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 use.
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
228 * The rmw is racy on a preemptible kernel but this is acceptable, so
229 * avoid this_cpu_add()'s irq-disable overhead.
231 raw_cpu_inc(s->cpu_slab->stat[si]);
235 /********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
239 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
241 return s->node[node];
244 /* Verify that a pointer has an address that is valid within a slab page */
245 static inline int check_valid_pointer(struct kmem_cache *s,
246 struct page *page, const void *object)
253 base = page_address(page);
254 if (object < base || object >= base + page->objects * s->size ||
255 (object - base) % s->size) {
262 static inline void *get_freepointer(struct kmem_cache *s, void *object)
264 return *(void **)(object + s->offset);
267 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
269 prefetch(object + s->offset);
272 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
276 #ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
279 p = get_freepointer(s, object);
284 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
286 *(void **)(object + s->offset) = fp;
289 /* Loop over all objects in a slab */
290 #define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
294 /* Determine object index from a given position */
295 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
297 return (p - addr) / s->size;
300 static inline size_t slab_ksize(const struct kmem_cache *s)
302 #ifdef CONFIG_SLUB_DEBUG
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
308 return s->object_size;
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
319 * Else we can use all the padding etc for the allocation
324 static inline int order_objects(int order, unsigned long size, int reserved)
326 return ((PAGE_SIZE << order) - reserved) / size;
329 static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
339 static inline int oo_order(struct kmem_cache_order_objects x)
341 return x.x >> OO_SHIFT;
344 static inline int oo_objects(struct kmem_cache_order_objects x)
346 return x.x & OO_MASK;
350 * Per slab locking using the pagelock
352 static __always_inline void slab_lock(struct page *page)
354 bit_spin_lock(PG_locked, &page->flags);
357 static __always_inline void slab_unlock(struct page *page)
359 __bit_spin_unlock(PG_locked, &page->flags);
362 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
365 tmp.counters = counters_new;
367 * page->counters can cover frozen/inuse/objects as well
368 * as page->_count. If we assign to ->counters directly
369 * we run the risk of losing updates to page->_count, so
370 * be careful and only assign to the fields we need.
372 page->frozen = tmp.frozen;
373 page->inuse = tmp.inuse;
374 page->objects = tmp.objects;
377 /* Interrupts must be disabled (for the fallback code to work right) */
378 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
379 void *freelist_old, unsigned long counters_old,
380 void *freelist_new, unsigned long counters_new,
383 VM_BUG_ON(!irqs_disabled());
384 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
385 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
386 if (s->flags & __CMPXCHG_DOUBLE) {
387 if (cmpxchg_double(&page->freelist, &page->counters,
388 freelist_old, counters_old,
389 freelist_new, counters_new))
395 if (page->freelist == freelist_old &&
396 page->counters == counters_old) {
397 page->freelist = freelist_new;
398 set_page_slub_counters(page, counters_new);
406 stat(s, CMPXCHG_DOUBLE_FAIL);
408 #ifdef SLUB_DEBUG_CMPXCHG
409 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
415 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
416 void *freelist_old, unsigned long counters_old,
417 void *freelist_new, unsigned long counters_new,
420 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
421 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
422 if (s->flags & __CMPXCHG_DOUBLE) {
423 if (cmpxchg_double(&page->freelist, &page->counters,
424 freelist_old, counters_old,
425 freelist_new, counters_new))
432 local_irq_save(flags);
434 if (page->freelist == freelist_old &&
435 page->counters == counters_old) {
436 page->freelist = freelist_new;
437 set_page_slub_counters(page, counters_new);
439 local_irq_restore(flags);
443 local_irq_restore(flags);
447 stat(s, CMPXCHG_DOUBLE_FAIL);
449 #ifdef SLUB_DEBUG_CMPXCHG
450 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
456 #ifdef CONFIG_SLUB_DEBUG
458 * Determine a map of object in use on a page.
460 * Node listlock must be held to guarantee that the page does
461 * not vanish from under us.
463 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
466 void *addr = page_address(page);
468 for (p = page->freelist; p; p = get_freepointer(s, p))
469 set_bit(slab_index(p, s, addr), map);
475 #ifdef CONFIG_SLUB_DEBUG_ON
476 static int slub_debug = DEBUG_DEFAULT_FLAGS;
478 static int slub_debug;
481 static char *slub_debug_slabs;
482 static int disable_higher_order_debug;
487 static void print_section(char *text, u8 *addr, unsigned int length)
489 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
493 static struct track *get_track(struct kmem_cache *s, void *object,
494 enum track_item alloc)
499 p = object + s->offset + sizeof(void *);
501 p = object + s->inuse;
506 static void set_track(struct kmem_cache *s, void *object,
507 enum track_item alloc, unsigned long addr)
509 struct track *p = get_track(s, object, alloc);
512 #ifdef CONFIG_STACKTRACE
513 struct stack_trace trace;
516 trace.nr_entries = 0;
517 trace.max_entries = TRACK_ADDRS_COUNT;
518 trace.entries = p->addrs;
520 save_stack_trace(&trace);
522 /* See rant in lockdep.c */
523 if (trace.nr_entries != 0 &&
524 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
527 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
531 p->cpu = smp_processor_id();
532 p->pid = current->pid;
535 memset(p, 0, sizeof(struct track));
538 static void init_tracking(struct kmem_cache *s, void *object)
540 if (!(s->flags & SLAB_STORE_USER))
543 set_track(s, object, TRACK_FREE, 0UL);
544 set_track(s, object, TRACK_ALLOC, 0UL);
547 static void print_track(const char *s, struct track *t)
552 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
553 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
554 #ifdef CONFIG_STACKTRACE
557 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
559 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
566 static void print_tracking(struct kmem_cache *s, void *object)
568 if (!(s->flags & SLAB_STORE_USER))
571 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
572 print_track("Freed", get_track(s, object, TRACK_FREE));
575 static void print_page_info(struct page *page)
578 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
579 page, page->objects, page->inuse, page->freelist, page->flags);
583 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
589 vsnprintf(buf, sizeof(buf), fmt, args);
591 printk(KERN_ERR "========================================"
592 "=====================================\n");
593 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
594 printk(KERN_ERR "----------------------------------------"
595 "-------------------------------------\n\n");
597 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
600 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
606 vsnprintf(buf, sizeof(buf), fmt, args);
608 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
611 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
613 unsigned int off; /* Offset of last byte */
614 u8 *addr = page_address(page);
616 print_tracking(s, p);
618 print_page_info(page);
620 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
621 p, p - addr, get_freepointer(s, p));
624 print_section("Bytes b4 ", p - 16, 16);
626 print_section("Object ", p, min_t(unsigned long, s->object_size,
628 if (s->flags & SLAB_RED_ZONE)
629 print_section("Redzone ", p + s->object_size,
630 s->inuse - s->object_size);
633 off = s->offset + sizeof(void *);
637 if (s->flags & SLAB_STORE_USER)
638 off += 2 * sizeof(struct track);
641 /* Beginning of the filler is the free pointer */
642 print_section("Padding ", p + off, s->size - off);
647 static void object_err(struct kmem_cache *s, struct page *page,
648 u8 *object, char *reason)
650 slab_bug(s, "%s", reason);
651 print_trailer(s, page, object);
654 static void slab_err(struct kmem_cache *s, struct page *page,
655 const char *fmt, ...)
661 vsnprintf(buf, sizeof(buf), fmt, args);
663 slab_bug(s, "%s", buf);
664 print_page_info(page);
668 static void init_object(struct kmem_cache *s, void *object, u8 val)
672 if (s->flags & __OBJECT_POISON) {
673 memset(p, POISON_FREE, s->object_size - 1);
674 p[s->object_size - 1] = POISON_END;
677 if (s->flags & SLAB_RED_ZONE)
678 memset(p + s->object_size, val, s->inuse - s->object_size);
681 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
682 void *from, void *to)
684 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
685 memset(from, data, to - from);
688 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
689 u8 *object, char *what,
690 u8 *start, unsigned int value, unsigned int bytes)
695 fault = memchr_inv(start, value, bytes);
700 while (end > fault && end[-1] == value)
703 slab_bug(s, "%s overwritten", what);
704 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
705 fault, end - 1, fault[0], value);
706 print_trailer(s, page, object);
708 restore_bytes(s, what, value, fault, end);
716 * Bytes of the object to be managed.
717 * If the freepointer may overlay the object then the free
718 * pointer is the first word of the object.
720 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
723 * object + s->object_size
724 * Padding to reach word boundary. This is also used for Redzoning.
725 * Padding is extended by another word if Redzoning is enabled and
726 * object_size == inuse.
728 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
729 * 0xcc (RED_ACTIVE) for objects in use.
732 * Meta data starts here.
734 * A. Free pointer (if we cannot overwrite object on free)
735 * B. Tracking data for SLAB_STORE_USER
736 * C. Padding to reach required alignment boundary or at mininum
737 * one word if debugging is on to be able to detect writes
738 * before the word boundary.
740 * Padding is done using 0x5a (POISON_INUSE)
743 * Nothing is used beyond s->size.
745 * If slabcaches are merged then the object_size and inuse boundaries are mostly
746 * ignored. And therefore no slab options that rely on these boundaries
747 * may be used with merged slabcaches.
750 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
752 unsigned long off = s->inuse; /* The end of info */
755 /* Freepointer is placed after the object. */
756 off += sizeof(void *);
758 if (s->flags & SLAB_STORE_USER)
759 /* We also have user information there */
760 off += 2 * sizeof(struct track);
765 return check_bytes_and_report(s, page, p, "Object padding",
766 p + off, POISON_INUSE, s->size - off);
769 /* Check the pad bytes at the end of a slab page */
770 static int slab_pad_check(struct kmem_cache *s, struct page *page)
778 if (!(s->flags & SLAB_POISON))
781 start = page_address(page);
782 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
783 end = start + length;
784 remainder = length % s->size;
788 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
791 while (end > fault && end[-1] == POISON_INUSE)
794 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
795 print_section("Padding ", end - remainder, remainder);
797 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
801 static int check_object(struct kmem_cache *s, struct page *page,
802 void *object, u8 val)
805 u8 *endobject = object + s->object_size;
807 if (s->flags & SLAB_RED_ZONE) {
808 if (!check_bytes_and_report(s, page, object, "Redzone",
809 endobject, val, s->inuse - s->object_size))
812 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
813 check_bytes_and_report(s, page, p, "Alignment padding",
814 endobject, POISON_INUSE,
815 s->inuse - s->object_size);
819 if (s->flags & SLAB_POISON) {
820 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
821 (!check_bytes_and_report(s, page, p, "Poison", p,
822 POISON_FREE, s->object_size - 1) ||
823 !check_bytes_and_report(s, page, p, "Poison",
824 p + s->object_size - 1, POISON_END, 1)))
827 * check_pad_bytes cleans up on its own.
829 check_pad_bytes(s, page, p);
832 if (!s->offset && val == SLUB_RED_ACTIVE)
834 * Object and freepointer overlap. Cannot check
835 * freepointer while object is allocated.
839 /* Check free pointer validity */
840 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
841 object_err(s, page, p, "Freepointer corrupt");
843 * No choice but to zap it and thus lose the remainder
844 * of the free objects in this slab. May cause
845 * another error because the object count is now wrong.
847 set_freepointer(s, p, NULL);
853 static int check_slab(struct kmem_cache *s, struct page *page)
857 VM_BUG_ON(!irqs_disabled());
859 if (!PageSlab(page)) {
860 slab_err(s, page, "Not a valid slab page");
864 maxobj = order_objects(compound_order(page), s->size, s->reserved);
865 if (page->objects > maxobj) {
866 slab_err(s, page, "objects %u > max %u",
867 s->name, page->objects, maxobj);
870 if (page->inuse > page->objects) {
871 slab_err(s, page, "inuse %u > max %u",
872 s->name, page->inuse, page->objects);
875 /* Slab_pad_check fixes things up after itself */
876 slab_pad_check(s, page);
881 * Determine if a certain object on a page is on the freelist. Must hold the
882 * slab lock to guarantee that the chains are in a consistent state.
884 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
889 unsigned long max_objects;
892 while (fp && nr <= page->objects) {
895 if (!check_valid_pointer(s, page, fp)) {
897 object_err(s, page, object,
898 "Freechain corrupt");
899 set_freepointer(s, object, NULL);
901 slab_err(s, page, "Freepointer corrupt");
902 page->freelist = NULL;
903 page->inuse = page->objects;
904 slab_fix(s, "Freelist cleared");
910 fp = get_freepointer(s, object);
914 max_objects = order_objects(compound_order(page), s->size, s->reserved);
915 if (max_objects > MAX_OBJS_PER_PAGE)
916 max_objects = MAX_OBJS_PER_PAGE;
918 if (page->objects != max_objects) {
919 slab_err(s, page, "Wrong number of objects. Found %d but "
920 "should be %d", page->objects, max_objects);
921 page->objects = max_objects;
922 slab_fix(s, "Number of objects adjusted.");
924 if (page->inuse != page->objects - nr) {
925 slab_err(s, page, "Wrong object count. Counter is %d but "
926 "counted were %d", page->inuse, page->objects - nr);
927 page->inuse = page->objects - nr;
928 slab_fix(s, "Object count adjusted.");
930 return search == NULL;
933 static void trace(struct kmem_cache *s, struct page *page, void *object,
936 if (s->flags & SLAB_TRACE) {
937 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
939 alloc ? "alloc" : "free",
944 print_section("Object ", (void *)object,
952 * Hooks for other subsystems that check memory allocations. In a typical
953 * production configuration these hooks all should produce no code at all.
955 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
957 kmemleak_alloc(ptr, size, 1, flags);
960 static inline void kfree_hook(const void *x)
965 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
967 flags &= gfp_allowed_mask;
968 lockdep_trace_alloc(flags);
969 might_sleep_if(flags & __GFP_WAIT);
971 return should_failslab(s->object_size, flags, s->flags);
974 static inline void slab_post_alloc_hook(struct kmem_cache *s,
975 gfp_t flags, void *object)
977 flags &= gfp_allowed_mask;
978 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
979 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
982 static inline void slab_free_hook(struct kmem_cache *s, void *x)
984 kmemleak_free_recursive(x, s->flags);
987 * Trouble is that we may no longer disable interrupts in the fast path
988 * So in order to make the debug calls that expect irqs to be
989 * disabled we need to disable interrupts temporarily.
991 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
995 local_irq_save(flags);
996 kmemcheck_slab_free(s, x, s->object_size);
997 debug_check_no_locks_freed(x, s->object_size);
998 local_irq_restore(flags);
1001 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1002 debug_check_no_obj_freed(x, s->object_size);
1006 * Tracking of fully allocated slabs for debugging purposes.
1008 static void add_full(struct kmem_cache *s,
1009 struct kmem_cache_node *n, struct page *page)
1011 if (!(s->flags & SLAB_STORE_USER))
1014 lockdep_assert_held(&n->list_lock);
1015 list_add(&page->lru, &n->full);
1018 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1020 if (!(s->flags & SLAB_STORE_USER))
1023 lockdep_assert_held(&n->list_lock);
1024 list_del(&page->lru);
1027 /* Tracking of the number of slabs for debugging purposes */
1028 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1030 struct kmem_cache_node *n = get_node(s, node);
1032 return atomic_long_read(&n->nr_slabs);
1035 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1037 return atomic_long_read(&n->nr_slabs);
1040 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1042 struct kmem_cache_node *n = get_node(s, node);
1045 * May be called early in order to allocate a slab for the
1046 * kmem_cache_node structure. Solve the chicken-egg
1047 * dilemma by deferring the increment of the count during
1048 * bootstrap (see early_kmem_cache_node_alloc).
1051 atomic_long_inc(&n->nr_slabs);
1052 atomic_long_add(objects, &n->total_objects);
1055 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1057 struct kmem_cache_node *n = get_node(s, node);
1059 atomic_long_dec(&n->nr_slabs);
1060 atomic_long_sub(objects, &n->total_objects);
1063 /* Object debug checks for alloc/free paths */
1064 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1067 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1070 init_object(s, object, SLUB_RED_INACTIVE);
1071 init_tracking(s, object);
1074 static noinline int alloc_debug_processing(struct kmem_cache *s,
1076 void *object, unsigned long addr)
1078 if (!check_slab(s, page))
1081 if (!check_valid_pointer(s, page, object)) {
1082 object_err(s, page, object, "Freelist Pointer check fails");
1086 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1089 /* Success perform special debug activities for allocs */
1090 if (s->flags & SLAB_STORE_USER)
1091 set_track(s, object, TRACK_ALLOC, addr);
1092 trace(s, page, object, 1);
1093 init_object(s, object, SLUB_RED_ACTIVE);
1097 if (PageSlab(page)) {
1099 * If this is a slab page then lets do the best we can
1100 * to avoid issues in the future. Marking all objects
1101 * as used avoids touching the remaining objects.
1103 slab_fix(s, "Marking all objects used");
1104 page->inuse = page->objects;
1105 page->freelist = NULL;
1110 static noinline struct kmem_cache_node *free_debug_processing(
1111 struct kmem_cache *s, struct page *page, void *object,
1112 unsigned long addr, unsigned long *flags)
1114 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1116 spin_lock_irqsave(&n->list_lock, *flags);
1119 if (!check_slab(s, page))
1122 if (!check_valid_pointer(s, page, object)) {
1123 slab_err(s, page, "Invalid object pointer 0x%p", object);
1127 if (on_freelist(s, page, object)) {
1128 object_err(s, page, object, "Object already free");
1132 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1135 if (unlikely(s != page->slab_cache)) {
1136 if (!PageSlab(page)) {
1137 slab_err(s, page, "Attempt to free object(0x%p) "
1138 "outside of slab", object);
1139 } else if (!page->slab_cache) {
1141 "SLUB <none>: no slab for object 0x%p.\n",
1145 object_err(s, page, object,
1146 "page slab pointer corrupt.");
1150 if (s->flags & SLAB_STORE_USER)
1151 set_track(s, object, TRACK_FREE, addr);
1152 trace(s, page, object, 0);
1153 init_object(s, object, SLUB_RED_INACTIVE);
1157 * Keep node_lock to preserve integrity
1158 * until the object is actually freed
1164 spin_unlock_irqrestore(&n->list_lock, *flags);
1165 slab_fix(s, "Object at 0x%p not freed", object);
1169 static int __init setup_slub_debug(char *str)
1171 slub_debug = DEBUG_DEFAULT_FLAGS;
1172 if (*str++ != '=' || !*str)
1174 * No options specified. Switch on full debugging.
1180 * No options but restriction on slabs. This means full
1181 * debugging for slabs matching a pattern.
1185 if (tolower(*str) == 'o') {
1187 * Avoid enabling debugging on caches if its minimum order
1188 * would increase as a result.
1190 disable_higher_order_debug = 1;
1197 * Switch off all debugging measures.
1202 * Determine which debug features should be switched on
1204 for (; *str && *str != ','; str++) {
1205 switch (tolower(*str)) {
1207 slub_debug |= SLAB_DEBUG_FREE;
1210 slub_debug |= SLAB_RED_ZONE;
1213 slub_debug |= SLAB_POISON;
1216 slub_debug |= SLAB_STORE_USER;
1219 slub_debug |= SLAB_TRACE;
1222 slub_debug |= SLAB_FAILSLAB;
1225 printk(KERN_ERR "slub_debug option '%c' "
1226 "unknown. skipped\n", *str);
1232 slub_debug_slabs = str + 1;
1237 __setup("slub_debug", setup_slub_debug);
1239 static unsigned long kmem_cache_flags(unsigned long object_size,
1240 unsigned long flags, const char *name,
1241 void (*ctor)(void *))
1244 * Enable debugging if selected on the kernel commandline.
1246 if (slub_debug && (!slub_debug_slabs || (name &&
1247 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1248 flags |= slub_debug;
1253 static inline void setup_object_debug(struct kmem_cache *s,
1254 struct page *page, void *object) {}
1256 static inline int alloc_debug_processing(struct kmem_cache *s,
1257 struct page *page, void *object, unsigned long addr) { return 0; }
1259 static inline struct kmem_cache_node *free_debug_processing(
1260 struct kmem_cache *s, struct page *page, void *object,
1261 unsigned long addr, unsigned long *flags) { return NULL; }
1263 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1265 static inline int check_object(struct kmem_cache *s, struct page *page,
1266 void *object, u8 val) { return 1; }
1267 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1268 struct page *page) {}
1269 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1270 struct page *page) {}
1271 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1272 unsigned long flags, const char *name,
1273 void (*ctor)(void *))
1277 #define slub_debug 0
1279 #define disable_higher_order_debug 0
1281 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1283 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1285 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1287 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1290 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1292 kmemleak_alloc(ptr, size, 1, flags);
1295 static inline void kfree_hook(const void *x)
1300 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1303 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1306 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1307 flags & gfp_allowed_mask);
1310 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1312 kmemleak_free_recursive(x, s->flags);
1315 #endif /* CONFIG_SLUB_DEBUG */
1318 * Slab allocation and freeing
1320 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1321 struct kmem_cache_order_objects oo)
1323 int order = oo_order(oo);
1325 flags |= __GFP_NOTRACK;
1327 if (node == NUMA_NO_NODE)
1328 return alloc_pages(flags, order);
1330 return alloc_pages_exact_node(node, flags, order);
1333 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1336 struct kmem_cache_order_objects oo = s->oo;
1339 flags &= gfp_allowed_mask;
1341 if (flags & __GFP_WAIT)
1344 flags |= s->allocflags;
1347 * Let the initial higher-order allocation fail under memory pressure
1348 * so we fall-back to the minimum order allocation.
1350 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1352 page = alloc_slab_page(alloc_gfp, node, oo);
1353 if (unlikely(!page)) {
1356 * Allocation may have failed due to fragmentation.
1357 * Try a lower order alloc if possible
1359 page = alloc_slab_page(flags, node, oo);
1362 stat(s, ORDER_FALLBACK);
1365 if (kmemcheck_enabled && page
1366 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1367 int pages = 1 << oo_order(oo);
1369 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1372 * Objects from caches that have a constructor don't get
1373 * cleared when they're allocated, so we need to do it here.
1376 kmemcheck_mark_uninitialized_pages(page, pages);
1378 kmemcheck_mark_unallocated_pages(page, pages);
1381 if (flags & __GFP_WAIT)
1382 local_irq_disable();
1386 page->objects = oo_objects(oo);
1387 mod_zone_page_state(page_zone(page),
1388 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1389 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1395 static void setup_object(struct kmem_cache *s, struct page *page,
1398 setup_object_debug(s, page, object);
1399 if (unlikely(s->ctor))
1403 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1411 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1413 page = allocate_slab(s,
1414 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1418 order = compound_order(page);
1419 inc_slabs_node(s, page_to_nid(page), page->objects);
1420 memcg_bind_pages(s, order);
1421 page->slab_cache = s;
1422 __SetPageSlab(page);
1423 if (page->pfmemalloc)
1424 SetPageSlabPfmemalloc(page);
1426 start = page_address(page);
1428 if (unlikely(s->flags & SLAB_POISON))
1429 memset(start, POISON_INUSE, PAGE_SIZE << order);
1432 for_each_object(p, s, start, page->objects) {
1433 setup_object(s, page, last);
1434 set_freepointer(s, last, p);
1437 setup_object(s, page, last);
1438 set_freepointer(s, last, NULL);
1440 page->freelist = start;
1441 page->inuse = page->objects;
1447 static void __free_slab(struct kmem_cache *s, struct page *page)
1449 int order = compound_order(page);
1450 int pages = 1 << order;
1452 if (kmem_cache_debug(s)) {
1455 slab_pad_check(s, page);
1456 for_each_object(p, s, page_address(page),
1458 check_object(s, page, p, SLUB_RED_INACTIVE);
1461 kmemcheck_free_shadow(page, compound_order(page));
1463 mod_zone_page_state(page_zone(page),
1464 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1465 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1468 __ClearPageSlabPfmemalloc(page);
1469 __ClearPageSlab(page);
1471 memcg_release_pages(s, order);
1472 page_mapcount_reset(page);
1473 if (current->reclaim_state)
1474 current->reclaim_state->reclaimed_slab += pages;
1475 __free_memcg_kmem_pages(page, order);
1478 #define need_reserve_slab_rcu \
1479 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1481 static void rcu_free_slab(struct rcu_head *h)
1485 if (need_reserve_slab_rcu)
1486 page = virt_to_head_page(h);
1488 page = container_of((struct list_head *)h, struct page, lru);
1490 __free_slab(page->slab_cache, page);
1493 static void free_slab(struct kmem_cache *s, struct page *page)
1495 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1496 struct rcu_head *head;
1498 if (need_reserve_slab_rcu) {
1499 int order = compound_order(page);
1500 int offset = (PAGE_SIZE << order) - s->reserved;
1502 VM_BUG_ON(s->reserved != sizeof(*head));
1503 head = page_address(page) + offset;
1506 * RCU free overloads the RCU head over the LRU
1508 head = (void *)&page->lru;
1511 call_rcu(head, rcu_free_slab);
1513 __free_slab(s, page);
1516 static void discard_slab(struct kmem_cache *s, struct page *page)
1518 dec_slabs_node(s, page_to_nid(page), page->objects);
1523 * Management of partially allocated slabs.
1526 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1529 if (tail == DEACTIVATE_TO_TAIL)
1530 list_add_tail(&page->lru, &n->partial);
1532 list_add(&page->lru, &n->partial);
1535 static inline void add_partial(struct kmem_cache_node *n,
1536 struct page *page, int tail)
1538 lockdep_assert_held(&n->list_lock);
1539 __add_partial(n, page, tail);
1543 __remove_partial(struct kmem_cache_node *n, struct page *page)
1545 list_del(&page->lru);
1549 static inline void remove_partial(struct kmem_cache_node *n,
1552 lockdep_assert_held(&n->list_lock);
1553 __remove_partial(n, page);
1557 * Remove slab from the partial list, freeze it and
1558 * return the pointer to the freelist.
1560 * Returns a list of objects or NULL if it fails.
1562 static inline void *acquire_slab(struct kmem_cache *s,
1563 struct kmem_cache_node *n, struct page *page,
1564 int mode, int *objects)
1567 unsigned long counters;
1570 lockdep_assert_held(&n->list_lock);
1573 * Zap the freelist and set the frozen bit.
1574 * The old freelist is the list of objects for the
1575 * per cpu allocation list.
1577 freelist = page->freelist;
1578 counters = page->counters;
1579 new.counters = counters;
1580 *objects = new.objects - new.inuse;
1582 new.inuse = page->objects;
1583 new.freelist = NULL;
1585 new.freelist = freelist;
1588 VM_BUG_ON(new.frozen);
1591 if (!__cmpxchg_double_slab(s, page,
1593 new.freelist, new.counters,
1597 remove_partial(n, page);
1602 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1603 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1606 * Try to allocate a partial slab from a specific node.
1608 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1609 struct kmem_cache_cpu *c, gfp_t flags)
1611 struct page *page, *page2;
1612 void *object = NULL;
1617 * Racy check. If we mistakenly see no partial slabs then we
1618 * just allocate an empty slab. If we mistakenly try to get a
1619 * partial slab and there is none available then get_partials()
1622 if (!n || !n->nr_partial)
1625 spin_lock(&n->list_lock);
1626 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1629 if (!pfmemalloc_match(page, flags))
1632 t = acquire_slab(s, n, page, object == NULL, &objects);
1636 available += objects;
1639 stat(s, ALLOC_FROM_PARTIAL);
1642 put_cpu_partial(s, page, 0);
1643 stat(s, CPU_PARTIAL_NODE);
1645 if (!kmem_cache_has_cpu_partial(s)
1646 || available > s->cpu_partial / 2)
1650 spin_unlock(&n->list_lock);
1655 * Get a page from somewhere. Search in increasing NUMA distances.
1657 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1658 struct kmem_cache_cpu *c)
1661 struct zonelist *zonelist;
1664 enum zone_type high_zoneidx = gfp_zone(flags);
1666 unsigned int cpuset_mems_cookie;
1669 * The defrag ratio allows a configuration of the tradeoffs between
1670 * inter node defragmentation and node local allocations. A lower
1671 * defrag_ratio increases the tendency to do local allocations
1672 * instead of attempting to obtain partial slabs from other nodes.
1674 * If the defrag_ratio is set to 0 then kmalloc() always
1675 * returns node local objects. If the ratio is higher then kmalloc()
1676 * may return off node objects because partial slabs are obtained
1677 * from other nodes and filled up.
1679 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1680 * defrag_ratio = 1000) then every (well almost) allocation will
1681 * first attempt to defrag slab caches on other nodes. This means
1682 * scanning over all nodes to look for partial slabs which may be
1683 * expensive if we do it every time we are trying to find a slab
1684 * with available objects.
1686 if (!s->remote_node_defrag_ratio ||
1687 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1691 cpuset_mems_cookie = read_mems_allowed_begin();
1692 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1693 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1694 struct kmem_cache_node *n;
1696 n = get_node(s, zone_to_nid(zone));
1698 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1699 n->nr_partial > s->min_partial) {
1700 object = get_partial_node(s, n, c, flags);
1703 * Don't check read_mems_allowed_retry()
1704 * here - if mems_allowed was updated in
1705 * parallel, that was a harmless race
1706 * between allocation and the cpuset
1713 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1719 * Get a partial page, lock it and return it.
1721 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1722 struct kmem_cache_cpu *c)
1725 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1727 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1728 if (object || node != NUMA_NO_NODE)
1731 return get_any_partial(s, flags, c);
1734 #ifdef CONFIG_PREEMPT
1736 * Calculate the next globally unique transaction for disambiguiation
1737 * during cmpxchg. The transactions start with the cpu number and are then
1738 * incremented by CONFIG_NR_CPUS.
1740 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1743 * No preemption supported therefore also no need to check for
1749 static inline unsigned long next_tid(unsigned long tid)
1751 return tid + TID_STEP;
1754 static inline unsigned int tid_to_cpu(unsigned long tid)
1756 return tid % TID_STEP;
1759 static inline unsigned long tid_to_event(unsigned long tid)
1761 return tid / TID_STEP;
1764 static inline unsigned int init_tid(int cpu)
1769 static inline void note_cmpxchg_failure(const char *n,
1770 const struct kmem_cache *s, unsigned long tid)
1772 #ifdef SLUB_DEBUG_CMPXCHG
1773 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1775 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1777 #ifdef CONFIG_PREEMPT
1778 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1779 printk("due to cpu change %d -> %d\n",
1780 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1783 if (tid_to_event(tid) != tid_to_event(actual_tid))
1784 printk("due to cpu running other code. Event %ld->%ld\n",
1785 tid_to_event(tid), tid_to_event(actual_tid));
1787 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1788 actual_tid, tid, next_tid(tid));
1790 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1793 static void init_kmem_cache_cpus(struct kmem_cache *s)
1797 for_each_possible_cpu(cpu)
1798 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1802 * Remove the cpu slab
1804 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1807 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1808 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1810 enum slab_modes l = M_NONE, m = M_NONE;
1812 int tail = DEACTIVATE_TO_HEAD;
1816 if (page->freelist) {
1817 stat(s, DEACTIVATE_REMOTE_FREES);
1818 tail = DEACTIVATE_TO_TAIL;
1822 * Stage one: Free all available per cpu objects back
1823 * to the page freelist while it is still frozen. Leave the
1826 * There is no need to take the list->lock because the page
1829 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1831 unsigned long counters;
1834 prior = page->freelist;
1835 counters = page->counters;
1836 set_freepointer(s, freelist, prior);
1837 new.counters = counters;
1839 VM_BUG_ON(!new.frozen);
1841 } while (!__cmpxchg_double_slab(s, page,
1843 freelist, new.counters,
1844 "drain percpu freelist"));
1846 freelist = nextfree;
1850 * Stage two: Ensure that the page is unfrozen while the
1851 * list presence reflects the actual number of objects
1854 * We setup the list membership and then perform a cmpxchg
1855 * with the count. If there is a mismatch then the page
1856 * is not unfrozen but the page is on the wrong list.
1858 * Then we restart the process which may have to remove
1859 * the page from the list that we just put it on again
1860 * because the number of objects in the slab may have
1865 old.freelist = page->freelist;
1866 old.counters = page->counters;
1867 VM_BUG_ON(!old.frozen);
1869 /* Determine target state of the slab */
1870 new.counters = old.counters;
1873 set_freepointer(s, freelist, old.freelist);
1874 new.freelist = freelist;
1876 new.freelist = old.freelist;
1880 if (!new.inuse && n->nr_partial > s->min_partial)
1882 else if (new.freelist) {
1887 * Taking the spinlock removes the possiblity
1888 * that acquire_slab() will see a slab page that
1891 spin_lock(&n->list_lock);
1895 if (kmem_cache_debug(s) && !lock) {
1898 * This also ensures that the scanning of full
1899 * slabs from diagnostic functions will not see
1902 spin_lock(&n->list_lock);
1910 remove_partial(n, page);
1912 else if (l == M_FULL)
1914 remove_full(s, n, page);
1916 if (m == M_PARTIAL) {
1918 add_partial(n, page, tail);
1921 } else if (m == M_FULL) {
1923 stat(s, DEACTIVATE_FULL);
1924 add_full(s, n, page);
1930 if (!__cmpxchg_double_slab(s, page,
1931 old.freelist, old.counters,
1932 new.freelist, new.counters,
1937 spin_unlock(&n->list_lock);
1940 stat(s, DEACTIVATE_EMPTY);
1941 discard_slab(s, page);
1947 * Unfreeze all the cpu partial slabs.
1949 * This function must be called with interrupts disabled
1950 * for the cpu using c (or some other guarantee must be there
1951 * to guarantee no concurrent accesses).
1953 static void unfreeze_partials(struct kmem_cache *s,
1954 struct kmem_cache_cpu *c)
1956 #ifdef CONFIG_SLUB_CPU_PARTIAL
1957 struct kmem_cache_node *n = NULL, *n2 = NULL;
1958 struct page *page, *discard_page = NULL;
1960 while ((page = c->partial)) {
1964 c->partial = page->next;
1966 n2 = get_node(s, page_to_nid(page));
1969 spin_unlock(&n->list_lock);
1972 spin_lock(&n->list_lock);
1977 old.freelist = page->freelist;
1978 old.counters = page->counters;
1979 VM_BUG_ON(!old.frozen);
1981 new.counters = old.counters;
1982 new.freelist = old.freelist;
1986 } while (!__cmpxchg_double_slab(s, page,
1987 old.freelist, old.counters,
1988 new.freelist, new.counters,
1989 "unfreezing slab"));
1991 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1992 page->next = discard_page;
1993 discard_page = page;
1995 add_partial(n, page, DEACTIVATE_TO_TAIL);
1996 stat(s, FREE_ADD_PARTIAL);
2001 spin_unlock(&n->list_lock);
2003 while (discard_page) {
2004 page = discard_page;
2005 discard_page = discard_page->next;
2007 stat(s, DEACTIVATE_EMPTY);
2008 discard_slab(s, page);
2015 * Put a page that was just frozen (in __slab_free) into a partial page
2016 * slot if available. This is done without interrupts disabled and without
2017 * preemption disabled. The cmpxchg is racy and may put the partial page
2018 * onto a random cpus partial slot.
2020 * If we did not find a slot then simply move all the partials to the
2021 * per node partial list.
2023 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2025 #ifdef CONFIG_SLUB_CPU_PARTIAL
2026 struct page *oldpage;
2033 oldpage = this_cpu_read(s->cpu_slab->partial);
2036 pobjects = oldpage->pobjects;
2037 pages = oldpage->pages;
2038 if (drain && pobjects > s->cpu_partial) {
2039 unsigned long flags;
2041 * partial array is full. Move the existing
2042 * set to the per node partial list.
2044 local_irq_save(flags);
2045 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2046 local_irq_restore(flags);
2050 stat(s, CPU_PARTIAL_DRAIN);
2055 pobjects += page->objects - page->inuse;
2057 page->pages = pages;
2058 page->pobjects = pobjects;
2059 page->next = oldpage;
2061 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2066 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2068 stat(s, CPUSLAB_FLUSH);
2069 deactivate_slab(s, c->page, c->freelist);
2071 c->tid = next_tid(c->tid);
2079 * Called from IPI handler with interrupts disabled.
2081 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2083 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2089 unfreeze_partials(s, c);
2093 static void flush_cpu_slab(void *d)
2095 struct kmem_cache *s = d;
2097 __flush_cpu_slab(s, smp_processor_id());
2100 static bool has_cpu_slab(int cpu, void *info)
2102 struct kmem_cache *s = info;
2103 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2105 return c->page || c->partial;
2108 static void flush_all(struct kmem_cache *s)
2110 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2114 * Check if the objects in a per cpu structure fit numa
2115 * locality expectations.
2117 static inline int node_match(struct page *page, int node)
2120 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2126 static int count_free(struct page *page)
2128 return page->objects - page->inuse;
2131 static unsigned long count_partial(struct kmem_cache_node *n,
2132 int (*get_count)(struct page *))
2134 unsigned long flags;
2135 unsigned long x = 0;
2138 spin_lock_irqsave(&n->list_lock, flags);
2139 list_for_each_entry(page, &n->partial, lru)
2140 x += get_count(page);
2141 spin_unlock_irqrestore(&n->list_lock, flags);
2145 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2147 #ifdef CONFIG_SLUB_DEBUG
2148 return atomic_long_read(&n->total_objects);
2154 static noinline void
2155 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2160 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2162 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2163 "default order: %d, min order: %d\n", s->name, s->object_size,
2164 s->size, oo_order(s->oo), oo_order(s->min));
2166 if (oo_order(s->min) > get_order(s->object_size))
2167 printk(KERN_WARNING " %s debugging increased min order, use "
2168 "slub_debug=O to disable.\n", s->name);
2170 for_each_online_node(node) {
2171 struct kmem_cache_node *n = get_node(s, node);
2172 unsigned long nr_slabs;
2173 unsigned long nr_objs;
2174 unsigned long nr_free;
2179 nr_free = count_partial(n, count_free);
2180 nr_slabs = node_nr_slabs(n);
2181 nr_objs = node_nr_objs(n);
2184 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2185 node, nr_slabs, nr_objs, nr_free);
2189 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2190 int node, struct kmem_cache_cpu **pc)
2193 struct kmem_cache_cpu *c = *pc;
2196 freelist = get_partial(s, flags, node, c);
2201 page = new_slab(s, flags, node);
2203 c = __this_cpu_ptr(s->cpu_slab);
2208 * No other reference to the page yet so we can
2209 * muck around with it freely without cmpxchg
2211 freelist = page->freelist;
2212 page->freelist = NULL;
2214 stat(s, ALLOC_SLAB);
2223 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2225 if (unlikely(PageSlabPfmemalloc(page)))
2226 return gfp_pfmemalloc_allowed(gfpflags);
2232 * Check the page->freelist of a page and either transfer the freelist to the
2233 * per cpu freelist or deactivate the page.
2235 * The page is still frozen if the return value is not NULL.
2237 * If this function returns NULL then the page has been unfrozen.
2239 * This function must be called with interrupt disabled.
2241 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2244 unsigned long counters;
2248 freelist = page->freelist;
2249 counters = page->counters;
2251 new.counters = counters;
2252 VM_BUG_ON(!new.frozen);
2254 new.inuse = page->objects;
2255 new.frozen = freelist != NULL;
2257 } while (!__cmpxchg_double_slab(s, page,
2266 * Slow path. The lockless freelist is empty or we need to perform
2269 * Processing is still very fast if new objects have been freed to the
2270 * regular freelist. In that case we simply take over the regular freelist
2271 * as the lockless freelist and zap the regular freelist.
2273 * If that is not working then we fall back to the partial lists. We take the
2274 * first element of the freelist as the object to allocate now and move the
2275 * rest of the freelist to the lockless freelist.
2277 * And if we were unable to get a new slab from the partial slab lists then
2278 * we need to allocate a new slab. This is the slowest path since it involves
2279 * a call to the page allocator and the setup of a new slab.
2281 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2282 unsigned long addr, struct kmem_cache_cpu *c)
2286 unsigned long flags;
2288 local_irq_save(flags);
2289 #ifdef CONFIG_PREEMPT
2291 * We may have been preempted and rescheduled on a different
2292 * cpu before disabling interrupts. Need to reload cpu area
2295 c = this_cpu_ptr(s->cpu_slab);
2303 if (unlikely(!node_match(page, node))) {
2304 stat(s, ALLOC_NODE_MISMATCH);
2305 deactivate_slab(s, page, c->freelist);
2312 * By rights, we should be searching for a slab page that was
2313 * PFMEMALLOC but right now, we are losing the pfmemalloc
2314 * information when the page leaves the per-cpu allocator
2316 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2317 deactivate_slab(s, page, c->freelist);
2323 /* must check again c->freelist in case of cpu migration or IRQ */
2324 freelist = c->freelist;
2328 stat(s, ALLOC_SLOWPATH);
2330 freelist = get_freelist(s, page);
2334 stat(s, DEACTIVATE_BYPASS);
2338 stat(s, ALLOC_REFILL);
2342 * freelist is pointing to the list of objects to be used.
2343 * page is pointing to the page from which the objects are obtained.
2344 * That page must be frozen for per cpu allocations to work.
2346 VM_BUG_ON(!c->page->frozen);
2347 c->freelist = get_freepointer(s, freelist);
2348 c->tid = next_tid(c->tid);
2349 local_irq_restore(flags);
2355 page = c->page = c->partial;
2356 c->partial = page->next;
2357 stat(s, CPU_PARTIAL_ALLOC);
2362 freelist = new_slab_objects(s, gfpflags, node, &c);
2364 if (unlikely(!freelist)) {
2365 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2366 slab_out_of_memory(s, gfpflags, node);
2368 local_irq_restore(flags);
2373 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2376 /* Only entered in the debug case */
2377 if (kmem_cache_debug(s) &&
2378 !alloc_debug_processing(s, page, freelist, addr))
2379 goto new_slab; /* Slab failed checks. Next slab needed */
2381 deactivate_slab(s, page, get_freepointer(s, freelist));
2384 local_irq_restore(flags);
2389 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2390 * have the fastpath folded into their functions. So no function call
2391 * overhead for requests that can be satisfied on the fastpath.
2393 * The fastpath works by first checking if the lockless freelist can be used.
2394 * If not then __slab_alloc is called for slow processing.
2396 * Otherwise we can simply pick the next object from the lockless free list.
2398 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2399 gfp_t gfpflags, int node, unsigned long addr)
2402 struct kmem_cache_cpu *c;
2406 if (slab_pre_alloc_hook(s, gfpflags))
2409 s = memcg_kmem_get_cache(s, gfpflags);
2412 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2413 * enabled. We may switch back and forth between cpus while
2414 * reading from one cpu area. That does not matter as long
2415 * as we end up on the original cpu again when doing the cmpxchg.
2417 * Preemption is disabled for the retrieval of the tid because that
2418 * must occur from the current processor. We cannot allow rescheduling
2419 * on a different processor between the determination of the pointer
2420 * and the retrieval of the tid.
2423 c = __this_cpu_ptr(s->cpu_slab);
2426 * The transaction ids are globally unique per cpu and per operation on
2427 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2428 * occurs on the right processor and that there was no operation on the
2429 * linked list in between.
2434 object = c->freelist;
2436 if (unlikely(!object || !node_match(page, node)))
2437 object = __slab_alloc(s, gfpflags, node, addr, c);
2440 void *next_object = get_freepointer_safe(s, object);
2443 * The cmpxchg will only match if there was no additional
2444 * operation and if we are on the right processor.
2446 * The cmpxchg does the following atomically (without lock
2448 * 1. Relocate first pointer to the current per cpu area.
2449 * 2. Verify that tid and freelist have not been changed
2450 * 3. If they were not changed replace tid and freelist
2452 * Since this is without lock semantics the protection is only
2453 * against code executing on this cpu *not* from access by
2456 if (unlikely(!this_cpu_cmpxchg_double(
2457 s->cpu_slab->freelist, s->cpu_slab->tid,
2459 next_object, next_tid(tid)))) {
2461 note_cmpxchg_failure("slab_alloc", s, tid);
2464 prefetch_freepointer(s, next_object);
2465 stat(s, ALLOC_FASTPATH);
2468 if (unlikely(gfpflags & __GFP_ZERO) && object)
2469 memset(object, 0, s->object_size);
2471 slab_post_alloc_hook(s, gfpflags, object);
2476 static __always_inline void *slab_alloc(struct kmem_cache *s,
2477 gfp_t gfpflags, unsigned long addr)
2479 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2482 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2484 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2486 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2491 EXPORT_SYMBOL(kmem_cache_alloc);
2493 #ifdef CONFIG_TRACING
2494 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2496 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2497 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2500 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2504 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2506 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2508 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2509 s->object_size, s->size, gfpflags, node);
2513 EXPORT_SYMBOL(kmem_cache_alloc_node);
2515 #ifdef CONFIG_TRACING
2516 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2518 int node, size_t size)
2520 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2522 trace_kmalloc_node(_RET_IP_, ret,
2523 size, s->size, gfpflags, node);
2526 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2531 * Slow patch handling. This may still be called frequently since objects
2532 * have a longer lifetime than the cpu slabs in most processing loads.
2534 * So we still attempt to reduce cache line usage. Just take the slab
2535 * lock and free the item. If there is no additional partial page
2536 * handling required then we can return immediately.
2538 static void __slab_free(struct kmem_cache *s, struct page *page,
2539 void *x, unsigned long addr)
2542 void **object = (void *)x;
2545 unsigned long counters;
2546 struct kmem_cache_node *n = NULL;
2547 unsigned long uninitialized_var(flags);
2549 stat(s, FREE_SLOWPATH);
2551 if (kmem_cache_debug(s) &&
2552 !(n = free_debug_processing(s, page, x, addr, &flags)))
2557 spin_unlock_irqrestore(&n->list_lock, flags);
2560 prior = page->freelist;
2561 counters = page->counters;
2562 set_freepointer(s, object, prior);
2563 new.counters = counters;
2564 was_frozen = new.frozen;
2566 if ((!new.inuse || !prior) && !was_frozen) {
2568 if (kmem_cache_has_cpu_partial(s) && !prior) {
2571 * Slab was on no list before and will be
2573 * We can defer the list move and instead
2578 } else { /* Needs to be taken off a list */
2580 n = get_node(s, page_to_nid(page));
2582 * Speculatively acquire the list_lock.
2583 * If the cmpxchg does not succeed then we may
2584 * drop the list_lock without any processing.
2586 * Otherwise the list_lock will synchronize with
2587 * other processors updating the list of slabs.
2589 spin_lock_irqsave(&n->list_lock, flags);
2594 } while (!cmpxchg_double_slab(s, page,
2596 object, new.counters,
2602 * If we just froze the page then put it onto the
2603 * per cpu partial list.
2605 if (new.frozen && !was_frozen) {
2606 put_cpu_partial(s, page, 1);
2607 stat(s, CPU_PARTIAL_FREE);
2610 * The list lock was not taken therefore no list
2611 * activity can be necessary.
2614 stat(s, FREE_FROZEN);
2618 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2622 * Objects left in the slab. If it was not on the partial list before
2625 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2626 if (kmem_cache_debug(s))
2627 remove_full(s, n, page);
2628 add_partial(n, page, DEACTIVATE_TO_TAIL);
2629 stat(s, FREE_ADD_PARTIAL);
2631 spin_unlock_irqrestore(&n->list_lock, flags);
2637 * Slab on the partial list.
2639 remove_partial(n, page);
2640 stat(s, FREE_REMOVE_PARTIAL);
2642 /* Slab must be on the full list */
2643 remove_full(s, n, page);
2646 spin_unlock_irqrestore(&n->list_lock, flags);
2648 discard_slab(s, page);
2652 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2653 * can perform fastpath freeing without additional function calls.
2655 * The fastpath is only possible if we are freeing to the current cpu slab
2656 * of this processor. This typically the case if we have just allocated
2659 * If fastpath is not possible then fall back to __slab_free where we deal
2660 * with all sorts of special processing.
2662 static __always_inline void slab_free(struct kmem_cache *s,
2663 struct page *page, void *x, unsigned long addr)
2665 void **object = (void *)x;
2666 struct kmem_cache_cpu *c;
2669 slab_free_hook(s, x);
2673 * Determine the currently cpus per cpu slab.
2674 * The cpu may change afterward. However that does not matter since
2675 * data is retrieved via this pointer. If we are on the same cpu
2676 * during the cmpxchg then the free will succedd.
2679 c = __this_cpu_ptr(s->cpu_slab);
2684 if (likely(page == c->page)) {
2685 set_freepointer(s, object, c->freelist);
2687 if (unlikely(!this_cpu_cmpxchg_double(
2688 s->cpu_slab->freelist, s->cpu_slab->tid,
2690 object, next_tid(tid)))) {
2692 note_cmpxchg_failure("slab_free", s, tid);
2695 stat(s, FREE_FASTPATH);
2697 __slab_free(s, page, x, addr);
2701 void kmem_cache_free(struct kmem_cache *s, void *x)
2703 s = cache_from_obj(s, x);
2706 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2707 trace_kmem_cache_free(_RET_IP_, x);
2709 EXPORT_SYMBOL(kmem_cache_free);
2712 * Object placement in a slab is made very easy because we always start at
2713 * offset 0. If we tune the size of the object to the alignment then we can
2714 * get the required alignment by putting one properly sized object after
2717 * Notice that the allocation order determines the sizes of the per cpu
2718 * caches. Each processor has always one slab available for allocations.
2719 * Increasing the allocation order reduces the number of times that slabs
2720 * must be moved on and off the partial lists and is therefore a factor in
2725 * Mininum / Maximum order of slab pages. This influences locking overhead
2726 * and slab fragmentation. A higher order reduces the number of partial slabs
2727 * and increases the number of allocations possible without having to
2728 * take the list_lock.
2730 static int slub_min_order;
2731 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2732 static int slub_min_objects;
2735 * Merge control. If this is set then no merging of slab caches will occur.
2736 * (Could be removed. This was introduced to pacify the merge skeptics.)
2738 static int slub_nomerge;
2741 * Calculate the order of allocation given an slab object size.
2743 * The order of allocation has significant impact on performance and other
2744 * system components. Generally order 0 allocations should be preferred since
2745 * order 0 does not cause fragmentation in the page allocator. Larger objects
2746 * be problematic to put into order 0 slabs because there may be too much
2747 * unused space left. We go to a higher order if more than 1/16th of the slab
2750 * In order to reach satisfactory performance we must ensure that a minimum
2751 * number of objects is in one slab. Otherwise we may generate too much
2752 * activity on the partial lists which requires taking the list_lock. This is
2753 * less a concern for large slabs though which are rarely used.
2755 * slub_max_order specifies the order where we begin to stop considering the
2756 * number of objects in a slab as critical. If we reach slub_max_order then
2757 * we try to keep the page order as low as possible. So we accept more waste
2758 * of space in favor of a small page order.
2760 * Higher order allocations also allow the placement of more objects in a
2761 * slab and thereby reduce object handling overhead. If the user has
2762 * requested a higher mininum order then we start with that one instead of
2763 * the smallest order which will fit the object.
2765 static inline int slab_order(int size, int min_objects,
2766 int max_order, int fract_leftover, int reserved)
2770 int min_order = slub_min_order;
2772 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2773 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2775 for (order = max(min_order,
2776 fls(min_objects * size - 1) - PAGE_SHIFT);
2777 order <= max_order; order++) {
2779 unsigned long slab_size = PAGE_SIZE << order;
2781 if (slab_size < min_objects * size + reserved)
2784 rem = (slab_size - reserved) % size;
2786 if (rem <= slab_size / fract_leftover)
2794 static inline int calculate_order(int size, int reserved)
2802 * Attempt to find best configuration for a slab. This
2803 * works by first attempting to generate a layout with
2804 * the best configuration and backing off gradually.
2806 * First we reduce the acceptable waste in a slab. Then
2807 * we reduce the minimum objects required in a slab.
2809 min_objects = slub_min_objects;
2811 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2812 max_objects = order_objects(slub_max_order, size, reserved);
2813 min_objects = min(min_objects, max_objects);
2815 while (min_objects > 1) {
2817 while (fraction >= 4) {
2818 order = slab_order(size, min_objects,
2819 slub_max_order, fraction, reserved);
2820 if (order <= slub_max_order)
2828 * We were unable to place multiple objects in a slab. Now
2829 * lets see if we can place a single object there.
2831 order = slab_order(size, 1, slub_max_order, 1, reserved);
2832 if (order <= slub_max_order)
2836 * Doh this slab cannot be placed using slub_max_order.
2838 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2839 if (order < MAX_ORDER)
2845 init_kmem_cache_node(struct kmem_cache_node *n)
2848 spin_lock_init(&n->list_lock);
2849 INIT_LIST_HEAD(&n->partial);
2850 #ifdef CONFIG_SLUB_DEBUG
2851 atomic_long_set(&n->nr_slabs, 0);
2852 atomic_long_set(&n->total_objects, 0);
2853 INIT_LIST_HEAD(&n->full);
2857 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2859 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2860 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2863 * Must align to double word boundary for the double cmpxchg
2864 * instructions to work; see __pcpu_double_call_return_bool().
2866 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2867 2 * sizeof(void *));
2872 init_kmem_cache_cpus(s);
2877 static struct kmem_cache *kmem_cache_node;
2880 * No kmalloc_node yet so do it by hand. We know that this is the first
2881 * slab on the node for this slabcache. There are no concurrent accesses
2884 * Note that this function only works on the kmem_cache_node
2885 * when allocating for the kmem_cache_node. This is used for bootstrapping
2886 * memory on a fresh node that has no slab structures yet.
2888 static void early_kmem_cache_node_alloc(int node)
2891 struct kmem_cache_node *n;
2893 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2895 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2898 if (page_to_nid(page) != node) {
2899 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2901 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2902 "in order to be able to continue\n");
2907 page->freelist = get_freepointer(kmem_cache_node, n);
2910 kmem_cache_node->node[node] = n;
2911 #ifdef CONFIG_SLUB_DEBUG
2912 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2913 init_tracking(kmem_cache_node, n);
2915 init_kmem_cache_node(n);
2916 inc_slabs_node(kmem_cache_node, node, page->objects);
2919 * No locks need to be taken here as it has just been
2920 * initialized and there is no concurrent access.
2922 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2925 static void free_kmem_cache_nodes(struct kmem_cache *s)
2929 for_each_node_state(node, N_NORMAL_MEMORY) {
2930 struct kmem_cache_node *n = s->node[node];
2933 kmem_cache_free(kmem_cache_node, n);
2935 s->node[node] = NULL;
2939 static int init_kmem_cache_nodes(struct kmem_cache *s)
2943 for_each_node_state(node, N_NORMAL_MEMORY) {
2944 struct kmem_cache_node *n;
2946 if (slab_state == DOWN) {
2947 early_kmem_cache_node_alloc(node);
2950 n = kmem_cache_alloc_node(kmem_cache_node,
2954 free_kmem_cache_nodes(s);
2959 init_kmem_cache_node(n);
2964 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2966 if (min < MIN_PARTIAL)
2968 else if (min > MAX_PARTIAL)
2970 s->min_partial = min;
2974 * calculate_sizes() determines the order and the distribution of data within
2977 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2979 unsigned long flags = s->flags;
2980 unsigned long size = s->object_size;
2984 * Round up object size to the next word boundary. We can only
2985 * place the free pointer at word boundaries and this determines
2986 * the possible location of the free pointer.
2988 size = ALIGN(size, sizeof(void *));
2990 #ifdef CONFIG_SLUB_DEBUG
2992 * Determine if we can poison the object itself. If the user of
2993 * the slab may touch the object after free or before allocation
2994 * then we should never poison the object itself.
2996 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2998 s->flags |= __OBJECT_POISON;
3000 s->flags &= ~__OBJECT_POISON;
3004 * If we are Redzoning then check if there is some space between the
3005 * end of the object and the free pointer. If not then add an
3006 * additional word to have some bytes to store Redzone information.
3008 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3009 size += sizeof(void *);
3013 * With that we have determined the number of bytes in actual use
3014 * by the object. This is the potential offset to the free pointer.
3018 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3021 * Relocate free pointer after the object if it is not
3022 * permitted to overwrite the first word of the object on
3025 * This is the case if we do RCU, have a constructor or
3026 * destructor or are poisoning the objects.
3029 size += sizeof(void *);
3032 #ifdef CONFIG_SLUB_DEBUG
3033 if (flags & SLAB_STORE_USER)
3035 * Need to store information about allocs and frees after
3038 size += 2 * sizeof(struct track);
3040 if (flags & SLAB_RED_ZONE)
3042 * Add some empty padding so that we can catch
3043 * overwrites from earlier objects rather than let
3044 * tracking information or the free pointer be
3045 * corrupted if a user writes before the start
3048 size += sizeof(void *);
3052 * SLUB stores one object immediately after another beginning from
3053 * offset 0. In order to align the objects we have to simply size
3054 * each object to conform to the alignment.
3056 size = ALIGN(size, s->align);
3058 if (forced_order >= 0)
3059 order = forced_order;
3061 order = calculate_order(size, s->reserved);
3068 s->allocflags |= __GFP_COMP;
3070 if (s->flags & SLAB_CACHE_DMA)
3071 s->allocflags |= GFP_DMA;
3073 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3074 s->allocflags |= __GFP_RECLAIMABLE;
3077 * Determine the number of objects per slab
3079 s->oo = oo_make(order, size, s->reserved);
3080 s->min = oo_make(get_order(size), size, s->reserved);
3081 if (oo_objects(s->oo) > oo_objects(s->max))
3084 return !!oo_objects(s->oo);
3087 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3089 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3092 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3093 s->reserved = sizeof(struct rcu_head);
3095 if (!calculate_sizes(s, -1))
3097 if (disable_higher_order_debug) {
3099 * Disable debugging flags that store metadata if the min slab
3102 if (get_order(s->size) > get_order(s->object_size)) {
3103 s->flags &= ~DEBUG_METADATA_FLAGS;
3105 if (!calculate_sizes(s, -1))
3110 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3111 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3112 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3113 /* Enable fast mode */
3114 s->flags |= __CMPXCHG_DOUBLE;
3118 * The larger the object size is, the more pages we want on the partial
3119 * list to avoid pounding the page allocator excessively.
3121 set_min_partial(s, ilog2(s->size) / 2);
3124 * cpu_partial determined the maximum number of objects kept in the
3125 * per cpu partial lists of a processor.
3127 * Per cpu partial lists mainly contain slabs that just have one
3128 * object freed. If they are used for allocation then they can be
3129 * filled up again with minimal effort. The slab will never hit the
3130 * per node partial lists and therefore no locking will be required.
3132 * This setting also determines
3134 * A) The number of objects from per cpu partial slabs dumped to the
3135 * per node list when we reach the limit.
3136 * B) The number of objects in cpu partial slabs to extract from the
3137 * per node list when we run out of per cpu objects. We only fetch
3138 * 50% to keep some capacity around for frees.
3140 if (!kmem_cache_has_cpu_partial(s))
3142 else if (s->size >= PAGE_SIZE)
3144 else if (s->size >= 1024)
3146 else if (s->size >= 256)
3147 s->cpu_partial = 13;
3149 s->cpu_partial = 30;
3152 s->remote_node_defrag_ratio = 1000;
3154 if (!init_kmem_cache_nodes(s))
3157 if (alloc_kmem_cache_cpus(s))
3160 free_kmem_cache_nodes(s);
3162 if (flags & SLAB_PANIC)
3163 panic("Cannot create slab %s size=%lu realsize=%u "
3164 "order=%u offset=%u flags=%lx\n",
3165 s->name, (unsigned long)s->size, s->size,
3166 oo_order(s->oo), s->offset, flags);
3170 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3173 #ifdef CONFIG_SLUB_DEBUG
3174 void *addr = page_address(page);
3176 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3177 sizeof(long), GFP_ATOMIC);
3180 slab_err(s, page, text, s->name);
3183 get_map(s, page, map);
3184 for_each_object(p, s, addr, page->objects) {
3186 if (!test_bit(slab_index(p, s, addr), map)) {
3187 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3189 print_tracking(s, p);
3198 * Attempt to free all partial slabs on a node.
3199 * This is called from kmem_cache_close(). We must be the last thread
3200 * using the cache and therefore we do not need to lock anymore.
3202 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3204 struct page *page, *h;
3206 list_for_each_entry_safe(page, h, &n->partial, lru) {
3208 __remove_partial(n, page);
3209 discard_slab(s, page);
3211 list_slab_objects(s, page,
3212 "Objects remaining in %s on kmem_cache_close()");
3218 * Release all resources used by a slab cache.
3220 static inline int kmem_cache_close(struct kmem_cache *s)
3225 /* Attempt to free all objects */
3226 for_each_node_state(node, N_NORMAL_MEMORY) {
3227 struct kmem_cache_node *n = get_node(s, node);
3230 if (n->nr_partial || slabs_node(s, node))
3233 free_percpu(s->cpu_slab);
3234 free_kmem_cache_nodes(s);
3238 int __kmem_cache_shutdown(struct kmem_cache *s)
3240 int rc = kmem_cache_close(s);
3244 * Since slab_attr_store may take the slab_mutex, we should
3245 * release the lock while removing the sysfs entry in order to
3246 * avoid a deadlock. Because this is pretty much the last
3247 * operation we do and the lock will be released shortly after
3248 * that in slab_common.c, we could just move sysfs_slab_remove
3249 * to a later point in common code. We should do that when we
3250 * have a common sysfs framework for all allocators.
3252 mutex_unlock(&slab_mutex);
3253 sysfs_slab_remove(s);
3254 mutex_lock(&slab_mutex);
3260 /********************************************************************
3262 *******************************************************************/
3264 static int __init setup_slub_min_order(char *str)
3266 get_option(&str, &slub_min_order);
3271 __setup("slub_min_order=", setup_slub_min_order);
3273 static int __init setup_slub_max_order(char *str)
3275 get_option(&str, &slub_max_order);
3276 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3281 __setup("slub_max_order=", setup_slub_max_order);
3283 static int __init setup_slub_min_objects(char *str)
3285 get_option(&str, &slub_min_objects);
3290 __setup("slub_min_objects=", setup_slub_min_objects);
3292 static int __init setup_slub_nomerge(char *str)
3298 __setup("slub_nomerge", setup_slub_nomerge);
3300 void *__kmalloc(size_t size, gfp_t flags)
3302 struct kmem_cache *s;
3305 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3306 return kmalloc_large(size, flags);
3308 s = kmalloc_slab(size, flags);
3310 if (unlikely(ZERO_OR_NULL_PTR(s)))
3313 ret = slab_alloc(s, flags, _RET_IP_);
3315 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3319 EXPORT_SYMBOL(__kmalloc);
3322 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3327 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3328 page = alloc_pages_node(node, flags, get_order(size));
3330 ptr = page_address(page);
3332 kmalloc_large_node_hook(ptr, size, flags);
3336 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3338 struct kmem_cache *s;
3341 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3342 ret = kmalloc_large_node(size, flags, node);
3344 trace_kmalloc_node(_RET_IP_, ret,
3345 size, PAGE_SIZE << get_order(size),
3351 s = kmalloc_slab(size, flags);
3353 if (unlikely(ZERO_OR_NULL_PTR(s)))
3356 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3358 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3362 EXPORT_SYMBOL(__kmalloc_node);
3365 size_t ksize(const void *object)
3369 if (unlikely(object == ZERO_SIZE_PTR))
3372 page = virt_to_head_page(object);
3374 if (unlikely(!PageSlab(page))) {
3375 WARN_ON(!PageCompound(page));
3376 return PAGE_SIZE << compound_order(page);
3379 return slab_ksize(page->slab_cache);
3381 EXPORT_SYMBOL(ksize);
3383 void kfree(const void *x)
3386 void *object = (void *)x;
3388 trace_kfree(_RET_IP_, x);
3390 if (unlikely(ZERO_OR_NULL_PTR(x)))
3393 page = virt_to_head_page(x);
3394 if (unlikely(!PageSlab(page))) {
3395 BUG_ON(!PageCompound(page));
3397 __free_memcg_kmem_pages(page, compound_order(page));
3400 slab_free(page->slab_cache, page, object, _RET_IP_);
3402 EXPORT_SYMBOL(kfree);
3405 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3406 * the remaining slabs by the number of items in use. The slabs with the
3407 * most items in use come first. New allocations will then fill those up
3408 * and thus they can be removed from the partial lists.
3410 * The slabs with the least items are placed last. This results in them
3411 * being allocated from last increasing the chance that the last objects
3412 * are freed in them.
3414 int kmem_cache_shrink(struct kmem_cache *s)
3418 struct kmem_cache_node *n;
3421 int objects = oo_objects(s->max);
3422 struct list_head *slabs_by_inuse =
3423 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3424 unsigned long flags;
3426 if (!slabs_by_inuse)
3430 for_each_node_state(node, N_NORMAL_MEMORY) {
3431 n = get_node(s, node);
3436 for (i = 0; i < objects; i++)
3437 INIT_LIST_HEAD(slabs_by_inuse + i);
3439 spin_lock_irqsave(&n->list_lock, flags);
3442 * Build lists indexed by the items in use in each slab.
3444 * Note that concurrent frees may occur while we hold the
3445 * list_lock. page->inuse here is the upper limit.
3447 list_for_each_entry_safe(page, t, &n->partial, lru) {
3448 list_move(&page->lru, slabs_by_inuse + page->inuse);
3454 * Rebuild the partial list with the slabs filled up most
3455 * first and the least used slabs at the end.
3457 for (i = objects - 1; i > 0; i--)
3458 list_splice(slabs_by_inuse + i, n->partial.prev);
3460 spin_unlock_irqrestore(&n->list_lock, flags);
3462 /* Release empty slabs */
3463 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3464 discard_slab(s, page);
3467 kfree(slabs_by_inuse);
3470 EXPORT_SYMBOL(kmem_cache_shrink);
3472 static int slab_mem_going_offline_callback(void *arg)
3474 struct kmem_cache *s;
3476 mutex_lock(&slab_mutex);
3477 list_for_each_entry(s, &slab_caches, list)
3478 kmem_cache_shrink(s);
3479 mutex_unlock(&slab_mutex);
3484 static void slab_mem_offline_callback(void *arg)
3486 struct kmem_cache_node *n;
3487 struct kmem_cache *s;
3488 struct memory_notify *marg = arg;
3491 offline_node = marg->status_change_nid_normal;
3494 * If the node still has available memory. we need kmem_cache_node
3497 if (offline_node < 0)
3500 mutex_lock(&slab_mutex);
3501 list_for_each_entry(s, &slab_caches, list) {
3502 n = get_node(s, offline_node);
3505 * if n->nr_slabs > 0, slabs still exist on the node
3506 * that is going down. We were unable to free them,
3507 * and offline_pages() function shouldn't call this
3508 * callback. So, we must fail.
3510 BUG_ON(slabs_node(s, offline_node));
3512 s->node[offline_node] = NULL;
3513 kmem_cache_free(kmem_cache_node, n);
3516 mutex_unlock(&slab_mutex);
3519 static int slab_mem_going_online_callback(void *arg)
3521 struct kmem_cache_node *n;
3522 struct kmem_cache *s;
3523 struct memory_notify *marg = arg;
3524 int nid = marg->status_change_nid_normal;
3528 * If the node's memory is already available, then kmem_cache_node is
3529 * already created. Nothing to do.
3535 * We are bringing a node online. No memory is available yet. We must
3536 * allocate a kmem_cache_node structure in order to bring the node
3539 mutex_lock(&slab_mutex);
3540 list_for_each_entry(s, &slab_caches, list) {
3542 * XXX: kmem_cache_alloc_node will fallback to other nodes
3543 * since memory is not yet available from the node that
3546 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3551 init_kmem_cache_node(n);
3555 mutex_unlock(&slab_mutex);
3559 static int slab_memory_callback(struct notifier_block *self,
3560 unsigned long action, void *arg)
3565 case MEM_GOING_ONLINE:
3566 ret = slab_mem_going_online_callback(arg);
3568 case MEM_GOING_OFFLINE:
3569 ret = slab_mem_going_offline_callback(arg);
3572 case MEM_CANCEL_ONLINE:
3573 slab_mem_offline_callback(arg);
3576 case MEM_CANCEL_OFFLINE:
3580 ret = notifier_from_errno(ret);
3586 static struct notifier_block slab_memory_callback_nb = {
3587 .notifier_call = slab_memory_callback,
3588 .priority = SLAB_CALLBACK_PRI,
3591 /********************************************************************
3592 * Basic setup of slabs
3593 *******************************************************************/
3596 * Used for early kmem_cache structures that were allocated using
3597 * the page allocator. Allocate them properly then fix up the pointers
3598 * that may be pointing to the wrong kmem_cache structure.
3601 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3604 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3606 memcpy(s, static_cache, kmem_cache->object_size);
3609 * This runs very early, and only the boot processor is supposed to be
3610 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3613 __flush_cpu_slab(s, smp_processor_id());
3614 for_each_node_state(node, N_NORMAL_MEMORY) {
3615 struct kmem_cache_node *n = get_node(s, node);
3619 list_for_each_entry(p, &n->partial, lru)
3622 #ifdef CONFIG_SLUB_DEBUG
3623 list_for_each_entry(p, &n->full, lru)
3628 list_add(&s->list, &slab_caches);
3632 void __init kmem_cache_init(void)
3634 static __initdata struct kmem_cache boot_kmem_cache,
3635 boot_kmem_cache_node;
3637 if (debug_guardpage_minorder())
3640 kmem_cache_node = &boot_kmem_cache_node;
3641 kmem_cache = &boot_kmem_cache;
3643 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3644 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3646 register_hotmemory_notifier(&slab_memory_callback_nb);
3648 /* Able to allocate the per node structures */
3649 slab_state = PARTIAL;
3651 create_boot_cache(kmem_cache, "kmem_cache",
3652 offsetof(struct kmem_cache, node) +
3653 nr_node_ids * sizeof(struct kmem_cache_node *),
3654 SLAB_HWCACHE_ALIGN);
3656 kmem_cache = bootstrap(&boot_kmem_cache);
3659 * Allocate kmem_cache_node properly from the kmem_cache slab.
3660 * kmem_cache_node is separately allocated so no need to
3661 * update any list pointers.
3663 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3665 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3666 create_kmalloc_caches(0);
3669 register_cpu_notifier(&slab_notifier);
3673 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3674 " CPUs=%d, Nodes=%d\n",
3676 slub_min_order, slub_max_order, slub_min_objects,
3677 nr_cpu_ids, nr_node_ids);
3680 void __init kmem_cache_init_late(void)
3685 * Find a mergeable slab cache
3687 static int slab_unmergeable(struct kmem_cache *s)
3689 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3692 if (!is_root_cache(s))
3699 * We may have set a slab to be unmergeable during bootstrap.
3701 if (s->refcount < 0)
3707 static struct kmem_cache *find_mergeable(size_t size, size_t align,
3708 unsigned long flags, const char *name, void (*ctor)(void *))
3710 struct kmem_cache *s;
3712 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3718 size = ALIGN(size, sizeof(void *));
3719 align = calculate_alignment(flags, align, size);
3720 size = ALIGN(size, align);
3721 flags = kmem_cache_flags(size, flags, name, NULL);
3723 list_for_each_entry(s, &slab_caches, list) {
3724 if (slab_unmergeable(s))
3730 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3733 * Check if alignment is compatible.
3734 * Courtesy of Adrian Drzewiecki
3736 if ((s->size & ~(align - 1)) != s->size)
3739 if (s->size - size >= sizeof(void *))
3748 __kmem_cache_alias(const char *name, size_t size, size_t align,
3749 unsigned long flags, void (*ctor)(void *))
3751 struct kmem_cache *s;
3753 s = find_mergeable(size, align, flags, name, ctor);
3756 struct kmem_cache *c;
3761 * Adjust the object sizes so that we clear
3762 * the complete object on kzalloc.
3764 s->object_size = max(s->object_size, (int)size);
3765 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3767 for_each_memcg_cache_index(i) {
3768 c = cache_from_memcg_idx(s, i);
3771 c->object_size = s->object_size;
3772 c->inuse = max_t(int, c->inuse,
3773 ALIGN(size, sizeof(void *)));
3776 if (sysfs_slab_alias(s, name)) {
3785 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3789 err = kmem_cache_open(s, flags);
3793 /* Mutex is not taken during early boot */
3794 if (slab_state <= UP)
3797 memcg_propagate_slab_attrs(s);
3798 err = sysfs_slab_add(s);
3800 kmem_cache_close(s);
3807 * Use the cpu notifier to insure that the cpu slabs are flushed when
3810 static int slab_cpuup_callback(struct notifier_block *nfb,
3811 unsigned long action, void *hcpu)
3813 long cpu = (long)hcpu;
3814 struct kmem_cache *s;
3815 unsigned long flags;
3818 case CPU_UP_CANCELED:
3819 case CPU_UP_CANCELED_FROZEN:
3821 case CPU_DEAD_FROZEN:
3822 mutex_lock(&slab_mutex);
3823 list_for_each_entry(s, &slab_caches, list) {
3824 local_irq_save(flags);
3825 __flush_cpu_slab(s, cpu);
3826 local_irq_restore(flags);
3828 mutex_unlock(&slab_mutex);
3836 static struct notifier_block slab_notifier = {
3837 .notifier_call = slab_cpuup_callback
3842 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3844 struct kmem_cache *s;
3847 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3848 return kmalloc_large(size, gfpflags);
3850 s = kmalloc_slab(size, gfpflags);
3852 if (unlikely(ZERO_OR_NULL_PTR(s)))
3855 ret = slab_alloc(s, gfpflags, caller);
3857 /* Honor the call site pointer we received. */
3858 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3864 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3865 int node, unsigned long caller)
3867 struct kmem_cache *s;
3870 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3871 ret = kmalloc_large_node(size, gfpflags, node);
3873 trace_kmalloc_node(caller, ret,
3874 size, PAGE_SIZE << get_order(size),
3880 s = kmalloc_slab(size, gfpflags);
3882 if (unlikely(ZERO_OR_NULL_PTR(s)))
3885 ret = slab_alloc_node(s, gfpflags, node, caller);
3887 /* Honor the call site pointer we received. */
3888 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3895 static int count_inuse(struct page *page)
3900 static int count_total(struct page *page)
3902 return page->objects;
3906 #ifdef CONFIG_SLUB_DEBUG
3907 static int validate_slab(struct kmem_cache *s, struct page *page,
3911 void *addr = page_address(page);
3913 if (!check_slab(s, page) ||
3914 !on_freelist(s, page, NULL))
3917 /* Now we know that a valid freelist exists */
3918 bitmap_zero(map, page->objects);
3920 get_map(s, page, map);
3921 for_each_object(p, s, addr, page->objects) {
3922 if (test_bit(slab_index(p, s, addr), map))
3923 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3927 for_each_object(p, s, addr, page->objects)
3928 if (!test_bit(slab_index(p, s, addr), map))
3929 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3934 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3938 validate_slab(s, page, map);
3942 static int validate_slab_node(struct kmem_cache *s,
3943 struct kmem_cache_node *n, unsigned long *map)
3945 unsigned long count = 0;
3947 unsigned long flags;
3949 spin_lock_irqsave(&n->list_lock, flags);
3951 list_for_each_entry(page, &n->partial, lru) {
3952 validate_slab_slab(s, page, map);
3955 if (count != n->nr_partial)
3956 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3957 "counter=%ld\n", s->name, count, n->nr_partial);
3959 if (!(s->flags & SLAB_STORE_USER))
3962 list_for_each_entry(page, &n->full, lru) {
3963 validate_slab_slab(s, page, map);
3966 if (count != atomic_long_read(&n->nr_slabs))
3967 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3968 "counter=%ld\n", s->name, count,
3969 atomic_long_read(&n->nr_slabs));
3972 spin_unlock_irqrestore(&n->list_lock, flags);
3976 static long validate_slab_cache(struct kmem_cache *s)
3979 unsigned long count = 0;
3980 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3981 sizeof(unsigned long), GFP_KERNEL);
3987 for_each_node_state(node, N_NORMAL_MEMORY) {
3988 struct kmem_cache_node *n = get_node(s, node);
3990 count += validate_slab_node(s, n, map);
3996 * Generate lists of code addresses where slabcache objects are allocated
4001 unsigned long count;
4008 DECLARE_BITMAP(cpus, NR_CPUS);
4014 unsigned long count;
4015 struct location *loc;
4018 static void free_loc_track(struct loc_track *t)
4021 free_pages((unsigned long)t->loc,
4022 get_order(sizeof(struct location) * t->max));
4025 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4030 order = get_order(sizeof(struct location) * max);
4032 l = (void *)__get_free_pages(flags, order);
4037 memcpy(l, t->loc, sizeof(struct location) * t->count);
4045 static int add_location(struct loc_track *t, struct kmem_cache *s,
4046 const struct track *track)
4048 long start, end, pos;
4050 unsigned long caddr;
4051 unsigned long age = jiffies - track->when;
4057 pos = start + (end - start + 1) / 2;
4060 * There is nothing at "end". If we end up there
4061 * we need to add something to before end.
4066 caddr = t->loc[pos].addr;
4067 if (track->addr == caddr) {
4073 if (age < l->min_time)
4075 if (age > l->max_time)
4078 if (track->pid < l->min_pid)
4079 l->min_pid = track->pid;
4080 if (track->pid > l->max_pid)
4081 l->max_pid = track->pid;
4083 cpumask_set_cpu(track->cpu,
4084 to_cpumask(l->cpus));
4086 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4090 if (track->addr < caddr)
4097 * Not found. Insert new tracking element.
4099 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4105 (t->count - pos) * sizeof(struct location));
4108 l->addr = track->addr;
4112 l->min_pid = track->pid;
4113 l->max_pid = track->pid;
4114 cpumask_clear(to_cpumask(l->cpus));
4115 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4116 nodes_clear(l->nodes);
4117 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4121 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4122 struct page *page, enum track_item alloc,
4125 void *addr = page_address(page);
4128 bitmap_zero(map, page->objects);
4129 get_map(s, page, map);
4131 for_each_object(p, s, addr, page->objects)
4132 if (!test_bit(slab_index(p, s, addr), map))
4133 add_location(t, s, get_track(s, p, alloc));
4136 static int list_locations(struct kmem_cache *s, char *buf,
4137 enum track_item alloc)
4141 struct loc_track t = { 0, 0, NULL };
4143 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4144 sizeof(unsigned long), GFP_KERNEL);
4146 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4149 return sprintf(buf, "Out of memory\n");
4151 /* Push back cpu slabs */
4154 for_each_node_state(node, N_NORMAL_MEMORY) {
4155 struct kmem_cache_node *n = get_node(s, node);
4156 unsigned long flags;
4159 if (!atomic_long_read(&n->nr_slabs))
4162 spin_lock_irqsave(&n->list_lock, flags);
4163 list_for_each_entry(page, &n->partial, lru)
4164 process_slab(&t, s, page, alloc, map);
4165 list_for_each_entry(page, &n->full, lru)
4166 process_slab(&t, s, page, alloc, map);
4167 spin_unlock_irqrestore(&n->list_lock, flags);
4170 for (i = 0; i < t.count; i++) {
4171 struct location *l = &t.loc[i];
4173 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4175 len += sprintf(buf + len, "%7ld ", l->count);
4178 len += sprintf(buf + len, "%pS", (void *)l->addr);
4180 len += sprintf(buf + len, "<not-available>");
4182 if (l->sum_time != l->min_time) {
4183 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4185 (long)div_u64(l->sum_time, l->count),
4188 len += sprintf(buf + len, " age=%ld",
4191 if (l->min_pid != l->max_pid)
4192 len += sprintf(buf + len, " pid=%ld-%ld",
4193 l->min_pid, l->max_pid);
4195 len += sprintf(buf + len, " pid=%ld",
4198 if (num_online_cpus() > 1 &&
4199 !cpumask_empty(to_cpumask(l->cpus)) &&
4200 len < PAGE_SIZE - 60) {
4201 len += sprintf(buf + len, " cpus=");
4202 len += cpulist_scnprintf(buf + len,
4203 PAGE_SIZE - len - 50,
4204 to_cpumask(l->cpus));
4207 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4208 len < PAGE_SIZE - 60) {
4209 len += sprintf(buf + len, " nodes=");
4210 len += nodelist_scnprintf(buf + len,
4211 PAGE_SIZE - len - 50,
4215 len += sprintf(buf + len, "\n");
4221 len += sprintf(buf, "No data\n");
4226 #ifdef SLUB_RESILIENCY_TEST
4227 static void resiliency_test(void)
4231 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4233 printk(KERN_ERR "SLUB resiliency testing\n");
4234 printk(KERN_ERR "-----------------------\n");
4235 printk(KERN_ERR "A. Corruption after allocation\n");
4237 p = kzalloc(16, GFP_KERNEL);
4239 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4240 " 0x12->0x%p\n\n", p + 16);
4242 validate_slab_cache(kmalloc_caches[4]);
4244 /* Hmmm... The next two are dangerous */
4245 p = kzalloc(32, GFP_KERNEL);
4246 p[32 + sizeof(void *)] = 0x34;
4247 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4248 " 0x34 -> -0x%p\n", p);
4250 "If allocated object is overwritten then not detectable\n\n");
4252 validate_slab_cache(kmalloc_caches[5]);
4253 p = kzalloc(64, GFP_KERNEL);
4254 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4256 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4259 "If allocated object is overwritten then not detectable\n\n");
4260 validate_slab_cache(kmalloc_caches[6]);
4262 printk(KERN_ERR "\nB. Corruption after free\n");
4263 p = kzalloc(128, GFP_KERNEL);
4266 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4267 validate_slab_cache(kmalloc_caches[7]);
4269 p = kzalloc(256, GFP_KERNEL);
4272 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4274 validate_slab_cache(kmalloc_caches[8]);
4276 p = kzalloc(512, GFP_KERNEL);
4279 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4280 validate_slab_cache(kmalloc_caches[9]);
4284 static void resiliency_test(void) {};
4289 enum slab_stat_type {
4290 SL_ALL, /* All slabs */
4291 SL_PARTIAL, /* Only partially allocated slabs */
4292 SL_CPU, /* Only slabs used for cpu caches */
4293 SL_OBJECTS, /* Determine allocated objects not slabs */
4294 SL_TOTAL /* Determine object capacity not slabs */
4297 #define SO_ALL (1 << SL_ALL)
4298 #define SO_PARTIAL (1 << SL_PARTIAL)
4299 #define SO_CPU (1 << SL_CPU)
4300 #define SO_OBJECTS (1 << SL_OBJECTS)
4301 #define SO_TOTAL (1 << SL_TOTAL)
4303 static ssize_t show_slab_objects(struct kmem_cache *s,
4304 char *buf, unsigned long flags)
4306 unsigned long total = 0;
4309 unsigned long *nodes;
4311 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4315 if (flags & SO_CPU) {
4318 for_each_possible_cpu(cpu) {
4319 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4324 page = ACCESS_ONCE(c->page);
4328 node = page_to_nid(page);
4329 if (flags & SO_TOTAL)
4331 else if (flags & SO_OBJECTS)
4339 page = ACCESS_ONCE(c->partial);
4341 node = page_to_nid(page);
4342 if (flags & SO_TOTAL)
4344 else if (flags & SO_OBJECTS)
4354 lock_memory_hotplug();
4355 #ifdef CONFIG_SLUB_DEBUG
4356 if (flags & SO_ALL) {
4357 for_each_node_state(node, N_NORMAL_MEMORY) {
4358 struct kmem_cache_node *n = get_node(s, node);
4360 if (flags & SO_TOTAL)
4361 x = atomic_long_read(&n->total_objects);
4362 else if (flags & SO_OBJECTS)
4363 x = atomic_long_read(&n->total_objects) -
4364 count_partial(n, count_free);
4366 x = atomic_long_read(&n->nr_slabs);
4373 if (flags & SO_PARTIAL) {
4374 for_each_node_state(node, N_NORMAL_MEMORY) {
4375 struct kmem_cache_node *n = get_node(s, node);
4377 if (flags & SO_TOTAL)
4378 x = count_partial(n, count_total);
4379 else if (flags & SO_OBJECTS)
4380 x = count_partial(n, count_inuse);
4387 x = sprintf(buf, "%lu", total);
4389 for_each_node_state(node, N_NORMAL_MEMORY)
4391 x += sprintf(buf + x, " N%d=%lu",
4394 unlock_memory_hotplug();
4396 return x + sprintf(buf + x, "\n");
4399 #ifdef CONFIG_SLUB_DEBUG
4400 static int any_slab_objects(struct kmem_cache *s)
4404 for_each_online_node(node) {
4405 struct kmem_cache_node *n = get_node(s, node);
4410 if (atomic_long_read(&n->total_objects))
4417 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4418 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4420 struct slab_attribute {
4421 struct attribute attr;
4422 ssize_t (*show)(struct kmem_cache *s, char *buf);
4423 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4426 #define SLAB_ATTR_RO(_name) \
4427 static struct slab_attribute _name##_attr = \
4428 __ATTR(_name, 0400, _name##_show, NULL)
4430 #define SLAB_ATTR(_name) \
4431 static struct slab_attribute _name##_attr = \
4432 __ATTR(_name, 0600, _name##_show, _name##_store)
4434 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4436 return sprintf(buf, "%d\n", s->size);
4438 SLAB_ATTR_RO(slab_size);
4440 static ssize_t align_show(struct kmem_cache *s, char *buf)
4442 return sprintf(buf, "%d\n", s->align);
4444 SLAB_ATTR_RO(align);
4446 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4448 return sprintf(buf, "%d\n", s->object_size);
4450 SLAB_ATTR_RO(object_size);
4452 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4454 return sprintf(buf, "%d\n", oo_objects(s->oo));
4456 SLAB_ATTR_RO(objs_per_slab);
4458 static ssize_t order_store(struct kmem_cache *s,
4459 const char *buf, size_t length)
4461 unsigned long order;
4464 err = kstrtoul(buf, 10, &order);
4468 if (order > slub_max_order || order < slub_min_order)
4471 calculate_sizes(s, order);
4475 static ssize_t order_show(struct kmem_cache *s, char *buf)
4477 return sprintf(buf, "%d\n", oo_order(s->oo));
4481 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4483 return sprintf(buf, "%lu\n", s->min_partial);
4486 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4492 err = kstrtoul(buf, 10, &min);
4496 set_min_partial(s, min);
4499 SLAB_ATTR(min_partial);
4501 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4503 return sprintf(buf, "%u\n", s->cpu_partial);
4506 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4509 unsigned long objects;
4512 err = kstrtoul(buf, 10, &objects);
4515 if (objects && !kmem_cache_has_cpu_partial(s))
4518 s->cpu_partial = objects;
4522 SLAB_ATTR(cpu_partial);
4524 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4528 return sprintf(buf, "%pS\n", s->ctor);
4532 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4534 return sprintf(buf, "%d\n", s->refcount - 1);
4536 SLAB_ATTR_RO(aliases);
4538 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4540 return show_slab_objects(s, buf, SO_PARTIAL);
4542 SLAB_ATTR_RO(partial);
4544 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4546 return show_slab_objects(s, buf, SO_CPU);
4548 SLAB_ATTR_RO(cpu_slabs);
4550 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4552 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4554 SLAB_ATTR_RO(objects);
4556 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4558 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4560 SLAB_ATTR_RO(objects_partial);
4562 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4569 for_each_online_cpu(cpu) {
4570 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4573 pages += page->pages;
4574 objects += page->pobjects;
4578 len = sprintf(buf, "%d(%d)", objects, pages);
4581 for_each_online_cpu(cpu) {
4582 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4584 if (page && len < PAGE_SIZE - 20)
4585 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4586 page->pobjects, page->pages);
4589 return len + sprintf(buf + len, "\n");
4591 SLAB_ATTR_RO(slabs_cpu_partial);
4593 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4595 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4598 static ssize_t reclaim_account_store(struct kmem_cache *s,
4599 const char *buf, size_t length)
4601 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4603 s->flags |= SLAB_RECLAIM_ACCOUNT;
4606 SLAB_ATTR(reclaim_account);
4608 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4610 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4612 SLAB_ATTR_RO(hwcache_align);
4614 #ifdef CONFIG_ZONE_DMA
4615 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4617 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4619 SLAB_ATTR_RO(cache_dma);
4622 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4624 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4626 SLAB_ATTR_RO(destroy_by_rcu);
4628 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4630 return sprintf(buf, "%d\n", s->reserved);
4632 SLAB_ATTR_RO(reserved);
4634 #ifdef CONFIG_SLUB_DEBUG
4635 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4637 return show_slab_objects(s, buf, SO_ALL);
4639 SLAB_ATTR_RO(slabs);
4641 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4643 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4645 SLAB_ATTR_RO(total_objects);
4647 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4649 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4652 static ssize_t sanity_checks_store(struct kmem_cache *s,
4653 const char *buf, size_t length)
4655 s->flags &= ~SLAB_DEBUG_FREE;
4656 if (buf[0] == '1') {
4657 s->flags &= ~__CMPXCHG_DOUBLE;
4658 s->flags |= SLAB_DEBUG_FREE;
4662 SLAB_ATTR(sanity_checks);
4664 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4666 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4669 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4672 s->flags &= ~SLAB_TRACE;
4673 if (buf[0] == '1') {
4674 s->flags &= ~__CMPXCHG_DOUBLE;
4675 s->flags |= SLAB_TRACE;
4681 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4683 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4686 static ssize_t red_zone_store(struct kmem_cache *s,
4687 const char *buf, size_t length)
4689 if (any_slab_objects(s))
4692 s->flags &= ~SLAB_RED_ZONE;
4693 if (buf[0] == '1') {
4694 s->flags &= ~__CMPXCHG_DOUBLE;
4695 s->flags |= SLAB_RED_ZONE;
4697 calculate_sizes(s, -1);
4700 SLAB_ATTR(red_zone);
4702 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4704 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4707 static ssize_t poison_store(struct kmem_cache *s,
4708 const char *buf, size_t length)
4710 if (any_slab_objects(s))
4713 s->flags &= ~SLAB_POISON;
4714 if (buf[0] == '1') {
4715 s->flags &= ~__CMPXCHG_DOUBLE;
4716 s->flags |= SLAB_POISON;
4718 calculate_sizes(s, -1);
4723 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4725 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4728 static ssize_t store_user_store(struct kmem_cache *s,
4729 const char *buf, size_t length)
4731 if (any_slab_objects(s))
4734 s->flags &= ~SLAB_STORE_USER;
4735 if (buf[0] == '1') {
4736 s->flags &= ~__CMPXCHG_DOUBLE;
4737 s->flags |= SLAB_STORE_USER;
4739 calculate_sizes(s, -1);
4742 SLAB_ATTR(store_user);
4744 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4749 static ssize_t validate_store(struct kmem_cache *s,
4750 const char *buf, size_t length)
4754 if (buf[0] == '1') {
4755 ret = validate_slab_cache(s);
4761 SLAB_ATTR(validate);
4763 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4765 if (!(s->flags & SLAB_STORE_USER))
4767 return list_locations(s, buf, TRACK_ALLOC);
4769 SLAB_ATTR_RO(alloc_calls);
4771 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4773 if (!(s->flags & SLAB_STORE_USER))
4775 return list_locations(s, buf, TRACK_FREE);
4777 SLAB_ATTR_RO(free_calls);
4778 #endif /* CONFIG_SLUB_DEBUG */
4780 #ifdef CONFIG_FAILSLAB
4781 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4783 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4786 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4789 s->flags &= ~SLAB_FAILSLAB;
4791 s->flags |= SLAB_FAILSLAB;
4794 SLAB_ATTR(failslab);
4797 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4802 static ssize_t shrink_store(struct kmem_cache *s,
4803 const char *buf, size_t length)
4805 if (buf[0] == '1') {
4806 int rc = kmem_cache_shrink(s);
4817 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4819 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4822 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4823 const char *buf, size_t length)
4825 unsigned long ratio;
4828 err = kstrtoul(buf, 10, &ratio);
4833 s->remote_node_defrag_ratio = ratio * 10;
4837 SLAB_ATTR(remote_node_defrag_ratio);
4840 #ifdef CONFIG_SLUB_STATS
4841 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4843 unsigned long sum = 0;
4846 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4851 for_each_online_cpu(cpu) {
4852 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4858 len = sprintf(buf, "%lu", sum);
4861 for_each_online_cpu(cpu) {
4862 if (data[cpu] && len < PAGE_SIZE - 20)
4863 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4867 return len + sprintf(buf + len, "\n");
4870 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4874 for_each_online_cpu(cpu)
4875 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4878 #define STAT_ATTR(si, text) \
4879 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4881 return show_stat(s, buf, si); \
4883 static ssize_t text##_store(struct kmem_cache *s, \
4884 const char *buf, size_t length) \
4886 if (buf[0] != '0') \
4888 clear_stat(s, si); \
4893 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4894 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4895 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4896 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4897 STAT_ATTR(FREE_FROZEN, free_frozen);
4898 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4899 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4900 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4901 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4902 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4903 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4904 STAT_ATTR(FREE_SLAB, free_slab);
4905 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4906 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4907 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4908 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4909 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4910 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4911 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4912 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4913 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4914 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4915 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4916 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4917 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4918 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4921 static struct attribute *slab_attrs[] = {
4922 &slab_size_attr.attr,
4923 &object_size_attr.attr,
4924 &objs_per_slab_attr.attr,
4926 &min_partial_attr.attr,
4927 &cpu_partial_attr.attr,
4929 &objects_partial_attr.attr,
4931 &cpu_slabs_attr.attr,
4935 &hwcache_align_attr.attr,
4936 &reclaim_account_attr.attr,
4937 &destroy_by_rcu_attr.attr,
4939 &reserved_attr.attr,
4940 &slabs_cpu_partial_attr.attr,
4941 #ifdef CONFIG_SLUB_DEBUG
4942 &total_objects_attr.attr,
4944 &sanity_checks_attr.attr,
4946 &red_zone_attr.attr,
4948 &store_user_attr.attr,
4949 &validate_attr.attr,
4950 &alloc_calls_attr.attr,
4951 &free_calls_attr.attr,
4953 #ifdef CONFIG_ZONE_DMA
4954 &cache_dma_attr.attr,
4957 &remote_node_defrag_ratio_attr.attr,
4959 #ifdef CONFIG_SLUB_STATS
4960 &alloc_fastpath_attr.attr,
4961 &alloc_slowpath_attr.attr,
4962 &free_fastpath_attr.attr,
4963 &free_slowpath_attr.attr,
4964 &free_frozen_attr.attr,
4965 &free_add_partial_attr.attr,
4966 &free_remove_partial_attr.attr,
4967 &alloc_from_partial_attr.attr,
4968 &alloc_slab_attr.attr,
4969 &alloc_refill_attr.attr,
4970 &alloc_node_mismatch_attr.attr,
4971 &free_slab_attr.attr,
4972 &cpuslab_flush_attr.attr,
4973 &deactivate_full_attr.attr,
4974 &deactivate_empty_attr.attr,
4975 &deactivate_to_head_attr.attr,
4976 &deactivate_to_tail_attr.attr,
4977 &deactivate_remote_frees_attr.attr,
4978 &deactivate_bypass_attr.attr,
4979 &order_fallback_attr.attr,
4980 &cmpxchg_double_fail_attr.attr,
4981 &cmpxchg_double_cpu_fail_attr.attr,
4982 &cpu_partial_alloc_attr.attr,
4983 &cpu_partial_free_attr.attr,
4984 &cpu_partial_node_attr.attr,
4985 &cpu_partial_drain_attr.attr,
4987 #ifdef CONFIG_FAILSLAB
4988 &failslab_attr.attr,
4994 static struct attribute_group slab_attr_group = {
4995 .attrs = slab_attrs,
4998 static ssize_t slab_attr_show(struct kobject *kobj,
4999 struct attribute *attr,
5002 struct slab_attribute *attribute;
5003 struct kmem_cache *s;
5006 attribute = to_slab_attr(attr);
5009 if (!attribute->show)
5012 err = attribute->show(s, buf);
5017 static ssize_t slab_attr_store(struct kobject *kobj,
5018 struct attribute *attr,
5019 const char *buf, size_t len)
5021 struct slab_attribute *attribute;
5022 struct kmem_cache *s;
5025 attribute = to_slab_attr(attr);
5028 if (!attribute->store)
5031 err = attribute->store(s, buf, len);
5032 #ifdef CONFIG_MEMCG_KMEM
5033 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5036 mutex_lock(&slab_mutex);
5037 if (s->max_attr_size < len)
5038 s->max_attr_size = len;
5041 * This is a best effort propagation, so this function's return
5042 * value will be determined by the parent cache only. This is
5043 * basically because not all attributes will have a well
5044 * defined semantics for rollbacks - most of the actions will
5045 * have permanent effects.
5047 * Returning the error value of any of the children that fail
5048 * is not 100 % defined, in the sense that users seeing the
5049 * error code won't be able to know anything about the state of
5052 * Only returning the error code for the parent cache at least
5053 * has well defined semantics. The cache being written to
5054 * directly either failed or succeeded, in which case we loop
5055 * through the descendants with best-effort propagation.
5057 for_each_memcg_cache_index(i) {
5058 struct kmem_cache *c = cache_from_memcg_idx(s, i);
5060 attribute->store(c, buf, len);
5062 mutex_unlock(&slab_mutex);
5068 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5070 #ifdef CONFIG_MEMCG_KMEM
5072 char *buffer = NULL;
5074 if (!is_root_cache(s))
5078 * This mean this cache had no attribute written. Therefore, no point
5079 * in copying default values around
5081 if (!s->max_attr_size)
5084 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5087 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5089 if (!attr || !attr->store || !attr->show)
5093 * It is really bad that we have to allocate here, so we will
5094 * do it only as a fallback. If we actually allocate, though,
5095 * we can just use the allocated buffer until the end.
5097 * Most of the slub attributes will tend to be very small in
5098 * size, but sysfs allows buffers up to a page, so they can
5099 * theoretically happen.
5103 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5106 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5107 if (WARN_ON(!buffer))
5112 attr->show(s->memcg_params->root_cache, buf);
5113 attr->store(s, buf, strlen(buf));
5117 free_page((unsigned long)buffer);
5121 static const struct sysfs_ops slab_sysfs_ops = {
5122 .show = slab_attr_show,
5123 .store = slab_attr_store,
5126 static struct kobj_type slab_ktype = {
5127 .sysfs_ops = &slab_sysfs_ops,
5130 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5132 struct kobj_type *ktype = get_ktype(kobj);
5134 if (ktype == &slab_ktype)
5139 static const struct kset_uevent_ops slab_uevent_ops = {
5140 .filter = uevent_filter,
5143 static struct kset *slab_kset;
5145 static inline struct kset *cache_kset(struct kmem_cache *s)
5147 #ifdef CONFIG_MEMCG_KMEM
5148 if (!is_root_cache(s))
5149 return s->memcg_params->root_cache->memcg_kset;
5154 #define ID_STR_LENGTH 64
5156 /* Create a unique string id for a slab cache:
5158 * Format :[flags-]size
5160 static char *create_unique_id(struct kmem_cache *s)
5162 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5169 * First flags affecting slabcache operations. We will only
5170 * get here for aliasable slabs so we do not need to support
5171 * too many flags. The flags here must cover all flags that
5172 * are matched during merging to guarantee that the id is
5175 if (s->flags & SLAB_CACHE_DMA)
5177 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5179 if (s->flags & SLAB_DEBUG_FREE)
5181 if (!(s->flags & SLAB_NOTRACK))
5185 p += sprintf(p, "%07d", s->size);
5187 #ifdef CONFIG_MEMCG_KMEM
5188 if (!is_root_cache(s))
5189 p += sprintf(p, "-%08d",
5190 memcg_cache_id(s->memcg_params->memcg));
5193 BUG_ON(p > name + ID_STR_LENGTH - 1);
5197 static int sysfs_slab_add(struct kmem_cache *s)
5201 int unmergeable = slab_unmergeable(s);
5205 * Slabcache can never be merged so we can use the name proper.
5206 * This is typically the case for debug situations. In that
5207 * case we can catch duplicate names easily.
5209 sysfs_remove_link(&slab_kset->kobj, s->name);
5213 * Create a unique name for the slab as a target
5216 name = create_unique_id(s);
5219 s->kobj.kset = cache_kset(s);
5220 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5224 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5228 #ifdef CONFIG_MEMCG_KMEM
5229 if (is_root_cache(s)) {
5230 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5231 if (!s->memcg_kset) {
5238 kobject_uevent(&s->kobj, KOBJ_ADD);
5240 /* Setup first alias */
5241 sysfs_slab_alias(s, s->name);
5248 kobject_del(&s->kobj);
5250 kobject_put(&s->kobj);
5254 static void sysfs_slab_remove(struct kmem_cache *s)
5256 if (slab_state < FULL)
5258 * Sysfs has not been setup yet so no need to remove the
5263 #ifdef CONFIG_MEMCG_KMEM
5264 kset_unregister(s->memcg_kset);
5266 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5267 kobject_del(&s->kobj);
5268 kobject_put(&s->kobj);
5272 * Need to buffer aliases during bootup until sysfs becomes
5273 * available lest we lose that information.
5275 struct saved_alias {
5276 struct kmem_cache *s;
5278 struct saved_alias *next;
5281 static struct saved_alias *alias_list;
5283 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5285 struct saved_alias *al;
5287 if (slab_state == FULL) {
5289 * If we have a leftover link then remove it.
5291 sysfs_remove_link(&slab_kset->kobj, name);
5292 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5295 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5301 al->next = alias_list;
5306 static int __init slab_sysfs_init(void)
5308 struct kmem_cache *s;
5311 mutex_lock(&slab_mutex);
5313 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5315 mutex_unlock(&slab_mutex);
5316 printk(KERN_ERR "Cannot register slab subsystem.\n");
5322 list_for_each_entry(s, &slab_caches, list) {
5323 err = sysfs_slab_add(s);
5325 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5326 " to sysfs\n", s->name);
5329 while (alias_list) {
5330 struct saved_alias *al = alias_list;
5332 alias_list = alias_list->next;
5333 err = sysfs_slab_alias(al->s, al->name);
5335 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5336 " %s to sysfs\n", al->name);
5340 mutex_unlock(&slab_mutex);
5345 __initcall(slab_sysfs_init);
5346 #endif /* CONFIG_SYSFS */
5349 * The /proc/slabinfo ABI
5351 #ifdef CONFIG_SLABINFO
5352 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5354 unsigned long nr_slabs = 0;
5355 unsigned long nr_objs = 0;
5356 unsigned long nr_free = 0;
5359 for_each_online_node(node) {
5360 struct kmem_cache_node *n = get_node(s, node);
5365 nr_slabs += node_nr_slabs(n);
5366 nr_objs += node_nr_objs(n);
5367 nr_free += count_partial(n, count_free);
5370 sinfo->active_objs = nr_objs - nr_free;
5371 sinfo->num_objs = nr_objs;
5372 sinfo->active_slabs = nr_slabs;
5373 sinfo->num_slabs = nr_slabs;
5374 sinfo->objects_per_slab = oo_objects(s->oo);
5375 sinfo->cache_order = oo_order(s->oo);
5378 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5382 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5383 size_t count, loff_t *ppos)
5387 #endif /* CONFIG_SLABINFO */