2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
34 #include <linux/memcontrol.h>
36 #include <trace/events/kmem.h>
42 * 1. slab_mutex (Global Mutex)
44 * 3. slab_lock(page) (Only on some arches and for debugging)
48 * The role of the slab_mutex is to protect the list of all the slabs
49 * and to synchronize major metadata changes to slab cache structures.
51 * The slab_lock is only used for debugging and on arches that do not
52 * have the ability to do a cmpxchg_double. It only protects the second
53 * double word in the page struct. Meaning
54 * A. page->freelist -> List of object free in a page
55 * B. page->counters -> Counters of objects
56 * C. page->frozen -> frozen state
58 * If a slab is frozen then it is exempt from list management. It is not
59 * on any list. The processor that froze the slab is the one who can
60 * perform list operations on the page. Other processors may put objects
61 * onto the freelist but the processor that froze the slab is the only
62 * one that can retrieve the objects from the page's freelist.
64 * The list_lock protects the partial and full list on each node and
65 * the partial slab counter. If taken then no new slabs may be added or
66 * removed from the lists nor make the number of partial slabs be modified.
67 * (Note that the total number of slabs is an atomic value that may be
68 * modified without taking the list lock).
70 * The list_lock is a centralized lock and thus we avoid taking it as
71 * much as possible. As long as SLUB does not have to handle partial
72 * slabs, operations can continue without any centralized lock. F.e.
73 * allocating a long series of objects that fill up slabs does not require
75 * Interrupts are disabled during allocation and deallocation in order to
76 * make the slab allocator safe to use in the context of an irq. In addition
77 * interrupts are disabled to ensure that the processor does not change
78 * while handling per_cpu slabs, due to kernel preemption.
80 * SLUB assigns one slab for allocation to each processor.
81 * Allocations only occur from these slabs called cpu slabs.
83 * Slabs with free elements are kept on a partial list and during regular
84 * operations no list for full slabs is used. If an object in a full slab is
85 * freed then the slab will show up again on the partial lists.
86 * We track full slabs for debugging purposes though because otherwise we
87 * cannot scan all objects.
89 * Slabs are freed when they become empty. Teardown and setup is
90 * minimal so we rely on the page allocators per cpu caches for
91 * fast frees and allocs.
93 * Overloading of page flags that are otherwise used for LRU management.
95 * PageActive The slab is frozen and exempt from list processing.
96 * This means that the slab is dedicated to a purpose
97 * such as satisfying allocations for a specific
98 * processor. Objects may be freed in the slab while
99 * it is frozen but slab_free will then skip the usual
100 * list operations. It is up to the processor holding
101 * the slab to integrate the slab into the slab lists
102 * when the slab is no longer needed.
104 * One use of this flag is to mark slabs that are
105 * used for allocations. Then such a slab becomes a cpu
106 * slab. The cpu slab may be equipped with an additional
107 * freelist that allows lockless access to
108 * free objects in addition to the regular freelist
109 * that requires the slab lock.
111 * PageError Slab requires special handling due to debug
112 * options set. This moves slab handling out of
113 * the fast path and disables lockless freelists.
116 static inline int kmem_cache_debug(struct kmem_cache *s)
118 #ifdef CONFIG_SLUB_DEBUG
119 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
125 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127 #ifdef CONFIG_SLUB_CPU_PARTIAL
128 return !kmem_cache_debug(s);
135 * Issues still to be resolved:
137 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139 * - Variable sizing of the per node arrays
142 /* Enable to test recovery from slab corruption on boot */
143 #undef SLUB_RESILIENCY_TEST
145 /* Enable to log cmpxchg failures */
146 #undef SLUB_DEBUG_CMPXCHG
149 * Mininum number of partial slabs. These will be left on the partial
150 * lists even if they are empty. kmem_cache_shrink may reclaim them.
152 #define MIN_PARTIAL 5
155 * Maximum number of desirable partial slabs.
156 * The existence of more partial slabs makes kmem_cache_shrink
157 * sort the partial list by the number of objects in the.
159 #define MAX_PARTIAL 10
161 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
162 SLAB_POISON | SLAB_STORE_USER)
165 * Debugging flags that require metadata to be stored in the slab. These get
166 * disabled when slub_debug=O is used and a cache's min order increases with
169 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
172 * Set of flags that will prevent slab merging
174 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
175 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
178 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
179 SLAB_CACHE_DMA | SLAB_NOTRACK)
182 #define OO_MASK ((1 << OO_SHIFT) - 1)
183 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
185 /* Internal SLUB flags */
186 #define __OBJECT_POISON 0x80000000UL /* Poison object */
187 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
190 static struct notifier_block slab_notifier;
194 * Tracking user of a slab.
196 #define TRACK_ADDRS_COUNT 16
198 unsigned long addr; /* Called from address */
199 #ifdef CONFIG_STACKTRACE
200 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202 int cpu; /* Was running on cpu */
203 int pid; /* Pid context */
204 unsigned long when; /* When did the operation occur */
207 enum track_item { TRACK_ALLOC, TRACK_FREE };
210 static int sysfs_slab_add(struct kmem_cache *);
211 static int sysfs_slab_alias(struct kmem_cache *, const char *);
212 static void sysfs_slab_remove(struct kmem_cache *);
213 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
215 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
216 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
218 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
220 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
223 static inline void stat(const struct kmem_cache *s, enum stat_item si)
225 #ifdef CONFIG_SLUB_STATS
226 __this_cpu_inc(s->cpu_slab->stat[si]);
230 /********************************************************************
231 * Core slab cache functions
232 *******************************************************************/
234 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
236 return s->node[node];
239 /* Verify that a pointer has an address that is valid within a slab page */
240 static inline int check_valid_pointer(struct kmem_cache *s,
241 struct page *page, const void *object)
248 base = page_address(page);
249 if (object < base || object >= base + page->objects * s->size ||
250 (object - base) % s->size) {
257 static inline void *get_freepointer(struct kmem_cache *s, void *object)
259 return *(void **)(object + s->offset);
262 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
264 prefetch(object + s->offset);
267 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
271 #ifdef CONFIG_DEBUG_PAGEALLOC
272 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
274 p = get_freepointer(s, object);
279 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
281 *(void **)(object + s->offset) = fp;
284 /* Loop over all objects in a slab */
285 #define for_each_object(__p, __s, __addr, __objects) \
286 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
289 /* Determine object index from a given position */
290 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292 return (p - addr) / s->size;
295 static inline size_t slab_ksize(const struct kmem_cache *s)
297 #ifdef CONFIG_SLUB_DEBUG
299 * Debugging requires use of the padding between object
300 * and whatever may come after it.
302 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
303 return s->object_size;
307 * If we have the need to store the freelist pointer
308 * back there or track user information then we can
309 * only use the space before that information.
311 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
314 * Else we can use all the padding etc for the allocation
319 static inline int order_objects(int order, unsigned long size, int reserved)
321 return ((PAGE_SIZE << order) - reserved) / size;
324 static inline struct kmem_cache_order_objects oo_make(int order,
325 unsigned long size, int reserved)
327 struct kmem_cache_order_objects x = {
328 (order << OO_SHIFT) + order_objects(order, size, reserved)
334 static inline int oo_order(struct kmem_cache_order_objects x)
336 return x.x >> OO_SHIFT;
339 static inline int oo_objects(struct kmem_cache_order_objects x)
341 return x.x & OO_MASK;
345 * Per slab locking using the pagelock
347 static __always_inline void slab_lock(struct page *page)
349 bit_spin_lock(PG_locked, &page->flags);
352 static __always_inline void slab_unlock(struct page *page)
354 __bit_spin_unlock(PG_locked, &page->flags);
357 /* Interrupts must be disabled (for the fallback code to work right) */
358 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
359 void *freelist_old, unsigned long counters_old,
360 void *freelist_new, unsigned long counters_new,
363 VM_BUG_ON(!irqs_disabled());
364 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
365 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
366 if (s->flags & __CMPXCHG_DOUBLE) {
367 if (cmpxchg_double(&page->freelist, &page->counters,
368 freelist_old, counters_old,
369 freelist_new, counters_new))
375 if (page->freelist == freelist_old && page->counters == counters_old) {
376 page->freelist = freelist_new;
377 page->counters = counters_new;
385 stat(s, CMPXCHG_DOUBLE_FAIL);
387 #ifdef SLUB_DEBUG_CMPXCHG
388 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
394 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
395 void *freelist_old, unsigned long counters_old,
396 void *freelist_new, unsigned long counters_new,
399 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
400 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
401 if (s->flags & __CMPXCHG_DOUBLE) {
402 if (cmpxchg_double(&page->freelist, &page->counters,
403 freelist_old, counters_old,
404 freelist_new, counters_new))
411 local_irq_save(flags);
413 if (page->freelist == freelist_old && page->counters == counters_old) {
414 page->freelist = freelist_new;
415 page->counters = counters_new;
417 local_irq_restore(flags);
421 local_irq_restore(flags);
425 stat(s, CMPXCHG_DOUBLE_FAIL);
427 #ifdef SLUB_DEBUG_CMPXCHG
428 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
434 #ifdef CONFIG_SLUB_DEBUG
436 * Determine a map of object in use on a page.
438 * Node listlock must be held to guarantee that the page does
439 * not vanish from under us.
441 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
444 void *addr = page_address(page);
446 for (p = page->freelist; p; p = get_freepointer(s, p))
447 set_bit(slab_index(p, s, addr), map);
453 #ifdef CONFIG_SLUB_DEBUG_ON
454 static int slub_debug = DEBUG_DEFAULT_FLAGS;
456 static int slub_debug;
459 static char *slub_debug_slabs;
460 static int disable_higher_order_debug;
465 static void print_section(char *text, u8 *addr, unsigned int length)
467 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
471 static struct track *get_track(struct kmem_cache *s, void *object,
472 enum track_item alloc)
477 p = object + s->offset + sizeof(void *);
479 p = object + s->inuse;
484 static void set_track(struct kmem_cache *s, void *object,
485 enum track_item alloc, unsigned long addr)
487 struct track *p = get_track(s, object, alloc);
490 #ifdef CONFIG_STACKTRACE
491 struct stack_trace trace;
494 trace.nr_entries = 0;
495 trace.max_entries = TRACK_ADDRS_COUNT;
496 trace.entries = p->addrs;
498 save_stack_trace(&trace);
500 /* See rant in lockdep.c */
501 if (trace.nr_entries != 0 &&
502 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
505 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
509 p->cpu = smp_processor_id();
510 p->pid = current->pid;
513 memset(p, 0, sizeof(struct track));
516 static void init_tracking(struct kmem_cache *s, void *object)
518 if (!(s->flags & SLAB_STORE_USER))
521 set_track(s, object, TRACK_FREE, 0UL);
522 set_track(s, object, TRACK_ALLOC, 0UL);
525 static void print_track(const char *s, struct track *t)
530 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
531 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
532 #ifdef CONFIG_STACKTRACE
535 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
537 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
544 static void print_tracking(struct kmem_cache *s, void *object)
546 if (!(s->flags & SLAB_STORE_USER))
549 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
550 print_track("Freed", get_track(s, object, TRACK_FREE));
553 static void print_page_info(struct page *page)
555 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
556 page, page->objects, page->inuse, page->freelist, page->flags);
560 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
566 vsnprintf(buf, sizeof(buf), fmt, args);
568 printk(KERN_ERR "========================================"
569 "=====================================\n");
570 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
571 printk(KERN_ERR "----------------------------------------"
572 "-------------------------------------\n\n");
574 add_taint(TAINT_BAD_PAGE);
577 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
583 vsnprintf(buf, sizeof(buf), fmt, args);
585 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
588 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
590 unsigned int off; /* Offset of last byte */
591 u8 *addr = page_address(page);
593 print_tracking(s, p);
595 print_page_info(page);
597 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
598 p, p - addr, get_freepointer(s, p));
601 print_section("Bytes b4 ", p - 16, 16);
603 print_section("Object ", p, min_t(unsigned long, s->object_size,
605 if (s->flags & SLAB_RED_ZONE)
606 print_section("Redzone ", p + s->object_size,
607 s->inuse - s->object_size);
610 off = s->offset + sizeof(void *);
614 if (s->flags & SLAB_STORE_USER)
615 off += 2 * sizeof(struct track);
618 /* Beginning of the filler is the free pointer */
619 print_section("Padding ", p + off, s->size - off);
624 static void object_err(struct kmem_cache *s, struct page *page,
625 u8 *object, char *reason)
627 slab_bug(s, "%s", reason);
628 print_trailer(s, page, object);
631 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
637 vsnprintf(buf, sizeof(buf), fmt, args);
639 slab_bug(s, "%s", buf);
640 print_page_info(page);
644 static void init_object(struct kmem_cache *s, void *object, u8 val)
648 if (s->flags & __OBJECT_POISON) {
649 memset(p, POISON_FREE, s->object_size - 1);
650 p[s->object_size - 1] = POISON_END;
653 if (s->flags & SLAB_RED_ZONE)
654 memset(p + s->object_size, val, s->inuse - s->object_size);
657 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
658 void *from, void *to)
660 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
661 memset(from, data, to - from);
664 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
665 u8 *object, char *what,
666 u8 *start, unsigned int value, unsigned int bytes)
671 fault = memchr_inv(start, value, bytes);
676 while (end > fault && end[-1] == value)
679 slab_bug(s, "%s overwritten", what);
680 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
681 fault, end - 1, fault[0], value);
682 print_trailer(s, page, object);
684 restore_bytes(s, what, value, fault, end);
692 * Bytes of the object to be managed.
693 * If the freepointer may overlay the object then the free
694 * pointer is the first word of the object.
696 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
699 * object + s->object_size
700 * Padding to reach word boundary. This is also used for Redzoning.
701 * Padding is extended by another word if Redzoning is enabled and
702 * object_size == inuse.
704 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
705 * 0xcc (RED_ACTIVE) for objects in use.
708 * Meta data starts here.
710 * A. Free pointer (if we cannot overwrite object on free)
711 * B. Tracking data for SLAB_STORE_USER
712 * C. Padding to reach required alignment boundary or at mininum
713 * one word if debugging is on to be able to detect writes
714 * before the word boundary.
716 * Padding is done using 0x5a (POISON_INUSE)
719 * Nothing is used beyond s->size.
721 * If slabcaches are merged then the object_size and inuse boundaries are mostly
722 * ignored. And therefore no slab options that rely on these boundaries
723 * may be used with merged slabcaches.
726 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
728 unsigned long off = s->inuse; /* The end of info */
731 /* Freepointer is placed after the object. */
732 off += sizeof(void *);
734 if (s->flags & SLAB_STORE_USER)
735 /* We also have user information there */
736 off += 2 * sizeof(struct track);
741 return check_bytes_and_report(s, page, p, "Object padding",
742 p + off, POISON_INUSE, s->size - off);
745 /* Check the pad bytes at the end of a slab page */
746 static int slab_pad_check(struct kmem_cache *s, struct page *page)
754 if (!(s->flags & SLAB_POISON))
757 start = page_address(page);
758 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
759 end = start + length;
760 remainder = length % s->size;
764 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
767 while (end > fault && end[-1] == POISON_INUSE)
770 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
771 print_section("Padding ", end - remainder, remainder);
773 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
777 static int check_object(struct kmem_cache *s, struct page *page,
778 void *object, u8 val)
781 u8 *endobject = object + s->object_size;
783 if (s->flags & SLAB_RED_ZONE) {
784 if (!check_bytes_and_report(s, page, object, "Redzone",
785 endobject, val, s->inuse - s->object_size))
788 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
789 check_bytes_and_report(s, page, p, "Alignment padding",
790 endobject, POISON_INUSE, s->inuse - s->object_size);
794 if (s->flags & SLAB_POISON) {
795 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
796 (!check_bytes_and_report(s, page, p, "Poison", p,
797 POISON_FREE, s->object_size - 1) ||
798 !check_bytes_and_report(s, page, p, "Poison",
799 p + s->object_size - 1, POISON_END, 1)))
802 * check_pad_bytes cleans up on its own.
804 check_pad_bytes(s, page, p);
807 if (!s->offset && val == SLUB_RED_ACTIVE)
809 * Object and freepointer overlap. Cannot check
810 * freepointer while object is allocated.
814 /* Check free pointer validity */
815 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
816 object_err(s, page, p, "Freepointer corrupt");
818 * No choice but to zap it and thus lose the remainder
819 * of the free objects in this slab. May cause
820 * another error because the object count is now wrong.
822 set_freepointer(s, p, NULL);
828 static int check_slab(struct kmem_cache *s, struct page *page)
832 VM_BUG_ON(!irqs_disabled());
834 if (!PageSlab(page)) {
835 slab_err(s, page, "Not a valid slab page");
839 maxobj = order_objects(compound_order(page), s->size, s->reserved);
840 if (page->objects > maxobj) {
841 slab_err(s, page, "objects %u > max %u",
842 s->name, page->objects, maxobj);
845 if (page->inuse > page->objects) {
846 slab_err(s, page, "inuse %u > max %u",
847 s->name, page->inuse, page->objects);
850 /* Slab_pad_check fixes things up after itself */
851 slab_pad_check(s, page);
856 * Determine if a certain object on a page is on the freelist. Must hold the
857 * slab lock to guarantee that the chains are in a consistent state.
859 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
864 unsigned long max_objects;
867 while (fp && nr <= page->objects) {
870 if (!check_valid_pointer(s, page, fp)) {
872 object_err(s, page, object,
873 "Freechain corrupt");
874 set_freepointer(s, object, NULL);
877 slab_err(s, page, "Freepointer corrupt");
878 page->freelist = NULL;
879 page->inuse = page->objects;
880 slab_fix(s, "Freelist cleared");
886 fp = get_freepointer(s, object);
890 max_objects = order_objects(compound_order(page), s->size, s->reserved);
891 if (max_objects > MAX_OBJS_PER_PAGE)
892 max_objects = MAX_OBJS_PER_PAGE;
894 if (page->objects != max_objects) {
895 slab_err(s, page, "Wrong number of objects. Found %d but "
896 "should be %d", page->objects, max_objects);
897 page->objects = max_objects;
898 slab_fix(s, "Number of objects adjusted.");
900 if (page->inuse != page->objects - nr) {
901 slab_err(s, page, "Wrong object count. Counter is %d but "
902 "counted were %d", page->inuse, page->objects - nr);
903 page->inuse = page->objects - nr;
904 slab_fix(s, "Object count adjusted.");
906 return search == NULL;
909 static void trace(struct kmem_cache *s, struct page *page, void *object,
912 if (s->flags & SLAB_TRACE) {
913 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
915 alloc ? "alloc" : "free",
920 print_section("Object ", (void *)object, s->object_size);
927 * Hooks for other subsystems that check memory allocations. In a typical
928 * production configuration these hooks all should produce no code at all.
930 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
932 flags &= gfp_allowed_mask;
933 lockdep_trace_alloc(flags);
934 might_sleep_if(flags & __GFP_WAIT);
936 return should_failslab(s->object_size, flags, s->flags);
939 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
941 flags &= gfp_allowed_mask;
942 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
943 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
946 static inline void slab_free_hook(struct kmem_cache *s, void *x)
948 kmemleak_free_recursive(x, s->flags);
951 * Trouble is that we may no longer disable interupts in the fast path
952 * So in order to make the debug calls that expect irqs to be
953 * disabled we need to disable interrupts temporarily.
955 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
959 local_irq_save(flags);
960 kmemcheck_slab_free(s, x, s->object_size);
961 debug_check_no_locks_freed(x, s->object_size);
962 local_irq_restore(flags);
965 if (!(s->flags & SLAB_DEBUG_OBJECTS))
966 debug_check_no_obj_freed(x, s->object_size);
970 * Tracking of fully allocated slabs for debugging purposes.
972 * list_lock must be held.
974 static void add_full(struct kmem_cache *s,
975 struct kmem_cache_node *n, struct page *page)
977 if (!(s->flags & SLAB_STORE_USER))
980 list_add(&page->lru, &n->full);
984 * list_lock must be held.
986 static void remove_full(struct kmem_cache *s, struct page *page)
988 if (!(s->flags & SLAB_STORE_USER))
991 list_del(&page->lru);
994 /* Tracking of the number of slabs for debugging purposes */
995 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
997 struct kmem_cache_node *n = get_node(s, node);
999 return atomic_long_read(&n->nr_slabs);
1002 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1004 return atomic_long_read(&n->nr_slabs);
1007 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1009 struct kmem_cache_node *n = get_node(s, node);
1012 * May be called early in order to allocate a slab for the
1013 * kmem_cache_node structure. Solve the chicken-egg
1014 * dilemma by deferring the increment of the count during
1015 * bootstrap (see early_kmem_cache_node_alloc).
1018 atomic_long_inc(&n->nr_slabs);
1019 atomic_long_add(objects, &n->total_objects);
1022 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1024 struct kmem_cache_node *n = get_node(s, node);
1026 atomic_long_dec(&n->nr_slabs);
1027 atomic_long_sub(objects, &n->total_objects);
1030 /* Object debug checks for alloc/free paths */
1031 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1034 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1037 init_object(s, object, SLUB_RED_INACTIVE);
1038 init_tracking(s, object);
1041 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1042 void *object, unsigned long addr)
1044 if (!check_slab(s, page))
1047 if (!check_valid_pointer(s, page, object)) {
1048 object_err(s, page, object, "Freelist Pointer check fails");
1052 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1055 /* Success perform special debug activities for allocs */
1056 if (s->flags & SLAB_STORE_USER)
1057 set_track(s, object, TRACK_ALLOC, addr);
1058 trace(s, page, object, 1);
1059 init_object(s, object, SLUB_RED_ACTIVE);
1063 if (PageSlab(page)) {
1065 * If this is a slab page then lets do the best we can
1066 * to avoid issues in the future. Marking all objects
1067 * as used avoids touching the remaining objects.
1069 slab_fix(s, "Marking all objects used");
1070 page->inuse = page->objects;
1071 page->freelist = NULL;
1076 static noinline struct kmem_cache_node *free_debug_processing(
1077 struct kmem_cache *s, struct page *page, void *object,
1078 unsigned long addr, unsigned long *flags)
1080 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1082 spin_lock_irqsave(&n->list_lock, *flags);
1085 if (!check_slab(s, page))
1088 if (!check_valid_pointer(s, page, object)) {
1089 slab_err(s, page, "Invalid object pointer 0x%p", object);
1093 if (on_freelist(s, page, object)) {
1094 object_err(s, page, object, "Object already free");
1098 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1101 if (unlikely(s != page->slab_cache)) {
1102 if (!PageSlab(page)) {
1103 slab_err(s, page, "Attempt to free object(0x%p) "
1104 "outside of slab", object);
1105 } else if (!page->slab_cache) {
1107 "SLUB <none>: no slab for object 0x%p.\n",
1111 object_err(s, page, object,
1112 "page slab pointer corrupt.");
1116 if (s->flags & SLAB_STORE_USER)
1117 set_track(s, object, TRACK_FREE, addr);
1118 trace(s, page, object, 0);
1119 init_object(s, object, SLUB_RED_INACTIVE);
1123 * Keep node_lock to preserve integrity
1124 * until the object is actually freed
1130 spin_unlock_irqrestore(&n->list_lock, *flags);
1131 slab_fix(s, "Object at 0x%p not freed", object);
1135 static int __init setup_slub_debug(char *str)
1137 slub_debug = DEBUG_DEFAULT_FLAGS;
1138 if (*str++ != '=' || !*str)
1140 * No options specified. Switch on full debugging.
1146 * No options but restriction on slabs. This means full
1147 * debugging for slabs matching a pattern.
1151 if (tolower(*str) == 'o') {
1153 * Avoid enabling debugging on caches if its minimum order
1154 * would increase as a result.
1156 disable_higher_order_debug = 1;
1163 * Switch off all debugging measures.
1168 * Determine which debug features should be switched on
1170 for (; *str && *str != ','; str++) {
1171 switch (tolower(*str)) {
1173 slub_debug |= SLAB_DEBUG_FREE;
1176 slub_debug |= SLAB_RED_ZONE;
1179 slub_debug |= SLAB_POISON;
1182 slub_debug |= SLAB_STORE_USER;
1185 slub_debug |= SLAB_TRACE;
1188 slub_debug |= SLAB_FAILSLAB;
1191 printk(KERN_ERR "slub_debug option '%c' "
1192 "unknown. skipped\n", *str);
1198 slub_debug_slabs = str + 1;
1203 __setup("slub_debug", setup_slub_debug);
1205 static unsigned long kmem_cache_flags(unsigned long object_size,
1206 unsigned long flags, const char *name,
1207 void (*ctor)(void *))
1210 * Enable debugging if selected on the kernel commandline.
1212 if (slub_debug && (!slub_debug_slabs ||
1213 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1214 flags |= slub_debug;
1219 static inline void setup_object_debug(struct kmem_cache *s,
1220 struct page *page, void *object) {}
1222 static inline int alloc_debug_processing(struct kmem_cache *s,
1223 struct page *page, void *object, unsigned long addr) { return 0; }
1225 static inline struct kmem_cache_node *free_debug_processing(
1226 struct kmem_cache *s, struct page *page, void *object,
1227 unsigned long addr, unsigned long *flags) { return NULL; }
1229 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1231 static inline int check_object(struct kmem_cache *s, struct page *page,
1232 void *object, u8 val) { return 1; }
1233 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1234 struct page *page) {}
1235 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1236 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1237 unsigned long flags, const char *name,
1238 void (*ctor)(void *))
1242 #define slub_debug 0
1244 #define disable_higher_order_debug 0
1246 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1248 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1250 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1252 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1255 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1258 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1261 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1263 #endif /* CONFIG_SLUB_DEBUG */
1266 * Slab allocation and freeing
1268 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1269 struct kmem_cache_order_objects oo)
1271 int order = oo_order(oo);
1273 flags |= __GFP_NOTRACK;
1275 if (node == NUMA_NO_NODE)
1276 return alloc_pages(flags, order);
1278 return alloc_pages_exact_node(node, flags, order);
1281 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1284 struct kmem_cache_order_objects oo = s->oo;
1287 flags &= gfp_allowed_mask;
1289 if (flags & __GFP_WAIT)
1292 flags |= s->allocflags;
1295 * Let the initial higher-order allocation fail under memory pressure
1296 * so we fall-back to the minimum order allocation.
1298 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1300 page = alloc_slab_page(alloc_gfp, node, oo);
1301 if (unlikely(!page)) {
1304 * Allocation may have failed due to fragmentation.
1305 * Try a lower order alloc if possible
1307 page = alloc_slab_page(flags, node, oo);
1310 stat(s, ORDER_FALLBACK);
1313 if (kmemcheck_enabled && page
1314 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1315 int pages = 1 << oo_order(oo);
1317 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1320 * Objects from caches that have a constructor don't get
1321 * cleared when they're allocated, so we need to do it here.
1324 kmemcheck_mark_uninitialized_pages(page, pages);
1326 kmemcheck_mark_unallocated_pages(page, pages);
1329 if (flags & __GFP_WAIT)
1330 local_irq_disable();
1334 page->objects = oo_objects(oo);
1335 mod_zone_page_state(page_zone(page),
1336 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1337 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1343 static void setup_object(struct kmem_cache *s, struct page *page,
1346 setup_object_debug(s, page, object);
1347 if (unlikely(s->ctor))
1351 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1359 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1361 page = allocate_slab(s,
1362 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1366 order = compound_order(page);
1367 inc_slabs_node(s, page_to_nid(page), page->objects);
1368 memcg_bind_pages(s, order);
1369 page->slab_cache = s;
1370 __SetPageSlab(page);
1371 if (page->pfmemalloc)
1372 SetPageSlabPfmemalloc(page);
1374 start = page_address(page);
1376 if (unlikely(s->flags & SLAB_POISON))
1377 memset(start, POISON_INUSE, PAGE_SIZE << order);
1380 for_each_object(p, s, start, page->objects) {
1381 setup_object(s, page, last);
1382 set_freepointer(s, last, p);
1385 setup_object(s, page, last);
1386 set_freepointer(s, last, NULL);
1388 page->freelist = start;
1389 page->inuse = page->objects;
1395 static void __free_slab(struct kmem_cache *s, struct page *page)
1397 int order = compound_order(page);
1398 int pages = 1 << order;
1400 if (kmem_cache_debug(s)) {
1403 slab_pad_check(s, page);
1404 for_each_object(p, s, page_address(page),
1406 check_object(s, page, p, SLUB_RED_INACTIVE);
1409 kmemcheck_free_shadow(page, compound_order(page));
1411 mod_zone_page_state(page_zone(page),
1412 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1413 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1416 __ClearPageSlabPfmemalloc(page);
1417 __ClearPageSlab(page);
1419 memcg_release_pages(s, order);
1420 reset_page_mapcount(page);
1421 if (current->reclaim_state)
1422 current->reclaim_state->reclaimed_slab += pages;
1423 __free_memcg_kmem_pages(page, order);
1426 #define need_reserve_slab_rcu \
1427 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1429 static void rcu_free_slab(struct rcu_head *h)
1433 if (need_reserve_slab_rcu)
1434 page = virt_to_head_page(h);
1436 page = container_of((struct list_head *)h, struct page, lru);
1438 __free_slab(page->slab_cache, page);
1441 static void free_slab(struct kmem_cache *s, struct page *page)
1443 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1444 struct rcu_head *head;
1446 if (need_reserve_slab_rcu) {
1447 int order = compound_order(page);
1448 int offset = (PAGE_SIZE << order) - s->reserved;
1450 VM_BUG_ON(s->reserved != sizeof(*head));
1451 head = page_address(page) + offset;
1454 * RCU free overloads the RCU head over the LRU
1456 head = (void *)&page->lru;
1459 call_rcu(head, rcu_free_slab);
1461 __free_slab(s, page);
1464 static void discard_slab(struct kmem_cache *s, struct page *page)
1466 dec_slabs_node(s, page_to_nid(page), page->objects);
1471 * Management of partially allocated slabs.
1473 * list_lock must be held.
1475 static inline void add_partial(struct kmem_cache_node *n,
1476 struct page *page, int tail)
1479 if (tail == DEACTIVATE_TO_TAIL)
1480 list_add_tail(&page->lru, &n->partial);
1482 list_add(&page->lru, &n->partial);
1486 * list_lock must be held.
1488 static inline void remove_partial(struct kmem_cache_node *n,
1491 list_del(&page->lru);
1496 * Remove slab from the partial list, freeze it and
1497 * return the pointer to the freelist.
1499 * Returns a list of objects or NULL if it fails.
1501 * Must hold list_lock since we modify the partial list.
1503 static inline void *acquire_slab(struct kmem_cache *s,
1504 struct kmem_cache_node *n, struct page *page,
1505 int mode, int *objects)
1508 unsigned long counters;
1512 * Zap the freelist and set the frozen bit.
1513 * The old freelist is the list of objects for the
1514 * per cpu allocation list.
1516 freelist = page->freelist;
1517 counters = page->counters;
1518 new.counters = counters;
1519 *objects = new.objects - new.inuse;
1521 new.inuse = page->objects;
1522 new.freelist = NULL;
1524 new.freelist = freelist;
1527 VM_BUG_ON(new.frozen);
1530 if (!__cmpxchg_double_slab(s, page,
1532 new.freelist, new.counters,
1536 remove_partial(n, page);
1541 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1542 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1545 * Try to allocate a partial slab from a specific node.
1547 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1548 struct kmem_cache_cpu *c, gfp_t flags)
1550 struct page *page, *page2;
1551 void *object = NULL;
1556 * Racy check. If we mistakenly see no partial slabs then we
1557 * just allocate an empty slab. If we mistakenly try to get a
1558 * partial slab and there is none available then get_partials()
1561 if (!n || !n->nr_partial)
1564 spin_lock(&n->list_lock);
1565 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1568 if (!pfmemalloc_match(page, flags))
1571 t = acquire_slab(s, n, page, object == NULL, &objects);
1575 available += objects;
1578 stat(s, ALLOC_FROM_PARTIAL);
1581 put_cpu_partial(s, page, 0);
1582 stat(s, CPU_PARTIAL_NODE);
1584 if (!kmem_cache_has_cpu_partial(s)
1585 || available > s->cpu_partial / 2)
1589 spin_unlock(&n->list_lock);
1594 * Get a page from somewhere. Search in increasing NUMA distances.
1596 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1597 struct kmem_cache_cpu *c)
1600 struct zonelist *zonelist;
1603 enum zone_type high_zoneidx = gfp_zone(flags);
1605 unsigned int cpuset_mems_cookie;
1608 * The defrag ratio allows a configuration of the tradeoffs between
1609 * inter node defragmentation and node local allocations. A lower
1610 * defrag_ratio increases the tendency to do local allocations
1611 * instead of attempting to obtain partial slabs from other nodes.
1613 * If the defrag_ratio is set to 0 then kmalloc() always
1614 * returns node local objects. If the ratio is higher then kmalloc()
1615 * may return off node objects because partial slabs are obtained
1616 * from other nodes and filled up.
1618 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1619 * defrag_ratio = 1000) then every (well almost) allocation will
1620 * first attempt to defrag slab caches on other nodes. This means
1621 * scanning over all nodes to look for partial slabs which may be
1622 * expensive if we do it every time we are trying to find a slab
1623 * with available objects.
1625 if (!s->remote_node_defrag_ratio ||
1626 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1630 cpuset_mems_cookie = get_mems_allowed();
1631 zonelist = node_zonelist(slab_node(), flags);
1632 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1633 struct kmem_cache_node *n;
1635 n = get_node(s, zone_to_nid(zone));
1637 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1638 n->nr_partial > s->min_partial) {
1639 object = get_partial_node(s, n, c, flags);
1642 * Return the object even if
1643 * put_mems_allowed indicated that
1644 * the cpuset mems_allowed was
1645 * updated in parallel. It's a
1646 * harmless race between the alloc
1647 * and the cpuset update.
1649 put_mems_allowed(cpuset_mems_cookie);
1654 } while (!put_mems_allowed(cpuset_mems_cookie));
1660 * Get a partial page, lock it and return it.
1662 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1663 struct kmem_cache_cpu *c)
1666 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1668 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1669 if (object || node != NUMA_NO_NODE)
1672 return get_any_partial(s, flags, c);
1675 #ifdef CONFIG_PREEMPT
1677 * Calculate the next globally unique transaction for disambiguiation
1678 * during cmpxchg. The transactions start with the cpu number and are then
1679 * incremented by CONFIG_NR_CPUS.
1681 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1684 * No preemption supported therefore also no need to check for
1690 static inline unsigned long next_tid(unsigned long tid)
1692 return tid + TID_STEP;
1695 static inline unsigned int tid_to_cpu(unsigned long tid)
1697 return tid % TID_STEP;
1700 static inline unsigned long tid_to_event(unsigned long tid)
1702 return tid / TID_STEP;
1705 static inline unsigned int init_tid(int cpu)
1710 static inline void note_cmpxchg_failure(const char *n,
1711 const struct kmem_cache *s, unsigned long tid)
1713 #ifdef SLUB_DEBUG_CMPXCHG
1714 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1716 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1718 #ifdef CONFIG_PREEMPT
1719 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1720 printk("due to cpu change %d -> %d\n",
1721 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1724 if (tid_to_event(tid) != tid_to_event(actual_tid))
1725 printk("due to cpu running other code. Event %ld->%ld\n",
1726 tid_to_event(tid), tid_to_event(actual_tid));
1728 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1729 actual_tid, tid, next_tid(tid));
1731 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1734 static void init_kmem_cache_cpus(struct kmem_cache *s)
1738 for_each_possible_cpu(cpu)
1739 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1743 * Remove the cpu slab
1745 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1747 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1748 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1750 enum slab_modes l = M_NONE, m = M_NONE;
1752 int tail = DEACTIVATE_TO_HEAD;
1756 if (page->freelist) {
1757 stat(s, DEACTIVATE_REMOTE_FREES);
1758 tail = DEACTIVATE_TO_TAIL;
1762 * Stage one: Free all available per cpu objects back
1763 * to the page freelist while it is still frozen. Leave the
1766 * There is no need to take the list->lock because the page
1769 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1771 unsigned long counters;
1774 prior = page->freelist;
1775 counters = page->counters;
1776 set_freepointer(s, freelist, prior);
1777 new.counters = counters;
1779 VM_BUG_ON(!new.frozen);
1781 } while (!__cmpxchg_double_slab(s, page,
1783 freelist, new.counters,
1784 "drain percpu freelist"));
1786 freelist = nextfree;
1790 * Stage two: Ensure that the page is unfrozen while the
1791 * list presence reflects the actual number of objects
1794 * We setup the list membership and then perform a cmpxchg
1795 * with the count. If there is a mismatch then the page
1796 * is not unfrozen but the page is on the wrong list.
1798 * Then we restart the process which may have to remove
1799 * the page from the list that we just put it on again
1800 * because the number of objects in the slab may have
1805 old.freelist = page->freelist;
1806 old.counters = page->counters;
1807 VM_BUG_ON(!old.frozen);
1809 /* Determine target state of the slab */
1810 new.counters = old.counters;
1813 set_freepointer(s, freelist, old.freelist);
1814 new.freelist = freelist;
1816 new.freelist = old.freelist;
1820 if (!new.inuse && n->nr_partial > s->min_partial)
1822 else if (new.freelist) {
1827 * Taking the spinlock removes the possiblity
1828 * that acquire_slab() will see a slab page that
1831 spin_lock(&n->list_lock);
1835 if (kmem_cache_debug(s) && !lock) {
1838 * This also ensures that the scanning of full
1839 * slabs from diagnostic functions will not see
1842 spin_lock(&n->list_lock);
1850 remove_partial(n, page);
1852 else if (l == M_FULL)
1854 remove_full(s, page);
1856 if (m == M_PARTIAL) {
1858 add_partial(n, page, tail);
1861 } else if (m == M_FULL) {
1863 stat(s, DEACTIVATE_FULL);
1864 add_full(s, n, page);
1870 if (!__cmpxchg_double_slab(s, page,
1871 old.freelist, old.counters,
1872 new.freelist, new.counters,
1877 spin_unlock(&n->list_lock);
1880 stat(s, DEACTIVATE_EMPTY);
1881 discard_slab(s, page);
1887 * Unfreeze all the cpu partial slabs.
1889 * This function must be called with interrupts disabled
1890 * for the cpu using c (or some other guarantee must be there
1891 * to guarantee no concurrent accesses).
1893 static void unfreeze_partials(struct kmem_cache *s,
1894 struct kmem_cache_cpu *c)
1896 #ifdef CONFIG_SLUB_CPU_PARTIAL
1897 struct kmem_cache_node *n = NULL, *n2 = NULL;
1898 struct page *page, *discard_page = NULL;
1900 while ((page = c->partial)) {
1904 c->partial = page->next;
1906 n2 = get_node(s, page_to_nid(page));
1909 spin_unlock(&n->list_lock);
1912 spin_lock(&n->list_lock);
1917 old.freelist = page->freelist;
1918 old.counters = page->counters;
1919 VM_BUG_ON(!old.frozen);
1921 new.counters = old.counters;
1922 new.freelist = old.freelist;
1926 } while (!__cmpxchg_double_slab(s, page,
1927 old.freelist, old.counters,
1928 new.freelist, new.counters,
1929 "unfreezing slab"));
1931 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1932 page->next = discard_page;
1933 discard_page = page;
1935 add_partial(n, page, DEACTIVATE_TO_TAIL);
1936 stat(s, FREE_ADD_PARTIAL);
1941 spin_unlock(&n->list_lock);
1943 while (discard_page) {
1944 page = discard_page;
1945 discard_page = discard_page->next;
1947 stat(s, DEACTIVATE_EMPTY);
1948 discard_slab(s, page);
1955 * Put a page that was just frozen (in __slab_free) into a partial page
1956 * slot if available. This is done without interrupts disabled and without
1957 * preemption disabled. The cmpxchg is racy and may put the partial page
1958 * onto a random cpus partial slot.
1960 * If we did not find a slot then simply move all the partials to the
1961 * per node partial list.
1963 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1965 #ifdef CONFIG_SLUB_CPU_PARTIAL
1966 struct page *oldpage;
1970 if (!s->cpu_partial)
1976 oldpage = this_cpu_read(s->cpu_slab->partial);
1979 pobjects = oldpage->pobjects;
1980 pages = oldpage->pages;
1981 if (drain && pobjects > s->cpu_partial) {
1982 unsigned long flags;
1984 * partial array is full. Move the existing
1985 * set to the per node partial list.
1987 local_irq_save(flags);
1988 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1989 local_irq_restore(flags);
1993 stat(s, CPU_PARTIAL_DRAIN);
1998 pobjects += page->objects - page->inuse;
2000 page->pages = pages;
2001 page->pobjects = pobjects;
2002 page->next = oldpage;
2004 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2008 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2010 stat(s, CPUSLAB_FLUSH);
2011 deactivate_slab(s, c->page, c->freelist);
2013 c->tid = next_tid(c->tid);
2021 * Called from IPI handler with interrupts disabled.
2023 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2025 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2031 unfreeze_partials(s, c);
2035 static void flush_cpu_slab(void *d)
2037 struct kmem_cache *s = d;
2039 __flush_cpu_slab(s, smp_processor_id());
2042 static bool has_cpu_slab(int cpu, void *info)
2044 struct kmem_cache *s = info;
2045 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2047 return c->page || c->partial;
2050 static void flush_all(struct kmem_cache *s)
2052 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2056 * Check if the objects in a per cpu structure fit numa
2057 * locality expectations.
2059 static inline int node_match(struct page *page, int node)
2062 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2068 static int count_free(struct page *page)
2070 return page->objects - page->inuse;
2073 static unsigned long count_partial(struct kmem_cache_node *n,
2074 int (*get_count)(struct page *))
2076 unsigned long flags;
2077 unsigned long x = 0;
2080 spin_lock_irqsave(&n->list_lock, flags);
2081 list_for_each_entry(page, &n->partial, lru)
2082 x += get_count(page);
2083 spin_unlock_irqrestore(&n->list_lock, flags);
2087 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2089 #ifdef CONFIG_SLUB_DEBUG
2090 return atomic_long_read(&n->total_objects);
2096 static noinline void
2097 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2102 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2104 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2105 "default order: %d, min order: %d\n", s->name, s->object_size,
2106 s->size, oo_order(s->oo), oo_order(s->min));
2108 if (oo_order(s->min) > get_order(s->object_size))
2109 printk(KERN_WARNING " %s debugging increased min order, use "
2110 "slub_debug=O to disable.\n", s->name);
2112 for_each_online_node(node) {
2113 struct kmem_cache_node *n = get_node(s, node);
2114 unsigned long nr_slabs;
2115 unsigned long nr_objs;
2116 unsigned long nr_free;
2121 nr_free = count_partial(n, count_free);
2122 nr_slabs = node_nr_slabs(n);
2123 nr_objs = node_nr_objs(n);
2126 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2127 node, nr_slabs, nr_objs, nr_free);
2131 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2132 int node, struct kmem_cache_cpu **pc)
2135 struct kmem_cache_cpu *c = *pc;
2138 freelist = get_partial(s, flags, node, c);
2143 page = new_slab(s, flags, node);
2145 c = __this_cpu_ptr(s->cpu_slab);
2150 * No other reference to the page yet so we can
2151 * muck around with it freely without cmpxchg
2153 freelist = page->freelist;
2154 page->freelist = NULL;
2156 stat(s, ALLOC_SLAB);
2165 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2167 if (unlikely(PageSlabPfmemalloc(page)))
2168 return gfp_pfmemalloc_allowed(gfpflags);
2174 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2175 * or deactivate the page.
2177 * The page is still frozen if the return value is not NULL.
2179 * If this function returns NULL then the page has been unfrozen.
2181 * This function must be called with interrupt disabled.
2183 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2186 unsigned long counters;
2190 freelist = page->freelist;
2191 counters = page->counters;
2193 new.counters = counters;
2194 VM_BUG_ON(!new.frozen);
2196 new.inuse = page->objects;
2197 new.frozen = freelist != NULL;
2199 } while (!__cmpxchg_double_slab(s, page,
2208 * Slow path. The lockless freelist is empty or we need to perform
2211 * Processing is still very fast if new objects have been freed to the
2212 * regular freelist. In that case we simply take over the regular freelist
2213 * as the lockless freelist and zap the regular freelist.
2215 * If that is not working then we fall back to the partial lists. We take the
2216 * first element of the freelist as the object to allocate now and move the
2217 * rest of the freelist to the lockless freelist.
2219 * And if we were unable to get a new slab from the partial slab lists then
2220 * we need to allocate a new slab. This is the slowest path since it involves
2221 * a call to the page allocator and the setup of a new slab.
2223 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2224 unsigned long addr, struct kmem_cache_cpu *c)
2228 unsigned long flags;
2230 local_irq_save(flags);
2231 #ifdef CONFIG_PREEMPT
2233 * We may have been preempted and rescheduled on a different
2234 * cpu before disabling interrupts. Need to reload cpu area
2237 c = this_cpu_ptr(s->cpu_slab);
2245 if (unlikely(!node_match(page, node))) {
2246 stat(s, ALLOC_NODE_MISMATCH);
2247 deactivate_slab(s, page, c->freelist);
2254 * By rights, we should be searching for a slab page that was
2255 * PFMEMALLOC but right now, we are losing the pfmemalloc
2256 * information when the page leaves the per-cpu allocator
2258 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2259 deactivate_slab(s, page, c->freelist);
2265 /* must check again c->freelist in case of cpu migration or IRQ */
2266 freelist = c->freelist;
2270 stat(s, ALLOC_SLOWPATH);
2272 freelist = get_freelist(s, page);
2276 stat(s, DEACTIVATE_BYPASS);
2280 stat(s, ALLOC_REFILL);
2284 * freelist is pointing to the list of objects to be used.
2285 * page is pointing to the page from which the objects are obtained.
2286 * That page must be frozen for per cpu allocations to work.
2288 VM_BUG_ON(!c->page->frozen);
2289 c->freelist = get_freepointer(s, freelist);
2290 c->tid = next_tid(c->tid);
2291 local_irq_restore(flags);
2297 page = c->page = c->partial;
2298 c->partial = page->next;
2299 stat(s, CPU_PARTIAL_ALLOC);
2304 freelist = new_slab_objects(s, gfpflags, node, &c);
2306 if (unlikely(!freelist)) {
2307 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2308 slab_out_of_memory(s, gfpflags, node);
2310 local_irq_restore(flags);
2315 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2318 /* Only entered in the debug case */
2319 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2320 goto new_slab; /* Slab failed checks. Next slab needed */
2322 deactivate_slab(s, page, get_freepointer(s, freelist));
2325 local_irq_restore(flags);
2330 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2331 * have the fastpath folded into their functions. So no function call
2332 * overhead for requests that can be satisfied on the fastpath.
2334 * The fastpath works by first checking if the lockless freelist can be used.
2335 * If not then __slab_alloc is called for slow processing.
2337 * Otherwise we can simply pick the next object from the lockless free list.
2339 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2340 gfp_t gfpflags, int node, unsigned long addr)
2343 struct kmem_cache_cpu *c;
2347 if (slab_pre_alloc_hook(s, gfpflags))
2350 s = memcg_kmem_get_cache(s, gfpflags);
2353 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2354 * enabled. We may switch back and forth between cpus while
2355 * reading from one cpu area. That does not matter as long
2356 * as we end up on the original cpu again when doing the cmpxchg.
2358 * Preemption is disabled for the retrieval of the tid because that
2359 * must occur from the current processor. We cannot allow rescheduling
2360 * on a different processor between the determination of the pointer
2361 * and the retrieval of the tid.
2364 c = __this_cpu_ptr(s->cpu_slab);
2367 * The transaction ids are globally unique per cpu and per operation on
2368 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2369 * occurs on the right processor and that there was no operation on the
2370 * linked list in between.
2375 object = c->freelist;
2377 if (unlikely(!object || !node_match(page, node)))
2378 object = __slab_alloc(s, gfpflags, node, addr, c);
2381 void *next_object = get_freepointer_safe(s, object);
2384 * The cmpxchg will only match if there was no additional
2385 * operation and if we are on the right processor.
2387 * The cmpxchg does the following atomically (without lock semantics!)
2388 * 1. Relocate first pointer to the current per cpu area.
2389 * 2. Verify that tid and freelist have not been changed
2390 * 3. If they were not changed replace tid and freelist
2392 * Since this is without lock semantics the protection is only against
2393 * code executing on this cpu *not* from access by other cpus.
2395 if (unlikely(!this_cpu_cmpxchg_double(
2396 s->cpu_slab->freelist, s->cpu_slab->tid,
2398 next_object, next_tid(tid)))) {
2400 note_cmpxchg_failure("slab_alloc", s, tid);
2403 prefetch_freepointer(s, next_object);
2404 stat(s, ALLOC_FASTPATH);
2407 if (unlikely(gfpflags & __GFP_ZERO) && object)
2408 memset(object, 0, s->object_size);
2410 slab_post_alloc_hook(s, gfpflags, object);
2415 static __always_inline void *slab_alloc(struct kmem_cache *s,
2416 gfp_t gfpflags, unsigned long addr)
2418 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2421 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2423 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2425 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2429 EXPORT_SYMBOL(kmem_cache_alloc);
2431 #ifdef CONFIG_TRACING
2432 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2434 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2435 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2438 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2440 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2442 void *ret = kmalloc_order(size, flags, order);
2443 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2446 EXPORT_SYMBOL(kmalloc_order_trace);
2450 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2452 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2454 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2455 s->object_size, s->size, gfpflags, node);
2459 EXPORT_SYMBOL(kmem_cache_alloc_node);
2461 #ifdef CONFIG_TRACING
2462 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2464 int node, size_t size)
2466 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2468 trace_kmalloc_node(_RET_IP_, ret,
2469 size, s->size, gfpflags, node);
2472 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2477 * Slow patch handling. This may still be called frequently since objects
2478 * have a longer lifetime than the cpu slabs in most processing loads.
2480 * So we still attempt to reduce cache line usage. Just take the slab
2481 * lock and free the item. If there is no additional partial page
2482 * handling required then we can return immediately.
2484 static void __slab_free(struct kmem_cache *s, struct page *page,
2485 void *x, unsigned long addr)
2488 void **object = (void *)x;
2491 unsigned long counters;
2492 struct kmem_cache_node *n = NULL;
2493 unsigned long uninitialized_var(flags);
2495 stat(s, FREE_SLOWPATH);
2497 if (kmem_cache_debug(s) &&
2498 !(n = free_debug_processing(s, page, x, addr, &flags)))
2503 spin_unlock_irqrestore(&n->list_lock, flags);
2506 prior = page->freelist;
2507 counters = page->counters;
2508 set_freepointer(s, object, prior);
2509 new.counters = counters;
2510 was_frozen = new.frozen;
2512 if ((!new.inuse || !prior) && !was_frozen) {
2514 if (kmem_cache_has_cpu_partial(s) && !prior)
2517 * Slab was on no list before and will be partially empty
2518 * We can defer the list move and instead freeze it.
2522 else { /* Needs to be taken off a list */
2524 n = get_node(s, page_to_nid(page));
2526 * Speculatively acquire the list_lock.
2527 * If the cmpxchg does not succeed then we may
2528 * drop the list_lock without any processing.
2530 * Otherwise the list_lock will synchronize with
2531 * other processors updating the list of slabs.
2533 spin_lock_irqsave(&n->list_lock, flags);
2538 } while (!cmpxchg_double_slab(s, page,
2540 object, new.counters,
2546 * If we just froze the page then put it onto the
2547 * per cpu partial list.
2549 if (new.frozen && !was_frozen) {
2550 put_cpu_partial(s, page, 1);
2551 stat(s, CPU_PARTIAL_FREE);
2554 * The list lock was not taken therefore no list
2555 * activity can be necessary.
2558 stat(s, FREE_FROZEN);
2562 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2566 * Objects left in the slab. If it was not on the partial list before
2569 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2570 if (kmem_cache_debug(s))
2571 remove_full(s, page);
2572 add_partial(n, page, DEACTIVATE_TO_TAIL);
2573 stat(s, FREE_ADD_PARTIAL);
2575 spin_unlock_irqrestore(&n->list_lock, flags);
2581 * Slab on the partial list.
2583 remove_partial(n, page);
2584 stat(s, FREE_REMOVE_PARTIAL);
2586 /* Slab must be on the full list */
2587 remove_full(s, page);
2589 spin_unlock_irqrestore(&n->list_lock, flags);
2591 discard_slab(s, page);
2595 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2596 * can perform fastpath freeing without additional function calls.
2598 * The fastpath is only possible if we are freeing to the current cpu slab
2599 * of this processor. This typically the case if we have just allocated
2602 * If fastpath is not possible then fall back to __slab_free where we deal
2603 * with all sorts of special processing.
2605 static __always_inline void slab_free(struct kmem_cache *s,
2606 struct page *page, void *x, unsigned long addr)
2608 void **object = (void *)x;
2609 struct kmem_cache_cpu *c;
2612 slab_free_hook(s, x);
2616 * Determine the currently cpus per cpu slab.
2617 * The cpu may change afterward. However that does not matter since
2618 * data is retrieved via this pointer. If we are on the same cpu
2619 * during the cmpxchg then the free will succedd.
2622 c = __this_cpu_ptr(s->cpu_slab);
2627 if (likely(page == c->page)) {
2628 set_freepointer(s, object, c->freelist);
2630 if (unlikely(!this_cpu_cmpxchg_double(
2631 s->cpu_slab->freelist, s->cpu_slab->tid,
2633 object, next_tid(tid)))) {
2635 note_cmpxchg_failure("slab_free", s, tid);
2638 stat(s, FREE_FASTPATH);
2640 __slab_free(s, page, x, addr);
2644 void kmem_cache_free(struct kmem_cache *s, void *x)
2646 s = cache_from_obj(s, x);
2649 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2650 trace_kmem_cache_free(_RET_IP_, x);
2652 EXPORT_SYMBOL(kmem_cache_free);
2655 * Object placement in a slab is made very easy because we always start at
2656 * offset 0. If we tune the size of the object to the alignment then we can
2657 * get the required alignment by putting one properly sized object after
2660 * Notice that the allocation order determines the sizes of the per cpu
2661 * caches. Each processor has always one slab available for allocations.
2662 * Increasing the allocation order reduces the number of times that slabs
2663 * must be moved on and off the partial lists and is therefore a factor in
2668 * Mininum / Maximum order of slab pages. This influences locking overhead
2669 * and slab fragmentation. A higher order reduces the number of partial slabs
2670 * and increases the number of allocations possible without having to
2671 * take the list_lock.
2673 static int slub_min_order;
2674 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2675 static int slub_min_objects;
2678 * Merge control. If this is set then no merging of slab caches will occur.
2679 * (Could be removed. This was introduced to pacify the merge skeptics.)
2681 static int slub_nomerge;
2684 * Calculate the order of allocation given an slab object size.
2686 * The order of allocation has significant impact on performance and other
2687 * system components. Generally order 0 allocations should be preferred since
2688 * order 0 does not cause fragmentation in the page allocator. Larger objects
2689 * be problematic to put into order 0 slabs because there may be too much
2690 * unused space left. We go to a higher order if more than 1/16th of the slab
2693 * In order to reach satisfactory performance we must ensure that a minimum
2694 * number of objects is in one slab. Otherwise we may generate too much
2695 * activity on the partial lists which requires taking the list_lock. This is
2696 * less a concern for large slabs though which are rarely used.
2698 * slub_max_order specifies the order where we begin to stop considering the
2699 * number of objects in a slab as critical. If we reach slub_max_order then
2700 * we try to keep the page order as low as possible. So we accept more waste
2701 * of space in favor of a small page order.
2703 * Higher order allocations also allow the placement of more objects in a
2704 * slab and thereby reduce object handling overhead. If the user has
2705 * requested a higher mininum order then we start with that one instead of
2706 * the smallest order which will fit the object.
2708 static inline int slab_order(int size, int min_objects,
2709 int max_order, int fract_leftover, int reserved)
2713 int min_order = slub_min_order;
2715 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2716 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2718 for (order = max(min_order,
2719 fls(min_objects * size - 1) - PAGE_SHIFT);
2720 order <= max_order; order++) {
2722 unsigned long slab_size = PAGE_SIZE << order;
2724 if (slab_size < min_objects * size + reserved)
2727 rem = (slab_size - reserved) % size;
2729 if (rem <= slab_size / fract_leftover)
2737 static inline int calculate_order(int size, int reserved)
2745 * Attempt to find best configuration for a slab. This
2746 * works by first attempting to generate a layout with
2747 * the best configuration and backing off gradually.
2749 * First we reduce the acceptable waste in a slab. Then
2750 * we reduce the minimum objects required in a slab.
2752 min_objects = slub_min_objects;
2754 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2755 max_objects = order_objects(slub_max_order, size, reserved);
2756 min_objects = min(min_objects, max_objects);
2758 while (min_objects > 1) {
2760 while (fraction >= 4) {
2761 order = slab_order(size, min_objects,
2762 slub_max_order, fraction, reserved);
2763 if (order <= slub_max_order)
2771 * We were unable to place multiple objects in a slab. Now
2772 * lets see if we can place a single object there.
2774 order = slab_order(size, 1, slub_max_order, 1, reserved);
2775 if (order <= slub_max_order)
2779 * Doh this slab cannot be placed using slub_max_order.
2781 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2782 if (order < MAX_ORDER)
2788 init_kmem_cache_node(struct kmem_cache_node *n)
2791 spin_lock_init(&n->list_lock);
2792 INIT_LIST_HEAD(&n->partial);
2793 #ifdef CONFIG_SLUB_DEBUG
2794 atomic_long_set(&n->nr_slabs, 0);
2795 atomic_long_set(&n->total_objects, 0);
2796 INIT_LIST_HEAD(&n->full);
2800 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2802 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2803 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2806 * Must align to double word boundary for the double cmpxchg
2807 * instructions to work; see __pcpu_double_call_return_bool().
2809 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2810 2 * sizeof(void *));
2815 init_kmem_cache_cpus(s);
2820 static struct kmem_cache *kmem_cache_node;
2823 * No kmalloc_node yet so do it by hand. We know that this is the first
2824 * slab on the node for this slabcache. There are no concurrent accesses
2827 * Note that this function only works on the kmalloc_node_cache
2828 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2829 * memory on a fresh node that has no slab structures yet.
2831 static void early_kmem_cache_node_alloc(int node)
2834 struct kmem_cache_node *n;
2836 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2838 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2841 if (page_to_nid(page) != node) {
2842 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2844 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2845 "in order to be able to continue\n");
2850 page->freelist = get_freepointer(kmem_cache_node, n);
2853 kmem_cache_node->node[node] = n;
2854 #ifdef CONFIG_SLUB_DEBUG
2855 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2856 init_tracking(kmem_cache_node, n);
2858 init_kmem_cache_node(n);
2859 inc_slabs_node(kmem_cache_node, node, page->objects);
2861 add_partial(n, page, DEACTIVATE_TO_HEAD);
2864 static void free_kmem_cache_nodes(struct kmem_cache *s)
2868 for_each_node_state(node, N_NORMAL_MEMORY) {
2869 struct kmem_cache_node *n = s->node[node];
2872 kmem_cache_free(kmem_cache_node, n);
2874 s->node[node] = NULL;
2878 static int init_kmem_cache_nodes(struct kmem_cache *s)
2882 for_each_node_state(node, N_NORMAL_MEMORY) {
2883 struct kmem_cache_node *n;
2885 if (slab_state == DOWN) {
2886 early_kmem_cache_node_alloc(node);
2889 n = kmem_cache_alloc_node(kmem_cache_node,
2893 free_kmem_cache_nodes(s);
2898 init_kmem_cache_node(n);
2903 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2905 if (min < MIN_PARTIAL)
2907 else if (min > MAX_PARTIAL)
2909 s->min_partial = min;
2913 * calculate_sizes() determines the order and the distribution of data within
2916 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2918 unsigned long flags = s->flags;
2919 unsigned long size = s->object_size;
2923 * Round up object size to the next word boundary. We can only
2924 * place the free pointer at word boundaries and this determines
2925 * the possible location of the free pointer.
2927 size = ALIGN(size, sizeof(void *));
2929 #ifdef CONFIG_SLUB_DEBUG
2931 * Determine if we can poison the object itself. If the user of
2932 * the slab may touch the object after free or before allocation
2933 * then we should never poison the object itself.
2935 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2937 s->flags |= __OBJECT_POISON;
2939 s->flags &= ~__OBJECT_POISON;
2943 * If we are Redzoning then check if there is some space between the
2944 * end of the object and the free pointer. If not then add an
2945 * additional word to have some bytes to store Redzone information.
2947 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2948 size += sizeof(void *);
2952 * With that we have determined the number of bytes in actual use
2953 * by the object. This is the potential offset to the free pointer.
2957 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2960 * Relocate free pointer after the object if it is not
2961 * permitted to overwrite the first word of the object on
2964 * This is the case if we do RCU, have a constructor or
2965 * destructor or are poisoning the objects.
2968 size += sizeof(void *);
2971 #ifdef CONFIG_SLUB_DEBUG
2972 if (flags & SLAB_STORE_USER)
2974 * Need to store information about allocs and frees after
2977 size += 2 * sizeof(struct track);
2979 if (flags & SLAB_RED_ZONE)
2981 * Add some empty padding so that we can catch
2982 * overwrites from earlier objects rather than let
2983 * tracking information or the free pointer be
2984 * corrupted if a user writes before the start
2987 size += sizeof(void *);
2991 * SLUB stores one object immediately after another beginning from
2992 * offset 0. In order to align the objects we have to simply size
2993 * each object to conform to the alignment.
2995 size = ALIGN(size, s->align);
2997 if (forced_order >= 0)
2998 order = forced_order;
3000 order = calculate_order(size, s->reserved);
3007 s->allocflags |= __GFP_COMP;
3009 if (s->flags & SLAB_CACHE_DMA)
3010 s->allocflags |= GFP_DMA;
3012 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3013 s->allocflags |= __GFP_RECLAIMABLE;
3016 * Determine the number of objects per slab
3018 s->oo = oo_make(order, size, s->reserved);
3019 s->min = oo_make(get_order(size), size, s->reserved);
3020 if (oo_objects(s->oo) > oo_objects(s->max))
3023 return !!oo_objects(s->oo);
3026 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3028 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3031 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3032 s->reserved = sizeof(struct rcu_head);
3034 if (!calculate_sizes(s, -1))
3036 if (disable_higher_order_debug) {
3038 * Disable debugging flags that store metadata if the min slab
3041 if (get_order(s->size) > get_order(s->object_size)) {
3042 s->flags &= ~DEBUG_METADATA_FLAGS;
3044 if (!calculate_sizes(s, -1))
3049 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3050 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3051 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3052 /* Enable fast mode */
3053 s->flags |= __CMPXCHG_DOUBLE;
3057 * The larger the object size is, the more pages we want on the partial
3058 * list to avoid pounding the page allocator excessively.
3060 set_min_partial(s, ilog2(s->size) / 2);
3063 * cpu_partial determined the maximum number of objects kept in the
3064 * per cpu partial lists of a processor.
3066 * Per cpu partial lists mainly contain slabs that just have one
3067 * object freed. If they are used for allocation then they can be
3068 * filled up again with minimal effort. The slab will never hit the
3069 * per node partial lists and therefore no locking will be required.
3071 * This setting also determines
3073 * A) The number of objects from per cpu partial slabs dumped to the
3074 * per node list when we reach the limit.
3075 * B) The number of objects in cpu partial slabs to extract from the
3076 * per node list when we run out of per cpu objects. We only fetch 50%
3077 * to keep some capacity around for frees.
3079 if (!kmem_cache_has_cpu_partial(s))
3081 else if (s->size >= PAGE_SIZE)
3083 else if (s->size >= 1024)
3085 else if (s->size >= 256)
3086 s->cpu_partial = 13;
3088 s->cpu_partial = 30;
3091 s->remote_node_defrag_ratio = 1000;
3093 if (!init_kmem_cache_nodes(s))
3096 if (alloc_kmem_cache_cpus(s))
3099 free_kmem_cache_nodes(s);
3101 if (flags & SLAB_PANIC)
3102 panic("Cannot create slab %s size=%lu realsize=%u "
3103 "order=%u offset=%u flags=%lx\n",
3104 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3109 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3112 #ifdef CONFIG_SLUB_DEBUG
3113 void *addr = page_address(page);
3115 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3116 sizeof(long), GFP_ATOMIC);
3119 slab_err(s, page, text, s->name);
3122 get_map(s, page, map);
3123 for_each_object(p, s, addr, page->objects) {
3125 if (!test_bit(slab_index(p, s, addr), map)) {
3126 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3128 print_tracking(s, p);
3137 * Attempt to free all partial slabs on a node.
3138 * This is called from kmem_cache_close(). We must be the last thread
3139 * using the cache and therefore we do not need to lock anymore.
3141 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3143 struct page *page, *h;
3145 list_for_each_entry_safe(page, h, &n->partial, lru) {
3147 remove_partial(n, page);
3148 discard_slab(s, page);
3150 list_slab_objects(s, page,
3151 "Objects remaining in %s on kmem_cache_close()");
3157 * Release all resources used by a slab cache.
3159 static inline int kmem_cache_close(struct kmem_cache *s)
3164 /* Attempt to free all objects */
3165 for_each_node_state(node, N_NORMAL_MEMORY) {
3166 struct kmem_cache_node *n = get_node(s, node);
3169 if (n->nr_partial || slabs_node(s, node))
3172 free_percpu(s->cpu_slab);
3173 free_kmem_cache_nodes(s);
3177 int __kmem_cache_shutdown(struct kmem_cache *s)
3179 int rc = kmem_cache_close(s);
3183 * We do the same lock strategy around sysfs_slab_add, see
3184 * __kmem_cache_create. Because this is pretty much the last
3185 * operation we do and the lock will be released shortly after
3186 * that in slab_common.c, we could just move sysfs_slab_remove
3187 * to a later point in common code. We should do that when we
3188 * have a common sysfs framework for all allocators.
3190 mutex_unlock(&slab_mutex);
3191 sysfs_slab_remove(s);
3192 mutex_lock(&slab_mutex);
3198 /********************************************************************
3200 *******************************************************************/
3202 static int __init setup_slub_min_order(char *str)
3204 get_option(&str, &slub_min_order);
3209 __setup("slub_min_order=", setup_slub_min_order);
3211 static int __init setup_slub_max_order(char *str)
3213 get_option(&str, &slub_max_order);
3214 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3219 __setup("slub_max_order=", setup_slub_max_order);
3221 static int __init setup_slub_min_objects(char *str)
3223 get_option(&str, &slub_min_objects);
3228 __setup("slub_min_objects=", setup_slub_min_objects);
3230 static int __init setup_slub_nomerge(char *str)
3236 __setup("slub_nomerge", setup_slub_nomerge);
3238 void *__kmalloc(size_t size, gfp_t flags)
3240 struct kmem_cache *s;
3243 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3244 return kmalloc_large(size, flags);
3246 s = kmalloc_slab(size, flags);
3248 if (unlikely(ZERO_OR_NULL_PTR(s)))
3251 ret = slab_alloc(s, flags, _RET_IP_);
3253 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3257 EXPORT_SYMBOL(__kmalloc);
3260 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3265 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3266 page = alloc_pages_node(node, flags, get_order(size));
3268 ptr = page_address(page);
3270 kmemleak_alloc(ptr, size, 1, flags);
3274 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3276 struct kmem_cache *s;
3279 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3280 ret = kmalloc_large_node(size, flags, node);
3282 trace_kmalloc_node(_RET_IP_, ret,
3283 size, PAGE_SIZE << get_order(size),
3289 s = kmalloc_slab(size, flags);
3291 if (unlikely(ZERO_OR_NULL_PTR(s)))
3294 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3296 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3300 EXPORT_SYMBOL(__kmalloc_node);
3303 size_t ksize(const void *object)
3307 if (unlikely(object == ZERO_SIZE_PTR))
3310 page = virt_to_head_page(object);
3312 if (unlikely(!PageSlab(page))) {
3313 WARN_ON(!PageCompound(page));
3314 return PAGE_SIZE << compound_order(page);
3317 return slab_ksize(page->slab_cache);
3319 EXPORT_SYMBOL(ksize);
3321 #ifdef CONFIG_SLUB_DEBUG
3322 bool verify_mem_not_deleted(const void *x)
3325 void *object = (void *)x;
3326 unsigned long flags;
3329 if (unlikely(ZERO_OR_NULL_PTR(x)))
3332 local_irq_save(flags);
3334 page = virt_to_head_page(x);
3335 if (unlikely(!PageSlab(page))) {
3336 /* maybe it was from stack? */
3342 if (on_freelist(page->slab_cache, page, object)) {
3343 object_err(page->slab_cache, page, object, "Object is on free-list");
3351 local_irq_restore(flags);
3354 EXPORT_SYMBOL(verify_mem_not_deleted);
3357 void kfree(const void *x)
3360 void *object = (void *)x;
3362 trace_kfree(_RET_IP_, x);
3364 if (unlikely(ZERO_OR_NULL_PTR(x)))
3367 page = virt_to_head_page(x);
3368 if (unlikely(!PageSlab(page))) {
3369 BUG_ON(!PageCompound(page));
3371 __free_memcg_kmem_pages(page, compound_order(page));
3374 slab_free(page->slab_cache, page, object, _RET_IP_);
3376 EXPORT_SYMBOL(kfree);
3379 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3380 * the remaining slabs by the number of items in use. The slabs with the
3381 * most items in use come first. New allocations will then fill those up
3382 * and thus they can be removed from the partial lists.
3384 * The slabs with the least items are placed last. This results in them
3385 * being allocated from last increasing the chance that the last objects
3386 * are freed in them.
3388 int kmem_cache_shrink(struct kmem_cache *s)
3392 struct kmem_cache_node *n;
3395 int objects = oo_objects(s->max);
3396 struct list_head *slabs_by_inuse =
3397 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3398 unsigned long flags;
3400 if (!slabs_by_inuse)
3404 for_each_node_state(node, N_NORMAL_MEMORY) {
3405 n = get_node(s, node);
3410 for (i = 0; i < objects; i++)
3411 INIT_LIST_HEAD(slabs_by_inuse + i);
3413 spin_lock_irqsave(&n->list_lock, flags);
3416 * Build lists indexed by the items in use in each slab.
3418 * Note that concurrent frees may occur while we hold the
3419 * list_lock. page->inuse here is the upper limit.
3421 list_for_each_entry_safe(page, t, &n->partial, lru) {
3422 list_move(&page->lru, slabs_by_inuse + page->inuse);
3428 * Rebuild the partial list with the slabs filled up most
3429 * first and the least used slabs at the end.
3431 for (i = objects - 1; i > 0; i--)
3432 list_splice(slabs_by_inuse + i, n->partial.prev);
3434 spin_unlock_irqrestore(&n->list_lock, flags);
3436 /* Release empty slabs */
3437 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3438 discard_slab(s, page);
3441 kfree(slabs_by_inuse);
3444 EXPORT_SYMBOL(kmem_cache_shrink);
3446 #if defined(CONFIG_MEMORY_HOTPLUG)
3447 static int slab_mem_going_offline_callback(void *arg)
3449 struct kmem_cache *s;
3451 mutex_lock(&slab_mutex);
3452 list_for_each_entry(s, &slab_caches, list)
3453 kmem_cache_shrink(s);
3454 mutex_unlock(&slab_mutex);
3459 static void slab_mem_offline_callback(void *arg)
3461 struct kmem_cache_node *n;
3462 struct kmem_cache *s;
3463 struct memory_notify *marg = arg;
3466 offline_node = marg->status_change_nid_normal;
3469 * If the node still has available memory. we need kmem_cache_node
3472 if (offline_node < 0)
3475 mutex_lock(&slab_mutex);
3476 list_for_each_entry(s, &slab_caches, list) {
3477 n = get_node(s, offline_node);
3480 * if n->nr_slabs > 0, slabs still exist on the node
3481 * that is going down. We were unable to free them,
3482 * and offline_pages() function shouldn't call this
3483 * callback. So, we must fail.
3485 BUG_ON(slabs_node(s, offline_node));
3487 s->node[offline_node] = NULL;
3488 kmem_cache_free(kmem_cache_node, n);
3491 mutex_unlock(&slab_mutex);
3494 static int slab_mem_going_online_callback(void *arg)
3496 struct kmem_cache_node *n;
3497 struct kmem_cache *s;
3498 struct memory_notify *marg = arg;
3499 int nid = marg->status_change_nid_normal;
3503 * If the node's memory is already available, then kmem_cache_node is
3504 * already created. Nothing to do.
3510 * We are bringing a node online. No memory is available yet. We must
3511 * allocate a kmem_cache_node structure in order to bring the node
3514 mutex_lock(&slab_mutex);
3515 list_for_each_entry(s, &slab_caches, list) {
3517 * XXX: kmem_cache_alloc_node will fallback to other nodes
3518 * since memory is not yet available from the node that
3521 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3526 init_kmem_cache_node(n);
3530 mutex_unlock(&slab_mutex);
3534 static int slab_memory_callback(struct notifier_block *self,
3535 unsigned long action, void *arg)
3540 case MEM_GOING_ONLINE:
3541 ret = slab_mem_going_online_callback(arg);
3543 case MEM_GOING_OFFLINE:
3544 ret = slab_mem_going_offline_callback(arg);
3547 case MEM_CANCEL_ONLINE:
3548 slab_mem_offline_callback(arg);
3551 case MEM_CANCEL_OFFLINE:
3555 ret = notifier_from_errno(ret);
3561 #endif /* CONFIG_MEMORY_HOTPLUG */
3563 /********************************************************************
3564 * Basic setup of slabs
3565 *******************************************************************/
3568 * Used for early kmem_cache structures that were allocated using
3569 * the page allocator. Allocate them properly then fix up the pointers
3570 * that may be pointing to the wrong kmem_cache structure.
3573 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3576 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3578 memcpy(s, static_cache, kmem_cache->object_size);
3581 * This runs very early, and only the boot processor is supposed to be
3582 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3585 __flush_cpu_slab(s, smp_processor_id());
3586 for_each_node_state(node, N_NORMAL_MEMORY) {
3587 struct kmem_cache_node *n = get_node(s, node);
3591 list_for_each_entry(p, &n->partial, lru)
3594 #ifdef CONFIG_SLUB_DEBUG
3595 list_for_each_entry(p, &n->full, lru)
3600 list_add(&s->list, &slab_caches);
3604 void __init kmem_cache_init(void)
3606 static __initdata struct kmem_cache boot_kmem_cache,
3607 boot_kmem_cache_node;
3609 if (debug_guardpage_minorder())
3612 kmem_cache_node = &boot_kmem_cache_node;
3613 kmem_cache = &boot_kmem_cache;
3615 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3616 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3618 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3620 /* Able to allocate the per node structures */
3621 slab_state = PARTIAL;
3623 create_boot_cache(kmem_cache, "kmem_cache",
3624 offsetof(struct kmem_cache, node) +
3625 nr_node_ids * sizeof(struct kmem_cache_node *),
3626 SLAB_HWCACHE_ALIGN);
3628 kmem_cache = bootstrap(&boot_kmem_cache);
3631 * Allocate kmem_cache_node properly from the kmem_cache slab.
3632 * kmem_cache_node is separately allocated so no need to
3633 * update any list pointers.
3635 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3637 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3638 create_kmalloc_caches(0);
3641 register_cpu_notifier(&slab_notifier);
3645 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3646 " CPUs=%d, Nodes=%d\n",
3648 slub_min_order, slub_max_order, slub_min_objects,
3649 nr_cpu_ids, nr_node_ids);
3652 void __init kmem_cache_init_late(void)
3657 * Find a mergeable slab cache
3659 static int slab_unmergeable(struct kmem_cache *s)
3661 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3668 * We may have set a slab to be unmergeable during bootstrap.
3670 if (s->refcount < 0)
3676 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3677 size_t align, unsigned long flags, const char *name,
3678 void (*ctor)(void *))
3680 struct kmem_cache *s;
3682 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3688 size = ALIGN(size, sizeof(void *));
3689 align = calculate_alignment(flags, align, size);
3690 size = ALIGN(size, align);
3691 flags = kmem_cache_flags(size, flags, name, NULL);
3693 list_for_each_entry(s, &slab_caches, list) {
3694 if (slab_unmergeable(s))
3700 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3703 * Check if alignment is compatible.
3704 * Courtesy of Adrian Drzewiecki
3706 if ((s->size & ~(align - 1)) != s->size)
3709 if (s->size - size >= sizeof(void *))
3712 if (!cache_match_memcg(s, memcg))
3721 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3722 size_t align, unsigned long flags, void (*ctor)(void *))
3724 struct kmem_cache *s;
3726 s = find_mergeable(memcg, size, align, flags, name, ctor);
3730 * Adjust the object sizes so that we clear
3731 * the complete object on kzalloc.
3733 s->object_size = max(s->object_size, (int)size);
3734 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3736 if (sysfs_slab_alias(s, name)) {
3745 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3749 err = kmem_cache_open(s, flags);
3753 /* Mutex is not taken during early boot */
3754 if (slab_state <= UP)
3757 memcg_propagate_slab_attrs(s);
3758 mutex_unlock(&slab_mutex);
3759 err = sysfs_slab_add(s);
3760 mutex_lock(&slab_mutex);
3763 kmem_cache_close(s);
3770 * Use the cpu notifier to insure that the cpu slabs are flushed when
3773 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3774 unsigned long action, void *hcpu)
3776 long cpu = (long)hcpu;
3777 struct kmem_cache *s;
3778 unsigned long flags;
3781 case CPU_UP_CANCELED:
3782 case CPU_UP_CANCELED_FROZEN:
3784 case CPU_DEAD_FROZEN:
3785 mutex_lock(&slab_mutex);
3786 list_for_each_entry(s, &slab_caches, list) {
3787 local_irq_save(flags);
3788 __flush_cpu_slab(s, cpu);
3789 local_irq_restore(flags);
3791 mutex_unlock(&slab_mutex);
3799 static struct notifier_block __cpuinitdata slab_notifier = {
3800 .notifier_call = slab_cpuup_callback
3805 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3807 struct kmem_cache *s;
3810 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3811 return kmalloc_large(size, gfpflags);
3813 s = kmalloc_slab(size, gfpflags);
3815 if (unlikely(ZERO_OR_NULL_PTR(s)))
3818 ret = slab_alloc(s, gfpflags, caller);
3820 /* Honor the call site pointer we received. */
3821 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3827 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3828 int node, unsigned long caller)
3830 struct kmem_cache *s;
3833 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3834 ret = kmalloc_large_node(size, gfpflags, node);
3836 trace_kmalloc_node(caller, ret,
3837 size, PAGE_SIZE << get_order(size),
3843 s = kmalloc_slab(size, gfpflags);
3845 if (unlikely(ZERO_OR_NULL_PTR(s)))
3848 ret = slab_alloc_node(s, gfpflags, node, caller);
3850 /* Honor the call site pointer we received. */
3851 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3858 static int count_inuse(struct page *page)
3863 static int count_total(struct page *page)
3865 return page->objects;
3869 #ifdef CONFIG_SLUB_DEBUG
3870 static int validate_slab(struct kmem_cache *s, struct page *page,
3874 void *addr = page_address(page);
3876 if (!check_slab(s, page) ||
3877 !on_freelist(s, page, NULL))
3880 /* Now we know that a valid freelist exists */
3881 bitmap_zero(map, page->objects);
3883 get_map(s, page, map);
3884 for_each_object(p, s, addr, page->objects) {
3885 if (test_bit(slab_index(p, s, addr), map))
3886 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3890 for_each_object(p, s, addr, page->objects)
3891 if (!test_bit(slab_index(p, s, addr), map))
3892 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3897 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3901 validate_slab(s, page, map);
3905 static int validate_slab_node(struct kmem_cache *s,
3906 struct kmem_cache_node *n, unsigned long *map)
3908 unsigned long count = 0;
3910 unsigned long flags;
3912 spin_lock_irqsave(&n->list_lock, flags);
3914 list_for_each_entry(page, &n->partial, lru) {
3915 validate_slab_slab(s, page, map);
3918 if (count != n->nr_partial)
3919 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3920 "counter=%ld\n", s->name, count, n->nr_partial);
3922 if (!(s->flags & SLAB_STORE_USER))
3925 list_for_each_entry(page, &n->full, lru) {
3926 validate_slab_slab(s, page, map);
3929 if (count != atomic_long_read(&n->nr_slabs))
3930 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3931 "counter=%ld\n", s->name, count,
3932 atomic_long_read(&n->nr_slabs));
3935 spin_unlock_irqrestore(&n->list_lock, flags);
3939 static long validate_slab_cache(struct kmem_cache *s)
3942 unsigned long count = 0;
3943 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3944 sizeof(unsigned long), GFP_KERNEL);
3950 for_each_node_state(node, N_NORMAL_MEMORY) {
3951 struct kmem_cache_node *n = get_node(s, node);
3953 count += validate_slab_node(s, n, map);
3959 * Generate lists of code addresses where slabcache objects are allocated
3964 unsigned long count;
3971 DECLARE_BITMAP(cpus, NR_CPUS);
3977 unsigned long count;
3978 struct location *loc;
3981 static void free_loc_track(struct loc_track *t)
3984 free_pages((unsigned long)t->loc,
3985 get_order(sizeof(struct location) * t->max));
3988 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3993 order = get_order(sizeof(struct location) * max);
3995 l = (void *)__get_free_pages(flags, order);
4000 memcpy(l, t->loc, sizeof(struct location) * t->count);
4008 static int add_location(struct loc_track *t, struct kmem_cache *s,
4009 const struct track *track)
4011 long start, end, pos;
4013 unsigned long caddr;
4014 unsigned long age = jiffies - track->when;
4020 pos = start + (end - start + 1) / 2;
4023 * There is nothing at "end". If we end up there
4024 * we need to add something to before end.
4029 caddr = t->loc[pos].addr;
4030 if (track->addr == caddr) {
4036 if (age < l->min_time)
4038 if (age > l->max_time)
4041 if (track->pid < l->min_pid)
4042 l->min_pid = track->pid;
4043 if (track->pid > l->max_pid)
4044 l->max_pid = track->pid;
4046 cpumask_set_cpu(track->cpu,
4047 to_cpumask(l->cpus));
4049 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4053 if (track->addr < caddr)
4060 * Not found. Insert new tracking element.
4062 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4068 (t->count - pos) * sizeof(struct location));
4071 l->addr = track->addr;
4075 l->min_pid = track->pid;
4076 l->max_pid = track->pid;
4077 cpumask_clear(to_cpumask(l->cpus));
4078 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4079 nodes_clear(l->nodes);
4080 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4084 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4085 struct page *page, enum track_item alloc,
4088 void *addr = page_address(page);
4091 bitmap_zero(map, page->objects);
4092 get_map(s, page, map);
4094 for_each_object(p, s, addr, page->objects)
4095 if (!test_bit(slab_index(p, s, addr), map))
4096 add_location(t, s, get_track(s, p, alloc));
4099 static int list_locations(struct kmem_cache *s, char *buf,
4100 enum track_item alloc)
4104 struct loc_track t = { 0, 0, NULL };
4106 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4107 sizeof(unsigned long), GFP_KERNEL);
4109 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4112 return sprintf(buf, "Out of memory\n");
4114 /* Push back cpu slabs */
4117 for_each_node_state(node, N_NORMAL_MEMORY) {
4118 struct kmem_cache_node *n = get_node(s, node);
4119 unsigned long flags;
4122 if (!atomic_long_read(&n->nr_slabs))
4125 spin_lock_irqsave(&n->list_lock, flags);
4126 list_for_each_entry(page, &n->partial, lru)
4127 process_slab(&t, s, page, alloc, map);
4128 list_for_each_entry(page, &n->full, lru)
4129 process_slab(&t, s, page, alloc, map);
4130 spin_unlock_irqrestore(&n->list_lock, flags);
4133 for (i = 0; i < t.count; i++) {
4134 struct location *l = &t.loc[i];
4136 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4138 len += sprintf(buf + len, "%7ld ", l->count);
4141 len += sprintf(buf + len, "%pS", (void *)l->addr);
4143 len += sprintf(buf + len, "<not-available>");
4145 if (l->sum_time != l->min_time) {
4146 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4148 (long)div_u64(l->sum_time, l->count),
4151 len += sprintf(buf + len, " age=%ld",
4154 if (l->min_pid != l->max_pid)
4155 len += sprintf(buf + len, " pid=%ld-%ld",
4156 l->min_pid, l->max_pid);
4158 len += sprintf(buf + len, " pid=%ld",
4161 if (num_online_cpus() > 1 &&
4162 !cpumask_empty(to_cpumask(l->cpus)) &&
4163 len < PAGE_SIZE - 60) {
4164 len += sprintf(buf + len, " cpus=");
4165 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4166 to_cpumask(l->cpus));
4169 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4170 len < PAGE_SIZE - 60) {
4171 len += sprintf(buf + len, " nodes=");
4172 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4176 len += sprintf(buf + len, "\n");
4182 len += sprintf(buf, "No data\n");
4187 #ifdef SLUB_RESILIENCY_TEST
4188 static void resiliency_test(void)
4192 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4194 printk(KERN_ERR "SLUB resiliency testing\n");
4195 printk(KERN_ERR "-----------------------\n");
4196 printk(KERN_ERR "A. Corruption after allocation\n");
4198 p = kzalloc(16, GFP_KERNEL);
4200 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4201 " 0x12->0x%p\n\n", p + 16);
4203 validate_slab_cache(kmalloc_caches[4]);
4205 /* Hmmm... The next two are dangerous */
4206 p = kzalloc(32, GFP_KERNEL);
4207 p[32 + sizeof(void *)] = 0x34;
4208 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4209 " 0x34 -> -0x%p\n", p);
4211 "If allocated object is overwritten then not detectable\n\n");
4213 validate_slab_cache(kmalloc_caches[5]);
4214 p = kzalloc(64, GFP_KERNEL);
4215 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4217 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4220 "If allocated object is overwritten then not detectable\n\n");
4221 validate_slab_cache(kmalloc_caches[6]);
4223 printk(KERN_ERR "\nB. Corruption after free\n");
4224 p = kzalloc(128, GFP_KERNEL);
4227 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4228 validate_slab_cache(kmalloc_caches[7]);
4230 p = kzalloc(256, GFP_KERNEL);
4233 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4235 validate_slab_cache(kmalloc_caches[8]);
4237 p = kzalloc(512, GFP_KERNEL);
4240 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4241 validate_slab_cache(kmalloc_caches[9]);
4245 static void resiliency_test(void) {};
4250 enum slab_stat_type {
4251 SL_ALL, /* All slabs */
4252 SL_PARTIAL, /* Only partially allocated slabs */
4253 SL_CPU, /* Only slabs used for cpu caches */
4254 SL_OBJECTS, /* Determine allocated objects not slabs */
4255 SL_TOTAL /* Determine object capacity not slabs */
4258 #define SO_ALL (1 << SL_ALL)
4259 #define SO_PARTIAL (1 << SL_PARTIAL)
4260 #define SO_CPU (1 << SL_CPU)
4261 #define SO_OBJECTS (1 << SL_OBJECTS)
4262 #define SO_TOTAL (1 << SL_TOTAL)
4264 static ssize_t show_slab_objects(struct kmem_cache *s,
4265 char *buf, unsigned long flags)
4267 unsigned long total = 0;
4270 unsigned long *nodes;
4271 unsigned long *per_cpu;
4273 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4276 per_cpu = nodes + nr_node_ids;
4278 if (flags & SO_CPU) {
4281 for_each_possible_cpu(cpu) {
4282 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4286 page = ACCESS_ONCE(c->page);
4290 node = page_to_nid(page);
4291 if (flags & SO_TOTAL)
4293 else if (flags & SO_OBJECTS)
4301 page = ACCESS_ONCE(c->partial);
4312 lock_memory_hotplug();
4313 #ifdef CONFIG_SLUB_DEBUG
4314 if (flags & SO_ALL) {
4315 for_each_node_state(node, N_NORMAL_MEMORY) {
4316 struct kmem_cache_node *n = get_node(s, node);
4318 if (flags & SO_TOTAL)
4319 x = atomic_long_read(&n->total_objects);
4320 else if (flags & SO_OBJECTS)
4321 x = atomic_long_read(&n->total_objects) -
4322 count_partial(n, count_free);
4325 x = atomic_long_read(&n->nr_slabs);
4332 if (flags & SO_PARTIAL) {
4333 for_each_node_state(node, N_NORMAL_MEMORY) {
4334 struct kmem_cache_node *n = get_node(s, node);
4336 if (flags & SO_TOTAL)
4337 x = count_partial(n, count_total);
4338 else if (flags & SO_OBJECTS)
4339 x = count_partial(n, count_inuse);
4346 x = sprintf(buf, "%lu", total);
4348 for_each_node_state(node, N_NORMAL_MEMORY)
4350 x += sprintf(buf + x, " N%d=%lu",
4353 unlock_memory_hotplug();
4355 return x + sprintf(buf + x, "\n");
4358 #ifdef CONFIG_SLUB_DEBUG
4359 static int any_slab_objects(struct kmem_cache *s)
4363 for_each_online_node(node) {
4364 struct kmem_cache_node *n = get_node(s, node);
4369 if (atomic_long_read(&n->total_objects))
4376 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4377 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4379 struct slab_attribute {
4380 struct attribute attr;
4381 ssize_t (*show)(struct kmem_cache *s, char *buf);
4382 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4385 #define SLAB_ATTR_RO(_name) \
4386 static struct slab_attribute _name##_attr = \
4387 __ATTR(_name, 0400, _name##_show, NULL)
4389 #define SLAB_ATTR(_name) \
4390 static struct slab_attribute _name##_attr = \
4391 __ATTR(_name, 0600, _name##_show, _name##_store)
4393 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4395 return sprintf(buf, "%d\n", s->size);
4397 SLAB_ATTR_RO(slab_size);
4399 static ssize_t align_show(struct kmem_cache *s, char *buf)
4401 return sprintf(buf, "%d\n", s->align);
4403 SLAB_ATTR_RO(align);
4405 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4407 return sprintf(buf, "%d\n", s->object_size);
4409 SLAB_ATTR_RO(object_size);
4411 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4413 return sprintf(buf, "%d\n", oo_objects(s->oo));
4415 SLAB_ATTR_RO(objs_per_slab);
4417 static ssize_t order_store(struct kmem_cache *s,
4418 const char *buf, size_t length)
4420 unsigned long order;
4423 err = strict_strtoul(buf, 10, &order);
4427 if (order > slub_max_order || order < slub_min_order)
4430 calculate_sizes(s, order);
4434 static ssize_t order_show(struct kmem_cache *s, char *buf)
4436 return sprintf(buf, "%d\n", oo_order(s->oo));
4440 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4442 return sprintf(buf, "%lu\n", s->min_partial);
4445 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4451 err = strict_strtoul(buf, 10, &min);
4455 set_min_partial(s, min);
4458 SLAB_ATTR(min_partial);
4460 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4462 return sprintf(buf, "%u\n", s->cpu_partial);
4465 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4468 unsigned long objects;
4471 err = strict_strtoul(buf, 10, &objects);
4474 if (objects && !kmem_cache_has_cpu_partial(s))
4477 s->cpu_partial = objects;
4481 SLAB_ATTR(cpu_partial);
4483 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4487 return sprintf(buf, "%pS\n", s->ctor);
4491 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4493 return sprintf(buf, "%d\n", s->refcount - 1);
4495 SLAB_ATTR_RO(aliases);
4497 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4499 return show_slab_objects(s, buf, SO_PARTIAL);
4501 SLAB_ATTR_RO(partial);
4503 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4505 return show_slab_objects(s, buf, SO_CPU);
4507 SLAB_ATTR_RO(cpu_slabs);
4509 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4511 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4513 SLAB_ATTR_RO(objects);
4515 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4517 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4519 SLAB_ATTR_RO(objects_partial);
4521 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4528 for_each_online_cpu(cpu) {
4529 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4532 pages += page->pages;
4533 objects += page->pobjects;
4537 len = sprintf(buf, "%d(%d)", objects, pages);
4540 for_each_online_cpu(cpu) {
4541 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4543 if (page && len < PAGE_SIZE - 20)
4544 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4545 page->pobjects, page->pages);
4548 return len + sprintf(buf + len, "\n");
4550 SLAB_ATTR_RO(slabs_cpu_partial);
4552 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4554 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4557 static ssize_t reclaim_account_store(struct kmem_cache *s,
4558 const char *buf, size_t length)
4560 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4562 s->flags |= SLAB_RECLAIM_ACCOUNT;
4565 SLAB_ATTR(reclaim_account);
4567 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4569 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4571 SLAB_ATTR_RO(hwcache_align);
4573 #ifdef CONFIG_ZONE_DMA
4574 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4578 SLAB_ATTR_RO(cache_dma);
4581 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4583 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4585 SLAB_ATTR_RO(destroy_by_rcu);
4587 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4589 return sprintf(buf, "%d\n", s->reserved);
4591 SLAB_ATTR_RO(reserved);
4593 #ifdef CONFIG_SLUB_DEBUG
4594 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4596 return show_slab_objects(s, buf, SO_ALL);
4598 SLAB_ATTR_RO(slabs);
4600 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4602 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4604 SLAB_ATTR_RO(total_objects);
4606 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4608 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4611 static ssize_t sanity_checks_store(struct kmem_cache *s,
4612 const char *buf, size_t length)
4614 s->flags &= ~SLAB_DEBUG_FREE;
4615 if (buf[0] == '1') {
4616 s->flags &= ~__CMPXCHG_DOUBLE;
4617 s->flags |= SLAB_DEBUG_FREE;
4621 SLAB_ATTR(sanity_checks);
4623 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4625 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4628 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4631 s->flags &= ~SLAB_TRACE;
4632 if (buf[0] == '1') {
4633 s->flags &= ~__CMPXCHG_DOUBLE;
4634 s->flags |= SLAB_TRACE;
4640 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4642 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4645 static ssize_t red_zone_store(struct kmem_cache *s,
4646 const char *buf, size_t length)
4648 if (any_slab_objects(s))
4651 s->flags &= ~SLAB_RED_ZONE;
4652 if (buf[0] == '1') {
4653 s->flags &= ~__CMPXCHG_DOUBLE;
4654 s->flags |= SLAB_RED_ZONE;
4656 calculate_sizes(s, -1);
4659 SLAB_ATTR(red_zone);
4661 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4663 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4666 static ssize_t poison_store(struct kmem_cache *s,
4667 const char *buf, size_t length)
4669 if (any_slab_objects(s))
4672 s->flags &= ~SLAB_POISON;
4673 if (buf[0] == '1') {
4674 s->flags &= ~__CMPXCHG_DOUBLE;
4675 s->flags |= SLAB_POISON;
4677 calculate_sizes(s, -1);
4682 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4684 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4687 static ssize_t store_user_store(struct kmem_cache *s,
4688 const char *buf, size_t length)
4690 if (any_slab_objects(s))
4693 s->flags &= ~SLAB_STORE_USER;
4694 if (buf[0] == '1') {
4695 s->flags &= ~__CMPXCHG_DOUBLE;
4696 s->flags |= SLAB_STORE_USER;
4698 calculate_sizes(s, -1);
4701 SLAB_ATTR(store_user);
4703 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4708 static ssize_t validate_store(struct kmem_cache *s,
4709 const char *buf, size_t length)
4713 if (buf[0] == '1') {
4714 ret = validate_slab_cache(s);
4720 SLAB_ATTR(validate);
4722 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4724 if (!(s->flags & SLAB_STORE_USER))
4726 return list_locations(s, buf, TRACK_ALLOC);
4728 SLAB_ATTR_RO(alloc_calls);
4730 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4732 if (!(s->flags & SLAB_STORE_USER))
4734 return list_locations(s, buf, TRACK_FREE);
4736 SLAB_ATTR_RO(free_calls);
4737 #endif /* CONFIG_SLUB_DEBUG */
4739 #ifdef CONFIG_FAILSLAB
4740 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4742 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4745 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4748 s->flags &= ~SLAB_FAILSLAB;
4750 s->flags |= SLAB_FAILSLAB;
4753 SLAB_ATTR(failslab);
4756 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4761 static ssize_t shrink_store(struct kmem_cache *s,
4762 const char *buf, size_t length)
4764 if (buf[0] == '1') {
4765 int rc = kmem_cache_shrink(s);
4776 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4778 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4781 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4782 const char *buf, size_t length)
4784 unsigned long ratio;
4787 err = strict_strtoul(buf, 10, &ratio);
4792 s->remote_node_defrag_ratio = ratio * 10;
4796 SLAB_ATTR(remote_node_defrag_ratio);
4799 #ifdef CONFIG_SLUB_STATS
4800 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4802 unsigned long sum = 0;
4805 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4810 for_each_online_cpu(cpu) {
4811 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4817 len = sprintf(buf, "%lu", sum);
4820 for_each_online_cpu(cpu) {
4821 if (data[cpu] && len < PAGE_SIZE - 20)
4822 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4826 return len + sprintf(buf + len, "\n");
4829 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4833 for_each_online_cpu(cpu)
4834 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4837 #define STAT_ATTR(si, text) \
4838 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4840 return show_stat(s, buf, si); \
4842 static ssize_t text##_store(struct kmem_cache *s, \
4843 const char *buf, size_t length) \
4845 if (buf[0] != '0') \
4847 clear_stat(s, si); \
4852 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4853 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4854 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4855 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4856 STAT_ATTR(FREE_FROZEN, free_frozen);
4857 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4858 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4859 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4860 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4861 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4862 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4863 STAT_ATTR(FREE_SLAB, free_slab);
4864 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4865 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4866 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4867 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4868 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4869 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4870 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4871 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4872 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4873 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4874 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4875 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4876 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4877 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4880 static struct attribute *slab_attrs[] = {
4881 &slab_size_attr.attr,
4882 &object_size_attr.attr,
4883 &objs_per_slab_attr.attr,
4885 &min_partial_attr.attr,
4886 &cpu_partial_attr.attr,
4888 &objects_partial_attr.attr,
4890 &cpu_slabs_attr.attr,
4894 &hwcache_align_attr.attr,
4895 &reclaim_account_attr.attr,
4896 &destroy_by_rcu_attr.attr,
4898 &reserved_attr.attr,
4899 &slabs_cpu_partial_attr.attr,
4900 #ifdef CONFIG_SLUB_DEBUG
4901 &total_objects_attr.attr,
4903 &sanity_checks_attr.attr,
4905 &red_zone_attr.attr,
4907 &store_user_attr.attr,
4908 &validate_attr.attr,
4909 &alloc_calls_attr.attr,
4910 &free_calls_attr.attr,
4912 #ifdef CONFIG_ZONE_DMA
4913 &cache_dma_attr.attr,
4916 &remote_node_defrag_ratio_attr.attr,
4918 #ifdef CONFIG_SLUB_STATS
4919 &alloc_fastpath_attr.attr,
4920 &alloc_slowpath_attr.attr,
4921 &free_fastpath_attr.attr,
4922 &free_slowpath_attr.attr,
4923 &free_frozen_attr.attr,
4924 &free_add_partial_attr.attr,
4925 &free_remove_partial_attr.attr,
4926 &alloc_from_partial_attr.attr,
4927 &alloc_slab_attr.attr,
4928 &alloc_refill_attr.attr,
4929 &alloc_node_mismatch_attr.attr,
4930 &free_slab_attr.attr,
4931 &cpuslab_flush_attr.attr,
4932 &deactivate_full_attr.attr,
4933 &deactivate_empty_attr.attr,
4934 &deactivate_to_head_attr.attr,
4935 &deactivate_to_tail_attr.attr,
4936 &deactivate_remote_frees_attr.attr,
4937 &deactivate_bypass_attr.attr,
4938 &order_fallback_attr.attr,
4939 &cmpxchg_double_fail_attr.attr,
4940 &cmpxchg_double_cpu_fail_attr.attr,
4941 &cpu_partial_alloc_attr.attr,
4942 &cpu_partial_free_attr.attr,
4943 &cpu_partial_node_attr.attr,
4944 &cpu_partial_drain_attr.attr,
4946 #ifdef CONFIG_FAILSLAB
4947 &failslab_attr.attr,
4953 static struct attribute_group slab_attr_group = {
4954 .attrs = slab_attrs,
4957 static ssize_t slab_attr_show(struct kobject *kobj,
4958 struct attribute *attr,
4961 struct slab_attribute *attribute;
4962 struct kmem_cache *s;
4965 attribute = to_slab_attr(attr);
4968 if (!attribute->show)
4971 err = attribute->show(s, buf);
4976 static ssize_t slab_attr_store(struct kobject *kobj,
4977 struct attribute *attr,
4978 const char *buf, size_t len)
4980 struct slab_attribute *attribute;
4981 struct kmem_cache *s;
4984 attribute = to_slab_attr(attr);
4987 if (!attribute->store)
4990 err = attribute->store(s, buf, len);
4991 #ifdef CONFIG_MEMCG_KMEM
4992 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4995 mutex_lock(&slab_mutex);
4996 if (s->max_attr_size < len)
4997 s->max_attr_size = len;
5000 * This is a best effort propagation, so this function's return
5001 * value will be determined by the parent cache only. This is
5002 * basically because not all attributes will have a well
5003 * defined semantics for rollbacks - most of the actions will
5004 * have permanent effects.
5006 * Returning the error value of any of the children that fail
5007 * is not 100 % defined, in the sense that users seeing the
5008 * error code won't be able to know anything about the state of
5011 * Only returning the error code for the parent cache at least
5012 * has well defined semantics. The cache being written to
5013 * directly either failed or succeeded, in which case we loop
5014 * through the descendants with best-effort propagation.
5016 for_each_memcg_cache_index(i) {
5017 struct kmem_cache *c = cache_from_memcg(s, i);
5019 attribute->store(c, buf, len);
5021 mutex_unlock(&slab_mutex);
5027 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5029 #ifdef CONFIG_MEMCG_KMEM
5031 char *buffer = NULL;
5033 if (!is_root_cache(s))
5037 * This mean this cache had no attribute written. Therefore, no point
5038 * in copying default values around
5040 if (!s->max_attr_size)
5043 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5046 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5048 if (!attr || !attr->store || !attr->show)
5052 * It is really bad that we have to allocate here, so we will
5053 * do it only as a fallback. If we actually allocate, though,
5054 * we can just use the allocated buffer until the end.
5056 * Most of the slub attributes will tend to be very small in
5057 * size, but sysfs allows buffers up to a page, so they can
5058 * theoretically happen.
5062 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5065 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5066 if (WARN_ON(!buffer))
5071 attr->show(s->memcg_params->root_cache, buf);
5072 attr->store(s, buf, strlen(buf));
5076 free_page((unsigned long)buffer);
5080 static const struct sysfs_ops slab_sysfs_ops = {
5081 .show = slab_attr_show,
5082 .store = slab_attr_store,
5085 static struct kobj_type slab_ktype = {
5086 .sysfs_ops = &slab_sysfs_ops,
5089 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5091 struct kobj_type *ktype = get_ktype(kobj);
5093 if (ktype == &slab_ktype)
5098 static const struct kset_uevent_ops slab_uevent_ops = {
5099 .filter = uevent_filter,
5102 static struct kset *slab_kset;
5104 #define ID_STR_LENGTH 64
5106 /* Create a unique string id for a slab cache:
5108 * Format :[flags-]size
5110 static char *create_unique_id(struct kmem_cache *s)
5112 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5119 * First flags affecting slabcache operations. We will only
5120 * get here for aliasable slabs so we do not need to support
5121 * too many flags. The flags here must cover all flags that
5122 * are matched during merging to guarantee that the id is
5125 if (s->flags & SLAB_CACHE_DMA)
5127 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5129 if (s->flags & SLAB_DEBUG_FREE)
5131 if (!(s->flags & SLAB_NOTRACK))
5135 p += sprintf(p, "%07d", s->size);
5137 #ifdef CONFIG_MEMCG_KMEM
5138 if (!is_root_cache(s))
5139 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5142 BUG_ON(p > name + ID_STR_LENGTH - 1);
5146 static int sysfs_slab_add(struct kmem_cache *s)
5150 int unmergeable = slab_unmergeable(s);
5154 * Slabcache can never be merged so we can use the name proper.
5155 * This is typically the case for debug situations. In that
5156 * case we can catch duplicate names easily.
5158 sysfs_remove_link(&slab_kset->kobj, s->name);
5162 * Create a unique name for the slab as a target
5165 name = create_unique_id(s);
5168 s->kobj.kset = slab_kset;
5169 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5171 kobject_put(&s->kobj);
5175 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5177 kobject_del(&s->kobj);
5178 kobject_put(&s->kobj);
5181 kobject_uevent(&s->kobj, KOBJ_ADD);
5183 /* Setup first alias */
5184 sysfs_slab_alias(s, s->name);
5190 static void sysfs_slab_remove(struct kmem_cache *s)
5192 if (slab_state < FULL)
5194 * Sysfs has not been setup yet so no need to remove the
5199 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5200 kobject_del(&s->kobj);
5201 kobject_put(&s->kobj);
5205 * Need to buffer aliases during bootup until sysfs becomes
5206 * available lest we lose that information.
5208 struct saved_alias {
5209 struct kmem_cache *s;
5211 struct saved_alias *next;
5214 static struct saved_alias *alias_list;
5216 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5218 struct saved_alias *al;
5220 if (slab_state == FULL) {
5222 * If we have a leftover link then remove it.
5224 sysfs_remove_link(&slab_kset->kobj, name);
5225 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5228 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5234 al->next = alias_list;
5239 static int __init slab_sysfs_init(void)
5241 struct kmem_cache *s;
5244 mutex_lock(&slab_mutex);
5246 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5248 mutex_unlock(&slab_mutex);
5249 printk(KERN_ERR "Cannot register slab subsystem.\n");
5255 list_for_each_entry(s, &slab_caches, list) {
5256 err = sysfs_slab_add(s);
5258 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5259 " to sysfs\n", s->name);
5262 while (alias_list) {
5263 struct saved_alias *al = alias_list;
5265 alias_list = alias_list->next;
5266 err = sysfs_slab_alias(al->s, al->name);
5268 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5269 " %s to sysfs\n", al->name);
5273 mutex_unlock(&slab_mutex);
5278 __initcall(slab_sysfs_init);
5279 #endif /* CONFIG_SYSFS */
5282 * The /proc/slabinfo ABI
5284 #ifdef CONFIG_SLABINFO
5285 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5287 unsigned long nr_slabs = 0;
5288 unsigned long nr_objs = 0;
5289 unsigned long nr_free = 0;
5292 for_each_online_node(node) {
5293 struct kmem_cache_node *n = get_node(s, node);
5298 nr_slabs += node_nr_slabs(n);
5299 nr_objs += node_nr_objs(n);
5300 nr_free += count_partial(n, count_free);
5303 sinfo->active_objs = nr_objs - nr_free;
5304 sinfo->num_objs = nr_objs;
5305 sinfo->active_slabs = nr_slabs;
5306 sinfo->num_slabs = nr_slabs;
5307 sinfo->objects_per_slab = oo_objects(s->oo);
5308 sinfo->cache_order = oo_order(s->oo);
5311 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5315 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5316 size_t count, loff_t *ppos)
5320 #endif /* CONFIG_SLABINFO */