2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
35 #include <trace/events/kmem.h>
41 * 1. slab_mutex (Global Mutex)
43 * 3. slab_lock(page) (Only on some arches and for debugging)
47 * The role of the slab_mutex is to protect the list of all the slabs
48 * and to synchronize major metadata changes to slab cache structures.
50 * The slab_lock is only used for debugging and on arches that do not
51 * have the ability to do a cmpxchg_double. It only protects the second
52 * double word in the page struct. Meaning
53 * A. page->freelist -> List of object free in a page
54 * B. page->counters -> Counters of objects
55 * C. page->frozen -> frozen state
57 * If a slab is frozen then it is exempt from list management. It is not
58 * on any list. The processor that froze the slab is the one who can
59 * perform list operations on the page. Other processors may put objects
60 * onto the freelist but the processor that froze the slab is the only
61 * one that can retrieve the objects from the page's freelist.
63 * The list_lock protects the partial and full list on each node and
64 * the partial slab counter. If taken then no new slabs may be added or
65 * removed from the lists nor make the number of partial slabs be modified.
66 * (Note that the total number of slabs is an atomic value that may be
67 * modified without taking the list lock).
69 * The list_lock is a centralized lock and thus we avoid taking it as
70 * much as possible. As long as SLUB does not have to handle partial
71 * slabs, operations can continue without any centralized lock. F.e.
72 * allocating a long series of objects that fill up slabs does not require
74 * Interrupts are disabled during allocation and deallocation in order to
75 * make the slab allocator safe to use in the context of an irq. In addition
76 * interrupts are disabled to ensure that the processor does not change
77 * while handling per_cpu slabs, due to kernel preemption.
79 * SLUB assigns one slab for allocation to each processor.
80 * Allocations only occur from these slabs called cpu slabs.
82 * Slabs with free elements are kept on a partial list and during regular
83 * operations no list for full slabs is used. If an object in a full slab is
84 * freed then the slab will show up again on the partial lists.
85 * We track full slabs for debugging purposes though because otherwise we
86 * cannot scan all objects.
88 * Slabs are freed when they become empty. Teardown and setup is
89 * minimal so we rely on the page allocators per cpu caches for
90 * fast frees and allocs.
92 * Overloading of page flags that are otherwise used for LRU management.
94 * PageActive The slab is frozen and exempt from list processing.
95 * This means that the slab is dedicated to a purpose
96 * such as satisfying allocations for a specific
97 * processor. Objects may be freed in the slab while
98 * it is frozen but slab_free will then skip the usual
99 * list operations. It is up to the processor holding
100 * the slab to integrate the slab into the slab lists
101 * when the slab is no longer needed.
103 * One use of this flag is to mark slabs that are
104 * used for allocations. Then such a slab becomes a cpu
105 * slab. The cpu slab may be equipped with an additional
106 * freelist that allows lockless access to
107 * free objects in addition to the regular freelist
108 * that requires the slab lock.
110 * PageError Slab requires special handling due to debug
111 * options set. This moves slab handling out of
112 * the fast path and disables lockless freelists.
115 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
116 SLAB_TRACE | SLAB_DEBUG_FREE)
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
128 * Issues still to be resolved:
130 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
132 * - Variable sizing of the per node arrays
135 /* Enable to test recovery from slab corruption on boot */
136 #undef SLUB_RESILIENCY_TEST
138 /* Enable to log cmpxchg failures */
139 #undef SLUB_DEBUG_CMPXCHG
142 * Mininum number of partial slabs. These will be left on the partial
143 * lists even if they are empty. kmem_cache_shrink may reclaim them.
145 #define MIN_PARTIAL 5
148 * Maximum number of desirable partial slabs.
149 * The existence of more partial slabs makes kmem_cache_shrink
150 * sort the partial list by the number of objects in the.
152 #define MAX_PARTIAL 10
154 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_STORE_USER)
158 * Debugging flags that require metadata to be stored in the slab. These get
159 * disabled when slub_debug=O is used and a cache's min order increases with
162 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
165 * Set of flags that will prevent slab merging
167 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
168 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
171 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
172 SLAB_CACHE_DMA | SLAB_NOTRACK)
175 #define OO_MASK ((1 << OO_SHIFT) - 1)
176 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
178 /* Internal SLUB flags */
179 #define __OBJECT_POISON 0x80000000UL /* Poison object */
180 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static int kmem_size = sizeof(struct kmem_cache);
185 static struct notifier_block slab_notifier;
189 * Tracking user of a slab.
191 #define TRACK_ADDRS_COUNT 16
193 unsigned long addr; /* Called from address */
194 #ifdef CONFIG_STACKTRACE
195 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
197 int cpu; /* Was running on cpu */
198 int pid; /* Pid context */
199 unsigned long when; /* When did the operation occur */
202 enum track_item { TRACK_ALLOC, TRACK_FREE };
205 static int sysfs_slab_add(struct kmem_cache *);
206 static int sysfs_slab_alias(struct kmem_cache *, const char *);
207 static void sysfs_slab_remove(struct kmem_cache *);
210 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
211 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
213 static inline void sysfs_slab_remove(struct kmem_cache *s)
221 static inline void stat(const struct kmem_cache *s, enum stat_item si)
223 #ifdef CONFIG_SLUB_STATS
224 __this_cpu_inc(s->cpu_slab->stat[si]);
228 /********************************************************************
229 * Core slab cache functions
230 *******************************************************************/
232 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
234 return s->node[node];
237 /* Verify that a pointer has an address that is valid within a slab page */
238 static inline int check_valid_pointer(struct kmem_cache *s,
239 struct page *page, const void *object)
246 base = page_address(page);
247 if (object < base || object >= base + page->objects * s->size ||
248 (object - base) % s->size) {
255 static inline void *get_freepointer(struct kmem_cache *s, void *object)
257 return *(void **)(object + s->offset);
260 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
262 prefetch(object + s->offset);
265 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
269 #ifdef CONFIG_DEBUG_PAGEALLOC
270 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
272 p = get_freepointer(s, object);
277 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
279 *(void **)(object + s->offset) = fp;
282 /* Loop over all objects in a slab */
283 #define for_each_object(__p, __s, __addr, __objects) \
284 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
287 /* Determine object index from a given position */
288 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
290 return (p - addr) / s->size;
293 static inline size_t slab_ksize(const struct kmem_cache *s)
295 #ifdef CONFIG_SLUB_DEBUG
297 * Debugging requires use of the padding between object
298 * and whatever may come after it.
300 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
301 return s->object_size;
305 * If we have the need to store the freelist pointer
306 * back there or track user information then we can
307 * only use the space before that information.
309 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
312 * Else we can use all the padding etc for the allocation
317 static inline int order_objects(int order, unsigned long size, int reserved)
319 return ((PAGE_SIZE << order) - reserved) / size;
322 static inline struct kmem_cache_order_objects oo_make(int order,
323 unsigned long size, int reserved)
325 struct kmem_cache_order_objects x = {
326 (order << OO_SHIFT) + order_objects(order, size, reserved)
332 static inline int oo_order(struct kmem_cache_order_objects x)
334 return x.x >> OO_SHIFT;
337 static inline int oo_objects(struct kmem_cache_order_objects x)
339 return x.x & OO_MASK;
343 * Per slab locking using the pagelock
345 static __always_inline void slab_lock(struct page *page)
347 bit_spin_lock(PG_locked, &page->flags);
350 static __always_inline void slab_unlock(struct page *page)
352 __bit_spin_unlock(PG_locked, &page->flags);
355 /* Interrupts must be disabled (for the fallback code to work right) */
356 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
357 void *freelist_old, unsigned long counters_old,
358 void *freelist_new, unsigned long counters_new,
361 VM_BUG_ON(!irqs_disabled());
362 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
363 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
364 if (s->flags & __CMPXCHG_DOUBLE) {
365 if (cmpxchg_double(&page->freelist, &page->counters,
366 freelist_old, counters_old,
367 freelist_new, counters_new))
373 if (page->freelist == freelist_old && page->counters == counters_old) {
374 page->freelist = freelist_new;
375 page->counters = counters_new;
383 stat(s, CMPXCHG_DOUBLE_FAIL);
385 #ifdef SLUB_DEBUG_CMPXCHG
386 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
392 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
393 void *freelist_old, unsigned long counters_old,
394 void *freelist_new, unsigned long counters_new,
397 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
398 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
399 if (s->flags & __CMPXCHG_DOUBLE) {
400 if (cmpxchg_double(&page->freelist, &page->counters,
401 freelist_old, counters_old,
402 freelist_new, counters_new))
409 local_irq_save(flags);
411 if (page->freelist == freelist_old && page->counters == counters_old) {
412 page->freelist = freelist_new;
413 page->counters = counters_new;
415 local_irq_restore(flags);
419 local_irq_restore(flags);
423 stat(s, CMPXCHG_DOUBLE_FAIL);
425 #ifdef SLUB_DEBUG_CMPXCHG
426 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432 #ifdef CONFIG_SLUB_DEBUG
434 * Determine a map of object in use on a page.
436 * Node listlock must be held to guarantee that the page does
437 * not vanish from under us.
439 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
442 void *addr = page_address(page);
444 for (p = page->freelist; p; p = get_freepointer(s, p))
445 set_bit(slab_index(p, s, addr), map);
451 #ifdef CONFIG_SLUB_DEBUG_ON
452 static int slub_debug = DEBUG_DEFAULT_FLAGS;
454 static int slub_debug;
457 static char *slub_debug_slabs;
458 static int disable_higher_order_debug;
463 static void print_section(char *text, u8 *addr, unsigned int length)
465 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
469 static struct track *get_track(struct kmem_cache *s, void *object,
470 enum track_item alloc)
475 p = object + s->offset + sizeof(void *);
477 p = object + s->inuse;
482 static void set_track(struct kmem_cache *s, void *object,
483 enum track_item alloc, unsigned long addr)
485 struct track *p = get_track(s, object, alloc);
488 #ifdef CONFIG_STACKTRACE
489 struct stack_trace trace;
492 trace.nr_entries = 0;
493 trace.max_entries = TRACK_ADDRS_COUNT;
494 trace.entries = p->addrs;
496 save_stack_trace(&trace);
498 /* See rant in lockdep.c */
499 if (trace.nr_entries != 0 &&
500 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
503 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
507 p->cpu = smp_processor_id();
508 p->pid = current->pid;
511 memset(p, 0, sizeof(struct track));
514 static void init_tracking(struct kmem_cache *s, void *object)
516 if (!(s->flags & SLAB_STORE_USER))
519 set_track(s, object, TRACK_FREE, 0UL);
520 set_track(s, object, TRACK_ALLOC, 0UL);
523 static void print_track(const char *s, struct track *t)
528 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
529 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
530 #ifdef CONFIG_STACKTRACE
533 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
535 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
542 static void print_tracking(struct kmem_cache *s, void *object)
544 if (!(s->flags & SLAB_STORE_USER))
547 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
548 print_track("Freed", get_track(s, object, TRACK_FREE));
551 static void print_page_info(struct page *page)
553 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
554 page, page->objects, page->inuse, page->freelist, page->flags);
558 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
564 vsnprintf(buf, sizeof(buf), fmt, args);
566 printk(KERN_ERR "========================================"
567 "=====================================\n");
568 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
569 printk(KERN_ERR "----------------------------------------"
570 "-------------------------------------\n\n");
572 add_taint(TAINT_BAD_PAGE);
575 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
581 vsnprintf(buf, sizeof(buf), fmt, args);
583 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
586 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
588 unsigned int off; /* Offset of last byte */
589 u8 *addr = page_address(page);
591 print_tracking(s, p);
593 print_page_info(page);
595 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
596 p, p - addr, get_freepointer(s, p));
599 print_section("Bytes b4 ", p - 16, 16);
601 print_section("Object ", p, min_t(unsigned long, s->object_size,
603 if (s->flags & SLAB_RED_ZONE)
604 print_section("Redzone ", p + s->object_size,
605 s->inuse - s->object_size);
608 off = s->offset + sizeof(void *);
612 if (s->flags & SLAB_STORE_USER)
613 off += 2 * sizeof(struct track);
616 /* Beginning of the filler is the free pointer */
617 print_section("Padding ", p + off, s->size - off);
622 static void object_err(struct kmem_cache *s, struct page *page,
623 u8 *object, char *reason)
625 slab_bug(s, "%s", reason);
626 print_trailer(s, page, object);
629 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
635 vsnprintf(buf, sizeof(buf), fmt, args);
637 slab_bug(s, "%s", buf);
638 print_page_info(page);
642 static void init_object(struct kmem_cache *s, void *object, u8 val)
646 if (s->flags & __OBJECT_POISON) {
647 memset(p, POISON_FREE, s->object_size - 1);
648 p[s->object_size - 1] = POISON_END;
651 if (s->flags & SLAB_RED_ZONE)
652 memset(p + s->object_size, val, s->inuse - s->object_size);
655 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
656 void *from, void *to)
658 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
659 memset(from, data, to - from);
662 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
663 u8 *object, char *what,
664 u8 *start, unsigned int value, unsigned int bytes)
669 fault = memchr_inv(start, value, bytes);
674 while (end > fault && end[-1] == value)
677 slab_bug(s, "%s overwritten", what);
678 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
679 fault, end - 1, fault[0], value);
680 print_trailer(s, page, object);
682 restore_bytes(s, what, value, fault, end);
690 * Bytes of the object to be managed.
691 * If the freepointer may overlay the object then the free
692 * pointer is the first word of the object.
694 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
697 * object + s->object_size
698 * Padding to reach word boundary. This is also used for Redzoning.
699 * Padding is extended by another word if Redzoning is enabled and
700 * object_size == inuse.
702 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
703 * 0xcc (RED_ACTIVE) for objects in use.
706 * Meta data starts here.
708 * A. Free pointer (if we cannot overwrite object on free)
709 * B. Tracking data for SLAB_STORE_USER
710 * C. Padding to reach required alignment boundary or at mininum
711 * one word if debugging is on to be able to detect writes
712 * before the word boundary.
714 * Padding is done using 0x5a (POISON_INUSE)
717 * Nothing is used beyond s->size.
719 * If slabcaches are merged then the object_size and inuse boundaries are mostly
720 * ignored. And therefore no slab options that rely on these boundaries
721 * may be used with merged slabcaches.
724 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
726 unsigned long off = s->inuse; /* The end of info */
729 /* Freepointer is placed after the object. */
730 off += sizeof(void *);
732 if (s->flags & SLAB_STORE_USER)
733 /* We also have user information there */
734 off += 2 * sizeof(struct track);
739 return check_bytes_and_report(s, page, p, "Object padding",
740 p + off, POISON_INUSE, s->size - off);
743 /* Check the pad bytes at the end of a slab page */
744 static int slab_pad_check(struct kmem_cache *s, struct page *page)
752 if (!(s->flags & SLAB_POISON))
755 start = page_address(page);
756 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
757 end = start + length;
758 remainder = length % s->size;
762 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
765 while (end > fault && end[-1] == POISON_INUSE)
768 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
769 print_section("Padding ", end - remainder, remainder);
771 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
775 static int check_object(struct kmem_cache *s, struct page *page,
776 void *object, u8 val)
779 u8 *endobject = object + s->object_size;
781 if (s->flags & SLAB_RED_ZONE) {
782 if (!check_bytes_and_report(s, page, object, "Redzone",
783 endobject, val, s->inuse - s->object_size))
786 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
787 check_bytes_and_report(s, page, p, "Alignment padding",
788 endobject, POISON_INUSE, s->inuse - s->object_size);
792 if (s->flags & SLAB_POISON) {
793 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
794 (!check_bytes_and_report(s, page, p, "Poison", p,
795 POISON_FREE, s->object_size - 1) ||
796 !check_bytes_and_report(s, page, p, "Poison",
797 p + s->object_size - 1, POISON_END, 1)))
800 * check_pad_bytes cleans up on its own.
802 check_pad_bytes(s, page, p);
805 if (!s->offset && val == SLUB_RED_ACTIVE)
807 * Object and freepointer overlap. Cannot check
808 * freepointer while object is allocated.
812 /* Check free pointer validity */
813 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
814 object_err(s, page, p, "Freepointer corrupt");
816 * No choice but to zap it and thus lose the remainder
817 * of the free objects in this slab. May cause
818 * another error because the object count is now wrong.
820 set_freepointer(s, p, NULL);
826 static int check_slab(struct kmem_cache *s, struct page *page)
830 VM_BUG_ON(!irqs_disabled());
832 if (!PageSlab(page)) {
833 slab_err(s, page, "Not a valid slab page");
837 maxobj = order_objects(compound_order(page), s->size, s->reserved);
838 if (page->objects > maxobj) {
839 slab_err(s, page, "objects %u > max %u",
840 s->name, page->objects, maxobj);
843 if (page->inuse > page->objects) {
844 slab_err(s, page, "inuse %u > max %u",
845 s->name, page->inuse, page->objects);
848 /* Slab_pad_check fixes things up after itself */
849 slab_pad_check(s, page);
854 * Determine if a certain object on a page is on the freelist. Must hold the
855 * slab lock to guarantee that the chains are in a consistent state.
857 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
862 unsigned long max_objects;
865 while (fp && nr <= page->objects) {
868 if (!check_valid_pointer(s, page, fp)) {
870 object_err(s, page, object,
871 "Freechain corrupt");
872 set_freepointer(s, object, NULL);
875 slab_err(s, page, "Freepointer corrupt");
876 page->freelist = NULL;
877 page->inuse = page->objects;
878 slab_fix(s, "Freelist cleared");
884 fp = get_freepointer(s, object);
888 max_objects = order_objects(compound_order(page), s->size, s->reserved);
889 if (max_objects > MAX_OBJS_PER_PAGE)
890 max_objects = MAX_OBJS_PER_PAGE;
892 if (page->objects != max_objects) {
893 slab_err(s, page, "Wrong number of objects. Found %d but "
894 "should be %d", page->objects, max_objects);
895 page->objects = max_objects;
896 slab_fix(s, "Number of objects adjusted.");
898 if (page->inuse != page->objects - nr) {
899 slab_err(s, page, "Wrong object count. Counter is %d but "
900 "counted were %d", page->inuse, page->objects - nr);
901 page->inuse = page->objects - nr;
902 slab_fix(s, "Object count adjusted.");
904 return search == NULL;
907 static void trace(struct kmem_cache *s, struct page *page, void *object,
910 if (s->flags & SLAB_TRACE) {
911 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
913 alloc ? "alloc" : "free",
918 print_section("Object ", (void *)object, s->object_size);
925 * Hooks for other subsystems that check memory allocations. In a typical
926 * production configuration these hooks all should produce no code at all.
928 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
930 flags &= gfp_allowed_mask;
931 lockdep_trace_alloc(flags);
932 might_sleep_if(flags & __GFP_WAIT);
934 return should_failslab(s->object_size, flags, s->flags);
937 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
939 flags &= gfp_allowed_mask;
940 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
941 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
944 static inline void slab_free_hook(struct kmem_cache *s, void *x)
946 kmemleak_free_recursive(x, s->flags);
949 * Trouble is that we may no longer disable interupts in the fast path
950 * So in order to make the debug calls that expect irqs to be
951 * disabled we need to disable interrupts temporarily.
953 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
957 local_irq_save(flags);
958 kmemcheck_slab_free(s, x, s->object_size);
959 debug_check_no_locks_freed(x, s->object_size);
960 local_irq_restore(flags);
963 if (!(s->flags & SLAB_DEBUG_OBJECTS))
964 debug_check_no_obj_freed(x, s->object_size);
968 * Tracking of fully allocated slabs for debugging purposes.
970 * list_lock must be held.
972 static void add_full(struct kmem_cache *s,
973 struct kmem_cache_node *n, struct page *page)
975 if (!(s->flags & SLAB_STORE_USER))
978 list_add(&page->lru, &n->full);
982 * list_lock must be held.
984 static void remove_full(struct kmem_cache *s, struct page *page)
986 if (!(s->flags & SLAB_STORE_USER))
989 list_del(&page->lru);
992 /* Tracking of the number of slabs for debugging purposes */
993 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
995 struct kmem_cache_node *n = get_node(s, node);
997 return atomic_long_read(&n->nr_slabs);
1000 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1002 return atomic_long_read(&n->nr_slabs);
1005 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1007 struct kmem_cache_node *n = get_node(s, node);
1010 * May be called early in order to allocate a slab for the
1011 * kmem_cache_node structure. Solve the chicken-egg
1012 * dilemma by deferring the increment of the count during
1013 * bootstrap (see early_kmem_cache_node_alloc).
1016 atomic_long_inc(&n->nr_slabs);
1017 atomic_long_add(objects, &n->total_objects);
1020 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1022 struct kmem_cache_node *n = get_node(s, node);
1024 atomic_long_dec(&n->nr_slabs);
1025 atomic_long_sub(objects, &n->total_objects);
1028 /* Object debug checks for alloc/free paths */
1029 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1032 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1035 init_object(s, object, SLUB_RED_INACTIVE);
1036 init_tracking(s, object);
1039 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1040 void *object, unsigned long addr)
1042 if (!check_slab(s, page))
1045 if (!check_valid_pointer(s, page, object)) {
1046 object_err(s, page, object, "Freelist Pointer check fails");
1050 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1053 /* Success perform special debug activities for allocs */
1054 if (s->flags & SLAB_STORE_USER)
1055 set_track(s, object, TRACK_ALLOC, addr);
1056 trace(s, page, object, 1);
1057 init_object(s, object, SLUB_RED_ACTIVE);
1061 if (PageSlab(page)) {
1063 * If this is a slab page then lets do the best we can
1064 * to avoid issues in the future. Marking all objects
1065 * as used avoids touching the remaining objects.
1067 slab_fix(s, "Marking all objects used");
1068 page->inuse = page->objects;
1069 page->freelist = NULL;
1074 static noinline struct kmem_cache_node *free_debug_processing(
1075 struct kmem_cache *s, struct page *page, void *object,
1076 unsigned long addr, unsigned long *flags)
1078 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1080 spin_lock_irqsave(&n->list_lock, *flags);
1083 if (!check_slab(s, page))
1086 if (!check_valid_pointer(s, page, object)) {
1087 slab_err(s, page, "Invalid object pointer 0x%p", object);
1091 if (on_freelist(s, page, object)) {
1092 object_err(s, page, object, "Object already free");
1096 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1099 if (unlikely(s != page->slab)) {
1100 if (!PageSlab(page)) {
1101 slab_err(s, page, "Attempt to free object(0x%p) "
1102 "outside of slab", object);
1103 } else if (!page->slab) {
1105 "SLUB <none>: no slab for object 0x%p.\n",
1109 object_err(s, page, object,
1110 "page slab pointer corrupt.");
1114 if (s->flags & SLAB_STORE_USER)
1115 set_track(s, object, TRACK_FREE, addr);
1116 trace(s, page, object, 0);
1117 init_object(s, object, SLUB_RED_INACTIVE);
1121 * Keep node_lock to preserve integrity
1122 * until the object is actually freed
1128 spin_unlock_irqrestore(&n->list_lock, *flags);
1129 slab_fix(s, "Object at 0x%p not freed", object);
1133 static int __init setup_slub_debug(char *str)
1135 slub_debug = DEBUG_DEFAULT_FLAGS;
1136 if (*str++ != '=' || !*str)
1138 * No options specified. Switch on full debugging.
1144 * No options but restriction on slabs. This means full
1145 * debugging for slabs matching a pattern.
1149 if (tolower(*str) == 'o') {
1151 * Avoid enabling debugging on caches if its minimum order
1152 * would increase as a result.
1154 disable_higher_order_debug = 1;
1161 * Switch off all debugging measures.
1166 * Determine which debug features should be switched on
1168 for (; *str && *str != ','; str++) {
1169 switch (tolower(*str)) {
1171 slub_debug |= SLAB_DEBUG_FREE;
1174 slub_debug |= SLAB_RED_ZONE;
1177 slub_debug |= SLAB_POISON;
1180 slub_debug |= SLAB_STORE_USER;
1183 slub_debug |= SLAB_TRACE;
1186 slub_debug |= SLAB_FAILSLAB;
1189 printk(KERN_ERR "slub_debug option '%c' "
1190 "unknown. skipped\n", *str);
1196 slub_debug_slabs = str + 1;
1201 __setup("slub_debug", setup_slub_debug);
1203 static unsigned long kmem_cache_flags(unsigned long object_size,
1204 unsigned long flags, const char *name,
1205 void (*ctor)(void *))
1208 * Enable debugging if selected on the kernel commandline.
1210 if (slub_debug && (!slub_debug_slabs ||
1211 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1212 flags |= slub_debug;
1217 static inline void setup_object_debug(struct kmem_cache *s,
1218 struct page *page, void *object) {}
1220 static inline int alloc_debug_processing(struct kmem_cache *s,
1221 struct page *page, void *object, unsigned long addr) { return 0; }
1223 static inline struct kmem_cache_node *free_debug_processing(
1224 struct kmem_cache *s, struct page *page, void *object,
1225 unsigned long addr, unsigned long *flags) { return NULL; }
1227 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1229 static inline int check_object(struct kmem_cache *s, struct page *page,
1230 void *object, u8 val) { return 1; }
1231 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1232 struct page *page) {}
1233 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1234 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1235 unsigned long flags, const char *name,
1236 void (*ctor)(void *))
1240 #define slub_debug 0
1242 #define disable_higher_order_debug 0
1244 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1246 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1248 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1250 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1253 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1256 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1259 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1261 #endif /* CONFIG_SLUB_DEBUG */
1264 * Slab allocation and freeing
1266 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1267 struct kmem_cache_order_objects oo)
1269 int order = oo_order(oo);
1271 flags |= __GFP_NOTRACK;
1273 if (node == NUMA_NO_NODE)
1274 return alloc_pages(flags, order);
1276 return alloc_pages_exact_node(node, flags, order);
1279 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1282 struct kmem_cache_order_objects oo = s->oo;
1285 flags &= gfp_allowed_mask;
1287 if (flags & __GFP_WAIT)
1290 flags |= s->allocflags;
1293 * Let the initial higher-order allocation fail under memory pressure
1294 * so we fall-back to the minimum order allocation.
1296 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1298 page = alloc_slab_page(alloc_gfp, node, oo);
1299 if (unlikely(!page)) {
1302 * Allocation may have failed due to fragmentation.
1303 * Try a lower order alloc if possible
1305 page = alloc_slab_page(flags, node, oo);
1308 stat(s, ORDER_FALLBACK);
1311 if (kmemcheck_enabled && page
1312 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1313 int pages = 1 << oo_order(oo);
1315 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1318 * Objects from caches that have a constructor don't get
1319 * cleared when they're allocated, so we need to do it here.
1322 kmemcheck_mark_uninitialized_pages(page, pages);
1324 kmemcheck_mark_unallocated_pages(page, pages);
1327 if (flags & __GFP_WAIT)
1328 local_irq_disable();
1332 page->objects = oo_objects(oo);
1333 mod_zone_page_state(page_zone(page),
1334 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1335 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1341 static void setup_object(struct kmem_cache *s, struct page *page,
1344 setup_object_debug(s, page, object);
1345 if (unlikely(s->ctor))
1349 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1356 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1358 page = allocate_slab(s,
1359 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1363 inc_slabs_node(s, page_to_nid(page), page->objects);
1365 __SetPageSlab(page);
1366 if (page->pfmemalloc)
1367 SetPageSlabPfmemalloc(page);
1369 start = page_address(page);
1371 if (unlikely(s->flags & SLAB_POISON))
1372 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1375 for_each_object(p, s, start, page->objects) {
1376 setup_object(s, page, last);
1377 set_freepointer(s, last, p);
1380 setup_object(s, page, last);
1381 set_freepointer(s, last, NULL);
1383 page->freelist = start;
1384 page->inuse = page->objects;
1390 static void __free_slab(struct kmem_cache *s, struct page *page)
1392 int order = compound_order(page);
1393 int pages = 1 << order;
1395 if (kmem_cache_debug(s)) {
1398 slab_pad_check(s, page);
1399 for_each_object(p, s, page_address(page),
1401 check_object(s, page, p, SLUB_RED_INACTIVE);
1404 kmemcheck_free_shadow(page, compound_order(page));
1406 mod_zone_page_state(page_zone(page),
1407 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1408 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1411 __ClearPageSlabPfmemalloc(page);
1412 __ClearPageSlab(page);
1413 reset_page_mapcount(page);
1414 if (current->reclaim_state)
1415 current->reclaim_state->reclaimed_slab += pages;
1416 __free_pages(page, order);
1419 #define need_reserve_slab_rcu \
1420 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1422 static void rcu_free_slab(struct rcu_head *h)
1426 if (need_reserve_slab_rcu)
1427 page = virt_to_head_page(h);
1429 page = container_of((struct list_head *)h, struct page, lru);
1431 __free_slab(page->slab, page);
1434 static void free_slab(struct kmem_cache *s, struct page *page)
1436 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1437 struct rcu_head *head;
1439 if (need_reserve_slab_rcu) {
1440 int order = compound_order(page);
1441 int offset = (PAGE_SIZE << order) - s->reserved;
1443 VM_BUG_ON(s->reserved != sizeof(*head));
1444 head = page_address(page) + offset;
1447 * RCU free overloads the RCU head over the LRU
1449 head = (void *)&page->lru;
1452 call_rcu(head, rcu_free_slab);
1454 __free_slab(s, page);
1457 static void discard_slab(struct kmem_cache *s, struct page *page)
1459 dec_slabs_node(s, page_to_nid(page), page->objects);
1464 * Management of partially allocated slabs.
1466 * list_lock must be held.
1468 static inline void add_partial(struct kmem_cache_node *n,
1469 struct page *page, int tail)
1472 if (tail == DEACTIVATE_TO_TAIL)
1473 list_add_tail(&page->lru, &n->partial);
1475 list_add(&page->lru, &n->partial);
1479 * list_lock must be held.
1481 static inline void remove_partial(struct kmem_cache_node *n,
1484 list_del(&page->lru);
1489 * Remove slab from the partial list, freeze it and
1490 * return the pointer to the freelist.
1492 * Returns a list of objects or NULL if it fails.
1494 * Must hold list_lock since we modify the partial list.
1496 static inline void *acquire_slab(struct kmem_cache *s,
1497 struct kmem_cache_node *n, struct page *page,
1501 unsigned long counters;
1505 * Zap the freelist and set the frozen bit.
1506 * The old freelist is the list of objects for the
1507 * per cpu allocation list.
1509 freelist = page->freelist;
1510 counters = page->counters;
1511 new.counters = counters;
1513 new.inuse = page->objects;
1514 new.freelist = NULL;
1516 new.freelist = freelist;
1519 VM_BUG_ON(new.frozen);
1522 if (!__cmpxchg_double_slab(s, page,
1524 new.freelist, new.counters,
1528 remove_partial(n, page);
1533 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1534 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1537 * Try to allocate a partial slab from a specific node.
1539 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1540 struct kmem_cache_cpu *c, gfp_t flags)
1542 struct page *page, *page2;
1543 void *object = NULL;
1546 * Racy check. If we mistakenly see no partial slabs then we
1547 * just allocate an empty slab. If we mistakenly try to get a
1548 * partial slab and there is none available then get_partials()
1551 if (!n || !n->nr_partial)
1554 spin_lock(&n->list_lock);
1555 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1559 if (!pfmemalloc_match(page, flags))
1562 t = acquire_slab(s, n, page, object == NULL);
1568 stat(s, ALLOC_FROM_PARTIAL);
1570 available = page->objects - page->inuse;
1572 available = put_cpu_partial(s, page, 0);
1573 stat(s, CPU_PARTIAL_NODE);
1575 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1579 spin_unlock(&n->list_lock);
1584 * Get a page from somewhere. Search in increasing NUMA distances.
1586 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1587 struct kmem_cache_cpu *c)
1590 struct zonelist *zonelist;
1593 enum zone_type high_zoneidx = gfp_zone(flags);
1595 unsigned int cpuset_mems_cookie;
1598 * The defrag ratio allows a configuration of the tradeoffs between
1599 * inter node defragmentation and node local allocations. A lower
1600 * defrag_ratio increases the tendency to do local allocations
1601 * instead of attempting to obtain partial slabs from other nodes.
1603 * If the defrag_ratio is set to 0 then kmalloc() always
1604 * returns node local objects. If the ratio is higher then kmalloc()
1605 * may return off node objects because partial slabs are obtained
1606 * from other nodes and filled up.
1608 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1609 * defrag_ratio = 1000) then every (well almost) allocation will
1610 * first attempt to defrag slab caches on other nodes. This means
1611 * scanning over all nodes to look for partial slabs which may be
1612 * expensive if we do it every time we are trying to find a slab
1613 * with available objects.
1615 if (!s->remote_node_defrag_ratio ||
1616 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1620 cpuset_mems_cookie = get_mems_allowed();
1621 zonelist = node_zonelist(slab_node(), flags);
1622 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1623 struct kmem_cache_node *n;
1625 n = get_node(s, zone_to_nid(zone));
1627 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1628 n->nr_partial > s->min_partial) {
1629 object = get_partial_node(s, n, c, flags);
1632 * Return the object even if
1633 * put_mems_allowed indicated that
1634 * the cpuset mems_allowed was
1635 * updated in parallel. It's a
1636 * harmless race between the alloc
1637 * and the cpuset update.
1639 put_mems_allowed(cpuset_mems_cookie);
1644 } while (!put_mems_allowed(cpuset_mems_cookie));
1650 * Get a partial page, lock it and return it.
1652 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1653 struct kmem_cache_cpu *c)
1656 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1658 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1659 if (object || node != NUMA_NO_NODE)
1662 return get_any_partial(s, flags, c);
1665 #ifdef CONFIG_PREEMPT
1667 * Calculate the next globally unique transaction for disambiguiation
1668 * during cmpxchg. The transactions start with the cpu number and are then
1669 * incremented by CONFIG_NR_CPUS.
1671 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1674 * No preemption supported therefore also no need to check for
1680 static inline unsigned long next_tid(unsigned long tid)
1682 return tid + TID_STEP;
1685 static inline unsigned int tid_to_cpu(unsigned long tid)
1687 return tid % TID_STEP;
1690 static inline unsigned long tid_to_event(unsigned long tid)
1692 return tid / TID_STEP;
1695 static inline unsigned int init_tid(int cpu)
1700 static inline void note_cmpxchg_failure(const char *n,
1701 const struct kmem_cache *s, unsigned long tid)
1703 #ifdef SLUB_DEBUG_CMPXCHG
1704 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1706 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1708 #ifdef CONFIG_PREEMPT
1709 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1710 printk("due to cpu change %d -> %d\n",
1711 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1714 if (tid_to_event(tid) != tid_to_event(actual_tid))
1715 printk("due to cpu running other code. Event %ld->%ld\n",
1716 tid_to_event(tid), tid_to_event(actual_tid));
1718 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1719 actual_tid, tid, next_tid(tid));
1721 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1724 static void init_kmem_cache_cpus(struct kmem_cache *s)
1728 for_each_possible_cpu(cpu)
1729 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1733 * Remove the cpu slab
1735 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1737 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1738 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1740 enum slab_modes l = M_NONE, m = M_NONE;
1742 int tail = DEACTIVATE_TO_HEAD;
1746 if (page->freelist) {
1747 stat(s, DEACTIVATE_REMOTE_FREES);
1748 tail = DEACTIVATE_TO_TAIL;
1752 * Stage one: Free all available per cpu objects back
1753 * to the page freelist while it is still frozen. Leave the
1756 * There is no need to take the list->lock because the page
1759 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1761 unsigned long counters;
1764 prior = page->freelist;
1765 counters = page->counters;
1766 set_freepointer(s, freelist, prior);
1767 new.counters = counters;
1769 VM_BUG_ON(!new.frozen);
1771 } while (!__cmpxchg_double_slab(s, page,
1773 freelist, new.counters,
1774 "drain percpu freelist"));
1776 freelist = nextfree;
1780 * Stage two: Ensure that the page is unfrozen while the
1781 * list presence reflects the actual number of objects
1784 * We setup the list membership and then perform a cmpxchg
1785 * with the count. If there is a mismatch then the page
1786 * is not unfrozen but the page is on the wrong list.
1788 * Then we restart the process which may have to remove
1789 * the page from the list that we just put it on again
1790 * because the number of objects in the slab may have
1795 old.freelist = page->freelist;
1796 old.counters = page->counters;
1797 VM_BUG_ON(!old.frozen);
1799 /* Determine target state of the slab */
1800 new.counters = old.counters;
1803 set_freepointer(s, freelist, old.freelist);
1804 new.freelist = freelist;
1806 new.freelist = old.freelist;
1810 if (!new.inuse && n->nr_partial > s->min_partial)
1812 else if (new.freelist) {
1817 * Taking the spinlock removes the possiblity
1818 * that acquire_slab() will see a slab page that
1821 spin_lock(&n->list_lock);
1825 if (kmem_cache_debug(s) && !lock) {
1828 * This also ensures that the scanning of full
1829 * slabs from diagnostic functions will not see
1832 spin_lock(&n->list_lock);
1840 remove_partial(n, page);
1842 else if (l == M_FULL)
1844 remove_full(s, page);
1846 if (m == M_PARTIAL) {
1848 add_partial(n, page, tail);
1851 } else if (m == M_FULL) {
1853 stat(s, DEACTIVATE_FULL);
1854 add_full(s, n, page);
1860 if (!__cmpxchg_double_slab(s, page,
1861 old.freelist, old.counters,
1862 new.freelist, new.counters,
1867 spin_unlock(&n->list_lock);
1870 stat(s, DEACTIVATE_EMPTY);
1871 discard_slab(s, page);
1877 * Unfreeze all the cpu partial slabs.
1879 * This function must be called with interrupt disabled.
1881 static void unfreeze_partials(struct kmem_cache *s)
1883 struct kmem_cache_node *n = NULL, *n2 = NULL;
1884 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1885 struct page *page, *discard_page = NULL;
1887 while ((page = c->partial)) {
1891 c->partial = page->next;
1893 n2 = get_node(s, page_to_nid(page));
1896 spin_unlock(&n->list_lock);
1899 spin_lock(&n->list_lock);
1904 old.freelist = page->freelist;
1905 old.counters = page->counters;
1906 VM_BUG_ON(!old.frozen);
1908 new.counters = old.counters;
1909 new.freelist = old.freelist;
1913 } while (!__cmpxchg_double_slab(s, page,
1914 old.freelist, old.counters,
1915 new.freelist, new.counters,
1916 "unfreezing slab"));
1918 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1919 page->next = discard_page;
1920 discard_page = page;
1922 add_partial(n, page, DEACTIVATE_TO_TAIL);
1923 stat(s, FREE_ADD_PARTIAL);
1928 spin_unlock(&n->list_lock);
1930 while (discard_page) {
1931 page = discard_page;
1932 discard_page = discard_page->next;
1934 stat(s, DEACTIVATE_EMPTY);
1935 discard_slab(s, page);
1941 * Put a page that was just frozen (in __slab_free) into a partial page
1942 * slot if available. This is done without interrupts disabled and without
1943 * preemption disabled. The cmpxchg is racy and may put the partial page
1944 * onto a random cpus partial slot.
1946 * If we did not find a slot then simply move all the partials to the
1947 * per node partial list.
1949 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1951 struct page *oldpage;
1958 oldpage = this_cpu_read(s->cpu_slab->partial);
1961 pobjects = oldpage->pobjects;
1962 pages = oldpage->pages;
1963 if (drain && pobjects > s->cpu_partial) {
1964 unsigned long flags;
1966 * partial array is full. Move the existing
1967 * set to the per node partial list.
1969 local_irq_save(flags);
1970 unfreeze_partials(s);
1971 local_irq_restore(flags);
1975 stat(s, CPU_PARTIAL_DRAIN);
1980 pobjects += page->objects - page->inuse;
1982 page->pages = pages;
1983 page->pobjects = pobjects;
1984 page->next = oldpage;
1986 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1990 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1992 stat(s, CPUSLAB_FLUSH);
1993 deactivate_slab(s, c->page, c->freelist);
1995 c->tid = next_tid(c->tid);
2003 * Called from IPI handler with interrupts disabled.
2005 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2007 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2013 unfreeze_partials(s);
2017 static void flush_cpu_slab(void *d)
2019 struct kmem_cache *s = d;
2021 __flush_cpu_slab(s, smp_processor_id());
2024 static bool has_cpu_slab(int cpu, void *info)
2026 struct kmem_cache *s = info;
2027 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2029 return c->page || c->partial;
2032 static void flush_all(struct kmem_cache *s)
2034 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2038 * Check if the objects in a per cpu structure fit numa
2039 * locality expectations.
2041 static inline int node_match(struct page *page, int node)
2044 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2050 static int count_free(struct page *page)
2052 return page->objects - page->inuse;
2055 static unsigned long count_partial(struct kmem_cache_node *n,
2056 int (*get_count)(struct page *))
2058 unsigned long flags;
2059 unsigned long x = 0;
2062 spin_lock_irqsave(&n->list_lock, flags);
2063 list_for_each_entry(page, &n->partial, lru)
2064 x += get_count(page);
2065 spin_unlock_irqrestore(&n->list_lock, flags);
2069 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2071 #ifdef CONFIG_SLUB_DEBUG
2072 return atomic_long_read(&n->total_objects);
2078 static noinline void
2079 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2084 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2086 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2087 "default order: %d, min order: %d\n", s->name, s->object_size,
2088 s->size, oo_order(s->oo), oo_order(s->min));
2090 if (oo_order(s->min) > get_order(s->object_size))
2091 printk(KERN_WARNING " %s debugging increased min order, use "
2092 "slub_debug=O to disable.\n", s->name);
2094 for_each_online_node(node) {
2095 struct kmem_cache_node *n = get_node(s, node);
2096 unsigned long nr_slabs;
2097 unsigned long nr_objs;
2098 unsigned long nr_free;
2103 nr_free = count_partial(n, count_free);
2104 nr_slabs = node_nr_slabs(n);
2105 nr_objs = node_nr_objs(n);
2108 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2109 node, nr_slabs, nr_objs, nr_free);
2113 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2114 int node, struct kmem_cache_cpu **pc)
2117 struct kmem_cache_cpu *c = *pc;
2120 freelist = get_partial(s, flags, node, c);
2125 page = new_slab(s, flags, node);
2127 c = __this_cpu_ptr(s->cpu_slab);
2132 * No other reference to the page yet so we can
2133 * muck around with it freely without cmpxchg
2135 freelist = page->freelist;
2136 page->freelist = NULL;
2138 stat(s, ALLOC_SLAB);
2147 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2149 if (unlikely(PageSlabPfmemalloc(page)))
2150 return gfp_pfmemalloc_allowed(gfpflags);
2156 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2157 * or deactivate the page.
2159 * The page is still frozen if the return value is not NULL.
2161 * If this function returns NULL then the page has been unfrozen.
2163 * This function must be called with interrupt disabled.
2165 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2168 unsigned long counters;
2172 freelist = page->freelist;
2173 counters = page->counters;
2175 new.counters = counters;
2176 VM_BUG_ON(!new.frozen);
2178 new.inuse = page->objects;
2179 new.frozen = freelist != NULL;
2181 } while (!__cmpxchg_double_slab(s, page,
2190 * Slow path. The lockless freelist is empty or we need to perform
2193 * Processing is still very fast if new objects have been freed to the
2194 * regular freelist. In that case we simply take over the regular freelist
2195 * as the lockless freelist and zap the regular freelist.
2197 * If that is not working then we fall back to the partial lists. We take the
2198 * first element of the freelist as the object to allocate now and move the
2199 * rest of the freelist to the lockless freelist.
2201 * And if we were unable to get a new slab from the partial slab lists then
2202 * we need to allocate a new slab. This is the slowest path since it involves
2203 * a call to the page allocator and the setup of a new slab.
2205 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2206 unsigned long addr, struct kmem_cache_cpu *c)
2210 unsigned long flags;
2212 local_irq_save(flags);
2213 #ifdef CONFIG_PREEMPT
2215 * We may have been preempted and rescheduled on a different
2216 * cpu before disabling interrupts. Need to reload cpu area
2219 c = this_cpu_ptr(s->cpu_slab);
2227 if (unlikely(!node_match(page, node))) {
2228 stat(s, ALLOC_NODE_MISMATCH);
2229 deactivate_slab(s, page, c->freelist);
2236 * By rights, we should be searching for a slab page that was
2237 * PFMEMALLOC but right now, we are losing the pfmemalloc
2238 * information when the page leaves the per-cpu allocator
2240 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2241 deactivate_slab(s, page, c->freelist);
2247 /* must check again c->freelist in case of cpu migration or IRQ */
2248 freelist = c->freelist;
2252 stat(s, ALLOC_SLOWPATH);
2254 freelist = get_freelist(s, page);
2258 stat(s, DEACTIVATE_BYPASS);
2262 stat(s, ALLOC_REFILL);
2266 * freelist is pointing to the list of objects to be used.
2267 * page is pointing to the page from which the objects are obtained.
2268 * That page must be frozen for per cpu allocations to work.
2270 VM_BUG_ON(!c->page->frozen);
2271 c->freelist = get_freepointer(s, freelist);
2272 c->tid = next_tid(c->tid);
2273 local_irq_restore(flags);
2279 page = c->page = c->partial;
2280 c->partial = page->next;
2281 stat(s, CPU_PARTIAL_ALLOC);
2286 freelist = new_slab_objects(s, gfpflags, node, &c);
2288 if (unlikely(!freelist)) {
2289 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2290 slab_out_of_memory(s, gfpflags, node);
2292 local_irq_restore(flags);
2297 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2300 /* Only entered in the debug case */
2301 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2302 goto new_slab; /* Slab failed checks. Next slab needed */
2304 deactivate_slab(s, page, get_freepointer(s, freelist));
2307 local_irq_restore(flags);
2312 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2313 * have the fastpath folded into their functions. So no function call
2314 * overhead for requests that can be satisfied on the fastpath.
2316 * The fastpath works by first checking if the lockless freelist can be used.
2317 * If not then __slab_alloc is called for slow processing.
2319 * Otherwise we can simply pick the next object from the lockless free list.
2321 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2322 gfp_t gfpflags, int node, unsigned long addr)
2325 struct kmem_cache_cpu *c;
2329 if (slab_pre_alloc_hook(s, gfpflags))
2335 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2336 * enabled. We may switch back and forth between cpus while
2337 * reading from one cpu area. That does not matter as long
2338 * as we end up on the original cpu again when doing the cmpxchg.
2340 c = __this_cpu_ptr(s->cpu_slab);
2343 * The transaction ids are globally unique per cpu and per operation on
2344 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2345 * occurs on the right processor and that there was no operation on the
2346 * linked list in between.
2351 object = c->freelist;
2353 if (unlikely(!object || !node_match(page, node)))
2354 object = __slab_alloc(s, gfpflags, node, addr, c);
2357 void *next_object = get_freepointer_safe(s, object);
2360 * The cmpxchg will only match if there was no additional
2361 * operation and if we are on the right processor.
2363 * The cmpxchg does the following atomically (without lock semantics!)
2364 * 1. Relocate first pointer to the current per cpu area.
2365 * 2. Verify that tid and freelist have not been changed
2366 * 3. If they were not changed replace tid and freelist
2368 * Since this is without lock semantics the protection is only against
2369 * code executing on this cpu *not* from access by other cpus.
2371 if (unlikely(!this_cpu_cmpxchg_double(
2372 s->cpu_slab->freelist, s->cpu_slab->tid,
2374 next_object, next_tid(tid)))) {
2376 note_cmpxchg_failure("slab_alloc", s, tid);
2379 prefetch_freepointer(s, next_object);
2380 stat(s, ALLOC_FASTPATH);
2383 if (unlikely(gfpflags & __GFP_ZERO) && object)
2384 memset(object, 0, s->object_size);
2386 slab_post_alloc_hook(s, gfpflags, object);
2391 static __always_inline void *slab_alloc(struct kmem_cache *s,
2392 gfp_t gfpflags, unsigned long addr)
2394 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2397 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2399 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2401 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2405 EXPORT_SYMBOL(kmem_cache_alloc);
2407 #ifdef CONFIG_TRACING
2408 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2410 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2411 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2414 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2416 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2418 void *ret = kmalloc_order(size, flags, order);
2419 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2422 EXPORT_SYMBOL(kmalloc_order_trace);
2426 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2428 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2430 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2431 s->object_size, s->size, gfpflags, node);
2435 EXPORT_SYMBOL(kmem_cache_alloc_node);
2437 #ifdef CONFIG_TRACING
2438 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2440 int node, size_t size)
2442 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2444 trace_kmalloc_node(_RET_IP_, ret,
2445 size, s->size, gfpflags, node);
2448 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2453 * Slow patch handling. This may still be called frequently since objects
2454 * have a longer lifetime than the cpu slabs in most processing loads.
2456 * So we still attempt to reduce cache line usage. Just take the slab
2457 * lock and free the item. If there is no additional partial page
2458 * handling required then we can return immediately.
2460 static void __slab_free(struct kmem_cache *s, struct page *page,
2461 void *x, unsigned long addr)
2464 void **object = (void *)x;
2468 unsigned long counters;
2469 struct kmem_cache_node *n = NULL;
2470 unsigned long uninitialized_var(flags);
2472 stat(s, FREE_SLOWPATH);
2474 if (kmem_cache_debug(s) &&
2475 !(n = free_debug_processing(s, page, x, addr, &flags)))
2479 prior = page->freelist;
2480 counters = page->counters;
2481 set_freepointer(s, object, prior);
2482 new.counters = counters;
2483 was_frozen = new.frozen;
2485 if ((!new.inuse || !prior) && !was_frozen && !n) {
2487 if (!kmem_cache_debug(s) && !prior)
2490 * Slab was on no list before and will be partially empty
2491 * We can defer the list move and instead freeze it.
2495 else { /* Needs to be taken off a list */
2497 n = get_node(s, page_to_nid(page));
2499 * Speculatively acquire the list_lock.
2500 * If the cmpxchg does not succeed then we may
2501 * drop the list_lock without any processing.
2503 * Otherwise the list_lock will synchronize with
2504 * other processors updating the list of slabs.
2506 spin_lock_irqsave(&n->list_lock, flags);
2512 } while (!cmpxchg_double_slab(s, page,
2514 object, new.counters,
2520 * If we just froze the page then put it onto the
2521 * per cpu partial list.
2523 if (new.frozen && !was_frozen) {
2524 put_cpu_partial(s, page, 1);
2525 stat(s, CPU_PARTIAL_FREE);
2528 * The list lock was not taken therefore no list
2529 * activity can be necessary.
2532 stat(s, FREE_FROZEN);
2537 * was_frozen may have been set after we acquired the list_lock in
2538 * an earlier loop. So we need to check it here again.
2541 stat(s, FREE_FROZEN);
2543 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2547 * Objects left in the slab. If it was not on the partial list before
2550 if (unlikely(!prior)) {
2551 remove_full(s, page);
2552 add_partial(n, page, DEACTIVATE_TO_TAIL);
2553 stat(s, FREE_ADD_PARTIAL);
2556 spin_unlock_irqrestore(&n->list_lock, flags);
2562 * Slab on the partial list.
2564 remove_partial(n, page);
2565 stat(s, FREE_REMOVE_PARTIAL);
2567 /* Slab must be on the full list */
2568 remove_full(s, page);
2570 spin_unlock_irqrestore(&n->list_lock, flags);
2572 discard_slab(s, page);
2576 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2577 * can perform fastpath freeing without additional function calls.
2579 * The fastpath is only possible if we are freeing to the current cpu slab
2580 * of this processor. This typically the case if we have just allocated
2583 * If fastpath is not possible then fall back to __slab_free where we deal
2584 * with all sorts of special processing.
2586 static __always_inline void slab_free(struct kmem_cache *s,
2587 struct page *page, void *x, unsigned long addr)
2589 void **object = (void *)x;
2590 struct kmem_cache_cpu *c;
2593 slab_free_hook(s, x);
2597 * Determine the currently cpus per cpu slab.
2598 * The cpu may change afterward. However that does not matter since
2599 * data is retrieved via this pointer. If we are on the same cpu
2600 * during the cmpxchg then the free will succedd.
2602 c = __this_cpu_ptr(s->cpu_slab);
2607 if (likely(page == c->page)) {
2608 set_freepointer(s, object, c->freelist);
2610 if (unlikely(!this_cpu_cmpxchg_double(
2611 s->cpu_slab->freelist, s->cpu_slab->tid,
2613 object, next_tid(tid)))) {
2615 note_cmpxchg_failure("slab_free", s, tid);
2618 stat(s, FREE_FASTPATH);
2620 __slab_free(s, page, x, addr);
2624 void kmem_cache_free(struct kmem_cache *s, void *x)
2628 page = virt_to_head_page(x);
2630 slab_free(s, page, x, _RET_IP_);
2632 trace_kmem_cache_free(_RET_IP_, x);
2634 EXPORT_SYMBOL(kmem_cache_free);
2637 * Object placement in a slab is made very easy because we always start at
2638 * offset 0. If we tune the size of the object to the alignment then we can
2639 * get the required alignment by putting one properly sized object after
2642 * Notice that the allocation order determines the sizes of the per cpu
2643 * caches. Each processor has always one slab available for allocations.
2644 * Increasing the allocation order reduces the number of times that slabs
2645 * must be moved on and off the partial lists and is therefore a factor in
2650 * Mininum / Maximum order of slab pages. This influences locking overhead
2651 * and slab fragmentation. A higher order reduces the number of partial slabs
2652 * and increases the number of allocations possible without having to
2653 * take the list_lock.
2655 static int slub_min_order;
2656 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2657 static int slub_min_objects;
2660 * Merge control. If this is set then no merging of slab caches will occur.
2661 * (Could be removed. This was introduced to pacify the merge skeptics.)
2663 static int slub_nomerge;
2666 * Calculate the order of allocation given an slab object size.
2668 * The order of allocation has significant impact on performance and other
2669 * system components. Generally order 0 allocations should be preferred since
2670 * order 0 does not cause fragmentation in the page allocator. Larger objects
2671 * be problematic to put into order 0 slabs because there may be too much
2672 * unused space left. We go to a higher order if more than 1/16th of the slab
2675 * In order to reach satisfactory performance we must ensure that a minimum
2676 * number of objects is in one slab. Otherwise we may generate too much
2677 * activity on the partial lists which requires taking the list_lock. This is
2678 * less a concern for large slabs though which are rarely used.
2680 * slub_max_order specifies the order where we begin to stop considering the
2681 * number of objects in a slab as critical. If we reach slub_max_order then
2682 * we try to keep the page order as low as possible. So we accept more waste
2683 * of space in favor of a small page order.
2685 * Higher order allocations also allow the placement of more objects in a
2686 * slab and thereby reduce object handling overhead. If the user has
2687 * requested a higher mininum order then we start with that one instead of
2688 * the smallest order which will fit the object.
2690 static inline int slab_order(int size, int min_objects,
2691 int max_order, int fract_leftover, int reserved)
2695 int min_order = slub_min_order;
2697 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2698 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2700 for (order = max(min_order,
2701 fls(min_objects * size - 1) - PAGE_SHIFT);
2702 order <= max_order; order++) {
2704 unsigned long slab_size = PAGE_SIZE << order;
2706 if (slab_size < min_objects * size + reserved)
2709 rem = (slab_size - reserved) % size;
2711 if (rem <= slab_size / fract_leftover)
2719 static inline int calculate_order(int size, int reserved)
2727 * Attempt to find best configuration for a slab. This
2728 * works by first attempting to generate a layout with
2729 * the best configuration and backing off gradually.
2731 * First we reduce the acceptable waste in a slab. Then
2732 * we reduce the minimum objects required in a slab.
2734 min_objects = slub_min_objects;
2736 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2737 max_objects = order_objects(slub_max_order, size, reserved);
2738 min_objects = min(min_objects, max_objects);
2740 while (min_objects > 1) {
2742 while (fraction >= 4) {
2743 order = slab_order(size, min_objects,
2744 slub_max_order, fraction, reserved);
2745 if (order <= slub_max_order)
2753 * We were unable to place multiple objects in a slab. Now
2754 * lets see if we can place a single object there.
2756 order = slab_order(size, 1, slub_max_order, 1, reserved);
2757 if (order <= slub_max_order)
2761 * Doh this slab cannot be placed using slub_max_order.
2763 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2764 if (order < MAX_ORDER)
2770 * Figure out what the alignment of the objects will be.
2772 static unsigned long calculate_alignment(unsigned long flags,
2773 unsigned long align, unsigned long size)
2776 * If the user wants hardware cache aligned objects then follow that
2777 * suggestion if the object is sufficiently large.
2779 * The hardware cache alignment cannot override the specified
2780 * alignment though. If that is greater then use it.
2782 if (flags & SLAB_HWCACHE_ALIGN) {
2783 unsigned long ralign = cache_line_size();
2784 while (size <= ralign / 2)
2786 align = max(align, ralign);
2789 if (align < ARCH_SLAB_MINALIGN)
2790 align = ARCH_SLAB_MINALIGN;
2792 return ALIGN(align, sizeof(void *));
2796 init_kmem_cache_node(struct kmem_cache_node *n)
2799 spin_lock_init(&n->list_lock);
2800 INIT_LIST_HEAD(&n->partial);
2801 #ifdef CONFIG_SLUB_DEBUG
2802 atomic_long_set(&n->nr_slabs, 0);
2803 atomic_long_set(&n->total_objects, 0);
2804 INIT_LIST_HEAD(&n->full);
2808 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2810 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2811 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2814 * Must align to double word boundary for the double cmpxchg
2815 * instructions to work; see __pcpu_double_call_return_bool().
2817 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2818 2 * sizeof(void *));
2823 init_kmem_cache_cpus(s);
2828 static struct kmem_cache *kmem_cache_node;
2831 * No kmalloc_node yet so do it by hand. We know that this is the first
2832 * slab on the node for this slabcache. There are no concurrent accesses
2835 * Note that this function only works on the kmalloc_node_cache
2836 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2837 * memory on a fresh node that has no slab structures yet.
2839 static void early_kmem_cache_node_alloc(int node)
2842 struct kmem_cache_node *n;
2844 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2846 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2849 if (page_to_nid(page) != node) {
2850 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2852 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2853 "in order to be able to continue\n");
2858 page->freelist = get_freepointer(kmem_cache_node, n);
2861 kmem_cache_node->node[node] = n;
2862 #ifdef CONFIG_SLUB_DEBUG
2863 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2864 init_tracking(kmem_cache_node, n);
2866 init_kmem_cache_node(n);
2867 inc_slabs_node(kmem_cache_node, node, page->objects);
2869 add_partial(n, page, DEACTIVATE_TO_HEAD);
2872 static void free_kmem_cache_nodes(struct kmem_cache *s)
2876 for_each_node_state(node, N_NORMAL_MEMORY) {
2877 struct kmem_cache_node *n = s->node[node];
2880 kmem_cache_free(kmem_cache_node, n);
2882 s->node[node] = NULL;
2886 static int init_kmem_cache_nodes(struct kmem_cache *s)
2890 for_each_node_state(node, N_NORMAL_MEMORY) {
2891 struct kmem_cache_node *n;
2893 if (slab_state == DOWN) {
2894 early_kmem_cache_node_alloc(node);
2897 n = kmem_cache_alloc_node(kmem_cache_node,
2901 free_kmem_cache_nodes(s);
2906 init_kmem_cache_node(n);
2911 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2913 if (min < MIN_PARTIAL)
2915 else if (min > MAX_PARTIAL)
2917 s->min_partial = min;
2921 * calculate_sizes() determines the order and the distribution of data within
2924 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2926 unsigned long flags = s->flags;
2927 unsigned long size = s->object_size;
2928 unsigned long align = s->align;
2932 * Round up object size to the next word boundary. We can only
2933 * place the free pointer at word boundaries and this determines
2934 * the possible location of the free pointer.
2936 size = ALIGN(size, sizeof(void *));
2938 #ifdef CONFIG_SLUB_DEBUG
2940 * Determine if we can poison the object itself. If the user of
2941 * the slab may touch the object after free or before allocation
2942 * then we should never poison the object itself.
2944 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2946 s->flags |= __OBJECT_POISON;
2948 s->flags &= ~__OBJECT_POISON;
2952 * If we are Redzoning then check if there is some space between the
2953 * end of the object and the free pointer. If not then add an
2954 * additional word to have some bytes to store Redzone information.
2956 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2957 size += sizeof(void *);
2961 * With that we have determined the number of bytes in actual use
2962 * by the object. This is the potential offset to the free pointer.
2966 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2969 * Relocate free pointer after the object if it is not
2970 * permitted to overwrite the first word of the object on
2973 * This is the case if we do RCU, have a constructor or
2974 * destructor or are poisoning the objects.
2977 size += sizeof(void *);
2980 #ifdef CONFIG_SLUB_DEBUG
2981 if (flags & SLAB_STORE_USER)
2983 * Need to store information about allocs and frees after
2986 size += 2 * sizeof(struct track);
2988 if (flags & SLAB_RED_ZONE)
2990 * Add some empty padding so that we can catch
2991 * overwrites from earlier objects rather than let
2992 * tracking information or the free pointer be
2993 * corrupted if a user writes before the start
2996 size += sizeof(void *);
3000 * Determine the alignment based on various parameters that the
3001 * user specified and the dynamic determination of cache line size
3004 align = calculate_alignment(flags, align, s->object_size);
3008 * SLUB stores one object immediately after another beginning from
3009 * offset 0. In order to align the objects we have to simply size
3010 * each object to conform to the alignment.
3012 size = ALIGN(size, align);
3014 if (forced_order >= 0)
3015 order = forced_order;
3017 order = calculate_order(size, s->reserved);
3024 s->allocflags |= __GFP_COMP;
3026 if (s->flags & SLAB_CACHE_DMA)
3027 s->allocflags |= SLUB_DMA;
3029 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3030 s->allocflags |= __GFP_RECLAIMABLE;
3033 * Determine the number of objects per slab
3035 s->oo = oo_make(order, size, s->reserved);
3036 s->min = oo_make(get_order(size), size, s->reserved);
3037 if (oo_objects(s->oo) > oo_objects(s->max))
3040 return !!oo_objects(s->oo);
3044 static int kmem_cache_open(struct kmem_cache *s,
3045 const char *name, size_t size,
3046 size_t align, unsigned long flags,
3047 void (*ctor)(void *))
3049 memset(s, 0, kmem_size);
3052 s->object_size = size;
3054 s->flags = kmem_cache_flags(size, flags, name, ctor);
3057 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3058 s->reserved = sizeof(struct rcu_head);
3060 if (!calculate_sizes(s, -1))
3062 if (disable_higher_order_debug) {
3064 * Disable debugging flags that store metadata if the min slab
3067 if (get_order(s->size) > get_order(s->object_size)) {
3068 s->flags &= ~DEBUG_METADATA_FLAGS;
3070 if (!calculate_sizes(s, -1))
3075 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3076 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3077 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3078 /* Enable fast mode */
3079 s->flags |= __CMPXCHG_DOUBLE;
3083 * The larger the object size is, the more pages we want on the partial
3084 * list to avoid pounding the page allocator excessively.
3086 set_min_partial(s, ilog2(s->size) / 2);
3089 * cpu_partial determined the maximum number of objects kept in the
3090 * per cpu partial lists of a processor.
3092 * Per cpu partial lists mainly contain slabs that just have one
3093 * object freed. If they are used for allocation then they can be
3094 * filled up again with minimal effort. The slab will never hit the
3095 * per node partial lists and therefore no locking will be required.
3097 * This setting also determines
3099 * A) The number of objects from per cpu partial slabs dumped to the
3100 * per node list when we reach the limit.
3101 * B) The number of objects in cpu partial slabs to extract from the
3102 * per node list when we run out of per cpu objects. We only fetch 50%
3103 * to keep some capacity around for frees.
3105 if (kmem_cache_debug(s))
3107 else if (s->size >= PAGE_SIZE)
3109 else if (s->size >= 1024)
3111 else if (s->size >= 256)
3112 s->cpu_partial = 13;
3114 s->cpu_partial = 30;
3118 s->remote_node_defrag_ratio = 1000;
3120 if (!init_kmem_cache_nodes(s))
3123 if (alloc_kmem_cache_cpus(s))
3126 free_kmem_cache_nodes(s);
3128 if (flags & SLAB_PANIC)
3129 panic("Cannot create slab %s size=%lu realsize=%u "
3130 "order=%u offset=%u flags=%lx\n",
3131 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3137 * Determine the size of a slab object
3139 unsigned int kmem_cache_size(struct kmem_cache *s)
3141 return s->object_size;
3143 EXPORT_SYMBOL(kmem_cache_size);
3145 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3148 #ifdef CONFIG_SLUB_DEBUG
3149 void *addr = page_address(page);
3151 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3152 sizeof(long), GFP_ATOMIC);
3155 slab_err(s, page, "%s", text);
3158 get_map(s, page, map);
3159 for_each_object(p, s, addr, page->objects) {
3161 if (!test_bit(slab_index(p, s, addr), map)) {
3162 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3164 print_tracking(s, p);
3173 * Attempt to free all partial slabs on a node.
3174 * This is called from kmem_cache_close(). We must be the last thread
3175 * using the cache and therefore we do not need to lock anymore.
3177 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3179 struct page *page, *h;
3181 list_for_each_entry_safe(page, h, &n->partial, lru) {
3183 remove_partial(n, page);
3184 discard_slab(s, page);
3186 list_slab_objects(s, page,
3187 "Objects remaining on kmem_cache_close()");
3193 * Release all resources used by a slab cache.
3195 static inline int kmem_cache_close(struct kmem_cache *s)
3200 free_percpu(s->cpu_slab);
3201 /* Attempt to free all objects */
3202 for_each_node_state(node, N_NORMAL_MEMORY) {
3203 struct kmem_cache_node *n = get_node(s, node);
3206 if (n->nr_partial || slabs_node(s, node))
3209 free_kmem_cache_nodes(s);
3214 * Close a cache and release the kmem_cache structure
3215 * (must be used for caches created using kmem_cache_create)
3217 void kmem_cache_destroy(struct kmem_cache *s)
3219 mutex_lock(&slab_mutex);
3223 mutex_unlock(&slab_mutex);
3224 if (kmem_cache_close(s)) {
3225 printk(KERN_ERR "SLUB %s: %s called for cache that "
3226 "still has objects.\n", s->name, __func__);
3229 if (s->flags & SLAB_DESTROY_BY_RCU)
3231 sysfs_slab_remove(s);
3233 mutex_unlock(&slab_mutex);
3235 EXPORT_SYMBOL(kmem_cache_destroy);
3237 /********************************************************************
3239 *******************************************************************/
3241 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3242 EXPORT_SYMBOL(kmalloc_caches);
3244 static struct kmem_cache *kmem_cache;
3246 #ifdef CONFIG_ZONE_DMA
3247 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3250 static int __init setup_slub_min_order(char *str)
3252 get_option(&str, &slub_min_order);
3257 __setup("slub_min_order=", setup_slub_min_order);
3259 static int __init setup_slub_max_order(char *str)
3261 get_option(&str, &slub_max_order);
3262 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3267 __setup("slub_max_order=", setup_slub_max_order);
3269 static int __init setup_slub_min_objects(char *str)
3271 get_option(&str, &slub_min_objects);
3276 __setup("slub_min_objects=", setup_slub_min_objects);
3278 static int __init setup_slub_nomerge(char *str)
3284 __setup("slub_nomerge", setup_slub_nomerge);
3286 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3287 int size, unsigned int flags)
3289 struct kmem_cache *s;
3291 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3294 * This function is called with IRQs disabled during early-boot on
3295 * single CPU so there's no need to take slab_mutex here.
3297 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3301 list_add(&s->list, &slab_caches);
3305 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3310 * Conversion table for small slabs sizes / 8 to the index in the
3311 * kmalloc array. This is necessary for slabs < 192 since we have non power
3312 * of two cache sizes there. The size of larger slabs can be determined using
3315 static s8 size_index[24] = {
3342 static inline int size_index_elem(size_t bytes)
3344 return (bytes - 1) / 8;
3347 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3353 return ZERO_SIZE_PTR;
3355 index = size_index[size_index_elem(size)];
3357 index = fls(size - 1);
3359 #ifdef CONFIG_ZONE_DMA
3360 if (unlikely((flags & SLUB_DMA)))
3361 return kmalloc_dma_caches[index];
3364 return kmalloc_caches[index];
3367 void *__kmalloc(size_t size, gfp_t flags)
3369 struct kmem_cache *s;
3372 if (unlikely(size > SLUB_MAX_SIZE))
3373 return kmalloc_large(size, flags);
3375 s = get_slab(size, flags);
3377 if (unlikely(ZERO_OR_NULL_PTR(s)))
3380 ret = slab_alloc(s, flags, _RET_IP_);
3382 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3386 EXPORT_SYMBOL(__kmalloc);
3389 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3394 flags |= __GFP_COMP | __GFP_NOTRACK;
3395 page = alloc_pages_node(node, flags, get_order(size));
3397 ptr = page_address(page);
3399 kmemleak_alloc(ptr, size, 1, flags);
3403 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3405 struct kmem_cache *s;
3408 if (unlikely(size > SLUB_MAX_SIZE)) {
3409 ret = kmalloc_large_node(size, flags, node);
3411 trace_kmalloc_node(_RET_IP_, ret,
3412 size, PAGE_SIZE << get_order(size),
3418 s = get_slab(size, flags);
3420 if (unlikely(ZERO_OR_NULL_PTR(s)))
3423 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3425 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3429 EXPORT_SYMBOL(__kmalloc_node);
3432 size_t ksize(const void *object)
3436 if (unlikely(object == ZERO_SIZE_PTR))
3439 page = virt_to_head_page(object);
3441 if (unlikely(!PageSlab(page))) {
3442 WARN_ON(!PageCompound(page));
3443 return PAGE_SIZE << compound_order(page);
3446 return slab_ksize(page->slab);
3448 EXPORT_SYMBOL(ksize);
3450 #ifdef CONFIG_SLUB_DEBUG
3451 bool verify_mem_not_deleted(const void *x)
3454 void *object = (void *)x;
3455 unsigned long flags;
3458 if (unlikely(ZERO_OR_NULL_PTR(x)))
3461 local_irq_save(flags);
3463 page = virt_to_head_page(x);
3464 if (unlikely(!PageSlab(page))) {
3465 /* maybe it was from stack? */
3471 if (on_freelist(page->slab, page, object)) {
3472 object_err(page->slab, page, object, "Object is on free-list");
3480 local_irq_restore(flags);
3483 EXPORT_SYMBOL(verify_mem_not_deleted);
3486 void kfree(const void *x)
3489 void *object = (void *)x;
3491 trace_kfree(_RET_IP_, x);
3493 if (unlikely(ZERO_OR_NULL_PTR(x)))
3496 page = virt_to_head_page(x);
3497 if (unlikely(!PageSlab(page))) {
3498 BUG_ON(!PageCompound(page));
3500 __free_pages(page, compound_order(page));
3503 slab_free(page->slab, page, object, _RET_IP_);
3505 EXPORT_SYMBOL(kfree);
3508 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3509 * the remaining slabs by the number of items in use. The slabs with the
3510 * most items in use come first. New allocations will then fill those up
3511 * and thus they can be removed from the partial lists.
3513 * The slabs with the least items are placed last. This results in them
3514 * being allocated from last increasing the chance that the last objects
3515 * are freed in them.
3517 int kmem_cache_shrink(struct kmem_cache *s)
3521 struct kmem_cache_node *n;
3524 int objects = oo_objects(s->max);
3525 struct list_head *slabs_by_inuse =
3526 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3527 unsigned long flags;
3529 if (!slabs_by_inuse)
3533 for_each_node_state(node, N_NORMAL_MEMORY) {
3534 n = get_node(s, node);
3539 for (i = 0; i < objects; i++)
3540 INIT_LIST_HEAD(slabs_by_inuse + i);
3542 spin_lock_irqsave(&n->list_lock, flags);
3545 * Build lists indexed by the items in use in each slab.
3547 * Note that concurrent frees may occur while we hold the
3548 * list_lock. page->inuse here is the upper limit.
3550 list_for_each_entry_safe(page, t, &n->partial, lru) {
3551 list_move(&page->lru, slabs_by_inuse + page->inuse);
3557 * Rebuild the partial list with the slabs filled up most
3558 * first and the least used slabs at the end.
3560 for (i = objects - 1; i > 0; i--)
3561 list_splice(slabs_by_inuse + i, n->partial.prev);
3563 spin_unlock_irqrestore(&n->list_lock, flags);
3565 /* Release empty slabs */
3566 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3567 discard_slab(s, page);
3570 kfree(slabs_by_inuse);
3573 EXPORT_SYMBOL(kmem_cache_shrink);
3575 #if defined(CONFIG_MEMORY_HOTPLUG)
3576 static int slab_mem_going_offline_callback(void *arg)
3578 struct kmem_cache *s;
3580 mutex_lock(&slab_mutex);
3581 list_for_each_entry(s, &slab_caches, list)
3582 kmem_cache_shrink(s);
3583 mutex_unlock(&slab_mutex);
3588 static void slab_mem_offline_callback(void *arg)
3590 struct kmem_cache_node *n;
3591 struct kmem_cache *s;
3592 struct memory_notify *marg = arg;
3595 offline_node = marg->status_change_nid;
3598 * If the node still has available memory. we need kmem_cache_node
3601 if (offline_node < 0)
3604 mutex_lock(&slab_mutex);
3605 list_for_each_entry(s, &slab_caches, list) {
3606 n = get_node(s, offline_node);
3609 * if n->nr_slabs > 0, slabs still exist on the node
3610 * that is going down. We were unable to free them,
3611 * and offline_pages() function shouldn't call this
3612 * callback. So, we must fail.
3614 BUG_ON(slabs_node(s, offline_node));
3616 s->node[offline_node] = NULL;
3617 kmem_cache_free(kmem_cache_node, n);
3620 mutex_unlock(&slab_mutex);
3623 static int slab_mem_going_online_callback(void *arg)
3625 struct kmem_cache_node *n;
3626 struct kmem_cache *s;
3627 struct memory_notify *marg = arg;
3628 int nid = marg->status_change_nid;
3632 * If the node's memory is already available, then kmem_cache_node is
3633 * already created. Nothing to do.
3639 * We are bringing a node online. No memory is available yet. We must
3640 * allocate a kmem_cache_node structure in order to bring the node
3643 mutex_lock(&slab_mutex);
3644 list_for_each_entry(s, &slab_caches, list) {
3646 * XXX: kmem_cache_alloc_node will fallback to other nodes
3647 * since memory is not yet available from the node that
3650 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3655 init_kmem_cache_node(n);
3659 mutex_unlock(&slab_mutex);
3663 static int slab_memory_callback(struct notifier_block *self,
3664 unsigned long action, void *arg)
3669 case MEM_GOING_ONLINE:
3670 ret = slab_mem_going_online_callback(arg);
3672 case MEM_GOING_OFFLINE:
3673 ret = slab_mem_going_offline_callback(arg);
3676 case MEM_CANCEL_ONLINE:
3677 slab_mem_offline_callback(arg);
3680 case MEM_CANCEL_OFFLINE:
3684 ret = notifier_from_errno(ret);
3690 #endif /* CONFIG_MEMORY_HOTPLUG */
3692 /********************************************************************
3693 * Basic setup of slabs
3694 *******************************************************************/
3697 * Used for early kmem_cache structures that were allocated using
3698 * the page allocator
3701 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3705 list_add(&s->list, &slab_caches);
3708 for_each_node_state(node, N_NORMAL_MEMORY) {
3709 struct kmem_cache_node *n = get_node(s, node);
3713 list_for_each_entry(p, &n->partial, lru)
3716 #ifdef CONFIG_SLUB_DEBUG
3717 list_for_each_entry(p, &n->full, lru)
3724 void __init kmem_cache_init(void)
3728 struct kmem_cache *temp_kmem_cache;
3730 struct kmem_cache *temp_kmem_cache_node;
3731 unsigned long kmalloc_size;
3733 if (debug_guardpage_minorder())
3736 kmem_size = offsetof(struct kmem_cache, node) +
3737 nr_node_ids * sizeof(struct kmem_cache_node *);
3739 /* Allocate two kmem_caches from the page allocator */
3740 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3741 order = get_order(2 * kmalloc_size);
3742 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3745 * Must first have the slab cache available for the allocations of the
3746 * struct kmem_cache_node's. There is special bootstrap code in
3747 * kmem_cache_open for slab_state == DOWN.
3749 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3751 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3752 sizeof(struct kmem_cache_node),
3753 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3755 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3757 /* Able to allocate the per node structures */
3758 slab_state = PARTIAL;
3760 temp_kmem_cache = kmem_cache;
3761 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3762 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3763 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3764 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3767 * Allocate kmem_cache_node properly from the kmem_cache slab.
3768 * kmem_cache_node is separately allocated so no need to
3769 * update any list pointers.
3771 temp_kmem_cache_node = kmem_cache_node;
3773 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3774 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3776 kmem_cache_bootstrap_fixup(kmem_cache_node);
3779 kmem_cache_bootstrap_fixup(kmem_cache);
3781 /* Free temporary boot structure */
3782 free_pages((unsigned long)temp_kmem_cache, order);
3784 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3787 * Patch up the size_index table if we have strange large alignment
3788 * requirements for the kmalloc array. This is only the case for
3789 * MIPS it seems. The standard arches will not generate any code here.
3791 * Largest permitted alignment is 256 bytes due to the way we
3792 * handle the index determination for the smaller caches.
3794 * Make sure that nothing crazy happens if someone starts tinkering
3795 * around with ARCH_KMALLOC_MINALIGN
3797 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3798 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3800 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3801 int elem = size_index_elem(i);
3802 if (elem >= ARRAY_SIZE(size_index))
3804 size_index[elem] = KMALLOC_SHIFT_LOW;
3807 if (KMALLOC_MIN_SIZE == 64) {
3809 * The 96 byte size cache is not used if the alignment
3812 for (i = 64 + 8; i <= 96; i += 8)
3813 size_index[size_index_elem(i)] = 7;
3814 } else if (KMALLOC_MIN_SIZE == 128) {
3816 * The 192 byte sized cache is not used if the alignment
3817 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3820 for (i = 128 + 8; i <= 192; i += 8)
3821 size_index[size_index_elem(i)] = 8;
3824 /* Caches that are not of the two-to-the-power-of size */
3825 if (KMALLOC_MIN_SIZE <= 32) {
3826 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3830 if (KMALLOC_MIN_SIZE <= 64) {
3831 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3835 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3836 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3842 /* Provide the correct kmalloc names now that the caches are up */
3843 if (KMALLOC_MIN_SIZE <= 32) {
3844 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3845 BUG_ON(!kmalloc_caches[1]->name);
3848 if (KMALLOC_MIN_SIZE <= 64) {
3849 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3850 BUG_ON(!kmalloc_caches[2]->name);
3853 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3854 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3857 kmalloc_caches[i]->name = s;
3861 register_cpu_notifier(&slab_notifier);
3864 #ifdef CONFIG_ZONE_DMA
3865 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3866 struct kmem_cache *s = kmalloc_caches[i];
3869 char *name = kasprintf(GFP_NOWAIT,
3870 "dma-kmalloc-%d", s->object_size);
3873 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3874 s->object_size, SLAB_CACHE_DMA);
3879 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3880 " CPUs=%d, Nodes=%d\n",
3881 caches, cache_line_size(),
3882 slub_min_order, slub_max_order, slub_min_objects,
3883 nr_cpu_ids, nr_node_ids);
3886 void __init kmem_cache_init_late(void)
3891 * Find a mergeable slab cache
3893 static int slab_unmergeable(struct kmem_cache *s)
3895 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3902 * We may have set a slab to be unmergeable during bootstrap.
3904 if (s->refcount < 0)
3910 static struct kmem_cache *find_mergeable(size_t size,
3911 size_t align, unsigned long flags, const char *name,
3912 void (*ctor)(void *))
3914 struct kmem_cache *s;
3916 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3922 size = ALIGN(size, sizeof(void *));
3923 align = calculate_alignment(flags, align, size);
3924 size = ALIGN(size, align);
3925 flags = kmem_cache_flags(size, flags, name, NULL);
3927 list_for_each_entry(s, &slab_caches, list) {
3928 if (slab_unmergeable(s))
3934 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3937 * Check if alignment is compatible.
3938 * Courtesy of Adrian Drzewiecki
3940 if ((s->size & ~(align - 1)) != s->size)
3943 if (s->size - size >= sizeof(void *))
3951 struct kmem_cache *__kmem_cache_create(const char *name, size_t size,
3952 size_t align, unsigned long flags, void (*ctor)(void *))
3954 struct kmem_cache *s;
3957 s = find_mergeable(size, align, flags, name, ctor);
3961 * Adjust the object sizes so that we clear
3962 * the complete object on kzalloc.
3964 s->object_size = max(s->object_size, (int)size);
3965 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3967 if (sysfs_slab_alias(s, name)) {
3974 n = kstrdup(name, GFP_KERNEL);
3978 s = kmalloc(kmem_size, GFP_KERNEL);
3980 if (kmem_cache_open(s, n,
3981 size, align, flags, ctor)) {
3984 list_add(&s->list, &slab_caches);
3985 mutex_unlock(&slab_mutex);
3986 r = sysfs_slab_add(s);
3987 mutex_lock(&slab_mutex);
3993 kmem_cache_close(s);
4003 * Use the cpu notifier to insure that the cpu slabs are flushed when
4006 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
4007 unsigned long action, void *hcpu)
4009 long cpu = (long)hcpu;
4010 struct kmem_cache *s;
4011 unsigned long flags;
4014 case CPU_UP_CANCELED:
4015 case CPU_UP_CANCELED_FROZEN:
4017 case CPU_DEAD_FROZEN:
4018 mutex_lock(&slab_mutex);
4019 list_for_each_entry(s, &slab_caches, list) {
4020 local_irq_save(flags);
4021 __flush_cpu_slab(s, cpu);
4022 local_irq_restore(flags);
4024 mutex_unlock(&slab_mutex);
4032 static struct notifier_block __cpuinitdata slab_notifier = {
4033 .notifier_call = slab_cpuup_callback
4038 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4040 struct kmem_cache *s;
4043 if (unlikely(size > SLUB_MAX_SIZE))
4044 return kmalloc_large(size, gfpflags);
4046 s = get_slab(size, gfpflags);
4048 if (unlikely(ZERO_OR_NULL_PTR(s)))
4051 ret = slab_alloc(s, gfpflags, caller);
4053 /* Honor the call site pointer we received. */
4054 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4060 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4061 int node, unsigned long caller)
4063 struct kmem_cache *s;
4066 if (unlikely(size > SLUB_MAX_SIZE)) {
4067 ret = kmalloc_large_node(size, gfpflags, node);
4069 trace_kmalloc_node(caller, ret,
4070 size, PAGE_SIZE << get_order(size),
4076 s = get_slab(size, gfpflags);
4078 if (unlikely(ZERO_OR_NULL_PTR(s)))
4081 ret = slab_alloc_node(s, gfpflags, node, caller);
4083 /* Honor the call site pointer we received. */
4084 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4091 static int count_inuse(struct page *page)
4096 static int count_total(struct page *page)
4098 return page->objects;
4102 #ifdef CONFIG_SLUB_DEBUG
4103 static int validate_slab(struct kmem_cache *s, struct page *page,
4107 void *addr = page_address(page);
4109 if (!check_slab(s, page) ||
4110 !on_freelist(s, page, NULL))
4113 /* Now we know that a valid freelist exists */
4114 bitmap_zero(map, page->objects);
4116 get_map(s, page, map);
4117 for_each_object(p, s, addr, page->objects) {
4118 if (test_bit(slab_index(p, s, addr), map))
4119 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4123 for_each_object(p, s, addr, page->objects)
4124 if (!test_bit(slab_index(p, s, addr), map))
4125 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4130 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4134 validate_slab(s, page, map);
4138 static int validate_slab_node(struct kmem_cache *s,
4139 struct kmem_cache_node *n, unsigned long *map)
4141 unsigned long count = 0;
4143 unsigned long flags;
4145 spin_lock_irqsave(&n->list_lock, flags);
4147 list_for_each_entry(page, &n->partial, lru) {
4148 validate_slab_slab(s, page, map);
4151 if (count != n->nr_partial)
4152 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4153 "counter=%ld\n", s->name, count, n->nr_partial);
4155 if (!(s->flags & SLAB_STORE_USER))
4158 list_for_each_entry(page, &n->full, lru) {
4159 validate_slab_slab(s, page, map);
4162 if (count != atomic_long_read(&n->nr_slabs))
4163 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4164 "counter=%ld\n", s->name, count,
4165 atomic_long_read(&n->nr_slabs));
4168 spin_unlock_irqrestore(&n->list_lock, flags);
4172 static long validate_slab_cache(struct kmem_cache *s)
4175 unsigned long count = 0;
4176 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4177 sizeof(unsigned long), GFP_KERNEL);
4183 for_each_node_state(node, N_NORMAL_MEMORY) {
4184 struct kmem_cache_node *n = get_node(s, node);
4186 count += validate_slab_node(s, n, map);
4192 * Generate lists of code addresses where slabcache objects are allocated
4197 unsigned long count;
4204 DECLARE_BITMAP(cpus, NR_CPUS);
4210 unsigned long count;
4211 struct location *loc;
4214 static void free_loc_track(struct loc_track *t)
4217 free_pages((unsigned long)t->loc,
4218 get_order(sizeof(struct location) * t->max));
4221 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4226 order = get_order(sizeof(struct location) * max);
4228 l = (void *)__get_free_pages(flags, order);
4233 memcpy(l, t->loc, sizeof(struct location) * t->count);
4241 static int add_location(struct loc_track *t, struct kmem_cache *s,
4242 const struct track *track)
4244 long start, end, pos;
4246 unsigned long caddr;
4247 unsigned long age = jiffies - track->when;
4253 pos = start + (end - start + 1) / 2;
4256 * There is nothing at "end". If we end up there
4257 * we need to add something to before end.
4262 caddr = t->loc[pos].addr;
4263 if (track->addr == caddr) {
4269 if (age < l->min_time)
4271 if (age > l->max_time)
4274 if (track->pid < l->min_pid)
4275 l->min_pid = track->pid;
4276 if (track->pid > l->max_pid)
4277 l->max_pid = track->pid;
4279 cpumask_set_cpu(track->cpu,
4280 to_cpumask(l->cpus));
4282 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4286 if (track->addr < caddr)
4293 * Not found. Insert new tracking element.
4295 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4301 (t->count - pos) * sizeof(struct location));
4304 l->addr = track->addr;
4308 l->min_pid = track->pid;
4309 l->max_pid = track->pid;
4310 cpumask_clear(to_cpumask(l->cpus));
4311 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4312 nodes_clear(l->nodes);
4313 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4317 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4318 struct page *page, enum track_item alloc,
4321 void *addr = page_address(page);
4324 bitmap_zero(map, page->objects);
4325 get_map(s, page, map);
4327 for_each_object(p, s, addr, page->objects)
4328 if (!test_bit(slab_index(p, s, addr), map))
4329 add_location(t, s, get_track(s, p, alloc));
4332 static int list_locations(struct kmem_cache *s, char *buf,
4333 enum track_item alloc)
4337 struct loc_track t = { 0, 0, NULL };
4339 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4340 sizeof(unsigned long), GFP_KERNEL);
4342 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4345 return sprintf(buf, "Out of memory\n");
4347 /* Push back cpu slabs */
4350 for_each_node_state(node, N_NORMAL_MEMORY) {
4351 struct kmem_cache_node *n = get_node(s, node);
4352 unsigned long flags;
4355 if (!atomic_long_read(&n->nr_slabs))
4358 spin_lock_irqsave(&n->list_lock, flags);
4359 list_for_each_entry(page, &n->partial, lru)
4360 process_slab(&t, s, page, alloc, map);
4361 list_for_each_entry(page, &n->full, lru)
4362 process_slab(&t, s, page, alloc, map);
4363 spin_unlock_irqrestore(&n->list_lock, flags);
4366 for (i = 0; i < t.count; i++) {
4367 struct location *l = &t.loc[i];
4369 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4371 len += sprintf(buf + len, "%7ld ", l->count);
4374 len += sprintf(buf + len, "%pS", (void *)l->addr);
4376 len += sprintf(buf + len, "<not-available>");
4378 if (l->sum_time != l->min_time) {
4379 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4381 (long)div_u64(l->sum_time, l->count),
4384 len += sprintf(buf + len, " age=%ld",
4387 if (l->min_pid != l->max_pid)
4388 len += sprintf(buf + len, " pid=%ld-%ld",
4389 l->min_pid, l->max_pid);
4391 len += sprintf(buf + len, " pid=%ld",
4394 if (num_online_cpus() > 1 &&
4395 !cpumask_empty(to_cpumask(l->cpus)) &&
4396 len < PAGE_SIZE - 60) {
4397 len += sprintf(buf + len, " cpus=");
4398 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4399 to_cpumask(l->cpus));
4402 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4403 len < PAGE_SIZE - 60) {
4404 len += sprintf(buf + len, " nodes=");
4405 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4409 len += sprintf(buf + len, "\n");
4415 len += sprintf(buf, "No data\n");
4420 #ifdef SLUB_RESILIENCY_TEST
4421 static void resiliency_test(void)
4425 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4427 printk(KERN_ERR "SLUB resiliency testing\n");
4428 printk(KERN_ERR "-----------------------\n");
4429 printk(KERN_ERR "A. Corruption after allocation\n");
4431 p = kzalloc(16, GFP_KERNEL);
4433 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4434 " 0x12->0x%p\n\n", p + 16);
4436 validate_slab_cache(kmalloc_caches[4]);
4438 /* Hmmm... The next two are dangerous */
4439 p = kzalloc(32, GFP_KERNEL);
4440 p[32 + sizeof(void *)] = 0x34;
4441 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4442 " 0x34 -> -0x%p\n", p);
4444 "If allocated object is overwritten then not detectable\n\n");
4446 validate_slab_cache(kmalloc_caches[5]);
4447 p = kzalloc(64, GFP_KERNEL);
4448 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4450 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4453 "If allocated object is overwritten then not detectable\n\n");
4454 validate_slab_cache(kmalloc_caches[6]);
4456 printk(KERN_ERR "\nB. Corruption after free\n");
4457 p = kzalloc(128, GFP_KERNEL);
4460 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4461 validate_slab_cache(kmalloc_caches[7]);
4463 p = kzalloc(256, GFP_KERNEL);
4466 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4468 validate_slab_cache(kmalloc_caches[8]);
4470 p = kzalloc(512, GFP_KERNEL);
4473 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4474 validate_slab_cache(kmalloc_caches[9]);
4478 static void resiliency_test(void) {};
4483 enum slab_stat_type {
4484 SL_ALL, /* All slabs */
4485 SL_PARTIAL, /* Only partially allocated slabs */
4486 SL_CPU, /* Only slabs used for cpu caches */
4487 SL_OBJECTS, /* Determine allocated objects not slabs */
4488 SL_TOTAL /* Determine object capacity not slabs */
4491 #define SO_ALL (1 << SL_ALL)
4492 #define SO_PARTIAL (1 << SL_PARTIAL)
4493 #define SO_CPU (1 << SL_CPU)
4494 #define SO_OBJECTS (1 << SL_OBJECTS)
4495 #define SO_TOTAL (1 << SL_TOTAL)
4497 static ssize_t show_slab_objects(struct kmem_cache *s,
4498 char *buf, unsigned long flags)
4500 unsigned long total = 0;
4503 unsigned long *nodes;
4504 unsigned long *per_cpu;
4506 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4509 per_cpu = nodes + nr_node_ids;
4511 if (flags & SO_CPU) {
4514 for_each_possible_cpu(cpu) {
4515 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4519 page = ACCESS_ONCE(c->page);
4523 node = page_to_nid(page);
4524 if (flags & SO_TOTAL)
4526 else if (flags & SO_OBJECTS)
4534 page = ACCESS_ONCE(c->partial);
4545 lock_memory_hotplug();
4546 #ifdef CONFIG_SLUB_DEBUG
4547 if (flags & SO_ALL) {
4548 for_each_node_state(node, N_NORMAL_MEMORY) {
4549 struct kmem_cache_node *n = get_node(s, node);
4551 if (flags & SO_TOTAL)
4552 x = atomic_long_read(&n->total_objects);
4553 else if (flags & SO_OBJECTS)
4554 x = atomic_long_read(&n->total_objects) -
4555 count_partial(n, count_free);
4558 x = atomic_long_read(&n->nr_slabs);
4565 if (flags & SO_PARTIAL) {
4566 for_each_node_state(node, N_NORMAL_MEMORY) {
4567 struct kmem_cache_node *n = get_node(s, node);
4569 if (flags & SO_TOTAL)
4570 x = count_partial(n, count_total);
4571 else if (flags & SO_OBJECTS)
4572 x = count_partial(n, count_inuse);
4579 x = sprintf(buf, "%lu", total);
4581 for_each_node_state(node, N_NORMAL_MEMORY)
4583 x += sprintf(buf + x, " N%d=%lu",
4586 unlock_memory_hotplug();
4588 return x + sprintf(buf + x, "\n");
4591 #ifdef CONFIG_SLUB_DEBUG
4592 static int any_slab_objects(struct kmem_cache *s)
4596 for_each_online_node(node) {
4597 struct kmem_cache_node *n = get_node(s, node);
4602 if (atomic_long_read(&n->total_objects))
4609 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4610 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4612 struct slab_attribute {
4613 struct attribute attr;
4614 ssize_t (*show)(struct kmem_cache *s, char *buf);
4615 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4618 #define SLAB_ATTR_RO(_name) \
4619 static struct slab_attribute _name##_attr = \
4620 __ATTR(_name, 0400, _name##_show, NULL)
4622 #define SLAB_ATTR(_name) \
4623 static struct slab_attribute _name##_attr = \
4624 __ATTR(_name, 0600, _name##_show, _name##_store)
4626 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4628 return sprintf(buf, "%d\n", s->size);
4630 SLAB_ATTR_RO(slab_size);
4632 static ssize_t align_show(struct kmem_cache *s, char *buf)
4634 return sprintf(buf, "%d\n", s->align);
4636 SLAB_ATTR_RO(align);
4638 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4640 return sprintf(buf, "%d\n", s->object_size);
4642 SLAB_ATTR_RO(object_size);
4644 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4646 return sprintf(buf, "%d\n", oo_objects(s->oo));
4648 SLAB_ATTR_RO(objs_per_slab);
4650 static ssize_t order_store(struct kmem_cache *s,
4651 const char *buf, size_t length)
4653 unsigned long order;
4656 err = strict_strtoul(buf, 10, &order);
4660 if (order > slub_max_order || order < slub_min_order)
4663 calculate_sizes(s, order);
4667 static ssize_t order_show(struct kmem_cache *s, char *buf)
4669 return sprintf(buf, "%d\n", oo_order(s->oo));
4673 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4675 return sprintf(buf, "%lu\n", s->min_partial);
4678 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4684 err = strict_strtoul(buf, 10, &min);
4688 set_min_partial(s, min);
4691 SLAB_ATTR(min_partial);
4693 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4695 return sprintf(buf, "%u\n", s->cpu_partial);
4698 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4701 unsigned long objects;
4704 err = strict_strtoul(buf, 10, &objects);
4707 if (objects && kmem_cache_debug(s))
4710 s->cpu_partial = objects;
4714 SLAB_ATTR(cpu_partial);
4716 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4720 return sprintf(buf, "%pS\n", s->ctor);
4724 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4726 return sprintf(buf, "%d\n", s->refcount - 1);
4728 SLAB_ATTR_RO(aliases);
4730 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4732 return show_slab_objects(s, buf, SO_PARTIAL);
4734 SLAB_ATTR_RO(partial);
4736 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4738 return show_slab_objects(s, buf, SO_CPU);
4740 SLAB_ATTR_RO(cpu_slabs);
4742 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4744 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4746 SLAB_ATTR_RO(objects);
4748 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4750 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4752 SLAB_ATTR_RO(objects_partial);
4754 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4761 for_each_online_cpu(cpu) {
4762 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4765 pages += page->pages;
4766 objects += page->pobjects;
4770 len = sprintf(buf, "%d(%d)", objects, pages);
4773 for_each_online_cpu(cpu) {
4774 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4776 if (page && len < PAGE_SIZE - 20)
4777 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4778 page->pobjects, page->pages);
4781 return len + sprintf(buf + len, "\n");
4783 SLAB_ATTR_RO(slabs_cpu_partial);
4785 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4787 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4790 static ssize_t reclaim_account_store(struct kmem_cache *s,
4791 const char *buf, size_t length)
4793 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4795 s->flags |= SLAB_RECLAIM_ACCOUNT;
4798 SLAB_ATTR(reclaim_account);
4800 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4802 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4804 SLAB_ATTR_RO(hwcache_align);
4806 #ifdef CONFIG_ZONE_DMA
4807 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4809 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4811 SLAB_ATTR_RO(cache_dma);
4814 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4816 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4818 SLAB_ATTR_RO(destroy_by_rcu);
4820 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4822 return sprintf(buf, "%d\n", s->reserved);
4824 SLAB_ATTR_RO(reserved);
4826 #ifdef CONFIG_SLUB_DEBUG
4827 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4829 return show_slab_objects(s, buf, SO_ALL);
4831 SLAB_ATTR_RO(slabs);
4833 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4835 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4837 SLAB_ATTR_RO(total_objects);
4839 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4841 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4844 static ssize_t sanity_checks_store(struct kmem_cache *s,
4845 const char *buf, size_t length)
4847 s->flags &= ~SLAB_DEBUG_FREE;
4848 if (buf[0] == '1') {
4849 s->flags &= ~__CMPXCHG_DOUBLE;
4850 s->flags |= SLAB_DEBUG_FREE;
4854 SLAB_ATTR(sanity_checks);
4856 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4858 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4861 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4864 s->flags &= ~SLAB_TRACE;
4865 if (buf[0] == '1') {
4866 s->flags &= ~__CMPXCHG_DOUBLE;
4867 s->flags |= SLAB_TRACE;
4873 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4875 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4878 static ssize_t red_zone_store(struct kmem_cache *s,
4879 const char *buf, size_t length)
4881 if (any_slab_objects(s))
4884 s->flags &= ~SLAB_RED_ZONE;
4885 if (buf[0] == '1') {
4886 s->flags &= ~__CMPXCHG_DOUBLE;
4887 s->flags |= SLAB_RED_ZONE;
4889 calculate_sizes(s, -1);
4892 SLAB_ATTR(red_zone);
4894 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4896 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4899 static ssize_t poison_store(struct kmem_cache *s,
4900 const char *buf, size_t length)
4902 if (any_slab_objects(s))
4905 s->flags &= ~SLAB_POISON;
4906 if (buf[0] == '1') {
4907 s->flags &= ~__CMPXCHG_DOUBLE;
4908 s->flags |= SLAB_POISON;
4910 calculate_sizes(s, -1);
4915 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4917 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4920 static ssize_t store_user_store(struct kmem_cache *s,
4921 const char *buf, size_t length)
4923 if (any_slab_objects(s))
4926 s->flags &= ~SLAB_STORE_USER;
4927 if (buf[0] == '1') {
4928 s->flags &= ~__CMPXCHG_DOUBLE;
4929 s->flags |= SLAB_STORE_USER;
4931 calculate_sizes(s, -1);
4934 SLAB_ATTR(store_user);
4936 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4941 static ssize_t validate_store(struct kmem_cache *s,
4942 const char *buf, size_t length)
4946 if (buf[0] == '1') {
4947 ret = validate_slab_cache(s);
4953 SLAB_ATTR(validate);
4955 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4957 if (!(s->flags & SLAB_STORE_USER))
4959 return list_locations(s, buf, TRACK_ALLOC);
4961 SLAB_ATTR_RO(alloc_calls);
4963 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4965 if (!(s->flags & SLAB_STORE_USER))
4967 return list_locations(s, buf, TRACK_FREE);
4969 SLAB_ATTR_RO(free_calls);
4970 #endif /* CONFIG_SLUB_DEBUG */
4972 #ifdef CONFIG_FAILSLAB
4973 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4975 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4978 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4981 s->flags &= ~SLAB_FAILSLAB;
4983 s->flags |= SLAB_FAILSLAB;
4986 SLAB_ATTR(failslab);
4989 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4994 static ssize_t shrink_store(struct kmem_cache *s,
4995 const char *buf, size_t length)
4997 if (buf[0] == '1') {
4998 int rc = kmem_cache_shrink(s);
5009 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5011 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5014 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5015 const char *buf, size_t length)
5017 unsigned long ratio;
5020 err = strict_strtoul(buf, 10, &ratio);
5025 s->remote_node_defrag_ratio = ratio * 10;
5029 SLAB_ATTR(remote_node_defrag_ratio);
5032 #ifdef CONFIG_SLUB_STATS
5033 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5035 unsigned long sum = 0;
5038 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5043 for_each_online_cpu(cpu) {
5044 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5050 len = sprintf(buf, "%lu", sum);
5053 for_each_online_cpu(cpu) {
5054 if (data[cpu] && len < PAGE_SIZE - 20)
5055 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5059 return len + sprintf(buf + len, "\n");
5062 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5066 for_each_online_cpu(cpu)
5067 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5070 #define STAT_ATTR(si, text) \
5071 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5073 return show_stat(s, buf, si); \
5075 static ssize_t text##_store(struct kmem_cache *s, \
5076 const char *buf, size_t length) \
5078 if (buf[0] != '0') \
5080 clear_stat(s, si); \
5085 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5086 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5087 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5088 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5089 STAT_ATTR(FREE_FROZEN, free_frozen);
5090 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5091 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5092 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5093 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5094 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5095 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5096 STAT_ATTR(FREE_SLAB, free_slab);
5097 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5098 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5099 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5100 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5101 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5102 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5103 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5104 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5105 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5106 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5107 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5108 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5109 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5110 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5113 static struct attribute *slab_attrs[] = {
5114 &slab_size_attr.attr,
5115 &object_size_attr.attr,
5116 &objs_per_slab_attr.attr,
5118 &min_partial_attr.attr,
5119 &cpu_partial_attr.attr,
5121 &objects_partial_attr.attr,
5123 &cpu_slabs_attr.attr,
5127 &hwcache_align_attr.attr,
5128 &reclaim_account_attr.attr,
5129 &destroy_by_rcu_attr.attr,
5131 &reserved_attr.attr,
5132 &slabs_cpu_partial_attr.attr,
5133 #ifdef CONFIG_SLUB_DEBUG
5134 &total_objects_attr.attr,
5136 &sanity_checks_attr.attr,
5138 &red_zone_attr.attr,
5140 &store_user_attr.attr,
5141 &validate_attr.attr,
5142 &alloc_calls_attr.attr,
5143 &free_calls_attr.attr,
5145 #ifdef CONFIG_ZONE_DMA
5146 &cache_dma_attr.attr,
5149 &remote_node_defrag_ratio_attr.attr,
5151 #ifdef CONFIG_SLUB_STATS
5152 &alloc_fastpath_attr.attr,
5153 &alloc_slowpath_attr.attr,
5154 &free_fastpath_attr.attr,
5155 &free_slowpath_attr.attr,
5156 &free_frozen_attr.attr,
5157 &free_add_partial_attr.attr,
5158 &free_remove_partial_attr.attr,
5159 &alloc_from_partial_attr.attr,
5160 &alloc_slab_attr.attr,
5161 &alloc_refill_attr.attr,
5162 &alloc_node_mismatch_attr.attr,
5163 &free_slab_attr.attr,
5164 &cpuslab_flush_attr.attr,
5165 &deactivate_full_attr.attr,
5166 &deactivate_empty_attr.attr,
5167 &deactivate_to_head_attr.attr,
5168 &deactivate_to_tail_attr.attr,
5169 &deactivate_remote_frees_attr.attr,
5170 &deactivate_bypass_attr.attr,
5171 &order_fallback_attr.attr,
5172 &cmpxchg_double_fail_attr.attr,
5173 &cmpxchg_double_cpu_fail_attr.attr,
5174 &cpu_partial_alloc_attr.attr,
5175 &cpu_partial_free_attr.attr,
5176 &cpu_partial_node_attr.attr,
5177 &cpu_partial_drain_attr.attr,
5179 #ifdef CONFIG_FAILSLAB
5180 &failslab_attr.attr,
5186 static struct attribute_group slab_attr_group = {
5187 .attrs = slab_attrs,
5190 static ssize_t slab_attr_show(struct kobject *kobj,
5191 struct attribute *attr,
5194 struct slab_attribute *attribute;
5195 struct kmem_cache *s;
5198 attribute = to_slab_attr(attr);
5201 if (!attribute->show)
5204 err = attribute->show(s, buf);
5209 static ssize_t slab_attr_store(struct kobject *kobj,
5210 struct attribute *attr,
5211 const char *buf, size_t len)
5213 struct slab_attribute *attribute;
5214 struct kmem_cache *s;
5217 attribute = to_slab_attr(attr);
5220 if (!attribute->store)
5223 err = attribute->store(s, buf, len);
5228 static void kmem_cache_release(struct kobject *kobj)
5230 struct kmem_cache *s = to_slab(kobj);
5236 static const struct sysfs_ops slab_sysfs_ops = {
5237 .show = slab_attr_show,
5238 .store = slab_attr_store,
5241 static struct kobj_type slab_ktype = {
5242 .sysfs_ops = &slab_sysfs_ops,
5243 .release = kmem_cache_release
5246 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5248 struct kobj_type *ktype = get_ktype(kobj);
5250 if (ktype == &slab_ktype)
5255 static const struct kset_uevent_ops slab_uevent_ops = {
5256 .filter = uevent_filter,
5259 static struct kset *slab_kset;
5261 #define ID_STR_LENGTH 64
5263 /* Create a unique string id for a slab cache:
5265 * Format :[flags-]size
5267 static char *create_unique_id(struct kmem_cache *s)
5269 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5276 * First flags affecting slabcache operations. We will only
5277 * get here for aliasable slabs so we do not need to support
5278 * too many flags. The flags here must cover all flags that
5279 * are matched during merging to guarantee that the id is
5282 if (s->flags & SLAB_CACHE_DMA)
5284 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5286 if (s->flags & SLAB_DEBUG_FREE)
5288 if (!(s->flags & SLAB_NOTRACK))
5292 p += sprintf(p, "%07d", s->size);
5293 BUG_ON(p > name + ID_STR_LENGTH - 1);
5297 static int sysfs_slab_add(struct kmem_cache *s)
5303 if (slab_state < FULL)
5304 /* Defer until later */
5307 unmergeable = slab_unmergeable(s);
5310 * Slabcache can never be merged so we can use the name proper.
5311 * This is typically the case for debug situations. In that
5312 * case we can catch duplicate names easily.
5314 sysfs_remove_link(&slab_kset->kobj, s->name);
5318 * Create a unique name for the slab as a target
5321 name = create_unique_id(s);
5324 s->kobj.kset = slab_kset;
5325 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5327 kobject_put(&s->kobj);
5331 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5333 kobject_del(&s->kobj);
5334 kobject_put(&s->kobj);
5337 kobject_uevent(&s->kobj, KOBJ_ADD);
5339 /* Setup first alias */
5340 sysfs_slab_alias(s, s->name);
5346 static void sysfs_slab_remove(struct kmem_cache *s)
5348 if (slab_state < FULL)
5350 * Sysfs has not been setup yet so no need to remove the
5355 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5356 kobject_del(&s->kobj);
5357 kobject_put(&s->kobj);
5361 * Need to buffer aliases during bootup until sysfs becomes
5362 * available lest we lose that information.
5364 struct saved_alias {
5365 struct kmem_cache *s;
5367 struct saved_alias *next;
5370 static struct saved_alias *alias_list;
5372 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5374 struct saved_alias *al;
5376 if (slab_state == FULL) {
5378 * If we have a leftover link then remove it.
5380 sysfs_remove_link(&slab_kset->kobj, name);
5381 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5384 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5390 al->next = alias_list;
5395 static int __init slab_sysfs_init(void)
5397 struct kmem_cache *s;
5400 mutex_lock(&slab_mutex);
5402 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5404 mutex_unlock(&slab_mutex);
5405 printk(KERN_ERR "Cannot register slab subsystem.\n");
5411 list_for_each_entry(s, &slab_caches, list) {
5412 err = sysfs_slab_add(s);
5414 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5415 " to sysfs\n", s->name);
5418 while (alias_list) {
5419 struct saved_alias *al = alias_list;
5421 alias_list = alias_list->next;
5422 err = sysfs_slab_alias(al->s, al->name);
5424 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5425 " %s to sysfs\n", al->name);
5429 mutex_unlock(&slab_mutex);
5434 __initcall(slab_sysfs_init);
5435 #endif /* CONFIG_SYSFS */
5438 * The /proc/slabinfo ABI
5440 #ifdef CONFIG_SLABINFO
5441 static void print_slabinfo_header(struct seq_file *m)
5443 seq_puts(m, "slabinfo - version: 2.1\n");
5444 seq_puts(m, "# name <active_objs> <num_objs> <object_size> "
5445 "<objperslab> <pagesperslab>");
5446 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5447 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5451 static void *s_start(struct seq_file *m, loff_t *pos)
5455 mutex_lock(&slab_mutex);
5457 print_slabinfo_header(m);
5459 return seq_list_start(&slab_caches, *pos);
5462 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5464 return seq_list_next(p, &slab_caches, pos);
5467 static void s_stop(struct seq_file *m, void *p)
5469 mutex_unlock(&slab_mutex);
5472 static int s_show(struct seq_file *m, void *p)
5474 unsigned long nr_partials = 0;
5475 unsigned long nr_slabs = 0;
5476 unsigned long nr_inuse = 0;
5477 unsigned long nr_objs = 0;
5478 unsigned long nr_free = 0;
5479 struct kmem_cache *s;
5482 s = list_entry(p, struct kmem_cache, list);
5484 for_each_online_node(node) {
5485 struct kmem_cache_node *n = get_node(s, node);
5490 nr_partials += n->nr_partial;
5491 nr_slabs += atomic_long_read(&n->nr_slabs);
5492 nr_objs += atomic_long_read(&n->total_objects);
5493 nr_free += count_partial(n, count_free);
5496 nr_inuse = nr_objs - nr_free;
5498 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5499 nr_objs, s->size, oo_objects(s->oo),
5500 (1 << oo_order(s->oo)));
5501 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5502 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5508 static const struct seq_operations slabinfo_op = {
5515 static int slabinfo_open(struct inode *inode, struct file *file)
5517 return seq_open(file, &slabinfo_op);
5520 static const struct file_operations proc_slabinfo_operations = {
5521 .open = slabinfo_open,
5523 .llseek = seq_lseek,
5524 .release = seq_release,
5527 static int __init slab_proc_init(void)
5529 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5532 module_init(slab_proc_init);
5533 #endif /* CONFIG_SLABINFO */