2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 * Issues still to be resolved:
129 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
131 * - Variable sizing of the per node arrays
134 /* Enable to test recovery from slab corruption on boot */
135 #undef SLUB_RESILIENCY_TEST
137 /* Enable to log cmpxchg failures */
138 #undef SLUB_DEBUG_CMPXCHG
141 * Mininum number of partial slabs. These will be left on the partial
142 * lists even if they are empty. kmem_cache_shrink may reclaim them.
144 #define MIN_PARTIAL 5
147 * Maximum number of desirable partial slabs.
148 * The existence of more partial slabs makes kmem_cache_shrink
149 * sort the partial list by the number of objects in the.
151 #define MAX_PARTIAL 10
153 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
154 SLAB_POISON | SLAB_STORE_USER)
157 * Debugging flags that require metadata to be stored in the slab. These get
158 * disabled when slub_debug=O is used and a cache's min order increases with
161 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
164 * Set of flags that will prevent slab merging
166 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
167 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
170 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
171 SLAB_CACHE_DMA | SLAB_NOTRACK)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void sysfs_slab_remove(struct kmem_cache *);
205 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
207 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
208 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
210 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
212 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
215 static inline void stat(const struct kmem_cache *s, enum stat_item si)
217 #ifdef CONFIG_SLUB_STATS
218 __this_cpu_inc(s->cpu_slab->stat[si]);
222 /********************************************************************
223 * Core slab cache functions
224 *******************************************************************/
226 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
228 return s->node[node];
231 /* Verify that a pointer has an address that is valid within a slab page */
232 static inline int check_valid_pointer(struct kmem_cache *s,
233 struct page *page, const void *object)
240 base = page_address(page);
241 if (object < base || object >= base + page->objects * s->size ||
242 (object - base) % s->size) {
249 static inline void *get_freepointer(struct kmem_cache *s, void *object)
251 return *(void **)(object + s->offset);
254 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
256 prefetch(object + s->offset);
259 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
263 #ifdef CONFIG_DEBUG_PAGEALLOC
264 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
266 p = get_freepointer(s, object);
271 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
273 *(void **)(object + s->offset) = fp;
276 /* Loop over all objects in a slab */
277 #define for_each_object(__p, __s, __addr, __objects) \
278 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline size_t slab_ksize(const struct kmem_cache *s)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order, unsigned long size, int reserved)
313 return ((PAGE_SIZE << order) - reserved) / size;
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
326 static inline int oo_order(struct kmem_cache_order_objects x)
328 return x.x >> OO_SHIFT;
331 static inline int oo_objects(struct kmem_cache_order_objects x)
333 return x.x & OO_MASK;
337 * Per slab locking using the pagelock
339 static __always_inline void slab_lock(struct page *page)
341 bit_spin_lock(PG_locked, &page->flags);
344 static __always_inline void slab_unlock(struct page *page)
346 __bit_spin_unlock(PG_locked, &page->flags);
349 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
352 tmp.counters = counters_new;
354 * page->counters can cover frozen/inuse/objects as well
355 * as page->_count. If we assign to ->counters directly
356 * we run the risk of losing updates to page->_count, so
357 * be careful and only assign to the fields we need.
359 page->frozen = tmp.frozen;
360 page->inuse = tmp.inuse;
361 page->objects = tmp.objects;
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
382 if (page->freelist == freelist_old && page->counters == counters_old) {
383 page->freelist = freelist_new;
384 set_page_slub_counters(page, counters_new);
392 stat(s, CMPXCHG_DOUBLE_FAIL);
394 #ifdef SLUB_DEBUG_CMPXCHG
395 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
401 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
402 void *freelist_old, unsigned long counters_old,
403 void *freelist_new, unsigned long counters_new,
406 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
407 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
408 if (s->flags & __CMPXCHG_DOUBLE) {
409 if (cmpxchg_double(&page->freelist, &page->counters,
410 freelist_old, counters_old,
411 freelist_new, counters_new))
418 local_irq_save(flags);
420 if (page->freelist == freelist_old && page->counters == counters_old) {
421 page->freelist = freelist_new;
422 set_page_slub_counters(page, counters_new);
424 local_irq_restore(flags);
428 local_irq_restore(flags);
432 stat(s, CMPXCHG_DOUBLE_FAIL);
434 #ifdef SLUB_DEBUG_CMPXCHG
435 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
441 #ifdef CONFIG_SLUB_DEBUG
443 * Determine a map of object in use on a page.
445 * Node listlock must be held to guarantee that the page does
446 * not vanish from under us.
448 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
451 void *addr = page_address(page);
453 for (p = page->freelist; p; p = get_freepointer(s, p))
454 set_bit(slab_index(p, s, addr), map);
460 #ifdef CONFIG_SLUB_DEBUG_ON
461 static int slub_debug = DEBUG_DEFAULT_FLAGS;
463 static int slub_debug;
466 static char *slub_debug_slabs;
467 static int disable_higher_order_debug;
472 static void print_section(char *text, u8 *addr, unsigned int length)
474 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
478 static struct track *get_track(struct kmem_cache *s, void *object,
479 enum track_item alloc)
484 p = object + s->offset + sizeof(void *);
486 p = object + s->inuse;
491 static void set_track(struct kmem_cache *s, void *object,
492 enum track_item alloc, unsigned long addr)
494 struct track *p = get_track(s, object, alloc);
497 #ifdef CONFIG_STACKTRACE
498 struct stack_trace trace;
501 trace.nr_entries = 0;
502 trace.max_entries = TRACK_ADDRS_COUNT;
503 trace.entries = p->addrs;
505 save_stack_trace(&trace);
507 /* See rant in lockdep.c */
508 if (trace.nr_entries != 0 &&
509 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
512 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
516 p->cpu = smp_processor_id();
517 p->pid = current->pid;
520 memset(p, 0, sizeof(struct track));
523 static void init_tracking(struct kmem_cache *s, void *object)
525 if (!(s->flags & SLAB_STORE_USER))
528 set_track(s, object, TRACK_FREE, 0UL);
529 set_track(s, object, TRACK_ALLOC, 0UL);
532 static void print_track(const char *s, struct track *t)
537 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
538 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
539 #ifdef CONFIG_STACKTRACE
542 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
544 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
551 static void print_tracking(struct kmem_cache *s, void *object)
553 if (!(s->flags & SLAB_STORE_USER))
556 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
557 print_track("Freed", get_track(s, object, TRACK_FREE));
560 static void print_page_info(struct page *page)
562 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
563 page, page->objects, page->inuse, page->freelist, page->flags);
567 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
573 vsnprintf(buf, sizeof(buf), fmt, args);
575 printk(KERN_ERR "========================================"
576 "=====================================\n");
577 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
578 printk(KERN_ERR "----------------------------------------"
579 "-------------------------------------\n\n");
581 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
584 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
590 vsnprintf(buf, sizeof(buf), fmt, args);
592 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
595 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
597 unsigned int off; /* Offset of last byte */
598 u8 *addr = page_address(page);
600 print_tracking(s, p);
602 print_page_info(page);
604 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
605 p, p - addr, get_freepointer(s, p));
608 print_section("Bytes b4 ", p - 16, 16);
610 print_section("Object ", p, min_t(unsigned long, s->object_size,
612 if (s->flags & SLAB_RED_ZONE)
613 print_section("Redzone ", p + s->object_size,
614 s->inuse - s->object_size);
617 off = s->offset + sizeof(void *);
621 if (s->flags & SLAB_STORE_USER)
622 off += 2 * sizeof(struct track);
625 /* Beginning of the filler is the free pointer */
626 print_section("Padding ", p + off, s->size - off);
631 static void object_err(struct kmem_cache *s, struct page *page,
632 u8 *object, char *reason)
634 slab_bug(s, "%s", reason);
635 print_trailer(s, page, object);
638 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
644 vsnprintf(buf, sizeof(buf), fmt, args);
646 slab_bug(s, "%s", buf);
647 print_page_info(page);
651 static void init_object(struct kmem_cache *s, void *object, u8 val)
655 if (s->flags & __OBJECT_POISON) {
656 memset(p, POISON_FREE, s->object_size - 1);
657 p[s->object_size - 1] = POISON_END;
660 if (s->flags & SLAB_RED_ZONE)
661 memset(p + s->object_size, val, s->inuse - s->object_size);
664 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
665 void *from, void *to)
667 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
668 memset(from, data, to - from);
671 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
672 u8 *object, char *what,
673 u8 *start, unsigned int value, unsigned int bytes)
678 fault = memchr_inv(start, value, bytes);
683 while (end > fault && end[-1] == value)
686 slab_bug(s, "%s overwritten", what);
687 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
688 fault, end - 1, fault[0], value);
689 print_trailer(s, page, object);
691 restore_bytes(s, what, value, fault, end);
699 * Bytes of the object to be managed.
700 * If the freepointer may overlay the object then the free
701 * pointer is the first word of the object.
703 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
706 * object + s->object_size
707 * Padding to reach word boundary. This is also used for Redzoning.
708 * Padding is extended by another word if Redzoning is enabled and
709 * object_size == inuse.
711 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
712 * 0xcc (RED_ACTIVE) for objects in use.
715 * Meta data starts here.
717 * A. Free pointer (if we cannot overwrite object on free)
718 * B. Tracking data for SLAB_STORE_USER
719 * C. Padding to reach required alignment boundary or at mininum
720 * one word if debugging is on to be able to detect writes
721 * before the word boundary.
723 * Padding is done using 0x5a (POISON_INUSE)
726 * Nothing is used beyond s->size.
728 * If slabcaches are merged then the object_size and inuse boundaries are mostly
729 * ignored. And therefore no slab options that rely on these boundaries
730 * may be used with merged slabcaches.
733 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
735 unsigned long off = s->inuse; /* The end of info */
738 /* Freepointer is placed after the object. */
739 off += sizeof(void *);
741 if (s->flags & SLAB_STORE_USER)
742 /* We also have user information there */
743 off += 2 * sizeof(struct track);
748 return check_bytes_and_report(s, page, p, "Object padding",
749 p + off, POISON_INUSE, s->size - off);
752 /* Check the pad bytes at the end of a slab page */
753 static int slab_pad_check(struct kmem_cache *s, struct page *page)
761 if (!(s->flags & SLAB_POISON))
764 start = page_address(page);
765 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
766 end = start + length;
767 remainder = length % s->size;
771 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
774 while (end > fault && end[-1] == POISON_INUSE)
777 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
778 print_section("Padding ", end - remainder, remainder);
780 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
784 static int check_object(struct kmem_cache *s, struct page *page,
785 void *object, u8 val)
788 u8 *endobject = object + s->object_size;
790 if (s->flags & SLAB_RED_ZONE) {
791 if (!check_bytes_and_report(s, page, object, "Redzone",
792 endobject, val, s->inuse - s->object_size))
795 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
796 check_bytes_and_report(s, page, p, "Alignment padding",
797 endobject, POISON_INUSE, s->inuse - s->object_size);
801 if (s->flags & SLAB_POISON) {
802 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
803 (!check_bytes_and_report(s, page, p, "Poison", p,
804 POISON_FREE, s->object_size - 1) ||
805 !check_bytes_and_report(s, page, p, "Poison",
806 p + s->object_size - 1, POISON_END, 1)))
809 * check_pad_bytes cleans up on its own.
811 check_pad_bytes(s, page, p);
814 if (!s->offset && val == SLUB_RED_ACTIVE)
816 * Object and freepointer overlap. Cannot check
817 * freepointer while object is allocated.
821 /* Check free pointer validity */
822 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
823 object_err(s, page, p, "Freepointer corrupt");
825 * No choice but to zap it and thus lose the remainder
826 * of the free objects in this slab. May cause
827 * another error because the object count is now wrong.
829 set_freepointer(s, p, NULL);
835 static int check_slab(struct kmem_cache *s, struct page *page)
839 VM_BUG_ON(!irqs_disabled());
841 if (!PageSlab(page)) {
842 slab_err(s, page, "Not a valid slab page");
846 maxobj = order_objects(compound_order(page), s->size, s->reserved);
847 if (page->objects > maxobj) {
848 slab_err(s, page, "objects %u > max %u",
849 s->name, page->objects, maxobj);
852 if (page->inuse > page->objects) {
853 slab_err(s, page, "inuse %u > max %u",
854 s->name, page->inuse, page->objects);
857 /* Slab_pad_check fixes things up after itself */
858 slab_pad_check(s, page);
863 * Determine if a certain object on a page is on the freelist. Must hold the
864 * slab lock to guarantee that the chains are in a consistent state.
866 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
871 unsigned long max_objects;
874 while (fp && nr <= page->objects) {
877 if (!check_valid_pointer(s, page, fp)) {
879 object_err(s, page, object,
880 "Freechain corrupt");
881 set_freepointer(s, object, NULL);
884 slab_err(s, page, "Freepointer corrupt");
885 page->freelist = NULL;
886 page->inuse = page->objects;
887 slab_fix(s, "Freelist cleared");
893 fp = get_freepointer(s, object);
897 max_objects = order_objects(compound_order(page), s->size, s->reserved);
898 if (max_objects > MAX_OBJS_PER_PAGE)
899 max_objects = MAX_OBJS_PER_PAGE;
901 if (page->objects != max_objects) {
902 slab_err(s, page, "Wrong number of objects. Found %d but "
903 "should be %d", page->objects, max_objects);
904 page->objects = max_objects;
905 slab_fix(s, "Number of objects adjusted.");
907 if (page->inuse != page->objects - nr) {
908 slab_err(s, page, "Wrong object count. Counter is %d but "
909 "counted were %d", page->inuse, page->objects - nr);
910 page->inuse = page->objects - nr;
911 slab_fix(s, "Object count adjusted.");
913 return search == NULL;
916 static void trace(struct kmem_cache *s, struct page *page, void *object,
919 if (s->flags & SLAB_TRACE) {
920 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
922 alloc ? "alloc" : "free",
927 print_section("Object ", (void *)object, s->object_size);
934 * Hooks for other subsystems that check memory allocations. In a typical
935 * production configuration these hooks all should produce no code at all.
937 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
939 flags &= gfp_allowed_mask;
940 lockdep_trace_alloc(flags);
941 might_sleep_if(flags & __GFP_WAIT);
943 return should_failslab(s->object_size, flags, s->flags);
946 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
948 flags &= gfp_allowed_mask;
949 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
950 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
953 static inline void slab_free_hook(struct kmem_cache *s, void *x)
955 kmemleak_free_recursive(x, s->flags);
958 * Trouble is that we may no longer disable interupts in the fast path
959 * So in order to make the debug calls that expect irqs to be
960 * disabled we need to disable interrupts temporarily.
962 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
966 local_irq_save(flags);
967 kmemcheck_slab_free(s, x, s->object_size);
968 debug_check_no_locks_freed(x, s->object_size);
969 local_irq_restore(flags);
972 if (!(s->flags & SLAB_DEBUG_OBJECTS))
973 debug_check_no_obj_freed(x, s->object_size);
977 * Tracking of fully allocated slabs for debugging purposes.
979 * list_lock must be held.
981 static void add_full(struct kmem_cache *s,
982 struct kmem_cache_node *n, struct page *page)
984 if (!(s->flags & SLAB_STORE_USER))
987 list_add(&page->lru, &n->full);
991 * list_lock must be held.
993 static void remove_full(struct kmem_cache *s, struct page *page)
995 if (!(s->flags & SLAB_STORE_USER))
998 list_del(&page->lru);
1001 /* Tracking of the number of slabs for debugging purposes */
1002 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1004 struct kmem_cache_node *n = get_node(s, node);
1006 return atomic_long_read(&n->nr_slabs);
1009 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1011 return atomic_long_read(&n->nr_slabs);
1014 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1016 struct kmem_cache_node *n = get_node(s, node);
1019 * May be called early in order to allocate a slab for the
1020 * kmem_cache_node structure. Solve the chicken-egg
1021 * dilemma by deferring the increment of the count during
1022 * bootstrap (see early_kmem_cache_node_alloc).
1025 atomic_long_inc(&n->nr_slabs);
1026 atomic_long_add(objects, &n->total_objects);
1029 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1031 struct kmem_cache_node *n = get_node(s, node);
1033 atomic_long_dec(&n->nr_slabs);
1034 atomic_long_sub(objects, &n->total_objects);
1037 /* Object debug checks for alloc/free paths */
1038 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1041 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1044 init_object(s, object, SLUB_RED_INACTIVE);
1045 init_tracking(s, object);
1048 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1049 void *object, unsigned long addr)
1051 if (!check_slab(s, page))
1054 if (!check_valid_pointer(s, page, object)) {
1055 object_err(s, page, object, "Freelist Pointer check fails");
1059 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1062 /* Success perform special debug activities for allocs */
1063 if (s->flags & SLAB_STORE_USER)
1064 set_track(s, object, TRACK_ALLOC, addr);
1065 trace(s, page, object, 1);
1066 init_object(s, object, SLUB_RED_ACTIVE);
1070 if (PageSlab(page)) {
1072 * If this is a slab page then lets do the best we can
1073 * to avoid issues in the future. Marking all objects
1074 * as used avoids touching the remaining objects.
1076 slab_fix(s, "Marking all objects used");
1077 page->inuse = page->objects;
1078 page->freelist = NULL;
1083 static noinline struct kmem_cache_node *free_debug_processing(
1084 struct kmem_cache *s, struct page *page, void *object,
1085 unsigned long addr, unsigned long *flags)
1087 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1089 spin_lock_irqsave(&n->list_lock, *flags);
1092 if (!check_slab(s, page))
1095 if (!check_valid_pointer(s, page, object)) {
1096 slab_err(s, page, "Invalid object pointer 0x%p", object);
1100 if (on_freelist(s, page, object)) {
1101 object_err(s, page, object, "Object already free");
1105 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1108 if (unlikely(s != page->slab_cache)) {
1109 if (!PageSlab(page)) {
1110 slab_err(s, page, "Attempt to free object(0x%p) "
1111 "outside of slab", object);
1112 } else if (!page->slab_cache) {
1114 "SLUB <none>: no slab for object 0x%p.\n",
1118 object_err(s, page, object,
1119 "page slab pointer corrupt.");
1123 if (s->flags & SLAB_STORE_USER)
1124 set_track(s, object, TRACK_FREE, addr);
1125 trace(s, page, object, 0);
1126 init_object(s, object, SLUB_RED_INACTIVE);
1130 * Keep node_lock to preserve integrity
1131 * until the object is actually freed
1137 spin_unlock_irqrestore(&n->list_lock, *flags);
1138 slab_fix(s, "Object at 0x%p not freed", object);
1142 static int __init setup_slub_debug(char *str)
1144 slub_debug = DEBUG_DEFAULT_FLAGS;
1145 if (*str++ != '=' || !*str)
1147 * No options specified. Switch on full debugging.
1153 * No options but restriction on slabs. This means full
1154 * debugging for slabs matching a pattern.
1158 if (tolower(*str) == 'o') {
1160 * Avoid enabling debugging on caches if its minimum order
1161 * would increase as a result.
1163 disable_higher_order_debug = 1;
1170 * Switch off all debugging measures.
1175 * Determine which debug features should be switched on
1177 for (; *str && *str != ','; str++) {
1178 switch (tolower(*str)) {
1180 slub_debug |= SLAB_DEBUG_FREE;
1183 slub_debug |= SLAB_RED_ZONE;
1186 slub_debug |= SLAB_POISON;
1189 slub_debug |= SLAB_STORE_USER;
1192 slub_debug |= SLAB_TRACE;
1195 slub_debug |= SLAB_FAILSLAB;
1198 printk(KERN_ERR "slub_debug option '%c' "
1199 "unknown. skipped\n", *str);
1205 slub_debug_slabs = str + 1;
1210 __setup("slub_debug", setup_slub_debug);
1212 static unsigned long kmem_cache_flags(unsigned long object_size,
1213 unsigned long flags, const char *name,
1214 void (*ctor)(void *))
1217 * Enable debugging if selected on the kernel commandline.
1219 if (slub_debug && (!slub_debug_slabs || (name &&
1220 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1221 flags |= slub_debug;
1226 static inline void setup_object_debug(struct kmem_cache *s,
1227 struct page *page, void *object) {}
1229 static inline int alloc_debug_processing(struct kmem_cache *s,
1230 struct page *page, void *object, unsigned long addr) { return 0; }
1232 static inline struct kmem_cache_node *free_debug_processing(
1233 struct kmem_cache *s, struct page *page, void *object,
1234 unsigned long addr, unsigned long *flags) { return NULL; }
1236 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1238 static inline int check_object(struct kmem_cache *s, struct page *page,
1239 void *object, u8 val) { return 1; }
1240 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1241 struct page *page) {}
1242 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1243 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1244 unsigned long flags, const char *name,
1245 void (*ctor)(void *))
1249 #define slub_debug 0
1251 #define disable_higher_order_debug 0
1253 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1255 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1257 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1259 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1262 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1265 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1268 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1270 #endif /* CONFIG_SLUB_DEBUG */
1273 * Slab allocation and freeing
1275 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1276 struct kmem_cache_order_objects oo)
1278 int order = oo_order(oo);
1280 flags |= __GFP_NOTRACK;
1282 if (node == NUMA_NO_NODE)
1283 #ifndef CONFIG_SPRD_PAGERECORDER
1284 return alloc_pages(flags, order);
1286 return alloc_pages_nopagedebug(flags, order);
1289 return alloc_pages_exact_node(node, flags, order);
1292 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1295 struct kmem_cache_order_objects oo = s->oo;
1298 flags &= gfp_allowed_mask;
1300 if (flags & __GFP_WAIT)
1303 flags |= s->allocflags;
1306 * Let the initial higher-order allocation fail under memory pressure
1307 * so we fall-back to the minimum order allocation.
1309 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC
1310 | __GFP_NO_KSWAPD) & ~(__GFP_NOFAIL|__GFP_WAIT);
1312 page = alloc_slab_page(alloc_gfp, node, oo);
1313 if (unlikely(!page)) {
1316 * Allocation may have failed due to fragmentation.
1317 * Try a lower order alloc if possible
1319 page = alloc_slab_page(flags, node, oo);
1322 stat(s, ORDER_FALLBACK);
1325 if (kmemcheck_enabled && page
1326 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1327 int pages = 1 << oo_order(oo);
1329 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1332 * Objects from caches that have a constructor don't get
1333 * cleared when they're allocated, so we need to do it here.
1336 kmemcheck_mark_uninitialized_pages(page, pages);
1338 kmemcheck_mark_unallocated_pages(page, pages);
1341 if (flags & __GFP_WAIT)
1342 local_irq_disable();
1346 page->objects = oo_objects(oo);
1347 mod_zone_page_state(page_zone(page),
1348 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1349 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1355 static void setup_object(struct kmem_cache *s, struct page *page,
1358 setup_object_debug(s, page, object);
1359 if (unlikely(s->ctor))
1363 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1371 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1373 page = allocate_slab(s,
1374 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1378 order = compound_order(page);
1379 inc_slabs_node(s, page_to_nid(page), page->objects);
1380 memcg_bind_pages(s, order);
1381 page->slab_cache = s;
1382 __SetPageSlab(page);
1383 if (page->pfmemalloc)
1384 SetPageSlabPfmemalloc(page);
1386 start = page_address(page);
1388 if (unlikely(s->flags & SLAB_POISON))
1389 memset(start, POISON_INUSE, PAGE_SIZE << order);
1392 for_each_object(p, s, start, page->objects) {
1393 setup_object(s, page, last);
1394 set_freepointer(s, last, p);
1397 setup_object(s, page, last);
1398 set_freepointer(s, last, NULL);
1400 page->freelist = start;
1401 page->inuse = page->objects;
1407 static void __free_slab(struct kmem_cache *s, struct page *page)
1409 int order = compound_order(page);
1410 int pages = 1 << order;
1412 if (kmem_cache_debug(s)) {
1415 slab_pad_check(s, page);
1416 for_each_object(p, s, page_address(page),
1418 check_object(s, page, p, SLUB_RED_INACTIVE);
1421 kmemcheck_free_shadow(page, compound_order(page));
1423 mod_zone_page_state(page_zone(page),
1424 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1425 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1428 __ClearPageSlabPfmemalloc(page);
1429 __ClearPageSlab(page);
1431 memcg_release_pages(s, order);
1432 page_mapcount_reset(page);
1433 if (current->reclaim_state)
1434 current->reclaim_state->reclaimed_slab += pages;
1436 #ifndef CONFIG_SPRD_PAGERECORDER
1437 __free_memcg_kmem_pages(page, order);
1439 __free_memcg_kmem_pages_nopagedebug(page, order);
1443 #define need_reserve_slab_rcu \
1444 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1446 static void rcu_free_slab(struct rcu_head *h)
1450 if (need_reserve_slab_rcu)
1451 page = virt_to_head_page(h);
1453 page = container_of((struct list_head *)h, struct page, lru);
1455 __free_slab(page->slab_cache, page);
1458 static void free_slab(struct kmem_cache *s, struct page *page)
1460 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1461 struct rcu_head *head;
1463 if (need_reserve_slab_rcu) {
1464 int order = compound_order(page);
1465 int offset = (PAGE_SIZE << order) - s->reserved;
1467 VM_BUG_ON(s->reserved != sizeof(*head));
1468 head = page_address(page) + offset;
1471 * RCU free overloads the RCU head over the LRU
1473 head = (void *)&page->lru;
1476 call_rcu(head, rcu_free_slab);
1478 __free_slab(s, page);
1481 static void discard_slab(struct kmem_cache *s, struct page *page)
1483 dec_slabs_node(s, page_to_nid(page), page->objects);
1488 * Management of partially allocated slabs.
1490 * list_lock must be held.
1492 static inline void add_partial(struct kmem_cache_node *n,
1493 struct page *page, int tail)
1496 if (tail == DEACTIVATE_TO_TAIL)
1497 list_add_tail(&page->lru, &n->partial);
1499 list_add(&page->lru, &n->partial);
1503 * list_lock must be held.
1505 static inline void remove_partial(struct kmem_cache_node *n,
1508 list_del(&page->lru);
1513 * Remove slab from the partial list, freeze it and
1514 * return the pointer to the freelist.
1516 * Returns a list of objects or NULL if it fails.
1518 * Must hold list_lock since we modify the partial list.
1520 static inline void *acquire_slab(struct kmem_cache *s,
1521 struct kmem_cache_node *n, struct page *page,
1522 int mode, int *objects)
1525 unsigned long counters;
1529 * Zap the freelist and set the frozen bit.
1530 * The old freelist is the list of objects for the
1531 * per cpu allocation list.
1533 freelist = page->freelist;
1534 counters = page->counters;
1535 new.counters = counters;
1536 *objects = new.objects - new.inuse;
1538 new.inuse = page->objects;
1539 new.freelist = NULL;
1541 new.freelist = freelist;
1544 VM_BUG_ON(new.frozen);
1547 if (!__cmpxchg_double_slab(s, page,
1549 new.freelist, new.counters,
1553 remove_partial(n, page);
1558 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1559 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1562 * Try to allocate a partial slab from a specific node.
1564 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1565 struct kmem_cache_cpu *c, gfp_t flags)
1567 struct page *page, *page2;
1568 void *object = NULL;
1573 * Racy check. If we mistakenly see no partial slabs then we
1574 * just allocate an empty slab. If we mistakenly try to get a
1575 * partial slab and there is none available then get_partials()
1578 if (!n || !n->nr_partial)
1581 spin_lock(&n->list_lock);
1582 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1585 if (!pfmemalloc_match(page, flags))
1588 t = acquire_slab(s, n, page, object == NULL, &objects);
1592 available += objects;
1595 stat(s, ALLOC_FROM_PARTIAL);
1598 put_cpu_partial(s, page, 0);
1599 stat(s, CPU_PARTIAL_NODE);
1601 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1605 spin_unlock(&n->list_lock);
1610 * Get a page from somewhere. Search in increasing NUMA distances.
1612 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1613 struct kmem_cache_cpu *c)
1616 struct zonelist *zonelist;
1619 enum zone_type high_zoneidx = gfp_zone(flags);
1621 unsigned int cpuset_mems_cookie;
1624 * The defrag ratio allows a configuration of the tradeoffs between
1625 * inter node defragmentation and node local allocations. A lower
1626 * defrag_ratio increases the tendency to do local allocations
1627 * instead of attempting to obtain partial slabs from other nodes.
1629 * If the defrag_ratio is set to 0 then kmalloc() always
1630 * returns node local objects. If the ratio is higher then kmalloc()
1631 * may return off node objects because partial slabs are obtained
1632 * from other nodes and filled up.
1634 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1635 * defrag_ratio = 1000) then every (well almost) allocation will
1636 * first attempt to defrag slab caches on other nodes. This means
1637 * scanning over all nodes to look for partial slabs which may be
1638 * expensive if we do it every time we are trying to find a slab
1639 * with available objects.
1641 if (!s->remote_node_defrag_ratio ||
1642 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1646 cpuset_mems_cookie = get_mems_allowed();
1647 zonelist = node_zonelist(slab_node(), flags);
1648 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1649 struct kmem_cache_node *n;
1651 n = get_node(s, zone_to_nid(zone));
1653 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1654 n->nr_partial > s->min_partial) {
1655 object = get_partial_node(s, n, c, flags);
1658 * Return the object even if
1659 * put_mems_allowed indicated that
1660 * the cpuset mems_allowed was
1661 * updated in parallel. It's a
1662 * harmless race between the alloc
1663 * and the cpuset update.
1665 put_mems_allowed(cpuset_mems_cookie);
1670 } while (!put_mems_allowed(cpuset_mems_cookie));
1676 * Get a partial page, lock it and return it.
1678 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1679 struct kmem_cache_cpu *c)
1682 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1684 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1685 if (object || node != NUMA_NO_NODE)
1688 return get_any_partial(s, flags, c);
1691 #ifdef CONFIG_PREEMPT
1693 * Calculate the next globally unique transaction for disambiguiation
1694 * during cmpxchg. The transactions start with the cpu number and are then
1695 * incremented by CONFIG_NR_CPUS.
1697 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1700 * No preemption supported therefore also no need to check for
1706 static inline unsigned long next_tid(unsigned long tid)
1708 return tid + TID_STEP;
1711 static inline unsigned int tid_to_cpu(unsigned long tid)
1713 return tid % TID_STEP;
1716 static inline unsigned long tid_to_event(unsigned long tid)
1718 return tid / TID_STEP;
1721 static inline unsigned int init_tid(int cpu)
1726 static inline void note_cmpxchg_failure(const char *n,
1727 const struct kmem_cache *s, unsigned long tid)
1729 #ifdef SLUB_DEBUG_CMPXCHG
1730 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1732 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1734 #ifdef CONFIG_PREEMPT
1735 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1736 printk("due to cpu change %d -> %d\n",
1737 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1740 if (tid_to_event(tid) != tid_to_event(actual_tid))
1741 printk("due to cpu running other code. Event %ld->%ld\n",
1742 tid_to_event(tid), tid_to_event(actual_tid));
1744 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1745 actual_tid, tid, next_tid(tid));
1747 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1750 static void init_kmem_cache_cpus(struct kmem_cache *s)
1754 for_each_possible_cpu(cpu)
1755 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1759 * Remove the cpu slab
1761 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1763 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1764 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1766 enum slab_modes l = M_NONE, m = M_NONE;
1768 int tail = DEACTIVATE_TO_HEAD;
1772 if (page->freelist) {
1773 stat(s, DEACTIVATE_REMOTE_FREES);
1774 tail = DEACTIVATE_TO_TAIL;
1778 * Stage one: Free all available per cpu objects back
1779 * to the page freelist while it is still frozen. Leave the
1782 * There is no need to take the list->lock because the page
1785 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1787 unsigned long counters;
1790 prior = page->freelist;
1791 counters = page->counters;
1792 set_freepointer(s, freelist, prior);
1793 new.counters = counters;
1795 VM_BUG_ON(!new.frozen);
1797 } while (!__cmpxchg_double_slab(s, page,
1799 freelist, new.counters,
1800 "drain percpu freelist"));
1802 freelist = nextfree;
1806 * Stage two: Ensure that the page is unfrozen while the
1807 * list presence reflects the actual number of objects
1810 * We setup the list membership and then perform a cmpxchg
1811 * with the count. If there is a mismatch then the page
1812 * is not unfrozen but the page is on the wrong list.
1814 * Then we restart the process which may have to remove
1815 * the page from the list that we just put it on again
1816 * because the number of objects in the slab may have
1821 old.freelist = page->freelist;
1822 old.counters = page->counters;
1823 VM_BUG_ON(!old.frozen);
1825 /* Determine target state of the slab */
1826 new.counters = old.counters;
1829 set_freepointer(s, freelist, old.freelist);
1830 new.freelist = freelist;
1832 new.freelist = old.freelist;
1836 if (!new.inuse && n->nr_partial > s->min_partial)
1838 else if (new.freelist) {
1843 * Taking the spinlock removes the possiblity
1844 * that acquire_slab() will see a slab page that
1847 spin_lock(&n->list_lock);
1851 if (kmem_cache_debug(s) && !lock) {
1854 * This also ensures that the scanning of full
1855 * slabs from diagnostic functions will not see
1858 spin_lock(&n->list_lock);
1866 remove_partial(n, page);
1868 else if (l == M_FULL)
1870 remove_full(s, page);
1872 if (m == M_PARTIAL) {
1874 add_partial(n, page, tail);
1877 } else if (m == M_FULL) {
1879 stat(s, DEACTIVATE_FULL);
1880 add_full(s, n, page);
1886 if (!__cmpxchg_double_slab(s, page,
1887 old.freelist, old.counters,
1888 new.freelist, new.counters,
1893 spin_unlock(&n->list_lock);
1896 stat(s, DEACTIVATE_EMPTY);
1897 discard_slab(s, page);
1903 * Unfreeze all the cpu partial slabs.
1905 * This function must be called with interrupts disabled
1906 * for the cpu using c (or some other guarantee must be there
1907 * to guarantee no concurrent accesses).
1909 static void unfreeze_partials(struct kmem_cache *s,
1910 struct kmem_cache_cpu *c)
1912 struct kmem_cache_node *n = NULL, *n2 = NULL;
1913 struct page *page, *discard_page = NULL;
1915 while ((page = c->partial)) {
1919 c->partial = page->next;
1921 n2 = get_node(s, page_to_nid(page));
1924 spin_unlock(&n->list_lock);
1927 spin_lock(&n->list_lock);
1932 old.freelist = page->freelist;
1933 old.counters = page->counters;
1934 VM_BUG_ON(!old.frozen);
1936 new.counters = old.counters;
1937 new.freelist = old.freelist;
1941 } while (!__cmpxchg_double_slab(s, page,
1942 old.freelist, old.counters,
1943 new.freelist, new.counters,
1944 "unfreezing slab"));
1946 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1947 page->next = discard_page;
1948 discard_page = page;
1950 add_partial(n, page, DEACTIVATE_TO_TAIL);
1951 stat(s, FREE_ADD_PARTIAL);
1956 spin_unlock(&n->list_lock);
1958 while (discard_page) {
1959 page = discard_page;
1960 discard_page = discard_page->next;
1962 stat(s, DEACTIVATE_EMPTY);
1963 discard_slab(s, page);
1969 * Put a page that was just frozen (in __slab_free) into a partial page
1970 * slot if available. This is done without interrupts disabled and without
1971 * preemption disabled. The cmpxchg is racy and may put the partial page
1972 * onto a random cpus partial slot.
1974 * If we did not find a slot then simply move all the partials to the
1975 * per node partial list.
1977 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1979 struct page *oldpage;
1986 oldpage = this_cpu_read(s->cpu_slab->partial);
1989 pobjects = oldpage->pobjects;
1990 pages = oldpage->pages;
1991 if (drain && pobjects > s->cpu_partial) {
1992 unsigned long flags;
1994 * partial array is full. Move the existing
1995 * set to the per node partial list.
1997 local_irq_save(flags);
1998 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1999 local_irq_restore(flags);
2003 stat(s, CPU_PARTIAL_DRAIN);
2008 pobjects += page->objects - page->inuse;
2010 page->pages = pages;
2011 page->pobjects = pobjects;
2012 page->next = oldpage;
2014 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2017 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2019 stat(s, CPUSLAB_FLUSH);
2020 deactivate_slab(s, c->page, c->freelist);
2022 c->tid = next_tid(c->tid);
2030 * Called from IPI handler with interrupts disabled.
2032 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2034 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2040 unfreeze_partials(s, c);
2044 static void flush_cpu_slab(void *d)
2046 struct kmem_cache *s = d;
2048 __flush_cpu_slab(s, smp_processor_id());
2051 static bool has_cpu_slab(int cpu, void *info)
2053 struct kmem_cache *s = info;
2054 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2056 return c->page || c->partial;
2059 static void flush_all(struct kmem_cache *s)
2061 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2065 * Check if the objects in a per cpu structure fit numa
2066 * locality expectations.
2068 static inline int node_match(struct page *page, int node)
2071 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2077 static int count_free(struct page *page)
2079 return page->objects - page->inuse;
2082 static unsigned long count_partial(struct kmem_cache_node *n,
2083 int (*get_count)(struct page *))
2085 unsigned long flags;
2086 unsigned long x = 0;
2089 spin_lock_irqsave(&n->list_lock, flags);
2090 list_for_each_entry(page, &n->partial, lru)
2091 x += get_count(page);
2092 spin_unlock_irqrestore(&n->list_lock, flags);
2096 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2098 #ifdef CONFIG_SLUB_DEBUG
2099 return atomic_long_read(&n->total_objects);
2105 static noinline void
2106 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2111 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2113 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2114 "default order: %d, min order: %d\n", s->name, s->object_size,
2115 s->size, oo_order(s->oo), oo_order(s->min));
2117 if (oo_order(s->min) > get_order(s->object_size))
2118 printk(KERN_WARNING " %s debugging increased min order, use "
2119 "slub_debug=O to disable.\n", s->name);
2121 for_each_online_node(node) {
2122 struct kmem_cache_node *n = get_node(s, node);
2123 unsigned long nr_slabs;
2124 unsigned long nr_objs;
2125 unsigned long nr_free;
2130 nr_free = count_partial(n, count_free);
2131 nr_slabs = node_nr_slabs(n);
2132 nr_objs = node_nr_objs(n);
2135 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2136 node, nr_slabs, nr_objs, nr_free);
2140 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2141 int node, struct kmem_cache_cpu **pc)
2144 struct kmem_cache_cpu *c = *pc;
2147 freelist = get_partial(s, flags, node, c);
2152 page = new_slab(s, flags, node);
2154 c = __this_cpu_ptr(s->cpu_slab);
2159 * No other reference to the page yet so we can
2160 * muck around with it freely without cmpxchg
2162 freelist = page->freelist;
2163 page->freelist = NULL;
2165 stat(s, ALLOC_SLAB);
2174 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2176 if (unlikely(PageSlabPfmemalloc(page)))
2177 return gfp_pfmemalloc_allowed(gfpflags);
2183 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2184 * or deactivate the page.
2186 * The page is still frozen if the return value is not NULL.
2188 * If this function returns NULL then the page has been unfrozen.
2190 * This function must be called with interrupt disabled.
2192 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2195 unsigned long counters;
2199 freelist = page->freelist;
2200 counters = page->counters;
2202 new.counters = counters;
2203 VM_BUG_ON(!new.frozen);
2205 new.inuse = page->objects;
2206 new.frozen = freelist != NULL;
2208 } while (!__cmpxchg_double_slab(s, page,
2217 * Slow path. The lockless freelist is empty or we need to perform
2220 * Processing is still very fast if new objects have been freed to the
2221 * regular freelist. In that case we simply take over the regular freelist
2222 * as the lockless freelist and zap the regular freelist.
2224 * If that is not working then we fall back to the partial lists. We take the
2225 * first element of the freelist as the object to allocate now and move the
2226 * rest of the freelist to the lockless freelist.
2228 * And if we were unable to get a new slab from the partial slab lists then
2229 * we need to allocate a new slab. This is the slowest path since it involves
2230 * a call to the page allocator and the setup of a new slab.
2232 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2233 unsigned long addr, struct kmem_cache_cpu *c)
2237 unsigned long flags;
2239 local_irq_save(flags);
2240 #ifdef CONFIG_PREEMPT
2242 * We may have been preempted and rescheduled on a different
2243 * cpu before disabling interrupts. Need to reload cpu area
2246 c = this_cpu_ptr(s->cpu_slab);
2254 if (unlikely(!node_match(page, node))) {
2255 stat(s, ALLOC_NODE_MISMATCH);
2256 deactivate_slab(s, page, c->freelist);
2263 * By rights, we should be searching for a slab page that was
2264 * PFMEMALLOC but right now, we are losing the pfmemalloc
2265 * information when the page leaves the per-cpu allocator
2267 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2268 deactivate_slab(s, page, c->freelist);
2274 /* must check again c->freelist in case of cpu migration or IRQ */
2275 freelist = c->freelist;
2279 stat(s, ALLOC_SLOWPATH);
2281 freelist = get_freelist(s, page);
2285 stat(s, DEACTIVATE_BYPASS);
2289 stat(s, ALLOC_REFILL);
2293 * freelist is pointing to the list of objects to be used.
2294 * page is pointing to the page from which the objects are obtained.
2295 * That page must be frozen for per cpu allocations to work.
2297 VM_BUG_ON(!c->page->frozen);
2298 c->freelist = get_freepointer(s, freelist);
2299 c->tid = next_tid(c->tid);
2300 local_irq_restore(flags);
2306 page = c->page = c->partial;
2307 c->partial = page->next;
2308 stat(s, CPU_PARTIAL_ALLOC);
2313 freelist = new_slab_objects(s, gfpflags, node, &c);
2315 if (unlikely(!freelist)) {
2316 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2317 slab_out_of_memory(s, gfpflags, node);
2319 local_irq_restore(flags);
2324 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2327 /* Only entered in the debug case */
2328 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2329 goto new_slab; /* Slab failed checks. Next slab needed */
2331 deactivate_slab(s, page, get_freepointer(s, freelist));
2334 local_irq_restore(flags);
2339 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2340 * have the fastpath folded into their functions. So no function call
2341 * overhead for requests that can be satisfied on the fastpath.
2343 * The fastpath works by first checking if the lockless freelist can be used.
2344 * If not then __slab_alloc is called for slow processing.
2346 * Otherwise we can simply pick the next object from the lockless free list.
2348 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2349 gfp_t gfpflags, int node, unsigned long addr)
2352 struct kmem_cache_cpu *c;
2356 if (slab_pre_alloc_hook(s, gfpflags))
2359 s = memcg_kmem_get_cache(s, gfpflags);
2362 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2363 * enabled. We may switch back and forth between cpus while
2364 * reading from one cpu area. That does not matter as long
2365 * as we end up on the original cpu again when doing the cmpxchg.
2367 * Preemption is disabled for the retrieval of the tid because that
2368 * must occur from the current processor. We cannot allow rescheduling
2369 * on a different processor between the determination of the pointer
2370 * and the retrieval of the tid.
2373 c = __this_cpu_ptr(s->cpu_slab);
2376 * The transaction ids are globally unique per cpu and per operation on
2377 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2378 * occurs on the right processor and that there was no operation on the
2379 * linked list in between.
2384 object = c->freelist;
2386 if (unlikely(!object || !node_match(page, node)))
2387 object = __slab_alloc(s, gfpflags, node, addr, c);
2390 void *next_object = get_freepointer_safe(s, object);
2393 * The cmpxchg will only match if there was no additional
2394 * operation and if we are on the right processor.
2396 * The cmpxchg does the following atomically (without lock semantics!)
2397 * 1. Relocate first pointer to the current per cpu area.
2398 * 2. Verify that tid and freelist have not been changed
2399 * 3. If they were not changed replace tid and freelist
2401 * Since this is without lock semantics the protection is only against
2402 * code executing on this cpu *not* from access by other cpus.
2404 if (unlikely(!this_cpu_cmpxchg_double(
2405 s->cpu_slab->freelist, s->cpu_slab->tid,
2407 next_object, next_tid(tid)))) {
2409 note_cmpxchg_failure("slab_alloc", s, tid);
2412 prefetch_freepointer(s, next_object);
2413 stat(s, ALLOC_FASTPATH);
2416 if (unlikely(gfpflags & __GFP_ZERO) && object)
2417 memset(object, 0, s->object_size);
2419 slab_post_alloc_hook(s, gfpflags, object);
2424 static __always_inline void *slab_alloc(struct kmem_cache *s,
2425 gfp_t gfpflags, unsigned long addr)
2427 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2430 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2432 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2434 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2438 EXPORT_SYMBOL(kmem_cache_alloc);
2440 #ifdef CONFIG_TRACING
2441 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2443 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2444 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2447 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2449 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2451 void *ret = kmalloc_order(size, flags, order);
2452 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2455 EXPORT_SYMBOL(kmalloc_order_trace);
2459 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2461 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2463 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2464 s->object_size, s->size, gfpflags, node);
2468 EXPORT_SYMBOL(kmem_cache_alloc_node);
2470 #ifdef CONFIG_TRACING
2471 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2473 int node, size_t size)
2475 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2477 trace_kmalloc_node(_RET_IP_, ret,
2478 size, s->size, gfpflags, node);
2481 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2486 * Slow patch handling. This may still be called frequently since objects
2487 * have a longer lifetime than the cpu slabs in most processing loads.
2489 * So we still attempt to reduce cache line usage. Just take the slab
2490 * lock and free the item. If there is no additional partial page
2491 * handling required then we can return immediately.
2493 static void __slab_free(struct kmem_cache *s, struct page *page,
2494 void *x, unsigned long addr)
2497 void **object = (void *)x;
2500 unsigned long counters;
2501 struct kmem_cache_node *n = NULL;
2502 unsigned long uninitialized_var(flags);
2504 stat(s, FREE_SLOWPATH);
2506 if (kmem_cache_debug(s) &&
2507 !(n = free_debug_processing(s, page, x, addr, &flags)))
2512 spin_unlock_irqrestore(&n->list_lock, flags);
2515 prior = page->freelist;
2516 counters = page->counters;
2517 set_freepointer(s, object, prior);
2518 new.counters = counters;
2519 was_frozen = new.frozen;
2521 if ((!new.inuse || !prior) && !was_frozen) {
2523 if (!kmem_cache_debug(s) && !prior)
2526 * Slab was on no list before and will be partially empty
2527 * We can defer the list move and instead freeze it.
2531 else { /* Needs to be taken off a list */
2533 n = get_node(s, page_to_nid(page));
2535 * Speculatively acquire the list_lock.
2536 * If the cmpxchg does not succeed then we may
2537 * drop the list_lock without any processing.
2539 * Otherwise the list_lock will synchronize with
2540 * other processors updating the list of slabs.
2542 spin_lock_irqsave(&n->list_lock, flags);
2547 } while (!cmpxchg_double_slab(s, page,
2549 object, new.counters,
2555 * If we just froze the page then put it onto the
2556 * per cpu partial list.
2558 if (new.frozen && !was_frozen) {
2559 put_cpu_partial(s, page, 1);
2560 stat(s, CPU_PARTIAL_FREE);
2563 * The list lock was not taken therefore no list
2564 * activity can be necessary.
2567 stat(s, FREE_FROZEN);
2571 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2575 * Objects left in the slab. If it was not on the partial list before
2578 if (kmem_cache_debug(s) && unlikely(!prior)) {
2579 remove_full(s, page);
2580 add_partial(n, page, DEACTIVATE_TO_TAIL);
2581 stat(s, FREE_ADD_PARTIAL);
2583 spin_unlock_irqrestore(&n->list_lock, flags);
2589 * Slab on the partial list.
2591 remove_partial(n, page);
2592 stat(s, FREE_REMOVE_PARTIAL);
2594 /* Slab must be on the full list */
2595 remove_full(s, page);
2597 spin_unlock_irqrestore(&n->list_lock, flags);
2599 discard_slab(s, page);
2603 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2604 * can perform fastpath freeing without additional function calls.
2606 * The fastpath is only possible if we are freeing to the current cpu slab
2607 * of this processor. This typically the case if we have just allocated
2610 * If fastpath is not possible then fall back to __slab_free where we deal
2611 * with all sorts of special processing.
2613 static __always_inline void slab_free(struct kmem_cache *s,
2614 struct page *page, void *x, unsigned long addr)
2616 void **object = (void *)x;
2617 struct kmem_cache_cpu *c;
2620 slab_free_hook(s, x);
2624 * Determine the currently cpus per cpu slab.
2625 * The cpu may change afterward. However that does not matter since
2626 * data is retrieved via this pointer. If we are on the same cpu
2627 * during the cmpxchg then the free will succedd.
2630 c = __this_cpu_ptr(s->cpu_slab);
2635 if (likely(page == c->page)) {
2636 set_freepointer(s, object, c->freelist);
2638 if (unlikely(!this_cpu_cmpxchg_double(
2639 s->cpu_slab->freelist, s->cpu_slab->tid,
2641 object, next_tid(tid)))) {
2643 note_cmpxchg_failure("slab_free", s, tid);
2646 stat(s, FREE_FASTPATH);
2648 __slab_free(s, page, x, addr);
2652 void kmem_cache_free(struct kmem_cache *s, void *x)
2654 s = cache_from_obj(s, x);
2657 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2658 trace_kmem_cache_free(_RET_IP_, x);
2660 EXPORT_SYMBOL(kmem_cache_free);
2663 * Object placement in a slab is made very easy because we always start at
2664 * offset 0. If we tune the size of the object to the alignment then we can
2665 * get the required alignment by putting one properly sized object after
2668 * Notice that the allocation order determines the sizes of the per cpu
2669 * caches. Each processor has always one slab available for allocations.
2670 * Increasing the allocation order reduces the number of times that slabs
2671 * must be moved on and off the partial lists and is therefore a factor in
2676 * Mininum / Maximum order of slab pages. This influences locking overhead
2677 * and slab fragmentation. A higher order reduces the number of partial slabs
2678 * and increases the number of allocations possible without having to
2679 * take the list_lock.
2681 static int slub_min_order;
2682 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2683 static int slub_min_objects;
2686 * Merge control. If this is set then no merging of slab caches will occur.
2687 * (Could be removed. This was introduced to pacify the merge skeptics.)
2689 static int slub_nomerge;
2692 * Calculate the order of allocation given an slab object size.
2694 * The order of allocation has significant impact on performance and other
2695 * system components. Generally order 0 allocations should be preferred since
2696 * order 0 does not cause fragmentation in the page allocator. Larger objects
2697 * be problematic to put into order 0 slabs because there may be too much
2698 * unused space left. We go to a higher order if more than 1/16th of the slab
2701 * In order to reach satisfactory performance we must ensure that a minimum
2702 * number of objects is in one slab. Otherwise we may generate too much
2703 * activity on the partial lists which requires taking the list_lock. This is
2704 * less a concern for large slabs though which are rarely used.
2706 * slub_max_order specifies the order where we begin to stop considering the
2707 * number of objects in a slab as critical. If we reach slub_max_order then
2708 * we try to keep the page order as low as possible. So we accept more waste
2709 * of space in favor of a small page order.
2711 * Higher order allocations also allow the placement of more objects in a
2712 * slab and thereby reduce object handling overhead. If the user has
2713 * requested a higher mininum order then we start with that one instead of
2714 * the smallest order which will fit the object.
2716 static inline int slab_order(int size, int min_objects,
2717 int max_order, int fract_leftover, int reserved)
2721 int min_order = slub_min_order;
2723 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2724 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2726 for (order = max(min_order,
2727 fls(min_objects * size - 1) - PAGE_SHIFT);
2728 order <= max_order; order++) {
2730 unsigned long slab_size = PAGE_SIZE << order;
2732 if (slab_size < min_objects * size + reserved)
2735 rem = (slab_size - reserved) % size;
2737 if (rem <= slab_size / fract_leftover)
2745 static inline int calculate_order(int size, int reserved)
2753 * Attempt to find best configuration for a slab. This
2754 * works by first attempting to generate a layout with
2755 * the best configuration and backing off gradually.
2757 * First we reduce the acceptable waste in a slab. Then
2758 * we reduce the minimum objects required in a slab.
2760 min_objects = slub_min_objects;
2762 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2763 max_objects = order_objects(slub_max_order, size, reserved);
2764 min_objects = min(min_objects, max_objects);
2766 while (min_objects > 1) {
2768 while (fraction >= 4) {
2769 order = slab_order(size, min_objects,
2770 slub_max_order, fraction, reserved);
2771 if (order <= slub_max_order)
2779 * We were unable to place multiple objects in a slab. Now
2780 * lets see if we can place a single object there.
2782 order = slab_order(size, 1, slub_max_order, 1, reserved);
2783 if (order <= slub_max_order)
2787 * Doh this slab cannot be placed using slub_max_order.
2789 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2790 if (order < MAX_ORDER)
2796 init_kmem_cache_node(struct kmem_cache_node *n)
2799 spin_lock_init(&n->list_lock);
2800 INIT_LIST_HEAD(&n->partial);
2801 #ifdef CONFIG_SLUB_DEBUG
2802 atomic_long_set(&n->nr_slabs, 0);
2803 atomic_long_set(&n->total_objects, 0);
2804 INIT_LIST_HEAD(&n->full);
2808 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2810 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2811 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2814 * Must align to double word boundary for the double cmpxchg
2815 * instructions to work; see __pcpu_double_call_return_bool().
2817 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2818 2 * sizeof(void *));
2823 init_kmem_cache_cpus(s);
2828 static struct kmem_cache *kmem_cache_node;
2831 * No kmalloc_node yet so do it by hand. We know that this is the first
2832 * slab on the node for this slabcache. There are no concurrent accesses
2835 * Note that this function only works on the kmalloc_node_cache
2836 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2837 * memory on a fresh node that has no slab structures yet.
2839 static void early_kmem_cache_node_alloc(int node)
2842 struct kmem_cache_node *n;
2844 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2846 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2849 if (page_to_nid(page) != node) {
2850 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2852 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2853 "in order to be able to continue\n");
2858 page->freelist = get_freepointer(kmem_cache_node, n);
2861 kmem_cache_node->node[node] = n;
2862 #ifdef CONFIG_SLUB_DEBUG
2863 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2864 init_tracking(kmem_cache_node, n);
2866 init_kmem_cache_node(n);
2867 inc_slabs_node(kmem_cache_node, node, page->objects);
2869 add_partial(n, page, DEACTIVATE_TO_HEAD);
2872 static void free_kmem_cache_nodes(struct kmem_cache *s)
2876 for_each_node_state(node, N_NORMAL_MEMORY) {
2877 struct kmem_cache_node *n = s->node[node];
2880 kmem_cache_free(kmem_cache_node, n);
2882 s->node[node] = NULL;
2886 static int init_kmem_cache_nodes(struct kmem_cache *s)
2890 for_each_node_state(node, N_NORMAL_MEMORY) {
2891 struct kmem_cache_node *n;
2893 if (slab_state == DOWN) {
2894 early_kmem_cache_node_alloc(node);
2897 n = kmem_cache_alloc_node(kmem_cache_node,
2901 free_kmem_cache_nodes(s);
2906 init_kmem_cache_node(n);
2911 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2913 if (min < MIN_PARTIAL)
2915 else if (min > MAX_PARTIAL)
2917 s->min_partial = min;
2921 * calculate_sizes() determines the order and the distribution of data within
2924 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2926 unsigned long flags = s->flags;
2927 unsigned long size = s->object_size;
2931 * Round up object size to the next word boundary. We can only
2932 * place the free pointer at word boundaries and this determines
2933 * the possible location of the free pointer.
2935 size = ALIGN(size, sizeof(void *));
2937 #ifdef CONFIG_SLUB_DEBUG
2939 * Determine if we can poison the object itself. If the user of
2940 * the slab may touch the object after free or before allocation
2941 * then we should never poison the object itself.
2943 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2945 s->flags |= __OBJECT_POISON;
2947 s->flags &= ~__OBJECT_POISON;
2951 * If we are Redzoning then check if there is some space between the
2952 * end of the object and the free pointer. If not then add an
2953 * additional word to have some bytes to store Redzone information.
2955 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2956 size += sizeof(void *);
2960 * With that we have determined the number of bytes in actual use
2961 * by the object. This is the potential offset to the free pointer.
2965 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2968 * Relocate free pointer after the object if it is not
2969 * permitted to overwrite the first word of the object on
2972 * This is the case if we do RCU, have a constructor or
2973 * destructor or are poisoning the objects.
2976 size += sizeof(void *);
2979 #ifdef CONFIG_SLUB_DEBUG
2980 if (flags & SLAB_STORE_USER)
2982 * Need to store information about allocs and frees after
2985 size += 2 * sizeof(struct track);
2987 if (flags & SLAB_RED_ZONE)
2989 * Add some empty padding so that we can catch
2990 * overwrites from earlier objects rather than let
2991 * tracking information or the free pointer be
2992 * corrupted if a user writes before the start
2995 size += sizeof(void *);
2999 * SLUB stores one object immediately after another beginning from
3000 * offset 0. In order to align the objects we have to simply size
3001 * each object to conform to the alignment.
3003 size = ALIGN(size, s->align);
3005 if (forced_order >= 0)
3006 order = forced_order;
3008 order = calculate_order(size, s->reserved);
3015 s->allocflags |= __GFP_COMP;
3017 if (s->flags & SLAB_CACHE_DMA)
3018 s->allocflags |= GFP_DMA;
3020 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3021 s->allocflags |= __GFP_RECLAIMABLE;
3024 * Determine the number of objects per slab
3026 s->oo = oo_make(order, size, s->reserved);
3027 s->min = oo_make(get_order(size), size, s->reserved);
3028 if (oo_objects(s->oo) > oo_objects(s->max))
3031 return !!oo_objects(s->oo);
3034 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3036 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3039 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3040 s->reserved = sizeof(struct rcu_head);
3042 if (!calculate_sizes(s, -1))
3044 if (disable_higher_order_debug) {
3046 * Disable debugging flags that store metadata if the min slab
3049 if (get_order(s->size) > get_order(s->object_size)) {
3050 s->flags &= ~DEBUG_METADATA_FLAGS;
3052 if (!calculate_sizes(s, -1))
3057 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3058 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3059 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3060 /* Enable fast mode */
3061 s->flags |= __CMPXCHG_DOUBLE;
3065 * The larger the object size is, the more pages we want on the partial
3066 * list to avoid pounding the page allocator excessively.
3068 set_min_partial(s, ilog2(s->size) / 2);
3071 * cpu_partial determined the maximum number of objects kept in the
3072 * per cpu partial lists of a processor.
3074 * Per cpu partial lists mainly contain slabs that just have one
3075 * object freed. If they are used for allocation then they can be
3076 * filled up again with minimal effort. The slab will never hit the
3077 * per node partial lists and therefore no locking will be required.
3079 * This setting also determines
3081 * A) The number of objects from per cpu partial slabs dumped to the
3082 * per node list when we reach the limit.
3083 * B) The number of objects in cpu partial slabs to extract from the
3084 * per node list when we run out of per cpu objects. We only fetch 50%
3085 * to keep some capacity around for frees.
3087 if (kmem_cache_debug(s))
3089 else if (s->size >= PAGE_SIZE)
3091 else if (s->size >= 1024)
3093 else if (s->size >= 256)
3094 s->cpu_partial = 13;
3096 s->cpu_partial = 30;
3099 s->remote_node_defrag_ratio = 1000;
3101 if (!init_kmem_cache_nodes(s))
3104 if (alloc_kmem_cache_cpus(s))
3107 free_kmem_cache_nodes(s);
3109 if (flags & SLAB_PANIC)
3110 panic("Cannot create slab %s size=%lu realsize=%u "
3111 "order=%u offset=%u flags=%lx\n",
3112 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3117 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3120 #ifdef CONFIG_SLUB_DEBUG
3121 void *addr = page_address(page);
3123 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3124 sizeof(long), GFP_ATOMIC);
3127 slab_err(s, page, text, s->name);
3130 get_map(s, page, map);
3131 for_each_object(p, s, addr, page->objects) {
3133 if (!test_bit(slab_index(p, s, addr), map)) {
3134 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3136 print_tracking(s, p);
3145 * Attempt to free all partial slabs on a node.
3146 * This is called from kmem_cache_close(). We must be the last thread
3147 * using the cache and therefore we do not need to lock anymore.
3149 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3151 struct page *page, *h;
3153 list_for_each_entry_safe(page, h, &n->partial, lru) {
3155 remove_partial(n, page);
3156 discard_slab(s, page);
3158 list_slab_objects(s, page,
3159 "Objects remaining in %s on kmem_cache_close()");
3165 * Release all resources used by a slab cache.
3167 static inline int kmem_cache_close(struct kmem_cache *s)
3172 /* Attempt to free all objects */
3173 for_each_node_state(node, N_NORMAL_MEMORY) {
3174 struct kmem_cache_node *n = get_node(s, node);
3177 if (n->nr_partial || slabs_node(s, node))
3180 free_percpu(s->cpu_slab);
3181 free_kmem_cache_nodes(s);
3185 int __kmem_cache_shutdown(struct kmem_cache *s)
3187 int rc = kmem_cache_close(s);
3191 * We do the same lock strategy around sysfs_slab_add, see
3192 * __kmem_cache_create. Because this is pretty much the last
3193 * operation we do and the lock will be released shortly after
3194 * that in slab_common.c, we could just move sysfs_slab_remove
3195 * to a later point in common code. We should do that when we
3196 * have a common sysfs framework for all allocators.
3198 mutex_unlock(&slab_mutex);
3199 sysfs_slab_remove(s);
3200 mutex_lock(&slab_mutex);
3206 /********************************************************************
3208 *******************************************************************/
3210 static int __init setup_slub_min_order(char *str)
3212 get_option(&str, &slub_min_order);
3217 __setup("slub_min_order=", setup_slub_min_order);
3219 static int __init setup_slub_max_order(char *str)
3221 get_option(&str, &slub_max_order);
3222 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3227 __setup("slub_max_order=", setup_slub_max_order);
3229 static int __init setup_slub_min_objects(char *str)
3231 get_option(&str, &slub_min_objects);
3236 __setup("slub_min_objects=", setup_slub_min_objects);
3238 static int __init setup_slub_nomerge(char *str)
3244 __setup("slub_nomerge", setup_slub_nomerge);
3246 void *__kmalloc(size_t size, gfp_t flags)
3248 struct kmem_cache *s;
3251 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3252 return kmalloc_large(size, flags);
3254 s = kmalloc_slab(size, flags);
3256 if (unlikely(ZERO_OR_NULL_PTR(s)))
3259 ret = slab_alloc(s, flags, _RET_IP_);
3261 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3265 EXPORT_SYMBOL(__kmalloc);
3268 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3273 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3274 page = alloc_pages_node(node, flags, get_order(size));
3276 ptr = page_address(page);
3278 kmemleak_alloc(ptr, size, 1, flags);
3282 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3284 struct kmem_cache *s;
3287 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3288 ret = kmalloc_large_node(size, flags, node);
3290 trace_kmalloc_node(_RET_IP_, ret,
3291 size, PAGE_SIZE << get_order(size),
3297 s = kmalloc_slab(size, flags);
3299 if (unlikely(ZERO_OR_NULL_PTR(s)))
3302 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3304 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3308 EXPORT_SYMBOL(__kmalloc_node);
3311 size_t ksize(const void *object)
3315 if (unlikely(object == ZERO_SIZE_PTR))
3318 page = virt_to_head_page(object);
3320 if (unlikely(!PageSlab(page))) {
3321 WARN_ON(!PageCompound(page));
3322 return PAGE_SIZE << compound_order(page);
3325 return slab_ksize(page->slab_cache);
3327 EXPORT_SYMBOL(ksize);
3329 #ifdef CONFIG_SLUB_DEBUG
3330 bool verify_mem_not_deleted(const void *x)
3333 void *object = (void *)x;
3334 unsigned long flags;
3337 if (unlikely(ZERO_OR_NULL_PTR(x)))
3340 local_irq_save(flags);
3342 page = virt_to_head_page(x);
3343 if (unlikely(!PageSlab(page))) {
3344 /* maybe it was from stack? */
3350 if (on_freelist(page->slab_cache, page, object)) {
3351 object_err(page->slab_cache, page, object, "Object is on free-list");
3359 local_irq_restore(flags);
3362 EXPORT_SYMBOL(verify_mem_not_deleted);
3365 void kfree(const void *x)
3368 void *object = (void *)x;
3370 trace_kfree(_RET_IP_, x);
3372 if (unlikely(ZERO_OR_NULL_PTR(x)))
3375 page = virt_to_head_page(x);
3376 if (unlikely(!PageSlab(page))) {
3377 BUG_ON(!PageCompound(page));
3379 __free_memcg_kmem_pages(page, compound_order(page));
3382 slab_free(page->slab_cache, page, object, _RET_IP_);
3384 EXPORT_SYMBOL(kfree);
3387 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3388 * the remaining slabs by the number of items in use. The slabs with the
3389 * most items in use come first. New allocations will then fill those up
3390 * and thus they can be removed from the partial lists.
3392 * The slabs with the least items are placed last. This results in them
3393 * being allocated from last increasing the chance that the last objects
3394 * are freed in them.
3396 int kmem_cache_shrink(struct kmem_cache *s)
3400 struct kmem_cache_node *n;
3403 int objects = oo_objects(s->max);
3404 struct list_head *slabs_by_inuse =
3405 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3406 unsigned long flags;
3408 if (!slabs_by_inuse)
3412 for_each_node_state(node, N_NORMAL_MEMORY) {
3413 n = get_node(s, node);
3418 for (i = 0; i < objects; i++)
3419 INIT_LIST_HEAD(slabs_by_inuse + i);
3421 spin_lock_irqsave(&n->list_lock, flags);
3424 * Build lists indexed by the items in use in each slab.
3426 * Note that concurrent frees may occur while we hold the
3427 * list_lock. page->inuse here is the upper limit.
3429 list_for_each_entry_safe(page, t, &n->partial, lru) {
3430 list_move(&page->lru, slabs_by_inuse + page->inuse);
3436 * Rebuild the partial list with the slabs filled up most
3437 * first and the least used slabs at the end.
3439 for (i = objects - 1; i > 0; i--)
3440 list_splice(slabs_by_inuse + i, n->partial.prev);
3442 spin_unlock_irqrestore(&n->list_lock, flags);
3444 /* Release empty slabs */
3445 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3446 discard_slab(s, page);
3449 kfree(slabs_by_inuse);
3452 EXPORT_SYMBOL(kmem_cache_shrink);
3454 static int slab_mem_going_offline_callback(void *arg)
3456 struct kmem_cache *s;
3458 mutex_lock(&slab_mutex);
3459 list_for_each_entry(s, &slab_caches, list)
3460 kmem_cache_shrink(s);
3461 mutex_unlock(&slab_mutex);
3466 static void slab_mem_offline_callback(void *arg)
3468 struct kmem_cache_node *n;
3469 struct kmem_cache *s;
3470 struct memory_notify *marg = arg;
3473 offline_node = marg->status_change_nid_normal;
3476 * If the node still has available memory. we need kmem_cache_node
3479 if (offline_node < 0)
3482 mutex_lock(&slab_mutex);
3483 list_for_each_entry(s, &slab_caches, list) {
3484 n = get_node(s, offline_node);
3487 * if n->nr_slabs > 0, slabs still exist on the node
3488 * that is going down. We were unable to free them,
3489 * and offline_pages() function shouldn't call this
3490 * callback. So, we must fail.
3492 BUG_ON(slabs_node(s, offline_node));
3494 s->node[offline_node] = NULL;
3495 kmem_cache_free(kmem_cache_node, n);
3498 mutex_unlock(&slab_mutex);
3501 static int slab_mem_going_online_callback(void *arg)
3503 struct kmem_cache_node *n;
3504 struct kmem_cache *s;
3505 struct memory_notify *marg = arg;
3506 int nid = marg->status_change_nid_normal;
3510 * If the node's memory is already available, then kmem_cache_node is
3511 * already created. Nothing to do.
3517 * We are bringing a node online. No memory is available yet. We must
3518 * allocate a kmem_cache_node structure in order to bring the node
3521 mutex_lock(&slab_mutex);
3522 list_for_each_entry(s, &slab_caches, list) {
3524 * XXX: kmem_cache_alloc_node will fallback to other nodes
3525 * since memory is not yet available from the node that
3528 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3533 init_kmem_cache_node(n);
3537 mutex_unlock(&slab_mutex);
3541 static int slab_memory_callback(struct notifier_block *self,
3542 unsigned long action, void *arg)
3547 case MEM_GOING_ONLINE:
3548 ret = slab_mem_going_online_callback(arg);
3550 case MEM_GOING_OFFLINE:
3551 ret = slab_mem_going_offline_callback(arg);
3554 case MEM_CANCEL_ONLINE:
3555 slab_mem_offline_callback(arg);
3558 case MEM_CANCEL_OFFLINE:
3562 ret = notifier_from_errno(ret);
3568 static struct notifier_block slab_memory_callback_nb = {
3569 .notifier_call = slab_memory_callback,
3570 .priority = SLAB_CALLBACK_PRI,
3573 /********************************************************************
3574 * Basic setup of slabs
3575 *******************************************************************/
3578 * Used for early kmem_cache structures that were allocated using
3579 * the page allocator. Allocate them properly then fix up the pointers
3580 * that may be pointing to the wrong kmem_cache structure.
3583 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3586 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3588 memcpy(s, static_cache, kmem_cache->object_size);
3591 * This runs very early, and only the boot processor is supposed to be
3592 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3595 __flush_cpu_slab(s, smp_processor_id());
3596 for_each_node_state(node, N_NORMAL_MEMORY) {
3597 struct kmem_cache_node *n = get_node(s, node);
3601 list_for_each_entry(p, &n->partial, lru)
3604 #ifdef CONFIG_SLUB_DEBUG
3605 list_for_each_entry(p, &n->full, lru)
3610 list_add(&s->list, &slab_caches);
3614 void __init kmem_cache_init(void)
3616 static __initdata struct kmem_cache boot_kmem_cache,
3617 boot_kmem_cache_node;
3619 if (debug_guardpage_minorder())
3622 kmem_cache_node = &boot_kmem_cache_node;
3623 kmem_cache = &boot_kmem_cache;
3625 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3626 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3628 register_hotmemory_notifier(&slab_memory_callback_nb);
3630 /* Able to allocate the per node structures */
3631 slab_state = PARTIAL;
3633 create_boot_cache(kmem_cache, "kmem_cache",
3634 offsetof(struct kmem_cache, node) +
3635 nr_node_ids * sizeof(struct kmem_cache_node *),
3636 SLAB_HWCACHE_ALIGN);
3638 kmem_cache = bootstrap(&boot_kmem_cache);
3641 * Allocate kmem_cache_node properly from the kmem_cache slab.
3642 * kmem_cache_node is separately allocated so no need to
3643 * update any list pointers.
3645 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3647 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3648 create_kmalloc_caches(0);
3651 register_cpu_notifier(&slab_notifier);
3655 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3656 " CPUs=%d, Nodes=%d\n",
3658 slub_min_order, slub_max_order, slub_min_objects,
3659 nr_cpu_ids, nr_node_ids);
3662 void __init kmem_cache_init_late(void)
3667 * Find a mergeable slab cache
3669 static int slab_unmergeable(struct kmem_cache *s)
3671 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3678 * We may have set a slab to be unmergeable during bootstrap.
3680 if (s->refcount < 0)
3686 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3687 size_t align, unsigned long flags, const char *name,
3688 void (*ctor)(void *))
3690 struct kmem_cache *s;
3692 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3698 size = ALIGN(size, sizeof(void *));
3699 align = calculate_alignment(flags, align, size);
3700 size = ALIGN(size, align);
3701 flags = kmem_cache_flags(size, flags, name, NULL);
3703 list_for_each_entry(s, &slab_caches, list) {
3704 if (slab_unmergeable(s))
3710 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3713 * Check if alignment is compatible.
3714 * Courtesy of Adrian Drzewiecki
3716 if ((s->size & ~(align - 1)) != s->size)
3719 if (s->size - size >= sizeof(void *))
3722 if (!cache_match_memcg(s, memcg))
3731 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3732 size_t align, unsigned long flags, void (*ctor)(void *))
3734 struct kmem_cache *s;
3736 s = find_mergeable(memcg, size, align, flags, name, ctor);
3740 * Adjust the object sizes so that we clear
3741 * the complete object on kzalloc.
3743 s->object_size = max(s->object_size, (int)size);
3744 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3746 if (sysfs_slab_alias(s, name)) {
3755 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3759 err = kmem_cache_open(s, flags);
3763 /* Mutex is not taken during early boot */
3764 if (slab_state <= UP)
3767 memcg_propagate_slab_attrs(s);
3768 mutex_unlock(&slab_mutex);
3769 err = sysfs_slab_add(s);
3770 mutex_lock(&slab_mutex);
3773 kmem_cache_close(s);
3780 * Use the cpu notifier to insure that the cpu slabs are flushed when
3783 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3784 unsigned long action, void *hcpu)
3786 long cpu = (long)hcpu;
3787 struct kmem_cache *s;
3788 unsigned long flags;
3791 case CPU_UP_CANCELED:
3792 case CPU_UP_CANCELED_FROZEN:
3794 case CPU_DEAD_FROZEN:
3795 mutex_lock(&slab_mutex);
3796 list_for_each_entry(s, &slab_caches, list) {
3797 local_irq_save(flags);
3798 __flush_cpu_slab(s, cpu);
3799 local_irq_restore(flags);
3801 mutex_unlock(&slab_mutex);
3809 static struct notifier_block __cpuinitdata slab_notifier = {
3810 .notifier_call = slab_cpuup_callback
3815 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3817 struct kmem_cache *s;
3820 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3821 return kmalloc_large(size, gfpflags);
3823 s = kmalloc_slab(size, gfpflags);
3825 if (unlikely(ZERO_OR_NULL_PTR(s)))
3828 ret = slab_alloc(s, gfpflags, caller);
3830 /* Honor the call site pointer we received. */
3831 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3837 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3838 int node, unsigned long caller)
3840 struct kmem_cache *s;
3843 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3844 ret = kmalloc_large_node(size, gfpflags, node);
3846 trace_kmalloc_node(caller, ret,
3847 size, PAGE_SIZE << get_order(size),
3853 s = kmalloc_slab(size, gfpflags);
3855 if (unlikely(ZERO_OR_NULL_PTR(s)))
3858 ret = slab_alloc_node(s, gfpflags, node, caller);
3860 /* Honor the call site pointer we received. */
3861 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3868 static int count_inuse(struct page *page)
3873 static int count_total(struct page *page)
3875 return page->objects;
3879 #ifdef CONFIG_SLUB_DEBUG
3880 static int validate_slab(struct kmem_cache *s, struct page *page,
3884 void *addr = page_address(page);
3886 if (!check_slab(s, page) ||
3887 !on_freelist(s, page, NULL))
3890 /* Now we know that a valid freelist exists */
3891 bitmap_zero(map, page->objects);
3893 get_map(s, page, map);
3894 for_each_object(p, s, addr, page->objects) {
3895 if (test_bit(slab_index(p, s, addr), map))
3896 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3900 for_each_object(p, s, addr, page->objects)
3901 if (!test_bit(slab_index(p, s, addr), map))
3902 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3907 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3911 validate_slab(s, page, map);
3915 static int validate_slab_node(struct kmem_cache *s,
3916 struct kmem_cache_node *n, unsigned long *map)
3918 unsigned long count = 0;
3920 unsigned long flags;
3922 spin_lock_irqsave(&n->list_lock, flags);
3924 list_for_each_entry(page, &n->partial, lru) {
3925 validate_slab_slab(s, page, map);
3928 if (count != n->nr_partial)
3929 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3930 "counter=%ld\n", s->name, count, n->nr_partial);
3932 if (!(s->flags & SLAB_STORE_USER))
3935 list_for_each_entry(page, &n->full, lru) {
3936 validate_slab_slab(s, page, map);
3939 if (count != atomic_long_read(&n->nr_slabs))
3940 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3941 "counter=%ld\n", s->name, count,
3942 atomic_long_read(&n->nr_slabs));
3945 spin_unlock_irqrestore(&n->list_lock, flags);
3949 static long validate_slab_cache(struct kmem_cache *s)
3952 unsigned long count = 0;
3953 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3954 sizeof(unsigned long), GFP_KERNEL);
3960 for_each_node_state(node, N_NORMAL_MEMORY) {
3961 struct kmem_cache_node *n = get_node(s, node);
3963 count += validate_slab_node(s, n, map);
3969 * Generate lists of code addresses where slabcache objects are allocated
3974 unsigned long count;
3981 DECLARE_BITMAP(cpus, NR_CPUS);
3987 unsigned long count;
3988 struct location *loc;
3991 static void free_loc_track(struct loc_track *t)
3994 #ifndef CONFIG_SPRD_PAGERECORDER
3995 free_pages((unsigned long)t->loc,
3996 get_order(sizeof(struct location) * t->max));
3998 __free_pages_nopagedebug((unsigned long)t->loc,
3999 get_order(sizeof(struct location) * t->max));
4003 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4008 order = get_order(sizeof(struct location) * max);
4010 #ifndef CONFIG_SPRD_PAGERECORDER
4011 l = (void *)__get_free_pages(flags, order);
4013 l = (void *)__get_free_pages_nopagedebug(flags, order);
4020 memcpy(l, t->loc, sizeof(struct location) * t->count);
4028 static int add_location(struct loc_track *t, struct kmem_cache *s,
4029 const struct track *track)
4031 long start, end, pos;
4033 unsigned long caddr;
4034 unsigned long age = jiffies - track->when;
4040 pos = start + (end - start + 1) / 2;
4043 * There is nothing at "end". If we end up there
4044 * we need to add something to before end.
4049 caddr = t->loc[pos].addr;
4050 if (track->addr == caddr) {
4056 if (age < l->min_time)
4058 if (age > l->max_time)
4061 if (track->pid < l->min_pid)
4062 l->min_pid = track->pid;
4063 if (track->pid > l->max_pid)
4064 l->max_pid = track->pid;
4066 cpumask_set_cpu(track->cpu,
4067 to_cpumask(l->cpus));
4069 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4073 if (track->addr < caddr)
4080 * Not found. Insert new tracking element.
4082 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4088 (t->count - pos) * sizeof(struct location));
4091 l->addr = track->addr;
4095 l->min_pid = track->pid;
4096 l->max_pid = track->pid;
4097 cpumask_clear(to_cpumask(l->cpus));
4098 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4099 nodes_clear(l->nodes);
4100 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4104 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4105 struct page *page, enum track_item alloc,
4108 void *addr = page_address(page);
4111 bitmap_zero(map, page->objects);
4112 get_map(s, page, map);
4114 for_each_object(p, s, addr, page->objects)
4115 if (!test_bit(slab_index(p, s, addr), map))
4116 add_location(t, s, get_track(s, p, alloc));
4119 static int list_locations(struct kmem_cache *s, char *buf,
4120 enum track_item alloc)
4124 struct loc_track t = { 0, 0, NULL };
4126 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4127 sizeof(unsigned long), GFP_KERNEL);
4129 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4132 return sprintf(buf, "Out of memory\n");
4134 /* Push back cpu slabs */
4137 for_each_node_state(node, N_NORMAL_MEMORY) {
4138 struct kmem_cache_node *n = get_node(s, node);
4139 unsigned long flags;
4142 if (!atomic_long_read(&n->nr_slabs))
4145 spin_lock_irqsave(&n->list_lock, flags);
4146 list_for_each_entry(page, &n->partial, lru)
4147 process_slab(&t, s, page, alloc, map);
4148 list_for_each_entry(page, &n->full, lru)
4149 process_slab(&t, s, page, alloc, map);
4150 spin_unlock_irqrestore(&n->list_lock, flags);
4153 for (i = 0; i < t.count; i++) {
4154 struct location *l = &t.loc[i];
4156 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4158 len += sprintf(buf + len, "%7ld ", l->count);
4161 len += sprintf(buf + len, "%pS", (void *)l->addr);
4163 len += sprintf(buf + len, "<not-available>");
4165 if (l->sum_time != l->min_time) {
4166 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4168 (long)div_u64(l->sum_time, l->count),
4171 len += sprintf(buf + len, " age=%ld",
4174 if (l->min_pid != l->max_pid)
4175 len += sprintf(buf + len, " pid=%ld-%ld",
4176 l->min_pid, l->max_pid);
4178 len += sprintf(buf + len, " pid=%ld",
4181 if (num_online_cpus() > 1 &&
4182 !cpumask_empty(to_cpumask(l->cpus)) &&
4183 len < PAGE_SIZE - 60) {
4184 len += sprintf(buf + len, " cpus=");
4185 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4186 to_cpumask(l->cpus));
4189 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4190 len < PAGE_SIZE - 60) {
4191 len += sprintf(buf + len, " nodes=");
4192 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4196 len += sprintf(buf + len, "\n");
4202 len += sprintf(buf, "No data\n");
4207 #ifdef SLUB_RESILIENCY_TEST
4208 static void resiliency_test(void)
4212 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4214 printk(KERN_ERR "SLUB resiliency testing\n");
4215 printk(KERN_ERR "-----------------------\n");
4216 printk(KERN_ERR "A. Corruption after allocation\n");
4218 p = kzalloc(16, GFP_KERNEL);
4220 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4221 " 0x12->0x%p\n\n", p + 16);
4223 validate_slab_cache(kmalloc_caches[4]);
4225 /* Hmmm... The next two are dangerous */
4226 p = kzalloc(32, GFP_KERNEL);
4227 p[32 + sizeof(void *)] = 0x34;
4228 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4229 " 0x34 -> -0x%p\n", p);
4231 "If allocated object is overwritten then not detectable\n\n");
4233 validate_slab_cache(kmalloc_caches[5]);
4234 p = kzalloc(64, GFP_KERNEL);
4235 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4237 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4240 "If allocated object is overwritten then not detectable\n\n");
4241 validate_slab_cache(kmalloc_caches[6]);
4243 printk(KERN_ERR "\nB. Corruption after free\n");
4244 p = kzalloc(128, GFP_KERNEL);
4247 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4248 validate_slab_cache(kmalloc_caches[7]);
4250 p = kzalloc(256, GFP_KERNEL);
4253 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4255 validate_slab_cache(kmalloc_caches[8]);
4257 p = kzalloc(512, GFP_KERNEL);
4260 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4261 validate_slab_cache(kmalloc_caches[9]);
4265 static void resiliency_test(void) {};
4270 enum slab_stat_type {
4271 SL_ALL, /* All slabs */
4272 SL_PARTIAL, /* Only partially allocated slabs */
4273 SL_CPU, /* Only slabs used for cpu caches */
4274 SL_OBJECTS, /* Determine allocated objects not slabs */
4275 SL_TOTAL /* Determine object capacity not slabs */
4278 #define SO_ALL (1 << SL_ALL)
4279 #define SO_PARTIAL (1 << SL_PARTIAL)
4280 #define SO_CPU (1 << SL_CPU)
4281 #define SO_OBJECTS (1 << SL_OBJECTS)
4282 #define SO_TOTAL (1 << SL_TOTAL)
4284 static ssize_t show_slab_objects(struct kmem_cache *s,
4285 char *buf, unsigned long flags)
4287 unsigned long total = 0;
4290 unsigned long *nodes;
4291 unsigned long *per_cpu;
4293 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4296 per_cpu = nodes + nr_node_ids;
4298 if (flags & SO_CPU) {
4301 for_each_possible_cpu(cpu) {
4302 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4306 page = ACCESS_ONCE(c->page);
4310 node = page_to_nid(page);
4311 if (flags & SO_TOTAL)
4313 else if (flags & SO_OBJECTS)
4321 page = ACCESS_ONCE(c->partial);
4323 node = page_to_nid(page);
4324 if (flags & SO_TOTAL)
4326 else if (flags & SO_OBJECTS)
4338 lock_memory_hotplug();
4339 #ifdef CONFIG_SLUB_DEBUG
4340 if (flags & SO_ALL) {
4341 for_each_node_state(node, N_NORMAL_MEMORY) {
4342 struct kmem_cache_node *n = get_node(s, node);
4344 if (flags & SO_TOTAL)
4345 x = atomic_long_read(&n->total_objects);
4346 else if (flags & SO_OBJECTS)
4347 x = atomic_long_read(&n->total_objects) -
4348 count_partial(n, count_free);
4351 x = atomic_long_read(&n->nr_slabs);
4358 if (flags & SO_PARTIAL) {
4359 for_each_node_state(node, N_NORMAL_MEMORY) {
4360 struct kmem_cache_node *n = get_node(s, node);
4362 if (flags & SO_TOTAL)
4363 x = count_partial(n, count_total);
4364 else if (flags & SO_OBJECTS)
4365 x = count_partial(n, count_inuse);
4372 x = sprintf(buf, "%lu", total);
4374 for_each_node_state(node, N_NORMAL_MEMORY)
4376 x += sprintf(buf + x, " N%d=%lu",
4379 unlock_memory_hotplug();
4381 return x + sprintf(buf + x, "\n");
4384 #ifdef CONFIG_SLUB_DEBUG
4385 static int any_slab_objects(struct kmem_cache *s)
4389 for_each_online_node(node) {
4390 struct kmem_cache_node *n = get_node(s, node);
4395 if (atomic_long_read(&n->total_objects))
4402 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4403 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4405 struct slab_attribute {
4406 struct attribute attr;
4407 ssize_t (*show)(struct kmem_cache *s, char *buf);
4408 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4411 #define SLAB_ATTR_RO(_name) \
4412 static struct slab_attribute _name##_attr = \
4413 __ATTR(_name, 0400, _name##_show, NULL)
4415 #define SLAB_ATTR(_name) \
4416 static struct slab_attribute _name##_attr = \
4417 __ATTR(_name, 0600, _name##_show, _name##_store)
4419 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4421 return sprintf(buf, "%d\n", s->size);
4423 SLAB_ATTR_RO(slab_size);
4425 static ssize_t align_show(struct kmem_cache *s, char *buf)
4427 return sprintf(buf, "%d\n", s->align);
4429 SLAB_ATTR_RO(align);
4431 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4433 return sprintf(buf, "%d\n", s->object_size);
4435 SLAB_ATTR_RO(object_size);
4437 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4439 return sprintf(buf, "%d\n", oo_objects(s->oo));
4441 SLAB_ATTR_RO(objs_per_slab);
4443 static ssize_t order_store(struct kmem_cache *s,
4444 const char *buf, size_t length)
4446 unsigned long order;
4449 err = strict_strtoul(buf, 10, &order);
4453 if (order > slub_max_order || order < slub_min_order)
4456 calculate_sizes(s, order);
4460 static ssize_t order_show(struct kmem_cache *s, char *buf)
4462 return sprintf(buf, "%d\n", oo_order(s->oo));
4466 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4468 return sprintf(buf, "%lu\n", s->min_partial);
4471 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4477 err = strict_strtoul(buf, 10, &min);
4481 set_min_partial(s, min);
4484 SLAB_ATTR(min_partial);
4486 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4488 return sprintf(buf, "%u\n", s->cpu_partial);
4491 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4494 unsigned long objects;
4497 err = strict_strtoul(buf, 10, &objects);
4500 if (objects && kmem_cache_debug(s))
4503 s->cpu_partial = objects;
4507 SLAB_ATTR(cpu_partial);
4509 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4513 return sprintf(buf, "%pS\n", s->ctor);
4517 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4519 return sprintf(buf, "%d\n", s->refcount - 1);
4521 SLAB_ATTR_RO(aliases);
4523 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4525 return show_slab_objects(s, buf, SO_PARTIAL);
4527 SLAB_ATTR_RO(partial);
4529 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4531 return show_slab_objects(s, buf, SO_CPU);
4533 SLAB_ATTR_RO(cpu_slabs);
4535 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4537 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4539 SLAB_ATTR_RO(objects);
4541 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4543 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4545 SLAB_ATTR_RO(objects_partial);
4547 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4554 for_each_online_cpu(cpu) {
4555 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4558 pages += page->pages;
4559 objects += page->pobjects;
4563 len = sprintf(buf, "%d(%d)", objects, pages);
4566 for_each_online_cpu(cpu) {
4567 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4569 if (page && len < PAGE_SIZE - 20)
4570 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4571 page->pobjects, page->pages);
4574 return len + sprintf(buf + len, "\n");
4576 SLAB_ATTR_RO(slabs_cpu_partial);
4578 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4580 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4583 static ssize_t reclaim_account_store(struct kmem_cache *s,
4584 const char *buf, size_t length)
4586 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4588 s->flags |= SLAB_RECLAIM_ACCOUNT;
4591 SLAB_ATTR(reclaim_account);
4593 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4595 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4597 SLAB_ATTR_RO(hwcache_align);
4599 #ifdef CONFIG_ZONE_DMA
4600 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4602 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4604 SLAB_ATTR_RO(cache_dma);
4607 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4609 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4611 SLAB_ATTR_RO(destroy_by_rcu);
4613 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4615 return sprintf(buf, "%d\n", s->reserved);
4617 SLAB_ATTR_RO(reserved);
4619 #ifdef CONFIG_SLUB_DEBUG
4620 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4622 return show_slab_objects(s, buf, SO_ALL);
4624 SLAB_ATTR_RO(slabs);
4626 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4628 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4630 SLAB_ATTR_RO(total_objects);
4632 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4634 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4637 static ssize_t sanity_checks_store(struct kmem_cache *s,
4638 const char *buf, size_t length)
4640 s->flags &= ~SLAB_DEBUG_FREE;
4641 if (buf[0] == '1') {
4642 s->flags &= ~__CMPXCHG_DOUBLE;
4643 s->flags |= SLAB_DEBUG_FREE;
4647 SLAB_ATTR(sanity_checks);
4649 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4651 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4654 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4657 s->flags &= ~SLAB_TRACE;
4658 if (buf[0] == '1') {
4659 s->flags &= ~__CMPXCHG_DOUBLE;
4660 s->flags |= SLAB_TRACE;
4666 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4668 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4671 static ssize_t red_zone_store(struct kmem_cache *s,
4672 const char *buf, size_t length)
4674 if (any_slab_objects(s))
4677 s->flags &= ~SLAB_RED_ZONE;
4678 if (buf[0] == '1') {
4679 s->flags &= ~__CMPXCHG_DOUBLE;
4680 s->flags |= SLAB_RED_ZONE;
4682 calculate_sizes(s, -1);
4685 SLAB_ATTR(red_zone);
4687 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4689 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4692 static ssize_t poison_store(struct kmem_cache *s,
4693 const char *buf, size_t length)
4695 if (any_slab_objects(s))
4698 s->flags &= ~SLAB_POISON;
4699 if (buf[0] == '1') {
4700 s->flags &= ~__CMPXCHG_DOUBLE;
4701 s->flags |= SLAB_POISON;
4703 calculate_sizes(s, -1);
4708 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4710 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4713 static ssize_t store_user_store(struct kmem_cache *s,
4714 const char *buf, size_t length)
4716 if (any_slab_objects(s))
4719 s->flags &= ~SLAB_STORE_USER;
4720 if (buf[0] == '1') {
4721 s->flags &= ~__CMPXCHG_DOUBLE;
4722 s->flags |= SLAB_STORE_USER;
4724 calculate_sizes(s, -1);
4727 SLAB_ATTR(store_user);
4729 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4734 static ssize_t validate_store(struct kmem_cache *s,
4735 const char *buf, size_t length)
4739 if (buf[0] == '1') {
4740 ret = validate_slab_cache(s);
4746 SLAB_ATTR(validate);
4748 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4750 if (!(s->flags & SLAB_STORE_USER))
4752 return list_locations(s, buf, TRACK_ALLOC);
4754 SLAB_ATTR_RO(alloc_calls);
4756 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4758 if (!(s->flags & SLAB_STORE_USER))
4760 return list_locations(s, buf, TRACK_FREE);
4762 SLAB_ATTR_RO(free_calls);
4763 #endif /* CONFIG_SLUB_DEBUG */
4765 #ifdef CONFIG_FAILSLAB
4766 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4768 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4771 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4774 s->flags &= ~SLAB_FAILSLAB;
4776 s->flags |= SLAB_FAILSLAB;
4779 SLAB_ATTR(failslab);
4782 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4787 static ssize_t shrink_store(struct kmem_cache *s,
4788 const char *buf, size_t length)
4790 if (buf[0] == '1') {
4791 int rc = kmem_cache_shrink(s);
4802 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4804 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4807 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4808 const char *buf, size_t length)
4810 unsigned long ratio;
4813 err = strict_strtoul(buf, 10, &ratio);
4818 s->remote_node_defrag_ratio = ratio * 10;
4822 SLAB_ATTR(remote_node_defrag_ratio);
4825 #ifdef CONFIG_SLUB_STATS
4826 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4828 unsigned long sum = 0;
4831 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4836 for_each_online_cpu(cpu) {
4837 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4843 len = sprintf(buf, "%lu", sum);
4846 for_each_online_cpu(cpu) {
4847 if (data[cpu] && len < PAGE_SIZE - 20)
4848 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4852 return len + sprintf(buf + len, "\n");
4855 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4859 for_each_online_cpu(cpu)
4860 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4863 #define STAT_ATTR(si, text) \
4864 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4866 return show_stat(s, buf, si); \
4868 static ssize_t text##_store(struct kmem_cache *s, \
4869 const char *buf, size_t length) \
4871 if (buf[0] != '0') \
4873 clear_stat(s, si); \
4878 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4879 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4880 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4881 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4882 STAT_ATTR(FREE_FROZEN, free_frozen);
4883 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4884 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4885 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4886 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4887 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4888 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4889 STAT_ATTR(FREE_SLAB, free_slab);
4890 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4891 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4892 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4893 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4894 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4895 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4896 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4897 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4898 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4899 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4900 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4901 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4902 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4903 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4906 static struct attribute *slab_attrs[] = {
4907 &slab_size_attr.attr,
4908 &object_size_attr.attr,
4909 &objs_per_slab_attr.attr,
4911 &min_partial_attr.attr,
4912 &cpu_partial_attr.attr,
4914 &objects_partial_attr.attr,
4916 &cpu_slabs_attr.attr,
4920 &hwcache_align_attr.attr,
4921 &reclaim_account_attr.attr,
4922 &destroy_by_rcu_attr.attr,
4924 &reserved_attr.attr,
4925 &slabs_cpu_partial_attr.attr,
4926 #ifdef CONFIG_SLUB_DEBUG
4927 &total_objects_attr.attr,
4929 &sanity_checks_attr.attr,
4931 &red_zone_attr.attr,
4933 &store_user_attr.attr,
4934 &validate_attr.attr,
4935 &alloc_calls_attr.attr,
4936 &free_calls_attr.attr,
4938 #ifdef CONFIG_ZONE_DMA
4939 &cache_dma_attr.attr,
4942 &remote_node_defrag_ratio_attr.attr,
4944 #ifdef CONFIG_SLUB_STATS
4945 &alloc_fastpath_attr.attr,
4946 &alloc_slowpath_attr.attr,
4947 &free_fastpath_attr.attr,
4948 &free_slowpath_attr.attr,
4949 &free_frozen_attr.attr,
4950 &free_add_partial_attr.attr,
4951 &free_remove_partial_attr.attr,
4952 &alloc_from_partial_attr.attr,
4953 &alloc_slab_attr.attr,
4954 &alloc_refill_attr.attr,
4955 &alloc_node_mismatch_attr.attr,
4956 &free_slab_attr.attr,
4957 &cpuslab_flush_attr.attr,
4958 &deactivate_full_attr.attr,
4959 &deactivate_empty_attr.attr,
4960 &deactivate_to_head_attr.attr,
4961 &deactivate_to_tail_attr.attr,
4962 &deactivate_remote_frees_attr.attr,
4963 &deactivate_bypass_attr.attr,
4964 &order_fallback_attr.attr,
4965 &cmpxchg_double_fail_attr.attr,
4966 &cmpxchg_double_cpu_fail_attr.attr,
4967 &cpu_partial_alloc_attr.attr,
4968 &cpu_partial_free_attr.attr,
4969 &cpu_partial_node_attr.attr,
4970 &cpu_partial_drain_attr.attr,
4972 #ifdef CONFIG_FAILSLAB
4973 &failslab_attr.attr,
4979 static struct attribute_group slab_attr_group = {
4980 .attrs = slab_attrs,
4983 static ssize_t slab_attr_show(struct kobject *kobj,
4984 struct attribute *attr,
4987 struct slab_attribute *attribute;
4988 struct kmem_cache *s;
4991 attribute = to_slab_attr(attr);
4994 if (!attribute->show)
4997 err = attribute->show(s, buf);
5002 static ssize_t slab_attr_store(struct kobject *kobj,
5003 struct attribute *attr,
5004 const char *buf, size_t len)
5006 struct slab_attribute *attribute;
5007 struct kmem_cache *s;
5010 attribute = to_slab_attr(attr);
5013 if (!attribute->store)
5016 err = attribute->store(s, buf, len);
5017 #ifdef CONFIG_MEMCG_KMEM
5018 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5021 mutex_lock(&slab_mutex);
5022 if (s->max_attr_size < len)
5023 s->max_attr_size = len;
5026 * This is a best effort propagation, so this function's return
5027 * value will be determined by the parent cache only. This is
5028 * basically because not all attributes will have a well
5029 * defined semantics for rollbacks - most of the actions will
5030 * have permanent effects.
5032 * Returning the error value of any of the children that fail
5033 * is not 100 % defined, in the sense that users seeing the
5034 * error code won't be able to know anything about the state of
5037 * Only returning the error code for the parent cache at least
5038 * has well defined semantics. The cache being written to
5039 * directly either failed or succeeded, in which case we loop
5040 * through the descendants with best-effort propagation.
5042 for_each_memcg_cache_index(i) {
5043 struct kmem_cache *c = cache_from_memcg(s, i);
5045 attribute->store(c, buf, len);
5047 mutex_unlock(&slab_mutex);
5053 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5055 #ifdef CONFIG_MEMCG_KMEM
5057 char *buffer = NULL;
5059 if (!is_root_cache(s))
5063 * This mean this cache had no attribute written. Therefore, no point
5064 * in copying default values around
5066 if (!s->max_attr_size)
5069 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5072 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5074 if (!attr || !attr->store || !attr->show)
5078 * It is really bad that we have to allocate here, so we will
5079 * do it only as a fallback. If we actually allocate, though,
5080 * we can just use the allocated buffer until the end.
5082 * Most of the slub attributes will tend to be very small in
5083 * size, but sysfs allows buffers up to a page, so they can
5084 * theoretically happen.
5088 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5091 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5092 if (WARN_ON(!buffer))
5097 attr->show(s->memcg_params->root_cache, buf);
5098 attr->store(s, buf, strlen(buf));
5102 free_page((unsigned long)buffer);
5106 static const struct sysfs_ops slab_sysfs_ops = {
5107 .show = slab_attr_show,
5108 .store = slab_attr_store,
5111 static struct kobj_type slab_ktype = {
5112 .sysfs_ops = &slab_sysfs_ops,
5115 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5117 struct kobj_type *ktype = get_ktype(kobj);
5119 if (ktype == &slab_ktype)
5124 static const struct kset_uevent_ops slab_uevent_ops = {
5125 .filter = uevent_filter,
5128 static struct kset *slab_kset;
5130 #define ID_STR_LENGTH 64
5132 /* Create a unique string id for a slab cache:
5134 * Format :[flags-]size
5136 static char *create_unique_id(struct kmem_cache *s)
5138 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5145 * First flags affecting slabcache operations. We will only
5146 * get here for aliasable slabs so we do not need to support
5147 * too many flags. The flags here must cover all flags that
5148 * are matched during merging to guarantee that the id is
5151 if (s->flags & SLAB_CACHE_DMA)
5153 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5155 if (s->flags & SLAB_DEBUG_FREE)
5157 if (!(s->flags & SLAB_NOTRACK))
5161 p += sprintf(p, "%07d", s->size);
5163 #ifdef CONFIG_MEMCG_KMEM
5164 if (!is_root_cache(s))
5165 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5168 BUG_ON(p > name + ID_STR_LENGTH - 1);
5172 static int sysfs_slab_add(struct kmem_cache *s)
5176 int unmergeable = slab_unmergeable(s);
5180 * Slabcache can never be merged so we can use the name proper.
5181 * This is typically the case for debug situations. In that
5182 * case we can catch duplicate names easily.
5184 sysfs_remove_link(&slab_kset->kobj, s->name);
5188 * Create a unique name for the slab as a target
5191 name = create_unique_id(s);
5194 s->kobj.kset = slab_kset;
5195 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5197 kobject_put(&s->kobj);
5201 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5203 kobject_del(&s->kobj);
5204 kobject_put(&s->kobj);
5207 kobject_uevent(&s->kobj, KOBJ_ADD);
5209 /* Setup first alias */
5210 sysfs_slab_alias(s, s->name);
5216 static void sysfs_slab_remove(struct kmem_cache *s)
5218 if (slab_state < FULL)
5220 * Sysfs has not been setup yet so no need to remove the
5225 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5226 kobject_del(&s->kobj);
5227 kobject_put(&s->kobj);
5231 * Need to buffer aliases during bootup until sysfs becomes
5232 * available lest we lose that information.
5234 struct saved_alias {
5235 struct kmem_cache *s;
5237 struct saved_alias *next;
5240 static struct saved_alias *alias_list;
5242 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5244 struct saved_alias *al;
5246 if (slab_state == FULL) {
5248 * If we have a leftover link then remove it.
5250 sysfs_remove_link(&slab_kset->kobj, name);
5251 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5254 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5260 al->next = alias_list;
5265 static int __init slab_sysfs_init(void)
5267 struct kmem_cache *s;
5270 mutex_lock(&slab_mutex);
5272 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5274 mutex_unlock(&slab_mutex);
5275 printk(KERN_ERR "Cannot register slab subsystem.\n");
5281 list_for_each_entry(s, &slab_caches, list) {
5282 err = sysfs_slab_add(s);
5284 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5285 " to sysfs\n", s->name);
5288 while (alias_list) {
5289 struct saved_alias *al = alias_list;
5291 alias_list = alias_list->next;
5292 err = sysfs_slab_alias(al->s, al->name);
5294 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5295 " %s to sysfs\n", al->name);
5299 mutex_unlock(&slab_mutex);
5304 __initcall(slab_sysfs_init);
5305 #endif /* CONFIG_SYSFS */
5308 * The /proc/slabinfo ABI
5310 #ifdef CONFIG_SLABINFO
5311 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5313 unsigned long nr_partials = 0;
5314 unsigned long nr_slabs = 0;
5315 unsigned long nr_objs = 0;
5316 unsigned long nr_free = 0;
5319 for_each_online_node(node) {
5320 struct kmem_cache_node *n = get_node(s, node);
5325 nr_partials += n->nr_partial;
5326 nr_slabs += atomic_long_read(&n->nr_slabs);
5327 nr_objs += atomic_long_read(&n->total_objects);
5328 nr_free += count_partial(n, count_free);
5331 sinfo->active_objs = nr_objs - nr_free;
5332 sinfo->num_objs = nr_objs;
5333 sinfo->active_slabs = nr_slabs;
5334 sinfo->num_slabs = nr_slabs;
5335 sinfo->objects_per_slab = oo_objects(s->oo);
5336 sinfo->cache_order = oo_order(s->oo);
5339 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5343 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5344 size_t count, loff_t *ppos)
5348 #endif /* CONFIG_SLABINFO */