2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
137 * Issues still to be resolved:
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
141 * - Variable sizing of the per node arrays
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
154 #define MIN_PARTIAL 5
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
161 #define MAX_PARTIAL 10
163 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
171 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
209 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
212 static inline void stat(const struct kmem_cache *s, enum stat_item si)
214 #ifdef CONFIG_SLUB_STATS
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
219 raw_cpu_inc(s->cpu_slab->stat[si]);
223 /********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
227 /* Verify that a pointer has an address that is valid within a slab page */
228 static inline int check_valid_pointer(struct kmem_cache *s,
229 struct page *page, const void *object)
236 base = page_address(page);
237 if (object < base || object >= base + page->objects * s->size ||
238 (object - base) % s->size) {
245 static inline void *get_freepointer(struct kmem_cache *s, void *object)
247 return *(void **)(object + s->offset);
250 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
252 prefetch(object + s->offset);
255 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
259 #ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
262 p = get_freepointer(s, object);
267 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
269 *(void **)(object + s->offset) = fp;
272 /* Loop over all objects in a slab */
273 #define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
277 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline size_t slab_ksize(const struct kmem_cache *s)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order, unsigned long size, int reserved)
313 return ((PAGE_SIZE << order) - reserved) / size;
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
326 static inline int oo_order(struct kmem_cache_order_objects x)
328 return x.x >> OO_SHIFT;
331 static inline int oo_objects(struct kmem_cache_order_objects x)
333 return x.x & OO_MASK;
337 * Per slab locking using the pagelock
339 static __always_inline void slab_lock(struct page *page)
341 VM_BUG_ON_PAGE(PageTail(page), page);
342 bit_spin_lock(PG_locked, &page->flags);
345 static __always_inline void slab_unlock(struct page *page)
347 VM_BUG_ON_PAGE(PageTail(page), page);
348 __bit_spin_unlock(PG_locked, &page->flags);
351 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
354 tmp.counters = counters_new;
356 * page->counters can cover frozen/inuse/objects as well
357 * as page->_count. If we assign to ->counters directly
358 * we run the risk of losing updates to page->_count, so
359 * be careful and only assign to the fields we need.
361 page->frozen = tmp.frozen;
362 page->inuse = tmp.inuse;
363 page->objects = tmp.objects;
366 /* Interrupts must be disabled (for the fallback code to work right) */
367 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
368 void *freelist_old, unsigned long counters_old,
369 void *freelist_new, unsigned long counters_new,
372 VM_BUG_ON(!irqs_disabled());
373 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
374 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
375 if (s->flags & __CMPXCHG_DOUBLE) {
376 if (cmpxchg_double(&page->freelist, &page->counters,
377 freelist_old, counters_old,
378 freelist_new, counters_new))
384 if (page->freelist == freelist_old &&
385 page->counters == counters_old) {
386 page->freelist = freelist_new;
387 set_page_slub_counters(page, counters_new);
395 stat(s, CMPXCHG_DOUBLE_FAIL);
397 #ifdef SLUB_DEBUG_CMPXCHG
398 pr_info("%s %s: cmpxchg double redo ", n, s->name);
404 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
405 void *freelist_old, unsigned long counters_old,
406 void *freelist_new, unsigned long counters_new,
409 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
410 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
411 if (s->flags & __CMPXCHG_DOUBLE) {
412 if (cmpxchg_double(&page->freelist, &page->counters,
413 freelist_old, counters_old,
414 freelist_new, counters_new))
421 local_irq_save(flags);
423 if (page->freelist == freelist_old &&
424 page->counters == counters_old) {
425 page->freelist = freelist_new;
426 set_page_slub_counters(page, counters_new);
428 local_irq_restore(flags);
432 local_irq_restore(flags);
436 stat(s, CMPXCHG_DOUBLE_FAIL);
438 #ifdef SLUB_DEBUG_CMPXCHG
439 pr_info("%s %s: cmpxchg double redo ", n, s->name);
445 #ifdef CONFIG_SLUB_DEBUG
447 * Determine a map of object in use on a page.
449 * Node listlock must be held to guarantee that the page does
450 * not vanish from under us.
452 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
455 void *addr = page_address(page);
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(slab_index(p, s, addr), map);
464 #if defined(CONFIG_SLUB_DEBUG_ON)
465 static int slub_debug = DEBUG_DEFAULT_FLAGS;
466 #elif defined(CONFIG_KASAN)
467 static int slub_debug = SLAB_STORE_USER;
469 static int slub_debug;
472 static char *slub_debug_slabs;
473 static int disable_higher_order_debug;
476 * slub is about to manipulate internal object metadata. This memory lies
477 * outside the range of the allocated object, so accessing it would normally
478 * be reported by kasan as a bounds error. metadata_access_enable() is used
479 * to tell kasan that these accesses are OK.
481 static inline void metadata_access_enable(void)
483 kasan_disable_current();
486 static inline void metadata_access_disable(void)
488 kasan_enable_current();
494 static void print_section(char *text, u8 *addr, unsigned int length)
496 metadata_access_enable();
497 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
499 metadata_access_disable();
502 static struct track *get_track(struct kmem_cache *s, void *object,
503 enum track_item alloc)
508 p = object + s->offset + sizeof(void *);
510 p = object + s->inuse;
515 static void set_track(struct kmem_cache *s, void *object,
516 enum track_item alloc, unsigned long addr)
518 struct track *p = get_track(s, object, alloc);
521 #ifdef CONFIG_STACKTRACE
522 struct stack_trace trace;
525 trace.nr_entries = 0;
526 trace.max_entries = TRACK_ADDRS_COUNT;
527 trace.entries = p->addrs;
529 metadata_access_enable();
530 save_stack_trace(&trace);
531 metadata_access_disable();
533 /* See rant in lockdep.c */
534 if (trace.nr_entries != 0 &&
535 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
538 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
542 p->cpu = smp_processor_id();
543 p->pid = current->pid;
546 memset(p, 0, sizeof(struct track));
549 static void init_tracking(struct kmem_cache *s, void *object)
551 if (!(s->flags & SLAB_STORE_USER))
554 set_track(s, object, TRACK_FREE, 0UL);
555 set_track(s, object, TRACK_ALLOC, 0UL);
558 static void print_track(const char *s, struct track *t)
563 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
564 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
565 #ifdef CONFIG_STACKTRACE
568 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
570 pr_err("\t%pS\n", (void *)t->addrs[i]);
577 static void print_tracking(struct kmem_cache *s, void *object)
579 if (!(s->flags & SLAB_STORE_USER))
582 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
583 print_track("Freed", get_track(s, object, TRACK_FREE));
586 static void print_page_info(struct page *page)
588 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
589 page, page->objects, page->inuse, page->freelist, page->flags);
593 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
595 struct va_format vaf;
601 pr_err("=============================================================================\n");
602 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
603 pr_err("-----------------------------------------------------------------------------\n\n");
605 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
609 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
611 struct va_format vaf;
617 pr_err("FIX %s: %pV\n", s->name, &vaf);
621 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
623 unsigned int off; /* Offset of last byte */
624 u8 *addr = page_address(page);
626 print_tracking(s, p);
628 print_page_info(page);
630 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
631 p, p - addr, get_freepointer(s, p));
634 print_section("Bytes b4 ", p - 16, 16);
636 print_section("Object ", p, min_t(unsigned long, s->object_size,
638 if (s->flags & SLAB_RED_ZONE)
639 print_section("Redzone ", p + s->object_size,
640 s->inuse - s->object_size);
643 off = s->offset + sizeof(void *);
647 if (s->flags & SLAB_STORE_USER)
648 off += 2 * sizeof(struct track);
651 /* Beginning of the filler is the free pointer */
652 print_section("Padding ", p + off, s->size - off);
657 void object_err(struct kmem_cache *s, struct page *page,
658 u8 *object, char *reason)
660 slab_bug(s, "%s", reason);
661 print_trailer(s, page, object);
664 static void slab_err(struct kmem_cache *s, struct page *page,
665 const char *fmt, ...)
671 vsnprintf(buf, sizeof(buf), fmt, args);
673 slab_bug(s, "%s", buf);
674 print_page_info(page);
678 static void init_object(struct kmem_cache *s, void *object, u8 val)
682 if (s->flags & __OBJECT_POISON) {
683 memset(p, POISON_FREE, s->object_size - 1);
684 p[s->object_size - 1] = POISON_END;
687 if (s->flags & SLAB_RED_ZONE)
688 memset(p + s->object_size, val, s->inuse - s->object_size);
691 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
692 void *from, void *to)
694 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
695 memset(from, data, to - from);
698 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
699 u8 *object, char *what,
700 u8 *start, unsigned int value, unsigned int bytes)
705 metadata_access_enable();
706 fault = memchr_inv(start, value, bytes);
707 metadata_access_disable();
712 while (end > fault && end[-1] == value)
715 slab_bug(s, "%s overwritten", what);
716 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
717 fault, end - 1, fault[0], value);
718 print_trailer(s, page, object);
720 restore_bytes(s, what, value, fault, end);
728 * Bytes of the object to be managed.
729 * If the freepointer may overlay the object then the free
730 * pointer is the first word of the object.
732 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
735 * object + s->object_size
736 * Padding to reach word boundary. This is also used for Redzoning.
737 * Padding is extended by another word if Redzoning is enabled and
738 * object_size == inuse.
740 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
741 * 0xcc (RED_ACTIVE) for objects in use.
744 * Meta data starts here.
746 * A. Free pointer (if we cannot overwrite object on free)
747 * B. Tracking data for SLAB_STORE_USER
748 * C. Padding to reach required alignment boundary or at mininum
749 * one word if debugging is on to be able to detect writes
750 * before the word boundary.
752 * Padding is done using 0x5a (POISON_INUSE)
755 * Nothing is used beyond s->size.
757 * If slabcaches are merged then the object_size and inuse boundaries are mostly
758 * ignored. And therefore no slab options that rely on these boundaries
759 * may be used with merged slabcaches.
762 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
764 unsigned long off = s->inuse; /* The end of info */
767 /* Freepointer is placed after the object. */
768 off += sizeof(void *);
770 if (s->flags & SLAB_STORE_USER)
771 /* We also have user information there */
772 off += 2 * sizeof(struct track);
777 return check_bytes_and_report(s, page, p, "Object padding",
778 p + off, POISON_INUSE, s->size - off);
781 /* Check the pad bytes at the end of a slab page */
782 static int slab_pad_check(struct kmem_cache *s, struct page *page)
790 if (!(s->flags & SLAB_POISON))
793 start = page_address(page);
794 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
795 end = start + length;
796 remainder = length % s->size;
800 metadata_access_enable();
801 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
802 metadata_access_disable();
805 while (end > fault && end[-1] == POISON_INUSE)
808 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
809 print_section("Padding ", end - remainder, remainder);
811 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
815 static int check_object(struct kmem_cache *s, struct page *page,
816 void *object, u8 val)
819 u8 *endobject = object + s->object_size;
821 if (s->flags & SLAB_RED_ZONE) {
822 if (!check_bytes_and_report(s, page, object, "Redzone",
823 endobject, val, s->inuse - s->object_size))
826 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
827 check_bytes_and_report(s, page, p, "Alignment padding",
828 endobject, POISON_INUSE,
829 s->inuse - s->object_size);
833 if (s->flags & SLAB_POISON) {
834 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
835 (!check_bytes_and_report(s, page, p, "Poison", p,
836 POISON_FREE, s->object_size - 1) ||
837 !check_bytes_and_report(s, page, p, "Poison",
838 p + s->object_size - 1, POISON_END, 1)))
841 * check_pad_bytes cleans up on its own.
843 check_pad_bytes(s, page, p);
846 if (!s->offset && val == SLUB_RED_ACTIVE)
848 * Object and freepointer overlap. Cannot check
849 * freepointer while object is allocated.
853 /* Check free pointer validity */
854 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
855 object_err(s, page, p, "Freepointer corrupt");
857 * No choice but to zap it and thus lose the remainder
858 * of the free objects in this slab. May cause
859 * another error because the object count is now wrong.
861 set_freepointer(s, p, NULL);
867 static int check_slab(struct kmem_cache *s, struct page *page)
871 VM_BUG_ON(!irqs_disabled());
873 if (!PageSlab(page)) {
874 slab_err(s, page, "Not a valid slab page");
878 maxobj = order_objects(compound_order(page), s->size, s->reserved);
879 if (page->objects > maxobj) {
880 slab_err(s, page, "objects %u > max %u",
881 page->objects, maxobj);
884 if (page->inuse > page->objects) {
885 slab_err(s, page, "inuse %u > max %u",
886 page->inuse, page->objects);
889 /* Slab_pad_check fixes things up after itself */
890 slab_pad_check(s, page);
895 * Determine if a certain object on a page is on the freelist. Must hold the
896 * slab lock to guarantee that the chains are in a consistent state.
898 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
906 while (fp && nr <= page->objects) {
909 if (!check_valid_pointer(s, page, fp)) {
911 object_err(s, page, object,
912 "Freechain corrupt");
913 set_freepointer(s, object, NULL);
915 slab_err(s, page, "Freepointer corrupt");
916 page->freelist = NULL;
917 page->inuse = page->objects;
918 slab_fix(s, "Freelist cleared");
924 fp = get_freepointer(s, object);
928 max_objects = order_objects(compound_order(page), s->size, s->reserved);
929 if (max_objects > MAX_OBJS_PER_PAGE)
930 max_objects = MAX_OBJS_PER_PAGE;
932 if (page->objects != max_objects) {
933 slab_err(s, page, "Wrong number of objects. Found %d but "
934 "should be %d", page->objects, max_objects);
935 page->objects = max_objects;
936 slab_fix(s, "Number of objects adjusted.");
938 if (page->inuse != page->objects - nr) {
939 slab_err(s, page, "Wrong object count. Counter is %d but "
940 "counted were %d", page->inuse, page->objects - nr);
941 page->inuse = page->objects - nr;
942 slab_fix(s, "Object count adjusted.");
944 return search == NULL;
947 static void trace(struct kmem_cache *s, struct page *page, void *object,
950 if (s->flags & SLAB_TRACE) {
951 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
953 alloc ? "alloc" : "free",
958 print_section("Object ", (void *)object,
966 * Tracking of fully allocated slabs for debugging purposes.
968 static void add_full(struct kmem_cache *s,
969 struct kmem_cache_node *n, struct page *page)
971 if (!(s->flags & SLAB_STORE_USER))
974 lockdep_assert_held(&n->list_lock);
975 list_add(&page->lru, &n->full);
978 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
980 if (!(s->flags & SLAB_STORE_USER))
983 lockdep_assert_held(&n->list_lock);
984 list_del(&page->lru);
987 /* Tracking of the number of slabs for debugging purposes */
988 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
990 struct kmem_cache_node *n = get_node(s, node);
992 return atomic_long_read(&n->nr_slabs);
995 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
997 return atomic_long_read(&n->nr_slabs);
1000 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1002 struct kmem_cache_node *n = get_node(s, node);
1005 * May be called early in order to allocate a slab for the
1006 * kmem_cache_node structure. Solve the chicken-egg
1007 * dilemma by deferring the increment of the count during
1008 * bootstrap (see early_kmem_cache_node_alloc).
1011 atomic_long_inc(&n->nr_slabs);
1012 atomic_long_add(objects, &n->total_objects);
1015 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1017 struct kmem_cache_node *n = get_node(s, node);
1019 atomic_long_dec(&n->nr_slabs);
1020 atomic_long_sub(objects, &n->total_objects);
1023 /* Object debug checks for alloc/free paths */
1024 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1027 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1030 init_object(s, object, SLUB_RED_INACTIVE);
1031 init_tracking(s, object);
1034 static noinline int alloc_debug_processing(struct kmem_cache *s,
1036 void *object, unsigned long addr)
1038 if (!check_slab(s, page))
1041 if (!check_valid_pointer(s, page, object)) {
1042 object_err(s, page, object, "Freelist Pointer check fails");
1046 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1049 /* Success perform special debug activities for allocs */
1050 if (s->flags & SLAB_STORE_USER)
1051 set_track(s, object, TRACK_ALLOC, addr);
1052 trace(s, page, object, 1);
1053 init_object(s, object, SLUB_RED_ACTIVE);
1057 if (PageSlab(page)) {
1059 * If this is a slab page then lets do the best we can
1060 * to avoid issues in the future. Marking all objects
1061 * as used avoids touching the remaining objects.
1063 slab_fix(s, "Marking all objects used");
1064 page->inuse = page->objects;
1065 page->freelist = NULL;
1070 /* Supports checking bulk free of a constructed freelist */
1071 static noinline struct kmem_cache_node *free_debug_processing(
1072 struct kmem_cache *s, struct page *page,
1073 void *head, void *tail, int bulk_cnt,
1074 unsigned long addr, unsigned long *flags)
1076 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1077 void *object = head;
1080 spin_lock_irqsave(&n->list_lock, *flags);
1083 if (!check_slab(s, page))
1089 if (!check_valid_pointer(s, page, object)) {
1090 slab_err(s, page, "Invalid object pointer 0x%p", object);
1094 if (on_freelist(s, page, object)) {
1095 object_err(s, page, object, "Object already free");
1099 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1102 if (unlikely(s != page->slab_cache)) {
1103 if (!PageSlab(page)) {
1104 slab_err(s, page, "Attempt to free object(0x%p) "
1105 "outside of slab", object);
1106 } else if (!page->slab_cache) {
1107 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1111 object_err(s, page, object,
1112 "page slab pointer corrupt.");
1116 if (s->flags & SLAB_STORE_USER)
1117 set_track(s, object, TRACK_FREE, addr);
1118 trace(s, page, object, 0);
1119 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1120 init_object(s, object, SLUB_RED_INACTIVE);
1122 /* Reached end of constructed freelist yet? */
1123 if (object != tail) {
1124 object = get_freepointer(s, object);
1128 if (cnt != bulk_cnt)
1129 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1134 * Keep node_lock to preserve integrity
1135 * until the object is actually freed
1141 spin_unlock_irqrestore(&n->list_lock, *flags);
1142 slab_fix(s, "Object at 0x%p not freed", object);
1146 static int __init setup_slub_debug(char *str)
1148 slub_debug = DEBUG_DEFAULT_FLAGS;
1149 if (*str++ != '=' || !*str)
1151 * No options specified. Switch on full debugging.
1157 * No options but restriction on slabs. This means full
1158 * debugging for slabs matching a pattern.
1165 * Switch off all debugging measures.
1170 * Determine which debug features should be switched on
1172 for (; *str && *str != ','; str++) {
1173 switch (tolower(*str)) {
1175 slub_debug |= SLAB_DEBUG_FREE;
1178 slub_debug |= SLAB_RED_ZONE;
1181 slub_debug |= SLAB_POISON;
1184 slub_debug |= SLAB_STORE_USER;
1187 slub_debug |= SLAB_TRACE;
1190 slub_debug |= SLAB_FAILSLAB;
1194 * Avoid enabling debugging on caches if its minimum
1195 * order would increase as a result.
1197 disable_higher_order_debug = 1;
1200 pr_err("slub_debug option '%c' unknown. skipped\n",
1207 slub_debug_slabs = str + 1;
1212 __setup("slub_debug", setup_slub_debug);
1214 unsigned long kmem_cache_flags(unsigned long object_size,
1215 unsigned long flags, const char *name,
1216 void (*ctor)(void *))
1219 * Enable debugging if selected on the kernel commandline.
1221 if (slub_debug && (!slub_debug_slabs || (name &&
1222 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1223 flags |= slub_debug;
1227 #else /* !CONFIG_SLUB_DEBUG */
1228 static inline void setup_object_debug(struct kmem_cache *s,
1229 struct page *page, void *object) {}
1231 static inline int alloc_debug_processing(struct kmem_cache *s,
1232 struct page *page, void *object, unsigned long addr) { return 0; }
1234 static inline struct kmem_cache_node *free_debug_processing(
1235 struct kmem_cache *s, struct page *page,
1236 void *head, void *tail, int bulk_cnt,
1237 unsigned long addr, unsigned long *flags) { return NULL; }
1239 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1241 static inline int check_object(struct kmem_cache *s, struct page *page,
1242 void *object, u8 val) { return 1; }
1243 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1244 struct page *page) {}
1245 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1246 struct page *page) {}
1247 unsigned long kmem_cache_flags(unsigned long object_size,
1248 unsigned long flags, const char *name,
1249 void (*ctor)(void *))
1253 #define slub_debug 0
1255 #define disable_higher_order_debug 0
1257 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1259 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1261 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1263 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1266 #endif /* CONFIG_SLUB_DEBUG */
1269 * Hooks for other subsystems that check memory allocations. In a typical
1270 * production configuration these hooks all should produce no code at all.
1272 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1274 kmemleak_alloc(ptr, size, 1, flags);
1275 kasan_kmalloc_large(ptr, size);
1278 static inline void kfree_hook(const void *x)
1281 kasan_kfree_large(x);
1284 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1287 flags &= gfp_allowed_mask;
1288 lockdep_trace_alloc(flags);
1289 might_sleep_if(gfpflags_allow_blocking(flags));
1291 if (should_failslab(s->object_size, flags, s->flags))
1294 return memcg_kmem_get_cache(s, flags);
1297 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1298 size_t size, void **p)
1302 flags &= gfp_allowed_mask;
1303 for (i = 0; i < size; i++) {
1304 void *object = p[i];
1306 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1307 kmemleak_alloc_recursive(object, s->object_size, 1,
1309 kasan_slab_alloc(s, object);
1311 memcg_kmem_put_cache(s);
1314 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1316 kmemleak_free_recursive(x, s->flags);
1319 * Trouble is that we may no longer disable interrupts in the fast path
1320 * So in order to make the debug calls that expect irqs to be
1321 * disabled we need to disable interrupts temporarily.
1323 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1325 unsigned long flags;
1327 local_irq_save(flags);
1328 kmemcheck_slab_free(s, x, s->object_size);
1329 debug_check_no_locks_freed(x, s->object_size);
1330 local_irq_restore(flags);
1333 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1334 debug_check_no_obj_freed(x, s->object_size);
1336 kasan_slab_free(s, x);
1339 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1340 void *head, void *tail)
1343 * Compiler cannot detect this function can be removed if slab_free_hook()
1344 * evaluates to nothing. Thus, catch all relevant config debug options here.
1346 #if defined(CONFIG_KMEMCHECK) || \
1347 defined(CONFIG_LOCKDEP) || \
1348 defined(CONFIG_DEBUG_KMEMLEAK) || \
1349 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1350 defined(CONFIG_KASAN)
1352 void *object = head;
1353 void *tail_obj = tail ? : head;
1356 slab_free_hook(s, object);
1357 } while ((object != tail_obj) &&
1358 (object = get_freepointer(s, object)));
1362 static void setup_object(struct kmem_cache *s, struct page *page,
1365 setup_object_debug(s, page, object);
1366 if (unlikely(s->ctor)) {
1367 kasan_unpoison_object_data(s, object);
1369 kasan_poison_object_data(s, object);
1374 * Slab allocation and freeing
1376 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1377 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1380 int order = oo_order(oo);
1382 flags |= __GFP_NOTRACK;
1384 if (node == NUMA_NO_NODE)
1385 page = alloc_pages(flags, order);
1387 page = __alloc_pages_node(node, flags, order);
1389 if (page && memcg_charge_slab(page, flags, order, s)) {
1390 __free_pages(page, order);
1397 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1400 struct kmem_cache_order_objects oo = s->oo;
1405 flags &= gfp_allowed_mask;
1407 if (gfpflags_allow_blocking(flags))
1410 flags |= s->allocflags;
1413 * Let the initial higher-order allocation fail under memory pressure
1414 * so we fall-back to the minimum order allocation.
1416 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1417 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1418 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1420 page = alloc_slab_page(s, alloc_gfp, node, oo);
1421 if (unlikely(!page)) {
1425 * Allocation may have failed due to fragmentation.
1426 * Try a lower order alloc if possible
1428 page = alloc_slab_page(s, alloc_gfp, node, oo);
1429 if (unlikely(!page))
1431 stat(s, ORDER_FALLBACK);
1434 if (kmemcheck_enabled &&
1435 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1436 int pages = 1 << oo_order(oo);
1438 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1441 * Objects from caches that have a constructor don't get
1442 * cleared when they're allocated, so we need to do it here.
1445 kmemcheck_mark_uninitialized_pages(page, pages);
1447 kmemcheck_mark_unallocated_pages(page, pages);
1450 page->objects = oo_objects(oo);
1452 order = compound_order(page);
1453 page->slab_cache = s;
1454 __SetPageSlab(page);
1455 if (page_is_pfmemalloc(page))
1456 SetPageSlabPfmemalloc(page);
1458 start = page_address(page);
1460 if (unlikely(s->flags & SLAB_POISON))
1461 memset(start, POISON_INUSE, PAGE_SIZE << order);
1463 kasan_poison_slab(page);
1465 for_each_object_idx(p, idx, s, start, page->objects) {
1466 setup_object(s, page, p);
1467 if (likely(idx < page->objects))
1468 set_freepointer(s, p, p + s->size);
1470 set_freepointer(s, p, NULL);
1473 page->freelist = start;
1474 page->inuse = page->objects;
1478 if (gfpflags_allow_blocking(flags))
1479 local_irq_disable();
1483 mod_zone_page_state(page_zone(page),
1484 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1485 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1488 inc_slabs_node(s, page_to_nid(page), page->objects);
1493 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1495 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1496 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1500 return allocate_slab(s,
1501 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1504 static void __free_slab(struct kmem_cache *s, struct page *page)
1506 int order = compound_order(page);
1507 int pages = 1 << order;
1509 if (kmem_cache_debug(s)) {
1512 slab_pad_check(s, page);
1513 for_each_object(p, s, page_address(page),
1515 check_object(s, page, p, SLUB_RED_INACTIVE);
1518 kmemcheck_free_shadow(page, compound_order(page));
1520 mod_zone_page_state(page_zone(page),
1521 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1522 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1525 __ClearPageSlabPfmemalloc(page);
1526 __ClearPageSlab(page);
1528 page_mapcount_reset(page);
1529 if (current->reclaim_state)
1530 current->reclaim_state->reclaimed_slab += pages;
1531 __free_kmem_pages(page, order);
1534 #define need_reserve_slab_rcu \
1535 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1537 static void rcu_free_slab(struct rcu_head *h)
1541 if (need_reserve_slab_rcu)
1542 page = virt_to_head_page(h);
1544 page = container_of((struct list_head *)h, struct page, lru);
1546 __free_slab(page->slab_cache, page);
1549 static void free_slab(struct kmem_cache *s, struct page *page)
1551 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1552 struct rcu_head *head;
1554 if (need_reserve_slab_rcu) {
1555 int order = compound_order(page);
1556 int offset = (PAGE_SIZE << order) - s->reserved;
1558 VM_BUG_ON(s->reserved != sizeof(*head));
1559 head = page_address(page) + offset;
1561 head = &page->rcu_head;
1564 call_rcu(head, rcu_free_slab);
1566 __free_slab(s, page);
1569 static void discard_slab(struct kmem_cache *s, struct page *page)
1571 dec_slabs_node(s, page_to_nid(page), page->objects);
1576 * Management of partially allocated slabs.
1579 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1582 if (tail == DEACTIVATE_TO_TAIL)
1583 list_add_tail(&page->lru, &n->partial);
1585 list_add(&page->lru, &n->partial);
1588 static inline void add_partial(struct kmem_cache_node *n,
1589 struct page *page, int tail)
1591 lockdep_assert_held(&n->list_lock);
1592 __add_partial(n, page, tail);
1595 static inline void remove_partial(struct kmem_cache_node *n,
1598 lockdep_assert_held(&n->list_lock);
1599 list_del(&page->lru);
1604 * Remove slab from the partial list, freeze it and
1605 * return the pointer to the freelist.
1607 * Returns a list of objects or NULL if it fails.
1609 static inline void *acquire_slab(struct kmem_cache *s,
1610 struct kmem_cache_node *n, struct page *page,
1611 int mode, int *objects)
1614 unsigned long counters;
1617 lockdep_assert_held(&n->list_lock);
1620 * Zap the freelist and set the frozen bit.
1621 * The old freelist is the list of objects for the
1622 * per cpu allocation list.
1624 freelist = page->freelist;
1625 counters = page->counters;
1626 new.counters = counters;
1627 *objects = new.objects - new.inuse;
1629 new.inuse = page->objects;
1630 new.freelist = NULL;
1632 new.freelist = freelist;
1635 VM_BUG_ON(new.frozen);
1638 if (!__cmpxchg_double_slab(s, page,
1640 new.freelist, new.counters,
1644 remove_partial(n, page);
1649 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1650 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1653 * Try to allocate a partial slab from a specific node.
1655 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1656 struct kmem_cache_cpu *c, gfp_t flags)
1658 struct page *page, *page2;
1659 void *object = NULL;
1664 * Racy check. If we mistakenly see no partial slabs then we
1665 * just allocate an empty slab. If we mistakenly try to get a
1666 * partial slab and there is none available then get_partials()
1669 if (!n || !n->nr_partial)
1672 spin_lock(&n->list_lock);
1673 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1676 if (!pfmemalloc_match(page, flags))
1679 t = acquire_slab(s, n, page, object == NULL, &objects);
1683 available += objects;
1686 stat(s, ALLOC_FROM_PARTIAL);
1689 put_cpu_partial(s, page, 0);
1690 stat(s, CPU_PARTIAL_NODE);
1692 if (!kmem_cache_has_cpu_partial(s)
1693 || available > s->cpu_partial / 2)
1697 spin_unlock(&n->list_lock);
1702 * Get a page from somewhere. Search in increasing NUMA distances.
1704 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1705 struct kmem_cache_cpu *c)
1708 struct zonelist *zonelist;
1711 enum zone_type high_zoneidx = gfp_zone(flags);
1713 unsigned int cpuset_mems_cookie;
1716 * The defrag ratio allows a configuration of the tradeoffs between
1717 * inter node defragmentation and node local allocations. A lower
1718 * defrag_ratio increases the tendency to do local allocations
1719 * instead of attempting to obtain partial slabs from other nodes.
1721 * If the defrag_ratio is set to 0 then kmalloc() always
1722 * returns node local objects. If the ratio is higher then kmalloc()
1723 * may return off node objects because partial slabs are obtained
1724 * from other nodes and filled up.
1726 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1727 * defrag_ratio = 1000) then every (well almost) allocation will
1728 * first attempt to defrag slab caches on other nodes. This means
1729 * scanning over all nodes to look for partial slabs which may be
1730 * expensive if we do it every time we are trying to find a slab
1731 * with available objects.
1733 if (!s->remote_node_defrag_ratio ||
1734 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1738 cpuset_mems_cookie = read_mems_allowed_begin();
1739 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1740 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1741 struct kmem_cache_node *n;
1743 n = get_node(s, zone_to_nid(zone));
1745 if (n && cpuset_zone_allowed(zone, flags) &&
1746 n->nr_partial > s->min_partial) {
1747 object = get_partial_node(s, n, c, flags);
1750 * Don't check read_mems_allowed_retry()
1751 * here - if mems_allowed was updated in
1752 * parallel, that was a harmless race
1753 * between allocation and the cpuset
1760 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1766 * Get a partial page, lock it and return it.
1768 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1769 struct kmem_cache_cpu *c)
1772 int searchnode = node;
1774 if (node == NUMA_NO_NODE)
1775 searchnode = numa_mem_id();
1776 else if (!node_present_pages(node))
1777 searchnode = node_to_mem_node(node);
1779 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1780 if (object || node != NUMA_NO_NODE)
1783 return get_any_partial(s, flags, c);
1786 #ifdef CONFIG_PREEMPT
1788 * Calculate the next globally unique transaction for disambiguiation
1789 * during cmpxchg. The transactions start with the cpu number and are then
1790 * incremented by CONFIG_NR_CPUS.
1792 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1795 * No preemption supported therefore also no need to check for
1801 static inline unsigned long next_tid(unsigned long tid)
1803 return tid + TID_STEP;
1806 static inline unsigned int tid_to_cpu(unsigned long tid)
1808 return tid % TID_STEP;
1811 static inline unsigned long tid_to_event(unsigned long tid)
1813 return tid / TID_STEP;
1816 static inline unsigned int init_tid(int cpu)
1821 static inline void note_cmpxchg_failure(const char *n,
1822 const struct kmem_cache *s, unsigned long tid)
1824 #ifdef SLUB_DEBUG_CMPXCHG
1825 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1827 pr_info("%s %s: cmpxchg redo ", n, s->name);
1829 #ifdef CONFIG_PREEMPT
1830 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1831 pr_warn("due to cpu change %d -> %d\n",
1832 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1835 if (tid_to_event(tid) != tid_to_event(actual_tid))
1836 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1837 tid_to_event(tid), tid_to_event(actual_tid));
1839 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1840 actual_tid, tid, next_tid(tid));
1842 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1845 static void init_kmem_cache_cpus(struct kmem_cache *s)
1849 for_each_possible_cpu(cpu)
1850 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1854 * Remove the cpu slab
1856 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1859 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1860 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1862 enum slab_modes l = M_NONE, m = M_NONE;
1864 int tail = DEACTIVATE_TO_HEAD;
1868 if (page->freelist) {
1869 stat(s, DEACTIVATE_REMOTE_FREES);
1870 tail = DEACTIVATE_TO_TAIL;
1874 * Stage one: Free all available per cpu objects back
1875 * to the page freelist while it is still frozen. Leave the
1878 * There is no need to take the list->lock because the page
1881 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1883 unsigned long counters;
1886 prior = page->freelist;
1887 counters = page->counters;
1888 set_freepointer(s, freelist, prior);
1889 new.counters = counters;
1891 VM_BUG_ON(!new.frozen);
1893 } while (!__cmpxchg_double_slab(s, page,
1895 freelist, new.counters,
1896 "drain percpu freelist"));
1898 freelist = nextfree;
1902 * Stage two: Ensure that the page is unfrozen while the
1903 * list presence reflects the actual number of objects
1906 * We setup the list membership and then perform a cmpxchg
1907 * with the count. If there is a mismatch then the page
1908 * is not unfrozen but the page is on the wrong list.
1910 * Then we restart the process which may have to remove
1911 * the page from the list that we just put it on again
1912 * because the number of objects in the slab may have
1917 old.freelist = page->freelist;
1918 old.counters = page->counters;
1919 VM_BUG_ON(!old.frozen);
1921 /* Determine target state of the slab */
1922 new.counters = old.counters;
1925 set_freepointer(s, freelist, old.freelist);
1926 new.freelist = freelist;
1928 new.freelist = old.freelist;
1932 if (!new.inuse && n->nr_partial >= s->min_partial)
1934 else if (new.freelist) {
1939 * Taking the spinlock removes the possiblity
1940 * that acquire_slab() will see a slab page that
1943 spin_lock(&n->list_lock);
1947 if (kmem_cache_debug(s) && !lock) {
1950 * This also ensures that the scanning of full
1951 * slabs from diagnostic functions will not see
1954 spin_lock(&n->list_lock);
1962 remove_partial(n, page);
1964 else if (l == M_FULL)
1966 remove_full(s, n, page);
1968 if (m == M_PARTIAL) {
1970 add_partial(n, page, tail);
1973 } else if (m == M_FULL) {
1975 stat(s, DEACTIVATE_FULL);
1976 add_full(s, n, page);
1982 if (!__cmpxchg_double_slab(s, page,
1983 old.freelist, old.counters,
1984 new.freelist, new.counters,
1989 spin_unlock(&n->list_lock);
1992 stat(s, DEACTIVATE_EMPTY);
1993 discard_slab(s, page);
1999 * Unfreeze all the cpu partial slabs.
2001 * This function must be called with interrupts disabled
2002 * for the cpu using c (or some other guarantee must be there
2003 * to guarantee no concurrent accesses).
2005 static void unfreeze_partials(struct kmem_cache *s,
2006 struct kmem_cache_cpu *c)
2008 #ifdef CONFIG_SLUB_CPU_PARTIAL
2009 struct kmem_cache_node *n = NULL, *n2 = NULL;
2010 struct page *page, *discard_page = NULL;
2012 while ((page = c->partial)) {
2016 c->partial = page->next;
2018 n2 = get_node(s, page_to_nid(page));
2021 spin_unlock(&n->list_lock);
2024 spin_lock(&n->list_lock);
2029 old.freelist = page->freelist;
2030 old.counters = page->counters;
2031 VM_BUG_ON(!old.frozen);
2033 new.counters = old.counters;
2034 new.freelist = old.freelist;
2038 } while (!__cmpxchg_double_slab(s, page,
2039 old.freelist, old.counters,
2040 new.freelist, new.counters,
2041 "unfreezing slab"));
2043 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2044 page->next = discard_page;
2045 discard_page = page;
2047 add_partial(n, page, DEACTIVATE_TO_TAIL);
2048 stat(s, FREE_ADD_PARTIAL);
2053 spin_unlock(&n->list_lock);
2055 while (discard_page) {
2056 page = discard_page;
2057 discard_page = discard_page->next;
2059 stat(s, DEACTIVATE_EMPTY);
2060 discard_slab(s, page);
2067 * Put a page that was just frozen (in __slab_free) into a partial page
2068 * slot if available. This is done without interrupts disabled and without
2069 * preemption disabled. The cmpxchg is racy and may put the partial page
2070 * onto a random cpus partial slot.
2072 * If we did not find a slot then simply move all the partials to the
2073 * per node partial list.
2075 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2077 #ifdef CONFIG_SLUB_CPU_PARTIAL
2078 struct page *oldpage;
2086 oldpage = this_cpu_read(s->cpu_slab->partial);
2089 pobjects = oldpage->pobjects;
2090 pages = oldpage->pages;
2091 if (drain && pobjects > s->cpu_partial) {
2092 unsigned long flags;
2094 * partial array is full. Move the existing
2095 * set to the per node partial list.
2097 local_irq_save(flags);
2098 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2099 local_irq_restore(flags);
2103 stat(s, CPU_PARTIAL_DRAIN);
2108 pobjects += page->objects - page->inuse;
2110 page->pages = pages;
2111 page->pobjects = pobjects;
2112 page->next = oldpage;
2114 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2116 if (unlikely(!s->cpu_partial)) {
2117 unsigned long flags;
2119 local_irq_save(flags);
2120 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2121 local_irq_restore(flags);
2127 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2129 stat(s, CPUSLAB_FLUSH);
2130 deactivate_slab(s, c->page, c->freelist);
2132 c->tid = next_tid(c->tid);
2140 * Called from IPI handler with interrupts disabled.
2142 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2144 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2150 unfreeze_partials(s, c);
2154 static void flush_cpu_slab(void *d)
2156 struct kmem_cache *s = d;
2158 __flush_cpu_slab(s, smp_processor_id());
2161 static bool has_cpu_slab(int cpu, void *info)
2163 struct kmem_cache *s = info;
2164 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2166 return c->page || c->partial;
2169 static void flush_all(struct kmem_cache *s)
2171 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2175 * Check if the objects in a per cpu structure fit numa
2176 * locality expectations.
2178 static inline int node_match(struct page *page, int node)
2181 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2187 #ifdef CONFIG_SLUB_DEBUG
2188 static int count_free(struct page *page)
2190 return page->objects - page->inuse;
2193 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2195 return atomic_long_read(&n->total_objects);
2197 #endif /* CONFIG_SLUB_DEBUG */
2199 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2200 static unsigned long count_partial(struct kmem_cache_node *n,
2201 int (*get_count)(struct page *))
2203 unsigned long flags;
2204 unsigned long x = 0;
2207 spin_lock_irqsave(&n->list_lock, flags);
2208 list_for_each_entry(page, &n->partial, lru)
2209 x += get_count(page);
2210 spin_unlock_irqrestore(&n->list_lock, flags);
2213 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2215 static noinline void
2216 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2218 #ifdef CONFIG_SLUB_DEBUG
2219 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2220 DEFAULT_RATELIMIT_BURST);
2222 struct kmem_cache_node *n;
2224 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2227 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2229 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2230 s->name, s->object_size, s->size, oo_order(s->oo),
2233 if (oo_order(s->min) > get_order(s->object_size))
2234 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2237 for_each_kmem_cache_node(s, node, n) {
2238 unsigned long nr_slabs;
2239 unsigned long nr_objs;
2240 unsigned long nr_free;
2242 nr_free = count_partial(n, count_free);
2243 nr_slabs = node_nr_slabs(n);
2244 nr_objs = node_nr_objs(n);
2246 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2247 node, nr_slabs, nr_objs, nr_free);
2252 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2253 int node, struct kmem_cache_cpu **pc)
2256 struct kmem_cache_cpu *c = *pc;
2259 freelist = get_partial(s, flags, node, c);
2264 page = new_slab(s, flags, node);
2266 c = raw_cpu_ptr(s->cpu_slab);
2271 * No other reference to the page yet so we can
2272 * muck around with it freely without cmpxchg
2274 freelist = page->freelist;
2275 page->freelist = NULL;
2277 stat(s, ALLOC_SLAB);
2286 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2288 if (unlikely(PageSlabPfmemalloc(page)))
2289 return gfp_pfmemalloc_allowed(gfpflags);
2295 * Check the page->freelist of a page and either transfer the freelist to the
2296 * per cpu freelist or deactivate the page.
2298 * The page is still frozen if the return value is not NULL.
2300 * If this function returns NULL then the page has been unfrozen.
2302 * This function must be called with interrupt disabled.
2304 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2307 unsigned long counters;
2311 freelist = page->freelist;
2312 counters = page->counters;
2314 new.counters = counters;
2315 VM_BUG_ON(!new.frozen);
2317 new.inuse = page->objects;
2318 new.frozen = freelist != NULL;
2320 } while (!__cmpxchg_double_slab(s, page,
2329 * Slow path. The lockless freelist is empty or we need to perform
2332 * Processing is still very fast if new objects have been freed to the
2333 * regular freelist. In that case we simply take over the regular freelist
2334 * as the lockless freelist and zap the regular freelist.
2336 * If that is not working then we fall back to the partial lists. We take the
2337 * first element of the freelist as the object to allocate now and move the
2338 * rest of the freelist to the lockless freelist.
2340 * And if we were unable to get a new slab from the partial slab lists then
2341 * we need to allocate a new slab. This is the slowest path since it involves
2342 * a call to the page allocator and the setup of a new slab.
2344 * Version of __slab_alloc to use when we know that interrupts are
2345 * already disabled (which is the case for bulk allocation).
2347 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2348 unsigned long addr, struct kmem_cache_cpu *c)
2358 if (unlikely(!node_match(page, node))) {
2359 int searchnode = node;
2361 if (node != NUMA_NO_NODE && !node_present_pages(node))
2362 searchnode = node_to_mem_node(node);
2364 if (unlikely(!node_match(page, searchnode))) {
2365 stat(s, ALLOC_NODE_MISMATCH);
2366 deactivate_slab(s, page, c->freelist);
2374 * By rights, we should be searching for a slab page that was
2375 * PFMEMALLOC but right now, we are losing the pfmemalloc
2376 * information when the page leaves the per-cpu allocator
2378 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2379 deactivate_slab(s, page, c->freelist);
2385 /* must check again c->freelist in case of cpu migration or IRQ */
2386 freelist = c->freelist;
2390 freelist = get_freelist(s, page);
2394 stat(s, DEACTIVATE_BYPASS);
2398 stat(s, ALLOC_REFILL);
2402 * freelist is pointing to the list of objects to be used.
2403 * page is pointing to the page from which the objects are obtained.
2404 * That page must be frozen for per cpu allocations to work.
2406 VM_BUG_ON(!c->page->frozen);
2407 c->freelist = get_freepointer(s, freelist);
2408 c->tid = next_tid(c->tid);
2414 page = c->page = c->partial;
2415 c->partial = page->next;
2416 stat(s, CPU_PARTIAL_ALLOC);
2421 freelist = new_slab_objects(s, gfpflags, node, &c);
2423 if (unlikely(!freelist)) {
2424 slab_out_of_memory(s, gfpflags, node);
2429 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2432 /* Only entered in the debug case */
2433 if (kmem_cache_debug(s) &&
2434 !alloc_debug_processing(s, page, freelist, addr))
2435 goto new_slab; /* Slab failed checks. Next slab needed */
2437 deactivate_slab(s, page, get_freepointer(s, freelist));
2444 * Another one that disabled interrupt and compensates for possible
2445 * cpu changes by refetching the per cpu area pointer.
2447 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2448 unsigned long addr, struct kmem_cache_cpu *c)
2451 unsigned long flags;
2453 local_irq_save(flags);
2454 #ifdef CONFIG_PREEMPT
2456 * We may have been preempted and rescheduled on a different
2457 * cpu before disabling interrupts. Need to reload cpu area
2460 c = this_cpu_ptr(s->cpu_slab);
2463 p = ___slab_alloc(s, gfpflags, node, addr, c);
2464 local_irq_restore(flags);
2469 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2470 * have the fastpath folded into their functions. So no function call
2471 * overhead for requests that can be satisfied on the fastpath.
2473 * The fastpath works by first checking if the lockless freelist can be used.
2474 * If not then __slab_alloc is called for slow processing.
2476 * Otherwise we can simply pick the next object from the lockless free list.
2478 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2479 gfp_t gfpflags, int node, unsigned long addr)
2482 struct kmem_cache_cpu *c;
2486 s = slab_pre_alloc_hook(s, gfpflags);
2491 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2492 * enabled. We may switch back and forth between cpus while
2493 * reading from one cpu area. That does not matter as long
2494 * as we end up on the original cpu again when doing the cmpxchg.
2496 * We should guarantee that tid and kmem_cache are retrieved on
2497 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2498 * to check if it is matched or not.
2501 tid = this_cpu_read(s->cpu_slab->tid);
2502 c = raw_cpu_ptr(s->cpu_slab);
2503 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2504 unlikely(tid != READ_ONCE(c->tid)));
2507 * Irqless object alloc/free algorithm used here depends on sequence
2508 * of fetching cpu_slab's data. tid should be fetched before anything
2509 * on c to guarantee that object and page associated with previous tid
2510 * won't be used with current tid. If we fetch tid first, object and
2511 * page could be one associated with next tid and our alloc/free
2512 * request will be failed. In this case, we will retry. So, no problem.
2517 * The transaction ids are globally unique per cpu and per operation on
2518 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2519 * occurs on the right processor and that there was no operation on the
2520 * linked list in between.
2523 object = c->freelist;
2525 if (unlikely(!object || !node_match(page, node))) {
2526 object = __slab_alloc(s, gfpflags, node, addr, c);
2527 stat(s, ALLOC_SLOWPATH);
2529 void *next_object = get_freepointer_safe(s, object);
2532 * The cmpxchg will only match if there was no additional
2533 * operation and if we are on the right processor.
2535 * The cmpxchg does the following atomically (without lock
2537 * 1. Relocate first pointer to the current per cpu area.
2538 * 2. Verify that tid and freelist have not been changed
2539 * 3. If they were not changed replace tid and freelist
2541 * Since this is without lock semantics the protection is only
2542 * against code executing on this cpu *not* from access by
2545 if (unlikely(!this_cpu_cmpxchg_double(
2546 s->cpu_slab->freelist, s->cpu_slab->tid,
2548 next_object, next_tid(tid)))) {
2550 note_cmpxchg_failure("slab_alloc", s, tid);
2553 prefetch_freepointer(s, next_object);
2554 stat(s, ALLOC_FASTPATH);
2557 if (unlikely(gfpflags & __GFP_ZERO) && object)
2558 memset(object, 0, s->object_size);
2560 slab_post_alloc_hook(s, gfpflags, 1, &object);
2565 static __always_inline void *slab_alloc(struct kmem_cache *s,
2566 gfp_t gfpflags, unsigned long addr)
2568 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2571 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2573 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2575 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2580 EXPORT_SYMBOL(kmem_cache_alloc);
2582 #ifdef CONFIG_TRACING
2583 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2585 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2586 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2587 kasan_kmalloc(s, ret, size);
2590 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2594 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2596 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2598 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2599 s->object_size, s->size, gfpflags, node);
2603 EXPORT_SYMBOL(kmem_cache_alloc_node);
2605 #ifdef CONFIG_TRACING
2606 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2608 int node, size_t size)
2610 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2612 trace_kmalloc_node(_RET_IP_, ret,
2613 size, s->size, gfpflags, node);
2615 kasan_kmalloc(s, ret, size);
2618 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2623 * Slow path handling. This may still be called frequently since objects
2624 * have a longer lifetime than the cpu slabs in most processing loads.
2626 * So we still attempt to reduce cache line usage. Just take the slab
2627 * lock and free the item. If there is no additional partial page
2628 * handling required then we can return immediately.
2630 static void __slab_free(struct kmem_cache *s, struct page *page,
2631 void *head, void *tail, int cnt,
2638 unsigned long counters;
2639 struct kmem_cache_node *n = NULL;
2640 unsigned long uninitialized_var(flags);
2642 stat(s, FREE_SLOWPATH);
2644 if (kmem_cache_debug(s) &&
2645 !(n = free_debug_processing(s, page, head, tail, cnt,
2651 spin_unlock_irqrestore(&n->list_lock, flags);
2654 prior = page->freelist;
2655 counters = page->counters;
2656 set_freepointer(s, tail, prior);
2657 new.counters = counters;
2658 was_frozen = new.frozen;
2660 if ((!new.inuse || !prior) && !was_frozen) {
2662 if (kmem_cache_has_cpu_partial(s) && !prior) {
2665 * Slab was on no list before and will be
2667 * We can defer the list move and instead
2672 } else { /* Needs to be taken off a list */
2674 n = get_node(s, page_to_nid(page));
2676 * Speculatively acquire the list_lock.
2677 * If the cmpxchg does not succeed then we may
2678 * drop the list_lock without any processing.
2680 * Otherwise the list_lock will synchronize with
2681 * other processors updating the list of slabs.
2683 spin_lock_irqsave(&n->list_lock, flags);
2688 } while (!cmpxchg_double_slab(s, page,
2696 * If we just froze the page then put it onto the
2697 * per cpu partial list.
2699 if (new.frozen && !was_frozen) {
2700 put_cpu_partial(s, page, 1);
2701 stat(s, CPU_PARTIAL_FREE);
2704 * The list lock was not taken therefore no list
2705 * activity can be necessary.
2708 stat(s, FREE_FROZEN);
2712 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2716 * Objects left in the slab. If it was not on the partial list before
2719 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2720 if (kmem_cache_debug(s))
2721 remove_full(s, n, page);
2722 add_partial(n, page, DEACTIVATE_TO_TAIL);
2723 stat(s, FREE_ADD_PARTIAL);
2725 spin_unlock_irqrestore(&n->list_lock, flags);
2731 * Slab on the partial list.
2733 remove_partial(n, page);
2734 stat(s, FREE_REMOVE_PARTIAL);
2736 /* Slab must be on the full list */
2737 remove_full(s, n, page);
2740 spin_unlock_irqrestore(&n->list_lock, flags);
2742 discard_slab(s, page);
2746 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2747 * can perform fastpath freeing without additional function calls.
2749 * The fastpath is only possible if we are freeing to the current cpu slab
2750 * of this processor. This typically the case if we have just allocated
2753 * If fastpath is not possible then fall back to __slab_free where we deal
2754 * with all sorts of special processing.
2756 * Bulk free of a freelist with several objects (all pointing to the
2757 * same page) possible by specifying head and tail ptr, plus objects
2758 * count (cnt). Bulk free indicated by tail pointer being set.
2760 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2761 void *head, void *tail, int cnt,
2764 void *tail_obj = tail ? : head;
2765 struct kmem_cache_cpu *c;
2768 slab_free_freelist_hook(s, head, tail);
2772 * Determine the currently cpus per cpu slab.
2773 * The cpu may change afterward. However that does not matter since
2774 * data is retrieved via this pointer. If we are on the same cpu
2775 * during the cmpxchg then the free will succeed.
2778 tid = this_cpu_read(s->cpu_slab->tid);
2779 c = raw_cpu_ptr(s->cpu_slab);
2780 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2781 unlikely(tid != READ_ONCE(c->tid)));
2783 /* Same with comment on barrier() in slab_alloc_node() */
2786 if (likely(page == c->page)) {
2787 set_freepointer(s, tail_obj, c->freelist);
2789 if (unlikely(!this_cpu_cmpxchg_double(
2790 s->cpu_slab->freelist, s->cpu_slab->tid,
2792 head, next_tid(tid)))) {
2794 note_cmpxchg_failure("slab_free", s, tid);
2797 stat(s, FREE_FASTPATH);
2799 __slab_free(s, page, head, tail_obj, cnt, addr);
2803 void kmem_cache_free(struct kmem_cache *s, void *x)
2805 s = cache_from_obj(s, x);
2808 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2809 trace_kmem_cache_free(_RET_IP_, x);
2811 EXPORT_SYMBOL(kmem_cache_free);
2813 struct detached_freelist {
2821 * This function progressively scans the array with free objects (with
2822 * a limited look ahead) and extract objects belonging to the same
2823 * page. It builds a detached freelist directly within the given
2824 * page/objects. This can happen without any need for
2825 * synchronization, because the objects are owned by running process.
2826 * The freelist is build up as a single linked list in the objects.
2827 * The idea is, that this detached freelist can then be bulk
2828 * transferred to the real freelist(s), but only requiring a single
2829 * synchronization primitive. Look ahead in the array is limited due
2830 * to performance reasons.
2832 static int build_detached_freelist(struct kmem_cache *s, size_t size,
2833 void **p, struct detached_freelist *df)
2835 size_t first_skipped_index = 0;
2839 /* Always re-init detached_freelist */
2844 } while (!object && size);
2849 /* Start new detached freelist */
2850 set_freepointer(s, object, NULL);
2851 df->page = virt_to_head_page(object);
2853 df->freelist = object;
2854 p[size] = NULL; /* mark object processed */
2860 continue; /* Skip processed objects */
2862 /* df->page is always set at this point */
2863 if (df->page == virt_to_head_page(object)) {
2864 /* Opportunity build freelist */
2865 set_freepointer(s, object, df->freelist);
2866 df->freelist = object;
2868 p[size] = NULL; /* mark object processed */
2873 /* Limit look ahead search */
2877 if (!first_skipped_index)
2878 first_skipped_index = size + 1;
2881 return first_skipped_index;
2885 /* Note that interrupts must be enabled when calling this function. */
2886 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
2892 struct detached_freelist df;
2893 struct kmem_cache *s;
2895 /* Support for memcg */
2896 s = cache_from_obj(orig_s, p[size - 1]);
2898 size = build_detached_freelist(s, size, p, &df);
2899 if (unlikely(!df.page))
2902 slab_free(s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
2903 } while (likely(size));
2905 EXPORT_SYMBOL(kmem_cache_free_bulk);
2907 /* Note that interrupts must be enabled when calling this function. */
2908 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2911 struct kmem_cache_cpu *c;
2914 /* memcg and kmem_cache debug support */
2915 s = slab_pre_alloc_hook(s, flags);
2919 * Drain objects in the per cpu slab, while disabling local
2920 * IRQs, which protects against PREEMPT and interrupts
2921 * handlers invoking normal fastpath.
2923 local_irq_disable();
2924 c = this_cpu_ptr(s->cpu_slab);
2926 for (i = 0; i < size; i++) {
2927 void *object = c->freelist;
2929 if (unlikely(!object)) {
2931 * Invoking slow path likely have side-effect
2932 * of re-populating per CPU c->freelist
2934 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2936 if (unlikely(!p[i]))
2939 c = this_cpu_ptr(s->cpu_slab);
2940 continue; /* goto for-loop */
2942 c->freelist = get_freepointer(s, object);
2945 c->tid = next_tid(c->tid);
2948 /* Clear memory outside IRQ disabled fastpath loop */
2949 if (unlikely(flags & __GFP_ZERO)) {
2952 for (j = 0; j < i; j++)
2953 memset(p[j], 0, s->object_size);
2956 /* memcg and kmem_cache debug support */
2957 slab_post_alloc_hook(s, flags, size, p);
2961 slab_post_alloc_hook(s, flags, i, p);
2962 __kmem_cache_free_bulk(s, i, p);
2965 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2969 * Object placement in a slab is made very easy because we always start at
2970 * offset 0. If we tune the size of the object to the alignment then we can
2971 * get the required alignment by putting one properly sized object after
2974 * Notice that the allocation order determines the sizes of the per cpu
2975 * caches. Each processor has always one slab available for allocations.
2976 * Increasing the allocation order reduces the number of times that slabs
2977 * must be moved on and off the partial lists and is therefore a factor in
2982 * Mininum / Maximum order of slab pages. This influences locking overhead
2983 * and slab fragmentation. A higher order reduces the number of partial slabs
2984 * and increases the number of allocations possible without having to
2985 * take the list_lock.
2987 static int slub_min_order;
2988 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2989 static int slub_min_objects;
2992 * Calculate the order of allocation given an slab object size.
2994 * The order of allocation has significant impact on performance and other
2995 * system components. Generally order 0 allocations should be preferred since
2996 * order 0 does not cause fragmentation in the page allocator. Larger objects
2997 * be problematic to put into order 0 slabs because there may be too much
2998 * unused space left. We go to a higher order if more than 1/16th of the slab
3001 * In order to reach satisfactory performance we must ensure that a minimum
3002 * number of objects is in one slab. Otherwise we may generate too much
3003 * activity on the partial lists which requires taking the list_lock. This is
3004 * less a concern for large slabs though which are rarely used.
3006 * slub_max_order specifies the order where we begin to stop considering the
3007 * number of objects in a slab as critical. If we reach slub_max_order then
3008 * we try to keep the page order as low as possible. So we accept more waste
3009 * of space in favor of a small page order.
3011 * Higher order allocations also allow the placement of more objects in a
3012 * slab and thereby reduce object handling overhead. If the user has
3013 * requested a higher mininum order then we start with that one instead of
3014 * the smallest order which will fit the object.
3016 static inline int slab_order(int size, int min_objects,
3017 int max_order, int fract_leftover, int reserved)
3021 int min_order = slub_min_order;
3023 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3024 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3026 for (order = max(min_order, get_order(min_objects * size + reserved));
3027 order <= max_order; order++) {
3029 unsigned long slab_size = PAGE_SIZE << order;
3031 rem = (slab_size - reserved) % size;
3033 if (rem <= slab_size / fract_leftover)
3040 static inline int calculate_order(int size, int reserved)
3048 * Attempt to find best configuration for a slab. This
3049 * works by first attempting to generate a layout with
3050 * the best configuration and backing off gradually.
3052 * First we increase the acceptable waste in a slab. Then
3053 * we reduce the minimum objects required in a slab.
3055 min_objects = slub_min_objects;
3057 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3058 max_objects = order_objects(slub_max_order, size, reserved);
3059 min_objects = min(min_objects, max_objects);
3061 while (min_objects > 1) {
3063 while (fraction >= 4) {
3064 order = slab_order(size, min_objects,
3065 slub_max_order, fraction, reserved);
3066 if (order <= slub_max_order)
3074 * We were unable to place multiple objects in a slab. Now
3075 * lets see if we can place a single object there.
3077 order = slab_order(size, 1, slub_max_order, 1, reserved);
3078 if (order <= slub_max_order)
3082 * Doh this slab cannot be placed using slub_max_order.
3084 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3085 if (order < MAX_ORDER)
3091 init_kmem_cache_node(struct kmem_cache_node *n)
3094 spin_lock_init(&n->list_lock);
3095 INIT_LIST_HEAD(&n->partial);
3096 #ifdef CONFIG_SLUB_DEBUG
3097 atomic_long_set(&n->nr_slabs, 0);
3098 atomic_long_set(&n->total_objects, 0);
3099 INIT_LIST_HEAD(&n->full);
3103 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3105 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3106 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3109 * Must align to double word boundary for the double cmpxchg
3110 * instructions to work; see __pcpu_double_call_return_bool().
3112 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3113 2 * sizeof(void *));
3118 init_kmem_cache_cpus(s);
3123 static struct kmem_cache *kmem_cache_node;
3126 * No kmalloc_node yet so do it by hand. We know that this is the first
3127 * slab on the node for this slabcache. There are no concurrent accesses
3130 * Note that this function only works on the kmem_cache_node
3131 * when allocating for the kmem_cache_node. This is used for bootstrapping
3132 * memory on a fresh node that has no slab structures yet.
3134 static void early_kmem_cache_node_alloc(int node)
3137 struct kmem_cache_node *n;
3139 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3141 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3144 if (page_to_nid(page) != node) {
3145 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3146 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3151 page->freelist = get_freepointer(kmem_cache_node, n);
3154 kmem_cache_node->node[node] = n;
3155 #ifdef CONFIG_SLUB_DEBUG
3156 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3157 init_tracking(kmem_cache_node, n);
3159 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3160 init_kmem_cache_node(n);
3161 inc_slabs_node(kmem_cache_node, node, page->objects);
3164 * No locks need to be taken here as it has just been
3165 * initialized and there is no concurrent access.
3167 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3170 static void free_kmem_cache_nodes(struct kmem_cache *s)
3173 struct kmem_cache_node *n;
3175 for_each_kmem_cache_node(s, node, n) {
3176 kmem_cache_free(kmem_cache_node, n);
3177 s->node[node] = NULL;
3181 void __kmem_cache_release(struct kmem_cache *s)
3183 free_percpu(s->cpu_slab);
3184 free_kmem_cache_nodes(s);
3187 static int init_kmem_cache_nodes(struct kmem_cache *s)
3191 for_each_node_state(node, N_NORMAL_MEMORY) {
3192 struct kmem_cache_node *n;
3194 if (slab_state == DOWN) {
3195 early_kmem_cache_node_alloc(node);
3198 n = kmem_cache_alloc_node(kmem_cache_node,
3202 free_kmem_cache_nodes(s);
3207 init_kmem_cache_node(n);
3212 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3214 if (min < MIN_PARTIAL)
3216 else if (min > MAX_PARTIAL)
3218 s->min_partial = min;
3222 * calculate_sizes() determines the order and the distribution of data within
3225 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3227 unsigned long flags = s->flags;
3228 unsigned long size = s->object_size;
3232 * Round up object size to the next word boundary. We can only
3233 * place the free pointer at word boundaries and this determines
3234 * the possible location of the free pointer.
3236 size = ALIGN(size, sizeof(void *));
3238 #ifdef CONFIG_SLUB_DEBUG
3240 * Determine if we can poison the object itself. If the user of
3241 * the slab may touch the object after free or before allocation
3242 * then we should never poison the object itself.
3244 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3246 s->flags |= __OBJECT_POISON;
3248 s->flags &= ~__OBJECT_POISON;
3252 * If we are Redzoning then check if there is some space between the
3253 * end of the object and the free pointer. If not then add an
3254 * additional word to have some bytes to store Redzone information.
3256 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3257 size += sizeof(void *);
3261 * With that we have determined the number of bytes in actual use
3262 * by the object. This is the potential offset to the free pointer.
3266 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3269 * Relocate free pointer after the object if it is not
3270 * permitted to overwrite the first word of the object on
3273 * This is the case if we do RCU, have a constructor or
3274 * destructor or are poisoning the objects.
3277 size += sizeof(void *);
3280 #ifdef CONFIG_SLUB_DEBUG
3281 if (flags & SLAB_STORE_USER)
3283 * Need to store information about allocs and frees after
3286 size += 2 * sizeof(struct track);
3288 if (flags & SLAB_RED_ZONE)
3290 * Add some empty padding so that we can catch
3291 * overwrites from earlier objects rather than let
3292 * tracking information or the free pointer be
3293 * corrupted if a user writes before the start
3296 size += sizeof(void *);
3300 * SLUB stores one object immediately after another beginning from
3301 * offset 0. In order to align the objects we have to simply size
3302 * each object to conform to the alignment.
3304 size = ALIGN(size, s->align);
3306 if (forced_order >= 0)
3307 order = forced_order;
3309 order = calculate_order(size, s->reserved);
3316 s->allocflags |= __GFP_COMP;
3318 if (s->flags & SLAB_CACHE_DMA)
3319 s->allocflags |= GFP_DMA;
3321 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3322 s->allocflags |= __GFP_RECLAIMABLE;
3325 * Determine the number of objects per slab
3327 s->oo = oo_make(order, size, s->reserved);
3328 s->min = oo_make(get_order(size), size, s->reserved);
3329 if (oo_objects(s->oo) > oo_objects(s->max))
3332 return !!oo_objects(s->oo);
3335 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3337 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3340 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3341 s->reserved = sizeof(struct rcu_head);
3343 if (!calculate_sizes(s, -1))
3345 if (disable_higher_order_debug) {
3347 * Disable debugging flags that store metadata if the min slab
3350 if (get_order(s->size) > get_order(s->object_size)) {
3351 s->flags &= ~DEBUG_METADATA_FLAGS;
3353 if (!calculate_sizes(s, -1))
3358 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3359 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3360 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3361 /* Enable fast mode */
3362 s->flags |= __CMPXCHG_DOUBLE;
3366 * The larger the object size is, the more pages we want on the partial
3367 * list to avoid pounding the page allocator excessively.
3369 set_min_partial(s, ilog2(s->size) / 2);
3372 * cpu_partial determined the maximum number of objects kept in the
3373 * per cpu partial lists of a processor.
3375 * Per cpu partial lists mainly contain slabs that just have one
3376 * object freed. If they are used for allocation then they can be
3377 * filled up again with minimal effort. The slab will never hit the
3378 * per node partial lists and therefore no locking will be required.
3380 * This setting also determines
3382 * A) The number of objects from per cpu partial slabs dumped to the
3383 * per node list when we reach the limit.
3384 * B) The number of objects in cpu partial slabs to extract from the
3385 * per node list when we run out of per cpu objects. We only fetch
3386 * 50% to keep some capacity around for frees.
3388 if (!kmem_cache_has_cpu_partial(s))
3390 else if (s->size >= PAGE_SIZE)
3392 else if (s->size >= 1024)
3394 else if (s->size >= 256)
3395 s->cpu_partial = 13;
3397 s->cpu_partial = 30;
3400 s->remote_node_defrag_ratio = 1000;
3402 if (!init_kmem_cache_nodes(s))
3405 if (alloc_kmem_cache_cpus(s))
3408 free_kmem_cache_nodes(s);
3410 if (flags & SLAB_PANIC)
3411 panic("Cannot create slab %s size=%lu realsize=%u "
3412 "order=%u offset=%u flags=%lx\n",
3413 s->name, (unsigned long)s->size, s->size,
3414 oo_order(s->oo), s->offset, flags);
3418 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3421 #ifdef CONFIG_SLUB_DEBUG
3422 void *addr = page_address(page);
3424 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3425 sizeof(long), GFP_ATOMIC);
3428 slab_err(s, page, text, s->name);
3431 get_map(s, page, map);
3432 for_each_object(p, s, addr, page->objects) {
3434 if (!test_bit(slab_index(p, s, addr), map)) {
3435 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3436 print_tracking(s, p);
3445 * Attempt to free all partial slabs on a node.
3446 * This is called from __kmem_cache_shutdown(). We must take list_lock
3447 * because sysfs file might still access partial list after the shutdowning.
3449 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3451 struct page *page, *h;
3453 BUG_ON(irqs_disabled());
3454 spin_lock_irq(&n->list_lock);
3455 list_for_each_entry_safe(page, h, &n->partial, lru) {
3457 remove_partial(n, page);
3458 discard_slab(s, page);
3460 list_slab_objects(s, page,
3461 "Objects remaining in %s on __kmem_cache_shutdown()");
3464 spin_unlock_irq(&n->list_lock);
3468 * Release all resources used by a slab cache.
3470 int __kmem_cache_shutdown(struct kmem_cache *s)
3473 struct kmem_cache_node *n;
3476 /* Attempt to free all objects */
3477 for_each_kmem_cache_node(s, node, n) {
3479 if (n->nr_partial || slabs_node(s, node))
3485 /********************************************************************
3487 *******************************************************************/
3489 static int __init setup_slub_min_order(char *str)
3491 get_option(&str, &slub_min_order);
3496 __setup("slub_min_order=", setup_slub_min_order);
3498 static int __init setup_slub_max_order(char *str)
3500 get_option(&str, &slub_max_order);
3501 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3506 __setup("slub_max_order=", setup_slub_max_order);
3508 static int __init setup_slub_min_objects(char *str)
3510 get_option(&str, &slub_min_objects);
3515 __setup("slub_min_objects=", setup_slub_min_objects);
3517 void *__kmalloc(size_t size, gfp_t flags)
3519 struct kmem_cache *s;
3522 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3523 return kmalloc_large(size, flags);
3525 s = kmalloc_slab(size, flags);
3527 if (unlikely(ZERO_OR_NULL_PTR(s)))
3530 ret = slab_alloc(s, flags, _RET_IP_);
3532 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3534 kasan_kmalloc(s, ret, size);
3538 EXPORT_SYMBOL(__kmalloc);
3541 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3546 flags |= __GFP_COMP | __GFP_NOTRACK;
3547 page = alloc_kmem_pages_node(node, flags, get_order(size));
3549 ptr = page_address(page);
3551 kmalloc_large_node_hook(ptr, size, flags);
3555 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3557 struct kmem_cache *s;
3560 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3561 ret = kmalloc_large_node(size, flags, node);
3563 trace_kmalloc_node(_RET_IP_, ret,
3564 size, PAGE_SIZE << get_order(size),
3570 s = kmalloc_slab(size, flags);
3572 if (unlikely(ZERO_OR_NULL_PTR(s)))
3575 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3577 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3579 kasan_kmalloc(s, ret, size);
3583 EXPORT_SYMBOL(__kmalloc_node);
3586 static size_t __ksize(const void *object)
3590 if (unlikely(object == ZERO_SIZE_PTR))
3593 page = virt_to_head_page(object);
3595 if (unlikely(!PageSlab(page))) {
3596 WARN_ON(!PageCompound(page));
3597 return PAGE_SIZE << compound_order(page);
3600 return slab_ksize(page->slab_cache);
3603 size_t ksize(const void *object)
3605 size_t size = __ksize(object);
3606 /* We assume that ksize callers could use whole allocated area,
3607 so we need unpoison this area. */
3608 kasan_krealloc(object, size);
3611 EXPORT_SYMBOL(ksize);
3613 void kfree(const void *x)
3616 void *object = (void *)x;
3618 trace_kfree(_RET_IP_, x);
3620 if (unlikely(ZERO_OR_NULL_PTR(x)))
3623 page = virt_to_head_page(x);
3624 if (unlikely(!PageSlab(page))) {
3625 BUG_ON(!PageCompound(page));
3627 __free_kmem_pages(page, compound_order(page));
3630 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3632 EXPORT_SYMBOL(kfree);
3634 #define SHRINK_PROMOTE_MAX 32
3637 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3638 * up most to the head of the partial lists. New allocations will then
3639 * fill those up and thus they can be removed from the partial lists.
3641 * The slabs with the least items are placed last. This results in them
3642 * being allocated from last increasing the chance that the last objects
3643 * are freed in them.
3645 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3649 struct kmem_cache_node *n;
3652 struct list_head discard;
3653 struct list_head promote[SHRINK_PROMOTE_MAX];
3654 unsigned long flags;
3659 * Disable empty slabs caching. Used to avoid pinning offline
3660 * memory cgroups by kmem pages that can be freed.
3666 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3667 * so we have to make sure the change is visible.
3669 kick_all_cpus_sync();
3673 for_each_kmem_cache_node(s, node, n) {
3674 INIT_LIST_HEAD(&discard);
3675 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3676 INIT_LIST_HEAD(promote + i);
3678 spin_lock_irqsave(&n->list_lock, flags);
3681 * Build lists of slabs to discard or promote.
3683 * Note that concurrent frees may occur while we hold the
3684 * list_lock. page->inuse here is the upper limit.
3686 list_for_each_entry_safe(page, t, &n->partial, lru) {
3687 int free = page->objects - page->inuse;
3689 /* Do not reread page->inuse */
3692 /* We do not keep full slabs on the list */
3695 if (free == page->objects) {
3696 list_move(&page->lru, &discard);
3698 } else if (free <= SHRINK_PROMOTE_MAX)
3699 list_move(&page->lru, promote + free - 1);
3703 * Promote the slabs filled up most to the head of the
3706 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3707 list_splice(promote + i, &n->partial);
3709 spin_unlock_irqrestore(&n->list_lock, flags);
3711 /* Release empty slabs */
3712 list_for_each_entry_safe(page, t, &discard, lru)
3713 discard_slab(s, page);
3715 if (slabs_node(s, node))
3722 static int slab_mem_going_offline_callback(void *arg)
3724 struct kmem_cache *s;
3726 mutex_lock(&slab_mutex);
3727 list_for_each_entry(s, &slab_caches, list)
3728 __kmem_cache_shrink(s, false);
3729 mutex_unlock(&slab_mutex);
3734 static void slab_mem_offline_callback(void *arg)
3736 struct kmem_cache_node *n;
3737 struct kmem_cache *s;
3738 struct memory_notify *marg = arg;
3741 offline_node = marg->status_change_nid_normal;
3744 * If the node still has available memory. we need kmem_cache_node
3747 if (offline_node < 0)
3750 mutex_lock(&slab_mutex);
3751 list_for_each_entry(s, &slab_caches, list) {
3752 n = get_node(s, offline_node);
3755 * if n->nr_slabs > 0, slabs still exist on the node
3756 * that is going down. We were unable to free them,
3757 * and offline_pages() function shouldn't call this
3758 * callback. So, we must fail.
3760 BUG_ON(slabs_node(s, offline_node));
3762 s->node[offline_node] = NULL;
3763 kmem_cache_free(kmem_cache_node, n);
3766 mutex_unlock(&slab_mutex);
3769 static int slab_mem_going_online_callback(void *arg)
3771 struct kmem_cache_node *n;
3772 struct kmem_cache *s;
3773 struct memory_notify *marg = arg;
3774 int nid = marg->status_change_nid_normal;
3778 * If the node's memory is already available, then kmem_cache_node is
3779 * already created. Nothing to do.
3785 * We are bringing a node online. No memory is available yet. We must
3786 * allocate a kmem_cache_node structure in order to bring the node
3789 mutex_lock(&slab_mutex);
3790 list_for_each_entry(s, &slab_caches, list) {
3792 * XXX: kmem_cache_alloc_node will fallback to other nodes
3793 * since memory is not yet available from the node that
3796 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3801 init_kmem_cache_node(n);
3805 mutex_unlock(&slab_mutex);
3809 static int slab_memory_callback(struct notifier_block *self,
3810 unsigned long action, void *arg)
3815 case MEM_GOING_ONLINE:
3816 ret = slab_mem_going_online_callback(arg);
3818 case MEM_GOING_OFFLINE:
3819 ret = slab_mem_going_offline_callback(arg);
3822 case MEM_CANCEL_ONLINE:
3823 slab_mem_offline_callback(arg);
3826 case MEM_CANCEL_OFFLINE:
3830 ret = notifier_from_errno(ret);
3836 static struct notifier_block slab_memory_callback_nb = {
3837 .notifier_call = slab_memory_callback,
3838 .priority = SLAB_CALLBACK_PRI,
3841 /********************************************************************
3842 * Basic setup of slabs
3843 *******************************************************************/
3846 * Used for early kmem_cache structures that were allocated using
3847 * the page allocator. Allocate them properly then fix up the pointers
3848 * that may be pointing to the wrong kmem_cache structure.
3851 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3854 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3855 struct kmem_cache_node *n;
3857 memcpy(s, static_cache, kmem_cache->object_size);
3860 * This runs very early, and only the boot processor is supposed to be
3861 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3864 __flush_cpu_slab(s, smp_processor_id());
3865 for_each_kmem_cache_node(s, node, n) {
3868 list_for_each_entry(p, &n->partial, lru)
3871 #ifdef CONFIG_SLUB_DEBUG
3872 list_for_each_entry(p, &n->full, lru)
3876 slab_init_memcg_params(s);
3877 list_add(&s->list, &slab_caches);
3881 void __init kmem_cache_init(void)
3883 static __initdata struct kmem_cache boot_kmem_cache,
3884 boot_kmem_cache_node;
3886 if (debug_guardpage_minorder())
3889 kmem_cache_node = &boot_kmem_cache_node;
3890 kmem_cache = &boot_kmem_cache;
3892 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3893 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3895 register_hotmemory_notifier(&slab_memory_callback_nb);
3897 /* Able to allocate the per node structures */
3898 slab_state = PARTIAL;
3900 create_boot_cache(kmem_cache, "kmem_cache",
3901 offsetof(struct kmem_cache, node) +
3902 nr_node_ids * sizeof(struct kmem_cache_node *),
3903 SLAB_HWCACHE_ALIGN);
3905 kmem_cache = bootstrap(&boot_kmem_cache);
3908 * Allocate kmem_cache_node properly from the kmem_cache slab.
3909 * kmem_cache_node is separately allocated so no need to
3910 * update any list pointers.
3912 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3914 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3915 setup_kmalloc_cache_index_table();
3916 create_kmalloc_caches(0);
3919 register_cpu_notifier(&slab_notifier);
3922 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3924 slub_min_order, slub_max_order, slub_min_objects,
3925 nr_cpu_ids, nr_node_ids);
3928 void __init kmem_cache_init_late(void)
3933 __kmem_cache_alias(const char *name, size_t size, size_t align,
3934 unsigned long flags, void (*ctor)(void *))
3936 struct kmem_cache *s, *c;
3938 s = find_mergeable(size, align, flags, name, ctor);
3943 * Adjust the object sizes so that we clear
3944 * the complete object on kzalloc.
3946 s->object_size = max(s->object_size, (int)size);
3947 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3949 for_each_memcg_cache(c, s) {
3950 c->object_size = s->object_size;
3951 c->inuse = max_t(int, c->inuse,
3952 ALIGN(size, sizeof(void *)));
3955 if (sysfs_slab_alias(s, name)) {
3964 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3968 err = kmem_cache_open(s, flags);
3972 /* Mutex is not taken during early boot */
3973 if (slab_state <= UP)
3976 memcg_propagate_slab_attrs(s);
3977 err = sysfs_slab_add(s);
3979 __kmem_cache_release(s);
3986 * Use the cpu notifier to insure that the cpu slabs are flushed when
3989 static int slab_cpuup_callback(struct notifier_block *nfb,
3990 unsigned long action, void *hcpu)
3992 long cpu = (long)hcpu;
3993 struct kmem_cache *s;
3994 unsigned long flags;
3997 case CPU_UP_CANCELED:
3998 case CPU_UP_CANCELED_FROZEN:
4000 case CPU_DEAD_FROZEN:
4001 mutex_lock(&slab_mutex);
4002 list_for_each_entry(s, &slab_caches, list) {
4003 local_irq_save(flags);
4004 __flush_cpu_slab(s, cpu);
4005 local_irq_restore(flags);
4007 mutex_unlock(&slab_mutex);
4015 static struct notifier_block slab_notifier = {
4016 .notifier_call = slab_cpuup_callback
4021 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4023 struct kmem_cache *s;
4026 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4027 return kmalloc_large(size, gfpflags);
4029 s = kmalloc_slab(size, gfpflags);
4031 if (unlikely(ZERO_OR_NULL_PTR(s)))
4034 ret = slab_alloc(s, gfpflags, caller);
4036 /* Honor the call site pointer we received. */
4037 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4043 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4044 int node, unsigned long caller)
4046 struct kmem_cache *s;
4049 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4050 ret = kmalloc_large_node(size, gfpflags, node);
4052 trace_kmalloc_node(caller, ret,
4053 size, PAGE_SIZE << get_order(size),
4059 s = kmalloc_slab(size, gfpflags);
4061 if (unlikely(ZERO_OR_NULL_PTR(s)))
4064 ret = slab_alloc_node(s, gfpflags, node, caller);
4066 /* Honor the call site pointer we received. */
4067 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4074 static int count_inuse(struct page *page)
4079 static int count_total(struct page *page)
4081 return page->objects;
4085 #ifdef CONFIG_SLUB_DEBUG
4086 static int validate_slab(struct kmem_cache *s, struct page *page,
4090 void *addr = page_address(page);
4092 if (!check_slab(s, page) ||
4093 !on_freelist(s, page, NULL))
4096 /* Now we know that a valid freelist exists */
4097 bitmap_zero(map, page->objects);
4099 get_map(s, page, map);
4100 for_each_object(p, s, addr, page->objects) {
4101 if (test_bit(slab_index(p, s, addr), map))
4102 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4106 for_each_object(p, s, addr, page->objects)
4107 if (!test_bit(slab_index(p, s, addr), map))
4108 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4113 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4117 validate_slab(s, page, map);
4121 static int validate_slab_node(struct kmem_cache *s,
4122 struct kmem_cache_node *n, unsigned long *map)
4124 unsigned long count = 0;
4126 unsigned long flags;
4128 spin_lock_irqsave(&n->list_lock, flags);
4130 list_for_each_entry(page, &n->partial, lru) {
4131 validate_slab_slab(s, page, map);
4134 if (count != n->nr_partial)
4135 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4136 s->name, count, n->nr_partial);
4138 if (!(s->flags & SLAB_STORE_USER))
4141 list_for_each_entry(page, &n->full, lru) {
4142 validate_slab_slab(s, page, map);
4145 if (count != atomic_long_read(&n->nr_slabs))
4146 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4147 s->name, count, atomic_long_read(&n->nr_slabs));
4150 spin_unlock_irqrestore(&n->list_lock, flags);
4154 static long validate_slab_cache(struct kmem_cache *s)
4157 unsigned long count = 0;
4158 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4159 sizeof(unsigned long), GFP_KERNEL);
4160 struct kmem_cache_node *n;
4166 for_each_kmem_cache_node(s, node, n)
4167 count += validate_slab_node(s, n, map);
4172 * Generate lists of code addresses where slabcache objects are allocated
4177 unsigned long count;
4184 DECLARE_BITMAP(cpus, NR_CPUS);
4190 unsigned long count;
4191 struct location *loc;
4194 static void free_loc_track(struct loc_track *t)
4197 free_pages((unsigned long)t->loc,
4198 get_order(sizeof(struct location) * t->max));
4201 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4206 order = get_order(sizeof(struct location) * max);
4208 l = (void *)__get_free_pages(flags, order);
4213 memcpy(l, t->loc, sizeof(struct location) * t->count);
4221 static int add_location(struct loc_track *t, struct kmem_cache *s,
4222 const struct track *track)
4224 long start, end, pos;
4226 unsigned long caddr;
4227 unsigned long age = jiffies - track->when;
4233 pos = start + (end - start + 1) / 2;
4236 * There is nothing at "end". If we end up there
4237 * we need to add something to before end.
4242 caddr = t->loc[pos].addr;
4243 if (track->addr == caddr) {
4249 if (age < l->min_time)
4251 if (age > l->max_time)
4254 if (track->pid < l->min_pid)
4255 l->min_pid = track->pid;
4256 if (track->pid > l->max_pid)
4257 l->max_pid = track->pid;
4259 cpumask_set_cpu(track->cpu,
4260 to_cpumask(l->cpus));
4262 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4266 if (track->addr < caddr)
4273 * Not found. Insert new tracking element.
4275 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4281 (t->count - pos) * sizeof(struct location));
4284 l->addr = track->addr;
4288 l->min_pid = track->pid;
4289 l->max_pid = track->pid;
4290 cpumask_clear(to_cpumask(l->cpus));
4291 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4292 nodes_clear(l->nodes);
4293 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4297 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4298 struct page *page, enum track_item alloc,
4301 void *addr = page_address(page);
4304 bitmap_zero(map, page->objects);
4305 get_map(s, page, map);
4307 for_each_object(p, s, addr, page->objects)
4308 if (!test_bit(slab_index(p, s, addr), map))
4309 add_location(t, s, get_track(s, p, alloc));
4312 static int list_locations(struct kmem_cache *s, char *buf,
4313 enum track_item alloc)
4317 struct loc_track t = { 0, 0, NULL };
4319 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4320 sizeof(unsigned long), GFP_KERNEL);
4321 struct kmem_cache_node *n;
4323 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4326 return sprintf(buf, "Out of memory\n");
4328 /* Push back cpu slabs */
4331 for_each_kmem_cache_node(s, node, n) {
4332 unsigned long flags;
4335 if (!atomic_long_read(&n->nr_slabs))
4338 spin_lock_irqsave(&n->list_lock, flags);
4339 list_for_each_entry(page, &n->partial, lru)
4340 process_slab(&t, s, page, alloc, map);
4341 list_for_each_entry(page, &n->full, lru)
4342 process_slab(&t, s, page, alloc, map);
4343 spin_unlock_irqrestore(&n->list_lock, flags);
4346 for (i = 0; i < t.count; i++) {
4347 struct location *l = &t.loc[i];
4349 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4351 len += sprintf(buf + len, "%7ld ", l->count);
4354 len += sprintf(buf + len, "%pS", (void *)l->addr);
4356 len += sprintf(buf + len, "<not-available>");
4358 if (l->sum_time != l->min_time) {
4359 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4361 (long)div_u64(l->sum_time, l->count),
4364 len += sprintf(buf + len, " age=%ld",
4367 if (l->min_pid != l->max_pid)
4368 len += sprintf(buf + len, " pid=%ld-%ld",
4369 l->min_pid, l->max_pid);
4371 len += sprintf(buf + len, " pid=%ld",
4374 if (num_online_cpus() > 1 &&
4375 !cpumask_empty(to_cpumask(l->cpus)) &&
4376 len < PAGE_SIZE - 60)
4377 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4379 cpumask_pr_args(to_cpumask(l->cpus)));
4381 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4382 len < PAGE_SIZE - 60)
4383 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4385 nodemask_pr_args(&l->nodes));
4387 len += sprintf(buf + len, "\n");
4393 len += sprintf(buf, "No data\n");
4398 #ifdef SLUB_RESILIENCY_TEST
4399 static void __init resiliency_test(void)
4403 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4405 pr_err("SLUB resiliency testing\n");
4406 pr_err("-----------------------\n");
4407 pr_err("A. Corruption after allocation\n");
4409 p = kzalloc(16, GFP_KERNEL);
4411 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4414 validate_slab_cache(kmalloc_caches[4]);
4416 /* Hmmm... The next two are dangerous */
4417 p = kzalloc(32, GFP_KERNEL);
4418 p[32 + sizeof(void *)] = 0x34;
4419 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4421 pr_err("If allocated object is overwritten then not detectable\n\n");
4423 validate_slab_cache(kmalloc_caches[5]);
4424 p = kzalloc(64, GFP_KERNEL);
4425 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4427 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4429 pr_err("If allocated object is overwritten then not detectable\n\n");
4430 validate_slab_cache(kmalloc_caches[6]);
4432 pr_err("\nB. Corruption after free\n");
4433 p = kzalloc(128, GFP_KERNEL);
4436 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4437 validate_slab_cache(kmalloc_caches[7]);
4439 p = kzalloc(256, GFP_KERNEL);
4442 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4443 validate_slab_cache(kmalloc_caches[8]);
4445 p = kzalloc(512, GFP_KERNEL);
4448 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4449 validate_slab_cache(kmalloc_caches[9]);
4453 static void resiliency_test(void) {};
4458 enum slab_stat_type {
4459 SL_ALL, /* All slabs */
4460 SL_PARTIAL, /* Only partially allocated slabs */
4461 SL_CPU, /* Only slabs used for cpu caches */
4462 SL_OBJECTS, /* Determine allocated objects not slabs */
4463 SL_TOTAL /* Determine object capacity not slabs */
4466 #define SO_ALL (1 << SL_ALL)
4467 #define SO_PARTIAL (1 << SL_PARTIAL)
4468 #define SO_CPU (1 << SL_CPU)
4469 #define SO_OBJECTS (1 << SL_OBJECTS)
4470 #define SO_TOTAL (1 << SL_TOTAL)
4472 static ssize_t show_slab_objects(struct kmem_cache *s,
4473 char *buf, unsigned long flags)
4475 unsigned long total = 0;
4478 unsigned long *nodes;
4480 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4484 if (flags & SO_CPU) {
4487 for_each_possible_cpu(cpu) {
4488 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4493 page = READ_ONCE(c->page);
4497 node = page_to_nid(page);
4498 if (flags & SO_TOTAL)
4500 else if (flags & SO_OBJECTS)
4508 page = READ_ONCE(c->partial);
4510 node = page_to_nid(page);
4511 if (flags & SO_TOTAL)
4513 else if (flags & SO_OBJECTS)
4524 #ifdef CONFIG_SLUB_DEBUG
4525 if (flags & SO_ALL) {
4526 struct kmem_cache_node *n;
4528 for_each_kmem_cache_node(s, node, n) {
4530 if (flags & SO_TOTAL)
4531 x = atomic_long_read(&n->total_objects);
4532 else if (flags & SO_OBJECTS)
4533 x = atomic_long_read(&n->total_objects) -
4534 count_partial(n, count_free);
4536 x = atomic_long_read(&n->nr_slabs);
4543 if (flags & SO_PARTIAL) {
4544 struct kmem_cache_node *n;
4546 for_each_kmem_cache_node(s, node, n) {
4547 if (flags & SO_TOTAL)
4548 x = count_partial(n, count_total);
4549 else if (flags & SO_OBJECTS)
4550 x = count_partial(n, count_inuse);
4557 x = sprintf(buf, "%lu", total);
4559 for (node = 0; node < nr_node_ids; node++)
4561 x += sprintf(buf + x, " N%d=%lu",
4566 return x + sprintf(buf + x, "\n");
4569 #ifdef CONFIG_SLUB_DEBUG
4570 static int any_slab_objects(struct kmem_cache *s)
4573 struct kmem_cache_node *n;
4575 for_each_kmem_cache_node(s, node, n)
4576 if (atomic_long_read(&n->total_objects))
4583 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4584 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4586 struct slab_attribute {
4587 struct attribute attr;
4588 ssize_t (*show)(struct kmem_cache *s, char *buf);
4589 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4592 #define SLAB_ATTR_RO(_name) \
4593 static struct slab_attribute _name##_attr = \
4594 __ATTR(_name, 0400, _name##_show, NULL)
4596 #define SLAB_ATTR(_name) \
4597 static struct slab_attribute _name##_attr = \
4598 __ATTR(_name, 0600, _name##_show, _name##_store)
4600 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4602 return sprintf(buf, "%d\n", s->size);
4604 SLAB_ATTR_RO(slab_size);
4606 static ssize_t align_show(struct kmem_cache *s, char *buf)
4608 return sprintf(buf, "%d\n", s->align);
4610 SLAB_ATTR_RO(align);
4612 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4614 return sprintf(buf, "%d\n", s->object_size);
4616 SLAB_ATTR_RO(object_size);
4618 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4620 return sprintf(buf, "%d\n", oo_objects(s->oo));
4622 SLAB_ATTR_RO(objs_per_slab);
4624 static ssize_t order_store(struct kmem_cache *s,
4625 const char *buf, size_t length)
4627 unsigned long order;
4630 err = kstrtoul(buf, 10, &order);
4634 if (order > slub_max_order || order < slub_min_order)
4637 calculate_sizes(s, order);
4641 static ssize_t order_show(struct kmem_cache *s, char *buf)
4643 return sprintf(buf, "%d\n", oo_order(s->oo));
4647 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4649 return sprintf(buf, "%lu\n", s->min_partial);
4652 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4658 err = kstrtoul(buf, 10, &min);
4662 set_min_partial(s, min);
4665 SLAB_ATTR(min_partial);
4667 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4669 return sprintf(buf, "%u\n", s->cpu_partial);
4672 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4675 unsigned long objects;
4678 err = kstrtoul(buf, 10, &objects);
4681 if (objects && !kmem_cache_has_cpu_partial(s))
4684 s->cpu_partial = objects;
4688 SLAB_ATTR(cpu_partial);
4690 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4694 return sprintf(buf, "%pS\n", s->ctor);
4698 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4700 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4702 SLAB_ATTR_RO(aliases);
4704 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4706 return show_slab_objects(s, buf, SO_PARTIAL);
4708 SLAB_ATTR_RO(partial);
4710 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4712 return show_slab_objects(s, buf, SO_CPU);
4714 SLAB_ATTR_RO(cpu_slabs);
4716 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4718 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4720 SLAB_ATTR_RO(objects);
4722 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4724 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4726 SLAB_ATTR_RO(objects_partial);
4728 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4735 for_each_online_cpu(cpu) {
4736 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4739 pages += page->pages;
4740 objects += page->pobjects;
4744 len = sprintf(buf, "%d(%d)", objects, pages);
4747 for_each_online_cpu(cpu) {
4748 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4750 if (page && len < PAGE_SIZE - 20)
4751 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4752 page->pobjects, page->pages);
4755 return len + sprintf(buf + len, "\n");
4757 SLAB_ATTR_RO(slabs_cpu_partial);
4759 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4761 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4764 static ssize_t reclaim_account_store(struct kmem_cache *s,
4765 const char *buf, size_t length)
4767 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4769 s->flags |= SLAB_RECLAIM_ACCOUNT;
4772 SLAB_ATTR(reclaim_account);
4774 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4776 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4778 SLAB_ATTR_RO(hwcache_align);
4780 #ifdef CONFIG_ZONE_DMA
4781 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4783 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4785 SLAB_ATTR_RO(cache_dma);
4788 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4790 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4792 SLAB_ATTR_RO(destroy_by_rcu);
4794 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4796 return sprintf(buf, "%d\n", s->reserved);
4798 SLAB_ATTR_RO(reserved);
4800 #ifdef CONFIG_SLUB_DEBUG
4801 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4803 return show_slab_objects(s, buf, SO_ALL);
4805 SLAB_ATTR_RO(slabs);
4807 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4809 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4811 SLAB_ATTR_RO(total_objects);
4813 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4818 static ssize_t sanity_checks_store(struct kmem_cache *s,
4819 const char *buf, size_t length)
4821 s->flags &= ~SLAB_DEBUG_FREE;
4822 if (buf[0] == '1') {
4823 s->flags &= ~__CMPXCHG_DOUBLE;
4824 s->flags |= SLAB_DEBUG_FREE;
4828 SLAB_ATTR(sanity_checks);
4830 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4832 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4835 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4839 * Tracing a merged cache is going to give confusing results
4840 * as well as cause other issues like converting a mergeable
4841 * cache into an umergeable one.
4843 if (s->refcount > 1)
4846 s->flags &= ~SLAB_TRACE;
4847 if (buf[0] == '1') {
4848 s->flags &= ~__CMPXCHG_DOUBLE;
4849 s->flags |= SLAB_TRACE;
4855 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4857 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4860 static ssize_t red_zone_store(struct kmem_cache *s,
4861 const char *buf, size_t length)
4863 if (any_slab_objects(s))
4866 s->flags &= ~SLAB_RED_ZONE;
4867 if (buf[0] == '1') {
4868 s->flags &= ~__CMPXCHG_DOUBLE;
4869 s->flags |= SLAB_RED_ZONE;
4871 calculate_sizes(s, -1);
4874 SLAB_ATTR(red_zone);
4876 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4878 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4881 static ssize_t poison_store(struct kmem_cache *s,
4882 const char *buf, size_t length)
4884 if (any_slab_objects(s))
4887 s->flags &= ~SLAB_POISON;
4888 if (buf[0] == '1') {
4889 s->flags &= ~__CMPXCHG_DOUBLE;
4890 s->flags |= SLAB_POISON;
4892 calculate_sizes(s, -1);
4897 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4899 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4902 static ssize_t store_user_store(struct kmem_cache *s,
4903 const char *buf, size_t length)
4905 if (any_slab_objects(s))
4908 s->flags &= ~SLAB_STORE_USER;
4909 if (buf[0] == '1') {
4910 s->flags &= ~__CMPXCHG_DOUBLE;
4911 s->flags |= SLAB_STORE_USER;
4913 calculate_sizes(s, -1);
4916 SLAB_ATTR(store_user);
4918 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4923 static ssize_t validate_store(struct kmem_cache *s,
4924 const char *buf, size_t length)
4928 if (buf[0] == '1') {
4929 ret = validate_slab_cache(s);
4935 SLAB_ATTR(validate);
4937 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4939 if (!(s->flags & SLAB_STORE_USER))
4941 return list_locations(s, buf, TRACK_ALLOC);
4943 SLAB_ATTR_RO(alloc_calls);
4945 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4947 if (!(s->flags & SLAB_STORE_USER))
4949 return list_locations(s, buf, TRACK_FREE);
4951 SLAB_ATTR_RO(free_calls);
4952 #endif /* CONFIG_SLUB_DEBUG */
4954 #ifdef CONFIG_FAILSLAB
4955 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4957 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4960 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4963 if (s->refcount > 1)
4966 s->flags &= ~SLAB_FAILSLAB;
4968 s->flags |= SLAB_FAILSLAB;
4971 SLAB_ATTR(failslab);
4974 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4979 static ssize_t shrink_store(struct kmem_cache *s,
4980 const char *buf, size_t length)
4983 kmem_cache_shrink(s);
4991 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4993 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4996 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4997 const char *buf, size_t length)
4999 unsigned long ratio;
5002 err = kstrtoul(buf, 10, &ratio);
5007 s->remote_node_defrag_ratio = ratio * 10;
5011 SLAB_ATTR(remote_node_defrag_ratio);
5014 #ifdef CONFIG_SLUB_STATS
5015 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5017 unsigned long sum = 0;
5020 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5025 for_each_online_cpu(cpu) {
5026 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5032 len = sprintf(buf, "%lu", sum);
5035 for_each_online_cpu(cpu) {
5036 if (data[cpu] && len < PAGE_SIZE - 20)
5037 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5041 return len + sprintf(buf + len, "\n");
5044 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5048 for_each_online_cpu(cpu)
5049 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5052 #define STAT_ATTR(si, text) \
5053 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5055 return show_stat(s, buf, si); \
5057 static ssize_t text##_store(struct kmem_cache *s, \
5058 const char *buf, size_t length) \
5060 if (buf[0] != '0') \
5062 clear_stat(s, si); \
5067 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5068 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5069 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5070 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5071 STAT_ATTR(FREE_FROZEN, free_frozen);
5072 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5073 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5074 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5075 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5076 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5077 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5078 STAT_ATTR(FREE_SLAB, free_slab);
5079 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5080 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5081 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5082 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5083 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5084 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5085 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5086 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5087 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5088 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5089 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5090 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5091 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5092 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5095 static struct attribute *slab_attrs[] = {
5096 &slab_size_attr.attr,
5097 &object_size_attr.attr,
5098 &objs_per_slab_attr.attr,
5100 &min_partial_attr.attr,
5101 &cpu_partial_attr.attr,
5103 &objects_partial_attr.attr,
5105 &cpu_slabs_attr.attr,
5109 &hwcache_align_attr.attr,
5110 &reclaim_account_attr.attr,
5111 &destroy_by_rcu_attr.attr,
5113 &reserved_attr.attr,
5114 &slabs_cpu_partial_attr.attr,
5115 #ifdef CONFIG_SLUB_DEBUG
5116 &total_objects_attr.attr,
5118 &sanity_checks_attr.attr,
5120 &red_zone_attr.attr,
5122 &store_user_attr.attr,
5123 &validate_attr.attr,
5124 &alloc_calls_attr.attr,
5125 &free_calls_attr.attr,
5127 #ifdef CONFIG_ZONE_DMA
5128 &cache_dma_attr.attr,
5131 &remote_node_defrag_ratio_attr.attr,
5133 #ifdef CONFIG_SLUB_STATS
5134 &alloc_fastpath_attr.attr,
5135 &alloc_slowpath_attr.attr,
5136 &free_fastpath_attr.attr,
5137 &free_slowpath_attr.attr,
5138 &free_frozen_attr.attr,
5139 &free_add_partial_attr.attr,
5140 &free_remove_partial_attr.attr,
5141 &alloc_from_partial_attr.attr,
5142 &alloc_slab_attr.attr,
5143 &alloc_refill_attr.attr,
5144 &alloc_node_mismatch_attr.attr,
5145 &free_slab_attr.attr,
5146 &cpuslab_flush_attr.attr,
5147 &deactivate_full_attr.attr,
5148 &deactivate_empty_attr.attr,
5149 &deactivate_to_head_attr.attr,
5150 &deactivate_to_tail_attr.attr,
5151 &deactivate_remote_frees_attr.attr,
5152 &deactivate_bypass_attr.attr,
5153 &order_fallback_attr.attr,
5154 &cmpxchg_double_fail_attr.attr,
5155 &cmpxchg_double_cpu_fail_attr.attr,
5156 &cpu_partial_alloc_attr.attr,
5157 &cpu_partial_free_attr.attr,
5158 &cpu_partial_node_attr.attr,
5159 &cpu_partial_drain_attr.attr,
5161 #ifdef CONFIG_FAILSLAB
5162 &failslab_attr.attr,
5168 static struct attribute_group slab_attr_group = {
5169 .attrs = slab_attrs,
5172 static ssize_t slab_attr_show(struct kobject *kobj,
5173 struct attribute *attr,
5176 struct slab_attribute *attribute;
5177 struct kmem_cache *s;
5180 attribute = to_slab_attr(attr);
5183 if (!attribute->show)
5186 err = attribute->show(s, buf);
5191 static ssize_t slab_attr_store(struct kobject *kobj,
5192 struct attribute *attr,
5193 const char *buf, size_t len)
5195 struct slab_attribute *attribute;
5196 struct kmem_cache *s;
5199 attribute = to_slab_attr(attr);
5202 if (!attribute->store)
5205 err = attribute->store(s, buf, len);
5207 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5208 struct kmem_cache *c;
5210 mutex_lock(&slab_mutex);
5211 if (s->max_attr_size < len)
5212 s->max_attr_size = len;
5215 * This is a best effort propagation, so this function's return
5216 * value will be determined by the parent cache only. This is
5217 * basically because not all attributes will have a well
5218 * defined semantics for rollbacks - most of the actions will
5219 * have permanent effects.
5221 * Returning the error value of any of the children that fail
5222 * is not 100 % defined, in the sense that users seeing the
5223 * error code won't be able to know anything about the state of
5226 * Only returning the error code for the parent cache at least
5227 * has well defined semantics. The cache being written to
5228 * directly either failed or succeeded, in which case we loop
5229 * through the descendants with best-effort propagation.
5231 for_each_memcg_cache(c, s)
5232 attribute->store(c, buf, len);
5233 mutex_unlock(&slab_mutex);
5239 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5243 char *buffer = NULL;
5244 struct kmem_cache *root_cache;
5246 if (is_root_cache(s))
5249 root_cache = s->memcg_params.root_cache;
5252 * This mean this cache had no attribute written. Therefore, no point
5253 * in copying default values around
5255 if (!root_cache->max_attr_size)
5258 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5261 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5263 if (!attr || !attr->store || !attr->show)
5267 * It is really bad that we have to allocate here, so we will
5268 * do it only as a fallback. If we actually allocate, though,
5269 * we can just use the allocated buffer until the end.
5271 * Most of the slub attributes will tend to be very small in
5272 * size, but sysfs allows buffers up to a page, so they can
5273 * theoretically happen.
5277 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5280 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5281 if (WARN_ON(!buffer))
5286 attr->show(root_cache, buf);
5287 attr->store(s, buf, strlen(buf));
5291 free_page((unsigned long)buffer);
5295 static void kmem_cache_release(struct kobject *k)
5297 slab_kmem_cache_release(to_slab(k));
5300 static const struct sysfs_ops slab_sysfs_ops = {
5301 .show = slab_attr_show,
5302 .store = slab_attr_store,
5305 static struct kobj_type slab_ktype = {
5306 .sysfs_ops = &slab_sysfs_ops,
5307 .release = kmem_cache_release,
5310 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5312 struct kobj_type *ktype = get_ktype(kobj);
5314 if (ktype == &slab_ktype)
5319 static const struct kset_uevent_ops slab_uevent_ops = {
5320 .filter = uevent_filter,
5323 static struct kset *slab_kset;
5325 static inline struct kset *cache_kset(struct kmem_cache *s)
5328 if (!is_root_cache(s))
5329 return s->memcg_params.root_cache->memcg_kset;
5334 #define ID_STR_LENGTH 64
5336 /* Create a unique string id for a slab cache:
5338 * Format :[flags-]size
5340 static char *create_unique_id(struct kmem_cache *s)
5342 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5349 * First flags affecting slabcache operations. We will only
5350 * get here for aliasable slabs so we do not need to support
5351 * too many flags. The flags here must cover all flags that
5352 * are matched during merging to guarantee that the id is
5355 if (s->flags & SLAB_CACHE_DMA)
5357 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5359 if (s->flags & SLAB_DEBUG_FREE)
5361 if (!(s->flags & SLAB_NOTRACK))
5363 if (s->flags & SLAB_ACCOUNT)
5367 p += sprintf(p, "%07d", s->size);
5369 BUG_ON(p > name + ID_STR_LENGTH - 1);
5373 static int sysfs_slab_add(struct kmem_cache *s)
5377 int unmergeable = slab_unmergeable(s);
5381 * Slabcache can never be merged so we can use the name proper.
5382 * This is typically the case for debug situations. In that
5383 * case we can catch duplicate names easily.
5385 sysfs_remove_link(&slab_kset->kobj, s->name);
5389 * Create a unique name for the slab as a target
5392 name = create_unique_id(s);
5395 s->kobj.kset = cache_kset(s);
5396 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5400 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5405 if (is_root_cache(s)) {
5406 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5407 if (!s->memcg_kset) {
5414 kobject_uevent(&s->kobj, KOBJ_ADD);
5416 /* Setup first alias */
5417 sysfs_slab_alias(s, s->name);
5424 kobject_del(&s->kobj);
5428 void sysfs_slab_remove(struct kmem_cache *s)
5430 if (slab_state < FULL)
5432 * Sysfs has not been setup yet so no need to remove the
5438 kset_unregister(s->memcg_kset);
5440 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5441 kobject_del(&s->kobj);
5442 kobject_put(&s->kobj);
5446 * Need to buffer aliases during bootup until sysfs becomes
5447 * available lest we lose that information.
5449 struct saved_alias {
5450 struct kmem_cache *s;
5452 struct saved_alias *next;
5455 static struct saved_alias *alias_list;
5457 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5459 struct saved_alias *al;
5461 if (slab_state == FULL) {
5463 * If we have a leftover link then remove it.
5465 sysfs_remove_link(&slab_kset->kobj, name);
5466 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5469 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5475 al->next = alias_list;
5480 static int __init slab_sysfs_init(void)
5482 struct kmem_cache *s;
5485 mutex_lock(&slab_mutex);
5487 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5489 mutex_unlock(&slab_mutex);
5490 pr_err("Cannot register slab subsystem.\n");
5496 list_for_each_entry(s, &slab_caches, list) {
5497 err = sysfs_slab_add(s);
5499 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5503 while (alias_list) {
5504 struct saved_alias *al = alias_list;
5506 alias_list = alias_list->next;
5507 err = sysfs_slab_alias(al->s, al->name);
5509 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5514 mutex_unlock(&slab_mutex);
5519 __initcall(slab_sysfs_init);
5520 #endif /* CONFIG_SYSFS */
5523 * The /proc/slabinfo ABI
5525 #ifdef CONFIG_SLABINFO
5526 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5528 unsigned long nr_slabs = 0;
5529 unsigned long nr_objs = 0;
5530 unsigned long nr_free = 0;
5532 struct kmem_cache_node *n;
5534 for_each_kmem_cache_node(s, node, n) {
5535 nr_slabs += node_nr_slabs(n);
5536 nr_objs += node_nr_objs(n);
5537 nr_free += count_partial(n, count_free);
5540 sinfo->active_objs = nr_objs - nr_free;
5541 sinfo->num_objs = nr_objs;
5542 sinfo->active_slabs = nr_slabs;
5543 sinfo->num_slabs = nr_slabs;
5544 sinfo->objects_per_slab = oo_objects(s->oo);
5545 sinfo->cache_order = oo_order(s->oo);
5548 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5552 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5553 size_t count, loff_t *ppos)
5557 #endif /* CONFIG_SLABINFO */