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/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 * Issues still to be resolved:
129 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
131 * - Variable sizing of the per node arrays
134 /* Enable to test recovery from slab corruption on boot */
135 #undef SLUB_RESILIENCY_TEST
137 /* Enable to log cmpxchg failures */
138 #undef SLUB_DEBUG_CMPXCHG
141 * Mininum number of partial slabs. These will be left on the partial
142 * lists even if they are empty. kmem_cache_shrink may reclaim them.
144 #define MIN_PARTIAL 5
147 * Maximum number of desirable partial slabs.
148 * The existence of more partial slabs makes kmem_cache_shrink
149 * sort the partial list by the number of objects in the.
151 #define MAX_PARTIAL 10
153 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
154 SLAB_POISON | SLAB_STORE_USER)
157 * Debugging flags that require metadata to be stored in the slab. These get
158 * disabled when slub_debug=O is used and a cache's min order increases with
161 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
164 * Set of flags that will prevent slab merging
166 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
167 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
170 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
171 SLAB_CACHE_DMA | SLAB_NOTRACK)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void sysfs_slab_remove(struct kmem_cache *);
205 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
207 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
208 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
210 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
212 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
215 static inline void stat(const struct kmem_cache *s, enum stat_item si)
217 #ifdef CONFIG_SLUB_STATS
218 __this_cpu_inc(s->cpu_slab->stat[si]);
222 /********************************************************************
223 * Core slab cache functions
224 *******************************************************************/
226 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
228 return s->node[node];
231 /* Verify that a pointer has an address that is valid within a slab page */
232 static inline int check_valid_pointer(struct kmem_cache *s,
233 struct page *page, const void *object)
240 base = page_address(page);
241 if (object < base || object >= base + page->objects * s->size ||
242 (object - base) % s->size) {
249 static inline void *get_freepointer(struct kmem_cache *s, void *object)
251 return *(void **)(object + s->offset);
254 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
256 prefetch(object + s->offset);
259 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
263 #ifdef CONFIG_DEBUG_PAGEALLOC
264 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
266 p = get_freepointer(s, object);
271 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
273 *(void **)(object + s->offset) = fp;
276 /* Loop over all objects in a slab */
277 #define for_each_object(__p, __s, __addr, __objects) \
278 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline size_t slab_ksize(const struct kmem_cache *s)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order, unsigned long size, int reserved)
313 return ((PAGE_SIZE << order) - reserved) / size;
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
326 static inline int oo_order(struct kmem_cache_order_objects x)
328 return x.x >> OO_SHIFT;
331 static inline int oo_objects(struct kmem_cache_order_objects x)
333 return x.x & OO_MASK;
337 * Per slab locking using the pagelock
339 static __always_inline void slab_lock(struct page *page)
341 bit_spin_lock(PG_locked, &page->flags);
344 static __always_inline void slab_unlock(struct page *page)
346 __bit_spin_unlock(PG_locked, &page->flags);
349 /* Interrupts must be disabled (for the fallback code to work right) */
350 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
351 void *freelist_old, unsigned long counters_old,
352 void *freelist_new, unsigned long counters_new,
355 VM_BUG_ON(!irqs_disabled());
356 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
357 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
358 if (s->flags & __CMPXCHG_DOUBLE) {
359 if (cmpxchg_double(&page->freelist, &page->counters,
360 freelist_old, counters_old,
361 freelist_new, counters_new))
367 if (page->freelist == freelist_old && page->counters == counters_old) {
368 page->freelist = freelist_new;
369 page->counters = counters_new;
377 stat(s, CMPXCHG_DOUBLE_FAIL);
379 #ifdef SLUB_DEBUG_CMPXCHG
380 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
386 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
387 void *freelist_old, unsigned long counters_old,
388 void *freelist_new, unsigned long counters_new,
391 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
392 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
393 if (s->flags & __CMPXCHG_DOUBLE) {
394 if (cmpxchg_double(&page->freelist, &page->counters,
395 freelist_old, counters_old,
396 freelist_new, counters_new))
403 local_irq_save(flags);
405 if (page->freelist == freelist_old && page->counters == counters_old) {
406 page->freelist = freelist_new;
407 page->counters = counters_new;
409 local_irq_restore(flags);
413 local_irq_restore(flags);
417 stat(s, CMPXCHG_DOUBLE_FAIL);
419 #ifdef SLUB_DEBUG_CMPXCHG
420 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
426 #ifdef CONFIG_SLUB_DEBUG
428 * Determine a map of object in use on a page.
430 * Node listlock must be held to guarantee that the page does
431 * not vanish from under us.
433 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
436 void *addr = page_address(page);
438 for (p = page->freelist; p; p = get_freepointer(s, p))
439 set_bit(slab_index(p, s, addr), map);
445 #ifdef CONFIG_SLUB_DEBUG_ON
446 static int slub_debug = DEBUG_DEFAULT_FLAGS;
448 static int slub_debug;
451 static char *slub_debug_slabs;
452 static int disable_higher_order_debug;
457 static void print_section(char *text, u8 *addr, unsigned int length)
459 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
463 static struct track *get_track(struct kmem_cache *s, void *object,
464 enum track_item alloc)
469 p = object + s->offset + sizeof(void *);
471 p = object + s->inuse;
476 static void set_track(struct kmem_cache *s, void *object,
477 enum track_item alloc, unsigned long addr)
479 struct track *p = get_track(s, object, alloc);
482 #ifdef CONFIG_STACKTRACE
483 struct stack_trace trace;
486 trace.nr_entries = 0;
487 trace.max_entries = TRACK_ADDRS_COUNT;
488 trace.entries = p->addrs;
490 save_stack_trace(&trace);
492 /* See rant in lockdep.c */
493 if (trace.nr_entries != 0 &&
494 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
497 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
501 p->cpu = smp_processor_id();
502 p->pid = current->pid;
505 memset(p, 0, sizeof(struct track));
508 static void init_tracking(struct kmem_cache *s, void *object)
510 if (!(s->flags & SLAB_STORE_USER))
513 set_track(s, object, TRACK_FREE, 0UL);
514 set_track(s, object, TRACK_ALLOC, 0UL);
517 static void print_track(const char *s, struct track *t)
522 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
523 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
524 #ifdef CONFIG_STACKTRACE
527 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
529 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
536 static void print_tracking(struct kmem_cache *s, void *object)
538 if (!(s->flags & SLAB_STORE_USER))
541 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
542 print_track("Freed", get_track(s, object, TRACK_FREE));
545 static void print_page_info(struct page *page)
547 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
548 page, page->objects, page->inuse, page->freelist, page->flags);
552 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
558 vsnprintf(buf, sizeof(buf), fmt, args);
560 printk(KERN_ERR "========================================"
561 "=====================================\n");
562 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
563 printk(KERN_ERR "----------------------------------------"
564 "-------------------------------------\n\n");
566 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
569 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
575 vsnprintf(buf, sizeof(buf), fmt, args);
577 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
580 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
582 unsigned int off; /* Offset of last byte */
583 u8 *addr = page_address(page);
585 print_tracking(s, p);
587 print_page_info(page);
589 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
590 p, p - addr, get_freepointer(s, p));
593 print_section("Bytes b4 ", p - 16, 16);
595 print_section("Object ", p, min_t(unsigned long, s->object_size,
597 if (s->flags & SLAB_RED_ZONE)
598 print_section("Redzone ", p + s->object_size,
599 s->inuse - s->object_size);
602 off = s->offset + sizeof(void *);
606 if (s->flags & SLAB_STORE_USER)
607 off += 2 * sizeof(struct track);
610 /* Beginning of the filler is the free pointer */
611 print_section("Padding ", p + off, s->size - off);
616 static void object_err(struct kmem_cache *s, struct page *page,
617 u8 *object, char *reason)
619 slab_bug(s, "%s", reason);
620 print_trailer(s, page, object);
623 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
629 vsnprintf(buf, sizeof(buf), fmt, args);
631 slab_bug(s, "%s", buf);
632 print_page_info(page);
636 static void init_object(struct kmem_cache *s, void *object, u8 val)
640 if (s->flags & __OBJECT_POISON) {
641 memset(p, POISON_FREE, s->object_size - 1);
642 p[s->object_size - 1] = POISON_END;
645 if (s->flags & SLAB_RED_ZONE)
646 memset(p + s->object_size, val, s->inuse - s->object_size);
649 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
650 void *from, void *to)
652 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
653 memset(from, data, to - from);
656 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
657 u8 *object, char *what,
658 u8 *start, unsigned int value, unsigned int bytes)
663 fault = memchr_inv(start, value, bytes);
668 while (end > fault && end[-1] == value)
671 slab_bug(s, "%s overwritten", what);
672 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
673 fault, end - 1, fault[0], value);
674 print_trailer(s, page, object);
676 restore_bytes(s, what, value, fault, end);
684 * Bytes of the object to be managed.
685 * If the freepointer may overlay the object then the free
686 * pointer is the first word of the object.
688 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
691 * object + s->object_size
692 * Padding to reach word boundary. This is also used for Redzoning.
693 * Padding is extended by another word if Redzoning is enabled and
694 * object_size == inuse.
696 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
697 * 0xcc (RED_ACTIVE) for objects in use.
700 * Meta data starts here.
702 * A. Free pointer (if we cannot overwrite object on free)
703 * B. Tracking data for SLAB_STORE_USER
704 * C. Padding to reach required alignment boundary or at mininum
705 * one word if debugging is on to be able to detect writes
706 * before the word boundary.
708 * Padding is done using 0x5a (POISON_INUSE)
711 * Nothing is used beyond s->size.
713 * If slabcaches are merged then the object_size and inuse boundaries are mostly
714 * ignored. And therefore no slab options that rely on these boundaries
715 * may be used with merged slabcaches.
718 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
720 unsigned long off = s->inuse; /* The end of info */
723 /* Freepointer is placed after the object. */
724 off += sizeof(void *);
726 if (s->flags & SLAB_STORE_USER)
727 /* We also have user information there */
728 off += 2 * sizeof(struct track);
733 return check_bytes_and_report(s, page, p, "Object padding",
734 p + off, POISON_INUSE, s->size - off);
737 /* Check the pad bytes at the end of a slab page */
738 static int slab_pad_check(struct kmem_cache *s, struct page *page)
746 if (!(s->flags & SLAB_POISON))
749 start = page_address(page);
750 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
751 end = start + length;
752 remainder = length % s->size;
756 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
759 while (end > fault && end[-1] == POISON_INUSE)
762 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
763 print_section("Padding ", end - remainder, remainder);
765 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
769 static int check_object(struct kmem_cache *s, struct page *page,
770 void *object, u8 val)
773 u8 *endobject = object + s->object_size;
775 if (s->flags & SLAB_RED_ZONE) {
776 if (!check_bytes_and_report(s, page, object, "Redzone",
777 endobject, val, s->inuse - s->object_size))
780 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
781 check_bytes_and_report(s, page, p, "Alignment padding",
782 endobject, POISON_INUSE, s->inuse - s->object_size);
786 if (s->flags & SLAB_POISON) {
787 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
788 (!check_bytes_and_report(s, page, p, "Poison", p,
789 POISON_FREE, s->object_size - 1) ||
790 !check_bytes_and_report(s, page, p, "Poison",
791 p + s->object_size - 1, POISON_END, 1)))
794 * check_pad_bytes cleans up on its own.
796 check_pad_bytes(s, page, p);
799 if (!s->offset && val == SLUB_RED_ACTIVE)
801 * Object and freepointer overlap. Cannot check
802 * freepointer while object is allocated.
806 /* Check free pointer validity */
807 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
808 object_err(s, page, p, "Freepointer corrupt");
810 * No choice but to zap it and thus lose the remainder
811 * of the free objects in this slab. May cause
812 * another error because the object count is now wrong.
814 set_freepointer(s, p, NULL);
820 static int check_slab(struct kmem_cache *s, struct page *page)
824 VM_BUG_ON(!irqs_disabled());
826 if (!PageSlab(page)) {
827 slab_err(s, page, "Not a valid slab page");
831 maxobj = order_objects(compound_order(page), s->size, s->reserved);
832 if (page->objects > maxobj) {
833 slab_err(s, page, "objects %u > max %u",
834 s->name, page->objects, maxobj);
837 if (page->inuse > page->objects) {
838 slab_err(s, page, "inuse %u > max %u",
839 s->name, page->inuse, page->objects);
842 /* Slab_pad_check fixes things up after itself */
843 slab_pad_check(s, page);
848 * Determine if a certain object on a page is on the freelist. Must hold the
849 * slab lock to guarantee that the chains are in a consistent state.
851 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
856 unsigned long max_objects;
859 while (fp && nr <= page->objects) {
862 if (!check_valid_pointer(s, page, fp)) {
864 object_err(s, page, object,
865 "Freechain corrupt");
866 set_freepointer(s, object, NULL);
869 slab_err(s, page, "Freepointer corrupt");
870 page->freelist = NULL;
871 page->inuse = page->objects;
872 slab_fix(s, "Freelist cleared");
878 fp = get_freepointer(s, object);
882 max_objects = order_objects(compound_order(page), s->size, s->reserved);
883 if (max_objects > MAX_OBJS_PER_PAGE)
884 max_objects = MAX_OBJS_PER_PAGE;
886 if (page->objects != max_objects) {
887 slab_err(s, page, "Wrong number of objects. Found %d but "
888 "should be %d", page->objects, max_objects);
889 page->objects = max_objects;
890 slab_fix(s, "Number of objects adjusted.");
892 if (page->inuse != page->objects - nr) {
893 slab_err(s, page, "Wrong object count. Counter is %d but "
894 "counted were %d", page->inuse, page->objects - nr);
895 page->inuse = page->objects - nr;
896 slab_fix(s, "Object count adjusted.");
898 return search == NULL;
901 static void trace(struct kmem_cache *s, struct page *page, void *object,
904 if (s->flags & SLAB_TRACE) {
905 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
907 alloc ? "alloc" : "free",
912 print_section("Object ", (void *)object, s->object_size);
919 * Hooks for other subsystems that check memory allocations. In a typical
920 * production configuration these hooks all should produce no code at all.
922 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
924 flags &= gfp_allowed_mask;
925 lockdep_trace_alloc(flags);
926 might_sleep_if(flags & __GFP_WAIT);
928 return should_failslab(s->object_size, flags, s->flags);
931 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
933 flags &= gfp_allowed_mask;
934 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
935 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
938 static inline void slab_free_hook(struct kmem_cache *s, void *x)
940 kmemleak_free_recursive(x, s->flags);
943 * Trouble is that we may no longer disable interupts in the fast path
944 * So in order to make the debug calls that expect irqs to be
945 * disabled we need to disable interrupts temporarily.
947 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
951 local_irq_save(flags);
952 kmemcheck_slab_free(s, x, s->object_size);
953 debug_check_no_locks_freed(x, s->object_size);
954 local_irq_restore(flags);
957 if (!(s->flags & SLAB_DEBUG_OBJECTS))
958 debug_check_no_obj_freed(x, s->object_size);
962 * Tracking of fully allocated slabs for debugging purposes.
964 * list_lock must be held.
966 static void add_full(struct kmem_cache *s,
967 struct kmem_cache_node *n, struct page *page)
969 if (!(s->flags & SLAB_STORE_USER))
972 list_add(&page->lru, &n->full);
976 * list_lock must be held.
978 static void remove_full(struct kmem_cache *s, struct page *page)
980 if (!(s->flags & SLAB_STORE_USER))
983 list_del(&page->lru);
986 /* Tracking of the number of slabs for debugging purposes */
987 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
989 struct kmem_cache_node *n = get_node(s, node);
991 return atomic_long_read(&n->nr_slabs);
994 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
996 return atomic_long_read(&n->nr_slabs);
999 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1001 struct kmem_cache_node *n = get_node(s, node);
1004 * May be called early in order to allocate a slab for the
1005 * kmem_cache_node structure. Solve the chicken-egg
1006 * dilemma by deferring the increment of the count during
1007 * bootstrap (see early_kmem_cache_node_alloc).
1010 atomic_long_inc(&n->nr_slabs);
1011 atomic_long_add(objects, &n->total_objects);
1014 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1016 struct kmem_cache_node *n = get_node(s, node);
1018 atomic_long_dec(&n->nr_slabs);
1019 atomic_long_sub(objects, &n->total_objects);
1022 /* Object debug checks for alloc/free paths */
1023 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1026 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1029 init_object(s, object, SLUB_RED_INACTIVE);
1030 init_tracking(s, object);
1033 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1034 void *object, unsigned long addr)
1036 if (!check_slab(s, page))
1039 if (!check_valid_pointer(s, page, object)) {
1040 object_err(s, page, object, "Freelist Pointer check fails");
1044 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1047 /* Success perform special debug activities for allocs */
1048 if (s->flags & SLAB_STORE_USER)
1049 set_track(s, object, TRACK_ALLOC, addr);
1050 trace(s, page, object, 1);
1051 init_object(s, object, SLUB_RED_ACTIVE);
1055 if (PageSlab(page)) {
1057 * If this is a slab page then lets do the best we can
1058 * to avoid issues in the future. Marking all objects
1059 * as used avoids touching the remaining objects.
1061 slab_fix(s, "Marking all objects used");
1062 page->inuse = page->objects;
1063 page->freelist = NULL;
1068 static noinline struct kmem_cache_node *free_debug_processing(
1069 struct kmem_cache *s, struct page *page, void *object,
1070 unsigned long addr, unsigned long *flags)
1072 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1074 spin_lock_irqsave(&n->list_lock, *flags);
1077 if (!check_slab(s, page))
1080 if (!check_valid_pointer(s, page, object)) {
1081 slab_err(s, page, "Invalid object pointer 0x%p", object);
1085 if (on_freelist(s, page, object)) {
1086 object_err(s, page, object, "Object already free");
1090 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1093 if (unlikely(s != page->slab_cache)) {
1094 if (!PageSlab(page)) {
1095 slab_err(s, page, "Attempt to free object(0x%p) "
1096 "outside of slab", object);
1097 } else if (!page->slab_cache) {
1099 "SLUB <none>: no slab for object 0x%p.\n",
1103 object_err(s, page, object,
1104 "page slab pointer corrupt.");
1108 if (s->flags & SLAB_STORE_USER)
1109 set_track(s, object, TRACK_FREE, addr);
1110 trace(s, page, object, 0);
1111 init_object(s, object, SLUB_RED_INACTIVE);
1115 * Keep node_lock to preserve integrity
1116 * until the object is actually freed
1122 spin_unlock_irqrestore(&n->list_lock, *flags);
1123 slab_fix(s, "Object at 0x%p not freed", object);
1127 static int __init setup_slub_debug(char *str)
1129 slub_debug = DEBUG_DEFAULT_FLAGS;
1130 if (*str++ != '=' || !*str)
1132 * No options specified. Switch on full debugging.
1138 * No options but restriction on slabs. This means full
1139 * debugging for slabs matching a pattern.
1143 if (tolower(*str) == 'o') {
1145 * Avoid enabling debugging on caches if its minimum order
1146 * would increase as a result.
1148 disable_higher_order_debug = 1;
1155 * Switch off all debugging measures.
1160 * Determine which debug features should be switched on
1162 for (; *str && *str != ','; str++) {
1163 switch (tolower(*str)) {
1165 slub_debug |= SLAB_DEBUG_FREE;
1168 slub_debug |= SLAB_RED_ZONE;
1171 slub_debug |= SLAB_POISON;
1174 slub_debug |= SLAB_STORE_USER;
1177 slub_debug |= SLAB_TRACE;
1180 slub_debug |= SLAB_FAILSLAB;
1183 printk(KERN_ERR "slub_debug option '%c' "
1184 "unknown. skipped\n", *str);
1190 slub_debug_slabs = str + 1;
1195 __setup("slub_debug", setup_slub_debug);
1197 static unsigned long kmem_cache_flags(unsigned long object_size,
1198 unsigned long flags, const char *name,
1199 void (*ctor)(void *))
1202 * Enable debugging if selected on the kernel commandline.
1204 if (slub_debug && (!slub_debug_slabs || (name &&
1205 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1206 flags |= slub_debug;
1211 static inline void setup_object_debug(struct kmem_cache *s,
1212 struct page *page, void *object) {}
1214 static inline int alloc_debug_processing(struct kmem_cache *s,
1215 struct page *page, void *object, unsigned long addr) { return 0; }
1217 static inline struct kmem_cache_node *free_debug_processing(
1218 struct kmem_cache *s, struct page *page, void *object,
1219 unsigned long addr, unsigned long *flags) { return NULL; }
1221 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1223 static inline int check_object(struct kmem_cache *s, struct page *page,
1224 void *object, u8 val) { return 1; }
1225 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1226 struct page *page) {}
1227 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1228 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1229 unsigned long flags, const char *name,
1230 void (*ctor)(void *))
1234 #define slub_debug 0
1236 #define disable_higher_order_debug 0
1238 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1240 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1242 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1244 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1247 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1250 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1253 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1255 #endif /* CONFIG_SLUB_DEBUG */
1258 * Slab allocation and freeing
1260 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1261 struct kmem_cache_order_objects oo)
1263 int order = oo_order(oo);
1265 flags |= __GFP_NOTRACK;
1267 if (node == NUMA_NO_NODE)
1268 #ifndef CONFIG_SPRD_PAGERECORDER
1269 return alloc_pages(flags, order);
1271 return alloc_pages_nopagedebug(flags, order);
1274 return alloc_pages_exact_node(node, flags, order);
1277 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1280 struct kmem_cache_order_objects oo = s->oo;
1283 flags &= gfp_allowed_mask;
1285 if (flags & __GFP_WAIT)
1288 flags |= s->allocflags;
1291 * Let the initial higher-order allocation fail under memory pressure
1292 * so we fall-back to the minimum order allocation.
1294 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC
1295 | __GFP_NO_KSWAPD) & ~(__GFP_NOFAIL|__GFP_WAIT);
1297 page = alloc_slab_page(alloc_gfp, node, oo);
1298 if (unlikely(!page)) {
1301 * Allocation may have failed due to fragmentation.
1302 * Try a lower order alloc if possible
1304 page = alloc_slab_page(flags, node, oo);
1307 stat(s, ORDER_FALLBACK);
1310 if (kmemcheck_enabled && page
1311 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1312 int pages = 1 << oo_order(oo);
1314 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1317 * Objects from caches that have a constructor don't get
1318 * cleared when they're allocated, so we need to do it here.
1321 kmemcheck_mark_uninitialized_pages(page, pages);
1323 kmemcheck_mark_unallocated_pages(page, pages);
1326 if (flags & __GFP_WAIT)
1327 local_irq_disable();
1331 page->objects = oo_objects(oo);
1332 mod_zone_page_state(page_zone(page),
1333 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1334 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1340 static void setup_object(struct kmem_cache *s, struct page *page,
1343 setup_object_debug(s, page, object);
1344 if (unlikely(s->ctor))
1348 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1356 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1358 page = allocate_slab(s,
1359 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1363 order = compound_order(page);
1364 inc_slabs_node(s, page_to_nid(page), page->objects);
1365 memcg_bind_pages(s, order);
1366 page->slab_cache = s;
1367 __SetPageSlab(page);
1368 if (page->pfmemalloc)
1369 SetPageSlabPfmemalloc(page);
1371 start = page_address(page);
1373 if (unlikely(s->flags & SLAB_POISON))
1374 memset(start, POISON_INUSE, PAGE_SIZE << order);
1377 for_each_object(p, s, start, page->objects) {
1378 setup_object(s, page, last);
1379 set_freepointer(s, last, p);
1382 setup_object(s, page, last);
1383 set_freepointer(s, last, NULL);
1385 page->freelist = start;
1386 page->inuse = page->objects;
1392 static void __free_slab(struct kmem_cache *s, struct page *page)
1394 int order = compound_order(page);
1395 int pages = 1 << order;
1397 if (kmem_cache_debug(s)) {
1400 slab_pad_check(s, page);
1401 for_each_object(p, s, page_address(page),
1403 check_object(s, page, p, SLUB_RED_INACTIVE);
1406 kmemcheck_free_shadow(page, compound_order(page));
1408 mod_zone_page_state(page_zone(page),
1409 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1410 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1413 __ClearPageSlabPfmemalloc(page);
1414 __ClearPageSlab(page);
1416 memcg_release_pages(s, order);
1417 page_mapcount_reset(page);
1418 if (current->reclaim_state)
1419 current->reclaim_state->reclaimed_slab += pages;
1421 #ifndef CONFIG_SPRD_PAGERECORDER
1422 __free_memcg_kmem_pages(page, order);
1424 __free_memcg_kmem_pages_nopagedebug(page, order);
1428 #define need_reserve_slab_rcu \
1429 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1431 static void rcu_free_slab(struct rcu_head *h)
1435 if (need_reserve_slab_rcu)
1436 page = virt_to_head_page(h);
1438 page = container_of((struct list_head *)h, struct page, lru);
1440 __free_slab(page->slab_cache, page);
1443 static void free_slab(struct kmem_cache *s, struct page *page)
1445 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1446 struct rcu_head *head;
1448 if (need_reserve_slab_rcu) {
1449 int order = compound_order(page);
1450 int offset = (PAGE_SIZE << order) - s->reserved;
1452 VM_BUG_ON(s->reserved != sizeof(*head));
1453 head = page_address(page) + offset;
1456 * RCU free overloads the RCU head over the LRU
1458 head = (void *)&page->lru;
1461 call_rcu(head, rcu_free_slab);
1463 __free_slab(s, page);
1466 static void discard_slab(struct kmem_cache *s, struct page *page)
1468 dec_slabs_node(s, page_to_nid(page), page->objects);
1473 * Management of partially allocated slabs.
1475 * list_lock must be held.
1477 static inline void add_partial(struct kmem_cache_node *n,
1478 struct page *page, int tail)
1481 if (tail == DEACTIVATE_TO_TAIL)
1482 list_add_tail(&page->lru, &n->partial);
1484 list_add(&page->lru, &n->partial);
1488 * list_lock must be held.
1490 static inline void remove_partial(struct kmem_cache_node *n,
1493 list_del(&page->lru);
1498 * Remove slab from the partial list, freeze it and
1499 * return the pointer to the freelist.
1501 * Returns a list of objects or NULL if it fails.
1503 * Must hold list_lock since we modify the partial list.
1505 static inline void *acquire_slab(struct kmem_cache *s,
1506 struct kmem_cache_node *n, struct page *page,
1507 int mode, int *objects)
1510 unsigned long counters;
1514 * Zap the freelist and set the frozen bit.
1515 * The old freelist is the list of objects for the
1516 * per cpu allocation list.
1518 freelist = page->freelist;
1519 counters = page->counters;
1520 new.counters = counters;
1521 *objects = new.objects - new.inuse;
1523 new.inuse = page->objects;
1524 new.freelist = NULL;
1526 new.freelist = freelist;
1529 VM_BUG_ON(new.frozen);
1532 if (!__cmpxchg_double_slab(s, page,
1534 new.freelist, new.counters,
1538 remove_partial(n, page);
1543 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1544 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1547 * Try to allocate a partial slab from a specific node.
1549 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1550 struct kmem_cache_cpu *c, gfp_t flags)
1552 struct page *page, *page2;
1553 void *object = NULL;
1558 * Racy check. If we mistakenly see no partial slabs then we
1559 * just allocate an empty slab. If we mistakenly try to get a
1560 * partial slab and there is none available then get_partials()
1563 if (!n || !n->nr_partial)
1566 spin_lock(&n->list_lock);
1567 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1570 if (!pfmemalloc_match(page, flags))
1573 t = acquire_slab(s, n, page, object == NULL, &objects);
1577 available += objects;
1580 stat(s, ALLOC_FROM_PARTIAL);
1583 put_cpu_partial(s, page, 0);
1584 stat(s, CPU_PARTIAL_NODE);
1586 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1590 spin_unlock(&n->list_lock);
1595 * Get a page from somewhere. Search in increasing NUMA distances.
1597 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1598 struct kmem_cache_cpu *c)
1601 struct zonelist *zonelist;
1604 enum zone_type high_zoneidx = gfp_zone(flags);
1606 unsigned int cpuset_mems_cookie;
1609 * The defrag ratio allows a configuration of the tradeoffs between
1610 * inter node defragmentation and node local allocations. A lower
1611 * defrag_ratio increases the tendency to do local allocations
1612 * instead of attempting to obtain partial slabs from other nodes.
1614 * If the defrag_ratio is set to 0 then kmalloc() always
1615 * returns node local objects. If the ratio is higher then kmalloc()
1616 * may return off node objects because partial slabs are obtained
1617 * from other nodes and filled up.
1619 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1620 * defrag_ratio = 1000) then every (well almost) allocation will
1621 * first attempt to defrag slab caches on other nodes. This means
1622 * scanning over all nodes to look for partial slabs which may be
1623 * expensive if we do it every time we are trying to find a slab
1624 * with available objects.
1626 if (!s->remote_node_defrag_ratio ||
1627 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1631 cpuset_mems_cookie = get_mems_allowed();
1632 zonelist = node_zonelist(slab_node(), flags);
1633 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1634 struct kmem_cache_node *n;
1636 n = get_node(s, zone_to_nid(zone));
1638 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1639 n->nr_partial > s->min_partial) {
1640 object = get_partial_node(s, n, c, flags);
1643 * Return the object even if
1644 * put_mems_allowed indicated that
1645 * the cpuset mems_allowed was
1646 * updated in parallel. It's a
1647 * harmless race between the alloc
1648 * and the cpuset update.
1650 put_mems_allowed(cpuset_mems_cookie);
1655 } while (!put_mems_allowed(cpuset_mems_cookie));
1661 * Get a partial page, lock it and return it.
1663 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1664 struct kmem_cache_cpu *c)
1667 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1669 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1670 if (object || node != NUMA_NO_NODE)
1673 return get_any_partial(s, flags, c);
1676 #ifdef CONFIG_PREEMPT
1678 * Calculate the next globally unique transaction for disambiguiation
1679 * during cmpxchg. The transactions start with the cpu number and are then
1680 * incremented by CONFIG_NR_CPUS.
1682 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1685 * No preemption supported therefore also no need to check for
1691 static inline unsigned long next_tid(unsigned long tid)
1693 return tid + TID_STEP;
1696 static inline unsigned int tid_to_cpu(unsigned long tid)
1698 return tid % TID_STEP;
1701 static inline unsigned long tid_to_event(unsigned long tid)
1703 return tid / TID_STEP;
1706 static inline unsigned int init_tid(int cpu)
1711 static inline void note_cmpxchg_failure(const char *n,
1712 const struct kmem_cache *s, unsigned long tid)
1714 #ifdef SLUB_DEBUG_CMPXCHG
1715 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1717 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1719 #ifdef CONFIG_PREEMPT
1720 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1721 printk("due to cpu change %d -> %d\n",
1722 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1725 if (tid_to_event(tid) != tid_to_event(actual_tid))
1726 printk("due to cpu running other code. Event %ld->%ld\n",
1727 tid_to_event(tid), tid_to_event(actual_tid));
1729 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1730 actual_tid, tid, next_tid(tid));
1732 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1735 static void init_kmem_cache_cpus(struct kmem_cache *s)
1739 for_each_possible_cpu(cpu)
1740 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1744 * Remove the cpu slab
1746 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1748 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1749 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1751 enum slab_modes l = M_NONE, m = M_NONE;
1753 int tail = DEACTIVATE_TO_HEAD;
1757 if (page->freelist) {
1758 stat(s, DEACTIVATE_REMOTE_FREES);
1759 tail = DEACTIVATE_TO_TAIL;
1763 * Stage one: Free all available per cpu objects back
1764 * to the page freelist while it is still frozen. Leave the
1767 * There is no need to take the list->lock because the page
1770 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1772 unsigned long counters;
1775 prior = page->freelist;
1776 counters = page->counters;
1777 set_freepointer(s, freelist, prior);
1778 new.counters = counters;
1780 VM_BUG_ON(!new.frozen);
1782 } while (!__cmpxchg_double_slab(s, page,
1784 freelist, new.counters,
1785 "drain percpu freelist"));
1787 freelist = nextfree;
1791 * Stage two: Ensure that the page is unfrozen while the
1792 * list presence reflects the actual number of objects
1795 * We setup the list membership and then perform a cmpxchg
1796 * with the count. If there is a mismatch then the page
1797 * is not unfrozen but the page is on the wrong list.
1799 * Then we restart the process which may have to remove
1800 * the page from the list that we just put it on again
1801 * because the number of objects in the slab may have
1806 old.freelist = page->freelist;
1807 old.counters = page->counters;
1808 VM_BUG_ON(!old.frozen);
1810 /* Determine target state of the slab */
1811 new.counters = old.counters;
1814 set_freepointer(s, freelist, old.freelist);
1815 new.freelist = freelist;
1817 new.freelist = old.freelist;
1821 if (!new.inuse && n->nr_partial > s->min_partial)
1823 else if (new.freelist) {
1828 * Taking the spinlock removes the possiblity
1829 * that acquire_slab() will see a slab page that
1832 spin_lock(&n->list_lock);
1836 if (kmem_cache_debug(s) && !lock) {
1839 * This also ensures that the scanning of full
1840 * slabs from diagnostic functions will not see
1843 spin_lock(&n->list_lock);
1851 remove_partial(n, page);
1853 else if (l == M_FULL)
1855 remove_full(s, page);
1857 if (m == M_PARTIAL) {
1859 add_partial(n, page, tail);
1862 } else if (m == M_FULL) {
1864 stat(s, DEACTIVATE_FULL);
1865 add_full(s, n, page);
1871 if (!__cmpxchg_double_slab(s, page,
1872 old.freelist, old.counters,
1873 new.freelist, new.counters,
1878 spin_unlock(&n->list_lock);
1881 stat(s, DEACTIVATE_EMPTY);
1882 discard_slab(s, page);
1888 * Unfreeze all the cpu partial slabs.
1890 * This function must be called with interrupts disabled
1891 * for the cpu using c (or some other guarantee must be there
1892 * to guarantee no concurrent accesses).
1894 static void unfreeze_partials(struct kmem_cache *s,
1895 struct kmem_cache_cpu *c)
1897 struct kmem_cache_node *n = NULL, *n2 = NULL;
1898 struct page *page, *discard_page = NULL;
1900 while ((page = c->partial)) {
1904 c->partial = page->next;
1906 n2 = get_node(s, page_to_nid(page));
1909 spin_unlock(&n->list_lock);
1912 spin_lock(&n->list_lock);
1917 old.freelist = page->freelist;
1918 old.counters = page->counters;
1919 VM_BUG_ON(!old.frozen);
1921 new.counters = old.counters;
1922 new.freelist = old.freelist;
1926 } while (!__cmpxchg_double_slab(s, page,
1927 old.freelist, old.counters,
1928 new.freelist, new.counters,
1929 "unfreezing slab"));
1931 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1932 page->next = discard_page;
1933 discard_page = page;
1935 add_partial(n, page, DEACTIVATE_TO_TAIL);
1936 stat(s, FREE_ADD_PARTIAL);
1941 spin_unlock(&n->list_lock);
1943 while (discard_page) {
1944 page = discard_page;
1945 discard_page = discard_page->next;
1947 stat(s, DEACTIVATE_EMPTY);
1948 discard_slab(s, page);
1954 * Put a page that was just frozen (in __slab_free) into a partial page
1955 * slot if available. This is done without interrupts disabled and without
1956 * preemption disabled. The cmpxchg is racy and may put the partial page
1957 * onto a random cpus partial slot.
1959 * If we did not find a slot then simply move all the partials to the
1960 * per node partial list.
1962 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1964 struct page *oldpage;
1971 oldpage = this_cpu_read(s->cpu_slab->partial);
1974 pobjects = oldpage->pobjects;
1975 pages = oldpage->pages;
1976 if (drain && pobjects > s->cpu_partial) {
1977 unsigned long flags;
1979 * partial array is full. Move the existing
1980 * set to the per node partial list.
1982 local_irq_save(flags);
1983 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1984 local_irq_restore(flags);
1988 stat(s, CPU_PARTIAL_DRAIN);
1993 pobjects += page->objects - page->inuse;
1995 page->pages = pages;
1996 page->pobjects = pobjects;
1997 page->next = oldpage;
1999 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2002 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2004 stat(s, CPUSLAB_FLUSH);
2005 deactivate_slab(s, c->page, c->freelist);
2007 c->tid = next_tid(c->tid);
2015 * Called from IPI handler with interrupts disabled.
2017 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2019 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2025 unfreeze_partials(s, c);
2029 static void flush_cpu_slab(void *d)
2031 struct kmem_cache *s = d;
2033 __flush_cpu_slab(s, smp_processor_id());
2036 static bool has_cpu_slab(int cpu, void *info)
2038 struct kmem_cache *s = info;
2039 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2041 return c->page || c->partial;
2044 static void flush_all(struct kmem_cache *s)
2046 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2050 * Check if the objects in a per cpu structure fit numa
2051 * locality expectations.
2053 static inline int node_match(struct page *page, int node)
2056 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2062 static int count_free(struct page *page)
2064 return page->objects - page->inuse;
2067 static unsigned long count_partial(struct kmem_cache_node *n,
2068 int (*get_count)(struct page *))
2070 unsigned long flags;
2071 unsigned long x = 0;
2074 spin_lock_irqsave(&n->list_lock, flags);
2075 list_for_each_entry(page, &n->partial, lru)
2076 x += get_count(page);
2077 spin_unlock_irqrestore(&n->list_lock, flags);
2081 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2083 #ifdef CONFIG_SLUB_DEBUG
2084 return atomic_long_read(&n->total_objects);
2090 static noinline void
2091 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2096 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2098 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2099 "default order: %d, min order: %d\n", s->name, s->object_size,
2100 s->size, oo_order(s->oo), oo_order(s->min));
2102 if (oo_order(s->min) > get_order(s->object_size))
2103 printk(KERN_WARNING " %s debugging increased min order, use "
2104 "slub_debug=O to disable.\n", s->name);
2106 for_each_online_node(node) {
2107 struct kmem_cache_node *n = get_node(s, node);
2108 unsigned long nr_slabs;
2109 unsigned long nr_objs;
2110 unsigned long nr_free;
2115 nr_free = count_partial(n, count_free);
2116 nr_slabs = node_nr_slabs(n);
2117 nr_objs = node_nr_objs(n);
2120 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2121 node, nr_slabs, nr_objs, nr_free);
2125 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2126 int node, struct kmem_cache_cpu **pc)
2129 struct kmem_cache_cpu *c = *pc;
2132 freelist = get_partial(s, flags, node, c);
2137 page = new_slab(s, flags, node);
2139 c = __this_cpu_ptr(s->cpu_slab);
2144 * No other reference to the page yet so we can
2145 * muck around with it freely without cmpxchg
2147 freelist = page->freelist;
2148 page->freelist = NULL;
2150 stat(s, ALLOC_SLAB);
2159 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2161 if (unlikely(PageSlabPfmemalloc(page)))
2162 return gfp_pfmemalloc_allowed(gfpflags);
2168 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2169 * or deactivate the page.
2171 * The page is still frozen if the return value is not NULL.
2173 * If this function returns NULL then the page has been unfrozen.
2175 * This function must be called with interrupt disabled.
2177 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2180 unsigned long counters;
2184 freelist = page->freelist;
2185 counters = page->counters;
2187 new.counters = counters;
2188 VM_BUG_ON(!new.frozen);
2190 new.inuse = page->objects;
2191 new.frozen = freelist != NULL;
2193 } while (!__cmpxchg_double_slab(s, page,
2202 * Slow path. The lockless freelist is empty or we need to perform
2205 * Processing is still very fast if new objects have been freed to the
2206 * regular freelist. In that case we simply take over the regular freelist
2207 * as the lockless freelist and zap the regular freelist.
2209 * If that is not working then we fall back to the partial lists. We take the
2210 * first element of the freelist as the object to allocate now and move the
2211 * rest of the freelist to the lockless freelist.
2213 * And if we were unable to get a new slab from the partial slab lists then
2214 * we need to allocate a new slab. This is the slowest path since it involves
2215 * a call to the page allocator and the setup of a new slab.
2217 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2218 unsigned long addr, struct kmem_cache_cpu *c)
2222 unsigned long flags;
2224 local_irq_save(flags);
2225 #ifdef CONFIG_PREEMPT
2227 * We may have been preempted and rescheduled on a different
2228 * cpu before disabling interrupts. Need to reload cpu area
2231 c = this_cpu_ptr(s->cpu_slab);
2239 if (unlikely(!node_match(page, node))) {
2240 stat(s, ALLOC_NODE_MISMATCH);
2241 deactivate_slab(s, page, c->freelist);
2248 * By rights, we should be searching for a slab page that was
2249 * PFMEMALLOC but right now, we are losing the pfmemalloc
2250 * information when the page leaves the per-cpu allocator
2252 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2253 deactivate_slab(s, page, c->freelist);
2259 /* must check again c->freelist in case of cpu migration or IRQ */
2260 freelist = c->freelist;
2264 stat(s, ALLOC_SLOWPATH);
2266 freelist = get_freelist(s, page);
2270 stat(s, DEACTIVATE_BYPASS);
2274 stat(s, ALLOC_REFILL);
2278 * freelist is pointing to the list of objects to be used.
2279 * page is pointing to the page from which the objects are obtained.
2280 * That page must be frozen for per cpu allocations to work.
2282 VM_BUG_ON(!c->page->frozen);
2283 c->freelist = get_freepointer(s, freelist);
2284 c->tid = next_tid(c->tid);
2285 local_irq_restore(flags);
2291 page = c->page = c->partial;
2292 c->partial = page->next;
2293 stat(s, CPU_PARTIAL_ALLOC);
2298 freelist = new_slab_objects(s, gfpflags, node, &c);
2300 if (unlikely(!freelist)) {
2301 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2302 slab_out_of_memory(s, gfpflags, node);
2304 local_irq_restore(flags);
2309 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2312 /* Only entered in the debug case */
2313 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2314 goto new_slab; /* Slab failed checks. Next slab needed */
2316 deactivate_slab(s, page, get_freepointer(s, freelist));
2319 local_irq_restore(flags);
2324 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2325 * have the fastpath folded into their functions. So no function call
2326 * overhead for requests that can be satisfied on the fastpath.
2328 * The fastpath works by first checking if the lockless freelist can be used.
2329 * If not then __slab_alloc is called for slow processing.
2331 * Otherwise we can simply pick the next object from the lockless free list.
2333 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2334 gfp_t gfpflags, int node, unsigned long addr)
2337 struct kmem_cache_cpu *c;
2341 if (slab_pre_alloc_hook(s, gfpflags))
2344 s = memcg_kmem_get_cache(s, gfpflags);
2347 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2348 * enabled. We may switch back and forth between cpus while
2349 * reading from one cpu area. That does not matter as long
2350 * as we end up on the original cpu again when doing the cmpxchg.
2352 * Preemption is disabled for the retrieval of the tid because that
2353 * must occur from the current processor. We cannot allow rescheduling
2354 * on a different processor between the determination of the pointer
2355 * and the retrieval of the tid.
2358 c = __this_cpu_ptr(s->cpu_slab);
2361 * The transaction ids are globally unique per cpu and per operation on
2362 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2363 * occurs on the right processor and that there was no operation on the
2364 * linked list in between.
2369 object = c->freelist;
2371 if (unlikely(!object || !node_match(page, node)))
2372 object = __slab_alloc(s, gfpflags, node, addr, c);
2375 void *next_object = get_freepointer_safe(s, object);
2378 * The cmpxchg will only match if there was no additional
2379 * operation and if we are on the right processor.
2381 * The cmpxchg does the following atomically (without lock semantics!)
2382 * 1. Relocate first pointer to the current per cpu area.
2383 * 2. Verify that tid and freelist have not been changed
2384 * 3. If they were not changed replace tid and freelist
2386 * Since this is without lock semantics the protection is only against
2387 * code executing on this cpu *not* from access by other cpus.
2389 if (unlikely(!this_cpu_cmpxchg_double(
2390 s->cpu_slab->freelist, s->cpu_slab->tid,
2392 next_object, next_tid(tid)))) {
2394 note_cmpxchg_failure("slab_alloc", s, tid);
2397 prefetch_freepointer(s, next_object);
2398 stat(s, ALLOC_FASTPATH);
2401 if (unlikely(gfpflags & __GFP_ZERO) && object)
2402 memset(object, 0, s->object_size);
2404 slab_post_alloc_hook(s, gfpflags, object);
2409 static __always_inline void *slab_alloc(struct kmem_cache *s,
2410 gfp_t gfpflags, unsigned long addr)
2412 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2415 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2417 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2419 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2423 EXPORT_SYMBOL(kmem_cache_alloc);
2425 #ifdef CONFIG_TRACING
2426 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2428 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2429 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2432 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2434 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2436 void *ret = kmalloc_order(size, flags, order);
2437 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2440 EXPORT_SYMBOL(kmalloc_order_trace);
2444 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2446 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2448 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2449 s->object_size, s->size, gfpflags, node);
2453 EXPORT_SYMBOL(kmem_cache_alloc_node);
2455 #ifdef CONFIG_TRACING
2456 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2458 int node, size_t size)
2460 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2462 trace_kmalloc_node(_RET_IP_, ret,
2463 size, s->size, gfpflags, node);
2466 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2471 * Slow patch handling. This may still be called frequently since objects
2472 * have a longer lifetime than the cpu slabs in most processing loads.
2474 * So we still attempt to reduce cache line usage. Just take the slab
2475 * lock and free the item. If there is no additional partial page
2476 * handling required then we can return immediately.
2478 static void __slab_free(struct kmem_cache *s, struct page *page,
2479 void *x, unsigned long addr)
2482 void **object = (void *)x;
2485 unsigned long counters;
2486 struct kmem_cache_node *n = NULL;
2487 unsigned long uninitialized_var(flags);
2489 stat(s, FREE_SLOWPATH);
2491 if (kmem_cache_debug(s) &&
2492 !(n = free_debug_processing(s, page, x, addr, &flags)))
2497 spin_unlock_irqrestore(&n->list_lock, flags);
2500 prior = page->freelist;
2501 counters = page->counters;
2502 set_freepointer(s, object, prior);
2503 new.counters = counters;
2504 was_frozen = new.frozen;
2506 if ((!new.inuse || !prior) && !was_frozen) {
2508 if (!kmem_cache_debug(s) && !prior)
2511 * Slab was on no list before and will be partially empty
2512 * We can defer the list move and instead freeze it.
2516 else { /* Needs to be taken off a list */
2518 n = get_node(s, page_to_nid(page));
2520 * Speculatively acquire the list_lock.
2521 * If the cmpxchg does not succeed then we may
2522 * drop the list_lock without any processing.
2524 * Otherwise the list_lock will synchronize with
2525 * other processors updating the list of slabs.
2527 spin_lock_irqsave(&n->list_lock, flags);
2532 } while (!cmpxchg_double_slab(s, page,
2534 object, new.counters,
2540 * If we just froze the page then put it onto the
2541 * per cpu partial list.
2543 if (new.frozen && !was_frozen) {
2544 put_cpu_partial(s, page, 1);
2545 stat(s, CPU_PARTIAL_FREE);
2548 * The list lock was not taken therefore no list
2549 * activity can be necessary.
2552 stat(s, FREE_FROZEN);
2556 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2560 * Objects left in the slab. If it was not on the partial list before
2563 if (kmem_cache_debug(s) && unlikely(!prior)) {
2564 remove_full(s, page);
2565 add_partial(n, page, DEACTIVATE_TO_TAIL);
2566 stat(s, FREE_ADD_PARTIAL);
2568 spin_unlock_irqrestore(&n->list_lock, flags);
2574 * Slab on the partial list.
2576 remove_partial(n, page);
2577 stat(s, FREE_REMOVE_PARTIAL);
2579 /* Slab must be on the full list */
2580 remove_full(s, page);
2582 spin_unlock_irqrestore(&n->list_lock, flags);
2584 discard_slab(s, page);
2588 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2589 * can perform fastpath freeing without additional function calls.
2591 * The fastpath is only possible if we are freeing to the current cpu slab
2592 * of this processor. This typically the case if we have just allocated
2595 * If fastpath is not possible then fall back to __slab_free where we deal
2596 * with all sorts of special processing.
2598 static __always_inline void slab_free(struct kmem_cache *s,
2599 struct page *page, void *x, unsigned long addr)
2601 void **object = (void *)x;
2602 struct kmem_cache_cpu *c;
2605 slab_free_hook(s, x);
2609 * Determine the currently cpus per cpu slab.
2610 * The cpu may change afterward. However that does not matter since
2611 * data is retrieved via this pointer. If we are on the same cpu
2612 * during the cmpxchg then the free will succedd.
2615 c = __this_cpu_ptr(s->cpu_slab);
2620 if (likely(page == c->page)) {
2621 set_freepointer(s, object, c->freelist);
2623 if (unlikely(!this_cpu_cmpxchg_double(
2624 s->cpu_slab->freelist, s->cpu_slab->tid,
2626 object, next_tid(tid)))) {
2628 note_cmpxchg_failure("slab_free", s, tid);
2631 stat(s, FREE_FASTPATH);
2633 __slab_free(s, page, x, addr);
2637 void kmem_cache_free(struct kmem_cache *s, void *x)
2639 s = cache_from_obj(s, x);
2642 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2643 trace_kmem_cache_free(_RET_IP_, x);
2645 EXPORT_SYMBOL(kmem_cache_free);
2648 * Object placement in a slab is made very easy because we always start at
2649 * offset 0. If we tune the size of the object to the alignment then we can
2650 * get the required alignment by putting one properly sized object after
2653 * Notice that the allocation order determines the sizes of the per cpu
2654 * caches. Each processor has always one slab available for allocations.
2655 * Increasing the allocation order reduces the number of times that slabs
2656 * must be moved on and off the partial lists and is therefore a factor in
2661 * Mininum / Maximum order of slab pages. This influences locking overhead
2662 * and slab fragmentation. A higher order reduces the number of partial slabs
2663 * and increases the number of allocations possible without having to
2664 * take the list_lock.
2666 static int slub_min_order;
2667 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2668 static int slub_min_objects;
2671 * Merge control. If this is set then no merging of slab caches will occur.
2672 * (Could be removed. This was introduced to pacify the merge skeptics.)
2674 static int slub_nomerge;
2677 * Calculate the order of allocation given an slab object size.
2679 * The order of allocation has significant impact on performance and other
2680 * system components. Generally order 0 allocations should be preferred since
2681 * order 0 does not cause fragmentation in the page allocator. Larger objects
2682 * be problematic to put into order 0 slabs because there may be too much
2683 * unused space left. We go to a higher order if more than 1/16th of the slab
2686 * In order to reach satisfactory performance we must ensure that a minimum
2687 * number of objects is in one slab. Otherwise we may generate too much
2688 * activity on the partial lists which requires taking the list_lock. This is
2689 * less a concern for large slabs though which are rarely used.
2691 * slub_max_order specifies the order where we begin to stop considering the
2692 * number of objects in a slab as critical. If we reach slub_max_order then
2693 * we try to keep the page order as low as possible. So we accept more waste
2694 * of space in favor of a small page order.
2696 * Higher order allocations also allow the placement of more objects in a
2697 * slab and thereby reduce object handling overhead. If the user has
2698 * requested a higher mininum order then we start with that one instead of
2699 * the smallest order which will fit the object.
2701 static inline int slab_order(int size, int min_objects,
2702 int max_order, int fract_leftover, int reserved)
2706 int min_order = slub_min_order;
2708 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2709 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2711 for (order = max(min_order,
2712 fls(min_objects * size - 1) - PAGE_SHIFT);
2713 order <= max_order; order++) {
2715 unsigned long slab_size = PAGE_SIZE << order;
2717 if (slab_size < min_objects * size + reserved)
2720 rem = (slab_size - reserved) % size;
2722 if (rem <= slab_size / fract_leftover)
2730 static inline int calculate_order(int size, int reserved)
2738 * Attempt to find best configuration for a slab. This
2739 * works by first attempting to generate a layout with
2740 * the best configuration and backing off gradually.
2742 * First we reduce the acceptable waste in a slab. Then
2743 * we reduce the minimum objects required in a slab.
2745 min_objects = slub_min_objects;
2747 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2748 max_objects = order_objects(slub_max_order, size, reserved);
2749 min_objects = min(min_objects, max_objects);
2751 while (min_objects > 1) {
2753 while (fraction >= 4) {
2754 order = slab_order(size, min_objects,
2755 slub_max_order, fraction, reserved);
2756 if (order <= slub_max_order)
2764 * We were unable to place multiple objects in a slab. Now
2765 * lets see if we can place a single object there.
2767 order = slab_order(size, 1, slub_max_order, 1, reserved);
2768 if (order <= slub_max_order)
2772 * Doh this slab cannot be placed using slub_max_order.
2774 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2775 if (order < MAX_ORDER)
2781 init_kmem_cache_node(struct kmem_cache_node *n)
2784 spin_lock_init(&n->list_lock);
2785 INIT_LIST_HEAD(&n->partial);
2786 #ifdef CONFIG_SLUB_DEBUG
2787 atomic_long_set(&n->nr_slabs, 0);
2788 atomic_long_set(&n->total_objects, 0);
2789 INIT_LIST_HEAD(&n->full);
2793 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2795 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2796 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2799 * Must align to double word boundary for the double cmpxchg
2800 * instructions to work; see __pcpu_double_call_return_bool().
2802 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2803 2 * sizeof(void *));
2808 init_kmem_cache_cpus(s);
2813 static struct kmem_cache *kmem_cache_node;
2816 * No kmalloc_node yet so do it by hand. We know that this is the first
2817 * slab on the node for this slabcache. There are no concurrent accesses
2820 * Note that this function only works on the kmalloc_node_cache
2821 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2822 * memory on a fresh node that has no slab structures yet.
2824 static void early_kmem_cache_node_alloc(int node)
2827 struct kmem_cache_node *n;
2829 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2831 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2834 if (page_to_nid(page) != node) {
2835 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2837 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2838 "in order to be able to continue\n");
2843 page->freelist = get_freepointer(kmem_cache_node, n);
2846 kmem_cache_node->node[node] = n;
2847 #ifdef CONFIG_SLUB_DEBUG
2848 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2849 init_tracking(kmem_cache_node, n);
2851 init_kmem_cache_node(n);
2852 inc_slabs_node(kmem_cache_node, node, page->objects);
2854 add_partial(n, page, DEACTIVATE_TO_HEAD);
2857 static void free_kmem_cache_nodes(struct kmem_cache *s)
2861 for_each_node_state(node, N_NORMAL_MEMORY) {
2862 struct kmem_cache_node *n = s->node[node];
2865 kmem_cache_free(kmem_cache_node, n);
2867 s->node[node] = NULL;
2871 static int init_kmem_cache_nodes(struct kmem_cache *s)
2875 for_each_node_state(node, N_NORMAL_MEMORY) {
2876 struct kmem_cache_node *n;
2878 if (slab_state == DOWN) {
2879 early_kmem_cache_node_alloc(node);
2882 n = kmem_cache_alloc_node(kmem_cache_node,
2886 free_kmem_cache_nodes(s);
2891 init_kmem_cache_node(n);
2896 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2898 if (min < MIN_PARTIAL)
2900 else if (min > MAX_PARTIAL)
2902 s->min_partial = min;
2906 * calculate_sizes() determines the order and the distribution of data within
2909 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2911 unsigned long flags = s->flags;
2912 unsigned long size = s->object_size;
2916 * Round up object size to the next word boundary. We can only
2917 * place the free pointer at word boundaries and this determines
2918 * the possible location of the free pointer.
2920 size = ALIGN(size, sizeof(void *));
2922 #ifdef CONFIG_SLUB_DEBUG
2924 * Determine if we can poison the object itself. If the user of
2925 * the slab may touch the object after free or before allocation
2926 * then we should never poison the object itself.
2928 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2930 s->flags |= __OBJECT_POISON;
2932 s->flags &= ~__OBJECT_POISON;
2936 * If we are Redzoning then check if there is some space between the
2937 * end of the object and the free pointer. If not then add an
2938 * additional word to have some bytes to store Redzone information.
2940 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2941 size += sizeof(void *);
2945 * With that we have determined the number of bytes in actual use
2946 * by the object. This is the potential offset to the free pointer.
2950 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2953 * Relocate free pointer after the object if it is not
2954 * permitted to overwrite the first word of the object on
2957 * This is the case if we do RCU, have a constructor or
2958 * destructor or are poisoning the objects.
2961 size += sizeof(void *);
2964 #ifdef CONFIG_SLUB_DEBUG
2965 if (flags & SLAB_STORE_USER)
2967 * Need to store information about allocs and frees after
2970 size += 2 * sizeof(struct track);
2972 if (flags & SLAB_RED_ZONE)
2974 * Add some empty padding so that we can catch
2975 * overwrites from earlier objects rather than let
2976 * tracking information or the free pointer be
2977 * corrupted if a user writes before the start
2980 size += sizeof(void *);
2984 * SLUB stores one object immediately after another beginning from
2985 * offset 0. In order to align the objects we have to simply size
2986 * each object to conform to the alignment.
2988 size = ALIGN(size, s->align);
2990 if (forced_order >= 0)
2991 order = forced_order;
2993 order = calculate_order(size, s->reserved);
3000 s->allocflags |= __GFP_COMP;
3002 if (s->flags & SLAB_CACHE_DMA)
3003 s->allocflags |= GFP_DMA;
3005 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3006 s->allocflags |= __GFP_RECLAIMABLE;
3009 * Determine the number of objects per slab
3011 s->oo = oo_make(order, size, s->reserved);
3012 s->min = oo_make(get_order(size), size, s->reserved);
3013 if (oo_objects(s->oo) > oo_objects(s->max))
3016 return !!oo_objects(s->oo);
3019 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3021 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3024 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3025 s->reserved = sizeof(struct rcu_head);
3027 if (!calculate_sizes(s, -1))
3029 if (disable_higher_order_debug) {
3031 * Disable debugging flags that store metadata if the min slab
3034 if (get_order(s->size) > get_order(s->object_size)) {
3035 s->flags &= ~DEBUG_METADATA_FLAGS;
3037 if (!calculate_sizes(s, -1))
3042 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3043 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3044 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3045 /* Enable fast mode */
3046 s->flags |= __CMPXCHG_DOUBLE;
3050 * The larger the object size is, the more pages we want on the partial
3051 * list to avoid pounding the page allocator excessively.
3053 set_min_partial(s, ilog2(s->size) / 2);
3056 * cpu_partial determined the maximum number of objects kept in the
3057 * per cpu partial lists of a processor.
3059 * Per cpu partial lists mainly contain slabs that just have one
3060 * object freed. If they are used for allocation then they can be
3061 * filled up again with minimal effort. The slab will never hit the
3062 * per node partial lists and therefore no locking will be required.
3064 * This setting also determines
3066 * A) The number of objects from per cpu partial slabs dumped to the
3067 * per node list when we reach the limit.
3068 * B) The number of objects in cpu partial slabs to extract from the
3069 * per node list when we run out of per cpu objects. We only fetch 50%
3070 * to keep some capacity around for frees.
3072 if (kmem_cache_debug(s))
3074 else if (s->size >= PAGE_SIZE)
3076 else if (s->size >= 1024)
3078 else if (s->size >= 256)
3079 s->cpu_partial = 13;
3081 s->cpu_partial = 30;
3084 s->remote_node_defrag_ratio = 1000;
3086 if (!init_kmem_cache_nodes(s))
3089 if (alloc_kmem_cache_cpus(s))
3092 free_kmem_cache_nodes(s);
3094 if (flags & SLAB_PANIC)
3095 panic("Cannot create slab %s size=%lu realsize=%u "
3096 "order=%u offset=%u flags=%lx\n",
3097 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3102 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3105 #ifdef CONFIG_SLUB_DEBUG
3106 void *addr = page_address(page);
3108 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3109 sizeof(long), GFP_ATOMIC);
3112 slab_err(s, page, text, s->name);
3115 get_map(s, page, map);
3116 for_each_object(p, s, addr, page->objects) {
3118 if (!test_bit(slab_index(p, s, addr), map)) {
3119 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3121 print_tracking(s, p);
3130 * Attempt to free all partial slabs on a node.
3131 * This is called from kmem_cache_close(). We must be the last thread
3132 * using the cache and therefore we do not need to lock anymore.
3134 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3136 struct page *page, *h;
3138 list_for_each_entry_safe(page, h, &n->partial, lru) {
3140 remove_partial(n, page);
3141 discard_slab(s, page);
3143 list_slab_objects(s, page,
3144 "Objects remaining in %s on kmem_cache_close()");
3150 * Release all resources used by a slab cache.
3152 static inline int kmem_cache_close(struct kmem_cache *s)
3157 /* Attempt to free all objects */
3158 for_each_node_state(node, N_NORMAL_MEMORY) {
3159 struct kmem_cache_node *n = get_node(s, node);
3162 if (n->nr_partial || slabs_node(s, node))
3165 free_percpu(s->cpu_slab);
3166 free_kmem_cache_nodes(s);
3170 int __kmem_cache_shutdown(struct kmem_cache *s)
3172 int rc = kmem_cache_close(s);
3176 * We do the same lock strategy around sysfs_slab_add, see
3177 * __kmem_cache_create. Because this is pretty much the last
3178 * operation we do and the lock will be released shortly after
3179 * that in slab_common.c, we could just move sysfs_slab_remove
3180 * to a later point in common code. We should do that when we
3181 * have a common sysfs framework for all allocators.
3183 mutex_unlock(&slab_mutex);
3184 sysfs_slab_remove(s);
3185 mutex_lock(&slab_mutex);
3191 /********************************************************************
3193 *******************************************************************/
3195 static int __init setup_slub_min_order(char *str)
3197 get_option(&str, &slub_min_order);
3202 __setup("slub_min_order=", setup_slub_min_order);
3204 static int __init setup_slub_max_order(char *str)
3206 get_option(&str, &slub_max_order);
3207 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3212 __setup("slub_max_order=", setup_slub_max_order);
3214 static int __init setup_slub_min_objects(char *str)
3216 get_option(&str, &slub_min_objects);
3221 __setup("slub_min_objects=", setup_slub_min_objects);
3223 static int __init setup_slub_nomerge(char *str)
3229 __setup("slub_nomerge", setup_slub_nomerge);
3231 void *__kmalloc(size_t size, gfp_t flags)
3233 struct kmem_cache *s;
3236 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3237 return kmalloc_large(size, flags);
3239 s = kmalloc_slab(size, flags);
3241 if (unlikely(ZERO_OR_NULL_PTR(s)))
3244 ret = slab_alloc(s, flags, _RET_IP_);
3246 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3250 EXPORT_SYMBOL(__kmalloc);
3253 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3258 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3259 page = alloc_pages_node(node, flags, get_order(size));
3261 ptr = page_address(page);
3263 kmemleak_alloc(ptr, size, 1, flags);
3267 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3269 struct kmem_cache *s;
3272 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3273 ret = kmalloc_large_node(size, flags, node);
3275 trace_kmalloc_node(_RET_IP_, ret,
3276 size, PAGE_SIZE << get_order(size),
3282 s = kmalloc_slab(size, flags);
3284 if (unlikely(ZERO_OR_NULL_PTR(s)))
3287 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3289 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3293 EXPORT_SYMBOL(__kmalloc_node);
3296 size_t ksize(const void *object)
3300 if (unlikely(object == ZERO_SIZE_PTR))
3303 page = virt_to_head_page(object);
3305 if (unlikely(!PageSlab(page))) {
3306 WARN_ON(!PageCompound(page));
3307 return PAGE_SIZE << compound_order(page);
3310 return slab_ksize(page->slab_cache);
3312 EXPORT_SYMBOL(ksize);
3314 #ifdef CONFIG_SLUB_DEBUG
3315 bool verify_mem_not_deleted(const void *x)
3318 void *object = (void *)x;
3319 unsigned long flags;
3322 if (unlikely(ZERO_OR_NULL_PTR(x)))
3325 local_irq_save(flags);
3327 page = virt_to_head_page(x);
3328 if (unlikely(!PageSlab(page))) {
3329 /* maybe it was from stack? */
3335 if (on_freelist(page->slab_cache, page, object)) {
3336 object_err(page->slab_cache, page, object, "Object is on free-list");
3344 local_irq_restore(flags);
3347 EXPORT_SYMBOL(verify_mem_not_deleted);
3350 void kfree(const void *x)
3353 void *object = (void *)x;
3355 trace_kfree(_RET_IP_, x);
3357 if (unlikely(ZERO_OR_NULL_PTR(x)))
3360 page = virt_to_head_page(x);
3361 if (unlikely(!PageSlab(page))) {
3362 BUG_ON(!PageCompound(page));
3364 __free_memcg_kmem_pages(page, compound_order(page));
3367 slab_free(page->slab_cache, page, object, _RET_IP_);
3369 EXPORT_SYMBOL(kfree);
3372 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3373 * the remaining slabs by the number of items in use. The slabs with the
3374 * most items in use come first. New allocations will then fill those up
3375 * and thus they can be removed from the partial lists.
3377 * The slabs with the least items are placed last. This results in them
3378 * being allocated from last increasing the chance that the last objects
3379 * are freed in them.
3381 int kmem_cache_shrink(struct kmem_cache *s)
3385 struct kmem_cache_node *n;
3388 int objects = oo_objects(s->max);
3389 struct list_head *slabs_by_inuse =
3390 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3391 unsigned long flags;
3393 if (!slabs_by_inuse)
3397 for_each_node_state(node, N_NORMAL_MEMORY) {
3398 n = get_node(s, node);
3403 for (i = 0; i < objects; i++)
3404 INIT_LIST_HEAD(slabs_by_inuse + i);
3406 spin_lock_irqsave(&n->list_lock, flags);
3409 * Build lists indexed by the items in use in each slab.
3411 * Note that concurrent frees may occur while we hold the
3412 * list_lock. page->inuse here is the upper limit.
3414 list_for_each_entry_safe(page, t, &n->partial, lru) {
3415 list_move(&page->lru, slabs_by_inuse + page->inuse);
3421 * Rebuild the partial list with the slabs filled up most
3422 * first and the least used slabs at the end.
3424 for (i = objects - 1; i > 0; i--)
3425 list_splice(slabs_by_inuse + i, n->partial.prev);
3427 spin_unlock_irqrestore(&n->list_lock, flags);
3429 /* Release empty slabs */
3430 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3431 discard_slab(s, page);
3434 kfree(slabs_by_inuse);
3437 EXPORT_SYMBOL(kmem_cache_shrink);
3439 static int slab_mem_going_offline_callback(void *arg)
3441 struct kmem_cache *s;
3443 mutex_lock(&slab_mutex);
3444 list_for_each_entry(s, &slab_caches, list)
3445 kmem_cache_shrink(s);
3446 mutex_unlock(&slab_mutex);
3451 static void slab_mem_offline_callback(void *arg)
3453 struct kmem_cache_node *n;
3454 struct kmem_cache *s;
3455 struct memory_notify *marg = arg;
3458 offline_node = marg->status_change_nid_normal;
3461 * If the node still has available memory. we need kmem_cache_node
3464 if (offline_node < 0)
3467 mutex_lock(&slab_mutex);
3468 list_for_each_entry(s, &slab_caches, list) {
3469 n = get_node(s, offline_node);
3472 * if n->nr_slabs > 0, slabs still exist on the node
3473 * that is going down. We were unable to free them,
3474 * and offline_pages() function shouldn't call this
3475 * callback. So, we must fail.
3477 BUG_ON(slabs_node(s, offline_node));
3479 s->node[offline_node] = NULL;
3480 kmem_cache_free(kmem_cache_node, n);
3483 mutex_unlock(&slab_mutex);
3486 static int slab_mem_going_online_callback(void *arg)
3488 struct kmem_cache_node *n;
3489 struct kmem_cache *s;
3490 struct memory_notify *marg = arg;
3491 int nid = marg->status_change_nid_normal;
3495 * If the node's memory is already available, then kmem_cache_node is
3496 * already created. Nothing to do.
3502 * We are bringing a node online. No memory is available yet. We must
3503 * allocate a kmem_cache_node structure in order to bring the node
3506 mutex_lock(&slab_mutex);
3507 list_for_each_entry(s, &slab_caches, list) {
3509 * XXX: kmem_cache_alloc_node will fallback to other nodes
3510 * since memory is not yet available from the node that
3513 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3518 init_kmem_cache_node(n);
3522 mutex_unlock(&slab_mutex);
3526 static int slab_memory_callback(struct notifier_block *self,
3527 unsigned long action, void *arg)
3532 case MEM_GOING_ONLINE:
3533 ret = slab_mem_going_online_callback(arg);
3535 case MEM_GOING_OFFLINE:
3536 ret = slab_mem_going_offline_callback(arg);
3539 case MEM_CANCEL_ONLINE:
3540 slab_mem_offline_callback(arg);
3543 case MEM_CANCEL_OFFLINE:
3547 ret = notifier_from_errno(ret);
3553 static struct notifier_block slab_memory_callback_nb = {
3554 .notifier_call = slab_memory_callback,
3555 .priority = SLAB_CALLBACK_PRI,
3558 /********************************************************************
3559 * Basic setup of slabs
3560 *******************************************************************/
3563 * Used for early kmem_cache structures that were allocated using
3564 * the page allocator. Allocate them properly then fix up the pointers
3565 * that may be pointing to the wrong kmem_cache structure.
3568 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3571 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3573 memcpy(s, static_cache, kmem_cache->object_size);
3576 * This runs very early, and only the boot processor is supposed to be
3577 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3580 __flush_cpu_slab(s, smp_processor_id());
3581 for_each_node_state(node, N_NORMAL_MEMORY) {
3582 struct kmem_cache_node *n = get_node(s, node);
3586 list_for_each_entry(p, &n->partial, lru)
3589 #ifdef CONFIG_SLUB_DEBUG
3590 list_for_each_entry(p, &n->full, lru)
3595 list_add(&s->list, &slab_caches);
3599 void __init kmem_cache_init(void)
3601 static __initdata struct kmem_cache boot_kmem_cache,
3602 boot_kmem_cache_node;
3604 if (debug_guardpage_minorder())
3607 kmem_cache_node = &boot_kmem_cache_node;
3608 kmem_cache = &boot_kmem_cache;
3610 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3611 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3613 register_hotmemory_notifier(&slab_memory_callback_nb);
3615 /* Able to allocate the per node structures */
3616 slab_state = PARTIAL;
3618 create_boot_cache(kmem_cache, "kmem_cache",
3619 offsetof(struct kmem_cache, node) +
3620 nr_node_ids * sizeof(struct kmem_cache_node *),
3621 SLAB_HWCACHE_ALIGN);
3623 kmem_cache = bootstrap(&boot_kmem_cache);
3626 * Allocate kmem_cache_node properly from the kmem_cache slab.
3627 * kmem_cache_node is separately allocated so no need to
3628 * update any list pointers.
3630 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3632 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3633 create_kmalloc_caches(0);
3636 register_cpu_notifier(&slab_notifier);
3640 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3641 " CPUs=%d, Nodes=%d\n",
3643 slub_min_order, slub_max_order, slub_min_objects,
3644 nr_cpu_ids, nr_node_ids);
3647 void __init kmem_cache_init_late(void)
3652 * Find a mergeable slab cache
3654 static int slab_unmergeable(struct kmem_cache *s)
3656 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3663 * We may have set a slab to be unmergeable during bootstrap.
3665 if (s->refcount < 0)
3671 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3672 size_t align, unsigned long flags, const char *name,
3673 void (*ctor)(void *))
3675 struct kmem_cache *s;
3677 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3683 size = ALIGN(size, sizeof(void *));
3684 align = calculate_alignment(flags, align, size);
3685 size = ALIGN(size, align);
3686 flags = kmem_cache_flags(size, flags, name, NULL);
3688 list_for_each_entry(s, &slab_caches, list) {
3689 if (slab_unmergeable(s))
3695 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3698 * Check if alignment is compatible.
3699 * Courtesy of Adrian Drzewiecki
3701 if ((s->size & ~(align - 1)) != s->size)
3704 if (s->size - size >= sizeof(void *))
3707 if (!cache_match_memcg(s, memcg))
3716 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3717 size_t align, unsigned long flags, void (*ctor)(void *))
3719 struct kmem_cache *s;
3721 s = find_mergeable(memcg, size, align, flags, name, ctor);
3725 * Adjust the object sizes so that we clear
3726 * the complete object on kzalloc.
3728 s->object_size = max(s->object_size, (int)size);
3729 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3731 if (sysfs_slab_alias(s, name)) {
3740 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3744 err = kmem_cache_open(s, flags);
3748 /* Mutex is not taken during early boot */
3749 if (slab_state <= UP)
3752 memcg_propagate_slab_attrs(s);
3753 mutex_unlock(&slab_mutex);
3754 err = sysfs_slab_add(s);
3755 mutex_lock(&slab_mutex);
3758 kmem_cache_close(s);
3765 * Use the cpu notifier to insure that the cpu slabs are flushed when
3768 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3769 unsigned long action, void *hcpu)
3771 long cpu = (long)hcpu;
3772 struct kmem_cache *s;
3773 unsigned long flags;
3776 case CPU_UP_CANCELED:
3777 case CPU_UP_CANCELED_FROZEN:
3779 case CPU_DEAD_FROZEN:
3780 mutex_lock(&slab_mutex);
3781 list_for_each_entry(s, &slab_caches, list) {
3782 local_irq_save(flags);
3783 __flush_cpu_slab(s, cpu);
3784 local_irq_restore(flags);
3786 mutex_unlock(&slab_mutex);
3794 static struct notifier_block __cpuinitdata slab_notifier = {
3795 .notifier_call = slab_cpuup_callback
3800 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3802 struct kmem_cache *s;
3805 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3806 return kmalloc_large(size, gfpflags);
3808 s = kmalloc_slab(size, gfpflags);
3810 if (unlikely(ZERO_OR_NULL_PTR(s)))
3813 ret = slab_alloc(s, gfpflags, caller);
3815 /* Honor the call site pointer we received. */
3816 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3822 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3823 int node, unsigned long caller)
3825 struct kmem_cache *s;
3828 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3829 ret = kmalloc_large_node(size, gfpflags, node);
3831 trace_kmalloc_node(caller, ret,
3832 size, PAGE_SIZE << get_order(size),
3838 s = kmalloc_slab(size, gfpflags);
3840 if (unlikely(ZERO_OR_NULL_PTR(s)))
3843 ret = slab_alloc_node(s, gfpflags, node, caller);
3845 /* Honor the call site pointer we received. */
3846 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3853 static int count_inuse(struct page *page)
3858 static int count_total(struct page *page)
3860 return page->objects;
3864 #ifdef CONFIG_SLUB_DEBUG
3865 static int validate_slab(struct kmem_cache *s, struct page *page,
3869 void *addr = page_address(page);
3871 if (!check_slab(s, page) ||
3872 !on_freelist(s, page, NULL))
3875 /* Now we know that a valid freelist exists */
3876 bitmap_zero(map, page->objects);
3878 get_map(s, page, map);
3879 for_each_object(p, s, addr, page->objects) {
3880 if (test_bit(slab_index(p, s, addr), map))
3881 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3885 for_each_object(p, s, addr, page->objects)
3886 if (!test_bit(slab_index(p, s, addr), map))
3887 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3892 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3896 validate_slab(s, page, map);
3900 static int validate_slab_node(struct kmem_cache *s,
3901 struct kmem_cache_node *n, unsigned long *map)
3903 unsigned long count = 0;
3905 unsigned long flags;
3907 spin_lock_irqsave(&n->list_lock, flags);
3909 list_for_each_entry(page, &n->partial, lru) {
3910 validate_slab_slab(s, page, map);
3913 if (count != n->nr_partial)
3914 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3915 "counter=%ld\n", s->name, count, n->nr_partial);
3917 if (!(s->flags & SLAB_STORE_USER))
3920 list_for_each_entry(page, &n->full, lru) {
3921 validate_slab_slab(s, page, map);
3924 if (count != atomic_long_read(&n->nr_slabs))
3925 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3926 "counter=%ld\n", s->name, count,
3927 atomic_long_read(&n->nr_slabs));
3930 spin_unlock_irqrestore(&n->list_lock, flags);
3934 static long validate_slab_cache(struct kmem_cache *s)
3937 unsigned long count = 0;
3938 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3939 sizeof(unsigned long), GFP_KERNEL);
3945 for_each_node_state(node, N_NORMAL_MEMORY) {
3946 struct kmem_cache_node *n = get_node(s, node);
3948 count += validate_slab_node(s, n, map);
3954 * Generate lists of code addresses where slabcache objects are allocated
3959 unsigned long count;
3966 DECLARE_BITMAP(cpus, NR_CPUS);
3972 unsigned long count;
3973 struct location *loc;
3976 static void free_loc_track(struct loc_track *t)
3979 #ifndef CONFIG_SPRD_PAGERECORDER
3980 free_pages((unsigned long)t->loc,
3981 get_order(sizeof(struct location) * t->max));
3983 __free_pages_nopagedebug((unsigned long)t->loc,
3984 get_order(sizeof(struct location) * t->max));
3988 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3993 order = get_order(sizeof(struct location) * max);
3995 #ifndef CONFIG_SPRD_PAGERECORDER
3996 l = (void *)__get_free_pages(flags, order);
3998 l = (void *)__get_free_pages_nopagedebug(flags, order);
4005 memcpy(l, t->loc, sizeof(struct location) * t->count);
4013 static int add_location(struct loc_track *t, struct kmem_cache *s,
4014 const struct track *track)
4016 long start, end, pos;
4018 unsigned long caddr;
4019 unsigned long age = jiffies - track->when;
4025 pos = start + (end - start + 1) / 2;
4028 * There is nothing at "end". If we end up there
4029 * we need to add something to before end.
4034 caddr = t->loc[pos].addr;
4035 if (track->addr == caddr) {
4041 if (age < l->min_time)
4043 if (age > l->max_time)
4046 if (track->pid < l->min_pid)
4047 l->min_pid = track->pid;
4048 if (track->pid > l->max_pid)
4049 l->max_pid = track->pid;
4051 cpumask_set_cpu(track->cpu,
4052 to_cpumask(l->cpus));
4054 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4058 if (track->addr < caddr)
4065 * Not found. Insert new tracking element.
4067 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4073 (t->count - pos) * sizeof(struct location));
4076 l->addr = track->addr;
4080 l->min_pid = track->pid;
4081 l->max_pid = track->pid;
4082 cpumask_clear(to_cpumask(l->cpus));
4083 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4084 nodes_clear(l->nodes);
4085 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4089 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4090 struct page *page, enum track_item alloc,
4093 void *addr = page_address(page);
4096 bitmap_zero(map, page->objects);
4097 get_map(s, page, map);
4099 for_each_object(p, s, addr, page->objects)
4100 if (!test_bit(slab_index(p, s, addr), map))
4101 add_location(t, s, get_track(s, p, alloc));
4104 static int list_locations(struct kmem_cache *s, char *buf,
4105 enum track_item alloc)
4109 struct loc_track t = { 0, 0, NULL };
4111 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4112 sizeof(unsigned long), GFP_KERNEL);
4114 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4117 return sprintf(buf, "Out of memory\n");
4119 /* Push back cpu slabs */
4122 for_each_node_state(node, N_NORMAL_MEMORY) {
4123 struct kmem_cache_node *n = get_node(s, node);
4124 unsigned long flags;
4127 if (!atomic_long_read(&n->nr_slabs))
4130 spin_lock_irqsave(&n->list_lock, flags);
4131 list_for_each_entry(page, &n->partial, lru)
4132 process_slab(&t, s, page, alloc, map);
4133 list_for_each_entry(page, &n->full, lru)
4134 process_slab(&t, s, page, alloc, map);
4135 spin_unlock_irqrestore(&n->list_lock, flags);
4138 for (i = 0; i < t.count; i++) {
4139 struct location *l = &t.loc[i];
4141 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4143 len += sprintf(buf + len, "%7ld ", l->count);
4146 len += sprintf(buf + len, "%pS", (void *)l->addr);
4148 len += sprintf(buf + len, "<not-available>");
4150 if (l->sum_time != l->min_time) {
4151 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4153 (long)div_u64(l->sum_time, l->count),
4156 len += sprintf(buf + len, " age=%ld",
4159 if (l->min_pid != l->max_pid)
4160 len += sprintf(buf + len, " pid=%ld-%ld",
4161 l->min_pid, l->max_pid);
4163 len += sprintf(buf + len, " pid=%ld",
4166 if (num_online_cpus() > 1 &&
4167 !cpumask_empty(to_cpumask(l->cpus)) &&
4168 len < PAGE_SIZE - 60) {
4169 len += sprintf(buf + len, " cpus=");
4170 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4171 to_cpumask(l->cpus));
4174 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4175 len < PAGE_SIZE - 60) {
4176 len += sprintf(buf + len, " nodes=");
4177 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4181 len += sprintf(buf + len, "\n");
4187 len += sprintf(buf, "No data\n");
4192 #ifdef SLUB_RESILIENCY_TEST
4193 static void resiliency_test(void)
4197 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4199 printk(KERN_ERR "SLUB resiliency testing\n");
4200 printk(KERN_ERR "-----------------------\n");
4201 printk(KERN_ERR "A. Corruption after allocation\n");
4203 p = kzalloc(16, GFP_KERNEL);
4205 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4206 " 0x12->0x%p\n\n", p + 16);
4208 validate_slab_cache(kmalloc_caches[4]);
4210 /* Hmmm... The next two are dangerous */
4211 p = kzalloc(32, GFP_KERNEL);
4212 p[32 + sizeof(void *)] = 0x34;
4213 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4214 " 0x34 -> -0x%p\n", p);
4216 "If allocated object is overwritten then not detectable\n\n");
4218 validate_slab_cache(kmalloc_caches[5]);
4219 p = kzalloc(64, GFP_KERNEL);
4220 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4222 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4225 "If allocated object is overwritten then not detectable\n\n");
4226 validate_slab_cache(kmalloc_caches[6]);
4228 printk(KERN_ERR "\nB. Corruption after free\n");
4229 p = kzalloc(128, GFP_KERNEL);
4232 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4233 validate_slab_cache(kmalloc_caches[7]);
4235 p = kzalloc(256, GFP_KERNEL);
4238 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4240 validate_slab_cache(kmalloc_caches[8]);
4242 p = kzalloc(512, GFP_KERNEL);
4245 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4246 validate_slab_cache(kmalloc_caches[9]);
4250 static void resiliency_test(void) {};
4255 enum slab_stat_type {
4256 SL_ALL, /* All slabs */
4257 SL_PARTIAL, /* Only partially allocated slabs */
4258 SL_CPU, /* Only slabs used for cpu caches */
4259 SL_OBJECTS, /* Determine allocated objects not slabs */
4260 SL_TOTAL /* Determine object capacity not slabs */
4263 #define SO_ALL (1 << SL_ALL)
4264 #define SO_PARTIAL (1 << SL_PARTIAL)
4265 #define SO_CPU (1 << SL_CPU)
4266 #define SO_OBJECTS (1 << SL_OBJECTS)
4267 #define SO_TOTAL (1 << SL_TOTAL)
4269 static ssize_t show_slab_objects(struct kmem_cache *s,
4270 char *buf, unsigned long flags)
4272 unsigned long total = 0;
4275 unsigned long *nodes;
4276 unsigned long *per_cpu;
4278 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4281 per_cpu = nodes + nr_node_ids;
4283 if (flags & SO_CPU) {
4286 for_each_possible_cpu(cpu) {
4287 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4291 page = ACCESS_ONCE(c->page);
4295 node = page_to_nid(page);
4296 if (flags & SO_TOTAL)
4298 else if (flags & SO_OBJECTS)
4306 page = ACCESS_ONCE(c->partial);
4308 node = page_to_nid(page);
4309 if (flags & SO_TOTAL)
4311 else if (flags & SO_OBJECTS)
4323 lock_memory_hotplug();
4324 #ifdef CONFIG_SLUB_DEBUG
4325 if (flags & SO_ALL) {
4326 for_each_node_state(node, N_NORMAL_MEMORY) {
4327 struct kmem_cache_node *n = get_node(s, node);
4329 if (flags & SO_TOTAL)
4330 x = atomic_long_read(&n->total_objects);
4331 else if (flags & SO_OBJECTS)
4332 x = atomic_long_read(&n->total_objects) -
4333 count_partial(n, count_free);
4336 x = atomic_long_read(&n->nr_slabs);
4343 if (flags & SO_PARTIAL) {
4344 for_each_node_state(node, N_NORMAL_MEMORY) {
4345 struct kmem_cache_node *n = get_node(s, node);
4347 if (flags & SO_TOTAL)
4348 x = count_partial(n, count_total);
4349 else if (flags & SO_OBJECTS)
4350 x = count_partial(n, count_inuse);
4357 x = sprintf(buf, "%lu", total);
4359 for_each_node_state(node, N_NORMAL_MEMORY)
4361 x += sprintf(buf + x, " N%d=%lu",
4364 unlock_memory_hotplug();
4366 return x + sprintf(buf + x, "\n");
4369 #ifdef CONFIG_SLUB_DEBUG
4370 static int any_slab_objects(struct kmem_cache *s)
4374 for_each_online_node(node) {
4375 struct kmem_cache_node *n = get_node(s, node);
4380 if (atomic_long_read(&n->total_objects))
4387 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4388 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4390 struct slab_attribute {
4391 struct attribute attr;
4392 ssize_t (*show)(struct kmem_cache *s, char *buf);
4393 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4396 #define SLAB_ATTR_RO(_name) \
4397 static struct slab_attribute _name##_attr = \
4398 __ATTR(_name, 0400, _name##_show, NULL)
4400 #define SLAB_ATTR(_name) \
4401 static struct slab_attribute _name##_attr = \
4402 __ATTR(_name, 0600, _name##_show, _name##_store)
4404 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4406 return sprintf(buf, "%d\n", s->size);
4408 SLAB_ATTR_RO(slab_size);
4410 static ssize_t align_show(struct kmem_cache *s, char *buf)
4412 return sprintf(buf, "%d\n", s->align);
4414 SLAB_ATTR_RO(align);
4416 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4418 return sprintf(buf, "%d\n", s->object_size);
4420 SLAB_ATTR_RO(object_size);
4422 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4424 return sprintf(buf, "%d\n", oo_objects(s->oo));
4426 SLAB_ATTR_RO(objs_per_slab);
4428 static ssize_t order_store(struct kmem_cache *s,
4429 const char *buf, size_t length)
4431 unsigned long order;
4434 err = strict_strtoul(buf, 10, &order);
4438 if (order > slub_max_order || order < slub_min_order)
4441 calculate_sizes(s, order);
4445 static ssize_t order_show(struct kmem_cache *s, char *buf)
4447 return sprintf(buf, "%d\n", oo_order(s->oo));
4451 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4453 return sprintf(buf, "%lu\n", s->min_partial);
4456 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4462 err = strict_strtoul(buf, 10, &min);
4466 set_min_partial(s, min);
4469 SLAB_ATTR(min_partial);
4471 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4473 return sprintf(buf, "%u\n", s->cpu_partial);
4476 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4479 unsigned long objects;
4482 err = strict_strtoul(buf, 10, &objects);
4485 if (objects && kmem_cache_debug(s))
4488 s->cpu_partial = objects;
4492 SLAB_ATTR(cpu_partial);
4494 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4498 return sprintf(buf, "%pS\n", s->ctor);
4502 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4504 return sprintf(buf, "%d\n", s->refcount - 1);
4506 SLAB_ATTR_RO(aliases);
4508 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4510 return show_slab_objects(s, buf, SO_PARTIAL);
4512 SLAB_ATTR_RO(partial);
4514 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4516 return show_slab_objects(s, buf, SO_CPU);
4518 SLAB_ATTR_RO(cpu_slabs);
4520 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4522 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4524 SLAB_ATTR_RO(objects);
4526 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4528 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4530 SLAB_ATTR_RO(objects_partial);
4532 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4539 for_each_online_cpu(cpu) {
4540 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4543 pages += page->pages;
4544 objects += page->pobjects;
4548 len = sprintf(buf, "%d(%d)", objects, pages);
4551 for_each_online_cpu(cpu) {
4552 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4554 if (page && len < PAGE_SIZE - 20)
4555 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4556 page->pobjects, page->pages);
4559 return len + sprintf(buf + len, "\n");
4561 SLAB_ATTR_RO(slabs_cpu_partial);
4563 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4565 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4568 static ssize_t reclaim_account_store(struct kmem_cache *s,
4569 const char *buf, size_t length)
4571 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4573 s->flags |= SLAB_RECLAIM_ACCOUNT;
4576 SLAB_ATTR(reclaim_account);
4578 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4580 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4582 SLAB_ATTR_RO(hwcache_align);
4584 #ifdef CONFIG_ZONE_DMA
4585 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4587 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4589 SLAB_ATTR_RO(cache_dma);
4592 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4594 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4596 SLAB_ATTR_RO(destroy_by_rcu);
4598 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4600 return sprintf(buf, "%d\n", s->reserved);
4602 SLAB_ATTR_RO(reserved);
4604 #ifdef CONFIG_SLUB_DEBUG
4605 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4607 return show_slab_objects(s, buf, SO_ALL);
4609 SLAB_ATTR_RO(slabs);
4611 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4613 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4615 SLAB_ATTR_RO(total_objects);
4617 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4619 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4622 static ssize_t sanity_checks_store(struct kmem_cache *s,
4623 const char *buf, size_t length)
4625 s->flags &= ~SLAB_DEBUG_FREE;
4626 if (buf[0] == '1') {
4627 s->flags &= ~__CMPXCHG_DOUBLE;
4628 s->flags |= SLAB_DEBUG_FREE;
4632 SLAB_ATTR(sanity_checks);
4634 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4636 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4639 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4642 s->flags &= ~SLAB_TRACE;
4643 if (buf[0] == '1') {
4644 s->flags &= ~__CMPXCHG_DOUBLE;
4645 s->flags |= SLAB_TRACE;
4651 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4653 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4656 static ssize_t red_zone_store(struct kmem_cache *s,
4657 const char *buf, size_t length)
4659 if (any_slab_objects(s))
4662 s->flags &= ~SLAB_RED_ZONE;
4663 if (buf[0] == '1') {
4664 s->flags &= ~__CMPXCHG_DOUBLE;
4665 s->flags |= SLAB_RED_ZONE;
4667 calculate_sizes(s, -1);
4670 SLAB_ATTR(red_zone);
4672 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4674 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4677 static ssize_t poison_store(struct kmem_cache *s,
4678 const char *buf, size_t length)
4680 if (any_slab_objects(s))
4683 s->flags &= ~SLAB_POISON;
4684 if (buf[0] == '1') {
4685 s->flags &= ~__CMPXCHG_DOUBLE;
4686 s->flags |= SLAB_POISON;
4688 calculate_sizes(s, -1);
4693 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4695 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4698 static ssize_t store_user_store(struct kmem_cache *s,
4699 const char *buf, size_t length)
4701 if (any_slab_objects(s))
4704 s->flags &= ~SLAB_STORE_USER;
4705 if (buf[0] == '1') {
4706 s->flags &= ~__CMPXCHG_DOUBLE;
4707 s->flags |= SLAB_STORE_USER;
4709 calculate_sizes(s, -1);
4712 SLAB_ATTR(store_user);
4714 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4719 static ssize_t validate_store(struct kmem_cache *s,
4720 const char *buf, size_t length)
4724 if (buf[0] == '1') {
4725 ret = validate_slab_cache(s);
4731 SLAB_ATTR(validate);
4733 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4735 if (!(s->flags & SLAB_STORE_USER))
4737 return list_locations(s, buf, TRACK_ALLOC);
4739 SLAB_ATTR_RO(alloc_calls);
4741 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4743 if (!(s->flags & SLAB_STORE_USER))
4745 return list_locations(s, buf, TRACK_FREE);
4747 SLAB_ATTR_RO(free_calls);
4748 #endif /* CONFIG_SLUB_DEBUG */
4750 #ifdef CONFIG_FAILSLAB
4751 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4753 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4756 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4759 s->flags &= ~SLAB_FAILSLAB;
4761 s->flags |= SLAB_FAILSLAB;
4764 SLAB_ATTR(failslab);
4767 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4772 static ssize_t shrink_store(struct kmem_cache *s,
4773 const char *buf, size_t length)
4775 if (buf[0] == '1') {
4776 int rc = kmem_cache_shrink(s);
4787 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4789 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4792 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4793 const char *buf, size_t length)
4795 unsigned long ratio;
4798 err = strict_strtoul(buf, 10, &ratio);
4803 s->remote_node_defrag_ratio = ratio * 10;
4807 SLAB_ATTR(remote_node_defrag_ratio);
4810 #ifdef CONFIG_SLUB_STATS
4811 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4813 unsigned long sum = 0;
4816 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4821 for_each_online_cpu(cpu) {
4822 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4828 len = sprintf(buf, "%lu", sum);
4831 for_each_online_cpu(cpu) {
4832 if (data[cpu] && len < PAGE_SIZE - 20)
4833 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4837 return len + sprintf(buf + len, "\n");
4840 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4844 for_each_online_cpu(cpu)
4845 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4848 #define STAT_ATTR(si, text) \
4849 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4851 return show_stat(s, buf, si); \
4853 static ssize_t text##_store(struct kmem_cache *s, \
4854 const char *buf, size_t length) \
4856 if (buf[0] != '0') \
4858 clear_stat(s, si); \
4863 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4864 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4865 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4866 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4867 STAT_ATTR(FREE_FROZEN, free_frozen);
4868 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4869 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4870 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4871 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4872 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4873 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4874 STAT_ATTR(FREE_SLAB, free_slab);
4875 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4876 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4877 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4878 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4879 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4880 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4881 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4882 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4883 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4884 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4885 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4886 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4887 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4888 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4891 static struct attribute *slab_attrs[] = {
4892 &slab_size_attr.attr,
4893 &object_size_attr.attr,
4894 &objs_per_slab_attr.attr,
4896 &min_partial_attr.attr,
4897 &cpu_partial_attr.attr,
4899 &objects_partial_attr.attr,
4901 &cpu_slabs_attr.attr,
4905 &hwcache_align_attr.attr,
4906 &reclaim_account_attr.attr,
4907 &destroy_by_rcu_attr.attr,
4909 &reserved_attr.attr,
4910 &slabs_cpu_partial_attr.attr,
4911 #ifdef CONFIG_SLUB_DEBUG
4912 &total_objects_attr.attr,
4914 &sanity_checks_attr.attr,
4916 &red_zone_attr.attr,
4918 &store_user_attr.attr,
4919 &validate_attr.attr,
4920 &alloc_calls_attr.attr,
4921 &free_calls_attr.attr,
4923 #ifdef CONFIG_ZONE_DMA
4924 &cache_dma_attr.attr,
4927 &remote_node_defrag_ratio_attr.attr,
4929 #ifdef CONFIG_SLUB_STATS
4930 &alloc_fastpath_attr.attr,
4931 &alloc_slowpath_attr.attr,
4932 &free_fastpath_attr.attr,
4933 &free_slowpath_attr.attr,
4934 &free_frozen_attr.attr,
4935 &free_add_partial_attr.attr,
4936 &free_remove_partial_attr.attr,
4937 &alloc_from_partial_attr.attr,
4938 &alloc_slab_attr.attr,
4939 &alloc_refill_attr.attr,
4940 &alloc_node_mismatch_attr.attr,
4941 &free_slab_attr.attr,
4942 &cpuslab_flush_attr.attr,
4943 &deactivate_full_attr.attr,
4944 &deactivate_empty_attr.attr,
4945 &deactivate_to_head_attr.attr,
4946 &deactivate_to_tail_attr.attr,
4947 &deactivate_remote_frees_attr.attr,
4948 &deactivate_bypass_attr.attr,
4949 &order_fallback_attr.attr,
4950 &cmpxchg_double_fail_attr.attr,
4951 &cmpxchg_double_cpu_fail_attr.attr,
4952 &cpu_partial_alloc_attr.attr,
4953 &cpu_partial_free_attr.attr,
4954 &cpu_partial_node_attr.attr,
4955 &cpu_partial_drain_attr.attr,
4957 #ifdef CONFIG_FAILSLAB
4958 &failslab_attr.attr,
4964 static struct attribute_group slab_attr_group = {
4965 .attrs = slab_attrs,
4968 static ssize_t slab_attr_show(struct kobject *kobj,
4969 struct attribute *attr,
4972 struct slab_attribute *attribute;
4973 struct kmem_cache *s;
4976 attribute = to_slab_attr(attr);
4979 if (!attribute->show)
4982 err = attribute->show(s, buf);
4987 static ssize_t slab_attr_store(struct kobject *kobj,
4988 struct attribute *attr,
4989 const char *buf, size_t len)
4991 struct slab_attribute *attribute;
4992 struct kmem_cache *s;
4995 attribute = to_slab_attr(attr);
4998 if (!attribute->store)
5001 err = attribute->store(s, buf, len);
5002 #ifdef CONFIG_MEMCG_KMEM
5003 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5006 mutex_lock(&slab_mutex);
5007 if (s->max_attr_size < len)
5008 s->max_attr_size = len;
5011 * This is a best effort propagation, so this function's return
5012 * value will be determined by the parent cache only. This is
5013 * basically because not all attributes will have a well
5014 * defined semantics for rollbacks - most of the actions will
5015 * have permanent effects.
5017 * Returning the error value of any of the children that fail
5018 * is not 100 % defined, in the sense that users seeing the
5019 * error code won't be able to know anything about the state of
5022 * Only returning the error code for the parent cache at least
5023 * has well defined semantics. The cache being written to
5024 * directly either failed or succeeded, in which case we loop
5025 * through the descendants with best-effort propagation.
5027 for_each_memcg_cache_index(i) {
5028 struct kmem_cache *c = cache_from_memcg(s, i);
5030 attribute->store(c, buf, len);
5032 mutex_unlock(&slab_mutex);
5038 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5040 #ifdef CONFIG_MEMCG_KMEM
5042 char *buffer = NULL;
5044 if (!is_root_cache(s))
5048 * This mean this cache had no attribute written. Therefore, no point
5049 * in copying default values around
5051 if (!s->max_attr_size)
5054 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5057 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5059 if (!attr || !attr->store || !attr->show)
5063 * It is really bad that we have to allocate here, so we will
5064 * do it only as a fallback. If we actually allocate, though,
5065 * we can just use the allocated buffer until the end.
5067 * Most of the slub attributes will tend to be very small in
5068 * size, but sysfs allows buffers up to a page, so they can
5069 * theoretically happen.
5073 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5076 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5077 if (WARN_ON(!buffer))
5082 attr->show(s->memcg_params->root_cache, buf);
5083 attr->store(s, buf, strlen(buf));
5087 free_page((unsigned long)buffer);
5091 static const struct sysfs_ops slab_sysfs_ops = {
5092 .show = slab_attr_show,
5093 .store = slab_attr_store,
5096 static struct kobj_type slab_ktype = {
5097 .sysfs_ops = &slab_sysfs_ops,
5100 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5102 struct kobj_type *ktype = get_ktype(kobj);
5104 if (ktype == &slab_ktype)
5109 static const struct kset_uevent_ops slab_uevent_ops = {
5110 .filter = uevent_filter,
5113 static struct kset *slab_kset;
5115 #define ID_STR_LENGTH 64
5117 /* Create a unique string id for a slab cache:
5119 * Format :[flags-]size
5121 static char *create_unique_id(struct kmem_cache *s)
5123 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5130 * First flags affecting slabcache operations. We will only
5131 * get here for aliasable slabs so we do not need to support
5132 * too many flags. The flags here must cover all flags that
5133 * are matched during merging to guarantee that the id is
5136 if (s->flags & SLAB_CACHE_DMA)
5138 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5140 if (s->flags & SLAB_DEBUG_FREE)
5142 if (!(s->flags & SLAB_NOTRACK))
5146 p += sprintf(p, "%07d", s->size);
5148 #ifdef CONFIG_MEMCG_KMEM
5149 if (!is_root_cache(s))
5150 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5153 BUG_ON(p > name + ID_STR_LENGTH - 1);
5157 static int sysfs_slab_add(struct kmem_cache *s)
5161 int unmergeable = slab_unmergeable(s);
5165 * Slabcache can never be merged so we can use the name proper.
5166 * This is typically the case for debug situations. In that
5167 * case we can catch duplicate names easily.
5169 sysfs_remove_link(&slab_kset->kobj, s->name);
5173 * Create a unique name for the slab as a target
5176 name = create_unique_id(s);
5179 s->kobj.kset = slab_kset;
5180 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5182 kobject_put(&s->kobj);
5186 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5188 kobject_del(&s->kobj);
5189 kobject_put(&s->kobj);
5192 kobject_uevent(&s->kobj, KOBJ_ADD);
5194 /* Setup first alias */
5195 sysfs_slab_alias(s, s->name);
5201 static void sysfs_slab_remove(struct kmem_cache *s)
5203 if (slab_state < FULL)
5205 * Sysfs has not been setup yet so no need to remove the
5210 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5211 kobject_del(&s->kobj);
5212 kobject_put(&s->kobj);
5216 * Need to buffer aliases during bootup until sysfs becomes
5217 * available lest we lose that information.
5219 struct saved_alias {
5220 struct kmem_cache *s;
5222 struct saved_alias *next;
5225 static struct saved_alias *alias_list;
5227 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5229 struct saved_alias *al;
5231 if (slab_state == FULL) {
5233 * If we have a leftover link then remove it.
5235 sysfs_remove_link(&slab_kset->kobj, name);
5236 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5239 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5245 al->next = alias_list;
5250 static int __init slab_sysfs_init(void)
5252 struct kmem_cache *s;
5255 mutex_lock(&slab_mutex);
5257 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5259 mutex_unlock(&slab_mutex);
5260 printk(KERN_ERR "Cannot register slab subsystem.\n");
5266 list_for_each_entry(s, &slab_caches, list) {
5267 err = sysfs_slab_add(s);
5269 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5270 " to sysfs\n", s->name);
5273 while (alias_list) {
5274 struct saved_alias *al = alias_list;
5276 alias_list = alias_list->next;
5277 err = sysfs_slab_alias(al->s, al->name);
5279 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5280 " %s to sysfs\n", al->name);
5284 mutex_unlock(&slab_mutex);
5289 __initcall(slab_sysfs_init);
5290 #endif /* CONFIG_SYSFS */
5293 * The /proc/slabinfo ABI
5295 #ifdef CONFIG_SLABINFO
5296 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5298 unsigned long nr_partials = 0;
5299 unsigned long nr_slabs = 0;
5300 unsigned long nr_objs = 0;
5301 unsigned long nr_free = 0;
5304 for_each_online_node(node) {
5305 struct kmem_cache_node *n = get_node(s, node);
5310 nr_partials += n->nr_partial;
5311 nr_slabs += atomic_long_read(&n->nr_slabs);
5312 nr_objs += atomic_long_read(&n->total_objects);
5313 nr_free += count_partial(n, count_free);
5316 sinfo->active_objs = nr_objs - nr_free;
5317 sinfo->num_objs = nr_objs;
5318 sinfo->active_slabs = nr_slabs;
5319 sinfo->num_slabs = nr_slabs;
5320 sinfo->objects_per_slab = oo_objects(s->oo);
5321 sinfo->cache_order = oo_order(s->oo);
5324 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5328 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5329 size_t count, loff_t *ppos)
5333 #endif /* CONFIG_SLABINFO */