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 int order_objects(int order, unsigned long size, int reserved)
289 return ((PAGE_SIZE << order) - reserved) / size;
292 static inline struct kmem_cache_order_objects oo_make(int order,
293 unsigned long size, int reserved)
295 struct kmem_cache_order_objects x = {
296 (order << OO_SHIFT) + order_objects(order, size, reserved)
302 static inline int oo_order(struct kmem_cache_order_objects x)
304 return x.x >> OO_SHIFT;
307 static inline int oo_objects(struct kmem_cache_order_objects x)
309 return x.x & OO_MASK;
313 * Per slab locking using the pagelock
315 static __always_inline void slab_lock(struct page *page)
317 VM_BUG_ON_PAGE(PageTail(page), page);
318 bit_spin_lock(PG_locked, &page->flags);
321 static __always_inline void slab_unlock(struct page *page)
323 VM_BUG_ON_PAGE(PageTail(page), page);
324 __bit_spin_unlock(PG_locked, &page->flags);
327 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
330 tmp.counters = counters_new;
332 * page->counters can cover frozen/inuse/objects as well
333 * as page->_count. If we assign to ->counters directly
334 * we run the risk of losing updates to page->_count, so
335 * be careful and only assign to the fields we need.
337 page->frozen = tmp.frozen;
338 page->inuse = tmp.inuse;
339 page->objects = tmp.objects;
342 /* Interrupts must be disabled (for the fallback code to work right) */
343 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
344 void *freelist_old, unsigned long counters_old,
345 void *freelist_new, unsigned long counters_new,
348 VM_BUG_ON(!irqs_disabled());
349 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
350 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
351 if (s->flags & __CMPXCHG_DOUBLE) {
352 if (cmpxchg_double(&page->freelist, &page->counters,
353 freelist_old, counters_old,
354 freelist_new, counters_new))
360 if (page->freelist == freelist_old &&
361 page->counters == counters_old) {
362 page->freelist = freelist_new;
363 set_page_slub_counters(page, counters_new);
371 stat(s, CMPXCHG_DOUBLE_FAIL);
373 #ifdef SLUB_DEBUG_CMPXCHG
374 pr_info("%s %s: cmpxchg double redo ", n, s->name);
380 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
381 void *freelist_old, unsigned long counters_old,
382 void *freelist_new, unsigned long counters_new,
385 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
386 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
387 if (s->flags & __CMPXCHG_DOUBLE) {
388 if (cmpxchg_double(&page->freelist, &page->counters,
389 freelist_old, counters_old,
390 freelist_new, counters_new))
397 local_irq_save(flags);
399 if (page->freelist == freelist_old &&
400 page->counters == counters_old) {
401 page->freelist = freelist_new;
402 set_page_slub_counters(page, counters_new);
404 local_irq_restore(flags);
408 local_irq_restore(flags);
412 stat(s, CMPXCHG_DOUBLE_FAIL);
414 #ifdef SLUB_DEBUG_CMPXCHG
415 pr_info("%s %s: cmpxchg double redo ", n, s->name);
421 #ifdef CONFIG_SLUB_DEBUG
423 * Determine a map of object in use on a page.
425 * Node listlock must be held to guarantee that the page does
426 * not vanish from under us.
428 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
431 void *addr = page_address(page);
433 for (p = page->freelist; p; p = get_freepointer(s, p))
434 set_bit(slab_index(p, s, addr), map);
440 #if defined(CONFIG_SLUB_DEBUG_ON)
441 static int slub_debug = DEBUG_DEFAULT_FLAGS;
442 #elif defined(CONFIG_KASAN)
443 static int slub_debug = SLAB_STORE_USER;
445 static int slub_debug;
448 static char *slub_debug_slabs;
449 static int disable_higher_order_debug;
452 * slub is about to manipulate internal object metadata. This memory lies
453 * outside the range of the allocated object, so accessing it would normally
454 * be reported by kasan as a bounds error. metadata_access_enable() is used
455 * to tell kasan that these accesses are OK.
457 static inline void metadata_access_enable(void)
459 kasan_disable_current();
462 static inline void metadata_access_disable(void)
464 kasan_enable_current();
470 static void print_section(char *text, u8 *addr, unsigned int length)
472 metadata_access_enable();
473 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
475 metadata_access_disable();
478 static struct track *get_track(struct kmem_cache *s, void *object,
479 enum track_item alloc)
484 p = object + s->offset + sizeof(void *);
486 p = object + s->inuse;
491 static void set_track(struct kmem_cache *s, void *object,
492 enum track_item alloc, unsigned long addr)
494 struct track *p = get_track(s, object, alloc);
497 #ifdef CONFIG_STACKTRACE
498 struct stack_trace trace;
501 trace.nr_entries = 0;
502 trace.max_entries = TRACK_ADDRS_COUNT;
503 trace.entries = p->addrs;
505 metadata_access_enable();
506 save_stack_trace(&trace);
507 metadata_access_disable();
509 /* See rant in lockdep.c */
510 if (trace.nr_entries != 0 &&
511 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
514 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
518 p->cpu = smp_processor_id();
519 p->pid = current->pid;
522 memset(p, 0, sizeof(struct track));
525 static void init_tracking(struct kmem_cache *s, void *object)
527 if (!(s->flags & SLAB_STORE_USER))
530 set_track(s, object, TRACK_FREE, 0UL);
531 set_track(s, object, TRACK_ALLOC, 0UL);
534 static void print_track(const char *s, struct track *t)
539 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
540 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
541 #ifdef CONFIG_STACKTRACE
544 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
546 pr_err("\t%pS\n", (void *)t->addrs[i]);
553 static void print_tracking(struct kmem_cache *s, void *object)
555 if (!(s->flags & SLAB_STORE_USER))
558 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
559 print_track("Freed", get_track(s, object, TRACK_FREE));
562 static void print_page_info(struct page *page)
564 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
565 page, page->objects, page->inuse, page->freelist, page->flags);
569 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
571 struct va_format vaf;
577 pr_err("=============================================================================\n");
578 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
579 pr_err("-----------------------------------------------------------------------------\n\n");
581 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
585 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
587 struct va_format vaf;
593 pr_err("FIX %s: %pV\n", s->name, &vaf);
597 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
599 unsigned int off; /* Offset of last byte */
600 u8 *addr = page_address(page);
602 print_tracking(s, p);
604 print_page_info(page);
606 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607 p, p - addr, get_freepointer(s, p));
610 print_section("Bytes b4 ", p - 16, 16);
612 print_section("Object ", p, min_t(unsigned long, s->object_size,
614 if (s->flags & SLAB_RED_ZONE)
615 print_section("Redzone ", p + s->object_size,
616 s->inuse - s->object_size);
619 off = s->offset + sizeof(void *);
623 if (s->flags & SLAB_STORE_USER)
624 off += 2 * sizeof(struct track);
627 /* Beginning of the filler is the free pointer */
628 print_section("Padding ", p + off, s->size - off);
633 void object_err(struct kmem_cache *s, struct page *page,
634 u8 *object, char *reason)
636 slab_bug(s, "%s", reason);
637 print_trailer(s, page, object);
640 static void slab_err(struct kmem_cache *s, struct page *page,
641 const char *fmt, ...)
647 vsnprintf(buf, sizeof(buf), fmt, args);
649 slab_bug(s, "%s", buf);
650 print_page_info(page);
654 static void init_object(struct kmem_cache *s, void *object, u8 val)
658 if (s->flags & __OBJECT_POISON) {
659 memset(p, POISON_FREE, s->object_size - 1);
660 p[s->object_size - 1] = POISON_END;
663 if (s->flags & SLAB_RED_ZONE)
664 memset(p + s->object_size, val, s->inuse - s->object_size);
667 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
668 void *from, void *to)
670 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
671 memset(from, data, to - from);
674 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
675 u8 *object, char *what,
676 u8 *start, unsigned int value, unsigned int bytes)
681 metadata_access_enable();
682 fault = memchr_inv(start, value, bytes);
683 metadata_access_disable();
688 while (end > fault && end[-1] == value)
691 slab_bug(s, "%s overwritten", what);
692 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
693 fault, end - 1, fault[0], value);
694 print_trailer(s, page, object);
696 restore_bytes(s, what, value, fault, end);
704 * Bytes of the object to be managed.
705 * If the freepointer may overlay the object then the free
706 * pointer is the first word of the object.
708 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
711 * object + s->object_size
712 * Padding to reach word boundary. This is also used for Redzoning.
713 * Padding is extended by another word if Redzoning is enabled and
714 * object_size == inuse.
716 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
717 * 0xcc (RED_ACTIVE) for objects in use.
720 * Meta data starts here.
722 * A. Free pointer (if we cannot overwrite object on free)
723 * B. Tracking data for SLAB_STORE_USER
724 * C. Padding to reach required alignment boundary or at mininum
725 * one word if debugging is on to be able to detect writes
726 * before the word boundary.
728 * Padding is done using 0x5a (POISON_INUSE)
731 * Nothing is used beyond s->size.
733 * If slabcaches are merged then the object_size and inuse boundaries are mostly
734 * ignored. And therefore no slab options that rely on these boundaries
735 * may be used with merged slabcaches.
738 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
740 unsigned long off = s->inuse; /* The end of info */
743 /* Freepointer is placed after the object. */
744 off += sizeof(void *);
746 if (s->flags & SLAB_STORE_USER)
747 /* We also have user information there */
748 off += 2 * sizeof(struct track);
753 return check_bytes_and_report(s, page, p, "Object padding",
754 p + off, POISON_INUSE, s->size - off);
757 /* Check the pad bytes at the end of a slab page */
758 static int slab_pad_check(struct kmem_cache *s, struct page *page)
766 if (!(s->flags & SLAB_POISON))
769 start = page_address(page);
770 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
771 end = start + length;
772 remainder = length % s->size;
776 metadata_access_enable();
777 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
778 metadata_access_disable();
781 while (end > fault && end[-1] == POISON_INUSE)
784 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
785 print_section("Padding ", end - remainder, remainder);
787 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
791 static int check_object(struct kmem_cache *s, struct page *page,
792 void *object, u8 val)
795 u8 *endobject = object + s->object_size;
797 if (s->flags & SLAB_RED_ZONE) {
798 if (!check_bytes_and_report(s, page, object, "Redzone",
799 endobject, val, s->inuse - s->object_size))
802 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
803 check_bytes_and_report(s, page, p, "Alignment padding",
804 endobject, POISON_INUSE,
805 s->inuse - s->object_size);
809 if (s->flags & SLAB_POISON) {
810 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
811 (!check_bytes_and_report(s, page, p, "Poison", p,
812 POISON_FREE, s->object_size - 1) ||
813 !check_bytes_and_report(s, page, p, "Poison",
814 p + s->object_size - 1, POISON_END, 1)))
817 * check_pad_bytes cleans up on its own.
819 check_pad_bytes(s, page, p);
822 if (!s->offset && val == SLUB_RED_ACTIVE)
824 * Object and freepointer overlap. Cannot check
825 * freepointer while object is allocated.
829 /* Check free pointer validity */
830 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
831 object_err(s, page, p, "Freepointer corrupt");
833 * No choice but to zap it and thus lose the remainder
834 * of the free objects in this slab. May cause
835 * another error because the object count is now wrong.
837 set_freepointer(s, p, NULL);
843 static int check_slab(struct kmem_cache *s, struct page *page)
847 VM_BUG_ON(!irqs_disabled());
849 if (!PageSlab(page)) {
850 slab_err(s, page, "Not a valid slab page");
854 maxobj = order_objects(compound_order(page), s->size, s->reserved);
855 if (page->objects > maxobj) {
856 slab_err(s, page, "objects %u > max %u",
857 page->objects, maxobj);
860 if (page->inuse > page->objects) {
861 slab_err(s, page, "inuse %u > max %u",
862 page->inuse, page->objects);
865 /* Slab_pad_check fixes things up after itself */
866 slab_pad_check(s, page);
871 * Determine if a certain object on a page is on the freelist. Must hold the
872 * slab lock to guarantee that the chains are in a consistent state.
874 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
882 while (fp && nr <= page->objects) {
885 if (!check_valid_pointer(s, page, fp)) {
887 object_err(s, page, object,
888 "Freechain corrupt");
889 set_freepointer(s, object, NULL);
891 slab_err(s, page, "Freepointer corrupt");
892 page->freelist = NULL;
893 page->inuse = page->objects;
894 slab_fix(s, "Freelist cleared");
900 fp = get_freepointer(s, object);
904 max_objects = order_objects(compound_order(page), s->size, s->reserved);
905 if (max_objects > MAX_OBJS_PER_PAGE)
906 max_objects = MAX_OBJS_PER_PAGE;
908 if (page->objects != max_objects) {
909 slab_err(s, page, "Wrong number of objects. Found %d but "
910 "should be %d", page->objects, max_objects);
911 page->objects = max_objects;
912 slab_fix(s, "Number of objects adjusted.");
914 if (page->inuse != page->objects - nr) {
915 slab_err(s, page, "Wrong object count. Counter is %d but "
916 "counted were %d", page->inuse, page->objects - nr);
917 page->inuse = page->objects - nr;
918 slab_fix(s, "Object count adjusted.");
920 return search == NULL;
923 static void trace(struct kmem_cache *s, struct page *page, void *object,
926 if (s->flags & SLAB_TRACE) {
927 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
929 alloc ? "alloc" : "free",
934 print_section("Object ", (void *)object,
942 * Tracking of fully allocated slabs for debugging purposes.
944 static void add_full(struct kmem_cache *s,
945 struct kmem_cache_node *n, struct page *page)
947 if (!(s->flags & SLAB_STORE_USER))
950 lockdep_assert_held(&n->list_lock);
951 list_add(&page->lru, &n->full);
954 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
956 if (!(s->flags & SLAB_STORE_USER))
959 lockdep_assert_held(&n->list_lock);
960 list_del(&page->lru);
963 /* Tracking of the number of slabs for debugging purposes */
964 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
966 struct kmem_cache_node *n = get_node(s, node);
968 return atomic_long_read(&n->nr_slabs);
971 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
973 return atomic_long_read(&n->nr_slabs);
976 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
978 struct kmem_cache_node *n = get_node(s, node);
981 * May be called early in order to allocate a slab for the
982 * kmem_cache_node structure. Solve the chicken-egg
983 * dilemma by deferring the increment of the count during
984 * bootstrap (see early_kmem_cache_node_alloc).
987 atomic_long_inc(&n->nr_slabs);
988 atomic_long_add(objects, &n->total_objects);
991 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
993 struct kmem_cache_node *n = get_node(s, node);
995 atomic_long_dec(&n->nr_slabs);
996 atomic_long_sub(objects, &n->total_objects);
999 /* Object debug checks for alloc/free paths */
1000 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1003 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1006 init_object(s, object, SLUB_RED_INACTIVE);
1007 init_tracking(s, object);
1010 static noinline int alloc_debug_processing(struct kmem_cache *s,
1012 void *object, unsigned long addr)
1014 if (!check_slab(s, page))
1017 if (!check_valid_pointer(s, page, object)) {
1018 object_err(s, page, object, "Freelist Pointer check fails");
1022 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1025 /* Success perform special debug activities for allocs */
1026 if (s->flags & SLAB_STORE_USER)
1027 set_track(s, object, TRACK_ALLOC, addr);
1028 trace(s, page, object, 1);
1029 init_object(s, object, SLUB_RED_ACTIVE);
1033 if (PageSlab(page)) {
1035 * If this is a slab page then lets do the best we can
1036 * to avoid issues in the future. Marking all objects
1037 * as used avoids touching the remaining objects.
1039 slab_fix(s, "Marking all objects used");
1040 page->inuse = page->objects;
1041 page->freelist = NULL;
1046 /* Supports checking bulk free of a constructed freelist */
1047 static noinline int free_debug_processing(
1048 struct kmem_cache *s, struct page *page,
1049 void *head, void *tail, int bulk_cnt,
1052 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1053 void *object = head;
1055 unsigned long uninitialized_var(flags);
1058 spin_lock_irqsave(&n->list_lock, flags);
1061 if (!check_slab(s, page))
1067 if (!check_valid_pointer(s, page, object)) {
1068 slab_err(s, page, "Invalid object pointer 0x%p", object);
1072 if (on_freelist(s, page, object)) {
1073 object_err(s, page, object, "Object already free");
1077 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1080 if (unlikely(s != page->slab_cache)) {
1081 if (!PageSlab(page)) {
1082 slab_err(s, page, "Attempt to free object(0x%p) "
1083 "outside of slab", object);
1084 } else if (!page->slab_cache) {
1085 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1089 object_err(s, page, object,
1090 "page slab pointer corrupt.");
1094 if (s->flags & SLAB_STORE_USER)
1095 set_track(s, object, TRACK_FREE, addr);
1096 trace(s, page, object, 0);
1097 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1098 init_object(s, object, SLUB_RED_INACTIVE);
1100 /* Reached end of constructed freelist yet? */
1101 if (object != tail) {
1102 object = get_freepointer(s, object);
1108 if (cnt != bulk_cnt)
1109 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1113 spin_unlock_irqrestore(&n->list_lock, flags);
1115 slab_fix(s, "Object at 0x%p not freed", object);
1119 static int __init setup_slub_debug(char *str)
1121 slub_debug = DEBUG_DEFAULT_FLAGS;
1122 if (*str++ != '=' || !*str)
1124 * No options specified. Switch on full debugging.
1130 * No options but restriction on slabs. This means full
1131 * debugging for slabs matching a pattern.
1138 * Switch off all debugging measures.
1143 * Determine which debug features should be switched on
1145 for (; *str && *str != ','; str++) {
1146 switch (tolower(*str)) {
1148 slub_debug |= SLAB_DEBUG_FREE;
1151 slub_debug |= SLAB_RED_ZONE;
1154 slub_debug |= SLAB_POISON;
1157 slub_debug |= SLAB_STORE_USER;
1160 slub_debug |= SLAB_TRACE;
1163 slub_debug |= SLAB_FAILSLAB;
1167 * Avoid enabling debugging on caches if its minimum
1168 * order would increase as a result.
1170 disable_higher_order_debug = 1;
1173 pr_err("slub_debug option '%c' unknown. skipped\n",
1180 slub_debug_slabs = str + 1;
1185 __setup("slub_debug", setup_slub_debug);
1187 unsigned long kmem_cache_flags(unsigned long object_size,
1188 unsigned long flags, const char *name,
1189 void (*ctor)(void *))
1192 * Enable debugging if selected on the kernel commandline.
1194 if (slub_debug && (!slub_debug_slabs || (name &&
1195 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1196 flags |= slub_debug;
1200 #else /* !CONFIG_SLUB_DEBUG */
1201 static inline void setup_object_debug(struct kmem_cache *s,
1202 struct page *page, void *object) {}
1204 static inline int alloc_debug_processing(struct kmem_cache *s,
1205 struct page *page, void *object, unsigned long addr) { return 0; }
1207 static inline int free_debug_processing(
1208 struct kmem_cache *s, struct page *page,
1209 void *head, void *tail, int bulk_cnt,
1210 unsigned long addr) { return 0; }
1212 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1214 static inline int check_object(struct kmem_cache *s, struct page *page,
1215 void *object, u8 val) { return 1; }
1216 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1217 struct page *page) {}
1218 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1219 struct page *page) {}
1220 unsigned long kmem_cache_flags(unsigned long object_size,
1221 unsigned long flags, const char *name,
1222 void (*ctor)(void *))
1226 #define slub_debug 0
1228 #define disable_higher_order_debug 0
1230 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1232 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1234 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1236 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1239 #endif /* CONFIG_SLUB_DEBUG */
1242 * Hooks for other subsystems that check memory allocations. In a typical
1243 * production configuration these hooks all should produce no code at all.
1245 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1247 kmemleak_alloc(ptr, size, 1, flags);
1248 kasan_kmalloc_large(ptr, size);
1251 static inline void kfree_hook(const void *x)
1254 kasan_kfree_large(x);
1257 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1259 kmemleak_free_recursive(x, s->flags);
1262 * Trouble is that we may no longer disable interrupts in the fast path
1263 * So in order to make the debug calls that expect irqs to be
1264 * disabled we need to disable interrupts temporarily.
1266 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1268 unsigned long flags;
1270 local_irq_save(flags);
1271 kmemcheck_slab_free(s, x, s->object_size);
1272 debug_check_no_locks_freed(x, s->object_size);
1273 local_irq_restore(flags);
1276 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1277 debug_check_no_obj_freed(x, s->object_size);
1279 kasan_slab_free(s, x);
1282 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1283 void *head, void *tail)
1286 * Compiler cannot detect this function can be removed if slab_free_hook()
1287 * evaluates to nothing. Thus, catch all relevant config debug options here.
1289 #if defined(CONFIG_KMEMCHECK) || \
1290 defined(CONFIG_LOCKDEP) || \
1291 defined(CONFIG_DEBUG_KMEMLEAK) || \
1292 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1293 defined(CONFIG_KASAN)
1295 void *object = head;
1296 void *tail_obj = tail ? : head;
1299 slab_free_hook(s, object);
1300 } while ((object != tail_obj) &&
1301 (object = get_freepointer(s, object)));
1305 static void setup_object(struct kmem_cache *s, struct page *page,
1308 setup_object_debug(s, page, object);
1309 if (unlikely(s->ctor)) {
1310 kasan_unpoison_object_data(s, object);
1312 kasan_poison_object_data(s, object);
1317 * Slab allocation and freeing
1319 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1320 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1323 int order = oo_order(oo);
1325 flags |= __GFP_NOTRACK;
1327 if (node == NUMA_NO_NODE)
1328 page = alloc_pages(flags, order);
1330 page = __alloc_pages_node(node, flags, order);
1332 if (page && memcg_charge_slab(page, flags, order, s)) {
1333 __free_pages(page, order);
1340 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1343 struct kmem_cache_order_objects oo = s->oo;
1348 flags &= gfp_allowed_mask;
1350 if (gfpflags_allow_blocking(flags))
1353 flags |= s->allocflags;
1356 * Let the initial higher-order allocation fail under memory pressure
1357 * so we fall-back to the minimum order allocation.
1359 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1360 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1361 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1363 page = alloc_slab_page(s, alloc_gfp, node, oo);
1364 if (unlikely(!page)) {
1368 * Allocation may have failed due to fragmentation.
1369 * Try a lower order alloc if possible
1371 page = alloc_slab_page(s, alloc_gfp, node, oo);
1372 if (unlikely(!page))
1374 stat(s, ORDER_FALLBACK);
1377 if (kmemcheck_enabled &&
1378 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1379 int pages = 1 << oo_order(oo);
1381 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1384 * Objects from caches that have a constructor don't get
1385 * cleared when they're allocated, so we need to do it here.
1388 kmemcheck_mark_uninitialized_pages(page, pages);
1390 kmemcheck_mark_unallocated_pages(page, pages);
1393 page->objects = oo_objects(oo);
1395 order = compound_order(page);
1396 page->slab_cache = s;
1397 __SetPageSlab(page);
1398 if (page_is_pfmemalloc(page))
1399 SetPageSlabPfmemalloc(page);
1401 start = page_address(page);
1403 if (unlikely(s->flags & SLAB_POISON))
1404 memset(start, POISON_INUSE, PAGE_SIZE << order);
1406 kasan_poison_slab(page);
1408 for_each_object_idx(p, idx, s, start, page->objects) {
1409 setup_object(s, page, p);
1410 if (likely(idx < page->objects))
1411 set_freepointer(s, p, p + s->size);
1413 set_freepointer(s, p, NULL);
1416 page->freelist = start;
1417 page->inuse = page->objects;
1421 if (gfpflags_allow_blocking(flags))
1422 local_irq_disable();
1426 mod_zone_page_state(page_zone(page),
1427 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1428 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1431 inc_slabs_node(s, page_to_nid(page), page->objects);
1436 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1438 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1439 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1443 return allocate_slab(s,
1444 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1447 static void __free_slab(struct kmem_cache *s, struct page *page)
1449 int order = compound_order(page);
1450 int pages = 1 << order;
1452 if (kmem_cache_debug(s)) {
1455 slab_pad_check(s, page);
1456 for_each_object(p, s, page_address(page),
1458 check_object(s, page, p, SLUB_RED_INACTIVE);
1461 kmemcheck_free_shadow(page, compound_order(page));
1463 mod_zone_page_state(page_zone(page),
1464 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1465 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1468 __ClearPageSlabPfmemalloc(page);
1469 __ClearPageSlab(page);
1471 page_mapcount_reset(page);
1472 if (current->reclaim_state)
1473 current->reclaim_state->reclaimed_slab += pages;
1474 __free_kmem_pages(page, order);
1477 #define need_reserve_slab_rcu \
1478 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1480 static void rcu_free_slab(struct rcu_head *h)
1484 if (need_reserve_slab_rcu)
1485 page = virt_to_head_page(h);
1487 page = container_of((struct list_head *)h, struct page, lru);
1489 __free_slab(page->slab_cache, page);
1492 static void free_slab(struct kmem_cache *s, struct page *page)
1494 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1495 struct rcu_head *head;
1497 if (need_reserve_slab_rcu) {
1498 int order = compound_order(page);
1499 int offset = (PAGE_SIZE << order) - s->reserved;
1501 VM_BUG_ON(s->reserved != sizeof(*head));
1502 head = page_address(page) + offset;
1504 head = &page->rcu_head;
1507 call_rcu(head, rcu_free_slab);
1509 __free_slab(s, page);
1512 static void discard_slab(struct kmem_cache *s, struct page *page)
1514 dec_slabs_node(s, page_to_nid(page), page->objects);
1519 * Management of partially allocated slabs.
1522 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1525 if (tail == DEACTIVATE_TO_TAIL)
1526 list_add_tail(&page->lru, &n->partial);
1528 list_add(&page->lru, &n->partial);
1531 static inline void add_partial(struct kmem_cache_node *n,
1532 struct page *page, int tail)
1534 lockdep_assert_held(&n->list_lock);
1535 __add_partial(n, page, tail);
1538 static inline void remove_partial(struct kmem_cache_node *n,
1541 lockdep_assert_held(&n->list_lock);
1542 list_del(&page->lru);
1547 * Remove slab from the partial list, freeze it and
1548 * return the pointer to the freelist.
1550 * Returns a list of objects or NULL if it fails.
1552 static inline void *acquire_slab(struct kmem_cache *s,
1553 struct kmem_cache_node *n, struct page *page,
1554 int mode, int *objects)
1557 unsigned long counters;
1560 lockdep_assert_held(&n->list_lock);
1563 * Zap the freelist and set the frozen bit.
1564 * The old freelist is the list of objects for the
1565 * per cpu allocation list.
1567 freelist = page->freelist;
1568 counters = page->counters;
1569 new.counters = counters;
1570 *objects = new.objects - new.inuse;
1572 new.inuse = page->objects;
1573 new.freelist = NULL;
1575 new.freelist = freelist;
1578 VM_BUG_ON(new.frozen);
1581 if (!__cmpxchg_double_slab(s, page,
1583 new.freelist, new.counters,
1587 remove_partial(n, page);
1592 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1593 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1596 * Try to allocate a partial slab from a specific node.
1598 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1599 struct kmem_cache_cpu *c, gfp_t flags)
1601 struct page *page, *page2;
1602 void *object = NULL;
1607 * Racy check. If we mistakenly see no partial slabs then we
1608 * just allocate an empty slab. If we mistakenly try to get a
1609 * partial slab and there is none available then get_partials()
1612 if (!n || !n->nr_partial)
1615 spin_lock(&n->list_lock);
1616 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1619 if (!pfmemalloc_match(page, flags))
1622 t = acquire_slab(s, n, page, object == NULL, &objects);
1626 available += objects;
1629 stat(s, ALLOC_FROM_PARTIAL);
1632 put_cpu_partial(s, page, 0);
1633 stat(s, CPU_PARTIAL_NODE);
1635 if (!kmem_cache_has_cpu_partial(s)
1636 || available > s->cpu_partial / 2)
1640 spin_unlock(&n->list_lock);
1645 * Get a page from somewhere. Search in increasing NUMA distances.
1647 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1648 struct kmem_cache_cpu *c)
1651 struct zonelist *zonelist;
1654 enum zone_type high_zoneidx = gfp_zone(flags);
1656 unsigned int cpuset_mems_cookie;
1659 * The defrag ratio allows a configuration of the tradeoffs between
1660 * inter node defragmentation and node local allocations. A lower
1661 * defrag_ratio increases the tendency to do local allocations
1662 * instead of attempting to obtain partial slabs from other nodes.
1664 * If the defrag_ratio is set to 0 then kmalloc() always
1665 * returns node local objects. If the ratio is higher then kmalloc()
1666 * may return off node objects because partial slabs are obtained
1667 * from other nodes and filled up.
1669 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1670 * defrag_ratio = 1000) then every (well almost) allocation will
1671 * first attempt to defrag slab caches on other nodes. This means
1672 * scanning over all nodes to look for partial slabs which may be
1673 * expensive if we do it every time we are trying to find a slab
1674 * with available objects.
1676 if (!s->remote_node_defrag_ratio ||
1677 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1681 cpuset_mems_cookie = read_mems_allowed_begin();
1682 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1683 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1684 struct kmem_cache_node *n;
1686 n = get_node(s, zone_to_nid(zone));
1688 if (n && cpuset_zone_allowed(zone, flags) &&
1689 n->nr_partial > s->min_partial) {
1690 object = get_partial_node(s, n, c, flags);
1693 * Don't check read_mems_allowed_retry()
1694 * here - if mems_allowed was updated in
1695 * parallel, that was a harmless race
1696 * between allocation and the cpuset
1703 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1709 * Get a partial page, lock it and return it.
1711 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1712 struct kmem_cache_cpu *c)
1715 int searchnode = node;
1717 if (node == NUMA_NO_NODE)
1718 searchnode = numa_mem_id();
1719 else if (!node_present_pages(node))
1720 searchnode = node_to_mem_node(node);
1722 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1723 if (object || node != NUMA_NO_NODE)
1726 return get_any_partial(s, flags, c);
1729 #ifdef CONFIG_PREEMPT
1731 * Calculate the next globally unique transaction for disambiguiation
1732 * during cmpxchg. The transactions start with the cpu number and are then
1733 * incremented by CONFIG_NR_CPUS.
1735 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1738 * No preemption supported therefore also no need to check for
1744 static inline unsigned long next_tid(unsigned long tid)
1746 return tid + TID_STEP;
1749 static inline unsigned int tid_to_cpu(unsigned long tid)
1751 return tid % TID_STEP;
1754 static inline unsigned long tid_to_event(unsigned long tid)
1756 return tid / TID_STEP;
1759 static inline unsigned int init_tid(int cpu)
1764 static inline void note_cmpxchg_failure(const char *n,
1765 const struct kmem_cache *s, unsigned long tid)
1767 #ifdef SLUB_DEBUG_CMPXCHG
1768 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1770 pr_info("%s %s: cmpxchg redo ", n, s->name);
1772 #ifdef CONFIG_PREEMPT
1773 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1774 pr_warn("due to cpu change %d -> %d\n",
1775 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1778 if (tid_to_event(tid) != tid_to_event(actual_tid))
1779 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1780 tid_to_event(tid), tid_to_event(actual_tid));
1782 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1783 actual_tid, tid, next_tid(tid));
1785 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1788 static void init_kmem_cache_cpus(struct kmem_cache *s)
1792 for_each_possible_cpu(cpu)
1793 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1797 * Remove the cpu slab
1799 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1802 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1803 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1805 enum slab_modes l = M_NONE, m = M_NONE;
1807 int tail = DEACTIVATE_TO_HEAD;
1811 if (page->freelist) {
1812 stat(s, DEACTIVATE_REMOTE_FREES);
1813 tail = DEACTIVATE_TO_TAIL;
1817 * Stage one: Free all available per cpu objects back
1818 * to the page freelist while it is still frozen. Leave the
1821 * There is no need to take the list->lock because the page
1824 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1826 unsigned long counters;
1829 prior = page->freelist;
1830 counters = page->counters;
1831 set_freepointer(s, freelist, prior);
1832 new.counters = counters;
1834 VM_BUG_ON(!new.frozen);
1836 } while (!__cmpxchg_double_slab(s, page,
1838 freelist, new.counters,
1839 "drain percpu freelist"));
1841 freelist = nextfree;
1845 * Stage two: Ensure that the page is unfrozen while the
1846 * list presence reflects the actual number of objects
1849 * We setup the list membership and then perform a cmpxchg
1850 * with the count. If there is a mismatch then the page
1851 * is not unfrozen but the page is on the wrong list.
1853 * Then we restart the process which may have to remove
1854 * the page from the list that we just put it on again
1855 * because the number of objects in the slab may have
1860 old.freelist = page->freelist;
1861 old.counters = page->counters;
1862 VM_BUG_ON(!old.frozen);
1864 /* Determine target state of the slab */
1865 new.counters = old.counters;
1868 set_freepointer(s, freelist, old.freelist);
1869 new.freelist = freelist;
1871 new.freelist = old.freelist;
1875 if (!new.inuse && n->nr_partial >= s->min_partial)
1877 else if (new.freelist) {
1882 * Taking the spinlock removes the possiblity
1883 * that acquire_slab() will see a slab page that
1886 spin_lock(&n->list_lock);
1890 if (kmem_cache_debug(s) && !lock) {
1893 * This also ensures that the scanning of full
1894 * slabs from diagnostic functions will not see
1897 spin_lock(&n->list_lock);
1905 remove_partial(n, page);
1907 else if (l == M_FULL)
1909 remove_full(s, n, page);
1911 if (m == M_PARTIAL) {
1913 add_partial(n, page, tail);
1916 } else if (m == M_FULL) {
1918 stat(s, DEACTIVATE_FULL);
1919 add_full(s, n, page);
1925 if (!__cmpxchg_double_slab(s, page,
1926 old.freelist, old.counters,
1927 new.freelist, new.counters,
1932 spin_unlock(&n->list_lock);
1935 stat(s, DEACTIVATE_EMPTY);
1936 discard_slab(s, page);
1942 * Unfreeze all the cpu partial slabs.
1944 * This function must be called with interrupts disabled
1945 * for the cpu using c (or some other guarantee must be there
1946 * to guarantee no concurrent accesses).
1948 static void unfreeze_partials(struct kmem_cache *s,
1949 struct kmem_cache_cpu *c)
1951 #ifdef CONFIG_SLUB_CPU_PARTIAL
1952 struct kmem_cache_node *n = NULL, *n2 = NULL;
1953 struct page *page, *discard_page = NULL;
1955 while ((page = c->partial)) {
1959 c->partial = page->next;
1961 n2 = get_node(s, page_to_nid(page));
1964 spin_unlock(&n->list_lock);
1967 spin_lock(&n->list_lock);
1972 old.freelist = page->freelist;
1973 old.counters = page->counters;
1974 VM_BUG_ON(!old.frozen);
1976 new.counters = old.counters;
1977 new.freelist = old.freelist;
1981 } while (!__cmpxchg_double_slab(s, page,
1982 old.freelist, old.counters,
1983 new.freelist, new.counters,
1984 "unfreezing slab"));
1986 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1987 page->next = discard_page;
1988 discard_page = page;
1990 add_partial(n, page, DEACTIVATE_TO_TAIL);
1991 stat(s, FREE_ADD_PARTIAL);
1996 spin_unlock(&n->list_lock);
1998 while (discard_page) {
1999 page = discard_page;
2000 discard_page = discard_page->next;
2002 stat(s, DEACTIVATE_EMPTY);
2003 discard_slab(s, page);
2010 * Put a page that was just frozen (in __slab_free) into a partial page
2011 * slot if available. This is done without interrupts disabled and without
2012 * preemption disabled. The cmpxchg is racy and may put the partial page
2013 * onto a random cpus partial slot.
2015 * If we did not find a slot then simply move all the partials to the
2016 * per node partial list.
2018 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2020 #ifdef CONFIG_SLUB_CPU_PARTIAL
2021 struct page *oldpage;
2029 oldpage = this_cpu_read(s->cpu_slab->partial);
2032 pobjects = oldpage->pobjects;
2033 pages = oldpage->pages;
2034 if (drain && pobjects > s->cpu_partial) {
2035 unsigned long flags;
2037 * partial array is full. Move the existing
2038 * set to the per node partial list.
2040 local_irq_save(flags);
2041 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2042 local_irq_restore(flags);
2046 stat(s, CPU_PARTIAL_DRAIN);
2051 pobjects += page->objects - page->inuse;
2053 page->pages = pages;
2054 page->pobjects = pobjects;
2055 page->next = oldpage;
2057 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2059 if (unlikely(!s->cpu_partial)) {
2060 unsigned long flags;
2062 local_irq_save(flags);
2063 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2064 local_irq_restore(flags);
2070 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2072 stat(s, CPUSLAB_FLUSH);
2073 deactivate_slab(s, c->page, c->freelist);
2075 c->tid = next_tid(c->tid);
2083 * Called from IPI handler with interrupts disabled.
2085 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2087 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2093 unfreeze_partials(s, c);
2097 static void flush_cpu_slab(void *d)
2099 struct kmem_cache *s = d;
2101 __flush_cpu_slab(s, smp_processor_id());
2104 static bool has_cpu_slab(int cpu, void *info)
2106 struct kmem_cache *s = info;
2107 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2109 return c->page || c->partial;
2112 static void flush_all(struct kmem_cache *s)
2114 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2118 * Check if the objects in a per cpu structure fit numa
2119 * locality expectations.
2121 static inline int node_match(struct page *page, int node)
2124 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2130 #ifdef CONFIG_SLUB_DEBUG
2131 static int count_free(struct page *page)
2133 return page->objects - page->inuse;
2136 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2138 return atomic_long_read(&n->total_objects);
2140 #endif /* CONFIG_SLUB_DEBUG */
2142 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2143 static unsigned long count_partial(struct kmem_cache_node *n,
2144 int (*get_count)(struct page *))
2146 unsigned long flags;
2147 unsigned long x = 0;
2150 spin_lock_irqsave(&n->list_lock, flags);
2151 list_for_each_entry(page, &n->partial, lru)
2152 x += get_count(page);
2153 spin_unlock_irqrestore(&n->list_lock, flags);
2156 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2158 static noinline void
2159 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2161 #ifdef CONFIG_SLUB_DEBUG
2162 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2163 DEFAULT_RATELIMIT_BURST);
2165 struct kmem_cache_node *n;
2167 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2170 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2172 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2173 s->name, s->object_size, s->size, oo_order(s->oo),
2176 if (oo_order(s->min) > get_order(s->object_size))
2177 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2180 for_each_kmem_cache_node(s, node, n) {
2181 unsigned long nr_slabs;
2182 unsigned long nr_objs;
2183 unsigned long nr_free;
2185 nr_free = count_partial(n, count_free);
2186 nr_slabs = node_nr_slabs(n);
2187 nr_objs = node_nr_objs(n);
2189 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2190 node, nr_slabs, nr_objs, nr_free);
2195 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2196 int node, struct kmem_cache_cpu **pc)
2199 struct kmem_cache_cpu *c = *pc;
2202 freelist = get_partial(s, flags, node, c);
2207 page = new_slab(s, flags, node);
2209 c = raw_cpu_ptr(s->cpu_slab);
2214 * No other reference to the page yet so we can
2215 * muck around with it freely without cmpxchg
2217 freelist = page->freelist;
2218 page->freelist = NULL;
2220 stat(s, ALLOC_SLAB);
2229 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2231 if (unlikely(PageSlabPfmemalloc(page)))
2232 return gfp_pfmemalloc_allowed(gfpflags);
2238 * Check the page->freelist of a page and either transfer the freelist to the
2239 * per cpu freelist or deactivate the page.
2241 * The page is still frozen if the return value is not NULL.
2243 * If this function returns NULL then the page has been unfrozen.
2245 * This function must be called with interrupt disabled.
2247 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2250 unsigned long counters;
2254 freelist = page->freelist;
2255 counters = page->counters;
2257 new.counters = counters;
2258 VM_BUG_ON(!new.frozen);
2260 new.inuse = page->objects;
2261 new.frozen = freelist != NULL;
2263 } while (!__cmpxchg_double_slab(s, page,
2272 * Slow path. The lockless freelist is empty or we need to perform
2275 * Processing is still very fast if new objects have been freed to the
2276 * regular freelist. In that case we simply take over the regular freelist
2277 * as the lockless freelist and zap the regular freelist.
2279 * If that is not working then we fall back to the partial lists. We take the
2280 * first element of the freelist as the object to allocate now and move the
2281 * rest of the freelist to the lockless freelist.
2283 * And if we were unable to get a new slab from the partial slab lists then
2284 * we need to allocate a new slab. This is the slowest path since it involves
2285 * a call to the page allocator and the setup of a new slab.
2287 * Version of __slab_alloc to use when we know that interrupts are
2288 * already disabled (which is the case for bulk allocation).
2290 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2291 unsigned long addr, struct kmem_cache_cpu *c)
2301 if (unlikely(!node_match(page, node))) {
2302 int searchnode = node;
2304 if (node != NUMA_NO_NODE && !node_present_pages(node))
2305 searchnode = node_to_mem_node(node);
2307 if (unlikely(!node_match(page, searchnode))) {
2308 stat(s, ALLOC_NODE_MISMATCH);
2309 deactivate_slab(s, page, c->freelist);
2317 * By rights, we should be searching for a slab page that was
2318 * PFMEMALLOC but right now, we are losing the pfmemalloc
2319 * information when the page leaves the per-cpu allocator
2321 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2322 deactivate_slab(s, page, c->freelist);
2328 /* must check again c->freelist in case of cpu migration or IRQ */
2329 freelist = c->freelist;
2333 freelist = get_freelist(s, page);
2337 stat(s, DEACTIVATE_BYPASS);
2341 stat(s, ALLOC_REFILL);
2345 * freelist is pointing to the list of objects to be used.
2346 * page is pointing to the page from which the objects are obtained.
2347 * That page must be frozen for per cpu allocations to work.
2349 VM_BUG_ON(!c->page->frozen);
2350 c->freelist = get_freepointer(s, freelist);
2351 c->tid = next_tid(c->tid);
2357 page = c->page = c->partial;
2358 c->partial = page->next;
2359 stat(s, CPU_PARTIAL_ALLOC);
2364 freelist = new_slab_objects(s, gfpflags, node, &c);
2366 if (unlikely(!freelist)) {
2367 slab_out_of_memory(s, gfpflags, node);
2372 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2375 /* Only entered in the debug case */
2376 if (kmem_cache_debug(s) &&
2377 !alloc_debug_processing(s, page, freelist, addr))
2378 goto new_slab; /* Slab failed checks. Next slab needed */
2380 deactivate_slab(s, page, get_freepointer(s, freelist));
2387 * Another one that disabled interrupt and compensates for possible
2388 * cpu changes by refetching the per cpu area pointer.
2390 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2391 unsigned long addr, struct kmem_cache_cpu *c)
2394 unsigned long flags;
2396 local_irq_save(flags);
2397 #ifdef CONFIG_PREEMPT
2399 * We may have been preempted and rescheduled on a different
2400 * cpu before disabling interrupts. Need to reload cpu area
2403 c = this_cpu_ptr(s->cpu_slab);
2406 p = ___slab_alloc(s, gfpflags, node, addr, c);
2407 local_irq_restore(flags);
2412 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2413 * have the fastpath folded into their functions. So no function call
2414 * overhead for requests that can be satisfied on the fastpath.
2416 * The fastpath works by first checking if the lockless freelist can be used.
2417 * If not then __slab_alloc is called for slow processing.
2419 * Otherwise we can simply pick the next object from the lockless free list.
2421 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2422 gfp_t gfpflags, int node, unsigned long addr)
2425 struct kmem_cache_cpu *c;
2429 s = slab_pre_alloc_hook(s, gfpflags);
2434 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2435 * enabled. We may switch back and forth between cpus while
2436 * reading from one cpu area. That does not matter as long
2437 * as we end up on the original cpu again when doing the cmpxchg.
2439 * We should guarantee that tid and kmem_cache are retrieved on
2440 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2441 * to check if it is matched or not.
2444 tid = this_cpu_read(s->cpu_slab->tid);
2445 c = raw_cpu_ptr(s->cpu_slab);
2446 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2447 unlikely(tid != READ_ONCE(c->tid)));
2450 * Irqless object alloc/free algorithm used here depends on sequence
2451 * of fetching cpu_slab's data. tid should be fetched before anything
2452 * on c to guarantee that object and page associated with previous tid
2453 * won't be used with current tid. If we fetch tid first, object and
2454 * page could be one associated with next tid and our alloc/free
2455 * request will be failed. In this case, we will retry. So, no problem.
2460 * The transaction ids are globally unique per cpu and per operation on
2461 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2462 * occurs on the right processor and that there was no operation on the
2463 * linked list in between.
2466 object = c->freelist;
2468 if (unlikely(!object || !node_match(page, node))) {
2469 object = __slab_alloc(s, gfpflags, node, addr, c);
2470 stat(s, ALLOC_SLOWPATH);
2472 void *next_object = get_freepointer_safe(s, object);
2475 * The cmpxchg will only match if there was no additional
2476 * operation and if we are on the right processor.
2478 * The cmpxchg does the following atomically (without lock
2480 * 1. Relocate first pointer to the current per cpu area.
2481 * 2. Verify that tid and freelist have not been changed
2482 * 3. If they were not changed replace tid and freelist
2484 * Since this is without lock semantics the protection is only
2485 * against code executing on this cpu *not* from access by
2488 if (unlikely(!this_cpu_cmpxchg_double(
2489 s->cpu_slab->freelist, s->cpu_slab->tid,
2491 next_object, next_tid(tid)))) {
2493 note_cmpxchg_failure("slab_alloc", s, tid);
2496 prefetch_freepointer(s, next_object);
2497 stat(s, ALLOC_FASTPATH);
2500 if (unlikely(gfpflags & __GFP_ZERO) && object)
2501 memset(object, 0, s->object_size);
2503 slab_post_alloc_hook(s, gfpflags, 1, &object);
2508 static __always_inline void *slab_alloc(struct kmem_cache *s,
2509 gfp_t gfpflags, unsigned long addr)
2511 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2514 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2516 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2518 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2523 EXPORT_SYMBOL(kmem_cache_alloc);
2525 #ifdef CONFIG_TRACING
2526 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2528 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2529 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2530 kasan_kmalloc(s, ret, size);
2533 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2537 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2539 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2541 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2542 s->object_size, s->size, gfpflags, node);
2546 EXPORT_SYMBOL(kmem_cache_alloc_node);
2548 #ifdef CONFIG_TRACING
2549 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2551 int node, size_t size)
2553 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2555 trace_kmalloc_node(_RET_IP_, ret,
2556 size, s->size, gfpflags, node);
2558 kasan_kmalloc(s, ret, size);
2561 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2566 * Slow path handling. This may still be called frequently since objects
2567 * have a longer lifetime than the cpu slabs in most processing loads.
2569 * So we still attempt to reduce cache line usage. Just take the slab
2570 * lock and free the item. If there is no additional partial page
2571 * handling required then we can return immediately.
2573 static void __slab_free(struct kmem_cache *s, struct page *page,
2574 void *head, void *tail, int cnt,
2581 unsigned long counters;
2582 struct kmem_cache_node *n = NULL;
2583 unsigned long uninitialized_var(flags);
2585 stat(s, FREE_SLOWPATH);
2587 if (kmem_cache_debug(s) &&
2588 !free_debug_processing(s, page, head, tail, cnt, addr))
2593 spin_unlock_irqrestore(&n->list_lock, flags);
2596 prior = page->freelist;
2597 counters = page->counters;
2598 set_freepointer(s, tail, prior);
2599 new.counters = counters;
2600 was_frozen = new.frozen;
2602 if ((!new.inuse || !prior) && !was_frozen) {
2604 if (kmem_cache_has_cpu_partial(s) && !prior) {
2607 * Slab was on no list before and will be
2609 * We can defer the list move and instead
2614 } else { /* Needs to be taken off a list */
2616 n = get_node(s, page_to_nid(page));
2618 * Speculatively acquire the list_lock.
2619 * If the cmpxchg does not succeed then we may
2620 * drop the list_lock without any processing.
2622 * Otherwise the list_lock will synchronize with
2623 * other processors updating the list of slabs.
2625 spin_lock_irqsave(&n->list_lock, flags);
2630 } while (!cmpxchg_double_slab(s, page,
2638 * If we just froze the page then put it onto the
2639 * per cpu partial list.
2641 if (new.frozen && !was_frozen) {
2642 put_cpu_partial(s, page, 1);
2643 stat(s, CPU_PARTIAL_FREE);
2646 * The list lock was not taken therefore no list
2647 * activity can be necessary.
2650 stat(s, FREE_FROZEN);
2654 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2658 * Objects left in the slab. If it was not on the partial list before
2661 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2662 if (kmem_cache_debug(s))
2663 remove_full(s, n, page);
2664 add_partial(n, page, DEACTIVATE_TO_TAIL);
2665 stat(s, FREE_ADD_PARTIAL);
2667 spin_unlock_irqrestore(&n->list_lock, flags);
2673 * Slab on the partial list.
2675 remove_partial(n, page);
2676 stat(s, FREE_REMOVE_PARTIAL);
2678 /* Slab must be on the full list */
2679 remove_full(s, n, page);
2682 spin_unlock_irqrestore(&n->list_lock, flags);
2684 discard_slab(s, page);
2688 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2689 * can perform fastpath freeing without additional function calls.
2691 * The fastpath is only possible if we are freeing to the current cpu slab
2692 * of this processor. This typically the case if we have just allocated
2695 * If fastpath is not possible then fall back to __slab_free where we deal
2696 * with all sorts of special processing.
2698 * Bulk free of a freelist with several objects (all pointing to the
2699 * same page) possible by specifying head and tail ptr, plus objects
2700 * count (cnt). Bulk free indicated by tail pointer being set.
2702 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2703 void *head, void *tail, int cnt,
2706 void *tail_obj = tail ? : head;
2707 struct kmem_cache_cpu *c;
2710 slab_free_freelist_hook(s, head, tail);
2714 * Determine the currently cpus per cpu slab.
2715 * The cpu may change afterward. However that does not matter since
2716 * data is retrieved via this pointer. If we are on the same cpu
2717 * during the cmpxchg then the free will succeed.
2720 tid = this_cpu_read(s->cpu_slab->tid);
2721 c = raw_cpu_ptr(s->cpu_slab);
2722 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2723 unlikely(tid != READ_ONCE(c->tid)));
2725 /* Same with comment on barrier() in slab_alloc_node() */
2728 if (likely(page == c->page)) {
2729 set_freepointer(s, tail_obj, c->freelist);
2731 if (unlikely(!this_cpu_cmpxchg_double(
2732 s->cpu_slab->freelist, s->cpu_slab->tid,
2734 head, next_tid(tid)))) {
2736 note_cmpxchg_failure("slab_free", s, tid);
2739 stat(s, FREE_FASTPATH);
2741 __slab_free(s, page, head, tail_obj, cnt, addr);
2745 void kmem_cache_free(struct kmem_cache *s, void *x)
2747 s = cache_from_obj(s, x);
2750 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2751 trace_kmem_cache_free(_RET_IP_, x);
2753 EXPORT_SYMBOL(kmem_cache_free);
2755 struct detached_freelist {
2760 struct kmem_cache *s;
2764 * This function progressively scans the array with free objects (with
2765 * a limited look ahead) and extract objects belonging to the same
2766 * page. It builds a detached freelist directly within the given
2767 * page/objects. This can happen without any need for
2768 * synchronization, because the objects are owned by running process.
2769 * The freelist is build up as a single linked list in the objects.
2770 * The idea is, that this detached freelist can then be bulk
2771 * transferred to the real freelist(s), but only requiring a single
2772 * synchronization primitive. Look ahead in the array is limited due
2773 * to performance reasons.
2776 int build_detached_freelist(struct kmem_cache *s, size_t size,
2777 void **p, struct detached_freelist *df)
2779 size_t first_skipped_index = 0;
2784 /* Always re-init detached_freelist */
2789 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2790 } while (!object && size);
2795 page = virt_to_head_page(object);
2797 /* Handle kalloc'ed objects */
2798 if (unlikely(!PageSlab(page))) {
2799 BUG_ON(!PageCompound(page));
2801 __free_kmem_pages(page, compound_order(page));
2802 p[size] = NULL; /* mark object processed */
2805 /* Derive kmem_cache from object */
2806 df->s = page->slab_cache;
2808 df->s = cache_from_obj(s, object); /* Support for memcg */
2811 /* Start new detached freelist */
2813 set_freepointer(df->s, object, NULL);
2815 df->freelist = object;
2816 p[size] = NULL; /* mark object processed */
2822 continue; /* Skip processed objects */
2824 /* df->page is always set at this point */
2825 if (df->page == virt_to_head_page(object)) {
2826 /* Opportunity build freelist */
2827 set_freepointer(df->s, object, df->freelist);
2828 df->freelist = object;
2830 p[size] = NULL; /* mark object processed */
2835 /* Limit look ahead search */
2839 if (!first_skipped_index)
2840 first_skipped_index = size + 1;
2843 return first_skipped_index;
2846 /* Note that interrupts must be enabled when calling this function. */
2847 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2853 struct detached_freelist df;
2855 size = build_detached_freelist(s, size, p, &df);
2856 if (unlikely(!df.page))
2859 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2860 } while (likely(size));
2862 EXPORT_SYMBOL(kmem_cache_free_bulk);
2864 /* Note that interrupts must be enabled when calling this function. */
2865 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2868 struct kmem_cache_cpu *c;
2871 /* memcg and kmem_cache debug support */
2872 s = slab_pre_alloc_hook(s, flags);
2876 * Drain objects in the per cpu slab, while disabling local
2877 * IRQs, which protects against PREEMPT and interrupts
2878 * handlers invoking normal fastpath.
2880 local_irq_disable();
2881 c = this_cpu_ptr(s->cpu_slab);
2883 for (i = 0; i < size; i++) {
2884 void *object = c->freelist;
2886 if (unlikely(!object)) {
2888 * Invoking slow path likely have side-effect
2889 * of re-populating per CPU c->freelist
2891 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2893 if (unlikely(!p[i]))
2896 c = this_cpu_ptr(s->cpu_slab);
2897 continue; /* goto for-loop */
2899 c->freelist = get_freepointer(s, object);
2902 c->tid = next_tid(c->tid);
2905 /* Clear memory outside IRQ disabled fastpath loop */
2906 if (unlikely(flags & __GFP_ZERO)) {
2909 for (j = 0; j < i; j++)
2910 memset(p[j], 0, s->object_size);
2913 /* memcg and kmem_cache debug support */
2914 slab_post_alloc_hook(s, flags, size, p);
2918 slab_post_alloc_hook(s, flags, i, p);
2919 __kmem_cache_free_bulk(s, i, p);
2922 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2926 * Object placement in a slab is made very easy because we always start at
2927 * offset 0. If we tune the size of the object to the alignment then we can
2928 * get the required alignment by putting one properly sized object after
2931 * Notice that the allocation order determines the sizes of the per cpu
2932 * caches. Each processor has always one slab available for allocations.
2933 * Increasing the allocation order reduces the number of times that slabs
2934 * must be moved on and off the partial lists and is therefore a factor in
2939 * Mininum / Maximum order of slab pages. This influences locking overhead
2940 * and slab fragmentation. A higher order reduces the number of partial slabs
2941 * and increases the number of allocations possible without having to
2942 * take the list_lock.
2944 static int slub_min_order;
2945 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2946 static int slub_min_objects;
2949 * Calculate the order of allocation given an slab object size.
2951 * The order of allocation has significant impact on performance and other
2952 * system components. Generally order 0 allocations should be preferred since
2953 * order 0 does not cause fragmentation in the page allocator. Larger objects
2954 * be problematic to put into order 0 slabs because there may be too much
2955 * unused space left. We go to a higher order if more than 1/16th of the slab
2958 * In order to reach satisfactory performance we must ensure that a minimum
2959 * number of objects is in one slab. Otherwise we may generate too much
2960 * activity on the partial lists which requires taking the list_lock. This is
2961 * less a concern for large slabs though which are rarely used.
2963 * slub_max_order specifies the order where we begin to stop considering the
2964 * number of objects in a slab as critical. If we reach slub_max_order then
2965 * we try to keep the page order as low as possible. So we accept more waste
2966 * of space in favor of a small page order.
2968 * Higher order allocations also allow the placement of more objects in a
2969 * slab and thereby reduce object handling overhead. If the user has
2970 * requested a higher mininum order then we start with that one instead of
2971 * the smallest order which will fit the object.
2973 static inline int slab_order(int size, int min_objects,
2974 int max_order, int fract_leftover, int reserved)
2978 int min_order = slub_min_order;
2980 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2981 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2983 for (order = max(min_order, get_order(min_objects * size + reserved));
2984 order <= max_order; order++) {
2986 unsigned long slab_size = PAGE_SIZE << order;
2988 rem = (slab_size - reserved) % size;
2990 if (rem <= slab_size / fract_leftover)
2997 static inline int calculate_order(int size, int reserved)
3005 * Attempt to find best configuration for a slab. This
3006 * works by first attempting to generate a layout with
3007 * the best configuration and backing off gradually.
3009 * First we increase the acceptable waste in a slab. Then
3010 * we reduce the minimum objects required in a slab.
3012 min_objects = slub_min_objects;
3014 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3015 max_objects = order_objects(slub_max_order, size, reserved);
3016 min_objects = min(min_objects, max_objects);
3018 while (min_objects > 1) {
3020 while (fraction >= 4) {
3021 order = slab_order(size, min_objects,
3022 slub_max_order, fraction, reserved);
3023 if (order <= slub_max_order)
3031 * We were unable to place multiple objects in a slab. Now
3032 * lets see if we can place a single object there.
3034 order = slab_order(size, 1, slub_max_order, 1, reserved);
3035 if (order <= slub_max_order)
3039 * Doh this slab cannot be placed using slub_max_order.
3041 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3042 if (order < MAX_ORDER)
3048 init_kmem_cache_node(struct kmem_cache_node *n)
3051 spin_lock_init(&n->list_lock);
3052 INIT_LIST_HEAD(&n->partial);
3053 #ifdef CONFIG_SLUB_DEBUG
3054 atomic_long_set(&n->nr_slabs, 0);
3055 atomic_long_set(&n->total_objects, 0);
3056 INIT_LIST_HEAD(&n->full);
3060 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3062 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3063 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3066 * Must align to double word boundary for the double cmpxchg
3067 * instructions to work; see __pcpu_double_call_return_bool().
3069 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3070 2 * sizeof(void *));
3075 init_kmem_cache_cpus(s);
3080 static struct kmem_cache *kmem_cache_node;
3083 * No kmalloc_node yet so do it by hand. We know that this is the first
3084 * slab on the node for this slabcache. There are no concurrent accesses
3087 * Note that this function only works on the kmem_cache_node
3088 * when allocating for the kmem_cache_node. This is used for bootstrapping
3089 * memory on a fresh node that has no slab structures yet.
3091 static void early_kmem_cache_node_alloc(int node)
3094 struct kmem_cache_node *n;
3096 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3098 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3101 if (page_to_nid(page) != node) {
3102 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3103 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3108 page->freelist = get_freepointer(kmem_cache_node, n);
3111 kmem_cache_node->node[node] = n;
3112 #ifdef CONFIG_SLUB_DEBUG
3113 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3114 init_tracking(kmem_cache_node, n);
3116 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3117 init_kmem_cache_node(n);
3118 inc_slabs_node(kmem_cache_node, node, page->objects);
3121 * No locks need to be taken here as it has just been
3122 * initialized and there is no concurrent access.
3124 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3127 static void free_kmem_cache_nodes(struct kmem_cache *s)
3130 struct kmem_cache_node *n;
3132 for_each_kmem_cache_node(s, node, n) {
3133 kmem_cache_free(kmem_cache_node, n);
3134 s->node[node] = NULL;
3138 void __kmem_cache_release(struct kmem_cache *s)
3140 free_percpu(s->cpu_slab);
3141 free_kmem_cache_nodes(s);
3144 static int init_kmem_cache_nodes(struct kmem_cache *s)
3148 for_each_node_state(node, N_NORMAL_MEMORY) {
3149 struct kmem_cache_node *n;
3151 if (slab_state == DOWN) {
3152 early_kmem_cache_node_alloc(node);
3155 n = kmem_cache_alloc_node(kmem_cache_node,
3159 free_kmem_cache_nodes(s);
3164 init_kmem_cache_node(n);
3169 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3171 if (min < MIN_PARTIAL)
3173 else if (min > MAX_PARTIAL)
3175 s->min_partial = min;
3179 * calculate_sizes() determines the order and the distribution of data within
3182 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3184 unsigned long flags = s->flags;
3185 unsigned long size = s->object_size;
3189 * Round up object size to the next word boundary. We can only
3190 * place the free pointer at word boundaries and this determines
3191 * the possible location of the free pointer.
3193 size = ALIGN(size, sizeof(void *));
3195 #ifdef CONFIG_SLUB_DEBUG
3197 * Determine if we can poison the object itself. If the user of
3198 * the slab may touch the object after free or before allocation
3199 * then we should never poison the object itself.
3201 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3203 s->flags |= __OBJECT_POISON;
3205 s->flags &= ~__OBJECT_POISON;
3209 * If we are Redzoning then check if there is some space between the
3210 * end of the object and the free pointer. If not then add an
3211 * additional word to have some bytes to store Redzone information.
3213 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3214 size += sizeof(void *);
3218 * With that we have determined the number of bytes in actual use
3219 * by the object. This is the potential offset to the free pointer.
3223 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3226 * Relocate free pointer after the object if it is not
3227 * permitted to overwrite the first word of the object on
3230 * This is the case if we do RCU, have a constructor or
3231 * destructor or are poisoning the objects.
3234 size += sizeof(void *);
3237 #ifdef CONFIG_SLUB_DEBUG
3238 if (flags & SLAB_STORE_USER)
3240 * Need to store information about allocs and frees after
3243 size += 2 * sizeof(struct track);
3245 if (flags & SLAB_RED_ZONE)
3247 * Add some empty padding so that we can catch
3248 * overwrites from earlier objects rather than let
3249 * tracking information or the free pointer be
3250 * corrupted if a user writes before the start
3253 size += sizeof(void *);
3257 * SLUB stores one object immediately after another beginning from
3258 * offset 0. In order to align the objects we have to simply size
3259 * each object to conform to the alignment.
3261 size = ALIGN(size, s->align);
3263 if (forced_order >= 0)
3264 order = forced_order;
3266 order = calculate_order(size, s->reserved);
3273 s->allocflags |= __GFP_COMP;
3275 if (s->flags & SLAB_CACHE_DMA)
3276 s->allocflags |= GFP_DMA;
3278 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3279 s->allocflags |= __GFP_RECLAIMABLE;
3282 * Determine the number of objects per slab
3284 s->oo = oo_make(order, size, s->reserved);
3285 s->min = oo_make(get_order(size), size, s->reserved);
3286 if (oo_objects(s->oo) > oo_objects(s->max))
3289 return !!oo_objects(s->oo);
3292 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3294 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3297 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3298 s->reserved = sizeof(struct rcu_head);
3300 if (!calculate_sizes(s, -1))
3302 if (disable_higher_order_debug) {
3304 * Disable debugging flags that store metadata if the min slab
3307 if (get_order(s->size) > get_order(s->object_size)) {
3308 s->flags &= ~DEBUG_METADATA_FLAGS;
3310 if (!calculate_sizes(s, -1))
3315 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3316 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3317 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3318 /* Enable fast mode */
3319 s->flags |= __CMPXCHG_DOUBLE;
3323 * The larger the object size is, the more pages we want on the partial
3324 * list to avoid pounding the page allocator excessively.
3326 set_min_partial(s, ilog2(s->size) / 2);
3329 * cpu_partial determined the maximum number of objects kept in the
3330 * per cpu partial lists of a processor.
3332 * Per cpu partial lists mainly contain slabs that just have one
3333 * object freed. If they are used for allocation then they can be
3334 * filled up again with minimal effort. The slab will never hit the
3335 * per node partial lists and therefore no locking will be required.
3337 * This setting also determines
3339 * A) The number of objects from per cpu partial slabs dumped to the
3340 * per node list when we reach the limit.
3341 * B) The number of objects in cpu partial slabs to extract from the
3342 * per node list when we run out of per cpu objects. We only fetch
3343 * 50% to keep some capacity around for frees.
3345 if (!kmem_cache_has_cpu_partial(s))
3347 else if (s->size >= PAGE_SIZE)
3349 else if (s->size >= 1024)
3351 else if (s->size >= 256)
3352 s->cpu_partial = 13;
3354 s->cpu_partial = 30;
3357 s->remote_node_defrag_ratio = 1000;
3359 if (!init_kmem_cache_nodes(s))
3362 if (alloc_kmem_cache_cpus(s))
3365 free_kmem_cache_nodes(s);
3367 if (flags & SLAB_PANIC)
3368 panic("Cannot create slab %s size=%lu realsize=%u "
3369 "order=%u offset=%u flags=%lx\n",
3370 s->name, (unsigned long)s->size, s->size,
3371 oo_order(s->oo), s->offset, flags);
3375 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3378 #ifdef CONFIG_SLUB_DEBUG
3379 void *addr = page_address(page);
3381 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3382 sizeof(long), GFP_ATOMIC);
3385 slab_err(s, page, text, s->name);
3388 get_map(s, page, map);
3389 for_each_object(p, s, addr, page->objects) {
3391 if (!test_bit(slab_index(p, s, addr), map)) {
3392 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3393 print_tracking(s, p);
3402 * Attempt to free all partial slabs on a node.
3403 * This is called from __kmem_cache_shutdown(). We must take list_lock
3404 * because sysfs file might still access partial list after the shutdowning.
3406 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3408 struct page *page, *h;
3410 BUG_ON(irqs_disabled());
3411 spin_lock_irq(&n->list_lock);
3412 list_for_each_entry_safe(page, h, &n->partial, lru) {
3414 remove_partial(n, page);
3415 discard_slab(s, page);
3417 list_slab_objects(s, page,
3418 "Objects remaining in %s on __kmem_cache_shutdown()");
3421 spin_unlock_irq(&n->list_lock);
3425 * Release all resources used by a slab cache.
3427 int __kmem_cache_shutdown(struct kmem_cache *s)
3430 struct kmem_cache_node *n;
3433 /* Attempt to free all objects */
3434 for_each_kmem_cache_node(s, node, n) {
3436 if (n->nr_partial || slabs_node(s, node))
3442 /********************************************************************
3444 *******************************************************************/
3446 static int __init setup_slub_min_order(char *str)
3448 get_option(&str, &slub_min_order);
3453 __setup("slub_min_order=", setup_slub_min_order);
3455 static int __init setup_slub_max_order(char *str)
3457 get_option(&str, &slub_max_order);
3458 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3463 __setup("slub_max_order=", setup_slub_max_order);
3465 static int __init setup_slub_min_objects(char *str)
3467 get_option(&str, &slub_min_objects);
3472 __setup("slub_min_objects=", setup_slub_min_objects);
3474 void *__kmalloc(size_t size, gfp_t flags)
3476 struct kmem_cache *s;
3479 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3480 return kmalloc_large(size, flags);
3482 s = kmalloc_slab(size, flags);
3484 if (unlikely(ZERO_OR_NULL_PTR(s)))
3487 ret = slab_alloc(s, flags, _RET_IP_);
3489 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3491 kasan_kmalloc(s, ret, size);
3495 EXPORT_SYMBOL(__kmalloc);
3498 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3503 flags |= __GFP_COMP | __GFP_NOTRACK;
3504 page = alloc_kmem_pages_node(node, flags, get_order(size));
3506 ptr = page_address(page);
3508 kmalloc_large_node_hook(ptr, size, flags);
3512 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3514 struct kmem_cache *s;
3517 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3518 ret = kmalloc_large_node(size, flags, node);
3520 trace_kmalloc_node(_RET_IP_, ret,
3521 size, PAGE_SIZE << get_order(size),
3527 s = kmalloc_slab(size, flags);
3529 if (unlikely(ZERO_OR_NULL_PTR(s)))
3532 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3534 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3536 kasan_kmalloc(s, ret, size);
3540 EXPORT_SYMBOL(__kmalloc_node);
3543 static size_t __ksize(const void *object)
3547 if (unlikely(object == ZERO_SIZE_PTR))
3550 page = virt_to_head_page(object);
3552 if (unlikely(!PageSlab(page))) {
3553 WARN_ON(!PageCompound(page));
3554 return PAGE_SIZE << compound_order(page);
3557 return slab_ksize(page->slab_cache);
3560 size_t ksize(const void *object)
3562 size_t size = __ksize(object);
3563 /* We assume that ksize callers could use whole allocated area,
3564 so we need unpoison this area. */
3565 kasan_krealloc(object, size);
3568 EXPORT_SYMBOL(ksize);
3570 void kfree(const void *x)
3573 void *object = (void *)x;
3575 trace_kfree(_RET_IP_, x);
3577 if (unlikely(ZERO_OR_NULL_PTR(x)))
3580 page = virt_to_head_page(x);
3581 if (unlikely(!PageSlab(page))) {
3582 BUG_ON(!PageCompound(page));
3584 __free_kmem_pages(page, compound_order(page));
3587 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3589 EXPORT_SYMBOL(kfree);
3591 #define SHRINK_PROMOTE_MAX 32
3594 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3595 * up most to the head of the partial lists. New allocations will then
3596 * fill those up and thus they can be removed from the partial lists.
3598 * The slabs with the least items are placed last. This results in them
3599 * being allocated from last increasing the chance that the last objects
3600 * are freed in them.
3602 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3606 struct kmem_cache_node *n;
3609 struct list_head discard;
3610 struct list_head promote[SHRINK_PROMOTE_MAX];
3611 unsigned long flags;
3616 * Disable empty slabs caching. Used to avoid pinning offline
3617 * memory cgroups by kmem pages that can be freed.
3623 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3624 * so we have to make sure the change is visible.
3626 kick_all_cpus_sync();
3630 for_each_kmem_cache_node(s, node, n) {
3631 INIT_LIST_HEAD(&discard);
3632 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3633 INIT_LIST_HEAD(promote + i);
3635 spin_lock_irqsave(&n->list_lock, flags);
3638 * Build lists of slabs to discard or promote.
3640 * Note that concurrent frees may occur while we hold the
3641 * list_lock. page->inuse here is the upper limit.
3643 list_for_each_entry_safe(page, t, &n->partial, lru) {
3644 int free = page->objects - page->inuse;
3646 /* Do not reread page->inuse */
3649 /* We do not keep full slabs on the list */
3652 if (free == page->objects) {
3653 list_move(&page->lru, &discard);
3655 } else if (free <= SHRINK_PROMOTE_MAX)
3656 list_move(&page->lru, promote + free - 1);
3660 * Promote the slabs filled up most to the head of the
3663 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3664 list_splice(promote + i, &n->partial);
3666 spin_unlock_irqrestore(&n->list_lock, flags);
3668 /* Release empty slabs */
3669 list_for_each_entry_safe(page, t, &discard, lru)
3670 discard_slab(s, page);
3672 if (slabs_node(s, node))
3679 static int slab_mem_going_offline_callback(void *arg)
3681 struct kmem_cache *s;
3683 mutex_lock(&slab_mutex);
3684 list_for_each_entry(s, &slab_caches, list)
3685 __kmem_cache_shrink(s, false);
3686 mutex_unlock(&slab_mutex);
3691 static void slab_mem_offline_callback(void *arg)
3693 struct kmem_cache_node *n;
3694 struct kmem_cache *s;
3695 struct memory_notify *marg = arg;
3698 offline_node = marg->status_change_nid_normal;
3701 * If the node still has available memory. we need kmem_cache_node
3704 if (offline_node < 0)
3707 mutex_lock(&slab_mutex);
3708 list_for_each_entry(s, &slab_caches, list) {
3709 n = get_node(s, offline_node);
3712 * if n->nr_slabs > 0, slabs still exist on the node
3713 * that is going down. We were unable to free them,
3714 * and offline_pages() function shouldn't call this
3715 * callback. So, we must fail.
3717 BUG_ON(slabs_node(s, offline_node));
3719 s->node[offline_node] = NULL;
3720 kmem_cache_free(kmem_cache_node, n);
3723 mutex_unlock(&slab_mutex);
3726 static int slab_mem_going_online_callback(void *arg)
3728 struct kmem_cache_node *n;
3729 struct kmem_cache *s;
3730 struct memory_notify *marg = arg;
3731 int nid = marg->status_change_nid_normal;
3735 * If the node's memory is already available, then kmem_cache_node is
3736 * already created. Nothing to do.
3742 * We are bringing a node online. No memory is available yet. We must
3743 * allocate a kmem_cache_node structure in order to bring the node
3746 mutex_lock(&slab_mutex);
3747 list_for_each_entry(s, &slab_caches, list) {
3749 * XXX: kmem_cache_alloc_node will fallback to other nodes
3750 * since memory is not yet available from the node that
3753 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3758 init_kmem_cache_node(n);
3762 mutex_unlock(&slab_mutex);
3766 static int slab_memory_callback(struct notifier_block *self,
3767 unsigned long action, void *arg)
3772 case MEM_GOING_ONLINE:
3773 ret = slab_mem_going_online_callback(arg);
3775 case MEM_GOING_OFFLINE:
3776 ret = slab_mem_going_offline_callback(arg);
3779 case MEM_CANCEL_ONLINE:
3780 slab_mem_offline_callback(arg);
3783 case MEM_CANCEL_OFFLINE:
3787 ret = notifier_from_errno(ret);
3793 static struct notifier_block slab_memory_callback_nb = {
3794 .notifier_call = slab_memory_callback,
3795 .priority = SLAB_CALLBACK_PRI,
3798 /********************************************************************
3799 * Basic setup of slabs
3800 *******************************************************************/
3803 * Used for early kmem_cache structures that were allocated using
3804 * the page allocator. Allocate them properly then fix up the pointers
3805 * that may be pointing to the wrong kmem_cache structure.
3808 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3811 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3812 struct kmem_cache_node *n;
3814 memcpy(s, static_cache, kmem_cache->object_size);
3817 * This runs very early, and only the boot processor is supposed to be
3818 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3821 __flush_cpu_slab(s, smp_processor_id());
3822 for_each_kmem_cache_node(s, node, n) {
3825 list_for_each_entry(p, &n->partial, lru)
3828 #ifdef CONFIG_SLUB_DEBUG
3829 list_for_each_entry(p, &n->full, lru)
3833 slab_init_memcg_params(s);
3834 list_add(&s->list, &slab_caches);
3838 void __init kmem_cache_init(void)
3840 static __initdata struct kmem_cache boot_kmem_cache,
3841 boot_kmem_cache_node;
3843 if (debug_guardpage_minorder())
3846 kmem_cache_node = &boot_kmem_cache_node;
3847 kmem_cache = &boot_kmem_cache;
3849 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3850 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3852 register_hotmemory_notifier(&slab_memory_callback_nb);
3854 /* Able to allocate the per node structures */
3855 slab_state = PARTIAL;
3857 create_boot_cache(kmem_cache, "kmem_cache",
3858 offsetof(struct kmem_cache, node) +
3859 nr_node_ids * sizeof(struct kmem_cache_node *),
3860 SLAB_HWCACHE_ALIGN);
3862 kmem_cache = bootstrap(&boot_kmem_cache);
3865 * Allocate kmem_cache_node properly from the kmem_cache slab.
3866 * kmem_cache_node is separately allocated so no need to
3867 * update any list pointers.
3869 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3871 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3872 setup_kmalloc_cache_index_table();
3873 create_kmalloc_caches(0);
3876 register_cpu_notifier(&slab_notifier);
3879 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3881 slub_min_order, slub_max_order, slub_min_objects,
3882 nr_cpu_ids, nr_node_ids);
3885 void __init kmem_cache_init_late(void)
3890 __kmem_cache_alias(const char *name, size_t size, size_t align,
3891 unsigned long flags, void (*ctor)(void *))
3893 struct kmem_cache *s, *c;
3895 s = find_mergeable(size, align, flags, name, ctor);
3900 * Adjust the object sizes so that we clear
3901 * the complete object on kzalloc.
3903 s->object_size = max(s->object_size, (int)size);
3904 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3906 for_each_memcg_cache(c, s) {
3907 c->object_size = s->object_size;
3908 c->inuse = max_t(int, c->inuse,
3909 ALIGN(size, sizeof(void *)));
3912 if (sysfs_slab_alias(s, name)) {
3921 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3925 err = kmem_cache_open(s, flags);
3929 /* Mutex is not taken during early boot */
3930 if (slab_state <= UP)
3933 memcg_propagate_slab_attrs(s);
3934 err = sysfs_slab_add(s);
3936 __kmem_cache_release(s);
3943 * Use the cpu notifier to insure that the cpu slabs are flushed when
3946 static int slab_cpuup_callback(struct notifier_block *nfb,
3947 unsigned long action, void *hcpu)
3949 long cpu = (long)hcpu;
3950 struct kmem_cache *s;
3951 unsigned long flags;
3954 case CPU_UP_CANCELED:
3955 case CPU_UP_CANCELED_FROZEN:
3957 case CPU_DEAD_FROZEN:
3958 mutex_lock(&slab_mutex);
3959 list_for_each_entry(s, &slab_caches, list) {
3960 local_irq_save(flags);
3961 __flush_cpu_slab(s, cpu);
3962 local_irq_restore(flags);
3964 mutex_unlock(&slab_mutex);
3972 static struct notifier_block slab_notifier = {
3973 .notifier_call = slab_cpuup_callback
3978 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3980 struct kmem_cache *s;
3983 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3984 return kmalloc_large(size, gfpflags);
3986 s = kmalloc_slab(size, gfpflags);
3988 if (unlikely(ZERO_OR_NULL_PTR(s)))
3991 ret = slab_alloc(s, gfpflags, caller);
3993 /* Honor the call site pointer we received. */
3994 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4000 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4001 int node, unsigned long caller)
4003 struct kmem_cache *s;
4006 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4007 ret = kmalloc_large_node(size, gfpflags, node);
4009 trace_kmalloc_node(caller, ret,
4010 size, PAGE_SIZE << get_order(size),
4016 s = kmalloc_slab(size, gfpflags);
4018 if (unlikely(ZERO_OR_NULL_PTR(s)))
4021 ret = slab_alloc_node(s, gfpflags, node, caller);
4023 /* Honor the call site pointer we received. */
4024 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4031 static int count_inuse(struct page *page)
4036 static int count_total(struct page *page)
4038 return page->objects;
4042 #ifdef CONFIG_SLUB_DEBUG
4043 static int validate_slab(struct kmem_cache *s, struct page *page,
4047 void *addr = page_address(page);
4049 if (!check_slab(s, page) ||
4050 !on_freelist(s, page, NULL))
4053 /* Now we know that a valid freelist exists */
4054 bitmap_zero(map, page->objects);
4056 get_map(s, page, map);
4057 for_each_object(p, s, addr, page->objects) {
4058 if (test_bit(slab_index(p, s, addr), map))
4059 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4063 for_each_object(p, s, addr, page->objects)
4064 if (!test_bit(slab_index(p, s, addr), map))
4065 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4070 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4074 validate_slab(s, page, map);
4078 static int validate_slab_node(struct kmem_cache *s,
4079 struct kmem_cache_node *n, unsigned long *map)
4081 unsigned long count = 0;
4083 unsigned long flags;
4085 spin_lock_irqsave(&n->list_lock, flags);
4087 list_for_each_entry(page, &n->partial, lru) {
4088 validate_slab_slab(s, page, map);
4091 if (count != n->nr_partial)
4092 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4093 s->name, count, n->nr_partial);
4095 if (!(s->flags & SLAB_STORE_USER))
4098 list_for_each_entry(page, &n->full, lru) {
4099 validate_slab_slab(s, page, map);
4102 if (count != atomic_long_read(&n->nr_slabs))
4103 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4104 s->name, count, atomic_long_read(&n->nr_slabs));
4107 spin_unlock_irqrestore(&n->list_lock, flags);
4111 static long validate_slab_cache(struct kmem_cache *s)
4114 unsigned long count = 0;
4115 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4116 sizeof(unsigned long), GFP_KERNEL);
4117 struct kmem_cache_node *n;
4123 for_each_kmem_cache_node(s, node, n)
4124 count += validate_slab_node(s, n, map);
4129 * Generate lists of code addresses where slabcache objects are allocated
4134 unsigned long count;
4141 DECLARE_BITMAP(cpus, NR_CPUS);
4147 unsigned long count;
4148 struct location *loc;
4151 static void free_loc_track(struct loc_track *t)
4154 free_pages((unsigned long)t->loc,
4155 get_order(sizeof(struct location) * t->max));
4158 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4163 order = get_order(sizeof(struct location) * max);
4165 l = (void *)__get_free_pages(flags, order);
4170 memcpy(l, t->loc, sizeof(struct location) * t->count);
4178 static int add_location(struct loc_track *t, struct kmem_cache *s,
4179 const struct track *track)
4181 long start, end, pos;
4183 unsigned long caddr;
4184 unsigned long age = jiffies - track->when;
4190 pos = start + (end - start + 1) / 2;
4193 * There is nothing at "end". If we end up there
4194 * we need to add something to before end.
4199 caddr = t->loc[pos].addr;
4200 if (track->addr == caddr) {
4206 if (age < l->min_time)
4208 if (age > l->max_time)
4211 if (track->pid < l->min_pid)
4212 l->min_pid = track->pid;
4213 if (track->pid > l->max_pid)
4214 l->max_pid = track->pid;
4216 cpumask_set_cpu(track->cpu,
4217 to_cpumask(l->cpus));
4219 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4223 if (track->addr < caddr)
4230 * Not found. Insert new tracking element.
4232 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4238 (t->count - pos) * sizeof(struct location));
4241 l->addr = track->addr;
4245 l->min_pid = track->pid;
4246 l->max_pid = track->pid;
4247 cpumask_clear(to_cpumask(l->cpus));
4248 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4249 nodes_clear(l->nodes);
4250 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4254 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4255 struct page *page, enum track_item alloc,
4258 void *addr = page_address(page);
4261 bitmap_zero(map, page->objects);
4262 get_map(s, page, map);
4264 for_each_object(p, s, addr, page->objects)
4265 if (!test_bit(slab_index(p, s, addr), map))
4266 add_location(t, s, get_track(s, p, alloc));
4269 static int list_locations(struct kmem_cache *s, char *buf,
4270 enum track_item alloc)
4274 struct loc_track t = { 0, 0, NULL };
4276 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4277 sizeof(unsigned long), GFP_KERNEL);
4278 struct kmem_cache_node *n;
4280 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4283 return sprintf(buf, "Out of memory\n");
4285 /* Push back cpu slabs */
4288 for_each_kmem_cache_node(s, node, n) {
4289 unsigned long flags;
4292 if (!atomic_long_read(&n->nr_slabs))
4295 spin_lock_irqsave(&n->list_lock, flags);
4296 list_for_each_entry(page, &n->partial, lru)
4297 process_slab(&t, s, page, alloc, map);
4298 list_for_each_entry(page, &n->full, lru)
4299 process_slab(&t, s, page, alloc, map);
4300 spin_unlock_irqrestore(&n->list_lock, flags);
4303 for (i = 0; i < t.count; i++) {
4304 struct location *l = &t.loc[i];
4306 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4308 len += sprintf(buf + len, "%7ld ", l->count);
4311 len += sprintf(buf + len, "%pS", (void *)l->addr);
4313 len += sprintf(buf + len, "<not-available>");
4315 if (l->sum_time != l->min_time) {
4316 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4318 (long)div_u64(l->sum_time, l->count),
4321 len += sprintf(buf + len, " age=%ld",
4324 if (l->min_pid != l->max_pid)
4325 len += sprintf(buf + len, " pid=%ld-%ld",
4326 l->min_pid, l->max_pid);
4328 len += sprintf(buf + len, " pid=%ld",
4331 if (num_online_cpus() > 1 &&
4332 !cpumask_empty(to_cpumask(l->cpus)) &&
4333 len < PAGE_SIZE - 60)
4334 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4336 cpumask_pr_args(to_cpumask(l->cpus)));
4338 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4339 len < PAGE_SIZE - 60)
4340 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4342 nodemask_pr_args(&l->nodes));
4344 len += sprintf(buf + len, "\n");
4350 len += sprintf(buf, "No data\n");
4355 #ifdef SLUB_RESILIENCY_TEST
4356 static void __init resiliency_test(void)
4360 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4362 pr_err("SLUB resiliency testing\n");
4363 pr_err("-----------------------\n");
4364 pr_err("A. Corruption after allocation\n");
4366 p = kzalloc(16, GFP_KERNEL);
4368 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4371 validate_slab_cache(kmalloc_caches[4]);
4373 /* Hmmm... The next two are dangerous */
4374 p = kzalloc(32, GFP_KERNEL);
4375 p[32 + sizeof(void *)] = 0x34;
4376 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4378 pr_err("If allocated object is overwritten then not detectable\n\n");
4380 validate_slab_cache(kmalloc_caches[5]);
4381 p = kzalloc(64, GFP_KERNEL);
4382 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4384 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4386 pr_err("If allocated object is overwritten then not detectable\n\n");
4387 validate_slab_cache(kmalloc_caches[6]);
4389 pr_err("\nB. Corruption after free\n");
4390 p = kzalloc(128, GFP_KERNEL);
4393 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4394 validate_slab_cache(kmalloc_caches[7]);
4396 p = kzalloc(256, GFP_KERNEL);
4399 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4400 validate_slab_cache(kmalloc_caches[8]);
4402 p = kzalloc(512, GFP_KERNEL);
4405 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4406 validate_slab_cache(kmalloc_caches[9]);
4410 static void resiliency_test(void) {};
4415 enum slab_stat_type {
4416 SL_ALL, /* All slabs */
4417 SL_PARTIAL, /* Only partially allocated slabs */
4418 SL_CPU, /* Only slabs used for cpu caches */
4419 SL_OBJECTS, /* Determine allocated objects not slabs */
4420 SL_TOTAL /* Determine object capacity not slabs */
4423 #define SO_ALL (1 << SL_ALL)
4424 #define SO_PARTIAL (1 << SL_PARTIAL)
4425 #define SO_CPU (1 << SL_CPU)
4426 #define SO_OBJECTS (1 << SL_OBJECTS)
4427 #define SO_TOTAL (1 << SL_TOTAL)
4429 static ssize_t show_slab_objects(struct kmem_cache *s,
4430 char *buf, unsigned long flags)
4432 unsigned long total = 0;
4435 unsigned long *nodes;
4437 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4441 if (flags & SO_CPU) {
4444 for_each_possible_cpu(cpu) {
4445 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4450 page = READ_ONCE(c->page);
4454 node = page_to_nid(page);
4455 if (flags & SO_TOTAL)
4457 else if (flags & SO_OBJECTS)
4465 page = READ_ONCE(c->partial);
4467 node = page_to_nid(page);
4468 if (flags & SO_TOTAL)
4470 else if (flags & SO_OBJECTS)
4481 #ifdef CONFIG_SLUB_DEBUG
4482 if (flags & SO_ALL) {
4483 struct kmem_cache_node *n;
4485 for_each_kmem_cache_node(s, node, n) {
4487 if (flags & SO_TOTAL)
4488 x = atomic_long_read(&n->total_objects);
4489 else if (flags & SO_OBJECTS)
4490 x = atomic_long_read(&n->total_objects) -
4491 count_partial(n, count_free);
4493 x = atomic_long_read(&n->nr_slabs);
4500 if (flags & SO_PARTIAL) {
4501 struct kmem_cache_node *n;
4503 for_each_kmem_cache_node(s, node, n) {
4504 if (flags & SO_TOTAL)
4505 x = count_partial(n, count_total);
4506 else if (flags & SO_OBJECTS)
4507 x = count_partial(n, count_inuse);
4514 x = sprintf(buf, "%lu", total);
4516 for (node = 0; node < nr_node_ids; node++)
4518 x += sprintf(buf + x, " N%d=%lu",
4523 return x + sprintf(buf + x, "\n");
4526 #ifdef CONFIG_SLUB_DEBUG
4527 static int any_slab_objects(struct kmem_cache *s)
4530 struct kmem_cache_node *n;
4532 for_each_kmem_cache_node(s, node, n)
4533 if (atomic_long_read(&n->total_objects))
4540 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4541 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4543 struct slab_attribute {
4544 struct attribute attr;
4545 ssize_t (*show)(struct kmem_cache *s, char *buf);
4546 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4549 #define SLAB_ATTR_RO(_name) \
4550 static struct slab_attribute _name##_attr = \
4551 __ATTR(_name, 0400, _name##_show, NULL)
4553 #define SLAB_ATTR(_name) \
4554 static struct slab_attribute _name##_attr = \
4555 __ATTR(_name, 0600, _name##_show, _name##_store)
4557 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4559 return sprintf(buf, "%d\n", s->size);
4561 SLAB_ATTR_RO(slab_size);
4563 static ssize_t align_show(struct kmem_cache *s, char *buf)
4565 return sprintf(buf, "%d\n", s->align);
4567 SLAB_ATTR_RO(align);
4569 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4571 return sprintf(buf, "%d\n", s->object_size);
4573 SLAB_ATTR_RO(object_size);
4575 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4577 return sprintf(buf, "%d\n", oo_objects(s->oo));
4579 SLAB_ATTR_RO(objs_per_slab);
4581 static ssize_t order_store(struct kmem_cache *s,
4582 const char *buf, size_t length)
4584 unsigned long order;
4587 err = kstrtoul(buf, 10, &order);
4591 if (order > slub_max_order || order < slub_min_order)
4594 calculate_sizes(s, order);
4598 static ssize_t order_show(struct kmem_cache *s, char *buf)
4600 return sprintf(buf, "%d\n", oo_order(s->oo));
4604 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4606 return sprintf(buf, "%lu\n", s->min_partial);
4609 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4615 err = kstrtoul(buf, 10, &min);
4619 set_min_partial(s, min);
4622 SLAB_ATTR(min_partial);
4624 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4626 return sprintf(buf, "%u\n", s->cpu_partial);
4629 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4632 unsigned long objects;
4635 err = kstrtoul(buf, 10, &objects);
4638 if (objects && !kmem_cache_has_cpu_partial(s))
4641 s->cpu_partial = objects;
4645 SLAB_ATTR(cpu_partial);
4647 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4651 return sprintf(buf, "%pS\n", s->ctor);
4655 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4657 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4659 SLAB_ATTR_RO(aliases);
4661 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4663 return show_slab_objects(s, buf, SO_PARTIAL);
4665 SLAB_ATTR_RO(partial);
4667 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4669 return show_slab_objects(s, buf, SO_CPU);
4671 SLAB_ATTR_RO(cpu_slabs);
4673 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4675 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4677 SLAB_ATTR_RO(objects);
4679 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4681 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4683 SLAB_ATTR_RO(objects_partial);
4685 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4692 for_each_online_cpu(cpu) {
4693 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4696 pages += page->pages;
4697 objects += page->pobjects;
4701 len = sprintf(buf, "%d(%d)", objects, pages);
4704 for_each_online_cpu(cpu) {
4705 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4707 if (page && len < PAGE_SIZE - 20)
4708 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4709 page->pobjects, page->pages);
4712 return len + sprintf(buf + len, "\n");
4714 SLAB_ATTR_RO(slabs_cpu_partial);
4716 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4718 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4721 static ssize_t reclaim_account_store(struct kmem_cache *s,
4722 const char *buf, size_t length)
4724 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4726 s->flags |= SLAB_RECLAIM_ACCOUNT;
4729 SLAB_ATTR(reclaim_account);
4731 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4733 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4735 SLAB_ATTR_RO(hwcache_align);
4737 #ifdef CONFIG_ZONE_DMA
4738 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4740 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4742 SLAB_ATTR_RO(cache_dma);
4745 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4747 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4749 SLAB_ATTR_RO(destroy_by_rcu);
4751 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4753 return sprintf(buf, "%d\n", s->reserved);
4755 SLAB_ATTR_RO(reserved);
4757 #ifdef CONFIG_SLUB_DEBUG
4758 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4760 return show_slab_objects(s, buf, SO_ALL);
4762 SLAB_ATTR_RO(slabs);
4764 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4766 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4768 SLAB_ATTR_RO(total_objects);
4770 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4772 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4775 static ssize_t sanity_checks_store(struct kmem_cache *s,
4776 const char *buf, size_t length)
4778 s->flags &= ~SLAB_DEBUG_FREE;
4779 if (buf[0] == '1') {
4780 s->flags &= ~__CMPXCHG_DOUBLE;
4781 s->flags |= SLAB_DEBUG_FREE;
4785 SLAB_ATTR(sanity_checks);
4787 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4789 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4792 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4796 * Tracing a merged cache is going to give confusing results
4797 * as well as cause other issues like converting a mergeable
4798 * cache into an umergeable one.
4800 if (s->refcount > 1)
4803 s->flags &= ~SLAB_TRACE;
4804 if (buf[0] == '1') {
4805 s->flags &= ~__CMPXCHG_DOUBLE;
4806 s->flags |= SLAB_TRACE;
4812 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4814 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4817 static ssize_t red_zone_store(struct kmem_cache *s,
4818 const char *buf, size_t length)
4820 if (any_slab_objects(s))
4823 s->flags &= ~SLAB_RED_ZONE;
4824 if (buf[0] == '1') {
4825 s->flags &= ~__CMPXCHG_DOUBLE;
4826 s->flags |= SLAB_RED_ZONE;
4828 calculate_sizes(s, -1);
4831 SLAB_ATTR(red_zone);
4833 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4835 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4838 static ssize_t poison_store(struct kmem_cache *s,
4839 const char *buf, size_t length)
4841 if (any_slab_objects(s))
4844 s->flags &= ~SLAB_POISON;
4845 if (buf[0] == '1') {
4846 s->flags &= ~__CMPXCHG_DOUBLE;
4847 s->flags |= SLAB_POISON;
4849 calculate_sizes(s, -1);
4854 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4856 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4859 static ssize_t store_user_store(struct kmem_cache *s,
4860 const char *buf, size_t length)
4862 if (any_slab_objects(s))
4865 s->flags &= ~SLAB_STORE_USER;
4866 if (buf[0] == '1') {
4867 s->flags &= ~__CMPXCHG_DOUBLE;
4868 s->flags |= SLAB_STORE_USER;
4870 calculate_sizes(s, -1);
4873 SLAB_ATTR(store_user);
4875 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4880 static ssize_t validate_store(struct kmem_cache *s,
4881 const char *buf, size_t length)
4885 if (buf[0] == '1') {
4886 ret = validate_slab_cache(s);
4892 SLAB_ATTR(validate);
4894 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4896 if (!(s->flags & SLAB_STORE_USER))
4898 return list_locations(s, buf, TRACK_ALLOC);
4900 SLAB_ATTR_RO(alloc_calls);
4902 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4904 if (!(s->flags & SLAB_STORE_USER))
4906 return list_locations(s, buf, TRACK_FREE);
4908 SLAB_ATTR_RO(free_calls);
4909 #endif /* CONFIG_SLUB_DEBUG */
4911 #ifdef CONFIG_FAILSLAB
4912 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4914 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4917 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4920 if (s->refcount > 1)
4923 s->flags &= ~SLAB_FAILSLAB;
4925 s->flags |= SLAB_FAILSLAB;
4928 SLAB_ATTR(failslab);
4931 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4936 static ssize_t shrink_store(struct kmem_cache *s,
4937 const char *buf, size_t length)
4940 kmem_cache_shrink(s);
4948 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4950 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4953 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4954 const char *buf, size_t length)
4956 unsigned long ratio;
4959 err = kstrtoul(buf, 10, &ratio);
4964 s->remote_node_defrag_ratio = ratio * 10;
4968 SLAB_ATTR(remote_node_defrag_ratio);
4971 #ifdef CONFIG_SLUB_STATS
4972 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4974 unsigned long sum = 0;
4977 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4982 for_each_online_cpu(cpu) {
4983 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4989 len = sprintf(buf, "%lu", sum);
4992 for_each_online_cpu(cpu) {
4993 if (data[cpu] && len < PAGE_SIZE - 20)
4994 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4998 return len + sprintf(buf + len, "\n");
5001 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5005 for_each_online_cpu(cpu)
5006 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5009 #define STAT_ATTR(si, text) \
5010 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5012 return show_stat(s, buf, si); \
5014 static ssize_t text##_store(struct kmem_cache *s, \
5015 const char *buf, size_t length) \
5017 if (buf[0] != '0') \
5019 clear_stat(s, si); \
5024 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5025 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5026 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5027 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5028 STAT_ATTR(FREE_FROZEN, free_frozen);
5029 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5030 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5031 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5032 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5033 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5034 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5035 STAT_ATTR(FREE_SLAB, free_slab);
5036 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5037 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5038 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5039 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5040 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5041 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5042 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5043 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5044 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5045 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5046 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5047 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5048 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5049 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5052 static struct attribute *slab_attrs[] = {
5053 &slab_size_attr.attr,
5054 &object_size_attr.attr,
5055 &objs_per_slab_attr.attr,
5057 &min_partial_attr.attr,
5058 &cpu_partial_attr.attr,
5060 &objects_partial_attr.attr,
5062 &cpu_slabs_attr.attr,
5066 &hwcache_align_attr.attr,
5067 &reclaim_account_attr.attr,
5068 &destroy_by_rcu_attr.attr,
5070 &reserved_attr.attr,
5071 &slabs_cpu_partial_attr.attr,
5072 #ifdef CONFIG_SLUB_DEBUG
5073 &total_objects_attr.attr,
5075 &sanity_checks_attr.attr,
5077 &red_zone_attr.attr,
5079 &store_user_attr.attr,
5080 &validate_attr.attr,
5081 &alloc_calls_attr.attr,
5082 &free_calls_attr.attr,
5084 #ifdef CONFIG_ZONE_DMA
5085 &cache_dma_attr.attr,
5088 &remote_node_defrag_ratio_attr.attr,
5090 #ifdef CONFIG_SLUB_STATS
5091 &alloc_fastpath_attr.attr,
5092 &alloc_slowpath_attr.attr,
5093 &free_fastpath_attr.attr,
5094 &free_slowpath_attr.attr,
5095 &free_frozen_attr.attr,
5096 &free_add_partial_attr.attr,
5097 &free_remove_partial_attr.attr,
5098 &alloc_from_partial_attr.attr,
5099 &alloc_slab_attr.attr,
5100 &alloc_refill_attr.attr,
5101 &alloc_node_mismatch_attr.attr,
5102 &free_slab_attr.attr,
5103 &cpuslab_flush_attr.attr,
5104 &deactivate_full_attr.attr,
5105 &deactivate_empty_attr.attr,
5106 &deactivate_to_head_attr.attr,
5107 &deactivate_to_tail_attr.attr,
5108 &deactivate_remote_frees_attr.attr,
5109 &deactivate_bypass_attr.attr,
5110 &order_fallback_attr.attr,
5111 &cmpxchg_double_fail_attr.attr,
5112 &cmpxchg_double_cpu_fail_attr.attr,
5113 &cpu_partial_alloc_attr.attr,
5114 &cpu_partial_free_attr.attr,
5115 &cpu_partial_node_attr.attr,
5116 &cpu_partial_drain_attr.attr,
5118 #ifdef CONFIG_FAILSLAB
5119 &failslab_attr.attr,
5125 static struct attribute_group slab_attr_group = {
5126 .attrs = slab_attrs,
5129 static ssize_t slab_attr_show(struct kobject *kobj,
5130 struct attribute *attr,
5133 struct slab_attribute *attribute;
5134 struct kmem_cache *s;
5137 attribute = to_slab_attr(attr);
5140 if (!attribute->show)
5143 err = attribute->show(s, buf);
5148 static ssize_t slab_attr_store(struct kobject *kobj,
5149 struct attribute *attr,
5150 const char *buf, size_t len)
5152 struct slab_attribute *attribute;
5153 struct kmem_cache *s;
5156 attribute = to_slab_attr(attr);
5159 if (!attribute->store)
5162 err = attribute->store(s, buf, len);
5164 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5165 struct kmem_cache *c;
5167 mutex_lock(&slab_mutex);
5168 if (s->max_attr_size < len)
5169 s->max_attr_size = len;
5172 * This is a best effort propagation, so this function's return
5173 * value will be determined by the parent cache only. This is
5174 * basically because not all attributes will have a well
5175 * defined semantics for rollbacks - most of the actions will
5176 * have permanent effects.
5178 * Returning the error value of any of the children that fail
5179 * is not 100 % defined, in the sense that users seeing the
5180 * error code won't be able to know anything about the state of
5183 * Only returning the error code for the parent cache at least
5184 * has well defined semantics. The cache being written to
5185 * directly either failed or succeeded, in which case we loop
5186 * through the descendants with best-effort propagation.
5188 for_each_memcg_cache(c, s)
5189 attribute->store(c, buf, len);
5190 mutex_unlock(&slab_mutex);
5196 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5200 char *buffer = NULL;
5201 struct kmem_cache *root_cache;
5203 if (is_root_cache(s))
5206 root_cache = s->memcg_params.root_cache;
5209 * This mean this cache had no attribute written. Therefore, no point
5210 * in copying default values around
5212 if (!root_cache->max_attr_size)
5215 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5218 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5220 if (!attr || !attr->store || !attr->show)
5224 * It is really bad that we have to allocate here, so we will
5225 * do it only as a fallback. If we actually allocate, though,
5226 * we can just use the allocated buffer until the end.
5228 * Most of the slub attributes will tend to be very small in
5229 * size, but sysfs allows buffers up to a page, so they can
5230 * theoretically happen.
5234 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5237 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5238 if (WARN_ON(!buffer))
5243 attr->show(root_cache, buf);
5244 attr->store(s, buf, strlen(buf));
5248 free_page((unsigned long)buffer);
5252 static void kmem_cache_release(struct kobject *k)
5254 slab_kmem_cache_release(to_slab(k));
5257 static const struct sysfs_ops slab_sysfs_ops = {
5258 .show = slab_attr_show,
5259 .store = slab_attr_store,
5262 static struct kobj_type slab_ktype = {
5263 .sysfs_ops = &slab_sysfs_ops,
5264 .release = kmem_cache_release,
5267 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5269 struct kobj_type *ktype = get_ktype(kobj);
5271 if (ktype == &slab_ktype)
5276 static const struct kset_uevent_ops slab_uevent_ops = {
5277 .filter = uevent_filter,
5280 static struct kset *slab_kset;
5282 static inline struct kset *cache_kset(struct kmem_cache *s)
5285 if (!is_root_cache(s))
5286 return s->memcg_params.root_cache->memcg_kset;
5291 #define ID_STR_LENGTH 64
5293 /* Create a unique string id for a slab cache:
5295 * Format :[flags-]size
5297 static char *create_unique_id(struct kmem_cache *s)
5299 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5306 * First flags affecting slabcache operations. We will only
5307 * get here for aliasable slabs so we do not need to support
5308 * too many flags. The flags here must cover all flags that
5309 * are matched during merging to guarantee that the id is
5312 if (s->flags & SLAB_CACHE_DMA)
5314 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5316 if (s->flags & SLAB_DEBUG_FREE)
5318 if (!(s->flags & SLAB_NOTRACK))
5320 if (s->flags & SLAB_ACCOUNT)
5324 p += sprintf(p, "%07d", s->size);
5326 BUG_ON(p > name + ID_STR_LENGTH - 1);
5330 static int sysfs_slab_add(struct kmem_cache *s)
5334 int unmergeable = slab_unmergeable(s);
5338 * Slabcache can never be merged so we can use the name proper.
5339 * This is typically the case for debug situations. In that
5340 * case we can catch duplicate names easily.
5342 sysfs_remove_link(&slab_kset->kobj, s->name);
5346 * Create a unique name for the slab as a target
5349 name = create_unique_id(s);
5352 s->kobj.kset = cache_kset(s);
5353 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5357 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5362 if (is_root_cache(s)) {
5363 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5364 if (!s->memcg_kset) {
5371 kobject_uevent(&s->kobj, KOBJ_ADD);
5373 /* Setup first alias */
5374 sysfs_slab_alias(s, s->name);
5381 kobject_del(&s->kobj);
5385 void sysfs_slab_remove(struct kmem_cache *s)
5387 if (slab_state < FULL)
5389 * Sysfs has not been setup yet so no need to remove the
5395 kset_unregister(s->memcg_kset);
5397 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5398 kobject_del(&s->kobj);
5399 kobject_put(&s->kobj);
5403 * Need to buffer aliases during bootup until sysfs becomes
5404 * available lest we lose that information.
5406 struct saved_alias {
5407 struct kmem_cache *s;
5409 struct saved_alias *next;
5412 static struct saved_alias *alias_list;
5414 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5416 struct saved_alias *al;
5418 if (slab_state == FULL) {
5420 * If we have a leftover link then remove it.
5422 sysfs_remove_link(&slab_kset->kobj, name);
5423 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5426 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5432 al->next = alias_list;
5437 static int __init slab_sysfs_init(void)
5439 struct kmem_cache *s;
5442 mutex_lock(&slab_mutex);
5444 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5446 mutex_unlock(&slab_mutex);
5447 pr_err("Cannot register slab subsystem.\n");
5453 list_for_each_entry(s, &slab_caches, list) {
5454 err = sysfs_slab_add(s);
5456 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5460 while (alias_list) {
5461 struct saved_alias *al = alias_list;
5463 alias_list = alias_list->next;
5464 err = sysfs_slab_alias(al->s, al->name);
5466 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5471 mutex_unlock(&slab_mutex);
5476 __initcall(slab_sysfs_init);
5477 #endif /* CONFIG_SYSFS */
5480 * The /proc/slabinfo ABI
5482 #ifdef CONFIG_SLABINFO
5483 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5485 unsigned long nr_slabs = 0;
5486 unsigned long nr_objs = 0;
5487 unsigned long nr_free = 0;
5489 struct kmem_cache_node *n;
5491 for_each_kmem_cache_node(s, node, n) {
5492 nr_slabs += node_nr_slabs(n);
5493 nr_objs += node_nr_objs(n);
5494 nr_free += count_partial(n, count_free);
5497 sinfo->active_objs = nr_objs - nr_free;
5498 sinfo->num_objs = nr_objs;
5499 sinfo->active_slabs = nr_slabs;
5500 sinfo->num_slabs = nr_slabs;
5501 sinfo->objects_per_slab = oo_objects(s->oo);
5502 sinfo->cache_order = oo_order(s->oo);
5505 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5509 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5510 size_t count, loff_t *ppos)
5514 #endif /* CONFIG_SLABINFO */