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)
221 static inline void stat(const struct kmem_cache *s, enum stat_item si)
223 #ifdef CONFIG_SLUB_STATS
224 __this_cpu_inc(s->cpu_slab->stat[si]);
228 /********************************************************************
229 * Core slab cache functions
230 *******************************************************************/
232 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
234 return s->node[node];
237 /* Verify that a pointer has an address that is valid within a slab page */
238 static inline int check_valid_pointer(struct kmem_cache *s,
239 struct page *page, const void *object)
246 base = page_address(page);
247 if (object < base || object >= base + page->objects * s->size ||
248 (object - base) % s->size) {
255 static inline void *get_freepointer(struct kmem_cache *s, void *object)
257 return *(void **)(object + s->offset);
260 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
262 prefetch(object + s->offset);
265 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
269 #ifdef CONFIG_DEBUG_PAGEALLOC
270 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
272 p = get_freepointer(s, object);
277 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
279 *(void **)(object + s->offset) = fp;
282 /* Loop over all objects in a slab */
283 #define for_each_object(__p, __s, __addr, __objects) \
284 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
287 /* Determine object index from a given position */
288 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
290 return (p - addr) / s->size;
293 static inline size_t slab_ksize(const struct kmem_cache *s)
295 #ifdef CONFIG_SLUB_DEBUG
297 * Debugging requires use of the padding between object
298 * and whatever may come after it.
300 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
301 return s->object_size;
305 * If we have the need to store the freelist pointer
306 * back there or track user information then we can
307 * only use the space before that information.
309 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
312 * Else we can use all the padding etc for the allocation
317 static inline int order_objects(int order, unsigned long size, int reserved)
319 return ((PAGE_SIZE << order) - reserved) / size;
322 static inline struct kmem_cache_order_objects oo_make(int order,
323 unsigned long size, int reserved)
325 struct kmem_cache_order_objects x = {
326 (order << OO_SHIFT) + order_objects(order, size, reserved)
332 static inline int oo_order(struct kmem_cache_order_objects x)
334 return x.x >> OO_SHIFT;
337 static inline int oo_objects(struct kmem_cache_order_objects x)
339 return x.x & OO_MASK;
343 * Per slab locking using the pagelock
345 static __always_inline void slab_lock(struct page *page)
347 bit_spin_lock(PG_locked, &page->flags);
350 static __always_inline void slab_unlock(struct page *page)
352 __bit_spin_unlock(PG_locked, &page->flags);
355 /* Interrupts must be disabled (for the fallback code to work right) */
356 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
357 void *freelist_old, unsigned long counters_old,
358 void *freelist_new, unsigned long counters_new,
361 VM_BUG_ON(!irqs_disabled());
362 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
363 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
364 if (s->flags & __CMPXCHG_DOUBLE) {
365 if (cmpxchg_double(&page->freelist, &page->counters,
366 freelist_old, counters_old,
367 freelist_new, counters_new))
373 if (page->freelist == freelist_old && page->counters == counters_old) {
374 page->freelist = freelist_new;
375 page->counters = counters_new;
383 stat(s, CMPXCHG_DOUBLE_FAIL);
385 #ifdef SLUB_DEBUG_CMPXCHG
386 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
392 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
393 void *freelist_old, unsigned long counters_old,
394 void *freelist_new, unsigned long counters_new,
397 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
398 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
399 if (s->flags & __CMPXCHG_DOUBLE) {
400 if (cmpxchg_double(&page->freelist, &page->counters,
401 freelist_old, counters_old,
402 freelist_new, counters_new))
409 local_irq_save(flags);
411 if (page->freelist == freelist_old && page->counters == counters_old) {
412 page->freelist = freelist_new;
413 page->counters = counters_new;
415 local_irq_restore(flags);
419 local_irq_restore(flags);
423 stat(s, CMPXCHG_DOUBLE_FAIL);
425 #ifdef SLUB_DEBUG_CMPXCHG
426 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432 #ifdef CONFIG_SLUB_DEBUG
434 * Determine a map of object in use on a page.
436 * Node listlock must be held to guarantee that the page does
437 * not vanish from under us.
439 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
442 void *addr = page_address(page);
444 for (p = page->freelist; p; p = get_freepointer(s, p))
445 set_bit(slab_index(p, s, addr), map);
451 #ifdef CONFIG_SLUB_DEBUG_ON
452 static int slub_debug = DEBUG_DEFAULT_FLAGS;
454 static int slub_debug;
457 static char *slub_debug_slabs;
458 static int disable_higher_order_debug;
463 static void print_section(char *text, u8 *addr, unsigned int length)
465 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
469 static struct track *get_track(struct kmem_cache *s, void *object,
470 enum track_item alloc)
475 p = object + s->offset + sizeof(void *);
477 p = object + s->inuse;
482 static void set_track(struct kmem_cache *s, void *object,
483 enum track_item alloc, unsigned long addr)
485 struct track *p = get_track(s, object, alloc);
488 #ifdef CONFIG_STACKTRACE
489 struct stack_trace trace;
492 trace.nr_entries = 0;
493 trace.max_entries = TRACK_ADDRS_COUNT;
494 trace.entries = p->addrs;
496 save_stack_trace(&trace);
498 /* See rant in lockdep.c */
499 if (trace.nr_entries != 0 &&
500 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
503 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
507 p->cpu = smp_processor_id();
508 p->pid = current->pid;
511 memset(p, 0, sizeof(struct track));
514 static void init_tracking(struct kmem_cache *s, void *object)
516 if (!(s->flags & SLAB_STORE_USER))
519 set_track(s, object, TRACK_FREE, 0UL);
520 set_track(s, object, TRACK_ALLOC, 0UL);
523 static void print_track(const char *s, struct track *t)
528 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
529 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
530 #ifdef CONFIG_STACKTRACE
533 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
535 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
542 static void print_tracking(struct kmem_cache *s, void *object)
544 if (!(s->flags & SLAB_STORE_USER))
547 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
548 print_track("Freed", get_track(s, object, TRACK_FREE));
551 static void print_page_info(struct page *page)
553 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
554 page, page->objects, page->inuse, page->freelist, page->flags);
558 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
564 vsnprintf(buf, sizeof(buf), fmt, args);
566 printk(KERN_ERR "========================================"
567 "=====================================\n");
568 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
569 printk(KERN_ERR "----------------------------------------"
570 "-------------------------------------\n\n");
573 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
579 vsnprintf(buf, sizeof(buf), fmt, args);
581 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
584 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
586 unsigned int off; /* Offset of last byte */
587 u8 *addr = page_address(page);
589 print_tracking(s, p);
591 print_page_info(page);
593 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
594 p, p - addr, get_freepointer(s, p));
597 print_section("Bytes b4 ", p - 16, 16);
599 print_section("Object ", p, min_t(unsigned long, s->object_size,
601 if (s->flags & SLAB_RED_ZONE)
602 print_section("Redzone ", p + s->object_size,
603 s->inuse - s->object_size);
606 off = s->offset + sizeof(void *);
610 if (s->flags & SLAB_STORE_USER)
611 off += 2 * sizeof(struct track);
614 /* Beginning of the filler is the free pointer */
615 print_section("Padding ", p + off, s->size - off);
620 static void object_err(struct kmem_cache *s, struct page *page,
621 u8 *object, char *reason)
623 slab_bug(s, "%s", reason);
624 print_trailer(s, page, object);
627 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
633 vsnprintf(buf, sizeof(buf), fmt, args);
635 slab_bug(s, "%s", buf);
636 print_page_info(page);
640 static void init_object(struct kmem_cache *s, void *object, u8 val)
644 if (s->flags & __OBJECT_POISON) {
645 memset(p, POISON_FREE, s->object_size - 1);
646 p[s->object_size - 1] = POISON_END;
649 if (s->flags & SLAB_RED_ZONE)
650 memset(p + s->object_size, val, s->inuse - s->object_size);
653 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
654 void *from, void *to)
656 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
657 memset(from, data, to - from);
660 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
661 u8 *object, char *what,
662 u8 *start, unsigned int value, unsigned int bytes)
667 fault = memchr_inv(start, value, bytes);
672 while (end > fault && end[-1] == value)
675 slab_bug(s, "%s overwritten", what);
676 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
677 fault, end - 1, fault[0], value);
678 print_trailer(s, page, object);
680 restore_bytes(s, what, value, fault, end);
688 * Bytes of the object to be managed.
689 * If the freepointer may overlay the object then the free
690 * pointer is the first word of the object.
692 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
695 * object + s->object_size
696 * Padding to reach word boundary. This is also used for Redzoning.
697 * Padding is extended by another word if Redzoning is enabled and
698 * object_size == inuse.
700 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
701 * 0xcc (RED_ACTIVE) for objects in use.
704 * Meta data starts here.
706 * A. Free pointer (if we cannot overwrite object on free)
707 * B. Tracking data for SLAB_STORE_USER
708 * C. Padding to reach required alignment boundary or at mininum
709 * one word if debugging is on to be able to detect writes
710 * before the word boundary.
712 * Padding is done using 0x5a (POISON_INUSE)
715 * Nothing is used beyond s->size.
717 * If slabcaches are merged then the object_size and inuse boundaries are mostly
718 * ignored. And therefore no slab options that rely on these boundaries
719 * may be used with merged slabcaches.
722 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
724 unsigned long off = s->inuse; /* The end of info */
727 /* Freepointer is placed after the object. */
728 off += sizeof(void *);
730 if (s->flags & SLAB_STORE_USER)
731 /* We also have user information there */
732 off += 2 * sizeof(struct track);
737 return check_bytes_and_report(s, page, p, "Object padding",
738 p + off, POISON_INUSE, s->size - off);
741 /* Check the pad bytes at the end of a slab page */
742 static int slab_pad_check(struct kmem_cache *s, struct page *page)
750 if (!(s->flags & SLAB_POISON))
753 start = page_address(page);
754 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
755 end = start + length;
756 remainder = length % s->size;
760 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
763 while (end > fault && end[-1] == POISON_INUSE)
766 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
767 print_section("Padding ", end - remainder, remainder);
769 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
773 static int check_object(struct kmem_cache *s, struct page *page,
774 void *object, u8 val)
777 u8 *endobject = object + s->object_size;
779 if (s->flags & SLAB_RED_ZONE) {
780 if (!check_bytes_and_report(s, page, object, "Redzone",
781 endobject, val, s->inuse - s->object_size))
784 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
785 check_bytes_and_report(s, page, p, "Alignment padding",
786 endobject, POISON_INUSE, s->inuse - s->object_size);
790 if (s->flags & SLAB_POISON) {
791 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
792 (!check_bytes_and_report(s, page, p, "Poison", p,
793 POISON_FREE, s->object_size - 1) ||
794 !check_bytes_and_report(s, page, p, "Poison",
795 p + s->object_size - 1, POISON_END, 1)))
798 * check_pad_bytes cleans up on its own.
800 check_pad_bytes(s, page, p);
803 if (!s->offset && val == SLUB_RED_ACTIVE)
805 * Object and freepointer overlap. Cannot check
806 * freepointer while object is allocated.
810 /* Check free pointer validity */
811 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
812 object_err(s, page, p, "Freepointer corrupt");
814 * No choice but to zap it and thus lose the remainder
815 * of the free objects in this slab. May cause
816 * another error because the object count is now wrong.
818 set_freepointer(s, p, NULL);
824 static int check_slab(struct kmem_cache *s, struct page *page)
828 VM_BUG_ON(!irqs_disabled());
830 if (!PageSlab(page)) {
831 slab_err(s, page, "Not a valid slab page");
835 maxobj = order_objects(compound_order(page), s->size, s->reserved);
836 if (page->objects > maxobj) {
837 slab_err(s, page, "objects %u > max %u",
838 s->name, page->objects, maxobj);
841 if (page->inuse > page->objects) {
842 slab_err(s, page, "inuse %u > max %u",
843 s->name, page->inuse, page->objects);
846 /* Slab_pad_check fixes things up after itself */
847 slab_pad_check(s, page);
852 * Determine if a certain object on a page is on the freelist. Must hold the
853 * slab lock to guarantee that the chains are in a consistent state.
855 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
860 unsigned long max_objects;
863 while (fp && nr <= page->objects) {
866 if (!check_valid_pointer(s, page, fp)) {
868 object_err(s, page, object,
869 "Freechain corrupt");
870 set_freepointer(s, object, NULL);
873 slab_err(s, page, "Freepointer corrupt");
874 page->freelist = NULL;
875 page->inuse = page->objects;
876 slab_fix(s, "Freelist cleared");
882 fp = get_freepointer(s, object);
886 max_objects = order_objects(compound_order(page), s->size, s->reserved);
887 if (max_objects > MAX_OBJS_PER_PAGE)
888 max_objects = MAX_OBJS_PER_PAGE;
890 if (page->objects != max_objects) {
891 slab_err(s, page, "Wrong number of objects. Found %d but "
892 "should be %d", page->objects, max_objects);
893 page->objects = max_objects;
894 slab_fix(s, "Number of objects adjusted.");
896 if (page->inuse != page->objects - nr) {
897 slab_err(s, page, "Wrong object count. Counter is %d but "
898 "counted were %d", page->inuse, page->objects - nr);
899 page->inuse = page->objects - nr;
900 slab_fix(s, "Object count adjusted.");
902 return search == NULL;
905 static void trace(struct kmem_cache *s, struct page *page, void *object,
908 if (s->flags & SLAB_TRACE) {
909 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
911 alloc ? "alloc" : "free",
916 print_section("Object ", (void *)object, s->object_size);
923 * Hooks for other subsystems that check memory allocations. In a typical
924 * production configuration these hooks all should produce no code at all.
926 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
928 flags &= gfp_allowed_mask;
929 lockdep_trace_alloc(flags);
930 might_sleep_if(flags & __GFP_WAIT);
932 return should_failslab(s->object_size, flags, s->flags);
935 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
937 flags &= gfp_allowed_mask;
938 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
939 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
942 static inline void slab_free_hook(struct kmem_cache *s, void *x)
944 kmemleak_free_recursive(x, s->flags);
947 * Trouble is that we may no longer disable interupts in the fast path
948 * So in order to make the debug calls that expect irqs to be
949 * disabled we need to disable interrupts temporarily.
951 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
955 local_irq_save(flags);
956 kmemcheck_slab_free(s, x, s->object_size);
957 debug_check_no_locks_freed(x, s->object_size);
958 local_irq_restore(flags);
961 if (!(s->flags & SLAB_DEBUG_OBJECTS))
962 debug_check_no_obj_freed(x, s->object_size);
966 * Tracking of fully allocated slabs for debugging purposes.
968 * list_lock must be held.
970 static void add_full(struct kmem_cache *s,
971 struct kmem_cache_node *n, struct page *page)
973 if (!(s->flags & SLAB_STORE_USER))
976 list_add(&page->lru, &n->full);
980 * list_lock must be held.
982 static void remove_full(struct kmem_cache *s, struct page *page)
984 if (!(s->flags & SLAB_STORE_USER))
987 list_del(&page->lru);
990 /* Tracking of the number of slabs for debugging purposes */
991 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
993 struct kmem_cache_node *n = get_node(s, node);
995 return atomic_long_read(&n->nr_slabs);
998 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1000 return atomic_long_read(&n->nr_slabs);
1003 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1005 struct kmem_cache_node *n = get_node(s, node);
1008 * May be called early in order to allocate a slab for the
1009 * kmem_cache_node structure. Solve the chicken-egg
1010 * dilemma by deferring the increment of the count during
1011 * bootstrap (see early_kmem_cache_node_alloc).
1014 atomic_long_inc(&n->nr_slabs);
1015 atomic_long_add(objects, &n->total_objects);
1018 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1020 struct kmem_cache_node *n = get_node(s, node);
1022 atomic_long_dec(&n->nr_slabs);
1023 atomic_long_sub(objects, &n->total_objects);
1026 /* Object debug checks for alloc/free paths */
1027 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1030 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1033 init_object(s, object, SLUB_RED_INACTIVE);
1034 init_tracking(s, object);
1037 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1038 void *object, unsigned long addr)
1040 if (!check_slab(s, page))
1043 if (!check_valid_pointer(s, page, object)) {
1044 object_err(s, page, object, "Freelist Pointer check fails");
1048 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1051 /* Success perform special debug activities for allocs */
1052 if (s->flags & SLAB_STORE_USER)
1053 set_track(s, object, TRACK_ALLOC, addr);
1054 trace(s, page, object, 1);
1055 init_object(s, object, SLUB_RED_ACTIVE);
1059 if (PageSlab(page)) {
1061 * If this is a slab page then lets do the best we can
1062 * to avoid issues in the future. Marking all objects
1063 * as used avoids touching the remaining objects.
1065 slab_fix(s, "Marking all objects used");
1066 page->inuse = page->objects;
1067 page->freelist = NULL;
1072 static noinline int free_debug_processing(struct kmem_cache *s,
1073 struct page *page, void *object, unsigned long addr)
1075 unsigned long flags;
1078 local_irq_save(flags);
1081 if (!check_slab(s, page))
1084 if (!check_valid_pointer(s, page, object)) {
1085 slab_err(s, page, "Invalid object pointer 0x%p", object);
1089 if (on_freelist(s, page, object)) {
1090 object_err(s, page, object, "Object already free");
1094 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1097 if (unlikely(s != page->slab)) {
1098 if (!PageSlab(page)) {
1099 slab_err(s, page, "Attempt to free object(0x%p) "
1100 "outside of slab", object);
1101 } else if (!page->slab) {
1103 "SLUB <none>: no slab for object 0x%p.\n",
1107 object_err(s, page, object,
1108 "page slab pointer corrupt.");
1112 if (s->flags & SLAB_STORE_USER)
1113 set_track(s, object, TRACK_FREE, addr);
1114 trace(s, page, object, 0);
1115 init_object(s, object, SLUB_RED_INACTIVE);
1119 local_irq_restore(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 int free_debug_processing(struct kmem_cache *s,
1218 struct page *page, void *object, unsigned long addr) { return 0; }
1220 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1222 static inline int check_object(struct kmem_cache *s, struct page *page,
1223 void *object, u8 val) { return 1; }
1224 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1225 struct page *page) {}
1226 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1227 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1228 unsigned long flags, const char *name,
1229 void (*ctor)(void *))
1233 #define slub_debug 0
1235 #define disable_higher_order_debug 0
1237 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1239 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1241 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1243 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1246 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1249 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1252 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1254 #endif /* CONFIG_SLUB_DEBUG */
1257 * Slab allocation and freeing
1259 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1260 struct kmem_cache_order_objects oo)
1262 int order = oo_order(oo);
1264 flags |= __GFP_NOTRACK;
1266 if (node == NUMA_NO_NODE)
1267 return alloc_pages(flags, order);
1269 return alloc_pages_exact_node(node, flags, order);
1272 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1275 struct kmem_cache_order_objects oo = s->oo;
1278 flags &= gfp_allowed_mask;
1280 if (flags & __GFP_WAIT)
1283 flags |= s->allocflags;
1286 * Let the initial higher-order allocation fail under memory pressure
1287 * so we fall-back to the minimum order allocation.
1289 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1291 page = alloc_slab_page(alloc_gfp, node, oo);
1292 if (unlikely(!page)) {
1295 * Allocation may have failed due to fragmentation.
1296 * Try a lower order alloc if possible
1298 page = alloc_slab_page(flags, node, oo);
1301 stat(s, ORDER_FALLBACK);
1304 if (kmemcheck_enabled && page
1305 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1306 int pages = 1 << oo_order(oo);
1308 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1311 * Objects from caches that have a constructor don't get
1312 * cleared when they're allocated, so we need to do it here.
1315 kmemcheck_mark_uninitialized_pages(page, pages);
1317 kmemcheck_mark_unallocated_pages(page, pages);
1320 if (flags & __GFP_WAIT)
1321 local_irq_disable();
1325 page->objects = oo_objects(oo);
1326 mod_zone_page_state(page_zone(page),
1327 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1328 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1334 static void setup_object(struct kmem_cache *s, struct page *page,
1337 setup_object_debug(s, page, object);
1338 if (unlikely(s->ctor))
1342 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1349 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1351 page = allocate_slab(s,
1352 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1356 inc_slabs_node(s, page_to_nid(page), page->objects);
1358 __SetPageSlab(page);
1359 if (page->pfmemalloc)
1360 SetPageSlabPfmemalloc(page);
1362 start = page_address(page);
1364 if (unlikely(s->flags & SLAB_POISON))
1365 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1368 for_each_object(p, s, start, page->objects) {
1369 setup_object(s, page, last);
1370 set_freepointer(s, last, p);
1373 setup_object(s, page, last);
1374 set_freepointer(s, last, NULL);
1376 page->freelist = start;
1377 page->inuse = page->objects;
1383 static void __free_slab(struct kmem_cache *s, struct page *page)
1385 int order = compound_order(page);
1386 int pages = 1 << order;
1388 if (kmem_cache_debug(s)) {
1391 slab_pad_check(s, page);
1392 for_each_object(p, s, page_address(page),
1394 check_object(s, page, p, SLUB_RED_INACTIVE);
1397 kmemcheck_free_shadow(page, compound_order(page));
1399 mod_zone_page_state(page_zone(page),
1400 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1401 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1404 __ClearPageSlabPfmemalloc(page);
1405 __ClearPageSlab(page);
1406 reset_page_mapcount(page);
1407 if (current->reclaim_state)
1408 current->reclaim_state->reclaimed_slab += pages;
1409 __free_pages(page, order);
1412 #define need_reserve_slab_rcu \
1413 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1415 static void rcu_free_slab(struct rcu_head *h)
1419 if (need_reserve_slab_rcu)
1420 page = virt_to_head_page(h);
1422 page = container_of((struct list_head *)h, struct page, lru);
1424 __free_slab(page->slab, page);
1427 static void free_slab(struct kmem_cache *s, struct page *page)
1429 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1430 struct rcu_head *head;
1432 if (need_reserve_slab_rcu) {
1433 int order = compound_order(page);
1434 int offset = (PAGE_SIZE << order) - s->reserved;
1436 VM_BUG_ON(s->reserved != sizeof(*head));
1437 head = page_address(page) + offset;
1440 * RCU free overloads the RCU head over the LRU
1442 head = (void *)&page->lru;
1445 call_rcu(head, rcu_free_slab);
1447 __free_slab(s, page);
1450 static void discard_slab(struct kmem_cache *s, struct page *page)
1452 dec_slabs_node(s, page_to_nid(page), page->objects);
1457 * Management of partially allocated slabs.
1459 * list_lock must be held.
1461 static inline void add_partial(struct kmem_cache_node *n,
1462 struct page *page, int tail)
1465 if (tail == DEACTIVATE_TO_TAIL)
1466 list_add_tail(&page->lru, &n->partial);
1468 list_add(&page->lru, &n->partial);
1472 * list_lock must be held.
1474 static inline void remove_partial(struct kmem_cache_node *n,
1477 list_del(&page->lru);
1482 * Remove slab from the partial list, freeze it and
1483 * return the pointer to the freelist.
1485 * Returns a list of objects or NULL if it fails.
1487 * Must hold list_lock since we modify the partial list.
1489 static inline void *acquire_slab(struct kmem_cache *s,
1490 struct kmem_cache_node *n, struct page *page,
1494 unsigned long counters;
1498 * Zap the freelist and set the frozen bit.
1499 * The old freelist is the list of objects for the
1500 * per cpu allocation list.
1502 freelist = page->freelist;
1503 counters = page->counters;
1504 new.counters = counters;
1506 new.inuse = page->objects;
1507 new.freelist = NULL;
1509 new.freelist = freelist;
1512 VM_BUG_ON(new.frozen);
1515 if (!__cmpxchg_double_slab(s, page,
1517 new.freelist, new.counters,
1521 remove_partial(n, page);
1526 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1527 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1530 * Try to allocate a partial slab from a specific node.
1532 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1533 struct kmem_cache_cpu *c, gfp_t flags)
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) {
1552 if (!pfmemalloc_match(page, flags))
1555 t = acquire_slab(s, n, page, object == NULL);
1561 stat(s, ALLOC_FROM_PARTIAL);
1563 available = page->objects - page->inuse;
1565 available = put_cpu_partial(s, page, 0);
1566 stat(s, CPU_PARTIAL_NODE);
1568 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1572 spin_unlock(&n->list_lock);
1577 * Get a page from somewhere. Search in increasing NUMA distances.
1579 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1580 struct kmem_cache_cpu *c)
1583 struct zonelist *zonelist;
1586 enum zone_type high_zoneidx = gfp_zone(flags);
1588 unsigned int cpuset_mems_cookie;
1591 * The defrag ratio allows a configuration of the tradeoffs between
1592 * inter node defragmentation and node local allocations. A lower
1593 * defrag_ratio increases the tendency to do local allocations
1594 * instead of attempting to obtain partial slabs from other nodes.
1596 * If the defrag_ratio is set to 0 then kmalloc() always
1597 * returns node local objects. If the ratio is higher then kmalloc()
1598 * may return off node objects because partial slabs are obtained
1599 * from other nodes and filled up.
1601 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1602 * defrag_ratio = 1000) then every (well almost) allocation will
1603 * first attempt to defrag slab caches on other nodes. This means
1604 * scanning over all nodes to look for partial slabs which may be
1605 * expensive if we do it every time we are trying to find a slab
1606 * with available objects.
1608 if (!s->remote_node_defrag_ratio ||
1609 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1613 cpuset_mems_cookie = get_mems_allowed();
1614 zonelist = node_zonelist(slab_node(), flags);
1615 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1616 struct kmem_cache_node *n;
1618 n = get_node(s, zone_to_nid(zone));
1620 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1621 n->nr_partial > s->min_partial) {
1622 object = get_partial_node(s, n, c, flags);
1625 * Return the object even if
1626 * put_mems_allowed indicated that
1627 * the cpuset mems_allowed was
1628 * updated in parallel. It's a
1629 * harmless race between the alloc
1630 * and the cpuset update.
1632 put_mems_allowed(cpuset_mems_cookie);
1637 } while (!put_mems_allowed(cpuset_mems_cookie));
1643 * Get a partial page, lock it and return it.
1645 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1646 struct kmem_cache_cpu *c)
1649 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1651 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1652 if (object || node != NUMA_NO_NODE)
1655 return get_any_partial(s, flags, c);
1658 #ifdef CONFIG_PREEMPT
1660 * Calculate the next globally unique transaction for disambiguiation
1661 * during cmpxchg. The transactions start with the cpu number and are then
1662 * incremented by CONFIG_NR_CPUS.
1664 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1667 * No preemption supported therefore also no need to check for
1673 static inline unsigned long next_tid(unsigned long tid)
1675 return tid + TID_STEP;
1678 static inline unsigned int tid_to_cpu(unsigned long tid)
1680 return tid % TID_STEP;
1683 static inline unsigned long tid_to_event(unsigned long tid)
1685 return tid / TID_STEP;
1688 static inline unsigned int init_tid(int cpu)
1693 static inline void note_cmpxchg_failure(const char *n,
1694 const struct kmem_cache *s, unsigned long tid)
1696 #ifdef SLUB_DEBUG_CMPXCHG
1697 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1699 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1701 #ifdef CONFIG_PREEMPT
1702 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1703 printk("due to cpu change %d -> %d\n",
1704 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1707 if (tid_to_event(tid) != tid_to_event(actual_tid))
1708 printk("due to cpu running other code. Event %ld->%ld\n",
1709 tid_to_event(tid), tid_to_event(actual_tid));
1711 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1712 actual_tid, tid, next_tid(tid));
1714 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1717 void init_kmem_cache_cpus(struct kmem_cache *s)
1721 for_each_possible_cpu(cpu)
1722 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1726 * Remove the cpu slab
1728 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1730 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1731 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1733 enum slab_modes l = M_NONE, m = M_NONE;
1735 int tail = DEACTIVATE_TO_HEAD;
1739 if (page->freelist) {
1740 stat(s, DEACTIVATE_REMOTE_FREES);
1741 tail = DEACTIVATE_TO_TAIL;
1745 * Stage one: Free all available per cpu objects back
1746 * to the page freelist while it is still frozen. Leave the
1749 * There is no need to take the list->lock because the page
1752 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1754 unsigned long counters;
1757 prior = page->freelist;
1758 counters = page->counters;
1759 set_freepointer(s, freelist, prior);
1760 new.counters = counters;
1762 VM_BUG_ON(!new.frozen);
1764 } while (!__cmpxchg_double_slab(s, page,
1766 freelist, new.counters,
1767 "drain percpu freelist"));
1769 freelist = nextfree;
1773 * Stage two: Ensure that the page is unfrozen while the
1774 * list presence reflects the actual number of objects
1777 * We setup the list membership and then perform a cmpxchg
1778 * with the count. If there is a mismatch then the page
1779 * is not unfrozen but the page is on the wrong list.
1781 * Then we restart the process which may have to remove
1782 * the page from the list that we just put it on again
1783 * because the number of objects in the slab may have
1788 old.freelist = page->freelist;
1789 old.counters = page->counters;
1790 VM_BUG_ON(!old.frozen);
1792 /* Determine target state of the slab */
1793 new.counters = old.counters;
1796 set_freepointer(s, freelist, old.freelist);
1797 new.freelist = freelist;
1799 new.freelist = old.freelist;
1803 if (!new.inuse && n->nr_partial > s->min_partial)
1805 else if (new.freelist) {
1810 * Taking the spinlock removes the possiblity
1811 * that acquire_slab() will see a slab page that
1814 spin_lock(&n->list_lock);
1818 if (kmem_cache_debug(s) && !lock) {
1821 * This also ensures that the scanning of full
1822 * slabs from diagnostic functions will not see
1825 spin_lock(&n->list_lock);
1833 remove_partial(n, page);
1835 else if (l == M_FULL)
1837 remove_full(s, page);
1839 if (m == M_PARTIAL) {
1841 add_partial(n, page, tail);
1844 } else if (m == M_FULL) {
1846 stat(s, DEACTIVATE_FULL);
1847 add_full(s, n, page);
1853 if (!__cmpxchg_double_slab(s, page,
1854 old.freelist, old.counters,
1855 new.freelist, new.counters,
1860 spin_unlock(&n->list_lock);
1863 stat(s, DEACTIVATE_EMPTY);
1864 discard_slab(s, page);
1870 * Unfreeze all the cpu partial slabs.
1872 * This function must be called with interrupt disabled.
1874 static void unfreeze_partials(struct kmem_cache *s)
1876 struct kmem_cache_node *n = NULL, *n2 = NULL;
1877 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1878 struct page *page, *discard_page = NULL;
1880 while ((page = c->partial)) {
1884 c->partial = page->next;
1886 n2 = get_node(s, page_to_nid(page));
1889 spin_unlock(&n->list_lock);
1892 spin_lock(&n->list_lock);
1897 old.freelist = page->freelist;
1898 old.counters = page->counters;
1899 VM_BUG_ON(!old.frozen);
1901 new.counters = old.counters;
1902 new.freelist = old.freelist;
1906 } while (!__cmpxchg_double_slab(s, page,
1907 old.freelist, old.counters,
1908 new.freelist, new.counters,
1909 "unfreezing slab"));
1911 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1912 page->next = discard_page;
1913 discard_page = page;
1915 add_partial(n, page, DEACTIVATE_TO_TAIL);
1916 stat(s, FREE_ADD_PARTIAL);
1921 spin_unlock(&n->list_lock);
1923 while (discard_page) {
1924 page = discard_page;
1925 discard_page = discard_page->next;
1927 stat(s, DEACTIVATE_EMPTY);
1928 discard_slab(s, page);
1934 * Put a page that was just frozen (in __slab_free) into a partial page
1935 * slot if available. This is done without interrupts disabled and without
1936 * preemption disabled. The cmpxchg is racy and may put the partial page
1937 * onto a random cpus partial slot.
1939 * If we did not find a slot then simply move all the partials to the
1940 * per node partial list.
1942 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1944 struct page *oldpage;
1951 oldpage = this_cpu_read(s->cpu_slab->partial);
1954 pobjects = oldpage->pobjects;
1955 pages = oldpage->pages;
1956 if (drain && pobjects > s->cpu_partial) {
1957 unsigned long flags;
1959 * partial array is full. Move the existing
1960 * set to the per node partial list.
1962 local_irq_save(flags);
1963 unfreeze_partials(s);
1964 local_irq_restore(flags);
1967 stat(s, CPU_PARTIAL_DRAIN);
1972 pobjects += page->objects - page->inuse;
1974 page->pages = pages;
1975 page->pobjects = pobjects;
1976 page->next = oldpage;
1978 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1982 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1984 stat(s, CPUSLAB_FLUSH);
1985 deactivate_slab(s, c->page, c->freelist);
1987 c->tid = next_tid(c->tid);
1995 * Called from IPI handler with interrupts disabled.
1997 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1999 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2005 unfreeze_partials(s);
2009 static void flush_cpu_slab(void *d)
2011 struct kmem_cache *s = d;
2013 __flush_cpu_slab(s, smp_processor_id());
2016 static bool has_cpu_slab(int cpu, void *info)
2018 struct kmem_cache *s = info;
2019 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2021 return c->page || c->partial;
2024 static void flush_all(struct kmem_cache *s)
2026 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2030 * Check if the objects in a per cpu structure fit numa
2031 * locality expectations.
2033 static inline int node_match(struct page *page, int node)
2036 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2042 static int count_free(struct page *page)
2044 return page->objects - page->inuse;
2047 static unsigned long count_partial(struct kmem_cache_node *n,
2048 int (*get_count)(struct page *))
2050 unsigned long flags;
2051 unsigned long x = 0;
2054 spin_lock_irqsave(&n->list_lock, flags);
2055 list_for_each_entry(page, &n->partial, lru)
2056 x += get_count(page);
2057 spin_unlock_irqrestore(&n->list_lock, flags);
2061 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2063 #ifdef CONFIG_SLUB_DEBUG
2064 return atomic_long_read(&n->total_objects);
2070 static noinline void
2071 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2076 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2078 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2079 "default order: %d, min order: %d\n", s->name, s->object_size,
2080 s->size, oo_order(s->oo), oo_order(s->min));
2082 if (oo_order(s->min) > get_order(s->object_size))
2083 printk(KERN_WARNING " %s debugging increased min order, use "
2084 "slub_debug=O to disable.\n", s->name);
2086 for_each_online_node(node) {
2087 struct kmem_cache_node *n = get_node(s, node);
2088 unsigned long nr_slabs;
2089 unsigned long nr_objs;
2090 unsigned long nr_free;
2095 nr_free = count_partial(n, count_free);
2096 nr_slabs = node_nr_slabs(n);
2097 nr_objs = node_nr_objs(n);
2100 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2101 node, nr_slabs, nr_objs, nr_free);
2105 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2106 int node, struct kmem_cache_cpu **pc)
2109 struct kmem_cache_cpu *c = *pc;
2112 freelist = get_partial(s, flags, node, c);
2117 page = new_slab(s, flags, node);
2119 c = __this_cpu_ptr(s->cpu_slab);
2124 * No other reference to the page yet so we can
2125 * muck around with it freely without cmpxchg
2127 freelist = page->freelist;
2128 page->freelist = NULL;
2130 stat(s, ALLOC_SLAB);
2139 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2141 if (unlikely(PageSlabPfmemalloc(page)))
2142 return gfp_pfmemalloc_allowed(gfpflags);
2148 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2149 * or deactivate the page.
2151 * The page is still frozen if the return value is not NULL.
2153 * If this function returns NULL then the page has been unfrozen.
2155 * This function must be called with interrupt disabled.
2157 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2160 unsigned long counters;
2164 freelist = page->freelist;
2165 counters = page->counters;
2167 new.counters = counters;
2168 VM_BUG_ON(!new.frozen);
2170 new.inuse = page->objects;
2171 new.frozen = freelist != NULL;
2173 } while (!__cmpxchg_double_slab(s, page,
2182 * Slow path. The lockless freelist is empty or we need to perform
2185 * Processing is still very fast if new objects have been freed to the
2186 * regular freelist. In that case we simply take over the regular freelist
2187 * as the lockless freelist and zap the regular freelist.
2189 * If that is not working then we fall back to the partial lists. We take the
2190 * first element of the freelist as the object to allocate now and move the
2191 * rest of the freelist to the lockless freelist.
2193 * And if we were unable to get a new slab from the partial slab lists then
2194 * we need to allocate a new slab. This is the slowest path since it involves
2195 * a call to the page allocator and the setup of a new slab.
2197 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2198 unsigned long addr, struct kmem_cache_cpu *c)
2202 unsigned long flags;
2204 local_irq_save(flags);
2205 #ifdef CONFIG_PREEMPT
2207 * We may have been preempted and rescheduled on a different
2208 * cpu before disabling interrupts. Need to reload cpu area
2211 c = this_cpu_ptr(s->cpu_slab);
2219 if (unlikely(!node_match(page, node))) {
2220 stat(s, ALLOC_NODE_MISMATCH);
2221 deactivate_slab(s, page, c->freelist);
2228 * By rights, we should be searching for a slab page that was
2229 * PFMEMALLOC but right now, we are losing the pfmemalloc
2230 * information when the page leaves the per-cpu allocator
2232 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2233 deactivate_slab(s, page, c->freelist);
2239 /* must check again c->freelist in case of cpu migration or IRQ */
2240 freelist = c->freelist;
2244 stat(s, ALLOC_SLOWPATH);
2246 freelist = get_freelist(s, page);
2250 stat(s, DEACTIVATE_BYPASS);
2254 stat(s, ALLOC_REFILL);
2258 * freelist is pointing to the list of objects to be used.
2259 * page is pointing to the page from which the objects are obtained.
2260 * That page must be frozen for per cpu allocations to work.
2262 VM_BUG_ON(!c->page->frozen);
2263 c->freelist = get_freepointer(s, freelist);
2264 c->tid = next_tid(c->tid);
2265 local_irq_restore(flags);
2271 page = c->page = c->partial;
2272 c->partial = page->next;
2273 stat(s, CPU_PARTIAL_ALLOC);
2278 freelist = new_slab_objects(s, gfpflags, node, &c);
2280 if (unlikely(!freelist)) {
2281 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2282 slab_out_of_memory(s, gfpflags, node);
2284 local_irq_restore(flags);
2289 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2292 /* Only entered in the debug case */
2293 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2294 goto new_slab; /* Slab failed checks. Next slab needed */
2296 deactivate_slab(s, page, get_freepointer(s, freelist));
2299 local_irq_restore(flags);
2304 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2305 * have the fastpath folded into their functions. So no function call
2306 * overhead for requests that can be satisfied on the fastpath.
2308 * The fastpath works by first checking if the lockless freelist can be used.
2309 * If not then __slab_alloc is called for slow processing.
2311 * Otherwise we can simply pick the next object from the lockless free list.
2313 static __always_inline void *slab_alloc(struct kmem_cache *s,
2314 gfp_t gfpflags, int node, unsigned long addr)
2317 struct kmem_cache_cpu *c;
2321 if (slab_pre_alloc_hook(s, gfpflags))
2327 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2328 * enabled. We may switch back and forth between cpus while
2329 * reading from one cpu area. That does not matter as long
2330 * as we end up on the original cpu again when doing the cmpxchg.
2332 c = __this_cpu_ptr(s->cpu_slab);
2335 * The transaction ids are globally unique per cpu and per operation on
2336 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2337 * occurs on the right processor and that there was no operation on the
2338 * linked list in between.
2343 object = c->freelist;
2345 if (unlikely(!object || !node_match(page, node)))
2346 object = __slab_alloc(s, gfpflags, node, addr, c);
2349 void *next_object = get_freepointer_safe(s, object);
2352 * The cmpxchg will only match if there was no additional
2353 * operation and if we are on the right processor.
2355 * The cmpxchg does the following atomically (without lock semantics!)
2356 * 1. Relocate first pointer to the current per cpu area.
2357 * 2. Verify that tid and freelist have not been changed
2358 * 3. If they were not changed replace tid and freelist
2360 * Since this is without lock semantics the protection is only against
2361 * code executing on this cpu *not* from access by other cpus.
2363 if (unlikely(!this_cpu_cmpxchg_double(
2364 s->cpu_slab->freelist, s->cpu_slab->tid,
2366 next_object, next_tid(tid)))) {
2368 note_cmpxchg_failure("slab_alloc", s, tid);
2371 prefetch_freepointer(s, next_object);
2372 stat(s, ALLOC_FASTPATH);
2375 if (unlikely(gfpflags & __GFP_ZERO) && object)
2376 memset(object, 0, s->object_size);
2378 slab_post_alloc_hook(s, gfpflags, object);
2383 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2385 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2387 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2391 EXPORT_SYMBOL(kmem_cache_alloc);
2393 #ifdef CONFIG_TRACING
2394 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2396 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2397 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2400 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2402 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2404 void *ret = kmalloc_order(size, flags, order);
2405 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2408 EXPORT_SYMBOL(kmalloc_order_trace);
2412 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2414 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2416 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2417 s->object_size, s->size, gfpflags, node);
2421 EXPORT_SYMBOL(kmem_cache_alloc_node);
2423 #ifdef CONFIG_TRACING
2424 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2426 int node, size_t size)
2428 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2430 trace_kmalloc_node(_RET_IP_, ret,
2431 size, s->size, gfpflags, node);
2434 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2439 * Slow patch handling. This may still be called frequently since objects
2440 * have a longer lifetime than the cpu slabs in most processing loads.
2442 * So we still attempt to reduce cache line usage. Just take the slab
2443 * lock and free the item. If there is no additional partial page
2444 * handling required then we can return immediately.
2446 static void __slab_free(struct kmem_cache *s, struct page *page,
2447 void *x, unsigned long addr)
2450 void **object = (void *)x;
2454 unsigned long counters;
2455 struct kmem_cache_node *n = NULL;
2456 unsigned long uninitialized_var(flags);
2458 stat(s, FREE_SLOWPATH);
2460 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2464 prior = page->freelist;
2465 counters = page->counters;
2466 set_freepointer(s, object, prior);
2467 new.counters = counters;
2468 was_frozen = new.frozen;
2470 if ((!new.inuse || !prior) && !was_frozen && !n) {
2472 if (!kmem_cache_debug(s) && !prior)
2475 * Slab was on no list before and will be partially empty
2476 * We can defer the list move and instead freeze it.
2480 else { /* Needs to be taken off a list */
2482 n = get_node(s, page_to_nid(page));
2484 * Speculatively acquire the list_lock.
2485 * If the cmpxchg does not succeed then we may
2486 * drop the list_lock without any processing.
2488 * Otherwise the list_lock will synchronize with
2489 * other processors updating the list of slabs.
2491 spin_lock_irqsave(&n->list_lock, flags);
2497 } while (!cmpxchg_double_slab(s, page,
2499 object, new.counters,
2505 * If we just froze the page then put it onto the
2506 * per cpu partial list.
2508 if (new.frozen && !was_frozen) {
2509 put_cpu_partial(s, page, 1);
2510 stat(s, CPU_PARTIAL_FREE);
2513 * The list lock was not taken therefore no list
2514 * activity can be necessary.
2517 stat(s, FREE_FROZEN);
2522 * was_frozen may have been set after we acquired the list_lock in
2523 * an earlier loop. So we need to check it here again.
2526 stat(s, FREE_FROZEN);
2528 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2532 * Objects left in the slab. If it was not on the partial list before
2535 if (unlikely(!prior)) {
2536 remove_full(s, page);
2537 add_partial(n, page, DEACTIVATE_TO_TAIL);
2538 stat(s, FREE_ADD_PARTIAL);
2541 spin_unlock_irqrestore(&n->list_lock, flags);
2547 * Slab on the partial list.
2549 remove_partial(n, page);
2550 stat(s, FREE_REMOVE_PARTIAL);
2552 /* Slab must be on the full list */
2553 remove_full(s, page);
2555 spin_unlock_irqrestore(&n->list_lock, flags);
2557 discard_slab(s, page);
2561 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2562 * can perform fastpath freeing without additional function calls.
2564 * The fastpath is only possible if we are freeing to the current cpu slab
2565 * of this processor. This typically the case if we have just allocated
2568 * If fastpath is not possible then fall back to __slab_free where we deal
2569 * with all sorts of special processing.
2571 static __always_inline void slab_free(struct kmem_cache *s,
2572 struct page *page, void *x, unsigned long addr)
2574 void **object = (void *)x;
2575 struct kmem_cache_cpu *c;
2578 slab_free_hook(s, x);
2582 * Determine the currently cpus per cpu slab.
2583 * The cpu may change afterward. However that does not matter since
2584 * data is retrieved via this pointer. If we are on the same cpu
2585 * during the cmpxchg then the free will succedd.
2587 c = __this_cpu_ptr(s->cpu_slab);
2592 if (likely(page == c->page)) {
2593 set_freepointer(s, object, c->freelist);
2595 if (unlikely(!this_cpu_cmpxchg_double(
2596 s->cpu_slab->freelist, s->cpu_slab->tid,
2598 object, next_tid(tid)))) {
2600 note_cmpxchg_failure("slab_free", s, tid);
2603 stat(s, FREE_FASTPATH);
2605 __slab_free(s, page, x, addr);
2609 void kmem_cache_free(struct kmem_cache *s, void *x)
2613 page = virt_to_head_page(x);
2615 slab_free(s, page, x, _RET_IP_);
2617 trace_kmem_cache_free(_RET_IP_, x);
2619 EXPORT_SYMBOL(kmem_cache_free);
2622 * Object placement in a slab is made very easy because we always start at
2623 * offset 0. If we tune the size of the object to the alignment then we can
2624 * get the required alignment by putting one properly sized object after
2627 * Notice that the allocation order determines the sizes of the per cpu
2628 * caches. Each processor has always one slab available for allocations.
2629 * Increasing the allocation order reduces the number of times that slabs
2630 * must be moved on and off the partial lists and is therefore a factor in
2635 * Mininum / Maximum order of slab pages. This influences locking overhead
2636 * and slab fragmentation. A higher order reduces the number of partial slabs
2637 * and increases the number of allocations possible without having to
2638 * take the list_lock.
2640 static int slub_min_order;
2641 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2642 static int slub_min_objects;
2645 * Merge control. If this is set then no merging of slab caches will occur.
2646 * (Could be removed. This was introduced to pacify the merge skeptics.)
2648 static int slub_nomerge;
2651 * Calculate the order of allocation given an slab object size.
2653 * The order of allocation has significant impact on performance and other
2654 * system components. Generally order 0 allocations should be preferred since
2655 * order 0 does not cause fragmentation in the page allocator. Larger objects
2656 * be problematic to put into order 0 slabs because there may be too much
2657 * unused space left. We go to a higher order if more than 1/16th of the slab
2660 * In order to reach satisfactory performance we must ensure that a minimum
2661 * number of objects is in one slab. Otherwise we may generate too much
2662 * activity on the partial lists which requires taking the list_lock. This is
2663 * less a concern for large slabs though which are rarely used.
2665 * slub_max_order specifies the order where we begin to stop considering the
2666 * number of objects in a slab as critical. If we reach slub_max_order then
2667 * we try to keep the page order as low as possible. So we accept more waste
2668 * of space in favor of a small page order.
2670 * Higher order allocations also allow the placement of more objects in a
2671 * slab and thereby reduce object handling overhead. If the user has
2672 * requested a higher mininum order then we start with that one instead of
2673 * the smallest order which will fit the object.
2675 static inline int slab_order(int size, int min_objects,
2676 int max_order, int fract_leftover, int reserved)
2680 int min_order = slub_min_order;
2682 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2683 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2685 for (order = max(min_order,
2686 fls(min_objects * size - 1) - PAGE_SHIFT);
2687 order <= max_order; order++) {
2689 unsigned long slab_size = PAGE_SIZE << order;
2691 if (slab_size < min_objects * size + reserved)
2694 rem = (slab_size - reserved) % size;
2696 if (rem <= slab_size / fract_leftover)
2704 static inline int calculate_order(int size, int reserved)
2712 * Attempt to find best configuration for a slab. This
2713 * works by first attempting to generate a layout with
2714 * the best configuration and backing off gradually.
2716 * First we reduce the acceptable waste in a slab. Then
2717 * we reduce the minimum objects required in a slab.
2719 min_objects = slub_min_objects;
2721 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2722 max_objects = order_objects(slub_max_order, size, reserved);
2723 min_objects = min(min_objects, max_objects);
2725 while (min_objects > 1) {
2727 while (fraction >= 4) {
2728 order = slab_order(size, min_objects,
2729 slub_max_order, fraction, reserved);
2730 if (order <= slub_max_order)
2738 * We were unable to place multiple objects in a slab. Now
2739 * lets see if we can place a single object there.
2741 order = slab_order(size, 1, slub_max_order, 1, reserved);
2742 if (order <= slub_max_order)
2746 * Doh this slab cannot be placed using slub_max_order.
2748 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2749 if (order < MAX_ORDER)
2755 * Figure out what the alignment of the objects will be.
2757 static unsigned long calculate_alignment(unsigned long flags,
2758 unsigned long align, unsigned long size)
2761 * If the user wants hardware cache aligned objects then follow that
2762 * suggestion if the object is sufficiently large.
2764 * The hardware cache alignment cannot override the specified
2765 * alignment though. If that is greater then use it.
2767 if (flags & SLAB_HWCACHE_ALIGN) {
2768 unsigned long ralign = cache_line_size();
2769 while (size <= ralign / 2)
2771 align = max(align, ralign);
2774 if (align < ARCH_SLAB_MINALIGN)
2775 align = ARCH_SLAB_MINALIGN;
2777 return ALIGN(align, sizeof(void *));
2781 init_kmem_cache_node(struct kmem_cache_node *n)
2784 spin_lock_init(&n->list_lock);
2785 INIT_LIST_HEAD(&n->partial);
2786 #ifdef CONFIG_SLUB_DEBUG
2787 atomic_long_set(&n->nr_slabs, 0);
2788 atomic_long_set(&n->total_objects, 0);
2789 INIT_LIST_HEAD(&n->full);
2793 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2795 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2796 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2799 * Must align to double word boundary for the double cmpxchg
2800 * instructions to work; see __pcpu_double_call_return_bool().
2802 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2803 2 * sizeof(void *));
2808 init_kmem_cache_cpus(s);
2813 static struct kmem_cache *kmem_cache_node;
2816 * No kmalloc_node yet so do it by hand. We know that this is the first
2817 * slab on the node for this slabcache. There are no concurrent accesses
2820 * Note that this function only works on the kmalloc_node_cache
2821 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2822 * memory on a fresh node that has no slab structures yet.
2824 static void early_kmem_cache_node_alloc(int node)
2827 struct kmem_cache_node *n;
2829 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2831 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2834 if (page_to_nid(page) != node) {
2835 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2837 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2838 "in order to be able to continue\n");
2843 page->freelist = get_freepointer(kmem_cache_node, n);
2846 kmem_cache_node->node[node] = n;
2847 #ifdef CONFIG_SLUB_DEBUG
2848 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2849 init_tracking(kmem_cache_node, n);
2851 init_kmem_cache_node(n);
2852 inc_slabs_node(kmem_cache_node, node, page->objects);
2854 add_partial(n, page, DEACTIVATE_TO_HEAD);
2857 static void free_kmem_cache_nodes(struct kmem_cache *s)
2861 for_each_node_state(node, N_NORMAL_MEMORY) {
2862 struct kmem_cache_node *n = s->node[node];
2865 kmem_cache_free(kmem_cache_node, n);
2867 s->node[node] = NULL;
2871 static int init_kmem_cache_nodes(struct kmem_cache *s)
2875 for_each_node_state(node, N_NORMAL_MEMORY) {
2876 struct kmem_cache_node *n;
2878 if (slab_state == DOWN) {
2879 early_kmem_cache_node_alloc(node);
2882 n = kmem_cache_alloc_node(kmem_cache_node,
2886 free_kmem_cache_nodes(s);
2891 init_kmem_cache_node(n);
2896 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2898 if (min < MIN_PARTIAL)
2900 else if (min > MAX_PARTIAL)
2902 s->min_partial = min;
2906 * calculate_sizes() determines the order and the distribution of data within
2909 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2911 unsigned long flags = s->flags;
2912 unsigned long size = s->object_size;
2913 unsigned long align = s->align;
2917 * Round up object size to the next word boundary. We can only
2918 * place the free pointer at word boundaries and this determines
2919 * the possible location of the free pointer.
2921 size = ALIGN(size, sizeof(void *));
2923 #ifdef CONFIG_SLUB_DEBUG
2925 * Determine if we can poison the object itself. If the user of
2926 * the slab may touch the object after free or before allocation
2927 * then we should never poison the object itself.
2929 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2931 s->flags |= __OBJECT_POISON;
2933 s->flags &= ~__OBJECT_POISON;
2937 * If we are Redzoning then check if there is some space between the
2938 * end of the object and the free pointer. If not then add an
2939 * additional word to have some bytes to store Redzone information.
2941 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2942 size += sizeof(void *);
2946 * With that we have determined the number of bytes in actual use
2947 * by the object. This is the potential offset to the free pointer.
2951 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2954 * Relocate free pointer after the object if it is not
2955 * permitted to overwrite the first word of the object on
2958 * This is the case if we do RCU, have a constructor or
2959 * destructor or are poisoning the objects.
2962 size += sizeof(void *);
2965 #ifdef CONFIG_SLUB_DEBUG
2966 if (flags & SLAB_STORE_USER)
2968 * Need to store information about allocs and frees after
2971 size += 2 * sizeof(struct track);
2973 if (flags & SLAB_RED_ZONE)
2975 * Add some empty padding so that we can catch
2976 * overwrites from earlier objects rather than let
2977 * tracking information or the free pointer be
2978 * corrupted if a user writes before the start
2981 size += sizeof(void *);
2985 * Determine the alignment based on various parameters that the
2986 * user specified and the dynamic determination of cache line size
2989 align = calculate_alignment(flags, align, s->object_size);
2993 * SLUB stores one object immediately after another beginning from
2994 * offset 0. In order to align the objects we have to simply size
2995 * each object to conform to the alignment.
2997 size = ALIGN(size, align);
2999 if (forced_order >= 0)
3000 order = forced_order;
3002 order = calculate_order(size, s->reserved);
3009 s->allocflags |= __GFP_COMP;
3011 if (s->flags & SLAB_CACHE_DMA)
3012 s->allocflags |= SLUB_DMA;
3014 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3015 s->allocflags |= __GFP_RECLAIMABLE;
3018 * Determine the number of objects per slab
3020 s->oo = oo_make(order, size, s->reserved);
3021 s->min = oo_make(get_order(size), size, s->reserved);
3022 if (oo_objects(s->oo) > oo_objects(s->max))
3025 return !!oo_objects(s->oo);
3029 static int kmem_cache_open(struct kmem_cache *s,
3030 const char *name, size_t size,
3031 size_t align, unsigned long flags,
3032 void (*ctor)(void *))
3034 memset(s, 0, kmem_size);
3037 s->object_size = size;
3039 s->flags = kmem_cache_flags(size, flags, name, ctor);
3042 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3043 s->reserved = sizeof(struct rcu_head);
3045 if (!calculate_sizes(s, -1))
3047 if (disable_higher_order_debug) {
3049 * Disable debugging flags that store metadata if the min slab
3052 if (get_order(s->size) > get_order(s->object_size)) {
3053 s->flags &= ~DEBUG_METADATA_FLAGS;
3055 if (!calculate_sizes(s, -1))
3060 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3061 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3062 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3063 /* Enable fast mode */
3064 s->flags |= __CMPXCHG_DOUBLE;
3068 * The larger the object size is, the more pages we want on the partial
3069 * list to avoid pounding the page allocator excessively.
3071 set_min_partial(s, ilog2(s->size) / 2);
3074 * cpu_partial determined the maximum number of objects kept in the
3075 * per cpu partial lists of a processor.
3077 * Per cpu partial lists mainly contain slabs that just have one
3078 * object freed. If they are used for allocation then they can be
3079 * filled up again with minimal effort. The slab will never hit the
3080 * per node partial lists and therefore no locking will be required.
3082 * This setting also determines
3084 * A) The number of objects from per cpu partial slabs dumped to the
3085 * per node list when we reach the limit.
3086 * B) The number of objects in cpu partial slabs to extract from the
3087 * per node list when we run out of per cpu objects. We only fetch 50%
3088 * to keep some capacity around for frees.
3090 if (kmem_cache_debug(s))
3092 else if (s->size >= PAGE_SIZE)
3094 else if (s->size >= 1024)
3096 else if (s->size >= 256)
3097 s->cpu_partial = 13;
3099 s->cpu_partial = 30;
3103 s->remote_node_defrag_ratio = 1000;
3105 if (!init_kmem_cache_nodes(s))
3108 if (alloc_kmem_cache_cpus(s))
3111 free_kmem_cache_nodes(s);
3113 if (flags & SLAB_PANIC)
3114 panic("Cannot create slab %s size=%lu realsize=%u "
3115 "order=%u offset=%u flags=%lx\n",
3116 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3122 * Determine the size of a slab object
3124 unsigned int kmem_cache_size(struct kmem_cache *s)
3126 return s->object_size;
3128 EXPORT_SYMBOL(kmem_cache_size);
3130 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3133 #ifdef CONFIG_SLUB_DEBUG
3134 void *addr = page_address(page);
3136 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3137 sizeof(long), GFP_ATOMIC);
3140 slab_err(s, page, "%s", text);
3143 get_map(s, page, map);
3144 for_each_object(p, s, addr, page->objects) {
3146 if (!test_bit(slab_index(p, s, addr), map)) {
3147 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3149 print_tracking(s, p);
3158 * Attempt to free all partial slabs on a node.
3159 * This is called from kmem_cache_close(). We must be the last thread
3160 * using the cache and therefore we do not need to lock anymore.
3162 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3164 struct page *page, *h;
3166 list_for_each_entry_safe(page, h, &n->partial, lru) {
3168 remove_partial(n, page);
3169 discard_slab(s, page);
3171 list_slab_objects(s, page,
3172 "Objects remaining on kmem_cache_close()");
3178 * Release all resources used by a slab cache.
3180 static inline int kmem_cache_close(struct kmem_cache *s)
3185 free_percpu(s->cpu_slab);
3186 /* Attempt to free all objects */
3187 for_each_node_state(node, N_NORMAL_MEMORY) {
3188 struct kmem_cache_node *n = get_node(s, node);
3191 if (n->nr_partial || slabs_node(s, node))
3194 free_kmem_cache_nodes(s);
3199 * Close a cache and release the kmem_cache structure
3200 * (must be used for caches created using kmem_cache_create)
3202 void kmem_cache_destroy(struct kmem_cache *s)
3204 mutex_lock(&slab_mutex);
3208 mutex_unlock(&slab_mutex);
3209 if (kmem_cache_close(s)) {
3210 printk(KERN_ERR "SLUB %s: %s called for cache that "
3211 "still has objects.\n", s->name, __func__);
3214 if (s->flags & SLAB_DESTROY_BY_RCU)
3216 sysfs_slab_remove(s);
3218 mutex_unlock(&slab_mutex);
3220 EXPORT_SYMBOL(kmem_cache_destroy);
3222 /********************************************************************
3224 *******************************************************************/
3226 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3227 EXPORT_SYMBOL(kmalloc_caches);
3229 static struct kmem_cache *kmem_cache;
3231 #ifdef CONFIG_ZONE_DMA
3232 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3235 static int __init setup_slub_min_order(char *str)
3237 get_option(&str, &slub_min_order);
3242 __setup("slub_min_order=", setup_slub_min_order);
3244 static int __init setup_slub_max_order(char *str)
3246 get_option(&str, &slub_max_order);
3247 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3252 __setup("slub_max_order=", setup_slub_max_order);
3254 static int __init setup_slub_min_objects(char *str)
3256 get_option(&str, &slub_min_objects);
3261 __setup("slub_min_objects=", setup_slub_min_objects);
3263 static int __init setup_slub_nomerge(char *str)
3269 __setup("slub_nomerge", setup_slub_nomerge);
3271 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3272 int size, unsigned int flags)
3274 struct kmem_cache *s;
3276 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3279 * This function is called with IRQs disabled during early-boot on
3280 * single CPU so there's no need to take slab_mutex here.
3282 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3286 list_add(&s->list, &slab_caches);
3290 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3295 * Conversion table for small slabs sizes / 8 to the index in the
3296 * kmalloc array. This is necessary for slabs < 192 since we have non power
3297 * of two cache sizes there. The size of larger slabs can be determined using
3300 static s8 size_index[24] = {
3327 static inline int size_index_elem(size_t bytes)
3329 return (bytes - 1) / 8;
3332 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3338 return ZERO_SIZE_PTR;
3340 index = size_index[size_index_elem(size)];
3342 index = fls(size - 1);
3344 #ifdef CONFIG_ZONE_DMA
3345 if (unlikely((flags & SLUB_DMA)))
3346 return kmalloc_dma_caches[index];
3349 return kmalloc_caches[index];
3352 void *__kmalloc(size_t size, gfp_t flags)
3354 struct kmem_cache *s;
3357 if (unlikely(size > SLUB_MAX_SIZE))
3358 return kmalloc_large(size, flags);
3360 s = get_slab(size, flags);
3362 if (unlikely(ZERO_OR_NULL_PTR(s)))
3365 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3367 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3371 EXPORT_SYMBOL(__kmalloc);
3374 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3379 flags |= __GFP_COMP | __GFP_NOTRACK;
3380 page = alloc_pages_node(node, flags, get_order(size));
3382 ptr = page_address(page);
3384 kmemleak_alloc(ptr, size, 1, flags);
3388 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3390 struct kmem_cache *s;
3393 if (unlikely(size > SLUB_MAX_SIZE)) {
3394 ret = kmalloc_large_node(size, flags, node);
3396 trace_kmalloc_node(_RET_IP_, ret,
3397 size, PAGE_SIZE << get_order(size),
3403 s = get_slab(size, flags);
3405 if (unlikely(ZERO_OR_NULL_PTR(s)))
3408 ret = slab_alloc(s, flags, node, _RET_IP_);
3410 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3414 EXPORT_SYMBOL(__kmalloc_node);
3417 size_t ksize(const void *object)
3421 if (unlikely(object == ZERO_SIZE_PTR))
3424 page = virt_to_head_page(object);
3426 if (unlikely(!PageSlab(page))) {
3427 WARN_ON(!PageCompound(page));
3428 return PAGE_SIZE << compound_order(page);
3431 return slab_ksize(page->slab);
3433 EXPORT_SYMBOL(ksize);
3435 #ifdef CONFIG_SLUB_DEBUG
3436 bool verify_mem_not_deleted(const void *x)
3439 void *object = (void *)x;
3440 unsigned long flags;
3443 if (unlikely(ZERO_OR_NULL_PTR(x)))
3446 local_irq_save(flags);
3448 page = virt_to_head_page(x);
3449 if (unlikely(!PageSlab(page))) {
3450 /* maybe it was from stack? */
3456 if (on_freelist(page->slab, page, object)) {
3457 object_err(page->slab, page, object, "Object is on free-list");
3465 local_irq_restore(flags);
3468 EXPORT_SYMBOL(verify_mem_not_deleted);
3471 void kfree(const void *x)
3474 void *object = (void *)x;
3476 trace_kfree(_RET_IP_, x);
3478 if (unlikely(ZERO_OR_NULL_PTR(x)))
3481 page = virt_to_head_page(x);
3482 if (unlikely(!PageSlab(page))) {
3483 BUG_ON(!PageCompound(page));
3488 slab_free(page->slab, page, object, _RET_IP_);
3490 EXPORT_SYMBOL(kfree);
3493 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3494 * the remaining slabs by the number of items in use. The slabs with the
3495 * most items in use come first. New allocations will then fill those up
3496 * and thus they can be removed from the partial lists.
3498 * The slabs with the least items are placed last. This results in them
3499 * being allocated from last increasing the chance that the last objects
3500 * are freed in them.
3502 int kmem_cache_shrink(struct kmem_cache *s)
3506 struct kmem_cache_node *n;
3509 int objects = oo_objects(s->max);
3510 struct list_head *slabs_by_inuse =
3511 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3512 unsigned long flags;
3514 if (!slabs_by_inuse)
3518 for_each_node_state(node, N_NORMAL_MEMORY) {
3519 n = get_node(s, node);
3524 for (i = 0; i < objects; i++)
3525 INIT_LIST_HEAD(slabs_by_inuse + i);
3527 spin_lock_irqsave(&n->list_lock, flags);
3530 * Build lists indexed by the items in use in each slab.
3532 * Note that concurrent frees may occur while we hold the
3533 * list_lock. page->inuse here is the upper limit.
3535 list_for_each_entry_safe(page, t, &n->partial, lru) {
3536 list_move(&page->lru, slabs_by_inuse + page->inuse);
3542 * Rebuild the partial list with the slabs filled up most
3543 * first and the least used slabs at the end.
3545 for (i = objects - 1; i > 0; i--)
3546 list_splice(slabs_by_inuse + i, n->partial.prev);
3548 spin_unlock_irqrestore(&n->list_lock, flags);
3550 /* Release empty slabs */
3551 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3552 discard_slab(s, page);
3555 kfree(slabs_by_inuse);
3558 EXPORT_SYMBOL(kmem_cache_shrink);
3560 #if defined(CONFIG_MEMORY_HOTPLUG)
3561 static int slab_mem_going_offline_callback(void *arg)
3563 struct kmem_cache *s;
3565 mutex_lock(&slab_mutex);
3566 list_for_each_entry(s, &slab_caches, list)
3567 kmem_cache_shrink(s);
3568 mutex_unlock(&slab_mutex);
3573 static void slab_mem_offline_callback(void *arg)
3575 struct kmem_cache_node *n;
3576 struct kmem_cache *s;
3577 struct memory_notify *marg = arg;
3580 offline_node = marg->status_change_nid;
3583 * If the node still has available memory. we need kmem_cache_node
3586 if (offline_node < 0)
3589 mutex_lock(&slab_mutex);
3590 list_for_each_entry(s, &slab_caches, list) {
3591 n = get_node(s, offline_node);
3594 * if n->nr_slabs > 0, slabs still exist on the node
3595 * that is going down. We were unable to free them,
3596 * and offline_pages() function shouldn't call this
3597 * callback. So, we must fail.
3599 BUG_ON(slabs_node(s, offline_node));
3601 s->node[offline_node] = NULL;
3602 kmem_cache_free(kmem_cache_node, n);
3605 mutex_unlock(&slab_mutex);
3608 static int slab_mem_going_online_callback(void *arg)
3610 struct kmem_cache_node *n;
3611 struct kmem_cache *s;
3612 struct memory_notify *marg = arg;
3613 int nid = marg->status_change_nid;
3617 * If the node's memory is already available, then kmem_cache_node is
3618 * already created. Nothing to do.
3624 * We are bringing a node online. No memory is available yet. We must
3625 * allocate a kmem_cache_node structure in order to bring the node
3628 mutex_lock(&slab_mutex);
3629 list_for_each_entry(s, &slab_caches, list) {
3631 * XXX: kmem_cache_alloc_node will fallback to other nodes
3632 * since memory is not yet available from the node that
3635 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3640 init_kmem_cache_node(n);
3644 mutex_unlock(&slab_mutex);
3648 static int slab_memory_callback(struct notifier_block *self,
3649 unsigned long action, void *arg)
3654 case MEM_GOING_ONLINE:
3655 ret = slab_mem_going_online_callback(arg);
3657 case MEM_GOING_OFFLINE:
3658 ret = slab_mem_going_offline_callback(arg);
3661 case MEM_CANCEL_ONLINE:
3662 slab_mem_offline_callback(arg);
3665 case MEM_CANCEL_OFFLINE:
3669 ret = notifier_from_errno(ret);
3675 #endif /* CONFIG_MEMORY_HOTPLUG */
3677 /********************************************************************
3678 * Basic setup of slabs
3679 *******************************************************************/
3682 * Used for early kmem_cache structures that were allocated using
3683 * the page allocator
3686 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3690 list_add(&s->list, &slab_caches);
3693 for_each_node_state(node, N_NORMAL_MEMORY) {
3694 struct kmem_cache_node *n = get_node(s, node);
3698 list_for_each_entry(p, &n->partial, lru)
3701 #ifdef CONFIG_SLUB_DEBUG
3702 list_for_each_entry(p, &n->full, lru)
3709 void __init kmem_cache_init(void)
3713 struct kmem_cache *temp_kmem_cache;
3715 struct kmem_cache *temp_kmem_cache_node;
3716 unsigned long kmalloc_size;
3718 if (debug_guardpage_minorder())
3721 kmem_size = offsetof(struct kmem_cache, node) +
3722 nr_node_ids * sizeof(struct kmem_cache_node *);
3724 /* Allocate two kmem_caches from the page allocator */
3725 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3726 order = get_order(2 * kmalloc_size);
3727 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3730 * Must first have the slab cache available for the allocations of the
3731 * struct kmem_cache_node's. There is special bootstrap code in
3732 * kmem_cache_open for slab_state == DOWN.
3734 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3736 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3737 sizeof(struct kmem_cache_node),
3738 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3740 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3742 /* Able to allocate the per node structures */
3743 slab_state = PARTIAL;
3745 temp_kmem_cache = kmem_cache;
3746 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3747 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3748 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3749 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3752 * Allocate kmem_cache_node properly from the kmem_cache slab.
3753 * kmem_cache_node is separately allocated so no need to
3754 * update any list pointers.
3756 temp_kmem_cache_node = kmem_cache_node;
3758 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3759 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3761 kmem_cache_bootstrap_fixup(kmem_cache_node);
3764 kmem_cache_bootstrap_fixup(kmem_cache);
3766 /* Free temporary boot structure */
3767 free_pages((unsigned long)temp_kmem_cache, order);
3769 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3772 * Patch up the size_index table if we have strange large alignment
3773 * requirements for the kmalloc array. This is only the case for
3774 * MIPS it seems. The standard arches will not generate any code here.
3776 * Largest permitted alignment is 256 bytes due to the way we
3777 * handle the index determination for the smaller caches.
3779 * Make sure that nothing crazy happens if someone starts tinkering
3780 * around with ARCH_KMALLOC_MINALIGN
3782 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3783 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3785 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3786 int elem = size_index_elem(i);
3787 if (elem >= ARRAY_SIZE(size_index))
3789 size_index[elem] = KMALLOC_SHIFT_LOW;
3792 if (KMALLOC_MIN_SIZE == 64) {
3794 * The 96 byte size cache is not used if the alignment
3797 for (i = 64 + 8; i <= 96; i += 8)
3798 size_index[size_index_elem(i)] = 7;
3799 } else if (KMALLOC_MIN_SIZE == 128) {
3801 * The 192 byte sized cache is not used if the alignment
3802 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3805 for (i = 128 + 8; i <= 192; i += 8)
3806 size_index[size_index_elem(i)] = 8;
3809 /* Caches that are not of the two-to-the-power-of size */
3810 if (KMALLOC_MIN_SIZE <= 32) {
3811 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3815 if (KMALLOC_MIN_SIZE <= 64) {
3816 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3820 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3821 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3827 /* Provide the correct kmalloc names now that the caches are up */
3828 if (KMALLOC_MIN_SIZE <= 32) {
3829 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3830 BUG_ON(!kmalloc_caches[1]->name);
3833 if (KMALLOC_MIN_SIZE <= 64) {
3834 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3835 BUG_ON(!kmalloc_caches[2]->name);
3838 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3839 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3842 kmalloc_caches[i]->name = s;
3846 register_cpu_notifier(&slab_notifier);
3849 #ifdef CONFIG_ZONE_DMA
3850 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3851 struct kmem_cache *s = kmalloc_caches[i];
3854 char *name = kasprintf(GFP_NOWAIT,
3855 "dma-kmalloc-%d", s->object_size);
3858 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3859 s->object_size, SLAB_CACHE_DMA);
3864 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3865 " CPUs=%d, Nodes=%d\n",
3866 caches, cache_line_size(),
3867 slub_min_order, slub_max_order, slub_min_objects,
3868 nr_cpu_ids, nr_node_ids);
3871 void __init kmem_cache_init_late(void)
3876 * Find a mergeable slab cache
3878 static int slab_unmergeable(struct kmem_cache *s)
3880 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3887 * We may have set a slab to be unmergeable during bootstrap.
3889 if (s->refcount < 0)
3895 static struct kmem_cache *find_mergeable(size_t size,
3896 size_t align, unsigned long flags, const char *name,
3897 void (*ctor)(void *))
3899 struct kmem_cache *s;
3901 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3907 size = ALIGN(size, sizeof(void *));
3908 align = calculate_alignment(flags, align, size);
3909 size = ALIGN(size, align);
3910 flags = kmem_cache_flags(size, flags, name, NULL);
3912 list_for_each_entry(s, &slab_caches, list) {
3913 if (slab_unmergeable(s))
3919 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3922 * Check if alignment is compatible.
3923 * Courtesy of Adrian Drzewiecki
3925 if ((s->size & ~(align - 1)) != s->size)
3928 if (s->size - size >= sizeof(void *))
3936 struct kmem_cache *__kmem_cache_create(const char *name, size_t size,
3937 size_t align, unsigned long flags, void (*ctor)(void *))
3939 struct kmem_cache *s;
3942 s = find_mergeable(size, align, flags, name, ctor);
3946 * Adjust the object sizes so that we clear
3947 * the complete object on kzalloc.
3949 s->object_size = max(s->object_size, (int)size);
3950 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3952 if (sysfs_slab_alias(s, name)) {
3959 n = kstrdup(name, GFP_KERNEL);
3963 s = kmalloc(kmem_size, GFP_KERNEL);
3965 if (kmem_cache_open(s, n,
3966 size, align, flags, ctor)) {
3969 list_add(&s->list, &slab_caches);
3970 mutex_unlock(&slab_mutex);
3971 r = sysfs_slab_add(s);
3972 mutex_lock(&slab_mutex);
3978 kmem_cache_close(s);
3988 * Use the cpu notifier to insure that the cpu slabs are flushed when
3991 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3992 unsigned long action, void *hcpu)
3994 long cpu = (long)hcpu;
3995 struct kmem_cache *s;
3996 unsigned long flags;
3999 case CPU_UP_CANCELED:
4000 case CPU_UP_CANCELED_FROZEN:
4002 case CPU_DEAD_FROZEN:
4003 mutex_lock(&slab_mutex);
4004 list_for_each_entry(s, &slab_caches, list) {
4005 local_irq_save(flags);
4006 __flush_cpu_slab(s, cpu);
4007 local_irq_restore(flags);
4009 mutex_unlock(&slab_mutex);
4017 static struct notifier_block __cpuinitdata slab_notifier = {
4018 .notifier_call = slab_cpuup_callback
4023 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4025 struct kmem_cache *s;
4028 if (unlikely(size > SLUB_MAX_SIZE))
4029 return kmalloc_large(size, gfpflags);
4031 s = get_slab(size, gfpflags);
4033 if (unlikely(ZERO_OR_NULL_PTR(s)))
4036 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4038 /* Honor the call site pointer we received. */
4039 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4045 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4046 int node, unsigned long caller)
4048 struct kmem_cache *s;
4051 if (unlikely(size > SLUB_MAX_SIZE)) {
4052 ret = kmalloc_large_node(size, gfpflags, node);
4054 trace_kmalloc_node(caller, ret,
4055 size, PAGE_SIZE << get_order(size),
4061 s = get_slab(size, gfpflags);
4063 if (unlikely(ZERO_OR_NULL_PTR(s)))
4066 ret = slab_alloc(s, gfpflags, node, caller);
4068 /* Honor the call site pointer we received. */
4069 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4076 static int count_inuse(struct page *page)
4081 static int count_total(struct page *page)
4083 return page->objects;
4087 #ifdef CONFIG_SLUB_DEBUG
4088 static int validate_slab(struct kmem_cache *s, struct page *page,
4092 void *addr = page_address(page);
4094 if (!check_slab(s, page) ||
4095 !on_freelist(s, page, NULL))
4098 /* Now we know that a valid freelist exists */
4099 bitmap_zero(map, page->objects);
4101 get_map(s, page, map);
4102 for_each_object(p, s, addr, page->objects) {
4103 if (test_bit(slab_index(p, s, addr), map))
4104 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4108 for_each_object(p, s, addr, page->objects)
4109 if (!test_bit(slab_index(p, s, addr), map))
4110 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4115 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4119 validate_slab(s, page, map);
4123 static int validate_slab_node(struct kmem_cache *s,
4124 struct kmem_cache_node *n, unsigned long *map)
4126 unsigned long count = 0;
4128 unsigned long flags;
4130 spin_lock_irqsave(&n->list_lock, flags);
4132 list_for_each_entry(page, &n->partial, lru) {
4133 validate_slab_slab(s, page, map);
4136 if (count != n->nr_partial)
4137 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4138 "counter=%ld\n", s->name, count, n->nr_partial);
4140 if (!(s->flags & SLAB_STORE_USER))
4143 list_for_each_entry(page, &n->full, lru) {
4144 validate_slab_slab(s, page, map);
4147 if (count != atomic_long_read(&n->nr_slabs))
4148 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4149 "counter=%ld\n", s->name, count,
4150 atomic_long_read(&n->nr_slabs));
4153 spin_unlock_irqrestore(&n->list_lock, flags);
4157 static long validate_slab_cache(struct kmem_cache *s)
4160 unsigned long count = 0;
4161 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4162 sizeof(unsigned long), GFP_KERNEL);
4168 for_each_node_state(node, N_NORMAL_MEMORY) {
4169 struct kmem_cache_node *n = get_node(s, node);
4171 count += validate_slab_node(s, n, map);
4177 * Generate lists of code addresses where slabcache objects are allocated
4182 unsigned long count;
4189 DECLARE_BITMAP(cpus, NR_CPUS);
4195 unsigned long count;
4196 struct location *loc;
4199 static void free_loc_track(struct loc_track *t)
4202 free_pages((unsigned long)t->loc,
4203 get_order(sizeof(struct location) * t->max));
4206 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4211 order = get_order(sizeof(struct location) * max);
4213 l = (void *)__get_free_pages(flags, order);
4218 memcpy(l, t->loc, sizeof(struct location) * t->count);
4226 static int add_location(struct loc_track *t, struct kmem_cache *s,
4227 const struct track *track)
4229 long start, end, pos;
4231 unsigned long caddr;
4232 unsigned long age = jiffies - track->when;
4238 pos = start + (end - start + 1) / 2;
4241 * There is nothing at "end". If we end up there
4242 * we need to add something to before end.
4247 caddr = t->loc[pos].addr;
4248 if (track->addr == caddr) {
4254 if (age < l->min_time)
4256 if (age > l->max_time)
4259 if (track->pid < l->min_pid)
4260 l->min_pid = track->pid;
4261 if (track->pid > l->max_pid)
4262 l->max_pid = track->pid;
4264 cpumask_set_cpu(track->cpu,
4265 to_cpumask(l->cpus));
4267 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4271 if (track->addr < caddr)
4278 * Not found. Insert new tracking element.
4280 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4286 (t->count - pos) * sizeof(struct location));
4289 l->addr = track->addr;
4293 l->min_pid = track->pid;
4294 l->max_pid = track->pid;
4295 cpumask_clear(to_cpumask(l->cpus));
4296 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4297 nodes_clear(l->nodes);
4298 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4302 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4303 struct page *page, enum track_item alloc,
4306 void *addr = page_address(page);
4309 bitmap_zero(map, page->objects);
4310 get_map(s, page, map);
4312 for_each_object(p, s, addr, page->objects)
4313 if (!test_bit(slab_index(p, s, addr), map))
4314 add_location(t, s, get_track(s, p, alloc));
4317 static int list_locations(struct kmem_cache *s, char *buf,
4318 enum track_item alloc)
4322 struct loc_track t = { 0, 0, NULL };
4324 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4325 sizeof(unsigned long), GFP_KERNEL);
4327 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4330 return sprintf(buf, "Out of memory\n");
4332 /* Push back cpu slabs */
4335 for_each_node_state(node, N_NORMAL_MEMORY) {
4336 struct kmem_cache_node *n = get_node(s, node);
4337 unsigned long flags;
4340 if (!atomic_long_read(&n->nr_slabs))
4343 spin_lock_irqsave(&n->list_lock, flags);
4344 list_for_each_entry(page, &n->partial, lru)
4345 process_slab(&t, s, page, alloc, map);
4346 list_for_each_entry(page, &n->full, lru)
4347 process_slab(&t, s, page, alloc, map);
4348 spin_unlock_irqrestore(&n->list_lock, flags);
4351 for (i = 0; i < t.count; i++) {
4352 struct location *l = &t.loc[i];
4354 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4356 len += sprintf(buf + len, "%7ld ", l->count);
4359 len += sprintf(buf + len, "%pS", (void *)l->addr);
4361 len += sprintf(buf + len, "<not-available>");
4363 if (l->sum_time != l->min_time) {
4364 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4366 (long)div_u64(l->sum_time, l->count),
4369 len += sprintf(buf + len, " age=%ld",
4372 if (l->min_pid != l->max_pid)
4373 len += sprintf(buf + len, " pid=%ld-%ld",
4374 l->min_pid, l->max_pid);
4376 len += sprintf(buf + len, " pid=%ld",
4379 if (num_online_cpus() > 1 &&
4380 !cpumask_empty(to_cpumask(l->cpus)) &&
4381 len < PAGE_SIZE - 60) {
4382 len += sprintf(buf + len, " cpus=");
4383 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4384 to_cpumask(l->cpus));
4387 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4388 len < PAGE_SIZE - 60) {
4389 len += sprintf(buf + len, " nodes=");
4390 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4394 len += sprintf(buf + len, "\n");
4400 len += sprintf(buf, "No data\n");
4405 #ifdef SLUB_RESILIENCY_TEST
4406 static void resiliency_test(void)
4410 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4412 printk(KERN_ERR "SLUB resiliency testing\n");
4413 printk(KERN_ERR "-----------------------\n");
4414 printk(KERN_ERR "A. Corruption after allocation\n");
4416 p = kzalloc(16, GFP_KERNEL);
4418 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4419 " 0x12->0x%p\n\n", p + 16);
4421 validate_slab_cache(kmalloc_caches[4]);
4423 /* Hmmm... The next two are dangerous */
4424 p = kzalloc(32, GFP_KERNEL);
4425 p[32 + sizeof(void *)] = 0x34;
4426 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4427 " 0x34 -> -0x%p\n", p);
4429 "If allocated object is overwritten then not detectable\n\n");
4431 validate_slab_cache(kmalloc_caches[5]);
4432 p = kzalloc(64, GFP_KERNEL);
4433 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4435 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4438 "If allocated object is overwritten then not detectable\n\n");
4439 validate_slab_cache(kmalloc_caches[6]);
4441 printk(KERN_ERR "\nB. Corruption after free\n");
4442 p = kzalloc(128, GFP_KERNEL);
4445 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4446 validate_slab_cache(kmalloc_caches[7]);
4448 p = kzalloc(256, GFP_KERNEL);
4451 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4453 validate_slab_cache(kmalloc_caches[8]);
4455 p = kzalloc(512, GFP_KERNEL);
4458 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4459 validate_slab_cache(kmalloc_caches[9]);
4463 static void resiliency_test(void) {};
4468 enum slab_stat_type {
4469 SL_ALL, /* All slabs */
4470 SL_PARTIAL, /* Only partially allocated slabs */
4471 SL_CPU, /* Only slabs used for cpu caches */
4472 SL_OBJECTS, /* Determine allocated objects not slabs */
4473 SL_TOTAL /* Determine object capacity not slabs */
4476 #define SO_ALL (1 << SL_ALL)
4477 #define SO_PARTIAL (1 << SL_PARTIAL)
4478 #define SO_CPU (1 << SL_CPU)
4479 #define SO_OBJECTS (1 << SL_OBJECTS)
4480 #define SO_TOTAL (1 << SL_TOTAL)
4482 static ssize_t show_slab_objects(struct kmem_cache *s,
4483 char *buf, unsigned long flags)
4485 unsigned long total = 0;
4488 unsigned long *nodes;
4489 unsigned long *per_cpu;
4491 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4494 per_cpu = nodes + nr_node_ids;
4496 if (flags & SO_CPU) {
4499 for_each_possible_cpu(cpu) {
4500 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4504 page = ACCESS_ONCE(c->page);
4508 node = page_to_nid(page);
4509 if (flags & SO_TOTAL)
4511 else if (flags & SO_OBJECTS)
4519 page = ACCESS_ONCE(c->partial);
4530 lock_memory_hotplug();
4531 #ifdef CONFIG_SLUB_DEBUG
4532 if (flags & SO_ALL) {
4533 for_each_node_state(node, N_NORMAL_MEMORY) {
4534 struct kmem_cache_node *n = get_node(s, node);
4536 if (flags & SO_TOTAL)
4537 x = atomic_long_read(&n->total_objects);
4538 else if (flags & SO_OBJECTS)
4539 x = atomic_long_read(&n->total_objects) -
4540 count_partial(n, count_free);
4543 x = atomic_long_read(&n->nr_slabs);
4550 if (flags & SO_PARTIAL) {
4551 for_each_node_state(node, N_NORMAL_MEMORY) {
4552 struct kmem_cache_node *n = get_node(s, node);
4554 if (flags & SO_TOTAL)
4555 x = count_partial(n, count_total);
4556 else if (flags & SO_OBJECTS)
4557 x = count_partial(n, count_inuse);
4564 x = sprintf(buf, "%lu", total);
4566 for_each_node_state(node, N_NORMAL_MEMORY)
4568 x += sprintf(buf + x, " N%d=%lu",
4571 unlock_memory_hotplug();
4573 return x + sprintf(buf + x, "\n");
4576 #ifdef CONFIG_SLUB_DEBUG
4577 static int any_slab_objects(struct kmem_cache *s)
4581 for_each_online_node(node) {
4582 struct kmem_cache_node *n = get_node(s, node);
4587 if (atomic_long_read(&n->total_objects))
4594 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4595 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4597 struct slab_attribute {
4598 struct attribute attr;
4599 ssize_t (*show)(struct kmem_cache *s, char *buf);
4600 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4603 #define SLAB_ATTR_RO(_name) \
4604 static struct slab_attribute _name##_attr = \
4605 __ATTR(_name, 0400, _name##_show, NULL)
4607 #define SLAB_ATTR(_name) \
4608 static struct slab_attribute _name##_attr = \
4609 __ATTR(_name, 0600, _name##_show, _name##_store)
4611 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4613 return sprintf(buf, "%d\n", s->size);
4615 SLAB_ATTR_RO(slab_size);
4617 static ssize_t align_show(struct kmem_cache *s, char *buf)
4619 return sprintf(buf, "%d\n", s->align);
4621 SLAB_ATTR_RO(align);
4623 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4625 return sprintf(buf, "%d\n", s->object_size);
4627 SLAB_ATTR_RO(object_size);
4629 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4631 return sprintf(buf, "%d\n", oo_objects(s->oo));
4633 SLAB_ATTR_RO(objs_per_slab);
4635 static ssize_t order_store(struct kmem_cache *s,
4636 const char *buf, size_t length)
4638 unsigned long order;
4641 err = strict_strtoul(buf, 10, &order);
4645 if (order > slub_max_order || order < slub_min_order)
4648 calculate_sizes(s, order);
4652 static ssize_t order_show(struct kmem_cache *s, char *buf)
4654 return sprintf(buf, "%d\n", oo_order(s->oo));
4658 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4660 return sprintf(buf, "%lu\n", s->min_partial);
4663 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4669 err = strict_strtoul(buf, 10, &min);
4673 set_min_partial(s, min);
4676 SLAB_ATTR(min_partial);
4678 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4680 return sprintf(buf, "%u\n", s->cpu_partial);
4683 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4686 unsigned long objects;
4689 err = strict_strtoul(buf, 10, &objects);
4692 if (objects && kmem_cache_debug(s))
4695 s->cpu_partial = objects;
4699 SLAB_ATTR(cpu_partial);
4701 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4705 return sprintf(buf, "%pS\n", s->ctor);
4709 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4711 return sprintf(buf, "%d\n", s->refcount - 1);
4713 SLAB_ATTR_RO(aliases);
4715 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4717 return show_slab_objects(s, buf, SO_PARTIAL);
4719 SLAB_ATTR_RO(partial);
4721 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4723 return show_slab_objects(s, buf, SO_CPU);
4725 SLAB_ATTR_RO(cpu_slabs);
4727 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4729 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4731 SLAB_ATTR_RO(objects);
4733 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4735 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4737 SLAB_ATTR_RO(objects_partial);
4739 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4746 for_each_online_cpu(cpu) {
4747 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4750 pages += page->pages;
4751 objects += page->pobjects;
4755 len = sprintf(buf, "%d(%d)", objects, pages);
4758 for_each_online_cpu(cpu) {
4759 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4761 if (page && len < PAGE_SIZE - 20)
4762 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4763 page->pobjects, page->pages);
4766 return len + sprintf(buf + len, "\n");
4768 SLAB_ATTR_RO(slabs_cpu_partial);
4770 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4772 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4775 static ssize_t reclaim_account_store(struct kmem_cache *s,
4776 const char *buf, size_t length)
4778 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4780 s->flags |= SLAB_RECLAIM_ACCOUNT;
4783 SLAB_ATTR(reclaim_account);
4785 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4787 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4789 SLAB_ATTR_RO(hwcache_align);
4791 #ifdef CONFIG_ZONE_DMA
4792 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4794 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4796 SLAB_ATTR_RO(cache_dma);
4799 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4801 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4803 SLAB_ATTR_RO(destroy_by_rcu);
4805 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4807 return sprintf(buf, "%d\n", s->reserved);
4809 SLAB_ATTR_RO(reserved);
4811 #ifdef CONFIG_SLUB_DEBUG
4812 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4814 return show_slab_objects(s, buf, SO_ALL);
4816 SLAB_ATTR_RO(slabs);
4818 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4820 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4822 SLAB_ATTR_RO(total_objects);
4824 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4826 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4829 static ssize_t sanity_checks_store(struct kmem_cache *s,
4830 const char *buf, size_t length)
4832 s->flags &= ~SLAB_DEBUG_FREE;
4833 if (buf[0] == '1') {
4834 s->flags &= ~__CMPXCHG_DOUBLE;
4835 s->flags |= SLAB_DEBUG_FREE;
4839 SLAB_ATTR(sanity_checks);
4841 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4843 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4846 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4849 s->flags &= ~SLAB_TRACE;
4850 if (buf[0] == '1') {
4851 s->flags &= ~__CMPXCHG_DOUBLE;
4852 s->flags |= SLAB_TRACE;
4858 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4860 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4863 static ssize_t red_zone_store(struct kmem_cache *s,
4864 const char *buf, size_t length)
4866 if (any_slab_objects(s))
4869 s->flags &= ~SLAB_RED_ZONE;
4870 if (buf[0] == '1') {
4871 s->flags &= ~__CMPXCHG_DOUBLE;
4872 s->flags |= SLAB_RED_ZONE;
4874 calculate_sizes(s, -1);
4877 SLAB_ATTR(red_zone);
4879 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4881 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4884 static ssize_t poison_store(struct kmem_cache *s,
4885 const char *buf, size_t length)
4887 if (any_slab_objects(s))
4890 s->flags &= ~SLAB_POISON;
4891 if (buf[0] == '1') {
4892 s->flags &= ~__CMPXCHG_DOUBLE;
4893 s->flags |= SLAB_POISON;
4895 calculate_sizes(s, -1);
4900 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4902 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4905 static ssize_t store_user_store(struct kmem_cache *s,
4906 const char *buf, size_t length)
4908 if (any_slab_objects(s))
4911 s->flags &= ~SLAB_STORE_USER;
4912 if (buf[0] == '1') {
4913 s->flags &= ~__CMPXCHG_DOUBLE;
4914 s->flags |= SLAB_STORE_USER;
4916 calculate_sizes(s, -1);
4919 SLAB_ATTR(store_user);
4921 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4926 static ssize_t validate_store(struct kmem_cache *s,
4927 const char *buf, size_t length)
4931 if (buf[0] == '1') {
4932 ret = validate_slab_cache(s);
4938 SLAB_ATTR(validate);
4940 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4942 if (!(s->flags & SLAB_STORE_USER))
4944 return list_locations(s, buf, TRACK_ALLOC);
4946 SLAB_ATTR_RO(alloc_calls);
4948 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4950 if (!(s->flags & SLAB_STORE_USER))
4952 return list_locations(s, buf, TRACK_FREE);
4954 SLAB_ATTR_RO(free_calls);
4955 #endif /* CONFIG_SLUB_DEBUG */
4957 #ifdef CONFIG_FAILSLAB
4958 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4960 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4963 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4966 s->flags &= ~SLAB_FAILSLAB;
4968 s->flags |= SLAB_FAILSLAB;
4971 SLAB_ATTR(failslab);
4974 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4979 static ssize_t shrink_store(struct kmem_cache *s,
4980 const char *buf, size_t length)
4982 if (buf[0] == '1') {
4983 int rc = kmem_cache_shrink(s);
4994 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4996 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4999 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5000 const char *buf, size_t length)
5002 unsigned long ratio;
5005 err = strict_strtoul(buf, 10, &ratio);
5010 s->remote_node_defrag_ratio = ratio * 10;
5014 SLAB_ATTR(remote_node_defrag_ratio);
5017 #ifdef CONFIG_SLUB_STATS
5018 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5020 unsigned long sum = 0;
5023 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5028 for_each_online_cpu(cpu) {
5029 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5035 len = sprintf(buf, "%lu", sum);
5038 for_each_online_cpu(cpu) {
5039 if (data[cpu] && len < PAGE_SIZE - 20)
5040 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5044 return len + sprintf(buf + len, "\n");
5047 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5051 for_each_online_cpu(cpu)
5052 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5055 #define STAT_ATTR(si, text) \
5056 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5058 return show_stat(s, buf, si); \
5060 static ssize_t text##_store(struct kmem_cache *s, \
5061 const char *buf, size_t length) \
5063 if (buf[0] != '0') \
5065 clear_stat(s, si); \
5070 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5071 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5072 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5073 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5074 STAT_ATTR(FREE_FROZEN, free_frozen);
5075 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5076 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5077 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5078 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5079 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5080 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5081 STAT_ATTR(FREE_SLAB, free_slab);
5082 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5083 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5084 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5085 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5086 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5087 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5088 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5089 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5090 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5091 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5092 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5093 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5094 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5095 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5098 static struct attribute *slab_attrs[] = {
5099 &slab_size_attr.attr,
5100 &object_size_attr.attr,
5101 &objs_per_slab_attr.attr,
5103 &min_partial_attr.attr,
5104 &cpu_partial_attr.attr,
5106 &objects_partial_attr.attr,
5108 &cpu_slabs_attr.attr,
5112 &hwcache_align_attr.attr,
5113 &reclaim_account_attr.attr,
5114 &destroy_by_rcu_attr.attr,
5116 &reserved_attr.attr,
5117 &slabs_cpu_partial_attr.attr,
5118 #ifdef CONFIG_SLUB_DEBUG
5119 &total_objects_attr.attr,
5121 &sanity_checks_attr.attr,
5123 &red_zone_attr.attr,
5125 &store_user_attr.attr,
5126 &validate_attr.attr,
5127 &alloc_calls_attr.attr,
5128 &free_calls_attr.attr,
5130 #ifdef CONFIG_ZONE_DMA
5131 &cache_dma_attr.attr,
5134 &remote_node_defrag_ratio_attr.attr,
5136 #ifdef CONFIG_SLUB_STATS
5137 &alloc_fastpath_attr.attr,
5138 &alloc_slowpath_attr.attr,
5139 &free_fastpath_attr.attr,
5140 &free_slowpath_attr.attr,
5141 &free_frozen_attr.attr,
5142 &free_add_partial_attr.attr,
5143 &free_remove_partial_attr.attr,
5144 &alloc_from_partial_attr.attr,
5145 &alloc_slab_attr.attr,
5146 &alloc_refill_attr.attr,
5147 &alloc_node_mismatch_attr.attr,
5148 &free_slab_attr.attr,
5149 &cpuslab_flush_attr.attr,
5150 &deactivate_full_attr.attr,
5151 &deactivate_empty_attr.attr,
5152 &deactivate_to_head_attr.attr,
5153 &deactivate_to_tail_attr.attr,
5154 &deactivate_remote_frees_attr.attr,
5155 &deactivate_bypass_attr.attr,
5156 &order_fallback_attr.attr,
5157 &cmpxchg_double_fail_attr.attr,
5158 &cmpxchg_double_cpu_fail_attr.attr,
5159 &cpu_partial_alloc_attr.attr,
5160 &cpu_partial_free_attr.attr,
5161 &cpu_partial_node_attr.attr,
5162 &cpu_partial_drain_attr.attr,
5164 #ifdef CONFIG_FAILSLAB
5165 &failslab_attr.attr,
5171 static struct attribute_group slab_attr_group = {
5172 .attrs = slab_attrs,
5175 static ssize_t slab_attr_show(struct kobject *kobj,
5176 struct attribute *attr,
5179 struct slab_attribute *attribute;
5180 struct kmem_cache *s;
5183 attribute = to_slab_attr(attr);
5186 if (!attribute->show)
5189 err = attribute->show(s, buf);
5194 static ssize_t slab_attr_store(struct kobject *kobj,
5195 struct attribute *attr,
5196 const char *buf, size_t len)
5198 struct slab_attribute *attribute;
5199 struct kmem_cache *s;
5202 attribute = to_slab_attr(attr);
5205 if (!attribute->store)
5208 err = attribute->store(s, buf, len);
5213 static void kmem_cache_release(struct kobject *kobj)
5215 struct kmem_cache *s = to_slab(kobj);
5221 static const struct sysfs_ops slab_sysfs_ops = {
5222 .show = slab_attr_show,
5223 .store = slab_attr_store,
5226 static struct kobj_type slab_ktype = {
5227 .sysfs_ops = &slab_sysfs_ops,
5228 .release = kmem_cache_release
5231 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5233 struct kobj_type *ktype = get_ktype(kobj);
5235 if (ktype == &slab_ktype)
5240 static const struct kset_uevent_ops slab_uevent_ops = {
5241 .filter = uevent_filter,
5244 static struct kset *slab_kset;
5246 #define ID_STR_LENGTH 64
5248 /* Create a unique string id for a slab cache:
5250 * Format :[flags-]size
5252 static char *create_unique_id(struct kmem_cache *s)
5254 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5261 * First flags affecting slabcache operations. We will only
5262 * get here for aliasable slabs so we do not need to support
5263 * too many flags. The flags here must cover all flags that
5264 * are matched during merging to guarantee that the id is
5267 if (s->flags & SLAB_CACHE_DMA)
5269 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5271 if (s->flags & SLAB_DEBUG_FREE)
5273 if (!(s->flags & SLAB_NOTRACK))
5277 p += sprintf(p, "%07d", s->size);
5278 BUG_ON(p > name + ID_STR_LENGTH - 1);
5282 static int sysfs_slab_add(struct kmem_cache *s)
5288 if (slab_state < FULL)
5289 /* Defer until later */
5292 unmergeable = slab_unmergeable(s);
5295 * Slabcache can never be merged so we can use the name proper.
5296 * This is typically the case for debug situations. In that
5297 * case we can catch duplicate names easily.
5299 sysfs_remove_link(&slab_kset->kobj, s->name);
5303 * Create a unique name for the slab as a target
5306 name = create_unique_id(s);
5309 s->kobj.kset = slab_kset;
5310 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5312 kobject_put(&s->kobj);
5316 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5318 kobject_del(&s->kobj);
5319 kobject_put(&s->kobj);
5322 kobject_uevent(&s->kobj, KOBJ_ADD);
5324 /* Setup first alias */
5325 sysfs_slab_alias(s, s->name);
5331 static void sysfs_slab_remove(struct kmem_cache *s)
5333 if (slab_state < FULL)
5335 * Sysfs has not been setup yet so no need to remove the
5340 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5341 kobject_del(&s->kobj);
5342 kobject_put(&s->kobj);
5346 * Need to buffer aliases during bootup until sysfs becomes
5347 * available lest we lose that information.
5349 struct saved_alias {
5350 struct kmem_cache *s;
5352 struct saved_alias *next;
5355 static struct saved_alias *alias_list;
5357 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5359 struct saved_alias *al;
5361 if (slab_state == FULL) {
5363 * If we have a leftover link then remove it.
5365 sysfs_remove_link(&slab_kset->kobj, name);
5366 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5369 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5375 al->next = alias_list;
5380 static int __init slab_sysfs_init(void)
5382 struct kmem_cache *s;
5385 mutex_lock(&slab_mutex);
5387 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5389 mutex_unlock(&slab_mutex);
5390 printk(KERN_ERR "Cannot register slab subsystem.\n");
5396 list_for_each_entry(s, &slab_caches, list) {
5397 err = sysfs_slab_add(s);
5399 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5400 " to sysfs\n", s->name);
5403 while (alias_list) {
5404 struct saved_alias *al = alias_list;
5406 alias_list = alias_list->next;
5407 err = sysfs_slab_alias(al->s, al->name);
5409 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5410 " %s to sysfs\n", al->name);
5414 mutex_unlock(&slab_mutex);
5419 __initcall(slab_sysfs_init);
5420 #endif /* CONFIG_SYSFS */
5423 * The /proc/slabinfo ABI
5425 #ifdef CONFIG_SLABINFO
5426 static void print_slabinfo_header(struct seq_file *m)
5428 seq_puts(m, "slabinfo - version: 2.1\n");
5429 seq_puts(m, "# name <active_objs> <num_objs> <object_size> "
5430 "<objperslab> <pagesperslab>");
5431 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5432 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5436 static void *s_start(struct seq_file *m, loff_t *pos)
5440 mutex_lock(&slab_mutex);
5442 print_slabinfo_header(m);
5444 return seq_list_start(&slab_caches, *pos);
5447 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5449 return seq_list_next(p, &slab_caches, pos);
5452 static void s_stop(struct seq_file *m, void *p)
5454 mutex_unlock(&slab_mutex);
5457 static int s_show(struct seq_file *m, void *p)
5459 unsigned long nr_partials = 0;
5460 unsigned long nr_slabs = 0;
5461 unsigned long nr_inuse = 0;
5462 unsigned long nr_objs = 0;
5463 unsigned long nr_free = 0;
5464 struct kmem_cache *s;
5467 s = list_entry(p, struct kmem_cache, list);
5469 for_each_online_node(node) {
5470 struct kmem_cache_node *n = get_node(s, node);
5475 nr_partials += n->nr_partial;
5476 nr_slabs += atomic_long_read(&n->nr_slabs);
5477 nr_objs += atomic_long_read(&n->total_objects);
5478 nr_free += count_partial(n, count_free);
5481 nr_inuse = nr_objs - nr_free;
5483 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5484 nr_objs, s->size, oo_objects(s->oo),
5485 (1 << oo_order(s->oo)));
5486 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5487 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5493 static const struct seq_operations slabinfo_op = {
5500 static int slabinfo_open(struct inode *inode, struct file *file)
5502 return seq_open(file, &slabinfo_op);
5505 static const struct file_operations proc_slabinfo_operations = {
5506 .open = slabinfo_open,
5508 .llseek = seq_lseek,
5509 .release = seq_release,
5512 static int __init slab_proc_init(void)
5514 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5517 module_init(slab_proc_init);
5518 #endif /* CONFIG_SLABINFO */