2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
136 * Issues still to be resolved:
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 * - Variable sizing of the per node arrays
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 #define MIN_PARTIAL 5
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in the.
160 #define MAX_PARTIAL 10
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
173 * Set of flags that will prevent slab merging
175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
183 #define OO_MASK ((1 << OO_SHIFT) - 1)
184 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
186 /* Internal SLUB flags */
187 #define __OBJECT_POISON 0x80000000UL /* Poison object */
188 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
191 static struct notifier_block slab_notifier;
195 * Tracking user of a slab.
197 #define TRACK_ADDRS_COUNT 16
199 unsigned long addr; /* Called from address */
200 #ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
208 enum track_item { TRACK_ALLOC, TRACK_FREE };
211 static int sysfs_slab_add(struct kmem_cache *);
212 static int sysfs_slab_alias(struct kmem_cache *, const char *);
213 static void sysfs_slab_remove(struct kmem_cache *);
214 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
216 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
217 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
219 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
221 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
224 static inline void stat(const struct kmem_cache *s, enum stat_item si)
226 #ifdef CONFIG_SLUB_STATS
227 __this_cpu_inc(s->cpu_slab->stat[si]);
231 /********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
235 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
237 return s->node[node];
240 /* Verify that a pointer has an address that is valid within a slab page */
241 static inline int check_valid_pointer(struct kmem_cache *s,
242 struct page *page, const void *object)
249 base = page_address(page);
250 if (object < base || object >= base + page->objects * s->size ||
251 (object - base) % s->size) {
258 static inline void *get_freepointer(struct kmem_cache *s, void *object)
260 return *(void **)(object + s->offset);
263 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
265 prefetch(object + s->offset);
268 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
272 #ifdef CONFIG_DEBUG_PAGEALLOC
273 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
275 p = get_freepointer(s, object);
280 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
282 *(void **)(object + s->offset) = fp;
285 /* Loop over all objects in a slab */
286 #define for_each_object(__p, __s, __addr, __objects) \
287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
290 /* Determine object index from a given position */
291 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293 return (p - addr) / s->size;
296 static inline size_t slab_ksize(const struct kmem_cache *s)
298 #ifdef CONFIG_SLUB_DEBUG
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
304 return s->object_size;
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
315 * Else we can use all the padding etc for the allocation
320 static inline int order_objects(int order, unsigned long size, int reserved)
322 return ((PAGE_SIZE << order) - reserved) / size;
325 static inline struct kmem_cache_order_objects oo_make(int order,
326 unsigned long size, int reserved)
328 struct kmem_cache_order_objects x = {
329 (order << OO_SHIFT) + order_objects(order, size, reserved)
335 static inline int oo_order(struct kmem_cache_order_objects x)
337 return x.x >> OO_SHIFT;
340 static inline int oo_objects(struct kmem_cache_order_objects x)
342 return x.x & OO_MASK;
346 * Per slab locking using the pagelock
348 static __always_inline void slab_lock(struct page *page)
350 bit_spin_lock(PG_locked, &page->flags);
353 static __always_inline void slab_unlock(struct page *page)
355 __bit_spin_unlock(PG_locked, &page->flags);
358 /* Interrupts must be disabled (for the fallback code to work right) */
359 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
360 void *freelist_old, unsigned long counters_old,
361 void *freelist_new, unsigned long counters_new,
364 VM_BUG_ON(!irqs_disabled());
365 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
367 if (s->flags & __CMPXCHG_DOUBLE) {
368 if (cmpxchg_double(&page->freelist, &page->counters,
369 freelist_old, counters_old,
370 freelist_new, counters_new))
376 if (page->freelist == freelist_old &&
377 page->counters == counters_old) {
378 page->freelist = freelist_new;
379 page->counters = counters_new;
387 stat(s, CMPXCHG_DOUBLE_FAIL);
389 #ifdef SLUB_DEBUG_CMPXCHG
390 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
396 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
397 void *freelist_old, unsigned long counters_old,
398 void *freelist_new, unsigned long counters_new,
401 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
402 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
403 if (s->flags & __CMPXCHG_DOUBLE) {
404 if (cmpxchg_double(&page->freelist, &page->counters,
405 freelist_old, counters_old,
406 freelist_new, counters_new))
413 local_irq_save(flags);
415 if (page->freelist == freelist_old &&
416 page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
420 local_irq_restore(flags);
424 local_irq_restore(flags);
428 stat(s, CMPXCHG_DOUBLE_FAIL);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
437 #ifdef CONFIG_SLUB_DEBUG
439 * Determine a map of object in use on a page.
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
444 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
447 void *addr = page_address(page);
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
456 #ifdef CONFIG_SLUB_DEBUG_ON
457 static int slub_debug = DEBUG_DEFAULT_FLAGS;
459 static int slub_debug;
462 static char *slub_debug_slabs;
463 static int disable_higher_order_debug;
468 static void print_section(char *text, u8 *addr, unsigned int length)
470 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
474 static struct track *get_track(struct kmem_cache *s, void *object,
475 enum track_item alloc)
480 p = object + s->offset + sizeof(void *);
482 p = object + s->inuse;
487 static void set_track(struct kmem_cache *s, void *object,
488 enum track_item alloc, unsigned long addr)
490 struct track *p = get_track(s, object, alloc);
493 #ifdef CONFIG_STACKTRACE
494 struct stack_trace trace;
497 trace.nr_entries = 0;
498 trace.max_entries = TRACK_ADDRS_COUNT;
499 trace.entries = p->addrs;
501 save_stack_trace(&trace);
503 /* See rant in lockdep.c */
504 if (trace.nr_entries != 0 &&
505 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
508 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
512 p->cpu = smp_processor_id();
513 p->pid = current->pid;
516 memset(p, 0, sizeof(struct track));
519 static void init_tracking(struct kmem_cache *s, void *object)
521 if (!(s->flags & SLAB_STORE_USER))
524 set_track(s, object, TRACK_FREE, 0UL);
525 set_track(s, object, TRACK_ALLOC, 0UL);
528 static void print_track(const char *s, struct track *t)
533 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
534 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
535 #ifdef CONFIG_STACKTRACE
538 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
540 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
547 static void print_tracking(struct kmem_cache *s, void *object)
549 if (!(s->flags & SLAB_STORE_USER))
552 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
553 print_track("Freed", get_track(s, object, TRACK_FREE));
556 static void print_page_info(struct page *page)
559 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
560 page, page->objects, page->inuse, page->freelist, page->flags);
564 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
570 vsnprintf(buf, sizeof(buf), fmt, args);
572 printk(KERN_ERR "========================================"
573 "=====================================\n");
574 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
575 printk(KERN_ERR "----------------------------------------"
576 "-------------------------------------\n\n");
578 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
581 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
587 vsnprintf(buf, sizeof(buf), fmt, args);
589 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
592 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
594 unsigned int off; /* Offset of last byte */
595 u8 *addr = page_address(page);
597 print_tracking(s, p);
599 print_page_info(page);
601 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
602 p, p - addr, get_freepointer(s, p));
605 print_section("Bytes b4 ", p - 16, 16);
607 print_section("Object ", p, min_t(unsigned long, s->object_size,
609 if (s->flags & SLAB_RED_ZONE)
610 print_section("Redzone ", p + s->object_size,
611 s->inuse - s->object_size);
614 off = s->offset + sizeof(void *);
618 if (s->flags & SLAB_STORE_USER)
619 off += 2 * sizeof(struct track);
622 /* Beginning of the filler is the free pointer */
623 print_section("Padding ", p + off, s->size - off);
628 static void object_err(struct kmem_cache *s, struct page *page,
629 u8 *object, char *reason)
631 slab_bug(s, "%s", reason);
632 print_trailer(s, page, object);
635 static void slab_err(struct kmem_cache *s, struct page *page,
636 const char *fmt, ...)
642 vsnprintf(buf, sizeof(buf), fmt, args);
644 slab_bug(s, "%s", buf);
645 print_page_info(page);
649 static void init_object(struct kmem_cache *s, void *object, u8 val)
653 if (s->flags & __OBJECT_POISON) {
654 memset(p, POISON_FREE, s->object_size - 1);
655 p[s->object_size - 1] = POISON_END;
658 if (s->flags & SLAB_RED_ZONE)
659 memset(p + s->object_size, val, s->inuse - s->object_size);
662 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
663 void *from, void *to)
665 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
666 memset(from, data, to - from);
669 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
670 u8 *object, char *what,
671 u8 *start, unsigned int value, unsigned int bytes)
676 fault = memchr_inv(start, value, bytes);
681 while (end > fault && end[-1] == value)
684 slab_bug(s, "%s overwritten", what);
685 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
686 fault, end - 1, fault[0], value);
687 print_trailer(s, page, object);
689 restore_bytes(s, what, value, fault, end);
697 * Bytes of the object to be managed.
698 * If the freepointer may overlay the object then the free
699 * pointer is the first word of the object.
701 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
704 * object + s->object_size
705 * Padding to reach word boundary. This is also used for Redzoning.
706 * Padding is extended by another word if Redzoning is enabled and
707 * object_size == inuse.
709 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
710 * 0xcc (RED_ACTIVE) for objects in use.
713 * Meta data starts here.
715 * A. Free pointer (if we cannot overwrite object on free)
716 * B. Tracking data for SLAB_STORE_USER
717 * C. Padding to reach required alignment boundary or at mininum
718 * one word if debugging is on to be able to detect writes
719 * before the word boundary.
721 * Padding is done using 0x5a (POISON_INUSE)
724 * Nothing is used beyond s->size.
726 * If slabcaches are merged then the object_size and inuse boundaries are mostly
727 * ignored. And therefore no slab options that rely on these boundaries
728 * may be used with merged slabcaches.
731 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
733 unsigned long off = s->inuse; /* The end of info */
736 /* Freepointer is placed after the object. */
737 off += sizeof(void *);
739 if (s->flags & SLAB_STORE_USER)
740 /* We also have user information there */
741 off += 2 * sizeof(struct track);
746 return check_bytes_and_report(s, page, p, "Object padding",
747 p + off, POISON_INUSE, s->size - off);
750 /* Check the pad bytes at the end of a slab page */
751 static int slab_pad_check(struct kmem_cache *s, struct page *page)
759 if (!(s->flags & SLAB_POISON))
762 start = page_address(page);
763 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
764 end = start + length;
765 remainder = length % s->size;
769 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
772 while (end > fault && end[-1] == POISON_INUSE)
775 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
776 print_section("Padding ", end - remainder, remainder);
778 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
782 static int check_object(struct kmem_cache *s, struct page *page,
783 void *object, u8 val)
786 u8 *endobject = object + s->object_size;
788 if (s->flags & SLAB_RED_ZONE) {
789 if (!check_bytes_and_report(s, page, object, "Redzone",
790 endobject, val, s->inuse - s->object_size))
793 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
794 check_bytes_and_report(s, page, p, "Alignment padding",
795 endobject, POISON_INUSE,
796 s->inuse - s->object_size);
800 if (s->flags & SLAB_POISON) {
801 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
802 (!check_bytes_and_report(s, page, p, "Poison", p,
803 POISON_FREE, s->object_size - 1) ||
804 !check_bytes_and_report(s, page, p, "Poison",
805 p + s->object_size - 1, POISON_END, 1)))
808 * check_pad_bytes cleans up on its own.
810 check_pad_bytes(s, page, p);
813 if (!s->offset && val == SLUB_RED_ACTIVE)
815 * Object and freepointer overlap. Cannot check
816 * freepointer while object is allocated.
820 /* Check free pointer validity */
821 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
822 object_err(s, page, p, "Freepointer corrupt");
824 * No choice but to zap it and thus lose the remainder
825 * of the free objects in this slab. May cause
826 * another error because the object count is now wrong.
828 set_freepointer(s, p, NULL);
834 static int check_slab(struct kmem_cache *s, struct page *page)
838 VM_BUG_ON(!irqs_disabled());
840 if (!PageSlab(page)) {
841 slab_err(s, page, "Not a valid slab page");
845 maxobj = order_objects(compound_order(page), s->size, s->reserved);
846 if (page->objects > maxobj) {
847 slab_err(s, page, "objects %u > max %u",
848 s->name, page->objects, maxobj);
851 if (page->inuse > page->objects) {
852 slab_err(s, page, "inuse %u > max %u",
853 s->name, page->inuse, page->objects);
856 /* Slab_pad_check fixes things up after itself */
857 slab_pad_check(s, page);
862 * Determine if a certain object on a page is on the freelist. Must hold the
863 * slab lock to guarantee that the chains are in a consistent state.
865 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
870 unsigned long max_objects;
873 while (fp && nr <= page->objects) {
876 if (!check_valid_pointer(s, page, fp)) {
878 object_err(s, page, object,
879 "Freechain corrupt");
880 set_freepointer(s, object, NULL);
882 slab_err(s, page, "Freepointer corrupt");
883 page->freelist = NULL;
884 page->inuse = page->objects;
885 slab_fix(s, "Freelist cleared");
891 fp = get_freepointer(s, object);
895 max_objects = order_objects(compound_order(page), s->size, s->reserved);
896 if (max_objects > MAX_OBJS_PER_PAGE)
897 max_objects = MAX_OBJS_PER_PAGE;
899 if (page->objects != max_objects) {
900 slab_err(s, page, "Wrong number of objects. Found %d but "
901 "should be %d", page->objects, max_objects);
902 page->objects = max_objects;
903 slab_fix(s, "Number of objects adjusted.");
905 if (page->inuse != page->objects - nr) {
906 slab_err(s, page, "Wrong object count. Counter is %d but "
907 "counted were %d", page->inuse, page->objects - nr);
908 page->inuse = page->objects - nr;
909 slab_fix(s, "Object count adjusted.");
911 return search == NULL;
914 static void trace(struct kmem_cache *s, struct page *page, void *object,
917 if (s->flags & SLAB_TRACE) {
918 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
920 alloc ? "alloc" : "free",
925 print_section("Object ", (void *)object,
933 * Hooks for other subsystems that check memory allocations. In a typical
934 * production configuration these hooks all should produce no code at all.
936 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
938 flags &= gfp_allowed_mask;
939 lockdep_trace_alloc(flags);
940 might_sleep_if(flags & __GFP_WAIT);
942 return should_failslab(s->object_size, flags, s->flags);
945 static inline void slab_post_alloc_hook(struct kmem_cache *s,
946 gfp_t flags, void *object)
948 flags &= gfp_allowed_mask;
949 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
950 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
953 static inline void slab_free_hook(struct kmem_cache *s, void *x)
955 kmemleak_free_recursive(x, s->flags);
958 * Trouble is that we may no longer disable interupts in the fast path
959 * So in order to make the debug calls that expect irqs to be
960 * disabled we need to disable interrupts temporarily.
962 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
966 local_irq_save(flags);
967 kmemcheck_slab_free(s, x, s->object_size);
968 debug_check_no_locks_freed(x, s->object_size);
969 local_irq_restore(flags);
972 if (!(s->flags & SLAB_DEBUG_OBJECTS))
973 debug_check_no_obj_freed(x, s->object_size);
977 * Tracking of fully allocated slabs for debugging purposes.
979 * list_lock must be held.
981 static void add_full(struct kmem_cache *s,
982 struct kmem_cache_node *n, struct page *page)
984 if (!(s->flags & SLAB_STORE_USER))
987 list_add(&page->lru, &n->full);
991 * list_lock must be held.
993 static void remove_full(struct kmem_cache *s, struct page *page)
995 if (!(s->flags & SLAB_STORE_USER))
998 list_del(&page->lru);
1001 /* Tracking of the number of slabs for debugging purposes */
1002 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1004 struct kmem_cache_node *n = get_node(s, node);
1006 return atomic_long_read(&n->nr_slabs);
1009 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1011 return atomic_long_read(&n->nr_slabs);
1014 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1016 struct kmem_cache_node *n = get_node(s, node);
1019 * May be called early in order to allocate a slab for the
1020 * kmem_cache_node structure. Solve the chicken-egg
1021 * dilemma by deferring the increment of the count during
1022 * bootstrap (see early_kmem_cache_node_alloc).
1025 atomic_long_inc(&n->nr_slabs);
1026 atomic_long_add(objects, &n->total_objects);
1029 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1031 struct kmem_cache_node *n = get_node(s, node);
1033 atomic_long_dec(&n->nr_slabs);
1034 atomic_long_sub(objects, &n->total_objects);
1037 /* Object debug checks for alloc/free paths */
1038 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1041 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1044 init_object(s, object, SLUB_RED_INACTIVE);
1045 init_tracking(s, object);
1048 static noinline int alloc_debug_processing(struct kmem_cache *s,
1050 void *object, unsigned long addr)
1052 if (!check_slab(s, page))
1055 if (!check_valid_pointer(s, page, object)) {
1056 object_err(s, page, object, "Freelist Pointer check fails");
1060 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1063 /* Success perform special debug activities for allocs */
1064 if (s->flags & SLAB_STORE_USER)
1065 set_track(s, object, TRACK_ALLOC, addr);
1066 trace(s, page, object, 1);
1067 init_object(s, object, SLUB_RED_ACTIVE);
1071 if (PageSlab(page)) {
1073 * If this is a slab page then lets do the best we can
1074 * to avoid issues in the future. Marking all objects
1075 * as used avoids touching the remaining objects.
1077 slab_fix(s, "Marking all objects used");
1078 page->inuse = page->objects;
1079 page->freelist = NULL;
1084 static noinline struct kmem_cache_node *free_debug_processing(
1085 struct kmem_cache *s, struct page *page, void *object,
1086 unsigned long addr, unsigned long *flags)
1088 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1090 spin_lock_irqsave(&n->list_lock, *flags);
1093 if (!check_slab(s, page))
1096 if (!check_valid_pointer(s, page, object)) {
1097 slab_err(s, page, "Invalid object pointer 0x%p", object);
1101 if (on_freelist(s, page, object)) {
1102 object_err(s, page, object, "Object already free");
1106 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1109 if (unlikely(s != page->slab_cache)) {
1110 if (!PageSlab(page)) {
1111 slab_err(s, page, "Attempt to free object(0x%p) "
1112 "outside of slab", object);
1113 } else if (!page->slab_cache) {
1115 "SLUB <none>: no slab for object 0x%p.\n",
1119 object_err(s, page, object,
1120 "page slab pointer corrupt.");
1124 if (s->flags & SLAB_STORE_USER)
1125 set_track(s, object, TRACK_FREE, addr);
1126 trace(s, page, object, 0);
1127 init_object(s, object, SLUB_RED_INACTIVE);
1131 * Keep node_lock to preserve integrity
1132 * until the object is actually freed
1138 spin_unlock_irqrestore(&n->list_lock, *flags);
1139 slab_fix(s, "Object at 0x%p not freed", object);
1143 static int __init setup_slub_debug(char *str)
1145 slub_debug = DEBUG_DEFAULT_FLAGS;
1146 if (*str++ != '=' || !*str)
1148 * No options specified. Switch on full debugging.
1154 * No options but restriction on slabs. This means full
1155 * debugging for slabs matching a pattern.
1159 if (tolower(*str) == 'o') {
1161 * Avoid enabling debugging on caches if its minimum order
1162 * would increase as a result.
1164 disable_higher_order_debug = 1;
1171 * Switch off all debugging measures.
1176 * Determine which debug features should be switched on
1178 for (; *str && *str != ','; str++) {
1179 switch (tolower(*str)) {
1181 slub_debug |= SLAB_DEBUG_FREE;
1184 slub_debug |= SLAB_RED_ZONE;
1187 slub_debug |= SLAB_POISON;
1190 slub_debug |= SLAB_STORE_USER;
1193 slub_debug |= SLAB_TRACE;
1196 slub_debug |= SLAB_FAILSLAB;
1199 printk(KERN_ERR "slub_debug option '%c' "
1200 "unknown. skipped\n", *str);
1206 slub_debug_slabs = str + 1;
1211 __setup("slub_debug", setup_slub_debug);
1213 static unsigned long kmem_cache_flags(unsigned long object_size,
1214 unsigned long flags, const char *name,
1215 void (*ctor)(void *))
1218 * Enable debugging if selected on the kernel commandline.
1220 if (slub_debug && (!slub_debug_slabs ||
1221 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1222 flags |= slub_debug;
1227 static inline void setup_object_debug(struct kmem_cache *s,
1228 struct page *page, void *object) {}
1230 static inline int alloc_debug_processing(struct kmem_cache *s,
1231 struct page *page, void *object, unsigned long addr) { return 0; }
1233 static inline struct kmem_cache_node *free_debug_processing(
1234 struct kmem_cache *s, struct page *page, void *object,
1235 unsigned long addr, unsigned long *flags) { return NULL; }
1237 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1239 static inline int check_object(struct kmem_cache *s, struct page *page,
1240 void *object, u8 val) { return 1; }
1241 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1242 struct page *page) {}
1243 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1244 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1245 unsigned long flags, const char *name,
1246 void (*ctor)(void *))
1250 #define slub_debug 0
1252 #define disable_higher_order_debug 0
1254 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1256 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1258 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1260 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1263 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1266 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1269 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1271 #endif /* CONFIG_SLUB_DEBUG */
1274 * Slab allocation and freeing
1276 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1277 struct kmem_cache_order_objects oo)
1279 int order = oo_order(oo);
1281 flags |= __GFP_NOTRACK;
1283 if (node == NUMA_NO_NODE)
1284 return alloc_pages(flags, order);
1286 return alloc_pages_exact_node(node, flags, order);
1289 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1292 struct kmem_cache_order_objects oo = s->oo;
1295 flags &= gfp_allowed_mask;
1297 if (flags & __GFP_WAIT)
1300 flags |= s->allocflags;
1303 * Let the initial higher-order allocation fail under memory pressure
1304 * so we fall-back to the minimum order allocation.
1306 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1308 page = alloc_slab_page(alloc_gfp, node, oo);
1309 if (unlikely(!page)) {
1312 * Allocation may have failed due to fragmentation.
1313 * Try a lower order alloc if possible
1315 page = alloc_slab_page(flags, node, oo);
1318 stat(s, ORDER_FALLBACK);
1321 if (kmemcheck_enabled && page
1322 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1323 int pages = 1 << oo_order(oo);
1325 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1328 * Objects from caches that have a constructor don't get
1329 * cleared when they're allocated, so we need to do it here.
1332 kmemcheck_mark_uninitialized_pages(page, pages);
1334 kmemcheck_mark_unallocated_pages(page, pages);
1337 if (flags & __GFP_WAIT)
1338 local_irq_disable();
1342 page->objects = oo_objects(oo);
1343 mod_zone_page_state(page_zone(page),
1344 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1345 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1351 static void setup_object(struct kmem_cache *s, struct page *page,
1354 setup_object_debug(s, page, object);
1355 if (unlikely(s->ctor))
1359 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1367 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1369 page = allocate_slab(s,
1370 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1374 order = compound_order(page);
1375 inc_slabs_node(s, page_to_nid(page), page->objects);
1376 memcg_bind_pages(s, order);
1377 page->slab_cache = s;
1378 __SetPageSlab(page);
1379 if (page->pfmemalloc)
1380 SetPageSlabPfmemalloc(page);
1382 start = page_address(page);
1384 if (unlikely(s->flags & SLAB_POISON))
1385 memset(start, POISON_INUSE, PAGE_SIZE << order);
1388 for_each_object(p, s, start, page->objects) {
1389 setup_object(s, page, last);
1390 set_freepointer(s, last, p);
1393 setup_object(s, page, last);
1394 set_freepointer(s, last, NULL);
1396 page->freelist = start;
1397 page->inuse = page->objects;
1403 static void __free_slab(struct kmem_cache *s, struct page *page)
1405 int order = compound_order(page);
1406 int pages = 1 << order;
1408 if (kmem_cache_debug(s)) {
1411 slab_pad_check(s, page);
1412 for_each_object(p, s, page_address(page),
1414 check_object(s, page, p, SLUB_RED_INACTIVE);
1417 kmemcheck_free_shadow(page, compound_order(page));
1419 mod_zone_page_state(page_zone(page),
1420 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1421 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1424 __ClearPageSlabPfmemalloc(page);
1425 __ClearPageSlab(page);
1427 memcg_release_pages(s, order);
1428 page_mapcount_reset(page);
1429 if (current->reclaim_state)
1430 current->reclaim_state->reclaimed_slab += pages;
1431 __free_memcg_kmem_pages(page, order);
1434 #define need_reserve_slab_rcu \
1435 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1437 static void rcu_free_slab(struct rcu_head *h)
1441 if (need_reserve_slab_rcu)
1442 page = virt_to_head_page(h);
1444 page = container_of((struct list_head *)h, struct page, lru);
1446 __free_slab(page->slab_cache, page);
1449 static void free_slab(struct kmem_cache *s, struct page *page)
1451 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1452 struct rcu_head *head;
1454 if (need_reserve_slab_rcu) {
1455 int order = compound_order(page);
1456 int offset = (PAGE_SIZE << order) - s->reserved;
1458 VM_BUG_ON(s->reserved != sizeof(*head));
1459 head = page_address(page) + offset;
1462 * RCU free overloads the RCU head over the LRU
1464 head = (void *)&page->lru;
1467 call_rcu(head, rcu_free_slab);
1469 __free_slab(s, page);
1472 static void discard_slab(struct kmem_cache *s, struct page *page)
1474 dec_slabs_node(s, page_to_nid(page), page->objects);
1479 * Management of partially allocated slabs.
1481 * list_lock must be held.
1483 static inline void add_partial(struct kmem_cache_node *n,
1484 struct page *page, int tail)
1487 if (tail == DEACTIVATE_TO_TAIL)
1488 list_add_tail(&page->lru, &n->partial);
1490 list_add(&page->lru, &n->partial);
1494 * list_lock must be held.
1496 static inline void remove_partial(struct kmem_cache_node *n,
1499 list_del(&page->lru);
1504 * Remove slab from the partial list, freeze it and
1505 * return the pointer to the freelist.
1507 * Returns a list of objects or NULL if it fails.
1509 * Must hold list_lock since we modify the partial list.
1511 static inline void *acquire_slab(struct kmem_cache *s,
1512 struct kmem_cache_node *n, struct page *page,
1513 int mode, int *objects)
1516 unsigned long counters;
1520 * Zap the freelist and set the frozen bit.
1521 * The old freelist is the list of objects for the
1522 * per cpu allocation list.
1524 freelist = page->freelist;
1525 counters = page->counters;
1526 new.counters = counters;
1527 *objects = new.objects - new.inuse;
1529 new.inuse = page->objects;
1530 new.freelist = NULL;
1532 new.freelist = freelist;
1535 VM_BUG_ON(new.frozen);
1538 if (!__cmpxchg_double_slab(s, page,
1540 new.freelist, new.counters,
1544 remove_partial(n, page);
1549 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1550 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1553 * Try to allocate a partial slab from a specific node.
1555 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1556 struct kmem_cache_cpu *c, gfp_t flags)
1558 struct page *page, *page2;
1559 void *object = NULL;
1564 * Racy check. If we mistakenly see no partial slabs then we
1565 * just allocate an empty slab. If we mistakenly try to get a
1566 * partial slab and there is none available then get_partials()
1569 if (!n || !n->nr_partial)
1572 spin_lock(&n->list_lock);
1573 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1576 if (!pfmemalloc_match(page, flags))
1579 t = acquire_slab(s, n, page, object == NULL, &objects);
1583 available += objects;
1586 stat(s, ALLOC_FROM_PARTIAL);
1589 put_cpu_partial(s, page, 0);
1590 stat(s, CPU_PARTIAL_NODE);
1592 if (!kmem_cache_has_cpu_partial(s)
1593 || available > s->cpu_partial / 2)
1597 spin_unlock(&n->list_lock);
1602 * Get a page from somewhere. Search in increasing NUMA distances.
1604 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1605 struct kmem_cache_cpu *c)
1608 struct zonelist *zonelist;
1611 enum zone_type high_zoneidx = gfp_zone(flags);
1613 unsigned int cpuset_mems_cookie;
1616 * The defrag ratio allows a configuration of the tradeoffs between
1617 * inter node defragmentation and node local allocations. A lower
1618 * defrag_ratio increases the tendency to do local allocations
1619 * instead of attempting to obtain partial slabs from other nodes.
1621 * If the defrag_ratio is set to 0 then kmalloc() always
1622 * returns node local objects. If the ratio is higher then kmalloc()
1623 * may return off node objects because partial slabs are obtained
1624 * from other nodes and filled up.
1626 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1627 * defrag_ratio = 1000) then every (well almost) allocation will
1628 * first attempt to defrag slab caches on other nodes. This means
1629 * scanning over all nodes to look for partial slabs which may be
1630 * expensive if we do it every time we are trying to find a slab
1631 * with available objects.
1633 if (!s->remote_node_defrag_ratio ||
1634 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1638 cpuset_mems_cookie = get_mems_allowed();
1639 zonelist = node_zonelist(slab_node(), flags);
1640 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1641 struct kmem_cache_node *n;
1643 n = get_node(s, zone_to_nid(zone));
1645 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1646 n->nr_partial > s->min_partial) {
1647 object = get_partial_node(s, n, c, flags);
1650 * Return the object even if
1651 * put_mems_allowed indicated that
1652 * the cpuset mems_allowed was
1653 * updated in parallel. It's a
1654 * harmless race between the alloc
1655 * and the cpuset update.
1657 put_mems_allowed(cpuset_mems_cookie);
1662 } while (!put_mems_allowed(cpuset_mems_cookie));
1668 * Get a partial page, lock it and return it.
1670 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1671 struct kmem_cache_cpu *c)
1674 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1676 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1677 if (object || node != NUMA_NO_NODE)
1680 return get_any_partial(s, flags, c);
1683 #ifdef CONFIG_PREEMPT
1685 * Calculate the next globally unique transaction for disambiguiation
1686 * during cmpxchg. The transactions start with the cpu number and are then
1687 * incremented by CONFIG_NR_CPUS.
1689 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1692 * No preemption supported therefore also no need to check for
1698 static inline unsigned long next_tid(unsigned long tid)
1700 return tid + TID_STEP;
1703 static inline unsigned int tid_to_cpu(unsigned long tid)
1705 return tid % TID_STEP;
1708 static inline unsigned long tid_to_event(unsigned long tid)
1710 return tid / TID_STEP;
1713 static inline unsigned int init_tid(int cpu)
1718 static inline void note_cmpxchg_failure(const char *n,
1719 const struct kmem_cache *s, unsigned long tid)
1721 #ifdef SLUB_DEBUG_CMPXCHG
1722 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1724 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1726 #ifdef CONFIG_PREEMPT
1727 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1728 printk("due to cpu change %d -> %d\n",
1729 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1732 if (tid_to_event(tid) != tid_to_event(actual_tid))
1733 printk("due to cpu running other code. Event %ld->%ld\n",
1734 tid_to_event(tid), tid_to_event(actual_tid));
1736 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1737 actual_tid, tid, next_tid(tid));
1739 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1742 static void init_kmem_cache_cpus(struct kmem_cache *s)
1746 for_each_possible_cpu(cpu)
1747 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1751 * Remove the cpu slab
1753 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1756 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1757 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1759 enum slab_modes l = M_NONE, m = M_NONE;
1761 int tail = DEACTIVATE_TO_HEAD;
1765 if (page->freelist) {
1766 stat(s, DEACTIVATE_REMOTE_FREES);
1767 tail = DEACTIVATE_TO_TAIL;
1771 * Stage one: Free all available per cpu objects back
1772 * to the page freelist while it is still frozen. Leave the
1775 * There is no need to take the list->lock because the page
1778 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1780 unsigned long counters;
1783 prior = page->freelist;
1784 counters = page->counters;
1785 set_freepointer(s, freelist, prior);
1786 new.counters = counters;
1788 VM_BUG_ON(!new.frozen);
1790 } while (!__cmpxchg_double_slab(s, page,
1792 freelist, new.counters,
1793 "drain percpu freelist"));
1795 freelist = nextfree;
1799 * Stage two: Ensure that the page is unfrozen while the
1800 * list presence reflects the actual number of objects
1803 * We setup the list membership and then perform a cmpxchg
1804 * with the count. If there is a mismatch then the page
1805 * is not unfrozen but the page is on the wrong list.
1807 * Then we restart the process which may have to remove
1808 * the page from the list that we just put it on again
1809 * because the number of objects in the slab may have
1814 old.freelist = page->freelist;
1815 old.counters = page->counters;
1816 VM_BUG_ON(!old.frozen);
1818 /* Determine target state of the slab */
1819 new.counters = old.counters;
1822 set_freepointer(s, freelist, old.freelist);
1823 new.freelist = freelist;
1825 new.freelist = old.freelist;
1829 if (!new.inuse && n->nr_partial > s->min_partial)
1831 else if (new.freelist) {
1836 * Taking the spinlock removes the possiblity
1837 * that acquire_slab() will see a slab page that
1840 spin_lock(&n->list_lock);
1844 if (kmem_cache_debug(s) && !lock) {
1847 * This also ensures that the scanning of full
1848 * slabs from diagnostic functions will not see
1851 spin_lock(&n->list_lock);
1859 remove_partial(n, page);
1861 else if (l == M_FULL)
1863 remove_full(s, page);
1865 if (m == M_PARTIAL) {
1867 add_partial(n, page, tail);
1870 } else if (m == M_FULL) {
1872 stat(s, DEACTIVATE_FULL);
1873 add_full(s, n, page);
1879 if (!__cmpxchg_double_slab(s, page,
1880 old.freelist, old.counters,
1881 new.freelist, new.counters,
1886 spin_unlock(&n->list_lock);
1889 stat(s, DEACTIVATE_EMPTY);
1890 discard_slab(s, page);
1896 * Unfreeze all the cpu partial slabs.
1898 * This function must be called with interrupts disabled
1899 * for the cpu using c (or some other guarantee must be there
1900 * to guarantee no concurrent accesses).
1902 static void unfreeze_partials(struct kmem_cache *s,
1903 struct kmem_cache_cpu *c)
1905 #ifdef CONFIG_SLUB_CPU_PARTIAL
1906 struct kmem_cache_node *n = NULL, *n2 = NULL;
1907 struct page *page, *discard_page = NULL;
1909 while ((page = c->partial)) {
1913 c->partial = page->next;
1915 n2 = get_node(s, page_to_nid(page));
1918 spin_unlock(&n->list_lock);
1921 spin_lock(&n->list_lock);
1926 old.freelist = page->freelist;
1927 old.counters = page->counters;
1928 VM_BUG_ON(!old.frozen);
1930 new.counters = old.counters;
1931 new.freelist = old.freelist;
1935 } while (!__cmpxchg_double_slab(s, page,
1936 old.freelist, old.counters,
1937 new.freelist, new.counters,
1938 "unfreezing slab"));
1940 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1941 page->next = discard_page;
1942 discard_page = page;
1944 add_partial(n, page, DEACTIVATE_TO_TAIL);
1945 stat(s, FREE_ADD_PARTIAL);
1950 spin_unlock(&n->list_lock);
1952 while (discard_page) {
1953 page = discard_page;
1954 discard_page = discard_page->next;
1956 stat(s, DEACTIVATE_EMPTY);
1957 discard_slab(s, page);
1964 * Put a page that was just frozen (in __slab_free) into a partial page
1965 * slot if available. This is done without interrupts disabled and without
1966 * preemption disabled. The cmpxchg is racy and may put the partial page
1967 * onto a random cpus partial slot.
1969 * If we did not find a slot then simply move all the partials to the
1970 * per node partial list.
1972 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1974 #ifdef CONFIG_SLUB_CPU_PARTIAL
1975 struct page *oldpage;
1982 oldpage = this_cpu_read(s->cpu_slab->partial);
1985 pobjects = oldpage->pobjects;
1986 pages = oldpage->pages;
1987 if (drain && pobjects > s->cpu_partial) {
1988 unsigned long flags;
1990 * partial array is full. Move the existing
1991 * set to the per node partial list.
1993 local_irq_save(flags);
1994 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1995 local_irq_restore(flags);
1999 stat(s, CPU_PARTIAL_DRAIN);
2004 pobjects += page->objects - page->inuse;
2006 page->pages = pages;
2007 page->pobjects = pobjects;
2008 page->next = oldpage;
2010 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2015 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2017 stat(s, CPUSLAB_FLUSH);
2018 deactivate_slab(s, c->page, c->freelist);
2020 c->tid = next_tid(c->tid);
2028 * Called from IPI handler with interrupts disabled.
2030 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2032 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2038 unfreeze_partials(s, c);
2042 static void flush_cpu_slab(void *d)
2044 struct kmem_cache *s = d;
2046 __flush_cpu_slab(s, smp_processor_id());
2049 static bool has_cpu_slab(int cpu, void *info)
2051 struct kmem_cache *s = info;
2052 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2054 return c->page || c->partial;
2057 static void flush_all(struct kmem_cache *s)
2059 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2063 * Check if the objects in a per cpu structure fit numa
2064 * locality expectations.
2066 static inline int node_match(struct page *page, int node)
2069 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2075 static int count_free(struct page *page)
2077 return page->objects - page->inuse;
2080 static unsigned long count_partial(struct kmem_cache_node *n,
2081 int (*get_count)(struct page *))
2083 unsigned long flags;
2084 unsigned long x = 0;
2087 spin_lock_irqsave(&n->list_lock, flags);
2088 list_for_each_entry(page, &n->partial, lru)
2089 x += get_count(page);
2090 spin_unlock_irqrestore(&n->list_lock, flags);
2094 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2096 #ifdef CONFIG_SLUB_DEBUG
2097 return atomic_long_read(&n->total_objects);
2103 static noinline void
2104 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2109 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2111 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2112 "default order: %d, min order: %d\n", s->name, s->object_size,
2113 s->size, oo_order(s->oo), oo_order(s->min));
2115 if (oo_order(s->min) > get_order(s->object_size))
2116 printk(KERN_WARNING " %s debugging increased min order, use "
2117 "slub_debug=O to disable.\n", s->name);
2119 for_each_online_node(node) {
2120 struct kmem_cache_node *n = get_node(s, node);
2121 unsigned long nr_slabs;
2122 unsigned long nr_objs;
2123 unsigned long nr_free;
2128 nr_free = count_partial(n, count_free);
2129 nr_slabs = node_nr_slabs(n);
2130 nr_objs = node_nr_objs(n);
2133 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2134 node, nr_slabs, nr_objs, nr_free);
2138 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2139 int node, struct kmem_cache_cpu **pc)
2142 struct kmem_cache_cpu *c = *pc;
2145 freelist = get_partial(s, flags, node, c);
2150 page = new_slab(s, flags, node);
2152 c = __this_cpu_ptr(s->cpu_slab);
2157 * No other reference to the page yet so we can
2158 * muck around with it freely without cmpxchg
2160 freelist = page->freelist;
2161 page->freelist = NULL;
2163 stat(s, ALLOC_SLAB);
2172 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2174 if (unlikely(PageSlabPfmemalloc(page)))
2175 return gfp_pfmemalloc_allowed(gfpflags);
2181 * Check the page->freelist of a page and either transfer the freelist to the
2182 * per cpu freelist or deactivate the page.
2184 * The page is still frozen if the return value is not NULL.
2186 * If this function returns NULL then the page has been unfrozen.
2188 * This function must be called with interrupt disabled.
2190 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2193 unsigned long counters;
2197 freelist = page->freelist;
2198 counters = page->counters;
2200 new.counters = counters;
2201 VM_BUG_ON(!new.frozen);
2203 new.inuse = page->objects;
2204 new.frozen = freelist != NULL;
2206 } while (!__cmpxchg_double_slab(s, page,
2215 * Slow path. The lockless freelist is empty or we need to perform
2218 * Processing is still very fast if new objects have been freed to the
2219 * regular freelist. In that case we simply take over the regular freelist
2220 * as the lockless freelist and zap the regular freelist.
2222 * If that is not working then we fall back to the partial lists. We take the
2223 * first element of the freelist as the object to allocate now and move the
2224 * rest of the freelist to the lockless freelist.
2226 * And if we were unable to get a new slab from the partial slab lists then
2227 * we need to allocate a new slab. This is the slowest path since it involves
2228 * a call to the page allocator and the setup of a new slab.
2230 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2231 unsigned long addr, struct kmem_cache_cpu *c)
2235 unsigned long flags;
2237 local_irq_save(flags);
2238 #ifdef CONFIG_PREEMPT
2240 * We may have been preempted and rescheduled on a different
2241 * cpu before disabling interrupts. Need to reload cpu area
2244 c = this_cpu_ptr(s->cpu_slab);
2252 if (unlikely(!node_match(page, node))) {
2253 stat(s, ALLOC_NODE_MISMATCH);
2254 deactivate_slab(s, page, c->freelist);
2261 * By rights, we should be searching for a slab page that was
2262 * PFMEMALLOC but right now, we are losing the pfmemalloc
2263 * information when the page leaves the per-cpu allocator
2265 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2266 deactivate_slab(s, page, c->freelist);
2272 /* must check again c->freelist in case of cpu migration or IRQ */
2273 freelist = c->freelist;
2277 stat(s, ALLOC_SLOWPATH);
2279 freelist = get_freelist(s, page);
2283 stat(s, DEACTIVATE_BYPASS);
2287 stat(s, ALLOC_REFILL);
2291 * freelist is pointing to the list of objects to be used.
2292 * page is pointing to the page from which the objects are obtained.
2293 * That page must be frozen for per cpu allocations to work.
2295 VM_BUG_ON(!c->page->frozen);
2296 c->freelist = get_freepointer(s, freelist);
2297 c->tid = next_tid(c->tid);
2298 local_irq_restore(flags);
2304 page = c->page = c->partial;
2305 c->partial = page->next;
2306 stat(s, CPU_PARTIAL_ALLOC);
2311 freelist = new_slab_objects(s, gfpflags, node, &c);
2313 if (unlikely(!freelist)) {
2314 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2315 slab_out_of_memory(s, gfpflags, node);
2317 local_irq_restore(flags);
2322 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2325 /* Only entered in the debug case */
2326 if (kmem_cache_debug(s) &&
2327 !alloc_debug_processing(s, page, freelist, addr))
2328 goto new_slab; /* Slab failed checks. Next slab needed */
2330 deactivate_slab(s, page, get_freepointer(s, freelist));
2333 local_irq_restore(flags);
2338 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2339 * have the fastpath folded into their functions. So no function call
2340 * overhead for requests that can be satisfied on the fastpath.
2342 * The fastpath works by first checking if the lockless freelist can be used.
2343 * If not then __slab_alloc is called for slow processing.
2345 * Otherwise we can simply pick the next object from the lockless free list.
2347 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2348 gfp_t gfpflags, int node, unsigned long addr)
2351 struct kmem_cache_cpu *c;
2355 if (slab_pre_alloc_hook(s, gfpflags))
2358 s = memcg_kmem_get_cache(s, gfpflags);
2361 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2362 * enabled. We may switch back and forth between cpus while
2363 * reading from one cpu area. That does not matter as long
2364 * as we end up on the original cpu again when doing the cmpxchg.
2366 * Preemption is disabled for the retrieval of the tid because that
2367 * must occur from the current processor. We cannot allow rescheduling
2368 * on a different processor between the determination of the pointer
2369 * and the retrieval of the tid.
2372 c = __this_cpu_ptr(s->cpu_slab);
2375 * The transaction ids are globally unique per cpu and per operation on
2376 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2377 * occurs on the right processor and that there was no operation on the
2378 * linked list in between.
2383 object = c->freelist;
2385 if (unlikely(!object || !node_match(page, node)))
2386 object = __slab_alloc(s, gfpflags, node, addr, c);
2389 void *next_object = get_freepointer_safe(s, object);
2392 * The cmpxchg will only match if there was no additional
2393 * operation and if we are on the right processor.
2395 * The cmpxchg does the following atomically (without lock
2397 * 1. Relocate first pointer to the current per cpu area.
2398 * 2. Verify that tid and freelist have not been changed
2399 * 3. If they were not changed replace tid and freelist
2401 * Since this is without lock semantics the protection is only
2402 * against code executing on this cpu *not* from access by
2405 if (unlikely(!this_cpu_cmpxchg_double(
2406 s->cpu_slab->freelist, s->cpu_slab->tid,
2408 next_object, next_tid(tid)))) {
2410 note_cmpxchg_failure("slab_alloc", s, tid);
2413 prefetch_freepointer(s, next_object);
2414 stat(s, ALLOC_FASTPATH);
2417 if (unlikely(gfpflags & __GFP_ZERO) && object)
2418 memset(object, 0, s->object_size);
2420 slab_post_alloc_hook(s, gfpflags, object);
2425 static __always_inline void *slab_alloc(struct kmem_cache *s,
2426 gfp_t gfpflags, unsigned long addr)
2428 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2431 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2433 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2435 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2440 EXPORT_SYMBOL(kmem_cache_alloc);
2442 #ifdef CONFIG_TRACING
2443 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2445 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2446 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2449 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2453 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2455 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2457 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2458 s->object_size, s->size, gfpflags, node);
2462 EXPORT_SYMBOL(kmem_cache_alloc_node);
2464 #ifdef CONFIG_TRACING
2465 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2467 int node, size_t size)
2469 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2471 trace_kmalloc_node(_RET_IP_, ret,
2472 size, s->size, gfpflags, node);
2475 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2480 * Slow patch handling. This may still be called frequently since objects
2481 * have a longer lifetime than the cpu slabs in most processing loads.
2483 * So we still attempt to reduce cache line usage. Just take the slab
2484 * lock and free the item. If there is no additional partial page
2485 * handling required then we can return immediately.
2487 static void __slab_free(struct kmem_cache *s, struct page *page,
2488 void *x, unsigned long addr)
2491 void **object = (void *)x;
2494 unsigned long counters;
2495 struct kmem_cache_node *n = NULL;
2496 unsigned long uninitialized_var(flags);
2498 stat(s, FREE_SLOWPATH);
2500 if (kmem_cache_debug(s) &&
2501 !(n = free_debug_processing(s, page, x, addr, &flags)))
2506 spin_unlock_irqrestore(&n->list_lock, flags);
2509 prior = page->freelist;
2510 counters = page->counters;
2511 set_freepointer(s, object, prior);
2512 new.counters = counters;
2513 was_frozen = new.frozen;
2515 if ((!new.inuse || !prior) && !was_frozen) {
2517 if (kmem_cache_has_cpu_partial(s) && !prior)
2520 * Slab was on no list before and will be
2522 * We can defer the list move and instead
2527 else { /* Needs to be taken off a list */
2529 n = get_node(s, page_to_nid(page));
2531 * Speculatively acquire the list_lock.
2532 * If the cmpxchg does not succeed then we may
2533 * drop the list_lock without any processing.
2535 * Otherwise the list_lock will synchronize with
2536 * other processors updating the list of slabs.
2538 spin_lock_irqsave(&n->list_lock, flags);
2543 } while (!cmpxchg_double_slab(s, page,
2545 object, new.counters,
2551 * If we just froze the page then put it onto the
2552 * per cpu partial list.
2554 if (new.frozen && !was_frozen) {
2555 put_cpu_partial(s, page, 1);
2556 stat(s, CPU_PARTIAL_FREE);
2559 * The list lock was not taken therefore no list
2560 * activity can be necessary.
2563 stat(s, FREE_FROZEN);
2567 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2571 * Objects left in the slab. If it was not on the partial list before
2574 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2575 if (kmem_cache_debug(s))
2576 remove_full(s, page);
2577 add_partial(n, page, DEACTIVATE_TO_TAIL);
2578 stat(s, FREE_ADD_PARTIAL);
2580 spin_unlock_irqrestore(&n->list_lock, flags);
2586 * Slab on the partial list.
2588 remove_partial(n, page);
2589 stat(s, FREE_REMOVE_PARTIAL);
2591 /* Slab must be on the full list */
2592 remove_full(s, page);
2594 spin_unlock_irqrestore(&n->list_lock, flags);
2596 discard_slab(s, page);
2600 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2601 * can perform fastpath freeing without additional function calls.
2603 * The fastpath is only possible if we are freeing to the current cpu slab
2604 * of this processor. This typically the case if we have just allocated
2607 * If fastpath is not possible then fall back to __slab_free where we deal
2608 * with all sorts of special processing.
2610 static __always_inline void slab_free(struct kmem_cache *s,
2611 struct page *page, void *x, unsigned long addr)
2613 void **object = (void *)x;
2614 struct kmem_cache_cpu *c;
2617 slab_free_hook(s, x);
2621 * Determine the currently cpus per cpu slab.
2622 * The cpu may change afterward. However that does not matter since
2623 * data is retrieved via this pointer. If we are on the same cpu
2624 * during the cmpxchg then the free will succedd.
2627 c = __this_cpu_ptr(s->cpu_slab);
2632 if (likely(page == c->page)) {
2633 set_freepointer(s, object, c->freelist);
2635 if (unlikely(!this_cpu_cmpxchg_double(
2636 s->cpu_slab->freelist, s->cpu_slab->tid,
2638 object, next_tid(tid)))) {
2640 note_cmpxchg_failure("slab_free", s, tid);
2643 stat(s, FREE_FASTPATH);
2645 __slab_free(s, page, x, addr);
2649 void kmem_cache_free(struct kmem_cache *s, void *x)
2651 s = cache_from_obj(s, x);
2654 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2655 trace_kmem_cache_free(_RET_IP_, x);
2657 EXPORT_SYMBOL(kmem_cache_free);
2660 * Object placement in a slab is made very easy because we always start at
2661 * offset 0. If we tune the size of the object to the alignment then we can
2662 * get the required alignment by putting one properly sized object after
2665 * Notice that the allocation order determines the sizes of the per cpu
2666 * caches. Each processor has always one slab available for allocations.
2667 * Increasing the allocation order reduces the number of times that slabs
2668 * must be moved on and off the partial lists and is therefore a factor in
2673 * Mininum / Maximum order of slab pages. This influences locking overhead
2674 * and slab fragmentation. A higher order reduces the number of partial slabs
2675 * and increases the number of allocations possible without having to
2676 * take the list_lock.
2678 static int slub_min_order;
2679 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2680 static int slub_min_objects;
2683 * Merge control. If this is set then no merging of slab caches will occur.
2684 * (Could be removed. This was introduced to pacify the merge skeptics.)
2686 static int slub_nomerge;
2689 * Calculate the order of allocation given an slab object size.
2691 * The order of allocation has significant impact on performance and other
2692 * system components. Generally order 0 allocations should be preferred since
2693 * order 0 does not cause fragmentation in the page allocator. Larger objects
2694 * be problematic to put into order 0 slabs because there may be too much
2695 * unused space left. We go to a higher order if more than 1/16th of the slab
2698 * In order to reach satisfactory performance we must ensure that a minimum
2699 * number of objects is in one slab. Otherwise we may generate too much
2700 * activity on the partial lists which requires taking the list_lock. This is
2701 * less a concern for large slabs though which are rarely used.
2703 * slub_max_order specifies the order where we begin to stop considering the
2704 * number of objects in a slab as critical. If we reach slub_max_order then
2705 * we try to keep the page order as low as possible. So we accept more waste
2706 * of space in favor of a small page order.
2708 * Higher order allocations also allow the placement of more objects in a
2709 * slab and thereby reduce object handling overhead. If the user has
2710 * requested a higher mininum order then we start with that one instead of
2711 * the smallest order which will fit the object.
2713 static inline int slab_order(int size, int min_objects,
2714 int max_order, int fract_leftover, int reserved)
2718 int min_order = slub_min_order;
2720 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2721 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2723 for (order = max(min_order,
2724 fls(min_objects * size - 1) - PAGE_SHIFT);
2725 order <= max_order; order++) {
2727 unsigned long slab_size = PAGE_SIZE << order;
2729 if (slab_size < min_objects * size + reserved)
2732 rem = (slab_size - reserved) % size;
2734 if (rem <= slab_size / fract_leftover)
2742 static inline int calculate_order(int size, int reserved)
2750 * Attempt to find best configuration for a slab. This
2751 * works by first attempting to generate a layout with
2752 * the best configuration and backing off gradually.
2754 * First we reduce the acceptable waste in a slab. Then
2755 * we reduce the minimum objects required in a slab.
2757 min_objects = slub_min_objects;
2759 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2760 max_objects = order_objects(slub_max_order, size, reserved);
2761 min_objects = min(min_objects, max_objects);
2763 while (min_objects > 1) {
2765 while (fraction >= 4) {
2766 order = slab_order(size, min_objects,
2767 slub_max_order, fraction, reserved);
2768 if (order <= slub_max_order)
2776 * We were unable to place multiple objects in a slab. Now
2777 * lets see if we can place a single object there.
2779 order = slab_order(size, 1, slub_max_order, 1, reserved);
2780 if (order <= slub_max_order)
2784 * Doh this slab cannot be placed using slub_max_order.
2786 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2787 if (order < MAX_ORDER)
2793 init_kmem_cache_node(struct kmem_cache_node *n)
2796 spin_lock_init(&n->list_lock);
2797 INIT_LIST_HEAD(&n->partial);
2798 #ifdef CONFIG_SLUB_DEBUG
2799 atomic_long_set(&n->nr_slabs, 0);
2800 atomic_long_set(&n->total_objects, 0);
2801 INIT_LIST_HEAD(&n->full);
2805 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2807 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2808 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2811 * Must align to double word boundary for the double cmpxchg
2812 * instructions to work; see __pcpu_double_call_return_bool().
2814 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2815 2 * sizeof(void *));
2820 init_kmem_cache_cpus(s);
2825 static struct kmem_cache *kmem_cache_node;
2828 * No kmalloc_node yet so do it by hand. We know that this is the first
2829 * slab on the node for this slabcache. There are no concurrent accesses
2832 * Note that this function only works on the kmalloc_node_cache
2833 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2834 * memory on a fresh node that has no slab structures yet.
2836 static void early_kmem_cache_node_alloc(int node)
2839 struct kmem_cache_node *n;
2841 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2843 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2846 if (page_to_nid(page) != node) {
2847 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2849 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2850 "in order to be able to continue\n");
2855 page->freelist = get_freepointer(kmem_cache_node, n);
2858 kmem_cache_node->node[node] = n;
2859 #ifdef CONFIG_SLUB_DEBUG
2860 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2861 init_tracking(kmem_cache_node, n);
2863 init_kmem_cache_node(n);
2864 inc_slabs_node(kmem_cache_node, node, page->objects);
2866 add_partial(n, page, DEACTIVATE_TO_HEAD);
2869 static void free_kmem_cache_nodes(struct kmem_cache *s)
2873 for_each_node_state(node, N_NORMAL_MEMORY) {
2874 struct kmem_cache_node *n = s->node[node];
2877 kmem_cache_free(kmem_cache_node, n);
2879 s->node[node] = NULL;
2883 static int init_kmem_cache_nodes(struct kmem_cache *s)
2887 for_each_node_state(node, N_NORMAL_MEMORY) {
2888 struct kmem_cache_node *n;
2890 if (slab_state == DOWN) {
2891 early_kmem_cache_node_alloc(node);
2894 n = kmem_cache_alloc_node(kmem_cache_node,
2898 free_kmem_cache_nodes(s);
2903 init_kmem_cache_node(n);
2908 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2910 if (min < MIN_PARTIAL)
2912 else if (min > MAX_PARTIAL)
2914 s->min_partial = min;
2918 * calculate_sizes() determines the order and the distribution of data within
2921 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2923 unsigned long flags = s->flags;
2924 unsigned long size = s->object_size;
2928 * Round up object size to the next word boundary. We can only
2929 * place the free pointer at word boundaries and this determines
2930 * the possible location of the free pointer.
2932 size = ALIGN(size, sizeof(void *));
2934 #ifdef CONFIG_SLUB_DEBUG
2936 * Determine if we can poison the object itself. If the user of
2937 * the slab may touch the object after free or before allocation
2938 * then we should never poison the object itself.
2940 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2942 s->flags |= __OBJECT_POISON;
2944 s->flags &= ~__OBJECT_POISON;
2948 * If we are Redzoning then check if there is some space between the
2949 * end of the object and the free pointer. If not then add an
2950 * additional word to have some bytes to store Redzone information.
2952 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2953 size += sizeof(void *);
2957 * With that we have determined the number of bytes in actual use
2958 * by the object. This is the potential offset to the free pointer.
2962 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2965 * Relocate free pointer after the object if it is not
2966 * permitted to overwrite the first word of the object on
2969 * This is the case if we do RCU, have a constructor or
2970 * destructor or are poisoning the objects.
2973 size += sizeof(void *);
2976 #ifdef CONFIG_SLUB_DEBUG
2977 if (flags & SLAB_STORE_USER)
2979 * Need to store information about allocs and frees after
2982 size += 2 * sizeof(struct track);
2984 if (flags & SLAB_RED_ZONE)
2986 * Add some empty padding so that we can catch
2987 * overwrites from earlier objects rather than let
2988 * tracking information or the free pointer be
2989 * corrupted if a user writes before the start
2992 size += sizeof(void *);
2996 * SLUB stores one object immediately after another beginning from
2997 * offset 0. In order to align the objects we have to simply size
2998 * each object to conform to the alignment.
3000 size = ALIGN(size, s->align);
3002 if (forced_order >= 0)
3003 order = forced_order;
3005 order = calculate_order(size, s->reserved);
3012 s->allocflags |= __GFP_COMP;
3014 if (s->flags & SLAB_CACHE_DMA)
3015 s->allocflags |= GFP_DMA;
3017 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3018 s->allocflags |= __GFP_RECLAIMABLE;
3021 * Determine the number of objects per slab
3023 s->oo = oo_make(order, size, s->reserved);
3024 s->min = oo_make(get_order(size), size, s->reserved);
3025 if (oo_objects(s->oo) > oo_objects(s->max))
3028 return !!oo_objects(s->oo);
3031 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3033 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3036 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3037 s->reserved = sizeof(struct rcu_head);
3039 if (!calculate_sizes(s, -1))
3041 if (disable_higher_order_debug) {
3043 * Disable debugging flags that store metadata if the min slab
3046 if (get_order(s->size) > get_order(s->object_size)) {
3047 s->flags &= ~DEBUG_METADATA_FLAGS;
3049 if (!calculate_sizes(s, -1))
3054 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3055 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3056 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3057 /* Enable fast mode */
3058 s->flags |= __CMPXCHG_DOUBLE;
3062 * The larger the object size is, the more pages we want on the partial
3063 * list to avoid pounding the page allocator excessively.
3065 set_min_partial(s, ilog2(s->size) / 2);
3068 * cpu_partial determined the maximum number of objects kept in the
3069 * per cpu partial lists of a processor.
3071 * Per cpu partial lists mainly contain slabs that just have one
3072 * object freed. If they are used for allocation then they can be
3073 * filled up again with minimal effort. The slab will never hit the
3074 * per node partial lists and therefore no locking will be required.
3076 * This setting also determines
3078 * A) The number of objects from per cpu partial slabs dumped to the
3079 * per node list when we reach the limit.
3080 * B) The number of objects in cpu partial slabs to extract from the
3081 * per node list when we run out of per cpu objects. We only fetch
3082 * 50% to keep some capacity around for frees.
3084 if (!kmem_cache_has_cpu_partial(s))
3086 else if (s->size >= PAGE_SIZE)
3088 else if (s->size >= 1024)
3090 else if (s->size >= 256)
3091 s->cpu_partial = 13;
3093 s->cpu_partial = 30;
3096 s->remote_node_defrag_ratio = 1000;
3098 if (!init_kmem_cache_nodes(s))
3101 if (alloc_kmem_cache_cpus(s))
3104 free_kmem_cache_nodes(s);
3106 if (flags & SLAB_PANIC)
3107 panic("Cannot create slab %s size=%lu realsize=%u "
3108 "order=%u offset=%u flags=%lx\n",
3109 s->name, (unsigned long)s->size, s->size,
3110 oo_order(s->oo), s->offset, flags);
3114 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3117 #ifdef CONFIG_SLUB_DEBUG
3118 void *addr = page_address(page);
3120 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3121 sizeof(long), GFP_ATOMIC);
3124 slab_err(s, page, text, s->name);
3127 get_map(s, page, map);
3128 for_each_object(p, s, addr, page->objects) {
3130 if (!test_bit(slab_index(p, s, addr), map)) {
3131 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3133 print_tracking(s, p);
3142 * Attempt to free all partial slabs on a node.
3143 * This is called from kmem_cache_close(). We must be the last thread
3144 * using the cache and therefore we do not need to lock anymore.
3146 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3148 struct page *page, *h;
3150 list_for_each_entry_safe(page, h, &n->partial, lru) {
3152 remove_partial(n, page);
3153 discard_slab(s, page);
3155 list_slab_objects(s, page,
3156 "Objects remaining in %s on kmem_cache_close()");
3162 * Release all resources used by a slab cache.
3164 static inline int kmem_cache_close(struct kmem_cache *s)
3169 /* Attempt to free all objects */
3170 for_each_node_state(node, N_NORMAL_MEMORY) {
3171 struct kmem_cache_node *n = get_node(s, node);
3174 if (n->nr_partial || slabs_node(s, node))
3177 free_percpu(s->cpu_slab);
3178 free_kmem_cache_nodes(s);
3182 int __kmem_cache_shutdown(struct kmem_cache *s)
3184 int rc = kmem_cache_close(s);
3188 * We do the same lock strategy around sysfs_slab_add, see
3189 * __kmem_cache_create. Because this is pretty much the last
3190 * operation we do and the lock will be released shortly after
3191 * that in slab_common.c, we could just move sysfs_slab_remove
3192 * to a later point in common code. We should do that when we
3193 * have a common sysfs framework for all allocators.
3195 mutex_unlock(&slab_mutex);
3196 sysfs_slab_remove(s);
3197 mutex_lock(&slab_mutex);
3203 /********************************************************************
3205 *******************************************************************/
3207 static int __init setup_slub_min_order(char *str)
3209 get_option(&str, &slub_min_order);
3214 __setup("slub_min_order=", setup_slub_min_order);
3216 static int __init setup_slub_max_order(char *str)
3218 get_option(&str, &slub_max_order);
3219 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3224 __setup("slub_max_order=", setup_slub_max_order);
3226 static int __init setup_slub_min_objects(char *str)
3228 get_option(&str, &slub_min_objects);
3233 __setup("slub_min_objects=", setup_slub_min_objects);
3235 static int __init setup_slub_nomerge(char *str)
3241 __setup("slub_nomerge", setup_slub_nomerge);
3243 void *__kmalloc(size_t size, gfp_t flags)
3245 struct kmem_cache *s;
3248 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3249 return kmalloc_large(size, flags);
3251 s = kmalloc_slab(size, flags);
3253 if (unlikely(ZERO_OR_NULL_PTR(s)))
3256 ret = slab_alloc(s, flags, _RET_IP_);
3258 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3262 EXPORT_SYMBOL(__kmalloc);
3265 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3270 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3271 page = alloc_pages_node(node, flags, get_order(size));
3273 ptr = page_address(page);
3275 kmemleak_alloc(ptr, size, 1, flags);
3279 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3281 struct kmem_cache *s;
3284 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3285 ret = kmalloc_large_node(size, flags, node);
3287 trace_kmalloc_node(_RET_IP_, ret,
3288 size, PAGE_SIZE << get_order(size),
3294 s = kmalloc_slab(size, flags);
3296 if (unlikely(ZERO_OR_NULL_PTR(s)))
3299 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3301 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3305 EXPORT_SYMBOL(__kmalloc_node);
3308 size_t ksize(const void *object)
3312 if (unlikely(object == ZERO_SIZE_PTR))
3315 page = virt_to_head_page(object);
3317 if (unlikely(!PageSlab(page))) {
3318 WARN_ON(!PageCompound(page));
3319 return PAGE_SIZE << compound_order(page);
3322 return slab_ksize(page->slab_cache);
3324 EXPORT_SYMBOL(ksize);
3326 void kfree(const void *x)
3329 void *object = (void *)x;
3331 trace_kfree(_RET_IP_, x);
3333 if (unlikely(ZERO_OR_NULL_PTR(x)))
3336 page = virt_to_head_page(x);
3337 if (unlikely(!PageSlab(page))) {
3338 BUG_ON(!PageCompound(page));
3340 __free_memcg_kmem_pages(page, compound_order(page));
3343 slab_free(page->slab_cache, page, object, _RET_IP_);
3345 EXPORT_SYMBOL(kfree);
3348 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3349 * the remaining slabs by the number of items in use. The slabs with the
3350 * most items in use come first. New allocations will then fill those up
3351 * and thus they can be removed from the partial lists.
3353 * The slabs with the least items are placed last. This results in them
3354 * being allocated from last increasing the chance that the last objects
3355 * are freed in them.
3357 int kmem_cache_shrink(struct kmem_cache *s)
3361 struct kmem_cache_node *n;
3364 int objects = oo_objects(s->max);
3365 struct list_head *slabs_by_inuse =
3366 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3367 unsigned long flags;
3369 if (!slabs_by_inuse)
3373 for_each_node_state(node, N_NORMAL_MEMORY) {
3374 n = get_node(s, node);
3379 for (i = 0; i < objects; i++)
3380 INIT_LIST_HEAD(slabs_by_inuse + i);
3382 spin_lock_irqsave(&n->list_lock, flags);
3385 * Build lists indexed by the items in use in each slab.
3387 * Note that concurrent frees may occur while we hold the
3388 * list_lock. page->inuse here is the upper limit.
3390 list_for_each_entry_safe(page, t, &n->partial, lru) {
3391 list_move(&page->lru, slabs_by_inuse + page->inuse);
3397 * Rebuild the partial list with the slabs filled up most
3398 * first and the least used slabs at the end.
3400 for (i = objects - 1; i > 0; i--)
3401 list_splice(slabs_by_inuse + i, n->partial.prev);
3403 spin_unlock_irqrestore(&n->list_lock, flags);
3405 /* Release empty slabs */
3406 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3407 discard_slab(s, page);
3410 kfree(slabs_by_inuse);
3413 EXPORT_SYMBOL(kmem_cache_shrink);
3415 static int slab_mem_going_offline_callback(void *arg)
3417 struct kmem_cache *s;
3419 mutex_lock(&slab_mutex);
3420 list_for_each_entry(s, &slab_caches, list)
3421 kmem_cache_shrink(s);
3422 mutex_unlock(&slab_mutex);
3427 static void slab_mem_offline_callback(void *arg)
3429 struct kmem_cache_node *n;
3430 struct kmem_cache *s;
3431 struct memory_notify *marg = arg;
3434 offline_node = marg->status_change_nid_normal;
3437 * If the node still has available memory. we need kmem_cache_node
3440 if (offline_node < 0)
3443 mutex_lock(&slab_mutex);
3444 list_for_each_entry(s, &slab_caches, list) {
3445 n = get_node(s, offline_node);
3448 * if n->nr_slabs > 0, slabs still exist on the node
3449 * that is going down. We were unable to free them,
3450 * and offline_pages() function shouldn't call this
3451 * callback. So, we must fail.
3453 BUG_ON(slabs_node(s, offline_node));
3455 s->node[offline_node] = NULL;
3456 kmem_cache_free(kmem_cache_node, n);
3459 mutex_unlock(&slab_mutex);
3462 static int slab_mem_going_online_callback(void *arg)
3464 struct kmem_cache_node *n;
3465 struct kmem_cache *s;
3466 struct memory_notify *marg = arg;
3467 int nid = marg->status_change_nid_normal;
3471 * If the node's memory is already available, then kmem_cache_node is
3472 * already created. Nothing to do.
3478 * We are bringing a node online. No memory is available yet. We must
3479 * allocate a kmem_cache_node structure in order to bring the node
3482 mutex_lock(&slab_mutex);
3483 list_for_each_entry(s, &slab_caches, list) {
3485 * XXX: kmem_cache_alloc_node will fallback to other nodes
3486 * since memory is not yet available from the node that
3489 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3494 init_kmem_cache_node(n);
3498 mutex_unlock(&slab_mutex);
3502 static int slab_memory_callback(struct notifier_block *self,
3503 unsigned long action, void *arg)
3508 case MEM_GOING_ONLINE:
3509 ret = slab_mem_going_online_callback(arg);
3511 case MEM_GOING_OFFLINE:
3512 ret = slab_mem_going_offline_callback(arg);
3515 case MEM_CANCEL_ONLINE:
3516 slab_mem_offline_callback(arg);
3519 case MEM_CANCEL_OFFLINE:
3523 ret = notifier_from_errno(ret);
3529 static struct notifier_block slab_memory_callback_nb = {
3530 .notifier_call = slab_memory_callback,
3531 .priority = SLAB_CALLBACK_PRI,
3534 /********************************************************************
3535 * Basic setup of slabs
3536 *******************************************************************/
3539 * Used for early kmem_cache structures that were allocated using
3540 * the page allocator. Allocate them properly then fix up the pointers
3541 * that may be pointing to the wrong kmem_cache structure.
3544 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3547 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3549 memcpy(s, static_cache, kmem_cache->object_size);
3552 * This runs very early, and only the boot processor is supposed to be
3553 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3556 __flush_cpu_slab(s, smp_processor_id());
3557 for_each_node_state(node, N_NORMAL_MEMORY) {
3558 struct kmem_cache_node *n = get_node(s, node);
3562 list_for_each_entry(p, &n->partial, lru)
3565 #ifdef CONFIG_SLUB_DEBUG
3566 list_for_each_entry(p, &n->full, lru)
3571 list_add(&s->list, &slab_caches);
3575 void __init kmem_cache_init(void)
3577 static __initdata struct kmem_cache boot_kmem_cache,
3578 boot_kmem_cache_node;
3580 if (debug_guardpage_minorder())
3583 kmem_cache_node = &boot_kmem_cache_node;
3584 kmem_cache = &boot_kmem_cache;
3586 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3587 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3589 register_hotmemory_notifier(&slab_memory_callback_nb);
3591 /* Able to allocate the per node structures */
3592 slab_state = PARTIAL;
3594 create_boot_cache(kmem_cache, "kmem_cache",
3595 offsetof(struct kmem_cache, node) +
3596 nr_node_ids * sizeof(struct kmem_cache_node *),
3597 SLAB_HWCACHE_ALIGN);
3599 kmem_cache = bootstrap(&boot_kmem_cache);
3602 * Allocate kmem_cache_node properly from the kmem_cache slab.
3603 * kmem_cache_node is separately allocated so no need to
3604 * update any list pointers.
3606 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3608 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3609 create_kmalloc_caches(0);
3612 register_cpu_notifier(&slab_notifier);
3616 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3617 " CPUs=%d, Nodes=%d\n",
3619 slub_min_order, slub_max_order, slub_min_objects,
3620 nr_cpu_ids, nr_node_ids);
3623 void __init kmem_cache_init_late(void)
3628 * Find a mergeable slab cache
3630 static int slab_unmergeable(struct kmem_cache *s)
3632 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3639 * We may have set a slab to be unmergeable during bootstrap.
3641 if (s->refcount < 0)
3647 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3648 size_t align, unsigned long flags, const char *name,
3649 void (*ctor)(void *))
3651 struct kmem_cache *s;
3653 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3659 size = ALIGN(size, sizeof(void *));
3660 align = calculate_alignment(flags, align, size);
3661 size = ALIGN(size, align);
3662 flags = kmem_cache_flags(size, flags, name, NULL);
3664 list_for_each_entry(s, &slab_caches, list) {
3665 if (slab_unmergeable(s))
3671 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3674 * Check if alignment is compatible.
3675 * Courtesy of Adrian Drzewiecki
3677 if ((s->size & ~(align - 1)) != s->size)
3680 if (s->size - size >= sizeof(void *))
3683 if (!cache_match_memcg(s, memcg))
3692 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3693 size_t align, unsigned long flags, void (*ctor)(void *))
3695 struct kmem_cache *s;
3697 s = find_mergeable(memcg, size, align, flags, name, ctor);
3701 * Adjust the object sizes so that we clear
3702 * the complete object on kzalloc.
3704 s->object_size = max(s->object_size, (int)size);
3705 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3707 if (sysfs_slab_alias(s, name)) {
3716 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3720 err = kmem_cache_open(s, flags);
3724 /* Mutex is not taken during early boot */
3725 if (slab_state <= UP)
3728 memcg_propagate_slab_attrs(s);
3729 mutex_unlock(&slab_mutex);
3730 err = sysfs_slab_add(s);
3731 mutex_lock(&slab_mutex);
3734 kmem_cache_close(s);
3741 * Use the cpu notifier to insure that the cpu slabs are flushed when
3744 static int slab_cpuup_callback(struct notifier_block *nfb,
3745 unsigned long action, void *hcpu)
3747 long cpu = (long)hcpu;
3748 struct kmem_cache *s;
3749 unsigned long flags;
3752 case CPU_UP_CANCELED:
3753 case CPU_UP_CANCELED_FROZEN:
3755 case CPU_DEAD_FROZEN:
3756 mutex_lock(&slab_mutex);
3757 list_for_each_entry(s, &slab_caches, list) {
3758 local_irq_save(flags);
3759 __flush_cpu_slab(s, cpu);
3760 local_irq_restore(flags);
3762 mutex_unlock(&slab_mutex);
3770 static struct notifier_block slab_notifier = {
3771 .notifier_call = slab_cpuup_callback
3776 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3778 struct kmem_cache *s;
3781 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3782 return kmalloc_large(size, gfpflags);
3784 s = kmalloc_slab(size, gfpflags);
3786 if (unlikely(ZERO_OR_NULL_PTR(s)))
3789 ret = slab_alloc(s, gfpflags, caller);
3791 /* Honor the call site pointer we received. */
3792 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3798 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3799 int node, unsigned long caller)
3801 struct kmem_cache *s;
3804 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3805 ret = kmalloc_large_node(size, gfpflags, node);
3807 trace_kmalloc_node(caller, ret,
3808 size, PAGE_SIZE << get_order(size),
3814 s = kmalloc_slab(size, gfpflags);
3816 if (unlikely(ZERO_OR_NULL_PTR(s)))
3819 ret = slab_alloc_node(s, gfpflags, node, caller);
3821 /* Honor the call site pointer we received. */
3822 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3829 static int count_inuse(struct page *page)
3834 static int count_total(struct page *page)
3836 return page->objects;
3840 #ifdef CONFIG_SLUB_DEBUG
3841 static int validate_slab(struct kmem_cache *s, struct page *page,
3845 void *addr = page_address(page);
3847 if (!check_slab(s, page) ||
3848 !on_freelist(s, page, NULL))
3851 /* Now we know that a valid freelist exists */
3852 bitmap_zero(map, page->objects);
3854 get_map(s, page, map);
3855 for_each_object(p, s, addr, page->objects) {
3856 if (test_bit(slab_index(p, s, addr), map))
3857 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3861 for_each_object(p, s, addr, page->objects)
3862 if (!test_bit(slab_index(p, s, addr), map))
3863 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3868 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3872 validate_slab(s, page, map);
3876 static int validate_slab_node(struct kmem_cache *s,
3877 struct kmem_cache_node *n, unsigned long *map)
3879 unsigned long count = 0;
3881 unsigned long flags;
3883 spin_lock_irqsave(&n->list_lock, flags);
3885 list_for_each_entry(page, &n->partial, lru) {
3886 validate_slab_slab(s, page, map);
3889 if (count != n->nr_partial)
3890 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3891 "counter=%ld\n", s->name, count, n->nr_partial);
3893 if (!(s->flags & SLAB_STORE_USER))
3896 list_for_each_entry(page, &n->full, lru) {
3897 validate_slab_slab(s, page, map);
3900 if (count != atomic_long_read(&n->nr_slabs))
3901 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3902 "counter=%ld\n", s->name, count,
3903 atomic_long_read(&n->nr_slabs));
3906 spin_unlock_irqrestore(&n->list_lock, flags);
3910 static long validate_slab_cache(struct kmem_cache *s)
3913 unsigned long count = 0;
3914 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3915 sizeof(unsigned long), GFP_KERNEL);
3921 for_each_node_state(node, N_NORMAL_MEMORY) {
3922 struct kmem_cache_node *n = get_node(s, node);
3924 count += validate_slab_node(s, n, map);
3930 * Generate lists of code addresses where slabcache objects are allocated
3935 unsigned long count;
3942 DECLARE_BITMAP(cpus, NR_CPUS);
3948 unsigned long count;
3949 struct location *loc;
3952 static void free_loc_track(struct loc_track *t)
3955 free_pages((unsigned long)t->loc,
3956 get_order(sizeof(struct location) * t->max));
3959 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3964 order = get_order(sizeof(struct location) * max);
3966 l = (void *)__get_free_pages(flags, order);
3971 memcpy(l, t->loc, sizeof(struct location) * t->count);
3979 static int add_location(struct loc_track *t, struct kmem_cache *s,
3980 const struct track *track)
3982 long start, end, pos;
3984 unsigned long caddr;
3985 unsigned long age = jiffies - track->when;
3991 pos = start + (end - start + 1) / 2;
3994 * There is nothing at "end". If we end up there
3995 * we need to add something to before end.
4000 caddr = t->loc[pos].addr;
4001 if (track->addr == caddr) {
4007 if (age < l->min_time)
4009 if (age > l->max_time)
4012 if (track->pid < l->min_pid)
4013 l->min_pid = track->pid;
4014 if (track->pid > l->max_pid)
4015 l->max_pid = track->pid;
4017 cpumask_set_cpu(track->cpu,
4018 to_cpumask(l->cpus));
4020 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4024 if (track->addr < caddr)
4031 * Not found. Insert new tracking element.
4033 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4039 (t->count - pos) * sizeof(struct location));
4042 l->addr = track->addr;
4046 l->min_pid = track->pid;
4047 l->max_pid = track->pid;
4048 cpumask_clear(to_cpumask(l->cpus));
4049 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4050 nodes_clear(l->nodes);
4051 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4055 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4056 struct page *page, enum track_item alloc,
4059 void *addr = page_address(page);
4062 bitmap_zero(map, page->objects);
4063 get_map(s, page, map);
4065 for_each_object(p, s, addr, page->objects)
4066 if (!test_bit(slab_index(p, s, addr), map))
4067 add_location(t, s, get_track(s, p, alloc));
4070 static int list_locations(struct kmem_cache *s, char *buf,
4071 enum track_item alloc)
4075 struct loc_track t = { 0, 0, NULL };
4077 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4078 sizeof(unsigned long), GFP_KERNEL);
4080 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4083 return sprintf(buf, "Out of memory\n");
4085 /* Push back cpu slabs */
4088 for_each_node_state(node, N_NORMAL_MEMORY) {
4089 struct kmem_cache_node *n = get_node(s, node);
4090 unsigned long flags;
4093 if (!atomic_long_read(&n->nr_slabs))
4096 spin_lock_irqsave(&n->list_lock, flags);
4097 list_for_each_entry(page, &n->partial, lru)
4098 process_slab(&t, s, page, alloc, map);
4099 list_for_each_entry(page, &n->full, lru)
4100 process_slab(&t, s, page, alloc, map);
4101 spin_unlock_irqrestore(&n->list_lock, flags);
4104 for (i = 0; i < t.count; i++) {
4105 struct location *l = &t.loc[i];
4107 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4109 len += sprintf(buf + len, "%7ld ", l->count);
4112 len += sprintf(buf + len, "%pS", (void *)l->addr);
4114 len += sprintf(buf + len, "<not-available>");
4116 if (l->sum_time != l->min_time) {
4117 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4119 (long)div_u64(l->sum_time, l->count),
4122 len += sprintf(buf + len, " age=%ld",
4125 if (l->min_pid != l->max_pid)
4126 len += sprintf(buf + len, " pid=%ld-%ld",
4127 l->min_pid, l->max_pid);
4129 len += sprintf(buf + len, " pid=%ld",
4132 if (num_online_cpus() > 1 &&
4133 !cpumask_empty(to_cpumask(l->cpus)) &&
4134 len < PAGE_SIZE - 60) {
4135 len += sprintf(buf + len, " cpus=");
4136 len += cpulist_scnprintf(buf + len,
4137 PAGE_SIZE - len - 50,
4138 to_cpumask(l->cpus));
4141 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4142 len < PAGE_SIZE - 60) {
4143 len += sprintf(buf + len, " nodes=");
4144 len += nodelist_scnprintf(buf + len,
4145 PAGE_SIZE - len - 50,
4149 len += sprintf(buf + len, "\n");
4155 len += sprintf(buf, "No data\n");
4160 #ifdef SLUB_RESILIENCY_TEST
4161 static void resiliency_test(void)
4165 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4167 printk(KERN_ERR "SLUB resiliency testing\n");
4168 printk(KERN_ERR "-----------------------\n");
4169 printk(KERN_ERR "A. Corruption after allocation\n");
4171 p = kzalloc(16, GFP_KERNEL);
4173 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4174 " 0x12->0x%p\n\n", p + 16);
4176 validate_slab_cache(kmalloc_caches[4]);
4178 /* Hmmm... The next two are dangerous */
4179 p = kzalloc(32, GFP_KERNEL);
4180 p[32 + sizeof(void *)] = 0x34;
4181 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4182 " 0x34 -> -0x%p\n", p);
4184 "If allocated object is overwritten then not detectable\n\n");
4186 validate_slab_cache(kmalloc_caches[5]);
4187 p = kzalloc(64, GFP_KERNEL);
4188 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4190 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4193 "If allocated object is overwritten then not detectable\n\n");
4194 validate_slab_cache(kmalloc_caches[6]);
4196 printk(KERN_ERR "\nB. Corruption after free\n");
4197 p = kzalloc(128, GFP_KERNEL);
4200 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4201 validate_slab_cache(kmalloc_caches[7]);
4203 p = kzalloc(256, GFP_KERNEL);
4206 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4208 validate_slab_cache(kmalloc_caches[8]);
4210 p = kzalloc(512, GFP_KERNEL);
4213 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4214 validate_slab_cache(kmalloc_caches[9]);
4218 static void resiliency_test(void) {};
4223 enum slab_stat_type {
4224 SL_ALL, /* All slabs */
4225 SL_PARTIAL, /* Only partially allocated slabs */
4226 SL_CPU, /* Only slabs used for cpu caches */
4227 SL_OBJECTS, /* Determine allocated objects not slabs */
4228 SL_TOTAL /* Determine object capacity not slabs */
4231 #define SO_ALL (1 << SL_ALL)
4232 #define SO_PARTIAL (1 << SL_PARTIAL)
4233 #define SO_CPU (1 << SL_CPU)
4234 #define SO_OBJECTS (1 << SL_OBJECTS)
4235 #define SO_TOTAL (1 << SL_TOTAL)
4237 static ssize_t show_slab_objects(struct kmem_cache *s,
4238 char *buf, unsigned long flags)
4240 unsigned long total = 0;
4243 unsigned long *nodes;
4245 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4249 if (flags & SO_CPU) {
4252 for_each_possible_cpu(cpu) {
4253 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4258 page = ACCESS_ONCE(c->page);
4262 node = page_to_nid(page);
4263 if (flags & SO_TOTAL)
4265 else if (flags & SO_OBJECTS)
4273 page = ACCESS_ONCE(c->partial);
4282 lock_memory_hotplug();
4283 #ifdef CONFIG_SLUB_DEBUG
4284 if (flags & SO_ALL) {
4285 for_each_node_state(node, N_NORMAL_MEMORY) {
4286 struct kmem_cache_node *n = get_node(s, node);
4288 if (flags & SO_TOTAL)
4289 x = atomic_long_read(&n->total_objects);
4290 else if (flags & SO_OBJECTS)
4291 x = atomic_long_read(&n->total_objects) -
4292 count_partial(n, count_free);
4294 x = atomic_long_read(&n->nr_slabs);
4301 if (flags & SO_PARTIAL) {
4302 for_each_node_state(node, N_NORMAL_MEMORY) {
4303 struct kmem_cache_node *n = get_node(s, node);
4305 if (flags & SO_TOTAL)
4306 x = count_partial(n, count_total);
4307 else if (flags & SO_OBJECTS)
4308 x = count_partial(n, count_inuse);
4315 x = sprintf(buf, "%lu", total);
4317 for_each_node_state(node, N_NORMAL_MEMORY)
4319 x += sprintf(buf + x, " N%d=%lu",
4322 unlock_memory_hotplug();
4324 return x + sprintf(buf + x, "\n");
4327 #ifdef CONFIG_SLUB_DEBUG
4328 static int any_slab_objects(struct kmem_cache *s)
4332 for_each_online_node(node) {
4333 struct kmem_cache_node *n = get_node(s, node);
4338 if (atomic_long_read(&n->total_objects))
4345 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4346 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4348 struct slab_attribute {
4349 struct attribute attr;
4350 ssize_t (*show)(struct kmem_cache *s, char *buf);
4351 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4354 #define SLAB_ATTR_RO(_name) \
4355 static struct slab_attribute _name##_attr = \
4356 __ATTR(_name, 0400, _name##_show, NULL)
4358 #define SLAB_ATTR(_name) \
4359 static struct slab_attribute _name##_attr = \
4360 __ATTR(_name, 0600, _name##_show, _name##_store)
4362 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4364 return sprintf(buf, "%d\n", s->size);
4366 SLAB_ATTR_RO(slab_size);
4368 static ssize_t align_show(struct kmem_cache *s, char *buf)
4370 return sprintf(buf, "%d\n", s->align);
4372 SLAB_ATTR_RO(align);
4374 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4376 return sprintf(buf, "%d\n", s->object_size);
4378 SLAB_ATTR_RO(object_size);
4380 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4382 return sprintf(buf, "%d\n", oo_objects(s->oo));
4384 SLAB_ATTR_RO(objs_per_slab);
4386 static ssize_t order_store(struct kmem_cache *s,
4387 const char *buf, size_t length)
4389 unsigned long order;
4392 err = kstrtoul(buf, 10, &order);
4396 if (order > slub_max_order || order < slub_min_order)
4399 calculate_sizes(s, order);
4403 static ssize_t order_show(struct kmem_cache *s, char *buf)
4405 return sprintf(buf, "%d\n", oo_order(s->oo));
4409 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4411 return sprintf(buf, "%lu\n", s->min_partial);
4414 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4420 err = kstrtoul(buf, 10, &min);
4424 set_min_partial(s, min);
4427 SLAB_ATTR(min_partial);
4429 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4431 return sprintf(buf, "%u\n", s->cpu_partial);
4434 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4437 unsigned long objects;
4440 err = kstrtoul(buf, 10, &objects);
4443 if (objects && !kmem_cache_has_cpu_partial(s))
4446 s->cpu_partial = objects;
4450 SLAB_ATTR(cpu_partial);
4452 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4456 return sprintf(buf, "%pS\n", s->ctor);
4460 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4462 return sprintf(buf, "%d\n", s->refcount - 1);
4464 SLAB_ATTR_RO(aliases);
4466 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4468 return show_slab_objects(s, buf, SO_PARTIAL);
4470 SLAB_ATTR_RO(partial);
4472 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4474 return show_slab_objects(s, buf, SO_CPU);
4476 SLAB_ATTR_RO(cpu_slabs);
4478 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4480 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4482 SLAB_ATTR_RO(objects);
4484 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4486 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4488 SLAB_ATTR_RO(objects_partial);
4490 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4497 for_each_online_cpu(cpu) {
4498 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4501 pages += page->pages;
4502 objects += page->pobjects;
4506 len = sprintf(buf, "%d(%d)", objects, pages);
4509 for_each_online_cpu(cpu) {
4510 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4512 if (page && len < PAGE_SIZE - 20)
4513 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4514 page->pobjects, page->pages);
4517 return len + sprintf(buf + len, "\n");
4519 SLAB_ATTR_RO(slabs_cpu_partial);
4521 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4523 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4526 static ssize_t reclaim_account_store(struct kmem_cache *s,
4527 const char *buf, size_t length)
4529 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4531 s->flags |= SLAB_RECLAIM_ACCOUNT;
4534 SLAB_ATTR(reclaim_account);
4536 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4538 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4540 SLAB_ATTR_RO(hwcache_align);
4542 #ifdef CONFIG_ZONE_DMA
4543 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4545 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4547 SLAB_ATTR_RO(cache_dma);
4550 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4552 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4554 SLAB_ATTR_RO(destroy_by_rcu);
4556 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4558 return sprintf(buf, "%d\n", s->reserved);
4560 SLAB_ATTR_RO(reserved);
4562 #ifdef CONFIG_SLUB_DEBUG
4563 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4565 return show_slab_objects(s, buf, SO_ALL);
4567 SLAB_ATTR_RO(slabs);
4569 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4571 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4573 SLAB_ATTR_RO(total_objects);
4575 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4577 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4580 static ssize_t sanity_checks_store(struct kmem_cache *s,
4581 const char *buf, size_t length)
4583 s->flags &= ~SLAB_DEBUG_FREE;
4584 if (buf[0] == '1') {
4585 s->flags &= ~__CMPXCHG_DOUBLE;
4586 s->flags |= SLAB_DEBUG_FREE;
4590 SLAB_ATTR(sanity_checks);
4592 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4594 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4597 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4600 s->flags &= ~SLAB_TRACE;
4601 if (buf[0] == '1') {
4602 s->flags &= ~__CMPXCHG_DOUBLE;
4603 s->flags |= SLAB_TRACE;
4609 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4611 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4614 static ssize_t red_zone_store(struct kmem_cache *s,
4615 const char *buf, size_t length)
4617 if (any_slab_objects(s))
4620 s->flags &= ~SLAB_RED_ZONE;
4621 if (buf[0] == '1') {
4622 s->flags &= ~__CMPXCHG_DOUBLE;
4623 s->flags |= SLAB_RED_ZONE;
4625 calculate_sizes(s, -1);
4628 SLAB_ATTR(red_zone);
4630 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4632 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4635 static ssize_t poison_store(struct kmem_cache *s,
4636 const char *buf, size_t length)
4638 if (any_slab_objects(s))
4641 s->flags &= ~SLAB_POISON;
4642 if (buf[0] == '1') {
4643 s->flags &= ~__CMPXCHG_DOUBLE;
4644 s->flags |= SLAB_POISON;
4646 calculate_sizes(s, -1);
4651 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4653 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4656 static ssize_t store_user_store(struct kmem_cache *s,
4657 const char *buf, size_t length)
4659 if (any_slab_objects(s))
4662 s->flags &= ~SLAB_STORE_USER;
4663 if (buf[0] == '1') {
4664 s->flags &= ~__CMPXCHG_DOUBLE;
4665 s->flags |= SLAB_STORE_USER;
4667 calculate_sizes(s, -1);
4670 SLAB_ATTR(store_user);
4672 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4677 static ssize_t validate_store(struct kmem_cache *s,
4678 const char *buf, size_t length)
4682 if (buf[0] == '1') {
4683 ret = validate_slab_cache(s);
4689 SLAB_ATTR(validate);
4691 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4693 if (!(s->flags & SLAB_STORE_USER))
4695 return list_locations(s, buf, TRACK_ALLOC);
4697 SLAB_ATTR_RO(alloc_calls);
4699 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4701 if (!(s->flags & SLAB_STORE_USER))
4703 return list_locations(s, buf, TRACK_FREE);
4705 SLAB_ATTR_RO(free_calls);
4706 #endif /* CONFIG_SLUB_DEBUG */
4708 #ifdef CONFIG_FAILSLAB
4709 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4711 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4714 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4717 s->flags &= ~SLAB_FAILSLAB;
4719 s->flags |= SLAB_FAILSLAB;
4722 SLAB_ATTR(failslab);
4725 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4730 static ssize_t shrink_store(struct kmem_cache *s,
4731 const char *buf, size_t length)
4733 if (buf[0] == '1') {
4734 int rc = kmem_cache_shrink(s);
4745 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4747 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4750 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4751 const char *buf, size_t length)
4753 unsigned long ratio;
4756 err = kstrtoul(buf, 10, &ratio);
4761 s->remote_node_defrag_ratio = ratio * 10;
4765 SLAB_ATTR(remote_node_defrag_ratio);
4768 #ifdef CONFIG_SLUB_STATS
4769 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4771 unsigned long sum = 0;
4774 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4779 for_each_online_cpu(cpu) {
4780 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4786 len = sprintf(buf, "%lu", sum);
4789 for_each_online_cpu(cpu) {
4790 if (data[cpu] && len < PAGE_SIZE - 20)
4791 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4795 return len + sprintf(buf + len, "\n");
4798 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4802 for_each_online_cpu(cpu)
4803 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4806 #define STAT_ATTR(si, text) \
4807 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4809 return show_stat(s, buf, si); \
4811 static ssize_t text##_store(struct kmem_cache *s, \
4812 const char *buf, size_t length) \
4814 if (buf[0] != '0') \
4816 clear_stat(s, si); \
4821 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4822 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4823 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4824 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4825 STAT_ATTR(FREE_FROZEN, free_frozen);
4826 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4827 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4828 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4829 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4830 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4831 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4832 STAT_ATTR(FREE_SLAB, free_slab);
4833 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4834 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4835 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4836 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4837 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4838 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4839 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4840 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4841 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4842 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4843 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4844 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4845 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4846 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4849 static struct attribute *slab_attrs[] = {
4850 &slab_size_attr.attr,
4851 &object_size_attr.attr,
4852 &objs_per_slab_attr.attr,
4854 &min_partial_attr.attr,
4855 &cpu_partial_attr.attr,
4857 &objects_partial_attr.attr,
4859 &cpu_slabs_attr.attr,
4863 &hwcache_align_attr.attr,
4864 &reclaim_account_attr.attr,
4865 &destroy_by_rcu_attr.attr,
4867 &reserved_attr.attr,
4868 &slabs_cpu_partial_attr.attr,
4869 #ifdef CONFIG_SLUB_DEBUG
4870 &total_objects_attr.attr,
4872 &sanity_checks_attr.attr,
4874 &red_zone_attr.attr,
4876 &store_user_attr.attr,
4877 &validate_attr.attr,
4878 &alloc_calls_attr.attr,
4879 &free_calls_attr.attr,
4881 #ifdef CONFIG_ZONE_DMA
4882 &cache_dma_attr.attr,
4885 &remote_node_defrag_ratio_attr.attr,
4887 #ifdef CONFIG_SLUB_STATS
4888 &alloc_fastpath_attr.attr,
4889 &alloc_slowpath_attr.attr,
4890 &free_fastpath_attr.attr,
4891 &free_slowpath_attr.attr,
4892 &free_frozen_attr.attr,
4893 &free_add_partial_attr.attr,
4894 &free_remove_partial_attr.attr,
4895 &alloc_from_partial_attr.attr,
4896 &alloc_slab_attr.attr,
4897 &alloc_refill_attr.attr,
4898 &alloc_node_mismatch_attr.attr,
4899 &free_slab_attr.attr,
4900 &cpuslab_flush_attr.attr,
4901 &deactivate_full_attr.attr,
4902 &deactivate_empty_attr.attr,
4903 &deactivate_to_head_attr.attr,
4904 &deactivate_to_tail_attr.attr,
4905 &deactivate_remote_frees_attr.attr,
4906 &deactivate_bypass_attr.attr,
4907 &order_fallback_attr.attr,
4908 &cmpxchg_double_fail_attr.attr,
4909 &cmpxchg_double_cpu_fail_attr.attr,
4910 &cpu_partial_alloc_attr.attr,
4911 &cpu_partial_free_attr.attr,
4912 &cpu_partial_node_attr.attr,
4913 &cpu_partial_drain_attr.attr,
4915 #ifdef CONFIG_FAILSLAB
4916 &failslab_attr.attr,
4922 static struct attribute_group slab_attr_group = {
4923 .attrs = slab_attrs,
4926 static ssize_t slab_attr_show(struct kobject *kobj,
4927 struct attribute *attr,
4930 struct slab_attribute *attribute;
4931 struct kmem_cache *s;
4934 attribute = to_slab_attr(attr);
4937 if (!attribute->show)
4940 err = attribute->show(s, buf);
4945 static ssize_t slab_attr_store(struct kobject *kobj,
4946 struct attribute *attr,
4947 const char *buf, size_t len)
4949 struct slab_attribute *attribute;
4950 struct kmem_cache *s;
4953 attribute = to_slab_attr(attr);
4956 if (!attribute->store)
4959 err = attribute->store(s, buf, len);
4960 #ifdef CONFIG_MEMCG_KMEM
4961 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4964 mutex_lock(&slab_mutex);
4965 if (s->max_attr_size < len)
4966 s->max_attr_size = len;
4969 * This is a best effort propagation, so this function's return
4970 * value will be determined by the parent cache only. This is
4971 * basically because not all attributes will have a well
4972 * defined semantics for rollbacks - most of the actions will
4973 * have permanent effects.
4975 * Returning the error value of any of the children that fail
4976 * is not 100 % defined, in the sense that users seeing the
4977 * error code won't be able to know anything about the state of
4980 * Only returning the error code for the parent cache at least
4981 * has well defined semantics. The cache being written to
4982 * directly either failed or succeeded, in which case we loop
4983 * through the descendants with best-effort propagation.
4985 for_each_memcg_cache_index(i) {
4986 struct kmem_cache *c = cache_from_memcg(s, i);
4988 attribute->store(c, buf, len);
4990 mutex_unlock(&slab_mutex);
4996 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
4998 #ifdef CONFIG_MEMCG_KMEM
5000 char *buffer = NULL;
5002 if (!is_root_cache(s))
5006 * This mean this cache had no attribute written. Therefore, no point
5007 * in copying default values around
5009 if (!s->max_attr_size)
5012 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5015 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5017 if (!attr || !attr->store || !attr->show)
5021 * It is really bad that we have to allocate here, so we will
5022 * do it only as a fallback. If we actually allocate, though,
5023 * we can just use the allocated buffer until the end.
5025 * Most of the slub attributes will tend to be very small in
5026 * size, but sysfs allows buffers up to a page, so they can
5027 * theoretically happen.
5031 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5034 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5035 if (WARN_ON(!buffer))
5040 attr->show(s->memcg_params->root_cache, buf);
5041 attr->store(s, buf, strlen(buf));
5045 free_page((unsigned long)buffer);
5049 static const struct sysfs_ops slab_sysfs_ops = {
5050 .show = slab_attr_show,
5051 .store = slab_attr_store,
5054 static struct kobj_type slab_ktype = {
5055 .sysfs_ops = &slab_sysfs_ops,
5058 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5060 struct kobj_type *ktype = get_ktype(kobj);
5062 if (ktype == &slab_ktype)
5067 static const struct kset_uevent_ops slab_uevent_ops = {
5068 .filter = uevent_filter,
5071 static struct kset *slab_kset;
5073 #define ID_STR_LENGTH 64
5075 /* Create a unique string id for a slab cache:
5077 * Format :[flags-]size
5079 static char *create_unique_id(struct kmem_cache *s)
5081 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5088 * First flags affecting slabcache operations. We will only
5089 * get here for aliasable slabs so we do not need to support
5090 * too many flags. The flags here must cover all flags that
5091 * are matched during merging to guarantee that the id is
5094 if (s->flags & SLAB_CACHE_DMA)
5096 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5098 if (s->flags & SLAB_DEBUG_FREE)
5100 if (!(s->flags & SLAB_NOTRACK))
5104 p += sprintf(p, "%07d", s->size);
5106 #ifdef CONFIG_MEMCG_KMEM
5107 if (!is_root_cache(s))
5108 p += sprintf(p, "-%08d",
5109 memcg_cache_id(s->memcg_params->memcg));
5112 BUG_ON(p > name + ID_STR_LENGTH - 1);
5116 static int sysfs_slab_add(struct kmem_cache *s)
5120 int unmergeable = slab_unmergeable(s);
5124 * Slabcache can never be merged so we can use the name proper.
5125 * This is typically the case for debug situations. In that
5126 * case we can catch duplicate names easily.
5128 sysfs_remove_link(&slab_kset->kobj, s->name);
5132 * Create a unique name for the slab as a target
5135 name = create_unique_id(s);
5138 s->kobj.kset = slab_kset;
5139 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5141 kobject_put(&s->kobj);
5145 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5147 kobject_del(&s->kobj);
5148 kobject_put(&s->kobj);
5151 kobject_uevent(&s->kobj, KOBJ_ADD);
5153 /* Setup first alias */
5154 sysfs_slab_alias(s, s->name);
5160 static void sysfs_slab_remove(struct kmem_cache *s)
5162 if (slab_state < FULL)
5164 * Sysfs has not been setup yet so no need to remove the
5169 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5170 kobject_del(&s->kobj);
5171 kobject_put(&s->kobj);
5175 * Need to buffer aliases during bootup until sysfs becomes
5176 * available lest we lose that information.
5178 struct saved_alias {
5179 struct kmem_cache *s;
5181 struct saved_alias *next;
5184 static struct saved_alias *alias_list;
5186 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5188 struct saved_alias *al;
5190 if (slab_state == FULL) {
5192 * If we have a leftover link then remove it.
5194 sysfs_remove_link(&slab_kset->kobj, name);
5195 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5198 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5204 al->next = alias_list;
5209 static int __init slab_sysfs_init(void)
5211 struct kmem_cache *s;
5214 mutex_lock(&slab_mutex);
5216 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5218 mutex_unlock(&slab_mutex);
5219 printk(KERN_ERR "Cannot register slab subsystem.\n");
5225 list_for_each_entry(s, &slab_caches, list) {
5226 err = sysfs_slab_add(s);
5228 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5229 " to sysfs\n", s->name);
5232 while (alias_list) {
5233 struct saved_alias *al = alias_list;
5235 alias_list = alias_list->next;
5236 err = sysfs_slab_alias(al->s, al->name);
5238 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5239 " %s to sysfs\n", al->name);
5243 mutex_unlock(&slab_mutex);
5248 __initcall(slab_sysfs_init);
5249 #endif /* CONFIG_SYSFS */
5252 * The /proc/slabinfo ABI
5254 #ifdef CONFIG_SLABINFO
5255 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5257 unsigned long nr_slabs = 0;
5258 unsigned long nr_objs = 0;
5259 unsigned long nr_free = 0;
5262 for_each_online_node(node) {
5263 struct kmem_cache_node *n = get_node(s, node);
5268 nr_slabs += node_nr_slabs(n);
5269 nr_objs += node_nr_objs(n);
5270 nr_free += count_partial(n, count_free);
5273 sinfo->active_objs = nr_objs - nr_free;
5274 sinfo->num_objs = nr_objs;
5275 sinfo->active_slabs = nr_slabs;
5276 sinfo->num_slabs = nr_slabs;
5277 sinfo->objects_per_slab = oo_objects(s->oo);
5278 sinfo->cache_order = oo_order(s->oo);
5281 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5285 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5286 size_t count, loff_t *ppos)
5290 #endif /* CONFIG_SLABINFO */