2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
35 #include <trace/events/kmem.h>
41 * 1. slab_mutex (Global Mutex)
43 * 3. slab_lock(page) (Only on some arches and for debugging)
47 * The role of the slab_mutex is to protect the list of all the slabs
48 * and to synchronize major metadata changes to slab cache structures.
50 * The slab_lock is only used for debugging and on arches that do not
51 * have the ability to do a cmpxchg_double. It only protects the second
52 * double word in the page struct. Meaning
53 * A. page->freelist -> List of object free in a page
54 * B. page->counters -> Counters of objects
55 * C. page->frozen -> frozen state
57 * If a slab is frozen then it is exempt from list management. It is not
58 * on any list. The processor that froze the slab is the one who can
59 * perform list operations on the page. Other processors may put objects
60 * onto the freelist but the processor that froze the slab is the only
61 * one that can retrieve the objects from the page's freelist.
63 * The list_lock protects the partial and full list on each node and
64 * the partial slab counter. If taken then no new slabs may be added or
65 * removed from the lists nor make the number of partial slabs be modified.
66 * (Note that the total number of slabs is an atomic value that may be
67 * modified without taking the list lock).
69 * The list_lock is a centralized lock and thus we avoid taking it as
70 * much as possible. As long as SLUB does not have to handle partial
71 * slabs, operations can continue without any centralized lock. F.e.
72 * allocating a long series of objects that fill up slabs does not require
74 * Interrupts are disabled during allocation and deallocation in order to
75 * make the slab allocator safe to use in the context of an irq. In addition
76 * interrupts are disabled to ensure that the processor does not change
77 * while handling per_cpu slabs, due to kernel preemption.
79 * SLUB assigns one slab for allocation to each processor.
80 * Allocations only occur from these slabs called cpu slabs.
82 * Slabs with free elements are kept on a partial list and during regular
83 * operations no list for full slabs is used. If an object in a full slab is
84 * freed then the slab will show up again on the partial lists.
85 * We track full slabs for debugging purposes though because otherwise we
86 * cannot scan all objects.
88 * Slabs are freed when they become empty. Teardown and setup is
89 * minimal so we rely on the page allocators per cpu caches for
90 * fast frees and allocs.
92 * Overloading of page flags that are otherwise used for LRU management.
94 * PageActive The slab is frozen and exempt from list processing.
95 * This means that the slab is dedicated to a purpose
96 * such as satisfying allocations for a specific
97 * processor. Objects may be freed in the slab while
98 * it is frozen but slab_free will then skip the usual
99 * list operations. It is up to the processor holding
100 * the slab to integrate the slab into the slab lists
101 * when the slab is no longer needed.
103 * One use of this flag is to mark slabs that are
104 * used for allocations. Then such a slab becomes a cpu
105 * slab. The cpu slab may be equipped with an additional
106 * freelist that allows lockless access to
107 * free objects in addition to the regular freelist
108 * that requires the slab lock.
110 * PageError Slab requires special handling due to debug
111 * options set. This moves slab handling out of
112 * the fast path and disables lockless freelists.
115 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
116 SLAB_TRACE | SLAB_DEBUG_FREE)
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
128 * Issues still to be resolved:
130 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
132 * - Variable sizing of the per node arrays
135 /* Enable to test recovery from slab corruption on boot */
136 #undef SLUB_RESILIENCY_TEST
138 /* Enable to log cmpxchg failures */
139 #undef SLUB_DEBUG_CMPXCHG
142 * Mininum number of partial slabs. These will be left on the partial
143 * lists even if they are empty. kmem_cache_shrink may reclaim them.
145 #define MIN_PARTIAL 5
148 * Maximum number of desirable partial slabs.
149 * The existence of more partial slabs makes kmem_cache_shrink
150 * sort the partial list by the number of objects in the.
152 #define MAX_PARTIAL 10
154 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_STORE_USER)
158 * Debugging flags that require metadata to be stored in the slab. These get
159 * disabled when slub_debug=O is used and a cache's min order increases with
162 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
165 * Set of flags that will prevent slab merging
167 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
168 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
171 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
172 SLAB_CACHE_DMA | SLAB_NOTRACK)
175 #define OO_MASK ((1 << OO_SHIFT) - 1)
176 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
178 /* Internal SLUB flags */
179 #define __OBJECT_POISON 0x80000000UL /* Poison object */
180 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static int kmem_size = sizeof(struct kmem_cache);
185 static struct notifier_block slab_notifier;
189 * Tracking user of a slab.
191 #define TRACK_ADDRS_COUNT 16
193 unsigned long addr; /* Called from address */
194 #ifdef CONFIG_STACKTRACE
195 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
197 int cpu; /* Was running on cpu */
198 int pid; /* Pid context */
199 unsigned long when; /* When did the operation occur */
202 enum track_item { TRACK_ALLOC, TRACK_FREE };
205 static int sysfs_slab_add(struct kmem_cache *);
206 static int sysfs_slab_alias(struct kmem_cache *, const char *);
207 static void sysfs_slab_remove(struct kmem_cache *);
210 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
211 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
213 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
217 static inline void stat(const struct kmem_cache *s, enum stat_item si)
219 #ifdef CONFIG_SLUB_STATS
220 __this_cpu_inc(s->cpu_slab->stat[si]);
224 /********************************************************************
225 * Core slab cache functions
226 *******************************************************************/
228 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
230 return s->node[node];
233 /* Verify that a pointer has an address that is valid within a slab page */
234 static inline int check_valid_pointer(struct kmem_cache *s,
235 struct page *page, const void *object)
242 base = page_address(page);
243 if (object < base || object >= base + page->objects * s->size ||
244 (object - base) % s->size) {
251 static inline void *get_freepointer(struct kmem_cache *s, void *object)
253 return *(void **)(object + s->offset);
256 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
258 prefetch(object + s->offset);
261 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
265 #ifdef CONFIG_DEBUG_PAGEALLOC
266 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
268 p = get_freepointer(s, object);
273 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
275 *(void **)(object + s->offset) = fp;
278 /* Loop over all objects in a slab */
279 #define for_each_object(__p, __s, __addr, __objects) \
280 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
283 /* Determine object index from a given position */
284 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
286 return (p - addr) / s->size;
289 static inline size_t slab_ksize(const struct kmem_cache *s)
291 #ifdef CONFIG_SLUB_DEBUG
293 * Debugging requires use of the padding between object
294 * and whatever may come after it.
296 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
297 return s->object_size;
301 * If we have the need to store the freelist pointer
302 * back there or track user information then we can
303 * only use the space before that information.
305 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
308 * Else we can use all the padding etc for the allocation
313 static inline int order_objects(int order, unsigned long size, int reserved)
315 return ((PAGE_SIZE << order) - reserved) / size;
318 static inline struct kmem_cache_order_objects oo_make(int order,
319 unsigned long size, int reserved)
321 struct kmem_cache_order_objects x = {
322 (order << OO_SHIFT) + order_objects(order, size, reserved)
328 static inline int oo_order(struct kmem_cache_order_objects x)
330 return x.x >> OO_SHIFT;
333 static inline int oo_objects(struct kmem_cache_order_objects x)
335 return x.x & OO_MASK;
339 * Per slab locking using the pagelock
341 static __always_inline void slab_lock(struct page *page)
343 bit_spin_lock(PG_locked, &page->flags);
346 static __always_inline void slab_unlock(struct page *page)
348 __bit_spin_unlock(PG_locked, &page->flags);
351 /* Interrupts must be disabled (for the fallback code to work right) */
352 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
353 void *freelist_old, unsigned long counters_old,
354 void *freelist_new, unsigned long counters_new,
357 VM_BUG_ON(!irqs_disabled());
358 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
359 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
360 if (s->flags & __CMPXCHG_DOUBLE) {
361 if (cmpxchg_double(&page->freelist, &page->counters,
362 freelist_old, counters_old,
363 freelist_new, counters_new))
369 if (page->freelist == freelist_old && page->counters == counters_old) {
370 page->freelist = freelist_new;
371 page->counters = counters_new;
379 stat(s, CMPXCHG_DOUBLE_FAIL);
381 #ifdef SLUB_DEBUG_CMPXCHG
382 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
388 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
389 void *freelist_old, unsigned long counters_old,
390 void *freelist_new, unsigned long counters_new,
393 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
394 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
395 if (s->flags & __CMPXCHG_DOUBLE) {
396 if (cmpxchg_double(&page->freelist, &page->counters,
397 freelist_old, counters_old,
398 freelist_new, counters_new))
405 local_irq_save(flags);
407 if (page->freelist == freelist_old && page->counters == counters_old) {
408 page->freelist = freelist_new;
409 page->counters = counters_new;
411 local_irq_restore(flags);
415 local_irq_restore(flags);
419 stat(s, CMPXCHG_DOUBLE_FAIL);
421 #ifdef SLUB_DEBUG_CMPXCHG
422 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
428 #ifdef CONFIG_SLUB_DEBUG
430 * Determine a map of object in use on a page.
432 * Node listlock must be held to guarantee that the page does
433 * not vanish from under us.
435 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
438 void *addr = page_address(page);
440 for (p = page->freelist; p; p = get_freepointer(s, p))
441 set_bit(slab_index(p, s, addr), map);
447 #ifdef CONFIG_SLUB_DEBUG_ON
448 static int slub_debug = DEBUG_DEFAULT_FLAGS;
450 static int slub_debug;
453 static char *slub_debug_slabs;
454 static int disable_higher_order_debug;
459 static void print_section(char *text, u8 *addr, unsigned int length)
461 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
465 static struct track *get_track(struct kmem_cache *s, void *object,
466 enum track_item alloc)
471 p = object + s->offset + sizeof(void *);
473 p = object + s->inuse;
478 static void set_track(struct kmem_cache *s, void *object,
479 enum track_item alloc, unsigned long addr)
481 struct track *p = get_track(s, object, alloc);
484 #ifdef CONFIG_STACKTRACE
485 struct stack_trace trace;
488 trace.nr_entries = 0;
489 trace.max_entries = TRACK_ADDRS_COUNT;
490 trace.entries = p->addrs;
492 save_stack_trace(&trace);
494 /* See rant in lockdep.c */
495 if (trace.nr_entries != 0 &&
496 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
499 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
503 p->cpu = smp_processor_id();
504 p->pid = current->pid;
507 memset(p, 0, sizeof(struct track));
510 static void init_tracking(struct kmem_cache *s, void *object)
512 if (!(s->flags & SLAB_STORE_USER))
515 set_track(s, object, TRACK_FREE, 0UL);
516 set_track(s, object, TRACK_ALLOC, 0UL);
519 static void print_track(const char *s, struct track *t)
524 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
525 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
526 #ifdef CONFIG_STACKTRACE
529 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
531 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
538 static void print_tracking(struct kmem_cache *s, void *object)
540 if (!(s->flags & SLAB_STORE_USER))
543 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
544 print_track("Freed", get_track(s, object, TRACK_FREE));
547 static void print_page_info(struct page *page)
549 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
550 page, page->objects, page->inuse, page->freelist, page->flags);
554 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
560 vsnprintf(buf, sizeof(buf), fmt, args);
562 printk(KERN_ERR "========================================"
563 "=====================================\n");
564 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
565 printk(KERN_ERR "----------------------------------------"
566 "-------------------------------------\n\n");
568 add_taint(TAINT_BAD_PAGE);
571 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
577 vsnprintf(buf, sizeof(buf), fmt, args);
579 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
582 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
584 unsigned int off; /* Offset of last byte */
585 u8 *addr = page_address(page);
587 print_tracking(s, p);
589 print_page_info(page);
591 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
592 p, p - addr, get_freepointer(s, p));
595 print_section("Bytes b4 ", p - 16, 16);
597 print_section("Object ", p, min_t(unsigned long, s->object_size,
599 if (s->flags & SLAB_RED_ZONE)
600 print_section("Redzone ", p + s->object_size,
601 s->inuse - s->object_size);
604 off = s->offset + sizeof(void *);
608 if (s->flags & SLAB_STORE_USER)
609 off += 2 * sizeof(struct track);
612 /* Beginning of the filler is the free pointer */
613 print_section("Padding ", p + off, s->size - off);
618 static void object_err(struct kmem_cache *s, struct page *page,
619 u8 *object, char *reason)
621 slab_bug(s, "%s", reason);
622 print_trailer(s, page, object);
625 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
631 vsnprintf(buf, sizeof(buf), fmt, args);
633 slab_bug(s, "%s", buf);
634 print_page_info(page);
638 static void init_object(struct kmem_cache *s, void *object, u8 val)
642 if (s->flags & __OBJECT_POISON) {
643 memset(p, POISON_FREE, s->object_size - 1);
644 p[s->object_size - 1] = POISON_END;
647 if (s->flags & SLAB_RED_ZONE)
648 memset(p + s->object_size, val, s->inuse - s->object_size);
651 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
652 void *from, void *to)
654 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
655 memset(from, data, to - from);
658 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
659 u8 *object, char *what,
660 u8 *start, unsigned int value, unsigned int bytes)
665 fault = memchr_inv(start, value, bytes);
670 while (end > fault && end[-1] == value)
673 slab_bug(s, "%s overwritten", what);
674 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
675 fault, end - 1, fault[0], value);
676 print_trailer(s, page, object);
678 restore_bytes(s, what, value, fault, end);
686 * Bytes of the object to be managed.
687 * If the freepointer may overlay the object then the free
688 * pointer is the first word of the object.
690 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
693 * object + s->object_size
694 * Padding to reach word boundary. This is also used for Redzoning.
695 * Padding is extended by another word if Redzoning is enabled and
696 * object_size == inuse.
698 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
699 * 0xcc (RED_ACTIVE) for objects in use.
702 * Meta data starts here.
704 * A. Free pointer (if we cannot overwrite object on free)
705 * B. Tracking data for SLAB_STORE_USER
706 * C. Padding to reach required alignment boundary or at mininum
707 * one word if debugging is on to be able to detect writes
708 * before the word boundary.
710 * Padding is done using 0x5a (POISON_INUSE)
713 * Nothing is used beyond s->size.
715 * If slabcaches are merged then the object_size and inuse boundaries are mostly
716 * ignored. And therefore no slab options that rely on these boundaries
717 * may be used with merged slabcaches.
720 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
722 unsigned long off = s->inuse; /* The end of info */
725 /* Freepointer is placed after the object. */
726 off += sizeof(void *);
728 if (s->flags & SLAB_STORE_USER)
729 /* We also have user information there */
730 off += 2 * sizeof(struct track);
735 return check_bytes_and_report(s, page, p, "Object padding",
736 p + off, POISON_INUSE, s->size - off);
739 /* Check the pad bytes at the end of a slab page */
740 static int slab_pad_check(struct kmem_cache *s, struct page *page)
748 if (!(s->flags & SLAB_POISON))
751 start = page_address(page);
752 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
753 end = start + length;
754 remainder = length % s->size;
758 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
761 while (end > fault && end[-1] == POISON_INUSE)
764 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
765 print_section("Padding ", end - remainder, remainder);
767 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
771 static int check_object(struct kmem_cache *s, struct page *page,
772 void *object, u8 val)
775 u8 *endobject = object + s->object_size;
777 if (s->flags & SLAB_RED_ZONE) {
778 if (!check_bytes_and_report(s, page, object, "Redzone",
779 endobject, val, s->inuse - s->object_size))
782 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
783 check_bytes_and_report(s, page, p, "Alignment padding",
784 endobject, POISON_INUSE, s->inuse - s->object_size);
788 if (s->flags & SLAB_POISON) {
789 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
790 (!check_bytes_and_report(s, page, p, "Poison", p,
791 POISON_FREE, s->object_size - 1) ||
792 !check_bytes_and_report(s, page, p, "Poison",
793 p + s->object_size - 1, POISON_END, 1)))
796 * check_pad_bytes cleans up on its own.
798 check_pad_bytes(s, page, p);
801 if (!s->offset && val == SLUB_RED_ACTIVE)
803 * Object and freepointer overlap. Cannot check
804 * freepointer while object is allocated.
808 /* Check free pointer validity */
809 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
810 object_err(s, page, p, "Freepointer corrupt");
812 * No choice but to zap it and thus lose the remainder
813 * of the free objects in this slab. May cause
814 * another error because the object count is now wrong.
816 set_freepointer(s, p, NULL);
822 static int check_slab(struct kmem_cache *s, struct page *page)
826 VM_BUG_ON(!irqs_disabled());
828 if (!PageSlab(page)) {
829 slab_err(s, page, "Not a valid slab page");
833 maxobj = order_objects(compound_order(page), s->size, s->reserved);
834 if (page->objects > maxobj) {
835 slab_err(s, page, "objects %u > max %u",
836 s->name, page->objects, maxobj);
839 if (page->inuse > page->objects) {
840 slab_err(s, page, "inuse %u > max %u",
841 s->name, page->inuse, page->objects);
844 /* Slab_pad_check fixes things up after itself */
845 slab_pad_check(s, page);
850 * Determine if a certain object on a page is on the freelist. Must hold the
851 * slab lock to guarantee that the chains are in a consistent state.
853 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
858 unsigned long max_objects;
861 while (fp && nr <= page->objects) {
864 if (!check_valid_pointer(s, page, fp)) {
866 object_err(s, page, object,
867 "Freechain corrupt");
868 set_freepointer(s, object, NULL);
871 slab_err(s, page, "Freepointer corrupt");
872 page->freelist = NULL;
873 page->inuse = page->objects;
874 slab_fix(s, "Freelist cleared");
880 fp = get_freepointer(s, object);
884 max_objects = order_objects(compound_order(page), s->size, s->reserved);
885 if (max_objects > MAX_OBJS_PER_PAGE)
886 max_objects = MAX_OBJS_PER_PAGE;
888 if (page->objects != max_objects) {
889 slab_err(s, page, "Wrong number of objects. Found %d but "
890 "should be %d", page->objects, max_objects);
891 page->objects = max_objects;
892 slab_fix(s, "Number of objects adjusted.");
894 if (page->inuse != page->objects - nr) {
895 slab_err(s, page, "Wrong object count. Counter is %d but "
896 "counted were %d", page->inuse, page->objects - nr);
897 page->inuse = page->objects - nr;
898 slab_fix(s, "Object count adjusted.");
900 return search == NULL;
903 static void trace(struct kmem_cache *s, struct page *page, void *object,
906 if (s->flags & SLAB_TRACE) {
907 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
909 alloc ? "alloc" : "free",
914 print_section("Object ", (void *)object, s->object_size);
921 * Hooks for other subsystems that check memory allocations. In a typical
922 * production configuration these hooks all should produce no code at all.
924 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
926 flags &= gfp_allowed_mask;
927 lockdep_trace_alloc(flags);
928 might_sleep_if(flags & __GFP_WAIT);
930 return should_failslab(s->object_size, flags, s->flags);
933 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
935 flags &= gfp_allowed_mask;
936 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
937 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
940 static inline void slab_free_hook(struct kmem_cache *s, void *x)
942 kmemleak_free_recursive(x, s->flags);
945 * Trouble is that we may no longer disable interupts in the fast path
946 * So in order to make the debug calls that expect irqs to be
947 * disabled we need to disable interrupts temporarily.
949 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
953 local_irq_save(flags);
954 kmemcheck_slab_free(s, x, s->object_size);
955 debug_check_no_locks_freed(x, s->object_size);
956 local_irq_restore(flags);
959 if (!(s->flags & SLAB_DEBUG_OBJECTS))
960 debug_check_no_obj_freed(x, s->object_size);
964 * Tracking of fully allocated slabs for debugging purposes.
966 * list_lock must be held.
968 static void add_full(struct kmem_cache *s,
969 struct kmem_cache_node *n, struct page *page)
971 if (!(s->flags & SLAB_STORE_USER))
974 list_add(&page->lru, &n->full);
978 * list_lock must be held.
980 static void remove_full(struct kmem_cache *s, struct page *page)
982 if (!(s->flags & SLAB_STORE_USER))
985 list_del(&page->lru);
988 /* Tracking of the number of slabs for debugging purposes */
989 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
991 struct kmem_cache_node *n = get_node(s, node);
993 return atomic_long_read(&n->nr_slabs);
996 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
998 return atomic_long_read(&n->nr_slabs);
1001 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1003 struct kmem_cache_node *n = get_node(s, node);
1006 * May be called early in order to allocate a slab for the
1007 * kmem_cache_node structure. Solve the chicken-egg
1008 * dilemma by deferring the increment of the count during
1009 * bootstrap (see early_kmem_cache_node_alloc).
1012 atomic_long_inc(&n->nr_slabs);
1013 atomic_long_add(objects, &n->total_objects);
1016 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1018 struct kmem_cache_node *n = get_node(s, node);
1020 atomic_long_dec(&n->nr_slabs);
1021 atomic_long_sub(objects, &n->total_objects);
1024 /* Object debug checks for alloc/free paths */
1025 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1028 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1031 init_object(s, object, SLUB_RED_INACTIVE);
1032 init_tracking(s, object);
1035 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1036 void *object, unsigned long addr)
1038 if (!check_slab(s, page))
1041 if (!check_valid_pointer(s, page, object)) {
1042 object_err(s, page, object, "Freelist Pointer check fails");
1046 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1049 /* Success perform special debug activities for allocs */
1050 if (s->flags & SLAB_STORE_USER)
1051 set_track(s, object, TRACK_ALLOC, addr);
1052 trace(s, page, object, 1);
1053 init_object(s, object, SLUB_RED_ACTIVE);
1057 if (PageSlab(page)) {
1059 * If this is a slab page then lets do the best we can
1060 * to avoid issues in the future. Marking all objects
1061 * as used avoids touching the remaining objects.
1063 slab_fix(s, "Marking all objects used");
1064 page->inuse = page->objects;
1065 page->freelist = NULL;
1070 static noinline struct kmem_cache_node *free_debug_processing(
1071 struct kmem_cache *s, struct page *page, void *object,
1072 unsigned long addr, unsigned long *flags)
1074 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1076 spin_lock_irqsave(&n->list_lock, *flags);
1079 if (!check_slab(s, page))
1082 if (!check_valid_pointer(s, page, object)) {
1083 slab_err(s, page, "Invalid object pointer 0x%p", object);
1087 if (on_freelist(s, page, object)) {
1088 object_err(s, page, object, "Object already free");
1092 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1095 if (unlikely(s != page->slab)) {
1096 if (!PageSlab(page)) {
1097 slab_err(s, page, "Attempt to free object(0x%p) "
1098 "outside of slab", object);
1099 } else if (!page->slab) {
1101 "SLUB <none>: no slab for object 0x%p.\n",
1105 object_err(s, page, object,
1106 "page slab pointer corrupt.");
1110 if (s->flags & SLAB_STORE_USER)
1111 set_track(s, object, TRACK_FREE, addr);
1112 trace(s, page, object, 0);
1113 init_object(s, object, SLUB_RED_INACTIVE);
1117 * Keep node_lock to preserve integrity
1118 * until the object is actually freed
1124 spin_unlock_irqrestore(&n->list_lock, *flags);
1125 slab_fix(s, "Object at 0x%p not freed", object);
1129 static int __init setup_slub_debug(char *str)
1131 slub_debug = DEBUG_DEFAULT_FLAGS;
1132 if (*str++ != '=' || !*str)
1134 * No options specified. Switch on full debugging.
1140 * No options but restriction on slabs. This means full
1141 * debugging for slabs matching a pattern.
1145 if (tolower(*str) == 'o') {
1147 * Avoid enabling debugging on caches if its minimum order
1148 * would increase as a result.
1150 disable_higher_order_debug = 1;
1157 * Switch off all debugging measures.
1162 * Determine which debug features should be switched on
1164 for (; *str && *str != ','; str++) {
1165 switch (tolower(*str)) {
1167 slub_debug |= SLAB_DEBUG_FREE;
1170 slub_debug |= SLAB_RED_ZONE;
1173 slub_debug |= SLAB_POISON;
1176 slub_debug |= SLAB_STORE_USER;
1179 slub_debug |= SLAB_TRACE;
1182 slub_debug |= SLAB_FAILSLAB;
1185 printk(KERN_ERR "slub_debug option '%c' "
1186 "unknown. skipped\n", *str);
1192 slub_debug_slabs = str + 1;
1197 __setup("slub_debug", setup_slub_debug);
1199 static unsigned long kmem_cache_flags(unsigned long object_size,
1200 unsigned long flags, const char *name,
1201 void (*ctor)(void *))
1204 * Enable debugging if selected on the kernel commandline.
1206 if (slub_debug && (!slub_debug_slabs ||
1207 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1208 flags |= slub_debug;
1213 static inline void setup_object_debug(struct kmem_cache *s,
1214 struct page *page, void *object) {}
1216 static inline int alloc_debug_processing(struct kmem_cache *s,
1217 struct page *page, void *object, unsigned long addr) { return 0; }
1219 static inline struct kmem_cache_node *free_debug_processing(
1220 struct kmem_cache *s, struct page *page, void *object,
1221 unsigned long addr, unsigned long *flags) { return NULL; }
1223 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1225 static inline int check_object(struct kmem_cache *s, struct page *page,
1226 void *object, u8 val) { return 1; }
1227 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1228 struct page *page) {}
1229 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1230 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1231 unsigned long flags, const char *name,
1232 void (*ctor)(void *))
1236 #define slub_debug 0
1238 #define disable_higher_order_debug 0
1240 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1242 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1244 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1246 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1249 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1252 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1255 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1257 #endif /* CONFIG_SLUB_DEBUG */
1260 * Slab allocation and freeing
1262 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1263 struct kmem_cache_order_objects oo)
1265 int order = oo_order(oo);
1267 flags |= __GFP_NOTRACK;
1269 if (node == NUMA_NO_NODE)
1270 return alloc_pages(flags, order);
1272 return alloc_pages_exact_node(node, flags, order);
1275 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1278 struct kmem_cache_order_objects oo = s->oo;
1281 flags &= gfp_allowed_mask;
1283 if (flags & __GFP_WAIT)
1286 flags |= s->allocflags;
1289 * Let the initial higher-order allocation fail under memory pressure
1290 * so we fall-back to the minimum order allocation.
1292 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1294 page = alloc_slab_page(alloc_gfp, node, oo);
1295 if (unlikely(!page)) {
1298 * Allocation may have failed due to fragmentation.
1299 * Try a lower order alloc if possible
1301 page = alloc_slab_page(flags, node, oo);
1304 stat(s, ORDER_FALLBACK);
1307 if (kmemcheck_enabled && page
1308 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1309 int pages = 1 << oo_order(oo);
1311 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1314 * Objects from caches that have a constructor don't get
1315 * cleared when they're allocated, so we need to do it here.
1318 kmemcheck_mark_uninitialized_pages(page, pages);
1320 kmemcheck_mark_unallocated_pages(page, pages);
1323 if (flags & __GFP_WAIT)
1324 local_irq_disable();
1328 page->objects = oo_objects(oo);
1329 mod_zone_page_state(page_zone(page),
1330 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1331 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1337 static void setup_object(struct kmem_cache *s, struct page *page,
1340 setup_object_debug(s, page, object);
1341 if (unlikely(s->ctor))
1345 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1352 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1354 page = allocate_slab(s,
1355 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1359 inc_slabs_node(s, page_to_nid(page), page->objects);
1361 __SetPageSlab(page);
1362 if (page->pfmemalloc)
1363 SetPageSlabPfmemalloc(page);
1365 start = page_address(page);
1367 if (unlikely(s->flags & SLAB_POISON))
1368 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1371 for_each_object(p, s, start, page->objects) {
1372 setup_object(s, page, last);
1373 set_freepointer(s, last, p);
1376 setup_object(s, page, last);
1377 set_freepointer(s, last, NULL);
1379 page->freelist = start;
1380 page->inuse = page->objects;
1386 static void __free_slab(struct kmem_cache *s, struct page *page)
1388 int order = compound_order(page);
1389 int pages = 1 << order;
1391 if (kmem_cache_debug(s)) {
1394 slab_pad_check(s, page);
1395 for_each_object(p, s, page_address(page),
1397 check_object(s, page, p, SLUB_RED_INACTIVE);
1400 kmemcheck_free_shadow(page, compound_order(page));
1402 mod_zone_page_state(page_zone(page),
1403 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1404 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1407 __ClearPageSlabPfmemalloc(page);
1408 __ClearPageSlab(page);
1409 reset_page_mapcount(page);
1410 if (current->reclaim_state)
1411 current->reclaim_state->reclaimed_slab += pages;
1412 __free_pages(page, order);
1415 #define need_reserve_slab_rcu \
1416 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1418 static void rcu_free_slab(struct rcu_head *h)
1422 if (need_reserve_slab_rcu)
1423 page = virt_to_head_page(h);
1425 page = container_of((struct list_head *)h, struct page, lru);
1427 __free_slab(page->slab, page);
1430 static void free_slab(struct kmem_cache *s, struct page *page)
1432 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1433 struct rcu_head *head;
1435 if (need_reserve_slab_rcu) {
1436 int order = compound_order(page);
1437 int offset = (PAGE_SIZE << order) - s->reserved;
1439 VM_BUG_ON(s->reserved != sizeof(*head));
1440 head = page_address(page) + offset;
1443 * RCU free overloads the RCU head over the LRU
1445 head = (void *)&page->lru;
1448 call_rcu(head, rcu_free_slab);
1450 __free_slab(s, page);
1453 static void discard_slab(struct kmem_cache *s, struct page *page)
1455 dec_slabs_node(s, page_to_nid(page), page->objects);
1460 * Management of partially allocated slabs.
1462 * list_lock must be held.
1464 static inline void add_partial(struct kmem_cache_node *n,
1465 struct page *page, int tail)
1468 if (tail == DEACTIVATE_TO_TAIL)
1469 list_add_tail(&page->lru, &n->partial);
1471 list_add(&page->lru, &n->partial);
1475 * list_lock must be held.
1477 static inline void remove_partial(struct kmem_cache_node *n,
1480 list_del(&page->lru);
1485 * Remove slab from the partial list, freeze it and
1486 * return the pointer to the freelist.
1488 * Returns a list of objects or NULL if it fails.
1490 * Must hold list_lock since we modify the partial list.
1492 static inline void *acquire_slab(struct kmem_cache *s,
1493 struct kmem_cache_node *n, struct page *page,
1497 unsigned long counters;
1501 * Zap the freelist and set the frozen bit.
1502 * The old freelist is the list of objects for the
1503 * per cpu allocation list.
1505 freelist = page->freelist;
1506 counters = page->counters;
1507 new.counters = counters;
1509 new.inuse = page->objects;
1510 new.freelist = NULL;
1512 new.freelist = freelist;
1515 VM_BUG_ON(new.frozen);
1518 if (!__cmpxchg_double_slab(s, page,
1520 new.freelist, new.counters,
1524 remove_partial(n, page);
1529 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1530 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1533 * Try to allocate a partial slab from a specific node.
1535 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1536 struct kmem_cache_cpu *c, gfp_t flags)
1538 struct page *page, *page2;
1539 void *object = NULL;
1542 * Racy check. If we mistakenly see no partial slabs then we
1543 * just allocate an empty slab. If we mistakenly try to get a
1544 * partial slab and there is none available then get_partials()
1547 if (!n || !n->nr_partial)
1550 spin_lock(&n->list_lock);
1551 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1555 if (!pfmemalloc_match(page, flags))
1558 t = acquire_slab(s, n, page, object == NULL);
1564 stat(s, ALLOC_FROM_PARTIAL);
1566 available = page->objects - page->inuse;
1568 available = put_cpu_partial(s, page, 0);
1569 stat(s, CPU_PARTIAL_NODE);
1571 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1575 spin_unlock(&n->list_lock);
1580 * Get a page from somewhere. Search in increasing NUMA distances.
1582 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1583 struct kmem_cache_cpu *c)
1586 struct zonelist *zonelist;
1589 enum zone_type high_zoneidx = gfp_zone(flags);
1591 unsigned int cpuset_mems_cookie;
1594 * The defrag ratio allows a configuration of the tradeoffs between
1595 * inter node defragmentation and node local allocations. A lower
1596 * defrag_ratio increases the tendency to do local allocations
1597 * instead of attempting to obtain partial slabs from other nodes.
1599 * If the defrag_ratio is set to 0 then kmalloc() always
1600 * returns node local objects. If the ratio is higher then kmalloc()
1601 * may return off node objects because partial slabs are obtained
1602 * from other nodes and filled up.
1604 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1605 * defrag_ratio = 1000) then every (well almost) allocation will
1606 * first attempt to defrag slab caches on other nodes. This means
1607 * scanning over all nodes to look for partial slabs which may be
1608 * expensive if we do it every time we are trying to find a slab
1609 * with available objects.
1611 if (!s->remote_node_defrag_ratio ||
1612 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1616 cpuset_mems_cookie = get_mems_allowed();
1617 zonelist = node_zonelist(slab_node(), flags);
1618 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1619 struct kmem_cache_node *n;
1621 n = get_node(s, zone_to_nid(zone));
1623 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1624 n->nr_partial > s->min_partial) {
1625 object = get_partial_node(s, n, c, flags);
1628 * Return the object even if
1629 * put_mems_allowed indicated that
1630 * the cpuset mems_allowed was
1631 * updated in parallel. It's a
1632 * harmless race between the alloc
1633 * and the cpuset update.
1635 put_mems_allowed(cpuset_mems_cookie);
1640 } while (!put_mems_allowed(cpuset_mems_cookie));
1646 * Get a partial page, lock it and return it.
1648 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1649 struct kmem_cache_cpu *c)
1652 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1654 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1655 if (object || node != NUMA_NO_NODE)
1658 return get_any_partial(s, flags, c);
1661 #ifdef CONFIG_PREEMPT
1663 * Calculate the next globally unique transaction for disambiguiation
1664 * during cmpxchg. The transactions start with the cpu number and are then
1665 * incremented by CONFIG_NR_CPUS.
1667 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1670 * No preemption supported therefore also no need to check for
1676 static inline unsigned long next_tid(unsigned long tid)
1678 return tid + TID_STEP;
1681 static inline unsigned int tid_to_cpu(unsigned long tid)
1683 return tid % TID_STEP;
1686 static inline unsigned long tid_to_event(unsigned long tid)
1688 return tid / TID_STEP;
1691 static inline unsigned int init_tid(int cpu)
1696 static inline void note_cmpxchg_failure(const char *n,
1697 const struct kmem_cache *s, unsigned long tid)
1699 #ifdef SLUB_DEBUG_CMPXCHG
1700 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1702 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1704 #ifdef CONFIG_PREEMPT
1705 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1706 printk("due to cpu change %d -> %d\n",
1707 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1710 if (tid_to_event(tid) != tid_to_event(actual_tid))
1711 printk("due to cpu running other code. Event %ld->%ld\n",
1712 tid_to_event(tid), tid_to_event(actual_tid));
1714 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1715 actual_tid, tid, next_tid(tid));
1717 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1720 static void init_kmem_cache_cpus(struct kmem_cache *s)
1724 for_each_possible_cpu(cpu)
1725 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1729 * Remove the cpu slab
1731 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1733 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1734 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1736 enum slab_modes l = M_NONE, m = M_NONE;
1738 int tail = DEACTIVATE_TO_HEAD;
1742 if (page->freelist) {
1743 stat(s, DEACTIVATE_REMOTE_FREES);
1744 tail = DEACTIVATE_TO_TAIL;
1748 * Stage one: Free all available per cpu objects back
1749 * to the page freelist while it is still frozen. Leave the
1752 * There is no need to take the list->lock because the page
1755 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1757 unsigned long counters;
1760 prior = page->freelist;
1761 counters = page->counters;
1762 set_freepointer(s, freelist, prior);
1763 new.counters = counters;
1765 VM_BUG_ON(!new.frozen);
1767 } while (!__cmpxchg_double_slab(s, page,
1769 freelist, new.counters,
1770 "drain percpu freelist"));
1772 freelist = nextfree;
1776 * Stage two: Ensure that the page is unfrozen while the
1777 * list presence reflects the actual number of objects
1780 * We setup the list membership and then perform a cmpxchg
1781 * with the count. If there is a mismatch then the page
1782 * is not unfrozen but the page is on the wrong list.
1784 * Then we restart the process which may have to remove
1785 * the page from the list that we just put it on again
1786 * because the number of objects in the slab may have
1791 old.freelist = page->freelist;
1792 old.counters = page->counters;
1793 VM_BUG_ON(!old.frozen);
1795 /* Determine target state of the slab */
1796 new.counters = old.counters;
1799 set_freepointer(s, freelist, old.freelist);
1800 new.freelist = freelist;
1802 new.freelist = old.freelist;
1806 if (!new.inuse && n->nr_partial > s->min_partial)
1808 else if (new.freelist) {
1813 * Taking the spinlock removes the possiblity
1814 * that acquire_slab() will see a slab page that
1817 spin_lock(&n->list_lock);
1821 if (kmem_cache_debug(s) && !lock) {
1824 * This also ensures that the scanning of full
1825 * slabs from diagnostic functions will not see
1828 spin_lock(&n->list_lock);
1836 remove_partial(n, page);
1838 else if (l == M_FULL)
1840 remove_full(s, page);
1842 if (m == M_PARTIAL) {
1844 add_partial(n, page, tail);
1847 } else if (m == M_FULL) {
1849 stat(s, DEACTIVATE_FULL);
1850 add_full(s, n, page);
1856 if (!__cmpxchg_double_slab(s, page,
1857 old.freelist, old.counters,
1858 new.freelist, new.counters,
1863 spin_unlock(&n->list_lock);
1866 stat(s, DEACTIVATE_EMPTY);
1867 discard_slab(s, page);
1873 * Unfreeze all the cpu partial slabs.
1875 * This function must be called with interrupt disabled.
1877 static void unfreeze_partials(struct kmem_cache *s)
1879 struct kmem_cache_node *n = NULL, *n2 = NULL;
1880 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1881 struct page *page, *discard_page = NULL;
1883 while ((page = c->partial)) {
1887 c->partial = page->next;
1889 n2 = get_node(s, page_to_nid(page));
1892 spin_unlock(&n->list_lock);
1895 spin_lock(&n->list_lock);
1900 old.freelist = page->freelist;
1901 old.counters = page->counters;
1902 VM_BUG_ON(!old.frozen);
1904 new.counters = old.counters;
1905 new.freelist = old.freelist;
1909 } while (!__cmpxchg_double_slab(s, page,
1910 old.freelist, old.counters,
1911 new.freelist, new.counters,
1912 "unfreezing slab"));
1914 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1915 page->next = discard_page;
1916 discard_page = page;
1918 add_partial(n, page, DEACTIVATE_TO_TAIL);
1919 stat(s, FREE_ADD_PARTIAL);
1924 spin_unlock(&n->list_lock);
1926 while (discard_page) {
1927 page = discard_page;
1928 discard_page = discard_page->next;
1930 stat(s, DEACTIVATE_EMPTY);
1931 discard_slab(s, page);
1937 * Put a page that was just frozen (in __slab_free) into a partial page
1938 * slot if available. This is done without interrupts disabled and without
1939 * preemption disabled. The cmpxchg is racy and may put the partial page
1940 * onto a random cpus partial slot.
1942 * If we did not find a slot then simply move all the partials to the
1943 * per node partial list.
1945 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1947 struct page *oldpage;
1954 oldpage = this_cpu_read(s->cpu_slab->partial);
1957 pobjects = oldpage->pobjects;
1958 pages = oldpage->pages;
1959 if (drain && pobjects > s->cpu_partial) {
1960 unsigned long flags;
1962 * partial array is full. Move the existing
1963 * set to the per node partial list.
1965 local_irq_save(flags);
1966 unfreeze_partials(s);
1967 local_irq_restore(flags);
1971 stat(s, CPU_PARTIAL_DRAIN);
1976 pobjects += page->objects - page->inuse;
1978 page->pages = pages;
1979 page->pobjects = pobjects;
1980 page->next = oldpage;
1982 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1986 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1988 stat(s, CPUSLAB_FLUSH);
1989 deactivate_slab(s, c->page, c->freelist);
1991 c->tid = next_tid(c->tid);
1999 * Called from IPI handler with interrupts disabled.
2001 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2003 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2009 unfreeze_partials(s);
2013 static void flush_cpu_slab(void *d)
2015 struct kmem_cache *s = d;
2017 __flush_cpu_slab(s, smp_processor_id());
2020 static bool has_cpu_slab(int cpu, void *info)
2022 struct kmem_cache *s = info;
2023 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2025 return c->page || c->partial;
2028 static void flush_all(struct kmem_cache *s)
2030 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2034 * Check if the objects in a per cpu structure fit numa
2035 * locality expectations.
2037 static inline int node_match(struct page *page, int node)
2040 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2046 static int count_free(struct page *page)
2048 return page->objects - page->inuse;
2051 static unsigned long count_partial(struct kmem_cache_node *n,
2052 int (*get_count)(struct page *))
2054 unsigned long flags;
2055 unsigned long x = 0;
2058 spin_lock_irqsave(&n->list_lock, flags);
2059 list_for_each_entry(page, &n->partial, lru)
2060 x += get_count(page);
2061 spin_unlock_irqrestore(&n->list_lock, flags);
2065 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2067 #ifdef CONFIG_SLUB_DEBUG
2068 return atomic_long_read(&n->total_objects);
2074 static noinline void
2075 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2080 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2082 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2083 "default order: %d, min order: %d\n", s->name, s->object_size,
2084 s->size, oo_order(s->oo), oo_order(s->min));
2086 if (oo_order(s->min) > get_order(s->object_size))
2087 printk(KERN_WARNING " %s debugging increased min order, use "
2088 "slub_debug=O to disable.\n", s->name);
2090 for_each_online_node(node) {
2091 struct kmem_cache_node *n = get_node(s, node);
2092 unsigned long nr_slabs;
2093 unsigned long nr_objs;
2094 unsigned long nr_free;
2099 nr_free = count_partial(n, count_free);
2100 nr_slabs = node_nr_slabs(n);
2101 nr_objs = node_nr_objs(n);
2104 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2105 node, nr_slabs, nr_objs, nr_free);
2109 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2110 int node, struct kmem_cache_cpu **pc)
2113 struct kmem_cache_cpu *c = *pc;
2116 freelist = get_partial(s, flags, node, c);
2121 page = new_slab(s, flags, node);
2123 c = __this_cpu_ptr(s->cpu_slab);
2128 * No other reference to the page yet so we can
2129 * muck around with it freely without cmpxchg
2131 freelist = page->freelist;
2132 page->freelist = NULL;
2134 stat(s, ALLOC_SLAB);
2143 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2145 if (unlikely(PageSlabPfmemalloc(page)))
2146 return gfp_pfmemalloc_allowed(gfpflags);
2152 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2153 * or deactivate the page.
2155 * The page is still frozen if the return value is not NULL.
2157 * If this function returns NULL then the page has been unfrozen.
2159 * This function must be called with interrupt disabled.
2161 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2164 unsigned long counters;
2168 freelist = page->freelist;
2169 counters = page->counters;
2171 new.counters = counters;
2172 VM_BUG_ON(!new.frozen);
2174 new.inuse = page->objects;
2175 new.frozen = freelist != NULL;
2177 } while (!__cmpxchg_double_slab(s, page,
2186 * Slow path. The lockless freelist is empty or we need to perform
2189 * Processing is still very fast if new objects have been freed to the
2190 * regular freelist. In that case we simply take over the regular freelist
2191 * as the lockless freelist and zap the regular freelist.
2193 * If that is not working then we fall back to the partial lists. We take the
2194 * first element of the freelist as the object to allocate now and move the
2195 * rest of the freelist to the lockless freelist.
2197 * And if we were unable to get a new slab from the partial slab lists then
2198 * we need to allocate a new slab. This is the slowest path since it involves
2199 * a call to the page allocator and the setup of a new slab.
2201 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2202 unsigned long addr, struct kmem_cache_cpu *c)
2206 unsigned long flags;
2208 local_irq_save(flags);
2209 #ifdef CONFIG_PREEMPT
2211 * We may have been preempted and rescheduled on a different
2212 * cpu before disabling interrupts. Need to reload cpu area
2215 c = this_cpu_ptr(s->cpu_slab);
2223 if (unlikely(!node_match(page, node))) {
2224 stat(s, ALLOC_NODE_MISMATCH);
2225 deactivate_slab(s, page, c->freelist);
2232 * By rights, we should be searching for a slab page that was
2233 * PFMEMALLOC but right now, we are losing the pfmemalloc
2234 * information when the page leaves the per-cpu allocator
2236 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2237 deactivate_slab(s, page, c->freelist);
2243 /* must check again c->freelist in case of cpu migration or IRQ */
2244 freelist = c->freelist;
2248 stat(s, ALLOC_SLOWPATH);
2250 freelist = get_freelist(s, page);
2254 stat(s, DEACTIVATE_BYPASS);
2258 stat(s, ALLOC_REFILL);
2262 * freelist is pointing to the list of objects to be used.
2263 * page is pointing to the page from which the objects are obtained.
2264 * That page must be frozen for per cpu allocations to work.
2266 VM_BUG_ON(!c->page->frozen);
2267 c->freelist = get_freepointer(s, freelist);
2268 c->tid = next_tid(c->tid);
2269 local_irq_restore(flags);
2275 page = c->page = c->partial;
2276 c->partial = page->next;
2277 stat(s, CPU_PARTIAL_ALLOC);
2282 freelist = new_slab_objects(s, gfpflags, node, &c);
2284 if (unlikely(!freelist)) {
2285 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2286 slab_out_of_memory(s, gfpflags, node);
2288 local_irq_restore(flags);
2293 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2296 /* Only entered in the debug case */
2297 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2298 goto new_slab; /* Slab failed checks. Next slab needed */
2300 deactivate_slab(s, page, get_freepointer(s, freelist));
2303 local_irq_restore(flags);
2308 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2309 * have the fastpath folded into their functions. So no function call
2310 * overhead for requests that can be satisfied on the fastpath.
2312 * The fastpath works by first checking if the lockless freelist can be used.
2313 * If not then __slab_alloc is called for slow processing.
2315 * Otherwise we can simply pick the next object from the lockless free list.
2317 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2318 gfp_t gfpflags, int node, unsigned long addr)
2321 struct kmem_cache_cpu *c;
2325 if (slab_pre_alloc_hook(s, gfpflags))
2331 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2332 * enabled. We may switch back and forth between cpus while
2333 * reading from one cpu area. That does not matter as long
2334 * as we end up on the original cpu again when doing the cmpxchg.
2336 c = __this_cpu_ptr(s->cpu_slab);
2339 * The transaction ids are globally unique per cpu and per operation on
2340 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2341 * occurs on the right processor and that there was no operation on the
2342 * linked list in between.
2347 object = c->freelist;
2349 if (unlikely(!object || !node_match(page, node)))
2350 object = __slab_alloc(s, gfpflags, node, addr, c);
2353 void *next_object = get_freepointer_safe(s, object);
2356 * The cmpxchg will only match if there was no additional
2357 * operation and if we are on the right processor.
2359 * The cmpxchg does the following atomically (without lock semantics!)
2360 * 1. Relocate first pointer to the current per cpu area.
2361 * 2. Verify that tid and freelist have not been changed
2362 * 3. If they were not changed replace tid and freelist
2364 * Since this is without lock semantics the protection is only against
2365 * code executing on this cpu *not* from access by other cpus.
2367 if (unlikely(!this_cpu_cmpxchg_double(
2368 s->cpu_slab->freelist, s->cpu_slab->tid,
2370 next_object, next_tid(tid)))) {
2372 note_cmpxchg_failure("slab_alloc", s, tid);
2375 prefetch_freepointer(s, next_object);
2376 stat(s, ALLOC_FASTPATH);
2379 if (unlikely(gfpflags & __GFP_ZERO) && object)
2380 memset(object, 0, s->object_size);
2382 slab_post_alloc_hook(s, gfpflags, object);
2387 static __always_inline void *slab_alloc(struct kmem_cache *s,
2388 gfp_t gfpflags, unsigned long addr)
2390 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2393 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2395 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2397 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2401 EXPORT_SYMBOL(kmem_cache_alloc);
2403 #ifdef CONFIG_TRACING
2404 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2406 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2407 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2410 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2412 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2414 void *ret = kmalloc_order(size, flags, order);
2415 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2418 EXPORT_SYMBOL(kmalloc_order_trace);
2422 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2424 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2426 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2427 s->object_size, s->size, gfpflags, node);
2431 EXPORT_SYMBOL(kmem_cache_alloc_node);
2433 #ifdef CONFIG_TRACING
2434 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2436 int node, size_t size)
2438 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2440 trace_kmalloc_node(_RET_IP_, ret,
2441 size, s->size, gfpflags, node);
2444 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2449 * Slow patch handling. This may still be called frequently since objects
2450 * have a longer lifetime than the cpu slabs in most processing loads.
2452 * So we still attempt to reduce cache line usage. Just take the slab
2453 * lock and free the item. If there is no additional partial page
2454 * handling required then we can return immediately.
2456 static void __slab_free(struct kmem_cache *s, struct page *page,
2457 void *x, unsigned long addr)
2460 void **object = (void *)x;
2463 unsigned long counters;
2464 struct kmem_cache_node *n = NULL;
2465 unsigned long uninitialized_var(flags);
2467 stat(s, FREE_SLOWPATH);
2469 if (kmem_cache_debug(s) &&
2470 !(n = free_debug_processing(s, page, x, addr, &flags)))
2475 spin_unlock_irqrestore(&n->list_lock, flags);
2478 prior = page->freelist;
2479 counters = page->counters;
2480 set_freepointer(s, object, prior);
2481 new.counters = counters;
2482 was_frozen = new.frozen;
2484 if ((!new.inuse || !prior) && !was_frozen) {
2486 if (!kmem_cache_debug(s) && !prior)
2489 * Slab was on no list before and will be partially empty
2490 * We can defer the list move and instead freeze it.
2494 else { /* Needs to be taken off a list */
2496 n = get_node(s, page_to_nid(page));
2498 * Speculatively acquire the list_lock.
2499 * If the cmpxchg does not succeed then we may
2500 * drop the list_lock without any processing.
2502 * Otherwise the list_lock will synchronize with
2503 * other processors updating the list of slabs.
2505 spin_lock_irqsave(&n->list_lock, flags);
2510 } while (!cmpxchg_double_slab(s, page,
2512 object, new.counters,
2518 * If we just froze the page then put it onto the
2519 * per cpu partial list.
2521 if (new.frozen && !was_frozen) {
2522 put_cpu_partial(s, page, 1);
2523 stat(s, CPU_PARTIAL_FREE);
2526 * The list lock was not taken therefore no list
2527 * activity can be necessary.
2530 stat(s, FREE_FROZEN);
2534 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2538 * Objects left in the slab. If it was not on the partial list before
2541 if (kmem_cache_debug(s) && unlikely(!prior)) {
2542 remove_full(s, page);
2543 add_partial(n, page, DEACTIVATE_TO_TAIL);
2544 stat(s, FREE_ADD_PARTIAL);
2546 spin_unlock_irqrestore(&n->list_lock, flags);
2552 * Slab on the partial list.
2554 remove_partial(n, page);
2555 stat(s, FREE_REMOVE_PARTIAL);
2557 /* Slab must be on the full list */
2558 remove_full(s, page);
2560 spin_unlock_irqrestore(&n->list_lock, flags);
2562 discard_slab(s, page);
2566 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2567 * can perform fastpath freeing without additional function calls.
2569 * The fastpath is only possible if we are freeing to the current cpu slab
2570 * of this processor. This typically the case if we have just allocated
2573 * If fastpath is not possible then fall back to __slab_free where we deal
2574 * with all sorts of special processing.
2576 static __always_inline void slab_free(struct kmem_cache *s,
2577 struct page *page, void *x, unsigned long addr)
2579 void **object = (void *)x;
2580 struct kmem_cache_cpu *c;
2583 slab_free_hook(s, x);
2587 * Determine the currently cpus per cpu slab.
2588 * The cpu may change afterward. However that does not matter since
2589 * data is retrieved via this pointer. If we are on the same cpu
2590 * during the cmpxchg then the free will succedd.
2592 c = __this_cpu_ptr(s->cpu_slab);
2597 if (likely(page == c->page)) {
2598 set_freepointer(s, object, c->freelist);
2600 if (unlikely(!this_cpu_cmpxchg_double(
2601 s->cpu_slab->freelist, s->cpu_slab->tid,
2603 object, next_tid(tid)))) {
2605 note_cmpxchg_failure("slab_free", s, tid);
2608 stat(s, FREE_FASTPATH);
2610 __slab_free(s, page, x, addr);
2614 void kmem_cache_free(struct kmem_cache *s, void *x)
2618 page = virt_to_head_page(x);
2620 if (kmem_cache_debug(s) && page->slab != s) {
2621 pr_err("kmem_cache_free: Wrong slab cache. %s but object"
2622 " is from %s\n", page->slab->name, s->name);
2627 slab_free(s, page, x, _RET_IP_);
2629 trace_kmem_cache_free(_RET_IP_, x);
2631 EXPORT_SYMBOL(kmem_cache_free);
2634 * Object placement in a slab is made very easy because we always start at
2635 * offset 0. If we tune the size of the object to the alignment then we can
2636 * get the required alignment by putting one properly sized object after
2639 * Notice that the allocation order determines the sizes of the per cpu
2640 * caches. Each processor has always one slab available for allocations.
2641 * Increasing the allocation order reduces the number of times that slabs
2642 * must be moved on and off the partial lists and is therefore a factor in
2647 * Mininum / Maximum order of slab pages. This influences locking overhead
2648 * and slab fragmentation. A higher order reduces the number of partial slabs
2649 * and increases the number of allocations possible without having to
2650 * take the list_lock.
2652 static int slub_min_order;
2653 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2654 static int slub_min_objects;
2657 * Merge control. If this is set then no merging of slab caches will occur.
2658 * (Could be removed. This was introduced to pacify the merge skeptics.)
2660 static int slub_nomerge;
2663 * Calculate the order of allocation given an slab object size.
2665 * The order of allocation has significant impact on performance and other
2666 * system components. Generally order 0 allocations should be preferred since
2667 * order 0 does not cause fragmentation in the page allocator. Larger objects
2668 * be problematic to put into order 0 slabs because there may be too much
2669 * unused space left. We go to a higher order if more than 1/16th of the slab
2672 * In order to reach satisfactory performance we must ensure that a minimum
2673 * number of objects is in one slab. Otherwise we may generate too much
2674 * activity on the partial lists which requires taking the list_lock. This is
2675 * less a concern for large slabs though which are rarely used.
2677 * slub_max_order specifies the order where we begin to stop considering the
2678 * number of objects in a slab as critical. If we reach slub_max_order then
2679 * we try to keep the page order as low as possible. So we accept more waste
2680 * of space in favor of a small page order.
2682 * Higher order allocations also allow the placement of more objects in a
2683 * slab and thereby reduce object handling overhead. If the user has
2684 * requested a higher mininum order then we start with that one instead of
2685 * the smallest order which will fit the object.
2687 static inline int slab_order(int size, int min_objects,
2688 int max_order, int fract_leftover, int reserved)
2692 int min_order = slub_min_order;
2694 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2695 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2697 for (order = max(min_order,
2698 fls(min_objects * size - 1) - PAGE_SHIFT);
2699 order <= max_order; order++) {
2701 unsigned long slab_size = PAGE_SIZE << order;
2703 if (slab_size < min_objects * size + reserved)
2706 rem = (slab_size - reserved) % size;
2708 if (rem <= slab_size / fract_leftover)
2716 static inline int calculate_order(int size, int reserved)
2724 * Attempt to find best configuration for a slab. This
2725 * works by first attempting to generate a layout with
2726 * the best configuration and backing off gradually.
2728 * First we reduce the acceptable waste in a slab. Then
2729 * we reduce the minimum objects required in a slab.
2731 min_objects = slub_min_objects;
2733 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2734 max_objects = order_objects(slub_max_order, size, reserved);
2735 min_objects = min(min_objects, max_objects);
2737 while (min_objects > 1) {
2739 while (fraction >= 4) {
2740 order = slab_order(size, min_objects,
2741 slub_max_order, fraction, reserved);
2742 if (order <= slub_max_order)
2750 * We were unable to place multiple objects in a slab. Now
2751 * lets see if we can place a single object there.
2753 order = slab_order(size, 1, slub_max_order, 1, reserved);
2754 if (order <= slub_max_order)
2758 * Doh this slab cannot be placed using slub_max_order.
2760 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2761 if (order < MAX_ORDER)
2767 * Figure out what the alignment of the objects will be.
2769 static unsigned long calculate_alignment(unsigned long flags,
2770 unsigned long align, unsigned long size)
2773 * If the user wants hardware cache aligned objects then follow that
2774 * suggestion if the object is sufficiently large.
2776 * The hardware cache alignment cannot override the specified
2777 * alignment though. If that is greater then use it.
2779 if (flags & SLAB_HWCACHE_ALIGN) {
2780 unsigned long ralign = cache_line_size();
2781 while (size <= ralign / 2)
2783 align = max(align, ralign);
2786 if (align < ARCH_SLAB_MINALIGN)
2787 align = ARCH_SLAB_MINALIGN;
2789 return ALIGN(align, sizeof(void *));
2793 init_kmem_cache_node(struct kmem_cache_node *n)
2796 spin_lock_init(&n->list_lock);
2797 INIT_LIST_HEAD(&n->partial);
2798 #ifdef CONFIG_SLUB_DEBUG
2799 atomic_long_set(&n->nr_slabs, 0);
2800 atomic_long_set(&n->total_objects, 0);
2801 INIT_LIST_HEAD(&n->full);
2805 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2807 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2808 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2811 * Must align to double word boundary for the double cmpxchg
2812 * instructions to work; see __pcpu_double_call_return_bool().
2814 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2815 2 * sizeof(void *));
2820 init_kmem_cache_cpus(s);
2825 static struct kmem_cache *kmem_cache_node;
2828 * No kmalloc_node yet so do it by hand. We know that this is the first
2829 * slab on the node for this slabcache. There are no concurrent accesses
2832 * Note that this function only works on the kmalloc_node_cache
2833 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2834 * memory on a fresh node that has no slab structures yet.
2836 static void early_kmem_cache_node_alloc(int node)
2839 struct kmem_cache_node *n;
2841 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2843 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2846 if (page_to_nid(page) != node) {
2847 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2849 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2850 "in order to be able to continue\n");
2855 page->freelist = get_freepointer(kmem_cache_node, n);
2858 kmem_cache_node->node[node] = n;
2859 #ifdef CONFIG_SLUB_DEBUG
2860 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2861 init_tracking(kmem_cache_node, n);
2863 init_kmem_cache_node(n);
2864 inc_slabs_node(kmem_cache_node, node, page->objects);
2866 add_partial(n, page, DEACTIVATE_TO_HEAD);
2869 static void free_kmem_cache_nodes(struct kmem_cache *s)
2873 for_each_node_state(node, N_NORMAL_MEMORY) {
2874 struct kmem_cache_node *n = s->node[node];
2877 kmem_cache_free(kmem_cache_node, n);
2879 s->node[node] = NULL;
2883 static int init_kmem_cache_nodes(struct kmem_cache *s)
2887 for_each_node_state(node, N_NORMAL_MEMORY) {
2888 struct kmem_cache_node *n;
2890 if (slab_state == DOWN) {
2891 early_kmem_cache_node_alloc(node);
2894 n = kmem_cache_alloc_node(kmem_cache_node,
2898 free_kmem_cache_nodes(s);
2903 init_kmem_cache_node(n);
2908 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2910 if (min < MIN_PARTIAL)
2912 else if (min > MAX_PARTIAL)
2914 s->min_partial = min;
2918 * calculate_sizes() determines the order and the distribution of data within
2921 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2923 unsigned long flags = s->flags;
2924 unsigned long size = s->object_size;
2925 unsigned long align = s->align;
2929 * Round up object size to the next word boundary. We can only
2930 * place the free pointer at word boundaries and this determines
2931 * the possible location of the free pointer.
2933 size = ALIGN(size, sizeof(void *));
2935 #ifdef CONFIG_SLUB_DEBUG
2937 * Determine if we can poison the object itself. If the user of
2938 * the slab may touch the object after free or before allocation
2939 * then we should never poison the object itself.
2941 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2943 s->flags |= __OBJECT_POISON;
2945 s->flags &= ~__OBJECT_POISON;
2949 * If we are Redzoning then check if there is some space between the
2950 * end of the object and the free pointer. If not then add an
2951 * additional word to have some bytes to store Redzone information.
2953 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2954 size += sizeof(void *);
2958 * With that we have determined the number of bytes in actual use
2959 * by the object. This is the potential offset to the free pointer.
2963 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2966 * Relocate free pointer after the object if it is not
2967 * permitted to overwrite the first word of the object on
2970 * This is the case if we do RCU, have a constructor or
2971 * destructor or are poisoning the objects.
2974 size += sizeof(void *);
2977 #ifdef CONFIG_SLUB_DEBUG
2978 if (flags & SLAB_STORE_USER)
2980 * Need to store information about allocs and frees after
2983 size += 2 * sizeof(struct track);
2985 if (flags & SLAB_RED_ZONE)
2987 * Add some empty padding so that we can catch
2988 * overwrites from earlier objects rather than let
2989 * tracking information or the free pointer be
2990 * corrupted if a user writes before the start
2993 size += sizeof(void *);
2997 * Determine the alignment based on various parameters that the
2998 * user specified and the dynamic determination of cache line size
3001 align = calculate_alignment(flags, align, s->object_size);
3005 * SLUB stores one object immediately after another beginning from
3006 * offset 0. In order to align the objects we have to simply size
3007 * each object to conform to the alignment.
3009 size = ALIGN(size, align);
3011 if (forced_order >= 0)
3012 order = forced_order;
3014 order = calculate_order(size, s->reserved);
3021 s->allocflags |= __GFP_COMP;
3023 if (s->flags & SLAB_CACHE_DMA)
3024 s->allocflags |= SLUB_DMA;
3026 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3027 s->allocflags |= __GFP_RECLAIMABLE;
3030 * Determine the number of objects per slab
3032 s->oo = oo_make(order, size, s->reserved);
3033 s->min = oo_make(get_order(size), size, s->reserved);
3034 if (oo_objects(s->oo) > oo_objects(s->max))
3037 return !!oo_objects(s->oo);
3041 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3043 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3046 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3047 s->reserved = sizeof(struct rcu_head);
3049 if (!calculate_sizes(s, -1))
3051 if (disable_higher_order_debug) {
3053 * Disable debugging flags that store metadata if the min slab
3056 if (get_order(s->size) > get_order(s->object_size)) {
3057 s->flags &= ~DEBUG_METADATA_FLAGS;
3059 if (!calculate_sizes(s, -1))
3064 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3065 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3066 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3067 /* Enable fast mode */
3068 s->flags |= __CMPXCHG_DOUBLE;
3072 * The larger the object size is, the more pages we want on the partial
3073 * list to avoid pounding the page allocator excessively.
3075 set_min_partial(s, ilog2(s->size) / 2);
3078 * cpu_partial determined the maximum number of objects kept in the
3079 * per cpu partial lists of a processor.
3081 * Per cpu partial lists mainly contain slabs that just have one
3082 * object freed. If they are used for allocation then they can be
3083 * filled up again with minimal effort. The slab will never hit the
3084 * per node partial lists and therefore no locking will be required.
3086 * This setting also determines
3088 * A) The number of objects from per cpu partial slabs dumped to the
3089 * per node list when we reach the limit.
3090 * B) The number of objects in cpu partial slabs to extract from the
3091 * per node list when we run out of per cpu objects. We only fetch 50%
3092 * to keep some capacity around for frees.
3094 if (kmem_cache_debug(s))
3096 else if (s->size >= PAGE_SIZE)
3098 else if (s->size >= 1024)
3100 else if (s->size >= 256)
3101 s->cpu_partial = 13;
3103 s->cpu_partial = 30;
3106 s->remote_node_defrag_ratio = 1000;
3108 if (!init_kmem_cache_nodes(s))
3111 if (alloc_kmem_cache_cpus(s))
3114 free_kmem_cache_nodes(s);
3116 if (flags & SLAB_PANIC)
3117 panic("Cannot create slab %s size=%lu realsize=%u "
3118 "order=%u offset=%u flags=%lx\n",
3119 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3125 * Determine the size of a slab object
3127 unsigned int kmem_cache_size(struct kmem_cache *s)
3129 return s->object_size;
3131 EXPORT_SYMBOL(kmem_cache_size);
3133 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3136 #ifdef CONFIG_SLUB_DEBUG
3137 void *addr = page_address(page);
3139 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3140 sizeof(long), GFP_ATOMIC);
3143 slab_err(s, page, text, s->name);
3146 get_map(s, page, map);
3147 for_each_object(p, s, addr, page->objects) {
3149 if (!test_bit(slab_index(p, s, addr), map)) {
3150 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3152 print_tracking(s, p);
3161 * Attempt to free all partial slabs on a node.
3162 * This is called from kmem_cache_close(). We must be the last thread
3163 * using the cache and therefore we do not need to lock anymore.
3165 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3167 struct page *page, *h;
3169 list_for_each_entry_safe(page, h, &n->partial, lru) {
3171 remove_partial(n, page);
3172 discard_slab(s, page);
3174 list_slab_objects(s, page,
3175 "Objects remaining in %s on kmem_cache_close()");
3181 * Release all resources used by a slab cache.
3183 static inline int kmem_cache_close(struct kmem_cache *s)
3188 /* Attempt to free all objects */
3189 for_each_node_state(node, N_NORMAL_MEMORY) {
3190 struct kmem_cache_node *n = get_node(s, node);
3193 if (n->nr_partial || slabs_node(s, node))
3196 free_percpu(s->cpu_slab);
3197 free_kmem_cache_nodes(s);
3201 int __kmem_cache_shutdown(struct kmem_cache *s)
3203 int rc = kmem_cache_close(s);
3206 sysfs_slab_remove(s);
3211 /********************************************************************
3213 *******************************************************************/
3215 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3216 EXPORT_SYMBOL(kmalloc_caches);
3218 #ifdef CONFIG_ZONE_DMA
3219 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3222 static int __init setup_slub_min_order(char *str)
3224 get_option(&str, &slub_min_order);
3229 __setup("slub_min_order=", setup_slub_min_order);
3231 static int __init setup_slub_max_order(char *str)
3233 get_option(&str, &slub_max_order);
3234 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3239 __setup("slub_max_order=", setup_slub_max_order);
3241 static int __init setup_slub_min_objects(char *str)
3243 get_option(&str, &slub_min_objects);
3248 __setup("slub_min_objects=", setup_slub_min_objects);
3250 static int __init setup_slub_nomerge(char *str)
3256 __setup("slub_nomerge", setup_slub_nomerge);
3258 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3259 int size, unsigned int flags)
3261 struct kmem_cache *s;
3263 s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3266 s->size = s->object_size = size;
3267 s->align = ARCH_KMALLOC_MINALIGN;
3270 * This function is called with IRQs disabled during early-boot on
3271 * single CPU so there's no need to take slab_mutex here.
3273 if (kmem_cache_open(s, flags))
3276 list_add(&s->list, &slab_caches);
3280 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3285 * Conversion table for small slabs sizes / 8 to the index in the
3286 * kmalloc array. This is necessary for slabs < 192 since we have non power
3287 * of two cache sizes there. The size of larger slabs can be determined using
3290 static s8 size_index[24] = {
3317 static inline int size_index_elem(size_t bytes)
3319 return (bytes - 1) / 8;
3322 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3328 return ZERO_SIZE_PTR;
3330 index = size_index[size_index_elem(size)];
3332 index = fls(size - 1);
3334 #ifdef CONFIG_ZONE_DMA
3335 if (unlikely((flags & SLUB_DMA)))
3336 return kmalloc_dma_caches[index];
3339 return kmalloc_caches[index];
3342 void *__kmalloc(size_t size, gfp_t flags)
3344 struct kmem_cache *s;
3347 if (unlikely(size > SLUB_MAX_SIZE))
3348 return kmalloc_large(size, flags);
3350 s = get_slab(size, flags);
3352 if (unlikely(ZERO_OR_NULL_PTR(s)))
3355 ret = slab_alloc(s, flags, _RET_IP_);
3357 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3361 EXPORT_SYMBOL(__kmalloc);
3364 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3369 flags |= __GFP_COMP | __GFP_NOTRACK;
3370 page = alloc_pages_node(node, flags, get_order(size));
3372 ptr = page_address(page);
3374 kmemleak_alloc(ptr, size, 1, flags);
3378 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3380 struct kmem_cache *s;
3383 if (unlikely(size > SLUB_MAX_SIZE)) {
3384 ret = kmalloc_large_node(size, flags, node);
3386 trace_kmalloc_node(_RET_IP_, ret,
3387 size, PAGE_SIZE << get_order(size),
3393 s = get_slab(size, flags);
3395 if (unlikely(ZERO_OR_NULL_PTR(s)))
3398 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3400 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3404 EXPORT_SYMBOL(__kmalloc_node);
3407 size_t ksize(const void *object)
3411 if (unlikely(object == ZERO_SIZE_PTR))
3414 page = virt_to_head_page(object);
3416 if (unlikely(!PageSlab(page))) {
3417 WARN_ON(!PageCompound(page));
3418 return PAGE_SIZE << compound_order(page);
3421 return slab_ksize(page->slab);
3423 EXPORT_SYMBOL(ksize);
3425 #ifdef CONFIG_SLUB_DEBUG
3426 bool verify_mem_not_deleted(const void *x)
3429 void *object = (void *)x;
3430 unsigned long flags;
3433 if (unlikely(ZERO_OR_NULL_PTR(x)))
3436 local_irq_save(flags);
3438 page = virt_to_head_page(x);
3439 if (unlikely(!PageSlab(page))) {
3440 /* maybe it was from stack? */
3446 if (on_freelist(page->slab, page, object)) {
3447 object_err(page->slab, page, object, "Object is on free-list");
3455 local_irq_restore(flags);
3458 EXPORT_SYMBOL(verify_mem_not_deleted);
3461 void kfree(const void *x)
3464 void *object = (void *)x;
3466 trace_kfree(_RET_IP_, x);
3468 if (unlikely(ZERO_OR_NULL_PTR(x)))
3471 page = virt_to_head_page(x);
3472 if (unlikely(!PageSlab(page))) {
3473 BUG_ON(!PageCompound(page));
3475 __free_pages(page, compound_order(page));
3478 slab_free(page->slab, page, object, _RET_IP_);
3480 EXPORT_SYMBOL(kfree);
3483 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3484 * the remaining slabs by the number of items in use. The slabs with the
3485 * most items in use come first. New allocations will then fill those up
3486 * and thus they can be removed from the partial lists.
3488 * The slabs with the least items are placed last. This results in them
3489 * being allocated from last increasing the chance that the last objects
3490 * are freed in them.
3492 int kmem_cache_shrink(struct kmem_cache *s)
3496 struct kmem_cache_node *n;
3499 int objects = oo_objects(s->max);
3500 struct list_head *slabs_by_inuse =
3501 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3502 unsigned long flags;
3504 if (!slabs_by_inuse)
3508 for_each_node_state(node, N_NORMAL_MEMORY) {
3509 n = get_node(s, node);
3514 for (i = 0; i < objects; i++)
3515 INIT_LIST_HEAD(slabs_by_inuse + i);
3517 spin_lock_irqsave(&n->list_lock, flags);
3520 * Build lists indexed by the items in use in each slab.
3522 * Note that concurrent frees may occur while we hold the
3523 * list_lock. page->inuse here is the upper limit.
3525 list_for_each_entry_safe(page, t, &n->partial, lru) {
3526 list_move(&page->lru, slabs_by_inuse + page->inuse);
3532 * Rebuild the partial list with the slabs filled up most
3533 * first and the least used slabs at the end.
3535 for (i = objects - 1; i > 0; i--)
3536 list_splice(slabs_by_inuse + i, n->partial.prev);
3538 spin_unlock_irqrestore(&n->list_lock, flags);
3540 /* Release empty slabs */
3541 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3542 discard_slab(s, page);
3545 kfree(slabs_by_inuse);
3548 EXPORT_SYMBOL(kmem_cache_shrink);
3550 #if defined(CONFIG_MEMORY_HOTPLUG)
3551 static int slab_mem_going_offline_callback(void *arg)
3553 struct kmem_cache *s;
3555 mutex_lock(&slab_mutex);
3556 list_for_each_entry(s, &slab_caches, list)
3557 kmem_cache_shrink(s);
3558 mutex_unlock(&slab_mutex);
3563 static void slab_mem_offline_callback(void *arg)
3565 struct kmem_cache_node *n;
3566 struct kmem_cache *s;
3567 struct memory_notify *marg = arg;
3570 offline_node = marg->status_change_nid;
3573 * If the node still has available memory. we need kmem_cache_node
3576 if (offline_node < 0)
3579 mutex_lock(&slab_mutex);
3580 list_for_each_entry(s, &slab_caches, list) {
3581 n = get_node(s, offline_node);
3584 * if n->nr_slabs > 0, slabs still exist on the node
3585 * that is going down. We were unable to free them,
3586 * and offline_pages() function shouldn't call this
3587 * callback. So, we must fail.
3589 BUG_ON(slabs_node(s, offline_node));
3591 s->node[offline_node] = NULL;
3592 kmem_cache_free(kmem_cache_node, n);
3595 mutex_unlock(&slab_mutex);
3598 static int slab_mem_going_online_callback(void *arg)
3600 struct kmem_cache_node *n;
3601 struct kmem_cache *s;
3602 struct memory_notify *marg = arg;
3603 int nid = marg->status_change_nid;
3607 * If the node's memory is already available, then kmem_cache_node is
3608 * already created. Nothing to do.
3614 * We are bringing a node online. No memory is available yet. We must
3615 * allocate a kmem_cache_node structure in order to bring the node
3618 mutex_lock(&slab_mutex);
3619 list_for_each_entry(s, &slab_caches, list) {
3621 * XXX: kmem_cache_alloc_node will fallback to other nodes
3622 * since memory is not yet available from the node that
3625 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3630 init_kmem_cache_node(n);
3634 mutex_unlock(&slab_mutex);
3638 static int slab_memory_callback(struct notifier_block *self,
3639 unsigned long action, void *arg)
3644 case MEM_GOING_ONLINE:
3645 ret = slab_mem_going_online_callback(arg);
3647 case MEM_GOING_OFFLINE:
3648 ret = slab_mem_going_offline_callback(arg);
3651 case MEM_CANCEL_ONLINE:
3652 slab_mem_offline_callback(arg);
3655 case MEM_CANCEL_OFFLINE:
3659 ret = notifier_from_errno(ret);
3665 #endif /* CONFIG_MEMORY_HOTPLUG */
3667 /********************************************************************
3668 * Basic setup of slabs
3669 *******************************************************************/
3672 * Used for early kmem_cache structures that were allocated using
3673 * the page allocator
3676 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3680 list_add(&s->list, &slab_caches);
3683 for_each_node_state(node, N_NORMAL_MEMORY) {
3684 struct kmem_cache_node *n = get_node(s, node);
3688 list_for_each_entry(p, &n->partial, lru)
3691 #ifdef CONFIG_SLUB_DEBUG
3692 list_for_each_entry(p, &n->full, lru)
3699 void __init kmem_cache_init(void)
3703 struct kmem_cache *temp_kmem_cache;
3705 struct kmem_cache *temp_kmem_cache_node;
3706 unsigned long kmalloc_size;
3708 if (debug_guardpage_minorder())
3711 kmem_size = offsetof(struct kmem_cache, node) +
3712 nr_node_ids * sizeof(struct kmem_cache_node *);
3714 /* Allocate two kmem_caches from the page allocator */
3715 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3716 order = get_order(2 * kmalloc_size);
3717 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT | __GFP_ZERO, order);
3720 * Must first have the slab cache available for the allocations of the
3721 * struct kmem_cache_node's. There is special bootstrap code in
3722 * kmem_cache_open for slab_state == DOWN.
3724 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3726 kmem_cache_node->name = "kmem_cache_node";
3727 kmem_cache_node->size = kmem_cache_node->object_size =
3728 sizeof(struct kmem_cache_node);
3729 kmem_cache_open(kmem_cache_node, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3731 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3733 /* Able to allocate the per node structures */
3734 slab_state = PARTIAL;
3736 temp_kmem_cache = kmem_cache;
3737 kmem_cache->name = "kmem_cache";
3738 kmem_cache->size = kmem_cache->object_size = kmem_size;
3739 kmem_cache_open(kmem_cache, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3741 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3742 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3745 * Allocate kmem_cache_node properly from the kmem_cache slab.
3746 * kmem_cache_node is separately allocated so no need to
3747 * update any list pointers.
3749 temp_kmem_cache_node = kmem_cache_node;
3751 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3752 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3754 kmem_cache_bootstrap_fixup(kmem_cache_node);
3757 kmem_cache_bootstrap_fixup(kmem_cache);
3759 /* Free temporary boot structure */
3760 free_pages((unsigned long)temp_kmem_cache, order);
3762 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3765 * Patch up the size_index table if we have strange large alignment
3766 * requirements for the kmalloc array. This is only the case for
3767 * MIPS it seems. The standard arches will not generate any code here.
3769 * Largest permitted alignment is 256 bytes due to the way we
3770 * handle the index determination for the smaller caches.
3772 * Make sure that nothing crazy happens if someone starts tinkering
3773 * around with ARCH_KMALLOC_MINALIGN
3775 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3776 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3778 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3779 int elem = size_index_elem(i);
3780 if (elem >= ARRAY_SIZE(size_index))
3782 size_index[elem] = KMALLOC_SHIFT_LOW;
3785 if (KMALLOC_MIN_SIZE == 64) {
3787 * The 96 byte size cache is not used if the alignment
3790 for (i = 64 + 8; i <= 96; i += 8)
3791 size_index[size_index_elem(i)] = 7;
3792 } else if (KMALLOC_MIN_SIZE == 128) {
3794 * The 192 byte sized cache is not used if the alignment
3795 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3798 for (i = 128 + 8; i <= 192; i += 8)
3799 size_index[size_index_elem(i)] = 8;
3802 /* Caches that are not of the two-to-the-power-of size */
3803 if (KMALLOC_MIN_SIZE <= 32) {
3804 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3808 if (KMALLOC_MIN_SIZE <= 64) {
3809 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3813 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3814 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3820 /* Provide the correct kmalloc names now that the caches are up */
3821 if (KMALLOC_MIN_SIZE <= 32) {
3822 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3823 BUG_ON(!kmalloc_caches[1]->name);
3826 if (KMALLOC_MIN_SIZE <= 64) {
3827 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3828 BUG_ON(!kmalloc_caches[2]->name);
3831 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3832 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3835 kmalloc_caches[i]->name = s;
3839 register_cpu_notifier(&slab_notifier);
3842 #ifdef CONFIG_ZONE_DMA
3843 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3844 struct kmem_cache *s = kmalloc_caches[i];
3847 char *name = kasprintf(GFP_NOWAIT,
3848 "dma-kmalloc-%d", s->object_size);
3851 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3852 s->object_size, SLAB_CACHE_DMA);
3857 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3858 " CPUs=%d, Nodes=%d\n",
3859 caches, cache_line_size(),
3860 slub_min_order, slub_max_order, slub_min_objects,
3861 nr_cpu_ids, nr_node_ids);
3864 void __init kmem_cache_init_late(void)
3869 * Find a mergeable slab cache
3871 static int slab_unmergeable(struct kmem_cache *s)
3873 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3880 * We may have set a slab to be unmergeable during bootstrap.
3882 if (s->refcount < 0)
3888 static struct kmem_cache *find_mergeable(size_t size,
3889 size_t align, unsigned long flags, const char *name,
3890 void (*ctor)(void *))
3892 struct kmem_cache *s;
3894 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3900 size = ALIGN(size, sizeof(void *));
3901 align = calculate_alignment(flags, align, size);
3902 size = ALIGN(size, align);
3903 flags = kmem_cache_flags(size, flags, name, NULL);
3905 list_for_each_entry(s, &slab_caches, list) {
3906 if (slab_unmergeable(s))
3912 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3915 * Check if alignment is compatible.
3916 * Courtesy of Adrian Drzewiecki
3918 if ((s->size & ~(align - 1)) != s->size)
3921 if (s->size - size >= sizeof(void *))
3929 struct kmem_cache *__kmem_cache_alias(const char *name, size_t size,
3930 size_t align, unsigned long flags, void (*ctor)(void *))
3932 struct kmem_cache *s;
3934 s = find_mergeable(size, align, flags, name, ctor);
3938 * Adjust the object sizes so that we clear
3939 * the complete object on kzalloc.
3941 s->object_size = max(s->object_size, (int)size);
3942 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3944 if (sysfs_slab_alias(s, name)) {
3953 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3957 err = kmem_cache_open(s, flags);
3961 mutex_unlock(&slab_mutex);
3962 err = sysfs_slab_add(s);
3963 mutex_lock(&slab_mutex);
3966 kmem_cache_close(s);
3973 * Use the cpu notifier to insure that the cpu slabs are flushed when
3976 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3977 unsigned long action, void *hcpu)
3979 long cpu = (long)hcpu;
3980 struct kmem_cache *s;
3981 unsigned long flags;
3984 case CPU_UP_CANCELED:
3985 case CPU_UP_CANCELED_FROZEN:
3987 case CPU_DEAD_FROZEN:
3988 mutex_lock(&slab_mutex);
3989 list_for_each_entry(s, &slab_caches, list) {
3990 local_irq_save(flags);
3991 __flush_cpu_slab(s, cpu);
3992 local_irq_restore(flags);
3994 mutex_unlock(&slab_mutex);
4002 static struct notifier_block __cpuinitdata slab_notifier = {
4003 .notifier_call = slab_cpuup_callback
4008 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4010 struct kmem_cache *s;
4013 if (unlikely(size > SLUB_MAX_SIZE))
4014 return kmalloc_large(size, gfpflags);
4016 s = get_slab(size, gfpflags);
4018 if (unlikely(ZERO_OR_NULL_PTR(s)))
4021 ret = slab_alloc(s, gfpflags, caller);
4023 /* Honor the call site pointer we received. */
4024 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4030 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4031 int node, unsigned long caller)
4033 struct kmem_cache *s;
4036 if (unlikely(size > SLUB_MAX_SIZE)) {
4037 ret = kmalloc_large_node(size, gfpflags, node);
4039 trace_kmalloc_node(caller, ret,
4040 size, PAGE_SIZE << get_order(size),
4046 s = get_slab(size, gfpflags);
4048 if (unlikely(ZERO_OR_NULL_PTR(s)))
4051 ret = slab_alloc_node(s, gfpflags, node, caller);
4053 /* Honor the call site pointer we received. */
4054 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4061 static int count_inuse(struct page *page)
4066 static int count_total(struct page *page)
4068 return page->objects;
4072 #ifdef CONFIG_SLUB_DEBUG
4073 static int validate_slab(struct kmem_cache *s, struct page *page,
4077 void *addr = page_address(page);
4079 if (!check_slab(s, page) ||
4080 !on_freelist(s, page, NULL))
4083 /* Now we know that a valid freelist exists */
4084 bitmap_zero(map, page->objects);
4086 get_map(s, page, map);
4087 for_each_object(p, s, addr, page->objects) {
4088 if (test_bit(slab_index(p, s, addr), map))
4089 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4093 for_each_object(p, s, addr, page->objects)
4094 if (!test_bit(slab_index(p, s, addr), map))
4095 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4100 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4104 validate_slab(s, page, map);
4108 static int validate_slab_node(struct kmem_cache *s,
4109 struct kmem_cache_node *n, unsigned long *map)
4111 unsigned long count = 0;
4113 unsigned long flags;
4115 spin_lock_irqsave(&n->list_lock, flags);
4117 list_for_each_entry(page, &n->partial, lru) {
4118 validate_slab_slab(s, page, map);
4121 if (count != n->nr_partial)
4122 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4123 "counter=%ld\n", s->name, count, n->nr_partial);
4125 if (!(s->flags & SLAB_STORE_USER))
4128 list_for_each_entry(page, &n->full, lru) {
4129 validate_slab_slab(s, page, map);
4132 if (count != atomic_long_read(&n->nr_slabs))
4133 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4134 "counter=%ld\n", s->name, count,
4135 atomic_long_read(&n->nr_slabs));
4138 spin_unlock_irqrestore(&n->list_lock, flags);
4142 static long validate_slab_cache(struct kmem_cache *s)
4145 unsigned long count = 0;
4146 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4147 sizeof(unsigned long), GFP_KERNEL);
4153 for_each_node_state(node, N_NORMAL_MEMORY) {
4154 struct kmem_cache_node *n = get_node(s, node);
4156 count += validate_slab_node(s, n, map);
4162 * Generate lists of code addresses where slabcache objects are allocated
4167 unsigned long count;
4174 DECLARE_BITMAP(cpus, NR_CPUS);
4180 unsigned long count;
4181 struct location *loc;
4184 static void free_loc_track(struct loc_track *t)
4187 free_pages((unsigned long)t->loc,
4188 get_order(sizeof(struct location) * t->max));
4191 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4196 order = get_order(sizeof(struct location) * max);
4198 l = (void *)__get_free_pages(flags, order);
4203 memcpy(l, t->loc, sizeof(struct location) * t->count);
4211 static int add_location(struct loc_track *t, struct kmem_cache *s,
4212 const struct track *track)
4214 long start, end, pos;
4216 unsigned long caddr;
4217 unsigned long age = jiffies - track->when;
4223 pos = start + (end - start + 1) / 2;
4226 * There is nothing at "end". If we end up there
4227 * we need to add something to before end.
4232 caddr = t->loc[pos].addr;
4233 if (track->addr == caddr) {
4239 if (age < l->min_time)
4241 if (age > l->max_time)
4244 if (track->pid < l->min_pid)
4245 l->min_pid = track->pid;
4246 if (track->pid > l->max_pid)
4247 l->max_pid = track->pid;
4249 cpumask_set_cpu(track->cpu,
4250 to_cpumask(l->cpus));
4252 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4256 if (track->addr < caddr)
4263 * Not found. Insert new tracking element.
4265 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4271 (t->count - pos) * sizeof(struct location));
4274 l->addr = track->addr;
4278 l->min_pid = track->pid;
4279 l->max_pid = track->pid;
4280 cpumask_clear(to_cpumask(l->cpus));
4281 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4282 nodes_clear(l->nodes);
4283 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4287 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4288 struct page *page, enum track_item alloc,
4291 void *addr = page_address(page);
4294 bitmap_zero(map, page->objects);
4295 get_map(s, page, map);
4297 for_each_object(p, s, addr, page->objects)
4298 if (!test_bit(slab_index(p, s, addr), map))
4299 add_location(t, s, get_track(s, p, alloc));
4302 static int list_locations(struct kmem_cache *s, char *buf,
4303 enum track_item alloc)
4307 struct loc_track t = { 0, 0, NULL };
4309 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4310 sizeof(unsigned long), GFP_KERNEL);
4312 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4315 return sprintf(buf, "Out of memory\n");
4317 /* Push back cpu slabs */
4320 for_each_node_state(node, N_NORMAL_MEMORY) {
4321 struct kmem_cache_node *n = get_node(s, node);
4322 unsigned long flags;
4325 if (!atomic_long_read(&n->nr_slabs))
4328 spin_lock_irqsave(&n->list_lock, flags);
4329 list_for_each_entry(page, &n->partial, lru)
4330 process_slab(&t, s, page, alloc, map);
4331 list_for_each_entry(page, &n->full, lru)
4332 process_slab(&t, s, page, alloc, map);
4333 spin_unlock_irqrestore(&n->list_lock, flags);
4336 for (i = 0; i < t.count; i++) {
4337 struct location *l = &t.loc[i];
4339 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4341 len += sprintf(buf + len, "%7ld ", l->count);
4344 len += sprintf(buf + len, "%pS", (void *)l->addr);
4346 len += sprintf(buf + len, "<not-available>");
4348 if (l->sum_time != l->min_time) {
4349 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4351 (long)div_u64(l->sum_time, l->count),
4354 len += sprintf(buf + len, " age=%ld",
4357 if (l->min_pid != l->max_pid)
4358 len += sprintf(buf + len, " pid=%ld-%ld",
4359 l->min_pid, l->max_pid);
4361 len += sprintf(buf + len, " pid=%ld",
4364 if (num_online_cpus() > 1 &&
4365 !cpumask_empty(to_cpumask(l->cpus)) &&
4366 len < PAGE_SIZE - 60) {
4367 len += sprintf(buf + len, " cpus=");
4368 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4369 to_cpumask(l->cpus));
4372 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4373 len < PAGE_SIZE - 60) {
4374 len += sprintf(buf + len, " nodes=");
4375 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4379 len += sprintf(buf + len, "\n");
4385 len += sprintf(buf, "No data\n");
4390 #ifdef SLUB_RESILIENCY_TEST
4391 static void resiliency_test(void)
4395 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4397 printk(KERN_ERR "SLUB resiliency testing\n");
4398 printk(KERN_ERR "-----------------------\n");
4399 printk(KERN_ERR "A. Corruption after allocation\n");
4401 p = kzalloc(16, GFP_KERNEL);
4403 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4404 " 0x12->0x%p\n\n", p + 16);
4406 validate_slab_cache(kmalloc_caches[4]);
4408 /* Hmmm... The next two are dangerous */
4409 p = kzalloc(32, GFP_KERNEL);
4410 p[32 + sizeof(void *)] = 0x34;
4411 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4412 " 0x34 -> -0x%p\n", p);
4414 "If allocated object is overwritten then not detectable\n\n");
4416 validate_slab_cache(kmalloc_caches[5]);
4417 p = kzalloc(64, GFP_KERNEL);
4418 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4420 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4423 "If allocated object is overwritten then not detectable\n\n");
4424 validate_slab_cache(kmalloc_caches[6]);
4426 printk(KERN_ERR "\nB. Corruption after free\n");
4427 p = kzalloc(128, GFP_KERNEL);
4430 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4431 validate_slab_cache(kmalloc_caches[7]);
4433 p = kzalloc(256, GFP_KERNEL);
4436 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4438 validate_slab_cache(kmalloc_caches[8]);
4440 p = kzalloc(512, GFP_KERNEL);
4443 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4444 validate_slab_cache(kmalloc_caches[9]);
4448 static void resiliency_test(void) {};
4453 enum slab_stat_type {
4454 SL_ALL, /* All slabs */
4455 SL_PARTIAL, /* Only partially allocated slabs */
4456 SL_CPU, /* Only slabs used for cpu caches */
4457 SL_OBJECTS, /* Determine allocated objects not slabs */
4458 SL_TOTAL /* Determine object capacity not slabs */
4461 #define SO_ALL (1 << SL_ALL)
4462 #define SO_PARTIAL (1 << SL_PARTIAL)
4463 #define SO_CPU (1 << SL_CPU)
4464 #define SO_OBJECTS (1 << SL_OBJECTS)
4465 #define SO_TOTAL (1 << SL_TOTAL)
4467 static ssize_t show_slab_objects(struct kmem_cache *s,
4468 char *buf, unsigned long flags)
4470 unsigned long total = 0;
4473 unsigned long *nodes;
4474 unsigned long *per_cpu;
4476 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4479 per_cpu = nodes + nr_node_ids;
4481 if (flags & SO_CPU) {
4484 for_each_possible_cpu(cpu) {
4485 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4489 page = ACCESS_ONCE(c->page);
4493 node = page_to_nid(page);
4494 if (flags & SO_TOTAL)
4496 else if (flags & SO_OBJECTS)
4504 page = ACCESS_ONCE(c->partial);
4515 lock_memory_hotplug();
4516 #ifdef CONFIG_SLUB_DEBUG
4517 if (flags & SO_ALL) {
4518 for_each_node_state(node, N_NORMAL_MEMORY) {
4519 struct kmem_cache_node *n = get_node(s, node);
4521 if (flags & SO_TOTAL)
4522 x = atomic_long_read(&n->total_objects);
4523 else if (flags & SO_OBJECTS)
4524 x = atomic_long_read(&n->total_objects) -
4525 count_partial(n, count_free);
4528 x = atomic_long_read(&n->nr_slabs);
4535 if (flags & SO_PARTIAL) {
4536 for_each_node_state(node, N_NORMAL_MEMORY) {
4537 struct kmem_cache_node *n = get_node(s, node);
4539 if (flags & SO_TOTAL)
4540 x = count_partial(n, count_total);
4541 else if (flags & SO_OBJECTS)
4542 x = count_partial(n, count_inuse);
4549 x = sprintf(buf, "%lu", total);
4551 for_each_node_state(node, N_NORMAL_MEMORY)
4553 x += sprintf(buf + x, " N%d=%lu",
4556 unlock_memory_hotplug();
4558 return x + sprintf(buf + x, "\n");
4561 #ifdef CONFIG_SLUB_DEBUG
4562 static int any_slab_objects(struct kmem_cache *s)
4566 for_each_online_node(node) {
4567 struct kmem_cache_node *n = get_node(s, node);
4572 if (atomic_long_read(&n->total_objects))
4579 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4580 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4582 struct slab_attribute {
4583 struct attribute attr;
4584 ssize_t (*show)(struct kmem_cache *s, char *buf);
4585 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4588 #define SLAB_ATTR_RO(_name) \
4589 static struct slab_attribute _name##_attr = \
4590 __ATTR(_name, 0400, _name##_show, NULL)
4592 #define SLAB_ATTR(_name) \
4593 static struct slab_attribute _name##_attr = \
4594 __ATTR(_name, 0600, _name##_show, _name##_store)
4596 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4598 return sprintf(buf, "%d\n", s->size);
4600 SLAB_ATTR_RO(slab_size);
4602 static ssize_t align_show(struct kmem_cache *s, char *buf)
4604 return sprintf(buf, "%d\n", s->align);
4606 SLAB_ATTR_RO(align);
4608 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4610 return sprintf(buf, "%d\n", s->object_size);
4612 SLAB_ATTR_RO(object_size);
4614 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4616 return sprintf(buf, "%d\n", oo_objects(s->oo));
4618 SLAB_ATTR_RO(objs_per_slab);
4620 static ssize_t order_store(struct kmem_cache *s,
4621 const char *buf, size_t length)
4623 unsigned long order;
4626 err = strict_strtoul(buf, 10, &order);
4630 if (order > slub_max_order || order < slub_min_order)
4633 calculate_sizes(s, order);
4637 static ssize_t order_show(struct kmem_cache *s, char *buf)
4639 return sprintf(buf, "%d\n", oo_order(s->oo));
4643 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4645 return sprintf(buf, "%lu\n", s->min_partial);
4648 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4654 err = strict_strtoul(buf, 10, &min);
4658 set_min_partial(s, min);
4661 SLAB_ATTR(min_partial);
4663 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4665 return sprintf(buf, "%u\n", s->cpu_partial);
4668 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4671 unsigned long objects;
4674 err = strict_strtoul(buf, 10, &objects);
4677 if (objects && kmem_cache_debug(s))
4680 s->cpu_partial = objects;
4684 SLAB_ATTR(cpu_partial);
4686 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4690 return sprintf(buf, "%pS\n", s->ctor);
4694 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4696 return sprintf(buf, "%d\n", s->refcount - 1);
4698 SLAB_ATTR_RO(aliases);
4700 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4702 return show_slab_objects(s, buf, SO_PARTIAL);
4704 SLAB_ATTR_RO(partial);
4706 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4708 return show_slab_objects(s, buf, SO_CPU);
4710 SLAB_ATTR_RO(cpu_slabs);
4712 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4714 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4716 SLAB_ATTR_RO(objects);
4718 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4720 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4722 SLAB_ATTR_RO(objects_partial);
4724 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4731 for_each_online_cpu(cpu) {
4732 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4735 pages += page->pages;
4736 objects += page->pobjects;
4740 len = sprintf(buf, "%d(%d)", objects, pages);
4743 for_each_online_cpu(cpu) {
4744 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4746 if (page && len < PAGE_SIZE - 20)
4747 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4748 page->pobjects, page->pages);
4751 return len + sprintf(buf + len, "\n");
4753 SLAB_ATTR_RO(slabs_cpu_partial);
4755 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4757 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4760 static ssize_t reclaim_account_store(struct kmem_cache *s,
4761 const char *buf, size_t length)
4763 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4765 s->flags |= SLAB_RECLAIM_ACCOUNT;
4768 SLAB_ATTR(reclaim_account);
4770 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4772 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4774 SLAB_ATTR_RO(hwcache_align);
4776 #ifdef CONFIG_ZONE_DMA
4777 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4779 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4781 SLAB_ATTR_RO(cache_dma);
4784 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4786 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4788 SLAB_ATTR_RO(destroy_by_rcu);
4790 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4792 return sprintf(buf, "%d\n", s->reserved);
4794 SLAB_ATTR_RO(reserved);
4796 #ifdef CONFIG_SLUB_DEBUG
4797 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4799 return show_slab_objects(s, buf, SO_ALL);
4801 SLAB_ATTR_RO(slabs);
4803 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4805 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4807 SLAB_ATTR_RO(total_objects);
4809 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4811 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4814 static ssize_t sanity_checks_store(struct kmem_cache *s,
4815 const char *buf, size_t length)
4817 s->flags &= ~SLAB_DEBUG_FREE;
4818 if (buf[0] == '1') {
4819 s->flags &= ~__CMPXCHG_DOUBLE;
4820 s->flags |= SLAB_DEBUG_FREE;
4824 SLAB_ATTR(sanity_checks);
4826 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4828 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4831 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4834 s->flags &= ~SLAB_TRACE;
4835 if (buf[0] == '1') {
4836 s->flags &= ~__CMPXCHG_DOUBLE;
4837 s->flags |= SLAB_TRACE;
4843 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4845 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4848 static ssize_t red_zone_store(struct kmem_cache *s,
4849 const char *buf, size_t length)
4851 if (any_slab_objects(s))
4854 s->flags &= ~SLAB_RED_ZONE;
4855 if (buf[0] == '1') {
4856 s->flags &= ~__CMPXCHG_DOUBLE;
4857 s->flags |= SLAB_RED_ZONE;
4859 calculate_sizes(s, -1);
4862 SLAB_ATTR(red_zone);
4864 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4866 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4869 static ssize_t poison_store(struct kmem_cache *s,
4870 const char *buf, size_t length)
4872 if (any_slab_objects(s))
4875 s->flags &= ~SLAB_POISON;
4876 if (buf[0] == '1') {
4877 s->flags &= ~__CMPXCHG_DOUBLE;
4878 s->flags |= SLAB_POISON;
4880 calculate_sizes(s, -1);
4885 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4887 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4890 static ssize_t store_user_store(struct kmem_cache *s,
4891 const char *buf, size_t length)
4893 if (any_slab_objects(s))
4896 s->flags &= ~SLAB_STORE_USER;
4897 if (buf[0] == '1') {
4898 s->flags &= ~__CMPXCHG_DOUBLE;
4899 s->flags |= SLAB_STORE_USER;
4901 calculate_sizes(s, -1);
4904 SLAB_ATTR(store_user);
4906 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4911 static ssize_t validate_store(struct kmem_cache *s,
4912 const char *buf, size_t length)
4916 if (buf[0] == '1') {
4917 ret = validate_slab_cache(s);
4923 SLAB_ATTR(validate);
4925 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4927 if (!(s->flags & SLAB_STORE_USER))
4929 return list_locations(s, buf, TRACK_ALLOC);
4931 SLAB_ATTR_RO(alloc_calls);
4933 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4935 if (!(s->flags & SLAB_STORE_USER))
4937 return list_locations(s, buf, TRACK_FREE);
4939 SLAB_ATTR_RO(free_calls);
4940 #endif /* CONFIG_SLUB_DEBUG */
4942 #ifdef CONFIG_FAILSLAB
4943 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4945 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4948 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4951 s->flags &= ~SLAB_FAILSLAB;
4953 s->flags |= SLAB_FAILSLAB;
4956 SLAB_ATTR(failslab);
4959 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4964 static ssize_t shrink_store(struct kmem_cache *s,
4965 const char *buf, size_t length)
4967 if (buf[0] == '1') {
4968 int rc = kmem_cache_shrink(s);
4979 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4981 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4984 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4985 const char *buf, size_t length)
4987 unsigned long ratio;
4990 err = strict_strtoul(buf, 10, &ratio);
4995 s->remote_node_defrag_ratio = ratio * 10;
4999 SLAB_ATTR(remote_node_defrag_ratio);
5002 #ifdef CONFIG_SLUB_STATS
5003 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5005 unsigned long sum = 0;
5008 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5013 for_each_online_cpu(cpu) {
5014 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5020 len = sprintf(buf, "%lu", sum);
5023 for_each_online_cpu(cpu) {
5024 if (data[cpu] && len < PAGE_SIZE - 20)
5025 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5029 return len + sprintf(buf + len, "\n");
5032 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5036 for_each_online_cpu(cpu)
5037 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5040 #define STAT_ATTR(si, text) \
5041 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5043 return show_stat(s, buf, si); \
5045 static ssize_t text##_store(struct kmem_cache *s, \
5046 const char *buf, size_t length) \
5048 if (buf[0] != '0') \
5050 clear_stat(s, si); \
5055 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5056 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5057 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5058 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5059 STAT_ATTR(FREE_FROZEN, free_frozen);
5060 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5061 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5062 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5063 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5064 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5065 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5066 STAT_ATTR(FREE_SLAB, free_slab);
5067 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5068 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5069 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5070 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5071 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5072 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5073 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5074 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5075 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5076 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5077 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5078 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5079 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5080 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5083 static struct attribute *slab_attrs[] = {
5084 &slab_size_attr.attr,
5085 &object_size_attr.attr,
5086 &objs_per_slab_attr.attr,
5088 &min_partial_attr.attr,
5089 &cpu_partial_attr.attr,
5091 &objects_partial_attr.attr,
5093 &cpu_slabs_attr.attr,
5097 &hwcache_align_attr.attr,
5098 &reclaim_account_attr.attr,
5099 &destroy_by_rcu_attr.attr,
5101 &reserved_attr.attr,
5102 &slabs_cpu_partial_attr.attr,
5103 #ifdef CONFIG_SLUB_DEBUG
5104 &total_objects_attr.attr,
5106 &sanity_checks_attr.attr,
5108 &red_zone_attr.attr,
5110 &store_user_attr.attr,
5111 &validate_attr.attr,
5112 &alloc_calls_attr.attr,
5113 &free_calls_attr.attr,
5115 #ifdef CONFIG_ZONE_DMA
5116 &cache_dma_attr.attr,
5119 &remote_node_defrag_ratio_attr.attr,
5121 #ifdef CONFIG_SLUB_STATS
5122 &alloc_fastpath_attr.attr,
5123 &alloc_slowpath_attr.attr,
5124 &free_fastpath_attr.attr,
5125 &free_slowpath_attr.attr,
5126 &free_frozen_attr.attr,
5127 &free_add_partial_attr.attr,
5128 &free_remove_partial_attr.attr,
5129 &alloc_from_partial_attr.attr,
5130 &alloc_slab_attr.attr,
5131 &alloc_refill_attr.attr,
5132 &alloc_node_mismatch_attr.attr,
5133 &free_slab_attr.attr,
5134 &cpuslab_flush_attr.attr,
5135 &deactivate_full_attr.attr,
5136 &deactivate_empty_attr.attr,
5137 &deactivate_to_head_attr.attr,
5138 &deactivate_to_tail_attr.attr,
5139 &deactivate_remote_frees_attr.attr,
5140 &deactivate_bypass_attr.attr,
5141 &order_fallback_attr.attr,
5142 &cmpxchg_double_fail_attr.attr,
5143 &cmpxchg_double_cpu_fail_attr.attr,
5144 &cpu_partial_alloc_attr.attr,
5145 &cpu_partial_free_attr.attr,
5146 &cpu_partial_node_attr.attr,
5147 &cpu_partial_drain_attr.attr,
5149 #ifdef CONFIG_FAILSLAB
5150 &failslab_attr.attr,
5156 static struct attribute_group slab_attr_group = {
5157 .attrs = slab_attrs,
5160 static ssize_t slab_attr_show(struct kobject *kobj,
5161 struct attribute *attr,
5164 struct slab_attribute *attribute;
5165 struct kmem_cache *s;
5168 attribute = to_slab_attr(attr);
5171 if (!attribute->show)
5174 err = attribute->show(s, buf);
5179 static ssize_t slab_attr_store(struct kobject *kobj,
5180 struct attribute *attr,
5181 const char *buf, size_t len)
5183 struct slab_attribute *attribute;
5184 struct kmem_cache *s;
5187 attribute = to_slab_attr(attr);
5190 if (!attribute->store)
5193 err = attribute->store(s, buf, len);
5198 static const struct sysfs_ops slab_sysfs_ops = {
5199 .show = slab_attr_show,
5200 .store = slab_attr_store,
5203 static struct kobj_type slab_ktype = {
5204 .sysfs_ops = &slab_sysfs_ops,
5207 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5209 struct kobj_type *ktype = get_ktype(kobj);
5211 if (ktype == &slab_ktype)
5216 static const struct kset_uevent_ops slab_uevent_ops = {
5217 .filter = uevent_filter,
5220 static struct kset *slab_kset;
5222 #define ID_STR_LENGTH 64
5224 /* Create a unique string id for a slab cache:
5226 * Format :[flags-]size
5228 static char *create_unique_id(struct kmem_cache *s)
5230 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5237 * First flags affecting slabcache operations. We will only
5238 * get here for aliasable slabs so we do not need to support
5239 * too many flags. The flags here must cover all flags that
5240 * are matched during merging to guarantee that the id is
5243 if (s->flags & SLAB_CACHE_DMA)
5245 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5247 if (s->flags & SLAB_DEBUG_FREE)
5249 if (!(s->flags & SLAB_NOTRACK))
5253 p += sprintf(p, "%07d", s->size);
5254 BUG_ON(p > name + ID_STR_LENGTH - 1);
5258 static int sysfs_slab_add(struct kmem_cache *s)
5264 if (slab_state < FULL)
5265 /* Defer until later */
5268 unmergeable = slab_unmergeable(s);
5271 * Slabcache can never be merged so we can use the name proper.
5272 * This is typically the case for debug situations. In that
5273 * case we can catch duplicate names easily.
5275 sysfs_remove_link(&slab_kset->kobj, s->name);
5279 * Create a unique name for the slab as a target
5282 name = create_unique_id(s);
5285 s->kobj.kset = slab_kset;
5286 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5288 kobject_put(&s->kobj);
5292 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5294 kobject_del(&s->kobj);
5295 kobject_put(&s->kobj);
5298 kobject_uevent(&s->kobj, KOBJ_ADD);
5300 /* Setup first alias */
5301 sysfs_slab_alias(s, s->name);
5307 static void sysfs_slab_remove(struct kmem_cache *s)
5309 if (slab_state < FULL)
5311 * Sysfs has not been setup yet so no need to remove the
5316 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5317 kobject_del(&s->kobj);
5318 kobject_put(&s->kobj);
5322 * Need to buffer aliases during bootup until sysfs becomes
5323 * available lest we lose that information.
5325 struct saved_alias {
5326 struct kmem_cache *s;
5328 struct saved_alias *next;
5331 static struct saved_alias *alias_list;
5333 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5335 struct saved_alias *al;
5337 if (slab_state == FULL) {
5339 * If we have a leftover link then remove it.
5341 sysfs_remove_link(&slab_kset->kobj, name);
5342 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5345 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5351 al->next = alias_list;
5356 static int __init slab_sysfs_init(void)
5358 struct kmem_cache *s;
5361 mutex_lock(&slab_mutex);
5363 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5365 mutex_unlock(&slab_mutex);
5366 printk(KERN_ERR "Cannot register slab subsystem.\n");
5372 list_for_each_entry(s, &slab_caches, list) {
5373 err = sysfs_slab_add(s);
5375 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5376 " to sysfs\n", s->name);
5379 while (alias_list) {
5380 struct saved_alias *al = alias_list;
5382 alias_list = alias_list->next;
5383 err = sysfs_slab_alias(al->s, al->name);
5385 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5386 " %s to sysfs\n", al->name);
5390 mutex_unlock(&slab_mutex);
5395 __initcall(slab_sysfs_init);
5396 #endif /* CONFIG_SYSFS */
5399 * The /proc/slabinfo ABI
5401 #ifdef CONFIG_SLABINFO
5402 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5404 unsigned long nr_partials = 0;
5405 unsigned long nr_slabs = 0;
5406 unsigned long nr_objs = 0;
5407 unsigned long nr_free = 0;
5410 for_each_online_node(node) {
5411 struct kmem_cache_node *n = get_node(s, node);
5416 nr_partials += n->nr_partial;
5417 nr_slabs += atomic_long_read(&n->nr_slabs);
5418 nr_objs += atomic_long_read(&n->total_objects);
5419 nr_free += count_partial(n, count_free);
5422 sinfo->active_objs = nr_objs - nr_free;
5423 sinfo->num_objs = nr_objs;
5424 sinfo->active_slabs = nr_slabs;
5425 sinfo->num_slabs = nr_slabs;
5426 sinfo->objects_per_slab = oo_objects(s->oo);
5427 sinfo->cache_order = oo_order(s->oo);
5430 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5434 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5435 size_t count, loff_t *ppos)
5439 #endif /* CONFIG_SLABINFO */