2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
35 #include <trace/events/kmem.h>
41 * 1. slab_mutex (Global Mutex)
43 * 3. slab_lock(page) (Only on some arches and for debugging)
47 * The role of the slab_mutex is to protect the list of all the slabs
48 * and to synchronize major metadata changes to slab cache structures.
50 * The slab_lock is only used for debugging and on arches that do not
51 * have the ability to do a cmpxchg_double. It only protects the second
52 * double word in the page struct. Meaning
53 * A. page->freelist -> List of object free in a page
54 * B. page->counters -> Counters of objects
55 * C. page->frozen -> frozen state
57 * If a slab is frozen then it is exempt from list management. It is not
58 * on any list. The processor that froze the slab is the one who can
59 * perform list operations on the page. Other processors may put objects
60 * onto the freelist but the processor that froze the slab is the only
61 * one that can retrieve the objects from the page's freelist.
63 * The list_lock protects the partial and full list on each node and
64 * the partial slab counter. If taken then no new slabs may be added or
65 * removed from the lists nor make the number of partial slabs be modified.
66 * (Note that the total number of slabs is an atomic value that may be
67 * modified without taking the list lock).
69 * The list_lock is a centralized lock and thus we avoid taking it as
70 * much as possible. As long as SLUB does not have to handle partial
71 * slabs, operations can continue without any centralized lock. F.e.
72 * allocating a long series of objects that fill up slabs does not require
74 * Interrupts are disabled during allocation and deallocation in order to
75 * make the slab allocator safe to use in the context of an irq. In addition
76 * interrupts are disabled to ensure that the processor does not change
77 * while handling per_cpu slabs, due to kernel preemption.
79 * SLUB assigns one slab for allocation to each processor.
80 * Allocations only occur from these slabs called cpu slabs.
82 * Slabs with free elements are kept on a partial list and during regular
83 * operations no list for full slabs is used. If an object in a full slab is
84 * freed then the slab will show up again on the partial lists.
85 * We track full slabs for debugging purposes though because otherwise we
86 * cannot scan all objects.
88 * Slabs are freed when they become empty. Teardown and setup is
89 * minimal so we rely on the page allocators per cpu caches for
90 * fast frees and allocs.
92 * Overloading of page flags that are otherwise used for LRU management.
94 * PageActive The slab is frozen and exempt from list processing.
95 * This means that the slab is dedicated to a purpose
96 * such as satisfying allocations for a specific
97 * processor. Objects may be freed in the slab while
98 * it is frozen but slab_free will then skip the usual
99 * list operations. It is up to the processor holding
100 * the slab to integrate the slab into the slab lists
101 * when the slab is no longer needed.
103 * One use of this flag is to mark slabs that are
104 * used for allocations. Then such a slab becomes a cpu
105 * slab. The cpu slab may be equipped with an additional
106 * freelist that allows lockless access to
107 * free objects in addition to the regular freelist
108 * that requires the slab lock.
110 * PageError Slab requires special handling due to debug
111 * options set. This moves slab handling out of
112 * the fast path and disables lockless freelists.
115 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
116 SLAB_TRACE | SLAB_DEBUG_FREE)
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
128 * Issues still to be resolved:
130 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
132 * - Variable sizing of the per node arrays
135 /* Enable to test recovery from slab corruption on boot */
136 #undef SLUB_RESILIENCY_TEST
138 /* Enable to log cmpxchg failures */
139 #undef SLUB_DEBUG_CMPXCHG
142 * Mininum number of partial slabs. These will be left on the partial
143 * lists even if they are empty. kmem_cache_shrink may reclaim them.
145 #define MIN_PARTIAL 5
148 * Maximum number of desirable partial slabs.
149 * The existence of more partial slabs makes kmem_cache_shrink
150 * sort the partial list by the number of objects in the.
152 #define MAX_PARTIAL 10
154 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_STORE_USER)
158 * Debugging flags that require metadata to be stored in the slab. These get
159 * disabled when slub_debug=O is used and a cache's min order increases with
162 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
165 * Set of flags that will prevent slab merging
167 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
168 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
171 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
172 SLAB_CACHE_DMA | SLAB_NOTRACK)
175 #define OO_MASK ((1 << OO_SHIFT) - 1)
176 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
178 /* Internal SLUB flags */
179 #define __OBJECT_POISON 0x80000000UL /* Poison object */
180 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static int kmem_size = sizeof(struct kmem_cache);
185 static struct notifier_block slab_notifier;
189 * Tracking user of a slab.
191 #define TRACK_ADDRS_COUNT 16
193 unsigned long addr; /* Called from address */
194 #ifdef CONFIG_STACKTRACE
195 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
197 int cpu; /* Was running on cpu */
198 int pid; /* Pid context */
199 unsigned long when; /* When did the operation occur */
202 enum track_item { TRACK_ALLOC, TRACK_FREE };
205 static int sysfs_slab_add(struct kmem_cache *);
206 static int sysfs_slab_alias(struct kmem_cache *, const char *);
207 static void sysfs_slab_remove(struct kmem_cache *);
210 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
211 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
213 static inline void sysfs_slab_remove(struct kmem_cache *s)
220 static inline void stat(const struct kmem_cache *s, enum stat_item si)
222 #ifdef CONFIG_SLUB_STATS
223 __this_cpu_inc(s->cpu_slab->stat[si]);
227 /********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
231 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
233 return s->node[node];
236 /* Verify that a pointer has an address that is valid within a slab page */
237 static inline int check_valid_pointer(struct kmem_cache *s,
238 struct page *page, const void *object)
245 base = page_address(page);
246 if (object < base || object >= base + page->objects * s->size ||
247 (object - base) % s->size) {
254 static inline void *get_freepointer(struct kmem_cache *s, void *object)
256 return *(void **)(object + s->offset);
259 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
261 prefetch(object + s->offset);
264 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
268 #ifdef CONFIG_DEBUG_PAGEALLOC
269 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
271 p = get_freepointer(s, object);
276 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
278 *(void **)(object + s->offset) = fp;
281 /* Loop over all objects in a slab */
282 #define for_each_object(__p, __s, __addr, __objects) \
283 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
286 /* Determine object index from a given position */
287 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
289 return (p - addr) / s->size;
292 static inline size_t slab_ksize(const struct kmem_cache *s)
294 #ifdef CONFIG_SLUB_DEBUG
296 * Debugging requires use of the padding between object
297 * and whatever may come after it.
299 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
300 return s->object_size;
304 * If we have the need to store the freelist pointer
305 * back there or track user information then we can
306 * only use the space before that information.
308 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
311 * Else we can use all the padding etc for the allocation
316 static inline int order_objects(int order, unsigned long size, int reserved)
318 return ((PAGE_SIZE << order) - reserved) / size;
321 static inline struct kmem_cache_order_objects oo_make(int order,
322 unsigned long size, int reserved)
324 struct kmem_cache_order_objects x = {
325 (order << OO_SHIFT) + order_objects(order, size, reserved)
331 static inline int oo_order(struct kmem_cache_order_objects x)
333 return x.x >> OO_SHIFT;
336 static inline int oo_objects(struct kmem_cache_order_objects x)
338 return x.x & OO_MASK;
342 * Per slab locking using the pagelock
344 static __always_inline void slab_lock(struct page *page)
346 bit_spin_lock(PG_locked, &page->flags);
349 static __always_inline void slab_unlock(struct page *page)
351 __bit_spin_unlock(PG_locked, &page->flags);
354 /* Interrupts must be disabled (for the fallback code to work right) */
355 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
356 void *freelist_old, unsigned long counters_old,
357 void *freelist_new, unsigned long counters_new,
360 VM_BUG_ON(!irqs_disabled());
361 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
362 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
363 if (s->flags & __CMPXCHG_DOUBLE) {
364 if (cmpxchg_double(&page->freelist, &page->counters,
365 freelist_old, counters_old,
366 freelist_new, counters_new))
372 if (page->freelist == freelist_old && page->counters == counters_old) {
373 page->freelist = freelist_new;
374 page->counters = counters_new;
382 stat(s, CMPXCHG_DOUBLE_FAIL);
384 #ifdef SLUB_DEBUG_CMPXCHG
385 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
391 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
392 void *freelist_old, unsigned long counters_old,
393 void *freelist_new, unsigned long counters_new,
396 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
397 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
398 if (s->flags & __CMPXCHG_DOUBLE) {
399 if (cmpxchg_double(&page->freelist, &page->counters,
400 freelist_old, counters_old,
401 freelist_new, counters_new))
408 local_irq_save(flags);
410 if (page->freelist == freelist_old && page->counters == counters_old) {
411 page->freelist = freelist_new;
412 page->counters = counters_new;
414 local_irq_restore(flags);
418 local_irq_restore(flags);
422 stat(s, CMPXCHG_DOUBLE_FAIL);
424 #ifdef SLUB_DEBUG_CMPXCHG
425 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
431 #ifdef CONFIG_SLUB_DEBUG
433 * Determine a map of object in use on a page.
435 * Node listlock must be held to guarantee that the page does
436 * not vanish from under us.
438 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
441 void *addr = page_address(page);
443 for (p = page->freelist; p; p = get_freepointer(s, p))
444 set_bit(slab_index(p, s, addr), map);
450 #ifdef CONFIG_SLUB_DEBUG_ON
451 static int slub_debug = DEBUG_DEFAULT_FLAGS;
453 static int slub_debug;
456 static char *slub_debug_slabs;
457 static int disable_higher_order_debug;
462 static void print_section(char *text, u8 *addr, unsigned int length)
464 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
468 static struct track *get_track(struct kmem_cache *s, void *object,
469 enum track_item alloc)
474 p = object + s->offset + sizeof(void *);
476 p = object + s->inuse;
481 static void set_track(struct kmem_cache *s, void *object,
482 enum track_item alloc, unsigned long addr)
484 struct track *p = get_track(s, object, alloc);
487 #ifdef CONFIG_STACKTRACE
488 struct stack_trace trace;
491 trace.nr_entries = 0;
492 trace.max_entries = TRACK_ADDRS_COUNT;
493 trace.entries = p->addrs;
495 save_stack_trace(&trace);
497 /* See rant in lockdep.c */
498 if (trace.nr_entries != 0 &&
499 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
502 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
506 p->cpu = smp_processor_id();
507 p->pid = current->pid;
510 memset(p, 0, sizeof(struct track));
513 static void init_tracking(struct kmem_cache *s, void *object)
515 if (!(s->flags & SLAB_STORE_USER))
518 set_track(s, object, TRACK_FREE, 0UL);
519 set_track(s, object, TRACK_ALLOC, 0UL);
522 static void print_track(const char *s, struct track *t)
527 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
528 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
529 #ifdef CONFIG_STACKTRACE
532 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
534 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
541 static void print_tracking(struct kmem_cache *s, void *object)
543 if (!(s->flags & SLAB_STORE_USER))
546 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
547 print_track("Freed", get_track(s, object, TRACK_FREE));
550 static void print_page_info(struct page *page)
552 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
553 page, page->objects, page->inuse, page->freelist, page->flags);
557 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
563 vsnprintf(buf, sizeof(buf), fmt, args);
565 printk(KERN_ERR "========================================"
566 "=====================================\n");
567 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
568 printk(KERN_ERR "----------------------------------------"
569 "-------------------------------------\n\n");
572 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
578 vsnprintf(buf, sizeof(buf), fmt, args);
580 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
583 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
585 unsigned int off; /* Offset of last byte */
586 u8 *addr = page_address(page);
588 print_tracking(s, p);
590 print_page_info(page);
592 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
593 p, p - addr, get_freepointer(s, p));
596 print_section("Bytes b4 ", p - 16, 16);
598 print_section("Object ", p, min_t(unsigned long, s->object_size,
600 if (s->flags & SLAB_RED_ZONE)
601 print_section("Redzone ", p + s->object_size,
602 s->inuse - s->object_size);
605 off = s->offset + sizeof(void *);
609 if (s->flags & SLAB_STORE_USER)
610 off += 2 * sizeof(struct track);
613 /* Beginning of the filler is the free pointer */
614 print_section("Padding ", p + off, s->size - off);
619 static void object_err(struct kmem_cache *s, struct page *page,
620 u8 *object, char *reason)
622 slab_bug(s, "%s", reason);
623 print_trailer(s, page, object);
626 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
632 vsnprintf(buf, sizeof(buf), fmt, args);
634 slab_bug(s, "%s", buf);
635 print_page_info(page);
639 static void init_object(struct kmem_cache *s, void *object, u8 val)
643 if (s->flags & __OBJECT_POISON) {
644 memset(p, POISON_FREE, s->object_size - 1);
645 p[s->object_size - 1] = POISON_END;
648 if (s->flags & SLAB_RED_ZONE)
649 memset(p + s->object_size, val, s->inuse - s->object_size);
652 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
653 void *from, void *to)
655 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
656 memset(from, data, to - from);
659 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
660 u8 *object, char *what,
661 u8 *start, unsigned int value, unsigned int bytes)
666 fault = memchr_inv(start, value, bytes);
671 while (end > fault && end[-1] == value)
674 slab_bug(s, "%s overwritten", what);
675 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
676 fault, end - 1, fault[0], value);
677 print_trailer(s, page, object);
679 restore_bytes(s, what, value, fault, end);
687 * Bytes of the object to be managed.
688 * If the freepointer may overlay the object then the free
689 * pointer is the first word of the object.
691 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
694 * object + s->object_size
695 * Padding to reach word boundary. This is also used for Redzoning.
696 * Padding is extended by another word if Redzoning is enabled and
697 * object_size == inuse.
699 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
700 * 0xcc (RED_ACTIVE) for objects in use.
703 * Meta data starts here.
705 * A. Free pointer (if we cannot overwrite object on free)
706 * B. Tracking data for SLAB_STORE_USER
707 * C. Padding to reach required alignment boundary or at mininum
708 * one word if debugging is on to be able to detect writes
709 * before the word boundary.
711 * Padding is done using 0x5a (POISON_INUSE)
714 * Nothing is used beyond s->size.
716 * If slabcaches are merged then the object_size and inuse boundaries are mostly
717 * ignored. And therefore no slab options that rely on these boundaries
718 * may be used with merged slabcaches.
721 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
723 unsigned long off = s->inuse; /* The end of info */
726 /* Freepointer is placed after the object. */
727 off += sizeof(void *);
729 if (s->flags & SLAB_STORE_USER)
730 /* We also have user information there */
731 off += 2 * sizeof(struct track);
736 return check_bytes_and_report(s, page, p, "Object padding",
737 p + off, POISON_INUSE, s->size - off);
740 /* Check the pad bytes at the end of a slab page */
741 static int slab_pad_check(struct kmem_cache *s, struct page *page)
749 if (!(s->flags & SLAB_POISON))
752 start = page_address(page);
753 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
754 end = start + length;
755 remainder = length % s->size;
759 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
762 while (end > fault && end[-1] == POISON_INUSE)
765 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
766 print_section("Padding ", end - remainder, remainder);
768 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
772 static int check_object(struct kmem_cache *s, struct page *page,
773 void *object, u8 val)
776 u8 *endobject = object + s->object_size;
778 if (s->flags & SLAB_RED_ZONE) {
779 if (!check_bytes_and_report(s, page, object, "Redzone",
780 endobject, val, s->inuse - s->object_size))
783 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
784 check_bytes_and_report(s, page, p, "Alignment padding",
785 endobject, POISON_INUSE, s->inuse - s->object_size);
789 if (s->flags & SLAB_POISON) {
790 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
791 (!check_bytes_and_report(s, page, p, "Poison", p,
792 POISON_FREE, s->object_size - 1) ||
793 !check_bytes_and_report(s, page, p, "Poison",
794 p + s->object_size - 1, POISON_END, 1)))
797 * check_pad_bytes cleans up on its own.
799 check_pad_bytes(s, page, p);
802 if (!s->offset && val == SLUB_RED_ACTIVE)
804 * Object and freepointer overlap. Cannot check
805 * freepointer while object is allocated.
809 /* Check free pointer validity */
810 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
811 object_err(s, page, p, "Freepointer corrupt");
813 * No choice but to zap it and thus lose the remainder
814 * of the free objects in this slab. May cause
815 * another error because the object count is now wrong.
817 set_freepointer(s, p, NULL);
823 static int check_slab(struct kmem_cache *s, struct page *page)
827 VM_BUG_ON(!irqs_disabled());
829 if (!PageSlab(page)) {
830 slab_err(s, page, "Not a valid slab page");
834 maxobj = order_objects(compound_order(page), s->size, s->reserved);
835 if (page->objects > maxobj) {
836 slab_err(s, page, "objects %u > max %u",
837 s->name, page->objects, maxobj);
840 if (page->inuse > page->objects) {
841 slab_err(s, page, "inuse %u > max %u",
842 s->name, page->inuse, page->objects);
845 /* Slab_pad_check fixes things up after itself */
846 slab_pad_check(s, page);
851 * Determine if a certain object on a page is on the freelist. Must hold the
852 * slab lock to guarantee that the chains are in a consistent state.
854 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
859 unsigned long max_objects;
862 while (fp && nr <= page->objects) {
865 if (!check_valid_pointer(s, page, fp)) {
867 object_err(s, page, object,
868 "Freechain corrupt");
869 set_freepointer(s, object, NULL);
872 slab_err(s, page, "Freepointer corrupt");
873 page->freelist = NULL;
874 page->inuse = page->objects;
875 slab_fix(s, "Freelist cleared");
881 fp = get_freepointer(s, object);
885 max_objects = order_objects(compound_order(page), s->size, s->reserved);
886 if (max_objects > MAX_OBJS_PER_PAGE)
887 max_objects = MAX_OBJS_PER_PAGE;
889 if (page->objects != max_objects) {
890 slab_err(s, page, "Wrong number of objects. Found %d but "
891 "should be %d", page->objects, max_objects);
892 page->objects = max_objects;
893 slab_fix(s, "Number of objects adjusted.");
895 if (page->inuse != page->objects - nr) {
896 slab_err(s, page, "Wrong object count. Counter is %d but "
897 "counted were %d", page->inuse, page->objects - nr);
898 page->inuse = page->objects - nr;
899 slab_fix(s, "Object count adjusted.");
901 return search == NULL;
904 static void trace(struct kmem_cache *s, struct page *page, void *object,
907 if (s->flags & SLAB_TRACE) {
908 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
910 alloc ? "alloc" : "free",
915 print_section("Object ", (void *)object, s->object_size);
922 * Hooks for other subsystems that check memory allocations. In a typical
923 * production configuration these hooks all should produce no code at all.
925 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
927 flags &= gfp_allowed_mask;
928 lockdep_trace_alloc(flags);
929 might_sleep_if(flags & __GFP_WAIT);
931 return should_failslab(s->object_size, flags, s->flags);
934 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
936 flags &= gfp_allowed_mask;
937 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
938 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
941 static inline void slab_free_hook(struct kmem_cache *s, void *x)
943 kmemleak_free_recursive(x, s->flags);
946 * Trouble is that we may no longer disable interupts in the fast path
947 * So in order to make the debug calls that expect irqs to be
948 * disabled we need to disable interrupts temporarily.
950 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
954 local_irq_save(flags);
955 kmemcheck_slab_free(s, x, s->object_size);
956 debug_check_no_locks_freed(x, s->object_size);
957 local_irq_restore(flags);
960 if (!(s->flags & SLAB_DEBUG_OBJECTS))
961 debug_check_no_obj_freed(x, s->object_size);
965 * Tracking of fully allocated slabs for debugging purposes.
967 * list_lock must be held.
969 static void add_full(struct kmem_cache *s,
970 struct kmem_cache_node *n, struct page *page)
972 if (!(s->flags & SLAB_STORE_USER))
975 list_add(&page->lru, &n->full);
979 * list_lock must be held.
981 static void remove_full(struct kmem_cache *s, struct page *page)
983 if (!(s->flags & SLAB_STORE_USER))
986 list_del(&page->lru);
989 /* Tracking of the number of slabs for debugging purposes */
990 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
992 struct kmem_cache_node *n = get_node(s, node);
994 return atomic_long_read(&n->nr_slabs);
997 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
999 return atomic_long_read(&n->nr_slabs);
1002 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1004 struct kmem_cache_node *n = get_node(s, node);
1007 * May be called early in order to allocate a slab for the
1008 * kmem_cache_node structure. Solve the chicken-egg
1009 * dilemma by deferring the increment of the count during
1010 * bootstrap (see early_kmem_cache_node_alloc).
1013 atomic_long_inc(&n->nr_slabs);
1014 atomic_long_add(objects, &n->total_objects);
1017 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1019 struct kmem_cache_node *n = get_node(s, node);
1021 atomic_long_dec(&n->nr_slabs);
1022 atomic_long_sub(objects, &n->total_objects);
1025 /* Object debug checks for alloc/free paths */
1026 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1029 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1032 init_object(s, object, SLUB_RED_INACTIVE);
1033 init_tracking(s, object);
1036 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1037 void *object, unsigned long addr)
1039 if (!check_slab(s, page))
1042 if (!check_valid_pointer(s, page, object)) {
1043 object_err(s, page, object, "Freelist Pointer check fails");
1047 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1050 /* Success perform special debug activities for allocs */
1051 if (s->flags & SLAB_STORE_USER)
1052 set_track(s, object, TRACK_ALLOC, addr);
1053 trace(s, page, object, 1);
1054 init_object(s, object, SLUB_RED_ACTIVE);
1058 if (PageSlab(page)) {
1060 * If this is a slab page then lets do the best we can
1061 * to avoid issues in the future. Marking all objects
1062 * as used avoids touching the remaining objects.
1064 slab_fix(s, "Marking all objects used");
1065 page->inuse = page->objects;
1066 page->freelist = NULL;
1071 static noinline struct kmem_cache_node *free_debug_processing(
1072 struct kmem_cache *s, struct page *page, void *object,
1073 unsigned long addr, unsigned long *flags)
1075 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1077 spin_lock_irqsave(&n->list_lock, *flags);
1080 if (!check_slab(s, page))
1083 if (!check_valid_pointer(s, page, object)) {
1084 slab_err(s, page, "Invalid object pointer 0x%p", object);
1088 if (on_freelist(s, page, object)) {
1089 object_err(s, page, object, "Object already free");
1093 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1096 if (unlikely(s != page->slab)) {
1097 if (!PageSlab(page)) {
1098 slab_err(s, page, "Attempt to free object(0x%p) "
1099 "outside of slab", object);
1100 } else if (!page->slab) {
1102 "SLUB <none>: no slab for object 0x%p.\n",
1106 object_err(s, page, object,
1107 "page slab pointer corrupt.");
1111 if (s->flags & SLAB_STORE_USER)
1112 set_track(s, object, TRACK_FREE, addr);
1113 trace(s, page, object, 0);
1114 init_object(s, object, SLUB_RED_INACTIVE);
1118 * Keep node_lock to preserve integrity
1119 * until the object is actually freed
1125 spin_unlock_irqrestore(&n->list_lock, *flags);
1126 slab_fix(s, "Object at 0x%p not freed", object);
1130 static int __init setup_slub_debug(char *str)
1132 slub_debug = DEBUG_DEFAULT_FLAGS;
1133 if (*str++ != '=' || !*str)
1135 * No options specified. Switch on full debugging.
1141 * No options but restriction on slabs. This means full
1142 * debugging for slabs matching a pattern.
1146 if (tolower(*str) == 'o') {
1148 * Avoid enabling debugging on caches if its minimum order
1149 * would increase as a result.
1151 disable_higher_order_debug = 1;
1158 * Switch off all debugging measures.
1163 * Determine which debug features should be switched on
1165 for (; *str && *str != ','; str++) {
1166 switch (tolower(*str)) {
1168 slub_debug |= SLAB_DEBUG_FREE;
1171 slub_debug |= SLAB_RED_ZONE;
1174 slub_debug |= SLAB_POISON;
1177 slub_debug |= SLAB_STORE_USER;
1180 slub_debug |= SLAB_TRACE;
1183 slub_debug |= SLAB_FAILSLAB;
1186 printk(KERN_ERR "slub_debug option '%c' "
1187 "unknown. skipped\n", *str);
1193 slub_debug_slabs = str + 1;
1198 __setup("slub_debug", setup_slub_debug);
1200 static unsigned long kmem_cache_flags(unsigned long object_size,
1201 unsigned long flags, const char *name,
1202 void (*ctor)(void *))
1205 * Enable debugging if selected on the kernel commandline.
1207 if (slub_debug && (!slub_debug_slabs ||
1208 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1209 flags |= slub_debug;
1214 static inline void setup_object_debug(struct kmem_cache *s,
1215 struct page *page, void *object) {}
1217 static inline int alloc_debug_processing(struct kmem_cache *s,
1218 struct page *page, void *object, unsigned long addr) { return 0; }
1220 static inline struct kmem_cache_node *free_debug_processing(
1221 struct kmem_cache *s, struct page *page, void *object,
1222 unsigned long addr, unsigned long *flags) { return NULL; }
1224 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1226 static inline int check_object(struct kmem_cache *s, struct page *page,
1227 void *object, u8 val) { return 1; }
1228 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1229 struct page *page) {}
1230 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1231 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1232 unsigned long flags, const char *name,
1233 void (*ctor)(void *))
1237 #define slub_debug 0
1239 #define disable_higher_order_debug 0
1241 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1243 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1245 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1247 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1250 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1253 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1256 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1258 #endif /* CONFIG_SLUB_DEBUG */
1261 * Slab allocation and freeing
1263 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1264 struct kmem_cache_order_objects oo)
1266 int order = oo_order(oo);
1268 flags |= __GFP_NOTRACK;
1270 if (node == NUMA_NO_NODE)
1271 return alloc_pages(flags, order);
1273 return alloc_pages_exact_node(node, flags, order);
1276 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1279 struct kmem_cache_order_objects oo = s->oo;
1282 flags &= gfp_allowed_mask;
1284 if (flags & __GFP_WAIT)
1287 flags |= s->allocflags;
1290 * Let the initial higher-order allocation fail under memory pressure
1291 * so we fall-back to the minimum order allocation.
1293 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1295 page = alloc_slab_page(alloc_gfp, node, oo);
1296 if (unlikely(!page)) {
1299 * Allocation may have failed due to fragmentation.
1300 * Try a lower order alloc if possible
1302 page = alloc_slab_page(flags, node, oo);
1305 stat(s, ORDER_FALLBACK);
1308 if (kmemcheck_enabled && page
1309 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1310 int pages = 1 << oo_order(oo);
1312 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1315 * Objects from caches that have a constructor don't get
1316 * cleared when they're allocated, so we need to do it here.
1319 kmemcheck_mark_uninitialized_pages(page, pages);
1321 kmemcheck_mark_unallocated_pages(page, pages);
1324 if (flags & __GFP_WAIT)
1325 local_irq_disable();
1329 page->objects = oo_objects(oo);
1330 mod_zone_page_state(page_zone(page),
1331 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1332 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1338 static void setup_object(struct kmem_cache *s, struct page *page,
1341 setup_object_debug(s, page, object);
1342 if (unlikely(s->ctor))
1346 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1353 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1355 page = allocate_slab(s,
1356 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1360 inc_slabs_node(s, page_to_nid(page), page->objects);
1362 __SetPageSlab(page);
1363 if (page->pfmemalloc)
1364 SetPageSlabPfmemalloc(page);
1366 start = page_address(page);
1368 if (unlikely(s->flags & SLAB_POISON))
1369 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1372 for_each_object(p, s, start, page->objects) {
1373 setup_object(s, page, last);
1374 set_freepointer(s, last, p);
1377 setup_object(s, page, last);
1378 set_freepointer(s, last, NULL);
1380 page->freelist = start;
1381 page->inuse = page->objects;
1387 static void __free_slab(struct kmem_cache *s, struct page *page)
1389 int order = compound_order(page);
1390 int pages = 1 << order;
1392 if (kmem_cache_debug(s)) {
1395 slab_pad_check(s, page);
1396 for_each_object(p, s, page_address(page),
1398 check_object(s, page, p, SLUB_RED_INACTIVE);
1401 kmemcheck_free_shadow(page, compound_order(page));
1403 mod_zone_page_state(page_zone(page),
1404 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1405 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1408 __ClearPageSlabPfmemalloc(page);
1409 __ClearPageSlab(page);
1410 reset_page_mapcount(page);
1411 if (current->reclaim_state)
1412 current->reclaim_state->reclaimed_slab += pages;
1413 __free_pages(page, order);
1416 #define need_reserve_slab_rcu \
1417 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1419 static void rcu_free_slab(struct rcu_head *h)
1423 if (need_reserve_slab_rcu)
1424 page = virt_to_head_page(h);
1426 page = container_of((struct list_head *)h, struct page, lru);
1428 __free_slab(page->slab, page);
1431 static void free_slab(struct kmem_cache *s, struct page *page)
1433 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1434 struct rcu_head *head;
1436 if (need_reserve_slab_rcu) {
1437 int order = compound_order(page);
1438 int offset = (PAGE_SIZE << order) - s->reserved;
1440 VM_BUG_ON(s->reserved != sizeof(*head));
1441 head = page_address(page) + offset;
1444 * RCU free overloads the RCU head over the LRU
1446 head = (void *)&page->lru;
1449 call_rcu(head, rcu_free_slab);
1451 __free_slab(s, page);
1454 static void discard_slab(struct kmem_cache *s, struct page *page)
1456 dec_slabs_node(s, page_to_nid(page), page->objects);
1461 * Management of partially allocated slabs.
1463 * list_lock must be held.
1465 static inline void add_partial(struct kmem_cache_node *n,
1466 struct page *page, int tail)
1469 if (tail == DEACTIVATE_TO_TAIL)
1470 list_add_tail(&page->lru, &n->partial);
1472 list_add(&page->lru, &n->partial);
1476 * list_lock must be held.
1478 static inline void remove_partial(struct kmem_cache_node *n,
1481 list_del(&page->lru);
1486 * Remove slab from the partial list, freeze it and
1487 * return the pointer to the freelist.
1489 * Returns a list of objects or NULL if it fails.
1491 * Must hold list_lock since we modify the partial list.
1493 static inline void *acquire_slab(struct kmem_cache *s,
1494 struct kmem_cache_node *n, struct page *page,
1498 unsigned long counters;
1502 * Zap the freelist and set the frozen bit.
1503 * The old freelist is the list of objects for the
1504 * per cpu allocation list.
1506 freelist = page->freelist;
1507 counters = page->counters;
1508 new.counters = counters;
1510 new.inuse = page->objects;
1511 new.freelist = NULL;
1513 new.freelist = freelist;
1516 VM_BUG_ON(new.frozen);
1519 if (!__cmpxchg_double_slab(s, page,
1521 new.freelist, new.counters,
1525 remove_partial(n, page);
1530 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1533 * Try to allocate a partial slab from a specific node.
1535 static void *get_partial_node(struct kmem_cache *s,
1536 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1538 struct page *page, *page2;
1539 void *object = NULL;
1542 * Racy check. If we mistakenly see no partial slabs then we
1543 * just allocate an empty slab. If we mistakenly try to get a
1544 * partial slab and there is none available then get_partials()
1547 if (!n || !n->nr_partial)
1550 spin_lock(&n->list_lock);
1551 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1552 void *t = acquire_slab(s, n, page, object == NULL);
1560 stat(s, ALLOC_FROM_PARTIAL);
1562 available = page->objects - page->inuse;
1564 available = put_cpu_partial(s, page, 0);
1565 stat(s, CPU_PARTIAL_NODE);
1567 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1571 spin_unlock(&n->list_lock);
1576 * Get a page from somewhere. Search in increasing NUMA distances.
1578 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1579 struct kmem_cache_cpu *c)
1582 struct zonelist *zonelist;
1585 enum zone_type high_zoneidx = gfp_zone(flags);
1587 unsigned int cpuset_mems_cookie;
1590 * The defrag ratio allows a configuration of the tradeoffs between
1591 * inter node defragmentation and node local allocations. A lower
1592 * defrag_ratio increases the tendency to do local allocations
1593 * instead of attempting to obtain partial slabs from other nodes.
1595 * If the defrag_ratio is set to 0 then kmalloc() always
1596 * returns node local objects. If the ratio is higher then kmalloc()
1597 * may return off node objects because partial slabs are obtained
1598 * from other nodes and filled up.
1600 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1601 * defrag_ratio = 1000) then every (well almost) allocation will
1602 * first attempt to defrag slab caches on other nodes. This means
1603 * scanning over all nodes to look for partial slabs which may be
1604 * expensive if we do it every time we are trying to find a slab
1605 * with available objects.
1607 if (!s->remote_node_defrag_ratio ||
1608 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1612 cpuset_mems_cookie = get_mems_allowed();
1613 zonelist = node_zonelist(slab_node(), flags);
1614 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1615 struct kmem_cache_node *n;
1617 n = get_node(s, zone_to_nid(zone));
1619 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1620 n->nr_partial > s->min_partial) {
1621 object = get_partial_node(s, n, c);
1624 * Return the object even if
1625 * put_mems_allowed indicated that
1626 * the cpuset mems_allowed was
1627 * updated in parallel. It's a
1628 * harmless race between the alloc
1629 * and the cpuset update.
1631 put_mems_allowed(cpuset_mems_cookie);
1636 } while (!put_mems_allowed(cpuset_mems_cookie));
1642 * Get a partial page, lock it and return it.
1644 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1645 struct kmem_cache_cpu *c)
1648 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1650 object = get_partial_node(s, get_node(s, searchnode), c);
1651 if (object || node != NUMA_NO_NODE)
1654 return get_any_partial(s, flags, c);
1657 #ifdef CONFIG_PREEMPT
1659 * Calculate the next globally unique transaction for disambiguiation
1660 * during cmpxchg. The transactions start with the cpu number and are then
1661 * incremented by CONFIG_NR_CPUS.
1663 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1666 * No preemption supported therefore also no need to check for
1672 static inline unsigned long next_tid(unsigned long tid)
1674 return tid + TID_STEP;
1677 static inline unsigned int tid_to_cpu(unsigned long tid)
1679 return tid % TID_STEP;
1682 static inline unsigned long tid_to_event(unsigned long tid)
1684 return tid / TID_STEP;
1687 static inline unsigned int init_tid(int cpu)
1692 static inline void note_cmpxchg_failure(const char *n,
1693 const struct kmem_cache *s, unsigned long tid)
1695 #ifdef SLUB_DEBUG_CMPXCHG
1696 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1698 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1700 #ifdef CONFIG_PREEMPT
1701 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1702 printk("due to cpu change %d -> %d\n",
1703 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1706 if (tid_to_event(tid) != tid_to_event(actual_tid))
1707 printk("due to cpu running other code. Event %ld->%ld\n",
1708 tid_to_event(tid), tid_to_event(actual_tid));
1710 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1711 actual_tid, tid, next_tid(tid));
1713 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1716 void init_kmem_cache_cpus(struct kmem_cache *s)
1720 for_each_possible_cpu(cpu)
1721 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1725 * Remove the cpu slab
1727 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1729 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1730 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1732 enum slab_modes l = M_NONE, m = M_NONE;
1734 int tail = DEACTIVATE_TO_HEAD;
1738 if (page->freelist) {
1739 stat(s, DEACTIVATE_REMOTE_FREES);
1740 tail = DEACTIVATE_TO_TAIL;
1744 * Stage one: Free all available per cpu objects back
1745 * to the page freelist while it is still frozen. Leave the
1748 * There is no need to take the list->lock because the page
1751 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1753 unsigned long counters;
1756 prior = page->freelist;
1757 counters = page->counters;
1758 set_freepointer(s, freelist, prior);
1759 new.counters = counters;
1761 VM_BUG_ON(!new.frozen);
1763 } while (!__cmpxchg_double_slab(s, page,
1765 freelist, new.counters,
1766 "drain percpu freelist"));
1768 freelist = nextfree;
1772 * Stage two: Ensure that the page is unfrozen while the
1773 * list presence reflects the actual number of objects
1776 * We setup the list membership and then perform a cmpxchg
1777 * with the count. If there is a mismatch then the page
1778 * is not unfrozen but the page is on the wrong list.
1780 * Then we restart the process which may have to remove
1781 * the page from the list that we just put it on again
1782 * because the number of objects in the slab may have
1787 old.freelist = page->freelist;
1788 old.counters = page->counters;
1789 VM_BUG_ON(!old.frozen);
1791 /* Determine target state of the slab */
1792 new.counters = old.counters;
1795 set_freepointer(s, freelist, old.freelist);
1796 new.freelist = freelist;
1798 new.freelist = old.freelist;
1802 if (!new.inuse && n->nr_partial > s->min_partial)
1804 else if (new.freelist) {
1809 * Taking the spinlock removes the possiblity
1810 * that acquire_slab() will see a slab page that
1813 spin_lock(&n->list_lock);
1817 if (kmem_cache_debug(s) && !lock) {
1820 * This also ensures that the scanning of full
1821 * slabs from diagnostic functions will not see
1824 spin_lock(&n->list_lock);
1832 remove_partial(n, page);
1834 else if (l == M_FULL)
1836 remove_full(s, page);
1838 if (m == M_PARTIAL) {
1840 add_partial(n, page, tail);
1843 } else if (m == M_FULL) {
1845 stat(s, DEACTIVATE_FULL);
1846 add_full(s, n, page);
1852 if (!__cmpxchg_double_slab(s, page,
1853 old.freelist, old.counters,
1854 new.freelist, new.counters,
1859 spin_unlock(&n->list_lock);
1862 stat(s, DEACTIVATE_EMPTY);
1863 discard_slab(s, page);
1869 * Unfreeze all the cpu partial slabs.
1871 * This function must be called with interrupt disabled.
1873 static void unfreeze_partials(struct kmem_cache *s)
1875 struct kmem_cache_node *n = NULL, *n2 = NULL;
1876 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1877 struct page *page, *discard_page = NULL;
1879 while ((page = c->partial)) {
1883 c->partial = page->next;
1885 n2 = get_node(s, page_to_nid(page));
1888 spin_unlock(&n->list_lock);
1891 spin_lock(&n->list_lock);
1896 old.freelist = page->freelist;
1897 old.counters = page->counters;
1898 VM_BUG_ON(!old.frozen);
1900 new.counters = old.counters;
1901 new.freelist = old.freelist;
1905 } while (!__cmpxchg_double_slab(s, page,
1906 old.freelist, old.counters,
1907 new.freelist, new.counters,
1908 "unfreezing slab"));
1910 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1911 page->next = discard_page;
1912 discard_page = page;
1914 add_partial(n, page, DEACTIVATE_TO_TAIL);
1915 stat(s, FREE_ADD_PARTIAL);
1920 spin_unlock(&n->list_lock);
1922 while (discard_page) {
1923 page = discard_page;
1924 discard_page = discard_page->next;
1926 stat(s, DEACTIVATE_EMPTY);
1927 discard_slab(s, page);
1933 * Put a page that was just frozen (in __slab_free) into a partial page
1934 * slot if available. This is done without interrupts disabled and without
1935 * preemption disabled. The cmpxchg is racy and may put the partial page
1936 * onto a random cpus partial slot.
1938 * If we did not find a slot then simply move all the partials to the
1939 * per node partial list.
1941 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1943 struct page *oldpage;
1950 oldpage = this_cpu_read(s->cpu_slab->partial);
1953 pobjects = oldpage->pobjects;
1954 pages = oldpage->pages;
1955 if (drain && pobjects > s->cpu_partial) {
1956 unsigned long flags;
1958 * partial array is full. Move the existing
1959 * set to the per node partial list.
1961 local_irq_save(flags);
1962 unfreeze_partials(s);
1963 local_irq_restore(flags);
1967 stat(s, CPU_PARTIAL_DRAIN);
1972 pobjects += page->objects - page->inuse;
1974 page->pages = pages;
1975 page->pobjects = pobjects;
1976 page->next = oldpage;
1978 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1982 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1984 stat(s, CPUSLAB_FLUSH);
1985 deactivate_slab(s, c->page, c->freelist);
1987 c->tid = next_tid(c->tid);
1995 * Called from IPI handler with interrupts disabled.
1997 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1999 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2005 unfreeze_partials(s);
2009 static void flush_cpu_slab(void *d)
2011 struct kmem_cache *s = d;
2013 __flush_cpu_slab(s, smp_processor_id());
2016 static bool has_cpu_slab(int cpu, void *info)
2018 struct kmem_cache *s = info;
2019 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2021 return c->page || c->partial;
2024 static void flush_all(struct kmem_cache *s)
2026 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2030 * Check if the objects in a per cpu structure fit numa
2031 * locality expectations.
2033 static inline int node_match(struct page *page, int node)
2036 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2042 static int count_free(struct page *page)
2044 return page->objects - page->inuse;
2047 static unsigned long count_partial(struct kmem_cache_node *n,
2048 int (*get_count)(struct page *))
2050 unsigned long flags;
2051 unsigned long x = 0;
2054 spin_lock_irqsave(&n->list_lock, flags);
2055 list_for_each_entry(page, &n->partial, lru)
2056 x += get_count(page);
2057 spin_unlock_irqrestore(&n->list_lock, flags);
2061 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2063 #ifdef CONFIG_SLUB_DEBUG
2064 return atomic_long_read(&n->total_objects);
2070 static noinline void
2071 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2076 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2078 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2079 "default order: %d, min order: %d\n", s->name, s->object_size,
2080 s->size, oo_order(s->oo), oo_order(s->min));
2082 if (oo_order(s->min) > get_order(s->object_size))
2083 printk(KERN_WARNING " %s debugging increased min order, use "
2084 "slub_debug=O to disable.\n", s->name);
2086 for_each_online_node(node) {
2087 struct kmem_cache_node *n = get_node(s, node);
2088 unsigned long nr_slabs;
2089 unsigned long nr_objs;
2090 unsigned long nr_free;
2095 nr_free = count_partial(n, count_free);
2096 nr_slabs = node_nr_slabs(n);
2097 nr_objs = node_nr_objs(n);
2100 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2101 node, nr_slabs, nr_objs, nr_free);
2105 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2106 int node, struct kmem_cache_cpu **pc)
2109 struct kmem_cache_cpu *c = *pc;
2112 freelist = get_partial(s, flags, node, c);
2117 page = new_slab(s, flags, node);
2119 c = __this_cpu_ptr(s->cpu_slab);
2124 * No other reference to the page yet so we can
2125 * muck around with it freely without cmpxchg
2127 freelist = page->freelist;
2128 page->freelist = NULL;
2130 stat(s, ALLOC_SLAB);
2139 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2141 if (unlikely(PageSlabPfmemalloc(page)))
2142 return gfp_pfmemalloc_allowed(gfpflags);
2148 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2149 * or deactivate the page.
2151 * The page is still frozen if the return value is not NULL.
2153 * If this function returns NULL then the page has been unfrozen.
2155 * This function must be called with interrupt disabled.
2157 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2160 unsigned long counters;
2164 freelist = page->freelist;
2165 counters = page->counters;
2167 new.counters = counters;
2168 VM_BUG_ON(!new.frozen);
2170 new.inuse = page->objects;
2171 new.frozen = freelist != NULL;
2173 } while (!__cmpxchg_double_slab(s, page,
2182 * Slow path. The lockless freelist is empty or we need to perform
2185 * Processing is still very fast if new objects have been freed to the
2186 * regular freelist. In that case we simply take over the regular freelist
2187 * as the lockless freelist and zap the regular freelist.
2189 * If that is not working then we fall back to the partial lists. We take the
2190 * first element of the freelist as the object to allocate now and move the
2191 * rest of the freelist to the lockless freelist.
2193 * And if we were unable to get a new slab from the partial slab lists then
2194 * we need to allocate a new slab. This is the slowest path since it involves
2195 * a call to the page allocator and the setup of a new slab.
2197 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2198 unsigned long addr, struct kmem_cache_cpu *c)
2202 unsigned long flags;
2204 local_irq_save(flags);
2205 #ifdef CONFIG_PREEMPT
2207 * We may have been preempted and rescheduled on a different
2208 * cpu before disabling interrupts. Need to reload cpu area
2211 c = this_cpu_ptr(s->cpu_slab);
2219 if (unlikely(!node_match(page, node))) {
2220 stat(s, ALLOC_NODE_MISMATCH);
2221 deactivate_slab(s, page, c->freelist);
2228 * By rights, we should be searching for a slab page that was
2229 * PFMEMALLOC but right now, we are losing the pfmemalloc
2230 * information when the page leaves the per-cpu allocator
2232 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2233 deactivate_slab(s, page, c->freelist);
2239 /* must check again c->freelist in case of cpu migration or IRQ */
2240 freelist = c->freelist;
2244 stat(s, ALLOC_SLOWPATH);
2246 freelist = get_freelist(s, page);
2250 stat(s, DEACTIVATE_BYPASS);
2254 stat(s, ALLOC_REFILL);
2258 * freelist is pointing to the list of objects to be used.
2259 * page is pointing to the page from which the objects are obtained.
2260 * That page must be frozen for per cpu allocations to work.
2262 VM_BUG_ON(!c->page->frozen);
2263 c->freelist = get_freepointer(s, freelist);
2264 c->tid = next_tid(c->tid);
2265 local_irq_restore(flags);
2271 page = c->page = c->partial;
2272 c->partial = page->next;
2273 stat(s, CPU_PARTIAL_ALLOC);
2278 freelist = new_slab_objects(s, gfpflags, node, &c);
2280 if (unlikely(!freelist)) {
2281 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2282 slab_out_of_memory(s, gfpflags, node);
2284 local_irq_restore(flags);
2289 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2292 /* Only entered in the debug case */
2293 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2294 goto new_slab; /* Slab failed checks. Next slab needed */
2296 deactivate_slab(s, page, get_freepointer(s, freelist));
2299 local_irq_restore(flags);
2304 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2305 * have the fastpath folded into their functions. So no function call
2306 * overhead for requests that can be satisfied on the fastpath.
2308 * The fastpath works by first checking if the lockless freelist can be used.
2309 * If not then __slab_alloc is called for slow processing.
2311 * Otherwise we can simply pick the next object from the lockless free list.
2313 static __always_inline void *slab_alloc(struct kmem_cache *s,
2314 gfp_t gfpflags, int node, unsigned long addr)
2317 struct kmem_cache_cpu *c;
2321 if (slab_pre_alloc_hook(s, gfpflags))
2327 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2328 * enabled. We may switch back and forth between cpus while
2329 * reading from one cpu area. That does not matter as long
2330 * as we end up on the original cpu again when doing the cmpxchg.
2332 c = __this_cpu_ptr(s->cpu_slab);
2335 * The transaction ids are globally unique per cpu and per operation on
2336 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2337 * occurs on the right processor and that there was no operation on the
2338 * linked list in between.
2343 object = c->freelist;
2345 if (unlikely(!object || !node_match(page, node)))
2346 object = __slab_alloc(s, gfpflags, node, addr, c);
2349 void *next_object = get_freepointer_safe(s, object);
2352 * The cmpxchg will only match if there was no additional
2353 * operation and if we are on the right processor.
2355 * The cmpxchg does the following atomically (without lock semantics!)
2356 * 1. Relocate first pointer to the current per cpu area.
2357 * 2. Verify that tid and freelist have not been changed
2358 * 3. If they were not changed replace tid and freelist
2360 * Since this is without lock semantics the protection is only against
2361 * code executing on this cpu *not* from access by other cpus.
2363 if (unlikely(!this_cpu_cmpxchg_double(
2364 s->cpu_slab->freelist, s->cpu_slab->tid,
2366 next_object, next_tid(tid)))) {
2368 note_cmpxchg_failure("slab_alloc", s, tid);
2371 prefetch_freepointer(s, next_object);
2372 stat(s, ALLOC_FASTPATH);
2375 if (unlikely(gfpflags & __GFP_ZERO) && object)
2376 memset(object, 0, s->object_size);
2378 slab_post_alloc_hook(s, gfpflags, object);
2383 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2385 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2387 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2391 EXPORT_SYMBOL(kmem_cache_alloc);
2393 #ifdef CONFIG_TRACING
2394 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2396 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2397 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2400 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2402 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2404 void *ret = kmalloc_order(size, flags, order);
2405 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2408 EXPORT_SYMBOL(kmalloc_order_trace);
2412 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2414 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2416 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2417 s->object_size, s->size, gfpflags, node);
2421 EXPORT_SYMBOL(kmem_cache_alloc_node);
2423 #ifdef CONFIG_TRACING
2424 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2426 int node, size_t size)
2428 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2430 trace_kmalloc_node(_RET_IP_, ret,
2431 size, s->size, gfpflags, node);
2434 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2439 * Slow patch handling. This may still be called frequently since objects
2440 * have a longer lifetime than the cpu slabs in most processing loads.
2442 * So we still attempt to reduce cache line usage. Just take the slab
2443 * lock and free the item. If there is no additional partial page
2444 * handling required then we can return immediately.
2446 static void __slab_free(struct kmem_cache *s, struct page *page,
2447 void *x, unsigned long addr)
2450 void **object = (void *)x;
2454 unsigned long counters;
2455 struct kmem_cache_node *n = NULL;
2456 unsigned long uninitialized_var(flags);
2458 stat(s, FREE_SLOWPATH);
2460 if (kmem_cache_debug(s) &&
2461 !(n = free_debug_processing(s, page, x, addr, &flags)))
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 if (kmem_cache_debug(s) && page->slab != s) {
2617 pr_err("kmem_cache_free: Wrong slab cache. %s but object"
2618 " is from %s\n", page->slab->name, s->name);
2623 slab_free(s, page, x, _RET_IP_);
2625 trace_kmem_cache_free(_RET_IP_, x);
2627 EXPORT_SYMBOL(kmem_cache_free);
2630 * Object placement in a slab is made very easy because we always start at
2631 * offset 0. If we tune the size of the object to the alignment then we can
2632 * get the required alignment by putting one properly sized object after
2635 * Notice that the allocation order determines the sizes of the per cpu
2636 * caches. Each processor has always one slab available for allocations.
2637 * Increasing the allocation order reduces the number of times that slabs
2638 * must be moved on and off the partial lists and is therefore a factor in
2643 * Mininum / Maximum order of slab pages. This influences locking overhead
2644 * and slab fragmentation. A higher order reduces the number of partial slabs
2645 * and increases the number of allocations possible without having to
2646 * take the list_lock.
2648 static int slub_min_order;
2649 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2650 static int slub_min_objects;
2653 * Merge control. If this is set then no merging of slab caches will occur.
2654 * (Could be removed. This was introduced to pacify the merge skeptics.)
2656 static int slub_nomerge;
2659 * Calculate the order of allocation given an slab object size.
2661 * The order of allocation has significant impact on performance and other
2662 * system components. Generally order 0 allocations should be preferred since
2663 * order 0 does not cause fragmentation in the page allocator. Larger objects
2664 * be problematic to put into order 0 slabs because there may be too much
2665 * unused space left. We go to a higher order if more than 1/16th of the slab
2668 * In order to reach satisfactory performance we must ensure that a minimum
2669 * number of objects is in one slab. Otherwise we may generate too much
2670 * activity on the partial lists which requires taking the list_lock. This is
2671 * less a concern for large slabs though which are rarely used.
2673 * slub_max_order specifies the order where we begin to stop considering the
2674 * number of objects in a slab as critical. If we reach slub_max_order then
2675 * we try to keep the page order as low as possible. So we accept more waste
2676 * of space in favor of a small page order.
2678 * Higher order allocations also allow the placement of more objects in a
2679 * slab and thereby reduce object handling overhead. If the user has
2680 * requested a higher mininum order then we start with that one instead of
2681 * the smallest order which will fit the object.
2683 static inline int slab_order(int size, int min_objects,
2684 int max_order, int fract_leftover, int reserved)
2688 int min_order = slub_min_order;
2690 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2691 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2693 for (order = max(min_order,
2694 fls(min_objects * size - 1) - PAGE_SHIFT);
2695 order <= max_order; order++) {
2697 unsigned long slab_size = PAGE_SIZE << order;
2699 if (slab_size < min_objects * size + reserved)
2702 rem = (slab_size - reserved) % size;
2704 if (rem <= slab_size / fract_leftover)
2712 static inline int calculate_order(int size, int reserved)
2720 * Attempt to find best configuration for a slab. This
2721 * works by first attempting to generate a layout with
2722 * the best configuration and backing off gradually.
2724 * First we reduce the acceptable waste in a slab. Then
2725 * we reduce the minimum objects required in a slab.
2727 min_objects = slub_min_objects;
2729 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2730 max_objects = order_objects(slub_max_order, size, reserved);
2731 min_objects = min(min_objects, max_objects);
2733 while (min_objects > 1) {
2735 while (fraction >= 4) {
2736 order = slab_order(size, min_objects,
2737 slub_max_order, fraction, reserved);
2738 if (order <= slub_max_order)
2746 * We were unable to place multiple objects in a slab. Now
2747 * lets see if we can place a single object there.
2749 order = slab_order(size, 1, slub_max_order, 1, reserved);
2750 if (order <= slub_max_order)
2754 * Doh this slab cannot be placed using slub_max_order.
2756 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2757 if (order < MAX_ORDER)
2763 * Figure out what the alignment of the objects will be.
2765 static unsigned long calculate_alignment(unsigned long flags,
2766 unsigned long align, unsigned long size)
2769 * If the user wants hardware cache aligned objects then follow that
2770 * suggestion if the object is sufficiently large.
2772 * The hardware cache alignment cannot override the specified
2773 * alignment though. If that is greater then use it.
2775 if (flags & SLAB_HWCACHE_ALIGN) {
2776 unsigned long ralign = cache_line_size();
2777 while (size <= ralign / 2)
2779 align = max(align, ralign);
2782 if (align < ARCH_SLAB_MINALIGN)
2783 align = ARCH_SLAB_MINALIGN;
2785 return ALIGN(align, sizeof(void *));
2789 init_kmem_cache_node(struct kmem_cache_node *n)
2792 spin_lock_init(&n->list_lock);
2793 INIT_LIST_HEAD(&n->partial);
2794 #ifdef CONFIG_SLUB_DEBUG
2795 atomic_long_set(&n->nr_slabs, 0);
2796 atomic_long_set(&n->total_objects, 0);
2797 INIT_LIST_HEAD(&n->full);
2801 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2803 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2804 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2807 * Must align to double word boundary for the double cmpxchg
2808 * instructions to work; see __pcpu_double_call_return_bool().
2810 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2811 2 * sizeof(void *));
2816 init_kmem_cache_cpus(s);
2821 static struct kmem_cache *kmem_cache_node;
2824 * No kmalloc_node yet so do it by hand. We know that this is the first
2825 * slab on the node for this slabcache. There are no concurrent accesses
2828 * Note that this function only works on the kmalloc_node_cache
2829 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2830 * memory on a fresh node that has no slab structures yet.
2832 static void early_kmem_cache_node_alloc(int node)
2835 struct kmem_cache_node *n;
2837 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2839 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2842 if (page_to_nid(page) != node) {
2843 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2845 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2846 "in order to be able to continue\n");
2851 page->freelist = get_freepointer(kmem_cache_node, n);
2854 kmem_cache_node->node[node] = n;
2855 #ifdef CONFIG_SLUB_DEBUG
2856 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2857 init_tracking(kmem_cache_node, n);
2859 init_kmem_cache_node(n);
2860 inc_slabs_node(kmem_cache_node, node, page->objects);
2862 add_partial(n, page, DEACTIVATE_TO_HEAD);
2865 static void free_kmem_cache_nodes(struct kmem_cache *s)
2869 for_each_node_state(node, N_NORMAL_MEMORY) {
2870 struct kmem_cache_node *n = s->node[node];
2873 kmem_cache_free(kmem_cache_node, n);
2875 s->node[node] = NULL;
2879 static int init_kmem_cache_nodes(struct kmem_cache *s)
2883 for_each_node_state(node, N_NORMAL_MEMORY) {
2884 struct kmem_cache_node *n;
2886 if (slab_state == DOWN) {
2887 early_kmem_cache_node_alloc(node);
2890 n = kmem_cache_alloc_node(kmem_cache_node,
2894 free_kmem_cache_nodes(s);
2899 init_kmem_cache_node(n);
2904 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2906 if (min < MIN_PARTIAL)
2908 else if (min > MAX_PARTIAL)
2910 s->min_partial = min;
2914 * calculate_sizes() determines the order and the distribution of data within
2917 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2919 unsigned long flags = s->flags;
2920 unsigned long size = s->object_size;
2921 unsigned long align = s->align;
2925 * Round up object size to the next word boundary. We can only
2926 * place the free pointer at word boundaries and this determines
2927 * the possible location of the free pointer.
2929 size = ALIGN(size, sizeof(void *));
2931 #ifdef CONFIG_SLUB_DEBUG
2933 * Determine if we can poison the object itself. If the user of
2934 * the slab may touch the object after free or before allocation
2935 * then we should never poison the object itself.
2937 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2939 s->flags |= __OBJECT_POISON;
2941 s->flags &= ~__OBJECT_POISON;
2945 * If we are Redzoning then check if there is some space between the
2946 * end of the object and the free pointer. If not then add an
2947 * additional word to have some bytes to store Redzone information.
2949 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2950 size += sizeof(void *);
2954 * With that we have determined the number of bytes in actual use
2955 * by the object. This is the potential offset to the free pointer.
2959 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2962 * Relocate free pointer after the object if it is not
2963 * permitted to overwrite the first word of the object on
2966 * This is the case if we do RCU, have a constructor or
2967 * destructor or are poisoning the objects.
2970 size += sizeof(void *);
2973 #ifdef CONFIG_SLUB_DEBUG
2974 if (flags & SLAB_STORE_USER)
2976 * Need to store information about allocs and frees after
2979 size += 2 * sizeof(struct track);
2981 if (flags & SLAB_RED_ZONE)
2983 * Add some empty padding so that we can catch
2984 * overwrites from earlier objects rather than let
2985 * tracking information or the free pointer be
2986 * corrupted if a user writes before the start
2989 size += sizeof(void *);
2993 * Determine the alignment based on various parameters that the
2994 * user specified and the dynamic determination of cache line size
2997 align = calculate_alignment(flags, align, s->object_size);
3001 * SLUB stores one object immediately after another beginning from
3002 * offset 0. In order to align the objects we have to simply size
3003 * each object to conform to the alignment.
3005 size = ALIGN(size, align);
3007 if (forced_order >= 0)
3008 order = forced_order;
3010 order = calculate_order(size, s->reserved);
3017 s->allocflags |= __GFP_COMP;
3019 if (s->flags & SLAB_CACHE_DMA)
3020 s->allocflags |= SLUB_DMA;
3022 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3023 s->allocflags |= __GFP_RECLAIMABLE;
3026 * Determine the number of objects per slab
3028 s->oo = oo_make(order, size, s->reserved);
3029 s->min = oo_make(get_order(size), size, s->reserved);
3030 if (oo_objects(s->oo) > oo_objects(s->max))
3033 return !!oo_objects(s->oo);
3037 static int kmem_cache_open(struct kmem_cache *s,
3038 const char *name, size_t size,
3039 size_t align, unsigned long flags,
3040 void (*ctor)(void *))
3042 memset(s, 0, kmem_size);
3045 s->object_size = size;
3047 s->flags = kmem_cache_flags(size, flags, name, ctor);
3050 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3051 s->reserved = sizeof(struct rcu_head);
3053 if (!calculate_sizes(s, -1))
3055 if (disable_higher_order_debug) {
3057 * Disable debugging flags that store metadata if the min slab
3060 if (get_order(s->size) > get_order(s->object_size)) {
3061 s->flags &= ~DEBUG_METADATA_FLAGS;
3063 if (!calculate_sizes(s, -1))
3068 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3069 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3070 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3071 /* Enable fast mode */
3072 s->flags |= __CMPXCHG_DOUBLE;
3076 * The larger the object size is, the more pages we want on the partial
3077 * list to avoid pounding the page allocator excessively.
3079 set_min_partial(s, ilog2(s->size) / 2);
3082 * cpu_partial determined the maximum number of objects kept in the
3083 * per cpu partial lists of a processor.
3085 * Per cpu partial lists mainly contain slabs that just have one
3086 * object freed. If they are used for allocation then they can be
3087 * filled up again with minimal effort. The slab will never hit the
3088 * per node partial lists and therefore no locking will be required.
3090 * This setting also determines
3092 * A) The number of objects from per cpu partial slabs dumped to the
3093 * per node list when we reach the limit.
3094 * B) The number of objects in cpu partial slabs to extract from the
3095 * per node list when we run out of per cpu objects. We only fetch 50%
3096 * to keep some capacity around for frees.
3098 if (kmem_cache_debug(s))
3100 else if (s->size >= PAGE_SIZE)
3102 else if (s->size >= 1024)
3104 else if (s->size >= 256)
3105 s->cpu_partial = 13;
3107 s->cpu_partial = 30;
3111 s->remote_node_defrag_ratio = 1000;
3113 if (!init_kmem_cache_nodes(s))
3116 if (alloc_kmem_cache_cpus(s))
3119 free_kmem_cache_nodes(s);
3121 if (flags & SLAB_PANIC)
3122 panic("Cannot create slab %s size=%lu realsize=%u "
3123 "order=%u offset=%u flags=%lx\n",
3124 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3130 * Determine the size of a slab object
3132 unsigned int kmem_cache_size(struct kmem_cache *s)
3134 return s->object_size;
3136 EXPORT_SYMBOL(kmem_cache_size);
3138 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3141 #ifdef CONFIG_SLUB_DEBUG
3142 void *addr = page_address(page);
3144 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3145 sizeof(long), GFP_ATOMIC);
3148 slab_err(s, page, text, s->name);
3151 get_map(s, page, map);
3152 for_each_object(p, s, addr, page->objects) {
3154 if (!test_bit(slab_index(p, s, addr), map)) {
3155 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3157 print_tracking(s, p);
3166 * Attempt to free all partial slabs on a node.
3167 * This is called from kmem_cache_close(). We must be the last thread
3168 * using the cache and therefore we do not need to lock anymore.
3170 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3172 struct page *page, *h;
3174 list_for_each_entry_safe(page, h, &n->partial, lru) {
3176 remove_partial(n, page);
3177 discard_slab(s, page);
3179 list_slab_objects(s, page,
3180 "Objects remaining in %s on kmem_cache_close()");
3186 * Release all resources used by a slab cache.
3188 static inline int kmem_cache_close(struct kmem_cache *s)
3193 /* Attempt to free all objects */
3194 for_each_node_state(node, N_NORMAL_MEMORY) {
3195 struct kmem_cache_node *n = get_node(s, node);
3198 if (n->nr_partial || slabs_node(s, node))
3201 free_percpu(s->cpu_slab);
3202 free_kmem_cache_nodes(s);
3206 int __kmem_cache_shutdown(struct kmem_cache *s)
3208 int rc = kmem_cache_close(s);
3211 sysfs_slab_remove(s);
3216 /********************************************************************
3218 *******************************************************************/
3220 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3221 EXPORT_SYMBOL(kmalloc_caches);
3223 #ifdef CONFIG_ZONE_DMA
3224 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3227 static int __init setup_slub_min_order(char *str)
3229 get_option(&str, &slub_min_order);
3234 __setup("slub_min_order=", setup_slub_min_order);
3236 static int __init setup_slub_max_order(char *str)
3238 get_option(&str, &slub_max_order);
3239 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3244 __setup("slub_max_order=", setup_slub_max_order);
3246 static int __init setup_slub_min_objects(char *str)
3248 get_option(&str, &slub_min_objects);
3253 __setup("slub_min_objects=", setup_slub_min_objects);
3255 static int __init setup_slub_nomerge(char *str)
3261 __setup("slub_nomerge", setup_slub_nomerge);
3263 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3264 int size, unsigned int flags)
3266 struct kmem_cache *s;
3268 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3271 * This function is called with IRQs disabled during early-boot on
3272 * single CPU so there's no need to take slab_mutex here.
3274 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3278 list_add(&s->list, &slab_caches);
3282 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3287 * Conversion table for small slabs sizes / 8 to the index in the
3288 * kmalloc array. This is necessary for slabs < 192 since we have non power
3289 * of two cache sizes there. The size of larger slabs can be determined using
3292 static s8 size_index[24] = {
3319 static inline int size_index_elem(size_t bytes)
3321 return (bytes - 1) / 8;
3324 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3330 return ZERO_SIZE_PTR;
3332 index = size_index[size_index_elem(size)];
3334 index = fls(size - 1);
3336 #ifdef CONFIG_ZONE_DMA
3337 if (unlikely((flags & SLUB_DMA)))
3338 return kmalloc_dma_caches[index];
3341 return kmalloc_caches[index];
3344 void *__kmalloc(size_t size, gfp_t flags)
3346 struct kmem_cache *s;
3349 if (unlikely(size > SLUB_MAX_SIZE))
3350 return kmalloc_large(size, flags);
3352 s = get_slab(size, flags);
3354 if (unlikely(ZERO_OR_NULL_PTR(s)))
3357 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3359 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3363 EXPORT_SYMBOL(__kmalloc);
3366 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3371 flags |= __GFP_COMP | __GFP_NOTRACK;
3372 page = alloc_pages_node(node, flags, get_order(size));
3374 ptr = page_address(page);
3376 kmemleak_alloc(ptr, size, 1, flags);
3380 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3382 struct kmem_cache *s;
3385 if (unlikely(size > SLUB_MAX_SIZE)) {
3386 ret = kmalloc_large_node(size, flags, node);
3388 trace_kmalloc_node(_RET_IP_, ret,
3389 size, PAGE_SIZE << get_order(size),
3395 s = get_slab(size, flags);
3397 if (unlikely(ZERO_OR_NULL_PTR(s)))
3400 ret = slab_alloc(s, flags, node, _RET_IP_);
3402 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3406 EXPORT_SYMBOL(__kmalloc_node);
3409 size_t ksize(const void *object)
3413 if (unlikely(object == ZERO_SIZE_PTR))
3416 page = virt_to_head_page(object);
3418 if (unlikely(!PageSlab(page))) {
3419 WARN_ON(!PageCompound(page));
3420 return PAGE_SIZE << compound_order(page);
3423 return slab_ksize(page->slab);
3425 EXPORT_SYMBOL(ksize);
3427 #ifdef CONFIG_SLUB_DEBUG
3428 bool verify_mem_not_deleted(const void *x)
3431 void *object = (void *)x;
3432 unsigned long flags;
3435 if (unlikely(ZERO_OR_NULL_PTR(x)))
3438 local_irq_save(flags);
3440 page = virt_to_head_page(x);
3441 if (unlikely(!PageSlab(page))) {
3442 /* maybe it was from stack? */
3448 if (on_freelist(page->slab, page, object)) {
3449 object_err(page->slab, page, object, "Object is on free-list");
3457 local_irq_restore(flags);
3460 EXPORT_SYMBOL(verify_mem_not_deleted);
3463 void kfree(const void *x)
3466 void *object = (void *)x;
3468 trace_kfree(_RET_IP_, x);
3470 if (unlikely(ZERO_OR_NULL_PTR(x)))
3473 page = virt_to_head_page(x);
3474 if (unlikely(!PageSlab(page))) {
3475 BUG_ON(!PageCompound(page));
3477 __free_pages(page, compound_order(page));
3480 slab_free(page->slab, page, object, _RET_IP_);
3482 EXPORT_SYMBOL(kfree);
3485 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3486 * the remaining slabs by the number of items in use. The slabs with the
3487 * most items in use come first. New allocations will then fill those up
3488 * and thus they can be removed from the partial lists.
3490 * The slabs with the least items are placed last. This results in them
3491 * being allocated from last increasing the chance that the last objects
3492 * are freed in them.
3494 int kmem_cache_shrink(struct kmem_cache *s)
3498 struct kmem_cache_node *n;
3501 int objects = oo_objects(s->max);
3502 struct list_head *slabs_by_inuse =
3503 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3504 unsigned long flags;
3506 if (!slabs_by_inuse)
3510 for_each_node_state(node, N_NORMAL_MEMORY) {
3511 n = get_node(s, node);
3516 for (i = 0; i < objects; i++)
3517 INIT_LIST_HEAD(slabs_by_inuse + i);
3519 spin_lock_irqsave(&n->list_lock, flags);
3522 * Build lists indexed by the items in use in each slab.
3524 * Note that concurrent frees may occur while we hold the
3525 * list_lock. page->inuse here is the upper limit.
3527 list_for_each_entry_safe(page, t, &n->partial, lru) {
3528 list_move(&page->lru, slabs_by_inuse + page->inuse);
3534 * Rebuild the partial list with the slabs filled up most
3535 * first and the least used slabs at the end.
3537 for (i = objects - 1; i > 0; i--)
3538 list_splice(slabs_by_inuse + i, n->partial.prev);
3540 spin_unlock_irqrestore(&n->list_lock, flags);
3542 /* Release empty slabs */
3543 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3544 discard_slab(s, page);
3547 kfree(slabs_by_inuse);
3550 EXPORT_SYMBOL(kmem_cache_shrink);
3552 #if defined(CONFIG_MEMORY_HOTPLUG)
3553 static int slab_mem_going_offline_callback(void *arg)
3555 struct kmem_cache *s;
3557 mutex_lock(&slab_mutex);
3558 list_for_each_entry(s, &slab_caches, list)
3559 kmem_cache_shrink(s);
3560 mutex_unlock(&slab_mutex);
3565 static void slab_mem_offline_callback(void *arg)
3567 struct kmem_cache_node *n;
3568 struct kmem_cache *s;
3569 struct memory_notify *marg = arg;
3572 offline_node = marg->status_change_nid;
3575 * If the node still has available memory. we need kmem_cache_node
3578 if (offline_node < 0)
3581 mutex_lock(&slab_mutex);
3582 list_for_each_entry(s, &slab_caches, list) {
3583 n = get_node(s, offline_node);
3586 * if n->nr_slabs > 0, slabs still exist on the node
3587 * that is going down. We were unable to free them,
3588 * and offline_pages() function shouldn't call this
3589 * callback. So, we must fail.
3591 BUG_ON(slabs_node(s, offline_node));
3593 s->node[offline_node] = NULL;
3594 kmem_cache_free(kmem_cache_node, n);
3597 mutex_unlock(&slab_mutex);
3600 static int slab_mem_going_online_callback(void *arg)
3602 struct kmem_cache_node *n;
3603 struct kmem_cache *s;
3604 struct memory_notify *marg = arg;
3605 int nid = marg->status_change_nid;
3609 * If the node's memory is already available, then kmem_cache_node is
3610 * already created. Nothing to do.
3616 * We are bringing a node online. No memory is available yet. We must
3617 * allocate a kmem_cache_node structure in order to bring the node
3620 mutex_lock(&slab_mutex);
3621 list_for_each_entry(s, &slab_caches, list) {
3623 * XXX: kmem_cache_alloc_node will fallback to other nodes
3624 * since memory is not yet available from the node that
3627 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3632 init_kmem_cache_node(n);
3636 mutex_unlock(&slab_mutex);
3640 static int slab_memory_callback(struct notifier_block *self,
3641 unsigned long action, void *arg)
3646 case MEM_GOING_ONLINE:
3647 ret = slab_mem_going_online_callback(arg);
3649 case MEM_GOING_OFFLINE:
3650 ret = slab_mem_going_offline_callback(arg);
3653 case MEM_CANCEL_ONLINE:
3654 slab_mem_offline_callback(arg);
3657 case MEM_CANCEL_OFFLINE:
3661 ret = notifier_from_errno(ret);
3667 #endif /* CONFIG_MEMORY_HOTPLUG */
3669 /********************************************************************
3670 * Basic setup of slabs
3671 *******************************************************************/
3674 * Used for early kmem_cache structures that were allocated using
3675 * the page allocator
3678 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3682 list_add(&s->list, &slab_caches);
3685 for_each_node_state(node, N_NORMAL_MEMORY) {
3686 struct kmem_cache_node *n = get_node(s, node);
3690 list_for_each_entry(p, &n->partial, lru)
3693 #ifdef CONFIG_SLUB_DEBUG
3694 list_for_each_entry(p, &n->full, lru)
3701 void __init kmem_cache_init(void)
3705 struct kmem_cache *temp_kmem_cache;
3707 struct kmem_cache *temp_kmem_cache_node;
3708 unsigned long kmalloc_size;
3710 if (debug_guardpage_minorder())
3713 kmem_size = offsetof(struct kmem_cache, node) +
3714 nr_node_ids * sizeof(struct kmem_cache_node *);
3716 /* Allocate two kmem_caches from the page allocator */
3717 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3718 order = get_order(2 * kmalloc_size);
3719 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3722 * Must first have the slab cache available for the allocations of the
3723 * struct kmem_cache_node's. There is special bootstrap code in
3724 * kmem_cache_open for slab_state == DOWN.
3726 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3728 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3729 sizeof(struct kmem_cache_node),
3730 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3732 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3734 /* Able to allocate the per node structures */
3735 slab_state = PARTIAL;
3737 temp_kmem_cache = kmem_cache;
3738 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3739 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3740 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3741 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3744 * Allocate kmem_cache_node properly from the kmem_cache slab.
3745 * kmem_cache_node is separately allocated so no need to
3746 * update any list pointers.
3748 temp_kmem_cache_node = kmem_cache_node;
3750 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3751 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3753 kmem_cache_bootstrap_fixup(kmem_cache_node);
3756 kmem_cache_bootstrap_fixup(kmem_cache);
3758 /* Free temporary boot structure */
3759 free_pages((unsigned long)temp_kmem_cache, order);
3761 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3764 * Patch up the size_index table if we have strange large alignment
3765 * requirements for the kmalloc array. This is only the case for
3766 * MIPS it seems. The standard arches will not generate any code here.
3768 * Largest permitted alignment is 256 bytes due to the way we
3769 * handle the index determination for the smaller caches.
3771 * Make sure that nothing crazy happens if someone starts tinkering
3772 * around with ARCH_KMALLOC_MINALIGN
3774 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3775 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3777 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3778 int elem = size_index_elem(i);
3779 if (elem >= ARRAY_SIZE(size_index))
3781 size_index[elem] = KMALLOC_SHIFT_LOW;
3784 if (KMALLOC_MIN_SIZE == 64) {
3786 * The 96 byte size cache is not used if the alignment
3789 for (i = 64 + 8; i <= 96; i += 8)
3790 size_index[size_index_elem(i)] = 7;
3791 } else if (KMALLOC_MIN_SIZE == 128) {
3793 * The 192 byte sized cache is not used if the alignment
3794 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3797 for (i = 128 + 8; i <= 192; i += 8)
3798 size_index[size_index_elem(i)] = 8;
3801 /* Caches that are not of the two-to-the-power-of size */
3802 if (KMALLOC_MIN_SIZE <= 32) {
3803 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3807 if (KMALLOC_MIN_SIZE <= 64) {
3808 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3812 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3813 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3819 /* Provide the correct kmalloc names now that the caches are up */
3820 if (KMALLOC_MIN_SIZE <= 32) {
3821 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3822 BUG_ON(!kmalloc_caches[1]->name);
3825 if (KMALLOC_MIN_SIZE <= 64) {
3826 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3827 BUG_ON(!kmalloc_caches[2]->name);
3830 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3831 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3834 kmalloc_caches[i]->name = s;
3838 register_cpu_notifier(&slab_notifier);
3841 #ifdef CONFIG_ZONE_DMA
3842 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3843 struct kmem_cache *s = kmalloc_caches[i];
3846 char *name = kasprintf(GFP_NOWAIT,
3847 "dma-kmalloc-%d", s->object_size);
3850 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3851 s->object_size, SLAB_CACHE_DMA);
3856 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3857 " CPUs=%d, Nodes=%d\n",
3858 caches, cache_line_size(),
3859 slub_min_order, slub_max_order, slub_min_objects,
3860 nr_cpu_ids, nr_node_ids);
3863 void __init kmem_cache_init_late(void)
3868 * Find a mergeable slab cache
3870 static int slab_unmergeable(struct kmem_cache *s)
3872 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3879 * We may have set a slab to be unmergeable during bootstrap.
3881 if (s->refcount < 0)
3887 static struct kmem_cache *find_mergeable(size_t size,
3888 size_t align, unsigned long flags, const char *name,
3889 void (*ctor)(void *))
3891 struct kmem_cache *s;
3893 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3899 size = ALIGN(size, sizeof(void *));
3900 align = calculate_alignment(flags, align, size);
3901 size = ALIGN(size, align);
3902 flags = kmem_cache_flags(size, flags, name, NULL);
3904 list_for_each_entry(s, &slab_caches, list) {
3905 if (slab_unmergeable(s))
3911 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3914 * Check if alignment is compatible.
3915 * Courtesy of Adrian Drzewiecki
3917 if ((s->size & ~(align - 1)) != s->size)
3920 if (s->size - size >= sizeof(void *))
3928 struct kmem_cache *__kmem_cache_create(const char *name, size_t size,
3929 size_t align, unsigned long flags, void (*ctor)(void *))
3931 struct kmem_cache *s;
3934 s = find_mergeable(size, align, flags, name, ctor);
3938 * Adjust the object sizes so that we clear
3939 * the complete object on kzalloc.
3941 s->object_size = max(s->object_size, (int)size);
3942 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3944 if (sysfs_slab_alias(s, name)) {
3951 n = kstrdup(name, GFP_KERNEL);
3955 s = kmem_cache_alloc(kmem_cache, GFP_KERNEL);
3957 if (kmem_cache_open(s, n,
3958 size, align, flags, ctor)) {
3961 mutex_unlock(&slab_mutex);
3962 r = sysfs_slab_add(s);
3963 mutex_lock(&slab_mutex);
3968 kmem_cache_close(s);
3970 kmem_cache_free(kmem_cache, s);
3978 * Use the cpu notifier to insure that the cpu slabs are flushed when
3981 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3982 unsigned long action, void *hcpu)
3984 long cpu = (long)hcpu;
3985 struct kmem_cache *s;
3986 unsigned long flags;
3989 case CPU_UP_CANCELED:
3990 case CPU_UP_CANCELED_FROZEN:
3992 case CPU_DEAD_FROZEN:
3993 mutex_lock(&slab_mutex);
3994 list_for_each_entry(s, &slab_caches, list) {
3995 local_irq_save(flags);
3996 __flush_cpu_slab(s, cpu);
3997 local_irq_restore(flags);
3999 mutex_unlock(&slab_mutex);
4007 static struct notifier_block __cpuinitdata slab_notifier = {
4008 .notifier_call = slab_cpuup_callback
4013 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4015 struct kmem_cache *s;
4018 if (unlikely(size > SLUB_MAX_SIZE))
4019 return kmalloc_large(size, gfpflags);
4021 s = get_slab(size, gfpflags);
4023 if (unlikely(ZERO_OR_NULL_PTR(s)))
4026 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4028 /* Honor the call site pointer we received. */
4029 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4035 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4036 int node, unsigned long caller)
4038 struct kmem_cache *s;
4041 if (unlikely(size > SLUB_MAX_SIZE)) {
4042 ret = kmalloc_large_node(size, gfpflags, node);
4044 trace_kmalloc_node(caller, ret,
4045 size, PAGE_SIZE << get_order(size),
4051 s = get_slab(size, gfpflags);
4053 if (unlikely(ZERO_OR_NULL_PTR(s)))
4056 ret = slab_alloc(s, gfpflags, node, caller);
4058 /* Honor the call site pointer we received. */
4059 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4066 static int count_inuse(struct page *page)
4071 static int count_total(struct page *page)
4073 return page->objects;
4077 #ifdef CONFIG_SLUB_DEBUG
4078 static int validate_slab(struct kmem_cache *s, struct page *page,
4082 void *addr = page_address(page);
4084 if (!check_slab(s, page) ||
4085 !on_freelist(s, page, NULL))
4088 /* Now we know that a valid freelist exists */
4089 bitmap_zero(map, page->objects);
4091 get_map(s, page, map);
4092 for_each_object(p, s, addr, page->objects) {
4093 if (test_bit(slab_index(p, s, addr), map))
4094 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4098 for_each_object(p, s, addr, page->objects)
4099 if (!test_bit(slab_index(p, s, addr), map))
4100 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4105 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4109 validate_slab(s, page, map);
4113 static int validate_slab_node(struct kmem_cache *s,
4114 struct kmem_cache_node *n, unsigned long *map)
4116 unsigned long count = 0;
4118 unsigned long flags;
4120 spin_lock_irqsave(&n->list_lock, flags);
4122 list_for_each_entry(page, &n->partial, lru) {
4123 validate_slab_slab(s, page, map);
4126 if (count != n->nr_partial)
4127 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4128 "counter=%ld\n", s->name, count, n->nr_partial);
4130 if (!(s->flags & SLAB_STORE_USER))
4133 list_for_each_entry(page, &n->full, lru) {
4134 validate_slab_slab(s, page, map);
4137 if (count != atomic_long_read(&n->nr_slabs))
4138 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4139 "counter=%ld\n", s->name, count,
4140 atomic_long_read(&n->nr_slabs));
4143 spin_unlock_irqrestore(&n->list_lock, flags);
4147 static long validate_slab_cache(struct kmem_cache *s)
4150 unsigned long count = 0;
4151 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4152 sizeof(unsigned long), GFP_KERNEL);
4158 for_each_node_state(node, N_NORMAL_MEMORY) {
4159 struct kmem_cache_node *n = get_node(s, node);
4161 count += validate_slab_node(s, n, map);
4167 * Generate lists of code addresses where slabcache objects are allocated
4172 unsigned long count;
4179 DECLARE_BITMAP(cpus, NR_CPUS);
4185 unsigned long count;
4186 struct location *loc;
4189 static void free_loc_track(struct loc_track *t)
4192 free_pages((unsigned long)t->loc,
4193 get_order(sizeof(struct location) * t->max));
4196 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4201 order = get_order(sizeof(struct location) * max);
4203 l = (void *)__get_free_pages(flags, order);
4208 memcpy(l, t->loc, sizeof(struct location) * t->count);
4216 static int add_location(struct loc_track *t, struct kmem_cache *s,
4217 const struct track *track)
4219 long start, end, pos;
4221 unsigned long caddr;
4222 unsigned long age = jiffies - track->when;
4228 pos = start + (end - start + 1) / 2;
4231 * There is nothing at "end". If we end up there
4232 * we need to add something to before end.
4237 caddr = t->loc[pos].addr;
4238 if (track->addr == caddr) {
4244 if (age < l->min_time)
4246 if (age > l->max_time)
4249 if (track->pid < l->min_pid)
4250 l->min_pid = track->pid;
4251 if (track->pid > l->max_pid)
4252 l->max_pid = track->pid;
4254 cpumask_set_cpu(track->cpu,
4255 to_cpumask(l->cpus));
4257 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4261 if (track->addr < caddr)
4268 * Not found. Insert new tracking element.
4270 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4276 (t->count - pos) * sizeof(struct location));
4279 l->addr = track->addr;
4283 l->min_pid = track->pid;
4284 l->max_pid = track->pid;
4285 cpumask_clear(to_cpumask(l->cpus));
4286 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4287 nodes_clear(l->nodes);
4288 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4292 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4293 struct page *page, enum track_item alloc,
4296 void *addr = page_address(page);
4299 bitmap_zero(map, page->objects);
4300 get_map(s, page, map);
4302 for_each_object(p, s, addr, page->objects)
4303 if (!test_bit(slab_index(p, s, addr), map))
4304 add_location(t, s, get_track(s, p, alloc));
4307 static int list_locations(struct kmem_cache *s, char *buf,
4308 enum track_item alloc)
4312 struct loc_track t = { 0, 0, NULL };
4314 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4315 sizeof(unsigned long), GFP_KERNEL);
4317 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4320 return sprintf(buf, "Out of memory\n");
4322 /* Push back cpu slabs */
4325 for_each_node_state(node, N_NORMAL_MEMORY) {
4326 struct kmem_cache_node *n = get_node(s, node);
4327 unsigned long flags;
4330 if (!atomic_long_read(&n->nr_slabs))
4333 spin_lock_irqsave(&n->list_lock, flags);
4334 list_for_each_entry(page, &n->partial, lru)
4335 process_slab(&t, s, page, alloc, map);
4336 list_for_each_entry(page, &n->full, lru)
4337 process_slab(&t, s, page, alloc, map);
4338 spin_unlock_irqrestore(&n->list_lock, flags);
4341 for (i = 0; i < t.count; i++) {
4342 struct location *l = &t.loc[i];
4344 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4346 len += sprintf(buf + len, "%7ld ", l->count);
4349 len += sprintf(buf + len, "%pS", (void *)l->addr);
4351 len += sprintf(buf + len, "<not-available>");
4353 if (l->sum_time != l->min_time) {
4354 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4356 (long)div_u64(l->sum_time, l->count),
4359 len += sprintf(buf + len, " age=%ld",
4362 if (l->min_pid != l->max_pid)
4363 len += sprintf(buf + len, " pid=%ld-%ld",
4364 l->min_pid, l->max_pid);
4366 len += sprintf(buf + len, " pid=%ld",
4369 if (num_online_cpus() > 1 &&
4370 !cpumask_empty(to_cpumask(l->cpus)) &&
4371 len < PAGE_SIZE - 60) {
4372 len += sprintf(buf + len, " cpus=");
4373 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4374 to_cpumask(l->cpus));
4377 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4378 len < PAGE_SIZE - 60) {
4379 len += sprintf(buf + len, " nodes=");
4380 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4384 len += sprintf(buf + len, "\n");
4390 len += sprintf(buf, "No data\n");
4395 #ifdef SLUB_RESILIENCY_TEST
4396 static void resiliency_test(void)
4400 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4402 printk(KERN_ERR "SLUB resiliency testing\n");
4403 printk(KERN_ERR "-----------------------\n");
4404 printk(KERN_ERR "A. Corruption after allocation\n");
4406 p = kzalloc(16, GFP_KERNEL);
4408 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4409 " 0x12->0x%p\n\n", p + 16);
4411 validate_slab_cache(kmalloc_caches[4]);
4413 /* Hmmm... The next two are dangerous */
4414 p = kzalloc(32, GFP_KERNEL);
4415 p[32 + sizeof(void *)] = 0x34;
4416 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4417 " 0x34 -> -0x%p\n", p);
4419 "If allocated object is overwritten then not detectable\n\n");
4421 validate_slab_cache(kmalloc_caches[5]);
4422 p = kzalloc(64, GFP_KERNEL);
4423 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4425 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4428 "If allocated object is overwritten then not detectable\n\n");
4429 validate_slab_cache(kmalloc_caches[6]);
4431 printk(KERN_ERR "\nB. Corruption after free\n");
4432 p = kzalloc(128, GFP_KERNEL);
4435 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4436 validate_slab_cache(kmalloc_caches[7]);
4438 p = kzalloc(256, GFP_KERNEL);
4441 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4443 validate_slab_cache(kmalloc_caches[8]);
4445 p = kzalloc(512, GFP_KERNEL);
4448 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4449 validate_slab_cache(kmalloc_caches[9]);
4453 static void resiliency_test(void) {};
4458 enum slab_stat_type {
4459 SL_ALL, /* All slabs */
4460 SL_PARTIAL, /* Only partially allocated slabs */
4461 SL_CPU, /* Only slabs used for cpu caches */
4462 SL_OBJECTS, /* Determine allocated objects not slabs */
4463 SL_TOTAL /* Determine object capacity not slabs */
4466 #define SO_ALL (1 << SL_ALL)
4467 #define SO_PARTIAL (1 << SL_PARTIAL)
4468 #define SO_CPU (1 << SL_CPU)
4469 #define SO_OBJECTS (1 << SL_OBJECTS)
4470 #define SO_TOTAL (1 << SL_TOTAL)
4472 static ssize_t show_slab_objects(struct kmem_cache *s,
4473 char *buf, unsigned long flags)
4475 unsigned long total = 0;
4478 unsigned long *nodes;
4479 unsigned long *per_cpu;
4481 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4484 per_cpu = nodes + nr_node_ids;
4486 if (flags & SO_CPU) {
4489 for_each_possible_cpu(cpu) {
4490 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4494 page = ACCESS_ONCE(c->page);
4498 node = page_to_nid(page);
4499 if (flags & SO_TOTAL)
4501 else if (flags & SO_OBJECTS)
4509 page = ACCESS_ONCE(c->partial);
4520 lock_memory_hotplug();
4521 #ifdef CONFIG_SLUB_DEBUG
4522 if (flags & SO_ALL) {
4523 for_each_node_state(node, N_NORMAL_MEMORY) {
4524 struct kmem_cache_node *n = get_node(s, node);
4526 if (flags & SO_TOTAL)
4527 x = atomic_long_read(&n->total_objects);
4528 else if (flags & SO_OBJECTS)
4529 x = atomic_long_read(&n->total_objects) -
4530 count_partial(n, count_free);
4533 x = atomic_long_read(&n->nr_slabs);
4540 if (flags & SO_PARTIAL) {
4541 for_each_node_state(node, N_NORMAL_MEMORY) {
4542 struct kmem_cache_node *n = get_node(s, node);
4544 if (flags & SO_TOTAL)
4545 x = count_partial(n, count_total);
4546 else if (flags & SO_OBJECTS)
4547 x = count_partial(n, count_inuse);
4554 x = sprintf(buf, "%lu", total);
4556 for_each_node_state(node, N_NORMAL_MEMORY)
4558 x += sprintf(buf + x, " N%d=%lu",
4561 unlock_memory_hotplug();
4563 return x + sprintf(buf + x, "\n");
4566 #ifdef CONFIG_SLUB_DEBUG
4567 static int any_slab_objects(struct kmem_cache *s)
4571 for_each_online_node(node) {
4572 struct kmem_cache_node *n = get_node(s, node);
4577 if (atomic_long_read(&n->total_objects))
4584 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4585 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4587 struct slab_attribute {
4588 struct attribute attr;
4589 ssize_t (*show)(struct kmem_cache *s, char *buf);
4590 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4593 #define SLAB_ATTR_RO(_name) \
4594 static struct slab_attribute _name##_attr = \
4595 __ATTR(_name, 0400, _name##_show, NULL)
4597 #define SLAB_ATTR(_name) \
4598 static struct slab_attribute _name##_attr = \
4599 __ATTR(_name, 0600, _name##_show, _name##_store)
4601 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4603 return sprintf(buf, "%d\n", s->size);
4605 SLAB_ATTR_RO(slab_size);
4607 static ssize_t align_show(struct kmem_cache *s, char *buf)
4609 return sprintf(buf, "%d\n", s->align);
4611 SLAB_ATTR_RO(align);
4613 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4615 return sprintf(buf, "%d\n", s->object_size);
4617 SLAB_ATTR_RO(object_size);
4619 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4621 return sprintf(buf, "%d\n", oo_objects(s->oo));
4623 SLAB_ATTR_RO(objs_per_slab);
4625 static ssize_t order_store(struct kmem_cache *s,
4626 const char *buf, size_t length)
4628 unsigned long order;
4631 err = strict_strtoul(buf, 10, &order);
4635 if (order > slub_max_order || order < slub_min_order)
4638 calculate_sizes(s, order);
4642 static ssize_t order_show(struct kmem_cache *s, char *buf)
4644 return sprintf(buf, "%d\n", oo_order(s->oo));
4648 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4650 return sprintf(buf, "%lu\n", s->min_partial);
4653 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4659 err = strict_strtoul(buf, 10, &min);
4663 set_min_partial(s, min);
4666 SLAB_ATTR(min_partial);
4668 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4670 return sprintf(buf, "%u\n", s->cpu_partial);
4673 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4676 unsigned long objects;
4679 err = strict_strtoul(buf, 10, &objects);
4682 if (objects && kmem_cache_debug(s))
4685 s->cpu_partial = objects;
4689 SLAB_ATTR(cpu_partial);
4691 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4695 return sprintf(buf, "%pS\n", s->ctor);
4699 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4701 return sprintf(buf, "%d\n", s->refcount - 1);
4703 SLAB_ATTR_RO(aliases);
4705 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4707 return show_slab_objects(s, buf, SO_PARTIAL);
4709 SLAB_ATTR_RO(partial);
4711 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4713 return show_slab_objects(s, buf, SO_CPU);
4715 SLAB_ATTR_RO(cpu_slabs);
4717 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4719 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4721 SLAB_ATTR_RO(objects);
4723 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4725 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4727 SLAB_ATTR_RO(objects_partial);
4729 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4736 for_each_online_cpu(cpu) {
4737 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4740 pages += page->pages;
4741 objects += page->pobjects;
4745 len = sprintf(buf, "%d(%d)", objects, pages);
4748 for_each_online_cpu(cpu) {
4749 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4751 if (page && len < PAGE_SIZE - 20)
4752 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4753 page->pobjects, page->pages);
4756 return len + sprintf(buf + len, "\n");
4758 SLAB_ATTR_RO(slabs_cpu_partial);
4760 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4762 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4765 static ssize_t reclaim_account_store(struct kmem_cache *s,
4766 const char *buf, size_t length)
4768 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4770 s->flags |= SLAB_RECLAIM_ACCOUNT;
4773 SLAB_ATTR(reclaim_account);
4775 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4777 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4779 SLAB_ATTR_RO(hwcache_align);
4781 #ifdef CONFIG_ZONE_DMA
4782 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4784 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4786 SLAB_ATTR_RO(cache_dma);
4789 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4791 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4793 SLAB_ATTR_RO(destroy_by_rcu);
4795 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4797 return sprintf(buf, "%d\n", s->reserved);
4799 SLAB_ATTR_RO(reserved);
4801 #ifdef CONFIG_SLUB_DEBUG
4802 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4804 return show_slab_objects(s, buf, SO_ALL);
4806 SLAB_ATTR_RO(slabs);
4808 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4810 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4812 SLAB_ATTR_RO(total_objects);
4814 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4816 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4819 static ssize_t sanity_checks_store(struct kmem_cache *s,
4820 const char *buf, size_t length)
4822 s->flags &= ~SLAB_DEBUG_FREE;
4823 if (buf[0] == '1') {
4824 s->flags &= ~__CMPXCHG_DOUBLE;
4825 s->flags |= SLAB_DEBUG_FREE;
4829 SLAB_ATTR(sanity_checks);
4831 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4833 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4836 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4839 s->flags &= ~SLAB_TRACE;
4840 if (buf[0] == '1') {
4841 s->flags &= ~__CMPXCHG_DOUBLE;
4842 s->flags |= SLAB_TRACE;
4848 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4850 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4853 static ssize_t red_zone_store(struct kmem_cache *s,
4854 const char *buf, size_t length)
4856 if (any_slab_objects(s))
4859 s->flags &= ~SLAB_RED_ZONE;
4860 if (buf[0] == '1') {
4861 s->flags &= ~__CMPXCHG_DOUBLE;
4862 s->flags |= SLAB_RED_ZONE;
4864 calculate_sizes(s, -1);
4867 SLAB_ATTR(red_zone);
4869 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4871 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4874 static ssize_t poison_store(struct kmem_cache *s,
4875 const char *buf, size_t length)
4877 if (any_slab_objects(s))
4880 s->flags &= ~SLAB_POISON;
4881 if (buf[0] == '1') {
4882 s->flags &= ~__CMPXCHG_DOUBLE;
4883 s->flags |= SLAB_POISON;
4885 calculate_sizes(s, -1);
4890 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4892 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4895 static ssize_t store_user_store(struct kmem_cache *s,
4896 const char *buf, size_t length)
4898 if (any_slab_objects(s))
4901 s->flags &= ~SLAB_STORE_USER;
4902 if (buf[0] == '1') {
4903 s->flags &= ~__CMPXCHG_DOUBLE;
4904 s->flags |= SLAB_STORE_USER;
4906 calculate_sizes(s, -1);
4909 SLAB_ATTR(store_user);
4911 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4916 static ssize_t validate_store(struct kmem_cache *s,
4917 const char *buf, size_t length)
4921 if (buf[0] == '1') {
4922 ret = validate_slab_cache(s);
4928 SLAB_ATTR(validate);
4930 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4932 if (!(s->flags & SLAB_STORE_USER))
4934 return list_locations(s, buf, TRACK_ALLOC);
4936 SLAB_ATTR_RO(alloc_calls);
4938 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4940 if (!(s->flags & SLAB_STORE_USER))
4942 return list_locations(s, buf, TRACK_FREE);
4944 SLAB_ATTR_RO(free_calls);
4945 #endif /* CONFIG_SLUB_DEBUG */
4947 #ifdef CONFIG_FAILSLAB
4948 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4950 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4953 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4956 s->flags &= ~SLAB_FAILSLAB;
4958 s->flags |= SLAB_FAILSLAB;
4961 SLAB_ATTR(failslab);
4964 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4969 static ssize_t shrink_store(struct kmem_cache *s,
4970 const char *buf, size_t length)
4972 if (buf[0] == '1') {
4973 int rc = kmem_cache_shrink(s);
4984 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4986 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4989 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4990 const char *buf, size_t length)
4992 unsigned long ratio;
4995 err = strict_strtoul(buf, 10, &ratio);
5000 s->remote_node_defrag_ratio = ratio * 10;
5004 SLAB_ATTR(remote_node_defrag_ratio);
5007 #ifdef CONFIG_SLUB_STATS
5008 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5010 unsigned long sum = 0;
5013 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5018 for_each_online_cpu(cpu) {
5019 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5025 len = sprintf(buf, "%lu", sum);
5028 for_each_online_cpu(cpu) {
5029 if (data[cpu] && len < PAGE_SIZE - 20)
5030 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5034 return len + sprintf(buf + len, "\n");
5037 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5041 for_each_online_cpu(cpu)
5042 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5045 #define STAT_ATTR(si, text) \
5046 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5048 return show_stat(s, buf, si); \
5050 static ssize_t text##_store(struct kmem_cache *s, \
5051 const char *buf, size_t length) \
5053 if (buf[0] != '0') \
5055 clear_stat(s, si); \
5060 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5061 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5062 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5063 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5064 STAT_ATTR(FREE_FROZEN, free_frozen);
5065 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5066 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5067 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5068 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5069 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5070 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5071 STAT_ATTR(FREE_SLAB, free_slab);
5072 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5073 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5074 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5075 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5076 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5077 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5078 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5079 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5080 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5081 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5082 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5083 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5084 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5085 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5088 static struct attribute *slab_attrs[] = {
5089 &slab_size_attr.attr,
5090 &object_size_attr.attr,
5091 &objs_per_slab_attr.attr,
5093 &min_partial_attr.attr,
5094 &cpu_partial_attr.attr,
5096 &objects_partial_attr.attr,
5098 &cpu_slabs_attr.attr,
5102 &hwcache_align_attr.attr,
5103 &reclaim_account_attr.attr,
5104 &destroy_by_rcu_attr.attr,
5106 &reserved_attr.attr,
5107 &slabs_cpu_partial_attr.attr,
5108 #ifdef CONFIG_SLUB_DEBUG
5109 &total_objects_attr.attr,
5111 &sanity_checks_attr.attr,
5113 &red_zone_attr.attr,
5115 &store_user_attr.attr,
5116 &validate_attr.attr,
5117 &alloc_calls_attr.attr,
5118 &free_calls_attr.attr,
5120 #ifdef CONFIG_ZONE_DMA
5121 &cache_dma_attr.attr,
5124 &remote_node_defrag_ratio_attr.attr,
5126 #ifdef CONFIG_SLUB_STATS
5127 &alloc_fastpath_attr.attr,
5128 &alloc_slowpath_attr.attr,
5129 &free_fastpath_attr.attr,
5130 &free_slowpath_attr.attr,
5131 &free_frozen_attr.attr,
5132 &free_add_partial_attr.attr,
5133 &free_remove_partial_attr.attr,
5134 &alloc_from_partial_attr.attr,
5135 &alloc_slab_attr.attr,
5136 &alloc_refill_attr.attr,
5137 &alloc_node_mismatch_attr.attr,
5138 &free_slab_attr.attr,
5139 &cpuslab_flush_attr.attr,
5140 &deactivate_full_attr.attr,
5141 &deactivate_empty_attr.attr,
5142 &deactivate_to_head_attr.attr,
5143 &deactivate_to_tail_attr.attr,
5144 &deactivate_remote_frees_attr.attr,
5145 &deactivate_bypass_attr.attr,
5146 &order_fallback_attr.attr,
5147 &cmpxchg_double_fail_attr.attr,
5148 &cmpxchg_double_cpu_fail_attr.attr,
5149 &cpu_partial_alloc_attr.attr,
5150 &cpu_partial_free_attr.attr,
5151 &cpu_partial_node_attr.attr,
5152 &cpu_partial_drain_attr.attr,
5154 #ifdef CONFIG_FAILSLAB
5155 &failslab_attr.attr,
5161 static struct attribute_group slab_attr_group = {
5162 .attrs = slab_attrs,
5165 static ssize_t slab_attr_show(struct kobject *kobj,
5166 struct attribute *attr,
5169 struct slab_attribute *attribute;
5170 struct kmem_cache *s;
5173 attribute = to_slab_attr(attr);
5176 if (!attribute->show)
5179 err = attribute->show(s, buf);
5184 static ssize_t slab_attr_store(struct kobject *kobj,
5185 struct attribute *attr,
5186 const char *buf, size_t len)
5188 struct slab_attribute *attribute;
5189 struct kmem_cache *s;
5192 attribute = to_slab_attr(attr);
5195 if (!attribute->store)
5198 err = attribute->store(s, buf, len);
5203 static void kmem_cache_release(struct kobject *kobj)
5205 struct kmem_cache *s = to_slab(kobj);
5210 static const struct sysfs_ops slab_sysfs_ops = {
5211 .show = slab_attr_show,
5212 .store = slab_attr_store,
5215 static struct kobj_type slab_ktype = {
5216 .sysfs_ops = &slab_sysfs_ops,
5217 .release = kmem_cache_release
5220 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5222 struct kobj_type *ktype = get_ktype(kobj);
5224 if (ktype == &slab_ktype)
5229 static const struct kset_uevent_ops slab_uevent_ops = {
5230 .filter = uevent_filter,
5233 static struct kset *slab_kset;
5235 #define ID_STR_LENGTH 64
5237 /* Create a unique string id for a slab cache:
5239 * Format :[flags-]size
5241 static char *create_unique_id(struct kmem_cache *s)
5243 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5250 * First flags affecting slabcache operations. We will only
5251 * get here for aliasable slabs so we do not need to support
5252 * too many flags. The flags here must cover all flags that
5253 * are matched during merging to guarantee that the id is
5256 if (s->flags & SLAB_CACHE_DMA)
5258 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5260 if (s->flags & SLAB_DEBUG_FREE)
5262 if (!(s->flags & SLAB_NOTRACK))
5266 p += sprintf(p, "%07d", s->size);
5267 BUG_ON(p > name + ID_STR_LENGTH - 1);
5271 static int sysfs_slab_add(struct kmem_cache *s)
5277 if (slab_state < FULL)
5278 /* Defer until later */
5281 unmergeable = slab_unmergeable(s);
5284 * Slabcache can never be merged so we can use the name proper.
5285 * This is typically the case for debug situations. In that
5286 * case we can catch duplicate names easily.
5288 sysfs_remove_link(&slab_kset->kobj, s->name);
5292 * Create a unique name for the slab as a target
5295 name = create_unique_id(s);
5298 s->kobj.kset = slab_kset;
5299 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5301 kobject_put(&s->kobj);
5305 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5307 kobject_del(&s->kobj);
5308 kobject_put(&s->kobj);
5311 kobject_uevent(&s->kobj, KOBJ_ADD);
5313 /* Setup first alias */
5314 sysfs_slab_alias(s, s->name);
5320 static void sysfs_slab_remove(struct kmem_cache *s)
5322 if (slab_state < FULL)
5324 * Sysfs has not been setup yet so no need to remove the
5329 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5330 kobject_del(&s->kobj);
5331 kobject_put(&s->kobj);
5335 * Need to buffer aliases during bootup until sysfs becomes
5336 * available lest we lose that information.
5338 struct saved_alias {
5339 struct kmem_cache *s;
5341 struct saved_alias *next;
5344 static struct saved_alias *alias_list;
5346 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5348 struct saved_alias *al;
5350 if (slab_state == FULL) {
5352 * If we have a leftover link then remove it.
5354 sysfs_remove_link(&slab_kset->kobj, name);
5355 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5358 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5364 al->next = alias_list;
5369 static int __init slab_sysfs_init(void)
5371 struct kmem_cache *s;
5374 mutex_lock(&slab_mutex);
5376 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5378 mutex_unlock(&slab_mutex);
5379 printk(KERN_ERR "Cannot register slab subsystem.\n");
5385 list_for_each_entry(s, &slab_caches, list) {
5386 err = sysfs_slab_add(s);
5388 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5389 " to sysfs\n", s->name);
5392 while (alias_list) {
5393 struct saved_alias *al = alias_list;
5395 alias_list = alias_list->next;
5396 err = sysfs_slab_alias(al->s, al->name);
5398 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5399 " %s to sysfs\n", al->name);
5403 mutex_unlock(&slab_mutex);
5408 __initcall(slab_sysfs_init);
5409 #endif /* CONFIG_SYSFS */
5412 * The /proc/slabinfo ABI
5414 #ifdef CONFIG_SLABINFO
5415 static void print_slabinfo_header(struct seq_file *m)
5417 seq_puts(m, "slabinfo - version: 2.1\n");
5418 seq_puts(m, "# name <active_objs> <num_objs> <object_size> "
5419 "<objperslab> <pagesperslab>");
5420 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5421 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5425 static void *s_start(struct seq_file *m, loff_t *pos)
5429 mutex_lock(&slab_mutex);
5431 print_slabinfo_header(m);
5433 return seq_list_start(&slab_caches, *pos);
5436 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5438 return seq_list_next(p, &slab_caches, pos);
5441 static void s_stop(struct seq_file *m, void *p)
5443 mutex_unlock(&slab_mutex);
5446 static int s_show(struct seq_file *m, void *p)
5448 unsigned long nr_partials = 0;
5449 unsigned long nr_slabs = 0;
5450 unsigned long nr_inuse = 0;
5451 unsigned long nr_objs = 0;
5452 unsigned long nr_free = 0;
5453 struct kmem_cache *s;
5456 s = list_entry(p, struct kmem_cache, list);
5458 for_each_online_node(node) {
5459 struct kmem_cache_node *n = get_node(s, node);
5464 nr_partials += n->nr_partial;
5465 nr_slabs += atomic_long_read(&n->nr_slabs);
5466 nr_objs += atomic_long_read(&n->total_objects);
5467 nr_free += count_partial(n, count_free);
5470 nr_inuse = nr_objs - nr_free;
5472 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5473 nr_objs, s->size, oo_objects(s->oo),
5474 (1 << oo_order(s->oo)));
5475 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5476 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5482 static const struct seq_operations slabinfo_op = {
5489 static int slabinfo_open(struct inode *inode, struct file *file)
5491 return seq_open(file, &slabinfo_op);
5494 static const struct file_operations proc_slabinfo_operations = {
5495 .open = slabinfo_open,
5497 .llseek = seq_lseek,
5498 .release = seq_release,
5501 static int __init slab_proc_init(void)
5503 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5506 module_init(slab_proc_init);
5507 #endif /* CONFIG_SLABINFO */