fastpath folded into their functions. So no function call
2303 * overhead for requests that can be satisfied on the fastpath.
2305 * The fastpath works by first checking if the lockless freelist can be used.
2306 * If not then __slab_alloc is called for slow processing.
2308 * Otherwise we can simply pick the next object from the lockless free list.
2310 static __always_inline void *slab_alloc(struct kmem_cache *s,
2311 gfp_t gfpflags, int node, unsigned long addr)
2314 struct kmem_cache_cpu *c;
2318 if (slab_pre_alloc_hook(s, gfpflags))
2324 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2325 * enabled. We may switch back and forth between cpus while
2326 * reading from one cpu area. That does not matter as long
2327 * as we end up on the original cpu again when doing the cmpxchg.
2329 c = __this_cpu_ptr(s->cpu_slab);
2332 * The transaction ids are globally unique per cpu and per operation on
2333 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2334 * occurs on the right processor and that there was no operation on the
2335 * linked list in between.
2340 object = c->freelist;
2342 if (unlikely(!object || !node_match(page, node)))
2343 object = __slab_alloc(s, gfpflags, node, addr, c);
2346 void *next_object = get_freepointer_safe(s, object);
2349 * The cmpxchg will only match if there was no additional
2350 * operation and if we are on the right processor.
2352 * The cmpxchg does the following atomically (without lock semantics!)
2353 * 1. Relocate first pointer to the current per cpu area.
2354 * 2. Verify that tid and freelist have not been changed
2355 * 3. If they were not changed replace tid and freelist
2357 * Since this is without lock semantics the protection is only against
2358 * code executing on this cpu *not* from access by other cpus.
2360 if (unlikely(!this_cpu_cmpxchg_double(
2361 s->cpu_slab->freelist, s->cpu_slab->tid,
2363 next_object, next_tid(tid)))) {
2365 note_cmpxchg_failure("slab_alloc", s, tid);
2368 prefetch_freepointer(s, next_object);
2369 stat(s, ALLOC_FASTPATH);
2372 if (unlikely(gfpflags & __GFP_ZERO) && object)
2373 memset(object, 0, s->object_size);
2375 slab_post_alloc_hook(s, gfpflags, object);
2380 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2382 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2384 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2388 EXPORT_SYMBOL(kmem_cache_alloc);
2390 #ifdef CONFIG_TRACING
2391 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2393 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2394 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2397 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2399 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2401 void *ret = kmalloc_order(size, flags, order);
2402 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2405 EXPORT_SYMBOL(kmalloc_order_trace);
2409 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2411 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2413 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2414 s->object_size, s->size, gfpflags, node);
2418 EXPORT_SYMBOL(kmem_cache_alloc_node);
2420 #ifdef CONFIG_TRACING
2421 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2423 int node, size_t size)
2425 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2427 trace_kmalloc_node(_RET_IP_, ret,
2428 size, s->size, gfpflags, node);
2431 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2436 * Slow patch handling. This may still be called frequently since objects
2437 * have a longer lifetime than the cpu slabs in most processing loads.
2439 * So we still attempt to reduce cache line usage. Just take the slab
2440 * lock and free the item. If there is no additional partial page
2441 * handling required then we can return immediately.
2443 static void __slab_free(struct kmem_cache *s, struct page *page,
2444 void *x, unsigned long addr)
2447 void **object = (void *)x;
2451 unsigned long counters;
2452 struct kmem_cache_node *n = NULL;
2453 unsigned long uninitialized_var(flags);
2455 stat(s, FREE_SLOWPATH);
2457 if (kmem_cache_debug(s) &&
2458 !(n = free_debug_processing(s, page, x, addr, &flags)))
2462 prior = page->freelist;
2463 counters = page->counters;
2464 set_freepointer(s, object, prior);
2465 new.counters = counters;
2466 was_frozen = new.frozen;
2468 if ((!new.inuse || !prior) && !was_frozen && !n) {
2470 if (!kmem_cache_debug(s) && !prior)
2473 * Slab was on no list before and will be partially empty
2474 * We can defer the list move and instead freeze it.
2478 else { /* Needs to be taken off a list */
2480 n = get_node(s, page_to_nid(page));
2482 * Speculatively acquire the list_lock.
2483 * If the cmpxchg does not succeed then we may
2484 * drop the list_lock without any processing.
2486 * Otherwise the list_lock will synchronize with
2487 * other processors updating the list of slabs.
2489 spin_lock_irqsave(&n->list_lock, flags);
2495 } while (!cmpxchg_double_slab(s, page,
2497 object, new.counters,
2503 * If we just froze the page then put it onto the
2504 * per cpu partial list.
2506 if (new.frozen && !was_frozen) {
2507 put_cpu_partial(s, page, 1);
2508 stat(s, CPU_PARTIAL_FREE);
2511 * The list lock was not taken therefore no list
2512 * activity can be necessary.
2515 stat(s, FREE_FROZEN);
2520 * was_frozen may have been set after we acquired the list_lock in
2521 * an earlier loop. So we need to check it here again.
2524 stat(s, FREE_FROZEN);
2526 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2530 * Objects left in the slab. If it was not on the partial list before
2533 if (unlikely(!prior)) {
2534 remove_full(s, page);
2535 add_partial(n, page, DEACTIVATE_TO_TAIL);
2536 stat(s, FREE_ADD_PARTIAL);
2539 spin_unlock_irqrestore(&n->list_lock, flags);
2545 * Slab on the partial list.
2547 remove_partial(n, page);
2548 stat(s, FREE_REMOVE_PARTIAL);
2550 /* Slab must be on the full list */
2551 remove_full(s, page);
2553 spin_unlock_irqrestore(&n->list_lock, flags);
2555 discard_slab(s, page);
2559 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2560 * can perform fastpath freeing without additional function calls.
2562 * The fastpath is only possible if we are freeing to the current cpu slab
2563 * of this processor. This typically the case if we have just allocated
2566 * If fastpath is not possible then fall back to __slab_free where we deal
2567 * with all sorts of special processing.
2569 static __always_inline void slab_free(struct kmem_cache *s,
2570 struct page *page, void *x, unsigned long addr)
2572 void **object = (void *)x;
2573 struct kmem_cache_cpu *c;
2576 slab_free_hook(s, x);
2580 * Determine the currently cpus per cpu slab.
2581 * The cpu may change afterward. However that does not matter since
2582 * data is retrieved via this pointer. If we are on the same cpu
2583 * during the cmpxchg then the free will succedd.
2585 c = __this_cpu_ptr(s->cpu_slab);
2590 if (likely(page == c->page)) {
2591 set_freepointer(s, object, c->freelist);
2593 if (unlikely(!this_cpu_cmpxchg_double(
2594 s->cpu_slab->freelist, s->cpu_slab->tid,
2596 object, next_tid(tid)))) {
2598 note_cmpxchg_failure("slab_free", s, tid);
2601 stat(s, FREE_FASTPATH);
2603 __slab_free(s, page, x, addr);
2607 void kmem_cache_free(struct kmem_cache *s, void *x)
2611 page = virt_to_head_page(x);
2613 if (kmem_cache_debug(s) && page->slab != s) {
2614 pr_err("kmem_cache_free: Wrong slab cache. %s but object"
2615 " is from %s\n", page->slab->name, s->name);
2620 slab_free(s, page, x, _RET_IP_);
2622 trace_kmem_cache_free(_RET_IP_, x);
2624 EXPORT_SYMBOL(kmem_cache_free);
2627 * Object placement in a slab is made very easy because we always start at
2628 * offset 0. If we tune the size of the object to the alignment then we can
2629 * get the required alignment by putting one properly sized object after
2632 * Notice that the allocation order determines the sizes of the per cpu
2633 * caches. Each processor has always one slab available for allocations.
2634 * Increasing the allocation order reduces the number of times that slabs
2635 * must be moved on and off the partial lists and is therefore a factor in
2640 * Mininum / Maximum order of slab pages. This influences locking overhead
2641 * and slab fragmentation. A higher order reduces the number of partial slabs
2642 * and increases the number of allocations possible without having to
2643 * take the list_lock.
2645 static int slub_min_order;
2646 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2647 static int slub_min_objects;
2650 * Merge control. If this is set then no merging of slab caches will occur.
2651 * (Could be removed. This was introduced to pacify the merge skeptics.)
2653 static int slub_nomerge;
2656 * Calculate the order of allocation given an slab object size.
2658 * The order of allocation has significant impact on performance and other
2659 * system components. Generally order 0 allocations should be preferred since
2660 * order 0 does not cause fragmentation in the page allocator. Larger objects
2661 * be problematic to put into order 0 slabs because there may be too much
2662 * unused space left. We go to a higher order if more than 1/16th of the slab
2665 * In order to reach satisfactory performance we must ensure that a minimum
2666 * number of objects is in one slab. Otherwise we may generate too much
2667 * activity on the partial lists which requires taking the list_lock. This is
2668 * less a concern for large slabs though which are rarely used.
2670 * slub_max_order specifies the order where we begin to stop considering the
2671 * number of objects in a slab as critical. If we reach slub_max_order then
2672 * we try to keep the page order as low as possible. So we accept more waste
2673 * of space in favor of a small page order.
2675 * Higher order allocations also allow the placement of more objects in a
2676 * slab and thereby reduce object handling overhead. If the user has
2677 * requested a higher mininum order then we start with that one instead of
2678 * the smallest order which will fit the object.
2680 static inline int slab_order(int size, int min_objects,
2681 int max_order, int fract_leftover, int reserved)
2685 int min_order = slub_min_order;
2687 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2688 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2690 for (order = max(min_order,
2691 fls(min_objects * size - 1) - PAGE_SHIFT);
2692 order <= max_order; order++) {
2694 unsigned long slab_size = PAGE_SIZE << order;
2696 if (slab_size < min_objects * size + reserved)
2699 rem = (slab_size - reserved) % size;
2701 if (rem <= slab_size / fract_leftover)
2709 static inline int calculate_order(int size, int reserved)
2717 * Attempt to find best configuration for a slab. This
2718 * works by first attempting to generate a layout with
2719 * the best configuration and backing off gradually.
2721 * First we reduce the acceptable waste in a slab. Then
2722 * we reduce the minimum objects required in a slab.
2724 min_objects = slub_min_objects;
2726 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2727 max_objects = order_objects(slub_max_order, size, reserved);
2728 min_objects = min(min_objects, max_objects);
2730 while (min_objects > 1) {
2732 while (fraction >= 4) {
2733 order = slab_order(size, min_objects,
2734 slub_max_order, fraction, reserved);
2735 if (order <= slub_max_order)
2743 * We were unable to place multiple objects in a slab. Now
2744 * lets see if we can place a single object there.
2746 order = slab_order(size, 1, slub_max_order, 1, reserved);
2747 if (order <= slub_max_order)
2751 * Doh this slab cannot be placed using slub_max_order.
2753 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2754 if (order < MAX_ORDER)
2760 * Figure out what the alignment of the objects will be.
2762 static unsigned long calculate_alignment(unsigned long flags,
2763 unsigned long align, unsigned long size)
2766 * If the user wants hardware cache aligned objects then follow that
2767 * suggestion if the object is sufficiently large.
2769 * The hardware cache alignment cannot override the specified
2770 * alignment though. If that is greater then use it.
2772 if (flags & SLAB_HWCACHE_ALIGN) {
2773 unsigned long ralign = cache_line_size();
2774 while (size <= ralign / 2)
2776 align = max(align, ralign);
2779 if (align < ARCH_SLAB_MINALIGN)
2780 align = ARCH_SLAB_MINALIGN;
2782 return ALIGN(align, sizeof(void *));
2786 init_kmem_cache_node(struct kmem_cache_node *n)
2789 spin_lock_init(&n->list_lock);
2790 INIT_LIST_HEAD(&n->partial);
2791 #ifdef CONFIG_SLUB_DEBUG
2792 atomic_long_set(&n->nr_slabs, 0);
2793 atomic_long_set(&n->total_objects, 0);
2794 INIT_LIST_HEAD(&n->full);
2798 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2800 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2801 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2804 * Must align to double word boundary for the double cmpxchg
2805 * instructions to work; see __pcpu_double_call_return_bool().
2807 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2808 2 * sizeof(void *));
2813 init_kmem_cache_cpus(s);
2818 static struct kmem_cache *kmem_cache_node;
2821 * No kmalloc_node yet so do it by hand. We know that this is the first
2822 * slab on the node for this slabcache. There are no concurrent accesses
2825 * Note that this function only works on the kmalloc_node_cache
2826 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2827 * memory on a fresh node that has no slab structures yet.
2829 static void early_kmem_cache_node_alloc(int node)
2832 struct kmem_cache_node *n;
2834 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2836 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2839 if (page_to_nid(page) != node) {
2840 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2842 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2843 "in order to be able to continue\n");
2848 page->freelist = get_freepointer(kmem_cache_node, n);
2851 kmem_cache_node->node[node] = n;
2852 #ifdef CONFIG_SLUB_DEBUG
2853 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2854 init_tracking(kmem_cache_node, n);
2856 init_kmem_cache_node(n);
2857 inc_slabs_node(kmem_cache_node, node, page->objects);
2859 add_partial(n, page, DEACTIVATE_TO_HEAD);
2862 static void free_kmem_cache_nodes(struct kmem_cache *s)
2866 for_each_node_state(node, N_NORMAL_MEMORY) {
2867 struct kmem_cache_node *n = s->node[node];
2870 kmem_cache_free(kmem_cache_node, n);
2872 s->node[node] = NULL;
2876 static int init_kmem_cache_nodes(struct kmem_cache *s)
2880 for_each_node_state(node, N_NORMAL_MEMORY) {
2881 struct kmem_cache_node *n;
2883 if (slab_state == DOWN) {
2884 early_kmem_cache_node_alloc(node);
2887 n = kmem_cache_alloc_node(kmem_cache_node,
2891 free_kmem_cache_nodes(s);
2896 init_kmem_cache_node(n);
2901 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2903 if (min < MIN_PARTIAL)
2905 else if (min > MAX_PARTIAL)
2907 s->min_partial = min;
2911 * calculate_sizes() determines the order and the distribution of data within
2914 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2916 unsigned long flags = s->flags;
2917 unsigned long size = s->object_size;
2918 unsigned long align = s->align;
2922 * Round up object size to the next word boundary. We can only
2923 * place the free pointer at word boundaries and this determines
2924 * the possible location of the free pointer.
2926 size = ALIGN(size, sizeof(void *));
2928 #ifdef CONFIG_SLUB_DEBUG
2930 * Determine if we can poison the object itself. If the user of
2931 * the slab may touch the object after free or before allocation
2932 * then we should never poison the object itself.
2934 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2936 s->flags |= __OBJECT_POISON;
2938 s->flags &= ~__OBJECT_POISON;
2942 * If we are Redzoning then check if there is some space between the
2943 * end of the object and the free pointer. If not then add an
2944 * additional word to have some bytes to store Redzone information.
2946 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2947 size += sizeof(void *);
2951 * With that we have determined the number of bytes in actual use
2952 * by the object. This is the potential offset to the free pointer.
2956 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2959 * Relocate free pointer after the object if it is not
2960 * permitted to overwrite the first word of the object on
2963 * This is the case if we do RCU, have a constructor or
2964 * destructor or are poisoning the objects.
2967 size += sizeof(void *);
2970 #ifdef CONFIG_SLUB_DEBUG
2971 if (flags & SLAB_STORE_USER)
2973 * Need to store information about allocs and frees after
2976 size += 2 * sizeof(struct track);
2978 if (flags & SLAB_RED_ZONE)
2980 * Add some empty padding so that we can catch
2981 * overwrites from earlier objects rather than let
2982 * tracking information or the free pointer be
2983 * corrupted if a user writes before the start
2986 size += sizeof(void *);
2990 * Determine the alignment based on various parameters that the
2991 * user specified and the dynamic determination of cache line size
2994 align = calculate_alignment(flags, align, s->object_size);
2998 * SLUB stores one object immediately after another beginning from
2999 * offset 0. In order to align the objects we have to simply size
3000 * each object to conform to the alignment.
3002 size = ALIGN(size, align);
3004 if (forced_order >= 0)
3005 order = forced_order;
3007 order = calculate_order(size, s->reserved);
3014 s->allocflags |= __GFP_COMP;
3016 if (s->flags & SLAB_CACHE_DMA)
3017 s->allocflags |= SLUB_DMA;
3019 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3020 s->allocflags |= __GFP_RECLAIMABLE;
3023 * Determine the number of objects per slab
3025 s->oo = oo_make(order, size, s->reserved);
3026 s->min = oo_make(get_order(size), size, s->reserved);
3027 if (oo_objects(s->oo) > oo_objects(s->max))
3030 return !!oo_objects(s->oo);
3034 static int kmem_cache_open(struct kmem_cache *s,
3035 const char *name, size_t size,
3036 size_t align, unsigned long flags,
3037 void (*ctor)(void *))
3039 memset(s, 0, kmem_size);
3042 s->object_size = size;
3044 s->flags = kmem_cache_flags(size, flags, name, ctor);
3047 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3048 s->reserved = sizeof(struct rcu_head);
3050 if (!calculate_sizes(s, -1))
3052 if (disable_higher_order_debug) {
3054 * Disable debugging flags that store metadata if the min slab
3057 if (get_order(s->size) > get_order(s->object_size)) {
3058 s->flags &= ~DEBUG_METADATA_FLAGS;
3060 if (!calculate_sizes(s, -1))
3065 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3066 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3067 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3068 /* Enable fast mode */
3069 s->flags |= __CMPXCHG_DOUBLE;
3073 * The larger the object size is, the more pages we want on the partial
3074 * list to avoid pounding the page allocator excessively.
3076 set_min_partial(s, ilog2(s->size) / 2);
3079 * cpu_partial determined the maximum number of objects kept in the
3080 * per cpu partial lists of a processor.
3082 * Per cpu partial lists mainly contain slabs that just have one
3083 * object freed. If they are used for allocation then they can be
3084 * filled up again with minimal effort. The slab will never hit the
3085 * per node partial lists and therefore no locking will be required.
3087 * This setting also determines
3089 * A) The number of objects from per cpu partial slabs dumped to the
3090 * per node list when we reach the limit.
3091 * B) The number of objects in cpu partial slabs to extract from the
3092 * per node list when we run out of per cpu objects. We only fetch 50%
3093 * to keep some capacity around for frees.
3095 if (kmem_cache_debug(s))
3097 else if (s->size >= PAGE_SIZE)
3099 else if (s->size >= 1024)
3101 else if (s->size >= 256)
3102 s->cpu_partial = 13;
3104 s->cpu_partial = 30;
3108 s->remote_node_defrag_ratio = 1000;
3110 if (!init_kmem_cache_nodes(s))
3113 if (alloc_kmem_cache_cpus(s))
3116 free_kmem_cache_nodes(s);
3118 if (flags & SLAB_PANIC)
3119 panic("Cannot create slab %s size=%lu realsize=%u "
3120 "order=%u offset=%u flags=%lx\n",
3121 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3127 * Determine the size of a slab object
3129 unsigned int kmem_cache_size(struct kmem_cache *s)
3131 return s->object_size;
3133 EXPORT_SYMBOL(kmem_cache_size);
3135 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3138 #ifdef CONFIG_SLUB_DEBUG
3139 void *addr = page_address(page);
3141 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3142 sizeof(long), GFP_ATOMIC);
3145 slab_err(s, page, text, s->name);
3148 get_map(s, page, map);
3149 for_each_object(p, s, addr, page->objects) {
3151 if (!test_bit(slab_index(p, s, addr), map)) {
3152 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3154 print_tracking(s, p);
3163 * Attempt to free all partial slabs on a node.
3164 * This is called from kmem_cache_close(). We must be the last thread
3165 * using the cache and therefore we do not need to lock anymore.
3167 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3169 struct page *page, *h;
3171 list_for_each_entry_safe(page, h, &n->partial, lru) {
3173 remove_partial(n, page);
3174 discard_slab(s, page);
3176 list_slab_objects(s, page,
3177 "Objects remaining in %s on kmem_cache_close()");
3183 * Release all resources used by a slab cache.
3185 static inline int kmem_cache_close(struct kmem_cache *s)
3190 /* Attempt to free all objects */
3191 for_each_node_state(node, N_NORMAL_MEMORY) {
3192 struct kmem_cache_node *n = get_node(s, node);
3195 if (n->nr_partial || slabs_node(s, node))
3198 free_percpu(s->cpu_slab);
3199 free_kmem_cache_nodes(s);
3203 int __kmem_cache_shutdown(struct kmem_cache *s)
3205 int rc = kmem_cache_close(s);
3208 sysfs_slab_remove(s);
3213 /********************************************************************
3215 *******************************************************************/
3217 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3218 EXPORT_SYMBOL(kmalloc_caches);
3220 #ifdef CONFIG_ZONE_DMA
3221 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3224 static int __init setup_slub_min_order(char *str)
3226 get_option(&str, &slub_min_order);
3231 __setup("slub_min_order=", setup_slub_min_order);
3233 static int __init setup_slub_max_order(char *str)
3235 get_option(&str, &slub_max_order);
3236 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3241 __setup("slub_max_order=", setup_slub_max_order);
3243 static int __init setup_slub_min_objects(char *str)
3245 get_option(&str, &slub_min_objects);
3250 __setup("slub_min_objects=", setup_slub_min_objects);
3252 static int __init setup_slub_nomerge(char *str)
3258 __setup("slub_nomerge", setup_slub_nomerge);
3260 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3261 int size, unsigned int flags)
3263 struct kmem_cache *s;
3265 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3268 * This function is called with IRQs disabled during early-boot on
3269 * single CPU so there's no need to take slab_mutex here.
3271 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3275 list_add(&s->list, &slab_caches);
3279 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3284 * Conversion table for small slabs sizes / 8 to the index in the
3285 * kmalloc array. This is necessary for slabs < 192 since we have non power
3286 * of two cache sizes there. The size of larger slabs can be determined using
3289 static s8 size_index[24] = {
3316 static inline int size_index_elem(size_t bytes)
3318 return (bytes - 1) / 8;
3321 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3327 return ZERO_SIZE_PTR;
3329 index = size_index[size_index_elem(size)];
3331 index = fls(size - 1);
3333 #ifdef CONFIG_ZONE_DMA
3334 if (unlikely((flags & SLUB_DMA)))
3335 return kmalloc_dma_caches[index];
3338 return kmalloc_caches[index];
3341 void *__kmalloc(size_t size, gfp_t flags)
3343 struct kmem_cache *s;
3346 if (unlikely(size > SLUB_MAX_SIZE))
3347 return kmalloc_large(size, flags);
3349 s = get_slab(size, flags);
3351 if (unlikely(ZERO_OR_NULL_PTR(s)))
3354 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3356 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3360 EXPORT_SYMBOL(__kmalloc);
3363 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3368 flags |= __GFP_COMP | __GFP_NOTRACK;
3369 page = alloc_pages_node(node, flags, get_order(size));
3371 ptr = page_address(page);
3373 kmemleak_alloc(ptr, size, 1, flags);
3377 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3379 struct kmem_cache *s;
3382 if (unlikely(size > SLUB_MAX_SIZE)) {
3383 ret = kmalloc_large_node(size, flags, node);
3385 trace_kmalloc_node(_RET_IP_, ret,
3386 size, PAGE_SIZE << get_order(size),
3392 s = get_slab(size, flags);
3394 if (unlikely(ZERO_OR_NULL_PTR(s)))
3397 ret = slab_alloc(s, flags, node, _RET_IP_);
3399 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3403 EXPORT_SYMBOL(__kmalloc_node);
3406 size_t ksize(const void *object)
3410 if (unlikely(object == ZERO_SIZE_PTR))
3413 page = virt_to_head_page(object);
3415 if (unlikely(!PageSlab(page))) {
3416 WARN_ON(!PageCompound(page));
3417 return PAGE_SIZE << compound_order(page);
3420 return slab_ksize(page->slab);
3422 EXPORT_SYMBOL(ksize);
3424 #ifdef CONFIG_SLUB_DEBUG
3425 bool verify_mem_not_deleted(const void *x)
3428 void *object = (void *)x;
3429 unsigned long flags;
3432 if (unlikely(ZERO_OR_NULL_PTR(x)))
3435 local_irq_save(flags);
3437 page = virt_to_head_page(x);
3438 if (unlikely(!PageSlab(page))) {
3439 /* maybe it was from stack? */
3445 if (on_freelist(page->slab, page, object)) {
3446 object_err(page->slab, page, object, "Object is on free-list");
3454 local_irq_restore(flags);
3457 EXPORT_SYMBOL(verify_mem_not_deleted);
3460 void kfree(const void *x)
3463 void *object = (void *)x;
3465 trace_kfree(_RET_IP_, x);
3467 if (unlikely(ZERO_OR_NULL_PTR(x)))
3470 page = virt_to_head_page(x);
3471 if (unlikely(!PageSlab(page))) {
3472 BUG_ON(!PageCompound(page));
3474 __free_pages(page, compound_order(page));
3477 slab_free(page->slab, page, object, _RET_IP_);
3479 EXPORT_SYMBOL(kfree);
3482 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3483 * the remaining slabs by the number of items in use. The slabs with the
3484 * most items in use come first. New allocations will then fill those up
3485 * and thus they can be removed from the partial lists.
3487 * The slabs with the least items are placed last. This results in them
3488 * being allocated from last increasing the chance that the last objects
3489 * are freed in them.
3491 int kmem_cache_shrink(struct kmem_cache *s)
3495 struct kmem_cache_node *n;
3498 int objects = oo_objects(s->max);
3499 struct list_head *slabs_by_inuse =
3500 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3501 unsigned long flags;
3503 if (!slabs_by_inuse)
3507 for_each_node_state(node, N_NORMAL_MEMORY) {
3508 n = get_node(s, node);
3513 for (i = 0; i < objects; i++)
3514 INIT_LIST_HEAD(slabs_by_inuse + i);
3516 spin_lock_irqsave(&n->list_lock, flags);
3519 * Build lists indexed by the items in use in each slab.
3521 * Note that concurrent frees may occur while we hold the
3522 * list_lock. page->inuse here is the upper limit.
3524 list_for_each_entry_safe(page, t, &n->partial, lru) {
3525 list_move(&page->lru, slabs_by_inuse + page->inuse);
3531 * Rebuild the partial list with the slabs filled up most
3532 * first and the least used slabs at the end.
3534 for (i = objects - 1; i > 0; i--)
3535 list_splice(slabs_by_inuse + i, n->partial.prev);
3537 spin_unlock_irqrestore(&n->list_lock, flags);
3539 /* Release empty slabs */
3540 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3541 discard_slab(s, page);
3544 kfree(slabs_by_inuse);
3547 EXPORT_SYMBOL(kmem_cache_shrink);
3549 #if defined(CONFIG_MEMORY_HOTPLUG)
3550 static int slab_mem_going_offline_callback(void *arg)
3552 struct kmem_cache *s;
3554 mutex_lock(&slab_mutex);
3555 list_for_each_entry(s, &slab_caches, list)
3556 kmem_cache_shrink(s);
3557 mutex_unlock(&slab_mutex);
3562 static void slab_mem_offline_callback(void *arg)
3564 struct kmem_cache_node *n;
3565 struct kmem_cache *s;
3566 struct memory_notify *marg = arg;
3569 offline_node = marg->status_change_nid;
3572 * If the node still has available memory. we need kmem_cache_node
3575 if (offline_node < 0)
3578 mutex_lock(&slab_mutex);
3579 list_for_each_entry(s, &slab_caches, list) {
3580 n = get_node(s, offline_node);
3583 * if n->nr_slabs > 0, slabs still exist on the node
3584 * that is going down. We were unable to free them,
3585 * and offline_pages() function shouldn't call this
3586 * callback. So, we must fail.
3588 BUG_ON(slabs_node(s, offline_node));
3590 s->node[offline_node] = NULL;
3591 kmem_cache_free(kmem_cache_node, n);
3594 mutex_unlock(&slab_mutex);
3597 static int slab_mem_going_online_callback(void *arg)
3599 struct kmem_cache_node *n;
3600 struct kmem_cache *s;
3601 struct memory_notify *marg = arg;
3602 int nid = marg->status_change_nid;
3606 * If the node's memory is already available, then kmem_cache_node is
3607 * already created. Nothing to do.
3613 * We are bringing a node online. No memory is available yet. We must
3614 * allocate a kmem_cache_node structure in order to bring the node
3617 mutex_lock(&slab_mutex);
3618 list_for_each_entry(s, &slab_caches, list) {
3620 * XXX: kmem_cache_alloc_node will fallback to other nodes
3621 * since memory is not yet available from the node that
3624 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3629 init_kmem_cache_node(n);
3633 mutex_unlock(&slab_mutex);
3637 static int slab_memory_callback(struct notifier_block *self,
3638 unsigned long action, void *arg)
3643 case MEM_GOING_ONLINE:
3644 ret = slab_mem_going_online_callback(arg);
3646 case MEM_GOING_OFFLINE:
3647 ret = slab_mem_going_offline_callback(arg);
3650 case MEM_CANCEL_ONLINE:
3651 slab_mem_offline_callback(arg);
3654 case MEM_CANCEL_OFFLINE:
3658 ret = notifier_from_errno(ret);
3664 #endif /* CONFIG_MEMORY_HOTPLUG */
3666 /********************************************************************
3667 * Basic setup of slabs
3668 *******************************************************************/
3671 * Used for early kmem_cache structures that were allocated using
3672 * the page allocator
3675 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3679 list_add(&s->list, &slab_caches);
3682 for_each_node_state(node, N_NORMAL_MEMORY) {
3683 struct kmem_cache_node *n = get_node(s, node);
3687 list_for_each_entry(p, &n->partial, lru)
3690 #ifdef CONFIG_SLUB_DEBUG
3691 list_for_each_entry(p, &n->full, lru)
3698 void __init kmem_cache_init(void)
3702 struct kmem_cache *temp_kmem_cache;
3704 struct kmem_cache *temp_kmem_cache_node;
3705 unsigned long kmalloc_size;
3707 if (debug_guardpage_minorder())
3710 kmem_size = offsetof(struct kmem_cache, node) +
3711 nr_node_ids * sizeof(struct kmem_cache_node *);
3713 /* Allocate two kmem_caches from the page allocator */
3714 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3715 order = get_order(2 * kmalloc_size);
3716 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3719 * Must first have the slab cache available for the allocations of the
3720 * struct kmem_cache_node's. There is special bootstrap code in
3721 * kmem_cache_open for slab_state == DOWN.
3723 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3725 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3726 sizeof(struct kmem_cache_node),
3727 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3729 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3731 /* Able to allocate the per node structures */
3732 slab_state = PARTIAL;
3734 temp_kmem_cache = kmem_cache;
3735 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3736 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3737 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3738 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3741 * Allocate kmem_cache_node properly from the kmem_cache slab.
3742 * kmem_cache_node is separately allocated so no need to
3743 * update any list pointers.
3745 temp_kmem_cache_node = kmem_cache_node;
3747 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3748 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3750 kmem_cache_bootstrap_fixup(kmem_cache_node);
3753 kmem_cache_bootstrap_fixup(kmem_cache);
3755 /* Free temporary boot structure */
3756 free_pages((unsigned long)temp_kmem_cache, order);
3758 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3761 * Patch up the size_index table if we have strange large alignment
3762 * requirements for the kmalloc array. This is only the case for
3763 * MIPS it seems. The standard arches will not generate any code here.
3765 * Largest permitted alignment is 256 bytes due to the way we
3766 * handle the index determination for the smaller caches.
3768 * Make sure that nothing crazy happens if someone starts tinkering
3769 * around with ARCH_KMALLOC_MINALIGN
3771 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3772 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3774 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3775 int elem = size_index_elem(i);
3776 if (elem >= ARRAY_SIZE(size_index))
3778 size_index[elem] = KMALLOC_SHIFT_LOW;
3781 if (KMALLOC_MIN_SIZE == 64) {
3783 * The 96 byte size cache is not used if the alignment
3786 for (i = 64 + 8; i <= 96; i += 8)
3787 size_index[size_index_elem(i)] = 7;
3788 } else if (KMALLOC_MIN_SIZE == 128) {
3790 * The 192 byte sized cache is not used if the alignment
3791 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3794 for (i = 128 + 8; i <= 192; i += 8)
3795 size_index[size_index_elem(i)] = 8;
3798 /* Caches that are not of the two-to-the-power-of size */
3799 if (KMALLOC_MIN_SIZE <= 32) {
3800 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3804 if (KMALLOC_MIN_SIZE <= 64) {
3805 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3809 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3810 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3816 /* Provide the correct kmalloc names now that the caches are up */
3817 if (KMALLOC_MIN_SIZE <= 32) {
3818 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3819 BUG_ON(!kmalloc_caches[1]->name);
3822 if (KMALLOC_MIN_SIZE <= 64) {
3823 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3824 BUG_ON(!kmalloc_caches[2]->name);
3827 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3828 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3831 kmalloc_caches[i]->name = s;
3835 register_cpu_notifier(&slab_notifier);
3838 #ifdef CONFIG_ZONE_DMA
3839 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3840 struct kmem_cache *s = kmalloc_caches[i];
3843 char *name = kasprintf(GFP_NOWAIT,
3844 "dma-kmalloc-%d", s->object_size);
3847 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3848 s->object_size, SLAB_CACHE_DMA);
3853 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3854 " CPUs=%d, Nodes=%d\n",
3855 caches, cache_line_size(),
3856 slub_min_order, slub_max_order, slub_min_objects,
3857 nr_cpu_ids, nr_node_ids);
3860 void __init kmem_cache_init_late(void)
3865 * Find a mergeable slab cache
3867 static int slab_unmergeable(struct kmem_cache *s)
3869 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3876 * We may have set a slab to be unmergeable during bootstrap.
3878 if (s->refcount < 0)
3884 static struct kmem_cache *find_mergeable(size_t size,
3885 size_t align, unsigned long flags, const char *name,
3886 void (*ctor)(void *))
3888 struct kmem_cache *s;
3890 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3896 size = ALIGN(size, sizeof(void *));
3897 align = calculate_alignment(flags, align, size);
3898 size = ALIGN(size, align);
3899 flags = kmem_cache_flags(size, flags, name, NULL);
3901 list_for_each_entry(s, &slab_caches, list) {
3902 if (slab_unmergeable(s))
3908 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3911 * Check if alignment is compatible.
3912 * Courtesy of Adrian Drzewiecki
3914 if ((s->size & ~(align - 1)) != s->size)
3917 if (s->size - size >= sizeof(void *))
3925 struct kmem_cache *__kmem_cache_create(const char *name, size_t size,
3926 size_t align, unsigned long flags, void (*ctor)(void *))
3928 struct kmem_cache *s;
3930 s = find_mergeable(size, align, flags, name, ctor);
3934 * Adjust the object sizes so that we clear
3935 * the complete object on kzalloc.
3937 s->object_size = max(s->object_size, (int)size);
3938 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3940 if (sysfs_slab_alias(s, name)) {
3947 s = kmem_cache_alloc(kmem_cache, GFP_KERNEL);
3949 if (kmem_cache_open(s, name,
3950 size, align, flags, ctor)) {
3953 mutex_unlock(&slab_mutex);
3954 r = sysfs_slab_add(s);
3955 mutex_lock(&slab_mutex);
3960 kmem_cache_close(s);
3962 kmem_cache_free(kmem_cache, s);
3969 * Use the cpu notifier to insure that the cpu slabs are flushed when
3972 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3973 unsigned long action, void *hcpu)
3975 long cpu = (long)hcpu;
3976 struct kmem_cache *s;
3977 unsigned long flags;
3980 case CPU_UP_CANCELED:
3981 case CPU_UP_CANCELED_FROZEN:
3983 case CPU_DEAD_FROZEN:
3984 mutex_lock(&slab_mutex);
3985 list_for_each_entry(s, &slab_caches, list) {
3986 local_irq_save(flags);
3987 __flush_cpu_slab(s, cpu);
3988 local_irq_restore(flags);
3990 mutex_unlock(&slab_mutex);
3998 static struct notifier_block __cpuinitdata slab_notifier = {
3999 .notifier_call = slab_cpuup_callback
4004 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4006 struct kmem_cache *s;
4009 if (unlikely(size > SLUB_MAX_SIZE))
4010 return kmalloc_large(size, gfpflags);
4012 s = get_slab(size, gfpflags);
4014 if (unlikely(ZERO_OR_NULL_PTR(s)))
4017 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4019 /* Honor the call site pointer we received. */
4020 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4026 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4027 int node, unsigned long caller)
4029 struct kmem_cache *s;
4032 if (unlikely(size > SLUB_MAX_SIZE)) {
4033 ret = kmalloc_large_node(size, gfpflags, node);
4035 trace_kmalloc_node(caller, ret,
4036 size, PAGE_SIZE << get_order(size),
4042 s = get_slab(size, gfpflags);
4044 if (unlikely(ZERO_OR_NULL_PTR(s)))
4047 ret = slab_alloc(s, gfpflags, node, caller);
4049 /* Honor the call site pointer we received. */
4050 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4057 static int count_inuse(struct page *page)
4062 static int count_total(struct page *page)
4064 return page->objects;
4068 #ifdef CONFIG_SLUB_DEBUG
4069 static int validate_slab(struct kmem_cache *s, struct page *page,
4073 void *addr = page_address(page);
4075 if (!check_slab(s, page) ||
4076 !on_freelist(s, page, NULL))
4079 /* Now we know that a valid freelist exists */
4080 bitmap_zero(map, page->objects);
4082 get_map(s, page, map);
4083 for_each_object(p, s, addr, page->objects) {
4084 if (test_bit(slab_index(p, s, addr), map))
4085 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4089 for_each_object(p, s, addr, page->objects)
4090 if (!test_bit(slab_index(p, s, addr), map))
4091 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4096 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4100 validate_slab(s, page, map);
4104 static int validate_slab_node(struct kmem_cache *s,
4105 struct kmem_cache_node *n, unsigned long *map)
4107 unsigned long count = 0;
4109 unsigned long flags;
4111 spin_lock_irqsave(&n->list_lock, flags);
4113 list_for_each_entry(page, &n->partial, lru) {
4114 validate_slab_slab(s, page, map);
4117 if (count != n->nr_partial)
4118 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4119 "counter=%ld\n", s->name, count, n->nr_partial);
4121 if (!(s->flags & SLAB_STORE_USER))
4124 list_for_each_entry(page, &n->full, lru) {
4125 validate_slab_slab(s, page, map);
4128 if (count != atomic_long_read(&n->nr_slabs))
4129 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4130 "counter=%ld\n", s->name, count,
4131 atomic_long_read(&n->nr_slabs));
4134 spin_unlock_irqrestore(&n->list_lock, flags);
4138 static long validate_slab_cache(struct kmem_cache *s)
4141 unsigned long count = 0;
4142 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4143 sizeof(unsigned long), GFP_KERNEL);
4149 for_each_node_state(node, N_NORMAL_MEMORY) {
4150 struct kmem_cache_node *n = get_node(s, node);
4152 count += validate_slab_node(s, n, map);
4158 * Generate lists of code addresses where slabcache objects are allocated
4163 unsigned long count;
4170 DECLARE_BITMAP(cpus, NR_CPUS);
4176 unsigned long count;
4177 struct location *loc;
4180 static void free_loc_track(struct loc_track *t)
4183 free_pages((unsigned long)t->loc,
4184 get_order(sizeof(struct location) * t->max));
4187 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4192 order = get_order(sizeof(struct location) * max);
4194 l = (void *)__get_free_pages(flags, order);
4199 memcpy(l, t->loc, sizeof(struct location) * t->count);
4207 static int add_location(struct loc_track *t, struct kmem_cache *s,
4208 const struct track *track)
4210 long start, end, pos;
4212 unsigned long caddr;
4213 unsigned long age = jiffies - track->when;
4219 pos = start + (end - start + 1) / 2;
4222 * There is nothing at "end". If we end up there
4223 * we need to add something to before end.
4228 caddr = t->loc[pos].addr;
4229 if (track->addr == caddr) {
4235 if (age < l->min_time)
4237 if (age > l->max_time)
4240 if (track->pid < l->min_pid)
4241 l->min_pid = track->pid;
4242 if (track->pid > l->max_pid)
4243 l->max_pid = track->pid;
4245 cpumask_set_cpu(track->cpu,
4246 to_cpumask(l->cpus));
4248 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4252 if (track->addr < caddr)
4259 * Not found. Insert new tracking element.
4261 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4267 (t->count - pos) * sizeof(struct location));
4270 l->addr = track->addr;
4274 l->min_pid = track->pid;
4275 l->max_pid = track->pid;
4276 cpumask_clear(to_cpumask(l->cpus));
4277 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4278 nodes_clear(l->nodes);
4279 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4283 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4284 struct page *page, enum track_item alloc,
4287 void *addr = page_address(page);
4290 bitmap_zero(map, page->objects);
4291 get_map(s, page, map);
4293 for_each_object(p, s, addr, page->objects)
4294 if (!test_bit(slab_index(p, s, addr), map))
4295 add_location(t, s, get_track(s, p, alloc));
4298 static int list_locations(struct kmem_cache *s, char *buf,
4299 enum track_item alloc)
4303 struct loc_track t = { 0, 0, NULL };
4305 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4306 sizeof(unsigned long), GFP_KERNEL);
4308 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4311 return sprintf(buf, "Out of memory\n");
4313 /* Push back cpu slabs */
4316 for_each_node_state(node, N_NORMAL_MEMORY) {
4317 struct kmem_cache_node *n = get_node(s, node);
4318 unsigned long flags;
4321 if (!atomic_long_read(&n->nr_slabs))
4324 spin_lock_irqsave(&n->list_lock, flags);
4325 list_for_each_entry(page, &n->partial, lru)
4326 process_slab(&t, s, page, alloc, map);
4327 list_for_each_entry(page, &n->full, lru)
4328 process_slab(&t, s, page, alloc, map);
4329 spin_unlock_irqrestore(&n->list_lock, flags);
4332 for (i = 0; i < t.count; i++) {
4333 struct location *l = &t.loc[i];
4335 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4337 len += sprintf(buf + len, "%7ld ", l->count);
4340 len += sprintf(buf + len, "%pS", (void *)l->addr);
4342 len += sprintf(buf + len, "<not-available>");
4344 if (l->sum_time != l->min_time) {
4345 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4347 (long)div_u64(l->sum_time, l->count),
4350 len += sprintf(buf + len, " age=%ld",
4353 if (l->min_pid != l->max_pid)
4354 len += sprintf(buf + len, " pid=%ld-%ld",
4355 l->min_pid, l->max_pid);
4357 len += sprintf(buf + len, " pid=%ld",
4360 if (num_online_cpus() > 1 &&
4361 !cpumask_empty(to_cpumask(l->cpus)) &&
4362 len < PAGE_SIZE - 60) {
4363 len += sprintf(buf + len, " cpus=");
4364 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4365 to_cpumask(l->cpus));
4368 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4369 len < PAGE_SIZE - 60) {
4370 len += sprintf(buf + len, " nodes=");
4371 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4375 len += sprintf(buf + len, "\n");
4381 len += sprintf(buf, "No data\n");
4386 #ifdef SLUB_RESILIENCY_TEST
4387 static void resiliency_test(void)
4391 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4393 printk(KERN_ERR "SLUB resiliency testing\n");
4394 printk(KERN_ERR "-----------------------\n");
4395 printk(KERN_ERR "A. Corruption after allocation\n");
4397 p = kzalloc(16, GFP_KERNEL);
4399 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4400 " 0x12->0x%p\n\n", p + 16);
4402 validate_slab_cache(kmalloc_caches[4]);
4404 /* Hmmm... The next two are dangerous */
4405 p = kzalloc(32, GFP_KERNEL);
4406 p[32 + sizeof(void *)] = 0x34;
4407 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4408 " 0x34 -> -0x%p\n", p);
4410 "If allocated object is overwritten then not detectable\n\n");
4412 validate_slab_cache(kmalloc_caches[5]);
4413 p = kzalloc(64, GFP_KERNEL);
4414 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4416 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4419 "If allocated object is overwritten then not detectable\n\n");
4420 validate_slab_cache(kmalloc_caches[6]);
4422 printk(KERN_ERR "\nB. Corruption after free\n");
4423 p = kzalloc(128, GFP_KERNEL);
4426 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4427 validate_slab_cache(kmalloc_caches[7]);
4429 p = kzalloc(256, GFP_KERNEL);
4432 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4434 validate_slab_cache(kmalloc_caches[8]);
4436 p = kzalloc(512, GFP_KERNEL);
4439 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4440 validate_slab_cache(kmalloc_caches[9]);
4444 static void resiliency_test(void) {};
4449 enum slab_stat_type {
4450 SL_ALL, /* All slabs */
4451 SL_PARTIAL, /* Only partially allocated slabs */
4452 SL_CPU, /* Only slabs used for cpu caches */
4453 SL_OBJECTS, /* Determine allocated objects not slabs */
4454 SL_TOTAL /* Determine object capacity not slabs */
4457 #define SO_ALL (1 << SL_ALL)
4458 #define SO_PARTIAL (1 << SL_PARTIAL)
4459 #define SO_CPU (1 << SL_CPU)
4460 #define SO_OBJECTS (1 << SL_OBJECTS)
4461 #define SO_TOTAL (1 << SL_TOTAL)
4463 static ssize_t show_slab_objects(struct kmem_cache *s,
4464 char *buf, unsigned long flags)
4466 unsigned long total = 0;
4469 unsigned long *nodes;
4470 unsigned long *per_cpu;
4472 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4475 per_cpu = nodes + nr_node_ids;
4477 if (flags & SO_CPU) {
4480 for_each_possible_cpu(cpu) {
4481 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4485 page = ACCESS_ONCE(c->page);
4489 node = page_to_nid(page);
4490 if (flags & SO_TOTAL)
4492 else if (flags & SO_OBJECTS)
4500 page = ACCESS_ONCE(c->partial);
4511 lock_memory_hotplug();
4512 #ifdef CONFIG_SLUB_DEBUG
4513 if (flags & SO_ALL) {
4514 for_each_node_state(node, N_NORMAL_MEMORY) {
4515 struct kmem_cache_node *n = get_node(s, node);
4517 if (flags & SO_TOTAL)
4518 x = atomic_long_read(&n->total_objects);
4519 else if (flags & SO_OBJECTS)
4520 x = atomic_long_read(&n->total_objects) -
4521 count_partial(n, count_free);
4524 x = atomic_long_read(&n->nr_slabs);
4531 if (flags & SO_PARTIAL) {
4532 for_each_node_state(node, N_NORMAL_MEMORY) {
4533 struct kmem_cache_node *n = get_node(s, node);
4535 if (flags & SO_TOTAL)
4536 x = count_partial(n, count_total);
4537 else if (flags & SO_OBJECTS)
4538 x = count_partial(n, count_inuse);
4545 x = sprintf(buf, "%lu", total);
4547 for_each_node_state(node, N_NORMAL_MEMORY)
4549 x += sprintf(buf + x, " N%d=%lu",
4552 unlock_memory_hotplug();
4554 return x + sprintf(buf + x, "\n");
4557 #ifdef CONFIG_SLUB_DEBUG
4558 static int any_slab_objects(struct kmem_cache *s)
4562 for_each_online_node(node) {
4563 struct kmem_cache_node *n = get_node(s, node);
4568 if (atomic_long_read(&n->total_objects))
4575 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4576 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4578 struct slab_attribute {
4579 struct attribute attr;
4580 ssize_t (*show)(struct kmem_cache *s, char *buf);
4581 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4584 #define SLAB_ATTR_RO(_name) \
4585 static struct slab_attribute _name##_attr = \
4586 __ATTR(_name, 0400, _name##_show, NULL)
4588 #define SLAB_ATTR(_name) \
4589 static struct slab_attribute _name##_attr = \
4590 __ATTR(_name, 0600, _name##_show, _name##_store)
4592 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4594 return sprintf(buf, "%d\n", s->size);
4596 SLAB_ATTR_RO(slab_size);
4598 static ssize_t align_show(struct kmem_cache *s, char *buf)
4600 return sprintf(buf, "%d\n", s->align);
4602 SLAB_ATTR_RO(align);
4604 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4606 return sprintf(buf, "%d\n", s->object_size);
4608 SLAB_ATTR_RO(object_size);
4610 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4612 return sprintf(buf, "%d\n", oo_objects(s->oo));
4614 SLAB_ATTR_RO(objs_per_slab);
4616 static ssize_t order_store(struct kmem_cache *s,
4617 const char *buf, size_t length)
4619 unsigned long order;
4622 err = strict_strtoul(buf, 10, &order);
4626 if (order > slub_max_order || order < slub_min_order)
4629 calculate_sizes(s, order);
4633 static ssize_t order_show(struct kmem_cache *s, char *buf)
4635 return sprintf(buf, "%d\n", oo_order(s->oo));
4639 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4641 return sprintf(buf, "%lu\n", s->min_partial);
4644 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4650 err = strict_strtoul(buf, 10, &min);
4654 set_min_partial(s, min);
4657 SLAB_ATTR(min_partial);
4659 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4661 return sprintf(buf, "%u\n", s->cpu_partial);
4664 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4667 unsigned long objects;
4670 err = strict_strtoul(buf, 10, &objects);
4673 if (objects && kmem_cache_debug(s))
4676 s->cpu_partial = objects;
4680 SLAB_ATTR(cpu_partial);
4682 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4686 return sprintf(buf, "%pS\n", s->ctor);
4690 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4692 return sprintf(buf, "%d\n", s->refcount - 1);
4694 SLAB_ATTR_RO(aliases);
4696 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4698 return show_slab_objects(s, buf, SO_PARTIAL);
4700 SLAB_ATTR_RO(partial);
4702 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4704 return show_slab_objects(s, buf, SO_CPU);
4706 SLAB_ATTR_RO(cpu_slabs);
4708 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4710 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4712 SLAB_ATTR_RO(objects);
4714 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4716 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4718 SLAB_ATTR_RO(objects_partial);
4720 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4727 for_each_online_cpu(cpu) {
4728 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4731 pages += page->pages;
4732 objects += page->pobjects;
4736 len = sprintf(buf, "%d(%d)", objects, pages);
4739 for_each_online_cpu(cpu) {
4740 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4742 if (page && len < PAGE_SIZE - 20)
4743 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4744 page->pobjects, page->pages);
4747 return len + sprintf(buf + len, "\n");
4749 SLAB_ATTR_RO(slabs_cpu_partial);
4751 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4753 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4756 static ssize_t reclaim_account_store(struct kmem_cache *s,
4757 const char *buf, size_t length)
4759 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4761 s->flags |= SLAB_RECLAIM_ACCOUNT;
4764 SLAB_ATTR(reclaim_account);
4766 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4768 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4770 SLAB_ATTR_RO(hwcache_align);
4772 #ifdef CONFIG_ZONE_DMA
4773 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4775 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4777 SLAB_ATTR_RO(cache_dma);
4780 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4782 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4784 SLAB_ATTR_RO(destroy_by_rcu);
4786 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4788 return sprintf(buf, "%d\n", s->reserved);
4790 SLAB_ATTR_RO(reserved);
4792 #ifdef CONFIG_SLUB_DEBUG
4793 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4795 return show_slab_objects(s, buf, SO_ALL);
4797 SLAB_ATTR_RO(slabs);
4799 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4801 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4803 SLAB_ATTR_RO(total_objects);
4805 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4807 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4810 static ssize_t sanity_checks_store(struct kmem_cache *s,
4811 const char *buf, size_t length)
4813 s->flags &= ~SLAB_DEBUG_FREE;
4814 if (buf[0] == '1') {
4815 s->flags &= ~__CMPXCHG_DOUBLE;
4816 s->flags |= SLAB_DEBUG_FREE;
4820 SLAB_ATTR(sanity_checks);
4822 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4824 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4827 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4830 s->flags &= ~SLAB_TRACE;
4831 if (buf[0] == '1') {
4832 s->flags &= ~__CMPXCHG_DOUBLE;
4833 s->flags |= SLAB_TRACE;
4839 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4841 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4844 static ssize_t red_zone_store(struct kmem_cache *s,
4845 const char *buf, size_t length)
4847 if (any_slab_objects(s))
4850 s->flags &= ~SLAB_RED_ZONE;
4851 if (buf[0] == '1') {
4852 s->flags &= ~__CMPXCHG_DOUBLE;
4853 s->flags |= SLAB_RED_ZONE;
4855 calculate_sizes(s, -1);
4858 SLAB_ATTR(red_zone);
4860 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4862 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4865 static ssize_t poison_store(struct kmem_cache *s,
4866 const char *buf, size_t length)
4868 if (any_slab_objects(s))
4871 s->flags &= ~SLAB_POISON;
4872 if (buf[0] == '1') {
4873 s->flags &= ~__CMPXCHG_DOUBLE;
4874 s->flags |= SLAB_POISON;
4876 calculate_sizes(s, -1);
4881 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4883 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4886 static ssize_t store_user_store(struct kmem_cache *s,
4887 const char *buf, size_t length)
4889 if (any_slab_objects(s))
4892 s->flags &= ~SLAB_STORE_USER;
4893 if (buf[0] == '1') {
4894 s->flags &= ~__CMPXCHG_DOUBLE;
4895 s->flags |= SLAB_STORE_USER;
4897 calculate_sizes(s, -1);
4900 SLAB_ATTR(store_user);
4902 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4907 static ssize_t validate_store(struct kmem_cache *s,
4908 const char *buf, size_t length)
4912 if (buf[0] == '1') {
4913 ret = validate_slab_cache(s);
4919 SLAB_ATTR(validate);
4921 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4923 if (!(s->flags & SLAB_STORE_USER))
4925 return list_locations(s, buf, TRACK_ALLOC);
4927 SLAB_ATTR_RO(alloc_calls);
4929 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4931 if (!(s->flags & SLAB_STORE_USER))
4933 return list_locations(s, buf, TRACK_FREE);
4935 SLAB_ATTR_RO(free_calls);
4936 #endif /* CONFIG_SLUB_DEBUG */
4938 #ifdef CONFIG_FAILSLAB
4939 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4941 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4944 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4947 s->flags &= ~SLAB_FAILSLAB;
4949 s->flags |= SLAB_FAILSLAB;
4952 SLAB_ATTR(failslab);
4955 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4960 static ssize_t shrink_store(struct kmem_cache *s,
4961 const char *buf, size_t length)
4963 if (buf[0] == '1') {
4964 int rc = kmem_cache_shrink(s);
4975 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4977 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4980 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4981 const char *buf, size_t length)
4983 unsigned long ratio;
4986 err = strict_strtoul(buf, 10, &ratio);
4991 s->remote_node_defrag_ratio = ratio * 10;
4995 SLAB_ATTR(remote_node_defrag_ratio);
4998 #ifdef CONFIG_SLUB_STATS
4999 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5001 unsigned long sum = 0;
5004 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5009 for_each_online_cpu(cpu) {
5010 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5016 len = sprintf(buf, "%lu", sum);
5019 for_each_online_cpu(cpu) {
5020 if (data[cpu] && len < PAGE_SIZE - 20)
5021 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5025 return len + sprintf(buf + len, "\n");
5028 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5032 for_each_online_cpu(cpu)
5033 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5036 #define STAT_ATTR(si, text) \
5037 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5039 return show_stat(s, buf, si); \
5041 static ssize_t text##_store(struct kmem_cache *s, \
5042 const char *buf, size_t length) \
5044 if (buf[0] != '0') \
5046 clear_stat(s, si); \
5051 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5052 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5053 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5054 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5055 STAT_ATTR(FREE_FROZEN, free_frozen);
5056 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5057 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5058 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5059 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5060 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5061 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5062 STAT_ATTR(FREE_SLAB, free_slab);
5063 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5064 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5065 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5066 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5067 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5068 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5069 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5070 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5071 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5072 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5073 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5074 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5075 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5076 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5079 static struct attribute *slab_attrs[] = {
5080 &slab_size_attr.attr,
5081 &object_size_attr.attr,
5082 &objs_per_slab_attr.attr,
5084 &min_partial_attr.attr,
5085 &cpu_partial_attr.attr,
5087 &objects_partial_attr.attr,
5089 &cpu_slabs_attr.attr,
5093 &hwcache_align_attr.attr,
5094 &reclaim_account_attr.attr,
5095 &destroy_by_rcu_attr.attr,
5097 &reserved_attr.attr,
5098 &slabs_cpu_partial_attr.attr,
5099 #ifdef CONFIG_SLUB_DEBUG
5100 &total_objects_attr.attr,
5102 &sanity_checks_attr.attr,
5104 &red_zone_attr.attr,
5106 &store_user_attr.attr,
5107 &validate_attr.attr,
5108 &alloc_calls_attr.attr,
5109 &free_calls_attr.attr,
5111 #ifdef CONFIG_ZONE_DMA
5112 &cache_dma_attr.attr,
5115 &remote_node_defrag_ratio_attr.attr,
5117 #ifdef CONFIG_SLUB_STATS
5118 &alloc_fastpath_attr.attr,
5119 &alloc_slowpath_attr.attr,
5120 &free_fastpath_attr.attr,
5121 &free_slowpath_attr.attr,
5122 &free_frozen_attr.attr,
5123 &free_add_partial_attr.attr,
5124 &free_remove_partial_attr.attr,
5125 &alloc_from_partial_attr.attr,
5126 &alloc_slab_attr.attr,
5127 &alloc_refill_attr.attr,
5128 &alloc_node_mismatch_attr.attr,
5129 &free_slab_attr.attr,
5130 &cpuslab_flush_attr.attr,
5131 &deactivate_full_attr.attr,
5132 &deactivate_empty_attr.attr,
5133 &deactivate_to_head_attr.attr,
5134 &deactivate_to_tail_attr.attr,
5135 &deactivate_remote_frees_attr.attr,
5136 &deactivate_bypass_attr.attr,
5137 &order_fallback_attr.attr,
5138 &cmpxchg_double_fail_attr.attr,
5139 &cmpxchg_double_cpu_fail_attr.attr,
5140 &cpu_partial_alloc_attr.attr,
5141 &cpu_partial_free_attr.attr,
5142 &cpu_partial_node_attr.attr,
5143 &cpu_partial_drain_attr.attr,
5145 #ifdef CONFIG_FAILSLAB
5146 &failslab_attr.attr,
5152 static struct attribute_group slab_attr_group = {
5153 .attrs = slab_attrs,
5156 static ssize_t slab_attr_show(struct kobject *kobj,
5157 struct attribute *attr,
5160 struct slab_attribute *attribute;
5161 struct kmem_cache *s;
5164 attribute = to_slab_attr(attr);
5167 if (!attribute->show)
5170 err = attribute->show(s, buf);
5175 static ssize_t slab_attr_store(struct kobject *kobj,
5176 struct attribute *attr,
5177 const char *buf, size_t len)
5179 struct slab_attribute *attribute;
5180 struct kmem_cache *s;
5183 attribute = to_slab_attr(attr);
5186 if (!attribute->store)
5189 err = attribute->store(s, buf, len);
5194 static const struct sysfs_ops slab_sysfs_ops = {
5195 .show = slab_attr_show,
5196 .store = slab_attr_store,
5199 static struct kobj_type slab_ktype = {
5200 .sysfs_ops = &slab_sysfs_ops,
5203 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5205 struct kobj_type *ktype = get_ktype(kobj);
5207 if (ktype == &slab_ktype)
5212 static const struct kset_uevent_ops slab_uevent_ops = {
5213 .filter = uevent_filter,
5216 static struct kset *slab_kset;
5218 #define ID_STR_LENGTH 64
5220 /* Create a unique string id for a slab cache:
5222 * Format :[flags-]size
5224 static char *create_unique_id(struct kmem_cache *s)
5226 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5233 * First flags affecting slabcache operations. We will only
5234 * get here for aliasable slabs so we do not need to support
5235 * too many flags. The flags here must cover all flags that
5236 * are matched during merging to guarantee that the id is
5239 if (s->flags & SLAB_CACHE_DMA)
5241 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5243 if (s->flags & SLAB_DEBUG_FREE)
5245 if (!(s->flags & SLAB_NOTRACK))
5249 p += sprintf(p, "%07d", s->size);
5250 BUG_ON(p > name + ID_STR_LENGTH - 1);
5254 static int sysfs_slab_add(struct kmem_cache *s)
5260 if (slab_state < FULL)
5261 /* Defer until later */
5264 unmergeable = slab_unmergeable(s);
5267 * Slabcache can never be merged so we can use the name proper.
5268 * This is typically the case for debug situations. In that
5269 * case we can catch duplicate names easily.
5271 sysfs_remove_link(&slab_kset->kobj, s->name);
5275 * Create a unique name for the slab as a target
5278 name = create_unique_id(s);
5281 s->kobj.kset = slab_kset;
5282 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5284 kobject_put(&s->kobj);
5288 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5290 kobject_del(&s->kobj);
5291 kobject_put(&s->kobj);
5294 kobject_uevent(&s->kobj, KOBJ_ADD);
5296 /* Setup first alias */
5297 sysfs_slab_alias(s, s->name);
5303 static void sysfs_slab_remove(struct kmem_cache *s)
5305 if (slab_state < FULL)
5307 * Sysfs has not been setup yet so no need to remove the
5312 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5313 kobject_del(&s->kobj);
5314 kobject_put(&s->kobj);
5318 * Need to buffer aliases during bootup until sysfs becomes
5319 * available lest we lose that information.
5321 struct saved_alias {
5322 struct kmem_cache *s;
5324 struct saved_alias *next;
5327 static struct saved_alias *alias_list;
5329 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5331 struct saved_alias *al;
5333 if (slab_state == FULL) {
5335 * If we have a leftover link then remove it.
5337 sysfs_remove_link(&slab_kset->kobj, name);
5338 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5341 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5347 al->next = alias_list;
5352 static int __init slab_sysfs_init(void)
5354 struct kmem_cache *s;
5357 mutex_lock(&slab_mutex);
5359 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5361 mutex_unlock(&slab_mutex);
5362 printk(KERN_ERR "Cannot register slab subsystem.\n");
5368 list_for_each_entry(s, &slab_caches, list) {
5369 err = sysfs_slab_add(s);
5371 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5372 " to sysfs\n", s->name);
5375 while (alias_list) {
5376 struct saved_alias *al = alias_list;
5378 alias_list = alias_list->next;
5379 err = sysfs_slab_alias(al->s, al->name);
5381 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5382 " %s to sysfs\n", al->name);
5386 mutex_unlock(&slab_mutex);
5391 __initcall(slab_sysfs_init);
5392 #endif /* CONFIG_SYSFS */
5395 * The /proc/slabinfo ABI
5397 #ifdef CONFIG_SLABINFO
5398 static void print_slabinfo_header(struct seq_file *m)
5400 seq_puts(m, "slabinfo - version: 2.1\n");
5401 seq_puts(m, "# name <active_objs> <num_objs> <object_size> "
5402 "<objperslab> <pagesperslab>");
5403 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5404 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5408 static void *s_start(struct seq_file *m, loff_t *pos)
5412 mutex_lock(&slab_mutex);
5414 print_slabinfo_header(m);
5416 return seq_list_start(&slab_caches, *pos);
5419 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5421 return seq_list_next(p, &slab_caches, pos);
5424 static void s_stop(struct seq_file *m, void *p)
5426 mutex_unlock(&slab_mutex);
5429 static int s_show(struct seq_file *m, void *p)
5431 unsigned long nr_partials = 0;
5432 unsigned long nr_slabs = 0;
5433 unsigned long nr_inuse = 0;
5434 unsigned long nr_objs = 0;
5435 unsigned long nr_free = 0;
5436 struct kmem_cache *s;
5439 s = list_entry(p, struct kmem_cache, list);
5441 for_each_online_node(node) {
5442 struct kmem_cache_node *n = get_node(s, node);
5447 nr_partials += n->nr_partial;
5448 nr_slabs += atomic_long_read(&n->nr_slabs);
5449 nr_objs += atomic_long_read(&n->total_objects);
5450 nr_free += count_partial(n, count_free);
5453 nr_inuse = nr_objs - nr_free;
5455 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5456 nr_objs, s->size, oo_objects(s->oo),
5457 (1 << oo_order(s->oo)));
5458 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5459 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5465 static const struct seq_operations slabinfo_op = {
5472 static int slabinfo_open(struct inode *inode, struct file *file)
5474 return seq_open(file, &slabinfo_op);
5477 static const struct file_operations proc_slabinfo_operations = {
5478 .open = slabinfo_open,
5480 .llseek = seq_lseek,
5481 .release = seq_release,
5484 static int __init slab_proc_init(void)
5486 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5489 module_init(slab_proc_init);
5490 #endif /* CONFIG_SLABINFO */