1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.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:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * page->frozen The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 void *fixup_red_left(struct kmem_cache *s, void *p)
128 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
129 p += s->red_left_pad;
134 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
136 #ifdef CONFIG_SLUB_CPU_PARTIAL
137 return !kmem_cache_debug(s);
144 * Issues still to be resolved:
146 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 * - Variable sizing of the per node arrays
151 /* Enable to test recovery from slab corruption on boot */
152 #undef SLUB_RESILIENCY_TEST
154 /* Enable to log cmpxchg failures */
155 #undef SLUB_DEBUG_CMPXCHG
158 * Mininum number of partial slabs. These will be left on the partial
159 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 #define MIN_PARTIAL 5
164 * Maximum number of desirable partial slabs.
165 * The existence of more partial slabs makes kmem_cache_shrink
166 * sort the partial list by the number of objects in use.
168 #define MAX_PARTIAL 10
170 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_STORE_USER)
174 * These debug flags cannot use CMPXCHG because there might be consistency
175 * issues when checking or reading debug information
177 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
182 * Debugging flags that require metadata to be stored in the slab. These get
183 * disabled when slub_debug=O is used and a cache's min order increases with
186 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
189 #define OO_MASK ((1 << OO_SHIFT) - 1)
190 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
192 /* Internal SLUB flags */
194 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
195 /* Use cmpxchg_double */
196 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
199 * Tracking user of a slab.
201 #define TRACK_ADDRS_COUNT 16
203 unsigned long addr; /* Called from address */
204 #ifdef CONFIG_STACKTRACE
205 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
207 int cpu; /* Was running on cpu */
208 int pid; /* Pid context */
209 unsigned long when; /* When did the operation occur */
212 enum track_item { TRACK_ALLOC, TRACK_FREE };
215 static int sysfs_slab_add(struct kmem_cache *);
216 static int sysfs_slab_alias(struct kmem_cache *, const char *);
217 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
218 static void sysfs_slab_remove(struct kmem_cache *s);
220 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
221 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
223 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
224 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
227 static inline void stat(const struct kmem_cache *s, enum stat_item si)
229 #ifdef CONFIG_SLUB_STATS
231 * The rmw is racy on a preemptible kernel but this is acceptable, so
232 * avoid this_cpu_add()'s irq-disable overhead.
234 raw_cpu_inc(s->cpu_slab->stat[si]);
238 /********************************************************************
239 * Core slab cache functions
240 *******************************************************************/
243 * Returns freelist pointer (ptr). With hardening, this is obfuscated
244 * with an XOR of the address where the pointer is held and a per-cache
247 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
248 unsigned long ptr_addr)
250 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
253 * Normally, this doesn't cause any issues, as both set_freepointer()
254 * and get_freepointer() are called with a pointer with the same tag.
255 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256 * example, when __free_slub() iterates over objects in a cache, it
257 * passes untagged pointers to check_object(). check_object() in turns
258 * calls get_freepointer() with an untagged pointer, which causes the
259 * freepointer to be restored incorrectly.
261 return (void *)((unsigned long)ptr ^ s->random ^
262 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
268 /* Returns the freelist pointer recorded at location ptr_addr. */
269 static inline void *freelist_dereference(const struct kmem_cache *s,
272 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
273 (unsigned long)ptr_addr);
276 static inline void *get_freepointer(struct kmem_cache *s, void *object)
278 return freelist_dereference(s, object + s->offset);
281 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
283 prefetch(object + s->offset);
286 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
288 unsigned long freepointer_addr;
291 if (!debug_pagealloc_enabled_static())
292 return get_freepointer(s, object);
294 freepointer_addr = (unsigned long)object + s->offset;
295 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
296 return freelist_ptr(s, p, freepointer_addr);
299 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
301 unsigned long freeptr_addr = (unsigned long)object + s->offset;
303 #ifdef CONFIG_SLAB_FREELIST_HARDENED
304 BUG_ON(object == fp); /* naive detection of double free or corruption */
307 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
310 /* Loop over all objects in a slab */
311 #define for_each_object(__p, __s, __addr, __objects) \
312 for (__p = fixup_red_left(__s, __addr); \
313 __p < (__addr) + (__objects) * (__s)->size; \
316 /* Determine object index from a given position */
317 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
319 return (kasan_reset_tag(p) - addr) / s->size;
322 static inline unsigned int order_objects(unsigned int order, unsigned int size)
324 return ((unsigned int)PAGE_SIZE << order) / size;
327 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
330 struct kmem_cache_order_objects x = {
331 (order << OO_SHIFT) + order_objects(order, size)
337 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
339 return x.x >> OO_SHIFT;
342 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
344 return x.x & OO_MASK;
348 * Per slab locking using the pagelock
350 static __always_inline void slab_lock(struct page *page)
352 VM_BUG_ON_PAGE(PageTail(page), page);
353 bit_spin_lock(PG_locked, &page->flags);
356 static __always_inline void slab_unlock(struct page *page)
358 VM_BUG_ON_PAGE(PageTail(page), page);
359 __bit_spin_unlock(PG_locked, &page->flags);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
368 VM_BUG_ON(!irqs_disabled());
369 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
370 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
371 if (s->flags & __CMPXCHG_DOUBLE) {
372 if (cmpxchg_double(&page->freelist, &page->counters,
373 freelist_old, counters_old,
374 freelist_new, counters_new))
380 if (page->freelist == freelist_old &&
381 page->counters == counters_old) {
382 page->freelist = freelist_new;
383 page->counters = counters_new;
391 stat(s, CMPXCHG_DOUBLE_FAIL);
393 #ifdef SLUB_DEBUG_CMPXCHG
394 pr_info("%s %s: cmpxchg double redo ", n, s->name);
400 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
401 void *freelist_old, unsigned long counters_old,
402 void *freelist_new, unsigned long counters_new,
405 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
406 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
407 if (s->flags & __CMPXCHG_DOUBLE) {
408 if (cmpxchg_double(&page->freelist, &page->counters,
409 freelist_old, counters_old,
410 freelist_new, counters_new))
417 local_irq_save(flags);
419 if (page->freelist == freelist_old &&
420 page->counters == counters_old) {
421 page->freelist = freelist_new;
422 page->counters = counters_new;
424 local_irq_restore(flags);
428 local_irq_restore(flags);
432 stat(s, CMPXCHG_DOUBLE_FAIL);
434 #ifdef SLUB_DEBUG_CMPXCHG
435 pr_info("%s %s: cmpxchg double redo ", n, s->name);
441 #ifdef CONFIG_SLUB_DEBUG
442 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
443 static DEFINE_SPINLOCK(object_map_lock);
446 * Determine a map of object in use on a page.
448 * Node listlock must be held to guarantee that the page does
449 * not vanish from under us.
451 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
452 __acquires(&object_map_lock)
455 void *addr = page_address(page);
457 VM_BUG_ON(!irqs_disabled());
459 spin_lock(&object_map_lock);
461 bitmap_zero(object_map, page->objects);
463 for (p = page->freelist; p; p = get_freepointer(s, p))
464 set_bit(slab_index(p, s, addr), object_map);
469 static void put_map(unsigned long *map) __releases(&object_map_lock)
471 VM_BUG_ON(map != object_map);
472 lockdep_assert_held(&object_map_lock);
474 spin_unlock(&object_map_lock);
477 static inline unsigned int size_from_object(struct kmem_cache *s)
479 if (s->flags & SLAB_RED_ZONE)
480 return s->size - s->red_left_pad;
485 static inline void *restore_red_left(struct kmem_cache *s, void *p)
487 if (s->flags & SLAB_RED_ZONE)
488 p -= s->red_left_pad;
496 #if defined(CONFIG_SLUB_DEBUG_ON)
497 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
499 static slab_flags_t slub_debug;
502 static char *slub_debug_slabs;
503 static int disable_higher_order_debug;
506 * slub is about to manipulate internal object metadata. This memory lies
507 * outside the range of the allocated object, so accessing it would normally
508 * be reported by kasan as a bounds error. metadata_access_enable() is used
509 * to tell kasan that these accesses are OK.
511 static inline void metadata_access_enable(void)
513 kasan_disable_current();
516 static inline void metadata_access_disable(void)
518 kasan_enable_current();
525 /* Verify that a pointer has an address that is valid within a slab page */
526 static inline int check_valid_pointer(struct kmem_cache *s,
527 struct page *page, void *object)
534 base = page_address(page);
535 object = kasan_reset_tag(object);
536 object = restore_red_left(s, object);
537 if (object < base || object >= base + page->objects * s->size ||
538 (object - base) % s->size) {
545 static void print_section(char *level, char *text, u8 *addr,
548 metadata_access_enable();
549 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
551 metadata_access_disable();
554 static struct track *get_track(struct kmem_cache *s, void *object,
555 enum track_item alloc)
560 p = object + s->offset + sizeof(void *);
562 p = object + s->inuse;
567 static void set_track(struct kmem_cache *s, void *object,
568 enum track_item alloc, unsigned long addr)
570 struct track *p = get_track(s, object, alloc);
573 #ifdef CONFIG_STACKTRACE
574 unsigned int nr_entries;
576 metadata_access_enable();
577 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
578 metadata_access_disable();
580 if (nr_entries < TRACK_ADDRS_COUNT)
581 p->addrs[nr_entries] = 0;
584 p->cpu = smp_processor_id();
585 p->pid = current->pid;
588 memset(p, 0, sizeof(struct track));
592 static void init_tracking(struct kmem_cache *s, void *object)
594 if (!(s->flags & SLAB_STORE_USER))
597 set_track(s, object, TRACK_FREE, 0UL);
598 set_track(s, object, TRACK_ALLOC, 0UL);
601 static void print_track(const char *s, struct track *t, unsigned long pr_time)
606 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
607 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
608 #ifdef CONFIG_STACKTRACE
611 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
613 pr_err("\t%pS\n", (void *)t->addrs[i]);
620 static void print_tracking(struct kmem_cache *s, void *object)
622 unsigned long pr_time = jiffies;
623 if (!(s->flags & SLAB_STORE_USER))
626 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
627 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
630 static void print_page_info(struct page *page)
632 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
633 page, page->objects, page->inuse, page->freelist, page->flags);
637 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
639 struct va_format vaf;
645 pr_err("=============================================================================\n");
646 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
647 pr_err("-----------------------------------------------------------------------------\n\n");
649 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
653 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
655 struct va_format vaf;
661 pr_err("FIX %s: %pV\n", s->name, &vaf);
665 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
667 unsigned int off; /* Offset of last byte */
668 u8 *addr = page_address(page);
670 print_tracking(s, p);
672 print_page_info(page);
674 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
675 p, p - addr, get_freepointer(s, p));
677 if (s->flags & SLAB_RED_ZONE)
678 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
680 else if (p > addr + 16)
681 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
683 print_section(KERN_ERR, "Object ", p,
684 min_t(unsigned int, s->object_size, PAGE_SIZE));
685 if (s->flags & SLAB_RED_ZONE)
686 print_section(KERN_ERR, "Redzone ", p + s->object_size,
687 s->inuse - s->object_size);
690 off = s->offset + sizeof(void *);
694 if (s->flags & SLAB_STORE_USER)
695 off += 2 * sizeof(struct track);
697 off += kasan_metadata_size(s);
699 if (off != size_from_object(s))
700 /* Beginning of the filler is the free pointer */
701 print_section(KERN_ERR, "Padding ", p + off,
702 size_from_object(s) - off);
707 void object_err(struct kmem_cache *s, struct page *page,
708 u8 *object, char *reason)
710 slab_bug(s, "%s", reason);
711 print_trailer(s, page, object);
714 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
715 const char *fmt, ...)
721 vsnprintf(buf, sizeof(buf), fmt, args);
723 slab_bug(s, "%s", buf);
724 print_page_info(page);
728 static void init_object(struct kmem_cache *s, void *object, u8 val)
732 if (s->flags & SLAB_RED_ZONE)
733 memset(p - s->red_left_pad, val, s->red_left_pad);
735 if (s->flags & __OBJECT_POISON) {
736 memset(p, POISON_FREE, s->object_size - 1);
737 p[s->object_size - 1] = POISON_END;
740 if (s->flags & SLAB_RED_ZONE)
741 memset(p + s->object_size, val, s->inuse - s->object_size);
744 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
745 void *from, void *to)
747 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
748 memset(from, data, to - from);
751 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
752 u8 *object, char *what,
753 u8 *start, unsigned int value, unsigned int bytes)
757 u8 *addr = page_address(page);
759 metadata_access_enable();
760 fault = memchr_inv(start, value, bytes);
761 metadata_access_disable();
766 while (end > fault && end[-1] == value)
769 slab_bug(s, "%s overwritten", what);
770 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
771 fault, end - 1, fault - addr,
773 print_trailer(s, page, object);
775 restore_bytes(s, what, value, fault, end);
783 * Bytes of the object to be managed.
784 * If the freepointer may overlay the object then the free
785 * pointer is the first word of the object.
787 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
790 * object + s->object_size
791 * Padding to reach word boundary. This is also used for Redzoning.
792 * Padding is extended by another word if Redzoning is enabled and
793 * object_size == inuse.
795 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
796 * 0xcc (RED_ACTIVE) for objects in use.
799 * Meta data starts here.
801 * A. Free pointer (if we cannot overwrite object on free)
802 * B. Tracking data for SLAB_STORE_USER
803 * C. Padding to reach required alignment boundary or at mininum
804 * one word if debugging is on to be able to detect writes
805 * before the word boundary.
807 * Padding is done using 0x5a (POISON_INUSE)
810 * Nothing is used beyond s->size.
812 * If slabcaches are merged then the object_size and inuse boundaries are mostly
813 * ignored. And therefore no slab options that rely on these boundaries
814 * may be used with merged slabcaches.
817 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
819 unsigned long off = s->inuse; /* The end of info */
822 /* Freepointer is placed after the object. */
823 off += sizeof(void *);
825 if (s->flags & SLAB_STORE_USER)
826 /* We also have user information there */
827 off += 2 * sizeof(struct track);
829 off += kasan_metadata_size(s);
831 if (size_from_object(s) == off)
834 return check_bytes_and_report(s, page, p, "Object padding",
835 p + off, POISON_INUSE, size_from_object(s) - off);
838 /* Check the pad bytes at the end of a slab page */
839 static int slab_pad_check(struct kmem_cache *s, struct page *page)
848 if (!(s->flags & SLAB_POISON))
851 start = page_address(page);
852 length = page_size(page);
853 end = start + length;
854 remainder = length % s->size;
858 pad = end - remainder;
859 metadata_access_enable();
860 fault = memchr_inv(pad, POISON_INUSE, remainder);
861 metadata_access_disable();
864 while (end > fault && end[-1] == POISON_INUSE)
867 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
868 fault, end - 1, fault - start);
869 print_section(KERN_ERR, "Padding ", pad, remainder);
871 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
875 static int check_object(struct kmem_cache *s, struct page *page,
876 void *object, u8 val)
879 u8 *endobject = object + s->object_size;
881 if (s->flags & SLAB_RED_ZONE) {
882 if (!check_bytes_and_report(s, page, object, "Redzone",
883 object - s->red_left_pad, val, s->red_left_pad))
886 if (!check_bytes_and_report(s, page, object, "Redzone",
887 endobject, val, s->inuse - s->object_size))
890 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
891 check_bytes_and_report(s, page, p, "Alignment padding",
892 endobject, POISON_INUSE,
893 s->inuse - s->object_size);
897 if (s->flags & SLAB_POISON) {
898 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
899 (!check_bytes_and_report(s, page, p, "Poison", p,
900 POISON_FREE, s->object_size - 1) ||
901 !check_bytes_and_report(s, page, p, "Poison",
902 p + s->object_size - 1, POISON_END, 1)))
905 * check_pad_bytes cleans up on its own.
907 check_pad_bytes(s, page, p);
910 if (!s->offset && val == SLUB_RED_ACTIVE)
912 * Object and freepointer overlap. Cannot check
913 * freepointer while object is allocated.
917 /* Check free pointer validity */
918 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
919 object_err(s, page, p, "Freepointer corrupt");
921 * No choice but to zap it and thus lose the remainder
922 * of the free objects in this slab. May cause
923 * another error because the object count is now wrong.
925 set_freepointer(s, p, NULL);
931 static int check_slab(struct kmem_cache *s, struct page *page)
935 VM_BUG_ON(!irqs_disabled());
937 if (!PageSlab(page)) {
938 slab_err(s, page, "Not a valid slab page");
942 maxobj = order_objects(compound_order(page), s->size);
943 if (page->objects > maxobj) {
944 slab_err(s, page, "objects %u > max %u",
945 page->objects, maxobj);
948 if (page->inuse > page->objects) {
949 slab_err(s, page, "inuse %u > max %u",
950 page->inuse, page->objects);
953 /* Slab_pad_check fixes things up after itself */
954 slab_pad_check(s, page);
959 * Determine if a certain object on a page is on the freelist. Must hold the
960 * slab lock to guarantee that the chains are in a consistent state.
962 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
970 while (fp && nr <= page->objects) {
973 if (!check_valid_pointer(s, page, fp)) {
975 object_err(s, page, object,
976 "Freechain corrupt");
977 set_freepointer(s, object, NULL);
979 slab_err(s, page, "Freepointer corrupt");
980 page->freelist = NULL;
981 page->inuse = page->objects;
982 slab_fix(s, "Freelist cleared");
988 fp = get_freepointer(s, object);
992 max_objects = order_objects(compound_order(page), s->size);
993 if (max_objects > MAX_OBJS_PER_PAGE)
994 max_objects = MAX_OBJS_PER_PAGE;
996 if (page->objects != max_objects) {
997 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
998 page->objects, max_objects);
999 page->objects = max_objects;
1000 slab_fix(s, "Number of objects adjusted.");
1002 if (page->inuse != page->objects - nr) {
1003 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1004 page->inuse, page->objects - nr);
1005 page->inuse = page->objects - nr;
1006 slab_fix(s, "Object count adjusted.");
1008 return search == NULL;
1011 static void trace(struct kmem_cache *s, struct page *page, void *object,
1014 if (s->flags & SLAB_TRACE) {
1015 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1017 alloc ? "alloc" : "free",
1018 object, page->inuse,
1022 print_section(KERN_INFO, "Object ", (void *)object,
1030 * Tracking of fully allocated slabs for debugging purposes.
1032 static void add_full(struct kmem_cache *s,
1033 struct kmem_cache_node *n, struct page *page)
1035 if (!(s->flags & SLAB_STORE_USER))
1038 lockdep_assert_held(&n->list_lock);
1039 list_add(&page->slab_list, &n->full);
1042 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1044 if (!(s->flags & SLAB_STORE_USER))
1047 lockdep_assert_held(&n->list_lock);
1048 list_del(&page->slab_list);
1051 /* Tracking of the number of slabs for debugging purposes */
1052 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1054 struct kmem_cache_node *n = get_node(s, node);
1056 return atomic_long_read(&n->nr_slabs);
1059 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1061 return atomic_long_read(&n->nr_slabs);
1064 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1066 struct kmem_cache_node *n = get_node(s, node);
1069 * May be called early in order to allocate a slab for the
1070 * kmem_cache_node structure. Solve the chicken-egg
1071 * dilemma by deferring the increment of the count during
1072 * bootstrap (see early_kmem_cache_node_alloc).
1075 atomic_long_inc(&n->nr_slabs);
1076 atomic_long_add(objects, &n->total_objects);
1079 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1081 struct kmem_cache_node *n = get_node(s, node);
1083 atomic_long_dec(&n->nr_slabs);
1084 atomic_long_sub(objects, &n->total_objects);
1087 /* Object debug checks for alloc/free paths */
1088 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1091 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1094 init_object(s, object, SLUB_RED_INACTIVE);
1095 init_tracking(s, object);
1099 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1101 if (!(s->flags & SLAB_POISON))
1104 metadata_access_enable();
1105 memset(addr, POISON_INUSE, page_size(page));
1106 metadata_access_disable();
1109 static inline int alloc_consistency_checks(struct kmem_cache *s,
1110 struct page *page, void *object)
1112 if (!check_slab(s, page))
1115 if (!check_valid_pointer(s, page, object)) {
1116 object_err(s, page, object, "Freelist Pointer check fails");
1120 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1126 static noinline int alloc_debug_processing(struct kmem_cache *s,
1128 void *object, unsigned long addr)
1130 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1131 if (!alloc_consistency_checks(s, page, object))
1135 /* Success perform special debug activities for allocs */
1136 if (s->flags & SLAB_STORE_USER)
1137 set_track(s, object, TRACK_ALLOC, addr);
1138 trace(s, page, object, 1);
1139 init_object(s, object, SLUB_RED_ACTIVE);
1143 if (PageSlab(page)) {
1145 * If this is a slab page then lets do the best we can
1146 * to avoid issues in the future. Marking all objects
1147 * as used avoids touching the remaining objects.
1149 slab_fix(s, "Marking all objects used");
1150 page->inuse = page->objects;
1151 page->freelist = NULL;
1156 static inline int free_consistency_checks(struct kmem_cache *s,
1157 struct page *page, void *object, unsigned long addr)
1159 if (!check_valid_pointer(s, page, object)) {
1160 slab_err(s, page, "Invalid object pointer 0x%p", object);
1164 if (on_freelist(s, page, object)) {
1165 object_err(s, page, object, "Object already free");
1169 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1172 if (unlikely(s != page->slab_cache)) {
1173 if (!PageSlab(page)) {
1174 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1176 } else if (!page->slab_cache) {
1177 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1181 object_err(s, page, object,
1182 "page slab pointer corrupt.");
1188 /* Supports checking bulk free of a constructed freelist */
1189 static noinline int free_debug_processing(
1190 struct kmem_cache *s, struct page *page,
1191 void *head, void *tail, int bulk_cnt,
1194 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1195 void *object = head;
1197 unsigned long uninitialized_var(flags);
1200 spin_lock_irqsave(&n->list_lock, flags);
1203 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1204 if (!check_slab(s, page))
1211 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1212 if (!free_consistency_checks(s, page, object, addr))
1216 if (s->flags & SLAB_STORE_USER)
1217 set_track(s, object, TRACK_FREE, addr);
1218 trace(s, page, object, 0);
1219 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1220 init_object(s, object, SLUB_RED_INACTIVE);
1222 /* Reached end of constructed freelist yet? */
1223 if (object != tail) {
1224 object = get_freepointer(s, object);
1230 if (cnt != bulk_cnt)
1231 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1235 spin_unlock_irqrestore(&n->list_lock, flags);
1237 slab_fix(s, "Object at 0x%p not freed", object);
1241 static int __init setup_slub_debug(char *str)
1243 slub_debug = DEBUG_DEFAULT_FLAGS;
1244 if (*str++ != '=' || !*str)
1246 * No options specified. Switch on full debugging.
1252 * No options but restriction on slabs. This means full
1253 * debugging for slabs matching a pattern.
1260 * Switch off all debugging measures.
1265 * Determine which debug features should be switched on
1267 for (; *str && *str != ','; str++) {
1268 switch (tolower(*str)) {
1270 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1273 slub_debug |= SLAB_RED_ZONE;
1276 slub_debug |= SLAB_POISON;
1279 slub_debug |= SLAB_STORE_USER;
1282 slub_debug |= SLAB_TRACE;
1285 slub_debug |= SLAB_FAILSLAB;
1289 * Avoid enabling debugging on caches if its minimum
1290 * order would increase as a result.
1292 disable_higher_order_debug = 1;
1295 pr_err("slub_debug option '%c' unknown. skipped\n",
1302 slub_debug_slabs = str + 1;
1304 if ((static_branch_unlikely(&init_on_alloc) ||
1305 static_branch_unlikely(&init_on_free)) &&
1306 (slub_debug & SLAB_POISON))
1307 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1311 __setup("slub_debug", setup_slub_debug);
1314 * kmem_cache_flags - apply debugging options to the cache
1315 * @object_size: the size of an object without meta data
1316 * @flags: flags to set
1317 * @name: name of the cache
1318 * @ctor: constructor function
1320 * Debug option(s) are applied to @flags. In addition to the debug
1321 * option(s), if a slab name (or multiple) is specified i.e.
1322 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1323 * then only the select slabs will receive the debug option(s).
1325 slab_flags_t kmem_cache_flags(unsigned int object_size,
1326 slab_flags_t flags, const char *name,
1327 void (*ctor)(void *))
1332 /* If slub_debug = 0, it folds into the if conditional. */
1333 if (!slub_debug_slabs)
1334 return flags | slub_debug;
1337 iter = slub_debug_slabs;
1342 end = strchrnul(iter, ',');
1344 glob = strnchr(iter, end - iter, '*');
1346 cmplen = glob - iter;
1348 cmplen = max_t(size_t, len, (end - iter));
1350 if (!strncmp(name, iter, cmplen)) {
1351 flags |= slub_debug;
1362 #else /* !CONFIG_SLUB_DEBUG */
1363 static inline void setup_object_debug(struct kmem_cache *s,
1364 struct page *page, void *object) {}
1366 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1368 static inline int alloc_debug_processing(struct kmem_cache *s,
1369 struct page *page, void *object, unsigned long addr) { return 0; }
1371 static inline int free_debug_processing(
1372 struct kmem_cache *s, struct page *page,
1373 void *head, void *tail, int bulk_cnt,
1374 unsigned long addr) { return 0; }
1376 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1378 static inline int check_object(struct kmem_cache *s, struct page *page,
1379 void *object, u8 val) { return 1; }
1380 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1381 struct page *page) {}
1382 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1383 struct page *page) {}
1384 slab_flags_t kmem_cache_flags(unsigned int object_size,
1385 slab_flags_t flags, const char *name,
1386 void (*ctor)(void *))
1390 #define slub_debug 0
1392 #define disable_higher_order_debug 0
1394 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1396 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1398 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1400 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1403 #endif /* CONFIG_SLUB_DEBUG */
1406 * Hooks for other subsystems that check memory allocations. In a typical
1407 * production configuration these hooks all should produce no code at all.
1409 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1411 ptr = kasan_kmalloc_large(ptr, size, flags);
1412 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1413 kmemleak_alloc(ptr, size, 1, flags);
1417 static __always_inline void kfree_hook(void *x)
1420 kasan_kfree_large(x, _RET_IP_);
1423 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1425 kmemleak_free_recursive(x, s->flags);
1428 * Trouble is that we may no longer disable interrupts in the fast path
1429 * So in order to make the debug calls that expect irqs to be
1430 * disabled we need to disable interrupts temporarily.
1432 #ifdef CONFIG_LOCKDEP
1434 unsigned long flags;
1436 local_irq_save(flags);
1437 debug_check_no_locks_freed(x, s->object_size);
1438 local_irq_restore(flags);
1441 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1442 debug_check_no_obj_freed(x, s->object_size);
1444 /* KASAN might put x into memory quarantine, delaying its reuse */
1445 return kasan_slab_free(s, x, _RET_IP_);
1448 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1449 void **head, void **tail)
1454 void *old_tail = *tail ? *tail : *head;
1457 /* Head and tail of the reconstructed freelist */
1463 next = get_freepointer(s, object);
1465 if (slab_want_init_on_free(s)) {
1467 * Clear the object and the metadata, but don't touch
1470 memset(object, 0, s->object_size);
1471 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1473 memset((char *)object + s->inuse, 0,
1474 s->size - s->inuse - rsize);
1477 /* If object's reuse doesn't have to be delayed */
1478 if (!slab_free_hook(s, object)) {
1479 /* Move object to the new freelist */
1480 set_freepointer(s, object, *head);
1485 } while (object != old_tail);
1490 return *head != NULL;
1493 static void *setup_object(struct kmem_cache *s, struct page *page,
1496 setup_object_debug(s, page, object);
1497 object = kasan_init_slab_obj(s, object);
1498 if (unlikely(s->ctor)) {
1499 kasan_unpoison_object_data(s, object);
1501 kasan_poison_object_data(s, object);
1507 * Slab allocation and freeing
1509 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1510 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1513 unsigned int order = oo_order(oo);
1515 if (node == NUMA_NO_NODE)
1516 page = alloc_pages(flags, order);
1518 page = __alloc_pages_node(node, flags, order);
1520 if (page && charge_slab_page(page, flags, order, s)) {
1521 __free_pages(page, order);
1528 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1529 /* Pre-initialize the random sequence cache */
1530 static int init_cache_random_seq(struct kmem_cache *s)
1532 unsigned int count = oo_objects(s->oo);
1535 /* Bailout if already initialised */
1539 err = cache_random_seq_create(s, count, GFP_KERNEL);
1541 pr_err("SLUB: Unable to initialize free list for %s\n",
1546 /* Transform to an offset on the set of pages */
1547 if (s->random_seq) {
1550 for (i = 0; i < count; i++)
1551 s->random_seq[i] *= s->size;
1556 /* Initialize each random sequence freelist per cache */
1557 static void __init init_freelist_randomization(void)
1559 struct kmem_cache *s;
1561 mutex_lock(&slab_mutex);
1563 list_for_each_entry(s, &slab_caches, list)
1564 init_cache_random_seq(s);
1566 mutex_unlock(&slab_mutex);
1569 /* Get the next entry on the pre-computed freelist randomized */
1570 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1571 unsigned long *pos, void *start,
1572 unsigned long page_limit,
1573 unsigned long freelist_count)
1578 * If the target page allocation failed, the number of objects on the
1579 * page might be smaller than the usual size defined by the cache.
1582 idx = s->random_seq[*pos];
1584 if (*pos >= freelist_count)
1586 } while (unlikely(idx >= page_limit));
1588 return (char *)start + idx;
1591 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1592 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1597 unsigned long idx, pos, page_limit, freelist_count;
1599 if (page->objects < 2 || !s->random_seq)
1602 freelist_count = oo_objects(s->oo);
1603 pos = get_random_int() % freelist_count;
1605 page_limit = page->objects * s->size;
1606 start = fixup_red_left(s, page_address(page));
1608 /* First entry is used as the base of the freelist */
1609 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1611 cur = setup_object(s, page, cur);
1612 page->freelist = cur;
1614 for (idx = 1; idx < page->objects; idx++) {
1615 next = next_freelist_entry(s, page, &pos, start, page_limit,
1617 next = setup_object(s, page, next);
1618 set_freepointer(s, cur, next);
1621 set_freepointer(s, cur, NULL);
1626 static inline int init_cache_random_seq(struct kmem_cache *s)
1630 static inline void init_freelist_randomization(void) { }
1631 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1635 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1637 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1640 struct kmem_cache_order_objects oo = s->oo;
1642 void *start, *p, *next;
1646 flags &= gfp_allowed_mask;
1648 if (gfpflags_allow_blocking(flags))
1651 flags |= s->allocflags;
1654 * Let the initial higher-order allocation fail under memory pressure
1655 * so we fall-back to the minimum order allocation.
1657 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1658 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1659 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1661 page = alloc_slab_page(s, alloc_gfp, node, oo);
1662 if (unlikely(!page)) {
1666 * Allocation may have failed due to fragmentation.
1667 * Try a lower order alloc if possible
1669 page = alloc_slab_page(s, alloc_gfp, node, oo);
1670 if (unlikely(!page))
1672 stat(s, ORDER_FALLBACK);
1675 page->objects = oo_objects(oo);
1677 page->slab_cache = s;
1678 __SetPageSlab(page);
1679 if (page_is_pfmemalloc(page))
1680 SetPageSlabPfmemalloc(page);
1682 kasan_poison_slab(page);
1684 start = page_address(page);
1686 setup_page_debug(s, page, start);
1688 shuffle = shuffle_freelist(s, page);
1691 start = fixup_red_left(s, start);
1692 start = setup_object(s, page, start);
1693 page->freelist = start;
1694 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1696 next = setup_object(s, page, next);
1697 set_freepointer(s, p, next);
1700 set_freepointer(s, p, NULL);
1703 page->inuse = page->objects;
1707 if (gfpflags_allow_blocking(flags))
1708 local_irq_disable();
1712 inc_slabs_node(s, page_to_nid(page), page->objects);
1717 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1719 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1720 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1721 flags &= ~GFP_SLAB_BUG_MASK;
1722 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1723 invalid_mask, &invalid_mask, flags, &flags);
1727 return allocate_slab(s,
1728 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1731 static void __free_slab(struct kmem_cache *s, struct page *page)
1733 int order = compound_order(page);
1734 int pages = 1 << order;
1736 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1739 slab_pad_check(s, page);
1740 for_each_object(p, s, page_address(page),
1742 check_object(s, page, p, SLUB_RED_INACTIVE);
1745 __ClearPageSlabPfmemalloc(page);
1746 __ClearPageSlab(page);
1748 page->mapping = NULL;
1749 if (current->reclaim_state)
1750 current->reclaim_state->reclaimed_slab += pages;
1751 uncharge_slab_page(page, order, s);
1752 __free_pages(page, order);
1755 static void rcu_free_slab(struct rcu_head *h)
1757 struct page *page = container_of(h, struct page, rcu_head);
1759 __free_slab(page->slab_cache, page);
1762 static void free_slab(struct kmem_cache *s, struct page *page)
1764 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1765 call_rcu(&page->rcu_head, rcu_free_slab);
1767 __free_slab(s, page);
1770 static void discard_slab(struct kmem_cache *s, struct page *page)
1772 dec_slabs_node(s, page_to_nid(page), page->objects);
1777 * Management of partially allocated slabs.
1780 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1783 if (tail == DEACTIVATE_TO_TAIL)
1784 list_add_tail(&page->slab_list, &n->partial);
1786 list_add(&page->slab_list, &n->partial);
1789 static inline void add_partial(struct kmem_cache_node *n,
1790 struct page *page, int tail)
1792 lockdep_assert_held(&n->list_lock);
1793 __add_partial(n, page, tail);
1796 static inline void remove_partial(struct kmem_cache_node *n,
1799 lockdep_assert_held(&n->list_lock);
1800 list_del(&page->slab_list);
1805 * Remove slab from the partial list, freeze it and
1806 * return the pointer to the freelist.
1808 * Returns a list of objects or NULL if it fails.
1810 static inline void *acquire_slab(struct kmem_cache *s,
1811 struct kmem_cache_node *n, struct page *page,
1812 int mode, int *objects)
1815 unsigned long counters;
1818 lockdep_assert_held(&n->list_lock);
1821 * Zap the freelist and set the frozen bit.
1822 * The old freelist is the list of objects for the
1823 * per cpu allocation list.
1825 freelist = page->freelist;
1826 counters = page->counters;
1827 new.counters = counters;
1828 *objects = new.objects - new.inuse;
1830 new.inuse = page->objects;
1831 new.freelist = NULL;
1833 new.freelist = freelist;
1836 VM_BUG_ON(new.frozen);
1839 if (!__cmpxchg_double_slab(s, page,
1841 new.freelist, new.counters,
1845 remove_partial(n, page);
1850 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1851 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1854 * Try to allocate a partial slab from a specific node.
1856 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1857 struct kmem_cache_cpu *c, gfp_t flags)
1859 struct page *page, *page2;
1860 void *object = NULL;
1861 unsigned int available = 0;
1865 * Racy check. If we mistakenly see no partial slabs then we
1866 * just allocate an empty slab. If we mistakenly try to get a
1867 * partial slab and there is none available then get_partials()
1870 if (!n || !n->nr_partial)
1873 spin_lock(&n->list_lock);
1874 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1877 if (!pfmemalloc_match(page, flags))
1880 t = acquire_slab(s, n, page, object == NULL, &objects);
1884 available += objects;
1887 stat(s, ALLOC_FROM_PARTIAL);
1890 put_cpu_partial(s, page, 0);
1891 stat(s, CPU_PARTIAL_NODE);
1893 if (!kmem_cache_has_cpu_partial(s)
1894 || available > slub_cpu_partial(s) / 2)
1898 spin_unlock(&n->list_lock);
1903 * Get a page from somewhere. Search in increasing NUMA distances.
1905 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1906 struct kmem_cache_cpu *c)
1909 struct zonelist *zonelist;
1912 enum zone_type high_zoneidx = gfp_zone(flags);
1914 unsigned int cpuset_mems_cookie;
1917 * The defrag ratio allows a configuration of the tradeoffs between
1918 * inter node defragmentation and node local allocations. A lower
1919 * defrag_ratio increases the tendency to do local allocations
1920 * instead of attempting to obtain partial slabs from other nodes.
1922 * If the defrag_ratio is set to 0 then kmalloc() always
1923 * returns node local objects. If the ratio is higher then kmalloc()
1924 * may return off node objects because partial slabs are obtained
1925 * from other nodes and filled up.
1927 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1928 * (which makes defrag_ratio = 1000) then every (well almost)
1929 * allocation will first attempt to defrag slab caches on other nodes.
1930 * This means scanning over all nodes to look for partial slabs which
1931 * may be expensive if we do it every time we are trying to find a slab
1932 * with available objects.
1934 if (!s->remote_node_defrag_ratio ||
1935 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1939 cpuset_mems_cookie = read_mems_allowed_begin();
1940 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1941 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1942 struct kmem_cache_node *n;
1944 n = get_node(s, zone_to_nid(zone));
1946 if (n && cpuset_zone_allowed(zone, flags) &&
1947 n->nr_partial > s->min_partial) {
1948 object = get_partial_node(s, n, c, flags);
1951 * Don't check read_mems_allowed_retry()
1952 * here - if mems_allowed was updated in
1953 * parallel, that was a harmless race
1954 * between allocation and the cpuset
1961 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1962 #endif /* CONFIG_NUMA */
1967 * Get a partial page, lock it and return it.
1969 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1970 struct kmem_cache_cpu *c)
1973 int searchnode = node;
1975 if (node == NUMA_NO_NODE)
1976 searchnode = numa_mem_id();
1978 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1979 if (object || node != NUMA_NO_NODE)
1982 return get_any_partial(s, flags, c);
1985 #ifdef CONFIG_PREEMPTION
1987 * Calculate the next globally unique transaction for disambiguiation
1988 * during cmpxchg. The transactions start with the cpu number and are then
1989 * incremented by CONFIG_NR_CPUS.
1991 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1994 * No preemption supported therefore also no need to check for
2000 static inline unsigned long next_tid(unsigned long tid)
2002 return tid + TID_STEP;
2005 #ifdef SLUB_DEBUG_CMPXCHG
2006 static inline unsigned int tid_to_cpu(unsigned long tid)
2008 return tid % TID_STEP;
2011 static inline unsigned long tid_to_event(unsigned long tid)
2013 return tid / TID_STEP;
2017 static inline unsigned int init_tid(int cpu)
2022 static inline void note_cmpxchg_failure(const char *n,
2023 const struct kmem_cache *s, unsigned long tid)
2025 #ifdef SLUB_DEBUG_CMPXCHG
2026 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2028 pr_info("%s %s: cmpxchg redo ", n, s->name);
2030 #ifdef CONFIG_PREEMPTION
2031 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2032 pr_warn("due to cpu change %d -> %d\n",
2033 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2036 if (tid_to_event(tid) != tid_to_event(actual_tid))
2037 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2038 tid_to_event(tid), tid_to_event(actual_tid));
2040 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2041 actual_tid, tid, next_tid(tid));
2043 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2046 static void init_kmem_cache_cpus(struct kmem_cache *s)
2050 for_each_possible_cpu(cpu)
2051 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2055 * Remove the cpu slab
2057 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2058 void *freelist, struct kmem_cache_cpu *c)
2060 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2061 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2063 enum slab_modes l = M_NONE, m = M_NONE;
2065 int tail = DEACTIVATE_TO_HEAD;
2069 if (page->freelist) {
2070 stat(s, DEACTIVATE_REMOTE_FREES);
2071 tail = DEACTIVATE_TO_TAIL;
2075 * Stage one: Free all available per cpu objects back
2076 * to the page freelist while it is still frozen. Leave the
2079 * There is no need to take the list->lock because the page
2082 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2084 unsigned long counters;
2087 prior = page->freelist;
2088 counters = page->counters;
2089 set_freepointer(s, freelist, prior);
2090 new.counters = counters;
2092 VM_BUG_ON(!new.frozen);
2094 } while (!__cmpxchg_double_slab(s, page,
2096 freelist, new.counters,
2097 "drain percpu freelist"));
2099 freelist = nextfree;
2103 * Stage two: Ensure that the page is unfrozen while the
2104 * list presence reflects the actual number of objects
2107 * We setup the list membership and then perform a cmpxchg
2108 * with the count. If there is a mismatch then the page
2109 * is not unfrozen but the page is on the wrong list.
2111 * Then we restart the process which may have to remove
2112 * the page from the list that we just put it on again
2113 * because the number of objects in the slab may have
2118 old.freelist = page->freelist;
2119 old.counters = page->counters;
2120 VM_BUG_ON(!old.frozen);
2122 /* Determine target state of the slab */
2123 new.counters = old.counters;
2126 set_freepointer(s, freelist, old.freelist);
2127 new.freelist = freelist;
2129 new.freelist = old.freelist;
2133 if (!new.inuse && n->nr_partial >= s->min_partial)
2135 else if (new.freelist) {
2140 * Taking the spinlock removes the possibility
2141 * that acquire_slab() will see a slab page that
2144 spin_lock(&n->list_lock);
2148 if (kmem_cache_debug(s) && !lock) {
2151 * This also ensures that the scanning of full
2152 * slabs from diagnostic functions will not see
2155 spin_lock(&n->list_lock);
2161 remove_partial(n, page);
2162 else if (l == M_FULL)
2163 remove_full(s, n, page);
2166 add_partial(n, page, tail);
2167 else if (m == M_FULL)
2168 add_full(s, n, page);
2172 if (!__cmpxchg_double_slab(s, page,
2173 old.freelist, old.counters,
2174 new.freelist, new.counters,
2179 spin_unlock(&n->list_lock);
2183 else if (m == M_FULL)
2184 stat(s, DEACTIVATE_FULL);
2185 else if (m == M_FREE) {
2186 stat(s, DEACTIVATE_EMPTY);
2187 discard_slab(s, page);
2196 * Unfreeze all the cpu partial slabs.
2198 * This function must be called with interrupts disabled
2199 * for the cpu using c (or some other guarantee must be there
2200 * to guarantee no concurrent accesses).
2202 static void unfreeze_partials(struct kmem_cache *s,
2203 struct kmem_cache_cpu *c)
2205 #ifdef CONFIG_SLUB_CPU_PARTIAL
2206 struct kmem_cache_node *n = NULL, *n2 = NULL;
2207 struct page *page, *discard_page = NULL;
2209 while ((page = slub_percpu_partial(c))) {
2213 slub_set_percpu_partial(c, page);
2215 n2 = get_node(s, page_to_nid(page));
2218 spin_unlock(&n->list_lock);
2221 spin_lock(&n->list_lock);
2226 old.freelist = page->freelist;
2227 old.counters = page->counters;
2228 VM_BUG_ON(!old.frozen);
2230 new.counters = old.counters;
2231 new.freelist = old.freelist;
2235 } while (!__cmpxchg_double_slab(s, page,
2236 old.freelist, old.counters,
2237 new.freelist, new.counters,
2238 "unfreezing slab"));
2240 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2241 page->next = discard_page;
2242 discard_page = page;
2244 add_partial(n, page, DEACTIVATE_TO_TAIL);
2245 stat(s, FREE_ADD_PARTIAL);
2250 spin_unlock(&n->list_lock);
2252 while (discard_page) {
2253 page = discard_page;
2254 discard_page = discard_page->next;
2256 stat(s, DEACTIVATE_EMPTY);
2257 discard_slab(s, page);
2260 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2264 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2265 * partial page slot if available.
2267 * If we did not find a slot then simply move all the partials to the
2268 * per node partial list.
2270 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2272 #ifdef CONFIG_SLUB_CPU_PARTIAL
2273 struct page *oldpage;
2281 oldpage = this_cpu_read(s->cpu_slab->partial);
2284 pobjects = oldpage->pobjects;
2285 pages = oldpage->pages;
2286 if (drain && pobjects > slub_cpu_partial(s)) {
2287 unsigned long flags;
2289 * partial array is full. Move the existing
2290 * set to the per node partial list.
2292 local_irq_save(flags);
2293 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2294 local_irq_restore(flags);
2298 stat(s, CPU_PARTIAL_DRAIN);
2303 pobjects += page->objects - page->inuse;
2305 page->pages = pages;
2306 page->pobjects = pobjects;
2307 page->next = oldpage;
2309 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2311 if (unlikely(!slub_cpu_partial(s))) {
2312 unsigned long flags;
2314 local_irq_save(flags);
2315 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2316 local_irq_restore(flags);
2319 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2322 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2324 stat(s, CPUSLAB_FLUSH);
2325 deactivate_slab(s, c->page, c->freelist, c);
2327 c->tid = next_tid(c->tid);
2333 * Called from IPI handler with interrupts disabled.
2335 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2337 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2342 unfreeze_partials(s, c);
2345 static void flush_cpu_slab(void *d)
2347 struct kmem_cache *s = d;
2349 __flush_cpu_slab(s, smp_processor_id());
2352 static bool has_cpu_slab(int cpu, void *info)
2354 struct kmem_cache *s = info;
2355 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2357 return c->page || slub_percpu_partial(c);
2360 static void flush_all(struct kmem_cache *s)
2362 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2366 * Use the cpu notifier to insure that the cpu slabs are flushed when
2369 static int slub_cpu_dead(unsigned int cpu)
2371 struct kmem_cache *s;
2372 unsigned long flags;
2374 mutex_lock(&slab_mutex);
2375 list_for_each_entry(s, &slab_caches, list) {
2376 local_irq_save(flags);
2377 __flush_cpu_slab(s, cpu);
2378 local_irq_restore(flags);
2380 mutex_unlock(&slab_mutex);
2385 * Check if the objects in a per cpu structure fit numa
2386 * locality expectations.
2388 static inline int node_match(struct page *page, int node)
2391 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2397 #ifdef CONFIG_SLUB_DEBUG
2398 static int count_free(struct page *page)
2400 return page->objects - page->inuse;
2403 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2405 return atomic_long_read(&n->total_objects);
2407 #endif /* CONFIG_SLUB_DEBUG */
2409 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2410 static unsigned long count_partial(struct kmem_cache_node *n,
2411 int (*get_count)(struct page *))
2413 unsigned long flags;
2414 unsigned long x = 0;
2417 spin_lock_irqsave(&n->list_lock, flags);
2418 list_for_each_entry(page, &n->partial, slab_list)
2419 x += get_count(page);
2420 spin_unlock_irqrestore(&n->list_lock, flags);
2423 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2425 static noinline void
2426 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2428 #ifdef CONFIG_SLUB_DEBUG
2429 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2430 DEFAULT_RATELIMIT_BURST);
2432 struct kmem_cache_node *n;
2434 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2437 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2438 nid, gfpflags, &gfpflags);
2439 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2440 s->name, s->object_size, s->size, oo_order(s->oo),
2443 if (oo_order(s->min) > get_order(s->object_size))
2444 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2447 for_each_kmem_cache_node(s, node, n) {
2448 unsigned long nr_slabs;
2449 unsigned long nr_objs;
2450 unsigned long nr_free;
2452 nr_free = count_partial(n, count_free);
2453 nr_slabs = node_nr_slabs(n);
2454 nr_objs = node_nr_objs(n);
2456 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2457 node, nr_slabs, nr_objs, nr_free);
2462 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2463 int node, struct kmem_cache_cpu **pc)
2466 struct kmem_cache_cpu *c = *pc;
2469 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2471 freelist = get_partial(s, flags, node, c);
2476 page = new_slab(s, flags, node);
2478 c = raw_cpu_ptr(s->cpu_slab);
2483 * No other reference to the page yet so we can
2484 * muck around with it freely without cmpxchg
2486 freelist = page->freelist;
2487 page->freelist = NULL;
2489 stat(s, ALLOC_SLAB);
2497 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2499 if (unlikely(PageSlabPfmemalloc(page)))
2500 return gfp_pfmemalloc_allowed(gfpflags);
2506 * Check the page->freelist of a page and either transfer the freelist to the
2507 * per cpu freelist or deactivate the page.
2509 * The page is still frozen if the return value is not NULL.
2511 * If this function returns NULL then the page has been unfrozen.
2513 * This function must be called with interrupt disabled.
2515 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2518 unsigned long counters;
2522 freelist = page->freelist;
2523 counters = page->counters;
2525 new.counters = counters;
2526 VM_BUG_ON(!new.frozen);
2528 new.inuse = page->objects;
2529 new.frozen = freelist != NULL;
2531 } while (!__cmpxchg_double_slab(s, page,
2540 * Slow path. The lockless freelist is empty or we need to perform
2543 * Processing is still very fast if new objects have been freed to the
2544 * regular freelist. In that case we simply take over the regular freelist
2545 * as the lockless freelist and zap the regular freelist.
2547 * If that is not working then we fall back to the partial lists. We take the
2548 * first element of the freelist as the object to allocate now and move the
2549 * rest of the freelist to the lockless freelist.
2551 * And if we were unable to get a new slab from the partial slab lists then
2552 * we need to allocate a new slab. This is the slowest path since it involves
2553 * a call to the page allocator and the setup of a new slab.
2555 * Version of __slab_alloc to use when we know that interrupts are
2556 * already disabled (which is the case for bulk allocation).
2558 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2559 unsigned long addr, struct kmem_cache_cpu *c)
2567 * if the node is not online or has no normal memory, just
2568 * ignore the node constraint
2570 if (unlikely(node != NUMA_NO_NODE &&
2571 !node_state(node, N_NORMAL_MEMORY)))
2572 node = NUMA_NO_NODE;
2577 if (unlikely(!node_match(page, node))) {
2579 * same as above but node_match() being false already
2580 * implies node != NUMA_NO_NODE
2582 if (!node_state(node, N_NORMAL_MEMORY)) {
2583 node = NUMA_NO_NODE;
2586 stat(s, ALLOC_NODE_MISMATCH);
2587 deactivate_slab(s, page, c->freelist, c);
2593 * By rights, we should be searching for a slab page that was
2594 * PFMEMALLOC but right now, we are losing the pfmemalloc
2595 * information when the page leaves the per-cpu allocator
2597 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2598 deactivate_slab(s, page, c->freelist, c);
2602 /* must check again c->freelist in case of cpu migration or IRQ */
2603 freelist = c->freelist;
2607 freelist = get_freelist(s, page);
2611 stat(s, DEACTIVATE_BYPASS);
2615 stat(s, ALLOC_REFILL);
2619 * freelist is pointing to the list of objects to be used.
2620 * page is pointing to the page from which the objects are obtained.
2621 * That page must be frozen for per cpu allocations to work.
2623 VM_BUG_ON(!c->page->frozen);
2624 c->freelist = get_freepointer(s, freelist);
2625 c->tid = next_tid(c->tid);
2630 if (slub_percpu_partial(c)) {
2631 page = c->page = slub_percpu_partial(c);
2632 slub_set_percpu_partial(c, page);
2633 stat(s, CPU_PARTIAL_ALLOC);
2637 freelist = new_slab_objects(s, gfpflags, node, &c);
2639 if (unlikely(!freelist)) {
2640 slab_out_of_memory(s, gfpflags, node);
2645 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2648 /* Only entered in the debug case */
2649 if (kmem_cache_debug(s) &&
2650 !alloc_debug_processing(s, page, freelist, addr))
2651 goto new_slab; /* Slab failed checks. Next slab needed */
2653 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2658 * Another one that disabled interrupt and compensates for possible
2659 * cpu changes by refetching the per cpu area pointer.
2661 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2662 unsigned long addr, struct kmem_cache_cpu *c)
2665 unsigned long flags;
2667 local_irq_save(flags);
2668 #ifdef CONFIG_PREEMPTION
2670 * We may have been preempted and rescheduled on a different
2671 * cpu before disabling interrupts. Need to reload cpu area
2674 c = this_cpu_ptr(s->cpu_slab);
2677 p = ___slab_alloc(s, gfpflags, node, addr, c);
2678 local_irq_restore(flags);
2683 * If the object has been wiped upon free, make sure it's fully initialized by
2684 * zeroing out freelist pointer.
2686 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2689 if (unlikely(slab_want_init_on_free(s)) && obj)
2690 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2694 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2695 * have the fastpath folded into their functions. So no function call
2696 * overhead for requests that can be satisfied on the fastpath.
2698 * The fastpath works by first checking if the lockless freelist can be used.
2699 * If not then __slab_alloc is called for slow processing.
2701 * Otherwise we can simply pick the next object from the lockless free list.
2703 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2704 gfp_t gfpflags, int node, unsigned long addr)
2707 struct kmem_cache_cpu *c;
2711 s = slab_pre_alloc_hook(s, gfpflags);
2716 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2717 * enabled. We may switch back and forth between cpus while
2718 * reading from one cpu area. That does not matter as long
2719 * as we end up on the original cpu again when doing the cmpxchg.
2721 * We should guarantee that tid and kmem_cache are retrieved on
2722 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2723 * to check if it is matched or not.
2726 tid = this_cpu_read(s->cpu_slab->tid);
2727 c = raw_cpu_ptr(s->cpu_slab);
2728 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2729 unlikely(tid != READ_ONCE(c->tid)));
2732 * Irqless object alloc/free algorithm used here depends on sequence
2733 * of fetching cpu_slab's data. tid should be fetched before anything
2734 * on c to guarantee that object and page associated with previous tid
2735 * won't be used with current tid. If we fetch tid first, object and
2736 * page could be one associated with next tid and our alloc/free
2737 * request will be failed. In this case, we will retry. So, no problem.
2742 * The transaction ids are globally unique per cpu and per operation on
2743 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2744 * occurs on the right processor and that there was no operation on the
2745 * linked list in between.
2748 object = c->freelist;
2750 if (unlikely(!object || !node_match(page, node))) {
2751 object = __slab_alloc(s, gfpflags, node, addr, c);
2752 stat(s, ALLOC_SLOWPATH);
2754 void *next_object = get_freepointer_safe(s, object);
2757 * The cmpxchg will only match if there was no additional
2758 * operation and if we are on the right processor.
2760 * The cmpxchg does the following atomically (without lock
2762 * 1. Relocate first pointer to the current per cpu area.
2763 * 2. Verify that tid and freelist have not been changed
2764 * 3. If they were not changed replace tid and freelist
2766 * Since this is without lock semantics the protection is only
2767 * against code executing on this cpu *not* from access by
2770 if (unlikely(!this_cpu_cmpxchg_double(
2771 s->cpu_slab->freelist, s->cpu_slab->tid,
2773 next_object, next_tid(tid)))) {
2775 note_cmpxchg_failure("slab_alloc", s, tid);
2778 prefetch_freepointer(s, next_object);
2779 stat(s, ALLOC_FASTPATH);
2782 maybe_wipe_obj_freeptr(s, object);
2784 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2785 memset(object, 0, s->object_size);
2787 slab_post_alloc_hook(s, gfpflags, 1, &object);
2792 static __always_inline void *slab_alloc(struct kmem_cache *s,
2793 gfp_t gfpflags, unsigned long addr)
2795 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2798 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2800 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2802 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2807 EXPORT_SYMBOL(kmem_cache_alloc);
2809 #ifdef CONFIG_TRACING
2810 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2812 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2813 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2814 ret = kasan_kmalloc(s, ret, size, gfpflags);
2817 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2821 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2823 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2825 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2826 s->object_size, s->size, gfpflags, node);
2830 EXPORT_SYMBOL(kmem_cache_alloc_node);
2832 #ifdef CONFIG_TRACING
2833 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2835 int node, size_t size)
2837 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2839 trace_kmalloc_node(_RET_IP_, ret,
2840 size, s->size, gfpflags, node);
2842 ret = kasan_kmalloc(s, ret, size, gfpflags);
2845 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2847 #endif /* CONFIG_NUMA */
2850 * Slow path handling. This may still be called frequently since objects
2851 * have a longer lifetime than the cpu slabs in most processing loads.
2853 * So we still attempt to reduce cache line usage. Just take the slab
2854 * lock and free the item. If there is no additional partial page
2855 * handling required then we can return immediately.
2857 static void __slab_free(struct kmem_cache *s, struct page *page,
2858 void *head, void *tail, int cnt,
2865 unsigned long counters;
2866 struct kmem_cache_node *n = NULL;
2867 unsigned long uninitialized_var(flags);
2869 stat(s, FREE_SLOWPATH);
2871 if (kmem_cache_debug(s) &&
2872 !free_debug_processing(s, page, head, tail, cnt, addr))
2877 spin_unlock_irqrestore(&n->list_lock, flags);
2880 prior = page->freelist;
2881 counters = page->counters;
2882 set_freepointer(s, tail, prior);
2883 new.counters = counters;
2884 was_frozen = new.frozen;
2886 if ((!new.inuse || !prior) && !was_frozen) {
2888 if (kmem_cache_has_cpu_partial(s) && !prior) {
2891 * Slab was on no list before and will be
2893 * We can defer the list move and instead
2898 } else { /* Needs to be taken off a list */
2900 n = get_node(s, page_to_nid(page));
2902 * Speculatively acquire the list_lock.
2903 * If the cmpxchg does not succeed then we may
2904 * drop the list_lock without any processing.
2906 * Otherwise the list_lock will synchronize with
2907 * other processors updating the list of slabs.
2909 spin_lock_irqsave(&n->list_lock, flags);
2914 } while (!cmpxchg_double_slab(s, page,
2922 * If we just froze the page then put it onto the
2923 * per cpu partial list.
2925 if (new.frozen && !was_frozen) {
2926 put_cpu_partial(s, page, 1);
2927 stat(s, CPU_PARTIAL_FREE);
2930 * The list lock was not taken therefore no list
2931 * activity can be necessary.
2934 stat(s, FREE_FROZEN);
2938 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2942 * Objects left in the slab. If it was not on the partial list before
2945 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2946 remove_full(s, n, page);
2947 add_partial(n, page, DEACTIVATE_TO_TAIL);
2948 stat(s, FREE_ADD_PARTIAL);
2950 spin_unlock_irqrestore(&n->list_lock, flags);
2956 * Slab on the partial list.
2958 remove_partial(n, page);
2959 stat(s, FREE_REMOVE_PARTIAL);
2961 /* Slab must be on the full list */
2962 remove_full(s, n, page);
2965 spin_unlock_irqrestore(&n->list_lock, flags);
2967 discard_slab(s, page);
2971 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2972 * can perform fastpath freeing without additional function calls.
2974 * The fastpath is only possible if we are freeing to the current cpu slab
2975 * of this processor. This typically the case if we have just allocated
2978 * If fastpath is not possible then fall back to __slab_free where we deal
2979 * with all sorts of special processing.
2981 * Bulk free of a freelist with several objects (all pointing to the
2982 * same page) possible by specifying head and tail ptr, plus objects
2983 * count (cnt). Bulk free indicated by tail pointer being set.
2985 static __always_inline void do_slab_free(struct kmem_cache *s,
2986 struct page *page, void *head, void *tail,
2987 int cnt, unsigned long addr)
2989 void *tail_obj = tail ? : head;
2990 struct kmem_cache_cpu *c;
2994 * Determine the currently cpus per cpu slab.
2995 * The cpu may change afterward. However that does not matter since
2996 * data is retrieved via this pointer. If we are on the same cpu
2997 * during the cmpxchg then the free will succeed.
3000 tid = this_cpu_read(s->cpu_slab->tid);
3001 c = raw_cpu_ptr(s->cpu_slab);
3002 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3003 unlikely(tid != READ_ONCE(c->tid)));
3005 /* Same with comment on barrier() in slab_alloc_node() */
3008 if (likely(page == c->page)) {
3009 void **freelist = READ_ONCE(c->freelist);
3011 set_freepointer(s, tail_obj, freelist);
3013 if (unlikely(!this_cpu_cmpxchg_double(
3014 s->cpu_slab->freelist, s->cpu_slab->tid,
3016 head, next_tid(tid)))) {
3018 note_cmpxchg_failure("slab_free", s, tid);
3021 stat(s, FREE_FASTPATH);
3023 __slab_free(s, page, head, tail_obj, cnt, addr);
3027 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3028 void *head, void *tail, int cnt,
3032 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3033 * to remove objects, whose reuse must be delayed.
3035 if (slab_free_freelist_hook(s, &head, &tail))
3036 do_slab_free(s, page, head, tail, cnt, addr);
3039 #ifdef CONFIG_KASAN_GENERIC
3040 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3042 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3046 void kmem_cache_free(struct kmem_cache *s, void *x)
3048 s = cache_from_obj(s, x);
3051 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3052 trace_kmem_cache_free(_RET_IP_, x);
3054 EXPORT_SYMBOL(kmem_cache_free);
3056 struct detached_freelist {
3061 struct kmem_cache *s;
3065 * This function progressively scans the array with free objects (with
3066 * a limited look ahead) and extract objects belonging to the same
3067 * page. It builds a detached freelist directly within the given
3068 * page/objects. This can happen without any need for
3069 * synchronization, because the objects are owned by running process.
3070 * The freelist is build up as a single linked list in the objects.
3071 * The idea is, that this detached freelist can then be bulk
3072 * transferred to the real freelist(s), but only requiring a single
3073 * synchronization primitive. Look ahead in the array is limited due
3074 * to performance reasons.
3077 int build_detached_freelist(struct kmem_cache *s, size_t size,
3078 void **p, struct detached_freelist *df)
3080 size_t first_skipped_index = 0;
3085 /* Always re-init detached_freelist */
3090 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3091 } while (!object && size);
3096 page = virt_to_head_page(object);
3098 /* Handle kalloc'ed objects */
3099 if (unlikely(!PageSlab(page))) {
3100 BUG_ON(!PageCompound(page));
3102 __free_pages(page, compound_order(page));
3103 p[size] = NULL; /* mark object processed */
3106 /* Derive kmem_cache from object */
3107 df->s = page->slab_cache;
3109 df->s = cache_from_obj(s, object); /* Support for memcg */
3112 /* Start new detached freelist */
3114 set_freepointer(df->s, object, NULL);
3116 df->freelist = object;
3117 p[size] = NULL; /* mark object processed */
3123 continue; /* Skip processed objects */
3125 /* df->page is always set at this point */
3126 if (df->page == virt_to_head_page(object)) {
3127 /* Opportunity build freelist */
3128 set_freepointer(df->s, object, df->freelist);
3129 df->freelist = object;
3131 p[size] = NULL; /* mark object processed */
3136 /* Limit look ahead search */
3140 if (!first_skipped_index)
3141 first_skipped_index = size + 1;
3144 return first_skipped_index;
3147 /* Note that interrupts must be enabled when calling this function. */
3148 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3154 struct detached_freelist df;
3156 size = build_detached_freelist(s, size, p, &df);
3160 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3161 } while (likely(size));
3163 EXPORT_SYMBOL(kmem_cache_free_bulk);
3165 /* Note that interrupts must be enabled when calling this function. */
3166 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3169 struct kmem_cache_cpu *c;
3172 /* memcg and kmem_cache debug support */
3173 s = slab_pre_alloc_hook(s, flags);
3177 * Drain objects in the per cpu slab, while disabling local
3178 * IRQs, which protects against PREEMPT and interrupts
3179 * handlers invoking normal fastpath.
3181 local_irq_disable();
3182 c = this_cpu_ptr(s->cpu_slab);
3184 for (i = 0; i < size; i++) {
3185 void *object = c->freelist;
3187 if (unlikely(!object)) {
3189 * We may have removed an object from c->freelist using
3190 * the fastpath in the previous iteration; in that case,
3191 * c->tid has not been bumped yet.
3192 * Since ___slab_alloc() may reenable interrupts while
3193 * allocating memory, we should bump c->tid now.
3195 c->tid = next_tid(c->tid);
3198 * Invoking slow path likely have side-effect
3199 * of re-populating per CPU c->freelist
3201 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3203 if (unlikely(!p[i]))
3206 c = this_cpu_ptr(s->cpu_slab);
3207 maybe_wipe_obj_freeptr(s, p[i]);
3209 continue; /* goto for-loop */
3211 c->freelist = get_freepointer(s, object);
3213 maybe_wipe_obj_freeptr(s, p[i]);
3215 c->tid = next_tid(c->tid);
3218 /* Clear memory outside IRQ disabled fastpath loop */
3219 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3222 for (j = 0; j < i; j++)
3223 memset(p[j], 0, s->object_size);
3226 /* memcg and kmem_cache debug support */
3227 slab_post_alloc_hook(s, flags, size, p);
3231 slab_post_alloc_hook(s, flags, i, p);
3232 __kmem_cache_free_bulk(s, i, p);
3235 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3239 * Object placement in a slab is made very easy because we always start at
3240 * offset 0. If we tune the size of the object to the alignment then we can
3241 * get the required alignment by putting one properly sized object after
3244 * Notice that the allocation order determines the sizes of the per cpu
3245 * caches. Each processor has always one slab available for allocations.
3246 * Increasing the allocation order reduces the number of times that slabs
3247 * must be moved on and off the partial lists and is therefore a factor in
3252 * Mininum / Maximum order of slab pages. This influences locking overhead
3253 * and slab fragmentation. A higher order reduces the number of partial slabs
3254 * and increases the number of allocations possible without having to
3255 * take the list_lock.
3257 static unsigned int slub_min_order;
3258 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3259 static unsigned int slub_min_objects;
3262 * Calculate the order of allocation given an slab object size.
3264 * The order of allocation has significant impact on performance and other
3265 * system components. Generally order 0 allocations should be preferred since
3266 * order 0 does not cause fragmentation in the page allocator. Larger objects
3267 * be problematic to put into order 0 slabs because there may be too much
3268 * unused space left. We go to a higher order if more than 1/16th of the slab
3271 * In order to reach satisfactory performance we must ensure that a minimum
3272 * number of objects is in one slab. Otherwise we may generate too much
3273 * activity on the partial lists which requires taking the list_lock. This is
3274 * less a concern for large slabs though which are rarely used.
3276 * slub_max_order specifies the order where we begin to stop considering the
3277 * number of objects in a slab as critical. If we reach slub_max_order then
3278 * we try to keep the page order as low as possible. So we accept more waste
3279 * of space in favor of a small page order.
3281 * Higher order allocations also allow the placement of more objects in a
3282 * slab and thereby reduce object handling overhead. If the user has
3283 * requested a higher mininum order then we start with that one instead of
3284 * the smallest order which will fit the object.
3286 static inline unsigned int slab_order(unsigned int size,
3287 unsigned int min_objects, unsigned int max_order,
3288 unsigned int fract_leftover)
3290 unsigned int min_order = slub_min_order;
3293 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3294 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3296 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3297 order <= max_order; order++) {
3299 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3302 rem = slab_size % size;
3304 if (rem <= slab_size / fract_leftover)
3311 static inline int calculate_order(unsigned int size)
3314 unsigned int min_objects;
3315 unsigned int max_objects;
3318 * Attempt to find best configuration for a slab. This
3319 * works by first attempting to generate a layout with
3320 * the best configuration and backing off gradually.
3322 * First we increase the acceptable waste in a slab. Then
3323 * we reduce the minimum objects required in a slab.
3325 min_objects = slub_min_objects;
3327 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3328 max_objects = order_objects(slub_max_order, size);
3329 min_objects = min(min_objects, max_objects);
3331 while (min_objects > 1) {
3332 unsigned int fraction;
3335 while (fraction >= 4) {
3336 order = slab_order(size, min_objects,
3337 slub_max_order, fraction);
3338 if (order <= slub_max_order)
3346 * We were unable to place multiple objects in a slab. Now
3347 * lets see if we can place a single object there.
3349 order = slab_order(size, 1, slub_max_order, 1);
3350 if (order <= slub_max_order)
3354 * Doh this slab cannot be placed using slub_max_order.
3356 order = slab_order(size, 1, MAX_ORDER, 1);
3357 if (order < MAX_ORDER)
3363 init_kmem_cache_node(struct kmem_cache_node *n)
3366 spin_lock_init(&n->list_lock);
3367 INIT_LIST_HEAD(&n->partial);
3368 #ifdef CONFIG_SLUB_DEBUG
3369 atomic_long_set(&n->nr_slabs, 0);
3370 atomic_long_set(&n->total_objects, 0);
3371 INIT_LIST_HEAD(&n->full);
3375 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3377 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3378 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3381 * Must align to double word boundary for the double cmpxchg
3382 * instructions to work; see __pcpu_double_call_return_bool().
3384 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3385 2 * sizeof(void *));
3390 init_kmem_cache_cpus(s);
3395 static struct kmem_cache *kmem_cache_node;
3398 * No kmalloc_node yet so do it by hand. We know that this is the first
3399 * slab on the node for this slabcache. There are no concurrent accesses
3402 * Note that this function only works on the kmem_cache_node
3403 * when allocating for the kmem_cache_node. This is used for bootstrapping
3404 * memory on a fresh node that has no slab structures yet.
3406 static void early_kmem_cache_node_alloc(int node)
3409 struct kmem_cache_node *n;
3411 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3413 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3416 if (page_to_nid(page) != node) {
3417 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3418 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3423 #ifdef CONFIG_SLUB_DEBUG
3424 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3425 init_tracking(kmem_cache_node, n);
3427 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3429 page->freelist = get_freepointer(kmem_cache_node, n);
3432 kmem_cache_node->node[node] = n;
3433 init_kmem_cache_node(n);
3434 inc_slabs_node(kmem_cache_node, node, page->objects);
3437 * No locks need to be taken here as it has just been
3438 * initialized and there is no concurrent access.
3440 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3443 static void free_kmem_cache_nodes(struct kmem_cache *s)
3446 struct kmem_cache_node *n;
3448 for_each_kmem_cache_node(s, node, n) {
3449 s->node[node] = NULL;
3450 kmem_cache_free(kmem_cache_node, n);
3454 void __kmem_cache_release(struct kmem_cache *s)
3456 cache_random_seq_destroy(s);
3457 free_percpu(s->cpu_slab);
3458 free_kmem_cache_nodes(s);
3461 static int init_kmem_cache_nodes(struct kmem_cache *s)
3465 for_each_node_state(node, N_NORMAL_MEMORY) {
3466 struct kmem_cache_node *n;
3468 if (slab_state == DOWN) {
3469 early_kmem_cache_node_alloc(node);
3472 n = kmem_cache_alloc_node(kmem_cache_node,
3476 free_kmem_cache_nodes(s);
3480 init_kmem_cache_node(n);
3486 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3488 if (min < MIN_PARTIAL)
3490 else if (min > MAX_PARTIAL)
3492 s->min_partial = min;
3495 static void set_cpu_partial(struct kmem_cache *s)
3497 #ifdef CONFIG_SLUB_CPU_PARTIAL
3499 * cpu_partial determined the maximum number of objects kept in the
3500 * per cpu partial lists of a processor.
3502 * Per cpu partial lists mainly contain slabs that just have one
3503 * object freed. If they are used for allocation then they can be
3504 * filled up again with minimal effort. The slab will never hit the
3505 * per node partial lists and therefore no locking will be required.
3507 * This setting also determines
3509 * A) The number of objects from per cpu partial slabs dumped to the
3510 * per node list when we reach the limit.
3511 * B) The number of objects in cpu partial slabs to extract from the
3512 * per node list when we run out of per cpu objects. We only fetch
3513 * 50% to keep some capacity around for frees.
3515 if (!kmem_cache_has_cpu_partial(s))
3516 slub_set_cpu_partial(s, 0);
3517 else if (s->size >= PAGE_SIZE)
3518 slub_set_cpu_partial(s, 2);
3519 else if (s->size >= 1024)
3520 slub_set_cpu_partial(s, 6);
3521 else if (s->size >= 256)
3522 slub_set_cpu_partial(s, 13);
3524 slub_set_cpu_partial(s, 30);
3529 * calculate_sizes() determines the order and the distribution of data within
3532 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3534 slab_flags_t flags = s->flags;
3535 unsigned int size = s->object_size;
3539 * Round up object size to the next word boundary. We can only
3540 * place the free pointer at word boundaries and this determines
3541 * the possible location of the free pointer.
3543 size = ALIGN(size, sizeof(void *));
3545 #ifdef CONFIG_SLUB_DEBUG
3547 * Determine if we can poison the object itself. If the user of
3548 * the slab may touch the object after free or before allocation
3549 * then we should never poison the object itself.
3551 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3553 s->flags |= __OBJECT_POISON;
3555 s->flags &= ~__OBJECT_POISON;
3559 * If we are Redzoning then check if there is some space between the
3560 * end of the object and the free pointer. If not then add an
3561 * additional word to have some bytes to store Redzone information.
3563 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3564 size += sizeof(void *);
3568 * With that we have determined the number of bytes in actual use
3569 * by the object. This is the potential offset to the free pointer.
3573 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3576 * Relocate free pointer after the object if it is not
3577 * permitted to overwrite the first word of the object on
3580 * This is the case if we do RCU, have a constructor or
3581 * destructor or are poisoning the objects.
3584 size += sizeof(void *);
3585 } else if (size > sizeof(void *)) {
3587 * Store freelist pointer near middle of object to keep
3588 * it away from the edges of the object to avoid small
3589 * sized over/underflows from neighboring allocations.
3591 s->offset = ALIGN(size / 2, sizeof(void *));
3594 #ifdef CONFIG_SLUB_DEBUG
3595 if (flags & SLAB_STORE_USER)
3597 * Need to store information about allocs and frees after
3600 size += 2 * sizeof(struct track);
3603 kasan_cache_create(s, &size, &s->flags);
3604 #ifdef CONFIG_SLUB_DEBUG
3605 if (flags & SLAB_RED_ZONE) {
3607 * Add some empty padding so that we can catch
3608 * overwrites from earlier objects rather than let
3609 * tracking information or the free pointer be
3610 * corrupted if a user writes before the start
3613 size += sizeof(void *);
3615 s->red_left_pad = sizeof(void *);
3616 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3617 size += s->red_left_pad;
3622 * SLUB stores one object immediately after another beginning from
3623 * offset 0. In order to align the objects we have to simply size
3624 * each object to conform to the alignment.
3626 size = ALIGN(size, s->align);
3628 if (forced_order >= 0)
3629 order = forced_order;
3631 order = calculate_order(size);
3638 s->allocflags |= __GFP_COMP;
3640 if (s->flags & SLAB_CACHE_DMA)
3641 s->allocflags |= GFP_DMA;
3643 if (s->flags & SLAB_CACHE_DMA32)
3644 s->allocflags |= GFP_DMA32;
3646 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3647 s->allocflags |= __GFP_RECLAIMABLE;
3650 * Determine the number of objects per slab
3652 s->oo = oo_make(order, size);
3653 s->min = oo_make(get_order(size), size);
3654 if (oo_objects(s->oo) > oo_objects(s->max))
3657 return !!oo_objects(s->oo);
3660 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3662 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3663 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3664 s->random = get_random_long();
3667 if (!calculate_sizes(s, -1))
3669 if (disable_higher_order_debug) {
3671 * Disable debugging flags that store metadata if the min slab
3674 if (get_order(s->size) > get_order(s->object_size)) {
3675 s->flags &= ~DEBUG_METADATA_FLAGS;
3677 if (!calculate_sizes(s, -1))
3682 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3683 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3684 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3685 /* Enable fast mode */
3686 s->flags |= __CMPXCHG_DOUBLE;
3690 * The larger the object size is, the more pages we want on the partial
3691 * list to avoid pounding the page allocator excessively.
3693 set_min_partial(s, ilog2(s->size) / 2);
3698 s->remote_node_defrag_ratio = 1000;
3701 /* Initialize the pre-computed randomized freelist if slab is up */
3702 if (slab_state >= UP) {
3703 if (init_cache_random_seq(s))
3707 if (!init_kmem_cache_nodes(s))
3710 if (alloc_kmem_cache_cpus(s))
3713 free_kmem_cache_nodes(s);
3718 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3721 #ifdef CONFIG_SLUB_DEBUG
3722 void *addr = page_address(page);
3726 slab_err(s, page, text, s->name);
3729 map = get_map(s, page);
3730 for_each_object(p, s, addr, page->objects) {
3732 if (!test_bit(slab_index(p, s, addr), map)) {
3733 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3734 print_tracking(s, p);
3744 * Attempt to free all partial slabs on a node.
3745 * This is called from __kmem_cache_shutdown(). We must take list_lock
3746 * because sysfs file might still access partial list after the shutdowning.
3748 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3751 struct page *page, *h;
3753 BUG_ON(irqs_disabled());
3754 spin_lock_irq(&n->list_lock);
3755 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3757 remove_partial(n, page);
3758 list_add(&page->slab_list, &discard);
3760 list_slab_objects(s, page,
3761 "Objects remaining in %s on __kmem_cache_shutdown()");
3764 spin_unlock_irq(&n->list_lock);
3766 list_for_each_entry_safe(page, h, &discard, slab_list)
3767 discard_slab(s, page);
3770 bool __kmem_cache_empty(struct kmem_cache *s)
3773 struct kmem_cache_node *n;
3775 for_each_kmem_cache_node(s, node, n)
3776 if (n->nr_partial || slabs_node(s, node))
3782 * Release all resources used by a slab cache.
3784 int __kmem_cache_shutdown(struct kmem_cache *s)
3787 struct kmem_cache_node *n;
3790 /* Attempt to free all objects */
3791 for_each_kmem_cache_node(s, node, n) {
3793 if (n->nr_partial || slabs_node(s, node))
3796 sysfs_slab_remove(s);
3800 /********************************************************************
3802 *******************************************************************/
3804 static int __init setup_slub_min_order(char *str)
3806 get_option(&str, (int *)&slub_min_order);
3811 __setup("slub_min_order=", setup_slub_min_order);
3813 static int __init setup_slub_max_order(char *str)
3815 get_option(&str, (int *)&slub_max_order);
3816 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3821 __setup("slub_max_order=", setup_slub_max_order);
3823 static int __init setup_slub_min_objects(char *str)
3825 get_option(&str, (int *)&slub_min_objects);
3830 __setup("slub_min_objects=", setup_slub_min_objects);
3832 void *__kmalloc(size_t size, gfp_t flags)
3834 struct kmem_cache *s;
3837 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3838 return kmalloc_large(size, flags);
3840 s = kmalloc_slab(size, flags);
3842 if (unlikely(ZERO_OR_NULL_PTR(s)))
3845 ret = slab_alloc(s, flags, _RET_IP_);
3847 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3849 ret = kasan_kmalloc(s, ret, size, flags);
3853 EXPORT_SYMBOL(__kmalloc);
3856 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3860 unsigned int order = get_order(size);
3862 flags |= __GFP_COMP;
3863 page = alloc_pages_node(node, flags, order);
3865 ptr = page_address(page);
3866 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3870 return kmalloc_large_node_hook(ptr, size, flags);
3873 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3875 struct kmem_cache *s;
3878 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3879 ret = kmalloc_large_node(size, flags, node);
3881 trace_kmalloc_node(_RET_IP_, ret,
3882 size, PAGE_SIZE << get_order(size),
3888 s = kmalloc_slab(size, flags);
3890 if (unlikely(ZERO_OR_NULL_PTR(s)))
3893 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3895 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3897 ret = kasan_kmalloc(s, ret, size, flags);
3901 EXPORT_SYMBOL(__kmalloc_node);
3902 #endif /* CONFIG_NUMA */
3904 #ifdef CONFIG_HARDENED_USERCOPY
3906 * Rejects incorrectly sized objects and objects that are to be copied
3907 * to/from userspace but do not fall entirely within the containing slab
3908 * cache's usercopy region.
3910 * Returns NULL if check passes, otherwise const char * to name of cache
3911 * to indicate an error.
3913 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3916 struct kmem_cache *s;
3917 unsigned int offset;
3920 ptr = kasan_reset_tag(ptr);
3922 /* Find object and usable object size. */
3923 s = page->slab_cache;
3925 /* Reject impossible pointers. */
3926 if (ptr < page_address(page))
3927 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3930 /* Find offset within object. */
3931 offset = (ptr - page_address(page)) % s->size;
3933 /* Adjust for redzone and reject if within the redzone. */
3934 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3935 if (offset < s->red_left_pad)
3936 usercopy_abort("SLUB object in left red zone",
3937 s->name, to_user, offset, n);
3938 offset -= s->red_left_pad;
3941 /* Allow address range falling entirely within usercopy region. */
3942 if (offset >= s->useroffset &&
3943 offset - s->useroffset <= s->usersize &&
3944 n <= s->useroffset - offset + s->usersize)
3948 * If the copy is still within the allocated object, produce
3949 * a warning instead of rejecting the copy. This is intended
3950 * to be a temporary method to find any missing usercopy
3953 object_size = slab_ksize(s);
3954 if (usercopy_fallback &&
3955 offset <= object_size && n <= object_size - offset) {
3956 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3960 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3962 #endif /* CONFIG_HARDENED_USERCOPY */
3964 size_t __ksize(const void *object)
3968 if (unlikely(object == ZERO_SIZE_PTR))
3971 page = virt_to_head_page(object);
3973 if (unlikely(!PageSlab(page))) {
3974 WARN_ON(!PageCompound(page));
3975 return page_size(page);
3978 return slab_ksize(page->slab_cache);
3980 EXPORT_SYMBOL(__ksize);
3982 void kfree(const void *x)
3985 void *object = (void *)x;
3987 trace_kfree(_RET_IP_, x);
3989 if (unlikely(ZERO_OR_NULL_PTR(x)))
3992 page = virt_to_head_page(x);
3993 if (unlikely(!PageSlab(page))) {
3994 unsigned int order = compound_order(page);
3996 BUG_ON(!PageCompound(page));
3998 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
4000 __free_pages(page, order);
4003 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4005 EXPORT_SYMBOL(kfree);
4007 #define SHRINK_PROMOTE_MAX 32
4010 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4011 * up most to the head of the partial lists. New allocations will then
4012 * fill those up and thus they can be removed from the partial lists.
4014 * The slabs with the least items are placed last. This results in them
4015 * being allocated from last increasing the chance that the last objects
4016 * are freed in them.
4018 int __kmem_cache_shrink(struct kmem_cache *s)
4022 struct kmem_cache_node *n;
4025 struct list_head discard;
4026 struct list_head promote[SHRINK_PROMOTE_MAX];
4027 unsigned long flags;
4031 for_each_kmem_cache_node(s, node, n) {
4032 INIT_LIST_HEAD(&discard);
4033 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4034 INIT_LIST_HEAD(promote + i);
4036 spin_lock_irqsave(&n->list_lock, flags);
4039 * Build lists of slabs to discard or promote.
4041 * Note that concurrent frees may occur while we hold the
4042 * list_lock. page->inuse here is the upper limit.
4044 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4045 int free = page->objects - page->inuse;
4047 /* Do not reread page->inuse */
4050 /* We do not keep full slabs on the list */
4053 if (free == page->objects) {
4054 list_move(&page->slab_list, &discard);
4056 } else if (free <= SHRINK_PROMOTE_MAX)
4057 list_move(&page->slab_list, promote + free - 1);
4061 * Promote the slabs filled up most to the head of the
4064 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4065 list_splice(promote + i, &n->partial);
4067 spin_unlock_irqrestore(&n->list_lock, flags);
4069 /* Release empty slabs */
4070 list_for_each_entry_safe(page, t, &discard, slab_list)
4071 discard_slab(s, page);
4073 if (slabs_node(s, node))
4081 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4084 * Called with all the locks held after a sched RCU grace period.
4085 * Even if @s becomes empty after shrinking, we can't know that @s
4086 * doesn't have allocations already in-flight and thus can't
4087 * destroy @s until the associated memcg is released.
4089 * However, let's remove the sysfs files for empty caches here.
4090 * Each cache has a lot of interface files which aren't
4091 * particularly useful for empty draining caches; otherwise, we can
4092 * easily end up with millions of unnecessary sysfs files on
4093 * systems which have a lot of memory and transient cgroups.
4095 if (!__kmem_cache_shrink(s))
4096 sysfs_slab_remove(s);
4099 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4102 * Disable empty slabs caching. Used to avoid pinning offline
4103 * memory cgroups by kmem pages that can be freed.
4105 slub_set_cpu_partial(s, 0);
4108 #endif /* CONFIG_MEMCG */
4110 static int slab_mem_going_offline_callback(void *arg)
4112 struct kmem_cache *s;
4114 mutex_lock(&slab_mutex);
4115 list_for_each_entry(s, &slab_caches, list)
4116 __kmem_cache_shrink(s);
4117 mutex_unlock(&slab_mutex);
4122 static void slab_mem_offline_callback(void *arg)
4124 struct kmem_cache_node *n;
4125 struct kmem_cache *s;
4126 struct memory_notify *marg = arg;
4129 offline_node = marg->status_change_nid_normal;
4132 * If the node still has available memory. we need kmem_cache_node
4135 if (offline_node < 0)
4138 mutex_lock(&slab_mutex);
4139 list_for_each_entry(s, &slab_caches, list) {
4140 n = get_node(s, offline_node);
4143 * if n->nr_slabs > 0, slabs still exist on the node
4144 * that is going down. We were unable to free them,
4145 * and offline_pages() function shouldn't call this
4146 * callback. So, we must fail.
4148 BUG_ON(slabs_node(s, offline_node));
4150 s->node[offline_node] = NULL;
4151 kmem_cache_free(kmem_cache_node, n);
4154 mutex_unlock(&slab_mutex);
4157 static int slab_mem_going_online_callback(void *arg)
4159 struct kmem_cache_node *n;
4160 struct kmem_cache *s;
4161 struct memory_notify *marg = arg;
4162 int nid = marg->status_change_nid_normal;
4166 * If the node's memory is already available, then kmem_cache_node is
4167 * already created. Nothing to do.
4173 * We are bringing a node online. No memory is available yet. We must
4174 * allocate a kmem_cache_node structure in order to bring the node
4177 mutex_lock(&slab_mutex);
4178 list_for_each_entry(s, &slab_caches, list) {
4180 * XXX: kmem_cache_alloc_node will fallback to other nodes
4181 * since memory is not yet available from the node that
4184 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4189 init_kmem_cache_node(n);
4193 mutex_unlock(&slab_mutex);
4197 static int slab_memory_callback(struct notifier_block *self,
4198 unsigned long action, void *arg)
4203 case MEM_GOING_ONLINE:
4204 ret = slab_mem_going_online_callback(arg);
4206 case MEM_GOING_OFFLINE:
4207 ret = slab_mem_going_offline_callback(arg);
4210 case MEM_CANCEL_ONLINE:
4211 slab_mem_offline_callback(arg);
4214 case MEM_CANCEL_OFFLINE:
4218 ret = notifier_from_errno(ret);
4224 static struct notifier_block slab_memory_callback_nb = {
4225 .notifier_call = slab_memory_callback,
4226 .priority = SLAB_CALLBACK_PRI,
4229 /********************************************************************
4230 * Basic setup of slabs
4231 *******************************************************************/
4234 * Used for early kmem_cache structures that were allocated using
4235 * the page allocator. Allocate them properly then fix up the pointers
4236 * that may be pointing to the wrong kmem_cache structure.
4239 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4242 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4243 struct kmem_cache_node *n;
4245 memcpy(s, static_cache, kmem_cache->object_size);
4248 * This runs very early, and only the boot processor is supposed to be
4249 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4252 __flush_cpu_slab(s, smp_processor_id());
4253 for_each_kmem_cache_node(s, node, n) {
4256 list_for_each_entry(p, &n->partial, slab_list)
4259 #ifdef CONFIG_SLUB_DEBUG
4260 list_for_each_entry(p, &n->full, slab_list)
4264 slab_init_memcg_params(s);
4265 list_add(&s->list, &slab_caches);
4266 memcg_link_cache(s, NULL);
4270 void __init kmem_cache_init(void)
4272 static __initdata struct kmem_cache boot_kmem_cache,
4273 boot_kmem_cache_node;
4275 if (debug_guardpage_minorder())
4278 kmem_cache_node = &boot_kmem_cache_node;
4279 kmem_cache = &boot_kmem_cache;
4281 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4282 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4284 register_hotmemory_notifier(&slab_memory_callback_nb);
4286 /* Able to allocate the per node structures */
4287 slab_state = PARTIAL;
4289 create_boot_cache(kmem_cache, "kmem_cache",
4290 offsetof(struct kmem_cache, node) +
4291 nr_node_ids * sizeof(struct kmem_cache_node *),
4292 SLAB_HWCACHE_ALIGN, 0, 0);
4294 kmem_cache = bootstrap(&boot_kmem_cache);
4295 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4297 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4298 setup_kmalloc_cache_index_table();
4299 create_kmalloc_caches(0);
4301 /* Setup random freelists for each cache */
4302 init_freelist_randomization();
4304 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4307 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4309 slub_min_order, slub_max_order, slub_min_objects,
4310 nr_cpu_ids, nr_node_ids);
4313 void __init kmem_cache_init_late(void)
4318 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4319 slab_flags_t flags, void (*ctor)(void *))
4321 struct kmem_cache *s, *c;
4323 s = find_mergeable(size, align, flags, name, ctor);
4328 * Adjust the object sizes so that we clear
4329 * the complete object on kzalloc.
4331 s->object_size = max(s->object_size, size);
4332 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4334 for_each_memcg_cache(c, s) {
4335 c->object_size = s->object_size;
4336 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4339 if (sysfs_slab_alias(s, name)) {
4348 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4352 err = kmem_cache_open(s, flags);
4356 /* Mutex is not taken during early boot */
4357 if (slab_state <= UP)
4360 memcg_propagate_slab_attrs(s);
4361 err = sysfs_slab_add(s);
4363 __kmem_cache_release(s);
4368 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4370 struct kmem_cache *s;
4373 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4374 return kmalloc_large(size, gfpflags);
4376 s = kmalloc_slab(size, gfpflags);
4378 if (unlikely(ZERO_OR_NULL_PTR(s)))
4381 ret = slab_alloc(s, gfpflags, caller);
4383 /* Honor the call site pointer we received. */
4384 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4390 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4391 int node, unsigned long caller)
4393 struct kmem_cache *s;
4396 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4397 ret = kmalloc_large_node(size, gfpflags, node);
4399 trace_kmalloc_node(caller, ret,
4400 size, PAGE_SIZE << get_order(size),
4406 s = kmalloc_slab(size, gfpflags);
4408 if (unlikely(ZERO_OR_NULL_PTR(s)))
4411 ret = slab_alloc_node(s, gfpflags, node, caller);
4413 /* Honor the call site pointer we received. */
4414 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4421 static int count_inuse(struct page *page)
4426 static int count_total(struct page *page)
4428 return page->objects;
4432 #ifdef CONFIG_SLUB_DEBUG
4433 static void validate_slab(struct kmem_cache *s, struct page *page)
4436 void *addr = page_address(page);
4441 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4444 /* Now we know that a valid freelist exists */
4445 map = get_map(s, page);
4446 for_each_object(p, s, addr, page->objects) {
4447 u8 val = test_bit(slab_index(p, s, addr), map) ?
4448 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4450 if (!check_object(s, page, p, val))
4458 static int validate_slab_node(struct kmem_cache *s,
4459 struct kmem_cache_node *n)
4461 unsigned long count = 0;
4463 unsigned long flags;
4465 spin_lock_irqsave(&n->list_lock, flags);
4467 list_for_each_entry(page, &n->partial, slab_list) {
4468 validate_slab(s, page);
4471 if (count != n->nr_partial)
4472 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4473 s->name, count, n->nr_partial);
4475 if (!(s->flags & SLAB_STORE_USER))
4478 list_for_each_entry(page, &n->full, slab_list) {
4479 validate_slab(s, page);
4482 if (count != atomic_long_read(&n->nr_slabs))
4483 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4484 s->name, count, atomic_long_read(&n->nr_slabs));
4487 spin_unlock_irqrestore(&n->list_lock, flags);
4491 static long validate_slab_cache(struct kmem_cache *s)
4494 unsigned long count = 0;
4495 struct kmem_cache_node *n;
4498 for_each_kmem_cache_node(s, node, n)
4499 count += validate_slab_node(s, n);
4504 * Generate lists of code addresses where slabcache objects are allocated
4509 unsigned long count;
4516 DECLARE_BITMAP(cpus, NR_CPUS);
4522 unsigned long count;
4523 struct location *loc;
4526 static void free_loc_track(struct loc_track *t)
4529 free_pages((unsigned long)t->loc,
4530 get_order(sizeof(struct location) * t->max));
4533 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4538 order = get_order(sizeof(struct location) * max);
4540 l = (void *)__get_free_pages(flags, order);
4545 memcpy(l, t->loc, sizeof(struct location) * t->count);
4553 static int add_location(struct loc_track *t, struct kmem_cache *s,
4554 const struct track *track)
4556 long start, end, pos;
4558 unsigned long caddr;
4559 unsigned long age = jiffies - track->when;
4565 pos = start + (end - start + 1) / 2;
4568 * There is nothing at "end". If we end up there
4569 * we need to add something to before end.
4574 caddr = t->loc[pos].addr;
4575 if (track->addr == caddr) {
4581 if (age < l->min_time)
4583 if (age > l->max_time)
4586 if (track->pid < l->min_pid)
4587 l->min_pid = track->pid;
4588 if (track->pid > l->max_pid)
4589 l->max_pid = track->pid;
4591 cpumask_set_cpu(track->cpu,
4592 to_cpumask(l->cpus));
4594 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4598 if (track->addr < caddr)
4605 * Not found. Insert new tracking element.
4607 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4613 (t->count - pos) * sizeof(struct location));
4616 l->addr = track->addr;
4620 l->min_pid = track->pid;
4621 l->max_pid = track->pid;
4622 cpumask_clear(to_cpumask(l->cpus));
4623 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4624 nodes_clear(l->nodes);
4625 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4629 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4630 struct page *page, enum track_item alloc)
4632 void *addr = page_address(page);
4636 map = get_map(s, page);
4637 for_each_object(p, s, addr, page->objects)
4638 if (!test_bit(slab_index(p, s, addr), map))
4639 add_location(t, s, get_track(s, p, alloc));
4643 static int list_locations(struct kmem_cache *s, char *buf,
4644 enum track_item alloc)
4648 struct loc_track t = { 0, 0, NULL };
4650 struct kmem_cache_node *n;
4652 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4654 return sprintf(buf, "Out of memory\n");
4656 /* Push back cpu slabs */
4659 for_each_kmem_cache_node(s, node, n) {
4660 unsigned long flags;
4663 if (!atomic_long_read(&n->nr_slabs))
4666 spin_lock_irqsave(&n->list_lock, flags);
4667 list_for_each_entry(page, &n->partial, slab_list)
4668 process_slab(&t, s, page, alloc);
4669 list_for_each_entry(page, &n->full, slab_list)
4670 process_slab(&t, s, page, alloc);
4671 spin_unlock_irqrestore(&n->list_lock, flags);
4674 for (i = 0; i < t.count; i++) {
4675 struct location *l = &t.loc[i];
4677 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4679 len += sprintf(buf + len, "%7ld ", l->count);
4682 len += sprintf(buf + len, "%pS", (void *)l->addr);
4684 len += sprintf(buf + len, "<not-available>");
4686 if (l->sum_time != l->min_time) {
4687 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4689 (long)div_u64(l->sum_time, l->count),
4692 len += sprintf(buf + len, " age=%ld",
4695 if (l->min_pid != l->max_pid)
4696 len += sprintf(buf + len, " pid=%ld-%ld",
4697 l->min_pid, l->max_pid);
4699 len += sprintf(buf + len, " pid=%ld",
4702 if (num_online_cpus() > 1 &&
4703 !cpumask_empty(to_cpumask(l->cpus)) &&
4704 len < PAGE_SIZE - 60)
4705 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4707 cpumask_pr_args(to_cpumask(l->cpus)));
4709 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4710 len < PAGE_SIZE - 60)
4711 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4713 nodemask_pr_args(&l->nodes));
4715 len += sprintf(buf + len, "\n");
4720 len += sprintf(buf, "No data\n");
4723 #endif /* CONFIG_SLUB_DEBUG */
4725 #ifdef SLUB_RESILIENCY_TEST
4726 static void __init resiliency_test(void)
4729 int type = KMALLOC_NORMAL;
4731 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4733 pr_err("SLUB resiliency testing\n");
4734 pr_err("-----------------------\n");
4735 pr_err("A. Corruption after allocation\n");
4737 p = kzalloc(16, GFP_KERNEL);
4739 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4742 validate_slab_cache(kmalloc_caches[type][4]);
4744 /* Hmmm... The next two are dangerous */
4745 p = kzalloc(32, GFP_KERNEL);
4746 p[32 + sizeof(void *)] = 0x34;
4747 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4749 pr_err("If allocated object is overwritten then not detectable\n\n");
4751 validate_slab_cache(kmalloc_caches[type][5]);
4752 p = kzalloc(64, GFP_KERNEL);
4753 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4755 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4757 pr_err("If allocated object is overwritten then not detectable\n\n");
4758 validate_slab_cache(kmalloc_caches[type][6]);
4760 pr_err("\nB. Corruption after free\n");
4761 p = kzalloc(128, GFP_KERNEL);
4764 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4765 validate_slab_cache(kmalloc_caches[type][7]);
4767 p = kzalloc(256, GFP_KERNEL);
4770 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4771 validate_slab_cache(kmalloc_caches[type][8]);
4773 p = kzalloc(512, GFP_KERNEL);
4776 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4777 validate_slab_cache(kmalloc_caches[type][9]);
4781 static void resiliency_test(void) {};
4783 #endif /* SLUB_RESILIENCY_TEST */
4786 enum slab_stat_type {
4787 SL_ALL, /* All slabs */
4788 SL_PARTIAL, /* Only partially allocated slabs */
4789 SL_CPU, /* Only slabs used for cpu caches */
4790 SL_OBJECTS, /* Determine allocated objects not slabs */
4791 SL_TOTAL /* Determine object capacity not slabs */
4794 #define SO_ALL (1 << SL_ALL)
4795 #define SO_PARTIAL (1 << SL_PARTIAL)
4796 #define SO_CPU (1 << SL_CPU)
4797 #define SO_OBJECTS (1 << SL_OBJECTS)
4798 #define SO_TOTAL (1 << SL_TOTAL)
4801 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4803 static int __init setup_slub_memcg_sysfs(char *str)
4807 if (get_option(&str, &v) > 0)
4808 memcg_sysfs_enabled = v;
4813 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4816 static ssize_t show_slab_objects(struct kmem_cache *s,
4817 char *buf, unsigned long flags)
4819 unsigned long total = 0;
4822 unsigned long *nodes;
4824 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4828 if (flags & SO_CPU) {
4831 for_each_possible_cpu(cpu) {
4832 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4837 page = READ_ONCE(c->page);
4841 node = page_to_nid(page);
4842 if (flags & SO_TOTAL)
4844 else if (flags & SO_OBJECTS)
4852 page = slub_percpu_partial_read_once(c);
4854 node = page_to_nid(page);
4855 if (flags & SO_TOTAL)
4857 else if (flags & SO_OBJECTS)
4868 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4869 * already held which will conflict with an existing lock order:
4871 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4873 * We don't really need mem_hotplug_lock (to hold off
4874 * slab_mem_going_offline_callback) here because slab's memory hot
4875 * unplug code doesn't destroy the kmem_cache->node[] data.
4878 #ifdef CONFIG_SLUB_DEBUG
4879 if (flags & SO_ALL) {
4880 struct kmem_cache_node *n;
4882 for_each_kmem_cache_node(s, node, n) {
4884 if (flags & SO_TOTAL)
4885 x = atomic_long_read(&n->total_objects);
4886 else if (flags & SO_OBJECTS)
4887 x = atomic_long_read(&n->total_objects) -
4888 count_partial(n, count_free);
4890 x = atomic_long_read(&n->nr_slabs);
4897 if (flags & SO_PARTIAL) {
4898 struct kmem_cache_node *n;
4900 for_each_kmem_cache_node(s, node, n) {
4901 if (flags & SO_TOTAL)
4902 x = count_partial(n, count_total);
4903 else if (flags & SO_OBJECTS)
4904 x = count_partial(n, count_inuse);
4911 x = sprintf(buf, "%lu", total);
4913 for (node = 0; node < nr_node_ids; node++)
4915 x += sprintf(buf + x, " N%d=%lu",
4919 return x + sprintf(buf + x, "\n");
4922 #ifdef CONFIG_SLUB_DEBUG
4923 static int any_slab_objects(struct kmem_cache *s)
4926 struct kmem_cache_node *n;
4928 for_each_kmem_cache_node(s, node, n)
4929 if (atomic_long_read(&n->total_objects))
4936 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4937 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4939 struct slab_attribute {
4940 struct attribute attr;
4941 ssize_t (*show)(struct kmem_cache *s, char *buf);
4942 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4945 #define SLAB_ATTR_RO(_name) \
4946 static struct slab_attribute _name##_attr = \
4947 __ATTR(_name, 0400, _name##_show, NULL)
4949 #define SLAB_ATTR(_name) \
4950 static struct slab_attribute _name##_attr = \
4951 __ATTR(_name, 0600, _name##_show, _name##_store)
4953 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4955 return sprintf(buf, "%u\n", s->size);
4957 SLAB_ATTR_RO(slab_size);
4959 static ssize_t align_show(struct kmem_cache *s, char *buf)
4961 return sprintf(buf, "%u\n", s->align);
4963 SLAB_ATTR_RO(align);
4965 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4967 return sprintf(buf, "%u\n", s->object_size);
4969 SLAB_ATTR_RO(object_size);
4971 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4973 return sprintf(buf, "%u\n", oo_objects(s->oo));
4975 SLAB_ATTR_RO(objs_per_slab);
4977 static ssize_t order_store(struct kmem_cache *s,
4978 const char *buf, size_t length)
4983 err = kstrtouint(buf, 10, &order);
4987 if (order > slub_max_order || order < slub_min_order)
4990 calculate_sizes(s, order);
4994 static ssize_t order_show(struct kmem_cache *s, char *buf)
4996 return sprintf(buf, "%u\n", oo_order(s->oo));
5000 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5002 return sprintf(buf, "%lu\n", s->min_partial);
5005 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5011 err = kstrtoul(buf, 10, &min);
5015 set_min_partial(s, min);
5018 SLAB_ATTR(min_partial);
5020 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5022 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5025 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5028 unsigned int objects;
5031 err = kstrtouint(buf, 10, &objects);
5034 if (objects && !kmem_cache_has_cpu_partial(s))
5037 slub_set_cpu_partial(s, objects);
5041 SLAB_ATTR(cpu_partial);
5043 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5047 return sprintf(buf, "%pS\n", s->ctor);
5051 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5053 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5055 SLAB_ATTR_RO(aliases);
5057 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5059 return show_slab_objects(s, buf, SO_PARTIAL);
5061 SLAB_ATTR_RO(partial);
5063 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5065 return show_slab_objects(s, buf, SO_CPU);
5067 SLAB_ATTR_RO(cpu_slabs);
5069 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5071 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5073 SLAB_ATTR_RO(objects);
5075 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5077 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5079 SLAB_ATTR_RO(objects_partial);
5081 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5088 for_each_online_cpu(cpu) {
5091 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5094 pages += page->pages;
5095 objects += page->pobjects;
5099 len = sprintf(buf, "%d(%d)", objects, pages);
5102 for_each_online_cpu(cpu) {
5105 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5107 if (page && len < PAGE_SIZE - 20)
5108 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5109 page->pobjects, page->pages);
5112 return len + sprintf(buf + len, "\n");
5114 SLAB_ATTR_RO(slabs_cpu_partial);
5116 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5118 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5121 static ssize_t reclaim_account_store(struct kmem_cache *s,
5122 const char *buf, size_t length)
5124 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5126 s->flags |= SLAB_RECLAIM_ACCOUNT;
5129 SLAB_ATTR(reclaim_account);
5131 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5133 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5135 SLAB_ATTR_RO(hwcache_align);
5137 #ifdef CONFIG_ZONE_DMA
5138 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5140 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5142 SLAB_ATTR_RO(cache_dma);
5145 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5147 return sprintf(buf, "%u\n", s->usersize);
5149 SLAB_ATTR_RO(usersize);
5151 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5153 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5155 SLAB_ATTR_RO(destroy_by_rcu);
5157 #ifdef CONFIG_SLUB_DEBUG
5158 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5160 return show_slab_objects(s, buf, SO_ALL);
5162 SLAB_ATTR_RO(slabs);
5164 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5166 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5168 SLAB_ATTR_RO(total_objects);
5170 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5172 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5175 static ssize_t sanity_checks_store(struct kmem_cache *s,
5176 const char *buf, size_t length)
5178 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5179 if (buf[0] == '1') {
5180 s->flags &= ~__CMPXCHG_DOUBLE;
5181 s->flags |= SLAB_CONSISTENCY_CHECKS;
5185 SLAB_ATTR(sanity_checks);
5187 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5189 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5192 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5196 * Tracing a merged cache is going to give confusing results
5197 * as well as cause other issues like converting a mergeable
5198 * cache into an umergeable one.
5200 if (s->refcount > 1)
5203 s->flags &= ~SLAB_TRACE;
5204 if (buf[0] == '1') {
5205 s->flags &= ~__CMPXCHG_DOUBLE;
5206 s->flags |= SLAB_TRACE;
5212 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5214 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5217 static ssize_t red_zone_store(struct kmem_cache *s,
5218 const char *buf, size_t length)
5220 if (any_slab_objects(s))
5223 s->flags &= ~SLAB_RED_ZONE;
5224 if (buf[0] == '1') {
5225 s->flags |= SLAB_RED_ZONE;
5227 calculate_sizes(s, -1);
5230 SLAB_ATTR(red_zone);
5232 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5234 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5237 static ssize_t poison_store(struct kmem_cache *s,
5238 const char *buf, size_t length)
5240 if (any_slab_objects(s))
5243 s->flags &= ~SLAB_POISON;
5244 if (buf[0] == '1') {
5245 s->flags |= SLAB_POISON;
5247 calculate_sizes(s, -1);
5252 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5254 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5257 static ssize_t store_user_store(struct kmem_cache *s,
5258 const char *buf, size_t length)
5260 if (any_slab_objects(s))
5263 s->flags &= ~SLAB_STORE_USER;
5264 if (buf[0] == '1') {
5265 s->flags &= ~__CMPXCHG_DOUBLE;
5266 s->flags |= SLAB_STORE_USER;
5268 calculate_sizes(s, -1);
5271 SLAB_ATTR(store_user);
5273 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5278 static ssize_t validate_store(struct kmem_cache *s,
5279 const char *buf, size_t length)
5283 if (buf[0] == '1') {
5284 ret = validate_slab_cache(s);
5290 SLAB_ATTR(validate);
5292 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5294 if (!(s->flags & SLAB_STORE_USER))
5296 return list_locations(s, buf, TRACK_ALLOC);
5298 SLAB_ATTR_RO(alloc_calls);
5300 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5302 if (!(s->flags & SLAB_STORE_USER))
5304 return list_locations(s, buf, TRACK_FREE);
5306 SLAB_ATTR_RO(free_calls);
5307 #endif /* CONFIG_SLUB_DEBUG */
5309 #ifdef CONFIG_FAILSLAB
5310 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5312 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5315 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5318 if (s->refcount > 1)
5321 s->flags &= ~SLAB_FAILSLAB;
5323 s->flags |= SLAB_FAILSLAB;
5326 SLAB_ATTR(failslab);
5329 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5334 static ssize_t shrink_store(struct kmem_cache *s,
5335 const char *buf, size_t length)
5338 kmem_cache_shrink_all(s);
5346 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5348 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5351 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5352 const char *buf, size_t length)
5357 err = kstrtouint(buf, 10, &ratio);
5363 s->remote_node_defrag_ratio = ratio * 10;
5367 SLAB_ATTR(remote_node_defrag_ratio);
5370 #ifdef CONFIG_SLUB_STATS
5371 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5373 unsigned long sum = 0;
5376 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5381 for_each_online_cpu(cpu) {
5382 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5388 len = sprintf(buf, "%lu", sum);
5391 for_each_online_cpu(cpu) {
5392 if (data[cpu] && len < PAGE_SIZE - 20)
5393 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5397 return len + sprintf(buf + len, "\n");
5400 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5404 for_each_online_cpu(cpu)
5405 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5408 #define STAT_ATTR(si, text) \
5409 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5411 return show_stat(s, buf, si); \
5413 static ssize_t text##_store(struct kmem_cache *s, \
5414 const char *buf, size_t length) \
5416 if (buf[0] != '0') \
5418 clear_stat(s, si); \
5423 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5424 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5425 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5426 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5427 STAT_ATTR(FREE_FROZEN, free_frozen);
5428 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5429 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5430 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5431 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5432 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5433 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5434 STAT_ATTR(FREE_SLAB, free_slab);
5435 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5436 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5437 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5438 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5439 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5440 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5441 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5442 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5443 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5444 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5445 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5446 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5447 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5448 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5449 #endif /* CONFIG_SLUB_STATS */
5451 static struct attribute *slab_attrs[] = {
5452 &slab_size_attr.attr,
5453 &object_size_attr.attr,
5454 &objs_per_slab_attr.attr,
5456 &min_partial_attr.attr,
5457 &cpu_partial_attr.attr,
5459 &objects_partial_attr.attr,
5461 &cpu_slabs_attr.attr,
5465 &hwcache_align_attr.attr,
5466 &reclaim_account_attr.attr,
5467 &destroy_by_rcu_attr.attr,
5469 &slabs_cpu_partial_attr.attr,
5470 #ifdef CONFIG_SLUB_DEBUG
5471 &total_objects_attr.attr,
5473 &sanity_checks_attr.attr,
5475 &red_zone_attr.attr,
5477 &store_user_attr.attr,
5478 &validate_attr.attr,
5479 &alloc_calls_attr.attr,
5480 &free_calls_attr.attr,
5482 #ifdef CONFIG_ZONE_DMA
5483 &cache_dma_attr.attr,
5486 &remote_node_defrag_ratio_attr.attr,
5488 #ifdef CONFIG_SLUB_STATS
5489 &alloc_fastpath_attr.attr,
5490 &alloc_slowpath_attr.attr,
5491 &free_fastpath_attr.attr,
5492 &free_slowpath_attr.attr,
5493 &free_frozen_attr.attr,
5494 &free_add_partial_attr.attr,
5495 &free_remove_partial_attr.attr,
5496 &alloc_from_partial_attr.attr,
5497 &alloc_slab_attr.attr,
5498 &alloc_refill_attr.attr,
5499 &alloc_node_mismatch_attr.attr,
5500 &free_slab_attr.attr,
5501 &cpuslab_flush_attr.attr,
5502 &deactivate_full_attr.attr,
5503 &deactivate_empty_attr.attr,
5504 &deactivate_to_head_attr.attr,
5505 &deactivate_to_tail_attr.attr,
5506 &deactivate_remote_frees_attr.attr,
5507 &deactivate_bypass_attr.attr,
5508 &order_fallback_attr.attr,
5509 &cmpxchg_double_fail_attr.attr,
5510 &cmpxchg_double_cpu_fail_attr.attr,
5511 &cpu_partial_alloc_attr.attr,
5512 &cpu_partial_free_attr.attr,
5513 &cpu_partial_node_attr.attr,
5514 &cpu_partial_drain_attr.attr,
5516 #ifdef CONFIG_FAILSLAB
5517 &failslab_attr.attr,
5519 &usersize_attr.attr,
5524 static const struct attribute_group slab_attr_group = {
5525 .attrs = slab_attrs,
5528 static ssize_t slab_attr_show(struct kobject *kobj,
5529 struct attribute *attr,
5532 struct slab_attribute *attribute;
5533 struct kmem_cache *s;
5536 attribute = to_slab_attr(attr);
5539 if (!attribute->show)
5542 err = attribute->show(s, buf);
5547 static ssize_t slab_attr_store(struct kobject *kobj,
5548 struct attribute *attr,
5549 const char *buf, size_t len)
5551 struct slab_attribute *attribute;
5552 struct kmem_cache *s;
5555 attribute = to_slab_attr(attr);
5558 if (!attribute->store)
5561 err = attribute->store(s, buf, len);
5563 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5564 struct kmem_cache *c;
5566 mutex_lock(&slab_mutex);
5567 if (s->max_attr_size < len)
5568 s->max_attr_size = len;
5571 * This is a best effort propagation, so this function's return
5572 * value will be determined by the parent cache only. This is
5573 * basically because not all attributes will have a well
5574 * defined semantics for rollbacks - most of the actions will
5575 * have permanent effects.
5577 * Returning the error value of any of the children that fail
5578 * is not 100 % defined, in the sense that users seeing the
5579 * error code won't be able to know anything about the state of
5582 * Only returning the error code for the parent cache at least
5583 * has well defined semantics. The cache being written to
5584 * directly either failed or succeeded, in which case we loop
5585 * through the descendants with best-effort propagation.
5587 for_each_memcg_cache(c, s)
5588 attribute->store(c, buf, len);
5589 mutex_unlock(&slab_mutex);
5595 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5599 char *buffer = NULL;
5600 struct kmem_cache *root_cache;
5602 if (is_root_cache(s))
5605 root_cache = s->memcg_params.root_cache;
5608 * This mean this cache had no attribute written. Therefore, no point
5609 * in copying default values around
5611 if (!root_cache->max_attr_size)
5614 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5617 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5620 if (!attr || !attr->store || !attr->show)
5624 * It is really bad that we have to allocate here, so we will
5625 * do it only as a fallback. If we actually allocate, though,
5626 * we can just use the allocated buffer until the end.
5628 * Most of the slub attributes will tend to be very small in
5629 * size, but sysfs allows buffers up to a page, so they can
5630 * theoretically happen.
5634 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5637 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5638 if (WARN_ON(!buffer))
5643 len = attr->show(root_cache, buf);
5645 attr->store(s, buf, len);
5649 free_page((unsigned long)buffer);
5650 #endif /* CONFIG_MEMCG */
5653 static void kmem_cache_release(struct kobject *k)
5655 slab_kmem_cache_release(to_slab(k));
5658 static const struct sysfs_ops slab_sysfs_ops = {
5659 .show = slab_attr_show,
5660 .store = slab_attr_store,
5663 static struct kobj_type slab_ktype = {
5664 .sysfs_ops = &slab_sysfs_ops,
5665 .release = kmem_cache_release,
5668 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5670 struct kobj_type *ktype = get_ktype(kobj);
5672 if (ktype == &slab_ktype)
5677 static const struct kset_uevent_ops slab_uevent_ops = {
5678 .filter = uevent_filter,
5681 static struct kset *slab_kset;
5683 static inline struct kset *cache_kset(struct kmem_cache *s)
5686 if (!is_root_cache(s))
5687 return s->memcg_params.root_cache->memcg_kset;
5692 #define ID_STR_LENGTH 64
5694 /* Create a unique string id for a slab cache:
5696 * Format :[flags-]size
5698 static char *create_unique_id(struct kmem_cache *s)
5700 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5707 * First flags affecting slabcache operations. We will only
5708 * get here for aliasable slabs so we do not need to support
5709 * too many flags. The flags here must cover all flags that
5710 * are matched during merging to guarantee that the id is
5713 if (s->flags & SLAB_CACHE_DMA)
5715 if (s->flags & SLAB_CACHE_DMA32)
5717 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5719 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5721 if (s->flags & SLAB_ACCOUNT)
5725 p += sprintf(p, "%07u", s->size);
5727 BUG_ON(p > name + ID_STR_LENGTH - 1);
5731 static void sysfs_slab_remove_workfn(struct work_struct *work)
5733 struct kmem_cache *s =
5734 container_of(work, struct kmem_cache, kobj_remove_work);
5736 if (!s->kobj.state_in_sysfs)
5738 * For a memcg cache, this may be called during
5739 * deactivation and again on shutdown. Remove only once.
5740 * A cache is never shut down before deactivation is
5741 * complete, so no need to worry about synchronization.
5746 kset_unregister(s->memcg_kset);
5748 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5750 kobject_put(&s->kobj);
5753 static int sysfs_slab_add(struct kmem_cache *s)
5757 struct kset *kset = cache_kset(s);
5758 int unmergeable = slab_unmergeable(s);
5760 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5763 kobject_init(&s->kobj, &slab_ktype);
5767 if (!unmergeable && disable_higher_order_debug &&
5768 (slub_debug & DEBUG_METADATA_FLAGS))
5773 * Slabcache can never be merged so we can use the name proper.
5774 * This is typically the case for debug situations. In that
5775 * case we can catch duplicate names easily.
5777 sysfs_remove_link(&slab_kset->kobj, s->name);
5781 * Create a unique name for the slab as a target
5784 name = create_unique_id(s);
5787 s->kobj.kset = kset;
5788 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5792 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5797 if (is_root_cache(s) && memcg_sysfs_enabled) {
5798 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5799 if (!s->memcg_kset) {
5806 kobject_uevent(&s->kobj, KOBJ_ADD);
5808 /* Setup first alias */
5809 sysfs_slab_alias(s, s->name);
5816 kobject_del(&s->kobj);
5820 static void sysfs_slab_remove(struct kmem_cache *s)
5822 if (slab_state < FULL)
5824 * Sysfs has not been setup yet so no need to remove the
5829 kobject_get(&s->kobj);
5830 schedule_work(&s->kobj_remove_work);
5833 void sysfs_slab_unlink(struct kmem_cache *s)
5835 if (slab_state >= FULL)
5836 kobject_del(&s->kobj);
5839 void sysfs_slab_release(struct kmem_cache *s)
5841 if (slab_state >= FULL)
5842 kobject_put(&s->kobj);
5846 * Need to buffer aliases during bootup until sysfs becomes
5847 * available lest we lose that information.
5849 struct saved_alias {
5850 struct kmem_cache *s;
5852 struct saved_alias *next;
5855 static struct saved_alias *alias_list;
5857 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5859 struct saved_alias *al;
5861 if (slab_state == FULL) {
5863 * If we have a leftover link then remove it.
5865 sysfs_remove_link(&slab_kset->kobj, name);
5866 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5869 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5875 al->next = alias_list;
5880 static int __init slab_sysfs_init(void)
5882 struct kmem_cache *s;
5885 mutex_lock(&slab_mutex);
5887 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5889 mutex_unlock(&slab_mutex);
5890 pr_err("Cannot register slab subsystem.\n");
5896 list_for_each_entry(s, &slab_caches, list) {
5897 err = sysfs_slab_add(s);
5899 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5903 while (alias_list) {
5904 struct saved_alias *al = alias_list;
5906 alias_list = alias_list->next;
5907 err = sysfs_slab_alias(al->s, al->name);
5909 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5914 mutex_unlock(&slab_mutex);
5919 __initcall(slab_sysfs_init);
5920 #endif /* CONFIG_SYSFS */
5923 * The /proc/slabinfo ABI
5925 #ifdef CONFIG_SLUB_DEBUG
5926 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5928 unsigned long nr_slabs = 0;
5929 unsigned long nr_objs = 0;
5930 unsigned long nr_free = 0;
5932 struct kmem_cache_node *n;
5934 for_each_kmem_cache_node(s, node, n) {
5935 nr_slabs += node_nr_slabs(n);
5936 nr_objs += node_nr_objs(n);
5937 nr_free += count_partial(n, count_free);
5940 sinfo->active_objs = nr_objs - nr_free;
5941 sinfo->num_objs = nr_objs;
5942 sinfo->active_slabs = nr_slabs;
5943 sinfo->num_slabs = nr_slabs;
5944 sinfo->objects_per_slab = oo_objects(s->oo);
5945 sinfo->cache_order = oo_order(s->oo);
5948 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5952 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5953 size_t count, loff_t *ppos)
5957 #endif /* CONFIG_SLUB_DEBUG */