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/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
137 * Issues still to be resolved:
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
141 * - Variable sizing of the per node arrays
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
154 #define MIN_PARTIAL 5
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
161 #define MAX_PARTIAL 10
163 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
171 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
209 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
212 static inline void stat(const struct kmem_cache *s, enum stat_item si)
214 #ifdef CONFIG_SLUB_STATS
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
219 raw_cpu_inc(s->cpu_slab->stat[si]);
223 /********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
227 /* Verify that a pointer has an address that is valid within a slab page */
228 static inline int check_valid_pointer(struct kmem_cache *s,
229 struct page *page, const void *object)
236 base = page_address(page);
237 if (object < base || object >= base + page->objects * s->size ||
238 (object - base) % s->size) {
245 static inline void *get_freepointer(struct kmem_cache *s, void *object)
247 return *(void **)(object + s->offset);
250 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
252 prefetch(object + s->offset);
255 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
259 #ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
262 p = get_freepointer(s, object);
267 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
269 *(void **)(object + s->offset) = fp;
272 /* Loop over all objects in a slab */
273 #define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
277 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline size_t slab_ksize(const struct kmem_cache *s)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order, unsigned long size, int reserved)
313 return ((PAGE_SIZE << order) - reserved) / size;
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
326 static inline int oo_order(struct kmem_cache_order_objects x)
328 return x.x >> OO_SHIFT;
331 static inline int oo_objects(struct kmem_cache_order_objects x)
333 return x.x & OO_MASK;
337 * Per slab locking using the pagelock
339 static __always_inline void slab_lock(struct page *page)
341 bit_spin_lock(PG_locked, &page->flags);
344 static __always_inline void slab_unlock(struct page *page)
346 __bit_spin_unlock(PG_locked, &page->flags);
349 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
352 tmp.counters = counters_new;
354 * page->counters can cover frozen/inuse/objects as well
355 * as page->_count. If we assign to ->counters directly
356 * we run the risk of losing updates to page->_count, so
357 * be careful and only assign to the fields we need.
359 page->frozen = tmp.frozen;
360 page->inuse = tmp.inuse;
361 page->objects = tmp.objects;
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 set_page_slub_counters(page, counters_new);
393 stat(s, CMPXCHG_DOUBLE_FAIL);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
402 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
419 local_irq_save(flags);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 set_page_slub_counters(page, counters_new);
426 local_irq_restore(flags);
430 local_irq_restore(flags);
434 stat(s, CMPXCHG_DOUBLE_FAIL);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453 void *addr = page_address(page);
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
462 #if defined(CONFIG_SLUB_DEBUG_ON)
463 static int slub_debug = DEBUG_DEFAULT_FLAGS;
464 #elif defined(CONFIG_KASAN)
465 static int slub_debug = SLAB_STORE_USER;
467 static int slub_debug;
470 static char *slub_debug_slabs;
471 static int disable_higher_order_debug;
474 * slub is about to manipulate internal object metadata. This memory lies
475 * outside the range of the allocated object, so accessing it would normally
476 * be reported by kasan as a bounds error. metadata_access_enable() is used
477 * to tell kasan that these accesses are OK.
479 static inline void metadata_access_enable(void)
481 kasan_disable_current();
484 static inline void metadata_access_disable(void)
486 kasan_enable_current();
492 static void print_section(char *text, u8 *addr, unsigned int length)
494 metadata_access_enable();
495 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
497 metadata_access_disable();
500 static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
506 p = object + s->offset + sizeof(void *);
508 p = object + s->inuse;
513 static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
516 struct track *p = get_track(s, object, alloc);
519 #ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
527 metadata_access_enable();
528 save_stack_trace(&trace);
529 metadata_access_disable();
531 /* See rant in lockdep.c */
532 if (trace.nr_entries != 0 &&
533 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
536 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
540 p->cpu = smp_processor_id();
541 p->pid = current->pid;
544 memset(p, 0, sizeof(struct track));
547 static void init_tracking(struct kmem_cache *s, void *object)
549 if (!(s->flags & SLAB_STORE_USER))
552 set_track(s, object, TRACK_FREE, 0UL);
553 set_track(s, object, TRACK_ALLOC, 0UL);
556 static void print_track(const char *s, struct track *t)
561 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
562 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
563 #ifdef CONFIG_STACKTRACE
566 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
568 pr_err("\t%pS\n", (void *)t->addrs[i]);
575 static void print_tracking(struct kmem_cache *s, void *object)
577 if (!(s->flags & SLAB_STORE_USER))
580 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
581 print_track("Freed", get_track(s, object, TRACK_FREE));
584 static void print_page_info(struct page *page)
586 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
587 page, page->objects, page->inuse, page->freelist, page->flags);
591 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
593 struct va_format vaf;
599 pr_err("=============================================================================\n");
600 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
601 pr_err("-----------------------------------------------------------------------------\n\n");
603 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
607 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
609 struct va_format vaf;
615 pr_err("FIX %s: %pV\n", s->name, &vaf);
619 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
621 unsigned int off; /* Offset of last byte */
622 u8 *addr = page_address(page);
624 print_tracking(s, p);
626 print_page_info(page);
628 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
629 p, p - addr, get_freepointer(s, p));
632 print_section("Bytes b4 ", p - 16, 16);
634 print_section("Object ", p, min_t(unsigned long, s->object_size,
636 if (s->flags & SLAB_RED_ZONE)
637 print_section("Redzone ", p + s->object_size,
638 s->inuse - s->object_size);
641 off = s->offset + sizeof(void *);
645 if (s->flags & SLAB_STORE_USER)
646 off += 2 * sizeof(struct track);
649 /* Beginning of the filler is the free pointer */
650 print_section("Padding ", p + off, s->size - off);
655 void object_err(struct kmem_cache *s, struct page *page,
656 u8 *object, char *reason)
658 slab_bug(s, "%s", reason);
659 print_trailer(s, page, object);
662 static void slab_err(struct kmem_cache *s, struct page *page,
663 const char *fmt, ...)
669 vsnprintf(buf, sizeof(buf), fmt, args);
671 slab_bug(s, "%s", buf);
672 print_page_info(page);
676 static void init_object(struct kmem_cache *s, void *object, u8 val)
680 if (s->flags & __OBJECT_POISON) {
681 memset(p, POISON_FREE, s->object_size - 1);
682 p[s->object_size - 1] = POISON_END;
685 if (s->flags & SLAB_RED_ZONE)
686 memset(p + s->object_size, val, s->inuse - s->object_size);
689 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
690 void *from, void *to)
692 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
693 memset(from, data, to - from);
696 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
697 u8 *object, char *what,
698 u8 *start, unsigned int value, unsigned int bytes)
703 metadata_access_enable();
704 fault = memchr_inv(start, value, bytes);
705 metadata_access_disable();
710 while (end > fault && end[-1] == value)
713 slab_bug(s, "%s overwritten", what);
714 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
715 fault, end - 1, fault[0], value);
716 print_trailer(s, page, object);
718 restore_bytes(s, what, value, fault, end);
726 * Bytes of the object to be managed.
727 * If the freepointer may overlay the object then the free
728 * pointer is the first word of the object.
730 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
733 * object + s->object_size
734 * Padding to reach word boundary. This is also used for Redzoning.
735 * Padding is extended by another word if Redzoning is enabled and
736 * object_size == inuse.
738 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
739 * 0xcc (RED_ACTIVE) for objects in use.
742 * Meta data starts here.
744 * A. Free pointer (if we cannot overwrite object on free)
745 * B. Tracking data for SLAB_STORE_USER
746 * C. Padding to reach required alignment boundary or at mininum
747 * one word if debugging is on to be able to detect writes
748 * before the word boundary.
750 * Padding is done using 0x5a (POISON_INUSE)
753 * Nothing is used beyond s->size.
755 * If slabcaches are merged then the object_size and inuse boundaries are mostly
756 * ignored. And therefore no slab options that rely on these boundaries
757 * may be used with merged slabcaches.
760 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
762 unsigned long off = s->inuse; /* The end of info */
765 /* Freepointer is placed after the object. */
766 off += sizeof(void *);
768 if (s->flags & SLAB_STORE_USER)
769 /* We also have user information there */
770 off += 2 * sizeof(struct track);
775 return check_bytes_and_report(s, page, p, "Object padding",
776 p + off, POISON_INUSE, s->size - off);
779 /* Check the pad bytes at the end of a slab page */
780 static int slab_pad_check(struct kmem_cache *s, struct page *page)
788 if (!(s->flags & SLAB_POISON))
791 start = page_address(page);
792 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
793 end = start + length;
794 remainder = length % s->size;
798 metadata_access_enable();
799 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
800 metadata_access_disable();
803 while (end > fault && end[-1] == POISON_INUSE)
806 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
807 print_section("Padding ", end - remainder, remainder);
809 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
813 static int check_object(struct kmem_cache *s, struct page *page,
814 void *object, u8 val)
817 u8 *endobject = object + s->object_size;
819 if (s->flags & SLAB_RED_ZONE) {
820 if (!check_bytes_and_report(s, page, object, "Redzone",
821 endobject, val, s->inuse - s->object_size))
824 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
825 check_bytes_and_report(s, page, p, "Alignment padding",
826 endobject, POISON_INUSE,
827 s->inuse - s->object_size);
831 if (s->flags & SLAB_POISON) {
832 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
833 (!check_bytes_and_report(s, page, p, "Poison", p,
834 POISON_FREE, s->object_size - 1) ||
835 !check_bytes_and_report(s, page, p, "Poison",
836 p + s->object_size - 1, POISON_END, 1)))
839 * check_pad_bytes cleans up on its own.
841 check_pad_bytes(s, page, p);
844 if (!s->offset && val == SLUB_RED_ACTIVE)
846 * Object and freepointer overlap. Cannot check
847 * freepointer while object is allocated.
851 /* Check free pointer validity */
852 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
853 object_err(s, page, p, "Freepointer corrupt");
855 * No choice but to zap it and thus lose the remainder
856 * of the free objects in this slab. May cause
857 * another error because the object count is now wrong.
859 set_freepointer(s, p, NULL);
865 static int check_slab(struct kmem_cache *s, struct page *page)
869 VM_BUG_ON(!irqs_disabled());
871 if (!PageSlab(page)) {
872 slab_err(s, page, "Not a valid slab page");
876 maxobj = order_objects(compound_order(page), s->size, s->reserved);
877 if (page->objects > maxobj) {
878 slab_err(s, page, "objects %u > max %u",
879 page->objects, maxobj);
882 if (page->inuse > page->objects) {
883 slab_err(s, page, "inuse %u > max %u",
884 page->inuse, page->objects);
887 /* Slab_pad_check fixes things up after itself */
888 slab_pad_check(s, page);
893 * Determine if a certain object on a page is on the freelist. Must hold the
894 * slab lock to guarantee that the chains are in a consistent state.
896 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
904 while (fp && nr <= page->objects) {
907 if (!check_valid_pointer(s, page, fp)) {
909 object_err(s, page, object,
910 "Freechain corrupt");
911 set_freepointer(s, object, NULL);
913 slab_err(s, page, "Freepointer corrupt");
914 page->freelist = NULL;
915 page->inuse = page->objects;
916 slab_fix(s, "Freelist cleared");
922 fp = get_freepointer(s, object);
926 max_objects = order_objects(compound_order(page), s->size, s->reserved);
927 if (max_objects > MAX_OBJS_PER_PAGE)
928 max_objects = MAX_OBJS_PER_PAGE;
930 if (page->objects != max_objects) {
931 slab_err(s, page, "Wrong number of objects. Found %d but "
932 "should be %d", page->objects, max_objects);
933 page->objects = max_objects;
934 slab_fix(s, "Number of objects adjusted.");
936 if (page->inuse != page->objects - nr) {
937 slab_err(s, page, "Wrong object count. Counter is %d but "
938 "counted were %d", page->inuse, page->objects - nr);
939 page->inuse = page->objects - nr;
940 slab_fix(s, "Object count adjusted.");
942 return search == NULL;
945 static void trace(struct kmem_cache *s, struct page *page, void *object,
948 if (s->flags & SLAB_TRACE) {
949 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
951 alloc ? "alloc" : "free",
956 print_section("Object ", (void *)object,
964 * Tracking of fully allocated slabs for debugging purposes.
966 static void add_full(struct kmem_cache *s,
967 struct kmem_cache_node *n, struct page *page)
969 if (!(s->flags & SLAB_STORE_USER))
972 lockdep_assert_held(&n->list_lock);
973 list_add(&page->lru, &n->full);
976 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
978 if (!(s->flags & SLAB_STORE_USER))
981 lockdep_assert_held(&n->list_lock);
982 list_del(&page->lru);
985 /* Tracking of the number of slabs for debugging purposes */
986 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
988 struct kmem_cache_node *n = get_node(s, node);
990 return atomic_long_read(&n->nr_slabs);
993 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
995 return atomic_long_read(&n->nr_slabs);
998 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1000 struct kmem_cache_node *n = get_node(s, node);
1003 * May be called early in order to allocate a slab for the
1004 * kmem_cache_node structure. Solve the chicken-egg
1005 * dilemma by deferring the increment of the count during
1006 * bootstrap (see early_kmem_cache_node_alloc).
1009 atomic_long_inc(&n->nr_slabs);
1010 atomic_long_add(objects, &n->total_objects);
1013 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1015 struct kmem_cache_node *n = get_node(s, node);
1017 atomic_long_dec(&n->nr_slabs);
1018 atomic_long_sub(objects, &n->total_objects);
1021 /* Object debug checks for alloc/free paths */
1022 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1025 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1028 init_object(s, object, SLUB_RED_INACTIVE);
1029 init_tracking(s, object);
1032 static noinline int alloc_debug_processing(struct kmem_cache *s,
1034 void *object, unsigned long addr)
1036 if (!check_slab(s, page))
1039 if (!check_valid_pointer(s, page, object)) {
1040 object_err(s, page, object, "Freelist Pointer check fails");
1044 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1047 /* Success perform special debug activities for allocs */
1048 if (s->flags & SLAB_STORE_USER)
1049 set_track(s, object, TRACK_ALLOC, addr);
1050 trace(s, page, object, 1);
1051 init_object(s, object, SLUB_RED_ACTIVE);
1055 if (PageSlab(page)) {
1057 * If this is a slab page then lets do the best we can
1058 * to avoid issues in the future. Marking all objects
1059 * as used avoids touching the remaining objects.
1061 slab_fix(s, "Marking all objects used");
1062 page->inuse = page->objects;
1063 page->freelist = NULL;
1068 static noinline struct kmem_cache_node *free_debug_processing(
1069 struct kmem_cache *s, struct page *page, void *object,
1070 unsigned long addr, unsigned long *flags)
1072 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1074 spin_lock_irqsave(&n->list_lock, *flags);
1077 if (!check_slab(s, page))
1080 if (!check_valid_pointer(s, page, object)) {
1081 slab_err(s, page, "Invalid object pointer 0x%p", object);
1085 if (on_freelist(s, page, object)) {
1086 object_err(s, page, object, "Object already free");
1090 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1093 if (unlikely(s != page->slab_cache)) {
1094 if (!PageSlab(page)) {
1095 slab_err(s, page, "Attempt to free object(0x%p) "
1096 "outside of slab", object);
1097 } else if (!page->slab_cache) {
1098 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1102 object_err(s, page, object,
1103 "page slab pointer corrupt.");
1107 if (s->flags & SLAB_STORE_USER)
1108 set_track(s, object, TRACK_FREE, addr);
1109 trace(s, page, object, 0);
1110 init_object(s, object, SLUB_RED_INACTIVE);
1114 * Keep node_lock to preserve integrity
1115 * until the object is actually freed
1121 spin_unlock_irqrestore(&n->list_lock, *flags);
1122 slab_fix(s, "Object at 0x%p not freed", object);
1126 static int __init setup_slub_debug(char *str)
1128 slub_debug = DEBUG_DEFAULT_FLAGS;
1129 if (*str++ != '=' || !*str)
1131 * No options specified. Switch on full debugging.
1137 * No options but restriction on slabs. This means full
1138 * debugging for slabs matching a pattern.
1145 * Switch off all debugging measures.
1150 * Determine which debug features should be switched on
1152 for (; *str && *str != ','; str++) {
1153 switch (tolower(*str)) {
1155 slub_debug |= SLAB_DEBUG_FREE;
1158 slub_debug |= SLAB_RED_ZONE;
1161 slub_debug |= SLAB_POISON;
1164 slub_debug |= SLAB_STORE_USER;
1167 slub_debug |= SLAB_TRACE;
1170 slub_debug |= SLAB_FAILSLAB;
1174 * Avoid enabling debugging on caches if its minimum
1175 * order would increase as a result.
1177 disable_higher_order_debug = 1;
1180 pr_err("slub_debug option '%c' unknown. skipped\n",
1187 slub_debug_slabs = str + 1;
1192 __setup("slub_debug", setup_slub_debug);
1194 unsigned long kmem_cache_flags(unsigned long object_size,
1195 unsigned long flags, const char *name,
1196 void (*ctor)(void *))
1199 * Enable debugging if selected on the kernel commandline.
1201 if (slub_debug && (!slub_debug_slabs || (name &&
1202 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1203 flags |= slub_debug;
1207 #else /* !CONFIG_SLUB_DEBUG */
1208 static inline void setup_object_debug(struct kmem_cache *s,
1209 struct page *page, void *object) {}
1211 static inline int alloc_debug_processing(struct kmem_cache *s,
1212 struct page *page, void *object, unsigned long addr) { return 0; }
1214 static inline struct kmem_cache_node *free_debug_processing(
1215 struct kmem_cache *s, struct page *page, void *object,
1216 unsigned long addr, unsigned long *flags) { return NULL; }
1218 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1220 static inline int check_object(struct kmem_cache *s, struct page *page,
1221 void *object, u8 val) { return 1; }
1222 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1223 struct page *page) {}
1224 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1225 struct page *page) {}
1226 unsigned long kmem_cache_flags(unsigned long object_size,
1227 unsigned long flags, const char *name,
1228 void (*ctor)(void *))
1232 #define slub_debug 0
1234 #define disable_higher_order_debug 0
1236 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1238 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1240 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1242 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1245 #endif /* CONFIG_SLUB_DEBUG */
1248 * Hooks for other subsystems that check memory allocations. In a typical
1249 * production configuration these hooks all should produce no code at all.
1251 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1253 kmemleak_alloc(ptr, size, 1, flags);
1254 kasan_kmalloc_large(ptr, size);
1257 static inline void kfree_hook(const void *x)
1260 kasan_kfree_large(x);
1263 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1266 flags &= gfp_allowed_mask;
1267 lockdep_trace_alloc(flags);
1268 might_sleep_if(gfpflags_allow_blocking(flags));
1270 if (should_failslab(s->object_size, flags, s->flags))
1273 return memcg_kmem_get_cache(s, flags);
1276 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1277 gfp_t flags, void *object)
1279 flags &= gfp_allowed_mask;
1280 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1281 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1282 memcg_kmem_put_cache(s);
1283 kasan_slab_alloc(s, object);
1286 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1288 kmemleak_free_recursive(x, s->flags);
1291 * Trouble is that we may no longer disable interrupts in the fast path
1292 * So in order to make the debug calls that expect irqs to be
1293 * disabled we need to disable interrupts temporarily.
1295 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1297 unsigned long flags;
1299 local_irq_save(flags);
1300 kmemcheck_slab_free(s, x, s->object_size);
1301 debug_check_no_locks_freed(x, s->object_size);
1302 local_irq_restore(flags);
1305 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1306 debug_check_no_obj_freed(x, s->object_size);
1308 kasan_slab_free(s, x);
1311 static void setup_object(struct kmem_cache *s, struct page *page,
1314 setup_object_debug(s, page, object);
1315 if (unlikely(s->ctor)) {
1316 kasan_unpoison_object_data(s, object);
1318 kasan_poison_object_data(s, object);
1323 * Slab allocation and freeing
1325 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1326 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1329 int order = oo_order(oo);
1331 flags |= __GFP_NOTRACK;
1333 if (node == NUMA_NO_NODE)
1334 page = alloc_pages(flags, order);
1336 page = __alloc_pages_node(node, flags, order);
1338 if (page && memcg_charge_slab(page, flags, order, s)) {
1339 __free_pages(page, order);
1346 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1349 struct kmem_cache_order_objects oo = s->oo;
1354 flags &= gfp_allowed_mask;
1356 if (gfpflags_allow_blocking(flags))
1359 flags |= s->allocflags;
1362 * Let the initial higher-order allocation fail under memory pressure
1363 * so we fall-back to the minimum order allocation.
1365 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1366 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1367 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1369 page = alloc_slab_page(s, alloc_gfp, node, oo);
1370 if (unlikely(!page)) {
1374 * Allocation may have failed due to fragmentation.
1375 * Try a lower order alloc if possible
1377 page = alloc_slab_page(s, alloc_gfp, node, oo);
1378 if (unlikely(!page))
1380 stat(s, ORDER_FALLBACK);
1383 if (kmemcheck_enabled &&
1384 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1385 int pages = 1 << oo_order(oo);
1387 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1390 * Objects from caches that have a constructor don't get
1391 * cleared when they're allocated, so we need to do it here.
1394 kmemcheck_mark_uninitialized_pages(page, pages);
1396 kmemcheck_mark_unallocated_pages(page, pages);
1399 page->objects = oo_objects(oo);
1401 order = compound_order(page);
1402 page->slab_cache = s;
1403 __SetPageSlab(page);
1404 if (page_is_pfmemalloc(page))
1405 SetPageSlabPfmemalloc(page);
1407 start = page_address(page);
1409 if (unlikely(s->flags & SLAB_POISON))
1410 memset(start, POISON_INUSE, PAGE_SIZE << order);
1412 kasan_poison_slab(page);
1414 for_each_object_idx(p, idx, s, start, page->objects) {
1415 setup_object(s, page, p);
1416 if (likely(idx < page->objects))
1417 set_freepointer(s, p, p + s->size);
1419 set_freepointer(s, p, NULL);
1422 page->freelist = start;
1423 page->inuse = page->objects;
1427 if (gfpflags_allow_blocking(flags))
1428 local_irq_disable();
1432 mod_zone_page_state(page_zone(page),
1433 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1434 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1437 inc_slabs_node(s, page_to_nid(page), page->objects);
1442 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1444 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1445 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1449 return allocate_slab(s,
1450 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1453 static void __free_slab(struct kmem_cache *s, struct page *page)
1455 int order = compound_order(page);
1456 int pages = 1 << order;
1458 if (kmem_cache_debug(s)) {
1461 slab_pad_check(s, page);
1462 for_each_object(p, s, page_address(page),
1464 check_object(s, page, p, SLUB_RED_INACTIVE);
1467 kmemcheck_free_shadow(page, compound_order(page));
1469 mod_zone_page_state(page_zone(page),
1470 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1471 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1474 __ClearPageSlabPfmemalloc(page);
1475 __ClearPageSlab(page);
1477 page_mapcount_reset(page);
1478 if (current->reclaim_state)
1479 current->reclaim_state->reclaimed_slab += pages;
1480 __free_kmem_pages(page, order);
1483 #define need_reserve_slab_rcu \
1484 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1486 static void rcu_free_slab(struct rcu_head *h)
1490 if (need_reserve_slab_rcu)
1491 page = virt_to_head_page(h);
1493 page = container_of((struct list_head *)h, struct page, lru);
1495 __free_slab(page->slab_cache, page);
1498 static void free_slab(struct kmem_cache *s, struct page *page)
1500 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1501 struct rcu_head *head;
1503 if (need_reserve_slab_rcu) {
1504 int order = compound_order(page);
1505 int offset = (PAGE_SIZE << order) - s->reserved;
1507 VM_BUG_ON(s->reserved != sizeof(*head));
1508 head = page_address(page) + offset;
1510 head = &page->rcu_head;
1513 call_rcu(head, rcu_free_slab);
1515 __free_slab(s, page);
1518 static void discard_slab(struct kmem_cache *s, struct page *page)
1520 dec_slabs_node(s, page_to_nid(page), page->objects);
1525 * Management of partially allocated slabs.
1528 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1531 if (tail == DEACTIVATE_TO_TAIL)
1532 list_add_tail(&page->lru, &n->partial);
1534 list_add(&page->lru, &n->partial);
1537 static inline void add_partial(struct kmem_cache_node *n,
1538 struct page *page, int tail)
1540 lockdep_assert_held(&n->list_lock);
1541 __add_partial(n, page, tail);
1545 __remove_partial(struct kmem_cache_node *n, struct page *page)
1547 list_del(&page->lru);
1551 static inline void remove_partial(struct kmem_cache_node *n,
1554 lockdep_assert_held(&n->list_lock);
1555 __remove_partial(n, page);
1559 * Remove slab from the partial list, freeze it and
1560 * return the pointer to the freelist.
1562 * Returns a list of objects or NULL if it fails.
1564 static inline void *acquire_slab(struct kmem_cache *s,
1565 struct kmem_cache_node *n, struct page *page,
1566 int mode, int *objects)
1569 unsigned long counters;
1572 lockdep_assert_held(&n->list_lock);
1575 * Zap the freelist and set the frozen bit.
1576 * The old freelist is the list of objects for the
1577 * per cpu allocation list.
1579 freelist = page->freelist;
1580 counters = page->counters;
1581 new.counters = counters;
1582 *objects = new.objects - new.inuse;
1584 new.inuse = page->objects;
1585 new.freelist = NULL;
1587 new.freelist = freelist;
1590 VM_BUG_ON(new.frozen);
1593 if (!__cmpxchg_double_slab(s, page,
1595 new.freelist, new.counters,
1599 remove_partial(n, page);
1604 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1605 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1608 * Try to allocate a partial slab from a specific node.
1610 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1611 struct kmem_cache_cpu *c, gfp_t flags)
1613 struct page *page, *page2;
1614 void *object = NULL;
1619 * Racy check. If we mistakenly see no partial slabs then we
1620 * just allocate an empty slab. If we mistakenly try to get a
1621 * partial slab and there is none available then get_partials()
1624 if (!n || !n->nr_partial)
1627 spin_lock(&n->list_lock);
1628 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1631 if (!pfmemalloc_match(page, flags))
1634 t = acquire_slab(s, n, page, object == NULL, &objects);
1638 available += objects;
1641 stat(s, ALLOC_FROM_PARTIAL);
1644 put_cpu_partial(s, page, 0);
1645 stat(s, CPU_PARTIAL_NODE);
1647 if (!kmem_cache_has_cpu_partial(s)
1648 || available > s->cpu_partial / 2)
1652 spin_unlock(&n->list_lock);
1657 * Get a page from somewhere. Search in increasing NUMA distances.
1659 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1660 struct kmem_cache_cpu *c)
1663 struct zonelist *zonelist;
1666 enum zone_type high_zoneidx = gfp_zone(flags);
1668 unsigned int cpuset_mems_cookie;
1671 * The defrag ratio allows a configuration of the tradeoffs between
1672 * inter node defragmentation and node local allocations. A lower
1673 * defrag_ratio increases the tendency to do local allocations
1674 * instead of attempting to obtain partial slabs from other nodes.
1676 * If the defrag_ratio is set to 0 then kmalloc() always
1677 * returns node local objects. If the ratio is higher then kmalloc()
1678 * may return off node objects because partial slabs are obtained
1679 * from other nodes and filled up.
1681 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1682 * defrag_ratio = 1000) then every (well almost) allocation will
1683 * first attempt to defrag slab caches on other nodes. This means
1684 * scanning over all nodes to look for partial slabs which may be
1685 * expensive if we do it every time we are trying to find a slab
1686 * with available objects.
1688 if (!s->remote_node_defrag_ratio ||
1689 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1693 cpuset_mems_cookie = read_mems_allowed_begin();
1694 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1695 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1696 struct kmem_cache_node *n;
1698 n = get_node(s, zone_to_nid(zone));
1700 if (n && cpuset_zone_allowed(zone, flags) &&
1701 n->nr_partial > s->min_partial) {
1702 object = get_partial_node(s, n, c, flags);
1705 * Don't check read_mems_allowed_retry()
1706 * here - if mems_allowed was updated in
1707 * parallel, that was a harmless race
1708 * between allocation and the cpuset
1715 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1721 * Get a partial page, lock it and return it.
1723 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1724 struct kmem_cache_cpu *c)
1727 int searchnode = node;
1729 if (node == NUMA_NO_NODE)
1730 searchnode = numa_mem_id();
1731 else if (!node_present_pages(node))
1732 searchnode = node_to_mem_node(node);
1734 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1735 if (object || node != NUMA_NO_NODE)
1738 return get_any_partial(s, flags, c);
1741 #ifdef CONFIG_PREEMPT
1743 * Calculate the next globally unique transaction for disambiguiation
1744 * during cmpxchg. The transactions start with the cpu number and are then
1745 * incremented by CONFIG_NR_CPUS.
1747 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1750 * No preemption supported therefore also no need to check for
1756 static inline unsigned long next_tid(unsigned long tid)
1758 return tid + TID_STEP;
1761 static inline unsigned int tid_to_cpu(unsigned long tid)
1763 return tid % TID_STEP;
1766 static inline unsigned long tid_to_event(unsigned long tid)
1768 return tid / TID_STEP;
1771 static inline unsigned int init_tid(int cpu)
1776 static inline void note_cmpxchg_failure(const char *n,
1777 const struct kmem_cache *s, unsigned long tid)
1779 #ifdef SLUB_DEBUG_CMPXCHG
1780 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1782 pr_info("%s %s: cmpxchg redo ", n, s->name);
1784 #ifdef CONFIG_PREEMPT
1785 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1786 pr_warn("due to cpu change %d -> %d\n",
1787 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1790 if (tid_to_event(tid) != tid_to_event(actual_tid))
1791 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1792 tid_to_event(tid), tid_to_event(actual_tid));
1794 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1795 actual_tid, tid, next_tid(tid));
1797 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1800 static void init_kmem_cache_cpus(struct kmem_cache *s)
1804 for_each_possible_cpu(cpu)
1805 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1809 * Remove the cpu slab
1811 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1814 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1815 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1817 enum slab_modes l = M_NONE, m = M_NONE;
1819 int tail = DEACTIVATE_TO_HEAD;
1823 if (page->freelist) {
1824 stat(s, DEACTIVATE_REMOTE_FREES);
1825 tail = DEACTIVATE_TO_TAIL;
1829 * Stage one: Free all available per cpu objects back
1830 * to the page freelist while it is still frozen. Leave the
1833 * There is no need to take the list->lock because the page
1836 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1838 unsigned long counters;
1841 prior = page->freelist;
1842 counters = page->counters;
1843 set_freepointer(s, freelist, prior);
1844 new.counters = counters;
1846 VM_BUG_ON(!new.frozen);
1848 } while (!__cmpxchg_double_slab(s, page,
1850 freelist, new.counters,
1851 "drain percpu freelist"));
1853 freelist = nextfree;
1857 * Stage two: Ensure that the page is unfrozen while the
1858 * list presence reflects the actual number of objects
1861 * We setup the list membership and then perform a cmpxchg
1862 * with the count. If there is a mismatch then the page
1863 * is not unfrozen but the page is on the wrong list.
1865 * Then we restart the process which may have to remove
1866 * the page from the list that we just put it on again
1867 * because the number of objects in the slab may have
1872 old.freelist = page->freelist;
1873 old.counters = page->counters;
1874 VM_BUG_ON(!old.frozen);
1876 /* Determine target state of the slab */
1877 new.counters = old.counters;
1880 set_freepointer(s, freelist, old.freelist);
1881 new.freelist = freelist;
1883 new.freelist = old.freelist;
1887 if (!new.inuse && n->nr_partial >= s->min_partial)
1889 else if (new.freelist) {
1894 * Taking the spinlock removes the possiblity
1895 * that acquire_slab() will see a slab page that
1898 spin_lock(&n->list_lock);
1902 if (kmem_cache_debug(s) && !lock) {
1905 * This also ensures that the scanning of full
1906 * slabs from diagnostic functions will not see
1909 spin_lock(&n->list_lock);
1917 remove_partial(n, page);
1919 else if (l == M_FULL)
1921 remove_full(s, n, page);
1923 if (m == M_PARTIAL) {
1925 add_partial(n, page, tail);
1928 } else if (m == M_FULL) {
1930 stat(s, DEACTIVATE_FULL);
1931 add_full(s, n, page);
1937 if (!__cmpxchg_double_slab(s, page,
1938 old.freelist, old.counters,
1939 new.freelist, new.counters,
1944 spin_unlock(&n->list_lock);
1947 stat(s, DEACTIVATE_EMPTY);
1948 discard_slab(s, page);
1954 * Unfreeze all the cpu partial slabs.
1956 * This function must be called with interrupts disabled
1957 * for the cpu using c (or some other guarantee must be there
1958 * to guarantee no concurrent accesses).
1960 static void unfreeze_partials(struct kmem_cache *s,
1961 struct kmem_cache_cpu *c)
1963 #ifdef CONFIG_SLUB_CPU_PARTIAL
1964 struct kmem_cache_node *n = NULL, *n2 = NULL;
1965 struct page *page, *discard_page = NULL;
1967 while ((page = c->partial)) {
1971 c->partial = page->next;
1973 n2 = get_node(s, page_to_nid(page));
1976 spin_unlock(&n->list_lock);
1979 spin_lock(&n->list_lock);
1984 old.freelist = page->freelist;
1985 old.counters = page->counters;
1986 VM_BUG_ON(!old.frozen);
1988 new.counters = old.counters;
1989 new.freelist = old.freelist;
1993 } while (!__cmpxchg_double_slab(s, page,
1994 old.freelist, old.counters,
1995 new.freelist, new.counters,
1996 "unfreezing slab"));
1998 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1999 page->next = discard_page;
2000 discard_page = page;
2002 add_partial(n, page, DEACTIVATE_TO_TAIL);
2003 stat(s, FREE_ADD_PARTIAL);
2008 spin_unlock(&n->list_lock);
2010 while (discard_page) {
2011 page = discard_page;
2012 discard_page = discard_page->next;
2014 stat(s, DEACTIVATE_EMPTY);
2015 discard_slab(s, page);
2022 * Put a page that was just frozen (in __slab_free) into a partial page
2023 * slot if available. This is done without interrupts disabled and without
2024 * preemption disabled. The cmpxchg is racy and may put the partial page
2025 * onto a random cpus partial slot.
2027 * If we did not find a slot then simply move all the partials to the
2028 * per node partial list.
2030 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2032 #ifdef CONFIG_SLUB_CPU_PARTIAL
2033 struct page *oldpage;
2041 oldpage = this_cpu_read(s->cpu_slab->partial);
2044 pobjects = oldpage->pobjects;
2045 pages = oldpage->pages;
2046 if (drain && pobjects > s->cpu_partial) {
2047 unsigned long flags;
2049 * partial array is full. Move the existing
2050 * set to the per node partial list.
2052 local_irq_save(flags);
2053 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2054 local_irq_restore(flags);
2058 stat(s, CPU_PARTIAL_DRAIN);
2063 pobjects += page->objects - page->inuse;
2065 page->pages = pages;
2066 page->pobjects = pobjects;
2067 page->next = oldpage;
2069 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2071 if (unlikely(!s->cpu_partial)) {
2072 unsigned long flags;
2074 local_irq_save(flags);
2075 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2076 local_irq_restore(flags);
2082 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2084 stat(s, CPUSLAB_FLUSH);
2085 deactivate_slab(s, c->page, c->freelist);
2087 c->tid = next_tid(c->tid);
2095 * Called from IPI handler with interrupts disabled.
2097 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2099 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2105 unfreeze_partials(s, c);
2109 static void flush_cpu_slab(void *d)
2111 struct kmem_cache *s = d;
2113 __flush_cpu_slab(s, smp_processor_id());
2116 static bool has_cpu_slab(int cpu, void *info)
2118 struct kmem_cache *s = info;
2119 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2121 return c->page || c->partial;
2124 static void flush_all(struct kmem_cache *s)
2126 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2130 * Check if the objects in a per cpu structure fit numa
2131 * locality expectations.
2133 static inline int node_match(struct page *page, int node)
2136 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2142 #ifdef CONFIG_SLUB_DEBUG
2143 static int count_free(struct page *page)
2145 return page->objects - page->inuse;
2148 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2150 return atomic_long_read(&n->total_objects);
2152 #endif /* CONFIG_SLUB_DEBUG */
2154 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2155 static unsigned long count_partial(struct kmem_cache_node *n,
2156 int (*get_count)(struct page *))
2158 unsigned long flags;
2159 unsigned long x = 0;
2162 spin_lock_irqsave(&n->list_lock, flags);
2163 list_for_each_entry(page, &n->partial, lru)
2164 x += get_count(page);
2165 spin_unlock_irqrestore(&n->list_lock, flags);
2168 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2170 static noinline void
2171 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2173 #ifdef CONFIG_SLUB_DEBUG
2174 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2175 DEFAULT_RATELIMIT_BURST);
2177 struct kmem_cache_node *n;
2179 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2182 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2184 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2185 s->name, s->object_size, s->size, oo_order(s->oo),
2188 if (oo_order(s->min) > get_order(s->object_size))
2189 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2192 for_each_kmem_cache_node(s, node, n) {
2193 unsigned long nr_slabs;
2194 unsigned long nr_objs;
2195 unsigned long nr_free;
2197 nr_free = count_partial(n, count_free);
2198 nr_slabs = node_nr_slabs(n);
2199 nr_objs = node_nr_objs(n);
2201 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2202 node, nr_slabs, nr_objs, nr_free);
2207 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2208 int node, struct kmem_cache_cpu **pc)
2211 struct kmem_cache_cpu *c = *pc;
2214 freelist = get_partial(s, flags, node, c);
2219 page = new_slab(s, flags, node);
2221 c = raw_cpu_ptr(s->cpu_slab);
2226 * No other reference to the page yet so we can
2227 * muck around with it freely without cmpxchg
2229 freelist = page->freelist;
2230 page->freelist = NULL;
2232 stat(s, ALLOC_SLAB);
2241 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2243 if (unlikely(PageSlabPfmemalloc(page)))
2244 return gfp_pfmemalloc_allowed(gfpflags);
2250 * Check the page->freelist of a page and either transfer the freelist to the
2251 * per cpu freelist or deactivate the page.
2253 * The page is still frozen if the return value is not NULL.
2255 * If this function returns NULL then the page has been unfrozen.
2257 * This function must be called with interrupt disabled.
2259 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2262 unsigned long counters;
2266 freelist = page->freelist;
2267 counters = page->counters;
2269 new.counters = counters;
2270 VM_BUG_ON(!new.frozen);
2272 new.inuse = page->objects;
2273 new.frozen = freelist != NULL;
2275 } while (!__cmpxchg_double_slab(s, page,
2284 * Slow path. The lockless freelist is empty or we need to perform
2287 * Processing is still very fast if new objects have been freed to the
2288 * regular freelist. In that case we simply take over the regular freelist
2289 * as the lockless freelist and zap the regular freelist.
2291 * If that is not working then we fall back to the partial lists. We take the
2292 * first element of the freelist as the object to allocate now and move the
2293 * rest of the freelist to the lockless freelist.
2295 * And if we were unable to get a new slab from the partial slab lists then
2296 * we need to allocate a new slab. This is the slowest path since it involves
2297 * a call to the page allocator and the setup of a new slab.
2299 * Version of __slab_alloc to use when we know that interrupts are
2300 * already disabled (which is the case for bulk allocation).
2302 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2303 unsigned long addr, struct kmem_cache_cpu *c)
2313 if (unlikely(!node_match(page, node))) {
2314 int searchnode = node;
2316 if (node != NUMA_NO_NODE && !node_present_pages(node))
2317 searchnode = node_to_mem_node(node);
2319 if (unlikely(!node_match(page, searchnode))) {
2320 stat(s, ALLOC_NODE_MISMATCH);
2321 deactivate_slab(s, page, c->freelist);
2329 * By rights, we should be searching for a slab page that was
2330 * PFMEMALLOC but right now, we are losing the pfmemalloc
2331 * information when the page leaves the per-cpu allocator
2333 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2334 deactivate_slab(s, page, c->freelist);
2340 /* must check again c->freelist in case of cpu migration or IRQ */
2341 freelist = c->freelist;
2345 freelist = get_freelist(s, page);
2349 stat(s, DEACTIVATE_BYPASS);
2353 stat(s, ALLOC_REFILL);
2357 * freelist is pointing to the list of objects to be used.
2358 * page is pointing to the page from which the objects are obtained.
2359 * That page must be frozen for per cpu allocations to work.
2361 VM_BUG_ON(!c->page->frozen);
2362 c->freelist = get_freepointer(s, freelist);
2363 c->tid = next_tid(c->tid);
2369 page = c->page = c->partial;
2370 c->partial = page->next;
2371 stat(s, CPU_PARTIAL_ALLOC);
2376 freelist = new_slab_objects(s, gfpflags, node, &c);
2378 if (unlikely(!freelist)) {
2379 slab_out_of_memory(s, gfpflags, node);
2384 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2387 /* Only entered in the debug case */
2388 if (kmem_cache_debug(s) &&
2389 !alloc_debug_processing(s, page, freelist, addr))
2390 goto new_slab; /* Slab failed checks. Next slab needed */
2392 deactivate_slab(s, page, get_freepointer(s, freelist));
2399 * Another one that disabled interrupt and compensates for possible
2400 * cpu changes by refetching the per cpu area pointer.
2402 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2403 unsigned long addr, struct kmem_cache_cpu *c)
2406 unsigned long flags;
2408 local_irq_save(flags);
2409 #ifdef CONFIG_PREEMPT
2411 * We may have been preempted and rescheduled on a different
2412 * cpu before disabling interrupts. Need to reload cpu area
2415 c = this_cpu_ptr(s->cpu_slab);
2418 p = ___slab_alloc(s, gfpflags, node, addr, c);
2419 local_irq_restore(flags);
2424 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2425 * have the fastpath folded into their functions. So no function call
2426 * overhead for requests that can be satisfied on the fastpath.
2428 * The fastpath works by first checking if the lockless freelist can be used.
2429 * If not then __slab_alloc is called for slow processing.
2431 * Otherwise we can simply pick the next object from the lockless free list.
2433 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2434 gfp_t gfpflags, int node, unsigned long addr)
2437 struct kmem_cache_cpu *c;
2441 s = slab_pre_alloc_hook(s, gfpflags);
2446 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2447 * enabled. We may switch back and forth between cpus while
2448 * reading from one cpu area. That does not matter as long
2449 * as we end up on the original cpu again when doing the cmpxchg.
2451 * We should guarantee that tid and kmem_cache are retrieved on
2452 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2453 * to check if it is matched or not.
2456 tid = this_cpu_read(s->cpu_slab->tid);
2457 c = raw_cpu_ptr(s->cpu_slab);
2458 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2459 unlikely(tid != READ_ONCE(c->tid)));
2462 * Irqless object alloc/free algorithm used here depends on sequence
2463 * of fetching cpu_slab's data. tid should be fetched before anything
2464 * on c to guarantee that object and page associated with previous tid
2465 * won't be used with current tid. If we fetch tid first, object and
2466 * page could be one associated with next tid and our alloc/free
2467 * request will be failed. In this case, we will retry. So, no problem.
2472 * The transaction ids are globally unique per cpu and per operation on
2473 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2474 * occurs on the right processor and that there was no operation on the
2475 * linked list in between.
2478 object = c->freelist;
2480 if (unlikely(!object || !node_match(page, node))) {
2481 object = __slab_alloc(s, gfpflags, node, addr, c);
2482 stat(s, ALLOC_SLOWPATH);
2484 void *next_object = get_freepointer_safe(s, object);
2487 * The cmpxchg will only match if there was no additional
2488 * operation and if we are on the right processor.
2490 * The cmpxchg does the following atomically (without lock
2492 * 1. Relocate first pointer to the current per cpu area.
2493 * 2. Verify that tid and freelist have not been changed
2494 * 3. If they were not changed replace tid and freelist
2496 * Since this is without lock semantics the protection is only
2497 * against code executing on this cpu *not* from access by
2500 if (unlikely(!this_cpu_cmpxchg_double(
2501 s->cpu_slab->freelist, s->cpu_slab->tid,
2503 next_object, next_tid(tid)))) {
2505 note_cmpxchg_failure("slab_alloc", s, tid);
2508 prefetch_freepointer(s, next_object);
2509 stat(s, ALLOC_FASTPATH);
2512 if (unlikely(gfpflags & __GFP_ZERO) && object)
2513 memset(object, 0, s->object_size);
2515 slab_post_alloc_hook(s, gfpflags, object);
2520 static __always_inline void *slab_alloc(struct kmem_cache *s,
2521 gfp_t gfpflags, unsigned long addr)
2523 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2526 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2528 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2530 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2535 EXPORT_SYMBOL(kmem_cache_alloc);
2537 #ifdef CONFIG_TRACING
2538 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2540 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2541 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2542 kasan_kmalloc(s, ret, size);
2545 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2549 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2551 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2553 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2554 s->object_size, s->size, gfpflags, node);
2558 EXPORT_SYMBOL(kmem_cache_alloc_node);
2560 #ifdef CONFIG_TRACING
2561 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2563 int node, size_t size)
2565 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2567 trace_kmalloc_node(_RET_IP_, ret,
2568 size, s->size, gfpflags, node);
2570 kasan_kmalloc(s, ret, size);
2573 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2578 * Slow path handling. This may still be called frequently since objects
2579 * have a longer lifetime than the cpu slabs in most processing loads.
2581 * So we still attempt to reduce cache line usage. Just take the slab
2582 * lock and free the item. If there is no additional partial page
2583 * handling required then we can return immediately.
2585 static void __slab_free(struct kmem_cache *s, struct page *page,
2586 void *x, unsigned long addr)
2589 void **object = (void *)x;
2592 unsigned long counters;
2593 struct kmem_cache_node *n = NULL;
2594 unsigned long uninitialized_var(flags);
2596 stat(s, FREE_SLOWPATH);
2598 if (kmem_cache_debug(s) &&
2599 !(n = free_debug_processing(s, page, x, addr, &flags)))
2604 spin_unlock_irqrestore(&n->list_lock, flags);
2607 prior = page->freelist;
2608 counters = page->counters;
2609 set_freepointer(s, object, prior);
2610 new.counters = counters;
2611 was_frozen = new.frozen;
2613 if ((!new.inuse || !prior) && !was_frozen) {
2615 if (kmem_cache_has_cpu_partial(s) && !prior) {
2618 * Slab was on no list before and will be
2620 * We can defer the list move and instead
2625 } else { /* Needs to be taken off a list */
2627 n = get_node(s, page_to_nid(page));
2629 * Speculatively acquire the list_lock.
2630 * If the cmpxchg does not succeed then we may
2631 * drop the list_lock without any processing.
2633 * Otherwise the list_lock will synchronize with
2634 * other processors updating the list of slabs.
2636 spin_lock_irqsave(&n->list_lock, flags);
2641 } while (!cmpxchg_double_slab(s, page,
2643 object, new.counters,
2649 * If we just froze the page then put it onto the
2650 * per cpu partial list.
2652 if (new.frozen && !was_frozen) {
2653 put_cpu_partial(s, page, 1);
2654 stat(s, CPU_PARTIAL_FREE);
2657 * The list lock was not taken therefore no list
2658 * activity can be necessary.
2661 stat(s, FREE_FROZEN);
2665 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2669 * Objects left in the slab. If it was not on the partial list before
2672 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2673 if (kmem_cache_debug(s))
2674 remove_full(s, n, page);
2675 add_partial(n, page, DEACTIVATE_TO_TAIL);
2676 stat(s, FREE_ADD_PARTIAL);
2678 spin_unlock_irqrestore(&n->list_lock, flags);
2684 * Slab on the partial list.
2686 remove_partial(n, page);
2687 stat(s, FREE_REMOVE_PARTIAL);
2689 /* Slab must be on the full list */
2690 remove_full(s, n, page);
2693 spin_unlock_irqrestore(&n->list_lock, flags);
2695 discard_slab(s, page);
2699 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2700 * can perform fastpath freeing without additional function calls.
2702 * The fastpath is only possible if we are freeing to the current cpu slab
2703 * of this processor. This typically the case if we have just allocated
2706 * If fastpath is not possible then fall back to __slab_free where we deal
2707 * with all sorts of special processing.
2709 static __always_inline void slab_free(struct kmem_cache *s,
2710 struct page *page, void *x, unsigned long addr)
2712 void **object = (void *)x;
2713 struct kmem_cache_cpu *c;
2716 slab_free_hook(s, x);
2720 * Determine the currently cpus per cpu slab.
2721 * The cpu may change afterward. However that does not matter since
2722 * data is retrieved via this pointer. If we are on the same cpu
2723 * during the cmpxchg then the free will succeed.
2726 tid = this_cpu_read(s->cpu_slab->tid);
2727 c = raw_cpu_ptr(s->cpu_slab);
2728 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2729 unlikely(tid != READ_ONCE(c->tid)));
2731 /* Same with comment on barrier() in slab_alloc_node() */
2734 if (likely(page == c->page)) {
2735 set_freepointer(s, object, c->freelist);
2737 if (unlikely(!this_cpu_cmpxchg_double(
2738 s->cpu_slab->freelist, s->cpu_slab->tid,
2740 object, next_tid(tid)))) {
2742 note_cmpxchg_failure("slab_free", s, tid);
2745 stat(s, FREE_FASTPATH);
2747 __slab_free(s, page, x, addr);
2751 void kmem_cache_free(struct kmem_cache *s, void *x)
2753 s = cache_from_obj(s, x);
2756 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2757 trace_kmem_cache_free(_RET_IP_, x);
2759 EXPORT_SYMBOL(kmem_cache_free);
2761 /* Note that interrupts must be enabled when calling this function. */
2762 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2764 struct kmem_cache_cpu *c;
2768 local_irq_disable();
2769 c = this_cpu_ptr(s->cpu_slab);
2771 for (i = 0; i < size; i++) {
2772 void *object = p[i];
2775 /* kmem cache debug support */
2776 s = cache_from_obj(s, object);
2779 slab_free_hook(s, object);
2781 page = virt_to_head_page(object);
2783 if (c->page == page) {
2784 /* Fastpath: local CPU free */
2785 set_freepointer(s, object, c->freelist);
2786 c->freelist = object;
2788 c->tid = next_tid(c->tid);
2790 /* Slowpath: overhead locked cmpxchg_double_slab */
2791 __slab_free(s, page, object, _RET_IP_);
2792 local_irq_disable();
2793 c = this_cpu_ptr(s->cpu_slab);
2797 c->tid = next_tid(c->tid);
2800 EXPORT_SYMBOL(kmem_cache_free_bulk);
2802 /* Note that interrupts must be enabled when calling this function. */
2803 bool kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2806 struct kmem_cache_cpu *c;
2810 * Drain objects in the per cpu slab, while disabling local
2811 * IRQs, which protects against PREEMPT and interrupts
2812 * handlers invoking normal fastpath.
2814 local_irq_disable();
2815 c = this_cpu_ptr(s->cpu_slab);
2817 for (i = 0; i < size; i++) {
2818 void *object = c->freelist;
2820 if (unlikely(!object)) {
2822 * Invoking slow path likely have side-effect
2823 * of re-populating per CPU c->freelist
2825 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2827 if (unlikely(!p[i]))
2830 c = this_cpu_ptr(s->cpu_slab);
2831 continue; /* goto for-loop */
2834 /* kmem_cache debug support */
2835 s = slab_pre_alloc_hook(s, flags);
2839 c->freelist = get_freepointer(s, object);
2842 /* kmem_cache debug support */
2843 slab_post_alloc_hook(s, flags, object);
2845 c->tid = next_tid(c->tid);
2848 /* Clear memory outside IRQ disabled fastpath loop */
2849 if (unlikely(flags & __GFP_ZERO)) {
2852 for (j = 0; j < i; j++)
2853 memset(p[j], 0, s->object_size);
2859 __kmem_cache_free_bulk(s, i, p);
2863 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2867 * Object placement in a slab is made very easy because we always start at
2868 * offset 0. If we tune the size of the object to the alignment then we can
2869 * get the required alignment by putting one properly sized object after
2872 * Notice that the allocation order determines the sizes of the per cpu
2873 * caches. Each processor has always one slab available for allocations.
2874 * Increasing the allocation order reduces the number of times that slabs
2875 * must be moved on and off the partial lists and is therefore a factor in
2880 * Mininum / Maximum order of slab pages. This influences locking overhead
2881 * and slab fragmentation. A higher order reduces the number of partial slabs
2882 * and increases the number of allocations possible without having to
2883 * take the list_lock.
2885 static int slub_min_order;
2886 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2887 static int slub_min_objects;
2890 * Calculate the order of allocation given an slab object size.
2892 * The order of allocation has significant impact on performance and other
2893 * system components. Generally order 0 allocations should be preferred since
2894 * order 0 does not cause fragmentation in the page allocator. Larger objects
2895 * be problematic to put into order 0 slabs because there may be too much
2896 * unused space left. We go to a higher order if more than 1/16th of the slab
2899 * In order to reach satisfactory performance we must ensure that a minimum
2900 * number of objects is in one slab. Otherwise we may generate too much
2901 * activity on the partial lists which requires taking the list_lock. This is
2902 * less a concern for large slabs though which are rarely used.
2904 * slub_max_order specifies the order where we begin to stop considering the
2905 * number of objects in a slab as critical. If we reach slub_max_order then
2906 * we try to keep the page order as low as possible. So we accept more waste
2907 * of space in favor of a small page order.
2909 * Higher order allocations also allow the placement of more objects in a
2910 * slab and thereby reduce object handling overhead. If the user has
2911 * requested a higher mininum order then we start with that one instead of
2912 * the smallest order which will fit the object.
2914 static inline int slab_order(int size, int min_objects,
2915 int max_order, int fract_leftover, int reserved)
2919 int min_order = slub_min_order;
2921 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2922 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2924 for (order = max(min_order, get_order(min_objects * size + reserved));
2925 order <= max_order; order++) {
2927 unsigned long slab_size = PAGE_SIZE << order;
2929 rem = (slab_size - reserved) % size;
2931 if (rem <= slab_size / fract_leftover)
2938 static inline int calculate_order(int size, int reserved)
2946 * Attempt to find best configuration for a slab. This
2947 * works by first attempting to generate a layout with
2948 * the best configuration and backing off gradually.
2950 * First we increase the acceptable waste in a slab. Then
2951 * we reduce the minimum objects required in a slab.
2953 min_objects = slub_min_objects;
2955 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2956 max_objects = order_objects(slub_max_order, size, reserved);
2957 min_objects = min(min_objects, max_objects);
2959 while (min_objects > 1) {
2961 while (fraction >= 4) {
2962 order = slab_order(size, min_objects,
2963 slub_max_order, fraction, reserved);
2964 if (order <= slub_max_order)
2972 * We were unable to place multiple objects in a slab. Now
2973 * lets see if we can place a single object there.
2975 order = slab_order(size, 1, slub_max_order, 1, reserved);
2976 if (order <= slub_max_order)
2980 * Doh this slab cannot be placed using slub_max_order.
2982 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2983 if (order < MAX_ORDER)
2989 init_kmem_cache_node(struct kmem_cache_node *n)
2992 spin_lock_init(&n->list_lock);
2993 INIT_LIST_HEAD(&n->partial);
2994 #ifdef CONFIG_SLUB_DEBUG
2995 atomic_long_set(&n->nr_slabs, 0);
2996 atomic_long_set(&n->total_objects, 0);
2997 INIT_LIST_HEAD(&n->full);
3001 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3003 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3004 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3007 * Must align to double word boundary for the double cmpxchg
3008 * instructions to work; see __pcpu_double_call_return_bool().
3010 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3011 2 * sizeof(void *));
3016 init_kmem_cache_cpus(s);
3021 static struct kmem_cache *kmem_cache_node;
3024 * No kmalloc_node yet so do it by hand. We know that this is the first
3025 * slab on the node for this slabcache. There are no concurrent accesses
3028 * Note that this function only works on the kmem_cache_node
3029 * when allocating for the kmem_cache_node. This is used for bootstrapping
3030 * memory on a fresh node that has no slab structures yet.
3032 static void early_kmem_cache_node_alloc(int node)
3035 struct kmem_cache_node *n;
3037 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3039 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3042 if (page_to_nid(page) != node) {
3043 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3044 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3049 page->freelist = get_freepointer(kmem_cache_node, n);
3052 kmem_cache_node->node[node] = n;
3053 #ifdef CONFIG_SLUB_DEBUG
3054 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3055 init_tracking(kmem_cache_node, n);
3057 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3058 init_kmem_cache_node(n);
3059 inc_slabs_node(kmem_cache_node, node, page->objects);
3062 * No locks need to be taken here as it has just been
3063 * initialized and there is no concurrent access.
3065 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3068 static void free_kmem_cache_nodes(struct kmem_cache *s)
3071 struct kmem_cache_node *n;
3073 for_each_kmem_cache_node(s, node, n) {
3074 kmem_cache_free(kmem_cache_node, n);
3075 s->node[node] = NULL;
3079 static int init_kmem_cache_nodes(struct kmem_cache *s)
3083 for_each_node_state(node, N_NORMAL_MEMORY) {
3084 struct kmem_cache_node *n;
3086 if (slab_state == DOWN) {
3087 early_kmem_cache_node_alloc(node);
3090 n = kmem_cache_alloc_node(kmem_cache_node,
3094 free_kmem_cache_nodes(s);
3099 init_kmem_cache_node(n);
3104 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3106 if (min < MIN_PARTIAL)
3108 else if (min > MAX_PARTIAL)
3110 s->min_partial = min;
3114 * calculate_sizes() determines the order and the distribution of data within
3117 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3119 unsigned long flags = s->flags;
3120 unsigned long size = s->object_size;
3124 * Round up object size to the next word boundary. We can only
3125 * place the free pointer at word boundaries and this determines
3126 * the possible location of the free pointer.
3128 size = ALIGN(size, sizeof(void *));
3130 #ifdef CONFIG_SLUB_DEBUG
3132 * Determine if we can poison the object itself. If the user of
3133 * the slab may touch the object after free or before allocation
3134 * then we should never poison the object itself.
3136 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3138 s->flags |= __OBJECT_POISON;
3140 s->flags &= ~__OBJECT_POISON;
3144 * If we are Redzoning then check if there is some space between the
3145 * end of the object and the free pointer. If not then add an
3146 * additional word to have some bytes to store Redzone information.
3148 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3149 size += sizeof(void *);
3153 * With that we have determined the number of bytes in actual use
3154 * by the object. This is the potential offset to the free pointer.
3158 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3161 * Relocate free pointer after the object if it is not
3162 * permitted to overwrite the first word of the object on
3165 * This is the case if we do RCU, have a constructor or
3166 * destructor or are poisoning the objects.
3169 size += sizeof(void *);
3172 #ifdef CONFIG_SLUB_DEBUG
3173 if (flags & SLAB_STORE_USER)
3175 * Need to store information about allocs and frees after
3178 size += 2 * sizeof(struct track);
3180 if (flags & SLAB_RED_ZONE)
3182 * Add some empty padding so that we can catch
3183 * overwrites from earlier objects rather than let
3184 * tracking information or the free pointer be
3185 * corrupted if a user writes before the start
3188 size += sizeof(void *);
3192 * SLUB stores one object immediately after another beginning from
3193 * offset 0. In order to align the objects we have to simply size
3194 * each object to conform to the alignment.
3196 size = ALIGN(size, s->align);
3198 if (forced_order >= 0)
3199 order = forced_order;
3201 order = calculate_order(size, s->reserved);
3208 s->allocflags |= __GFP_COMP;
3210 if (s->flags & SLAB_CACHE_DMA)
3211 s->allocflags |= GFP_DMA;
3213 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3214 s->allocflags |= __GFP_RECLAIMABLE;
3217 * Determine the number of objects per slab
3219 s->oo = oo_make(order, size, s->reserved);
3220 s->min = oo_make(get_order(size), size, s->reserved);
3221 if (oo_objects(s->oo) > oo_objects(s->max))
3224 return !!oo_objects(s->oo);
3227 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3229 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3232 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3233 s->reserved = sizeof(struct rcu_head);
3235 if (!calculate_sizes(s, -1))
3237 if (disable_higher_order_debug) {
3239 * Disable debugging flags that store metadata if the min slab
3242 if (get_order(s->size) > get_order(s->object_size)) {
3243 s->flags &= ~DEBUG_METADATA_FLAGS;
3245 if (!calculate_sizes(s, -1))
3250 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3251 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3252 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3253 /* Enable fast mode */
3254 s->flags |= __CMPXCHG_DOUBLE;
3258 * The larger the object size is, the more pages we want on the partial
3259 * list to avoid pounding the page allocator excessively.
3261 set_min_partial(s, ilog2(s->size) / 2);
3264 * cpu_partial determined the maximum number of objects kept in the
3265 * per cpu partial lists of a processor.
3267 * Per cpu partial lists mainly contain slabs that just have one
3268 * object freed. If they are used for allocation then they can be
3269 * filled up again with minimal effort. The slab will never hit the
3270 * per node partial lists and therefore no locking will be required.
3272 * This setting also determines
3274 * A) The number of objects from per cpu partial slabs dumped to the
3275 * per node list when we reach the limit.
3276 * B) The number of objects in cpu partial slabs to extract from the
3277 * per node list when we run out of per cpu objects. We only fetch
3278 * 50% to keep some capacity around for frees.
3280 if (!kmem_cache_has_cpu_partial(s))
3282 else if (s->size >= PAGE_SIZE)
3284 else if (s->size >= 1024)
3286 else if (s->size >= 256)
3287 s->cpu_partial = 13;
3289 s->cpu_partial = 30;
3292 s->remote_node_defrag_ratio = 1000;
3294 if (!init_kmem_cache_nodes(s))
3297 if (alloc_kmem_cache_cpus(s))
3300 free_kmem_cache_nodes(s);
3302 if (flags & SLAB_PANIC)
3303 panic("Cannot create slab %s size=%lu realsize=%u "
3304 "order=%u offset=%u flags=%lx\n",
3305 s->name, (unsigned long)s->size, s->size,
3306 oo_order(s->oo), s->offset, flags);
3310 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3313 #ifdef CONFIG_SLUB_DEBUG
3314 void *addr = page_address(page);
3316 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3317 sizeof(long), GFP_ATOMIC);
3320 slab_err(s, page, text, s->name);
3323 get_map(s, page, map);
3324 for_each_object(p, s, addr, page->objects) {
3326 if (!test_bit(slab_index(p, s, addr), map)) {
3327 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3328 print_tracking(s, p);
3337 * Attempt to free all partial slabs on a node.
3338 * This is called from kmem_cache_close(). We must be the last thread
3339 * using the cache and therefore we do not need to lock anymore.
3341 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3343 struct page *page, *h;
3345 list_for_each_entry_safe(page, h, &n->partial, lru) {
3347 __remove_partial(n, page);
3348 discard_slab(s, page);
3350 list_slab_objects(s, page,
3351 "Objects remaining in %s on kmem_cache_close()");
3357 * Release all resources used by a slab cache.
3359 static inline int kmem_cache_close(struct kmem_cache *s)
3362 struct kmem_cache_node *n;
3365 /* Attempt to free all objects */
3366 for_each_kmem_cache_node(s, node, n) {
3368 if (n->nr_partial || slabs_node(s, node))
3371 free_percpu(s->cpu_slab);
3372 free_kmem_cache_nodes(s);
3376 int __kmem_cache_shutdown(struct kmem_cache *s)
3378 return kmem_cache_close(s);
3381 /********************************************************************
3383 *******************************************************************/
3385 static int __init setup_slub_min_order(char *str)
3387 get_option(&str, &slub_min_order);
3392 __setup("slub_min_order=", setup_slub_min_order);
3394 static int __init setup_slub_max_order(char *str)
3396 get_option(&str, &slub_max_order);
3397 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3402 __setup("slub_max_order=", setup_slub_max_order);
3404 static int __init setup_slub_min_objects(char *str)
3406 get_option(&str, &slub_min_objects);
3411 __setup("slub_min_objects=", setup_slub_min_objects);
3413 void *__kmalloc(size_t size, gfp_t flags)
3415 struct kmem_cache *s;
3418 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3419 return kmalloc_large(size, flags);
3421 s = kmalloc_slab(size, flags);
3423 if (unlikely(ZERO_OR_NULL_PTR(s)))
3426 ret = slab_alloc(s, flags, _RET_IP_);
3428 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3430 kasan_kmalloc(s, ret, size);
3434 EXPORT_SYMBOL(__kmalloc);
3437 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3442 flags |= __GFP_COMP | __GFP_NOTRACK;
3443 page = alloc_kmem_pages_node(node, flags, get_order(size));
3445 ptr = page_address(page);
3447 kmalloc_large_node_hook(ptr, size, flags);
3451 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3453 struct kmem_cache *s;
3456 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3457 ret = kmalloc_large_node(size, flags, node);
3459 trace_kmalloc_node(_RET_IP_, ret,
3460 size, PAGE_SIZE << get_order(size),
3466 s = kmalloc_slab(size, flags);
3468 if (unlikely(ZERO_OR_NULL_PTR(s)))
3471 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3473 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3475 kasan_kmalloc(s, ret, size);
3479 EXPORT_SYMBOL(__kmalloc_node);
3482 static size_t __ksize(const void *object)
3486 if (unlikely(object == ZERO_SIZE_PTR))
3489 page = virt_to_head_page(object);
3491 if (unlikely(!PageSlab(page))) {
3492 WARN_ON(!PageCompound(page));
3493 return PAGE_SIZE << compound_order(page);
3496 return slab_ksize(page->slab_cache);
3499 size_t ksize(const void *object)
3501 size_t size = __ksize(object);
3502 /* We assume that ksize callers could use whole allocated area,
3503 so we need unpoison this area. */
3504 kasan_krealloc(object, size);
3507 EXPORT_SYMBOL(ksize);
3509 void kfree(const void *x)
3512 void *object = (void *)x;
3514 trace_kfree(_RET_IP_, x);
3516 if (unlikely(ZERO_OR_NULL_PTR(x)))
3519 page = virt_to_head_page(x);
3520 if (unlikely(!PageSlab(page))) {
3521 BUG_ON(!PageCompound(page));
3523 __free_kmem_pages(page, compound_order(page));
3526 slab_free(page->slab_cache, page, object, _RET_IP_);
3528 EXPORT_SYMBOL(kfree);
3530 #define SHRINK_PROMOTE_MAX 32
3533 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3534 * up most to the head of the partial lists. New allocations will then
3535 * fill those up and thus they can be removed from the partial lists.
3537 * The slabs with the least items are placed last. This results in them
3538 * being allocated from last increasing the chance that the last objects
3539 * are freed in them.
3541 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3545 struct kmem_cache_node *n;
3548 struct list_head discard;
3549 struct list_head promote[SHRINK_PROMOTE_MAX];
3550 unsigned long flags;
3555 * Disable empty slabs caching. Used to avoid pinning offline
3556 * memory cgroups by kmem pages that can be freed.
3562 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3563 * so we have to make sure the change is visible.
3565 kick_all_cpus_sync();
3569 for_each_kmem_cache_node(s, node, n) {
3570 INIT_LIST_HEAD(&discard);
3571 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3572 INIT_LIST_HEAD(promote + i);
3574 spin_lock_irqsave(&n->list_lock, flags);
3577 * Build lists of slabs to discard or promote.
3579 * Note that concurrent frees may occur while we hold the
3580 * list_lock. page->inuse here is the upper limit.
3582 list_for_each_entry_safe(page, t, &n->partial, lru) {
3583 int free = page->objects - page->inuse;
3585 /* Do not reread page->inuse */
3588 /* We do not keep full slabs on the list */
3591 if (free == page->objects) {
3592 list_move(&page->lru, &discard);
3594 } else if (free <= SHRINK_PROMOTE_MAX)
3595 list_move(&page->lru, promote + free - 1);
3599 * Promote the slabs filled up most to the head of the
3602 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3603 list_splice(promote + i, &n->partial);
3605 spin_unlock_irqrestore(&n->list_lock, flags);
3607 /* Release empty slabs */
3608 list_for_each_entry_safe(page, t, &discard, lru)
3609 discard_slab(s, page);
3611 if (slabs_node(s, node))
3618 static int slab_mem_going_offline_callback(void *arg)
3620 struct kmem_cache *s;
3622 mutex_lock(&slab_mutex);
3623 list_for_each_entry(s, &slab_caches, list)
3624 __kmem_cache_shrink(s, false);
3625 mutex_unlock(&slab_mutex);
3630 static void slab_mem_offline_callback(void *arg)
3632 struct kmem_cache_node *n;
3633 struct kmem_cache *s;
3634 struct memory_notify *marg = arg;
3637 offline_node = marg->status_change_nid_normal;
3640 * If the node still has available memory. we need kmem_cache_node
3643 if (offline_node < 0)
3646 mutex_lock(&slab_mutex);
3647 list_for_each_entry(s, &slab_caches, list) {
3648 n = get_node(s, offline_node);
3651 * if n->nr_slabs > 0, slabs still exist on the node
3652 * that is going down. We were unable to free them,
3653 * and offline_pages() function shouldn't call this
3654 * callback. So, we must fail.
3656 BUG_ON(slabs_node(s, offline_node));
3658 s->node[offline_node] = NULL;
3659 kmem_cache_free(kmem_cache_node, n);
3662 mutex_unlock(&slab_mutex);
3665 static int slab_mem_going_online_callback(void *arg)
3667 struct kmem_cache_node *n;
3668 struct kmem_cache *s;
3669 struct memory_notify *marg = arg;
3670 int nid = marg->status_change_nid_normal;
3674 * If the node's memory is already available, then kmem_cache_node is
3675 * already created. Nothing to do.
3681 * We are bringing a node online. No memory is available yet. We must
3682 * allocate a kmem_cache_node structure in order to bring the node
3685 mutex_lock(&slab_mutex);
3686 list_for_each_entry(s, &slab_caches, list) {
3688 * XXX: kmem_cache_alloc_node will fallback to other nodes
3689 * since memory is not yet available from the node that
3692 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3697 init_kmem_cache_node(n);
3701 mutex_unlock(&slab_mutex);
3705 static int slab_memory_callback(struct notifier_block *self,
3706 unsigned long action, void *arg)
3711 case MEM_GOING_ONLINE:
3712 ret = slab_mem_going_online_callback(arg);
3714 case MEM_GOING_OFFLINE:
3715 ret = slab_mem_going_offline_callback(arg);
3718 case MEM_CANCEL_ONLINE:
3719 slab_mem_offline_callback(arg);
3722 case MEM_CANCEL_OFFLINE:
3726 ret = notifier_from_errno(ret);
3732 static struct notifier_block slab_memory_callback_nb = {
3733 .notifier_call = slab_memory_callback,
3734 .priority = SLAB_CALLBACK_PRI,
3737 /********************************************************************
3738 * Basic setup of slabs
3739 *******************************************************************/
3742 * Used for early kmem_cache structures that were allocated using
3743 * the page allocator. Allocate them properly then fix up the pointers
3744 * that may be pointing to the wrong kmem_cache structure.
3747 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3750 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3751 struct kmem_cache_node *n;
3753 memcpy(s, static_cache, kmem_cache->object_size);
3756 * This runs very early, and only the boot processor is supposed to be
3757 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3760 __flush_cpu_slab(s, smp_processor_id());
3761 for_each_kmem_cache_node(s, node, n) {
3764 list_for_each_entry(p, &n->partial, lru)
3767 #ifdef CONFIG_SLUB_DEBUG
3768 list_for_each_entry(p, &n->full, lru)
3772 slab_init_memcg_params(s);
3773 list_add(&s->list, &slab_caches);
3777 void __init kmem_cache_init(void)
3779 static __initdata struct kmem_cache boot_kmem_cache,
3780 boot_kmem_cache_node;
3782 if (debug_guardpage_minorder())
3785 kmem_cache_node = &boot_kmem_cache_node;
3786 kmem_cache = &boot_kmem_cache;
3788 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3789 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3791 register_hotmemory_notifier(&slab_memory_callback_nb);
3793 /* Able to allocate the per node structures */
3794 slab_state = PARTIAL;
3796 create_boot_cache(kmem_cache, "kmem_cache",
3797 offsetof(struct kmem_cache, node) +
3798 nr_node_ids * sizeof(struct kmem_cache_node *),
3799 SLAB_HWCACHE_ALIGN);
3801 kmem_cache = bootstrap(&boot_kmem_cache);
3804 * Allocate kmem_cache_node properly from the kmem_cache slab.
3805 * kmem_cache_node is separately allocated so no need to
3806 * update any list pointers.
3808 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3810 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3811 setup_kmalloc_cache_index_table();
3812 create_kmalloc_caches(0);
3815 register_cpu_notifier(&slab_notifier);
3818 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3820 slub_min_order, slub_max_order, slub_min_objects,
3821 nr_cpu_ids, nr_node_ids);
3824 void __init kmem_cache_init_late(void)
3829 __kmem_cache_alias(const char *name, size_t size, size_t align,
3830 unsigned long flags, void (*ctor)(void *))
3832 struct kmem_cache *s, *c;
3834 s = find_mergeable(size, align, flags, name, ctor);
3839 * Adjust the object sizes so that we clear
3840 * the complete object on kzalloc.
3842 s->object_size = max(s->object_size, (int)size);
3843 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3845 for_each_memcg_cache(c, s) {
3846 c->object_size = s->object_size;
3847 c->inuse = max_t(int, c->inuse,
3848 ALIGN(size, sizeof(void *)));
3851 if (sysfs_slab_alias(s, name)) {
3860 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3864 err = kmem_cache_open(s, flags);
3868 /* Mutex is not taken during early boot */
3869 if (slab_state <= UP)
3872 memcg_propagate_slab_attrs(s);
3873 err = sysfs_slab_add(s);
3875 kmem_cache_close(s);
3882 * Use the cpu notifier to insure that the cpu slabs are flushed when
3885 static int slab_cpuup_callback(struct notifier_block *nfb,
3886 unsigned long action, void *hcpu)
3888 long cpu = (long)hcpu;
3889 struct kmem_cache *s;
3890 unsigned long flags;
3893 case CPU_UP_CANCELED:
3894 case CPU_UP_CANCELED_FROZEN:
3896 case CPU_DEAD_FROZEN:
3897 mutex_lock(&slab_mutex);
3898 list_for_each_entry(s, &slab_caches, list) {
3899 local_irq_save(flags);
3900 __flush_cpu_slab(s, cpu);
3901 local_irq_restore(flags);
3903 mutex_unlock(&slab_mutex);
3911 static struct notifier_block slab_notifier = {
3912 .notifier_call = slab_cpuup_callback
3917 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3919 struct kmem_cache *s;
3922 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3923 return kmalloc_large(size, gfpflags);
3925 s = kmalloc_slab(size, gfpflags);
3927 if (unlikely(ZERO_OR_NULL_PTR(s)))
3930 ret = slab_alloc(s, gfpflags, caller);
3932 /* Honor the call site pointer we received. */
3933 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3939 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3940 int node, unsigned long caller)
3942 struct kmem_cache *s;
3945 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3946 ret = kmalloc_large_node(size, gfpflags, node);
3948 trace_kmalloc_node(caller, ret,
3949 size, PAGE_SIZE << get_order(size),
3955 s = kmalloc_slab(size, gfpflags);
3957 if (unlikely(ZERO_OR_NULL_PTR(s)))
3960 ret = slab_alloc_node(s, gfpflags, node, caller);
3962 /* Honor the call site pointer we received. */
3963 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3970 static int count_inuse(struct page *page)
3975 static int count_total(struct page *page)
3977 return page->objects;
3981 #ifdef CONFIG_SLUB_DEBUG
3982 static int validate_slab(struct kmem_cache *s, struct page *page,
3986 void *addr = page_address(page);
3988 if (!check_slab(s, page) ||
3989 !on_freelist(s, page, NULL))
3992 /* Now we know that a valid freelist exists */
3993 bitmap_zero(map, page->objects);
3995 get_map(s, page, map);
3996 for_each_object(p, s, addr, page->objects) {
3997 if (test_bit(slab_index(p, s, addr), map))
3998 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4002 for_each_object(p, s, addr, page->objects)
4003 if (!test_bit(slab_index(p, s, addr), map))
4004 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4009 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4013 validate_slab(s, page, map);
4017 static int validate_slab_node(struct kmem_cache *s,
4018 struct kmem_cache_node *n, unsigned long *map)
4020 unsigned long count = 0;
4022 unsigned long flags;
4024 spin_lock_irqsave(&n->list_lock, flags);
4026 list_for_each_entry(page, &n->partial, lru) {
4027 validate_slab_slab(s, page, map);
4030 if (count != n->nr_partial)
4031 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4032 s->name, count, n->nr_partial);
4034 if (!(s->flags & SLAB_STORE_USER))
4037 list_for_each_entry(page, &n->full, lru) {
4038 validate_slab_slab(s, page, map);
4041 if (count != atomic_long_read(&n->nr_slabs))
4042 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4043 s->name, count, atomic_long_read(&n->nr_slabs));
4046 spin_unlock_irqrestore(&n->list_lock, flags);
4050 static long validate_slab_cache(struct kmem_cache *s)
4053 unsigned long count = 0;
4054 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4055 sizeof(unsigned long), GFP_KERNEL);
4056 struct kmem_cache_node *n;
4062 for_each_kmem_cache_node(s, node, n)
4063 count += validate_slab_node(s, n, map);
4068 * Generate lists of code addresses where slabcache objects are allocated
4073 unsigned long count;
4080 DECLARE_BITMAP(cpus, NR_CPUS);
4086 unsigned long count;
4087 struct location *loc;
4090 static void free_loc_track(struct loc_track *t)
4093 free_pages((unsigned long)t->loc,
4094 get_order(sizeof(struct location) * t->max));
4097 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4102 order = get_order(sizeof(struct location) * max);
4104 l = (void *)__get_free_pages(flags, order);
4109 memcpy(l, t->loc, sizeof(struct location) * t->count);
4117 static int add_location(struct loc_track *t, struct kmem_cache *s,
4118 const struct track *track)
4120 long start, end, pos;
4122 unsigned long caddr;
4123 unsigned long age = jiffies - track->when;
4129 pos = start + (end - start + 1) / 2;
4132 * There is nothing at "end". If we end up there
4133 * we need to add something to before end.
4138 caddr = t->loc[pos].addr;
4139 if (track->addr == caddr) {
4145 if (age < l->min_time)
4147 if (age > l->max_time)
4150 if (track->pid < l->min_pid)
4151 l->min_pid = track->pid;
4152 if (track->pid > l->max_pid)
4153 l->max_pid = track->pid;
4155 cpumask_set_cpu(track->cpu,
4156 to_cpumask(l->cpus));
4158 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4162 if (track->addr < caddr)
4169 * Not found. Insert new tracking element.
4171 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4177 (t->count - pos) * sizeof(struct location));
4180 l->addr = track->addr;
4184 l->min_pid = track->pid;
4185 l->max_pid = track->pid;
4186 cpumask_clear(to_cpumask(l->cpus));
4187 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4188 nodes_clear(l->nodes);
4189 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4193 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4194 struct page *page, enum track_item alloc,
4197 void *addr = page_address(page);
4200 bitmap_zero(map, page->objects);
4201 get_map(s, page, map);
4203 for_each_object(p, s, addr, page->objects)
4204 if (!test_bit(slab_index(p, s, addr), map))
4205 add_location(t, s, get_track(s, p, alloc));
4208 static int list_locations(struct kmem_cache *s, char *buf,
4209 enum track_item alloc)
4213 struct loc_track t = { 0, 0, NULL };
4215 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4216 sizeof(unsigned long), GFP_KERNEL);
4217 struct kmem_cache_node *n;
4219 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4222 return sprintf(buf, "Out of memory\n");
4224 /* Push back cpu slabs */
4227 for_each_kmem_cache_node(s, node, n) {
4228 unsigned long flags;
4231 if (!atomic_long_read(&n->nr_slabs))
4234 spin_lock_irqsave(&n->list_lock, flags);
4235 list_for_each_entry(page, &n->partial, lru)
4236 process_slab(&t, s, page, alloc, map);
4237 list_for_each_entry(page, &n->full, lru)
4238 process_slab(&t, s, page, alloc, map);
4239 spin_unlock_irqrestore(&n->list_lock, flags);
4242 for (i = 0; i < t.count; i++) {
4243 struct location *l = &t.loc[i];
4245 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4247 len += sprintf(buf + len, "%7ld ", l->count);
4250 len += sprintf(buf + len, "%pS", (void *)l->addr);
4252 len += sprintf(buf + len, "<not-available>");
4254 if (l->sum_time != l->min_time) {
4255 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4257 (long)div_u64(l->sum_time, l->count),
4260 len += sprintf(buf + len, " age=%ld",
4263 if (l->min_pid != l->max_pid)
4264 len += sprintf(buf + len, " pid=%ld-%ld",
4265 l->min_pid, l->max_pid);
4267 len += sprintf(buf + len, " pid=%ld",
4270 if (num_online_cpus() > 1 &&
4271 !cpumask_empty(to_cpumask(l->cpus)) &&
4272 len < PAGE_SIZE - 60)
4273 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4275 cpumask_pr_args(to_cpumask(l->cpus)));
4277 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4278 len < PAGE_SIZE - 60)
4279 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4281 nodemask_pr_args(&l->nodes));
4283 len += sprintf(buf + len, "\n");
4289 len += sprintf(buf, "No data\n");
4294 #ifdef SLUB_RESILIENCY_TEST
4295 static void __init resiliency_test(void)
4299 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4301 pr_err("SLUB resiliency testing\n");
4302 pr_err("-----------------------\n");
4303 pr_err("A. Corruption after allocation\n");
4305 p = kzalloc(16, GFP_KERNEL);
4307 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4310 validate_slab_cache(kmalloc_caches[4]);
4312 /* Hmmm... The next two are dangerous */
4313 p = kzalloc(32, GFP_KERNEL);
4314 p[32 + sizeof(void *)] = 0x34;
4315 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4317 pr_err("If allocated object is overwritten then not detectable\n\n");
4319 validate_slab_cache(kmalloc_caches[5]);
4320 p = kzalloc(64, GFP_KERNEL);
4321 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4323 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4325 pr_err("If allocated object is overwritten then not detectable\n\n");
4326 validate_slab_cache(kmalloc_caches[6]);
4328 pr_err("\nB. Corruption after free\n");
4329 p = kzalloc(128, GFP_KERNEL);
4332 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4333 validate_slab_cache(kmalloc_caches[7]);
4335 p = kzalloc(256, GFP_KERNEL);
4338 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4339 validate_slab_cache(kmalloc_caches[8]);
4341 p = kzalloc(512, GFP_KERNEL);
4344 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4345 validate_slab_cache(kmalloc_caches[9]);
4349 static void resiliency_test(void) {};
4354 enum slab_stat_type {
4355 SL_ALL, /* All slabs */
4356 SL_PARTIAL, /* Only partially allocated slabs */
4357 SL_CPU, /* Only slabs used for cpu caches */
4358 SL_OBJECTS, /* Determine allocated objects not slabs */
4359 SL_TOTAL /* Determine object capacity not slabs */
4362 #define SO_ALL (1 << SL_ALL)
4363 #define SO_PARTIAL (1 << SL_PARTIAL)
4364 #define SO_CPU (1 << SL_CPU)
4365 #define SO_OBJECTS (1 << SL_OBJECTS)
4366 #define SO_TOTAL (1 << SL_TOTAL)
4368 static ssize_t show_slab_objects(struct kmem_cache *s,
4369 char *buf, unsigned long flags)
4371 unsigned long total = 0;
4374 unsigned long *nodes;
4376 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4380 if (flags & SO_CPU) {
4383 for_each_possible_cpu(cpu) {
4384 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4389 page = READ_ONCE(c->page);
4393 node = page_to_nid(page);
4394 if (flags & SO_TOTAL)
4396 else if (flags & SO_OBJECTS)
4404 page = READ_ONCE(c->partial);
4406 node = page_to_nid(page);
4407 if (flags & SO_TOTAL)
4409 else if (flags & SO_OBJECTS)
4420 #ifdef CONFIG_SLUB_DEBUG
4421 if (flags & SO_ALL) {
4422 struct kmem_cache_node *n;
4424 for_each_kmem_cache_node(s, node, n) {
4426 if (flags & SO_TOTAL)
4427 x = atomic_long_read(&n->total_objects);
4428 else if (flags & SO_OBJECTS)
4429 x = atomic_long_read(&n->total_objects) -
4430 count_partial(n, count_free);
4432 x = atomic_long_read(&n->nr_slabs);
4439 if (flags & SO_PARTIAL) {
4440 struct kmem_cache_node *n;
4442 for_each_kmem_cache_node(s, node, n) {
4443 if (flags & SO_TOTAL)
4444 x = count_partial(n, count_total);
4445 else if (flags & SO_OBJECTS)
4446 x = count_partial(n, count_inuse);
4453 x = sprintf(buf, "%lu", total);
4455 for (node = 0; node < nr_node_ids; node++)
4457 x += sprintf(buf + x, " N%d=%lu",
4462 return x + sprintf(buf + x, "\n");
4465 #ifdef CONFIG_SLUB_DEBUG
4466 static int any_slab_objects(struct kmem_cache *s)
4469 struct kmem_cache_node *n;
4471 for_each_kmem_cache_node(s, node, n)
4472 if (atomic_long_read(&n->total_objects))
4479 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4480 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4482 struct slab_attribute {
4483 struct attribute attr;
4484 ssize_t (*show)(struct kmem_cache *s, char *buf);
4485 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4488 #define SLAB_ATTR_RO(_name) \
4489 static struct slab_attribute _name##_attr = \
4490 __ATTR(_name, 0400, _name##_show, NULL)
4492 #define SLAB_ATTR(_name) \
4493 static struct slab_attribute _name##_attr = \
4494 __ATTR(_name, 0600, _name##_show, _name##_store)
4496 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4498 return sprintf(buf, "%d\n", s->size);
4500 SLAB_ATTR_RO(slab_size);
4502 static ssize_t align_show(struct kmem_cache *s, char *buf)
4504 return sprintf(buf, "%d\n", s->align);
4506 SLAB_ATTR_RO(align);
4508 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4510 return sprintf(buf, "%d\n", s->object_size);
4512 SLAB_ATTR_RO(object_size);
4514 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4516 return sprintf(buf, "%d\n", oo_objects(s->oo));
4518 SLAB_ATTR_RO(objs_per_slab);
4520 static ssize_t order_store(struct kmem_cache *s,
4521 const char *buf, size_t length)
4523 unsigned long order;
4526 err = kstrtoul(buf, 10, &order);
4530 if (order > slub_max_order || order < slub_min_order)
4533 calculate_sizes(s, order);
4537 static ssize_t order_show(struct kmem_cache *s, char *buf)
4539 return sprintf(buf, "%d\n", oo_order(s->oo));
4543 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4545 return sprintf(buf, "%lu\n", s->min_partial);
4548 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4554 err = kstrtoul(buf, 10, &min);
4558 set_min_partial(s, min);
4561 SLAB_ATTR(min_partial);
4563 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4565 return sprintf(buf, "%u\n", s->cpu_partial);
4568 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4571 unsigned long objects;
4574 err = kstrtoul(buf, 10, &objects);
4577 if (objects && !kmem_cache_has_cpu_partial(s))
4580 s->cpu_partial = objects;
4584 SLAB_ATTR(cpu_partial);
4586 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4590 return sprintf(buf, "%pS\n", s->ctor);
4594 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4596 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4598 SLAB_ATTR_RO(aliases);
4600 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4602 return show_slab_objects(s, buf, SO_PARTIAL);
4604 SLAB_ATTR_RO(partial);
4606 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4608 return show_slab_objects(s, buf, SO_CPU);
4610 SLAB_ATTR_RO(cpu_slabs);
4612 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4614 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4616 SLAB_ATTR_RO(objects);
4618 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4620 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4622 SLAB_ATTR_RO(objects_partial);
4624 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4631 for_each_online_cpu(cpu) {
4632 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4635 pages += page->pages;
4636 objects += page->pobjects;
4640 len = sprintf(buf, "%d(%d)", objects, pages);
4643 for_each_online_cpu(cpu) {
4644 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4646 if (page && len < PAGE_SIZE - 20)
4647 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4648 page->pobjects, page->pages);
4651 return len + sprintf(buf + len, "\n");
4653 SLAB_ATTR_RO(slabs_cpu_partial);
4655 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4657 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4660 static ssize_t reclaim_account_store(struct kmem_cache *s,
4661 const char *buf, size_t length)
4663 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4665 s->flags |= SLAB_RECLAIM_ACCOUNT;
4668 SLAB_ATTR(reclaim_account);
4670 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4672 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4674 SLAB_ATTR_RO(hwcache_align);
4676 #ifdef CONFIG_ZONE_DMA
4677 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4679 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4681 SLAB_ATTR_RO(cache_dma);
4684 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4686 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4688 SLAB_ATTR_RO(destroy_by_rcu);
4690 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4692 return sprintf(buf, "%d\n", s->reserved);
4694 SLAB_ATTR_RO(reserved);
4696 #ifdef CONFIG_SLUB_DEBUG
4697 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4699 return show_slab_objects(s, buf, SO_ALL);
4701 SLAB_ATTR_RO(slabs);
4703 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4705 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4707 SLAB_ATTR_RO(total_objects);
4709 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4711 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4714 static ssize_t sanity_checks_store(struct kmem_cache *s,
4715 const char *buf, size_t length)
4717 s->flags &= ~SLAB_DEBUG_FREE;
4718 if (buf[0] == '1') {
4719 s->flags &= ~__CMPXCHG_DOUBLE;
4720 s->flags |= SLAB_DEBUG_FREE;
4724 SLAB_ATTR(sanity_checks);
4726 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4728 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4731 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4735 * Tracing a merged cache is going to give confusing results
4736 * as well as cause other issues like converting a mergeable
4737 * cache into an umergeable one.
4739 if (s->refcount > 1)
4742 s->flags &= ~SLAB_TRACE;
4743 if (buf[0] == '1') {
4744 s->flags &= ~__CMPXCHG_DOUBLE;
4745 s->flags |= SLAB_TRACE;
4751 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4753 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4756 static ssize_t red_zone_store(struct kmem_cache *s,
4757 const char *buf, size_t length)
4759 if (any_slab_objects(s))
4762 s->flags &= ~SLAB_RED_ZONE;
4763 if (buf[0] == '1') {
4764 s->flags &= ~__CMPXCHG_DOUBLE;
4765 s->flags |= SLAB_RED_ZONE;
4767 calculate_sizes(s, -1);
4770 SLAB_ATTR(red_zone);
4772 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4774 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4777 static ssize_t poison_store(struct kmem_cache *s,
4778 const char *buf, size_t length)
4780 if (any_slab_objects(s))
4783 s->flags &= ~SLAB_POISON;
4784 if (buf[0] == '1') {
4785 s->flags &= ~__CMPXCHG_DOUBLE;
4786 s->flags |= SLAB_POISON;
4788 calculate_sizes(s, -1);
4793 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4795 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4798 static ssize_t store_user_store(struct kmem_cache *s,
4799 const char *buf, size_t length)
4801 if (any_slab_objects(s))
4804 s->flags &= ~SLAB_STORE_USER;
4805 if (buf[0] == '1') {
4806 s->flags &= ~__CMPXCHG_DOUBLE;
4807 s->flags |= SLAB_STORE_USER;
4809 calculate_sizes(s, -1);
4812 SLAB_ATTR(store_user);
4814 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4819 static ssize_t validate_store(struct kmem_cache *s,
4820 const char *buf, size_t length)
4824 if (buf[0] == '1') {
4825 ret = validate_slab_cache(s);
4831 SLAB_ATTR(validate);
4833 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4835 if (!(s->flags & SLAB_STORE_USER))
4837 return list_locations(s, buf, TRACK_ALLOC);
4839 SLAB_ATTR_RO(alloc_calls);
4841 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4843 if (!(s->flags & SLAB_STORE_USER))
4845 return list_locations(s, buf, TRACK_FREE);
4847 SLAB_ATTR_RO(free_calls);
4848 #endif /* CONFIG_SLUB_DEBUG */
4850 #ifdef CONFIG_FAILSLAB
4851 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4853 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4856 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4859 if (s->refcount > 1)
4862 s->flags &= ~SLAB_FAILSLAB;
4864 s->flags |= SLAB_FAILSLAB;
4867 SLAB_ATTR(failslab);
4870 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4875 static ssize_t shrink_store(struct kmem_cache *s,
4876 const char *buf, size_t length)
4879 kmem_cache_shrink(s);
4887 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4889 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4892 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4893 const char *buf, size_t length)
4895 unsigned long ratio;
4898 err = kstrtoul(buf, 10, &ratio);
4903 s->remote_node_defrag_ratio = ratio * 10;
4907 SLAB_ATTR(remote_node_defrag_ratio);
4910 #ifdef CONFIG_SLUB_STATS
4911 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4913 unsigned long sum = 0;
4916 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4921 for_each_online_cpu(cpu) {
4922 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4928 len = sprintf(buf, "%lu", sum);
4931 for_each_online_cpu(cpu) {
4932 if (data[cpu] && len < PAGE_SIZE - 20)
4933 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4937 return len + sprintf(buf + len, "\n");
4940 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4944 for_each_online_cpu(cpu)
4945 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4948 #define STAT_ATTR(si, text) \
4949 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4951 return show_stat(s, buf, si); \
4953 static ssize_t text##_store(struct kmem_cache *s, \
4954 const char *buf, size_t length) \
4956 if (buf[0] != '0') \
4958 clear_stat(s, si); \
4963 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4964 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4965 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4966 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4967 STAT_ATTR(FREE_FROZEN, free_frozen);
4968 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4969 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4970 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4971 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4972 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4973 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4974 STAT_ATTR(FREE_SLAB, free_slab);
4975 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4976 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4977 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4978 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4979 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4980 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4981 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4982 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4983 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4984 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4985 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4986 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4987 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4988 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4991 static struct attribute *slab_attrs[] = {
4992 &slab_size_attr.attr,
4993 &object_size_attr.attr,
4994 &objs_per_slab_attr.attr,
4996 &min_partial_attr.attr,
4997 &cpu_partial_attr.attr,
4999 &objects_partial_attr.attr,
5001 &cpu_slabs_attr.attr,
5005 &hwcache_align_attr.attr,
5006 &reclaim_account_attr.attr,
5007 &destroy_by_rcu_attr.attr,
5009 &reserved_attr.attr,
5010 &slabs_cpu_partial_attr.attr,
5011 #ifdef CONFIG_SLUB_DEBUG
5012 &total_objects_attr.attr,
5014 &sanity_checks_attr.attr,
5016 &red_zone_attr.attr,
5018 &store_user_attr.attr,
5019 &validate_attr.attr,
5020 &alloc_calls_attr.attr,
5021 &free_calls_attr.attr,
5023 #ifdef CONFIG_ZONE_DMA
5024 &cache_dma_attr.attr,
5027 &remote_node_defrag_ratio_attr.attr,
5029 #ifdef CONFIG_SLUB_STATS
5030 &alloc_fastpath_attr.attr,
5031 &alloc_slowpath_attr.attr,
5032 &free_fastpath_attr.attr,
5033 &free_slowpath_attr.attr,
5034 &free_frozen_attr.attr,
5035 &free_add_partial_attr.attr,
5036 &free_remove_partial_attr.attr,
5037 &alloc_from_partial_attr.attr,
5038 &alloc_slab_attr.attr,
5039 &alloc_refill_attr.attr,
5040 &alloc_node_mismatch_attr.attr,
5041 &free_slab_attr.attr,
5042 &cpuslab_flush_attr.attr,
5043 &deactivate_full_attr.attr,
5044 &deactivate_empty_attr.attr,
5045 &deactivate_to_head_attr.attr,
5046 &deactivate_to_tail_attr.attr,
5047 &deactivate_remote_frees_attr.attr,
5048 &deactivate_bypass_attr.attr,
5049 &order_fallback_attr.attr,
5050 &cmpxchg_double_fail_attr.attr,
5051 &cmpxchg_double_cpu_fail_attr.attr,
5052 &cpu_partial_alloc_attr.attr,
5053 &cpu_partial_free_attr.attr,
5054 &cpu_partial_node_attr.attr,
5055 &cpu_partial_drain_attr.attr,
5057 #ifdef CONFIG_FAILSLAB
5058 &failslab_attr.attr,
5064 static struct attribute_group slab_attr_group = {
5065 .attrs = slab_attrs,
5068 static ssize_t slab_attr_show(struct kobject *kobj,
5069 struct attribute *attr,
5072 struct slab_attribute *attribute;
5073 struct kmem_cache *s;
5076 attribute = to_slab_attr(attr);
5079 if (!attribute->show)
5082 err = attribute->show(s, buf);
5087 static ssize_t slab_attr_store(struct kobject *kobj,
5088 struct attribute *attr,
5089 const char *buf, size_t len)
5091 struct slab_attribute *attribute;
5092 struct kmem_cache *s;
5095 attribute = to_slab_attr(attr);
5098 if (!attribute->store)
5101 err = attribute->store(s, buf, len);
5102 #ifdef CONFIG_MEMCG_KMEM
5103 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5104 struct kmem_cache *c;
5106 mutex_lock(&slab_mutex);
5107 if (s->max_attr_size < len)
5108 s->max_attr_size = len;
5111 * This is a best effort propagation, so this function's return
5112 * value will be determined by the parent cache only. This is
5113 * basically because not all attributes will have a well
5114 * defined semantics for rollbacks - most of the actions will
5115 * have permanent effects.
5117 * Returning the error value of any of the children that fail
5118 * is not 100 % defined, in the sense that users seeing the
5119 * error code won't be able to know anything about the state of
5122 * Only returning the error code for the parent cache at least
5123 * has well defined semantics. The cache being written to
5124 * directly either failed or succeeded, in which case we loop
5125 * through the descendants with best-effort propagation.
5127 for_each_memcg_cache(c, s)
5128 attribute->store(c, buf, len);
5129 mutex_unlock(&slab_mutex);
5135 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5137 #ifdef CONFIG_MEMCG_KMEM
5139 char *buffer = NULL;
5140 struct kmem_cache *root_cache;
5142 if (is_root_cache(s))
5145 root_cache = s->memcg_params.root_cache;
5148 * This mean this cache had no attribute written. Therefore, no point
5149 * in copying default values around
5151 if (!root_cache->max_attr_size)
5154 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5157 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5159 if (!attr || !attr->store || !attr->show)
5163 * It is really bad that we have to allocate here, so we will
5164 * do it only as a fallback. If we actually allocate, though,
5165 * we can just use the allocated buffer until the end.
5167 * Most of the slub attributes will tend to be very small in
5168 * size, but sysfs allows buffers up to a page, so they can
5169 * theoretically happen.
5173 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5176 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5177 if (WARN_ON(!buffer))
5182 attr->show(root_cache, buf);
5183 attr->store(s, buf, strlen(buf));
5187 free_page((unsigned long)buffer);
5191 static void kmem_cache_release(struct kobject *k)
5193 slab_kmem_cache_release(to_slab(k));
5196 static const struct sysfs_ops slab_sysfs_ops = {
5197 .show = slab_attr_show,
5198 .store = slab_attr_store,
5201 static struct kobj_type slab_ktype = {
5202 .sysfs_ops = &slab_sysfs_ops,
5203 .release = kmem_cache_release,
5206 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5208 struct kobj_type *ktype = get_ktype(kobj);
5210 if (ktype == &slab_ktype)
5215 static const struct kset_uevent_ops slab_uevent_ops = {
5216 .filter = uevent_filter,
5219 static struct kset *slab_kset;
5221 static inline struct kset *cache_kset(struct kmem_cache *s)
5223 #ifdef CONFIG_MEMCG_KMEM
5224 if (!is_root_cache(s))
5225 return s->memcg_params.root_cache->memcg_kset;
5230 #define ID_STR_LENGTH 64
5232 /* Create a unique string id for a slab cache:
5234 * Format :[flags-]size
5236 static char *create_unique_id(struct kmem_cache *s)
5238 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5245 * First flags affecting slabcache operations. We will only
5246 * get here for aliasable slabs so we do not need to support
5247 * too many flags. The flags here must cover all flags that
5248 * are matched during merging to guarantee that the id is
5251 if (s->flags & SLAB_CACHE_DMA)
5253 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5255 if (s->flags & SLAB_DEBUG_FREE)
5257 if (!(s->flags & SLAB_NOTRACK))
5261 p += sprintf(p, "%07d", s->size);
5263 BUG_ON(p > name + ID_STR_LENGTH - 1);
5267 static int sysfs_slab_add(struct kmem_cache *s)
5271 int unmergeable = slab_unmergeable(s);
5275 * Slabcache can never be merged so we can use the name proper.
5276 * This is typically the case for debug situations. In that
5277 * case we can catch duplicate names easily.
5279 sysfs_remove_link(&slab_kset->kobj, s->name);
5283 * Create a unique name for the slab as a target
5286 name = create_unique_id(s);
5289 s->kobj.kset = cache_kset(s);
5290 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5294 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5298 #ifdef CONFIG_MEMCG_KMEM
5299 if (is_root_cache(s)) {
5300 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5301 if (!s->memcg_kset) {
5308 kobject_uevent(&s->kobj, KOBJ_ADD);
5310 /* Setup first alias */
5311 sysfs_slab_alias(s, s->name);
5318 kobject_del(&s->kobj);
5322 void sysfs_slab_remove(struct kmem_cache *s)
5324 if (slab_state < FULL)
5326 * Sysfs has not been setup yet so no need to remove the
5331 #ifdef CONFIG_MEMCG_KMEM
5332 kset_unregister(s->memcg_kset);
5334 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5335 kobject_del(&s->kobj);
5336 kobject_put(&s->kobj);
5340 * Need to buffer aliases during bootup until sysfs becomes
5341 * available lest we lose that information.
5343 struct saved_alias {
5344 struct kmem_cache *s;
5346 struct saved_alias *next;
5349 static struct saved_alias *alias_list;
5351 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5353 struct saved_alias *al;
5355 if (slab_state == FULL) {
5357 * If we have a leftover link then remove it.
5359 sysfs_remove_link(&slab_kset->kobj, name);
5360 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5363 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5369 al->next = alias_list;
5374 static int __init slab_sysfs_init(void)
5376 struct kmem_cache *s;
5379 mutex_lock(&slab_mutex);
5381 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5383 mutex_unlock(&slab_mutex);
5384 pr_err("Cannot register slab subsystem.\n");
5390 list_for_each_entry(s, &slab_caches, list) {
5391 err = sysfs_slab_add(s);
5393 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5397 while (alias_list) {
5398 struct saved_alias *al = alias_list;
5400 alias_list = alias_list->next;
5401 err = sysfs_slab_alias(al->s, al->name);
5403 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5408 mutex_unlock(&slab_mutex);
5413 __initcall(slab_sysfs_init);
5414 #endif /* CONFIG_SYSFS */
5417 * The /proc/slabinfo ABI
5419 #ifdef CONFIG_SLABINFO
5420 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5422 unsigned long nr_slabs = 0;
5423 unsigned long nr_objs = 0;
5424 unsigned long nr_free = 0;
5426 struct kmem_cache_node *n;
5428 for_each_kmem_cache_node(s, node, n) {
5429 nr_slabs += node_nr_slabs(n);
5430 nr_objs += node_nr_objs(n);
5431 nr_free += count_partial(n, count_free);
5434 sinfo->active_objs = nr_objs - nr_free;
5435 sinfo->num_objs = nr_objs;
5436 sinfo->active_slabs = nr_slabs;
5437 sinfo->num_slabs = nr_slabs;
5438 sinfo->objects_per_slab = oo_objects(s->oo);
5439 sinfo->cache_order = oo_order(s->oo);
5442 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5446 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5447 size_t count, loff_t *ppos)
5451 #endif /* CONFIG_SLABINFO */