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 #ifdef CONFIG_SLUB_DEBUG
118 #ifdef CONFIG_SLUB_DEBUG_ON
119 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
121 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
125 static inline bool kmem_cache_debug(struct kmem_cache *s)
127 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
130 void *fixup_red_left(struct kmem_cache *s, void *p)
132 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
133 p += s->red_left_pad;
138 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
140 #ifdef CONFIG_SLUB_CPU_PARTIAL
141 return !kmem_cache_debug(s);
148 * Issues still to be resolved:
150 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
152 * - Variable sizing of the per node arrays
155 /* Enable to test recovery from slab corruption on boot */
156 #undef SLUB_RESILIENCY_TEST
158 /* Enable to log cmpxchg failures */
159 #undef SLUB_DEBUG_CMPXCHG
162 * Mininum number of partial slabs. These will be left on the partial
163 * lists even if they are empty. kmem_cache_shrink may reclaim them.
165 #define MIN_PARTIAL 5
168 * Maximum number of desirable partial slabs.
169 * The existence of more partial slabs makes kmem_cache_shrink
170 * sort the partial list by the number of objects in use.
172 #define MAX_PARTIAL 10
174 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_STORE_USER)
178 * These debug flags cannot use CMPXCHG because there might be consistency
179 * issues when checking or reading debug information
181 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
186 * Debugging flags that require metadata to be stored in the slab. These get
187 * disabled when slub_debug=O is used and a cache's min order increases with
190 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
193 #define OO_MASK ((1 << OO_SHIFT) - 1)
194 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
196 /* Internal SLUB flags */
198 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
199 /* Use cmpxchg_double */
200 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
203 * Tracking user of a slab.
205 #define TRACK_ADDRS_COUNT 16
207 unsigned long addr; /* Called from address */
208 #ifdef CONFIG_STACKTRACE
209 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
211 int cpu; /* Was running on cpu */
212 int pid; /* Pid context */
213 unsigned long when; /* When did the operation occur */
216 enum track_item { TRACK_ALLOC, TRACK_FREE };
219 static int sysfs_slab_add(struct kmem_cache *);
220 static int sysfs_slab_alias(struct kmem_cache *, const char *);
221 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
222 static void sysfs_slab_remove(struct kmem_cache *s);
224 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
225 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
227 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
228 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
231 static inline void stat(const struct kmem_cache *s, enum stat_item si)
233 #ifdef CONFIG_SLUB_STATS
235 * The rmw is racy on a preemptible kernel but this is acceptable, so
236 * avoid this_cpu_add()'s irq-disable overhead.
238 raw_cpu_inc(s->cpu_slab->stat[si]);
242 /********************************************************************
243 * Core slab cache functions
244 *******************************************************************/
247 * Returns freelist pointer (ptr). With hardening, this is obfuscated
248 * with an XOR of the address where the pointer is held and a per-cache
251 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
252 unsigned long ptr_addr)
254 #ifdef CONFIG_SLAB_FREELIST_HARDENED
256 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
257 * Normally, this doesn't cause any issues, as both set_freepointer()
258 * and get_freepointer() are called with a pointer with the same tag.
259 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
260 * example, when __free_slub() iterates over objects in a cache, it
261 * passes untagged pointers to check_object(). check_object() in turns
262 * calls get_freepointer() with an untagged pointer, which causes the
263 * freepointer to be restored incorrectly.
265 return (void *)((unsigned long)ptr ^ s->random ^
266 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
272 /* Returns the freelist pointer recorded at location ptr_addr. */
273 static inline void *freelist_dereference(const struct kmem_cache *s,
276 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
277 (unsigned long)ptr_addr);
280 static inline void *get_freepointer(struct kmem_cache *s, void *object)
282 return freelist_dereference(s, object + s->offset);
285 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
287 prefetch(object + s->offset);
290 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
292 unsigned long freepointer_addr;
295 if (!debug_pagealloc_enabled_static())
296 return get_freepointer(s, object);
298 freepointer_addr = (unsigned long)object + s->offset;
299 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
300 return freelist_ptr(s, p, freepointer_addr);
303 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
305 unsigned long freeptr_addr = (unsigned long)object + s->offset;
307 #ifdef CONFIG_SLAB_FREELIST_HARDENED
308 BUG_ON(object == fp); /* naive detection of double free or corruption */
311 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
314 /* Loop over all objects in a slab */
315 #define for_each_object(__p, __s, __addr, __objects) \
316 for (__p = fixup_red_left(__s, __addr); \
317 __p < (__addr) + (__objects) * (__s)->size; \
320 static inline unsigned int order_objects(unsigned int order, unsigned int size)
322 return ((unsigned int)PAGE_SIZE << order) / size;
325 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
328 struct kmem_cache_order_objects x = {
329 (order << OO_SHIFT) + order_objects(order, size)
335 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
337 return x.x >> OO_SHIFT;
340 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
342 return x.x & OO_MASK;
346 * Per slab locking using the pagelock
348 static __always_inline void slab_lock(struct page *page)
350 VM_BUG_ON_PAGE(PageTail(page), page);
351 bit_spin_lock(PG_locked, &page->flags);
354 static __always_inline void slab_unlock(struct page *page)
356 VM_BUG_ON_PAGE(PageTail(page), page);
357 __bit_spin_unlock(PG_locked, &page->flags);
360 /* Interrupts must be disabled (for the fallback code to work right) */
361 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
362 void *freelist_old, unsigned long counters_old,
363 void *freelist_new, unsigned long counters_new,
366 VM_BUG_ON(!irqs_disabled());
367 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
368 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
369 if (s->flags & __CMPXCHG_DOUBLE) {
370 if (cmpxchg_double(&page->freelist, &page->counters,
371 freelist_old, counters_old,
372 freelist_new, counters_new))
378 if (page->freelist == freelist_old &&
379 page->counters == counters_old) {
380 page->freelist = freelist_new;
381 page->counters = counters_new;
389 stat(s, CMPXCHG_DOUBLE_FAIL);
391 #ifdef SLUB_DEBUG_CMPXCHG
392 pr_info("%s %s: cmpxchg double redo ", n, s->name);
398 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
399 void *freelist_old, unsigned long counters_old,
400 void *freelist_new, unsigned long counters_new,
403 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
404 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
405 if (s->flags & __CMPXCHG_DOUBLE) {
406 if (cmpxchg_double(&page->freelist, &page->counters,
407 freelist_old, counters_old,
408 freelist_new, counters_new))
415 local_irq_save(flags);
417 if (page->freelist == freelist_old &&
418 page->counters == counters_old) {
419 page->freelist = freelist_new;
420 page->counters = counters_new;
422 local_irq_restore(flags);
426 local_irq_restore(flags);
430 stat(s, CMPXCHG_DOUBLE_FAIL);
432 #ifdef SLUB_DEBUG_CMPXCHG
433 pr_info("%s %s: cmpxchg double redo ", n, s->name);
439 #ifdef CONFIG_SLUB_DEBUG
440 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
441 static DEFINE_SPINLOCK(object_map_lock);
444 * Determine a map of object in use on a page.
446 * Node listlock must be held to guarantee that the page does
447 * not vanish from under us.
449 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
450 __acquires(&object_map_lock)
453 void *addr = page_address(page);
455 VM_BUG_ON(!irqs_disabled());
457 spin_lock(&object_map_lock);
459 bitmap_zero(object_map, page->objects);
461 for (p = page->freelist; p; p = get_freepointer(s, p))
462 set_bit(__obj_to_index(s, addr, p), object_map);
467 static void put_map(unsigned long *map) __releases(&object_map_lock)
469 VM_BUG_ON(map != object_map);
470 spin_unlock(&object_map_lock);
473 static inline unsigned int size_from_object(struct kmem_cache *s)
475 if (s->flags & SLAB_RED_ZONE)
476 return s->size - s->red_left_pad;
481 static inline void *restore_red_left(struct kmem_cache *s, void *p)
483 if (s->flags & SLAB_RED_ZONE)
484 p -= s->red_left_pad;
492 #if defined(CONFIG_SLUB_DEBUG_ON)
493 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
495 static slab_flags_t slub_debug;
498 static char *slub_debug_string;
499 static int disable_higher_order_debug;
502 * slub is about to manipulate internal object metadata. This memory lies
503 * outside the range of the allocated object, so accessing it would normally
504 * be reported by kasan as a bounds error. metadata_access_enable() is used
505 * to tell kasan that these accesses are OK.
507 static inline void metadata_access_enable(void)
509 kasan_disable_current();
512 static inline void metadata_access_disable(void)
514 kasan_enable_current();
521 /* Verify that a pointer has an address that is valid within a slab page */
522 static inline int check_valid_pointer(struct kmem_cache *s,
523 struct page *page, void *object)
530 base = page_address(page);
531 object = kasan_reset_tag(object);
532 object = restore_red_left(s, object);
533 if (object < base || object >= base + page->objects * s->size ||
534 (object - base) % s->size) {
541 static void print_section(char *level, char *text, u8 *addr,
544 metadata_access_enable();
545 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
547 metadata_access_disable();
551 * See comment in calculate_sizes().
553 static inline bool freeptr_outside_object(struct kmem_cache *s)
555 return s->offset >= s->inuse;
559 * Return offset of the end of info block which is inuse + free pointer if
560 * not overlapping with object.
562 static inline unsigned int get_info_end(struct kmem_cache *s)
564 if (freeptr_outside_object(s))
565 return s->inuse + sizeof(void *);
570 static struct track *get_track(struct kmem_cache *s, void *object,
571 enum track_item alloc)
575 p = object + get_info_end(s);
580 static void set_track(struct kmem_cache *s, void *object,
581 enum track_item alloc, unsigned long addr)
583 struct track *p = get_track(s, object, alloc);
586 #ifdef CONFIG_STACKTRACE
587 unsigned int nr_entries;
589 metadata_access_enable();
590 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
591 metadata_access_disable();
593 if (nr_entries < TRACK_ADDRS_COUNT)
594 p->addrs[nr_entries] = 0;
597 p->cpu = smp_processor_id();
598 p->pid = current->pid;
601 memset(p, 0, sizeof(struct track));
605 static void init_tracking(struct kmem_cache *s, void *object)
607 if (!(s->flags & SLAB_STORE_USER))
610 set_track(s, object, TRACK_FREE, 0UL);
611 set_track(s, object, TRACK_ALLOC, 0UL);
614 static void print_track(const char *s, struct track *t, unsigned long pr_time)
619 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
620 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
621 #ifdef CONFIG_STACKTRACE
624 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
626 pr_err("\t%pS\n", (void *)t->addrs[i]);
633 void print_tracking(struct kmem_cache *s, void *object)
635 unsigned long pr_time = jiffies;
636 if (!(s->flags & SLAB_STORE_USER))
639 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
640 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
643 static void print_page_info(struct page *page)
645 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
646 page, page->objects, page->inuse, page->freelist, page->flags);
650 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
652 struct va_format vaf;
658 pr_err("=============================================================================\n");
659 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
660 pr_err("-----------------------------------------------------------------------------\n\n");
662 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
666 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
668 struct va_format vaf;
674 pr_err("FIX %s: %pV\n", s->name, &vaf);
678 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
679 void *freelist, void *nextfree)
681 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
682 !check_valid_pointer(s, page, nextfree)) {
683 object_err(s, page, freelist, "Freechain corrupt");
685 slab_fix(s, "Isolate corrupted freechain");
692 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
694 unsigned int off; /* Offset of last byte */
695 u8 *addr = page_address(page);
697 print_tracking(s, p);
699 print_page_info(page);
701 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
702 p, p - addr, get_freepointer(s, p));
704 if (s->flags & SLAB_RED_ZONE)
705 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
707 else if (p > addr + 16)
708 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
710 print_section(KERN_ERR, "Object ", p,
711 min_t(unsigned int, s->object_size, PAGE_SIZE));
712 if (s->flags & SLAB_RED_ZONE)
713 print_section(KERN_ERR, "Redzone ", p + s->object_size,
714 s->inuse - s->object_size);
716 off = get_info_end(s);
718 if (s->flags & SLAB_STORE_USER)
719 off += 2 * sizeof(struct track);
721 off += kasan_metadata_size(s);
723 if (off != size_from_object(s))
724 /* Beginning of the filler is the free pointer */
725 print_section(KERN_ERR, "Padding ", p + off,
726 size_from_object(s) - off);
731 void object_err(struct kmem_cache *s, struct page *page,
732 u8 *object, char *reason)
734 slab_bug(s, "%s", reason);
735 print_trailer(s, page, object);
738 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
739 const char *fmt, ...)
745 vsnprintf(buf, sizeof(buf), fmt, args);
747 slab_bug(s, "%s", buf);
748 print_page_info(page);
752 static void init_object(struct kmem_cache *s, void *object, u8 val)
756 if (s->flags & SLAB_RED_ZONE)
757 memset(p - s->red_left_pad, val, s->red_left_pad);
759 if (s->flags & __OBJECT_POISON) {
760 memset(p, POISON_FREE, s->object_size - 1);
761 p[s->object_size - 1] = POISON_END;
764 if (s->flags & SLAB_RED_ZONE)
765 memset(p + s->object_size, val, s->inuse - s->object_size);
768 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
769 void *from, void *to)
771 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
772 memset(from, data, to - from);
775 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
776 u8 *object, char *what,
777 u8 *start, unsigned int value, unsigned int bytes)
781 u8 *addr = page_address(page);
783 metadata_access_enable();
784 fault = memchr_inv(start, value, bytes);
785 metadata_access_disable();
790 while (end > fault && end[-1] == value)
793 slab_bug(s, "%s overwritten", what);
794 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
795 fault, end - 1, fault - addr,
797 print_trailer(s, page, object);
799 restore_bytes(s, what, value, fault, end);
807 * Bytes of the object to be managed.
808 * If the freepointer may overlay the object then the free
809 * pointer is at the middle of the object.
811 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
814 * object + s->object_size
815 * Padding to reach word boundary. This is also used for Redzoning.
816 * Padding is extended by another word if Redzoning is enabled and
817 * object_size == inuse.
819 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
820 * 0xcc (RED_ACTIVE) for objects in use.
823 * Meta data starts here.
825 * A. Free pointer (if we cannot overwrite object on free)
826 * B. Tracking data for SLAB_STORE_USER
827 * C. Padding to reach required alignment boundary or at mininum
828 * one word if debugging is on to be able to detect writes
829 * before the word boundary.
831 * Padding is done using 0x5a (POISON_INUSE)
834 * Nothing is used beyond s->size.
836 * If slabcaches are merged then the object_size and inuse boundaries are mostly
837 * ignored. And therefore no slab options that rely on these boundaries
838 * may be used with merged slabcaches.
841 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
843 unsigned long off = get_info_end(s); /* The end of info */
845 if (s->flags & SLAB_STORE_USER)
846 /* We also have user information there */
847 off += 2 * sizeof(struct track);
849 off += kasan_metadata_size(s);
851 if (size_from_object(s) == off)
854 return check_bytes_and_report(s, page, p, "Object padding",
855 p + off, POISON_INUSE, size_from_object(s) - off);
858 /* Check the pad bytes at the end of a slab page */
859 static int slab_pad_check(struct kmem_cache *s, struct page *page)
868 if (!(s->flags & SLAB_POISON))
871 start = page_address(page);
872 length = page_size(page);
873 end = start + length;
874 remainder = length % s->size;
878 pad = end - remainder;
879 metadata_access_enable();
880 fault = memchr_inv(pad, POISON_INUSE, remainder);
881 metadata_access_disable();
884 while (end > fault && end[-1] == POISON_INUSE)
887 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
888 fault, end - 1, fault - start);
889 print_section(KERN_ERR, "Padding ", pad, remainder);
891 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
895 static int check_object(struct kmem_cache *s, struct page *page,
896 void *object, u8 val)
899 u8 *endobject = object + s->object_size;
901 if (s->flags & SLAB_RED_ZONE) {
902 if (!check_bytes_and_report(s, page, object, "Redzone",
903 object - s->red_left_pad, val, s->red_left_pad))
906 if (!check_bytes_and_report(s, page, object, "Redzone",
907 endobject, val, s->inuse - s->object_size))
910 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
911 check_bytes_and_report(s, page, p, "Alignment padding",
912 endobject, POISON_INUSE,
913 s->inuse - s->object_size);
917 if (s->flags & SLAB_POISON) {
918 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
919 (!check_bytes_and_report(s, page, p, "Poison", p,
920 POISON_FREE, s->object_size - 1) ||
921 !check_bytes_and_report(s, page, p, "Poison",
922 p + s->object_size - 1, POISON_END, 1)))
925 * check_pad_bytes cleans up on its own.
927 check_pad_bytes(s, page, p);
930 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
932 * Object and freepointer overlap. Cannot check
933 * freepointer while object is allocated.
937 /* Check free pointer validity */
938 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
939 object_err(s, page, p, "Freepointer corrupt");
941 * No choice but to zap it and thus lose the remainder
942 * of the free objects in this slab. May cause
943 * another error because the object count is now wrong.
945 set_freepointer(s, p, NULL);
951 static int check_slab(struct kmem_cache *s, struct page *page)
955 VM_BUG_ON(!irqs_disabled());
957 if (!PageSlab(page)) {
958 slab_err(s, page, "Not a valid slab page");
962 maxobj = order_objects(compound_order(page), s->size);
963 if (page->objects > maxobj) {
964 slab_err(s, page, "objects %u > max %u",
965 page->objects, maxobj);
968 if (page->inuse > page->objects) {
969 slab_err(s, page, "inuse %u > max %u",
970 page->inuse, page->objects);
973 /* Slab_pad_check fixes things up after itself */
974 slab_pad_check(s, page);
979 * Determine if a certain object on a page is on the freelist. Must hold the
980 * slab lock to guarantee that the chains are in a consistent state.
982 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
990 while (fp && nr <= page->objects) {
993 if (!check_valid_pointer(s, page, fp)) {
995 object_err(s, page, object,
996 "Freechain corrupt");
997 set_freepointer(s, object, NULL);
999 slab_err(s, page, "Freepointer corrupt");
1000 page->freelist = NULL;
1001 page->inuse = page->objects;
1002 slab_fix(s, "Freelist cleared");
1008 fp = get_freepointer(s, object);
1012 max_objects = order_objects(compound_order(page), s->size);
1013 if (max_objects > MAX_OBJS_PER_PAGE)
1014 max_objects = MAX_OBJS_PER_PAGE;
1016 if (page->objects != max_objects) {
1017 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1018 page->objects, max_objects);
1019 page->objects = max_objects;
1020 slab_fix(s, "Number of objects adjusted.");
1022 if (page->inuse != page->objects - nr) {
1023 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1024 page->inuse, page->objects - nr);
1025 page->inuse = page->objects - nr;
1026 slab_fix(s, "Object count adjusted.");
1028 return search == NULL;
1031 static void trace(struct kmem_cache *s, struct page *page, void *object,
1034 if (s->flags & SLAB_TRACE) {
1035 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1037 alloc ? "alloc" : "free",
1038 object, page->inuse,
1042 print_section(KERN_INFO, "Object ", (void *)object,
1050 * Tracking of fully allocated slabs for debugging purposes.
1052 static void add_full(struct kmem_cache *s,
1053 struct kmem_cache_node *n, struct page *page)
1055 if (!(s->flags & SLAB_STORE_USER))
1058 lockdep_assert_held(&n->list_lock);
1059 list_add(&page->slab_list, &n->full);
1062 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1064 if (!(s->flags & SLAB_STORE_USER))
1067 lockdep_assert_held(&n->list_lock);
1068 list_del(&page->slab_list);
1071 /* Tracking of the number of slabs for debugging purposes */
1072 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1074 struct kmem_cache_node *n = get_node(s, node);
1076 return atomic_long_read(&n->nr_slabs);
1079 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1081 return atomic_long_read(&n->nr_slabs);
1084 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1086 struct kmem_cache_node *n = get_node(s, node);
1089 * May be called early in order to allocate a slab for the
1090 * kmem_cache_node structure. Solve the chicken-egg
1091 * dilemma by deferring the increment of the count during
1092 * bootstrap (see early_kmem_cache_node_alloc).
1095 atomic_long_inc(&n->nr_slabs);
1096 atomic_long_add(objects, &n->total_objects);
1099 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1101 struct kmem_cache_node *n = get_node(s, node);
1103 atomic_long_dec(&n->nr_slabs);
1104 atomic_long_sub(objects, &n->total_objects);
1107 /* Object debug checks for alloc/free paths */
1108 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1111 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1114 init_object(s, object, SLUB_RED_INACTIVE);
1115 init_tracking(s, object);
1119 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1121 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1124 metadata_access_enable();
1125 memset(addr, POISON_INUSE, page_size(page));
1126 metadata_access_disable();
1129 static inline int alloc_consistency_checks(struct kmem_cache *s,
1130 struct page *page, void *object)
1132 if (!check_slab(s, page))
1135 if (!check_valid_pointer(s, page, object)) {
1136 object_err(s, page, object, "Freelist Pointer check fails");
1140 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1146 static noinline int alloc_debug_processing(struct kmem_cache *s,
1148 void *object, unsigned long addr)
1150 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1151 if (!alloc_consistency_checks(s, page, object))
1155 /* Success perform special debug activities for allocs */
1156 if (s->flags & SLAB_STORE_USER)
1157 set_track(s, object, TRACK_ALLOC, addr);
1158 trace(s, page, object, 1);
1159 init_object(s, object, SLUB_RED_ACTIVE);
1163 if (PageSlab(page)) {
1165 * If this is a slab page then lets do the best we can
1166 * to avoid issues in the future. Marking all objects
1167 * as used avoids touching the remaining objects.
1169 slab_fix(s, "Marking all objects used");
1170 page->inuse = page->objects;
1171 page->freelist = NULL;
1176 static inline int free_consistency_checks(struct kmem_cache *s,
1177 struct page *page, void *object, unsigned long addr)
1179 if (!check_valid_pointer(s, page, object)) {
1180 slab_err(s, page, "Invalid object pointer 0x%p", object);
1184 if (on_freelist(s, page, object)) {
1185 object_err(s, page, object, "Object already free");
1189 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1192 if (unlikely(s != page->slab_cache)) {
1193 if (!PageSlab(page)) {
1194 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1196 } else if (!page->slab_cache) {
1197 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1201 object_err(s, page, object,
1202 "page slab pointer corrupt.");
1208 /* Supports checking bulk free of a constructed freelist */
1209 static noinline int free_debug_processing(
1210 struct kmem_cache *s, struct page *page,
1211 void *head, void *tail, int bulk_cnt,
1214 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1215 void *object = head;
1217 unsigned long flags;
1220 spin_lock_irqsave(&n->list_lock, flags);
1223 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1224 if (!check_slab(s, page))
1231 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1232 if (!free_consistency_checks(s, page, object, addr))
1236 if (s->flags & SLAB_STORE_USER)
1237 set_track(s, object, TRACK_FREE, addr);
1238 trace(s, page, object, 0);
1239 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1240 init_object(s, object, SLUB_RED_INACTIVE);
1242 /* Reached end of constructed freelist yet? */
1243 if (object != tail) {
1244 object = get_freepointer(s, object);
1250 if (cnt != bulk_cnt)
1251 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1255 spin_unlock_irqrestore(&n->list_lock, flags);
1257 slab_fix(s, "Object at 0x%p not freed", object);
1262 * Parse a block of slub_debug options. Blocks are delimited by ';'
1264 * @str: start of block
1265 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1266 * @slabs: return start of list of slabs, or NULL when there's no list
1267 * @init: assume this is initial parsing and not per-kmem-create parsing
1269 * returns the start of next block if there's any, or NULL
1272 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1274 bool higher_order_disable = false;
1276 /* Skip any completely empty blocks */
1277 while (*str && *str == ';')
1282 * No options but restriction on slabs. This means full
1283 * debugging for slabs matching a pattern.
1285 *flags = DEBUG_DEFAULT_FLAGS;
1290 /* Determine which debug features should be switched on */
1291 for (; *str && *str != ',' && *str != ';'; str++) {
1292 switch (tolower(*str)) {
1297 *flags |= SLAB_CONSISTENCY_CHECKS;
1300 *flags |= SLAB_RED_ZONE;
1303 *flags |= SLAB_POISON;
1306 *flags |= SLAB_STORE_USER;
1309 *flags |= SLAB_TRACE;
1312 *flags |= SLAB_FAILSLAB;
1316 * Avoid enabling debugging on caches if its minimum
1317 * order would increase as a result.
1319 higher_order_disable = true;
1323 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1332 /* Skip over the slab list */
1333 while (*str && *str != ';')
1336 /* Skip any completely empty blocks */
1337 while (*str && *str == ';')
1340 if (init && higher_order_disable)
1341 disable_higher_order_debug = 1;
1349 static int __init setup_slub_debug(char *str)
1354 bool global_slub_debug_changed = false;
1355 bool slab_list_specified = false;
1357 slub_debug = DEBUG_DEFAULT_FLAGS;
1358 if (*str++ != '=' || !*str)
1360 * No options specified. Switch on full debugging.
1366 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1370 global_slub_debug_changed = true;
1372 slab_list_specified = true;
1377 * For backwards compatibility, a single list of flags with list of
1378 * slabs means debugging is only enabled for those slabs, so the global
1379 * slub_debug should be 0. We can extended that to multiple lists as
1380 * long as there is no option specifying flags without a slab list.
1382 if (slab_list_specified) {
1383 if (!global_slub_debug_changed)
1385 slub_debug_string = saved_str;
1388 if (slub_debug != 0 || slub_debug_string)
1389 static_branch_enable(&slub_debug_enabled);
1390 if ((static_branch_unlikely(&init_on_alloc) ||
1391 static_branch_unlikely(&init_on_free)) &&
1392 (slub_debug & SLAB_POISON))
1393 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1397 __setup("slub_debug", setup_slub_debug);
1400 * kmem_cache_flags - apply debugging options to the cache
1401 * @object_size: the size of an object without meta data
1402 * @flags: flags to set
1403 * @name: name of the cache
1404 * @ctor: constructor function
1406 * Debug option(s) are applied to @flags. In addition to the debug
1407 * option(s), if a slab name (or multiple) is specified i.e.
1408 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1409 * then only the select slabs will receive the debug option(s).
1411 slab_flags_t kmem_cache_flags(unsigned int object_size,
1412 slab_flags_t flags, const char *name,
1413 void (*ctor)(void *))
1418 slab_flags_t block_flags;
1420 /* If slub_debug = 0, it folds into the if conditional. */
1421 if (!slub_debug_string)
1422 return flags | slub_debug;
1425 next_block = slub_debug_string;
1426 /* Go through all blocks of debug options, see if any matches our slab's name */
1427 while (next_block) {
1428 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1431 /* Found a block that has a slab list, search it */
1436 end = strchrnul(iter, ',');
1437 if (next_block && next_block < end)
1438 end = next_block - 1;
1440 glob = strnchr(iter, end - iter, '*');
1442 cmplen = glob - iter;
1444 cmplen = max_t(size_t, len, (end - iter));
1446 if (!strncmp(name, iter, cmplen)) {
1447 flags |= block_flags;
1451 if (!*end || *end == ';')
1459 #else /* !CONFIG_SLUB_DEBUG */
1460 static inline void setup_object_debug(struct kmem_cache *s,
1461 struct page *page, void *object) {}
1463 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1465 static inline int alloc_debug_processing(struct kmem_cache *s,
1466 struct page *page, void *object, unsigned long addr) { return 0; }
1468 static inline int free_debug_processing(
1469 struct kmem_cache *s, struct page *page,
1470 void *head, void *tail, int bulk_cnt,
1471 unsigned long addr) { return 0; }
1473 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1475 static inline int check_object(struct kmem_cache *s, struct page *page,
1476 void *object, u8 val) { return 1; }
1477 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1478 struct page *page) {}
1479 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1480 struct page *page) {}
1481 slab_flags_t kmem_cache_flags(unsigned int object_size,
1482 slab_flags_t flags, const char *name,
1483 void (*ctor)(void *))
1487 #define slub_debug 0
1489 #define disable_higher_order_debug 0
1491 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1493 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1495 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1497 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1500 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1501 void *freelist, void *nextfree)
1505 #endif /* CONFIG_SLUB_DEBUG */
1508 * Hooks for other subsystems that check memory allocations. In a typical
1509 * production configuration these hooks all should produce no code at all.
1511 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1513 ptr = kasan_kmalloc_large(ptr, size, flags);
1514 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1515 kmemleak_alloc(ptr, size, 1, flags);
1519 static __always_inline void kfree_hook(void *x)
1522 kasan_kfree_large(x, _RET_IP_);
1525 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1527 kmemleak_free_recursive(x, s->flags);
1530 * Trouble is that we may no longer disable interrupts in the fast path
1531 * So in order to make the debug calls that expect irqs to be
1532 * disabled we need to disable interrupts temporarily.
1534 #ifdef CONFIG_LOCKDEP
1536 unsigned long flags;
1538 local_irq_save(flags);
1539 debug_check_no_locks_freed(x, s->object_size);
1540 local_irq_restore(flags);
1543 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1544 debug_check_no_obj_freed(x, s->object_size);
1546 /* Use KCSAN to help debug racy use-after-free. */
1547 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1548 __kcsan_check_access(x, s->object_size,
1549 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1551 /* KASAN might put x into memory quarantine, delaying its reuse */
1552 return kasan_slab_free(s, x, _RET_IP_);
1555 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1556 void **head, void **tail)
1561 void *old_tail = *tail ? *tail : *head;
1564 /* Head and tail of the reconstructed freelist */
1570 next = get_freepointer(s, object);
1572 if (slab_want_init_on_free(s)) {
1574 * Clear the object and the metadata, but don't touch
1577 memset(object, 0, s->object_size);
1578 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1580 memset((char *)object + s->inuse, 0,
1581 s->size - s->inuse - rsize);
1584 /* If object's reuse doesn't have to be delayed */
1585 if (!slab_free_hook(s, object)) {
1586 /* Move object to the new freelist */
1587 set_freepointer(s, object, *head);
1592 } while (object != old_tail);
1597 return *head != NULL;
1600 static void *setup_object(struct kmem_cache *s, struct page *page,
1603 setup_object_debug(s, page, object);
1604 object = kasan_init_slab_obj(s, object);
1605 if (unlikely(s->ctor)) {
1606 kasan_unpoison_object_data(s, object);
1608 kasan_poison_object_data(s, object);
1614 * Slab allocation and freeing
1616 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1617 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1620 unsigned int order = oo_order(oo);
1622 if (node == NUMA_NO_NODE)
1623 page = alloc_pages(flags, order);
1625 page = __alloc_pages_node(node, flags, order);
1627 if (page && charge_slab_page(page, flags, order, s)) {
1628 __free_pages(page, order);
1635 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1636 /* Pre-initialize the random sequence cache */
1637 static int init_cache_random_seq(struct kmem_cache *s)
1639 unsigned int count = oo_objects(s->oo);
1642 /* Bailout if already initialised */
1646 err = cache_random_seq_create(s, count, GFP_KERNEL);
1648 pr_err("SLUB: Unable to initialize free list for %s\n",
1653 /* Transform to an offset on the set of pages */
1654 if (s->random_seq) {
1657 for (i = 0; i < count; i++)
1658 s->random_seq[i] *= s->size;
1663 /* Initialize each random sequence freelist per cache */
1664 static void __init init_freelist_randomization(void)
1666 struct kmem_cache *s;
1668 mutex_lock(&slab_mutex);
1670 list_for_each_entry(s, &slab_caches, list)
1671 init_cache_random_seq(s);
1673 mutex_unlock(&slab_mutex);
1676 /* Get the next entry on the pre-computed freelist randomized */
1677 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1678 unsigned long *pos, void *start,
1679 unsigned long page_limit,
1680 unsigned long freelist_count)
1685 * If the target page allocation failed, the number of objects on the
1686 * page might be smaller than the usual size defined by the cache.
1689 idx = s->random_seq[*pos];
1691 if (*pos >= freelist_count)
1693 } while (unlikely(idx >= page_limit));
1695 return (char *)start + idx;
1698 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1699 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1704 unsigned long idx, pos, page_limit, freelist_count;
1706 if (page->objects < 2 || !s->random_seq)
1709 freelist_count = oo_objects(s->oo);
1710 pos = get_random_int() % freelist_count;
1712 page_limit = page->objects * s->size;
1713 start = fixup_red_left(s, page_address(page));
1715 /* First entry is used as the base of the freelist */
1716 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1718 cur = setup_object(s, page, cur);
1719 page->freelist = cur;
1721 for (idx = 1; idx < page->objects; idx++) {
1722 next = next_freelist_entry(s, page, &pos, start, page_limit,
1724 next = setup_object(s, page, next);
1725 set_freepointer(s, cur, next);
1728 set_freepointer(s, cur, NULL);
1733 static inline int init_cache_random_seq(struct kmem_cache *s)
1737 static inline void init_freelist_randomization(void) { }
1738 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1742 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1744 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1747 struct kmem_cache_order_objects oo = s->oo;
1749 void *start, *p, *next;
1753 flags &= gfp_allowed_mask;
1755 if (gfpflags_allow_blocking(flags))
1758 flags |= s->allocflags;
1761 * Let the initial higher-order allocation fail under memory pressure
1762 * so we fall-back to the minimum order allocation.
1764 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1765 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1766 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1768 page = alloc_slab_page(s, alloc_gfp, node, oo);
1769 if (unlikely(!page)) {
1773 * Allocation may have failed due to fragmentation.
1774 * Try a lower order alloc if possible
1776 page = alloc_slab_page(s, alloc_gfp, node, oo);
1777 if (unlikely(!page))
1779 stat(s, ORDER_FALLBACK);
1782 page->objects = oo_objects(oo);
1784 page->slab_cache = s;
1785 __SetPageSlab(page);
1786 if (page_is_pfmemalloc(page))
1787 SetPageSlabPfmemalloc(page);
1789 kasan_poison_slab(page);
1791 start = page_address(page);
1793 setup_page_debug(s, page, start);
1795 shuffle = shuffle_freelist(s, page);
1798 start = fixup_red_left(s, start);
1799 start = setup_object(s, page, start);
1800 page->freelist = start;
1801 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1803 next = setup_object(s, page, next);
1804 set_freepointer(s, p, next);
1807 set_freepointer(s, p, NULL);
1810 page->inuse = page->objects;
1814 if (gfpflags_allow_blocking(flags))
1815 local_irq_disable();
1819 inc_slabs_node(s, page_to_nid(page), page->objects);
1824 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1826 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1827 flags = kmalloc_fix_flags(flags);
1829 return allocate_slab(s,
1830 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1833 static void __free_slab(struct kmem_cache *s, struct page *page)
1835 int order = compound_order(page);
1836 int pages = 1 << order;
1838 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1841 slab_pad_check(s, page);
1842 for_each_object(p, s, page_address(page),
1844 check_object(s, page, p, SLUB_RED_INACTIVE);
1847 __ClearPageSlabPfmemalloc(page);
1848 __ClearPageSlab(page);
1850 page->mapping = NULL;
1851 if (current->reclaim_state)
1852 current->reclaim_state->reclaimed_slab += pages;
1853 uncharge_slab_page(page, order, s);
1854 __free_pages(page, order);
1857 static void rcu_free_slab(struct rcu_head *h)
1859 struct page *page = container_of(h, struct page, rcu_head);
1861 __free_slab(page->slab_cache, page);
1864 static void free_slab(struct kmem_cache *s, struct page *page)
1866 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1867 call_rcu(&page->rcu_head, rcu_free_slab);
1869 __free_slab(s, page);
1872 static void discard_slab(struct kmem_cache *s, struct page *page)
1874 dec_slabs_node(s, page_to_nid(page), page->objects);
1879 * Management of partially allocated slabs.
1882 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1885 if (tail == DEACTIVATE_TO_TAIL)
1886 list_add_tail(&page->slab_list, &n->partial);
1888 list_add(&page->slab_list, &n->partial);
1891 static inline void add_partial(struct kmem_cache_node *n,
1892 struct page *page, int tail)
1894 lockdep_assert_held(&n->list_lock);
1895 __add_partial(n, page, tail);
1898 static inline void remove_partial(struct kmem_cache_node *n,
1901 lockdep_assert_held(&n->list_lock);
1902 list_del(&page->slab_list);
1907 * Remove slab from the partial list, freeze it and
1908 * return the pointer to the freelist.
1910 * Returns a list of objects or NULL if it fails.
1912 static inline void *acquire_slab(struct kmem_cache *s,
1913 struct kmem_cache_node *n, struct page *page,
1914 int mode, int *objects)
1917 unsigned long counters;
1920 lockdep_assert_held(&n->list_lock);
1923 * Zap the freelist and set the frozen bit.
1924 * The old freelist is the list of objects for the
1925 * per cpu allocation list.
1927 freelist = page->freelist;
1928 counters = page->counters;
1929 new.counters = counters;
1930 *objects = new.objects - new.inuse;
1932 new.inuse = page->objects;
1933 new.freelist = NULL;
1935 new.freelist = freelist;
1938 VM_BUG_ON(new.frozen);
1941 if (!__cmpxchg_double_slab(s, page,
1943 new.freelist, new.counters,
1947 remove_partial(n, page);
1952 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1953 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1956 * Try to allocate a partial slab from a specific node.
1958 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1959 struct kmem_cache_cpu *c, gfp_t flags)
1961 struct page *page, *page2;
1962 void *object = NULL;
1963 unsigned int available = 0;
1967 * Racy check. If we mistakenly see no partial slabs then we
1968 * just allocate an empty slab. If we mistakenly try to get a
1969 * partial slab and there is none available then get_partials()
1972 if (!n || !n->nr_partial)
1975 spin_lock(&n->list_lock);
1976 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1979 if (!pfmemalloc_match(page, flags))
1982 t = acquire_slab(s, n, page, object == NULL, &objects);
1986 available += objects;
1989 stat(s, ALLOC_FROM_PARTIAL);
1992 put_cpu_partial(s, page, 0);
1993 stat(s, CPU_PARTIAL_NODE);
1995 if (!kmem_cache_has_cpu_partial(s)
1996 || available > slub_cpu_partial(s) / 2)
2000 spin_unlock(&n->list_lock);
2005 * Get a page from somewhere. Search in increasing NUMA distances.
2007 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2008 struct kmem_cache_cpu *c)
2011 struct zonelist *zonelist;
2014 enum zone_type highest_zoneidx = gfp_zone(flags);
2016 unsigned int cpuset_mems_cookie;
2019 * The defrag ratio allows a configuration of the tradeoffs between
2020 * inter node defragmentation and node local allocations. A lower
2021 * defrag_ratio increases the tendency to do local allocations
2022 * instead of attempting to obtain partial slabs from other nodes.
2024 * If the defrag_ratio is set to 0 then kmalloc() always
2025 * returns node local objects. If the ratio is higher then kmalloc()
2026 * may return off node objects because partial slabs are obtained
2027 * from other nodes and filled up.
2029 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2030 * (which makes defrag_ratio = 1000) then every (well almost)
2031 * allocation will first attempt to defrag slab caches on other nodes.
2032 * This means scanning over all nodes to look for partial slabs which
2033 * may be expensive if we do it every time we are trying to find a slab
2034 * with available objects.
2036 if (!s->remote_node_defrag_ratio ||
2037 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2041 cpuset_mems_cookie = read_mems_allowed_begin();
2042 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2043 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2044 struct kmem_cache_node *n;
2046 n = get_node(s, zone_to_nid(zone));
2048 if (n && cpuset_zone_allowed(zone, flags) &&
2049 n->nr_partial > s->min_partial) {
2050 object = get_partial_node(s, n, c, flags);
2053 * Don't check read_mems_allowed_retry()
2054 * here - if mems_allowed was updated in
2055 * parallel, that was a harmless race
2056 * between allocation and the cpuset
2063 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2064 #endif /* CONFIG_NUMA */
2069 * Get a partial page, lock it and return it.
2071 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2072 struct kmem_cache_cpu *c)
2075 int searchnode = node;
2077 if (node == NUMA_NO_NODE)
2078 searchnode = numa_mem_id();
2080 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2081 if (object || node != NUMA_NO_NODE)
2084 return get_any_partial(s, flags, c);
2087 #ifdef CONFIG_PREEMPTION
2089 * Calculate the next globally unique transaction for disambiguation
2090 * during cmpxchg. The transactions start with the cpu number and are then
2091 * incremented by CONFIG_NR_CPUS.
2093 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2096 * No preemption supported therefore also no need to check for
2102 static inline unsigned long next_tid(unsigned long tid)
2104 return tid + TID_STEP;
2107 #ifdef SLUB_DEBUG_CMPXCHG
2108 static inline unsigned int tid_to_cpu(unsigned long tid)
2110 return tid % TID_STEP;
2113 static inline unsigned long tid_to_event(unsigned long tid)
2115 return tid / TID_STEP;
2119 static inline unsigned int init_tid(int cpu)
2124 static inline void note_cmpxchg_failure(const char *n,
2125 const struct kmem_cache *s, unsigned long tid)
2127 #ifdef SLUB_DEBUG_CMPXCHG
2128 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2130 pr_info("%s %s: cmpxchg redo ", n, s->name);
2132 #ifdef CONFIG_PREEMPTION
2133 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2134 pr_warn("due to cpu change %d -> %d\n",
2135 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2138 if (tid_to_event(tid) != tid_to_event(actual_tid))
2139 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2140 tid_to_event(tid), tid_to_event(actual_tid));
2142 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2143 actual_tid, tid, next_tid(tid));
2145 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2148 static void init_kmem_cache_cpus(struct kmem_cache *s)
2152 for_each_possible_cpu(cpu)
2153 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2157 * Remove the cpu slab
2159 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2160 void *freelist, struct kmem_cache_cpu *c)
2162 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2163 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2165 enum slab_modes l = M_NONE, m = M_NONE;
2167 int tail = DEACTIVATE_TO_HEAD;
2171 if (page->freelist) {
2172 stat(s, DEACTIVATE_REMOTE_FREES);
2173 tail = DEACTIVATE_TO_TAIL;
2177 * Stage one: Free all available per cpu objects back
2178 * to the page freelist while it is still frozen. Leave the
2181 * There is no need to take the list->lock because the page
2184 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2186 unsigned long counters;
2189 * If 'nextfree' is invalid, it is possible that the object at
2190 * 'freelist' is already corrupted. So isolate all objects
2191 * starting at 'freelist'.
2193 if (freelist_corrupted(s, page, freelist, nextfree))
2197 prior = page->freelist;
2198 counters = page->counters;
2199 set_freepointer(s, freelist, prior);
2200 new.counters = counters;
2202 VM_BUG_ON(!new.frozen);
2204 } while (!__cmpxchg_double_slab(s, page,
2206 freelist, new.counters,
2207 "drain percpu freelist"));
2209 freelist = nextfree;
2213 * Stage two: Ensure that the page is unfrozen while the
2214 * list presence reflects the actual number of objects
2217 * We setup the list membership and then perform a cmpxchg
2218 * with the count. If there is a mismatch then the page
2219 * is not unfrozen but the page is on the wrong list.
2221 * Then we restart the process which may have to remove
2222 * the page from the list that we just put it on again
2223 * because the number of objects in the slab may have
2228 old.freelist = page->freelist;
2229 old.counters = page->counters;
2230 VM_BUG_ON(!old.frozen);
2232 /* Determine target state of the slab */
2233 new.counters = old.counters;
2236 set_freepointer(s, freelist, old.freelist);
2237 new.freelist = freelist;
2239 new.freelist = old.freelist;
2243 if (!new.inuse && n->nr_partial >= s->min_partial)
2245 else if (new.freelist) {
2250 * Taking the spinlock removes the possibility
2251 * that acquire_slab() will see a slab page that
2254 spin_lock(&n->list_lock);
2258 if (kmem_cache_debug(s) && !lock) {
2261 * This also ensures that the scanning of full
2262 * slabs from diagnostic functions will not see
2265 spin_lock(&n->list_lock);
2271 remove_partial(n, page);
2272 else if (l == M_FULL)
2273 remove_full(s, n, page);
2276 add_partial(n, page, tail);
2277 else if (m == M_FULL)
2278 add_full(s, n, page);
2282 if (!__cmpxchg_double_slab(s, page,
2283 old.freelist, old.counters,
2284 new.freelist, new.counters,
2289 spin_unlock(&n->list_lock);
2293 else if (m == M_FULL)
2294 stat(s, DEACTIVATE_FULL);
2295 else if (m == M_FREE) {
2296 stat(s, DEACTIVATE_EMPTY);
2297 discard_slab(s, page);
2306 * Unfreeze all the cpu partial slabs.
2308 * This function must be called with interrupts disabled
2309 * for the cpu using c (or some other guarantee must be there
2310 * to guarantee no concurrent accesses).
2312 static void unfreeze_partials(struct kmem_cache *s,
2313 struct kmem_cache_cpu *c)
2315 #ifdef CONFIG_SLUB_CPU_PARTIAL
2316 struct kmem_cache_node *n = NULL, *n2 = NULL;
2317 struct page *page, *discard_page = NULL;
2319 while ((page = slub_percpu_partial(c))) {
2323 slub_set_percpu_partial(c, page);
2325 n2 = get_node(s, page_to_nid(page));
2328 spin_unlock(&n->list_lock);
2331 spin_lock(&n->list_lock);
2336 old.freelist = page->freelist;
2337 old.counters = page->counters;
2338 VM_BUG_ON(!old.frozen);
2340 new.counters = old.counters;
2341 new.freelist = old.freelist;
2345 } while (!__cmpxchg_double_slab(s, page,
2346 old.freelist, old.counters,
2347 new.freelist, new.counters,
2348 "unfreezing slab"));
2350 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2351 page->next = discard_page;
2352 discard_page = page;
2354 add_partial(n, page, DEACTIVATE_TO_TAIL);
2355 stat(s, FREE_ADD_PARTIAL);
2360 spin_unlock(&n->list_lock);
2362 while (discard_page) {
2363 page = discard_page;
2364 discard_page = discard_page->next;
2366 stat(s, DEACTIVATE_EMPTY);
2367 discard_slab(s, page);
2370 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2374 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2375 * partial page slot if available.
2377 * If we did not find a slot then simply move all the partials to the
2378 * per node partial list.
2380 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2382 #ifdef CONFIG_SLUB_CPU_PARTIAL
2383 struct page *oldpage;
2391 oldpage = this_cpu_read(s->cpu_slab->partial);
2394 pobjects = oldpage->pobjects;
2395 pages = oldpage->pages;
2396 if (drain && pobjects > slub_cpu_partial(s)) {
2397 unsigned long flags;
2399 * partial array is full. Move the existing
2400 * set to the per node partial list.
2402 local_irq_save(flags);
2403 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2404 local_irq_restore(flags);
2408 stat(s, CPU_PARTIAL_DRAIN);
2413 pobjects += page->objects - page->inuse;
2415 page->pages = pages;
2416 page->pobjects = pobjects;
2417 page->next = oldpage;
2419 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2421 if (unlikely(!slub_cpu_partial(s))) {
2422 unsigned long flags;
2424 local_irq_save(flags);
2425 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2426 local_irq_restore(flags);
2429 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2432 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2434 stat(s, CPUSLAB_FLUSH);
2435 deactivate_slab(s, c->page, c->freelist, c);
2437 c->tid = next_tid(c->tid);
2443 * Called from IPI handler with interrupts disabled.
2445 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2447 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2452 unfreeze_partials(s, c);
2455 static void flush_cpu_slab(void *d)
2457 struct kmem_cache *s = d;
2459 __flush_cpu_slab(s, smp_processor_id());
2462 static bool has_cpu_slab(int cpu, void *info)
2464 struct kmem_cache *s = info;
2465 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2467 return c->page || slub_percpu_partial(c);
2470 static void flush_all(struct kmem_cache *s)
2472 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2476 * Use the cpu notifier to insure that the cpu slabs are flushed when
2479 static int slub_cpu_dead(unsigned int cpu)
2481 struct kmem_cache *s;
2482 unsigned long flags;
2484 mutex_lock(&slab_mutex);
2485 list_for_each_entry(s, &slab_caches, list) {
2486 local_irq_save(flags);
2487 __flush_cpu_slab(s, cpu);
2488 local_irq_restore(flags);
2490 mutex_unlock(&slab_mutex);
2495 * Check if the objects in a per cpu structure fit numa
2496 * locality expectations.
2498 static inline int node_match(struct page *page, int node)
2501 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2507 #ifdef CONFIG_SLUB_DEBUG
2508 static int count_free(struct page *page)
2510 return page->objects - page->inuse;
2513 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2515 return atomic_long_read(&n->total_objects);
2517 #endif /* CONFIG_SLUB_DEBUG */
2519 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2520 static unsigned long count_partial(struct kmem_cache_node *n,
2521 int (*get_count)(struct page *))
2523 unsigned long flags;
2524 unsigned long x = 0;
2527 spin_lock_irqsave(&n->list_lock, flags);
2528 list_for_each_entry(page, &n->partial, slab_list)
2529 x += get_count(page);
2530 spin_unlock_irqrestore(&n->list_lock, flags);
2533 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2535 static noinline void
2536 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2538 #ifdef CONFIG_SLUB_DEBUG
2539 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2540 DEFAULT_RATELIMIT_BURST);
2542 struct kmem_cache_node *n;
2544 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2547 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2548 nid, gfpflags, &gfpflags);
2549 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2550 s->name, s->object_size, s->size, oo_order(s->oo),
2553 if (oo_order(s->min) > get_order(s->object_size))
2554 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2557 for_each_kmem_cache_node(s, node, n) {
2558 unsigned long nr_slabs;
2559 unsigned long nr_objs;
2560 unsigned long nr_free;
2562 nr_free = count_partial(n, count_free);
2563 nr_slabs = node_nr_slabs(n);
2564 nr_objs = node_nr_objs(n);
2566 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2567 node, nr_slabs, nr_objs, nr_free);
2572 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2573 int node, struct kmem_cache_cpu **pc)
2576 struct kmem_cache_cpu *c = *pc;
2579 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2581 freelist = get_partial(s, flags, node, c);
2586 page = new_slab(s, flags, node);
2588 c = raw_cpu_ptr(s->cpu_slab);
2593 * No other reference to the page yet so we can
2594 * muck around with it freely without cmpxchg
2596 freelist = page->freelist;
2597 page->freelist = NULL;
2599 stat(s, ALLOC_SLAB);
2607 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2609 if (unlikely(PageSlabPfmemalloc(page)))
2610 return gfp_pfmemalloc_allowed(gfpflags);
2616 * Check the page->freelist of a page and either transfer the freelist to the
2617 * per cpu freelist or deactivate the page.
2619 * The page is still frozen if the return value is not NULL.
2621 * If this function returns NULL then the page has been unfrozen.
2623 * This function must be called with interrupt disabled.
2625 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2628 unsigned long counters;
2632 freelist = page->freelist;
2633 counters = page->counters;
2635 new.counters = counters;
2636 VM_BUG_ON(!new.frozen);
2638 new.inuse = page->objects;
2639 new.frozen = freelist != NULL;
2641 } while (!__cmpxchg_double_slab(s, page,
2650 * Slow path. The lockless freelist is empty or we need to perform
2653 * Processing is still very fast if new objects have been freed to the
2654 * regular freelist. In that case we simply take over the regular freelist
2655 * as the lockless freelist and zap the regular freelist.
2657 * If that is not working then we fall back to the partial lists. We take the
2658 * first element of the freelist as the object to allocate now and move the
2659 * rest of the freelist to the lockless freelist.
2661 * And if we were unable to get a new slab from the partial slab lists then
2662 * we need to allocate a new slab. This is the slowest path since it involves
2663 * a call to the page allocator and the setup of a new slab.
2665 * Version of __slab_alloc to use when we know that interrupts are
2666 * already disabled (which is the case for bulk allocation).
2668 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2669 unsigned long addr, struct kmem_cache_cpu *c)
2677 * if the node is not online or has no normal memory, just
2678 * ignore the node constraint
2680 if (unlikely(node != NUMA_NO_NODE &&
2681 !node_state(node, N_NORMAL_MEMORY)))
2682 node = NUMA_NO_NODE;
2687 if (unlikely(!node_match(page, node))) {
2689 * same as above but node_match() being false already
2690 * implies node != NUMA_NO_NODE
2692 if (!node_state(node, N_NORMAL_MEMORY)) {
2693 node = NUMA_NO_NODE;
2696 stat(s, ALLOC_NODE_MISMATCH);
2697 deactivate_slab(s, page, c->freelist, c);
2703 * By rights, we should be searching for a slab page that was
2704 * PFMEMALLOC but right now, we are losing the pfmemalloc
2705 * information when the page leaves the per-cpu allocator
2707 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2708 deactivate_slab(s, page, c->freelist, c);
2712 /* must check again c->freelist in case of cpu migration or IRQ */
2713 freelist = c->freelist;
2717 freelist = get_freelist(s, page);
2721 stat(s, DEACTIVATE_BYPASS);
2725 stat(s, ALLOC_REFILL);
2729 * freelist is pointing to the list of objects to be used.
2730 * page is pointing to the page from which the objects are obtained.
2731 * That page must be frozen for per cpu allocations to work.
2733 VM_BUG_ON(!c->page->frozen);
2734 c->freelist = get_freepointer(s, freelist);
2735 c->tid = next_tid(c->tid);
2740 if (slub_percpu_partial(c)) {
2741 page = c->page = slub_percpu_partial(c);
2742 slub_set_percpu_partial(c, page);
2743 stat(s, CPU_PARTIAL_ALLOC);
2747 freelist = new_slab_objects(s, gfpflags, node, &c);
2749 if (unlikely(!freelist)) {
2750 slab_out_of_memory(s, gfpflags, node);
2755 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2758 /* Only entered in the debug case */
2759 if (kmem_cache_debug(s) &&
2760 !alloc_debug_processing(s, page, freelist, addr))
2761 goto new_slab; /* Slab failed checks. Next slab needed */
2763 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2768 * Another one that disabled interrupt and compensates for possible
2769 * cpu changes by refetching the per cpu area pointer.
2771 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2772 unsigned long addr, struct kmem_cache_cpu *c)
2775 unsigned long flags;
2777 local_irq_save(flags);
2778 #ifdef CONFIG_PREEMPTION
2780 * We may have been preempted and rescheduled on a different
2781 * cpu before disabling interrupts. Need to reload cpu area
2784 c = this_cpu_ptr(s->cpu_slab);
2787 p = ___slab_alloc(s, gfpflags, node, addr, c);
2788 local_irq_restore(flags);
2793 * If the object has been wiped upon free, make sure it's fully initialized by
2794 * zeroing out freelist pointer.
2796 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2799 if (unlikely(slab_want_init_on_free(s)) && obj)
2800 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2804 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2805 * have the fastpath folded into their functions. So no function call
2806 * overhead for requests that can be satisfied on the fastpath.
2808 * The fastpath works by first checking if the lockless freelist can be used.
2809 * If not then __slab_alloc is called for slow processing.
2811 * Otherwise we can simply pick the next object from the lockless free list.
2813 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2814 gfp_t gfpflags, int node, unsigned long addr)
2817 struct kmem_cache_cpu *c;
2821 s = slab_pre_alloc_hook(s, gfpflags);
2826 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2827 * enabled. We may switch back and forth between cpus while
2828 * reading from one cpu area. That does not matter as long
2829 * as we end up on the original cpu again when doing the cmpxchg.
2831 * We should guarantee that tid and kmem_cache are retrieved on
2832 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2833 * to check if it is matched or not.
2836 tid = this_cpu_read(s->cpu_slab->tid);
2837 c = raw_cpu_ptr(s->cpu_slab);
2838 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2839 unlikely(tid != READ_ONCE(c->tid)));
2842 * Irqless object alloc/free algorithm used here depends on sequence
2843 * of fetching cpu_slab's data. tid should be fetched before anything
2844 * on c to guarantee that object and page associated with previous tid
2845 * won't be used with current tid. If we fetch tid first, object and
2846 * page could be one associated with next tid and our alloc/free
2847 * request will be failed. In this case, we will retry. So, no problem.
2852 * The transaction ids are globally unique per cpu and per operation on
2853 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2854 * occurs on the right processor and that there was no operation on the
2855 * linked list in between.
2858 object = c->freelist;
2860 if (unlikely(!object || !node_match(page, node))) {
2861 object = __slab_alloc(s, gfpflags, node, addr, c);
2862 stat(s, ALLOC_SLOWPATH);
2864 void *next_object = get_freepointer_safe(s, object);
2867 * The cmpxchg will only match if there was no additional
2868 * operation and if we are on the right processor.
2870 * The cmpxchg does the following atomically (without lock
2872 * 1. Relocate first pointer to the current per cpu area.
2873 * 2. Verify that tid and freelist have not been changed
2874 * 3. If they were not changed replace tid and freelist
2876 * Since this is without lock semantics the protection is only
2877 * against code executing on this cpu *not* from access by
2880 if (unlikely(!this_cpu_cmpxchg_double(
2881 s->cpu_slab->freelist, s->cpu_slab->tid,
2883 next_object, next_tid(tid)))) {
2885 note_cmpxchg_failure("slab_alloc", s, tid);
2888 prefetch_freepointer(s, next_object);
2889 stat(s, ALLOC_FASTPATH);
2892 maybe_wipe_obj_freeptr(s, object);
2894 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2895 memset(object, 0, s->object_size);
2897 slab_post_alloc_hook(s, gfpflags, 1, &object);
2902 static __always_inline void *slab_alloc(struct kmem_cache *s,
2903 gfp_t gfpflags, unsigned long addr)
2905 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2908 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2910 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2912 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2917 EXPORT_SYMBOL(kmem_cache_alloc);
2919 #ifdef CONFIG_TRACING
2920 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2922 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2923 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2924 ret = kasan_kmalloc(s, ret, size, gfpflags);
2927 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2931 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2933 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2935 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2936 s->object_size, s->size, gfpflags, node);
2940 EXPORT_SYMBOL(kmem_cache_alloc_node);
2942 #ifdef CONFIG_TRACING
2943 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2945 int node, size_t size)
2947 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2949 trace_kmalloc_node(_RET_IP_, ret,
2950 size, s->size, gfpflags, node);
2952 ret = kasan_kmalloc(s, ret, size, gfpflags);
2955 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2957 #endif /* CONFIG_NUMA */
2960 * Slow path handling. This may still be called frequently since objects
2961 * have a longer lifetime than the cpu slabs in most processing loads.
2963 * So we still attempt to reduce cache line usage. Just take the slab
2964 * lock and free the item. If there is no additional partial page
2965 * handling required then we can return immediately.
2967 static void __slab_free(struct kmem_cache *s, struct page *page,
2968 void *head, void *tail, int cnt,
2975 unsigned long counters;
2976 struct kmem_cache_node *n = NULL;
2977 unsigned long flags;
2979 stat(s, FREE_SLOWPATH);
2981 if (kmem_cache_debug(s) &&
2982 !free_debug_processing(s, page, head, tail, cnt, addr))
2987 spin_unlock_irqrestore(&n->list_lock, flags);
2990 prior = page->freelist;
2991 counters = page->counters;
2992 set_freepointer(s, tail, prior);
2993 new.counters = counters;
2994 was_frozen = new.frozen;
2996 if ((!new.inuse || !prior) && !was_frozen) {
2998 if (kmem_cache_has_cpu_partial(s) && !prior) {
3001 * Slab was on no list before and will be
3003 * We can defer the list move and instead
3008 } else { /* Needs to be taken off a list */
3010 n = get_node(s, page_to_nid(page));
3012 * Speculatively acquire the list_lock.
3013 * If the cmpxchg does not succeed then we may
3014 * drop the list_lock without any processing.
3016 * Otherwise the list_lock will synchronize with
3017 * other processors updating the list of slabs.
3019 spin_lock_irqsave(&n->list_lock, flags);
3024 } while (!cmpxchg_double_slab(s, page,
3032 * If we just froze the page then put it onto the
3033 * per cpu partial list.
3035 if (new.frozen && !was_frozen) {
3036 put_cpu_partial(s, page, 1);
3037 stat(s, CPU_PARTIAL_FREE);
3040 * The list lock was not taken therefore no list
3041 * activity can be necessary.
3044 stat(s, FREE_FROZEN);
3048 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3052 * Objects left in the slab. If it was not on the partial list before
3055 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3056 remove_full(s, n, page);
3057 add_partial(n, page, DEACTIVATE_TO_TAIL);
3058 stat(s, FREE_ADD_PARTIAL);
3060 spin_unlock_irqrestore(&n->list_lock, flags);
3066 * Slab on the partial list.
3068 remove_partial(n, page);
3069 stat(s, FREE_REMOVE_PARTIAL);
3071 /* Slab must be on the full list */
3072 remove_full(s, n, page);
3075 spin_unlock_irqrestore(&n->list_lock, flags);
3077 discard_slab(s, page);
3081 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3082 * can perform fastpath freeing without additional function calls.
3084 * The fastpath is only possible if we are freeing to the current cpu slab
3085 * of this processor. This typically the case if we have just allocated
3088 * If fastpath is not possible then fall back to __slab_free where we deal
3089 * with all sorts of special processing.
3091 * Bulk free of a freelist with several objects (all pointing to the
3092 * same page) possible by specifying head and tail ptr, plus objects
3093 * count (cnt). Bulk free indicated by tail pointer being set.
3095 static __always_inline void do_slab_free(struct kmem_cache *s,
3096 struct page *page, void *head, void *tail,
3097 int cnt, unsigned long addr)
3099 void *tail_obj = tail ? : head;
3100 struct kmem_cache_cpu *c;
3104 * Determine the currently cpus per cpu slab.
3105 * The cpu may change afterward. However that does not matter since
3106 * data is retrieved via this pointer. If we are on the same cpu
3107 * during the cmpxchg then the free will succeed.
3110 tid = this_cpu_read(s->cpu_slab->tid);
3111 c = raw_cpu_ptr(s->cpu_slab);
3112 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3113 unlikely(tid != READ_ONCE(c->tid)));
3115 /* Same with comment on barrier() in slab_alloc_node() */
3118 if (likely(page == c->page)) {
3119 void **freelist = READ_ONCE(c->freelist);
3121 set_freepointer(s, tail_obj, freelist);
3123 if (unlikely(!this_cpu_cmpxchg_double(
3124 s->cpu_slab->freelist, s->cpu_slab->tid,
3126 head, next_tid(tid)))) {
3128 note_cmpxchg_failure("slab_free", s, tid);
3131 stat(s, FREE_FASTPATH);
3133 __slab_free(s, page, head, tail_obj, cnt, addr);
3137 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3138 void *head, void *tail, int cnt,
3142 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3143 * to remove objects, whose reuse must be delayed.
3145 if (slab_free_freelist_hook(s, &head, &tail))
3146 do_slab_free(s, page, head, tail, cnt, addr);
3149 #ifdef CONFIG_KASAN_GENERIC
3150 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3152 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3156 void kmem_cache_free(struct kmem_cache *s, void *x)
3158 s = cache_from_obj(s, x);
3161 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3162 trace_kmem_cache_free(_RET_IP_, x);
3164 EXPORT_SYMBOL(kmem_cache_free);
3166 struct detached_freelist {
3171 struct kmem_cache *s;
3175 * This function progressively scans the array with free objects (with
3176 * a limited look ahead) and extract objects belonging to the same
3177 * page. It builds a detached freelist directly within the given
3178 * page/objects. This can happen without any need for
3179 * synchronization, because the objects are owned by running process.
3180 * The freelist is build up as a single linked list in the objects.
3181 * The idea is, that this detached freelist can then be bulk
3182 * transferred to the real freelist(s), but only requiring a single
3183 * synchronization primitive. Look ahead in the array is limited due
3184 * to performance reasons.
3187 int build_detached_freelist(struct kmem_cache *s, size_t size,
3188 void **p, struct detached_freelist *df)
3190 size_t first_skipped_index = 0;
3195 /* Always re-init detached_freelist */
3200 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3201 } while (!object && size);
3206 page = virt_to_head_page(object);
3208 /* Handle kalloc'ed objects */
3209 if (unlikely(!PageSlab(page))) {
3210 BUG_ON(!PageCompound(page));
3212 __free_pages(page, compound_order(page));
3213 p[size] = NULL; /* mark object processed */
3216 /* Derive kmem_cache from object */
3217 df->s = page->slab_cache;
3219 df->s = cache_from_obj(s, object); /* Support for memcg */
3222 /* Start new detached freelist */
3224 set_freepointer(df->s, object, NULL);
3226 df->freelist = object;
3227 p[size] = NULL; /* mark object processed */
3233 continue; /* Skip processed objects */
3235 /* df->page is always set at this point */
3236 if (df->page == virt_to_head_page(object)) {
3237 /* Opportunity build freelist */
3238 set_freepointer(df->s, object, df->freelist);
3239 df->freelist = object;
3241 p[size] = NULL; /* mark object processed */
3246 /* Limit look ahead search */
3250 if (!first_skipped_index)
3251 first_skipped_index = size + 1;
3254 return first_skipped_index;
3257 /* Note that interrupts must be enabled when calling this function. */
3258 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3264 struct detached_freelist df;
3266 size = build_detached_freelist(s, size, p, &df);
3270 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3271 } while (likely(size));
3273 EXPORT_SYMBOL(kmem_cache_free_bulk);
3275 /* Note that interrupts must be enabled when calling this function. */
3276 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3279 struct kmem_cache_cpu *c;
3282 /* memcg and kmem_cache debug support */
3283 s = slab_pre_alloc_hook(s, flags);
3287 * Drain objects in the per cpu slab, while disabling local
3288 * IRQs, which protects against PREEMPT and interrupts
3289 * handlers invoking normal fastpath.
3291 local_irq_disable();
3292 c = this_cpu_ptr(s->cpu_slab);
3294 for (i = 0; i < size; i++) {
3295 void *object = c->freelist;
3297 if (unlikely(!object)) {
3299 * We may have removed an object from c->freelist using
3300 * the fastpath in the previous iteration; in that case,
3301 * c->tid has not been bumped yet.
3302 * Since ___slab_alloc() may reenable interrupts while
3303 * allocating memory, we should bump c->tid now.
3305 c->tid = next_tid(c->tid);
3308 * Invoking slow path likely have side-effect
3309 * of re-populating per CPU c->freelist
3311 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3313 if (unlikely(!p[i]))
3316 c = this_cpu_ptr(s->cpu_slab);
3317 maybe_wipe_obj_freeptr(s, p[i]);
3319 continue; /* goto for-loop */
3321 c->freelist = get_freepointer(s, object);
3323 maybe_wipe_obj_freeptr(s, p[i]);
3325 c->tid = next_tid(c->tid);
3328 /* Clear memory outside IRQ disabled fastpath loop */
3329 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3332 for (j = 0; j < i; j++)
3333 memset(p[j], 0, s->object_size);
3336 /* memcg and kmem_cache debug support */
3337 slab_post_alloc_hook(s, flags, size, p);
3341 slab_post_alloc_hook(s, flags, i, p);
3342 __kmem_cache_free_bulk(s, i, p);
3345 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3349 * Object placement in a slab is made very easy because we always start at
3350 * offset 0. If we tune the size of the object to the alignment then we can
3351 * get the required alignment by putting one properly sized object after
3354 * Notice that the allocation order determines the sizes of the per cpu
3355 * caches. Each processor has always one slab available for allocations.
3356 * Increasing the allocation order reduces the number of times that slabs
3357 * must be moved on and off the partial lists and is therefore a factor in
3362 * Mininum / Maximum order of slab pages. This influences locking overhead
3363 * and slab fragmentation. A higher order reduces the number of partial slabs
3364 * and increases the number of allocations possible without having to
3365 * take the list_lock.
3367 static unsigned int slub_min_order;
3368 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3369 static unsigned int slub_min_objects;
3372 * Calculate the order of allocation given an slab object size.
3374 * The order of allocation has significant impact on performance and other
3375 * system components. Generally order 0 allocations should be preferred since
3376 * order 0 does not cause fragmentation in the page allocator. Larger objects
3377 * be problematic to put into order 0 slabs because there may be too much
3378 * unused space left. We go to a higher order if more than 1/16th of the slab
3381 * In order to reach satisfactory performance we must ensure that a minimum
3382 * number of objects is in one slab. Otherwise we may generate too much
3383 * activity on the partial lists which requires taking the list_lock. This is
3384 * less a concern for large slabs though which are rarely used.
3386 * slub_max_order specifies the order where we begin to stop considering the
3387 * number of objects in a slab as critical. If we reach slub_max_order then
3388 * we try to keep the page order as low as possible. So we accept more waste
3389 * of space in favor of a small page order.
3391 * Higher order allocations also allow the placement of more objects in a
3392 * slab and thereby reduce object handling overhead. If the user has
3393 * requested a higher mininum order then we start with that one instead of
3394 * the smallest order which will fit the object.
3396 static inline unsigned int slab_order(unsigned int size,
3397 unsigned int min_objects, unsigned int max_order,
3398 unsigned int fract_leftover)
3400 unsigned int min_order = slub_min_order;
3403 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3404 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3406 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3407 order <= max_order; order++) {
3409 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3412 rem = slab_size % size;
3414 if (rem <= slab_size / fract_leftover)
3421 static inline int calculate_order(unsigned int size)
3424 unsigned int min_objects;
3425 unsigned int max_objects;
3428 * Attempt to find best configuration for a slab. This
3429 * works by first attempting to generate a layout with
3430 * the best configuration and backing off gradually.
3432 * First we increase the acceptable waste in a slab. Then
3433 * we reduce the minimum objects required in a slab.
3435 min_objects = slub_min_objects;
3437 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3438 max_objects = order_objects(slub_max_order, size);
3439 min_objects = min(min_objects, max_objects);
3441 while (min_objects > 1) {
3442 unsigned int fraction;
3445 while (fraction >= 4) {
3446 order = slab_order(size, min_objects,
3447 slub_max_order, fraction);
3448 if (order <= slub_max_order)
3456 * We were unable to place multiple objects in a slab. Now
3457 * lets see if we can place a single object there.
3459 order = slab_order(size, 1, slub_max_order, 1);
3460 if (order <= slub_max_order)
3464 * Doh this slab cannot be placed using slub_max_order.
3466 order = slab_order(size, 1, MAX_ORDER, 1);
3467 if (order < MAX_ORDER)
3473 init_kmem_cache_node(struct kmem_cache_node *n)
3476 spin_lock_init(&n->list_lock);
3477 INIT_LIST_HEAD(&n->partial);
3478 #ifdef CONFIG_SLUB_DEBUG
3479 atomic_long_set(&n->nr_slabs, 0);
3480 atomic_long_set(&n->total_objects, 0);
3481 INIT_LIST_HEAD(&n->full);
3485 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3487 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3488 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3491 * Must align to double word boundary for the double cmpxchg
3492 * instructions to work; see __pcpu_double_call_return_bool().
3494 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3495 2 * sizeof(void *));
3500 init_kmem_cache_cpus(s);
3505 static struct kmem_cache *kmem_cache_node;
3508 * No kmalloc_node yet so do it by hand. We know that this is the first
3509 * slab on the node for this slabcache. There are no concurrent accesses
3512 * Note that this function only works on the kmem_cache_node
3513 * when allocating for the kmem_cache_node. This is used for bootstrapping
3514 * memory on a fresh node that has no slab structures yet.
3516 static void early_kmem_cache_node_alloc(int node)
3519 struct kmem_cache_node *n;
3521 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3523 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3526 if (page_to_nid(page) != node) {
3527 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3528 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3533 #ifdef CONFIG_SLUB_DEBUG
3534 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3535 init_tracking(kmem_cache_node, n);
3537 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3539 page->freelist = get_freepointer(kmem_cache_node, n);
3542 kmem_cache_node->node[node] = n;
3543 init_kmem_cache_node(n);
3544 inc_slabs_node(kmem_cache_node, node, page->objects);
3547 * No locks need to be taken here as it has just been
3548 * initialized and there is no concurrent access.
3550 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3553 static void free_kmem_cache_nodes(struct kmem_cache *s)
3556 struct kmem_cache_node *n;
3558 for_each_kmem_cache_node(s, node, n) {
3559 s->node[node] = NULL;
3560 kmem_cache_free(kmem_cache_node, n);
3564 void __kmem_cache_release(struct kmem_cache *s)
3566 cache_random_seq_destroy(s);
3567 free_percpu(s->cpu_slab);
3568 free_kmem_cache_nodes(s);
3571 static int init_kmem_cache_nodes(struct kmem_cache *s)
3575 for_each_node_state(node, N_NORMAL_MEMORY) {
3576 struct kmem_cache_node *n;
3578 if (slab_state == DOWN) {
3579 early_kmem_cache_node_alloc(node);
3582 n = kmem_cache_alloc_node(kmem_cache_node,
3586 free_kmem_cache_nodes(s);
3590 init_kmem_cache_node(n);
3596 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3598 if (min < MIN_PARTIAL)
3600 else if (min > MAX_PARTIAL)
3602 s->min_partial = min;
3605 static void set_cpu_partial(struct kmem_cache *s)
3607 #ifdef CONFIG_SLUB_CPU_PARTIAL
3609 * cpu_partial determined the maximum number of objects kept in the
3610 * per cpu partial lists of a processor.
3612 * Per cpu partial lists mainly contain slabs that just have one
3613 * object freed. If they are used for allocation then they can be
3614 * filled up again with minimal effort. The slab will never hit the
3615 * per node partial lists and therefore no locking will be required.
3617 * This setting also determines
3619 * A) The number of objects from per cpu partial slabs dumped to the
3620 * per node list when we reach the limit.
3621 * B) The number of objects in cpu partial slabs to extract from the
3622 * per node list when we run out of per cpu objects. We only fetch
3623 * 50% to keep some capacity around for frees.
3625 if (!kmem_cache_has_cpu_partial(s))
3626 slub_set_cpu_partial(s, 0);
3627 else if (s->size >= PAGE_SIZE)
3628 slub_set_cpu_partial(s, 2);
3629 else if (s->size >= 1024)
3630 slub_set_cpu_partial(s, 6);
3631 else if (s->size >= 256)
3632 slub_set_cpu_partial(s, 13);
3634 slub_set_cpu_partial(s, 30);
3639 * calculate_sizes() determines the order and the distribution of data within
3642 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3644 slab_flags_t flags = s->flags;
3645 unsigned int size = s->object_size;
3646 unsigned int freepointer_area;
3650 * Round up object size to the next word boundary. We can only
3651 * place the free pointer at word boundaries and this determines
3652 * the possible location of the free pointer.
3654 size = ALIGN(size, sizeof(void *));
3656 * This is the area of the object where a freepointer can be
3657 * safely written. If redzoning adds more to the inuse size, we
3658 * can't use that portion for writing the freepointer, so
3659 * s->offset must be limited within this for the general case.
3661 freepointer_area = size;
3663 #ifdef CONFIG_SLUB_DEBUG
3665 * Determine if we can poison the object itself. If the user of
3666 * the slab may touch the object after free or before allocation
3667 * then we should never poison the object itself.
3669 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3671 s->flags |= __OBJECT_POISON;
3673 s->flags &= ~__OBJECT_POISON;
3677 * If we are Redzoning then check if there is some space between the
3678 * end of the object and the free pointer. If not then add an
3679 * additional word to have some bytes to store Redzone information.
3681 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3682 size += sizeof(void *);
3686 * With that we have determined the number of bytes in actual use
3687 * by the object. This is the potential offset to the free pointer.
3691 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3694 * Relocate free pointer after the object if it is not
3695 * permitted to overwrite the first word of the object on
3698 * This is the case if we do RCU, have a constructor or
3699 * destructor or are poisoning the objects.
3701 * The assumption that s->offset >= s->inuse means free
3702 * pointer is outside of the object is used in the
3703 * freeptr_outside_object() function. If that is no
3704 * longer true, the function needs to be modified.
3707 size += sizeof(void *);
3708 } else if (freepointer_area > sizeof(void *)) {
3710 * Store freelist pointer near middle of object to keep
3711 * it away from the edges of the object to avoid small
3712 * sized over/underflows from neighboring allocations.
3714 s->offset = ALIGN(freepointer_area / 2, sizeof(void *));
3717 #ifdef CONFIG_SLUB_DEBUG
3718 if (flags & SLAB_STORE_USER)
3720 * Need to store information about allocs and frees after
3723 size += 2 * sizeof(struct track);
3726 kasan_cache_create(s, &size, &s->flags);
3727 #ifdef CONFIG_SLUB_DEBUG
3728 if (flags & SLAB_RED_ZONE) {
3730 * Add some empty padding so that we can catch
3731 * overwrites from earlier objects rather than let
3732 * tracking information or the free pointer be
3733 * corrupted if a user writes before the start
3736 size += sizeof(void *);
3738 s->red_left_pad = sizeof(void *);
3739 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3740 size += s->red_left_pad;
3745 * SLUB stores one object immediately after another beginning from
3746 * offset 0. In order to align the objects we have to simply size
3747 * each object to conform to the alignment.
3749 size = ALIGN(size, s->align);
3751 s->reciprocal_size = reciprocal_value(size);
3752 if (forced_order >= 0)
3753 order = forced_order;
3755 order = calculate_order(size);
3762 s->allocflags |= __GFP_COMP;
3764 if (s->flags & SLAB_CACHE_DMA)
3765 s->allocflags |= GFP_DMA;
3767 if (s->flags & SLAB_CACHE_DMA32)
3768 s->allocflags |= GFP_DMA32;
3770 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3771 s->allocflags |= __GFP_RECLAIMABLE;
3774 * Determine the number of objects per slab
3776 s->oo = oo_make(order, size);
3777 s->min = oo_make(get_order(size), size);
3778 if (oo_objects(s->oo) > oo_objects(s->max))
3781 return !!oo_objects(s->oo);
3784 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3786 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3787 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3788 s->random = get_random_long();
3791 if (!calculate_sizes(s, -1))
3793 if (disable_higher_order_debug) {
3795 * Disable debugging flags that store metadata if the min slab
3798 if (get_order(s->size) > get_order(s->object_size)) {
3799 s->flags &= ~DEBUG_METADATA_FLAGS;
3801 if (!calculate_sizes(s, -1))
3806 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3807 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3808 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3809 /* Enable fast mode */
3810 s->flags |= __CMPXCHG_DOUBLE;
3814 * The larger the object size is, the more pages we want on the partial
3815 * list to avoid pounding the page allocator excessively.
3817 set_min_partial(s, ilog2(s->size) / 2);
3822 s->remote_node_defrag_ratio = 1000;
3825 /* Initialize the pre-computed randomized freelist if slab is up */
3826 if (slab_state >= UP) {
3827 if (init_cache_random_seq(s))
3831 if (!init_kmem_cache_nodes(s))
3834 if (alloc_kmem_cache_cpus(s))
3837 free_kmem_cache_nodes(s);
3842 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3845 #ifdef CONFIG_SLUB_DEBUG
3846 void *addr = page_address(page);
3850 slab_err(s, page, text, s->name);
3853 map = get_map(s, page);
3854 for_each_object(p, s, addr, page->objects) {
3856 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3857 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3858 print_tracking(s, p);
3867 * Attempt to free all partial slabs on a node.
3868 * This is called from __kmem_cache_shutdown(). We must take list_lock
3869 * because sysfs file might still access partial list after the shutdowning.
3871 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3874 struct page *page, *h;
3876 BUG_ON(irqs_disabled());
3877 spin_lock_irq(&n->list_lock);
3878 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3880 remove_partial(n, page);
3881 list_add(&page->slab_list, &discard);
3883 list_slab_objects(s, page,
3884 "Objects remaining in %s on __kmem_cache_shutdown()");
3887 spin_unlock_irq(&n->list_lock);
3889 list_for_each_entry_safe(page, h, &discard, slab_list)
3890 discard_slab(s, page);
3893 bool __kmem_cache_empty(struct kmem_cache *s)
3896 struct kmem_cache_node *n;
3898 for_each_kmem_cache_node(s, node, n)
3899 if (n->nr_partial || slabs_node(s, node))
3905 * Release all resources used by a slab cache.
3907 int __kmem_cache_shutdown(struct kmem_cache *s)
3910 struct kmem_cache_node *n;
3913 /* Attempt to free all objects */
3914 for_each_kmem_cache_node(s, node, n) {
3916 if (n->nr_partial || slabs_node(s, node))
3919 sysfs_slab_remove(s);
3923 /********************************************************************
3925 *******************************************************************/
3927 static int __init setup_slub_min_order(char *str)
3929 get_option(&str, (int *)&slub_min_order);
3934 __setup("slub_min_order=", setup_slub_min_order);
3936 static int __init setup_slub_max_order(char *str)
3938 get_option(&str, (int *)&slub_max_order);
3939 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3944 __setup("slub_max_order=", setup_slub_max_order);
3946 static int __init setup_slub_min_objects(char *str)
3948 get_option(&str, (int *)&slub_min_objects);
3953 __setup("slub_min_objects=", setup_slub_min_objects);
3955 void *__kmalloc(size_t size, gfp_t flags)
3957 struct kmem_cache *s;
3960 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3961 return kmalloc_large(size, flags);
3963 s = kmalloc_slab(size, flags);
3965 if (unlikely(ZERO_OR_NULL_PTR(s)))
3968 ret = slab_alloc(s, flags, _RET_IP_);
3970 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3972 ret = kasan_kmalloc(s, ret, size, flags);
3976 EXPORT_SYMBOL(__kmalloc);
3979 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3983 unsigned int order = get_order(size);
3985 flags |= __GFP_COMP;
3986 page = alloc_pages_node(node, flags, order);
3988 ptr = page_address(page);
3989 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
3990 PAGE_SIZE << order);
3993 return kmalloc_large_node_hook(ptr, size, flags);
3996 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3998 struct kmem_cache *s;
4001 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4002 ret = kmalloc_large_node(size, flags, node);
4004 trace_kmalloc_node(_RET_IP_, ret,
4005 size, PAGE_SIZE << get_order(size),
4011 s = kmalloc_slab(size, flags);
4013 if (unlikely(ZERO_OR_NULL_PTR(s)))
4016 ret = slab_alloc_node(s, flags, node, _RET_IP_);
4018 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4020 ret = kasan_kmalloc(s, ret, size, flags);
4024 EXPORT_SYMBOL(__kmalloc_node);
4025 #endif /* CONFIG_NUMA */
4027 #ifdef CONFIG_HARDENED_USERCOPY
4029 * Rejects incorrectly sized objects and objects that are to be copied
4030 * to/from userspace but do not fall entirely within the containing slab
4031 * cache's usercopy region.
4033 * Returns NULL if check passes, otherwise const char * to name of cache
4034 * to indicate an error.
4036 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4039 struct kmem_cache *s;
4040 unsigned int offset;
4043 ptr = kasan_reset_tag(ptr);
4045 /* Find object and usable object size. */
4046 s = page->slab_cache;
4048 /* Reject impossible pointers. */
4049 if (ptr < page_address(page))
4050 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4053 /* Find offset within object. */
4054 offset = (ptr - page_address(page)) % s->size;
4056 /* Adjust for redzone and reject if within the redzone. */
4057 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4058 if (offset < s->red_left_pad)
4059 usercopy_abort("SLUB object in left red zone",
4060 s->name, to_user, offset, n);
4061 offset -= s->red_left_pad;
4064 /* Allow address range falling entirely within usercopy region. */
4065 if (offset >= s->useroffset &&
4066 offset - s->useroffset <= s->usersize &&
4067 n <= s->useroffset - offset + s->usersize)
4071 * If the copy is still within the allocated object, produce
4072 * a warning instead of rejecting the copy. This is intended
4073 * to be a temporary method to find any missing usercopy
4076 object_size = slab_ksize(s);
4077 if (usercopy_fallback &&
4078 offset <= object_size && n <= object_size - offset) {
4079 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4083 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4085 #endif /* CONFIG_HARDENED_USERCOPY */
4087 size_t __ksize(const void *object)
4091 if (unlikely(object == ZERO_SIZE_PTR))
4094 page = virt_to_head_page(object);
4096 if (unlikely(!PageSlab(page))) {
4097 WARN_ON(!PageCompound(page));
4098 return page_size(page);
4101 return slab_ksize(page->slab_cache);
4103 EXPORT_SYMBOL(__ksize);
4105 void kfree(const void *x)
4108 void *object = (void *)x;
4110 trace_kfree(_RET_IP_, x);
4112 if (unlikely(ZERO_OR_NULL_PTR(x)))
4115 page = virt_to_head_page(x);
4116 if (unlikely(!PageSlab(page))) {
4117 unsigned int order = compound_order(page);
4119 BUG_ON(!PageCompound(page));
4121 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
4122 -(PAGE_SIZE << order));
4123 __free_pages(page, order);
4126 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4128 EXPORT_SYMBOL(kfree);
4130 #define SHRINK_PROMOTE_MAX 32
4133 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4134 * up most to the head of the partial lists. New allocations will then
4135 * fill those up and thus they can be removed from the partial lists.
4137 * The slabs with the least items are placed last. This results in them
4138 * being allocated from last increasing the chance that the last objects
4139 * are freed in them.
4141 int __kmem_cache_shrink(struct kmem_cache *s)
4145 struct kmem_cache_node *n;
4148 struct list_head discard;
4149 struct list_head promote[SHRINK_PROMOTE_MAX];
4150 unsigned long flags;
4154 for_each_kmem_cache_node(s, node, n) {
4155 INIT_LIST_HEAD(&discard);
4156 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4157 INIT_LIST_HEAD(promote + i);
4159 spin_lock_irqsave(&n->list_lock, flags);
4162 * Build lists of slabs to discard or promote.
4164 * Note that concurrent frees may occur while we hold the
4165 * list_lock. page->inuse here is the upper limit.
4167 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4168 int free = page->objects - page->inuse;
4170 /* Do not reread page->inuse */
4173 /* We do not keep full slabs on the list */
4176 if (free == page->objects) {
4177 list_move(&page->slab_list, &discard);
4179 } else if (free <= SHRINK_PROMOTE_MAX)
4180 list_move(&page->slab_list, promote + free - 1);
4184 * Promote the slabs filled up most to the head of the
4187 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4188 list_splice(promote + i, &n->partial);
4190 spin_unlock_irqrestore(&n->list_lock, flags);
4192 /* Release empty slabs */
4193 list_for_each_entry_safe(page, t, &discard, slab_list)
4194 discard_slab(s, page);
4196 if (slabs_node(s, node))
4204 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4207 * Called with all the locks held after a sched RCU grace period.
4208 * Even if @s becomes empty after shrinking, we can't know that @s
4209 * doesn't have allocations already in-flight and thus can't
4210 * destroy @s until the associated memcg is released.
4212 * However, let's remove the sysfs files for empty caches here.
4213 * Each cache has a lot of interface files which aren't
4214 * particularly useful for empty draining caches; otherwise, we can
4215 * easily end up with millions of unnecessary sysfs files on
4216 * systems which have a lot of memory and transient cgroups.
4218 if (!__kmem_cache_shrink(s))
4219 sysfs_slab_remove(s);
4222 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4225 * Disable empty slabs caching. Used to avoid pinning offline
4226 * memory cgroups by kmem pages that can be freed.
4228 slub_set_cpu_partial(s, 0);
4231 #endif /* CONFIG_MEMCG */
4233 static int slab_mem_going_offline_callback(void *arg)
4235 struct kmem_cache *s;
4237 mutex_lock(&slab_mutex);
4238 list_for_each_entry(s, &slab_caches, list)
4239 __kmem_cache_shrink(s);
4240 mutex_unlock(&slab_mutex);
4245 static void slab_mem_offline_callback(void *arg)
4247 struct kmem_cache_node *n;
4248 struct kmem_cache *s;
4249 struct memory_notify *marg = arg;
4252 offline_node = marg->status_change_nid_normal;
4255 * If the node still has available memory. we need kmem_cache_node
4258 if (offline_node < 0)
4261 mutex_lock(&slab_mutex);
4262 list_for_each_entry(s, &slab_caches, list) {
4263 n = get_node(s, offline_node);
4266 * if n->nr_slabs > 0, slabs still exist on the node
4267 * that is going down. We were unable to free them,
4268 * and offline_pages() function shouldn't call this
4269 * callback. So, we must fail.
4271 BUG_ON(slabs_node(s, offline_node));
4273 s->node[offline_node] = NULL;
4274 kmem_cache_free(kmem_cache_node, n);
4277 mutex_unlock(&slab_mutex);
4280 static int slab_mem_going_online_callback(void *arg)
4282 struct kmem_cache_node *n;
4283 struct kmem_cache *s;
4284 struct memory_notify *marg = arg;
4285 int nid = marg->status_change_nid_normal;
4289 * If the node's memory is already available, then kmem_cache_node is
4290 * already created. Nothing to do.
4296 * We are bringing a node online. No memory is available yet. We must
4297 * allocate a kmem_cache_node structure in order to bring the node
4300 mutex_lock(&slab_mutex);
4301 list_for_each_entry(s, &slab_caches, list) {
4303 * XXX: kmem_cache_alloc_node will fallback to other nodes
4304 * since memory is not yet available from the node that
4307 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4312 init_kmem_cache_node(n);
4316 mutex_unlock(&slab_mutex);
4320 static int slab_memory_callback(struct notifier_block *self,
4321 unsigned long action, void *arg)
4326 case MEM_GOING_ONLINE:
4327 ret = slab_mem_going_online_callback(arg);
4329 case MEM_GOING_OFFLINE:
4330 ret = slab_mem_going_offline_callback(arg);
4333 case MEM_CANCEL_ONLINE:
4334 slab_mem_offline_callback(arg);
4337 case MEM_CANCEL_OFFLINE:
4341 ret = notifier_from_errno(ret);
4347 static struct notifier_block slab_memory_callback_nb = {
4348 .notifier_call = slab_memory_callback,
4349 .priority = SLAB_CALLBACK_PRI,
4352 /********************************************************************
4353 * Basic setup of slabs
4354 *******************************************************************/
4357 * Used for early kmem_cache structures that were allocated using
4358 * the page allocator. Allocate them properly then fix up the pointers
4359 * that may be pointing to the wrong kmem_cache structure.
4362 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4365 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4366 struct kmem_cache_node *n;
4368 memcpy(s, static_cache, kmem_cache->object_size);
4371 * This runs very early, and only the boot processor is supposed to be
4372 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4375 __flush_cpu_slab(s, smp_processor_id());
4376 for_each_kmem_cache_node(s, node, n) {
4379 list_for_each_entry(p, &n->partial, slab_list)
4382 #ifdef CONFIG_SLUB_DEBUG
4383 list_for_each_entry(p, &n->full, slab_list)
4387 slab_init_memcg_params(s);
4388 list_add(&s->list, &slab_caches);
4389 memcg_link_cache(s, NULL);
4393 void __init kmem_cache_init(void)
4395 static __initdata struct kmem_cache boot_kmem_cache,
4396 boot_kmem_cache_node;
4398 if (debug_guardpage_minorder())
4401 kmem_cache_node = &boot_kmem_cache_node;
4402 kmem_cache = &boot_kmem_cache;
4404 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4405 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4407 register_hotmemory_notifier(&slab_memory_callback_nb);
4409 /* Able to allocate the per node structures */
4410 slab_state = PARTIAL;
4412 create_boot_cache(kmem_cache, "kmem_cache",
4413 offsetof(struct kmem_cache, node) +
4414 nr_node_ids * sizeof(struct kmem_cache_node *),
4415 SLAB_HWCACHE_ALIGN, 0, 0);
4417 kmem_cache = bootstrap(&boot_kmem_cache);
4418 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4420 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4421 setup_kmalloc_cache_index_table();
4422 create_kmalloc_caches(0);
4424 /* Setup random freelists for each cache */
4425 init_freelist_randomization();
4427 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4430 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4432 slub_min_order, slub_max_order, slub_min_objects,
4433 nr_cpu_ids, nr_node_ids);
4436 void __init kmem_cache_init_late(void)
4441 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4442 slab_flags_t flags, void (*ctor)(void *))
4444 struct kmem_cache *s, *c;
4446 s = find_mergeable(size, align, flags, name, ctor);
4451 * Adjust the object sizes so that we clear
4452 * the complete object on kzalloc.
4454 s->object_size = max(s->object_size, size);
4455 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4457 for_each_memcg_cache(c, s) {
4458 c->object_size = s->object_size;
4459 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4462 if (sysfs_slab_alias(s, name)) {
4471 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4475 err = kmem_cache_open(s, flags);
4479 /* Mutex is not taken during early boot */
4480 if (slab_state <= UP)
4483 memcg_propagate_slab_attrs(s);
4484 err = sysfs_slab_add(s);
4486 __kmem_cache_release(s);
4491 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4493 struct kmem_cache *s;
4496 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4497 return kmalloc_large(size, gfpflags);
4499 s = kmalloc_slab(size, gfpflags);
4501 if (unlikely(ZERO_OR_NULL_PTR(s)))
4504 ret = slab_alloc(s, gfpflags, caller);
4506 /* Honor the call site pointer we received. */
4507 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4511 EXPORT_SYMBOL(__kmalloc_track_caller);
4514 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4515 int node, unsigned long caller)
4517 struct kmem_cache *s;
4520 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4521 ret = kmalloc_large_node(size, gfpflags, node);
4523 trace_kmalloc_node(caller, ret,
4524 size, PAGE_SIZE << get_order(size),
4530 s = kmalloc_slab(size, gfpflags);
4532 if (unlikely(ZERO_OR_NULL_PTR(s)))
4535 ret = slab_alloc_node(s, gfpflags, node, caller);
4537 /* Honor the call site pointer we received. */
4538 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4542 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4546 static int count_inuse(struct page *page)
4551 static int count_total(struct page *page)
4553 return page->objects;
4557 #ifdef CONFIG_SLUB_DEBUG
4558 static void validate_slab(struct kmem_cache *s, struct page *page)
4561 void *addr = page_address(page);
4566 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4569 /* Now we know that a valid freelist exists */
4570 map = get_map(s, page);
4571 for_each_object(p, s, addr, page->objects) {
4572 u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
4573 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4575 if (!check_object(s, page, p, val))
4583 static int validate_slab_node(struct kmem_cache *s,
4584 struct kmem_cache_node *n)
4586 unsigned long count = 0;
4588 unsigned long flags;
4590 spin_lock_irqsave(&n->list_lock, flags);
4592 list_for_each_entry(page, &n->partial, slab_list) {
4593 validate_slab(s, page);
4596 if (count != n->nr_partial)
4597 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4598 s->name, count, n->nr_partial);
4600 if (!(s->flags & SLAB_STORE_USER))
4603 list_for_each_entry(page, &n->full, slab_list) {
4604 validate_slab(s, page);
4607 if (count != atomic_long_read(&n->nr_slabs))
4608 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4609 s->name, count, atomic_long_read(&n->nr_slabs));
4612 spin_unlock_irqrestore(&n->list_lock, flags);
4616 static long validate_slab_cache(struct kmem_cache *s)
4619 unsigned long count = 0;
4620 struct kmem_cache_node *n;
4623 for_each_kmem_cache_node(s, node, n)
4624 count += validate_slab_node(s, n);
4629 * Generate lists of code addresses where slabcache objects are allocated
4634 unsigned long count;
4641 DECLARE_BITMAP(cpus, NR_CPUS);
4647 unsigned long count;
4648 struct location *loc;
4651 static void free_loc_track(struct loc_track *t)
4654 free_pages((unsigned long)t->loc,
4655 get_order(sizeof(struct location) * t->max));
4658 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4663 order = get_order(sizeof(struct location) * max);
4665 l = (void *)__get_free_pages(flags, order);
4670 memcpy(l, t->loc, sizeof(struct location) * t->count);
4678 static int add_location(struct loc_track *t, struct kmem_cache *s,
4679 const struct track *track)
4681 long start, end, pos;
4683 unsigned long caddr;
4684 unsigned long age = jiffies - track->when;
4690 pos = start + (end - start + 1) / 2;
4693 * There is nothing at "end". If we end up there
4694 * we need to add something to before end.
4699 caddr = t->loc[pos].addr;
4700 if (track->addr == caddr) {
4706 if (age < l->min_time)
4708 if (age > l->max_time)
4711 if (track->pid < l->min_pid)
4712 l->min_pid = track->pid;
4713 if (track->pid > l->max_pid)
4714 l->max_pid = track->pid;
4716 cpumask_set_cpu(track->cpu,
4717 to_cpumask(l->cpus));
4719 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4723 if (track->addr < caddr)
4730 * Not found. Insert new tracking element.
4732 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4738 (t->count - pos) * sizeof(struct location));
4741 l->addr = track->addr;
4745 l->min_pid = track->pid;
4746 l->max_pid = track->pid;
4747 cpumask_clear(to_cpumask(l->cpus));
4748 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4749 nodes_clear(l->nodes);
4750 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4754 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4755 struct page *page, enum track_item alloc)
4757 void *addr = page_address(page);
4761 map = get_map(s, page);
4762 for_each_object(p, s, addr, page->objects)
4763 if (!test_bit(__obj_to_index(s, addr, p), map))
4764 add_location(t, s, get_track(s, p, alloc));
4768 static int list_locations(struct kmem_cache *s, char *buf,
4769 enum track_item alloc)
4773 struct loc_track t = { 0, 0, NULL };
4775 struct kmem_cache_node *n;
4777 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4779 return sprintf(buf, "Out of memory\n");
4781 /* Push back cpu slabs */
4784 for_each_kmem_cache_node(s, node, n) {
4785 unsigned long flags;
4788 if (!atomic_long_read(&n->nr_slabs))
4791 spin_lock_irqsave(&n->list_lock, flags);
4792 list_for_each_entry(page, &n->partial, slab_list)
4793 process_slab(&t, s, page, alloc);
4794 list_for_each_entry(page, &n->full, slab_list)
4795 process_slab(&t, s, page, alloc);
4796 spin_unlock_irqrestore(&n->list_lock, flags);
4799 for (i = 0; i < t.count; i++) {
4800 struct location *l = &t.loc[i];
4802 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4804 len += sprintf(buf + len, "%7ld ", l->count);
4807 len += sprintf(buf + len, "%pS", (void *)l->addr);
4809 len += sprintf(buf + len, "<not-available>");
4811 if (l->sum_time != l->min_time) {
4812 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4814 (long)div_u64(l->sum_time, l->count),
4817 len += sprintf(buf + len, " age=%ld",
4820 if (l->min_pid != l->max_pid)
4821 len += sprintf(buf + len, " pid=%ld-%ld",
4822 l->min_pid, l->max_pid);
4824 len += sprintf(buf + len, " pid=%ld",
4827 if (num_online_cpus() > 1 &&
4828 !cpumask_empty(to_cpumask(l->cpus)) &&
4829 len < PAGE_SIZE - 60)
4830 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4832 cpumask_pr_args(to_cpumask(l->cpus)));
4834 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4835 len < PAGE_SIZE - 60)
4836 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4838 nodemask_pr_args(&l->nodes));
4840 len += sprintf(buf + len, "\n");
4845 len += sprintf(buf, "No data\n");
4848 #endif /* CONFIG_SLUB_DEBUG */
4850 #ifdef SLUB_RESILIENCY_TEST
4851 static void __init resiliency_test(void)
4854 int type = KMALLOC_NORMAL;
4856 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4858 pr_err("SLUB resiliency testing\n");
4859 pr_err("-----------------------\n");
4860 pr_err("A. Corruption after allocation\n");
4862 p = kzalloc(16, GFP_KERNEL);
4864 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4867 validate_slab_cache(kmalloc_caches[type][4]);
4869 /* Hmmm... The next two are dangerous */
4870 p = kzalloc(32, GFP_KERNEL);
4871 p[32 + sizeof(void *)] = 0x34;
4872 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4874 pr_err("If allocated object is overwritten then not detectable\n\n");
4876 validate_slab_cache(kmalloc_caches[type][5]);
4877 p = kzalloc(64, GFP_KERNEL);
4878 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4880 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4882 pr_err("If allocated object is overwritten then not detectable\n\n");
4883 validate_slab_cache(kmalloc_caches[type][6]);
4885 pr_err("\nB. Corruption after free\n");
4886 p = kzalloc(128, GFP_KERNEL);
4889 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4890 validate_slab_cache(kmalloc_caches[type][7]);
4892 p = kzalloc(256, GFP_KERNEL);
4895 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4896 validate_slab_cache(kmalloc_caches[type][8]);
4898 p = kzalloc(512, GFP_KERNEL);
4901 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4902 validate_slab_cache(kmalloc_caches[type][9]);
4906 static void resiliency_test(void) {};
4908 #endif /* SLUB_RESILIENCY_TEST */
4911 enum slab_stat_type {
4912 SL_ALL, /* All slabs */
4913 SL_PARTIAL, /* Only partially allocated slabs */
4914 SL_CPU, /* Only slabs used for cpu caches */
4915 SL_OBJECTS, /* Determine allocated objects not slabs */
4916 SL_TOTAL /* Determine object capacity not slabs */
4919 #define SO_ALL (1 << SL_ALL)
4920 #define SO_PARTIAL (1 << SL_PARTIAL)
4921 #define SO_CPU (1 << SL_CPU)
4922 #define SO_OBJECTS (1 << SL_OBJECTS)
4923 #define SO_TOTAL (1 << SL_TOTAL)
4926 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4928 static int __init setup_slub_memcg_sysfs(char *str)
4932 if (get_option(&str, &v) > 0)
4933 memcg_sysfs_enabled = v;
4938 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4941 static ssize_t show_slab_objects(struct kmem_cache *s,
4942 char *buf, unsigned long flags)
4944 unsigned long total = 0;
4947 unsigned long *nodes;
4949 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4953 if (flags & SO_CPU) {
4956 for_each_possible_cpu(cpu) {
4957 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4962 page = READ_ONCE(c->page);
4966 node = page_to_nid(page);
4967 if (flags & SO_TOTAL)
4969 else if (flags & SO_OBJECTS)
4977 page = slub_percpu_partial_read_once(c);
4979 node = page_to_nid(page);
4980 if (flags & SO_TOTAL)
4982 else if (flags & SO_OBJECTS)
4993 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4994 * already held which will conflict with an existing lock order:
4996 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4998 * We don't really need mem_hotplug_lock (to hold off
4999 * slab_mem_going_offline_callback) here because slab's memory hot
5000 * unplug code doesn't destroy the kmem_cache->node[] data.
5003 #ifdef CONFIG_SLUB_DEBUG
5004 if (flags & SO_ALL) {
5005 struct kmem_cache_node *n;
5007 for_each_kmem_cache_node(s, node, n) {
5009 if (flags & SO_TOTAL)
5010 x = atomic_long_read(&n->total_objects);
5011 else if (flags & SO_OBJECTS)
5012 x = atomic_long_read(&n->total_objects) -
5013 count_partial(n, count_free);
5015 x = atomic_long_read(&n->nr_slabs);
5022 if (flags & SO_PARTIAL) {
5023 struct kmem_cache_node *n;
5025 for_each_kmem_cache_node(s, node, n) {
5026 if (flags & SO_TOTAL)
5027 x = count_partial(n, count_total);
5028 else if (flags & SO_OBJECTS)
5029 x = count_partial(n, count_inuse);
5036 x = sprintf(buf, "%lu", total);
5038 for (node = 0; node < nr_node_ids; node++)
5040 x += sprintf(buf + x, " N%d=%lu",
5044 return x + sprintf(buf + x, "\n");
5047 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5048 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5050 struct slab_attribute {
5051 struct attribute attr;
5052 ssize_t (*show)(struct kmem_cache *s, char *buf);
5053 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5056 #define SLAB_ATTR_RO(_name) \
5057 static struct slab_attribute _name##_attr = \
5058 __ATTR(_name, 0400, _name##_show, NULL)
5060 #define SLAB_ATTR(_name) \
5061 static struct slab_attribute _name##_attr = \
5062 __ATTR(_name, 0600, _name##_show, _name##_store)
5064 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5066 return sprintf(buf, "%u\n", s->size);
5068 SLAB_ATTR_RO(slab_size);
5070 static ssize_t align_show(struct kmem_cache *s, char *buf)
5072 return sprintf(buf, "%u\n", s->align);
5074 SLAB_ATTR_RO(align);
5076 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5078 return sprintf(buf, "%u\n", s->object_size);
5080 SLAB_ATTR_RO(object_size);
5082 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5084 return sprintf(buf, "%u\n", oo_objects(s->oo));
5086 SLAB_ATTR_RO(objs_per_slab);
5088 static ssize_t order_show(struct kmem_cache *s, char *buf)
5090 return sprintf(buf, "%u\n", oo_order(s->oo));
5092 SLAB_ATTR_RO(order);
5094 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5096 return sprintf(buf, "%lu\n", s->min_partial);
5099 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5105 err = kstrtoul(buf, 10, &min);
5109 set_min_partial(s, min);
5112 SLAB_ATTR(min_partial);
5114 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5116 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5119 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5122 unsigned int objects;
5125 err = kstrtouint(buf, 10, &objects);
5128 if (objects && !kmem_cache_has_cpu_partial(s))
5131 slub_set_cpu_partial(s, objects);
5135 SLAB_ATTR(cpu_partial);
5137 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5141 return sprintf(buf, "%pS\n", s->ctor);
5145 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5147 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5149 SLAB_ATTR_RO(aliases);
5151 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5153 return show_slab_objects(s, buf, SO_PARTIAL);
5155 SLAB_ATTR_RO(partial);
5157 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5159 return show_slab_objects(s, buf, SO_CPU);
5161 SLAB_ATTR_RO(cpu_slabs);
5163 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5165 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5167 SLAB_ATTR_RO(objects);
5169 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5171 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5173 SLAB_ATTR_RO(objects_partial);
5175 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5182 for_each_online_cpu(cpu) {
5185 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5188 pages += page->pages;
5189 objects += page->pobjects;
5193 len = sprintf(buf, "%d(%d)", objects, pages);
5196 for_each_online_cpu(cpu) {
5199 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5201 if (page && len < PAGE_SIZE - 20)
5202 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5203 page->pobjects, page->pages);
5206 return len + sprintf(buf + len, "\n");
5208 SLAB_ATTR_RO(slabs_cpu_partial);
5210 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5212 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5214 SLAB_ATTR_RO(reclaim_account);
5216 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5218 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5220 SLAB_ATTR_RO(hwcache_align);
5222 #ifdef CONFIG_ZONE_DMA
5223 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5225 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5227 SLAB_ATTR_RO(cache_dma);
5230 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5232 return sprintf(buf, "%u\n", s->usersize);
5234 SLAB_ATTR_RO(usersize);
5236 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5238 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5240 SLAB_ATTR_RO(destroy_by_rcu);
5242 #ifdef CONFIG_SLUB_DEBUG
5243 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5245 return show_slab_objects(s, buf, SO_ALL);
5247 SLAB_ATTR_RO(slabs);
5249 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5251 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5253 SLAB_ATTR_RO(total_objects);
5255 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5257 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5259 SLAB_ATTR_RO(sanity_checks);
5261 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5263 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5265 SLAB_ATTR_RO(trace);
5267 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5269 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5272 SLAB_ATTR_RO(red_zone);
5274 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5276 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5279 SLAB_ATTR_RO(poison);
5281 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5283 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5286 SLAB_ATTR_RO(store_user);
5288 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5293 static ssize_t validate_store(struct kmem_cache *s,
5294 const char *buf, size_t length)
5298 if (buf[0] == '1') {
5299 ret = validate_slab_cache(s);
5305 SLAB_ATTR(validate);
5307 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5309 if (!(s->flags & SLAB_STORE_USER))
5311 return list_locations(s, buf, TRACK_ALLOC);
5313 SLAB_ATTR_RO(alloc_calls);
5315 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5317 if (!(s->flags & SLAB_STORE_USER))
5319 return list_locations(s, buf, TRACK_FREE);
5321 SLAB_ATTR_RO(free_calls);
5322 #endif /* CONFIG_SLUB_DEBUG */
5324 #ifdef CONFIG_FAILSLAB
5325 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5327 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5329 SLAB_ATTR_RO(failslab);
5332 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5337 static ssize_t shrink_store(struct kmem_cache *s,
5338 const char *buf, size_t length)
5341 kmem_cache_shrink_all(s);
5349 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5351 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5354 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5355 const char *buf, size_t length)
5360 err = kstrtouint(buf, 10, &ratio);
5366 s->remote_node_defrag_ratio = ratio * 10;
5370 SLAB_ATTR(remote_node_defrag_ratio);
5373 #ifdef CONFIG_SLUB_STATS
5374 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5376 unsigned long sum = 0;
5379 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5384 for_each_online_cpu(cpu) {
5385 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5391 len = sprintf(buf, "%lu", sum);
5394 for_each_online_cpu(cpu) {
5395 if (data[cpu] && len < PAGE_SIZE - 20)
5396 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5400 return len + sprintf(buf + len, "\n");
5403 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5407 for_each_online_cpu(cpu)
5408 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5411 #define STAT_ATTR(si, text) \
5412 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5414 return show_stat(s, buf, si); \
5416 static ssize_t text##_store(struct kmem_cache *s, \
5417 const char *buf, size_t length) \
5419 if (buf[0] != '0') \
5421 clear_stat(s, si); \
5426 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5427 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5428 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5429 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5430 STAT_ATTR(FREE_FROZEN, free_frozen);
5431 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5432 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5433 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5434 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5435 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5436 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5437 STAT_ATTR(FREE_SLAB, free_slab);
5438 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5439 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5440 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5441 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5442 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5443 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5444 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5445 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5446 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5447 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5448 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5449 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5450 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5451 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5452 #endif /* CONFIG_SLUB_STATS */
5454 static struct attribute *slab_attrs[] = {
5455 &slab_size_attr.attr,
5456 &object_size_attr.attr,
5457 &objs_per_slab_attr.attr,
5459 &min_partial_attr.attr,
5460 &cpu_partial_attr.attr,
5462 &objects_partial_attr.attr,
5464 &cpu_slabs_attr.attr,
5468 &hwcache_align_attr.attr,
5469 &reclaim_account_attr.attr,
5470 &destroy_by_rcu_attr.attr,
5472 &slabs_cpu_partial_attr.attr,
5473 #ifdef CONFIG_SLUB_DEBUG
5474 &total_objects_attr.attr,
5476 &sanity_checks_attr.attr,
5478 &red_zone_attr.attr,
5480 &store_user_attr.attr,
5481 &validate_attr.attr,
5482 &alloc_calls_attr.attr,
5483 &free_calls_attr.attr,
5485 #ifdef CONFIG_ZONE_DMA
5486 &cache_dma_attr.attr,
5489 &remote_node_defrag_ratio_attr.attr,
5491 #ifdef CONFIG_SLUB_STATS
5492 &alloc_fastpath_attr.attr,
5493 &alloc_slowpath_attr.attr,
5494 &free_fastpath_attr.attr,
5495 &free_slowpath_attr.attr,
5496 &free_frozen_attr.attr,
5497 &free_add_partial_attr.attr,
5498 &free_remove_partial_attr.attr,
5499 &alloc_from_partial_attr.attr,
5500 &alloc_slab_attr.attr,
5501 &alloc_refill_attr.attr,
5502 &alloc_node_mismatch_attr.attr,
5503 &free_slab_attr.attr,
5504 &cpuslab_flush_attr.attr,
5505 &deactivate_full_attr.attr,
5506 &deactivate_empty_attr.attr,
5507 &deactivate_to_head_attr.attr,
5508 &deactivate_to_tail_attr.attr,
5509 &deactivate_remote_frees_attr.attr,
5510 &deactivate_bypass_attr.attr,
5511 &order_fallback_attr.attr,
5512 &cmpxchg_double_fail_attr.attr,
5513 &cmpxchg_double_cpu_fail_attr.attr,
5514 &cpu_partial_alloc_attr.attr,
5515 &cpu_partial_free_attr.attr,
5516 &cpu_partial_node_attr.attr,
5517 &cpu_partial_drain_attr.attr,
5519 #ifdef CONFIG_FAILSLAB
5520 &failslab_attr.attr,
5522 &usersize_attr.attr,
5527 static const struct attribute_group slab_attr_group = {
5528 .attrs = slab_attrs,
5531 static ssize_t slab_attr_show(struct kobject *kobj,
5532 struct attribute *attr,
5535 struct slab_attribute *attribute;
5536 struct kmem_cache *s;
5539 attribute = to_slab_attr(attr);
5542 if (!attribute->show)
5545 err = attribute->show(s, buf);
5550 static ssize_t slab_attr_store(struct kobject *kobj,
5551 struct attribute *attr,
5552 const char *buf, size_t len)
5554 struct slab_attribute *attribute;
5555 struct kmem_cache *s;
5558 attribute = to_slab_attr(attr);
5561 if (!attribute->store)
5564 err = attribute->store(s, buf, len);
5566 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5567 struct kmem_cache *c;
5569 mutex_lock(&slab_mutex);
5570 if (s->max_attr_size < len)
5571 s->max_attr_size = len;
5574 * This is a best effort propagation, so this function's return
5575 * value will be determined by the parent cache only. This is
5576 * basically because not all attributes will have a well
5577 * defined semantics for rollbacks - most of the actions will
5578 * have permanent effects.
5580 * Returning the error value of any of the children that fail
5581 * is not 100 % defined, in the sense that users seeing the
5582 * error code won't be able to know anything about the state of
5585 * Only returning the error code for the parent cache at least
5586 * has well defined semantics. The cache being written to
5587 * directly either failed or succeeded, in which case we loop
5588 * through the descendants with best-effort propagation.
5590 for_each_memcg_cache(c, s)
5591 attribute->store(c, buf, len);
5592 mutex_unlock(&slab_mutex);
5598 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5602 char *buffer = NULL;
5603 struct kmem_cache *root_cache;
5605 if (is_root_cache(s))
5608 root_cache = s->memcg_params.root_cache;
5611 * This mean this cache had no attribute written. Therefore, no point
5612 * in copying default values around
5614 if (!root_cache->max_attr_size)
5617 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5620 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5623 if (!attr || !attr->store || !attr->show)
5627 * It is really bad that we have to allocate here, so we will
5628 * do it only as a fallback. If we actually allocate, though,
5629 * we can just use the allocated buffer until the end.
5631 * Most of the slub attributes will tend to be very small in
5632 * size, but sysfs allows buffers up to a page, so they can
5633 * theoretically happen.
5637 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5638 !IS_ENABLED(CONFIG_SLUB_STATS))
5641 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5642 if (WARN_ON(!buffer))
5647 len = attr->show(root_cache, buf);
5649 attr->store(s, buf, len);
5653 free_page((unsigned long)buffer);
5654 #endif /* CONFIG_MEMCG */
5657 static void kmem_cache_release(struct kobject *k)
5659 slab_kmem_cache_release(to_slab(k));
5662 static const struct sysfs_ops slab_sysfs_ops = {
5663 .show = slab_attr_show,
5664 .store = slab_attr_store,
5667 static struct kobj_type slab_ktype = {
5668 .sysfs_ops = &slab_sysfs_ops,
5669 .release = kmem_cache_release,
5672 static struct kset *slab_kset;
5674 static inline struct kset *cache_kset(struct kmem_cache *s)
5677 if (!is_root_cache(s))
5678 return s->memcg_params.root_cache->memcg_kset;
5683 #define ID_STR_LENGTH 64
5685 /* Create a unique string id for a slab cache:
5687 * Format :[flags-]size
5689 static char *create_unique_id(struct kmem_cache *s)
5691 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5698 * First flags affecting slabcache operations. We will only
5699 * get here for aliasable slabs so we do not need to support
5700 * too many flags. The flags here must cover all flags that
5701 * are matched during merging to guarantee that the id is
5704 if (s->flags & SLAB_CACHE_DMA)
5706 if (s->flags & SLAB_CACHE_DMA32)
5708 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5710 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5712 if (s->flags & SLAB_ACCOUNT)
5716 p += sprintf(p, "%07u", s->size);
5718 BUG_ON(p > name + ID_STR_LENGTH - 1);
5722 static void sysfs_slab_remove_workfn(struct work_struct *work)
5724 struct kmem_cache *s =
5725 container_of(work, struct kmem_cache, kobj_remove_work);
5727 if (!s->kobj.state_in_sysfs)
5729 * For a memcg cache, this may be called during
5730 * deactivation and again on shutdown. Remove only once.
5731 * A cache is never shut down before deactivation is
5732 * complete, so no need to worry about synchronization.
5737 kset_unregister(s->memcg_kset);
5740 kobject_put(&s->kobj);
5743 static int sysfs_slab_add(struct kmem_cache *s)
5747 struct kset *kset = cache_kset(s);
5748 int unmergeable = slab_unmergeable(s);
5750 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5753 kobject_init(&s->kobj, &slab_ktype);
5757 if (!unmergeable && disable_higher_order_debug &&
5758 (slub_debug & DEBUG_METADATA_FLAGS))
5763 * Slabcache can never be merged so we can use the name proper.
5764 * This is typically the case for debug situations. In that
5765 * case we can catch duplicate names easily.
5767 sysfs_remove_link(&slab_kset->kobj, s->name);
5771 * Create a unique name for the slab as a target
5774 name = create_unique_id(s);
5777 s->kobj.kset = kset;
5778 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5780 kobject_put(&s->kobj);
5784 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5789 if (is_root_cache(s) && memcg_sysfs_enabled) {
5790 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5791 if (!s->memcg_kset) {
5799 /* Setup first alias */
5800 sysfs_slab_alias(s, s->name);
5807 kobject_del(&s->kobj);
5811 static void sysfs_slab_remove(struct kmem_cache *s)
5813 if (slab_state < FULL)
5815 * Sysfs has not been setup yet so no need to remove the
5820 kobject_get(&s->kobj);
5821 schedule_work(&s->kobj_remove_work);
5824 void sysfs_slab_unlink(struct kmem_cache *s)
5826 if (slab_state >= FULL)
5827 kobject_del(&s->kobj);
5830 void sysfs_slab_release(struct kmem_cache *s)
5832 if (slab_state >= FULL)
5833 kobject_put(&s->kobj);
5837 * Need to buffer aliases during bootup until sysfs becomes
5838 * available lest we lose that information.
5840 struct saved_alias {
5841 struct kmem_cache *s;
5843 struct saved_alias *next;
5846 static struct saved_alias *alias_list;
5848 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5850 struct saved_alias *al;
5852 if (slab_state == FULL) {
5854 * If we have a leftover link then remove it.
5856 sysfs_remove_link(&slab_kset->kobj, name);
5857 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5860 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5866 al->next = alias_list;
5871 static int __init slab_sysfs_init(void)
5873 struct kmem_cache *s;
5876 mutex_lock(&slab_mutex);
5878 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5880 mutex_unlock(&slab_mutex);
5881 pr_err("Cannot register slab subsystem.\n");
5887 list_for_each_entry(s, &slab_caches, list) {
5888 err = sysfs_slab_add(s);
5890 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5894 while (alias_list) {
5895 struct saved_alias *al = alias_list;
5897 alias_list = alias_list->next;
5898 err = sysfs_slab_alias(al->s, al->name);
5900 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5905 mutex_unlock(&slab_mutex);
5910 __initcall(slab_sysfs_init);
5911 #endif /* CONFIG_SYSFS */
5914 * The /proc/slabinfo ABI
5916 #ifdef CONFIG_SLUB_DEBUG
5917 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5919 unsigned long nr_slabs = 0;
5920 unsigned long nr_objs = 0;
5921 unsigned long nr_free = 0;
5923 struct kmem_cache_node *n;
5925 for_each_kmem_cache_node(s, node, n) {
5926 nr_slabs += node_nr_slabs(n);
5927 nr_objs += node_nr_objs(n);
5928 nr_free += count_partial(n, count_free);
5931 sinfo->active_objs = nr_objs - nr_free;
5932 sinfo->num_objs = nr_objs;
5933 sinfo->active_slabs = nr_slabs;
5934 sinfo->num_slabs = nr_slabs;
5935 sinfo->objects_per_slab = oo_objects(s->oo);
5936 sinfo->cache_order = oo_order(s->oo);
5939 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5943 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5944 size_t count, loff_t *ppos)
5948 #endif /* CONFIG_SLUB_DEBUG */