2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
31 #include <linux/stacktrace.h>
32 #include <linux/prefetch.h>
34 #include <trace/events/kmem.h>
38 * 1. slub_lock (Global Semaphore)
40 * 3. slab_lock(page) (Only on some arches and for debugging)
44 * The role of the slub_lock is to protect the list of all the slabs
45 * and to synchronize major metadata changes to slab cache structures.
47 * The slab_lock is only used for debugging and on arches that do not
48 * have the ability to do a cmpxchg_double. It only protects the second
49 * double word in the page struct. Meaning
50 * A. page->freelist -> List of object free in a page
51 * B. page->counters -> Counters of objects
52 * C. page->frozen -> frozen state
54 * If a slab is frozen then it is exempt from list management. It is not
55 * on any list. The processor that froze the slab is the one who can
56 * perform list operations on the page. Other processors may put objects
57 * onto the freelist but the processor that froze the slab is the only
58 * one that can retrieve the objects from the page's freelist.
60 * The list_lock protects the partial and full list on each node and
61 * the partial slab counter. If taken then no new slabs may be added or
62 * removed from the lists nor make the number of partial slabs be modified.
63 * (Note that the total number of slabs is an atomic value that may be
64 * modified without taking the list lock).
66 * The list_lock is a centralized lock and thus we avoid taking it as
67 * much as possible. As long as SLUB does not have to handle partial
68 * slabs, operations can continue without any centralized lock. F.e.
69 * allocating a long series of objects that fill up slabs does not require
71 * Interrupts are disabled during allocation and deallocation in order to
72 * make the slab allocator safe to use in the context of an irq. In addition
73 * interrupts are disabled to ensure that the processor does not change
74 * while handling per_cpu slabs, due to kernel preemption.
76 * SLUB assigns one slab for allocation to each processor.
77 * Allocations only occur from these slabs called cpu slabs.
79 * Slabs with free elements are kept on a partial list and during regular
80 * operations no list for full slabs is used. If an object in a full slab is
81 * freed then the slab will show up again on the partial lists.
82 * We track full slabs for debugging purposes though because otherwise we
83 * cannot scan all objects.
85 * Slabs are freed when they become empty. Teardown and setup is
86 * minimal so we rely on the page allocators per cpu caches for
87 * fast frees and allocs.
89 * Overloading of page flags that are otherwise used for LRU management.
91 * PageActive The slab is frozen and exempt from list processing.
92 * This means that the slab is dedicated to a purpose
93 * such as satisfying allocations for a specific
94 * processor. Objects may be freed in the slab while
95 * it is frozen but slab_free will then skip the usual
96 * list operations. It is up to the processor holding
97 * the slab to integrate the slab into the slab lists
98 * when the slab is no longer needed.
100 * One use of this flag is to mark slabs that are
101 * used for allocations. Then such a slab becomes a cpu
102 * slab. The cpu slab may be equipped with an additional
103 * freelist that allows lockless access to
104 * free objects in addition to the regular freelist
105 * that requires the slab lock.
107 * PageError Slab requires special handling due to debug
108 * options set. This moves slab handling out of
109 * the fast path and disables lockless freelists.
112 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
113 SLAB_TRACE | SLAB_DEBUG_FREE)
115 static inline int kmem_cache_debug(struct kmem_cache *s)
117 #ifdef CONFIG_SLUB_DEBUG
118 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
125 * Issues still to be resolved:
127 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
129 * - Variable sizing of the per node arrays
132 /* Enable to test recovery from slab corruption on boot */
133 #undef SLUB_RESILIENCY_TEST
135 /* Enable to log cmpxchg failures */
136 #undef SLUB_DEBUG_CMPXCHG
139 * Mininum number of partial slabs. These will be left on the partial
140 * lists even if they are empty. kmem_cache_shrink may reclaim them.
142 #define MIN_PARTIAL 5
145 * Maximum number of desirable partial slabs.
146 * The existence of more partial slabs makes kmem_cache_shrink
147 * sort the partial list by the number of objects in the.
149 #define MAX_PARTIAL 10
151 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
152 SLAB_POISON | SLAB_STORE_USER)
155 * Debugging flags that require metadata to be stored in the slab. These get
156 * disabled when slub_debug=O is used and a cache's min order increases with
159 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
162 * Set of flags that will prevent slab merging
164 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
165 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
168 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
169 SLAB_CACHE_DMA | SLAB_NOTRACK)
172 #define OO_MASK ((1 << OO_SHIFT) - 1)
173 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
175 /* Internal SLUB flags */
176 #define __OBJECT_POISON 0x80000000UL /* Poison object */
177 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
179 static int kmem_size = sizeof(struct kmem_cache);
182 static struct notifier_block slab_notifier;
186 DOWN, /* No slab functionality available */
187 PARTIAL, /* Kmem_cache_node works */
188 UP, /* Everything works but does not show up in sysfs */
192 /* A list of all slab caches on the system */
193 static DECLARE_RWSEM(slub_lock);
194 static LIST_HEAD(slab_caches);
197 * Tracking user of a slab.
199 #define TRACK_ADDRS_COUNT 16
201 unsigned long addr; /* Called from address */
202 #ifdef CONFIG_STACKTRACE
203 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
205 int cpu; /* Was running on cpu */
206 int pid; /* Pid context */
207 unsigned long when; /* When did the operation occur */
210 enum track_item { TRACK_ALLOC, TRACK_FREE };
213 static int sysfs_slab_add(struct kmem_cache *);
214 static int sysfs_slab_alias(struct kmem_cache *, const char *);
215 static void sysfs_slab_remove(struct kmem_cache *);
218 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
219 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
221 static inline void sysfs_slab_remove(struct kmem_cache *s)
229 static inline void stat(const struct kmem_cache *s, enum stat_item si)
231 #ifdef CONFIG_SLUB_STATS
232 __this_cpu_inc(s->cpu_slab->stat[si]);
236 /********************************************************************
237 * Core slab cache functions
238 *******************************************************************/
240 int slab_is_available(void)
242 return slab_state >= UP;
245 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
247 return s->node[node];
250 /* Verify that a pointer has an address that is valid within a slab page */
251 static inline int check_valid_pointer(struct kmem_cache *s,
252 struct page *page, const void *object)
259 base = page_address(page);
260 if (object < base || object >= base + page->objects * s->size ||
261 (object - base) % s->size) {
268 static inline void *get_freepointer(struct kmem_cache *s, void *object)
270 return *(void **)(object + s->offset);
273 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
275 prefetch(object + s->offset);
278 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
282 #ifdef CONFIG_DEBUG_PAGEALLOC
283 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
285 p = get_freepointer(s, object);
290 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
292 *(void **)(object + s->offset) = fp;
295 /* Loop over all objects in a slab */
296 #define for_each_object(__p, __s, __addr, __objects) \
297 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
300 /* Determine object index from a given position */
301 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
303 return (p - addr) / s->size;
306 static inline size_t slab_ksize(const struct kmem_cache *s)
308 #ifdef CONFIG_SLUB_DEBUG
310 * Debugging requires use of the padding between object
311 * and whatever may come after it.
313 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
318 * If we have the need to store the freelist pointer
319 * back there or track user information then we can
320 * only use the space before that information.
322 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
325 * Else we can use all the padding etc for the allocation
330 static inline int order_objects(int order, unsigned long size, int reserved)
332 return ((PAGE_SIZE << order) - reserved) / size;
335 static inline struct kmem_cache_order_objects oo_make(int order,
336 unsigned long size, int reserved)
338 struct kmem_cache_order_objects x = {
339 (order << OO_SHIFT) + order_objects(order, size, reserved)
345 static inline int oo_order(struct kmem_cache_order_objects x)
347 return x.x >> OO_SHIFT;
350 static inline int oo_objects(struct kmem_cache_order_objects x)
352 return x.x & OO_MASK;
356 * Per slab locking using the pagelock
358 static __always_inline void slab_lock(struct page *page)
360 bit_spin_lock(PG_locked, &page->flags);
363 static __always_inline void slab_unlock(struct page *page)
365 __bit_spin_unlock(PG_locked, &page->flags);
368 /* Interrupts must be disabled (for the fallback code to work right) */
369 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
370 void *freelist_old, unsigned long counters_old,
371 void *freelist_new, unsigned long counters_new,
374 VM_BUG_ON(!irqs_disabled());
375 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
376 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
377 if (s->flags & __CMPXCHG_DOUBLE) {
378 if (cmpxchg_double(&page->freelist, &page->counters,
379 freelist_old, counters_old,
380 freelist_new, counters_new))
386 if (page->freelist == freelist_old && page->counters == counters_old) {
387 page->freelist = freelist_new;
388 page->counters = counters_new;
396 stat(s, CMPXCHG_DOUBLE_FAIL);
398 #ifdef SLUB_DEBUG_CMPXCHG
399 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
405 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
406 void *freelist_old, unsigned long counters_old,
407 void *freelist_new, unsigned long counters_new,
410 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
411 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
412 if (s->flags & __CMPXCHG_DOUBLE) {
413 if (cmpxchg_double(&page->freelist, &page->counters,
414 freelist_old, counters_old,
415 freelist_new, counters_new))
422 local_irq_save(flags);
424 if (page->freelist == freelist_old && page->counters == counters_old) {
425 page->freelist = freelist_new;
426 page->counters = counters_new;
428 local_irq_restore(flags);
432 local_irq_restore(flags);
436 stat(s, CMPXCHG_DOUBLE_FAIL);
438 #ifdef SLUB_DEBUG_CMPXCHG
439 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
445 #ifdef CONFIG_SLUB_DEBUG
447 * Determine a map of object in use on a page.
449 * Node listlock must be held to guarantee that the page does
450 * not vanish from under us.
452 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
455 void *addr = page_address(page);
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(slab_index(p, s, addr), map);
464 #ifdef CONFIG_SLUB_DEBUG_ON
465 static int slub_debug = DEBUG_DEFAULT_FLAGS;
467 static int slub_debug;
470 static char *slub_debug_slabs;
471 static int disable_higher_order_debug;
476 static void print_section(char *text, u8 *addr, unsigned int length)
478 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
482 static struct track *get_track(struct kmem_cache *s, void *object,
483 enum track_item alloc)
488 p = object + s->offset + sizeof(void *);
490 p = object + s->inuse;
495 static void set_track(struct kmem_cache *s, void *object,
496 enum track_item alloc, unsigned long addr)
498 struct track *p = get_track(s, object, alloc);
501 #ifdef CONFIG_STACKTRACE
502 struct stack_trace trace;
505 trace.nr_entries = 0;
506 trace.max_entries = TRACK_ADDRS_COUNT;
507 trace.entries = p->addrs;
509 save_stack_trace(&trace);
511 /* See rant in lockdep.c */
512 if (trace.nr_entries != 0 &&
513 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
516 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
520 p->cpu = smp_processor_id();
521 p->pid = current->pid;
524 memset(p, 0, sizeof(struct track));
527 static void init_tracking(struct kmem_cache *s, void *object)
529 if (!(s->flags & SLAB_STORE_USER))
532 set_track(s, object, TRACK_FREE, 0UL);
533 set_track(s, object, TRACK_ALLOC, 0UL);
536 static void print_track(const char *s, struct track *t)
541 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
542 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
543 #ifdef CONFIG_STACKTRACE
546 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
548 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
555 static void print_tracking(struct kmem_cache *s, void *object)
557 if (!(s->flags & SLAB_STORE_USER))
560 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
561 print_track("Freed", get_track(s, object, TRACK_FREE));
564 static void print_page_info(struct page *page)
566 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
567 page, page->objects, page->inuse, page->freelist, page->flags);
571 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
577 vsnprintf(buf, sizeof(buf), fmt, args);
579 printk(KERN_ERR "========================================"
580 "=====================================\n");
581 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
582 printk(KERN_ERR "----------------------------------------"
583 "-------------------------------------\n\n");
586 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
592 vsnprintf(buf, sizeof(buf), fmt, args);
594 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
597 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
599 unsigned int off; /* Offset of last byte */
600 u8 *addr = page_address(page);
602 print_tracking(s, p);
604 print_page_info(page);
606 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607 p, p - addr, get_freepointer(s, p));
610 print_section("Bytes b4 ", p - 16, 16);
612 print_section("Object ", p, min_t(unsigned long, s->objsize,
614 if (s->flags & SLAB_RED_ZONE)
615 print_section("Redzone ", p + s->objsize,
616 s->inuse - s->objsize);
619 off = s->offset + sizeof(void *);
623 if (s->flags & SLAB_STORE_USER)
624 off += 2 * sizeof(struct track);
627 /* Beginning of the filler is the free pointer */
628 print_section("Padding ", p + off, s->size - off);
633 static void object_err(struct kmem_cache *s, struct page *page,
634 u8 *object, char *reason)
636 slab_bug(s, "%s", reason);
637 print_trailer(s, page, object);
640 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
646 vsnprintf(buf, sizeof(buf), fmt, args);
648 slab_bug(s, "%s", buf);
649 print_page_info(page);
653 static void init_object(struct kmem_cache *s, void *object, u8 val)
657 if (s->flags & __OBJECT_POISON) {
658 memset(p, POISON_FREE, s->objsize - 1);
659 p[s->objsize - 1] = POISON_END;
662 if (s->flags & SLAB_RED_ZONE)
663 memset(p + s->objsize, val, s->inuse - s->objsize);
666 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667 void *from, void *to)
669 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670 memset(from, data, to - from);
673 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674 u8 *object, char *what,
675 u8 *start, unsigned int value, unsigned int bytes)
680 fault = memchr_inv(start, value, bytes);
685 while (end > fault && end[-1] == value)
688 slab_bug(s, "%s overwritten", what);
689 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690 fault, end - 1, fault[0], value);
691 print_trailer(s, page, object);
693 restore_bytes(s, what, value, fault, end);
701 * Bytes of the object to be managed.
702 * If the freepointer may overlay the object then the free
703 * pointer is the first word of the object.
705 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
708 * object + s->objsize
709 * Padding to reach word boundary. This is also used for Redzoning.
710 * Padding is extended by another word if Redzoning is enabled and
713 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714 * 0xcc (RED_ACTIVE) for objects in use.
717 * Meta data starts here.
719 * A. Free pointer (if we cannot overwrite object on free)
720 * B. Tracking data for SLAB_STORE_USER
721 * C. Padding to reach required alignment boundary or at mininum
722 * one word if debugging is on to be able to detect writes
723 * before the word boundary.
725 * Padding is done using 0x5a (POISON_INUSE)
728 * Nothing is used beyond s->size.
730 * If slabcaches are merged then the objsize and inuse boundaries are mostly
731 * ignored. And therefore no slab options that rely on these boundaries
732 * may be used with merged slabcaches.
735 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
737 unsigned long off = s->inuse; /* The end of info */
740 /* Freepointer is placed after the object. */
741 off += sizeof(void *);
743 if (s->flags & SLAB_STORE_USER)
744 /* We also have user information there */
745 off += 2 * sizeof(struct track);
750 return check_bytes_and_report(s, page, p, "Object padding",
751 p + off, POISON_INUSE, s->size - off);
754 /* Check the pad bytes at the end of a slab page */
755 static int slab_pad_check(struct kmem_cache *s, struct page *page)
763 if (!(s->flags & SLAB_POISON))
766 start = page_address(page);
767 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768 end = start + length;
769 remainder = length % s->size;
773 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
776 while (end > fault && end[-1] == POISON_INUSE)
779 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780 print_section("Padding ", end - remainder, remainder);
782 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
786 static int check_object(struct kmem_cache *s, struct page *page,
787 void *object, u8 val)
790 u8 *endobject = object + s->objsize;
792 if (s->flags & SLAB_RED_ZONE) {
793 if (!check_bytes_and_report(s, page, object, "Redzone",
794 endobject, val, s->inuse - s->objsize))
797 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
798 check_bytes_and_report(s, page, p, "Alignment padding",
799 endobject, POISON_INUSE, s->inuse - s->objsize);
803 if (s->flags & SLAB_POISON) {
804 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
805 (!check_bytes_and_report(s, page, p, "Poison", p,
806 POISON_FREE, s->objsize - 1) ||
807 !check_bytes_and_report(s, page, p, "Poison",
808 p + s->objsize - 1, POISON_END, 1)))
811 * check_pad_bytes cleans up on its own.
813 check_pad_bytes(s, page, p);
816 if (!s->offset && val == SLUB_RED_ACTIVE)
818 * Object and freepointer overlap. Cannot check
819 * freepointer while object is allocated.
823 /* Check free pointer validity */
824 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
825 object_err(s, page, p, "Freepointer corrupt");
827 * No choice but to zap it and thus lose the remainder
828 * of the free objects in this slab. May cause
829 * another error because the object count is now wrong.
831 set_freepointer(s, p, NULL);
837 static int check_slab(struct kmem_cache *s, struct page *page)
841 VM_BUG_ON(!irqs_disabled());
843 if (!PageSlab(page)) {
844 slab_err(s, page, "Not a valid slab page");
848 maxobj = order_objects(compound_order(page), s->size, s->reserved);
849 if (page->objects > maxobj) {
850 slab_err(s, page, "objects %u > max %u",
851 s->name, page->objects, maxobj);
854 if (page->inuse > page->objects) {
855 slab_err(s, page, "inuse %u > max %u",
856 s->name, page->inuse, page->objects);
859 /* Slab_pad_check fixes things up after itself */
860 slab_pad_check(s, page);
865 * Determine if a certain object on a page is on the freelist. Must hold the
866 * slab lock to guarantee that the chains are in a consistent state.
868 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
873 unsigned long max_objects;
876 while (fp && nr <= page->objects) {
879 if (!check_valid_pointer(s, page, fp)) {
881 object_err(s, page, object,
882 "Freechain corrupt");
883 set_freepointer(s, object, NULL);
886 slab_err(s, page, "Freepointer corrupt");
887 page->freelist = NULL;
888 page->inuse = page->objects;
889 slab_fix(s, "Freelist cleared");
895 fp = get_freepointer(s, object);
899 max_objects = order_objects(compound_order(page), s->size, s->reserved);
900 if (max_objects > MAX_OBJS_PER_PAGE)
901 max_objects = MAX_OBJS_PER_PAGE;
903 if (page->objects != max_objects) {
904 slab_err(s, page, "Wrong number of objects. Found %d but "
905 "should be %d", page->objects, max_objects);
906 page->objects = max_objects;
907 slab_fix(s, "Number of objects adjusted.");
909 if (page->inuse != page->objects - nr) {
910 slab_err(s, page, "Wrong object count. Counter is %d but "
911 "counted were %d", page->inuse, page->objects - nr);
912 page->inuse = page->objects - nr;
913 slab_fix(s, "Object count adjusted.");
915 return search == NULL;
918 static void trace(struct kmem_cache *s, struct page *page, void *object,
921 if (s->flags & SLAB_TRACE) {
922 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
924 alloc ? "alloc" : "free",
929 print_section("Object ", (void *)object, s->objsize);
936 * Hooks for other subsystems that check memory allocations. In a typical
937 * production configuration these hooks all should produce no code at all.
939 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
941 flags &= gfp_allowed_mask;
942 lockdep_trace_alloc(flags);
943 might_sleep_if(flags & __GFP_WAIT);
945 return should_failslab(s->objsize, flags, s->flags);
948 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
950 flags &= gfp_allowed_mask;
951 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
952 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
955 static inline void slab_free_hook(struct kmem_cache *s, void *x)
957 kmemleak_free_recursive(x, s->flags);
960 * Trouble is that we may no longer disable interupts in the fast path
961 * So in order to make the debug calls that expect irqs to be
962 * disabled we need to disable interrupts temporarily.
964 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
968 local_irq_save(flags);
969 kmemcheck_slab_free(s, x, s->objsize);
970 debug_check_no_locks_freed(x, s->objsize);
971 local_irq_restore(flags);
974 if (!(s->flags & SLAB_DEBUG_OBJECTS))
975 debug_check_no_obj_freed(x, s->objsize);
979 * Tracking of fully allocated slabs for debugging purposes.
981 * list_lock must be held.
983 static void add_full(struct kmem_cache *s,
984 struct kmem_cache_node *n, struct page *page)
986 if (!(s->flags & SLAB_STORE_USER))
989 list_add(&page->lru, &n->full);
993 * list_lock must be held.
995 static void remove_full(struct kmem_cache *s, struct page *page)
997 if (!(s->flags & SLAB_STORE_USER))
1000 list_del(&page->lru);
1003 /* Tracking of the number of slabs for debugging purposes */
1004 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1006 struct kmem_cache_node *n = get_node(s, node);
1008 return atomic_long_read(&n->nr_slabs);
1011 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1013 return atomic_long_read(&n->nr_slabs);
1016 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1018 struct kmem_cache_node *n = get_node(s, node);
1021 * May be called early in order to allocate a slab for the
1022 * kmem_cache_node structure. Solve the chicken-egg
1023 * dilemma by deferring the increment of the count during
1024 * bootstrap (see early_kmem_cache_node_alloc).
1027 atomic_long_inc(&n->nr_slabs);
1028 atomic_long_add(objects, &n->total_objects);
1031 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1033 struct kmem_cache_node *n = get_node(s, node);
1035 atomic_long_dec(&n->nr_slabs);
1036 atomic_long_sub(objects, &n->total_objects);
1039 /* Object debug checks for alloc/free paths */
1040 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1043 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1046 init_object(s, object, SLUB_RED_INACTIVE);
1047 init_tracking(s, object);
1050 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1051 void *object, unsigned long addr)
1053 if (!check_slab(s, page))
1056 if (!check_valid_pointer(s, page, object)) {
1057 object_err(s, page, object, "Freelist Pointer check fails");
1061 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1064 /* Success perform special debug activities for allocs */
1065 if (s->flags & SLAB_STORE_USER)
1066 set_track(s, object, TRACK_ALLOC, addr);
1067 trace(s, page, object, 1);
1068 init_object(s, object, SLUB_RED_ACTIVE);
1072 if (PageSlab(page)) {
1074 * If this is a slab page then lets do the best we can
1075 * to avoid issues in the future. Marking all objects
1076 * as used avoids touching the remaining objects.
1078 slab_fix(s, "Marking all objects used");
1079 page->inuse = page->objects;
1080 page->freelist = NULL;
1085 static noinline int free_debug_processing(struct kmem_cache *s,
1086 struct page *page, void *object, unsigned long addr)
1088 unsigned long flags;
1091 local_irq_save(flags);
1094 if (!check_slab(s, page))
1097 if (!check_valid_pointer(s, page, object)) {
1098 slab_err(s, page, "Invalid object pointer 0x%p", object);
1102 if (on_freelist(s, page, object)) {
1103 object_err(s, page, object, "Object already free");
1107 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1110 if (unlikely(s != page->slab)) {
1111 if (!PageSlab(page)) {
1112 slab_err(s, page, "Attempt to free object(0x%p) "
1113 "outside of slab", object);
1114 } else if (!page->slab) {
1116 "SLUB <none>: no slab for object 0x%p.\n",
1120 object_err(s, page, object,
1121 "page slab pointer corrupt.");
1125 if (s->flags & SLAB_STORE_USER)
1126 set_track(s, object, TRACK_FREE, addr);
1127 trace(s, page, object, 0);
1128 init_object(s, object, SLUB_RED_INACTIVE);
1132 local_irq_restore(flags);
1136 slab_fix(s, "Object at 0x%p not freed", object);
1140 static int __init setup_slub_debug(char *str)
1142 slub_debug = DEBUG_DEFAULT_FLAGS;
1143 if (*str++ != '=' || !*str)
1145 * No options specified. Switch on full debugging.
1151 * No options but restriction on slabs. This means full
1152 * debugging for slabs matching a pattern.
1156 if (tolower(*str) == 'o') {
1158 * Avoid enabling debugging on caches if its minimum order
1159 * would increase as a result.
1161 disable_higher_order_debug = 1;
1168 * Switch off all debugging measures.
1173 * Determine which debug features should be switched on
1175 for (; *str && *str != ','; str++) {
1176 switch (tolower(*str)) {
1178 slub_debug |= SLAB_DEBUG_FREE;
1181 slub_debug |= SLAB_RED_ZONE;
1184 slub_debug |= SLAB_POISON;
1187 slub_debug |= SLAB_STORE_USER;
1190 slub_debug |= SLAB_TRACE;
1193 slub_debug |= SLAB_FAILSLAB;
1196 printk(KERN_ERR "slub_debug option '%c' "
1197 "unknown. skipped\n", *str);
1203 slub_debug_slabs = str + 1;
1208 __setup("slub_debug", setup_slub_debug);
1210 static unsigned long kmem_cache_flags(unsigned long objsize,
1211 unsigned long flags, const char *name,
1212 void (*ctor)(void *))
1215 * Enable debugging if selected on the kernel commandline.
1217 if (slub_debug && (!slub_debug_slabs ||
1218 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1219 flags |= slub_debug;
1224 static inline void setup_object_debug(struct kmem_cache *s,
1225 struct page *page, void *object) {}
1227 static inline int alloc_debug_processing(struct kmem_cache *s,
1228 struct page *page, void *object, unsigned long addr) { return 0; }
1230 static inline int free_debug_processing(struct kmem_cache *s,
1231 struct page *page, void *object, unsigned long addr) { return 0; }
1233 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1235 static inline int check_object(struct kmem_cache *s, struct page *page,
1236 void *object, u8 val) { return 1; }
1237 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1238 struct page *page) {}
1239 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1240 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1241 unsigned long flags, const char *name,
1242 void (*ctor)(void *))
1246 #define slub_debug 0
1248 #define disable_higher_order_debug 0
1250 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1252 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1254 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1256 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1259 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1262 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1265 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1267 #endif /* CONFIG_SLUB_DEBUG */
1270 * Slab allocation and freeing
1272 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1273 struct kmem_cache_order_objects oo)
1275 int order = oo_order(oo);
1277 flags |= __GFP_NOTRACK;
1279 if (node == NUMA_NO_NODE)
1280 return alloc_pages(flags, order);
1282 return alloc_pages_exact_node(node, flags, order);
1285 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1288 struct kmem_cache_order_objects oo = s->oo;
1291 flags &= gfp_allowed_mask;
1293 if (flags & __GFP_WAIT)
1296 flags |= s->allocflags;
1299 * Let the initial higher-order allocation fail under memory pressure
1300 * so we fall-back to the minimum order allocation.
1302 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1304 page = alloc_slab_page(alloc_gfp, node, oo);
1305 if (unlikely(!page)) {
1308 * Allocation may have failed due to fragmentation.
1309 * Try a lower order alloc if possible
1311 page = alloc_slab_page(flags, node, oo);
1314 stat(s, ORDER_FALLBACK);
1317 if (flags & __GFP_WAIT)
1318 local_irq_disable();
1323 if (kmemcheck_enabled
1324 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1325 int pages = 1 << oo_order(oo);
1327 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1330 * Objects from caches that have a constructor don't get
1331 * cleared when they're allocated, so we need to do it here.
1334 kmemcheck_mark_uninitialized_pages(page, pages);
1336 kmemcheck_mark_unallocated_pages(page, pages);
1339 page->objects = oo_objects(oo);
1340 mod_zone_page_state(page_zone(page),
1341 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1342 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1348 static void setup_object(struct kmem_cache *s, struct page *page,
1351 setup_object_debug(s, page, object);
1352 if (unlikely(s->ctor))
1356 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1363 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1365 page = allocate_slab(s,
1366 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1370 inc_slabs_node(s, page_to_nid(page), page->objects);
1372 __SetPageSlab(page);
1374 start = page_address(page);
1376 if (unlikely(s->flags & SLAB_POISON))
1377 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1380 for_each_object(p, s, start, page->objects) {
1381 setup_object(s, page, last);
1382 set_freepointer(s, last, p);
1385 setup_object(s, page, last);
1386 set_freepointer(s, last, NULL);
1388 page->freelist = start;
1389 page->inuse = page->objects;
1395 static void __free_slab(struct kmem_cache *s, struct page *page)
1397 int order = compound_order(page);
1398 int pages = 1 << order;
1400 if (kmem_cache_debug(s)) {
1403 slab_pad_check(s, page);
1404 for_each_object(p, s, page_address(page),
1406 check_object(s, page, p, SLUB_RED_INACTIVE);
1409 kmemcheck_free_shadow(page, compound_order(page));
1411 mod_zone_page_state(page_zone(page),
1412 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1413 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1416 __ClearPageSlab(page);
1417 reset_page_mapcount(page);
1418 if (current->reclaim_state)
1419 current->reclaim_state->reclaimed_slab += pages;
1420 __free_pages(page, order);
1423 #define need_reserve_slab_rcu \
1424 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1426 static void rcu_free_slab(struct rcu_head *h)
1430 if (need_reserve_slab_rcu)
1431 page = virt_to_head_page(h);
1433 page = container_of((struct list_head *)h, struct page, lru);
1435 __free_slab(page->slab, page);
1438 static void free_slab(struct kmem_cache *s, struct page *page)
1440 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1441 struct rcu_head *head;
1443 if (need_reserve_slab_rcu) {
1444 int order = compound_order(page);
1445 int offset = (PAGE_SIZE << order) - s->reserved;
1447 VM_BUG_ON(s->reserved != sizeof(*head));
1448 head = page_address(page) + offset;
1451 * RCU free overloads the RCU head over the LRU
1453 head = (void *)&page->lru;
1456 call_rcu(head, rcu_free_slab);
1458 __free_slab(s, page);
1461 static void discard_slab(struct kmem_cache *s, struct page *page)
1463 dec_slabs_node(s, page_to_nid(page), page->objects);
1468 * Management of partially allocated slabs.
1470 * list_lock must be held.
1472 static inline void add_partial(struct kmem_cache_node *n,
1473 struct page *page, int tail)
1476 if (tail == DEACTIVATE_TO_TAIL)
1477 list_add_tail(&page->lru, &n->partial);
1479 list_add(&page->lru, &n->partial);
1483 * list_lock must be held.
1485 static inline void remove_partial(struct kmem_cache_node *n,
1488 list_del(&page->lru);
1493 * Remove slab from the partial list, freeze it and
1494 * return the pointer to the freelist.
1496 * Returns a list of objects or NULL if it fails.
1498 * Must hold list_lock since we modify the partial list.
1500 static inline void *acquire_slab(struct kmem_cache *s,
1501 struct kmem_cache_node *n, struct page *page,
1505 unsigned long counters;
1509 * Zap the freelist and set the frozen bit.
1510 * The old freelist is the list of objects for the
1511 * per cpu allocation list.
1513 freelist = page->freelist;
1514 counters = page->counters;
1515 new.counters = counters;
1517 new.inuse = page->objects;
1518 new.freelist = NULL;
1520 new.freelist = freelist;
1523 VM_BUG_ON(new.frozen);
1526 if (!__cmpxchg_double_slab(s, page,
1528 new.freelist, new.counters,
1532 remove_partial(n, page);
1537 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1540 * Try to allocate a partial slab from a specific node.
1542 static void *get_partial_node(struct kmem_cache *s,
1543 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1545 struct page *page, *page2;
1546 void *object = NULL;
1549 * Racy check. If we mistakenly see no partial slabs then we
1550 * just allocate an empty slab. If we mistakenly try to get a
1551 * partial slab and there is none available then get_partials()
1554 if (!n || !n->nr_partial)
1557 spin_lock(&n->list_lock);
1558 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1559 void *t = acquire_slab(s, n, page, object == NULL);
1567 stat(s, ALLOC_FROM_PARTIAL);
1569 available = page->objects - page->inuse;
1571 available = put_cpu_partial(s, page, 0);
1572 stat(s, CPU_PARTIAL_NODE);
1574 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1578 spin_unlock(&n->list_lock);
1583 * Get a page from somewhere. Search in increasing NUMA distances.
1585 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1586 struct kmem_cache_cpu *c)
1589 struct zonelist *zonelist;
1592 enum zone_type high_zoneidx = gfp_zone(flags);
1594 unsigned int cpuset_mems_cookie;
1597 * The defrag ratio allows a configuration of the tradeoffs between
1598 * inter node defragmentation and node local allocations. A lower
1599 * defrag_ratio increases the tendency to do local allocations
1600 * instead of attempting to obtain partial slabs from other nodes.
1602 * If the defrag_ratio is set to 0 then kmalloc() always
1603 * returns node local objects. If the ratio is higher then kmalloc()
1604 * may return off node objects because partial slabs are obtained
1605 * from other nodes and filled up.
1607 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1608 * defrag_ratio = 1000) then every (well almost) allocation will
1609 * first attempt to defrag slab caches on other nodes. This means
1610 * scanning over all nodes to look for partial slabs which may be
1611 * expensive if we do it every time we are trying to find a slab
1612 * with available objects.
1614 if (!s->remote_node_defrag_ratio ||
1615 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1619 cpuset_mems_cookie = get_mems_allowed();
1620 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1621 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1622 struct kmem_cache_node *n;
1624 n = get_node(s, zone_to_nid(zone));
1626 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1627 n->nr_partial > s->min_partial) {
1628 object = get_partial_node(s, n, c);
1631 * Return the object even if
1632 * put_mems_allowed indicated that
1633 * the cpuset mems_allowed was
1634 * updated in parallel. It's a
1635 * harmless race between the alloc
1636 * and the cpuset update.
1638 put_mems_allowed(cpuset_mems_cookie);
1643 } while (!put_mems_allowed(cpuset_mems_cookie));
1649 * Get a partial page, lock it and return it.
1651 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1652 struct kmem_cache_cpu *c)
1655 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1657 object = get_partial_node(s, get_node(s, searchnode), c);
1658 if (object || node != NUMA_NO_NODE)
1661 return get_any_partial(s, flags, c);
1664 #ifdef CONFIG_PREEMPT
1666 * Calculate the next globally unique transaction for disambiguiation
1667 * during cmpxchg. The transactions start with the cpu number and are then
1668 * incremented by CONFIG_NR_CPUS.
1670 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1673 * No preemption supported therefore also no need to check for
1679 static inline unsigned long next_tid(unsigned long tid)
1681 return tid + TID_STEP;
1684 static inline unsigned int tid_to_cpu(unsigned long tid)
1686 return tid % TID_STEP;
1689 static inline unsigned long tid_to_event(unsigned long tid)
1691 return tid / TID_STEP;
1694 static inline unsigned int init_tid(int cpu)
1699 static inline void note_cmpxchg_failure(const char *n,
1700 const struct kmem_cache *s, unsigned long tid)
1702 #ifdef SLUB_DEBUG_CMPXCHG
1703 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1705 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1707 #ifdef CONFIG_PREEMPT
1708 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1709 printk("due to cpu change %d -> %d\n",
1710 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1713 if (tid_to_event(tid) != tid_to_event(actual_tid))
1714 printk("due to cpu running other code. Event %ld->%ld\n",
1715 tid_to_event(tid), tid_to_event(actual_tid));
1717 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1718 actual_tid, tid, next_tid(tid));
1720 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1723 void init_kmem_cache_cpus(struct kmem_cache *s)
1727 for_each_possible_cpu(cpu)
1728 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1732 * Remove the cpu slab
1734 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1736 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1737 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1739 enum slab_modes l = M_NONE, m = M_NONE;
1741 int tail = DEACTIVATE_TO_HEAD;
1745 if (page->freelist) {
1746 stat(s, DEACTIVATE_REMOTE_FREES);
1747 tail = DEACTIVATE_TO_TAIL;
1751 * Stage one: Free all available per cpu objects back
1752 * to the page freelist while it is still frozen. Leave the
1755 * There is no need to take the list->lock because the page
1758 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1760 unsigned long counters;
1763 prior = page->freelist;
1764 counters = page->counters;
1765 set_freepointer(s, freelist, prior);
1766 new.counters = counters;
1768 VM_BUG_ON(!new.frozen);
1770 } while (!__cmpxchg_double_slab(s, page,
1772 freelist, new.counters,
1773 "drain percpu freelist"));
1775 freelist = nextfree;
1779 * Stage two: Ensure that the page is unfrozen while the
1780 * list presence reflects the actual number of objects
1783 * We setup the list membership and then perform a cmpxchg
1784 * with the count. If there is a mismatch then the page
1785 * is not unfrozen but the page is on the wrong list.
1787 * Then we restart the process which may have to remove
1788 * the page from the list that we just put it on again
1789 * because the number of objects in the slab may have
1794 old.freelist = page->freelist;
1795 old.counters = page->counters;
1796 VM_BUG_ON(!old.frozen);
1798 /* Determine target state of the slab */
1799 new.counters = old.counters;
1802 set_freepointer(s, freelist, old.freelist);
1803 new.freelist = freelist;
1805 new.freelist = old.freelist;
1809 if (!new.inuse && n->nr_partial > s->min_partial)
1811 else if (new.freelist) {
1816 * Taking the spinlock removes the possiblity
1817 * that acquire_slab() will see a slab page that
1820 spin_lock(&n->list_lock);
1824 if (kmem_cache_debug(s) && !lock) {
1827 * This also ensures that the scanning of full
1828 * slabs from diagnostic functions will not see
1831 spin_lock(&n->list_lock);
1839 remove_partial(n, page);
1841 else if (l == M_FULL)
1843 remove_full(s, page);
1845 if (m == M_PARTIAL) {
1847 add_partial(n, page, tail);
1850 } else if (m == M_FULL) {
1852 stat(s, DEACTIVATE_FULL);
1853 add_full(s, n, page);
1859 if (!__cmpxchg_double_slab(s, page,
1860 old.freelist, old.counters,
1861 new.freelist, new.counters,
1866 spin_unlock(&n->list_lock);
1869 stat(s, DEACTIVATE_EMPTY);
1870 discard_slab(s, page);
1875 /* Unfreeze all the cpu partial slabs */
1876 static void unfreeze_partials(struct kmem_cache *s)
1878 struct kmem_cache_node *n = NULL;
1879 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1880 struct page *page, *discard_page = NULL;
1882 while ((page = c->partial)) {
1883 enum slab_modes { M_PARTIAL, M_FREE };
1884 enum slab_modes l, m;
1888 c->partial = page->next;
1893 old.freelist = page->freelist;
1894 old.counters = page->counters;
1895 VM_BUG_ON(!old.frozen);
1897 new.counters = old.counters;
1898 new.freelist = old.freelist;
1902 if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1905 struct kmem_cache_node *n2 = get_node(s,
1911 spin_unlock(&n->list_lock);
1914 spin_lock(&n->list_lock);
1919 if (l == M_PARTIAL) {
1920 remove_partial(n, page);
1921 stat(s, FREE_REMOVE_PARTIAL);
1923 add_partial(n, page,
1924 DEACTIVATE_TO_TAIL);
1925 stat(s, FREE_ADD_PARTIAL);
1931 } while (!cmpxchg_double_slab(s, page,
1932 old.freelist, old.counters,
1933 new.freelist, new.counters,
1934 "unfreezing slab"));
1937 page->next = discard_page;
1938 discard_page = page;
1943 spin_unlock(&n->list_lock);
1945 while (discard_page) {
1946 page = discard_page;
1947 discard_page = discard_page->next;
1949 stat(s, DEACTIVATE_EMPTY);
1950 discard_slab(s, page);
1956 * Put a page that was just frozen (in __slab_free) into a partial page
1957 * slot if available. This is done without interrupts disabled and without
1958 * preemption disabled. The cmpxchg is racy and may put the partial page
1959 * onto a random cpus partial slot.
1961 * If we did not find a slot then simply move all the partials to the
1962 * per node partial list.
1964 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1966 struct page *oldpage;
1973 oldpage = this_cpu_read(s->cpu_slab->partial);
1976 pobjects = oldpage->pobjects;
1977 pages = oldpage->pages;
1978 if (drain && pobjects > s->cpu_partial) {
1979 unsigned long flags;
1981 * partial array is full. Move the existing
1982 * set to the per node partial list.
1984 local_irq_save(flags);
1985 unfreeze_partials(s);
1986 local_irq_restore(flags);
1989 stat(s, CPU_PARTIAL_DRAIN);
1994 pobjects += page->objects - page->inuse;
1996 page->pages = pages;
1997 page->pobjects = pobjects;
1998 page->next = oldpage;
2000 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2004 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2006 stat(s, CPUSLAB_FLUSH);
2007 deactivate_slab(s, c->page, c->freelist);
2009 c->tid = next_tid(c->tid);
2017 * Called from IPI handler with interrupts disabled.
2019 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2021 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2027 unfreeze_partials(s);
2031 static void flush_cpu_slab(void *d)
2033 struct kmem_cache *s = d;
2035 __flush_cpu_slab(s, smp_processor_id());
2038 static bool has_cpu_slab(int cpu, void *info)
2040 struct kmem_cache *s = info;
2041 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2043 return c->page || c->partial;
2046 static void flush_all(struct kmem_cache *s)
2048 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2052 * Check if the objects in a per cpu structure fit numa
2053 * locality expectations.
2055 static inline int node_match(struct page *page, int node)
2058 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2064 static int count_free(struct page *page)
2066 return page->objects - page->inuse;
2069 static unsigned long count_partial(struct kmem_cache_node *n,
2070 int (*get_count)(struct page *))
2072 unsigned long flags;
2073 unsigned long x = 0;
2076 spin_lock_irqsave(&n->list_lock, flags);
2077 list_for_each_entry(page, &n->partial, lru)
2078 x += get_count(page);
2079 spin_unlock_irqrestore(&n->list_lock, flags);
2083 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2085 #ifdef CONFIG_SLUB_DEBUG
2086 return atomic_long_read(&n->total_objects);
2092 static noinline void
2093 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2098 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2100 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2101 "default order: %d, min order: %d\n", s->name, s->objsize,
2102 s->size, oo_order(s->oo), oo_order(s->min));
2104 if (oo_order(s->min) > get_order(s->objsize))
2105 printk(KERN_WARNING " %s debugging increased min order, use "
2106 "slub_debug=O to disable.\n", s->name);
2108 for_each_online_node(node) {
2109 struct kmem_cache_node *n = get_node(s, node);
2110 unsigned long nr_slabs;
2111 unsigned long nr_objs;
2112 unsigned long nr_free;
2117 nr_free = count_partial(n, count_free);
2118 nr_slabs = node_nr_slabs(n);
2119 nr_objs = node_nr_objs(n);
2122 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2123 node, nr_slabs, nr_objs, nr_free);
2127 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2128 int node, struct kmem_cache_cpu **pc)
2131 struct kmem_cache_cpu *c = *pc;
2134 freelist = get_partial(s, flags, node, c);
2139 page = new_slab(s, flags, node);
2141 c = __this_cpu_ptr(s->cpu_slab);
2146 * No other reference to the page yet so we can
2147 * muck around with it freely without cmpxchg
2149 freelist = page->freelist;
2150 page->freelist = NULL;
2152 stat(s, ALLOC_SLAB);
2162 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2163 * or deactivate the page.
2165 * The page is still frozen if the return value is not NULL.
2167 * If this function returns NULL then the page has been unfrozen.
2169 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2172 unsigned long counters;
2176 freelist = page->freelist;
2177 counters = page->counters;
2179 new.counters = counters;
2180 VM_BUG_ON(!new.frozen);
2182 new.inuse = page->objects;
2183 new.frozen = freelist != NULL;
2185 } while (!cmpxchg_double_slab(s, page,
2194 * Slow path. The lockless freelist is empty or we need to perform
2197 * Processing is still very fast if new objects have been freed to the
2198 * regular freelist. In that case we simply take over the regular freelist
2199 * as the lockless freelist and zap the regular freelist.
2201 * If that is not working then we fall back to the partial lists. We take the
2202 * first element of the freelist as the object to allocate now and move the
2203 * rest of the freelist to the lockless freelist.
2205 * And if we were unable to get a new slab from the partial slab lists then
2206 * we need to allocate a new slab. This is the slowest path since it involves
2207 * a call to the page allocator and the setup of a new slab.
2209 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2210 unsigned long addr, struct kmem_cache_cpu *c)
2214 unsigned long flags;
2216 local_irq_save(flags);
2217 #ifdef CONFIG_PREEMPT
2219 * We may have been preempted and rescheduled on a different
2220 * cpu before disabling interrupts. Need to reload cpu area
2223 c = this_cpu_ptr(s->cpu_slab);
2231 if (unlikely(!node_match(page, node))) {
2232 stat(s, ALLOC_NODE_MISMATCH);
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)))
2292 /* Only entered in the debug case */
2293 if (!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)))
2347 object = __slab_alloc(s, gfpflags, node, addr, c);
2350 void *next_object = get_freepointer_safe(s, object);
2353 * The cmpxchg will only match if there was no additional
2354 * operation and if we are on the right processor.
2356 * The cmpxchg does the following atomically (without lock semantics!)
2357 * 1. Relocate first pointer to the current per cpu area.
2358 * 2. Verify that tid and freelist have not been changed
2359 * 3. If they were not changed replace tid and freelist
2361 * Since this is without lock semantics the protection is only against
2362 * code executing on this cpu *not* from access by other cpus.
2364 if (unlikely(!this_cpu_cmpxchg_double(
2365 s->cpu_slab->freelist, s->cpu_slab->tid,
2367 next_object, next_tid(tid)))) {
2369 note_cmpxchg_failure("slab_alloc", s, tid);
2372 prefetch_freepointer(s, next_object);
2373 stat(s, ALLOC_FASTPATH);
2376 if (unlikely(gfpflags & __GFP_ZERO) && object)
2377 memset(object, 0, s->objsize);
2379 slab_post_alloc_hook(s, gfpflags, object);
2384 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2386 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2388 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2392 EXPORT_SYMBOL(kmem_cache_alloc);
2394 #ifdef CONFIG_TRACING
2395 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2397 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2398 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2401 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2403 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2405 void *ret = kmalloc_order(size, flags, order);
2406 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2409 EXPORT_SYMBOL(kmalloc_order_trace);
2413 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2415 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2417 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2418 s->objsize, s->size, gfpflags, node);
2422 EXPORT_SYMBOL(kmem_cache_alloc_node);
2424 #ifdef CONFIG_TRACING
2425 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2427 int node, size_t size)
2429 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2431 trace_kmalloc_node(_RET_IP_, ret,
2432 size, s->size, gfpflags, node);
2435 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2440 * Slow patch handling. This may still be called frequently since objects
2441 * have a longer lifetime than the cpu slabs in most processing loads.
2443 * So we still attempt to reduce cache line usage. Just take the slab
2444 * lock and free the item. If there is no additional partial page
2445 * handling required then we can return immediately.
2447 static void __slab_free(struct kmem_cache *s, struct page *page,
2448 void *x, unsigned long addr)
2451 void **object = (void *)x;
2455 unsigned long counters;
2456 struct kmem_cache_node *n = NULL;
2457 unsigned long uninitialized_var(flags);
2459 stat(s, FREE_SLOWPATH);
2461 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2465 prior = page->freelist;
2466 counters = page->counters;
2467 set_freepointer(s, object, prior);
2468 new.counters = counters;
2469 was_frozen = new.frozen;
2471 if ((!new.inuse || !prior) && !was_frozen && !n) {
2473 if (!kmem_cache_debug(s) && !prior)
2476 * Slab was on no list before and will be partially empty
2477 * We can defer the list move and instead freeze it.
2481 else { /* Needs to be taken off a list */
2483 n = get_node(s, page_to_nid(page));
2485 * Speculatively acquire the list_lock.
2486 * If the cmpxchg does not succeed then we may
2487 * drop the list_lock without any processing.
2489 * Otherwise the list_lock will synchronize with
2490 * other processors updating the list of slabs.
2492 spin_lock_irqsave(&n->list_lock, flags);
2498 } while (!cmpxchg_double_slab(s, page,
2500 object, new.counters,
2506 * If we just froze the page then put it onto the
2507 * per cpu partial list.
2509 if (new.frozen && !was_frozen) {
2510 put_cpu_partial(s, page, 1);
2511 stat(s, CPU_PARTIAL_FREE);
2514 * The list lock was not taken therefore no list
2515 * activity can be necessary.
2518 stat(s, FREE_FROZEN);
2523 * was_frozen may have been set after we acquired the list_lock in
2524 * an earlier loop. So we need to check it here again.
2527 stat(s, FREE_FROZEN);
2529 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2533 * Objects left in the slab. If it was not on the partial list before
2536 if (unlikely(!prior)) {
2537 remove_full(s, page);
2538 add_partial(n, page, DEACTIVATE_TO_TAIL);
2539 stat(s, FREE_ADD_PARTIAL);
2542 spin_unlock_irqrestore(&n->list_lock, flags);
2548 * Slab on the partial list.
2550 remove_partial(n, page);
2551 stat(s, FREE_REMOVE_PARTIAL);
2553 /* Slab must be on the full list */
2554 remove_full(s, page);
2556 spin_unlock_irqrestore(&n->list_lock, flags);
2558 discard_slab(s, page);
2562 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2563 * can perform fastpath freeing without additional function calls.
2565 * The fastpath is only possible if we are freeing to the current cpu slab
2566 * of this processor. This typically the case if we have just allocated
2569 * If fastpath is not possible then fall back to __slab_free where we deal
2570 * with all sorts of special processing.
2572 static __always_inline void slab_free(struct kmem_cache *s,
2573 struct page *page, void *x, unsigned long addr)
2575 void **object = (void *)x;
2576 struct kmem_cache_cpu *c;
2579 slab_free_hook(s, x);
2583 * Determine the currently cpus per cpu slab.
2584 * The cpu may change afterward. However that does not matter since
2585 * data is retrieved via this pointer. If we are on the same cpu
2586 * during the cmpxchg then the free will succedd.
2588 c = __this_cpu_ptr(s->cpu_slab);
2593 if (likely(page == c->page)) {
2594 set_freepointer(s, object, c->freelist);
2596 if (unlikely(!this_cpu_cmpxchg_double(
2597 s->cpu_slab->freelist, s->cpu_slab->tid,
2599 object, next_tid(tid)))) {
2601 note_cmpxchg_failure("slab_free", s, tid);
2604 stat(s, FREE_FASTPATH);
2606 __slab_free(s, page, x, addr);
2610 void kmem_cache_free(struct kmem_cache *s, void *x)
2614 page = virt_to_head_page(x);
2616 slab_free(s, page, x, _RET_IP_);
2618 trace_kmem_cache_free(_RET_IP_, x);
2620 EXPORT_SYMBOL(kmem_cache_free);
2623 * Object placement in a slab is made very easy because we always start at
2624 * offset 0. If we tune the size of the object to the alignment then we can
2625 * get the required alignment by putting one properly sized object after
2628 * Notice that the allocation order determines the sizes of the per cpu
2629 * caches. Each processor has always one slab available for allocations.
2630 * Increasing the allocation order reduces the number of times that slabs
2631 * must be moved on and off the partial lists and is therefore a factor in
2636 * Mininum / Maximum order of slab pages. This influences locking overhead
2637 * and slab fragmentation. A higher order reduces the number of partial slabs
2638 * and increases the number of allocations possible without having to
2639 * take the list_lock.
2641 static int slub_min_order;
2642 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2643 static int slub_min_objects;
2646 * Merge control. If this is set then no merging of slab caches will occur.
2647 * (Could be removed. This was introduced to pacify the merge skeptics.)
2649 static int slub_nomerge;
2652 * Calculate the order of allocation given an slab object size.
2654 * The order of allocation has significant impact on performance and other
2655 * system components. Generally order 0 allocations should be preferred since
2656 * order 0 does not cause fragmentation in the page allocator. Larger objects
2657 * be problematic to put into order 0 slabs because there may be too much
2658 * unused space left. We go to a higher order if more than 1/16th of the slab
2661 * In order to reach satisfactory performance we must ensure that a minimum
2662 * number of objects is in one slab. Otherwise we may generate too much
2663 * activity on the partial lists which requires taking the list_lock. This is
2664 * less a concern for large slabs though which are rarely used.
2666 * slub_max_order specifies the order where we begin to stop considering the
2667 * number of objects in a slab as critical. If we reach slub_max_order then
2668 * we try to keep the page order as low as possible. So we accept more waste
2669 * of space in favor of a small page order.
2671 * Higher order allocations also allow the placement of more objects in a
2672 * slab and thereby reduce object handling overhead. If the user has
2673 * requested a higher mininum order then we start with that one instead of
2674 * the smallest order which will fit the object.
2676 static inline int slab_order(int size, int min_objects,
2677 int max_order, int fract_leftover, int reserved)
2681 int min_order = slub_min_order;
2683 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2684 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2686 for (order = max(min_order,
2687 fls(min_objects * size - 1) - PAGE_SHIFT);
2688 order <= max_order; order++) {
2690 unsigned long slab_size = PAGE_SIZE << order;
2692 if (slab_size < min_objects * size + reserved)
2695 rem = (slab_size - reserved) % size;
2697 if (rem <= slab_size / fract_leftover)
2705 static inline int calculate_order(int size, int reserved)
2713 * Attempt to find best configuration for a slab. This
2714 * works by first attempting to generate a layout with
2715 * the best configuration and backing off gradually.
2717 * First we reduce the acceptable waste in a slab. Then
2718 * we reduce the minimum objects required in a slab.
2720 min_objects = slub_min_objects;
2722 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2723 max_objects = order_objects(slub_max_order, size, reserved);
2724 min_objects = min(min_objects, max_objects);
2726 while (min_objects > 1) {
2728 while (fraction >= 4) {
2729 order = slab_order(size, min_objects,
2730 slub_max_order, fraction, reserved);
2731 if (order <= slub_max_order)
2739 * We were unable to place multiple objects in a slab. Now
2740 * lets see if we can place a single object there.
2742 order = slab_order(size, 1, slub_max_order, 1, reserved);
2743 if (order <= slub_max_order)
2747 * Doh this slab cannot be placed using slub_max_order.
2749 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2750 if (order < MAX_ORDER)
2756 * Figure out what the alignment of the objects will be.
2758 static unsigned long calculate_alignment(unsigned long flags,
2759 unsigned long align, unsigned long size)
2762 * If the user wants hardware cache aligned objects then follow that
2763 * suggestion if the object is sufficiently large.
2765 * The hardware cache alignment cannot override the specified
2766 * alignment though. If that is greater then use it.
2768 if (flags & SLAB_HWCACHE_ALIGN) {
2769 unsigned long ralign = cache_line_size();
2770 while (size <= ralign / 2)
2772 align = max(align, ralign);
2775 if (align < ARCH_SLAB_MINALIGN)
2776 align = ARCH_SLAB_MINALIGN;
2778 return ALIGN(align, sizeof(void *));
2782 init_kmem_cache_node(struct kmem_cache_node *n)
2785 spin_lock_init(&n->list_lock);
2786 INIT_LIST_HEAD(&n->partial);
2787 #ifdef CONFIG_SLUB_DEBUG
2788 atomic_long_set(&n->nr_slabs, 0);
2789 atomic_long_set(&n->total_objects, 0);
2790 INIT_LIST_HEAD(&n->full);
2794 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2796 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2797 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2800 * Must align to double word boundary for the double cmpxchg
2801 * instructions to work; see __pcpu_double_call_return_bool().
2803 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2804 2 * sizeof(void *));
2809 init_kmem_cache_cpus(s);
2814 static struct kmem_cache *kmem_cache_node;
2817 * No kmalloc_node yet so do it by hand. We know that this is the first
2818 * slab on the node for this slabcache. There are no concurrent accesses
2821 * Note that this function only works on the kmalloc_node_cache
2822 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2823 * memory on a fresh node that has no slab structures yet.
2825 static void early_kmem_cache_node_alloc(int node)
2828 struct kmem_cache_node *n;
2830 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2832 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2835 if (page_to_nid(page) != node) {
2836 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2838 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2839 "in order to be able to continue\n");
2844 page->freelist = get_freepointer(kmem_cache_node, n);
2847 kmem_cache_node->node[node] = n;
2848 #ifdef CONFIG_SLUB_DEBUG
2849 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2850 init_tracking(kmem_cache_node, n);
2852 init_kmem_cache_node(n);
2853 inc_slabs_node(kmem_cache_node, node, page->objects);
2855 add_partial(n, page, DEACTIVATE_TO_HEAD);
2858 static void free_kmem_cache_nodes(struct kmem_cache *s)
2862 for_each_node_state(node, N_NORMAL_MEMORY) {
2863 struct kmem_cache_node *n = s->node[node];
2866 kmem_cache_free(kmem_cache_node, n);
2868 s->node[node] = NULL;
2872 static int init_kmem_cache_nodes(struct kmem_cache *s)
2876 for_each_node_state(node, N_NORMAL_MEMORY) {
2877 struct kmem_cache_node *n;
2879 if (slab_state == DOWN) {
2880 early_kmem_cache_node_alloc(node);
2883 n = kmem_cache_alloc_node(kmem_cache_node,
2887 free_kmem_cache_nodes(s);
2892 init_kmem_cache_node(n);
2897 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2899 if (min < MIN_PARTIAL)
2901 else if (min > MAX_PARTIAL)
2903 s->min_partial = min;
2907 * calculate_sizes() determines the order and the distribution of data within
2910 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2912 unsigned long flags = s->flags;
2913 unsigned long size = s->objsize;
2914 unsigned long align = s->align;
2918 * Round up object size to the next word boundary. We can only
2919 * place the free pointer at word boundaries and this determines
2920 * the possible location of the free pointer.
2922 size = ALIGN(size, sizeof(void *));
2924 #ifdef CONFIG_SLUB_DEBUG
2926 * Determine if we can poison the object itself. If the user of
2927 * the slab may touch the object after free or before allocation
2928 * then we should never poison the object itself.
2930 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2932 s->flags |= __OBJECT_POISON;
2934 s->flags &= ~__OBJECT_POISON;
2938 * If we are Redzoning then check if there is some space between the
2939 * end of the object and the free pointer. If not then add an
2940 * additional word to have some bytes to store Redzone information.
2942 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2943 size += sizeof(void *);
2947 * With that we have determined the number of bytes in actual use
2948 * by the object. This is the potential offset to the free pointer.
2952 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2955 * Relocate free pointer after the object if it is not
2956 * permitted to overwrite the first word of the object on
2959 * This is the case if we do RCU, have a constructor or
2960 * destructor or are poisoning the objects.
2963 size += sizeof(void *);
2966 #ifdef CONFIG_SLUB_DEBUG
2967 if (flags & SLAB_STORE_USER)
2969 * Need to store information about allocs and frees after
2972 size += 2 * sizeof(struct track);
2974 if (flags & SLAB_RED_ZONE)
2976 * Add some empty padding so that we can catch
2977 * overwrites from earlier objects rather than let
2978 * tracking information or the free pointer be
2979 * corrupted if a user writes before the start
2982 size += sizeof(void *);
2986 * Determine the alignment based on various parameters that the
2987 * user specified and the dynamic determination of cache line size
2990 align = calculate_alignment(flags, align, s->objsize);
2994 * SLUB stores one object immediately after another beginning from
2995 * offset 0. In order to align the objects we have to simply size
2996 * each object to conform to the alignment.
2998 size = ALIGN(size, align);
3000 if (forced_order >= 0)
3001 order = forced_order;
3003 order = calculate_order(size, s->reserved);
3010 s->allocflags |= __GFP_COMP;
3012 if (s->flags & SLAB_CACHE_DMA)
3013 s->allocflags |= SLUB_DMA;
3015 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3016 s->allocflags |= __GFP_RECLAIMABLE;
3019 * Determine the number of objects per slab
3021 s->oo = oo_make(order, size, s->reserved);
3022 s->min = oo_make(get_order(size), size, s->reserved);
3023 if (oo_objects(s->oo) > oo_objects(s->max))
3026 return !!oo_objects(s->oo);
3030 static int kmem_cache_open(struct kmem_cache *s,
3031 const char *name, size_t size,
3032 size_t align, unsigned long flags,
3033 void (*ctor)(void *))
3035 memset(s, 0, kmem_size);
3040 s->flags = kmem_cache_flags(size, flags, name, ctor);
3043 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3044 s->reserved = sizeof(struct rcu_head);
3046 if (!calculate_sizes(s, -1))
3048 if (disable_higher_order_debug) {
3050 * Disable debugging flags that store metadata if the min slab
3053 if (get_order(s->size) > get_order(s->objsize)) {
3054 s->flags &= ~DEBUG_METADATA_FLAGS;
3056 if (!calculate_sizes(s, -1))
3061 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3062 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3063 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3064 /* Enable fast mode */
3065 s->flags |= __CMPXCHG_DOUBLE;
3069 * The larger the object size is, the more pages we want on the partial
3070 * list to avoid pounding the page allocator excessively.
3072 set_min_partial(s, ilog2(s->size) / 2);
3075 * cpu_partial determined the maximum number of objects kept in the
3076 * per cpu partial lists of a processor.
3078 * Per cpu partial lists mainly contain slabs that just have one
3079 * object freed. If they are used for allocation then they can be
3080 * filled up again with minimal effort. The slab will never hit the
3081 * per node partial lists and therefore no locking will be required.
3083 * This setting also determines
3085 * A) The number of objects from per cpu partial slabs dumped to the
3086 * per node list when we reach the limit.
3087 * B) The number of objects in cpu partial slabs to extract from the
3088 * per node list when we run out of per cpu objects. We only fetch 50%
3089 * to keep some capacity around for frees.
3091 if (kmem_cache_debug(s))
3093 else if (s->size >= PAGE_SIZE)
3095 else if (s->size >= 1024)
3097 else if (s->size >= 256)
3098 s->cpu_partial = 13;
3100 s->cpu_partial = 30;
3104 s->remote_node_defrag_ratio = 1000;
3106 if (!init_kmem_cache_nodes(s))
3109 if (alloc_kmem_cache_cpus(s))
3112 free_kmem_cache_nodes(s);
3114 if (flags & SLAB_PANIC)
3115 panic("Cannot create slab %s size=%lu realsize=%u "
3116 "order=%u offset=%u flags=%lx\n",
3117 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3123 * Determine the size of a slab object
3125 unsigned int kmem_cache_size(struct kmem_cache *s)
3129 EXPORT_SYMBOL(kmem_cache_size);
3131 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3134 #ifdef CONFIG_SLUB_DEBUG
3135 void *addr = page_address(page);
3137 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3138 sizeof(long), GFP_ATOMIC);
3141 slab_err(s, page, "%s", text);
3144 get_map(s, page, map);
3145 for_each_object(p, s, addr, page->objects) {
3147 if (!test_bit(slab_index(p, s, addr), map)) {
3148 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3150 print_tracking(s, p);
3159 * Attempt to free all partial slabs on a node.
3160 * This is called from kmem_cache_close(). We must be the last thread
3161 * using the cache and therefore we do not need to lock anymore.
3163 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3165 struct page *page, *h;
3167 list_for_each_entry_safe(page, h, &n->partial, lru) {
3169 remove_partial(n, page);
3170 discard_slab(s, page);
3172 list_slab_objects(s, page,
3173 "Objects remaining on kmem_cache_close()");
3179 * Release all resources used by a slab cache.
3181 static inline int kmem_cache_close(struct kmem_cache *s)
3186 free_percpu(s->cpu_slab);
3187 /* Attempt to free all objects */
3188 for_each_node_state(node, N_NORMAL_MEMORY) {
3189 struct kmem_cache_node *n = get_node(s, node);
3192 if (n->nr_partial || slabs_node(s, node))
3195 free_kmem_cache_nodes(s);
3200 * Close a cache and release the kmem_cache structure
3201 * (must be used for caches created using kmem_cache_create)
3203 void kmem_cache_destroy(struct kmem_cache *s)
3205 down_write(&slub_lock);
3209 up_write(&slub_lock);
3210 if (kmem_cache_close(s)) {
3211 printk(KERN_ERR "SLUB %s: %s called for cache that "
3212 "still has objects.\n", s->name, __func__);
3215 if (s->flags & SLAB_DESTROY_BY_RCU)
3217 sysfs_slab_remove(s);
3219 up_write(&slub_lock);
3221 EXPORT_SYMBOL(kmem_cache_destroy);
3223 /********************************************************************
3225 *******************************************************************/
3227 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3228 EXPORT_SYMBOL(kmalloc_caches);
3230 static struct kmem_cache *kmem_cache;
3232 #ifdef CONFIG_ZONE_DMA
3233 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3236 static int __init setup_slub_min_order(char *str)
3238 get_option(&str, &slub_min_order);
3243 __setup("slub_min_order=", setup_slub_min_order);
3245 static int __init setup_slub_max_order(char *str)
3247 get_option(&str, &slub_max_order);
3248 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3253 __setup("slub_max_order=", setup_slub_max_order);
3255 static int __init setup_slub_min_objects(char *str)
3257 get_option(&str, &slub_min_objects);
3262 __setup("slub_min_objects=", setup_slub_min_objects);
3264 static int __init setup_slub_nomerge(char *str)
3270 __setup("slub_nomerge", setup_slub_nomerge);
3272 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3273 int size, unsigned int flags)
3275 struct kmem_cache *s;
3277 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3280 * This function is called with IRQs disabled during early-boot on
3281 * single CPU so there's no need to take slub_lock here.
3283 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3287 list_add(&s->list, &slab_caches);
3291 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3296 * Conversion table for small slabs sizes / 8 to the index in the
3297 * kmalloc array. This is necessary for slabs < 192 since we have non power
3298 * of two cache sizes there. The size of larger slabs can be determined using
3301 static s8 size_index[24] = {
3328 static inline int size_index_elem(size_t bytes)
3330 return (bytes - 1) / 8;
3333 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3339 return ZERO_SIZE_PTR;
3341 index = size_index[size_index_elem(size)];
3343 index = fls(size - 1);
3345 #ifdef CONFIG_ZONE_DMA
3346 if (unlikely((flags & SLUB_DMA)))
3347 return kmalloc_dma_caches[index];
3350 return kmalloc_caches[index];
3353 void *__kmalloc(size_t size, gfp_t flags)
3355 struct kmem_cache *s;
3358 if (unlikely(size > SLUB_MAX_SIZE))
3359 return kmalloc_large(size, flags);
3361 s = get_slab(size, flags);
3363 if (unlikely(ZERO_OR_NULL_PTR(s)))
3366 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3368 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3372 EXPORT_SYMBOL(__kmalloc);
3375 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3380 flags |= __GFP_COMP | __GFP_NOTRACK;
3381 page = alloc_pages_node(node, flags, get_order(size));
3383 ptr = page_address(page);
3385 kmemleak_alloc(ptr, size, 1, flags);
3389 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3391 struct kmem_cache *s;
3394 if (unlikely(size > SLUB_MAX_SIZE)) {
3395 ret = kmalloc_large_node(size, flags, node);
3397 trace_kmalloc_node(_RET_IP_, ret,
3398 size, PAGE_SIZE << get_order(size),
3404 s = get_slab(size, flags);
3406 if (unlikely(ZERO_OR_NULL_PTR(s)))
3409 ret = slab_alloc(s, flags, node, _RET_IP_);
3411 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3415 EXPORT_SYMBOL(__kmalloc_node);
3418 size_t ksize(const void *object)
3422 if (unlikely(object == ZERO_SIZE_PTR))
3425 page = virt_to_head_page(object);
3427 if (unlikely(!PageSlab(page))) {
3428 WARN_ON(!PageCompound(page));
3429 return PAGE_SIZE << compound_order(page);
3432 return slab_ksize(page->slab);
3434 EXPORT_SYMBOL(ksize);
3436 #ifdef CONFIG_SLUB_DEBUG
3437 bool verify_mem_not_deleted(const void *x)
3440 void *object = (void *)x;
3441 unsigned long flags;
3444 if (unlikely(ZERO_OR_NULL_PTR(x)))
3447 local_irq_save(flags);
3449 page = virt_to_head_page(x);
3450 if (unlikely(!PageSlab(page))) {
3451 /* maybe it was from stack? */
3457 if (on_freelist(page->slab, page, object)) {
3458 object_err(page->slab, page, object, "Object is on free-list");
3466 local_irq_restore(flags);
3469 EXPORT_SYMBOL(verify_mem_not_deleted);
3472 void kfree(const void *x)
3475 void *object = (void *)x;
3477 trace_kfree(_RET_IP_, x);
3479 if (unlikely(ZERO_OR_NULL_PTR(x)))
3482 page = virt_to_head_page(x);
3483 if (unlikely(!PageSlab(page))) {
3484 BUG_ON(!PageCompound(page));
3489 slab_free(page->slab, page, object, _RET_IP_);
3491 EXPORT_SYMBOL(kfree);
3494 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3495 * the remaining slabs by the number of items in use. The slabs with the
3496 * most items in use come first. New allocations will then fill those up
3497 * and thus they can be removed from the partial lists.
3499 * The slabs with the least items are placed last. This results in them
3500 * being allocated from last increasing the chance that the last objects
3501 * are freed in them.
3503 int kmem_cache_shrink(struct kmem_cache *s)
3507 struct kmem_cache_node *n;
3510 int objects = oo_objects(s->max);
3511 struct list_head *slabs_by_inuse =
3512 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3513 unsigned long flags;
3515 if (!slabs_by_inuse)
3519 for_each_node_state(node, N_NORMAL_MEMORY) {
3520 n = get_node(s, node);
3525 for (i = 0; i < objects; i++)
3526 INIT_LIST_HEAD(slabs_by_inuse + i);
3528 spin_lock_irqsave(&n->list_lock, flags);
3531 * Build lists indexed by the items in use in each slab.
3533 * Note that concurrent frees may occur while we hold the
3534 * list_lock. page->inuse here is the upper limit.
3536 list_for_each_entry_safe(page, t, &n->partial, lru) {
3537 list_move(&page->lru, slabs_by_inuse + page->inuse);
3543 * Rebuild the partial list with the slabs filled up most
3544 * first and the least used slabs at the end.
3546 for (i = objects - 1; i > 0; i--)
3547 list_splice(slabs_by_inuse + i, n->partial.prev);
3549 spin_unlock_irqrestore(&n->list_lock, flags);
3551 /* Release empty slabs */
3552 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3553 discard_slab(s, page);
3556 kfree(slabs_by_inuse);
3559 EXPORT_SYMBOL(kmem_cache_shrink);
3561 #if defined(CONFIG_MEMORY_HOTPLUG)
3562 static int slab_mem_going_offline_callback(void *arg)
3564 struct kmem_cache *s;
3566 down_read(&slub_lock);
3567 list_for_each_entry(s, &slab_caches, list)
3568 kmem_cache_shrink(s);
3569 up_read(&slub_lock);
3574 static void slab_mem_offline_callback(void *arg)
3576 struct kmem_cache_node *n;
3577 struct kmem_cache *s;
3578 struct memory_notify *marg = arg;
3581 offline_node = marg->status_change_nid;
3584 * If the node still has available memory. we need kmem_cache_node
3587 if (offline_node < 0)
3590 down_read(&slub_lock);
3591 list_for_each_entry(s, &slab_caches, list) {
3592 n = get_node(s, offline_node);
3595 * if n->nr_slabs > 0, slabs still exist on the node
3596 * that is going down. We were unable to free them,
3597 * and offline_pages() function shouldn't call this
3598 * callback. So, we must fail.
3600 BUG_ON(slabs_node(s, offline_node));
3602 s->node[offline_node] = NULL;
3603 kmem_cache_free(kmem_cache_node, n);
3606 up_read(&slub_lock);
3609 static int slab_mem_going_online_callback(void *arg)
3611 struct kmem_cache_node *n;
3612 struct kmem_cache *s;
3613 struct memory_notify *marg = arg;
3614 int nid = marg->status_change_nid;
3618 * If the node's memory is already available, then kmem_cache_node is
3619 * already created. Nothing to do.
3625 * We are bringing a node online. No memory is available yet. We must
3626 * allocate a kmem_cache_node structure in order to bring the node
3629 down_read(&slub_lock);
3630 list_for_each_entry(s, &slab_caches, list) {
3632 * XXX: kmem_cache_alloc_node will fallback to other nodes
3633 * since memory is not yet available from the node that
3636 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3641 init_kmem_cache_node(n);
3645 up_read(&slub_lock);
3649 static int slab_memory_callback(struct notifier_block *self,
3650 unsigned long action, void *arg)
3655 case MEM_GOING_ONLINE:
3656 ret = slab_mem_going_online_callback(arg);
3658 case MEM_GOING_OFFLINE:
3659 ret = slab_mem_going_offline_callback(arg);
3662 case MEM_CANCEL_ONLINE:
3663 slab_mem_offline_callback(arg);
3666 case MEM_CANCEL_OFFLINE:
3670 ret = notifier_from_errno(ret);
3676 #endif /* CONFIG_MEMORY_HOTPLUG */
3678 /********************************************************************
3679 * Basic setup of slabs
3680 *******************************************************************/
3683 * Used for early kmem_cache structures that were allocated using
3684 * the page allocator
3687 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3691 list_add(&s->list, &slab_caches);
3694 for_each_node_state(node, N_NORMAL_MEMORY) {
3695 struct kmem_cache_node *n = get_node(s, node);
3699 list_for_each_entry(p, &n->partial, lru)
3702 #ifdef CONFIG_SLUB_DEBUG
3703 list_for_each_entry(p, &n->full, lru)
3710 void __init kmem_cache_init(void)
3714 struct kmem_cache *temp_kmem_cache;
3716 struct kmem_cache *temp_kmem_cache_node;
3717 unsigned long kmalloc_size;
3719 if (debug_guardpage_minorder())
3722 kmem_size = offsetof(struct kmem_cache, node) +
3723 nr_node_ids * sizeof(struct kmem_cache_node *);
3725 /* Allocate two kmem_caches from the page allocator */
3726 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3727 order = get_order(2 * kmalloc_size);
3728 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3731 * Must first have the slab cache available for the allocations of the
3732 * struct kmem_cache_node's. There is special bootstrap code in
3733 * kmem_cache_open for slab_state == DOWN.
3735 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3737 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3738 sizeof(struct kmem_cache_node),
3739 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3741 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3743 /* Able to allocate the per node structures */
3744 slab_state = PARTIAL;
3746 temp_kmem_cache = kmem_cache;
3747 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3748 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3749 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3750 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3753 * Allocate kmem_cache_node properly from the kmem_cache slab.
3754 * kmem_cache_node is separately allocated so no need to
3755 * update any list pointers.
3757 temp_kmem_cache_node = kmem_cache_node;
3759 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3760 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3762 kmem_cache_bootstrap_fixup(kmem_cache_node);
3765 kmem_cache_bootstrap_fixup(kmem_cache);
3767 /* Free temporary boot structure */
3768 free_pages((unsigned long)temp_kmem_cache, order);
3770 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3773 * Patch up the size_index table if we have strange large alignment
3774 * requirements for the kmalloc array. This is only the case for
3775 * MIPS it seems. The standard arches will not generate any code here.
3777 * Largest permitted alignment is 256 bytes due to the way we
3778 * handle the index determination for the smaller caches.
3780 * Make sure that nothing crazy happens if someone starts tinkering
3781 * around with ARCH_KMALLOC_MINALIGN
3783 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3784 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3786 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3787 int elem = size_index_elem(i);
3788 if (elem >= ARRAY_SIZE(size_index))
3790 size_index[elem] = KMALLOC_SHIFT_LOW;
3793 if (KMALLOC_MIN_SIZE == 64) {
3795 * The 96 byte size cache is not used if the alignment
3798 for (i = 64 + 8; i <= 96; i += 8)
3799 size_index[size_index_elem(i)] = 7;
3800 } else if (KMALLOC_MIN_SIZE == 128) {
3802 * The 192 byte sized cache is not used if the alignment
3803 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3806 for (i = 128 + 8; i <= 192; i += 8)
3807 size_index[size_index_elem(i)] = 8;
3810 /* Caches that are not of the two-to-the-power-of size */
3811 if (KMALLOC_MIN_SIZE <= 32) {
3812 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3816 if (KMALLOC_MIN_SIZE <= 64) {
3817 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3821 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3822 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3828 /* Provide the correct kmalloc names now that the caches are up */
3829 if (KMALLOC_MIN_SIZE <= 32) {
3830 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3831 BUG_ON(!kmalloc_caches[1]->name);
3834 if (KMALLOC_MIN_SIZE <= 64) {
3835 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3836 BUG_ON(!kmalloc_caches[2]->name);
3839 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3840 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3843 kmalloc_caches[i]->name = s;
3847 register_cpu_notifier(&slab_notifier);
3850 #ifdef CONFIG_ZONE_DMA
3851 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3852 struct kmem_cache *s = kmalloc_caches[i];
3855 char *name = kasprintf(GFP_NOWAIT,
3856 "dma-kmalloc-%d", s->objsize);
3859 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3860 s->objsize, SLAB_CACHE_DMA);
3865 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3866 " CPUs=%d, Nodes=%d\n",
3867 caches, cache_line_size(),
3868 slub_min_order, slub_max_order, slub_min_objects,
3869 nr_cpu_ids, nr_node_ids);
3872 void __init kmem_cache_init_late(void)
3877 * Find a mergeable slab cache
3879 static int slab_unmergeable(struct kmem_cache *s)
3881 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3888 * We may have set a slab to be unmergeable during bootstrap.
3890 if (s->refcount < 0)
3896 static struct kmem_cache *find_mergeable(size_t size,
3897 size_t align, unsigned long flags, const char *name,
3898 void (*ctor)(void *))
3900 struct kmem_cache *s;
3902 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3908 size = ALIGN(size, sizeof(void *));
3909 align = calculate_alignment(flags, align, size);
3910 size = ALIGN(size, align);
3911 flags = kmem_cache_flags(size, flags, name, NULL);
3913 list_for_each_entry(s, &slab_caches, list) {
3914 if (slab_unmergeable(s))
3920 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3923 * Check if alignment is compatible.
3924 * Courtesy of Adrian Drzewiecki
3926 if ((s->size & ~(align - 1)) != s->size)
3929 if (s->size - size >= sizeof(void *))
3937 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3938 size_t align, unsigned long flags, void (*ctor)(void *))
3940 struct kmem_cache *s;
3946 down_write(&slub_lock);
3947 s = find_mergeable(size, align, flags, name, ctor);
3951 * Adjust the object sizes so that we clear
3952 * the complete object on kzalloc.
3954 s->objsize = max(s->objsize, (int)size);
3955 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3957 if (sysfs_slab_alias(s, name)) {
3961 up_write(&slub_lock);
3965 n = kstrdup(name, GFP_KERNEL);
3969 s = kmalloc(kmem_size, GFP_KERNEL);
3971 if (kmem_cache_open(s, n,
3972 size, align, flags, ctor)) {
3973 list_add(&s->list, &slab_caches);
3974 up_write(&slub_lock);
3975 if (sysfs_slab_add(s)) {
3976 down_write(&slub_lock);
3988 up_write(&slub_lock);
3990 if (flags & SLAB_PANIC)
3991 panic("Cannot create slabcache %s\n", name);
3996 EXPORT_SYMBOL(kmem_cache_create);
4000 * Use the cpu notifier to insure that the cpu slabs are flushed when
4003 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
4004 unsigned long action, void *hcpu)
4006 long cpu = (long)hcpu;
4007 struct kmem_cache *s;
4008 unsigned long flags;
4011 case CPU_UP_CANCELED:
4012 case CPU_UP_CANCELED_FROZEN:
4014 case CPU_DEAD_FROZEN:
4015 down_read(&slub_lock);
4016 list_for_each_entry(s, &slab_caches, list) {
4017 local_irq_save(flags);
4018 __flush_cpu_slab(s, cpu);
4019 local_irq_restore(flags);
4021 up_read(&slub_lock);
4029 static struct notifier_block __cpuinitdata slab_notifier = {
4030 .notifier_call = slab_cpuup_callback
4035 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4037 struct kmem_cache *s;
4040 if (unlikely(size > SLUB_MAX_SIZE))
4041 return kmalloc_large(size, gfpflags);
4043 s = get_slab(size, gfpflags);
4045 if (unlikely(ZERO_OR_NULL_PTR(s)))
4048 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4050 /* Honor the call site pointer we received. */
4051 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4057 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4058 int node, unsigned long caller)
4060 struct kmem_cache *s;
4063 if (unlikely(size > SLUB_MAX_SIZE)) {
4064 ret = kmalloc_large_node(size, gfpflags, node);
4066 trace_kmalloc_node(caller, ret,
4067 size, PAGE_SIZE << get_order(size),
4073 s = get_slab(size, gfpflags);
4075 if (unlikely(ZERO_OR_NULL_PTR(s)))
4078 ret = slab_alloc(s, gfpflags, node, caller);
4080 /* Honor the call site pointer we received. */
4081 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4088 static int count_inuse(struct page *page)
4093 static int count_total(struct page *page)
4095 return page->objects;
4099 #ifdef CONFIG_SLUB_DEBUG
4100 static int validate_slab(struct kmem_cache *s, struct page *page,
4104 void *addr = page_address(page);
4106 if (!check_slab(s, page) ||
4107 !on_freelist(s, page, NULL))
4110 /* Now we know that a valid freelist exists */
4111 bitmap_zero(map, page->objects);
4113 get_map(s, page, map);
4114 for_each_object(p, s, addr, page->objects) {
4115 if (test_bit(slab_index(p, s, addr), map))
4116 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4120 for_each_object(p, s, addr, page->objects)
4121 if (!test_bit(slab_index(p, s, addr), map))
4122 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4127 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4131 validate_slab(s, page, map);
4135 static int validate_slab_node(struct kmem_cache *s,
4136 struct kmem_cache_node *n, unsigned long *map)
4138 unsigned long count = 0;
4140 unsigned long flags;
4142 spin_lock_irqsave(&n->list_lock, flags);
4144 list_for_each_entry(page, &n->partial, lru) {
4145 validate_slab_slab(s, page, map);
4148 if (count != n->nr_partial)
4149 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4150 "counter=%ld\n", s->name, count, n->nr_partial);
4152 if (!(s->flags & SLAB_STORE_USER))
4155 list_for_each_entry(page, &n->full, lru) {
4156 validate_slab_slab(s, page, map);
4159 if (count != atomic_long_read(&n->nr_slabs))
4160 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4161 "counter=%ld\n", s->name, count,
4162 atomic_long_read(&n->nr_slabs));
4165 spin_unlock_irqrestore(&n->list_lock, flags);
4169 static long validate_slab_cache(struct kmem_cache *s)
4172 unsigned long count = 0;
4173 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4174 sizeof(unsigned long), GFP_KERNEL);
4180 for_each_node_state(node, N_NORMAL_MEMORY) {
4181 struct kmem_cache_node *n = get_node(s, node);
4183 count += validate_slab_node(s, n, map);
4189 * Generate lists of code addresses where slabcache objects are allocated
4194 unsigned long count;
4201 DECLARE_BITMAP(cpus, NR_CPUS);
4207 unsigned long count;
4208 struct location *loc;
4211 static void free_loc_track(struct loc_track *t)
4214 free_pages((unsigned long)t->loc,
4215 get_order(sizeof(struct location) * t->max));
4218 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4223 order = get_order(sizeof(struct location) * max);
4225 l = (void *)__get_free_pages(flags, order);
4230 memcpy(l, t->loc, sizeof(struct location) * t->count);
4238 static int add_location(struct loc_track *t, struct kmem_cache *s,
4239 const struct track *track)
4241 long start, end, pos;
4243 unsigned long caddr;
4244 unsigned long age = jiffies - track->when;
4250 pos = start + (end - start + 1) / 2;
4253 * There is nothing at "end". If we end up there
4254 * we need to add something to before end.
4259 caddr = t->loc[pos].addr;
4260 if (track->addr == caddr) {
4266 if (age < l->min_time)
4268 if (age > l->max_time)
4271 if (track->pid < l->min_pid)
4272 l->min_pid = track->pid;
4273 if (track->pid > l->max_pid)
4274 l->max_pid = track->pid;
4276 cpumask_set_cpu(track->cpu,
4277 to_cpumask(l->cpus));
4279 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4283 if (track->addr < caddr)
4290 * Not found. Insert new tracking element.
4292 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4298 (t->count - pos) * sizeof(struct location));
4301 l->addr = track->addr;
4305 l->min_pid = track->pid;
4306 l->max_pid = track->pid;
4307 cpumask_clear(to_cpumask(l->cpus));
4308 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4309 nodes_clear(l->nodes);
4310 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4314 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4315 struct page *page, enum track_item alloc,
4318 void *addr = page_address(page);
4321 bitmap_zero(map, page->objects);
4322 get_map(s, page, map);
4324 for_each_object(p, s, addr, page->objects)
4325 if (!test_bit(slab_index(p, s, addr), map))
4326 add_location(t, s, get_track(s, p, alloc));
4329 static int list_locations(struct kmem_cache *s, char *buf,
4330 enum track_item alloc)
4334 struct loc_track t = { 0, 0, NULL };
4336 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4337 sizeof(unsigned long), GFP_KERNEL);
4339 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4342 return sprintf(buf, "Out of memory\n");
4344 /* Push back cpu slabs */
4347 for_each_node_state(node, N_NORMAL_MEMORY) {
4348 struct kmem_cache_node *n = get_node(s, node);
4349 unsigned long flags;
4352 if (!atomic_long_read(&n->nr_slabs))
4355 spin_lock_irqsave(&n->list_lock, flags);
4356 list_for_each_entry(page, &n->partial, lru)
4357 process_slab(&t, s, page, alloc, map);
4358 list_for_each_entry(page, &n->full, lru)
4359 process_slab(&t, s, page, alloc, map);
4360 spin_unlock_irqrestore(&n->list_lock, flags);
4363 for (i = 0; i < t.count; i++) {
4364 struct location *l = &t.loc[i];
4366 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4368 len += sprintf(buf + len, "%7ld ", l->count);
4371 len += sprintf(buf + len, "%pS", (void *)l->addr);
4373 len += sprintf(buf + len, "<not-available>");
4375 if (l->sum_time != l->min_time) {
4376 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4378 (long)div_u64(l->sum_time, l->count),
4381 len += sprintf(buf + len, " age=%ld",
4384 if (l->min_pid != l->max_pid)
4385 len += sprintf(buf + len, " pid=%ld-%ld",
4386 l->min_pid, l->max_pid);
4388 len += sprintf(buf + len, " pid=%ld",
4391 if (num_online_cpus() > 1 &&
4392 !cpumask_empty(to_cpumask(l->cpus)) &&
4393 len < PAGE_SIZE - 60) {
4394 len += sprintf(buf + len, " cpus=");
4395 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4396 to_cpumask(l->cpus));
4399 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4400 len < PAGE_SIZE - 60) {
4401 len += sprintf(buf + len, " nodes=");
4402 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4406 len += sprintf(buf + len, "\n");
4412 len += sprintf(buf, "No data\n");
4417 #ifdef SLUB_RESILIENCY_TEST
4418 static void resiliency_test(void)
4422 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4424 printk(KERN_ERR "SLUB resiliency testing\n");
4425 printk(KERN_ERR "-----------------------\n");
4426 printk(KERN_ERR "A. Corruption after allocation\n");
4428 p = kzalloc(16, GFP_KERNEL);
4430 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4431 " 0x12->0x%p\n\n", p + 16);
4433 validate_slab_cache(kmalloc_caches[4]);
4435 /* Hmmm... The next two are dangerous */
4436 p = kzalloc(32, GFP_KERNEL);
4437 p[32 + sizeof(void *)] = 0x34;
4438 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4439 " 0x34 -> -0x%p\n", p);
4441 "If allocated object is overwritten then not detectable\n\n");
4443 validate_slab_cache(kmalloc_caches[5]);
4444 p = kzalloc(64, GFP_KERNEL);
4445 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4447 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4450 "If allocated object is overwritten then not detectable\n\n");
4451 validate_slab_cache(kmalloc_caches[6]);
4453 printk(KERN_ERR "\nB. Corruption after free\n");
4454 p = kzalloc(128, GFP_KERNEL);
4457 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4458 validate_slab_cache(kmalloc_caches[7]);
4460 p = kzalloc(256, GFP_KERNEL);
4463 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4465 validate_slab_cache(kmalloc_caches[8]);
4467 p = kzalloc(512, GFP_KERNEL);
4470 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4471 validate_slab_cache(kmalloc_caches[9]);
4475 static void resiliency_test(void) {};
4480 enum slab_stat_type {
4481 SL_ALL, /* All slabs */
4482 SL_PARTIAL, /* Only partially allocated slabs */
4483 SL_CPU, /* Only slabs used for cpu caches */
4484 SL_OBJECTS, /* Determine allocated objects not slabs */
4485 SL_TOTAL /* Determine object capacity not slabs */
4488 #define SO_ALL (1 << SL_ALL)
4489 #define SO_PARTIAL (1 << SL_PARTIAL)
4490 #define SO_CPU (1 << SL_CPU)
4491 #define SO_OBJECTS (1 << SL_OBJECTS)
4492 #define SO_TOTAL (1 << SL_TOTAL)
4494 static ssize_t show_slab_objects(struct kmem_cache *s,
4495 char *buf, unsigned long flags)
4497 unsigned long total = 0;
4500 unsigned long *nodes;
4501 unsigned long *per_cpu;
4503 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4506 per_cpu = nodes + nr_node_ids;
4508 if (flags & SO_CPU) {
4511 for_each_possible_cpu(cpu) {
4512 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4516 page = ACCESS_ONCE(c->page);
4520 node = page_to_nid(page);
4521 if (flags & SO_TOTAL)
4523 else if (flags & SO_OBJECTS)
4531 page = ACCESS_ONCE(c->partial);
4542 lock_memory_hotplug();
4543 #ifdef CONFIG_SLUB_DEBUG
4544 if (flags & SO_ALL) {
4545 for_each_node_state(node, N_NORMAL_MEMORY) {
4546 struct kmem_cache_node *n = get_node(s, node);
4548 if (flags & SO_TOTAL)
4549 x = atomic_long_read(&n->total_objects);
4550 else if (flags & SO_OBJECTS)
4551 x = atomic_long_read(&n->total_objects) -
4552 count_partial(n, count_free);
4555 x = atomic_long_read(&n->nr_slabs);
4562 if (flags & SO_PARTIAL) {
4563 for_each_node_state(node, N_NORMAL_MEMORY) {
4564 struct kmem_cache_node *n = get_node(s, node);
4566 if (flags & SO_TOTAL)
4567 x = count_partial(n, count_total);
4568 else if (flags & SO_OBJECTS)
4569 x = count_partial(n, count_inuse);
4576 x = sprintf(buf, "%lu", total);
4578 for_each_node_state(node, N_NORMAL_MEMORY)
4580 x += sprintf(buf + x, " N%d=%lu",
4583 unlock_memory_hotplug();
4585 return x + sprintf(buf + x, "\n");
4588 #ifdef CONFIG_SLUB_DEBUG
4589 static int any_slab_objects(struct kmem_cache *s)
4593 for_each_online_node(node) {
4594 struct kmem_cache_node *n = get_node(s, node);
4599 if (atomic_long_read(&n->total_objects))
4606 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4607 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4609 struct slab_attribute {
4610 struct attribute attr;
4611 ssize_t (*show)(struct kmem_cache *s, char *buf);
4612 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4615 #define SLAB_ATTR_RO(_name) \
4616 static struct slab_attribute _name##_attr = \
4617 __ATTR(_name, 0400, _name##_show, NULL)
4619 #define SLAB_ATTR(_name) \
4620 static struct slab_attribute _name##_attr = \
4621 __ATTR(_name, 0600, _name##_show, _name##_store)
4623 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4625 return sprintf(buf, "%d\n", s->size);
4627 SLAB_ATTR_RO(slab_size);
4629 static ssize_t align_show(struct kmem_cache *s, char *buf)
4631 return sprintf(buf, "%d\n", s->align);
4633 SLAB_ATTR_RO(align);
4635 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4637 return sprintf(buf, "%d\n", s->objsize);
4639 SLAB_ATTR_RO(object_size);
4641 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4643 return sprintf(buf, "%d\n", oo_objects(s->oo));
4645 SLAB_ATTR_RO(objs_per_slab);
4647 static ssize_t order_store(struct kmem_cache *s,
4648 const char *buf, size_t length)
4650 unsigned long order;
4653 err = strict_strtoul(buf, 10, &order);
4657 if (order > slub_max_order || order < slub_min_order)
4660 calculate_sizes(s, order);
4664 static ssize_t order_show(struct kmem_cache *s, char *buf)
4666 return sprintf(buf, "%d\n", oo_order(s->oo));
4670 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4672 return sprintf(buf, "%lu\n", s->min_partial);
4675 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4681 err = strict_strtoul(buf, 10, &min);
4685 set_min_partial(s, min);
4688 SLAB_ATTR(min_partial);
4690 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4692 return sprintf(buf, "%u\n", s->cpu_partial);
4695 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4698 unsigned long objects;
4701 err = strict_strtoul(buf, 10, &objects);
4704 if (objects && kmem_cache_debug(s))
4707 s->cpu_partial = objects;
4711 SLAB_ATTR(cpu_partial);
4713 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4717 return sprintf(buf, "%pS\n", s->ctor);
4721 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4723 return sprintf(buf, "%d\n", s->refcount - 1);
4725 SLAB_ATTR_RO(aliases);
4727 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4729 return show_slab_objects(s, buf, SO_PARTIAL);
4731 SLAB_ATTR_RO(partial);
4733 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4735 return show_slab_objects(s, buf, SO_CPU);
4737 SLAB_ATTR_RO(cpu_slabs);
4739 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4741 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4743 SLAB_ATTR_RO(objects);
4745 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4747 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4749 SLAB_ATTR_RO(objects_partial);
4751 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4758 for_each_online_cpu(cpu) {
4759 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4762 pages += page->pages;
4763 objects += page->pobjects;
4767 len = sprintf(buf, "%d(%d)", objects, pages);
4770 for_each_online_cpu(cpu) {
4771 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4773 if (page && len < PAGE_SIZE - 20)
4774 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4775 page->pobjects, page->pages);
4778 return len + sprintf(buf + len, "\n");
4780 SLAB_ATTR_RO(slabs_cpu_partial);
4782 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4784 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4787 static ssize_t reclaim_account_store(struct kmem_cache *s,
4788 const char *buf, size_t length)
4790 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4792 s->flags |= SLAB_RECLAIM_ACCOUNT;
4795 SLAB_ATTR(reclaim_account);
4797 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4799 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4801 SLAB_ATTR_RO(hwcache_align);
4803 #ifdef CONFIG_ZONE_DMA
4804 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4806 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4808 SLAB_ATTR_RO(cache_dma);
4811 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4813 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4815 SLAB_ATTR_RO(destroy_by_rcu);
4817 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4819 return sprintf(buf, "%d\n", s->reserved);
4821 SLAB_ATTR_RO(reserved);
4823 #ifdef CONFIG_SLUB_DEBUG
4824 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4826 return show_slab_objects(s, buf, SO_ALL);
4828 SLAB_ATTR_RO(slabs);
4830 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4832 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4834 SLAB_ATTR_RO(total_objects);
4836 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4838 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4841 static ssize_t sanity_checks_store(struct kmem_cache *s,
4842 const char *buf, size_t length)
4844 s->flags &= ~SLAB_DEBUG_FREE;
4845 if (buf[0] == '1') {
4846 s->flags &= ~__CMPXCHG_DOUBLE;
4847 s->flags |= SLAB_DEBUG_FREE;
4851 SLAB_ATTR(sanity_checks);
4853 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4855 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4858 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4861 s->flags &= ~SLAB_TRACE;
4862 if (buf[0] == '1') {
4863 s->flags &= ~__CMPXCHG_DOUBLE;
4864 s->flags |= SLAB_TRACE;
4870 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4872 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4875 static ssize_t red_zone_store(struct kmem_cache *s,
4876 const char *buf, size_t length)
4878 if (any_slab_objects(s))
4881 s->flags &= ~SLAB_RED_ZONE;
4882 if (buf[0] == '1') {
4883 s->flags &= ~__CMPXCHG_DOUBLE;
4884 s->flags |= SLAB_RED_ZONE;
4886 calculate_sizes(s, -1);
4889 SLAB_ATTR(red_zone);
4891 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4893 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4896 static ssize_t poison_store(struct kmem_cache *s,
4897 const char *buf, size_t length)
4899 if (any_slab_objects(s))
4902 s->flags &= ~SLAB_POISON;
4903 if (buf[0] == '1') {
4904 s->flags &= ~__CMPXCHG_DOUBLE;
4905 s->flags |= SLAB_POISON;
4907 calculate_sizes(s, -1);
4912 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4914 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4917 static ssize_t store_user_store(struct kmem_cache *s,
4918 const char *buf, size_t length)
4920 if (any_slab_objects(s))
4923 s->flags &= ~SLAB_STORE_USER;
4924 if (buf[0] == '1') {
4925 s->flags &= ~__CMPXCHG_DOUBLE;
4926 s->flags |= SLAB_STORE_USER;
4928 calculate_sizes(s, -1);
4931 SLAB_ATTR(store_user);
4933 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4938 static ssize_t validate_store(struct kmem_cache *s,
4939 const char *buf, size_t length)
4943 if (buf[0] == '1') {
4944 ret = validate_slab_cache(s);
4950 SLAB_ATTR(validate);
4952 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4954 if (!(s->flags & SLAB_STORE_USER))
4956 return list_locations(s, buf, TRACK_ALLOC);
4958 SLAB_ATTR_RO(alloc_calls);
4960 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4962 if (!(s->flags & SLAB_STORE_USER))
4964 return list_locations(s, buf, TRACK_FREE);
4966 SLAB_ATTR_RO(free_calls);
4967 #endif /* CONFIG_SLUB_DEBUG */
4969 #ifdef CONFIG_FAILSLAB
4970 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4972 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4975 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4978 s->flags &= ~SLAB_FAILSLAB;
4980 s->flags |= SLAB_FAILSLAB;
4983 SLAB_ATTR(failslab);
4986 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4991 static ssize_t shrink_store(struct kmem_cache *s,
4992 const char *buf, size_t length)
4994 if (buf[0] == '1') {
4995 int rc = kmem_cache_shrink(s);
5006 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5008 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5011 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5012 const char *buf, size_t length)
5014 unsigned long ratio;
5017 err = strict_strtoul(buf, 10, &ratio);
5022 s->remote_node_defrag_ratio = ratio * 10;
5026 SLAB_ATTR(remote_node_defrag_ratio);
5029 #ifdef CONFIG_SLUB_STATS
5030 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5032 unsigned long sum = 0;
5035 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5040 for_each_online_cpu(cpu) {
5041 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5047 len = sprintf(buf, "%lu", sum);
5050 for_each_online_cpu(cpu) {
5051 if (data[cpu] && len < PAGE_SIZE - 20)
5052 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5056 return len + sprintf(buf + len, "\n");
5059 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5063 for_each_online_cpu(cpu)
5064 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5067 #define STAT_ATTR(si, text) \
5068 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5070 return show_stat(s, buf, si); \
5072 static ssize_t text##_store(struct kmem_cache *s, \
5073 const char *buf, size_t length) \
5075 if (buf[0] != '0') \
5077 clear_stat(s, si); \
5082 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5083 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5084 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5085 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5086 STAT_ATTR(FREE_FROZEN, free_frozen);
5087 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5088 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5089 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5090 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5091 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5092 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5093 STAT_ATTR(FREE_SLAB, free_slab);
5094 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5095 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5096 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5097 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5098 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5099 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5100 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5101 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5102 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5103 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5104 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5105 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5106 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5107 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5110 static struct attribute *slab_attrs[] = {
5111 &slab_size_attr.attr,
5112 &object_size_attr.attr,
5113 &objs_per_slab_attr.attr,
5115 &min_partial_attr.attr,
5116 &cpu_partial_attr.attr,
5118 &objects_partial_attr.attr,
5120 &cpu_slabs_attr.attr,
5124 &hwcache_align_attr.attr,
5125 &reclaim_account_attr.attr,
5126 &destroy_by_rcu_attr.attr,
5128 &reserved_attr.attr,
5129 &slabs_cpu_partial_attr.attr,
5130 #ifdef CONFIG_SLUB_DEBUG
5131 &total_objects_attr.attr,
5133 &sanity_checks_attr.attr,
5135 &red_zone_attr.attr,
5137 &store_user_attr.attr,
5138 &validate_attr.attr,
5139 &alloc_calls_attr.attr,
5140 &free_calls_attr.attr,
5142 #ifdef CONFIG_ZONE_DMA
5143 &cache_dma_attr.attr,
5146 &remote_node_defrag_ratio_attr.attr,
5148 #ifdef CONFIG_SLUB_STATS
5149 &alloc_fastpath_attr.attr,
5150 &alloc_slowpath_attr.attr,
5151 &free_fastpath_attr.attr,
5152 &free_slowpath_attr.attr,
5153 &free_frozen_attr.attr,
5154 &free_add_partial_attr.attr,
5155 &free_remove_partial_attr.attr,
5156 &alloc_from_partial_attr.attr,
5157 &alloc_slab_attr.attr,
5158 &alloc_refill_attr.attr,
5159 &alloc_node_mismatch_attr.attr,
5160 &free_slab_attr.attr,
5161 &cpuslab_flush_attr.attr,
5162 &deactivate_full_attr.attr,
5163 &deactivate_empty_attr.attr,
5164 &deactivate_to_head_attr.attr,
5165 &deactivate_to_tail_attr.attr,
5166 &deactivate_remote_frees_attr.attr,
5167 &deactivate_bypass_attr.attr,
5168 &order_fallback_attr.attr,
5169 &cmpxchg_double_fail_attr.attr,
5170 &cmpxchg_double_cpu_fail_attr.attr,
5171 &cpu_partial_alloc_attr.attr,
5172 &cpu_partial_free_attr.attr,
5173 &cpu_partial_node_attr.attr,
5174 &cpu_partial_drain_attr.attr,
5176 #ifdef CONFIG_FAILSLAB
5177 &failslab_attr.attr,
5183 static struct attribute_group slab_attr_group = {
5184 .attrs = slab_attrs,
5187 static ssize_t slab_attr_show(struct kobject *kobj,
5188 struct attribute *attr,
5191 struct slab_attribute *attribute;
5192 struct kmem_cache *s;
5195 attribute = to_slab_attr(attr);
5198 if (!attribute->show)
5201 err = attribute->show(s, buf);
5206 static ssize_t slab_attr_store(struct kobject *kobj,
5207 struct attribute *attr,
5208 const char *buf, size_t len)
5210 struct slab_attribute *attribute;
5211 struct kmem_cache *s;
5214 attribute = to_slab_attr(attr);
5217 if (!attribute->store)
5220 err = attribute->store(s, buf, len);
5225 static void kmem_cache_release(struct kobject *kobj)
5227 struct kmem_cache *s = to_slab(kobj);
5233 static const struct sysfs_ops slab_sysfs_ops = {
5234 .show = slab_attr_show,
5235 .store = slab_attr_store,
5238 static struct kobj_type slab_ktype = {
5239 .sysfs_ops = &slab_sysfs_ops,
5240 .release = kmem_cache_release
5243 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5245 struct kobj_type *ktype = get_ktype(kobj);
5247 if (ktype == &slab_ktype)
5252 static const struct kset_uevent_ops slab_uevent_ops = {
5253 .filter = uevent_filter,
5256 static struct kset *slab_kset;
5258 #define ID_STR_LENGTH 64
5260 /* Create a unique string id for a slab cache:
5262 * Format :[flags-]size
5264 static char *create_unique_id(struct kmem_cache *s)
5266 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5273 * First flags affecting slabcache operations. We will only
5274 * get here for aliasable slabs so we do not need to support
5275 * too many flags. The flags here must cover all flags that
5276 * are matched during merging to guarantee that the id is
5279 if (s->flags & SLAB_CACHE_DMA)
5281 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5283 if (s->flags & SLAB_DEBUG_FREE)
5285 if (!(s->flags & SLAB_NOTRACK))
5289 p += sprintf(p, "%07d", s->size);
5290 BUG_ON(p > name + ID_STR_LENGTH - 1);
5294 static int sysfs_slab_add(struct kmem_cache *s)
5300 if (slab_state < SYSFS)
5301 /* Defer until later */
5304 unmergeable = slab_unmergeable(s);
5307 * Slabcache can never be merged so we can use the name proper.
5308 * This is typically the case for debug situations. In that
5309 * case we can catch duplicate names easily.
5311 sysfs_remove_link(&slab_kset->kobj, s->name);
5315 * Create a unique name for the slab as a target
5318 name = create_unique_id(s);
5321 s->kobj.kset = slab_kset;
5322 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5324 kobject_put(&s->kobj);
5328 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5330 kobject_del(&s->kobj);
5331 kobject_put(&s->kobj);
5334 kobject_uevent(&s->kobj, KOBJ_ADD);
5336 /* Setup first alias */
5337 sysfs_slab_alias(s, s->name);
5343 static void sysfs_slab_remove(struct kmem_cache *s)
5345 if (slab_state < SYSFS)
5347 * Sysfs has not been setup yet so no need to remove the
5352 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5353 kobject_del(&s->kobj);
5354 kobject_put(&s->kobj);
5358 * Need to buffer aliases during bootup until sysfs becomes
5359 * available lest we lose that information.
5361 struct saved_alias {
5362 struct kmem_cache *s;
5364 struct saved_alias *next;
5367 static struct saved_alias *alias_list;
5369 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5371 struct saved_alias *al;
5373 if (slab_state == SYSFS) {
5375 * If we have a leftover link then remove it.
5377 sysfs_remove_link(&slab_kset->kobj, name);
5378 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5381 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5387 al->next = alias_list;
5392 static int __init slab_sysfs_init(void)
5394 struct kmem_cache *s;
5397 down_write(&slub_lock);
5399 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5401 up_write(&slub_lock);
5402 printk(KERN_ERR "Cannot register slab subsystem.\n");
5408 list_for_each_entry(s, &slab_caches, list) {
5409 err = sysfs_slab_add(s);
5411 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5412 " to sysfs\n", s->name);
5415 while (alias_list) {
5416 struct saved_alias *al = alias_list;
5418 alias_list = alias_list->next;
5419 err = sysfs_slab_alias(al->s, al->name);
5421 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5422 " %s to sysfs\n", s->name);
5426 up_write(&slub_lock);
5431 __initcall(slab_sysfs_init);
5432 #endif /* CONFIG_SYSFS */
5435 * The /proc/slabinfo ABI
5437 #ifdef CONFIG_SLABINFO
5438 static void print_slabinfo_header(struct seq_file *m)
5440 seq_puts(m, "slabinfo - version: 2.1\n");
5441 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5442 "<objperslab> <pagesperslab>");
5443 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5444 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5448 static void *s_start(struct seq_file *m, loff_t *pos)
5452 down_read(&slub_lock);
5454 print_slabinfo_header(m);
5456 return seq_list_start(&slab_caches, *pos);
5459 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5461 return seq_list_next(p, &slab_caches, pos);
5464 static void s_stop(struct seq_file *m, void *p)
5466 up_read(&slub_lock);
5469 static int s_show(struct seq_file *m, void *p)
5471 unsigned long nr_partials = 0;
5472 unsigned long nr_slabs = 0;
5473 unsigned long nr_inuse = 0;
5474 unsigned long nr_objs = 0;
5475 unsigned long nr_free = 0;
5476 struct kmem_cache *s;
5479 s = list_entry(p, struct kmem_cache, list);
5481 for_each_online_node(node) {
5482 struct kmem_cache_node *n = get_node(s, node);
5487 nr_partials += n->nr_partial;
5488 nr_slabs += atomic_long_read(&n->nr_slabs);
5489 nr_objs += atomic_long_read(&n->total_objects);
5490 nr_free += count_partial(n, count_free);
5493 nr_inuse = nr_objs - nr_free;
5495 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5496 nr_objs, s->size, oo_objects(s->oo),
5497 (1 << oo_order(s->oo)));
5498 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5499 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5505 static const struct seq_operations slabinfo_op = {
5512 static int slabinfo_open(struct inode *inode, struct file *file)
5514 return seq_open(file, &slabinfo_op);
5517 static const struct file_operations proc_slabinfo_operations = {
5518 .open = slabinfo_open,
5520 .llseek = seq_lseek,
5521 .release = seq_release,
5524 static int __init slab_proc_init(void)
5526 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5529 module_init(slab_proc_init);
5530 #endif /* CONFIG_SLABINFO */