2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
34 #include <linux/memcontrol.h>
36 #include <trace/events/kmem.h>
42 * 1. slab_mutex (Global Mutex)
44 * 3. slab_lock(page) (Only on some arches and for debugging)
48 * The role of the slab_mutex is to protect the list of all the slabs
49 * and to synchronize major metadata changes to slab cache structures.
51 * The slab_lock is only used for debugging and on arches that do not
52 * have the ability to do a cmpxchg_double. It only protects the second
53 * double word in the page struct. Meaning
54 * A. page->freelist -> List of object free in a page
55 * B. page->counters -> Counters of objects
56 * C. page->frozen -> frozen state
58 * If a slab is frozen then it is exempt from list management. It is not
59 * on any list. The processor that froze the slab is the one who can
60 * perform list operations on the page. Other processors may put objects
61 * onto the freelist but the processor that froze the slab is the only
62 * one that can retrieve the objects from the page's freelist.
64 * The list_lock protects the partial and full list on each node and
65 * the partial slab counter. If taken then no new slabs may be added or
66 * removed from the lists nor make the number of partial slabs be modified.
67 * (Note that the total number of slabs is an atomic value that may be
68 * modified without taking the list lock).
70 * The list_lock is a centralized lock and thus we avoid taking it as
71 * much as possible. As long as SLUB does not have to handle partial
72 * slabs, operations can continue without any centralized lock. F.e.
73 * allocating a long series of objects that fill up slabs does not require
75 * Interrupts are disabled during allocation and deallocation in order to
76 * make the slab allocator safe to use in the context of an irq. In addition
77 * interrupts are disabled to ensure that the processor does not change
78 * while handling per_cpu slabs, due to kernel preemption.
80 * SLUB assigns one slab for allocation to each processor.
81 * Allocations only occur from these slabs called cpu slabs.
83 * Slabs with free elements are kept on a partial list and during regular
84 * operations no list for full slabs is used. If an object in a full slab is
85 * freed then the slab will show up again on the partial lists.
86 * We track full slabs for debugging purposes though because otherwise we
87 * cannot scan all objects.
89 * Slabs are freed when they become empty. Teardown and setup is
90 * minimal so we rely on the page allocators per cpu caches for
91 * fast frees and allocs.
93 * Overloading of page flags that are otherwise used for LRU management.
95 * PageActive The slab is frozen and exempt from list processing.
96 * This means that the slab is dedicated to a purpose
97 * such as satisfying allocations for a specific
98 * processor. Objects may be freed in the slab while
99 * it is frozen but slab_free will then skip the usual
100 * list operations. It is up to the processor holding
101 * the slab to integrate the slab into the slab lists
102 * when the slab is no longer needed.
104 * One use of this flag is to mark slabs that are
105 * used for allocations. Then such a slab becomes a cpu
106 * slab. The cpu slab may be equipped with an additional
107 * freelist that allows lockless access to
108 * free objects in addition to the regular freelist
109 * that requires the slab lock.
111 * PageError Slab requires special handling due to debug
112 * options set. This moves slab handling out of
113 * the fast path and disables lockless freelists.
116 static inline int kmem_cache_debug(struct kmem_cache *s)
118 #ifdef CONFIG_SLUB_DEBUG
119 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 * Issues still to be resolved:
128 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
130 * - Variable sizing of the per node arrays
133 /* Enable to test recovery from slab corruption on boot */
134 #undef SLUB_RESILIENCY_TEST
136 /* Enable to log cmpxchg failures */
137 #undef SLUB_DEBUG_CMPXCHG
140 * Mininum number of partial slabs. These will be left on the partial
141 * lists even if they are empty. kmem_cache_shrink may reclaim them.
143 #define MIN_PARTIAL 5
146 * Maximum number of desirable partial slabs.
147 * The existence of more partial slabs makes kmem_cache_shrink
148 * sort the partial list by the number of objects in the.
150 #define MAX_PARTIAL 10
152 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
153 SLAB_POISON | SLAB_STORE_USER)
156 * Debugging flags that require metadata to be stored in the slab. These get
157 * disabled when slub_debug=O is used and a cache's min order increases with
160 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
163 * Set of flags that will prevent slab merging
165 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
166 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
169 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
170 SLAB_CACHE_DMA | SLAB_NOTRACK)
173 #define OO_MASK ((1 << OO_SHIFT) - 1)
174 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
176 /* Internal SLUB flags */
177 #define __OBJECT_POISON 0x80000000UL /* Poison object */
178 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
181 static struct notifier_block slab_notifier;
185 * Tracking user of a slab.
187 #define TRACK_ADDRS_COUNT 16
189 unsigned long addr; /* Called from address */
190 #ifdef CONFIG_STACKTRACE
191 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
193 int cpu; /* Was running on cpu */
194 int pid; /* Pid context */
195 unsigned long when; /* When did the operation occur */
198 enum track_item { TRACK_ALLOC, TRACK_FREE };
201 static int sysfs_slab_add(struct kmem_cache *);
202 static int sysfs_slab_alias(struct kmem_cache *, const char *);
203 static void sysfs_slab_remove(struct kmem_cache *);
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
209 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
213 static inline void stat(const struct kmem_cache *s, enum stat_item si)
215 #ifdef CONFIG_SLUB_STATS
216 __this_cpu_inc(s->cpu_slab->stat[si]);
220 /********************************************************************
221 * Core slab cache functions
222 *******************************************************************/
224 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
226 return s->node[node];
229 /* Verify that a pointer has an address that is valid within a slab page */
230 static inline int check_valid_pointer(struct kmem_cache *s,
231 struct page *page, const void *object)
238 base = page_address(page);
239 if (object < base || object >= base + page->objects * s->size ||
240 (object - base) % s->size) {
247 static inline void *get_freepointer(struct kmem_cache *s, void *object)
249 return *(void **)(object + s->offset);
252 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
254 prefetch(object + s->offset);
257 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
261 #ifdef CONFIG_DEBUG_PAGEALLOC
262 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
264 p = get_freepointer(s, object);
269 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
271 *(void **)(object + s->offset) = fp;
274 /* Loop over all objects in a slab */
275 #define for_each_object(__p, __s, __addr, __objects) \
276 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
279 /* Determine object index from a given position */
280 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
282 return (p - addr) / s->size;
285 static inline size_t slab_ksize(const struct kmem_cache *s)
287 #ifdef CONFIG_SLUB_DEBUG
289 * Debugging requires use of the padding between object
290 * and whatever may come after it.
292 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
293 return s->object_size;
297 * If we have the need to store the freelist pointer
298 * back there or track user information then we can
299 * only use the space before that information.
301 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
304 * Else we can use all the padding etc for the allocation
309 static inline int order_objects(int order, unsigned long size, int reserved)
311 return ((PAGE_SIZE << order) - reserved) / size;
314 static inline struct kmem_cache_order_objects oo_make(int order,
315 unsigned long size, int reserved)
317 struct kmem_cache_order_objects x = {
318 (order << OO_SHIFT) + order_objects(order, size, reserved)
324 static inline int oo_order(struct kmem_cache_order_objects x)
326 return x.x >> OO_SHIFT;
329 static inline int oo_objects(struct kmem_cache_order_objects x)
331 return x.x & OO_MASK;
335 * Per slab locking using the pagelock
337 static __always_inline void slab_lock(struct page *page)
339 bit_spin_lock(PG_locked, &page->flags);
342 static __always_inline void slab_unlock(struct page *page)
344 __bit_spin_unlock(PG_locked, &page->flags);
347 /* Interrupts must be disabled (for the fallback code to work right) */
348 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
349 void *freelist_old, unsigned long counters_old,
350 void *freelist_new, unsigned long counters_new,
353 VM_BUG_ON(!irqs_disabled());
354 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
355 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
356 if (s->flags & __CMPXCHG_DOUBLE) {
357 if (cmpxchg_double(&page->freelist, &page->counters,
358 freelist_old, counters_old,
359 freelist_new, counters_new))
365 if (page->freelist == freelist_old && page->counters == counters_old) {
366 page->freelist = freelist_new;
367 page->counters = counters_new;
375 stat(s, CMPXCHG_DOUBLE_FAIL);
377 #ifdef SLUB_DEBUG_CMPXCHG
378 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
384 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
385 void *freelist_old, unsigned long counters_old,
386 void *freelist_new, unsigned long counters_new,
389 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
390 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
391 if (s->flags & __CMPXCHG_DOUBLE) {
392 if (cmpxchg_double(&page->freelist, &page->counters,
393 freelist_old, counters_old,
394 freelist_new, counters_new))
401 local_irq_save(flags);
403 if (page->freelist == freelist_old && page->counters == counters_old) {
404 page->freelist = freelist_new;
405 page->counters = counters_new;
407 local_irq_restore(flags);
411 local_irq_restore(flags);
415 stat(s, CMPXCHG_DOUBLE_FAIL);
417 #ifdef SLUB_DEBUG_CMPXCHG
418 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
424 #ifdef CONFIG_SLUB_DEBUG
426 * Determine a map of object in use on a page.
428 * Node listlock must be held to guarantee that the page does
429 * not vanish from under us.
431 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
434 void *addr = page_address(page);
436 for (p = page->freelist; p; p = get_freepointer(s, p))
437 set_bit(slab_index(p, s, addr), map);
443 #ifdef CONFIG_SLUB_DEBUG_ON
444 static int slub_debug = DEBUG_DEFAULT_FLAGS;
446 static int slub_debug;
449 static char *slub_debug_slabs;
450 static int disable_higher_order_debug;
455 static void print_section(char *text, u8 *addr, unsigned int length)
457 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
461 static struct track *get_track(struct kmem_cache *s, void *object,
462 enum track_item alloc)
467 p = object + s->offset + sizeof(void *);
469 p = object + s->inuse;
474 static void set_track(struct kmem_cache *s, void *object,
475 enum track_item alloc, unsigned long addr)
477 struct track *p = get_track(s, object, alloc);
480 #ifdef CONFIG_STACKTRACE
481 struct stack_trace trace;
484 trace.nr_entries = 0;
485 trace.max_entries = TRACK_ADDRS_COUNT;
486 trace.entries = p->addrs;
488 save_stack_trace(&trace);
490 /* See rant in lockdep.c */
491 if (trace.nr_entries != 0 &&
492 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
495 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
499 p->cpu = smp_processor_id();
500 p->pid = current->pid;
503 memset(p, 0, sizeof(struct track));
506 static void init_tracking(struct kmem_cache *s, void *object)
508 if (!(s->flags & SLAB_STORE_USER))
511 set_track(s, object, TRACK_FREE, 0UL);
512 set_track(s, object, TRACK_ALLOC, 0UL);
515 static void print_track(const char *s, struct track *t)
520 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
521 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
522 #ifdef CONFIG_STACKTRACE
525 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
527 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
534 static void print_tracking(struct kmem_cache *s, void *object)
536 if (!(s->flags & SLAB_STORE_USER))
539 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
540 print_track("Freed", get_track(s, object, TRACK_FREE));
543 static void print_page_info(struct page *page)
545 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
546 page, page->objects, page->inuse, page->freelist, page->flags);
550 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
556 vsnprintf(buf, sizeof(buf), fmt, args);
558 printk(KERN_ERR "========================================"
559 "=====================================\n");
560 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
561 printk(KERN_ERR "----------------------------------------"
562 "-------------------------------------\n\n");
564 add_taint(TAINT_BAD_PAGE);
567 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
573 vsnprintf(buf, sizeof(buf), fmt, args);
575 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
578 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
580 unsigned int off; /* Offset of last byte */
581 u8 *addr = page_address(page);
583 print_tracking(s, p);
585 print_page_info(page);
587 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
588 p, p - addr, get_freepointer(s, p));
591 print_section("Bytes b4 ", p - 16, 16);
593 print_section("Object ", p, min_t(unsigned long, s->object_size,
595 if (s->flags & SLAB_RED_ZONE)
596 print_section("Redzone ", p + s->object_size,
597 s->inuse - s->object_size);
600 off = s->offset + sizeof(void *);
604 if (s->flags & SLAB_STORE_USER)
605 off += 2 * sizeof(struct track);
608 /* Beginning of the filler is the free pointer */
609 print_section("Padding ", p + off, s->size - off);
614 static void object_err(struct kmem_cache *s, struct page *page,
615 u8 *object, char *reason)
617 slab_bug(s, "%s", reason);
618 print_trailer(s, page, object);
621 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
627 vsnprintf(buf, sizeof(buf), fmt, args);
629 slab_bug(s, "%s", buf);
630 print_page_info(page);
634 static void init_object(struct kmem_cache *s, void *object, u8 val)
638 if (s->flags & __OBJECT_POISON) {
639 memset(p, POISON_FREE, s->object_size - 1);
640 p[s->object_size - 1] = POISON_END;
643 if (s->flags & SLAB_RED_ZONE)
644 memset(p + s->object_size, val, s->inuse - s->object_size);
647 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
648 void *from, void *to)
650 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
651 memset(from, data, to - from);
654 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
655 u8 *object, char *what,
656 u8 *start, unsigned int value, unsigned int bytes)
661 fault = memchr_inv(start, value, bytes);
666 while (end > fault && end[-1] == value)
669 slab_bug(s, "%s overwritten", what);
670 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
671 fault, end - 1, fault[0], value);
672 print_trailer(s, page, object);
674 restore_bytes(s, what, value, fault, end);
682 * Bytes of the object to be managed.
683 * If the freepointer may overlay the object then the free
684 * pointer is the first word of the object.
686 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
689 * object + s->object_size
690 * Padding to reach word boundary. This is also used for Redzoning.
691 * Padding is extended by another word if Redzoning is enabled and
692 * object_size == inuse.
694 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
695 * 0xcc (RED_ACTIVE) for objects in use.
698 * Meta data starts here.
700 * A. Free pointer (if we cannot overwrite object on free)
701 * B. Tracking data for SLAB_STORE_USER
702 * C. Padding to reach required alignment boundary or at mininum
703 * one word if debugging is on to be able to detect writes
704 * before the word boundary.
706 * Padding is done using 0x5a (POISON_INUSE)
709 * Nothing is used beyond s->size.
711 * If slabcaches are merged then the object_size and inuse boundaries are mostly
712 * ignored. And therefore no slab options that rely on these boundaries
713 * may be used with merged slabcaches.
716 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
718 unsigned long off = s->inuse; /* The end of info */
721 /* Freepointer is placed after the object. */
722 off += sizeof(void *);
724 if (s->flags & SLAB_STORE_USER)
725 /* We also have user information there */
726 off += 2 * sizeof(struct track);
731 return check_bytes_and_report(s, page, p, "Object padding",
732 p + off, POISON_INUSE, s->size - off);
735 /* Check the pad bytes at the end of a slab page */
736 static int slab_pad_check(struct kmem_cache *s, struct page *page)
744 if (!(s->flags & SLAB_POISON))
747 start = page_address(page);
748 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
749 end = start + length;
750 remainder = length % s->size;
754 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
757 while (end > fault && end[-1] == POISON_INUSE)
760 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
761 print_section("Padding ", end - remainder, remainder);
763 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
767 static int check_object(struct kmem_cache *s, struct page *page,
768 void *object, u8 val)
771 u8 *endobject = object + s->object_size;
773 if (s->flags & SLAB_RED_ZONE) {
774 if (!check_bytes_and_report(s, page, object, "Redzone",
775 endobject, val, s->inuse - s->object_size))
778 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
779 check_bytes_and_report(s, page, p, "Alignment padding",
780 endobject, POISON_INUSE, s->inuse - s->object_size);
784 if (s->flags & SLAB_POISON) {
785 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
786 (!check_bytes_and_report(s, page, p, "Poison", p,
787 POISON_FREE, s->object_size - 1) ||
788 !check_bytes_and_report(s, page, p, "Poison",
789 p + s->object_size - 1, POISON_END, 1)))
792 * check_pad_bytes cleans up on its own.
794 check_pad_bytes(s, page, p);
797 if (!s->offset && val == SLUB_RED_ACTIVE)
799 * Object and freepointer overlap. Cannot check
800 * freepointer while object is allocated.
804 /* Check free pointer validity */
805 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
806 object_err(s, page, p, "Freepointer corrupt");
808 * No choice but to zap it and thus lose the remainder
809 * of the free objects in this slab. May cause
810 * another error because the object count is now wrong.
812 set_freepointer(s, p, NULL);
818 static int check_slab(struct kmem_cache *s, struct page *page)
822 VM_BUG_ON(!irqs_disabled());
824 if (!PageSlab(page)) {
825 slab_err(s, page, "Not a valid slab page");
829 maxobj = order_objects(compound_order(page), s->size, s->reserved);
830 if (page->objects > maxobj) {
831 slab_err(s, page, "objects %u > max %u",
832 s->name, page->objects, maxobj);
835 if (page->inuse > page->objects) {
836 slab_err(s, page, "inuse %u > max %u",
837 s->name, page->inuse, page->objects);
840 /* Slab_pad_check fixes things up after itself */
841 slab_pad_check(s, page);
846 * Determine if a certain object on a page is on the freelist. Must hold the
847 * slab lock to guarantee that the chains are in a consistent state.
849 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
854 unsigned long max_objects;
857 while (fp && nr <= page->objects) {
860 if (!check_valid_pointer(s, page, fp)) {
862 object_err(s, page, object,
863 "Freechain corrupt");
864 set_freepointer(s, object, NULL);
867 slab_err(s, page, "Freepointer corrupt");
868 page->freelist = NULL;
869 page->inuse = page->objects;
870 slab_fix(s, "Freelist cleared");
876 fp = get_freepointer(s, object);
880 max_objects = order_objects(compound_order(page), s->size, s->reserved);
881 if (max_objects > MAX_OBJS_PER_PAGE)
882 max_objects = MAX_OBJS_PER_PAGE;
884 if (page->objects != max_objects) {
885 slab_err(s, page, "Wrong number of objects. Found %d but "
886 "should be %d", page->objects, max_objects);
887 page->objects = max_objects;
888 slab_fix(s, "Number of objects adjusted.");
890 if (page->inuse != page->objects - nr) {
891 slab_err(s, page, "Wrong object count. Counter is %d but "
892 "counted were %d", page->inuse, page->objects - nr);
893 page->inuse = page->objects - nr;
894 slab_fix(s, "Object count adjusted.");
896 return search == NULL;
899 static void trace(struct kmem_cache *s, struct page *page, void *object,
902 if (s->flags & SLAB_TRACE) {
903 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
905 alloc ? "alloc" : "free",
910 print_section("Object ", (void *)object, s->object_size);
917 * Hooks for other subsystems that check memory allocations. In a typical
918 * production configuration these hooks all should produce no code at all.
920 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
922 flags &= gfp_allowed_mask;
923 lockdep_trace_alloc(flags);
924 might_sleep_if(flags & __GFP_WAIT);
926 return should_failslab(s->object_size, flags, s->flags);
929 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
931 flags &= gfp_allowed_mask;
932 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
933 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
936 static inline void slab_free_hook(struct kmem_cache *s, void *x)
938 kmemleak_free_recursive(x, s->flags);
941 * Trouble is that we may no longer disable interupts in the fast path
942 * So in order to make the debug calls that expect irqs to be
943 * disabled we need to disable interrupts temporarily.
945 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
949 local_irq_save(flags);
950 kmemcheck_slab_free(s, x, s->object_size);
951 debug_check_no_locks_freed(x, s->object_size);
952 local_irq_restore(flags);
955 if (!(s->flags & SLAB_DEBUG_OBJECTS))
956 debug_check_no_obj_freed(x, s->object_size);
960 * Tracking of fully allocated slabs for debugging purposes.
962 * list_lock must be held.
964 static void add_full(struct kmem_cache *s,
965 struct kmem_cache_node *n, struct page *page)
967 if (!(s->flags & SLAB_STORE_USER))
970 list_add(&page->lru, &n->full);
974 * list_lock must be held.
976 static void remove_full(struct kmem_cache *s, struct page *page)
978 if (!(s->flags & SLAB_STORE_USER))
981 list_del(&page->lru);
984 /* Tracking of the number of slabs for debugging purposes */
985 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
987 struct kmem_cache_node *n = get_node(s, node);
989 return atomic_long_read(&n->nr_slabs);
992 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
994 return atomic_long_read(&n->nr_slabs);
997 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
999 struct kmem_cache_node *n = get_node(s, node);
1002 * May be called early in order to allocate a slab for the
1003 * kmem_cache_node structure. Solve the chicken-egg
1004 * dilemma by deferring the increment of the count during
1005 * bootstrap (see early_kmem_cache_node_alloc).
1008 atomic_long_inc(&n->nr_slabs);
1009 atomic_long_add(objects, &n->total_objects);
1012 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1014 struct kmem_cache_node *n = get_node(s, node);
1016 atomic_long_dec(&n->nr_slabs);
1017 atomic_long_sub(objects, &n->total_objects);
1020 /* Object debug checks for alloc/free paths */
1021 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1024 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1027 init_object(s, object, SLUB_RED_INACTIVE);
1028 init_tracking(s, object);
1031 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1032 void *object, unsigned long addr)
1034 if (!check_slab(s, page))
1037 if (!check_valid_pointer(s, page, object)) {
1038 object_err(s, page, object, "Freelist Pointer check fails");
1042 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1045 /* Success perform special debug activities for allocs */
1046 if (s->flags & SLAB_STORE_USER)
1047 set_track(s, object, TRACK_ALLOC, addr);
1048 trace(s, page, object, 1);
1049 init_object(s, object, SLUB_RED_ACTIVE);
1053 if (PageSlab(page)) {
1055 * If this is a slab page then lets do the best we can
1056 * to avoid issues in the future. Marking all objects
1057 * as used avoids touching the remaining objects.
1059 slab_fix(s, "Marking all objects used");
1060 page->inuse = page->objects;
1061 page->freelist = NULL;
1066 static noinline struct kmem_cache_node *free_debug_processing(
1067 struct kmem_cache *s, struct page *page, void *object,
1068 unsigned long addr, unsigned long *flags)
1070 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1072 spin_lock_irqsave(&n->list_lock, *flags);
1075 if (!check_slab(s, page))
1078 if (!check_valid_pointer(s, page, object)) {
1079 slab_err(s, page, "Invalid object pointer 0x%p", object);
1083 if (on_freelist(s, page, object)) {
1084 object_err(s, page, object, "Object already free");
1088 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1091 if (unlikely(s != page->slab_cache)) {
1092 if (!PageSlab(page)) {
1093 slab_err(s, page, "Attempt to free object(0x%p) "
1094 "outside of slab", object);
1095 } else if (!page->slab_cache) {
1097 "SLUB <none>: no slab for object 0x%p.\n",
1101 object_err(s, page, object,
1102 "page slab pointer corrupt.");
1106 if (s->flags & SLAB_STORE_USER)
1107 set_track(s, object, TRACK_FREE, addr);
1108 trace(s, page, object, 0);
1109 init_object(s, object, SLUB_RED_INACTIVE);
1113 * Keep node_lock to preserve integrity
1114 * until the object is actually freed
1120 spin_unlock_irqrestore(&n->list_lock, *flags);
1121 slab_fix(s, "Object at 0x%p not freed", object);
1125 static int __init setup_slub_debug(char *str)
1127 slub_debug = DEBUG_DEFAULT_FLAGS;
1128 if (*str++ != '=' || !*str)
1130 * No options specified. Switch on full debugging.
1136 * No options but restriction on slabs. This means full
1137 * debugging for slabs matching a pattern.
1141 if (tolower(*str) == 'o') {
1143 * Avoid enabling debugging on caches if its minimum order
1144 * would increase as a result.
1146 disable_higher_order_debug = 1;
1153 * Switch off all debugging measures.
1158 * Determine which debug features should be switched on
1160 for (; *str && *str != ','; str++) {
1161 switch (tolower(*str)) {
1163 slub_debug |= SLAB_DEBUG_FREE;
1166 slub_debug |= SLAB_RED_ZONE;
1169 slub_debug |= SLAB_POISON;
1172 slub_debug |= SLAB_STORE_USER;
1175 slub_debug |= SLAB_TRACE;
1178 slub_debug |= SLAB_FAILSLAB;
1181 printk(KERN_ERR "slub_debug option '%c' "
1182 "unknown. skipped\n", *str);
1188 slub_debug_slabs = str + 1;
1193 __setup("slub_debug", setup_slub_debug);
1195 static unsigned long kmem_cache_flags(unsigned long object_size,
1196 unsigned long flags, const char *name,
1197 void (*ctor)(void *))
1200 * Enable debugging if selected on the kernel commandline.
1202 if (slub_debug && (!slub_debug_slabs ||
1203 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1204 flags |= slub_debug;
1209 static inline void setup_object_debug(struct kmem_cache *s,
1210 struct page *page, void *object) {}
1212 static inline int alloc_debug_processing(struct kmem_cache *s,
1213 struct page *page, void *object, unsigned long addr) { return 0; }
1215 static inline struct kmem_cache_node *free_debug_processing(
1216 struct kmem_cache *s, struct page *page, void *object,
1217 unsigned long addr, unsigned long *flags) { return NULL; }
1219 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1221 static inline int check_object(struct kmem_cache *s, struct page *page,
1222 void *object, u8 val) { return 1; }
1223 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1224 struct page *page) {}
1225 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1226 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1227 unsigned long flags, const char *name,
1228 void (*ctor)(void *))
1232 #define slub_debug 0
1234 #define disable_higher_order_debug 0
1236 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1238 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1240 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1242 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1245 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1248 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1251 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1253 #endif /* CONFIG_SLUB_DEBUG */
1256 * Slab allocation and freeing
1258 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1259 struct kmem_cache_order_objects oo)
1261 int order = oo_order(oo);
1263 flags |= __GFP_NOTRACK;
1265 if (node == NUMA_NO_NODE)
1266 return alloc_pages(flags, order);
1268 return alloc_pages_exact_node(node, flags, order);
1271 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1274 struct kmem_cache_order_objects oo = s->oo;
1277 flags &= gfp_allowed_mask;
1279 if (flags & __GFP_WAIT)
1282 flags |= s->allocflags;
1285 * Let the initial higher-order allocation fail under memory pressure
1286 * so we fall-back to the minimum order allocation.
1288 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1290 page = alloc_slab_page(alloc_gfp, node, oo);
1291 if (unlikely(!page)) {
1294 * Allocation may have failed due to fragmentation.
1295 * Try a lower order alloc if possible
1297 page = alloc_slab_page(flags, node, oo);
1300 stat(s, ORDER_FALLBACK);
1303 if (kmemcheck_enabled && page
1304 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1305 int pages = 1 << oo_order(oo);
1307 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1310 * Objects from caches that have a constructor don't get
1311 * cleared when they're allocated, so we need to do it here.
1314 kmemcheck_mark_uninitialized_pages(page, pages);
1316 kmemcheck_mark_unallocated_pages(page, pages);
1319 if (flags & __GFP_WAIT)
1320 local_irq_disable();
1324 page->objects = oo_objects(oo);
1325 mod_zone_page_state(page_zone(page),
1326 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1327 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1333 static void setup_object(struct kmem_cache *s, struct page *page,
1336 setup_object_debug(s, page, object);
1337 if (unlikely(s->ctor))
1341 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1348 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1350 page = allocate_slab(s,
1351 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1355 inc_slabs_node(s, page_to_nid(page), page->objects);
1356 page->slab_cache = s;
1357 __SetPageSlab(page);
1358 if (page->pfmemalloc)
1359 SetPageSlabPfmemalloc(page);
1361 start = page_address(page);
1363 if (unlikely(s->flags & SLAB_POISON))
1364 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1367 for_each_object(p, s, start, page->objects) {
1368 setup_object(s, page, last);
1369 set_freepointer(s, last, p);
1372 setup_object(s, page, last);
1373 set_freepointer(s, last, NULL);
1375 page->freelist = start;
1376 page->inuse = page->objects;
1382 static void __free_slab(struct kmem_cache *s, struct page *page)
1384 int order = compound_order(page);
1385 int pages = 1 << order;
1387 if (kmem_cache_debug(s)) {
1390 slab_pad_check(s, page);
1391 for_each_object(p, s, page_address(page),
1393 check_object(s, page, p, SLUB_RED_INACTIVE);
1396 kmemcheck_free_shadow(page, compound_order(page));
1398 mod_zone_page_state(page_zone(page),
1399 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1400 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1403 __ClearPageSlabPfmemalloc(page);
1404 __ClearPageSlab(page);
1405 reset_page_mapcount(page);
1406 if (current->reclaim_state)
1407 current->reclaim_state->reclaimed_slab += pages;
1408 __free_pages(page, order);
1411 #define need_reserve_slab_rcu \
1412 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1414 static void rcu_free_slab(struct rcu_head *h)
1418 if (need_reserve_slab_rcu)
1419 page = virt_to_head_page(h);
1421 page = container_of((struct list_head *)h, struct page, lru);
1423 __free_slab(page->slab_cache, page);
1426 static void free_slab(struct kmem_cache *s, struct page *page)
1428 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1429 struct rcu_head *head;
1431 if (need_reserve_slab_rcu) {
1432 int order = compound_order(page);
1433 int offset = (PAGE_SIZE << order) - s->reserved;
1435 VM_BUG_ON(s->reserved != sizeof(*head));
1436 head = page_address(page) + offset;
1439 * RCU free overloads the RCU head over the LRU
1441 head = (void *)&page->lru;
1444 call_rcu(head, rcu_free_slab);
1446 __free_slab(s, page);
1449 static void discard_slab(struct kmem_cache *s, struct page *page)
1451 dec_slabs_node(s, page_to_nid(page), page->objects);
1456 * Management of partially allocated slabs.
1458 * list_lock must be held.
1460 static inline void add_partial(struct kmem_cache_node *n,
1461 struct page *page, int tail)
1464 if (tail == DEACTIVATE_TO_TAIL)
1465 list_add_tail(&page->lru, &n->partial);
1467 list_add(&page->lru, &n->partial);
1471 * list_lock must be held.
1473 static inline void remove_partial(struct kmem_cache_node *n,
1476 list_del(&page->lru);
1481 * Remove slab from the partial list, freeze it and
1482 * return the pointer to the freelist.
1484 * Returns a list of objects or NULL if it fails.
1486 * Must hold list_lock since we modify the partial list.
1488 static inline void *acquire_slab(struct kmem_cache *s,
1489 struct kmem_cache_node *n, struct page *page,
1493 unsigned long counters;
1497 * Zap the freelist and set the frozen bit.
1498 * The old freelist is the list of objects for the
1499 * per cpu allocation list.
1501 freelist = page->freelist;
1502 counters = page->counters;
1503 new.counters = counters;
1505 new.inuse = page->objects;
1506 new.freelist = NULL;
1508 new.freelist = freelist;
1511 VM_BUG_ON(new.frozen);
1514 if (!__cmpxchg_double_slab(s, page,
1516 new.freelist, new.counters,
1520 remove_partial(n, page);
1525 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1526 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1529 * Try to allocate a partial slab from a specific node.
1531 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1532 struct kmem_cache_cpu *c, gfp_t flags)
1534 struct page *page, *page2;
1535 void *object = NULL;
1538 * Racy check. If we mistakenly see no partial slabs then we
1539 * just allocate an empty slab. If we mistakenly try to get a
1540 * partial slab and there is none available then get_partials()
1543 if (!n || !n->nr_partial)
1546 spin_lock(&n->list_lock);
1547 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1551 if (!pfmemalloc_match(page, flags))
1554 t = acquire_slab(s, n, page, object == NULL);
1560 stat(s, ALLOC_FROM_PARTIAL);
1562 available = page->objects - page->inuse;
1564 available = put_cpu_partial(s, page, 0);
1565 stat(s, CPU_PARTIAL_NODE);
1567 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1571 spin_unlock(&n->list_lock);
1576 * Get a page from somewhere. Search in increasing NUMA distances.
1578 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1579 struct kmem_cache_cpu *c)
1582 struct zonelist *zonelist;
1585 enum zone_type high_zoneidx = gfp_zone(flags);
1587 unsigned int cpuset_mems_cookie;
1590 * The defrag ratio allows a configuration of the tradeoffs between
1591 * inter node defragmentation and node local allocations. A lower
1592 * defrag_ratio increases the tendency to do local allocations
1593 * instead of attempting to obtain partial slabs from other nodes.
1595 * If the defrag_ratio is set to 0 then kmalloc() always
1596 * returns node local objects. If the ratio is higher then kmalloc()
1597 * may return off node objects because partial slabs are obtained
1598 * from other nodes and filled up.
1600 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1601 * defrag_ratio = 1000) then every (well almost) allocation will
1602 * first attempt to defrag slab caches on other nodes. This means
1603 * scanning over all nodes to look for partial slabs which may be
1604 * expensive if we do it every time we are trying to find a slab
1605 * with available objects.
1607 if (!s->remote_node_defrag_ratio ||
1608 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1612 cpuset_mems_cookie = get_mems_allowed();
1613 zonelist = node_zonelist(slab_node(), flags);
1614 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1615 struct kmem_cache_node *n;
1617 n = get_node(s, zone_to_nid(zone));
1619 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1620 n->nr_partial > s->min_partial) {
1621 object = get_partial_node(s, n, c, flags);
1624 * Return the object even if
1625 * put_mems_allowed indicated that
1626 * the cpuset mems_allowed was
1627 * updated in parallel. It's a
1628 * harmless race between the alloc
1629 * and the cpuset update.
1631 put_mems_allowed(cpuset_mems_cookie);
1636 } while (!put_mems_allowed(cpuset_mems_cookie));
1642 * Get a partial page, lock it and return it.
1644 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1645 struct kmem_cache_cpu *c)
1648 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1650 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1651 if (object || node != NUMA_NO_NODE)
1654 return get_any_partial(s, flags, c);
1657 #ifdef CONFIG_PREEMPT
1659 * Calculate the next globally unique transaction for disambiguiation
1660 * during cmpxchg. The transactions start with the cpu number and are then
1661 * incremented by CONFIG_NR_CPUS.
1663 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1666 * No preemption supported therefore also no need to check for
1672 static inline unsigned long next_tid(unsigned long tid)
1674 return tid + TID_STEP;
1677 static inline unsigned int tid_to_cpu(unsigned long tid)
1679 return tid % TID_STEP;
1682 static inline unsigned long tid_to_event(unsigned long tid)
1684 return tid / TID_STEP;
1687 static inline unsigned int init_tid(int cpu)
1692 static inline void note_cmpxchg_failure(const char *n,
1693 const struct kmem_cache *s, unsigned long tid)
1695 #ifdef SLUB_DEBUG_CMPXCHG
1696 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1698 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1700 #ifdef CONFIG_PREEMPT
1701 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1702 printk("due to cpu change %d -> %d\n",
1703 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1706 if (tid_to_event(tid) != tid_to_event(actual_tid))
1707 printk("due to cpu running other code. Event %ld->%ld\n",
1708 tid_to_event(tid), tid_to_event(actual_tid));
1710 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1711 actual_tid, tid, next_tid(tid));
1713 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1716 static void init_kmem_cache_cpus(struct kmem_cache *s)
1720 for_each_possible_cpu(cpu)
1721 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1725 * Remove the cpu slab
1727 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1729 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1730 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1732 enum slab_modes l = M_NONE, m = M_NONE;
1734 int tail = DEACTIVATE_TO_HEAD;
1738 if (page->freelist) {
1739 stat(s, DEACTIVATE_REMOTE_FREES);
1740 tail = DEACTIVATE_TO_TAIL;
1744 * Stage one: Free all available per cpu objects back
1745 * to the page freelist while it is still frozen. Leave the
1748 * There is no need to take the list->lock because the page
1751 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1753 unsigned long counters;
1756 prior = page->freelist;
1757 counters = page->counters;
1758 set_freepointer(s, freelist, prior);
1759 new.counters = counters;
1761 VM_BUG_ON(!new.frozen);
1763 } while (!__cmpxchg_double_slab(s, page,
1765 freelist, new.counters,
1766 "drain percpu freelist"));
1768 freelist = nextfree;
1772 * Stage two: Ensure that the page is unfrozen while the
1773 * list presence reflects the actual number of objects
1776 * We setup the list membership and then perform a cmpxchg
1777 * with the count. If there is a mismatch then the page
1778 * is not unfrozen but the page is on the wrong list.
1780 * Then we restart the process which may have to remove
1781 * the page from the list that we just put it on again
1782 * because the number of objects in the slab may have
1787 old.freelist = page->freelist;
1788 old.counters = page->counters;
1789 VM_BUG_ON(!old.frozen);
1791 /* Determine target state of the slab */
1792 new.counters = old.counters;
1795 set_freepointer(s, freelist, old.freelist);
1796 new.freelist = freelist;
1798 new.freelist = old.freelist;
1802 if (!new.inuse && n->nr_partial > s->min_partial)
1804 else if (new.freelist) {
1809 * Taking the spinlock removes the possiblity
1810 * that acquire_slab() will see a slab page that
1813 spin_lock(&n->list_lock);
1817 if (kmem_cache_debug(s) && !lock) {
1820 * This also ensures that the scanning of full
1821 * slabs from diagnostic functions will not see
1824 spin_lock(&n->list_lock);
1832 remove_partial(n, page);
1834 else if (l == M_FULL)
1836 remove_full(s, page);
1838 if (m == M_PARTIAL) {
1840 add_partial(n, page, tail);
1843 } else if (m == M_FULL) {
1845 stat(s, DEACTIVATE_FULL);
1846 add_full(s, n, page);
1852 if (!__cmpxchg_double_slab(s, page,
1853 old.freelist, old.counters,
1854 new.freelist, new.counters,
1859 spin_unlock(&n->list_lock);
1862 stat(s, DEACTIVATE_EMPTY);
1863 discard_slab(s, page);
1869 * Unfreeze all the cpu partial slabs.
1871 * This function must be called with interrupts disabled
1872 * for the cpu using c (or some other guarantee must be there
1873 * to guarantee no concurrent accesses).
1875 static void unfreeze_partials(struct kmem_cache *s,
1876 struct kmem_cache_cpu *c)
1878 struct kmem_cache_node *n = NULL, *n2 = NULL;
1879 struct page *page, *discard_page = NULL;
1881 while ((page = c->partial)) {
1885 c->partial = page->next;
1887 n2 = get_node(s, page_to_nid(page));
1890 spin_unlock(&n->list_lock);
1893 spin_lock(&n->list_lock);
1898 old.freelist = page->freelist;
1899 old.counters = page->counters;
1900 VM_BUG_ON(!old.frozen);
1902 new.counters = old.counters;
1903 new.freelist = old.freelist;
1907 } while (!__cmpxchg_double_slab(s, page,
1908 old.freelist, old.counters,
1909 new.freelist, new.counters,
1910 "unfreezing slab"));
1912 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1913 page->next = discard_page;
1914 discard_page = page;
1916 add_partial(n, page, DEACTIVATE_TO_TAIL);
1917 stat(s, FREE_ADD_PARTIAL);
1922 spin_unlock(&n->list_lock);
1924 while (discard_page) {
1925 page = discard_page;
1926 discard_page = discard_page->next;
1928 stat(s, DEACTIVATE_EMPTY);
1929 discard_slab(s, page);
1935 * Put a page that was just frozen (in __slab_free) into a partial page
1936 * slot if available. This is done without interrupts disabled and without
1937 * preemption disabled. The cmpxchg is racy and may put the partial page
1938 * onto a random cpus partial slot.
1940 * If we did not find a slot then simply move all the partials to the
1941 * per node partial list.
1943 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1945 struct page *oldpage;
1952 oldpage = this_cpu_read(s->cpu_slab->partial);
1955 pobjects = oldpage->pobjects;
1956 pages = oldpage->pages;
1957 if (drain && pobjects > s->cpu_partial) {
1958 unsigned long flags;
1960 * partial array is full. Move the existing
1961 * set to the per node partial list.
1963 local_irq_save(flags);
1964 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1965 local_irq_restore(flags);
1969 stat(s, CPU_PARTIAL_DRAIN);
1974 pobjects += page->objects - page->inuse;
1976 page->pages = pages;
1977 page->pobjects = pobjects;
1978 page->next = oldpage;
1980 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1984 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1986 stat(s, CPUSLAB_FLUSH);
1987 deactivate_slab(s, c->page, c->freelist);
1989 c->tid = next_tid(c->tid);
1997 * Called from IPI handler with interrupts disabled.
1999 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2001 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2007 unfreeze_partials(s, c);
2011 static void flush_cpu_slab(void *d)
2013 struct kmem_cache *s = d;
2015 __flush_cpu_slab(s, smp_processor_id());
2018 static bool has_cpu_slab(int cpu, void *info)
2020 struct kmem_cache *s = info;
2021 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2023 return c->page || c->partial;
2026 static void flush_all(struct kmem_cache *s)
2028 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2032 * Check if the objects in a per cpu structure fit numa
2033 * locality expectations.
2035 static inline int node_match(struct page *page, int node)
2038 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2044 static int count_free(struct page *page)
2046 return page->objects - page->inuse;
2049 static unsigned long count_partial(struct kmem_cache_node *n,
2050 int (*get_count)(struct page *))
2052 unsigned long flags;
2053 unsigned long x = 0;
2056 spin_lock_irqsave(&n->list_lock, flags);
2057 list_for_each_entry(page, &n->partial, lru)
2058 x += get_count(page);
2059 spin_unlock_irqrestore(&n->list_lock, flags);
2063 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2065 #ifdef CONFIG_SLUB_DEBUG
2066 return atomic_long_read(&n->total_objects);
2072 static noinline void
2073 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2078 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2080 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2081 "default order: %d, min order: %d\n", s->name, s->object_size,
2082 s->size, oo_order(s->oo), oo_order(s->min));
2084 if (oo_order(s->min) > get_order(s->object_size))
2085 printk(KERN_WARNING " %s debugging increased min order, use "
2086 "slub_debug=O to disable.\n", s->name);
2088 for_each_online_node(node) {
2089 struct kmem_cache_node *n = get_node(s, node);
2090 unsigned long nr_slabs;
2091 unsigned long nr_objs;
2092 unsigned long nr_free;
2097 nr_free = count_partial(n, count_free);
2098 nr_slabs = node_nr_slabs(n);
2099 nr_objs = node_nr_objs(n);
2102 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2103 node, nr_slabs, nr_objs, nr_free);
2107 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2108 int node, struct kmem_cache_cpu **pc)
2111 struct kmem_cache_cpu *c = *pc;
2114 freelist = get_partial(s, flags, node, c);
2119 page = new_slab(s, flags, node);
2121 c = __this_cpu_ptr(s->cpu_slab);
2126 * No other reference to the page yet so we can
2127 * muck around with it freely without cmpxchg
2129 freelist = page->freelist;
2130 page->freelist = NULL;
2132 stat(s, ALLOC_SLAB);
2141 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2143 if (unlikely(PageSlabPfmemalloc(page)))
2144 return gfp_pfmemalloc_allowed(gfpflags);
2150 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2151 * or deactivate the page.
2153 * The page is still frozen if the return value is not NULL.
2155 * If this function returns NULL then the page has been unfrozen.
2157 * This function must be called with interrupt disabled.
2159 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2162 unsigned long counters;
2166 freelist = page->freelist;
2167 counters = page->counters;
2169 new.counters = counters;
2170 VM_BUG_ON(!new.frozen);
2172 new.inuse = page->objects;
2173 new.frozen = freelist != NULL;
2175 } while (!__cmpxchg_double_slab(s, page,
2184 * Slow path. The lockless freelist is empty or we need to perform
2187 * Processing is still very fast if new objects have been freed to the
2188 * regular freelist. In that case we simply take over the regular freelist
2189 * as the lockless freelist and zap the regular freelist.
2191 * If that is not working then we fall back to the partial lists. We take the
2192 * first element of the freelist as the object to allocate now and move the
2193 * rest of the freelist to the lockless freelist.
2195 * And if we were unable to get a new slab from the partial slab lists then
2196 * we need to allocate a new slab. This is the slowest path since it involves
2197 * a call to the page allocator and the setup of a new slab.
2199 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2200 unsigned long addr, struct kmem_cache_cpu *c)
2204 unsigned long flags;
2206 local_irq_save(flags);
2207 #ifdef CONFIG_PREEMPT
2209 * We may have been preempted and rescheduled on a different
2210 * cpu before disabling interrupts. Need to reload cpu area
2213 c = this_cpu_ptr(s->cpu_slab);
2221 if (unlikely(!node_match(page, node))) {
2222 stat(s, ALLOC_NODE_MISMATCH);
2223 deactivate_slab(s, page, c->freelist);
2230 * By rights, we should be searching for a slab page that was
2231 * PFMEMALLOC but right now, we are losing the pfmemalloc
2232 * information when the page leaves the per-cpu allocator
2234 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2235 deactivate_slab(s, page, c->freelist);
2241 /* must check again c->freelist in case of cpu migration or IRQ */
2242 freelist = c->freelist;
2246 stat(s, ALLOC_SLOWPATH);
2248 freelist = get_freelist(s, page);
2252 stat(s, DEACTIVATE_BYPASS);
2256 stat(s, ALLOC_REFILL);
2260 * freelist is pointing to the list of objects to be used.
2261 * page is pointing to the page from which the objects are obtained.
2262 * That page must be frozen for per cpu allocations to work.
2264 VM_BUG_ON(!c->page->frozen);
2265 c->freelist = get_freepointer(s, freelist);
2266 c->tid = next_tid(c->tid);
2267 local_irq_restore(flags);
2273 page = c->page = c->partial;
2274 c->partial = page->next;
2275 stat(s, CPU_PARTIAL_ALLOC);
2280 freelist = new_slab_objects(s, gfpflags, node, &c);
2282 if (unlikely(!freelist)) {
2283 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2284 slab_out_of_memory(s, gfpflags, node);
2286 local_irq_restore(flags);
2291 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2294 /* Only entered in the debug case */
2295 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2296 goto new_slab; /* Slab failed checks. Next slab needed */
2298 deactivate_slab(s, page, get_freepointer(s, freelist));
2301 local_irq_restore(flags);
2306 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2307 * have the fastpath folded into their functions. So no function call
2308 * overhead for requests that can be satisfied on the fastpath.
2310 * The fastpath works by first checking if the lockless freelist can be used.
2311 * If not then __slab_alloc is called for slow processing.
2313 * Otherwise we can simply pick the next object from the lockless free list.
2315 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2316 gfp_t gfpflags, int node, unsigned long addr)
2319 struct kmem_cache_cpu *c;
2323 if (slab_pre_alloc_hook(s, gfpflags))
2329 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2330 * enabled. We may switch back and forth between cpus while
2331 * reading from one cpu area. That does not matter as long
2332 * as we end up on the original cpu again when doing the cmpxchg.
2334 c = __this_cpu_ptr(s->cpu_slab);
2337 * The transaction ids are globally unique per cpu and per operation on
2338 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2339 * occurs on the right processor and that there was no operation on the
2340 * linked list in between.
2345 object = c->freelist;
2347 if (unlikely(!object || !node_match(page, node)))
2348 object = __slab_alloc(s, gfpflags, node, addr, c);
2351 void *next_object = get_freepointer_safe(s, object);
2354 * The cmpxchg will only match if there was no additional
2355 * operation and if we are on the right processor.
2357 * The cmpxchg does the following atomically (without lock semantics!)
2358 * 1. Relocate first pointer to the current per cpu area.
2359 * 2. Verify that tid and freelist have not been changed
2360 * 3. If they were not changed replace tid and freelist
2362 * Since this is without lock semantics the protection is only against
2363 * code executing on this cpu *not* from access by other cpus.
2365 if (unlikely(!this_cpu_cmpxchg_double(
2366 s->cpu_slab->freelist, s->cpu_slab->tid,
2368 next_object, next_tid(tid)))) {
2370 note_cmpxchg_failure("slab_alloc", s, tid);
2373 prefetch_freepointer(s, next_object);
2374 stat(s, ALLOC_FASTPATH);
2377 if (unlikely(gfpflags & __GFP_ZERO) && object)
2378 memset(object, 0, s->object_size);
2380 slab_post_alloc_hook(s, gfpflags, object);
2385 static __always_inline void *slab_alloc(struct kmem_cache *s,
2386 gfp_t gfpflags, unsigned long addr)
2388 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2391 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2393 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2395 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2399 EXPORT_SYMBOL(kmem_cache_alloc);
2401 #ifdef CONFIG_TRACING
2402 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2404 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2405 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2408 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2410 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2412 void *ret = kmalloc_order(size, flags, order);
2413 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2416 EXPORT_SYMBOL(kmalloc_order_trace);
2420 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2422 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2424 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2425 s->object_size, s->size, gfpflags, node);
2429 EXPORT_SYMBOL(kmem_cache_alloc_node);
2431 #ifdef CONFIG_TRACING
2432 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2434 int node, size_t size)
2436 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2438 trace_kmalloc_node(_RET_IP_, ret,
2439 size, s->size, gfpflags, node);
2442 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2447 * Slow patch handling. This may still be called frequently since objects
2448 * have a longer lifetime than the cpu slabs in most processing loads.
2450 * So we still attempt to reduce cache line usage. Just take the slab
2451 * lock and free the item. If there is no additional partial page
2452 * handling required then we can return immediately.
2454 static void __slab_free(struct kmem_cache *s, struct page *page,
2455 void *x, unsigned long addr)
2458 void **object = (void *)x;
2461 unsigned long counters;
2462 struct kmem_cache_node *n = NULL;
2463 unsigned long uninitialized_var(flags);
2465 stat(s, FREE_SLOWPATH);
2467 if (kmem_cache_debug(s) &&
2468 !(n = free_debug_processing(s, page, x, addr, &flags)))
2473 spin_unlock_irqrestore(&n->list_lock, flags);
2476 prior = page->freelist;
2477 counters = page->counters;
2478 set_freepointer(s, object, prior);
2479 new.counters = counters;
2480 was_frozen = new.frozen;
2482 if ((!new.inuse || !prior) && !was_frozen) {
2484 if (!kmem_cache_debug(s) && !prior)
2487 * Slab was on no list before and will be partially empty
2488 * We can defer the list move and instead freeze it.
2492 else { /* Needs to be taken off a list */
2494 n = get_node(s, page_to_nid(page));
2496 * Speculatively acquire the list_lock.
2497 * If the cmpxchg does not succeed then we may
2498 * drop the list_lock without any processing.
2500 * Otherwise the list_lock will synchronize with
2501 * other processors updating the list of slabs.
2503 spin_lock_irqsave(&n->list_lock, flags);
2508 } while (!cmpxchg_double_slab(s, page,
2510 object, new.counters,
2516 * If we just froze the page then put it onto the
2517 * per cpu partial list.
2519 if (new.frozen && !was_frozen) {
2520 put_cpu_partial(s, page, 1);
2521 stat(s, CPU_PARTIAL_FREE);
2524 * The list lock was not taken therefore no list
2525 * activity can be necessary.
2528 stat(s, FREE_FROZEN);
2532 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2536 * Objects left in the slab. If it was not on the partial list before
2539 if (kmem_cache_debug(s) && unlikely(!prior)) {
2540 remove_full(s, page);
2541 add_partial(n, page, DEACTIVATE_TO_TAIL);
2542 stat(s, FREE_ADD_PARTIAL);
2544 spin_unlock_irqrestore(&n->list_lock, flags);
2550 * Slab on the partial list.
2552 remove_partial(n, page);
2553 stat(s, FREE_REMOVE_PARTIAL);
2555 /* Slab must be on the full list */
2556 remove_full(s, page);
2558 spin_unlock_irqrestore(&n->list_lock, flags);
2560 discard_slab(s, page);
2564 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2565 * can perform fastpath freeing without additional function calls.
2567 * The fastpath is only possible if we are freeing to the current cpu slab
2568 * of this processor. This typically the case if we have just allocated
2571 * If fastpath is not possible then fall back to __slab_free where we deal
2572 * with all sorts of special processing.
2574 static __always_inline void slab_free(struct kmem_cache *s,
2575 struct page *page, void *x, unsigned long addr)
2577 void **object = (void *)x;
2578 struct kmem_cache_cpu *c;
2581 slab_free_hook(s, x);
2585 * Determine the currently cpus per cpu slab.
2586 * The cpu may change afterward. However that does not matter since
2587 * data is retrieved via this pointer. If we are on the same cpu
2588 * during the cmpxchg then the free will succedd.
2590 c = __this_cpu_ptr(s->cpu_slab);
2595 if (likely(page == c->page)) {
2596 set_freepointer(s, object, c->freelist);
2598 if (unlikely(!this_cpu_cmpxchg_double(
2599 s->cpu_slab->freelist, s->cpu_slab->tid,
2601 object, next_tid(tid)))) {
2603 note_cmpxchg_failure("slab_free", s, tid);
2606 stat(s, FREE_FASTPATH);
2608 __slab_free(s, page, x, addr);
2612 void kmem_cache_free(struct kmem_cache *s, void *x)
2616 page = virt_to_head_page(x);
2618 if (kmem_cache_debug(s) && page->slab_cache != s) {
2619 pr_err("kmem_cache_free: Wrong slab cache. %s but object"
2620 " is from %s\n", page->slab_cache->name, s->name);
2625 slab_free(s, page, x, _RET_IP_);
2627 trace_kmem_cache_free(_RET_IP_, x);
2629 EXPORT_SYMBOL(kmem_cache_free);
2632 * Object placement in a slab is made very easy because we always start at
2633 * offset 0. If we tune the size of the object to the alignment then we can
2634 * get the required alignment by putting one properly sized object after
2637 * Notice that the allocation order determines the sizes of the per cpu
2638 * caches. Each processor has always one slab available for allocations.
2639 * Increasing the allocation order reduces the number of times that slabs
2640 * must be moved on and off the partial lists and is therefore a factor in
2645 * Mininum / Maximum order of slab pages. This influences locking overhead
2646 * and slab fragmentation. A higher order reduces the number of partial slabs
2647 * and increases the number of allocations possible without having to
2648 * take the list_lock.
2650 static int slub_min_order;
2651 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2652 static int slub_min_objects;
2655 * Merge control. If this is set then no merging of slab caches will occur.
2656 * (Could be removed. This was introduced to pacify the merge skeptics.)
2658 static int slub_nomerge;
2661 * Calculate the order of allocation given an slab object size.
2663 * The order of allocation has significant impact on performance and other
2664 * system components. Generally order 0 allocations should be preferred since
2665 * order 0 does not cause fragmentation in the page allocator. Larger objects
2666 * be problematic to put into order 0 slabs because there may be too much
2667 * unused space left. We go to a higher order if more than 1/16th of the slab
2670 * In order to reach satisfactory performance we must ensure that a minimum
2671 * number of objects is in one slab. Otherwise we may generate too much
2672 * activity on the partial lists which requires taking the list_lock. This is
2673 * less a concern for large slabs though which are rarely used.
2675 * slub_max_order specifies the order where we begin to stop considering the
2676 * number of objects in a slab as critical. If we reach slub_max_order then
2677 * we try to keep the page order as low as possible. So we accept more waste
2678 * of space in favor of a small page order.
2680 * Higher order allocations also allow the placement of more objects in a
2681 * slab and thereby reduce object handling overhead. If the user has
2682 * requested a higher mininum order then we start with that one instead of
2683 * the smallest order which will fit the object.
2685 static inline int slab_order(int size, int min_objects,
2686 int max_order, int fract_leftover, int reserved)
2690 int min_order = slub_min_order;
2692 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2693 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2695 for (order = max(min_order,
2696 fls(min_objects * size - 1) - PAGE_SHIFT);
2697 order <= max_order; order++) {
2699 unsigned long slab_size = PAGE_SIZE << order;
2701 if (slab_size < min_objects * size + reserved)
2704 rem = (slab_size - reserved) % size;
2706 if (rem <= slab_size / fract_leftover)
2714 static inline int calculate_order(int size, int reserved)
2722 * Attempt to find best configuration for a slab. This
2723 * works by first attempting to generate a layout with
2724 * the best configuration and backing off gradually.
2726 * First we reduce the acceptable waste in a slab. Then
2727 * we reduce the minimum objects required in a slab.
2729 min_objects = slub_min_objects;
2731 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2732 max_objects = order_objects(slub_max_order, size, reserved);
2733 min_objects = min(min_objects, max_objects);
2735 while (min_objects > 1) {
2737 while (fraction >= 4) {
2738 order = slab_order(size, min_objects,
2739 slub_max_order, fraction, reserved);
2740 if (order <= slub_max_order)
2748 * We were unable to place multiple objects in a slab. Now
2749 * lets see if we can place a single object there.
2751 order = slab_order(size, 1, slub_max_order, 1, reserved);
2752 if (order <= slub_max_order)
2756 * Doh this slab cannot be placed using slub_max_order.
2758 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2759 if (order < MAX_ORDER)
2765 init_kmem_cache_node(struct kmem_cache_node *n)
2768 spin_lock_init(&n->list_lock);
2769 INIT_LIST_HEAD(&n->partial);
2770 #ifdef CONFIG_SLUB_DEBUG
2771 atomic_long_set(&n->nr_slabs, 0);
2772 atomic_long_set(&n->total_objects, 0);
2773 INIT_LIST_HEAD(&n->full);
2777 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2779 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2780 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2783 * Must align to double word boundary for the double cmpxchg
2784 * instructions to work; see __pcpu_double_call_return_bool().
2786 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2787 2 * sizeof(void *));
2792 init_kmem_cache_cpus(s);
2797 static struct kmem_cache *kmem_cache_node;
2800 * No kmalloc_node yet so do it by hand. We know that this is the first
2801 * slab on the node for this slabcache. There are no concurrent accesses
2804 * Note that this function only works on the kmalloc_node_cache
2805 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2806 * memory on a fresh node that has no slab structures yet.
2808 static void early_kmem_cache_node_alloc(int node)
2811 struct kmem_cache_node *n;
2813 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2815 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2818 if (page_to_nid(page) != node) {
2819 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2821 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2822 "in order to be able to continue\n");
2827 page->freelist = get_freepointer(kmem_cache_node, n);
2830 kmem_cache_node->node[node] = n;
2831 #ifdef CONFIG_SLUB_DEBUG
2832 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2833 init_tracking(kmem_cache_node, n);
2835 init_kmem_cache_node(n);
2836 inc_slabs_node(kmem_cache_node, node, page->objects);
2838 add_partial(n, page, DEACTIVATE_TO_HEAD);
2841 static void free_kmem_cache_nodes(struct kmem_cache *s)
2845 for_each_node_state(node, N_NORMAL_MEMORY) {
2846 struct kmem_cache_node *n = s->node[node];
2849 kmem_cache_free(kmem_cache_node, n);
2851 s->node[node] = NULL;
2855 static int init_kmem_cache_nodes(struct kmem_cache *s)
2859 for_each_node_state(node, N_NORMAL_MEMORY) {
2860 struct kmem_cache_node *n;
2862 if (slab_state == DOWN) {
2863 early_kmem_cache_node_alloc(node);
2866 n = kmem_cache_alloc_node(kmem_cache_node,
2870 free_kmem_cache_nodes(s);
2875 init_kmem_cache_node(n);
2880 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2882 if (min < MIN_PARTIAL)
2884 else if (min > MAX_PARTIAL)
2886 s->min_partial = min;
2890 * calculate_sizes() determines the order and the distribution of data within
2893 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2895 unsigned long flags = s->flags;
2896 unsigned long size = s->object_size;
2900 * Round up object size to the next word boundary. We can only
2901 * place the free pointer at word boundaries and this determines
2902 * the possible location of the free pointer.
2904 size = ALIGN(size, sizeof(void *));
2906 #ifdef CONFIG_SLUB_DEBUG
2908 * Determine if we can poison the object itself. If the user of
2909 * the slab may touch the object after free or before allocation
2910 * then we should never poison the object itself.
2912 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2914 s->flags |= __OBJECT_POISON;
2916 s->flags &= ~__OBJECT_POISON;
2920 * If we are Redzoning then check if there is some space between the
2921 * end of the object and the free pointer. If not then add an
2922 * additional word to have some bytes to store Redzone information.
2924 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2925 size += sizeof(void *);
2929 * With that we have determined the number of bytes in actual use
2930 * by the object. This is the potential offset to the free pointer.
2934 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2937 * Relocate free pointer after the object if it is not
2938 * permitted to overwrite the first word of the object on
2941 * This is the case if we do RCU, have a constructor or
2942 * destructor or are poisoning the objects.
2945 size += sizeof(void *);
2948 #ifdef CONFIG_SLUB_DEBUG
2949 if (flags & SLAB_STORE_USER)
2951 * Need to store information about allocs and frees after
2954 size += 2 * sizeof(struct track);
2956 if (flags & SLAB_RED_ZONE)
2958 * Add some empty padding so that we can catch
2959 * overwrites from earlier objects rather than let
2960 * tracking information or the free pointer be
2961 * corrupted if a user writes before the start
2964 size += sizeof(void *);
2968 * SLUB stores one object immediately after another beginning from
2969 * offset 0. In order to align the objects we have to simply size
2970 * each object to conform to the alignment.
2972 size = ALIGN(size, s->align);
2974 if (forced_order >= 0)
2975 order = forced_order;
2977 order = calculate_order(size, s->reserved);
2984 s->allocflags |= __GFP_COMP;
2986 if (s->flags & SLAB_CACHE_DMA)
2987 s->allocflags |= SLUB_DMA;
2989 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2990 s->allocflags |= __GFP_RECLAIMABLE;
2993 * Determine the number of objects per slab
2995 s->oo = oo_make(order, size, s->reserved);
2996 s->min = oo_make(get_order(size), size, s->reserved);
2997 if (oo_objects(s->oo) > oo_objects(s->max))
3000 return !!oo_objects(s->oo);
3003 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3005 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3008 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3009 s->reserved = sizeof(struct rcu_head);
3011 if (!calculate_sizes(s, -1))
3013 if (disable_higher_order_debug) {
3015 * Disable debugging flags that store metadata if the min slab
3018 if (get_order(s->size) > get_order(s->object_size)) {
3019 s->flags &= ~DEBUG_METADATA_FLAGS;
3021 if (!calculate_sizes(s, -1))
3026 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3027 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3028 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3029 /* Enable fast mode */
3030 s->flags |= __CMPXCHG_DOUBLE;
3034 * The larger the object size is, the more pages we want on the partial
3035 * list to avoid pounding the page allocator excessively.
3037 set_min_partial(s, ilog2(s->size) / 2);
3040 * cpu_partial determined the maximum number of objects kept in the
3041 * per cpu partial lists of a processor.
3043 * Per cpu partial lists mainly contain slabs that just have one
3044 * object freed. If they are used for allocation then they can be
3045 * filled up again with minimal effort. The slab will never hit the
3046 * per node partial lists and therefore no locking will be required.
3048 * This setting also determines
3050 * A) The number of objects from per cpu partial slabs dumped to the
3051 * per node list when we reach the limit.
3052 * B) The number of objects in cpu partial slabs to extract from the
3053 * per node list when we run out of per cpu objects. We only fetch 50%
3054 * to keep some capacity around for frees.
3056 if (kmem_cache_debug(s))
3058 else if (s->size >= PAGE_SIZE)
3060 else if (s->size >= 1024)
3062 else if (s->size >= 256)
3063 s->cpu_partial = 13;
3065 s->cpu_partial = 30;
3068 s->remote_node_defrag_ratio = 1000;
3070 if (!init_kmem_cache_nodes(s))
3073 if (alloc_kmem_cache_cpus(s))
3076 free_kmem_cache_nodes(s);
3078 if (flags & SLAB_PANIC)
3079 panic("Cannot create slab %s size=%lu realsize=%u "
3080 "order=%u offset=%u flags=%lx\n",
3081 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3086 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3089 #ifdef CONFIG_SLUB_DEBUG
3090 void *addr = page_address(page);
3092 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3093 sizeof(long), GFP_ATOMIC);
3096 slab_err(s, page, text, s->name);
3099 get_map(s, page, map);
3100 for_each_object(p, s, addr, page->objects) {
3102 if (!test_bit(slab_index(p, s, addr), map)) {
3103 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3105 print_tracking(s, p);
3114 * Attempt to free all partial slabs on a node.
3115 * This is called from kmem_cache_close(). We must be the last thread
3116 * using the cache and therefore we do not need to lock anymore.
3118 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3120 struct page *page, *h;
3122 list_for_each_entry_safe(page, h, &n->partial, lru) {
3124 remove_partial(n, page);
3125 discard_slab(s, page);
3127 list_slab_objects(s, page,
3128 "Objects remaining in %s on kmem_cache_close()");
3134 * Release all resources used by a slab cache.
3136 static inline int kmem_cache_close(struct kmem_cache *s)
3141 /* Attempt to free all objects */
3142 for_each_node_state(node, N_NORMAL_MEMORY) {
3143 struct kmem_cache_node *n = get_node(s, node);
3146 if (n->nr_partial || slabs_node(s, node))
3149 free_percpu(s->cpu_slab);
3150 free_kmem_cache_nodes(s);
3154 int __kmem_cache_shutdown(struct kmem_cache *s)
3156 int rc = kmem_cache_close(s);
3159 sysfs_slab_remove(s);
3164 /********************************************************************
3166 *******************************************************************/
3168 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3169 EXPORT_SYMBOL(kmalloc_caches);
3171 #ifdef CONFIG_ZONE_DMA
3172 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3175 static int __init setup_slub_min_order(char *str)
3177 get_option(&str, &slub_min_order);
3182 __setup("slub_min_order=", setup_slub_min_order);
3184 static int __init setup_slub_max_order(char *str)
3186 get_option(&str, &slub_max_order);
3187 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3192 __setup("slub_max_order=", setup_slub_max_order);
3194 static int __init setup_slub_min_objects(char *str)
3196 get_option(&str, &slub_min_objects);
3201 __setup("slub_min_objects=", setup_slub_min_objects);
3203 static int __init setup_slub_nomerge(char *str)
3209 __setup("slub_nomerge", setup_slub_nomerge);
3212 * Conversion table for small slabs sizes / 8 to the index in the
3213 * kmalloc array. This is necessary for slabs < 192 since we have non power
3214 * of two cache sizes there. The size of larger slabs can be determined using
3217 static s8 size_index[24] = {
3244 static inline int size_index_elem(size_t bytes)
3246 return (bytes - 1) / 8;
3249 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3255 return ZERO_SIZE_PTR;
3257 index = size_index[size_index_elem(size)];
3259 index = fls(size - 1);
3261 #ifdef CONFIG_ZONE_DMA
3262 if (unlikely((flags & SLUB_DMA)))
3263 return kmalloc_dma_caches[index];
3266 return kmalloc_caches[index];
3269 void *__kmalloc(size_t size, gfp_t flags)
3271 struct kmem_cache *s;
3274 if (unlikely(size > SLUB_MAX_SIZE))
3275 return kmalloc_large(size, flags);
3277 s = get_slab(size, flags);
3279 if (unlikely(ZERO_OR_NULL_PTR(s)))
3282 ret = slab_alloc(s, flags, _RET_IP_);
3284 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3288 EXPORT_SYMBOL(__kmalloc);
3291 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3296 flags |= __GFP_COMP | __GFP_NOTRACK;
3297 page = alloc_pages_node(node, flags, get_order(size));
3299 ptr = page_address(page);
3301 kmemleak_alloc(ptr, size, 1, flags);
3305 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3307 struct kmem_cache *s;
3310 if (unlikely(size > SLUB_MAX_SIZE)) {
3311 ret = kmalloc_large_node(size, flags, node);
3313 trace_kmalloc_node(_RET_IP_, ret,
3314 size, PAGE_SIZE << get_order(size),
3320 s = get_slab(size, flags);
3322 if (unlikely(ZERO_OR_NULL_PTR(s)))
3325 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3327 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3331 EXPORT_SYMBOL(__kmalloc_node);
3334 size_t ksize(const void *object)
3338 if (unlikely(object == ZERO_SIZE_PTR))
3341 page = virt_to_head_page(object);
3343 if (unlikely(!PageSlab(page))) {
3344 WARN_ON(!PageCompound(page));
3345 return PAGE_SIZE << compound_order(page);
3348 return slab_ksize(page->slab_cache);
3350 EXPORT_SYMBOL(ksize);
3352 #ifdef CONFIG_SLUB_DEBUG
3353 bool verify_mem_not_deleted(const void *x)
3356 void *object = (void *)x;
3357 unsigned long flags;
3360 if (unlikely(ZERO_OR_NULL_PTR(x)))
3363 local_irq_save(flags);
3365 page = virt_to_head_page(x);
3366 if (unlikely(!PageSlab(page))) {
3367 /* maybe it was from stack? */
3373 if (on_freelist(page->slab_cache, page, object)) {
3374 object_err(page->slab_cache, page, object, "Object is on free-list");
3382 local_irq_restore(flags);
3385 EXPORT_SYMBOL(verify_mem_not_deleted);
3388 void kfree(const void *x)
3391 void *object = (void *)x;
3393 trace_kfree(_RET_IP_, x);
3395 if (unlikely(ZERO_OR_NULL_PTR(x)))
3398 page = virt_to_head_page(x);
3399 if (unlikely(!PageSlab(page))) {
3400 BUG_ON(!PageCompound(page));
3402 __free_pages(page, compound_order(page));
3405 slab_free(page->slab_cache, page, object, _RET_IP_);
3407 EXPORT_SYMBOL(kfree);
3410 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3411 * the remaining slabs by the number of items in use. The slabs with the
3412 * most items in use come first. New allocations will then fill those up
3413 * and thus they can be removed from the partial lists.
3415 * The slabs with the least items are placed last. This results in them
3416 * being allocated from last increasing the chance that the last objects
3417 * are freed in them.
3419 int kmem_cache_shrink(struct kmem_cache *s)
3423 struct kmem_cache_node *n;
3426 int objects = oo_objects(s->max);
3427 struct list_head *slabs_by_inuse =
3428 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3429 unsigned long flags;
3431 if (!slabs_by_inuse)
3435 for_each_node_state(node, N_NORMAL_MEMORY) {
3436 n = get_node(s, node);
3441 for (i = 0; i < objects; i++)
3442 INIT_LIST_HEAD(slabs_by_inuse + i);
3444 spin_lock_irqsave(&n->list_lock, flags);
3447 * Build lists indexed by the items in use in each slab.
3449 * Note that concurrent frees may occur while we hold the
3450 * list_lock. page->inuse here is the upper limit.
3452 list_for_each_entry_safe(page, t, &n->partial, lru) {
3453 list_move(&page->lru, slabs_by_inuse + page->inuse);
3459 * Rebuild the partial list with the slabs filled up most
3460 * first and the least used slabs at the end.
3462 for (i = objects - 1; i > 0; i--)
3463 list_splice(slabs_by_inuse + i, n->partial.prev);
3465 spin_unlock_irqrestore(&n->list_lock, flags);
3467 /* Release empty slabs */
3468 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3469 discard_slab(s, page);
3472 kfree(slabs_by_inuse);
3475 EXPORT_SYMBOL(kmem_cache_shrink);
3477 #if defined(CONFIG_MEMORY_HOTPLUG)
3478 static int slab_mem_going_offline_callback(void *arg)
3480 struct kmem_cache *s;
3482 mutex_lock(&slab_mutex);
3483 list_for_each_entry(s, &slab_caches, list)
3484 kmem_cache_shrink(s);
3485 mutex_unlock(&slab_mutex);
3490 static void slab_mem_offline_callback(void *arg)
3492 struct kmem_cache_node *n;
3493 struct kmem_cache *s;
3494 struct memory_notify *marg = arg;
3497 offline_node = marg->status_change_nid_normal;
3500 * If the node still has available memory. we need kmem_cache_node
3503 if (offline_node < 0)
3506 mutex_lock(&slab_mutex);
3507 list_for_each_entry(s, &slab_caches, list) {
3508 n = get_node(s, offline_node);
3511 * if n->nr_slabs > 0, slabs still exist on the node
3512 * that is going down. We were unable to free them,
3513 * and offline_pages() function shouldn't call this
3514 * callback. So, we must fail.
3516 BUG_ON(slabs_node(s, offline_node));
3518 s->node[offline_node] = NULL;
3519 kmem_cache_free(kmem_cache_node, n);
3522 mutex_unlock(&slab_mutex);
3525 static int slab_mem_going_online_callback(void *arg)
3527 struct kmem_cache_node *n;
3528 struct kmem_cache *s;
3529 struct memory_notify *marg = arg;
3530 int nid = marg->status_change_nid_normal;
3534 * If the node's memory is already available, then kmem_cache_node is
3535 * already created. Nothing to do.
3541 * We are bringing a node online. No memory is available yet. We must
3542 * allocate a kmem_cache_node structure in order to bring the node
3545 mutex_lock(&slab_mutex);
3546 list_for_each_entry(s, &slab_caches, list) {
3548 * XXX: kmem_cache_alloc_node will fallback to other nodes
3549 * since memory is not yet available from the node that
3552 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3557 init_kmem_cache_node(n);
3561 mutex_unlock(&slab_mutex);
3565 static int slab_memory_callback(struct notifier_block *self,
3566 unsigned long action, void *arg)
3571 case MEM_GOING_ONLINE:
3572 ret = slab_mem_going_online_callback(arg);
3574 case MEM_GOING_OFFLINE:
3575 ret = slab_mem_going_offline_callback(arg);
3578 case MEM_CANCEL_ONLINE:
3579 slab_mem_offline_callback(arg);
3582 case MEM_CANCEL_OFFLINE:
3586 ret = notifier_from_errno(ret);
3592 #endif /* CONFIG_MEMORY_HOTPLUG */
3594 /********************************************************************
3595 * Basic setup of slabs
3596 *******************************************************************/
3599 * Used for early kmem_cache structures that were allocated using
3600 * the page allocator. Allocate them properly then fix up the pointers
3601 * that may be pointing to the wrong kmem_cache structure.
3604 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3607 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3609 memcpy(s, static_cache, kmem_cache->object_size);
3611 for_each_node_state(node, N_NORMAL_MEMORY) {
3612 struct kmem_cache_node *n = get_node(s, node);
3616 list_for_each_entry(p, &n->partial, lru)
3619 #ifdef CONFIG_SLUB_DEBUG
3620 list_for_each_entry(p, &n->full, lru)
3625 list_add(&s->list, &slab_caches);
3629 void __init kmem_cache_init(void)
3631 static __initdata struct kmem_cache boot_kmem_cache,
3632 boot_kmem_cache_node;
3636 if (debug_guardpage_minorder())
3639 kmem_cache_node = &boot_kmem_cache_node;
3640 kmem_cache = &boot_kmem_cache;
3642 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3643 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3645 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3647 /* Able to allocate the per node structures */
3648 slab_state = PARTIAL;
3650 create_boot_cache(kmem_cache, "kmem_cache",
3651 offsetof(struct kmem_cache, node) +
3652 nr_node_ids * sizeof(struct kmem_cache_node *),
3653 SLAB_HWCACHE_ALIGN);
3655 kmem_cache = bootstrap(&boot_kmem_cache);
3658 * Allocate kmem_cache_node properly from the kmem_cache slab.
3659 * kmem_cache_node is separately allocated so no need to
3660 * update any list pointers.
3662 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3664 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3667 * Patch up the size_index table if we have strange large alignment
3668 * requirements for the kmalloc array. This is only the case for
3669 * MIPS it seems. The standard arches will not generate any code here.
3671 * Largest permitted alignment is 256 bytes due to the way we
3672 * handle the index determination for the smaller caches.
3674 * Make sure that nothing crazy happens if someone starts tinkering
3675 * around with ARCH_KMALLOC_MINALIGN
3677 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3678 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3680 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3681 int elem = size_index_elem(i);
3682 if (elem >= ARRAY_SIZE(size_index))
3684 size_index[elem] = KMALLOC_SHIFT_LOW;
3687 if (KMALLOC_MIN_SIZE == 64) {
3689 * The 96 byte size cache is not used if the alignment
3692 for (i = 64 + 8; i <= 96; i += 8)
3693 size_index[size_index_elem(i)] = 7;
3694 } else if (KMALLOC_MIN_SIZE == 128) {
3696 * The 192 byte sized cache is not used if the alignment
3697 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3700 for (i = 128 + 8; i <= 192; i += 8)
3701 size_index[size_index_elem(i)] = 8;
3704 /* Caches that are not of the two-to-the-power-of size */
3705 if (KMALLOC_MIN_SIZE <= 32) {
3706 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3710 if (KMALLOC_MIN_SIZE <= 64) {
3711 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3715 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3716 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3722 /* Provide the correct kmalloc names now that the caches are up */
3723 if (KMALLOC_MIN_SIZE <= 32) {
3724 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3725 BUG_ON(!kmalloc_caches[1]->name);
3728 if (KMALLOC_MIN_SIZE <= 64) {
3729 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3730 BUG_ON(!kmalloc_caches[2]->name);
3733 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3734 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3737 kmalloc_caches[i]->name = s;
3741 register_cpu_notifier(&slab_notifier);
3744 #ifdef CONFIG_ZONE_DMA
3745 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3746 struct kmem_cache *s = kmalloc_caches[i];
3749 char *name = kasprintf(GFP_NOWAIT,
3750 "dma-kmalloc-%d", s->object_size);
3753 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3754 s->object_size, SLAB_CACHE_DMA);
3759 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3760 " CPUs=%d, Nodes=%d\n",
3761 caches, cache_line_size(),
3762 slub_min_order, slub_max_order, slub_min_objects,
3763 nr_cpu_ids, nr_node_ids);
3766 void __init kmem_cache_init_late(void)
3771 * Find a mergeable slab cache
3773 static int slab_unmergeable(struct kmem_cache *s)
3775 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3782 * We may have set a slab to be unmergeable during bootstrap.
3784 if (s->refcount < 0)
3790 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3791 size_t align, unsigned long flags, const char *name,
3792 void (*ctor)(void *))
3794 struct kmem_cache *s;
3796 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3802 size = ALIGN(size, sizeof(void *));
3803 align = calculate_alignment(flags, align, size);
3804 size = ALIGN(size, align);
3805 flags = kmem_cache_flags(size, flags, name, NULL);
3807 list_for_each_entry(s, &slab_caches, list) {
3808 if (slab_unmergeable(s))
3814 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3817 * Check if alignment is compatible.
3818 * Courtesy of Adrian Drzewiecki
3820 if ((s->size & ~(align - 1)) != s->size)
3823 if (s->size - size >= sizeof(void *))
3826 if (!cache_match_memcg(s, memcg))
3835 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3836 size_t align, unsigned long flags, void (*ctor)(void *))
3838 struct kmem_cache *s;
3840 s = find_mergeable(memcg, size, align, flags, name, ctor);
3844 * Adjust the object sizes so that we clear
3845 * the complete object on kzalloc.
3847 s->object_size = max(s->object_size, (int)size);
3848 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3850 if (sysfs_slab_alias(s, name)) {
3859 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3863 err = kmem_cache_open(s, flags);
3867 /* Mutex is not taken during early boot */
3868 if (slab_state <= UP)
3871 mutex_unlock(&slab_mutex);
3872 err = sysfs_slab_add(s);
3873 mutex_lock(&slab_mutex);
3876 kmem_cache_close(s);
3883 * Use the cpu notifier to insure that the cpu slabs are flushed when
3886 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3887 unsigned long action, void *hcpu)
3889 long cpu = (long)hcpu;
3890 struct kmem_cache *s;
3891 unsigned long flags;
3894 case CPU_UP_CANCELED:
3895 case CPU_UP_CANCELED_FROZEN:
3897 case CPU_DEAD_FROZEN:
3898 mutex_lock(&slab_mutex);
3899 list_for_each_entry(s, &slab_caches, list) {
3900 local_irq_save(flags);
3901 __flush_cpu_slab(s, cpu);
3902 local_irq_restore(flags);
3904 mutex_unlock(&slab_mutex);
3912 static struct notifier_block __cpuinitdata slab_notifier = {
3913 .notifier_call = slab_cpuup_callback
3918 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3920 struct kmem_cache *s;
3923 if (unlikely(size > SLUB_MAX_SIZE))
3924 return kmalloc_large(size, gfpflags);
3926 s = get_slab(size, gfpflags);
3928 if (unlikely(ZERO_OR_NULL_PTR(s)))
3931 ret = slab_alloc(s, gfpflags, caller);
3933 /* Honor the call site pointer we received. */
3934 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3940 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3941 int node, unsigned long caller)
3943 struct kmem_cache *s;
3946 if (unlikely(size > SLUB_MAX_SIZE)) {
3947 ret = kmalloc_large_node(size, gfpflags, node);
3949 trace_kmalloc_node(caller, ret,
3950 size, PAGE_SIZE << get_order(size),
3956 s = get_slab(size, gfpflags);
3958 if (unlikely(ZERO_OR_NULL_PTR(s)))
3961 ret = slab_alloc_node(s, gfpflags, node, caller);
3963 /* Honor the call site pointer we received. */
3964 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3971 static int count_inuse(struct page *page)
3976 static int count_total(struct page *page)
3978 return page->objects;
3982 #ifdef CONFIG_SLUB_DEBUG
3983 static int validate_slab(struct kmem_cache *s, struct page *page,
3987 void *addr = page_address(page);
3989 if (!check_slab(s, page) ||
3990 !on_freelist(s, page, NULL))
3993 /* Now we know that a valid freelist exists */
3994 bitmap_zero(map, page->objects);
3996 get_map(s, page, map);
3997 for_each_object(p, s, addr, page->objects) {
3998 if (test_bit(slab_index(p, s, addr), map))
3999 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4003 for_each_object(p, s, addr, page->objects)
4004 if (!test_bit(slab_index(p, s, addr), map))
4005 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4010 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4014 validate_slab(s, page, map);
4018 static int validate_slab_node(struct kmem_cache *s,
4019 struct kmem_cache_node *n, unsigned long *map)
4021 unsigned long count = 0;
4023 unsigned long flags;
4025 spin_lock_irqsave(&n->list_lock, flags);
4027 list_for_each_entry(page, &n->partial, lru) {
4028 validate_slab_slab(s, page, map);
4031 if (count != n->nr_partial)
4032 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4033 "counter=%ld\n", s->name, count, n->nr_partial);
4035 if (!(s->flags & SLAB_STORE_USER))
4038 list_for_each_entry(page, &n->full, lru) {
4039 validate_slab_slab(s, page, map);
4042 if (count != atomic_long_read(&n->nr_slabs))
4043 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4044 "counter=%ld\n", s->name, count,
4045 atomic_long_read(&n->nr_slabs));
4048 spin_unlock_irqrestore(&n->list_lock, flags);
4052 static long validate_slab_cache(struct kmem_cache *s)
4055 unsigned long count = 0;
4056 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4057 sizeof(unsigned long), GFP_KERNEL);
4063 for_each_node_state(node, N_NORMAL_MEMORY) {
4064 struct kmem_cache_node *n = get_node(s, node);
4066 count += validate_slab_node(s, n, map);
4072 * Generate lists of code addresses where slabcache objects are allocated
4077 unsigned long count;
4084 DECLARE_BITMAP(cpus, NR_CPUS);
4090 unsigned long count;
4091 struct location *loc;
4094 static void free_loc_track(struct loc_track *t)
4097 free_pages((unsigned long)t->loc,
4098 get_order(sizeof(struct location) * t->max));
4101 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4106 order = get_order(sizeof(struct location) * max);
4108 l = (void *)__get_free_pages(flags, order);
4113 memcpy(l, t->loc, sizeof(struct location) * t->count);
4121 static int add_location(struct loc_track *t, struct kmem_cache *s,
4122 const struct track *track)
4124 long start, end, pos;
4126 unsigned long caddr;
4127 unsigned long age = jiffies - track->when;
4133 pos = start + (end - start + 1) / 2;
4136 * There is nothing at "end". If we end up there
4137 * we need to add something to before end.
4142 caddr = t->loc[pos].addr;
4143 if (track->addr == caddr) {
4149 if (age < l->min_time)
4151 if (age > l->max_time)
4154 if (track->pid < l->min_pid)
4155 l->min_pid = track->pid;
4156 if (track->pid > l->max_pid)
4157 l->max_pid = track->pid;
4159 cpumask_set_cpu(track->cpu,
4160 to_cpumask(l->cpus));
4162 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4166 if (track->addr < caddr)
4173 * Not found. Insert new tracking element.
4175 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4181 (t->count - pos) * sizeof(struct location));
4184 l->addr = track->addr;
4188 l->min_pid = track->pid;
4189 l->max_pid = track->pid;
4190 cpumask_clear(to_cpumask(l->cpus));
4191 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4192 nodes_clear(l->nodes);
4193 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4197 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4198 struct page *page, enum track_item alloc,
4201 void *addr = page_address(page);
4204 bitmap_zero(map, page->objects);
4205 get_map(s, page, map);
4207 for_each_object(p, s, addr, page->objects)
4208 if (!test_bit(slab_index(p, s, addr), map))
4209 add_location(t, s, get_track(s, p, alloc));
4212 static int list_locations(struct kmem_cache *s, char *buf,
4213 enum track_item alloc)
4217 struct loc_track t = { 0, 0, NULL };
4219 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4220 sizeof(unsigned long), GFP_KERNEL);
4222 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4225 return sprintf(buf, "Out of memory\n");
4227 /* Push back cpu slabs */
4230 for_each_node_state(node, N_NORMAL_MEMORY) {
4231 struct kmem_cache_node *n = get_node(s, node);
4232 unsigned long flags;
4235 if (!atomic_long_read(&n->nr_slabs))
4238 spin_lock_irqsave(&n->list_lock, flags);
4239 list_for_each_entry(page, &n->partial, lru)
4240 process_slab(&t, s, page, alloc, map);
4241 list_for_each_entry(page, &n->full, lru)
4242 process_slab(&t, s, page, alloc, map);
4243 spin_unlock_irqrestore(&n->list_lock, flags);
4246 for (i = 0; i < t.count; i++) {
4247 struct location *l = &t.loc[i];
4249 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4251 len += sprintf(buf + len, "%7ld ", l->count);
4254 len += sprintf(buf + len, "%pS", (void *)l->addr);
4256 len += sprintf(buf + len, "<not-available>");
4258 if (l->sum_time != l->min_time) {
4259 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4261 (long)div_u64(l->sum_time, l->count),
4264 len += sprintf(buf + len, " age=%ld",
4267 if (l->min_pid != l->max_pid)
4268 len += sprintf(buf + len, " pid=%ld-%ld",
4269 l->min_pid, l->max_pid);
4271 len += sprintf(buf + len, " pid=%ld",
4274 if (num_online_cpus() > 1 &&
4275 !cpumask_empty(to_cpumask(l->cpus)) &&
4276 len < PAGE_SIZE - 60) {
4277 len += sprintf(buf + len, " cpus=");
4278 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4279 to_cpumask(l->cpus));
4282 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4283 len < PAGE_SIZE - 60) {
4284 len += sprintf(buf + len, " nodes=");
4285 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4289 len += sprintf(buf + len, "\n");
4295 len += sprintf(buf, "No data\n");
4300 #ifdef SLUB_RESILIENCY_TEST
4301 static void resiliency_test(void)
4305 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4307 printk(KERN_ERR "SLUB resiliency testing\n");
4308 printk(KERN_ERR "-----------------------\n");
4309 printk(KERN_ERR "A. Corruption after allocation\n");
4311 p = kzalloc(16, GFP_KERNEL);
4313 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4314 " 0x12->0x%p\n\n", p + 16);
4316 validate_slab_cache(kmalloc_caches[4]);
4318 /* Hmmm... The next two are dangerous */
4319 p = kzalloc(32, GFP_KERNEL);
4320 p[32 + sizeof(void *)] = 0x34;
4321 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4322 " 0x34 -> -0x%p\n", p);
4324 "If allocated object is overwritten then not detectable\n\n");
4326 validate_slab_cache(kmalloc_caches[5]);
4327 p = kzalloc(64, GFP_KERNEL);
4328 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4330 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4333 "If allocated object is overwritten then not detectable\n\n");
4334 validate_slab_cache(kmalloc_caches[6]);
4336 printk(KERN_ERR "\nB. Corruption after free\n");
4337 p = kzalloc(128, GFP_KERNEL);
4340 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4341 validate_slab_cache(kmalloc_caches[7]);
4343 p = kzalloc(256, GFP_KERNEL);
4346 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4348 validate_slab_cache(kmalloc_caches[8]);
4350 p = kzalloc(512, GFP_KERNEL);
4353 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4354 validate_slab_cache(kmalloc_caches[9]);
4358 static void resiliency_test(void) {};
4363 enum slab_stat_type {
4364 SL_ALL, /* All slabs */
4365 SL_PARTIAL, /* Only partially allocated slabs */
4366 SL_CPU, /* Only slabs used for cpu caches */
4367 SL_OBJECTS, /* Determine allocated objects not slabs */
4368 SL_TOTAL /* Determine object capacity not slabs */
4371 #define SO_ALL (1 << SL_ALL)
4372 #define SO_PARTIAL (1 << SL_PARTIAL)
4373 #define SO_CPU (1 << SL_CPU)
4374 #define SO_OBJECTS (1 << SL_OBJECTS)
4375 #define SO_TOTAL (1 << SL_TOTAL)
4377 static ssize_t show_slab_objects(struct kmem_cache *s,
4378 char *buf, unsigned long flags)
4380 unsigned long total = 0;
4383 unsigned long *nodes;
4384 unsigned long *per_cpu;
4386 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4389 per_cpu = nodes + nr_node_ids;
4391 if (flags & SO_CPU) {
4394 for_each_possible_cpu(cpu) {
4395 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4399 page = ACCESS_ONCE(c->page);
4403 node = page_to_nid(page);
4404 if (flags & SO_TOTAL)
4406 else if (flags & SO_OBJECTS)
4414 page = ACCESS_ONCE(c->partial);
4425 lock_memory_hotplug();
4426 #ifdef CONFIG_SLUB_DEBUG
4427 if (flags & SO_ALL) {
4428 for_each_node_state(node, N_NORMAL_MEMORY) {
4429 struct kmem_cache_node *n = get_node(s, node);
4431 if (flags & SO_TOTAL)
4432 x = atomic_long_read(&n->total_objects);
4433 else if (flags & SO_OBJECTS)
4434 x = atomic_long_read(&n->total_objects) -
4435 count_partial(n, count_free);
4438 x = atomic_long_read(&n->nr_slabs);
4445 if (flags & SO_PARTIAL) {
4446 for_each_node_state(node, N_NORMAL_MEMORY) {
4447 struct kmem_cache_node *n = get_node(s, node);
4449 if (flags & SO_TOTAL)
4450 x = count_partial(n, count_total);
4451 else if (flags & SO_OBJECTS)
4452 x = count_partial(n, count_inuse);
4459 x = sprintf(buf, "%lu", total);
4461 for_each_node_state(node, N_NORMAL_MEMORY)
4463 x += sprintf(buf + x, " N%d=%lu",
4466 unlock_memory_hotplug();
4468 return x + sprintf(buf + x, "\n");
4471 #ifdef CONFIG_SLUB_DEBUG
4472 static int any_slab_objects(struct kmem_cache *s)
4476 for_each_online_node(node) {
4477 struct kmem_cache_node *n = get_node(s, node);
4482 if (atomic_long_read(&n->total_objects))
4489 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4490 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4492 struct slab_attribute {
4493 struct attribute attr;
4494 ssize_t (*show)(struct kmem_cache *s, char *buf);
4495 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4498 #define SLAB_ATTR_RO(_name) \
4499 static struct slab_attribute _name##_attr = \
4500 __ATTR(_name, 0400, _name##_show, NULL)
4502 #define SLAB_ATTR(_name) \
4503 static struct slab_attribute _name##_attr = \
4504 __ATTR(_name, 0600, _name##_show, _name##_store)
4506 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4508 return sprintf(buf, "%d\n", s->size);
4510 SLAB_ATTR_RO(slab_size);
4512 static ssize_t align_show(struct kmem_cache *s, char *buf)
4514 return sprintf(buf, "%d\n", s->align);
4516 SLAB_ATTR_RO(align);
4518 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4520 return sprintf(buf, "%d\n", s->object_size);
4522 SLAB_ATTR_RO(object_size);
4524 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4526 return sprintf(buf, "%d\n", oo_objects(s->oo));
4528 SLAB_ATTR_RO(objs_per_slab);
4530 static ssize_t order_store(struct kmem_cache *s,
4531 const char *buf, size_t length)
4533 unsigned long order;
4536 err = strict_strtoul(buf, 10, &order);
4540 if (order > slub_max_order || order < slub_min_order)
4543 calculate_sizes(s, order);
4547 static ssize_t order_show(struct kmem_cache *s, char *buf)
4549 return sprintf(buf, "%d\n", oo_order(s->oo));
4553 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4555 return sprintf(buf, "%lu\n", s->min_partial);
4558 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4564 err = strict_strtoul(buf, 10, &min);
4568 set_min_partial(s, min);
4571 SLAB_ATTR(min_partial);
4573 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4575 return sprintf(buf, "%u\n", s->cpu_partial);
4578 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4581 unsigned long objects;
4584 err = strict_strtoul(buf, 10, &objects);
4587 if (objects && kmem_cache_debug(s))
4590 s->cpu_partial = objects;
4594 SLAB_ATTR(cpu_partial);
4596 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4600 return sprintf(buf, "%pS\n", s->ctor);
4604 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4606 return sprintf(buf, "%d\n", s->refcount - 1);
4608 SLAB_ATTR_RO(aliases);
4610 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4612 return show_slab_objects(s, buf, SO_PARTIAL);
4614 SLAB_ATTR_RO(partial);
4616 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4618 return show_slab_objects(s, buf, SO_CPU);
4620 SLAB_ATTR_RO(cpu_slabs);
4622 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4624 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4626 SLAB_ATTR_RO(objects);
4628 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4630 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4632 SLAB_ATTR_RO(objects_partial);
4634 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4641 for_each_online_cpu(cpu) {
4642 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4645 pages += page->pages;
4646 objects += page->pobjects;
4650 len = sprintf(buf, "%d(%d)", objects, pages);
4653 for_each_online_cpu(cpu) {
4654 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4656 if (page && len < PAGE_SIZE - 20)
4657 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4658 page->pobjects, page->pages);
4661 return len + sprintf(buf + len, "\n");
4663 SLAB_ATTR_RO(slabs_cpu_partial);
4665 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4667 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4670 static ssize_t reclaim_account_store(struct kmem_cache *s,
4671 const char *buf, size_t length)
4673 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4675 s->flags |= SLAB_RECLAIM_ACCOUNT;
4678 SLAB_ATTR(reclaim_account);
4680 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4682 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4684 SLAB_ATTR_RO(hwcache_align);
4686 #ifdef CONFIG_ZONE_DMA
4687 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4689 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4691 SLAB_ATTR_RO(cache_dma);
4694 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4696 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4698 SLAB_ATTR_RO(destroy_by_rcu);
4700 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4702 return sprintf(buf, "%d\n", s->reserved);
4704 SLAB_ATTR_RO(reserved);
4706 #ifdef CONFIG_SLUB_DEBUG
4707 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4709 return show_slab_objects(s, buf, SO_ALL);
4711 SLAB_ATTR_RO(slabs);
4713 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4715 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4717 SLAB_ATTR_RO(total_objects);
4719 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4721 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4724 static ssize_t sanity_checks_store(struct kmem_cache *s,
4725 const char *buf, size_t length)
4727 s->flags &= ~SLAB_DEBUG_FREE;
4728 if (buf[0] == '1') {
4729 s->flags &= ~__CMPXCHG_DOUBLE;
4730 s->flags |= SLAB_DEBUG_FREE;
4734 SLAB_ATTR(sanity_checks);
4736 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4738 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4741 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4744 s->flags &= ~SLAB_TRACE;
4745 if (buf[0] == '1') {
4746 s->flags &= ~__CMPXCHG_DOUBLE;
4747 s->flags |= SLAB_TRACE;
4753 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4755 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4758 static ssize_t red_zone_store(struct kmem_cache *s,
4759 const char *buf, size_t length)
4761 if (any_slab_objects(s))
4764 s->flags &= ~SLAB_RED_ZONE;
4765 if (buf[0] == '1') {
4766 s->flags &= ~__CMPXCHG_DOUBLE;
4767 s->flags |= SLAB_RED_ZONE;
4769 calculate_sizes(s, -1);
4772 SLAB_ATTR(red_zone);
4774 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4776 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4779 static ssize_t poison_store(struct kmem_cache *s,
4780 const char *buf, size_t length)
4782 if (any_slab_objects(s))
4785 s->flags &= ~SLAB_POISON;
4786 if (buf[0] == '1') {
4787 s->flags &= ~__CMPXCHG_DOUBLE;
4788 s->flags |= SLAB_POISON;
4790 calculate_sizes(s, -1);
4795 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4797 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4800 static ssize_t store_user_store(struct kmem_cache *s,
4801 const char *buf, size_t length)
4803 if (any_slab_objects(s))
4806 s->flags &= ~SLAB_STORE_USER;
4807 if (buf[0] == '1') {
4808 s->flags &= ~__CMPXCHG_DOUBLE;
4809 s->flags |= SLAB_STORE_USER;
4811 calculate_sizes(s, -1);
4814 SLAB_ATTR(store_user);
4816 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4821 static ssize_t validate_store(struct kmem_cache *s,
4822 const char *buf, size_t length)
4826 if (buf[0] == '1') {
4827 ret = validate_slab_cache(s);
4833 SLAB_ATTR(validate);
4835 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4837 if (!(s->flags & SLAB_STORE_USER))
4839 return list_locations(s, buf, TRACK_ALLOC);
4841 SLAB_ATTR_RO(alloc_calls);
4843 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4845 if (!(s->flags & SLAB_STORE_USER))
4847 return list_locations(s, buf, TRACK_FREE);
4849 SLAB_ATTR_RO(free_calls);
4850 #endif /* CONFIG_SLUB_DEBUG */
4852 #ifdef CONFIG_FAILSLAB
4853 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4855 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4858 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4861 s->flags &= ~SLAB_FAILSLAB;
4863 s->flags |= SLAB_FAILSLAB;
4866 SLAB_ATTR(failslab);
4869 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4874 static ssize_t shrink_store(struct kmem_cache *s,
4875 const char *buf, size_t length)
4877 if (buf[0] == '1') {
4878 int rc = kmem_cache_shrink(s);
4889 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4891 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4894 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4895 const char *buf, size_t length)
4897 unsigned long ratio;
4900 err = strict_strtoul(buf, 10, &ratio);
4905 s->remote_node_defrag_ratio = ratio * 10;
4909 SLAB_ATTR(remote_node_defrag_ratio);
4912 #ifdef CONFIG_SLUB_STATS
4913 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4915 unsigned long sum = 0;
4918 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4923 for_each_online_cpu(cpu) {
4924 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4930 len = sprintf(buf, "%lu", sum);
4933 for_each_online_cpu(cpu) {
4934 if (data[cpu] && len < PAGE_SIZE - 20)
4935 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4939 return len + sprintf(buf + len, "\n");
4942 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4946 for_each_online_cpu(cpu)
4947 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4950 #define STAT_ATTR(si, text) \
4951 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4953 return show_stat(s, buf, si); \
4955 static ssize_t text##_store(struct kmem_cache *s, \
4956 const char *buf, size_t length) \
4958 if (buf[0] != '0') \
4960 clear_stat(s, si); \
4965 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4966 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4967 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4968 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4969 STAT_ATTR(FREE_FROZEN, free_frozen);
4970 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4971 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4972 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4973 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4974 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4975 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4976 STAT_ATTR(FREE_SLAB, free_slab);
4977 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4978 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4979 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4980 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4981 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4982 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4983 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4984 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4985 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4986 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4987 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4988 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4989 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4990 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4993 static struct attribute *slab_attrs[] = {
4994 &slab_size_attr.attr,
4995 &object_size_attr.attr,
4996 &objs_per_slab_attr.attr,
4998 &min_partial_attr.attr,
4999 &cpu_partial_attr.attr,
5001 &objects_partial_attr.attr,
5003 &cpu_slabs_attr.attr,
5007 &hwcache_align_attr.attr,
5008 &reclaim_account_attr.attr,
5009 &destroy_by_rcu_attr.attr,
5011 &reserved_attr.attr,
5012 &slabs_cpu_partial_attr.attr,
5013 #ifdef CONFIG_SLUB_DEBUG
5014 &total_objects_attr.attr,
5016 &sanity_checks_attr.attr,
5018 &red_zone_attr.attr,
5020 &store_user_attr.attr,
5021 &validate_attr.attr,
5022 &alloc_calls_attr.attr,
5023 &free_calls_attr.attr,
5025 #ifdef CONFIG_ZONE_DMA
5026 &cache_dma_attr.attr,
5029 &remote_node_defrag_ratio_attr.attr,
5031 #ifdef CONFIG_SLUB_STATS
5032 &alloc_fastpath_attr.attr,
5033 &alloc_slowpath_attr.attr,
5034 &free_fastpath_attr.attr,
5035 &free_slowpath_attr.attr,
5036 &free_frozen_attr.attr,
5037 &free_add_partial_attr.attr,
5038 &free_remove_partial_attr.attr,
5039 &alloc_from_partial_attr.attr,
5040 &alloc_slab_attr.attr,
5041 &alloc_refill_attr.attr,
5042 &alloc_node_mismatch_attr.attr,
5043 &free_slab_attr.attr,
5044 &cpuslab_flush_attr.attr,
5045 &deactivate_full_attr.attr,
5046 &deactivate_empty_attr.attr,
5047 &deactivate_to_head_attr.attr,
5048 &deactivate_to_tail_attr.attr,
5049 &deactivate_remote_frees_attr.attr,
5050 &deactivate_bypass_attr.attr,
5051 &order_fallback_attr.attr,
5052 &cmpxchg_double_fail_attr.attr,
5053 &cmpxchg_double_cpu_fail_attr.attr,
5054 &cpu_partial_alloc_attr.attr,
5055 &cpu_partial_free_attr.attr,
5056 &cpu_partial_node_attr.attr,
5057 &cpu_partial_drain_attr.attr,
5059 #ifdef CONFIG_FAILSLAB
5060 &failslab_attr.attr,
5066 static struct attribute_group slab_attr_group = {
5067 .attrs = slab_attrs,
5070 static ssize_t slab_attr_show(struct kobject *kobj,
5071 struct attribute *attr,
5074 struct slab_attribute *attribute;
5075 struct kmem_cache *s;
5078 attribute = to_slab_attr(attr);
5081 if (!attribute->show)
5084 err = attribute->show(s, buf);
5089 static ssize_t slab_attr_store(struct kobject *kobj,
5090 struct attribute *attr,
5091 const char *buf, size_t len)
5093 struct slab_attribute *attribute;
5094 struct kmem_cache *s;
5097 attribute = to_slab_attr(attr);
5100 if (!attribute->store)
5103 err = attribute->store(s, buf, len);
5108 static const struct sysfs_ops slab_sysfs_ops = {
5109 .show = slab_attr_show,
5110 .store = slab_attr_store,
5113 static struct kobj_type slab_ktype = {
5114 .sysfs_ops = &slab_sysfs_ops,
5117 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5119 struct kobj_type *ktype = get_ktype(kobj);
5121 if (ktype == &slab_ktype)
5126 static const struct kset_uevent_ops slab_uevent_ops = {
5127 .filter = uevent_filter,
5130 static struct kset *slab_kset;
5132 #define ID_STR_LENGTH 64
5134 /* Create a unique string id for a slab cache:
5136 * Format :[flags-]size
5138 static char *create_unique_id(struct kmem_cache *s)
5140 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5147 * First flags affecting slabcache operations. We will only
5148 * get here for aliasable slabs so we do not need to support
5149 * too many flags. The flags here must cover all flags that
5150 * are matched during merging to guarantee that the id is
5153 if (s->flags & SLAB_CACHE_DMA)
5155 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5157 if (s->flags & SLAB_DEBUG_FREE)
5159 if (!(s->flags & SLAB_NOTRACK))
5163 p += sprintf(p, "%07d", s->size);
5165 #ifdef CONFIG_MEMCG_KMEM
5166 if (!is_root_cache(s))
5167 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5170 BUG_ON(p > name + ID_STR_LENGTH - 1);
5174 static int sysfs_slab_add(struct kmem_cache *s)
5178 int unmergeable = slab_unmergeable(s);
5182 * Slabcache can never be merged so we can use the name proper.
5183 * This is typically the case for debug situations. In that
5184 * case we can catch duplicate names easily.
5186 sysfs_remove_link(&slab_kset->kobj, s->name);
5190 * Create a unique name for the slab as a target
5193 name = create_unique_id(s);
5196 s->kobj.kset = slab_kset;
5197 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5199 kobject_put(&s->kobj);
5203 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5205 kobject_del(&s->kobj);
5206 kobject_put(&s->kobj);
5209 kobject_uevent(&s->kobj, KOBJ_ADD);
5211 /* Setup first alias */
5212 sysfs_slab_alias(s, s->name);
5218 static void sysfs_slab_remove(struct kmem_cache *s)
5220 if (slab_state < FULL)
5222 * Sysfs has not been setup yet so no need to remove the
5227 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5228 kobject_del(&s->kobj);
5229 kobject_put(&s->kobj);
5233 * Need to buffer aliases during bootup until sysfs becomes
5234 * available lest we lose that information.
5236 struct saved_alias {
5237 struct kmem_cache *s;
5239 struct saved_alias *next;
5242 static struct saved_alias *alias_list;
5244 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5246 struct saved_alias *al;
5248 if (slab_state == FULL) {
5250 * If we have a leftover link then remove it.
5252 sysfs_remove_link(&slab_kset->kobj, name);
5253 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5256 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5262 al->next = alias_list;
5267 static int __init slab_sysfs_init(void)
5269 struct kmem_cache *s;
5272 mutex_lock(&slab_mutex);
5274 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5276 mutex_unlock(&slab_mutex);
5277 printk(KERN_ERR "Cannot register slab subsystem.\n");
5283 list_for_each_entry(s, &slab_caches, list) {
5284 err = sysfs_slab_add(s);
5286 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5287 " to sysfs\n", s->name);
5290 while (alias_list) {
5291 struct saved_alias *al = alias_list;
5293 alias_list = alias_list->next;
5294 err = sysfs_slab_alias(al->s, al->name);
5296 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5297 " %s to sysfs\n", al->name);
5301 mutex_unlock(&slab_mutex);
5306 __initcall(slab_sysfs_init);
5307 #endif /* CONFIG_SYSFS */
5310 * The /proc/slabinfo ABI
5312 #ifdef CONFIG_SLABINFO
5313 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5315 unsigned long nr_partials = 0;
5316 unsigned long nr_slabs = 0;
5317 unsigned long nr_objs = 0;
5318 unsigned long nr_free = 0;
5321 for_each_online_node(node) {
5322 struct kmem_cache_node *n = get_node(s, node);
5327 nr_partials += n->nr_partial;
5328 nr_slabs += atomic_long_read(&n->nr_slabs);
5329 nr_objs += atomic_long_read(&n->total_objects);
5330 nr_free += count_partial(n, count_free);
5333 sinfo->active_objs = nr_objs - nr_free;
5334 sinfo->num_objs = nr_objs;
5335 sinfo->active_slabs = nr_slabs;
5336 sinfo->num_slabs = nr_slabs;
5337 sinfo->objects_per_slab = oo_objects(s->oo);
5338 sinfo->cache_order = oo_order(s->oo);
5341 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5345 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5346 size_t count, loff_t *ppos)
5350 #endif /* CONFIG_SLABINFO */