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 void *fixup_red_left(struct kmem_cache *s, void *p)
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
137 #ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
145 * Issues still to be resolved:
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 * - Variable sizing of the per node arrays
152 /* Enable to test recovery from slab corruption on boot */
153 #undef SLUB_RESILIENCY_TEST
155 /* Enable to log cmpxchg failures */
156 #undef SLUB_DEBUG_CMPXCHG
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
162 #define MIN_PARTIAL 5
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
169 #define MAX_PARTIAL 10
171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
190 #define OO_MASK ((1 << OO_SHIFT) - 1)
191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
193 /* Internal SLUB flags */
194 #define __OBJECT_POISON 0x80000000UL /* Poison object */
195 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
198 static struct notifier_block slab_notifier;
202 * Tracking user of a slab.
204 #define TRACK_ADDRS_COUNT 16
206 unsigned long addr; /* Called from address */
207 #ifdef CONFIG_STACKTRACE
208 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
210 int cpu; /* Was running on cpu */
211 int pid; /* Pid context */
212 unsigned long when; /* When did the operation occur */
215 enum track_item { TRACK_ALLOC, TRACK_FREE };
218 static int sysfs_slab_add(struct kmem_cache *);
219 static int sysfs_slab_alias(struct kmem_cache *, const char *);
220 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
222 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s->cpu_slab->stat[si]);
239 /********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
243 static inline void *get_freepointer(struct kmem_cache *s, void *object)
245 return *(void **)(object + s->offset);
248 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
250 prefetch(object + s->offset);
253 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
257 #ifdef CONFIG_DEBUG_PAGEALLOC
258 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
260 p = get_freepointer(s, object);
265 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
267 *(void **)(object + s->offset) = fp;
270 /* Loop over all objects in a slab */
271 #define for_each_object(__p, __s, __addr, __objects) \
272 for (__p = fixup_red_left(__s, __addr); \
273 __p < (__addr) + (__objects) * (__s)->size; \
276 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
277 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
278 __idx <= __objects; \
279 __p += (__s)->size, __idx++)
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline int order_objects(int order, unsigned long size, int reserved)
289 return ((PAGE_SIZE << order) - reserved) / size;
292 static inline struct kmem_cache_order_objects oo_make(int order,
293 unsigned long size, int reserved)
295 struct kmem_cache_order_objects x = {
296 (order << OO_SHIFT) + order_objects(order, size, reserved)
302 static inline int oo_order(struct kmem_cache_order_objects x)
304 return x.x >> OO_SHIFT;
307 static inline int oo_objects(struct kmem_cache_order_objects x)
309 return x.x & OO_MASK;
313 * Per slab locking using the pagelock
315 static __always_inline void slab_lock(struct page *page)
317 VM_BUG_ON_PAGE(PageTail(page), page);
318 bit_spin_lock(PG_locked, &page->flags);
321 static __always_inline void slab_unlock(struct page *page)
323 VM_BUG_ON_PAGE(PageTail(page), page);
324 __bit_spin_unlock(PG_locked, &page->flags);
327 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
330 tmp.counters = counters_new;
332 * page->counters can cover frozen/inuse/objects as well
333 * as page->_count. If we assign to ->counters directly
334 * we run the risk of losing updates to page->_count, so
335 * be careful and only assign to the fields we need.
337 page->frozen = tmp.frozen;
338 page->inuse = tmp.inuse;
339 page->objects = tmp.objects;
342 /* Interrupts must be disabled (for the fallback code to work right) */
343 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
344 void *freelist_old, unsigned long counters_old,
345 void *freelist_new, unsigned long counters_new,
348 VM_BUG_ON(!irqs_disabled());
349 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
350 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
351 if (s->flags & __CMPXCHG_DOUBLE) {
352 if (cmpxchg_double(&page->freelist, &page->counters,
353 freelist_old, counters_old,
354 freelist_new, counters_new))
360 if (page->freelist == freelist_old &&
361 page->counters == counters_old) {
362 page->freelist = freelist_new;
363 set_page_slub_counters(page, counters_new);
371 stat(s, CMPXCHG_DOUBLE_FAIL);
373 #ifdef SLUB_DEBUG_CMPXCHG
374 pr_info("%s %s: cmpxchg double redo ", n, s->name);
380 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
381 void *freelist_old, unsigned long counters_old,
382 void *freelist_new, unsigned long counters_new,
385 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
386 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
387 if (s->flags & __CMPXCHG_DOUBLE) {
388 if (cmpxchg_double(&page->freelist, &page->counters,
389 freelist_old, counters_old,
390 freelist_new, counters_new))
397 local_irq_save(flags);
399 if (page->freelist == freelist_old &&
400 page->counters == counters_old) {
401 page->freelist = freelist_new;
402 set_page_slub_counters(page, counters_new);
404 local_irq_restore(flags);
408 local_irq_restore(flags);
412 stat(s, CMPXCHG_DOUBLE_FAIL);
414 #ifdef SLUB_DEBUG_CMPXCHG
415 pr_info("%s %s: cmpxchg double redo ", n, s->name);
421 #ifdef CONFIG_SLUB_DEBUG
423 * Determine a map of object in use on a page.
425 * Node listlock must be held to guarantee that the page does
426 * not vanish from under us.
428 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
431 void *addr = page_address(page);
433 for (p = page->freelist; p; p = get_freepointer(s, p))
434 set_bit(slab_index(p, s, addr), map);
437 static inline int size_from_object(struct kmem_cache *s)
439 if (s->flags & SLAB_RED_ZONE)
440 return s->size - s->red_left_pad;
445 static inline void *restore_red_left(struct kmem_cache *s, void *p)
447 if (s->flags & SLAB_RED_ZONE)
448 p -= s->red_left_pad;
456 #if defined(CONFIG_SLUB_DEBUG_ON)
457 static int slub_debug = DEBUG_DEFAULT_FLAGS;
458 #elif defined(CONFIG_KASAN)
459 static int slub_debug = SLAB_STORE_USER;
461 static int slub_debug;
464 static char *slub_debug_slabs;
465 static int disable_higher_order_debug;
468 * slub is about to manipulate internal object metadata. This memory lies
469 * outside the range of the allocated object, so accessing it would normally
470 * be reported by kasan as a bounds error. metadata_access_enable() is used
471 * to tell kasan that these accesses are OK.
473 static inline void metadata_access_enable(void)
475 kasan_disable_current();
478 static inline void metadata_access_disable(void)
480 kasan_enable_current();
487 /* Verify that a pointer has an address that is valid within a slab page */
488 static inline int check_valid_pointer(struct kmem_cache *s,
489 struct page *page, void *object)
496 base = page_address(page);
497 object = restore_red_left(s, object);
498 if (object < base || object >= base + page->objects * s->size ||
499 (object - base) % s->size) {
506 static void print_section(char *text, u8 *addr, unsigned int length)
508 metadata_access_enable();
509 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
511 metadata_access_disable();
514 static struct track *get_track(struct kmem_cache *s, void *object,
515 enum track_item alloc)
520 p = object + s->offset + sizeof(void *);
522 p = object + s->inuse;
527 static void set_track(struct kmem_cache *s, void *object,
528 enum track_item alloc, unsigned long addr)
530 struct track *p = get_track(s, object, alloc);
533 #ifdef CONFIG_STACKTRACE
534 struct stack_trace trace;
537 trace.nr_entries = 0;
538 trace.max_entries = TRACK_ADDRS_COUNT;
539 trace.entries = p->addrs;
541 metadata_access_enable();
542 save_stack_trace(&trace);
543 metadata_access_disable();
545 /* See rant in lockdep.c */
546 if (trace.nr_entries != 0 &&
547 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
550 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
554 p->cpu = smp_processor_id();
555 p->pid = current->pid;
558 memset(p, 0, sizeof(struct track));
561 static void init_tracking(struct kmem_cache *s, void *object)
563 if (!(s->flags & SLAB_STORE_USER))
566 set_track(s, object, TRACK_FREE, 0UL);
567 set_track(s, object, TRACK_ALLOC, 0UL);
570 static void print_track(const char *s, struct track *t)
575 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
576 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
577 #ifdef CONFIG_STACKTRACE
580 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
582 pr_err("\t%pS\n", (void *)t->addrs[i]);
589 static void print_tracking(struct kmem_cache *s, void *object)
591 if (!(s->flags & SLAB_STORE_USER))
594 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
595 print_track("Freed", get_track(s, object, TRACK_FREE));
598 static void print_page_info(struct page *page)
600 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
601 page, page->objects, page->inuse, page->freelist, page->flags);
605 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
607 struct va_format vaf;
613 pr_err("=============================================================================\n");
614 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
615 pr_err("-----------------------------------------------------------------------------\n\n");
617 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
621 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
623 struct va_format vaf;
629 pr_err("FIX %s: %pV\n", s->name, &vaf);
633 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
635 unsigned int off; /* Offset of last byte */
636 u8 *addr = page_address(page);
638 print_tracking(s, p);
640 print_page_info(page);
642 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
643 p, p - addr, get_freepointer(s, p));
645 if (s->flags & SLAB_RED_ZONE)
646 print_section("Redzone ", p - s->red_left_pad, s->red_left_pad);
647 else if (p > addr + 16)
648 print_section("Bytes b4 ", p - 16, 16);
650 print_section("Object ", p, min_t(unsigned long, s->object_size,
652 if (s->flags & SLAB_RED_ZONE)
653 print_section("Redzone ", p + s->object_size,
654 s->inuse - s->object_size);
657 off = s->offset + sizeof(void *);
661 if (s->flags & SLAB_STORE_USER)
662 off += 2 * sizeof(struct track);
664 if (off != size_from_object(s))
665 /* Beginning of the filler is the free pointer */
666 print_section("Padding ", p + off, size_from_object(s) - off);
671 void object_err(struct kmem_cache *s, struct page *page,
672 u8 *object, char *reason)
674 slab_bug(s, "%s", reason);
675 print_trailer(s, page, object);
678 static void slab_err(struct kmem_cache *s, struct page *page,
679 const char *fmt, ...)
685 vsnprintf(buf, sizeof(buf), fmt, args);
687 slab_bug(s, "%s", buf);
688 print_page_info(page);
692 static void init_object(struct kmem_cache *s, void *object, u8 val)
696 if (s->flags & SLAB_RED_ZONE)
697 memset(p - s->red_left_pad, val, s->red_left_pad);
699 if (s->flags & __OBJECT_POISON) {
700 memset(p, POISON_FREE, s->object_size - 1);
701 p[s->object_size - 1] = POISON_END;
704 if (s->flags & SLAB_RED_ZONE)
705 memset(p + s->object_size, val, s->inuse - s->object_size);
708 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
709 void *from, void *to)
711 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
712 memset(from, data, to - from);
715 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
716 u8 *object, char *what,
717 u8 *start, unsigned int value, unsigned int bytes)
722 metadata_access_enable();
723 fault = memchr_inv(start, value, bytes);
724 metadata_access_disable();
729 while (end > fault && end[-1] == value)
732 slab_bug(s, "%s overwritten", what);
733 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
734 fault, end - 1, fault[0], value);
735 print_trailer(s, page, object);
737 restore_bytes(s, what, value, fault, end);
745 * Bytes of the object to be managed.
746 * If the freepointer may overlay the object then the free
747 * pointer is the first word of the object.
749 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
752 * object + s->object_size
753 * Padding to reach word boundary. This is also used for Redzoning.
754 * Padding is extended by another word if Redzoning is enabled and
755 * object_size == inuse.
757 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
758 * 0xcc (RED_ACTIVE) for objects in use.
761 * Meta data starts here.
763 * A. Free pointer (if we cannot overwrite object on free)
764 * B. Tracking data for SLAB_STORE_USER
765 * C. Padding to reach required alignment boundary or at mininum
766 * one word if debugging is on to be able to detect writes
767 * before the word boundary.
769 * Padding is done using 0x5a (POISON_INUSE)
772 * Nothing is used beyond s->size.
774 * If slabcaches are merged then the object_size and inuse boundaries are mostly
775 * ignored. And therefore no slab options that rely on these boundaries
776 * may be used with merged slabcaches.
779 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
781 unsigned long off = s->inuse; /* The end of info */
784 /* Freepointer is placed after the object. */
785 off += sizeof(void *);
787 if (s->flags & SLAB_STORE_USER)
788 /* We also have user information there */
789 off += 2 * sizeof(struct track);
791 if (size_from_object(s) == off)
794 return check_bytes_and_report(s, page, p, "Object padding",
795 p + off, POISON_INUSE, size_from_object(s) - off);
798 /* Check the pad bytes at the end of a slab page */
799 static int slab_pad_check(struct kmem_cache *s, struct page *page)
807 if (!(s->flags & SLAB_POISON))
810 start = page_address(page);
811 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
812 end = start + length;
813 remainder = length % s->size;
817 metadata_access_enable();
818 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
819 metadata_access_disable();
822 while (end > fault && end[-1] == POISON_INUSE)
825 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
826 print_section("Padding ", end - remainder, remainder);
828 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
832 static int check_object(struct kmem_cache *s, struct page *page,
833 void *object, u8 val)
836 u8 *endobject = object + s->object_size;
838 if (s->flags & SLAB_RED_ZONE) {
839 if (!check_bytes_and_report(s, page, object, "Redzone",
840 object - s->red_left_pad, val, s->red_left_pad))
843 if (!check_bytes_and_report(s, page, object, "Redzone",
844 endobject, val, s->inuse - s->object_size))
847 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
848 check_bytes_and_report(s, page, p, "Alignment padding",
849 endobject, POISON_INUSE,
850 s->inuse - s->object_size);
854 if (s->flags & SLAB_POISON) {
855 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
856 (!check_bytes_and_report(s, page, p, "Poison", p,
857 POISON_FREE, s->object_size - 1) ||
858 !check_bytes_and_report(s, page, p, "Poison",
859 p + s->object_size - 1, POISON_END, 1)))
862 * check_pad_bytes cleans up on its own.
864 check_pad_bytes(s, page, p);
867 if (!s->offset && val == SLUB_RED_ACTIVE)
869 * Object and freepointer overlap. Cannot check
870 * freepointer while object is allocated.
874 /* Check free pointer validity */
875 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
876 object_err(s, page, p, "Freepointer corrupt");
878 * No choice but to zap it and thus lose the remainder
879 * of the free objects in this slab. May cause
880 * another error because the object count is now wrong.
882 set_freepointer(s, p, NULL);
888 static int check_slab(struct kmem_cache *s, struct page *page)
892 VM_BUG_ON(!irqs_disabled());
894 if (!PageSlab(page)) {
895 slab_err(s, page, "Not a valid slab page");
899 maxobj = order_objects(compound_order(page), s->size, s->reserved);
900 if (page->objects > maxobj) {
901 slab_err(s, page, "objects %u > max %u",
902 page->objects, maxobj);
905 if (page->inuse > page->objects) {
906 slab_err(s, page, "inuse %u > max %u",
907 page->inuse, page->objects);
910 /* Slab_pad_check fixes things up after itself */
911 slab_pad_check(s, page);
916 * Determine if a certain object on a page is on the freelist. Must hold the
917 * slab lock to guarantee that the chains are in a consistent state.
919 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
927 while (fp && nr <= page->objects) {
930 if (!check_valid_pointer(s, page, fp)) {
932 object_err(s, page, object,
933 "Freechain corrupt");
934 set_freepointer(s, object, NULL);
936 slab_err(s, page, "Freepointer corrupt");
937 page->freelist = NULL;
938 page->inuse = page->objects;
939 slab_fix(s, "Freelist cleared");
945 fp = get_freepointer(s, object);
949 max_objects = order_objects(compound_order(page), s->size, s->reserved);
950 if (max_objects > MAX_OBJS_PER_PAGE)
951 max_objects = MAX_OBJS_PER_PAGE;
953 if (page->objects != max_objects) {
954 slab_err(s, page, "Wrong number of objects. Found %d but "
955 "should be %d", page->objects, max_objects);
956 page->objects = max_objects;
957 slab_fix(s, "Number of objects adjusted.");
959 if (page->inuse != page->objects - nr) {
960 slab_err(s, page, "Wrong object count. Counter is %d but "
961 "counted were %d", page->inuse, page->objects - nr);
962 page->inuse = page->objects - nr;
963 slab_fix(s, "Object count adjusted.");
965 return search == NULL;
968 static void trace(struct kmem_cache *s, struct page *page, void *object,
971 if (s->flags & SLAB_TRACE) {
972 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
974 alloc ? "alloc" : "free",
979 print_section("Object ", (void *)object,
987 * Tracking of fully allocated slabs for debugging purposes.
989 static void add_full(struct kmem_cache *s,
990 struct kmem_cache_node *n, struct page *page)
992 if (!(s->flags & SLAB_STORE_USER))
995 lockdep_assert_held(&n->list_lock);
996 list_add(&page->lru, &n->full);
999 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1001 if (!(s->flags & SLAB_STORE_USER))
1004 lockdep_assert_held(&n->list_lock);
1005 list_del(&page->lru);
1008 /* Tracking of the number of slabs for debugging purposes */
1009 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1011 struct kmem_cache_node *n = get_node(s, node);
1013 return atomic_long_read(&n->nr_slabs);
1016 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1018 return atomic_long_read(&n->nr_slabs);
1021 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1023 struct kmem_cache_node *n = get_node(s, node);
1026 * May be called early in order to allocate a slab for the
1027 * kmem_cache_node structure. Solve the chicken-egg
1028 * dilemma by deferring the increment of the count during
1029 * bootstrap (see early_kmem_cache_node_alloc).
1032 atomic_long_inc(&n->nr_slabs);
1033 atomic_long_add(objects, &n->total_objects);
1036 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1038 struct kmem_cache_node *n = get_node(s, node);
1040 atomic_long_dec(&n->nr_slabs);
1041 atomic_long_sub(objects, &n->total_objects);
1044 /* Object debug checks for alloc/free paths */
1045 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1048 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1051 init_object(s, object, SLUB_RED_INACTIVE);
1052 init_tracking(s, object);
1055 static inline int alloc_consistency_checks(struct kmem_cache *s,
1057 void *object, unsigned long addr)
1059 if (!check_slab(s, page))
1062 if (!check_valid_pointer(s, page, object)) {
1063 object_err(s, page, object, "Freelist Pointer check fails");
1067 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1073 static noinline int alloc_debug_processing(struct kmem_cache *s,
1075 void *object, unsigned long addr)
1077 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1078 if (!alloc_consistency_checks(s, page, object, addr))
1082 /* Success perform special debug activities for allocs */
1083 if (s->flags & SLAB_STORE_USER)
1084 set_track(s, object, TRACK_ALLOC, addr);
1085 trace(s, page, object, 1);
1086 init_object(s, object, SLUB_RED_ACTIVE);
1090 if (PageSlab(page)) {
1092 * If this is a slab page then lets do the best we can
1093 * to avoid issues in the future. Marking all objects
1094 * as used avoids touching the remaining objects.
1096 slab_fix(s, "Marking all objects used");
1097 page->inuse = page->objects;
1098 page->freelist = NULL;
1103 static inline int free_consistency_checks(struct kmem_cache *s,
1104 struct page *page, void *object, unsigned long addr)
1106 if (!check_valid_pointer(s, page, object)) {
1107 slab_err(s, page, "Invalid object pointer 0x%p", object);
1111 if (on_freelist(s, page, object)) {
1112 object_err(s, page, object, "Object already free");
1116 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1119 if (unlikely(s != page->slab_cache)) {
1120 if (!PageSlab(page)) {
1121 slab_err(s, page, "Attempt to free object(0x%p) "
1122 "outside of slab", object);
1123 } else if (!page->slab_cache) {
1124 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1128 object_err(s, page, object,
1129 "page slab pointer corrupt.");
1135 /* Supports checking bulk free of a constructed freelist */
1136 static noinline int free_debug_processing(
1137 struct kmem_cache *s, struct page *page,
1138 void *head, void *tail, int bulk_cnt,
1141 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1142 void *object = head;
1144 unsigned long uninitialized_var(flags);
1147 spin_lock_irqsave(&n->list_lock, flags);
1150 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1151 if (!check_slab(s, page))
1158 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1159 if (!free_consistency_checks(s, page, object, addr))
1163 if (s->flags & SLAB_STORE_USER)
1164 set_track(s, object, TRACK_FREE, addr);
1165 trace(s, page, object, 0);
1166 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1167 init_object(s, object, SLUB_RED_INACTIVE);
1169 /* Reached end of constructed freelist yet? */
1170 if (object != tail) {
1171 object = get_freepointer(s, object);
1177 if (cnt != bulk_cnt)
1178 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1182 spin_unlock_irqrestore(&n->list_lock, flags);
1184 slab_fix(s, "Object at 0x%p not freed", object);
1188 static int __init setup_slub_debug(char *str)
1190 slub_debug = DEBUG_DEFAULT_FLAGS;
1191 if (*str++ != '=' || !*str)
1193 * No options specified. Switch on full debugging.
1199 * No options but restriction on slabs. This means full
1200 * debugging for slabs matching a pattern.
1207 * Switch off all debugging measures.
1212 * Determine which debug features should be switched on
1214 for (; *str && *str != ','; str++) {
1215 switch (tolower(*str)) {
1217 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1220 slub_debug |= SLAB_RED_ZONE;
1223 slub_debug |= SLAB_POISON;
1226 slub_debug |= SLAB_STORE_USER;
1229 slub_debug |= SLAB_TRACE;
1232 slub_debug |= SLAB_FAILSLAB;
1236 * Avoid enabling debugging on caches if its minimum
1237 * order would increase as a result.
1239 disable_higher_order_debug = 1;
1242 pr_err("slub_debug option '%c' unknown. skipped\n",
1249 slub_debug_slabs = str + 1;
1254 __setup("slub_debug", setup_slub_debug);
1256 unsigned long kmem_cache_flags(unsigned long object_size,
1257 unsigned long flags, const char *name,
1258 void (*ctor)(void *))
1261 * Enable debugging if selected on the kernel commandline.
1263 if (slub_debug && (!slub_debug_slabs || (name &&
1264 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1265 flags |= slub_debug;
1269 #else /* !CONFIG_SLUB_DEBUG */
1270 static inline void setup_object_debug(struct kmem_cache *s,
1271 struct page *page, void *object) {}
1273 static inline int alloc_debug_processing(struct kmem_cache *s,
1274 struct page *page, void *object, unsigned long addr) { return 0; }
1276 static inline int free_debug_processing(
1277 struct kmem_cache *s, struct page *page,
1278 void *head, void *tail, int bulk_cnt,
1279 unsigned long addr) { return 0; }
1281 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1283 static inline int check_object(struct kmem_cache *s, struct page *page,
1284 void *object, u8 val) { return 1; }
1285 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1286 struct page *page) {}
1287 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1288 struct page *page) {}
1289 unsigned long kmem_cache_flags(unsigned long object_size,
1290 unsigned long flags, const char *name,
1291 void (*ctor)(void *))
1295 #define slub_debug 0
1297 #define disable_higher_order_debug 0
1299 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1301 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1303 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1305 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1308 #endif /* CONFIG_SLUB_DEBUG */
1311 * Hooks for other subsystems that check memory allocations. In a typical
1312 * production configuration these hooks all should produce no code at all.
1314 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1316 kmemleak_alloc(ptr, size, 1, flags);
1317 kasan_kmalloc_large(ptr, size);
1320 static inline void kfree_hook(const void *x)
1323 kasan_kfree_large(x);
1326 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1328 kmemleak_free_recursive(x, s->flags);
1331 * Trouble is that we may no longer disable interrupts in the fast path
1332 * So in order to make the debug calls that expect irqs to be
1333 * disabled we need to disable interrupts temporarily.
1335 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1337 unsigned long flags;
1339 local_irq_save(flags);
1340 kmemcheck_slab_free(s, x, s->object_size);
1341 debug_check_no_locks_freed(x, s->object_size);
1342 local_irq_restore(flags);
1345 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1346 debug_check_no_obj_freed(x, s->object_size);
1348 kasan_slab_free(s, x);
1351 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1352 void *head, void *tail)
1355 * Compiler cannot detect this function can be removed if slab_free_hook()
1356 * evaluates to nothing. Thus, catch all relevant config debug options here.
1358 #if defined(CONFIG_KMEMCHECK) || \
1359 defined(CONFIG_LOCKDEP) || \
1360 defined(CONFIG_DEBUG_KMEMLEAK) || \
1361 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1362 defined(CONFIG_KASAN)
1364 void *object = head;
1365 void *tail_obj = tail ? : head;
1368 slab_free_hook(s, object);
1369 } while ((object != tail_obj) &&
1370 (object = get_freepointer(s, object)));
1374 static void setup_object(struct kmem_cache *s, struct page *page,
1377 setup_object_debug(s, page, object);
1378 if (unlikely(s->ctor)) {
1379 kasan_unpoison_object_data(s, object);
1381 kasan_poison_object_data(s, object);
1386 * Slab allocation and freeing
1388 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1389 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1392 int order = oo_order(oo);
1394 flags |= __GFP_NOTRACK;
1396 if (node == NUMA_NO_NODE)
1397 page = alloc_pages(flags, order);
1399 page = __alloc_pages_node(node, flags, order);
1401 if (page && memcg_charge_slab(page, flags, order, s)) {
1402 __free_pages(page, order);
1409 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1412 struct kmem_cache_order_objects oo = s->oo;
1417 flags &= gfp_allowed_mask;
1419 if (gfpflags_allow_blocking(flags))
1422 flags |= s->allocflags;
1425 * Let the initial higher-order allocation fail under memory pressure
1426 * so we fall-back to the minimum order allocation.
1428 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1429 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1430 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1432 page = alloc_slab_page(s, alloc_gfp, node, oo);
1433 if (unlikely(!page)) {
1437 * Allocation may have failed due to fragmentation.
1438 * Try a lower order alloc if possible
1440 page = alloc_slab_page(s, alloc_gfp, node, oo);
1441 if (unlikely(!page))
1443 stat(s, ORDER_FALLBACK);
1446 if (kmemcheck_enabled &&
1447 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1448 int pages = 1 << oo_order(oo);
1450 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1453 * Objects from caches that have a constructor don't get
1454 * cleared when they're allocated, so we need to do it here.
1457 kmemcheck_mark_uninitialized_pages(page, pages);
1459 kmemcheck_mark_unallocated_pages(page, pages);
1462 page->objects = oo_objects(oo);
1464 order = compound_order(page);
1465 page->slab_cache = s;
1466 __SetPageSlab(page);
1467 if (page_is_pfmemalloc(page))
1468 SetPageSlabPfmemalloc(page);
1470 start = page_address(page);
1472 if (unlikely(s->flags & SLAB_POISON))
1473 memset(start, POISON_INUSE, PAGE_SIZE << order);
1475 kasan_poison_slab(page);
1477 for_each_object_idx(p, idx, s, start, page->objects) {
1478 setup_object(s, page, p);
1479 if (likely(idx < page->objects))
1480 set_freepointer(s, p, p + s->size);
1482 set_freepointer(s, p, NULL);
1485 page->freelist = fixup_red_left(s, start);
1486 page->inuse = page->objects;
1490 if (gfpflags_allow_blocking(flags))
1491 local_irq_disable();
1495 mod_zone_page_state(page_zone(page),
1496 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1497 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1500 inc_slabs_node(s, page_to_nid(page), page->objects);
1505 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1507 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1508 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1512 return allocate_slab(s,
1513 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1516 static void __free_slab(struct kmem_cache *s, struct page *page)
1518 int order = compound_order(page);
1519 int pages = 1 << order;
1521 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1524 slab_pad_check(s, page);
1525 for_each_object(p, s, page_address(page),
1527 check_object(s, page, p, SLUB_RED_INACTIVE);
1530 kmemcheck_free_shadow(page, compound_order(page));
1532 mod_zone_page_state(page_zone(page),
1533 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1534 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1537 __ClearPageSlabPfmemalloc(page);
1538 __ClearPageSlab(page);
1540 page_mapcount_reset(page);
1541 if (current->reclaim_state)
1542 current->reclaim_state->reclaimed_slab += pages;
1543 memcg_uncharge_slab(page, order, s);
1544 __free_pages(page, order);
1547 #define need_reserve_slab_rcu \
1548 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1550 static void rcu_free_slab(struct rcu_head *h)
1554 if (need_reserve_slab_rcu)
1555 page = virt_to_head_page(h);
1557 page = container_of((struct list_head *)h, struct page, lru);
1559 __free_slab(page->slab_cache, page);
1562 static void free_slab(struct kmem_cache *s, struct page *page)
1564 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1565 struct rcu_head *head;
1567 if (need_reserve_slab_rcu) {
1568 int order = compound_order(page);
1569 int offset = (PAGE_SIZE << order) - s->reserved;
1571 VM_BUG_ON(s->reserved != sizeof(*head));
1572 head = page_address(page) + offset;
1574 head = &page->rcu_head;
1577 call_rcu(head, rcu_free_slab);
1579 __free_slab(s, page);
1582 static void discard_slab(struct kmem_cache *s, struct page *page)
1584 dec_slabs_node(s, page_to_nid(page), page->objects);
1589 * Management of partially allocated slabs.
1592 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1595 if (tail == DEACTIVATE_TO_TAIL)
1596 list_add_tail(&page->lru, &n->partial);
1598 list_add(&page->lru, &n->partial);
1601 static inline void add_partial(struct kmem_cache_node *n,
1602 struct page *page, int tail)
1604 lockdep_assert_held(&n->list_lock);
1605 __add_partial(n, page, tail);
1608 static inline void remove_partial(struct kmem_cache_node *n,
1611 lockdep_assert_held(&n->list_lock);
1612 list_del(&page->lru);
1617 * Remove slab from the partial list, freeze it and
1618 * return the pointer to the freelist.
1620 * Returns a list of objects or NULL if it fails.
1622 static inline void *acquire_slab(struct kmem_cache *s,
1623 struct kmem_cache_node *n, struct page *page,
1624 int mode, int *objects)
1627 unsigned long counters;
1630 lockdep_assert_held(&n->list_lock);
1633 * Zap the freelist and set the frozen bit.
1634 * The old freelist is the list of objects for the
1635 * per cpu allocation list.
1637 freelist = page->freelist;
1638 counters = page->counters;
1639 new.counters = counters;
1640 *objects = new.objects - new.inuse;
1642 new.inuse = page->objects;
1643 new.freelist = NULL;
1645 new.freelist = freelist;
1648 VM_BUG_ON(new.frozen);
1651 if (!__cmpxchg_double_slab(s, page,
1653 new.freelist, new.counters,
1657 remove_partial(n, page);
1662 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1663 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1666 * Try to allocate a partial slab from a specific node.
1668 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1669 struct kmem_cache_cpu *c, gfp_t flags)
1671 struct page *page, *page2;
1672 void *object = NULL;
1677 * Racy check. If we mistakenly see no partial slabs then we
1678 * just allocate an empty slab. If we mistakenly try to get a
1679 * partial slab and there is none available then get_partials()
1682 if (!n || !n->nr_partial)
1685 spin_lock(&n->list_lock);
1686 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1689 if (!pfmemalloc_match(page, flags))
1692 t = acquire_slab(s, n, page, object == NULL, &objects);
1696 available += objects;
1699 stat(s, ALLOC_FROM_PARTIAL);
1702 put_cpu_partial(s, page, 0);
1703 stat(s, CPU_PARTIAL_NODE);
1705 if (!kmem_cache_has_cpu_partial(s)
1706 || available > s->cpu_partial / 2)
1710 spin_unlock(&n->list_lock);
1715 * Get a page from somewhere. Search in increasing NUMA distances.
1717 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1718 struct kmem_cache_cpu *c)
1721 struct zonelist *zonelist;
1724 enum zone_type high_zoneidx = gfp_zone(flags);
1726 unsigned int cpuset_mems_cookie;
1729 * The defrag ratio allows a configuration of the tradeoffs between
1730 * inter node defragmentation and node local allocations. A lower
1731 * defrag_ratio increases the tendency to do local allocations
1732 * instead of attempting to obtain partial slabs from other nodes.
1734 * If the defrag_ratio is set to 0 then kmalloc() always
1735 * returns node local objects. If the ratio is higher then kmalloc()
1736 * may return off node objects because partial slabs are obtained
1737 * from other nodes and filled up.
1739 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1740 * defrag_ratio = 1000) then every (well almost) allocation will
1741 * first attempt to defrag slab caches on other nodes. This means
1742 * scanning over all nodes to look for partial slabs which may be
1743 * expensive if we do it every time we are trying to find a slab
1744 * with available objects.
1746 if (!s->remote_node_defrag_ratio ||
1747 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1751 cpuset_mems_cookie = read_mems_allowed_begin();
1752 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1753 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1754 struct kmem_cache_node *n;
1756 n = get_node(s, zone_to_nid(zone));
1758 if (n && cpuset_zone_allowed(zone, flags) &&
1759 n->nr_partial > s->min_partial) {
1760 object = get_partial_node(s, n, c, flags);
1763 * Don't check read_mems_allowed_retry()
1764 * here - if mems_allowed was updated in
1765 * parallel, that was a harmless race
1766 * between allocation and the cpuset
1773 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1779 * Get a partial page, lock it and return it.
1781 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1782 struct kmem_cache_cpu *c)
1785 int searchnode = node;
1787 if (node == NUMA_NO_NODE)
1788 searchnode = numa_mem_id();
1789 else if (!node_present_pages(node))
1790 searchnode = node_to_mem_node(node);
1792 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1793 if (object || node != NUMA_NO_NODE)
1796 return get_any_partial(s, flags, c);
1799 #ifdef CONFIG_PREEMPT
1801 * Calculate the next globally unique transaction for disambiguiation
1802 * during cmpxchg. The transactions start with the cpu number and are then
1803 * incremented by CONFIG_NR_CPUS.
1805 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1808 * No preemption supported therefore also no need to check for
1814 static inline unsigned long next_tid(unsigned long tid)
1816 return tid + TID_STEP;
1819 static inline unsigned int tid_to_cpu(unsigned long tid)
1821 return tid % TID_STEP;
1824 static inline unsigned long tid_to_event(unsigned long tid)
1826 return tid / TID_STEP;
1829 static inline unsigned int init_tid(int cpu)
1834 static inline void note_cmpxchg_failure(const char *n,
1835 const struct kmem_cache *s, unsigned long tid)
1837 #ifdef SLUB_DEBUG_CMPXCHG
1838 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1840 pr_info("%s %s: cmpxchg redo ", n, s->name);
1842 #ifdef CONFIG_PREEMPT
1843 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1844 pr_warn("due to cpu change %d -> %d\n",
1845 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1848 if (tid_to_event(tid) != tid_to_event(actual_tid))
1849 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1850 tid_to_event(tid), tid_to_event(actual_tid));
1852 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1853 actual_tid, tid, next_tid(tid));
1855 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1858 static void init_kmem_cache_cpus(struct kmem_cache *s)
1862 for_each_possible_cpu(cpu)
1863 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1867 * Remove the cpu slab
1869 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1872 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1873 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1875 enum slab_modes l = M_NONE, m = M_NONE;
1877 int tail = DEACTIVATE_TO_HEAD;
1881 if (page->freelist) {
1882 stat(s, DEACTIVATE_REMOTE_FREES);
1883 tail = DEACTIVATE_TO_TAIL;
1887 * Stage one: Free all available per cpu objects back
1888 * to the page freelist while it is still frozen. Leave the
1891 * There is no need to take the list->lock because the page
1894 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1896 unsigned long counters;
1899 prior = page->freelist;
1900 counters = page->counters;
1901 set_freepointer(s, freelist, prior);
1902 new.counters = counters;
1904 VM_BUG_ON(!new.frozen);
1906 } while (!__cmpxchg_double_slab(s, page,
1908 freelist, new.counters,
1909 "drain percpu freelist"));
1911 freelist = nextfree;
1915 * Stage two: Ensure that the page is unfrozen while the
1916 * list presence reflects the actual number of objects
1919 * We setup the list membership and then perform a cmpxchg
1920 * with the count. If there is a mismatch then the page
1921 * is not unfrozen but the page is on the wrong list.
1923 * Then we restart the process which may have to remove
1924 * the page from the list that we just put it on again
1925 * because the number of objects in the slab may have
1930 old.freelist = page->freelist;
1931 old.counters = page->counters;
1932 VM_BUG_ON(!old.frozen);
1934 /* Determine target state of the slab */
1935 new.counters = old.counters;
1938 set_freepointer(s, freelist, old.freelist);
1939 new.freelist = freelist;
1941 new.freelist = old.freelist;
1945 if (!new.inuse && n->nr_partial >= s->min_partial)
1947 else if (new.freelist) {
1952 * Taking the spinlock removes the possiblity
1953 * that acquire_slab() will see a slab page that
1956 spin_lock(&n->list_lock);
1960 if (kmem_cache_debug(s) && !lock) {
1963 * This also ensures that the scanning of full
1964 * slabs from diagnostic functions will not see
1967 spin_lock(&n->list_lock);
1975 remove_partial(n, page);
1977 else if (l == M_FULL)
1979 remove_full(s, n, page);
1981 if (m == M_PARTIAL) {
1983 add_partial(n, page, tail);
1986 } else if (m == M_FULL) {
1988 stat(s, DEACTIVATE_FULL);
1989 add_full(s, n, page);
1995 if (!__cmpxchg_double_slab(s, page,
1996 old.freelist, old.counters,
1997 new.freelist, new.counters,
2002 spin_unlock(&n->list_lock);
2005 stat(s, DEACTIVATE_EMPTY);
2006 discard_slab(s, page);
2012 * Unfreeze all the cpu partial slabs.
2014 * This function must be called with interrupts disabled
2015 * for the cpu using c (or some other guarantee must be there
2016 * to guarantee no concurrent accesses).
2018 static void unfreeze_partials(struct kmem_cache *s,
2019 struct kmem_cache_cpu *c)
2021 #ifdef CONFIG_SLUB_CPU_PARTIAL
2022 struct kmem_cache_node *n = NULL, *n2 = NULL;
2023 struct page *page, *discard_page = NULL;
2025 while ((page = c->partial)) {
2029 c->partial = page->next;
2031 n2 = get_node(s, page_to_nid(page));
2034 spin_unlock(&n->list_lock);
2037 spin_lock(&n->list_lock);
2042 old.freelist = page->freelist;
2043 old.counters = page->counters;
2044 VM_BUG_ON(!old.frozen);
2046 new.counters = old.counters;
2047 new.freelist = old.freelist;
2051 } while (!__cmpxchg_double_slab(s, page,
2052 old.freelist, old.counters,
2053 new.freelist, new.counters,
2054 "unfreezing slab"));
2056 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2057 page->next = discard_page;
2058 discard_page = page;
2060 add_partial(n, page, DEACTIVATE_TO_TAIL);
2061 stat(s, FREE_ADD_PARTIAL);
2066 spin_unlock(&n->list_lock);
2068 while (discard_page) {
2069 page = discard_page;
2070 discard_page = discard_page->next;
2072 stat(s, DEACTIVATE_EMPTY);
2073 discard_slab(s, page);
2080 * Put a page that was just frozen (in __slab_free) into a partial page
2081 * slot if available. This is done without interrupts disabled and without
2082 * preemption disabled. The cmpxchg is racy and may put the partial page
2083 * onto a random cpus partial slot.
2085 * If we did not find a slot then simply move all the partials to the
2086 * per node partial list.
2088 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2090 #ifdef CONFIG_SLUB_CPU_PARTIAL
2091 struct page *oldpage;
2099 oldpage = this_cpu_read(s->cpu_slab->partial);
2102 pobjects = oldpage->pobjects;
2103 pages = oldpage->pages;
2104 if (drain && pobjects > s->cpu_partial) {
2105 unsigned long flags;
2107 * partial array is full. Move the existing
2108 * set to the per node partial list.
2110 local_irq_save(flags);
2111 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2112 local_irq_restore(flags);
2116 stat(s, CPU_PARTIAL_DRAIN);
2121 pobjects += page->objects - page->inuse;
2123 page->pages = pages;
2124 page->pobjects = pobjects;
2125 page->next = oldpage;
2127 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2129 if (unlikely(!s->cpu_partial)) {
2130 unsigned long flags;
2132 local_irq_save(flags);
2133 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2134 local_irq_restore(flags);
2140 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2142 stat(s, CPUSLAB_FLUSH);
2143 deactivate_slab(s, c->page, c->freelist);
2145 c->tid = next_tid(c->tid);
2153 * Called from IPI handler with interrupts disabled.
2155 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2157 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2163 unfreeze_partials(s, c);
2167 static void flush_cpu_slab(void *d)
2169 struct kmem_cache *s = d;
2171 __flush_cpu_slab(s, smp_processor_id());
2174 static bool has_cpu_slab(int cpu, void *info)
2176 struct kmem_cache *s = info;
2177 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2179 return c->page || c->partial;
2182 static void flush_all(struct kmem_cache *s)
2184 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2188 * Check if the objects in a per cpu structure fit numa
2189 * locality expectations.
2191 static inline int node_match(struct page *page, int node)
2194 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2200 #ifdef CONFIG_SLUB_DEBUG
2201 static int count_free(struct page *page)
2203 return page->objects - page->inuse;
2206 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2208 return atomic_long_read(&n->total_objects);
2210 #endif /* CONFIG_SLUB_DEBUG */
2212 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2213 static unsigned long count_partial(struct kmem_cache_node *n,
2214 int (*get_count)(struct page *))
2216 unsigned long flags;
2217 unsigned long x = 0;
2220 spin_lock_irqsave(&n->list_lock, flags);
2221 list_for_each_entry(page, &n->partial, lru)
2222 x += get_count(page);
2223 spin_unlock_irqrestore(&n->list_lock, flags);
2226 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2228 static noinline void
2229 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2231 #ifdef CONFIG_SLUB_DEBUG
2232 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2233 DEFAULT_RATELIMIT_BURST);
2235 struct kmem_cache_node *n;
2237 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2240 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2241 nid, gfpflags, &gfpflags);
2242 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2243 s->name, s->object_size, s->size, oo_order(s->oo),
2246 if (oo_order(s->min) > get_order(s->object_size))
2247 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2250 for_each_kmem_cache_node(s, node, n) {
2251 unsigned long nr_slabs;
2252 unsigned long nr_objs;
2253 unsigned long nr_free;
2255 nr_free = count_partial(n, count_free);
2256 nr_slabs = node_nr_slabs(n);
2257 nr_objs = node_nr_objs(n);
2259 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2260 node, nr_slabs, nr_objs, nr_free);
2265 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2266 int node, struct kmem_cache_cpu **pc)
2269 struct kmem_cache_cpu *c = *pc;
2272 freelist = get_partial(s, flags, node, c);
2277 page = new_slab(s, flags, node);
2279 c = raw_cpu_ptr(s->cpu_slab);
2284 * No other reference to the page yet so we can
2285 * muck around with it freely without cmpxchg
2287 freelist = page->freelist;
2288 page->freelist = NULL;
2290 stat(s, ALLOC_SLAB);
2299 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2301 if (unlikely(PageSlabPfmemalloc(page)))
2302 return gfp_pfmemalloc_allowed(gfpflags);
2308 * Check the page->freelist of a page and either transfer the freelist to the
2309 * per cpu freelist or deactivate the page.
2311 * The page is still frozen if the return value is not NULL.
2313 * If this function returns NULL then the page has been unfrozen.
2315 * This function must be called with interrupt disabled.
2317 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2320 unsigned long counters;
2324 freelist = page->freelist;
2325 counters = page->counters;
2327 new.counters = counters;
2328 VM_BUG_ON(!new.frozen);
2330 new.inuse = page->objects;
2331 new.frozen = freelist != NULL;
2333 } while (!__cmpxchg_double_slab(s, page,
2342 * Slow path. The lockless freelist is empty or we need to perform
2345 * Processing is still very fast if new objects have been freed to the
2346 * regular freelist. In that case we simply take over the regular freelist
2347 * as the lockless freelist and zap the regular freelist.
2349 * If that is not working then we fall back to the partial lists. We take the
2350 * first element of the freelist as the object to allocate now and move the
2351 * rest of the freelist to the lockless freelist.
2353 * And if we were unable to get a new slab from the partial slab lists then
2354 * we need to allocate a new slab. This is the slowest path since it involves
2355 * a call to the page allocator and the setup of a new slab.
2357 * Version of __slab_alloc to use when we know that interrupts are
2358 * already disabled (which is the case for bulk allocation).
2360 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2361 unsigned long addr, struct kmem_cache_cpu *c)
2371 if (unlikely(!node_match(page, node))) {
2372 int searchnode = node;
2374 if (node != NUMA_NO_NODE && !node_present_pages(node))
2375 searchnode = node_to_mem_node(node);
2377 if (unlikely(!node_match(page, searchnode))) {
2378 stat(s, ALLOC_NODE_MISMATCH);
2379 deactivate_slab(s, page, c->freelist);
2387 * By rights, we should be searching for a slab page that was
2388 * PFMEMALLOC but right now, we are losing the pfmemalloc
2389 * information when the page leaves the per-cpu allocator
2391 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2392 deactivate_slab(s, page, c->freelist);
2398 /* must check again c->freelist in case of cpu migration or IRQ */
2399 freelist = c->freelist;
2403 freelist = get_freelist(s, page);
2407 stat(s, DEACTIVATE_BYPASS);
2411 stat(s, ALLOC_REFILL);
2415 * freelist is pointing to the list of objects to be used.
2416 * page is pointing to the page from which the objects are obtained.
2417 * That page must be frozen for per cpu allocations to work.
2419 VM_BUG_ON(!c->page->frozen);
2420 c->freelist = get_freepointer(s, freelist);
2421 c->tid = next_tid(c->tid);
2427 page = c->page = c->partial;
2428 c->partial = page->next;
2429 stat(s, CPU_PARTIAL_ALLOC);
2434 freelist = new_slab_objects(s, gfpflags, node, &c);
2436 if (unlikely(!freelist)) {
2437 slab_out_of_memory(s, gfpflags, node);
2442 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2445 /* Only entered in the debug case */
2446 if (kmem_cache_debug(s) &&
2447 !alloc_debug_processing(s, page, freelist, addr))
2448 goto new_slab; /* Slab failed checks. Next slab needed */
2450 deactivate_slab(s, page, get_freepointer(s, freelist));
2457 * Another one that disabled interrupt and compensates for possible
2458 * cpu changes by refetching the per cpu area pointer.
2460 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2461 unsigned long addr, struct kmem_cache_cpu *c)
2464 unsigned long flags;
2466 local_irq_save(flags);
2467 #ifdef CONFIG_PREEMPT
2469 * We may have been preempted and rescheduled on a different
2470 * cpu before disabling interrupts. Need to reload cpu area
2473 c = this_cpu_ptr(s->cpu_slab);
2476 p = ___slab_alloc(s, gfpflags, node, addr, c);
2477 local_irq_restore(flags);
2482 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2483 * have the fastpath folded into their functions. So no function call
2484 * overhead for requests that can be satisfied on the fastpath.
2486 * The fastpath works by first checking if the lockless freelist can be used.
2487 * If not then __slab_alloc is called for slow processing.
2489 * Otherwise we can simply pick the next object from the lockless free list.
2491 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2492 gfp_t gfpflags, int node, unsigned long addr)
2495 struct kmem_cache_cpu *c;
2499 s = slab_pre_alloc_hook(s, gfpflags);
2504 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2505 * enabled. We may switch back and forth between cpus while
2506 * reading from one cpu area. That does not matter as long
2507 * as we end up on the original cpu again when doing the cmpxchg.
2509 * We should guarantee that tid and kmem_cache are retrieved on
2510 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2511 * to check if it is matched or not.
2514 tid = this_cpu_read(s->cpu_slab->tid);
2515 c = raw_cpu_ptr(s->cpu_slab);
2516 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2517 unlikely(tid != READ_ONCE(c->tid)));
2520 * Irqless object alloc/free algorithm used here depends on sequence
2521 * of fetching cpu_slab's data. tid should be fetched before anything
2522 * on c to guarantee that object and page associated with previous tid
2523 * won't be used with current tid. If we fetch tid first, object and
2524 * page could be one associated with next tid and our alloc/free
2525 * request will be failed. In this case, we will retry. So, no problem.
2530 * The transaction ids are globally unique per cpu and per operation on
2531 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2532 * occurs on the right processor and that there was no operation on the
2533 * linked list in between.
2536 object = c->freelist;
2538 if (unlikely(!object || !node_match(page, node))) {
2539 object = __slab_alloc(s, gfpflags, node, addr, c);
2540 stat(s, ALLOC_SLOWPATH);
2542 void *next_object = get_freepointer_safe(s, object);
2545 * The cmpxchg will only match if there was no additional
2546 * operation and if we are on the right processor.
2548 * The cmpxchg does the following atomically (without lock
2550 * 1. Relocate first pointer to the current per cpu area.
2551 * 2. Verify that tid and freelist have not been changed
2552 * 3. If they were not changed replace tid and freelist
2554 * Since this is without lock semantics the protection is only
2555 * against code executing on this cpu *not* from access by
2558 if (unlikely(!this_cpu_cmpxchg_double(
2559 s->cpu_slab->freelist, s->cpu_slab->tid,
2561 next_object, next_tid(tid)))) {
2563 note_cmpxchg_failure("slab_alloc", s, tid);
2566 prefetch_freepointer(s, next_object);
2567 stat(s, ALLOC_FASTPATH);
2570 if (unlikely(gfpflags & __GFP_ZERO) && object)
2571 memset(object, 0, s->object_size);
2573 slab_post_alloc_hook(s, gfpflags, 1, &object);
2578 static __always_inline void *slab_alloc(struct kmem_cache *s,
2579 gfp_t gfpflags, unsigned long addr)
2581 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2584 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2586 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2588 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2593 EXPORT_SYMBOL(kmem_cache_alloc);
2595 #ifdef CONFIG_TRACING
2596 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2598 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2599 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2600 kasan_kmalloc(s, ret, size);
2603 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2607 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2609 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2611 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2612 s->object_size, s->size, gfpflags, node);
2616 EXPORT_SYMBOL(kmem_cache_alloc_node);
2618 #ifdef CONFIG_TRACING
2619 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2621 int node, size_t size)
2623 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2625 trace_kmalloc_node(_RET_IP_, ret,
2626 size, s->size, gfpflags, node);
2628 kasan_kmalloc(s, ret, size);
2631 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2636 * Slow path handling. This may still be called frequently since objects
2637 * have a longer lifetime than the cpu slabs in most processing loads.
2639 * So we still attempt to reduce cache line usage. Just take the slab
2640 * lock and free the item. If there is no additional partial page
2641 * handling required then we can return immediately.
2643 static void __slab_free(struct kmem_cache *s, struct page *page,
2644 void *head, void *tail, int cnt,
2651 unsigned long counters;
2652 struct kmem_cache_node *n = NULL;
2653 unsigned long uninitialized_var(flags);
2655 stat(s, FREE_SLOWPATH);
2657 if (kmem_cache_debug(s) &&
2658 !free_debug_processing(s, page, head, tail, cnt, addr))
2663 spin_unlock_irqrestore(&n->list_lock, flags);
2666 prior = page->freelist;
2667 counters = page->counters;
2668 set_freepointer(s, tail, prior);
2669 new.counters = counters;
2670 was_frozen = new.frozen;
2672 if ((!new.inuse || !prior) && !was_frozen) {
2674 if (kmem_cache_has_cpu_partial(s) && !prior) {
2677 * Slab was on no list before and will be
2679 * We can defer the list move and instead
2684 } else { /* Needs to be taken off a list */
2686 n = get_node(s, page_to_nid(page));
2688 * Speculatively acquire the list_lock.
2689 * If the cmpxchg does not succeed then we may
2690 * drop the list_lock without any processing.
2692 * Otherwise the list_lock will synchronize with
2693 * other processors updating the list of slabs.
2695 spin_lock_irqsave(&n->list_lock, flags);
2700 } while (!cmpxchg_double_slab(s, page,
2708 * If we just froze the page then put it onto the
2709 * per cpu partial list.
2711 if (new.frozen && !was_frozen) {
2712 put_cpu_partial(s, page, 1);
2713 stat(s, CPU_PARTIAL_FREE);
2716 * The list lock was not taken therefore no list
2717 * activity can be necessary.
2720 stat(s, FREE_FROZEN);
2724 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2728 * Objects left in the slab. If it was not on the partial list before
2731 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2732 if (kmem_cache_debug(s))
2733 remove_full(s, n, page);
2734 add_partial(n, page, DEACTIVATE_TO_TAIL);
2735 stat(s, FREE_ADD_PARTIAL);
2737 spin_unlock_irqrestore(&n->list_lock, flags);
2743 * Slab on the partial list.
2745 remove_partial(n, page);
2746 stat(s, FREE_REMOVE_PARTIAL);
2748 /* Slab must be on the full list */
2749 remove_full(s, n, page);
2752 spin_unlock_irqrestore(&n->list_lock, flags);
2754 discard_slab(s, page);
2758 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2759 * can perform fastpath freeing without additional function calls.
2761 * The fastpath is only possible if we are freeing to the current cpu slab
2762 * of this processor. This typically the case if we have just allocated
2765 * If fastpath is not possible then fall back to __slab_free where we deal
2766 * with all sorts of special processing.
2768 * Bulk free of a freelist with several objects (all pointing to the
2769 * same page) possible by specifying head and tail ptr, plus objects
2770 * count (cnt). Bulk free indicated by tail pointer being set.
2772 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2773 void *head, void *tail, int cnt,
2776 void *tail_obj = tail ? : head;
2777 struct kmem_cache_cpu *c;
2780 slab_free_freelist_hook(s, head, tail);
2784 * Determine the currently cpus per cpu slab.
2785 * The cpu may change afterward. However that does not matter since
2786 * data is retrieved via this pointer. If we are on the same cpu
2787 * during the cmpxchg then the free will succeed.
2790 tid = this_cpu_read(s->cpu_slab->tid);
2791 c = raw_cpu_ptr(s->cpu_slab);
2792 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2793 unlikely(tid != READ_ONCE(c->tid)));
2795 /* Same with comment on barrier() in slab_alloc_node() */
2798 if (likely(page == c->page)) {
2799 set_freepointer(s, tail_obj, c->freelist);
2801 if (unlikely(!this_cpu_cmpxchg_double(
2802 s->cpu_slab->freelist, s->cpu_slab->tid,
2804 head, next_tid(tid)))) {
2806 note_cmpxchg_failure("slab_free", s, tid);
2809 stat(s, FREE_FASTPATH);
2811 __slab_free(s, page, head, tail_obj, cnt, addr);
2815 void kmem_cache_free(struct kmem_cache *s, void *x)
2817 s = cache_from_obj(s, x);
2820 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2821 trace_kmem_cache_free(_RET_IP_, x);
2823 EXPORT_SYMBOL(kmem_cache_free);
2825 struct detached_freelist {
2830 struct kmem_cache *s;
2834 * This function progressively scans the array with free objects (with
2835 * a limited look ahead) and extract objects belonging to the same
2836 * page. It builds a detached freelist directly within the given
2837 * page/objects. This can happen without any need for
2838 * synchronization, because the objects are owned by running process.
2839 * The freelist is build up as a single linked list in the objects.
2840 * The idea is, that this detached freelist can then be bulk
2841 * transferred to the real freelist(s), but only requiring a single
2842 * synchronization primitive. Look ahead in the array is limited due
2843 * to performance reasons.
2846 int build_detached_freelist(struct kmem_cache *s, size_t size,
2847 void **p, struct detached_freelist *df)
2849 size_t first_skipped_index = 0;
2854 /* Always re-init detached_freelist */
2859 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2860 } while (!object && size);
2865 page = virt_to_head_page(object);
2867 /* Handle kalloc'ed objects */
2868 if (unlikely(!PageSlab(page))) {
2869 BUG_ON(!PageCompound(page));
2871 __free_kmem_pages(page, compound_order(page));
2872 p[size] = NULL; /* mark object processed */
2875 /* Derive kmem_cache from object */
2876 df->s = page->slab_cache;
2878 df->s = cache_from_obj(s, object); /* Support for memcg */
2881 /* Start new detached freelist */
2883 set_freepointer(df->s, object, NULL);
2885 df->freelist = object;
2886 p[size] = NULL; /* mark object processed */
2892 continue; /* Skip processed objects */
2894 /* df->page is always set at this point */
2895 if (df->page == virt_to_head_page(object)) {
2896 /* Opportunity build freelist */
2897 set_freepointer(df->s, object, df->freelist);
2898 df->freelist = object;
2900 p[size] = NULL; /* mark object processed */
2905 /* Limit look ahead search */
2909 if (!first_skipped_index)
2910 first_skipped_index = size + 1;
2913 return first_skipped_index;
2916 /* Note that interrupts must be enabled when calling this function. */
2917 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2923 struct detached_freelist df;
2925 size = build_detached_freelist(s, size, p, &df);
2926 if (unlikely(!df.page))
2929 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2930 } while (likely(size));
2932 EXPORT_SYMBOL(kmem_cache_free_bulk);
2934 /* Note that interrupts must be enabled when calling this function. */
2935 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2938 struct kmem_cache_cpu *c;
2941 /* memcg and kmem_cache debug support */
2942 s = slab_pre_alloc_hook(s, flags);
2946 * Drain objects in the per cpu slab, while disabling local
2947 * IRQs, which protects against PREEMPT and interrupts
2948 * handlers invoking normal fastpath.
2950 local_irq_disable();
2951 c = this_cpu_ptr(s->cpu_slab);
2953 for (i = 0; i < size; i++) {
2954 void *object = c->freelist;
2956 if (unlikely(!object)) {
2958 * Invoking slow path likely have side-effect
2959 * of re-populating per CPU c->freelist
2961 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2963 if (unlikely(!p[i]))
2966 c = this_cpu_ptr(s->cpu_slab);
2967 continue; /* goto for-loop */
2969 c->freelist = get_freepointer(s, object);
2972 c->tid = next_tid(c->tid);
2975 /* Clear memory outside IRQ disabled fastpath loop */
2976 if (unlikely(flags & __GFP_ZERO)) {
2979 for (j = 0; j < i; j++)
2980 memset(p[j], 0, s->object_size);
2983 /* memcg and kmem_cache debug support */
2984 slab_post_alloc_hook(s, flags, size, p);
2988 slab_post_alloc_hook(s, flags, i, p);
2989 __kmem_cache_free_bulk(s, i, p);
2992 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2996 * Object placement in a slab is made very easy because we always start at
2997 * offset 0. If we tune the size of the object to the alignment then we can
2998 * get the required alignment by putting one properly sized object after
3001 * Notice that the allocation order determines the sizes of the per cpu
3002 * caches. Each processor has always one slab available for allocations.
3003 * Increasing the allocation order reduces the number of times that slabs
3004 * must be moved on and off the partial lists and is therefore a factor in
3009 * Mininum / Maximum order of slab pages. This influences locking overhead
3010 * and slab fragmentation. A higher order reduces the number of partial slabs
3011 * and increases the number of allocations possible without having to
3012 * take the list_lock.
3014 static int slub_min_order;
3015 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3016 static int slub_min_objects;
3019 * Calculate the order of allocation given an slab object size.
3021 * The order of allocation has significant impact on performance and other
3022 * system components. Generally order 0 allocations should be preferred since
3023 * order 0 does not cause fragmentation in the page allocator. Larger objects
3024 * be problematic to put into order 0 slabs because there may be too much
3025 * unused space left. We go to a higher order if more than 1/16th of the slab
3028 * In order to reach satisfactory performance we must ensure that a minimum
3029 * number of objects is in one slab. Otherwise we may generate too much
3030 * activity on the partial lists which requires taking the list_lock. This is
3031 * less a concern for large slabs though which are rarely used.
3033 * slub_max_order specifies the order where we begin to stop considering the
3034 * number of objects in a slab as critical. If we reach slub_max_order then
3035 * we try to keep the page order as low as possible. So we accept more waste
3036 * of space in favor of a small page order.
3038 * Higher order allocations also allow the placement of more objects in a
3039 * slab and thereby reduce object handling overhead. If the user has
3040 * requested a higher mininum order then we start with that one instead of
3041 * the smallest order which will fit the object.
3043 static inline int slab_order(int size, int min_objects,
3044 int max_order, int fract_leftover, int reserved)
3048 int min_order = slub_min_order;
3050 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3051 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3053 for (order = max(min_order, get_order(min_objects * size + reserved));
3054 order <= max_order; order++) {
3056 unsigned long slab_size = PAGE_SIZE << order;
3058 rem = (slab_size - reserved) % size;
3060 if (rem <= slab_size / fract_leftover)
3067 static inline int calculate_order(int size, int reserved)
3075 * Attempt to find best configuration for a slab. This
3076 * works by first attempting to generate a layout with
3077 * the best configuration and backing off gradually.
3079 * First we increase the acceptable waste in a slab. Then
3080 * we reduce the minimum objects required in a slab.
3082 min_objects = slub_min_objects;
3084 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3085 max_objects = order_objects(slub_max_order, size, reserved);
3086 min_objects = min(min_objects, max_objects);
3088 while (min_objects > 1) {
3090 while (fraction >= 4) {
3091 order = slab_order(size, min_objects,
3092 slub_max_order, fraction, reserved);
3093 if (order <= slub_max_order)
3101 * We were unable to place multiple objects in a slab. Now
3102 * lets see if we can place a single object there.
3104 order = slab_order(size, 1, slub_max_order, 1, reserved);
3105 if (order <= slub_max_order)
3109 * Doh this slab cannot be placed using slub_max_order.
3111 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3112 if (order < MAX_ORDER)
3118 init_kmem_cache_node(struct kmem_cache_node *n)
3121 spin_lock_init(&n->list_lock);
3122 INIT_LIST_HEAD(&n->partial);
3123 #ifdef CONFIG_SLUB_DEBUG
3124 atomic_long_set(&n->nr_slabs, 0);
3125 atomic_long_set(&n->total_objects, 0);
3126 INIT_LIST_HEAD(&n->full);
3130 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3132 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3133 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3136 * Must align to double word boundary for the double cmpxchg
3137 * instructions to work; see __pcpu_double_call_return_bool().
3139 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3140 2 * sizeof(void *));
3145 init_kmem_cache_cpus(s);
3150 static struct kmem_cache *kmem_cache_node;
3153 * No kmalloc_node yet so do it by hand. We know that this is the first
3154 * slab on the node for this slabcache. There are no concurrent accesses
3157 * Note that this function only works on the kmem_cache_node
3158 * when allocating for the kmem_cache_node. This is used for bootstrapping
3159 * memory on a fresh node that has no slab structures yet.
3161 static void early_kmem_cache_node_alloc(int node)
3164 struct kmem_cache_node *n;
3166 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3168 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3171 if (page_to_nid(page) != node) {
3172 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3173 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3178 page->freelist = get_freepointer(kmem_cache_node, n);
3181 kmem_cache_node->node[node] = n;
3182 #ifdef CONFIG_SLUB_DEBUG
3183 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3184 init_tracking(kmem_cache_node, n);
3186 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3187 init_kmem_cache_node(n);
3188 inc_slabs_node(kmem_cache_node, node, page->objects);
3191 * No locks need to be taken here as it has just been
3192 * initialized and there is no concurrent access.
3194 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3197 static void free_kmem_cache_nodes(struct kmem_cache *s)
3200 struct kmem_cache_node *n;
3202 for_each_kmem_cache_node(s, node, n) {
3203 kmem_cache_free(kmem_cache_node, n);
3204 s->node[node] = NULL;
3208 void __kmem_cache_release(struct kmem_cache *s)
3210 free_percpu(s->cpu_slab);
3211 free_kmem_cache_nodes(s);
3214 static int init_kmem_cache_nodes(struct kmem_cache *s)
3218 for_each_node_state(node, N_NORMAL_MEMORY) {
3219 struct kmem_cache_node *n;
3221 if (slab_state == DOWN) {
3222 early_kmem_cache_node_alloc(node);
3225 n = kmem_cache_alloc_node(kmem_cache_node,
3229 free_kmem_cache_nodes(s);
3234 init_kmem_cache_node(n);
3239 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3241 if (min < MIN_PARTIAL)
3243 else if (min > MAX_PARTIAL)
3245 s->min_partial = min;
3249 * calculate_sizes() determines the order and the distribution of data within
3252 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3254 unsigned long flags = s->flags;
3255 unsigned long size = s->object_size;
3259 * Round up object size to the next word boundary. We can only
3260 * place the free pointer at word boundaries and this determines
3261 * the possible location of the free pointer.
3263 size = ALIGN(size, sizeof(void *));
3265 #ifdef CONFIG_SLUB_DEBUG
3267 * Determine if we can poison the object itself. If the user of
3268 * the slab may touch the object after free or before allocation
3269 * then we should never poison the object itself.
3271 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3273 s->flags |= __OBJECT_POISON;
3275 s->flags &= ~__OBJECT_POISON;
3279 * If we are Redzoning then check if there is some space between the
3280 * end of the object and the free pointer. If not then add an
3281 * additional word to have some bytes to store Redzone information.
3283 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3284 size += sizeof(void *);
3288 * With that we have determined the number of bytes in actual use
3289 * by the object. This is the potential offset to the free pointer.
3293 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3296 * Relocate free pointer after the object if it is not
3297 * permitted to overwrite the first word of the object on
3300 * This is the case if we do RCU, have a constructor or
3301 * destructor or are poisoning the objects.
3304 size += sizeof(void *);
3307 #ifdef CONFIG_SLUB_DEBUG
3308 if (flags & SLAB_STORE_USER)
3310 * Need to store information about allocs and frees after
3313 size += 2 * sizeof(struct track);
3315 if (flags & SLAB_RED_ZONE) {
3317 * Add some empty padding so that we can catch
3318 * overwrites from earlier objects rather than let
3319 * tracking information or the free pointer be
3320 * corrupted if a user writes before the start
3323 size += sizeof(void *);
3325 s->red_left_pad = sizeof(void *);
3326 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3327 size += s->red_left_pad;
3332 * SLUB stores one object immediately after another beginning from
3333 * offset 0. In order to align the objects we have to simply size
3334 * each object to conform to the alignment.
3336 size = ALIGN(size, s->align);
3338 if (forced_order >= 0)
3339 order = forced_order;
3341 order = calculate_order(size, s->reserved);
3348 s->allocflags |= __GFP_COMP;
3350 if (s->flags & SLAB_CACHE_DMA)
3351 s->allocflags |= GFP_DMA;
3353 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3354 s->allocflags |= __GFP_RECLAIMABLE;
3357 * Determine the number of objects per slab
3359 s->oo = oo_make(order, size, s->reserved);
3360 s->min = oo_make(get_order(size), size, s->reserved);
3361 if (oo_objects(s->oo) > oo_objects(s->max))
3364 return !!oo_objects(s->oo);
3367 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3369 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3372 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3373 s->reserved = sizeof(struct rcu_head);
3375 if (!calculate_sizes(s, -1))
3377 if (disable_higher_order_debug) {
3379 * Disable debugging flags that store metadata if the min slab
3382 if (get_order(s->size) > get_order(s->object_size)) {
3383 s->flags &= ~DEBUG_METADATA_FLAGS;
3385 if (!calculate_sizes(s, -1))
3390 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3391 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3392 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3393 /* Enable fast mode */
3394 s->flags |= __CMPXCHG_DOUBLE;
3398 * The larger the object size is, the more pages we want on the partial
3399 * list to avoid pounding the page allocator excessively.
3401 set_min_partial(s, ilog2(s->size) / 2);
3404 * cpu_partial determined the maximum number of objects kept in the
3405 * per cpu partial lists of a processor.
3407 * Per cpu partial lists mainly contain slabs that just have one
3408 * object freed. If they are used for allocation then they can be
3409 * filled up again with minimal effort. The slab will never hit the
3410 * per node partial lists and therefore no locking will be required.
3412 * This setting also determines
3414 * A) The number of objects from per cpu partial slabs dumped to the
3415 * per node list when we reach the limit.
3416 * B) The number of objects in cpu partial slabs to extract from the
3417 * per node list when we run out of per cpu objects. We only fetch
3418 * 50% to keep some capacity around for frees.
3420 if (!kmem_cache_has_cpu_partial(s))
3422 else if (s->size >= PAGE_SIZE)
3424 else if (s->size >= 1024)
3426 else if (s->size >= 256)
3427 s->cpu_partial = 13;
3429 s->cpu_partial = 30;
3432 s->remote_node_defrag_ratio = 1000;
3434 if (!init_kmem_cache_nodes(s))
3437 if (alloc_kmem_cache_cpus(s))
3440 free_kmem_cache_nodes(s);
3442 if (flags & SLAB_PANIC)
3443 panic("Cannot create slab %s size=%lu realsize=%u "
3444 "order=%u offset=%u flags=%lx\n",
3445 s->name, (unsigned long)s->size, s->size,
3446 oo_order(s->oo), s->offset, flags);
3450 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3453 #ifdef CONFIG_SLUB_DEBUG
3454 void *addr = page_address(page);
3456 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3457 sizeof(long), GFP_ATOMIC);
3460 slab_err(s, page, text, s->name);
3463 get_map(s, page, map);
3464 for_each_object(p, s, addr, page->objects) {
3466 if (!test_bit(slab_index(p, s, addr), map)) {
3467 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3468 print_tracking(s, p);
3477 * Attempt to free all partial slabs on a node.
3478 * This is called from __kmem_cache_shutdown(). We must take list_lock
3479 * because sysfs file might still access partial list after the shutdowning.
3481 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3483 struct page *page, *h;
3485 BUG_ON(irqs_disabled());
3486 spin_lock_irq(&n->list_lock);
3487 list_for_each_entry_safe(page, h, &n->partial, lru) {
3489 remove_partial(n, page);
3490 discard_slab(s, page);
3492 list_slab_objects(s, page,
3493 "Objects remaining in %s on __kmem_cache_shutdown()");
3496 spin_unlock_irq(&n->list_lock);
3500 * Release all resources used by a slab cache.
3502 int __kmem_cache_shutdown(struct kmem_cache *s)
3505 struct kmem_cache_node *n;
3508 /* Attempt to free all objects */
3509 for_each_kmem_cache_node(s, node, n) {
3511 if (n->nr_partial || slabs_node(s, node))
3517 /********************************************************************
3519 *******************************************************************/
3521 static int __init setup_slub_min_order(char *str)
3523 get_option(&str, &slub_min_order);
3528 __setup("slub_min_order=", setup_slub_min_order);
3530 static int __init setup_slub_max_order(char *str)
3532 get_option(&str, &slub_max_order);
3533 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3538 __setup("slub_max_order=", setup_slub_max_order);
3540 static int __init setup_slub_min_objects(char *str)
3542 get_option(&str, &slub_min_objects);
3547 __setup("slub_min_objects=", setup_slub_min_objects);
3549 void *__kmalloc(size_t size, gfp_t flags)
3551 struct kmem_cache *s;
3554 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3555 return kmalloc_large(size, flags);
3557 s = kmalloc_slab(size, flags);
3559 if (unlikely(ZERO_OR_NULL_PTR(s)))
3562 ret = slab_alloc(s, flags, _RET_IP_);
3564 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3566 kasan_kmalloc(s, ret, size);
3570 EXPORT_SYMBOL(__kmalloc);
3573 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3578 flags |= __GFP_COMP | __GFP_NOTRACK;
3579 page = alloc_kmem_pages_node(node, flags, get_order(size));
3581 ptr = page_address(page);
3583 kmalloc_large_node_hook(ptr, size, flags);
3587 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3589 struct kmem_cache *s;
3592 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3593 ret = kmalloc_large_node(size, flags, node);
3595 trace_kmalloc_node(_RET_IP_, ret,
3596 size, PAGE_SIZE << get_order(size),
3602 s = kmalloc_slab(size, flags);
3604 if (unlikely(ZERO_OR_NULL_PTR(s)))
3607 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3609 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3611 kasan_kmalloc(s, ret, size);
3615 EXPORT_SYMBOL(__kmalloc_node);
3618 static size_t __ksize(const void *object)
3622 if (unlikely(object == ZERO_SIZE_PTR))
3625 page = virt_to_head_page(object);
3627 if (unlikely(!PageSlab(page))) {
3628 WARN_ON(!PageCompound(page));
3629 return PAGE_SIZE << compound_order(page);
3632 return slab_ksize(page->slab_cache);
3635 size_t ksize(const void *object)
3637 size_t size = __ksize(object);
3638 /* We assume that ksize callers could use whole allocated area,
3639 so we need unpoison this area. */
3640 kasan_krealloc(object, size);
3643 EXPORT_SYMBOL(ksize);
3645 void kfree(const void *x)
3648 void *object = (void *)x;
3650 trace_kfree(_RET_IP_, x);
3652 if (unlikely(ZERO_OR_NULL_PTR(x)))
3655 page = virt_to_head_page(x);
3656 if (unlikely(!PageSlab(page))) {
3657 BUG_ON(!PageCompound(page));
3659 __free_kmem_pages(page, compound_order(page));
3662 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3664 EXPORT_SYMBOL(kfree);
3666 #define SHRINK_PROMOTE_MAX 32
3669 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3670 * up most to the head of the partial lists. New allocations will then
3671 * fill those up and thus they can be removed from the partial lists.
3673 * The slabs with the least items are placed last. This results in them
3674 * being allocated from last increasing the chance that the last objects
3675 * are freed in them.
3677 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3681 struct kmem_cache_node *n;
3684 struct list_head discard;
3685 struct list_head promote[SHRINK_PROMOTE_MAX];
3686 unsigned long flags;
3691 * Disable empty slabs caching. Used to avoid pinning offline
3692 * memory cgroups by kmem pages that can be freed.
3698 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3699 * so we have to make sure the change is visible.
3701 kick_all_cpus_sync();
3705 for_each_kmem_cache_node(s, node, n) {
3706 INIT_LIST_HEAD(&discard);
3707 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3708 INIT_LIST_HEAD(promote + i);
3710 spin_lock_irqsave(&n->list_lock, flags);
3713 * Build lists of slabs to discard or promote.
3715 * Note that concurrent frees may occur while we hold the
3716 * list_lock. page->inuse here is the upper limit.
3718 list_for_each_entry_safe(page, t, &n->partial, lru) {
3719 int free = page->objects - page->inuse;
3721 /* Do not reread page->inuse */
3724 /* We do not keep full slabs on the list */
3727 if (free == page->objects) {
3728 list_move(&page->lru, &discard);
3730 } else if (free <= SHRINK_PROMOTE_MAX)
3731 list_move(&page->lru, promote + free - 1);
3735 * Promote the slabs filled up most to the head of the
3738 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3739 list_splice(promote + i, &n->partial);
3741 spin_unlock_irqrestore(&n->list_lock, flags);
3743 /* Release empty slabs */
3744 list_for_each_entry_safe(page, t, &discard, lru)
3745 discard_slab(s, page);
3747 if (slabs_node(s, node))
3754 static int slab_mem_going_offline_callback(void *arg)
3756 struct kmem_cache *s;
3758 mutex_lock(&slab_mutex);
3759 list_for_each_entry(s, &slab_caches, list)
3760 __kmem_cache_shrink(s, false);
3761 mutex_unlock(&slab_mutex);
3766 static void slab_mem_offline_callback(void *arg)
3768 struct kmem_cache_node *n;
3769 struct kmem_cache *s;
3770 struct memory_notify *marg = arg;
3773 offline_node = marg->status_change_nid_normal;
3776 * If the node still has available memory. we need kmem_cache_node
3779 if (offline_node < 0)
3782 mutex_lock(&slab_mutex);
3783 list_for_each_entry(s, &slab_caches, list) {
3784 n = get_node(s, offline_node);
3787 * if n->nr_slabs > 0, slabs still exist on the node
3788 * that is going down. We were unable to free them,
3789 * and offline_pages() function shouldn't call this
3790 * callback. So, we must fail.
3792 BUG_ON(slabs_node(s, offline_node));
3794 s->node[offline_node] = NULL;
3795 kmem_cache_free(kmem_cache_node, n);
3798 mutex_unlock(&slab_mutex);
3801 static int slab_mem_going_online_callback(void *arg)
3803 struct kmem_cache_node *n;
3804 struct kmem_cache *s;
3805 struct memory_notify *marg = arg;
3806 int nid = marg->status_change_nid_normal;
3810 * If the node's memory is already available, then kmem_cache_node is
3811 * already created. Nothing to do.
3817 * We are bringing a node online. No memory is available yet. We must
3818 * allocate a kmem_cache_node structure in order to bring the node
3821 mutex_lock(&slab_mutex);
3822 list_for_each_entry(s, &slab_caches, list) {
3824 * XXX: kmem_cache_alloc_node will fallback to other nodes
3825 * since memory is not yet available from the node that
3828 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3833 init_kmem_cache_node(n);
3837 mutex_unlock(&slab_mutex);
3841 static int slab_memory_callback(struct notifier_block *self,
3842 unsigned long action, void *arg)
3847 case MEM_GOING_ONLINE:
3848 ret = slab_mem_going_online_callback(arg);
3850 case MEM_GOING_OFFLINE:
3851 ret = slab_mem_going_offline_callback(arg);
3854 case MEM_CANCEL_ONLINE:
3855 slab_mem_offline_callback(arg);
3858 case MEM_CANCEL_OFFLINE:
3862 ret = notifier_from_errno(ret);
3868 static struct notifier_block slab_memory_callback_nb = {
3869 .notifier_call = slab_memory_callback,
3870 .priority = SLAB_CALLBACK_PRI,
3873 /********************************************************************
3874 * Basic setup of slabs
3875 *******************************************************************/
3878 * Used for early kmem_cache structures that were allocated using
3879 * the page allocator. Allocate them properly then fix up the pointers
3880 * that may be pointing to the wrong kmem_cache structure.
3883 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3886 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3887 struct kmem_cache_node *n;
3889 memcpy(s, static_cache, kmem_cache->object_size);
3892 * This runs very early, and only the boot processor is supposed to be
3893 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3896 __flush_cpu_slab(s, smp_processor_id());
3897 for_each_kmem_cache_node(s, node, n) {
3900 list_for_each_entry(p, &n->partial, lru)
3903 #ifdef CONFIG_SLUB_DEBUG
3904 list_for_each_entry(p, &n->full, lru)
3908 slab_init_memcg_params(s);
3909 list_add(&s->list, &slab_caches);
3913 void __init kmem_cache_init(void)
3915 static __initdata struct kmem_cache boot_kmem_cache,
3916 boot_kmem_cache_node;
3918 if (debug_guardpage_minorder())
3921 kmem_cache_node = &boot_kmem_cache_node;
3922 kmem_cache = &boot_kmem_cache;
3924 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3925 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3927 register_hotmemory_notifier(&slab_memory_callback_nb);
3929 /* Able to allocate the per node structures */
3930 slab_state = PARTIAL;
3932 create_boot_cache(kmem_cache, "kmem_cache",
3933 offsetof(struct kmem_cache, node) +
3934 nr_node_ids * sizeof(struct kmem_cache_node *),
3935 SLAB_HWCACHE_ALIGN);
3937 kmem_cache = bootstrap(&boot_kmem_cache);
3940 * Allocate kmem_cache_node properly from the kmem_cache slab.
3941 * kmem_cache_node is separately allocated so no need to
3942 * update any list pointers.
3944 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3946 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3947 setup_kmalloc_cache_index_table();
3948 create_kmalloc_caches(0);
3951 register_cpu_notifier(&slab_notifier);
3954 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3956 slub_min_order, slub_max_order, slub_min_objects,
3957 nr_cpu_ids, nr_node_ids);
3960 void __init kmem_cache_init_late(void)
3965 __kmem_cache_alias(const char *name, size_t size, size_t align,
3966 unsigned long flags, void (*ctor)(void *))
3968 struct kmem_cache *s, *c;
3970 s = find_mergeable(size, align, flags, name, ctor);
3975 * Adjust the object sizes so that we clear
3976 * the complete object on kzalloc.
3978 s->object_size = max(s->object_size, (int)size);
3979 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3981 for_each_memcg_cache(c, s) {
3982 c->object_size = s->object_size;
3983 c->inuse = max_t(int, c->inuse,
3984 ALIGN(size, sizeof(void *)));
3987 if (sysfs_slab_alias(s, name)) {
3996 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
4000 err = kmem_cache_open(s, flags);
4004 /* Mutex is not taken during early boot */
4005 if (slab_state <= UP)
4008 memcg_propagate_slab_attrs(s);
4009 err = sysfs_slab_add(s);
4011 __kmem_cache_release(s);
4018 * Use the cpu notifier to insure that the cpu slabs are flushed when
4021 static int slab_cpuup_callback(struct notifier_block *nfb,
4022 unsigned long action, void *hcpu)
4024 long cpu = (long)hcpu;
4025 struct kmem_cache *s;
4026 unsigned long flags;
4029 case CPU_UP_CANCELED:
4030 case CPU_UP_CANCELED_FROZEN:
4032 case CPU_DEAD_FROZEN:
4033 mutex_lock(&slab_mutex);
4034 list_for_each_entry(s, &slab_caches, list) {
4035 local_irq_save(flags);
4036 __flush_cpu_slab(s, cpu);
4037 local_irq_restore(flags);
4039 mutex_unlock(&slab_mutex);
4047 static struct notifier_block slab_notifier = {
4048 .notifier_call = slab_cpuup_callback
4053 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4055 struct kmem_cache *s;
4058 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4059 return kmalloc_large(size, gfpflags);
4061 s = kmalloc_slab(size, gfpflags);
4063 if (unlikely(ZERO_OR_NULL_PTR(s)))
4066 ret = slab_alloc(s, gfpflags, caller);
4068 /* Honor the call site pointer we received. */
4069 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4075 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4076 int node, unsigned long caller)
4078 struct kmem_cache *s;
4081 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4082 ret = kmalloc_large_node(size, gfpflags, node);
4084 trace_kmalloc_node(caller, ret,
4085 size, PAGE_SIZE << get_order(size),
4091 s = kmalloc_slab(size, gfpflags);
4093 if (unlikely(ZERO_OR_NULL_PTR(s)))
4096 ret = slab_alloc_node(s, gfpflags, node, caller);
4098 /* Honor the call site pointer we received. */
4099 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4106 static int count_inuse(struct page *page)
4111 static int count_total(struct page *page)
4113 return page->objects;
4117 #ifdef CONFIG_SLUB_DEBUG
4118 static int validate_slab(struct kmem_cache *s, struct page *page,
4122 void *addr = page_address(page);
4124 if (!check_slab(s, page) ||
4125 !on_freelist(s, page, NULL))
4128 /* Now we know that a valid freelist exists */
4129 bitmap_zero(map, page->objects);
4131 get_map(s, page, map);
4132 for_each_object(p, s, addr, page->objects) {
4133 if (test_bit(slab_index(p, s, addr), map))
4134 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4138 for_each_object(p, s, addr, page->objects)
4139 if (!test_bit(slab_index(p, s, addr), map))
4140 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4145 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4149 validate_slab(s, page, map);
4153 static int validate_slab_node(struct kmem_cache *s,
4154 struct kmem_cache_node *n, unsigned long *map)
4156 unsigned long count = 0;
4158 unsigned long flags;
4160 spin_lock_irqsave(&n->list_lock, flags);
4162 list_for_each_entry(page, &n->partial, lru) {
4163 validate_slab_slab(s, page, map);
4166 if (count != n->nr_partial)
4167 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4168 s->name, count, n->nr_partial);
4170 if (!(s->flags & SLAB_STORE_USER))
4173 list_for_each_entry(page, &n->full, lru) {
4174 validate_slab_slab(s, page, map);
4177 if (count != atomic_long_read(&n->nr_slabs))
4178 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4179 s->name, count, atomic_long_read(&n->nr_slabs));
4182 spin_unlock_irqrestore(&n->list_lock, flags);
4186 static long validate_slab_cache(struct kmem_cache *s)
4189 unsigned long count = 0;
4190 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4191 sizeof(unsigned long), GFP_KERNEL);
4192 struct kmem_cache_node *n;
4198 for_each_kmem_cache_node(s, node, n)
4199 count += validate_slab_node(s, n, map);
4204 * Generate lists of code addresses where slabcache objects are allocated
4209 unsigned long count;
4216 DECLARE_BITMAP(cpus, NR_CPUS);
4222 unsigned long count;
4223 struct location *loc;
4226 static void free_loc_track(struct loc_track *t)
4229 free_pages((unsigned long)t->loc,
4230 get_order(sizeof(struct location) * t->max));
4233 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4238 order = get_order(sizeof(struct location) * max);
4240 l = (void *)__get_free_pages(flags, order);
4245 memcpy(l, t->loc, sizeof(struct location) * t->count);
4253 static int add_location(struct loc_track *t, struct kmem_cache *s,
4254 const struct track *track)
4256 long start, end, pos;
4258 unsigned long caddr;
4259 unsigned long age = jiffies - track->when;
4265 pos = start + (end - start + 1) / 2;
4268 * There is nothing at "end". If we end up there
4269 * we need to add something to before end.
4274 caddr = t->loc[pos].addr;
4275 if (track->addr == caddr) {
4281 if (age < l->min_time)
4283 if (age > l->max_time)
4286 if (track->pid < l->min_pid)
4287 l->min_pid = track->pid;
4288 if (track->pid > l->max_pid)
4289 l->max_pid = track->pid;
4291 cpumask_set_cpu(track->cpu,
4292 to_cpumask(l->cpus));
4294 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4298 if (track->addr < caddr)
4305 * Not found. Insert new tracking element.
4307 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4313 (t->count - pos) * sizeof(struct location));
4316 l->addr = track->addr;
4320 l->min_pid = track->pid;
4321 l->max_pid = track->pid;
4322 cpumask_clear(to_cpumask(l->cpus));
4323 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4324 nodes_clear(l->nodes);
4325 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4329 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4330 struct page *page, enum track_item alloc,
4333 void *addr = page_address(page);
4336 bitmap_zero(map, page->objects);
4337 get_map(s, page, map);
4339 for_each_object(p, s, addr, page->objects)
4340 if (!test_bit(slab_index(p, s, addr), map))
4341 add_location(t, s, get_track(s, p, alloc));
4344 static int list_locations(struct kmem_cache *s, char *buf,
4345 enum track_item alloc)
4349 struct loc_track t = { 0, 0, NULL };
4351 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4352 sizeof(unsigned long), GFP_KERNEL);
4353 struct kmem_cache_node *n;
4355 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4358 return sprintf(buf, "Out of memory\n");
4360 /* Push back cpu slabs */
4363 for_each_kmem_cache_node(s, node, n) {
4364 unsigned long flags;
4367 if (!atomic_long_read(&n->nr_slabs))
4370 spin_lock_irqsave(&n->list_lock, flags);
4371 list_for_each_entry(page, &n->partial, lru)
4372 process_slab(&t, s, page, alloc, map);
4373 list_for_each_entry(page, &n->full, lru)
4374 process_slab(&t, s, page, alloc, map);
4375 spin_unlock_irqrestore(&n->list_lock, flags);
4378 for (i = 0; i < t.count; i++) {
4379 struct location *l = &t.loc[i];
4381 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4383 len += sprintf(buf + len, "%7ld ", l->count);
4386 len += sprintf(buf + len, "%pS", (void *)l->addr);
4388 len += sprintf(buf + len, "<not-available>");
4390 if (l->sum_time != l->min_time) {
4391 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4393 (long)div_u64(l->sum_time, l->count),
4396 len += sprintf(buf + len, " age=%ld",
4399 if (l->min_pid != l->max_pid)
4400 len += sprintf(buf + len, " pid=%ld-%ld",
4401 l->min_pid, l->max_pid);
4403 len += sprintf(buf + len, " pid=%ld",
4406 if (num_online_cpus() > 1 &&
4407 !cpumask_empty(to_cpumask(l->cpus)) &&
4408 len < PAGE_SIZE - 60)
4409 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4411 cpumask_pr_args(to_cpumask(l->cpus)));
4413 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4414 len < PAGE_SIZE - 60)
4415 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4417 nodemask_pr_args(&l->nodes));
4419 len += sprintf(buf + len, "\n");
4425 len += sprintf(buf, "No data\n");
4430 #ifdef SLUB_RESILIENCY_TEST
4431 static void __init resiliency_test(void)
4435 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4437 pr_err("SLUB resiliency testing\n");
4438 pr_err("-----------------------\n");
4439 pr_err("A. Corruption after allocation\n");
4441 p = kzalloc(16, GFP_KERNEL);
4443 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4446 validate_slab_cache(kmalloc_caches[4]);
4448 /* Hmmm... The next two are dangerous */
4449 p = kzalloc(32, GFP_KERNEL);
4450 p[32 + sizeof(void *)] = 0x34;
4451 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4453 pr_err("If allocated object is overwritten then not detectable\n\n");
4455 validate_slab_cache(kmalloc_caches[5]);
4456 p = kzalloc(64, GFP_KERNEL);
4457 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4459 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4461 pr_err("If allocated object is overwritten then not detectable\n\n");
4462 validate_slab_cache(kmalloc_caches[6]);
4464 pr_err("\nB. Corruption after free\n");
4465 p = kzalloc(128, GFP_KERNEL);
4468 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4469 validate_slab_cache(kmalloc_caches[7]);
4471 p = kzalloc(256, GFP_KERNEL);
4474 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4475 validate_slab_cache(kmalloc_caches[8]);
4477 p = kzalloc(512, GFP_KERNEL);
4480 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4481 validate_slab_cache(kmalloc_caches[9]);
4485 static void resiliency_test(void) {};
4490 enum slab_stat_type {
4491 SL_ALL, /* All slabs */
4492 SL_PARTIAL, /* Only partially allocated slabs */
4493 SL_CPU, /* Only slabs used for cpu caches */
4494 SL_OBJECTS, /* Determine allocated objects not slabs */
4495 SL_TOTAL /* Determine object capacity not slabs */
4498 #define SO_ALL (1 << SL_ALL)
4499 #define SO_PARTIAL (1 << SL_PARTIAL)
4500 #define SO_CPU (1 << SL_CPU)
4501 #define SO_OBJECTS (1 << SL_OBJECTS)
4502 #define SO_TOTAL (1 << SL_TOTAL)
4504 static ssize_t show_slab_objects(struct kmem_cache *s,
4505 char *buf, unsigned long flags)
4507 unsigned long total = 0;
4510 unsigned long *nodes;
4512 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4516 if (flags & SO_CPU) {
4519 for_each_possible_cpu(cpu) {
4520 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4525 page = READ_ONCE(c->page);
4529 node = page_to_nid(page);
4530 if (flags & SO_TOTAL)
4532 else if (flags & SO_OBJECTS)
4540 page = READ_ONCE(c->partial);
4542 node = page_to_nid(page);
4543 if (flags & SO_TOTAL)
4545 else if (flags & SO_OBJECTS)
4556 #ifdef CONFIG_SLUB_DEBUG
4557 if (flags & SO_ALL) {
4558 struct kmem_cache_node *n;
4560 for_each_kmem_cache_node(s, node, n) {
4562 if (flags & SO_TOTAL)
4563 x = atomic_long_read(&n->total_objects);
4564 else if (flags & SO_OBJECTS)
4565 x = atomic_long_read(&n->total_objects) -
4566 count_partial(n, count_free);
4568 x = atomic_long_read(&n->nr_slabs);
4575 if (flags & SO_PARTIAL) {
4576 struct kmem_cache_node *n;
4578 for_each_kmem_cache_node(s, node, n) {
4579 if (flags & SO_TOTAL)
4580 x = count_partial(n, count_total);
4581 else if (flags & SO_OBJECTS)
4582 x = count_partial(n, count_inuse);
4589 x = sprintf(buf, "%lu", total);
4591 for (node = 0; node < nr_node_ids; node++)
4593 x += sprintf(buf + x, " N%d=%lu",
4598 return x + sprintf(buf + x, "\n");
4601 #ifdef CONFIG_SLUB_DEBUG
4602 static int any_slab_objects(struct kmem_cache *s)
4605 struct kmem_cache_node *n;
4607 for_each_kmem_cache_node(s, node, n)
4608 if (atomic_long_read(&n->total_objects))
4615 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4616 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4618 struct slab_attribute {
4619 struct attribute attr;
4620 ssize_t (*show)(struct kmem_cache *s, char *buf);
4621 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4624 #define SLAB_ATTR_RO(_name) \
4625 static struct slab_attribute _name##_attr = \
4626 __ATTR(_name, 0400, _name##_show, NULL)
4628 #define SLAB_ATTR(_name) \
4629 static struct slab_attribute _name##_attr = \
4630 __ATTR(_name, 0600, _name##_show, _name##_store)
4632 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4634 return sprintf(buf, "%d\n", s->size);
4636 SLAB_ATTR_RO(slab_size);
4638 static ssize_t align_show(struct kmem_cache *s, char *buf)
4640 return sprintf(buf, "%d\n", s->align);
4642 SLAB_ATTR_RO(align);
4644 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4646 return sprintf(buf, "%d\n", s->object_size);
4648 SLAB_ATTR_RO(object_size);
4650 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4652 return sprintf(buf, "%d\n", oo_objects(s->oo));
4654 SLAB_ATTR_RO(objs_per_slab);
4656 static ssize_t order_store(struct kmem_cache *s,
4657 const char *buf, size_t length)
4659 unsigned long order;
4662 err = kstrtoul(buf, 10, &order);
4666 if (order > slub_max_order || order < slub_min_order)
4669 calculate_sizes(s, order);
4673 static ssize_t order_show(struct kmem_cache *s, char *buf)
4675 return sprintf(buf, "%d\n", oo_order(s->oo));
4679 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4681 return sprintf(buf, "%lu\n", s->min_partial);
4684 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4690 err = kstrtoul(buf, 10, &min);
4694 set_min_partial(s, min);
4697 SLAB_ATTR(min_partial);
4699 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4701 return sprintf(buf, "%u\n", s->cpu_partial);
4704 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4707 unsigned long objects;
4710 err = kstrtoul(buf, 10, &objects);
4713 if (objects && !kmem_cache_has_cpu_partial(s))
4716 s->cpu_partial = objects;
4720 SLAB_ATTR(cpu_partial);
4722 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4726 return sprintf(buf, "%pS\n", s->ctor);
4730 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4732 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4734 SLAB_ATTR_RO(aliases);
4736 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4738 return show_slab_objects(s, buf, SO_PARTIAL);
4740 SLAB_ATTR_RO(partial);
4742 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4744 return show_slab_objects(s, buf, SO_CPU);
4746 SLAB_ATTR_RO(cpu_slabs);
4748 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4750 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4752 SLAB_ATTR_RO(objects);
4754 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4756 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4758 SLAB_ATTR_RO(objects_partial);
4760 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4767 for_each_online_cpu(cpu) {
4768 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4771 pages += page->pages;
4772 objects += page->pobjects;
4776 len = sprintf(buf, "%d(%d)", objects, pages);
4779 for_each_online_cpu(cpu) {
4780 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4782 if (page && len < PAGE_SIZE - 20)
4783 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4784 page->pobjects, page->pages);
4787 return len + sprintf(buf + len, "\n");
4789 SLAB_ATTR_RO(slabs_cpu_partial);
4791 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4793 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4796 static ssize_t reclaim_account_store(struct kmem_cache *s,
4797 const char *buf, size_t length)
4799 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4801 s->flags |= SLAB_RECLAIM_ACCOUNT;
4804 SLAB_ATTR(reclaim_account);
4806 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4808 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4810 SLAB_ATTR_RO(hwcache_align);
4812 #ifdef CONFIG_ZONE_DMA
4813 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4817 SLAB_ATTR_RO(cache_dma);
4820 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4822 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4824 SLAB_ATTR_RO(destroy_by_rcu);
4826 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4828 return sprintf(buf, "%d\n", s->reserved);
4830 SLAB_ATTR_RO(reserved);
4832 #ifdef CONFIG_SLUB_DEBUG
4833 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4835 return show_slab_objects(s, buf, SO_ALL);
4837 SLAB_ATTR_RO(slabs);
4839 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4841 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4843 SLAB_ATTR_RO(total_objects);
4845 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4847 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
4850 static ssize_t sanity_checks_store(struct kmem_cache *s,
4851 const char *buf, size_t length)
4853 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
4854 if (buf[0] == '1') {
4855 s->flags &= ~__CMPXCHG_DOUBLE;
4856 s->flags |= SLAB_CONSISTENCY_CHECKS;
4860 SLAB_ATTR(sanity_checks);
4862 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4864 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4867 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4871 * Tracing a merged cache is going to give confusing results
4872 * as well as cause other issues like converting a mergeable
4873 * cache into an umergeable one.
4875 if (s->refcount > 1)
4878 s->flags &= ~SLAB_TRACE;
4879 if (buf[0] == '1') {
4880 s->flags &= ~__CMPXCHG_DOUBLE;
4881 s->flags |= SLAB_TRACE;
4887 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4889 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4892 static ssize_t red_zone_store(struct kmem_cache *s,
4893 const char *buf, size_t length)
4895 if (any_slab_objects(s))
4898 s->flags &= ~SLAB_RED_ZONE;
4899 if (buf[0] == '1') {
4900 s->flags |= SLAB_RED_ZONE;
4902 calculate_sizes(s, -1);
4905 SLAB_ATTR(red_zone);
4907 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4909 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4912 static ssize_t poison_store(struct kmem_cache *s,
4913 const char *buf, size_t length)
4915 if (any_slab_objects(s))
4918 s->flags &= ~SLAB_POISON;
4919 if (buf[0] == '1') {
4920 s->flags |= SLAB_POISON;
4922 calculate_sizes(s, -1);
4927 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4929 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4932 static ssize_t store_user_store(struct kmem_cache *s,
4933 const char *buf, size_t length)
4935 if (any_slab_objects(s))
4938 s->flags &= ~SLAB_STORE_USER;
4939 if (buf[0] == '1') {
4940 s->flags &= ~__CMPXCHG_DOUBLE;
4941 s->flags |= SLAB_STORE_USER;
4943 calculate_sizes(s, -1);
4946 SLAB_ATTR(store_user);
4948 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4953 static ssize_t validate_store(struct kmem_cache *s,
4954 const char *buf, size_t length)
4958 if (buf[0] == '1') {
4959 ret = validate_slab_cache(s);
4965 SLAB_ATTR(validate);
4967 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4969 if (!(s->flags & SLAB_STORE_USER))
4971 return list_locations(s, buf, TRACK_ALLOC);
4973 SLAB_ATTR_RO(alloc_calls);
4975 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4977 if (!(s->flags & SLAB_STORE_USER))
4979 return list_locations(s, buf, TRACK_FREE);
4981 SLAB_ATTR_RO(free_calls);
4982 #endif /* CONFIG_SLUB_DEBUG */
4984 #ifdef CONFIG_FAILSLAB
4985 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4987 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4990 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4993 if (s->refcount > 1)
4996 s->flags &= ~SLAB_FAILSLAB;
4998 s->flags |= SLAB_FAILSLAB;
5001 SLAB_ATTR(failslab);
5004 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5009 static ssize_t shrink_store(struct kmem_cache *s,
5010 const char *buf, size_t length)
5013 kmem_cache_shrink(s);
5021 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5023 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5026 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5027 const char *buf, size_t length)
5029 unsigned long ratio;
5032 err = kstrtoul(buf, 10, &ratio);
5037 s->remote_node_defrag_ratio = ratio * 10;
5041 SLAB_ATTR(remote_node_defrag_ratio);
5044 #ifdef CONFIG_SLUB_STATS
5045 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5047 unsigned long sum = 0;
5050 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5055 for_each_online_cpu(cpu) {
5056 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5062 len = sprintf(buf, "%lu", sum);
5065 for_each_online_cpu(cpu) {
5066 if (data[cpu] && len < PAGE_SIZE - 20)
5067 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5071 return len + sprintf(buf + len, "\n");
5074 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5078 for_each_online_cpu(cpu)
5079 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5082 #define STAT_ATTR(si, text) \
5083 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5085 return show_stat(s, buf, si); \
5087 static ssize_t text##_store(struct kmem_cache *s, \
5088 const char *buf, size_t length) \
5090 if (buf[0] != '0') \
5092 clear_stat(s, si); \
5097 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5098 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5099 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5100 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5101 STAT_ATTR(FREE_FROZEN, free_frozen);
5102 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5103 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5104 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5105 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5106 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5107 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5108 STAT_ATTR(FREE_SLAB, free_slab);
5109 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5110 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5111 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5112 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5113 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5114 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5115 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5116 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5117 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5118 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5119 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5120 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5121 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5122 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5125 static struct attribute *slab_attrs[] = {
5126 &slab_size_attr.attr,
5127 &object_size_attr.attr,
5128 &objs_per_slab_attr.attr,
5130 &min_partial_attr.attr,
5131 &cpu_partial_attr.attr,
5133 &objects_partial_attr.attr,
5135 &cpu_slabs_attr.attr,
5139 &hwcache_align_attr.attr,
5140 &reclaim_account_attr.attr,
5141 &destroy_by_rcu_attr.attr,
5143 &reserved_attr.attr,
5144 &slabs_cpu_partial_attr.attr,
5145 #ifdef CONFIG_SLUB_DEBUG
5146 &total_objects_attr.attr,
5148 &sanity_checks_attr.attr,
5150 &red_zone_attr.attr,
5152 &store_user_attr.attr,
5153 &validate_attr.attr,
5154 &alloc_calls_attr.attr,
5155 &free_calls_attr.attr,
5157 #ifdef CONFIG_ZONE_DMA
5158 &cache_dma_attr.attr,
5161 &remote_node_defrag_ratio_attr.attr,
5163 #ifdef CONFIG_SLUB_STATS
5164 &alloc_fastpath_attr.attr,
5165 &alloc_slowpath_attr.attr,
5166 &free_fastpath_attr.attr,
5167 &free_slowpath_attr.attr,
5168 &free_frozen_attr.attr,
5169 &free_add_partial_attr.attr,
5170 &free_remove_partial_attr.attr,
5171 &alloc_from_partial_attr.attr,
5172 &alloc_slab_attr.attr,
5173 &alloc_refill_attr.attr,
5174 &alloc_node_mismatch_attr.attr,
5175 &free_slab_attr.attr,
5176 &cpuslab_flush_attr.attr,
5177 &deactivate_full_attr.attr,
5178 &deactivate_empty_attr.attr,
5179 &deactivate_to_head_attr.attr,
5180 &deactivate_to_tail_attr.attr,
5181 &deactivate_remote_frees_attr.attr,
5182 &deactivate_bypass_attr.attr,
5183 &order_fallback_attr.attr,
5184 &cmpxchg_double_fail_attr.attr,
5185 &cmpxchg_double_cpu_fail_attr.attr,
5186 &cpu_partial_alloc_attr.attr,
5187 &cpu_partial_free_attr.attr,
5188 &cpu_partial_node_attr.attr,
5189 &cpu_partial_drain_attr.attr,
5191 #ifdef CONFIG_FAILSLAB
5192 &failslab_attr.attr,
5198 static struct attribute_group slab_attr_group = {
5199 .attrs = slab_attrs,
5202 static ssize_t slab_attr_show(struct kobject *kobj,
5203 struct attribute *attr,
5206 struct slab_attribute *attribute;
5207 struct kmem_cache *s;
5210 attribute = to_slab_attr(attr);
5213 if (!attribute->show)
5216 err = attribute->show(s, buf);
5221 static ssize_t slab_attr_store(struct kobject *kobj,
5222 struct attribute *attr,
5223 const char *buf, size_t len)
5225 struct slab_attribute *attribute;
5226 struct kmem_cache *s;
5229 attribute = to_slab_attr(attr);
5232 if (!attribute->store)
5235 err = attribute->store(s, buf, len);
5237 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5238 struct kmem_cache *c;
5240 mutex_lock(&slab_mutex);
5241 if (s->max_attr_size < len)
5242 s->max_attr_size = len;
5245 * This is a best effort propagation, so this function's return
5246 * value will be determined by the parent cache only. This is
5247 * basically because not all attributes will have a well
5248 * defined semantics for rollbacks - most of the actions will
5249 * have permanent effects.
5251 * Returning the error value of any of the children that fail
5252 * is not 100 % defined, in the sense that users seeing the
5253 * error code won't be able to know anything about the state of
5256 * Only returning the error code for the parent cache at least
5257 * has well defined semantics. The cache being written to
5258 * directly either failed or succeeded, in which case we loop
5259 * through the descendants with best-effort propagation.
5261 for_each_memcg_cache(c, s)
5262 attribute->store(c, buf, len);
5263 mutex_unlock(&slab_mutex);
5269 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5273 char *buffer = NULL;
5274 struct kmem_cache *root_cache;
5276 if (is_root_cache(s))
5279 root_cache = s->memcg_params.root_cache;
5282 * This mean this cache had no attribute written. Therefore, no point
5283 * in copying default values around
5285 if (!root_cache->max_attr_size)
5288 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5291 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5293 if (!attr || !attr->store || !attr->show)
5297 * It is really bad that we have to allocate here, so we will
5298 * do it only as a fallback. If we actually allocate, though,
5299 * we can just use the allocated buffer until the end.
5301 * Most of the slub attributes will tend to be very small in
5302 * size, but sysfs allows buffers up to a page, so they can
5303 * theoretically happen.
5307 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5310 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5311 if (WARN_ON(!buffer))
5316 attr->show(root_cache, buf);
5317 attr->store(s, buf, strlen(buf));
5321 free_page((unsigned long)buffer);
5325 static void kmem_cache_release(struct kobject *k)
5327 slab_kmem_cache_release(to_slab(k));
5330 static const struct sysfs_ops slab_sysfs_ops = {
5331 .show = slab_attr_show,
5332 .store = slab_attr_store,
5335 static struct kobj_type slab_ktype = {
5336 .sysfs_ops = &slab_sysfs_ops,
5337 .release = kmem_cache_release,
5340 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5342 struct kobj_type *ktype = get_ktype(kobj);
5344 if (ktype == &slab_ktype)
5349 static const struct kset_uevent_ops slab_uevent_ops = {
5350 .filter = uevent_filter,
5353 static struct kset *slab_kset;
5355 static inline struct kset *cache_kset(struct kmem_cache *s)
5358 if (!is_root_cache(s))
5359 return s->memcg_params.root_cache->memcg_kset;
5364 #define ID_STR_LENGTH 64
5366 /* Create a unique string id for a slab cache:
5368 * Format :[flags-]size
5370 static char *create_unique_id(struct kmem_cache *s)
5372 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5379 * First flags affecting slabcache operations. We will only
5380 * get here for aliasable slabs so we do not need to support
5381 * too many flags. The flags here must cover all flags that
5382 * are matched during merging to guarantee that the id is
5385 if (s->flags & SLAB_CACHE_DMA)
5387 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5389 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5391 if (!(s->flags & SLAB_NOTRACK))
5393 if (s->flags & SLAB_ACCOUNT)
5397 p += sprintf(p, "%07d", s->size);
5399 BUG_ON(p > name + ID_STR_LENGTH - 1);
5403 static int sysfs_slab_add(struct kmem_cache *s)
5407 int unmergeable = slab_unmergeable(s);
5411 * Slabcache can never be merged so we can use the name proper.
5412 * This is typically the case for debug situations. In that
5413 * case we can catch duplicate names easily.
5415 sysfs_remove_link(&slab_kset->kobj, s->name);
5419 * Create a unique name for the slab as a target
5422 name = create_unique_id(s);
5425 s->kobj.kset = cache_kset(s);
5426 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5430 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5435 if (is_root_cache(s)) {
5436 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5437 if (!s->memcg_kset) {
5444 kobject_uevent(&s->kobj, KOBJ_ADD);
5446 /* Setup first alias */
5447 sysfs_slab_alias(s, s->name);
5454 kobject_del(&s->kobj);
5458 void sysfs_slab_remove(struct kmem_cache *s)
5460 if (slab_state < FULL)
5462 * Sysfs has not been setup yet so no need to remove the
5468 kset_unregister(s->memcg_kset);
5470 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5471 kobject_del(&s->kobj);
5472 kobject_put(&s->kobj);
5476 * Need to buffer aliases during bootup until sysfs becomes
5477 * available lest we lose that information.
5479 struct saved_alias {
5480 struct kmem_cache *s;
5482 struct saved_alias *next;
5485 static struct saved_alias *alias_list;
5487 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5489 struct saved_alias *al;
5491 if (slab_state == FULL) {
5493 * If we have a leftover link then remove it.
5495 sysfs_remove_link(&slab_kset->kobj, name);
5496 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5499 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5505 al->next = alias_list;
5510 static int __init slab_sysfs_init(void)
5512 struct kmem_cache *s;
5515 mutex_lock(&slab_mutex);
5517 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5519 mutex_unlock(&slab_mutex);
5520 pr_err("Cannot register slab subsystem.\n");
5526 list_for_each_entry(s, &slab_caches, list) {
5527 err = sysfs_slab_add(s);
5529 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5533 while (alias_list) {
5534 struct saved_alias *al = alias_list;
5536 alias_list = alias_list->next;
5537 err = sysfs_slab_alias(al->s, al->name);
5539 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5544 mutex_unlock(&slab_mutex);
5549 __initcall(slab_sysfs_init);
5550 #endif /* CONFIG_SYSFS */
5553 * The /proc/slabinfo ABI
5555 #ifdef CONFIG_SLABINFO
5556 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5558 unsigned long nr_slabs = 0;
5559 unsigned long nr_objs = 0;
5560 unsigned long nr_free = 0;
5562 struct kmem_cache_node *n;
5564 for_each_kmem_cache_node(s, node, n) {
5565 nr_slabs += node_nr_slabs(n);
5566 nr_objs += node_nr_objs(n);
5567 nr_free += count_partial(n, count_free);
5570 sinfo->active_objs = nr_objs - nr_free;
5571 sinfo->num_objs = nr_objs;
5572 sinfo->active_slabs = nr_slabs;
5573 sinfo->num_slabs = nr_slabs;
5574 sinfo->objects_per_slab = oo_objects(s->oo);
5575 sinfo->cache_order = oo_order(s->oo);
5578 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5582 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5583 size_t count, loff_t *ppos)
5587 #endif /* CONFIG_SLABINFO */