1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * page->frozen The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 void *fixup_red_left(struct kmem_cache *s, void *p)
128 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
129 p += s->red_left_pad;
134 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
136 #ifdef CONFIG_SLUB_CPU_PARTIAL
137 return !kmem_cache_debug(s);
144 * Issues still to be resolved:
146 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 * - Variable sizing of the per node arrays
151 /* Enable to test recovery from slab corruption on boot */
152 #undef SLUB_RESILIENCY_TEST
154 /* Enable to log cmpxchg failures */
155 #undef SLUB_DEBUG_CMPXCHG
158 * Mininum number of partial slabs. These will be left on the partial
159 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 #define MIN_PARTIAL 5
164 * Maximum number of desirable partial slabs.
165 * The existence of more partial slabs makes kmem_cache_shrink
166 * sort the partial list by the number of objects in use.
168 #define MAX_PARTIAL 10
170 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_STORE_USER)
174 * These debug flags cannot use CMPXCHG because there might be consistency
175 * issues when checking or reading debug information
177 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
182 * Debugging flags that require metadata to be stored in the slab. These get
183 * disabled when slub_debug=O is used and a cache's min order increases with
186 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
189 #define OO_MASK ((1 << OO_SHIFT) - 1)
190 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
192 /* Internal SLUB flags */
194 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
195 /* Use cmpxchg_double */
196 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
199 * Tracking user of a slab.
201 #define TRACK_ADDRS_COUNT 16
203 unsigned long addr; /* Called from address */
204 #ifdef CONFIG_STACKTRACE
205 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
207 int cpu; /* Was running on cpu */
208 int pid; /* Pid context */
209 unsigned long when; /* When did the operation occur */
212 enum track_item { TRACK_ALLOC, TRACK_FREE };
215 static int sysfs_slab_add(struct kmem_cache *);
216 static int sysfs_slab_alias(struct kmem_cache *, const char *);
217 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
218 static void sysfs_slab_remove(struct kmem_cache *s);
220 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
221 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
223 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
224 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
227 static inline void stat(const struct kmem_cache *s, enum stat_item si)
229 #ifdef CONFIG_SLUB_STATS
231 * The rmw is racy on a preemptible kernel but this is acceptable, so
232 * avoid this_cpu_add()'s irq-disable overhead.
234 raw_cpu_inc(s->cpu_slab->stat[si]);
238 /********************************************************************
239 * Core slab cache functions
240 *******************************************************************/
243 * Returns freelist pointer (ptr). With hardening, this is obfuscated
244 * with an XOR of the address where the pointer is held and a per-cache
247 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
248 unsigned long ptr_addr)
250 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
253 * Normally, this doesn't cause any issues, as both set_freepointer()
254 * and get_freepointer() are called with a pointer with the same tag.
255 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256 * example, when __free_slub() iterates over objects in a cache, it
257 * passes untagged pointers to check_object(). check_object() in turns
258 * calls get_freepointer() with an untagged pointer, which causes the
259 * freepointer to be restored incorrectly.
261 return (void *)((unsigned long)ptr ^ s->random ^
262 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
268 /* Returns the freelist pointer recorded at location ptr_addr. */
269 static inline void *freelist_dereference(const struct kmem_cache *s,
272 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
273 (unsigned long)ptr_addr);
276 static inline void *get_freepointer(struct kmem_cache *s, void *object)
278 return freelist_dereference(s, object + s->offset);
281 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
283 prefetch(object + s->offset);
286 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
288 unsigned long freepointer_addr;
291 if (!debug_pagealloc_enabled_static())
292 return get_freepointer(s, object);
294 freepointer_addr = (unsigned long)object + s->offset;
295 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
296 return freelist_ptr(s, p, freepointer_addr);
299 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
301 unsigned long freeptr_addr = (unsigned long)object + s->offset;
303 #ifdef CONFIG_SLAB_FREELIST_HARDENED
304 BUG_ON(object == fp); /* naive detection of double free or corruption */
307 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
310 /* Loop over all objects in a slab */
311 #define for_each_object(__p, __s, __addr, __objects) \
312 for (__p = fixup_red_left(__s, __addr); \
313 __p < (__addr) + (__objects) * (__s)->size; \
316 /* Determine object index from a given position */
317 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
319 return (kasan_reset_tag(p) - addr) / s->size;
322 static inline unsigned int order_objects(unsigned int order, unsigned int size)
324 return ((unsigned int)PAGE_SIZE << order) / size;
327 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
330 struct kmem_cache_order_objects x = {
331 (order << OO_SHIFT) + order_objects(order, size)
337 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
339 return x.x >> OO_SHIFT;
342 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
344 return x.x & OO_MASK;
348 * Per slab locking using the pagelock
350 static __always_inline void slab_lock(struct page *page)
352 VM_BUG_ON_PAGE(PageTail(page), page);
353 bit_spin_lock(PG_locked, &page->flags);
356 static __always_inline void slab_unlock(struct page *page)
358 VM_BUG_ON_PAGE(PageTail(page), page);
359 __bit_spin_unlock(PG_locked, &page->flags);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
368 VM_BUG_ON(!irqs_disabled());
369 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
370 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
371 if (s->flags & __CMPXCHG_DOUBLE) {
372 if (cmpxchg_double(&page->freelist, &page->counters,
373 freelist_old, counters_old,
374 freelist_new, counters_new))
380 if (page->freelist == freelist_old &&
381 page->counters == counters_old) {
382 page->freelist = freelist_new;
383 page->counters = counters_new;
391 stat(s, CMPXCHG_DOUBLE_FAIL);
393 #ifdef SLUB_DEBUG_CMPXCHG
394 pr_info("%s %s: cmpxchg double redo ", n, s->name);
400 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
401 void *freelist_old, unsigned long counters_old,
402 void *freelist_new, unsigned long counters_new,
405 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
406 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
407 if (s->flags & __CMPXCHG_DOUBLE) {
408 if (cmpxchg_double(&page->freelist, &page->counters,
409 freelist_old, counters_old,
410 freelist_new, counters_new))
417 local_irq_save(flags);
419 if (page->freelist == freelist_old &&
420 page->counters == counters_old) {
421 page->freelist = freelist_new;
422 page->counters = counters_new;
424 local_irq_restore(flags);
428 local_irq_restore(flags);
432 stat(s, CMPXCHG_DOUBLE_FAIL);
434 #ifdef SLUB_DEBUG_CMPXCHG
435 pr_info("%s %s: cmpxchg double redo ", n, s->name);
441 #ifdef CONFIG_SLUB_DEBUG
442 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
443 static DEFINE_SPINLOCK(object_map_lock);
446 * Determine a map of object in use on a page.
448 * Node listlock must be held to guarantee that the page does
449 * not vanish from under us.
451 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
454 void *addr = page_address(page);
456 VM_BUG_ON(!irqs_disabled());
458 spin_lock(&object_map_lock);
460 bitmap_zero(object_map, page->objects);
462 for (p = page->freelist; p; p = get_freepointer(s, p))
463 set_bit(slab_index(p, s, addr), object_map);
468 static void put_map(unsigned long *map)
470 VM_BUG_ON(map != object_map);
471 lockdep_assert_held(&object_map_lock);
473 spin_unlock(&object_map_lock);
476 static inline unsigned int size_from_object(struct kmem_cache *s)
478 if (s->flags & SLAB_RED_ZONE)
479 return s->size - s->red_left_pad;
484 static inline void *restore_red_left(struct kmem_cache *s, void *p)
486 if (s->flags & SLAB_RED_ZONE)
487 p -= s->red_left_pad;
495 #if defined(CONFIG_SLUB_DEBUG_ON)
496 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
498 static slab_flags_t slub_debug;
501 static char *slub_debug_slabs;
502 static int disable_higher_order_debug;
505 * slub is about to manipulate internal object metadata. This memory lies
506 * outside the range of the allocated object, so accessing it would normally
507 * be reported by kasan as a bounds error. metadata_access_enable() is used
508 * to tell kasan that these accesses are OK.
510 static inline void metadata_access_enable(void)
512 kasan_disable_current();
515 static inline void metadata_access_disable(void)
517 kasan_enable_current();
524 /* Verify that a pointer has an address that is valid within a slab page */
525 static inline int check_valid_pointer(struct kmem_cache *s,
526 struct page *page, void *object)
533 base = page_address(page);
534 object = kasan_reset_tag(object);
535 object = restore_red_left(s, object);
536 if (object < base || object >= base + page->objects * s->size ||
537 (object - base) % s->size) {
544 static void print_section(char *level, char *text, u8 *addr,
547 metadata_access_enable();
548 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
550 metadata_access_disable();
553 static struct track *get_track(struct kmem_cache *s, void *object,
554 enum track_item alloc)
559 p = object + s->offset + sizeof(void *);
561 p = object + s->inuse;
566 static void set_track(struct kmem_cache *s, void *object,
567 enum track_item alloc, unsigned long addr)
569 struct track *p = get_track(s, object, alloc);
572 #ifdef CONFIG_STACKTRACE
573 unsigned int nr_entries;
575 metadata_access_enable();
576 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
577 metadata_access_disable();
579 if (nr_entries < TRACK_ADDRS_COUNT)
580 p->addrs[nr_entries] = 0;
583 p->cpu = smp_processor_id();
584 p->pid = current->pid;
587 memset(p, 0, sizeof(struct track));
591 static void init_tracking(struct kmem_cache *s, void *object)
593 if (!(s->flags & SLAB_STORE_USER))
596 set_track(s, object, TRACK_FREE, 0UL);
597 set_track(s, object, TRACK_ALLOC, 0UL);
600 static void print_track(const char *s, struct track *t, unsigned long pr_time)
605 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
606 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
607 #ifdef CONFIG_STACKTRACE
610 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
612 pr_err("\t%pS\n", (void *)t->addrs[i]);
619 static void print_tracking(struct kmem_cache *s, void *object)
621 unsigned long pr_time = jiffies;
622 if (!(s->flags & SLAB_STORE_USER))
625 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
626 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
629 static void print_page_info(struct page *page)
631 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
632 page, page->objects, page->inuse, page->freelist, page->flags);
636 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
638 struct va_format vaf;
644 pr_err("=============================================================================\n");
645 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
646 pr_err("-----------------------------------------------------------------------------\n\n");
648 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
652 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
654 struct va_format vaf;
660 pr_err("FIX %s: %pV\n", s->name, &vaf);
664 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
666 unsigned int off; /* Offset of last byte */
667 u8 *addr = page_address(page);
669 print_tracking(s, p);
671 print_page_info(page);
673 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
674 p, p - addr, get_freepointer(s, p));
676 if (s->flags & SLAB_RED_ZONE)
677 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
679 else if (p > addr + 16)
680 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
682 print_section(KERN_ERR, "Object ", p,
683 min_t(unsigned int, s->object_size, PAGE_SIZE));
684 if (s->flags & SLAB_RED_ZONE)
685 print_section(KERN_ERR, "Redzone ", p + s->object_size,
686 s->inuse - s->object_size);
689 off = s->offset + sizeof(void *);
693 if (s->flags & SLAB_STORE_USER)
694 off += 2 * sizeof(struct track);
696 off += kasan_metadata_size(s);
698 if (off != size_from_object(s))
699 /* Beginning of the filler is the free pointer */
700 print_section(KERN_ERR, "Padding ", p + off,
701 size_from_object(s) - off);
706 void object_err(struct kmem_cache *s, struct page *page,
707 u8 *object, char *reason)
709 slab_bug(s, "%s", reason);
710 print_trailer(s, page, object);
713 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
714 const char *fmt, ...)
720 vsnprintf(buf, sizeof(buf), fmt, args);
722 slab_bug(s, "%s", buf);
723 print_page_info(page);
727 static void init_object(struct kmem_cache *s, void *object, u8 val)
731 if (s->flags & SLAB_RED_ZONE)
732 memset(p - s->red_left_pad, val, s->red_left_pad);
734 if (s->flags & __OBJECT_POISON) {
735 memset(p, POISON_FREE, s->object_size - 1);
736 p[s->object_size - 1] = POISON_END;
739 if (s->flags & SLAB_RED_ZONE)
740 memset(p + s->object_size, val, s->inuse - s->object_size);
743 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
744 void *from, void *to)
746 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
747 memset(from, data, to - from);
750 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
751 u8 *object, char *what,
752 u8 *start, unsigned int value, unsigned int bytes)
756 u8 *addr = page_address(page);
758 metadata_access_enable();
759 fault = memchr_inv(start, value, bytes);
760 metadata_access_disable();
765 while (end > fault && end[-1] == value)
768 slab_bug(s, "%s overwritten", what);
769 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
770 fault, end - 1, fault - addr,
772 print_trailer(s, page, object);
774 restore_bytes(s, what, value, fault, end);
782 * Bytes of the object to be managed.
783 * If the freepointer may overlay the object then the free
784 * pointer is the first word of the object.
786 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
789 * object + s->object_size
790 * Padding to reach word boundary. This is also used for Redzoning.
791 * Padding is extended by another word if Redzoning is enabled and
792 * object_size == inuse.
794 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
795 * 0xcc (RED_ACTIVE) for objects in use.
798 * Meta data starts here.
800 * A. Free pointer (if we cannot overwrite object on free)
801 * B. Tracking data for SLAB_STORE_USER
802 * C. Padding to reach required alignment boundary or at mininum
803 * one word if debugging is on to be able to detect writes
804 * before the word boundary.
806 * Padding is done using 0x5a (POISON_INUSE)
809 * Nothing is used beyond s->size.
811 * If slabcaches are merged then the object_size and inuse boundaries are mostly
812 * ignored. And therefore no slab options that rely on these boundaries
813 * may be used with merged slabcaches.
816 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
818 unsigned long off = s->inuse; /* The end of info */
821 /* Freepointer is placed after the object. */
822 off += sizeof(void *);
824 if (s->flags & SLAB_STORE_USER)
825 /* We also have user information there */
826 off += 2 * sizeof(struct track);
828 off += kasan_metadata_size(s);
830 if (size_from_object(s) == off)
833 return check_bytes_and_report(s, page, p, "Object padding",
834 p + off, POISON_INUSE, size_from_object(s) - off);
837 /* Check the pad bytes at the end of a slab page */
838 static int slab_pad_check(struct kmem_cache *s, struct page *page)
847 if (!(s->flags & SLAB_POISON))
850 start = page_address(page);
851 length = page_size(page);
852 end = start + length;
853 remainder = length % s->size;
857 pad = end - remainder;
858 metadata_access_enable();
859 fault = memchr_inv(pad, POISON_INUSE, remainder);
860 metadata_access_disable();
863 while (end > fault && end[-1] == POISON_INUSE)
866 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
867 fault, end - 1, fault - start);
868 print_section(KERN_ERR, "Padding ", pad, remainder);
870 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
874 static int check_object(struct kmem_cache *s, struct page *page,
875 void *object, u8 val)
878 u8 *endobject = object + s->object_size;
880 if (s->flags & SLAB_RED_ZONE) {
881 if (!check_bytes_and_report(s, page, object, "Redzone",
882 object - s->red_left_pad, val, s->red_left_pad))
885 if (!check_bytes_and_report(s, page, object, "Redzone",
886 endobject, val, s->inuse - s->object_size))
889 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
890 check_bytes_and_report(s, page, p, "Alignment padding",
891 endobject, POISON_INUSE,
892 s->inuse - s->object_size);
896 if (s->flags & SLAB_POISON) {
897 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
898 (!check_bytes_and_report(s, page, p, "Poison", p,
899 POISON_FREE, s->object_size - 1) ||
900 !check_bytes_and_report(s, page, p, "Poison",
901 p + s->object_size - 1, POISON_END, 1)))
904 * check_pad_bytes cleans up on its own.
906 check_pad_bytes(s, page, p);
909 if (!s->offset && val == SLUB_RED_ACTIVE)
911 * Object and freepointer overlap. Cannot check
912 * freepointer while object is allocated.
916 /* Check free pointer validity */
917 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
918 object_err(s, page, p, "Freepointer corrupt");
920 * No choice but to zap it and thus lose the remainder
921 * of the free objects in this slab. May cause
922 * another error because the object count is now wrong.
924 set_freepointer(s, p, NULL);
930 static int check_slab(struct kmem_cache *s, struct page *page)
934 VM_BUG_ON(!irqs_disabled());
936 if (!PageSlab(page)) {
937 slab_err(s, page, "Not a valid slab page");
941 maxobj = order_objects(compound_order(page), s->size);
942 if (page->objects > maxobj) {
943 slab_err(s, page, "objects %u > max %u",
944 page->objects, maxobj);
947 if (page->inuse > page->objects) {
948 slab_err(s, page, "inuse %u > max %u",
949 page->inuse, page->objects);
952 /* Slab_pad_check fixes things up after itself */
953 slab_pad_check(s, page);
958 * Determine if a certain object on a page is on the freelist. Must hold the
959 * slab lock to guarantee that the chains are in a consistent state.
961 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
969 while (fp && nr <= page->objects) {
972 if (!check_valid_pointer(s, page, fp)) {
974 object_err(s, page, object,
975 "Freechain corrupt");
976 set_freepointer(s, object, NULL);
978 slab_err(s, page, "Freepointer corrupt");
979 page->freelist = NULL;
980 page->inuse = page->objects;
981 slab_fix(s, "Freelist cleared");
987 fp = get_freepointer(s, object);
991 max_objects = order_objects(compound_order(page), s->size);
992 if (max_objects > MAX_OBJS_PER_PAGE)
993 max_objects = MAX_OBJS_PER_PAGE;
995 if (page->objects != max_objects) {
996 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
997 page->objects, max_objects);
998 page->objects = max_objects;
999 slab_fix(s, "Number of objects adjusted.");
1001 if (page->inuse != page->objects - nr) {
1002 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1003 page->inuse, page->objects - nr);
1004 page->inuse = page->objects - nr;
1005 slab_fix(s, "Object count adjusted.");
1007 return search == NULL;
1010 static void trace(struct kmem_cache *s, struct page *page, void *object,
1013 if (s->flags & SLAB_TRACE) {
1014 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1016 alloc ? "alloc" : "free",
1017 object, page->inuse,
1021 print_section(KERN_INFO, "Object ", (void *)object,
1029 * Tracking of fully allocated slabs for debugging purposes.
1031 static void add_full(struct kmem_cache *s,
1032 struct kmem_cache_node *n, struct page *page)
1034 if (!(s->flags & SLAB_STORE_USER))
1037 lockdep_assert_held(&n->list_lock);
1038 list_add(&page->slab_list, &n->full);
1041 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1043 if (!(s->flags & SLAB_STORE_USER))
1046 lockdep_assert_held(&n->list_lock);
1047 list_del(&page->slab_list);
1050 /* Tracking of the number of slabs for debugging purposes */
1051 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1053 struct kmem_cache_node *n = get_node(s, node);
1055 return atomic_long_read(&n->nr_slabs);
1058 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1060 return atomic_long_read(&n->nr_slabs);
1063 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1065 struct kmem_cache_node *n = get_node(s, node);
1068 * May be called early in order to allocate a slab for the
1069 * kmem_cache_node structure. Solve the chicken-egg
1070 * dilemma by deferring the increment of the count during
1071 * bootstrap (see early_kmem_cache_node_alloc).
1074 atomic_long_inc(&n->nr_slabs);
1075 atomic_long_add(objects, &n->total_objects);
1078 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1080 struct kmem_cache_node *n = get_node(s, node);
1082 atomic_long_dec(&n->nr_slabs);
1083 atomic_long_sub(objects, &n->total_objects);
1086 /* Object debug checks for alloc/free paths */
1087 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1090 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1093 init_object(s, object, SLUB_RED_INACTIVE);
1094 init_tracking(s, object);
1098 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1100 if (!(s->flags & SLAB_POISON))
1103 metadata_access_enable();
1104 memset(addr, POISON_INUSE, page_size(page));
1105 metadata_access_disable();
1108 static inline int alloc_consistency_checks(struct kmem_cache *s,
1109 struct page *page, void *object)
1111 if (!check_slab(s, page))
1114 if (!check_valid_pointer(s, page, object)) {
1115 object_err(s, page, object, "Freelist Pointer check fails");
1119 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1125 static noinline int alloc_debug_processing(struct kmem_cache *s,
1127 void *object, unsigned long addr)
1129 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1130 if (!alloc_consistency_checks(s, page, object))
1134 /* Success perform special debug activities for allocs */
1135 if (s->flags & SLAB_STORE_USER)
1136 set_track(s, object, TRACK_ALLOC, addr);
1137 trace(s, page, object, 1);
1138 init_object(s, object, SLUB_RED_ACTIVE);
1142 if (PageSlab(page)) {
1144 * If this is a slab page then lets do the best we can
1145 * to avoid issues in the future. Marking all objects
1146 * as used avoids touching the remaining objects.
1148 slab_fix(s, "Marking all objects used");
1149 page->inuse = page->objects;
1150 page->freelist = NULL;
1155 static inline int free_consistency_checks(struct kmem_cache *s,
1156 struct page *page, void *object, unsigned long addr)
1158 if (!check_valid_pointer(s, page, object)) {
1159 slab_err(s, page, "Invalid object pointer 0x%p", object);
1163 if (on_freelist(s, page, object)) {
1164 object_err(s, page, object, "Object already free");
1168 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1171 if (unlikely(s != page->slab_cache)) {
1172 if (!PageSlab(page)) {
1173 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1175 } else if (!page->slab_cache) {
1176 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1180 object_err(s, page, object,
1181 "page slab pointer corrupt.");
1187 /* Supports checking bulk free of a constructed freelist */
1188 static noinline int free_debug_processing(
1189 struct kmem_cache *s, struct page *page,
1190 void *head, void *tail, int bulk_cnt,
1193 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1194 void *object = head;
1196 unsigned long uninitialized_var(flags);
1199 spin_lock_irqsave(&n->list_lock, flags);
1202 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1203 if (!check_slab(s, page))
1210 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1211 if (!free_consistency_checks(s, page, object, addr))
1215 if (s->flags & SLAB_STORE_USER)
1216 set_track(s, object, TRACK_FREE, addr);
1217 trace(s, page, object, 0);
1218 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1219 init_object(s, object, SLUB_RED_INACTIVE);
1221 /* Reached end of constructed freelist yet? */
1222 if (object != tail) {
1223 object = get_freepointer(s, object);
1229 if (cnt != bulk_cnt)
1230 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1234 spin_unlock_irqrestore(&n->list_lock, flags);
1236 slab_fix(s, "Object at 0x%p not freed", object);
1240 static int __init setup_slub_debug(char *str)
1242 slub_debug = DEBUG_DEFAULT_FLAGS;
1243 if (*str++ != '=' || !*str)
1245 * No options specified. Switch on full debugging.
1251 * No options but restriction on slabs. This means full
1252 * debugging for slabs matching a pattern.
1259 * Switch off all debugging measures.
1264 * Determine which debug features should be switched on
1266 for (; *str && *str != ','; str++) {
1267 switch (tolower(*str)) {
1269 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1272 slub_debug |= SLAB_RED_ZONE;
1275 slub_debug |= SLAB_POISON;
1278 slub_debug |= SLAB_STORE_USER;
1281 slub_debug |= SLAB_TRACE;
1284 slub_debug |= SLAB_FAILSLAB;
1288 * Avoid enabling debugging on caches if its minimum
1289 * order would increase as a result.
1291 disable_higher_order_debug = 1;
1294 pr_err("slub_debug option '%c' unknown. skipped\n",
1301 slub_debug_slabs = str + 1;
1303 if ((static_branch_unlikely(&init_on_alloc) ||
1304 static_branch_unlikely(&init_on_free)) &&
1305 (slub_debug & SLAB_POISON))
1306 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1310 __setup("slub_debug", setup_slub_debug);
1313 * kmem_cache_flags - apply debugging options to the cache
1314 * @object_size: the size of an object without meta data
1315 * @flags: flags to set
1316 * @name: name of the cache
1317 * @ctor: constructor function
1319 * Debug option(s) are applied to @flags. In addition to the debug
1320 * option(s), if a slab name (or multiple) is specified i.e.
1321 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1322 * then only the select slabs will receive the debug option(s).
1324 slab_flags_t kmem_cache_flags(unsigned int object_size,
1325 slab_flags_t flags, const char *name,
1326 void (*ctor)(void *))
1331 /* If slub_debug = 0, it folds into the if conditional. */
1332 if (!slub_debug_slabs)
1333 return flags | slub_debug;
1336 iter = slub_debug_slabs;
1341 end = strchrnul(iter, ',');
1343 glob = strnchr(iter, end - iter, '*');
1345 cmplen = glob - iter;
1347 cmplen = max_t(size_t, len, (end - iter));
1349 if (!strncmp(name, iter, cmplen)) {
1350 flags |= slub_debug;
1361 #else /* !CONFIG_SLUB_DEBUG */
1362 static inline void setup_object_debug(struct kmem_cache *s,
1363 struct page *page, void *object) {}
1365 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1367 static inline int alloc_debug_processing(struct kmem_cache *s,
1368 struct page *page, void *object, unsigned long addr) { return 0; }
1370 static inline int free_debug_processing(
1371 struct kmem_cache *s, struct page *page,
1372 void *head, void *tail, int bulk_cnt,
1373 unsigned long addr) { return 0; }
1375 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1377 static inline int check_object(struct kmem_cache *s, struct page *page,
1378 void *object, u8 val) { return 1; }
1379 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1380 struct page *page) {}
1381 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1382 struct page *page) {}
1383 slab_flags_t kmem_cache_flags(unsigned int object_size,
1384 slab_flags_t flags, const char *name,
1385 void (*ctor)(void *))
1389 #define slub_debug 0
1391 #define disable_higher_order_debug 0
1393 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1395 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1397 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1399 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1402 #endif /* CONFIG_SLUB_DEBUG */
1405 * Hooks for other subsystems that check memory allocations. In a typical
1406 * production configuration these hooks all should produce no code at all.
1408 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1410 ptr = kasan_kmalloc_large(ptr, size, flags);
1411 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1412 kmemleak_alloc(ptr, size, 1, flags);
1416 static __always_inline void kfree_hook(void *x)
1419 kasan_kfree_large(x, _RET_IP_);
1422 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1424 kmemleak_free_recursive(x, s->flags);
1427 * Trouble is that we may no longer disable interrupts in the fast path
1428 * So in order to make the debug calls that expect irqs to be
1429 * disabled we need to disable interrupts temporarily.
1431 #ifdef CONFIG_LOCKDEP
1433 unsigned long flags;
1435 local_irq_save(flags);
1436 debug_check_no_locks_freed(x, s->object_size);
1437 local_irq_restore(flags);
1440 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1441 debug_check_no_obj_freed(x, s->object_size);
1443 /* KASAN might put x into memory quarantine, delaying its reuse */
1444 return kasan_slab_free(s, x, _RET_IP_);
1447 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1448 void **head, void **tail)
1453 void *old_tail = *tail ? *tail : *head;
1456 /* Head and tail of the reconstructed freelist */
1462 next = get_freepointer(s, object);
1464 if (slab_want_init_on_free(s)) {
1466 * Clear the object and the metadata, but don't touch
1469 memset(object, 0, s->object_size);
1470 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1472 memset((char *)object + s->inuse, 0,
1473 s->size - s->inuse - rsize);
1476 /* If object's reuse doesn't have to be delayed */
1477 if (!slab_free_hook(s, object)) {
1478 /* Move object to the new freelist */
1479 set_freepointer(s, object, *head);
1484 } while (object != old_tail);
1489 return *head != NULL;
1492 static void *setup_object(struct kmem_cache *s, struct page *page,
1495 setup_object_debug(s, page, object);
1496 object = kasan_init_slab_obj(s, object);
1497 if (unlikely(s->ctor)) {
1498 kasan_unpoison_object_data(s, object);
1500 kasan_poison_object_data(s, object);
1506 * Slab allocation and freeing
1508 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1509 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1512 unsigned int order = oo_order(oo);
1514 if (node == NUMA_NO_NODE)
1515 page = alloc_pages(flags, order);
1517 page = __alloc_pages_node(node, flags, order);
1519 if (page && charge_slab_page(page, flags, order, s)) {
1520 __free_pages(page, order);
1527 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1528 /* Pre-initialize the random sequence cache */
1529 static int init_cache_random_seq(struct kmem_cache *s)
1531 unsigned int count = oo_objects(s->oo);
1534 /* Bailout if already initialised */
1538 err = cache_random_seq_create(s, count, GFP_KERNEL);
1540 pr_err("SLUB: Unable to initialize free list for %s\n",
1545 /* Transform to an offset on the set of pages */
1546 if (s->random_seq) {
1549 for (i = 0; i < count; i++)
1550 s->random_seq[i] *= s->size;
1555 /* Initialize each random sequence freelist per cache */
1556 static void __init init_freelist_randomization(void)
1558 struct kmem_cache *s;
1560 mutex_lock(&slab_mutex);
1562 list_for_each_entry(s, &slab_caches, list)
1563 init_cache_random_seq(s);
1565 mutex_unlock(&slab_mutex);
1568 /* Get the next entry on the pre-computed freelist randomized */
1569 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1570 unsigned long *pos, void *start,
1571 unsigned long page_limit,
1572 unsigned long freelist_count)
1577 * If the target page allocation failed, the number of objects on the
1578 * page might be smaller than the usual size defined by the cache.
1581 idx = s->random_seq[*pos];
1583 if (*pos >= freelist_count)
1585 } while (unlikely(idx >= page_limit));
1587 return (char *)start + idx;
1590 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1591 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1596 unsigned long idx, pos, page_limit, freelist_count;
1598 if (page->objects < 2 || !s->random_seq)
1601 freelist_count = oo_objects(s->oo);
1602 pos = get_random_int() % freelist_count;
1604 page_limit = page->objects * s->size;
1605 start = fixup_red_left(s, page_address(page));
1607 /* First entry is used as the base of the freelist */
1608 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1610 cur = setup_object(s, page, cur);
1611 page->freelist = cur;
1613 for (idx = 1; idx < page->objects; idx++) {
1614 next = next_freelist_entry(s, page, &pos, start, page_limit,
1616 next = setup_object(s, page, next);
1617 set_freepointer(s, cur, next);
1620 set_freepointer(s, cur, NULL);
1625 static inline int init_cache_random_seq(struct kmem_cache *s)
1629 static inline void init_freelist_randomization(void) { }
1630 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1634 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1636 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1639 struct kmem_cache_order_objects oo = s->oo;
1641 void *start, *p, *next;
1645 flags &= gfp_allowed_mask;
1647 if (gfpflags_allow_blocking(flags))
1650 flags |= s->allocflags;
1653 * Let the initial higher-order allocation fail under memory pressure
1654 * so we fall-back to the minimum order allocation.
1656 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1657 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1658 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1660 page = alloc_slab_page(s, alloc_gfp, node, oo);
1661 if (unlikely(!page)) {
1665 * Allocation may have failed due to fragmentation.
1666 * Try a lower order alloc if possible
1668 page = alloc_slab_page(s, alloc_gfp, node, oo);
1669 if (unlikely(!page))
1671 stat(s, ORDER_FALLBACK);
1674 page->objects = oo_objects(oo);
1676 page->slab_cache = s;
1677 __SetPageSlab(page);
1678 if (page_is_pfmemalloc(page))
1679 SetPageSlabPfmemalloc(page);
1681 kasan_poison_slab(page);
1683 start = page_address(page);
1685 setup_page_debug(s, page, start);
1687 shuffle = shuffle_freelist(s, page);
1690 start = fixup_red_left(s, start);
1691 start = setup_object(s, page, start);
1692 page->freelist = start;
1693 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1695 next = setup_object(s, page, next);
1696 set_freepointer(s, p, next);
1699 set_freepointer(s, p, NULL);
1702 page->inuse = page->objects;
1706 if (gfpflags_allow_blocking(flags))
1707 local_irq_disable();
1711 inc_slabs_node(s, page_to_nid(page), page->objects);
1716 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1718 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1719 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1720 flags &= ~GFP_SLAB_BUG_MASK;
1721 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1722 invalid_mask, &invalid_mask, flags, &flags);
1726 return allocate_slab(s,
1727 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1730 static void __free_slab(struct kmem_cache *s, struct page *page)
1732 int order = compound_order(page);
1733 int pages = 1 << order;
1735 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1738 slab_pad_check(s, page);
1739 for_each_object(p, s, page_address(page),
1741 check_object(s, page, p, SLUB_RED_INACTIVE);
1744 __ClearPageSlabPfmemalloc(page);
1745 __ClearPageSlab(page);
1747 page->mapping = NULL;
1748 if (current->reclaim_state)
1749 current->reclaim_state->reclaimed_slab += pages;
1750 uncharge_slab_page(page, order, s);
1751 __free_pages(page, order);
1754 static void rcu_free_slab(struct rcu_head *h)
1756 struct page *page = container_of(h, struct page, rcu_head);
1758 __free_slab(page->slab_cache, page);
1761 static void free_slab(struct kmem_cache *s, struct page *page)
1763 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1764 call_rcu(&page->rcu_head, rcu_free_slab);
1766 __free_slab(s, page);
1769 static void discard_slab(struct kmem_cache *s, struct page *page)
1771 dec_slabs_node(s, page_to_nid(page), page->objects);
1776 * Management of partially allocated slabs.
1779 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1782 if (tail == DEACTIVATE_TO_TAIL)
1783 list_add_tail(&page->slab_list, &n->partial);
1785 list_add(&page->slab_list, &n->partial);
1788 static inline void add_partial(struct kmem_cache_node *n,
1789 struct page *page, int tail)
1791 lockdep_assert_held(&n->list_lock);
1792 __add_partial(n, page, tail);
1795 static inline void remove_partial(struct kmem_cache_node *n,
1798 lockdep_assert_held(&n->list_lock);
1799 list_del(&page->slab_list);
1804 * Remove slab from the partial list, freeze it and
1805 * return the pointer to the freelist.
1807 * Returns a list of objects or NULL if it fails.
1809 static inline void *acquire_slab(struct kmem_cache *s,
1810 struct kmem_cache_node *n, struct page *page,
1811 int mode, int *objects)
1814 unsigned long counters;
1817 lockdep_assert_held(&n->list_lock);
1820 * Zap the freelist and set the frozen bit.
1821 * The old freelist is the list of objects for the
1822 * per cpu allocation list.
1824 freelist = page->freelist;
1825 counters = page->counters;
1826 new.counters = counters;
1827 *objects = new.objects - new.inuse;
1829 new.inuse = page->objects;
1830 new.freelist = NULL;
1832 new.freelist = freelist;
1835 VM_BUG_ON(new.frozen);
1838 if (!__cmpxchg_double_slab(s, page,
1840 new.freelist, new.counters,
1844 remove_partial(n, page);
1849 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1850 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1853 * Try to allocate a partial slab from a specific node.
1855 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1856 struct kmem_cache_cpu *c, gfp_t flags)
1858 struct page *page, *page2;
1859 void *object = NULL;
1860 unsigned int available = 0;
1864 * Racy check. If we mistakenly see no partial slabs then we
1865 * just allocate an empty slab. If we mistakenly try to get a
1866 * partial slab and there is none available then get_partials()
1869 if (!n || !n->nr_partial)
1872 spin_lock(&n->list_lock);
1873 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1876 if (!pfmemalloc_match(page, flags))
1879 t = acquire_slab(s, n, page, object == NULL, &objects);
1883 available += objects;
1886 stat(s, ALLOC_FROM_PARTIAL);
1889 put_cpu_partial(s, page, 0);
1890 stat(s, CPU_PARTIAL_NODE);
1892 if (!kmem_cache_has_cpu_partial(s)
1893 || available > slub_cpu_partial(s) / 2)
1897 spin_unlock(&n->list_lock);
1902 * Get a page from somewhere. Search in increasing NUMA distances.
1904 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1905 struct kmem_cache_cpu *c)
1908 struct zonelist *zonelist;
1911 enum zone_type high_zoneidx = gfp_zone(flags);
1913 unsigned int cpuset_mems_cookie;
1916 * The defrag ratio allows a configuration of the tradeoffs between
1917 * inter node defragmentation and node local allocations. A lower
1918 * defrag_ratio increases the tendency to do local allocations
1919 * instead of attempting to obtain partial slabs from other nodes.
1921 * If the defrag_ratio is set to 0 then kmalloc() always
1922 * returns node local objects. If the ratio is higher then kmalloc()
1923 * may return off node objects because partial slabs are obtained
1924 * from other nodes and filled up.
1926 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1927 * (which makes defrag_ratio = 1000) then every (well almost)
1928 * allocation will first attempt to defrag slab caches on other nodes.
1929 * This means scanning over all nodes to look for partial slabs which
1930 * may be expensive if we do it every time we are trying to find a slab
1931 * with available objects.
1933 if (!s->remote_node_defrag_ratio ||
1934 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1938 cpuset_mems_cookie = read_mems_allowed_begin();
1939 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1940 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1941 struct kmem_cache_node *n;
1943 n = get_node(s, zone_to_nid(zone));
1945 if (n && cpuset_zone_allowed(zone, flags) &&
1946 n->nr_partial > s->min_partial) {
1947 object = get_partial_node(s, n, c, flags);
1950 * Don't check read_mems_allowed_retry()
1951 * here - if mems_allowed was updated in
1952 * parallel, that was a harmless race
1953 * between allocation and the cpuset
1960 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1961 #endif /* CONFIG_NUMA */
1966 * Get a partial page, lock it and return it.
1968 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1969 struct kmem_cache_cpu *c)
1972 int searchnode = node;
1974 if (node == NUMA_NO_NODE)
1975 searchnode = numa_mem_id();
1977 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1978 if (object || node != NUMA_NO_NODE)
1981 return get_any_partial(s, flags, c);
1984 #ifdef CONFIG_PREEMPTION
1986 * Calculate the next globally unique transaction for disambiguiation
1987 * during cmpxchg. The transactions start with the cpu number and are then
1988 * incremented by CONFIG_NR_CPUS.
1990 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1993 * No preemption supported therefore also no need to check for
1999 static inline unsigned long next_tid(unsigned long tid)
2001 return tid + TID_STEP;
2004 #ifdef SLUB_DEBUG_CMPXCHG
2005 static inline unsigned int tid_to_cpu(unsigned long tid)
2007 return tid % TID_STEP;
2010 static inline unsigned long tid_to_event(unsigned long tid)
2012 return tid / TID_STEP;
2016 static inline unsigned int init_tid(int cpu)
2021 static inline void note_cmpxchg_failure(const char *n,
2022 const struct kmem_cache *s, unsigned long tid)
2024 #ifdef SLUB_DEBUG_CMPXCHG
2025 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2027 pr_info("%s %s: cmpxchg redo ", n, s->name);
2029 #ifdef CONFIG_PREEMPTION
2030 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2031 pr_warn("due to cpu change %d -> %d\n",
2032 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2035 if (tid_to_event(tid) != tid_to_event(actual_tid))
2036 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2037 tid_to_event(tid), tid_to_event(actual_tid));
2039 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2040 actual_tid, tid, next_tid(tid));
2042 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2045 static void init_kmem_cache_cpus(struct kmem_cache *s)
2049 for_each_possible_cpu(cpu)
2050 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2054 * Remove the cpu slab
2056 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2057 void *freelist, struct kmem_cache_cpu *c)
2059 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2060 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2062 enum slab_modes l = M_NONE, m = M_NONE;
2064 int tail = DEACTIVATE_TO_HEAD;
2068 if (page->freelist) {
2069 stat(s, DEACTIVATE_REMOTE_FREES);
2070 tail = DEACTIVATE_TO_TAIL;
2074 * Stage one: Free all available per cpu objects back
2075 * to the page freelist while it is still frozen. Leave the
2078 * There is no need to take the list->lock because the page
2081 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2083 unsigned long counters;
2086 prior = page->freelist;
2087 counters = page->counters;
2088 set_freepointer(s, freelist, prior);
2089 new.counters = counters;
2091 VM_BUG_ON(!new.frozen);
2093 } while (!__cmpxchg_double_slab(s, page,
2095 freelist, new.counters,
2096 "drain percpu freelist"));
2098 freelist = nextfree;
2102 * Stage two: Ensure that the page is unfrozen while the
2103 * list presence reflects the actual number of objects
2106 * We setup the list membership and then perform a cmpxchg
2107 * with the count. If there is a mismatch then the page
2108 * is not unfrozen but the page is on the wrong list.
2110 * Then we restart the process which may have to remove
2111 * the page from the list that we just put it on again
2112 * because the number of objects in the slab may have
2117 old.freelist = page->freelist;
2118 old.counters = page->counters;
2119 VM_BUG_ON(!old.frozen);
2121 /* Determine target state of the slab */
2122 new.counters = old.counters;
2125 set_freepointer(s, freelist, old.freelist);
2126 new.freelist = freelist;
2128 new.freelist = old.freelist;
2132 if (!new.inuse && n->nr_partial >= s->min_partial)
2134 else if (new.freelist) {
2139 * Taking the spinlock removes the possibility
2140 * that acquire_slab() will see a slab page that
2143 spin_lock(&n->list_lock);
2147 if (kmem_cache_debug(s) && !lock) {
2150 * This also ensures that the scanning of full
2151 * slabs from diagnostic functions will not see
2154 spin_lock(&n->list_lock);
2160 remove_partial(n, page);
2161 else if (l == M_FULL)
2162 remove_full(s, n, page);
2165 add_partial(n, page, tail);
2166 else if (m == M_FULL)
2167 add_full(s, n, page);
2171 if (!__cmpxchg_double_slab(s, page,
2172 old.freelist, old.counters,
2173 new.freelist, new.counters,
2178 spin_unlock(&n->list_lock);
2182 else if (m == M_FULL)
2183 stat(s, DEACTIVATE_FULL);
2184 else if (m == M_FREE) {
2185 stat(s, DEACTIVATE_EMPTY);
2186 discard_slab(s, page);
2195 * Unfreeze all the cpu partial slabs.
2197 * This function must be called with interrupts disabled
2198 * for the cpu using c (or some other guarantee must be there
2199 * to guarantee no concurrent accesses).
2201 static void unfreeze_partials(struct kmem_cache *s,
2202 struct kmem_cache_cpu *c)
2204 #ifdef CONFIG_SLUB_CPU_PARTIAL
2205 struct kmem_cache_node *n = NULL, *n2 = NULL;
2206 struct page *page, *discard_page = NULL;
2208 while ((page = slub_percpu_partial(c))) {
2212 slub_set_percpu_partial(c, page);
2214 n2 = get_node(s, page_to_nid(page));
2217 spin_unlock(&n->list_lock);
2220 spin_lock(&n->list_lock);
2225 old.freelist = page->freelist;
2226 old.counters = page->counters;
2227 VM_BUG_ON(!old.frozen);
2229 new.counters = old.counters;
2230 new.freelist = old.freelist;
2234 } while (!__cmpxchg_double_slab(s, page,
2235 old.freelist, old.counters,
2236 new.freelist, new.counters,
2237 "unfreezing slab"));
2239 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2240 page->next = discard_page;
2241 discard_page = page;
2243 add_partial(n, page, DEACTIVATE_TO_TAIL);
2244 stat(s, FREE_ADD_PARTIAL);
2249 spin_unlock(&n->list_lock);
2251 while (discard_page) {
2252 page = discard_page;
2253 discard_page = discard_page->next;
2255 stat(s, DEACTIVATE_EMPTY);
2256 discard_slab(s, page);
2259 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2263 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2264 * partial page slot if available.
2266 * If we did not find a slot then simply move all the partials to the
2267 * per node partial list.
2269 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2271 #ifdef CONFIG_SLUB_CPU_PARTIAL
2272 struct page *oldpage;
2280 oldpage = this_cpu_read(s->cpu_slab->partial);
2283 pobjects = oldpage->pobjects;
2284 pages = oldpage->pages;
2285 if (drain && pobjects > slub_cpu_partial(s)) {
2286 unsigned long flags;
2288 * partial array is full. Move the existing
2289 * set to the per node partial list.
2291 local_irq_save(flags);
2292 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2293 local_irq_restore(flags);
2297 stat(s, CPU_PARTIAL_DRAIN);
2302 pobjects += page->objects - page->inuse;
2304 page->pages = pages;
2305 page->pobjects = pobjects;
2306 page->next = oldpage;
2308 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2310 if (unlikely(!slub_cpu_partial(s))) {
2311 unsigned long flags;
2313 local_irq_save(flags);
2314 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2315 local_irq_restore(flags);
2318 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2321 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2323 stat(s, CPUSLAB_FLUSH);
2324 deactivate_slab(s, c->page, c->freelist, c);
2326 c->tid = next_tid(c->tid);
2332 * Called from IPI handler with interrupts disabled.
2334 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2336 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2341 unfreeze_partials(s, c);
2344 static void flush_cpu_slab(void *d)
2346 struct kmem_cache *s = d;
2348 __flush_cpu_slab(s, smp_processor_id());
2351 static bool has_cpu_slab(int cpu, void *info)
2353 struct kmem_cache *s = info;
2354 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2356 return c->page || slub_percpu_partial(c);
2359 static void flush_all(struct kmem_cache *s)
2361 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2365 * Use the cpu notifier to insure that the cpu slabs are flushed when
2368 static int slub_cpu_dead(unsigned int cpu)
2370 struct kmem_cache *s;
2371 unsigned long flags;
2373 mutex_lock(&slab_mutex);
2374 list_for_each_entry(s, &slab_caches, list) {
2375 local_irq_save(flags);
2376 __flush_cpu_slab(s, cpu);
2377 local_irq_restore(flags);
2379 mutex_unlock(&slab_mutex);
2384 * Check if the objects in a per cpu structure fit numa
2385 * locality expectations.
2387 static inline int node_match(struct page *page, int node)
2390 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2396 #ifdef CONFIG_SLUB_DEBUG
2397 static int count_free(struct page *page)
2399 return page->objects - page->inuse;
2402 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2404 return atomic_long_read(&n->total_objects);
2406 #endif /* CONFIG_SLUB_DEBUG */
2408 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2409 static unsigned long count_partial(struct kmem_cache_node *n,
2410 int (*get_count)(struct page *))
2412 unsigned long flags;
2413 unsigned long x = 0;
2416 spin_lock_irqsave(&n->list_lock, flags);
2417 list_for_each_entry(page, &n->partial, slab_list)
2418 x += get_count(page);
2419 spin_unlock_irqrestore(&n->list_lock, flags);
2422 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2424 static noinline void
2425 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2427 #ifdef CONFIG_SLUB_DEBUG
2428 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2429 DEFAULT_RATELIMIT_BURST);
2431 struct kmem_cache_node *n;
2433 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2436 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2437 nid, gfpflags, &gfpflags);
2438 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2439 s->name, s->object_size, s->size, oo_order(s->oo),
2442 if (oo_order(s->min) > get_order(s->object_size))
2443 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2446 for_each_kmem_cache_node(s, node, n) {
2447 unsigned long nr_slabs;
2448 unsigned long nr_objs;
2449 unsigned long nr_free;
2451 nr_free = count_partial(n, count_free);
2452 nr_slabs = node_nr_slabs(n);
2453 nr_objs = node_nr_objs(n);
2455 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2456 node, nr_slabs, nr_objs, nr_free);
2461 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2462 int node, struct kmem_cache_cpu **pc)
2465 struct kmem_cache_cpu *c = *pc;
2468 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2470 freelist = get_partial(s, flags, node, c);
2475 page = new_slab(s, flags, node);
2477 c = raw_cpu_ptr(s->cpu_slab);
2482 * No other reference to the page yet so we can
2483 * muck around with it freely without cmpxchg
2485 freelist = page->freelist;
2486 page->freelist = NULL;
2488 stat(s, ALLOC_SLAB);
2496 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2498 if (unlikely(PageSlabPfmemalloc(page)))
2499 return gfp_pfmemalloc_allowed(gfpflags);
2505 * Check the page->freelist of a page and either transfer the freelist to the
2506 * per cpu freelist or deactivate the page.
2508 * The page is still frozen if the return value is not NULL.
2510 * If this function returns NULL then the page has been unfrozen.
2512 * This function must be called with interrupt disabled.
2514 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2517 unsigned long counters;
2521 freelist = page->freelist;
2522 counters = page->counters;
2524 new.counters = counters;
2525 VM_BUG_ON(!new.frozen);
2527 new.inuse = page->objects;
2528 new.frozen = freelist != NULL;
2530 } while (!__cmpxchg_double_slab(s, page,
2539 * Slow path. The lockless freelist is empty or we need to perform
2542 * Processing is still very fast if new objects have been freed to the
2543 * regular freelist. In that case we simply take over the regular freelist
2544 * as the lockless freelist and zap the regular freelist.
2546 * If that is not working then we fall back to the partial lists. We take the
2547 * first element of the freelist as the object to allocate now and move the
2548 * rest of the freelist to the lockless freelist.
2550 * And if we were unable to get a new slab from the partial slab lists then
2551 * we need to allocate a new slab. This is the slowest path since it involves
2552 * a call to the page allocator and the setup of a new slab.
2554 * Version of __slab_alloc to use when we know that interrupts are
2555 * already disabled (which is the case for bulk allocation).
2557 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2558 unsigned long addr, struct kmem_cache_cpu *c)
2566 * if the node is not online or has no normal memory, just
2567 * ignore the node constraint
2569 if (unlikely(node != NUMA_NO_NODE &&
2570 !node_state(node, N_NORMAL_MEMORY)))
2571 node = NUMA_NO_NODE;
2576 if (unlikely(!node_match(page, node))) {
2578 * same as above but node_match() being false already
2579 * implies node != NUMA_NO_NODE
2581 if (!node_state(node, N_NORMAL_MEMORY)) {
2582 node = NUMA_NO_NODE;
2585 stat(s, ALLOC_NODE_MISMATCH);
2586 deactivate_slab(s, page, c->freelist, c);
2592 * By rights, we should be searching for a slab page that was
2593 * PFMEMALLOC but right now, we are losing the pfmemalloc
2594 * information when the page leaves the per-cpu allocator
2596 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2597 deactivate_slab(s, page, c->freelist, c);
2601 /* must check again c->freelist in case of cpu migration or IRQ */
2602 freelist = c->freelist;
2606 freelist = get_freelist(s, page);
2610 stat(s, DEACTIVATE_BYPASS);
2614 stat(s, ALLOC_REFILL);
2618 * freelist is pointing to the list of objects to be used.
2619 * page is pointing to the page from which the objects are obtained.
2620 * That page must be frozen for per cpu allocations to work.
2622 VM_BUG_ON(!c->page->frozen);
2623 c->freelist = get_freepointer(s, freelist);
2624 c->tid = next_tid(c->tid);
2629 if (slub_percpu_partial(c)) {
2630 page = c->page = slub_percpu_partial(c);
2631 slub_set_percpu_partial(c, page);
2632 stat(s, CPU_PARTIAL_ALLOC);
2636 freelist = new_slab_objects(s, gfpflags, node, &c);
2638 if (unlikely(!freelist)) {
2639 slab_out_of_memory(s, gfpflags, node);
2644 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2647 /* Only entered in the debug case */
2648 if (kmem_cache_debug(s) &&
2649 !alloc_debug_processing(s, page, freelist, addr))
2650 goto new_slab; /* Slab failed checks. Next slab needed */
2652 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2657 * Another one that disabled interrupt and compensates for possible
2658 * cpu changes by refetching the per cpu area pointer.
2660 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2661 unsigned long addr, struct kmem_cache_cpu *c)
2664 unsigned long flags;
2666 local_irq_save(flags);
2667 #ifdef CONFIG_PREEMPTION
2669 * We may have been preempted and rescheduled on a different
2670 * cpu before disabling interrupts. Need to reload cpu area
2673 c = this_cpu_ptr(s->cpu_slab);
2676 p = ___slab_alloc(s, gfpflags, node, addr, c);
2677 local_irq_restore(flags);
2682 * If the object has been wiped upon free, make sure it's fully initialized by
2683 * zeroing out freelist pointer.
2685 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2688 if (unlikely(slab_want_init_on_free(s)) && obj)
2689 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2693 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2694 * have the fastpath folded into their functions. So no function call
2695 * overhead for requests that can be satisfied on the fastpath.
2697 * The fastpath works by first checking if the lockless freelist can be used.
2698 * If not then __slab_alloc is called for slow processing.
2700 * Otherwise we can simply pick the next object from the lockless free list.
2702 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2703 gfp_t gfpflags, int node, unsigned long addr)
2706 struct kmem_cache_cpu *c;
2710 s = slab_pre_alloc_hook(s, gfpflags);
2715 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2716 * enabled. We may switch back and forth between cpus while
2717 * reading from one cpu area. That does not matter as long
2718 * as we end up on the original cpu again when doing the cmpxchg.
2720 * We should guarantee that tid and kmem_cache are retrieved on
2721 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2722 * to check if it is matched or not.
2725 tid = this_cpu_read(s->cpu_slab->tid);
2726 c = raw_cpu_ptr(s->cpu_slab);
2727 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2728 unlikely(tid != READ_ONCE(c->tid)));
2731 * Irqless object alloc/free algorithm used here depends on sequence
2732 * of fetching cpu_slab's data. tid should be fetched before anything
2733 * on c to guarantee that object and page associated with previous tid
2734 * won't be used with current tid. If we fetch tid first, object and
2735 * page could be one associated with next tid and our alloc/free
2736 * request will be failed. In this case, we will retry. So, no problem.
2741 * The transaction ids are globally unique per cpu and per operation on
2742 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2743 * occurs on the right processor and that there was no operation on the
2744 * linked list in between.
2747 object = c->freelist;
2749 if (unlikely(!object || !node_match(page, node))) {
2750 object = __slab_alloc(s, gfpflags, node, addr, c);
2751 stat(s, ALLOC_SLOWPATH);
2753 void *next_object = get_freepointer_safe(s, object);
2756 * The cmpxchg will only match if there was no additional
2757 * operation and if we are on the right processor.
2759 * The cmpxchg does the following atomically (without lock
2761 * 1. Relocate first pointer to the current per cpu area.
2762 * 2. Verify that tid and freelist have not been changed
2763 * 3. If they were not changed replace tid and freelist
2765 * Since this is without lock semantics the protection is only
2766 * against code executing on this cpu *not* from access by
2769 if (unlikely(!this_cpu_cmpxchg_double(
2770 s->cpu_slab->freelist, s->cpu_slab->tid,
2772 next_object, next_tid(tid)))) {
2774 note_cmpxchg_failure("slab_alloc", s, tid);
2777 prefetch_freepointer(s, next_object);
2778 stat(s, ALLOC_FASTPATH);
2781 maybe_wipe_obj_freeptr(s, object);
2783 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2784 memset(object, 0, s->object_size);
2786 slab_post_alloc_hook(s, gfpflags, 1, &object);
2791 static __always_inline void *slab_alloc(struct kmem_cache *s,
2792 gfp_t gfpflags, unsigned long addr)
2794 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2797 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2799 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2801 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2806 EXPORT_SYMBOL(kmem_cache_alloc);
2808 #ifdef CONFIG_TRACING
2809 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2811 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2812 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2813 ret = kasan_kmalloc(s, ret, size, gfpflags);
2816 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2820 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2822 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2824 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2825 s->object_size, s->size, gfpflags, node);
2829 EXPORT_SYMBOL(kmem_cache_alloc_node);
2831 #ifdef CONFIG_TRACING
2832 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2834 int node, size_t size)
2836 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2838 trace_kmalloc_node(_RET_IP_, ret,
2839 size, s->size, gfpflags, node);
2841 ret = kasan_kmalloc(s, ret, size, gfpflags);
2844 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2846 #endif /* CONFIG_NUMA */
2849 * Slow path handling. This may still be called frequently since objects
2850 * have a longer lifetime than the cpu slabs in most processing loads.
2852 * So we still attempt to reduce cache line usage. Just take the slab
2853 * lock and free the item. If there is no additional partial page
2854 * handling required then we can return immediately.
2856 static void __slab_free(struct kmem_cache *s, struct page *page,
2857 void *head, void *tail, int cnt,
2864 unsigned long counters;
2865 struct kmem_cache_node *n = NULL;
2866 unsigned long uninitialized_var(flags);
2868 stat(s, FREE_SLOWPATH);
2870 if (kmem_cache_debug(s) &&
2871 !free_debug_processing(s, page, head, tail, cnt, addr))
2876 spin_unlock_irqrestore(&n->list_lock, flags);
2879 prior = page->freelist;
2880 counters = page->counters;
2881 set_freepointer(s, tail, prior);
2882 new.counters = counters;
2883 was_frozen = new.frozen;
2885 if ((!new.inuse || !prior) && !was_frozen) {
2887 if (kmem_cache_has_cpu_partial(s) && !prior) {
2890 * Slab was on no list before and will be
2892 * We can defer the list move and instead
2897 } else { /* Needs to be taken off a list */
2899 n = get_node(s, page_to_nid(page));
2901 * Speculatively acquire the list_lock.
2902 * If the cmpxchg does not succeed then we may
2903 * drop the list_lock without any processing.
2905 * Otherwise the list_lock will synchronize with
2906 * other processors updating the list of slabs.
2908 spin_lock_irqsave(&n->list_lock, flags);
2913 } while (!cmpxchg_double_slab(s, page,
2921 * If we just froze the page then put it onto the
2922 * per cpu partial list.
2924 if (new.frozen && !was_frozen) {
2925 put_cpu_partial(s, page, 1);
2926 stat(s, CPU_PARTIAL_FREE);
2929 * The list lock was not taken therefore no list
2930 * activity can be necessary.
2933 stat(s, FREE_FROZEN);
2937 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2941 * Objects left in the slab. If it was not on the partial list before
2944 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2945 remove_full(s, n, page);
2946 add_partial(n, page, DEACTIVATE_TO_TAIL);
2947 stat(s, FREE_ADD_PARTIAL);
2949 spin_unlock_irqrestore(&n->list_lock, flags);
2955 * Slab on the partial list.
2957 remove_partial(n, page);
2958 stat(s, FREE_REMOVE_PARTIAL);
2960 /* Slab must be on the full list */
2961 remove_full(s, n, page);
2964 spin_unlock_irqrestore(&n->list_lock, flags);
2966 discard_slab(s, page);
2970 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2971 * can perform fastpath freeing without additional function calls.
2973 * The fastpath is only possible if we are freeing to the current cpu slab
2974 * of this processor. This typically the case if we have just allocated
2977 * If fastpath is not possible then fall back to __slab_free where we deal
2978 * with all sorts of special processing.
2980 * Bulk free of a freelist with several objects (all pointing to the
2981 * same page) possible by specifying head and tail ptr, plus objects
2982 * count (cnt). Bulk free indicated by tail pointer being set.
2984 static __always_inline void do_slab_free(struct kmem_cache *s,
2985 struct page *page, void *head, void *tail,
2986 int cnt, unsigned long addr)
2988 void *tail_obj = tail ? : head;
2989 struct kmem_cache_cpu *c;
2993 * Determine the currently cpus per cpu slab.
2994 * The cpu may change afterward. However that does not matter since
2995 * data is retrieved via this pointer. If we are on the same cpu
2996 * during the cmpxchg then the free will succeed.
2999 tid = this_cpu_read(s->cpu_slab->tid);
3000 c = raw_cpu_ptr(s->cpu_slab);
3001 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3002 unlikely(tid != READ_ONCE(c->tid)));
3004 /* Same with comment on barrier() in slab_alloc_node() */
3007 if (likely(page == c->page)) {
3008 void **freelist = READ_ONCE(c->freelist);
3010 set_freepointer(s, tail_obj, freelist);
3012 if (unlikely(!this_cpu_cmpxchg_double(
3013 s->cpu_slab->freelist, s->cpu_slab->tid,
3015 head, next_tid(tid)))) {
3017 note_cmpxchg_failure("slab_free", s, tid);
3020 stat(s, FREE_FASTPATH);
3022 __slab_free(s, page, head, tail_obj, cnt, addr);
3026 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3027 void *head, void *tail, int cnt,
3031 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3032 * to remove objects, whose reuse must be delayed.
3034 if (slab_free_freelist_hook(s, &head, &tail))
3035 do_slab_free(s, page, head, tail, cnt, addr);
3038 #ifdef CONFIG_KASAN_GENERIC
3039 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3041 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3045 void kmem_cache_free(struct kmem_cache *s, void *x)
3047 s = cache_from_obj(s, x);
3050 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3051 trace_kmem_cache_free(_RET_IP_, x);
3053 EXPORT_SYMBOL(kmem_cache_free);
3055 struct detached_freelist {
3060 struct kmem_cache *s;
3064 * This function progressively scans the array with free objects (with
3065 * a limited look ahead) and extract objects belonging to the same
3066 * page. It builds a detached freelist directly within the given
3067 * page/objects. This can happen without any need for
3068 * synchronization, because the objects are owned by running process.
3069 * The freelist is build up as a single linked list in the objects.
3070 * The idea is, that this detached freelist can then be bulk
3071 * transferred to the real freelist(s), but only requiring a single
3072 * synchronization primitive. Look ahead in the array is limited due
3073 * to performance reasons.
3076 int build_detached_freelist(struct kmem_cache *s, size_t size,
3077 void **p, struct detached_freelist *df)
3079 size_t first_skipped_index = 0;
3084 /* Always re-init detached_freelist */
3089 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3090 } while (!object && size);
3095 page = virt_to_head_page(object);
3097 /* Handle kalloc'ed objects */
3098 if (unlikely(!PageSlab(page))) {
3099 BUG_ON(!PageCompound(page));
3101 __free_pages(page, compound_order(page));
3102 p[size] = NULL; /* mark object processed */
3105 /* Derive kmem_cache from object */
3106 df->s = page->slab_cache;
3108 df->s = cache_from_obj(s, object); /* Support for memcg */
3111 /* Start new detached freelist */
3113 set_freepointer(df->s, object, NULL);
3115 df->freelist = object;
3116 p[size] = NULL; /* mark object processed */
3122 continue; /* Skip processed objects */
3124 /* df->page is always set at this point */
3125 if (df->page == virt_to_head_page(object)) {
3126 /* Opportunity build freelist */
3127 set_freepointer(df->s, object, df->freelist);
3128 df->freelist = object;
3130 p[size] = NULL; /* mark object processed */
3135 /* Limit look ahead search */
3139 if (!first_skipped_index)
3140 first_skipped_index = size + 1;
3143 return first_skipped_index;
3146 /* Note that interrupts must be enabled when calling this function. */
3147 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3153 struct detached_freelist df;
3155 size = build_detached_freelist(s, size, p, &df);
3159 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3160 } while (likely(size));
3162 EXPORT_SYMBOL(kmem_cache_free_bulk);
3164 /* Note that interrupts must be enabled when calling this function. */
3165 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3168 struct kmem_cache_cpu *c;
3171 /* memcg and kmem_cache debug support */
3172 s = slab_pre_alloc_hook(s, flags);
3176 * Drain objects in the per cpu slab, while disabling local
3177 * IRQs, which protects against PREEMPT and interrupts
3178 * handlers invoking normal fastpath.
3180 local_irq_disable();
3181 c = this_cpu_ptr(s->cpu_slab);
3183 for (i = 0; i < size; i++) {
3184 void *object = c->freelist;
3186 if (unlikely(!object)) {
3188 * We may have removed an object from c->freelist using
3189 * the fastpath in the previous iteration; in that case,
3190 * c->tid has not been bumped yet.
3191 * Since ___slab_alloc() may reenable interrupts while
3192 * allocating memory, we should bump c->tid now.
3194 c->tid = next_tid(c->tid);
3197 * Invoking slow path likely have side-effect
3198 * of re-populating per CPU c->freelist
3200 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3202 if (unlikely(!p[i]))
3205 c = this_cpu_ptr(s->cpu_slab);
3206 maybe_wipe_obj_freeptr(s, p[i]);
3208 continue; /* goto for-loop */
3210 c->freelist = get_freepointer(s, object);
3212 maybe_wipe_obj_freeptr(s, p[i]);
3214 c->tid = next_tid(c->tid);
3217 /* Clear memory outside IRQ disabled fastpath loop */
3218 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3221 for (j = 0; j < i; j++)
3222 memset(p[j], 0, s->object_size);
3225 /* memcg and kmem_cache debug support */
3226 slab_post_alloc_hook(s, flags, size, p);
3230 slab_post_alloc_hook(s, flags, i, p);
3231 __kmem_cache_free_bulk(s, i, p);
3234 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3238 * Object placement in a slab is made very easy because we always start at
3239 * offset 0. If we tune the size of the object to the alignment then we can
3240 * get the required alignment by putting one properly sized object after
3243 * Notice that the allocation order determines the sizes of the per cpu
3244 * caches. Each processor has always one slab available for allocations.
3245 * Increasing the allocation order reduces the number of times that slabs
3246 * must be moved on and off the partial lists and is therefore a factor in
3251 * Mininum / Maximum order of slab pages. This influences locking overhead
3252 * and slab fragmentation. A higher order reduces the number of partial slabs
3253 * and increases the number of allocations possible without having to
3254 * take the list_lock.
3256 static unsigned int slub_min_order;
3257 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3258 static unsigned int slub_min_objects;
3261 * Calculate the order of allocation given an slab object size.
3263 * The order of allocation has significant impact on performance and other
3264 * system components. Generally order 0 allocations should be preferred since
3265 * order 0 does not cause fragmentation in the page allocator. Larger objects
3266 * be problematic to put into order 0 slabs because there may be too much
3267 * unused space left. We go to a higher order if more than 1/16th of the slab
3270 * In order to reach satisfactory performance we must ensure that a minimum
3271 * number of objects is in one slab. Otherwise we may generate too much
3272 * activity on the partial lists which requires taking the list_lock. This is
3273 * less a concern for large slabs though which are rarely used.
3275 * slub_max_order specifies the order where we begin to stop considering the
3276 * number of objects in a slab as critical. If we reach slub_max_order then
3277 * we try to keep the page order as low as possible. So we accept more waste
3278 * of space in favor of a small page order.
3280 * Higher order allocations also allow the placement of more objects in a
3281 * slab and thereby reduce object handling overhead. If the user has
3282 * requested a higher mininum order then we start with that one instead of
3283 * the smallest order which will fit the object.
3285 static inline unsigned int slab_order(unsigned int size,
3286 unsigned int min_objects, unsigned int max_order,
3287 unsigned int fract_leftover)
3289 unsigned int min_order = slub_min_order;
3292 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3293 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3295 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3296 order <= max_order; order++) {
3298 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3301 rem = slab_size % size;
3303 if (rem <= slab_size / fract_leftover)
3310 static inline int calculate_order(unsigned int size)
3313 unsigned int min_objects;
3314 unsigned int max_objects;
3317 * Attempt to find best configuration for a slab. This
3318 * works by first attempting to generate a layout with
3319 * the best configuration and backing off gradually.
3321 * First we increase the acceptable waste in a slab. Then
3322 * we reduce the minimum objects required in a slab.
3324 min_objects = slub_min_objects;
3326 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3327 max_objects = order_objects(slub_max_order, size);
3328 min_objects = min(min_objects, max_objects);
3330 while (min_objects > 1) {
3331 unsigned int fraction;
3334 while (fraction >= 4) {
3335 order = slab_order(size, min_objects,
3336 slub_max_order, fraction);
3337 if (order <= slub_max_order)
3345 * We were unable to place multiple objects in a slab. Now
3346 * lets see if we can place a single object there.
3348 order = slab_order(size, 1, slub_max_order, 1);
3349 if (order <= slub_max_order)
3353 * Doh this slab cannot be placed using slub_max_order.
3355 order = slab_order(size, 1, MAX_ORDER, 1);
3356 if (order < MAX_ORDER)
3362 init_kmem_cache_node(struct kmem_cache_node *n)
3365 spin_lock_init(&n->list_lock);
3366 INIT_LIST_HEAD(&n->partial);
3367 #ifdef CONFIG_SLUB_DEBUG
3368 atomic_long_set(&n->nr_slabs, 0);
3369 atomic_long_set(&n->total_objects, 0);
3370 INIT_LIST_HEAD(&n->full);
3374 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3376 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3377 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3380 * Must align to double word boundary for the double cmpxchg
3381 * instructions to work; see __pcpu_double_call_return_bool().
3383 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3384 2 * sizeof(void *));
3389 init_kmem_cache_cpus(s);
3394 static struct kmem_cache *kmem_cache_node;
3397 * No kmalloc_node yet so do it by hand. We know that this is the first
3398 * slab on the node for this slabcache. There are no concurrent accesses
3401 * Note that this function only works on the kmem_cache_node
3402 * when allocating for the kmem_cache_node. This is used for bootstrapping
3403 * memory on a fresh node that has no slab structures yet.
3405 static void early_kmem_cache_node_alloc(int node)
3408 struct kmem_cache_node *n;
3410 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3412 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3415 if (page_to_nid(page) != node) {
3416 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3417 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3422 #ifdef CONFIG_SLUB_DEBUG
3423 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3424 init_tracking(kmem_cache_node, n);
3426 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3428 page->freelist = get_freepointer(kmem_cache_node, n);
3431 kmem_cache_node->node[node] = n;
3432 init_kmem_cache_node(n);
3433 inc_slabs_node(kmem_cache_node, node, page->objects);
3436 * No locks need to be taken here as it has just been
3437 * initialized and there is no concurrent access.
3439 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3442 static void free_kmem_cache_nodes(struct kmem_cache *s)
3445 struct kmem_cache_node *n;
3447 for_each_kmem_cache_node(s, node, n) {
3448 s->node[node] = NULL;
3449 kmem_cache_free(kmem_cache_node, n);
3453 void __kmem_cache_release(struct kmem_cache *s)
3455 cache_random_seq_destroy(s);
3456 free_percpu(s->cpu_slab);
3457 free_kmem_cache_nodes(s);
3460 static int init_kmem_cache_nodes(struct kmem_cache *s)
3464 for_each_node_state(node, N_NORMAL_MEMORY) {
3465 struct kmem_cache_node *n;
3467 if (slab_state == DOWN) {
3468 early_kmem_cache_node_alloc(node);
3471 n = kmem_cache_alloc_node(kmem_cache_node,
3475 free_kmem_cache_nodes(s);
3479 init_kmem_cache_node(n);
3485 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3487 if (min < MIN_PARTIAL)
3489 else if (min > MAX_PARTIAL)
3491 s->min_partial = min;
3494 static void set_cpu_partial(struct kmem_cache *s)
3496 #ifdef CONFIG_SLUB_CPU_PARTIAL
3498 * cpu_partial determined the maximum number of objects kept in the
3499 * per cpu partial lists of a processor.
3501 * Per cpu partial lists mainly contain slabs that just have one
3502 * object freed. If they are used for allocation then they can be
3503 * filled up again with minimal effort. The slab will never hit the
3504 * per node partial lists and therefore no locking will be required.
3506 * This setting also determines
3508 * A) The number of objects from per cpu partial slabs dumped to the
3509 * per node list when we reach the limit.
3510 * B) The number of objects in cpu partial slabs to extract from the
3511 * per node list when we run out of per cpu objects. We only fetch
3512 * 50% to keep some capacity around for frees.
3514 if (!kmem_cache_has_cpu_partial(s))
3515 slub_set_cpu_partial(s, 0);
3516 else if (s->size >= PAGE_SIZE)
3517 slub_set_cpu_partial(s, 2);
3518 else if (s->size >= 1024)
3519 slub_set_cpu_partial(s, 6);
3520 else if (s->size >= 256)
3521 slub_set_cpu_partial(s, 13);
3523 slub_set_cpu_partial(s, 30);
3528 * calculate_sizes() determines the order and the distribution of data within
3531 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3533 slab_flags_t flags = s->flags;
3534 unsigned int size = s->object_size;
3538 * Round up object size to the next word boundary. We can only
3539 * place the free pointer at word boundaries and this determines
3540 * the possible location of the free pointer.
3542 size = ALIGN(size, sizeof(void *));
3544 #ifdef CONFIG_SLUB_DEBUG
3546 * Determine if we can poison the object itself. If the user of
3547 * the slab may touch the object after free or before allocation
3548 * then we should never poison the object itself.
3550 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3552 s->flags |= __OBJECT_POISON;
3554 s->flags &= ~__OBJECT_POISON;
3558 * If we are Redzoning then check if there is some space between the
3559 * end of the object and the free pointer. If not then add an
3560 * additional word to have some bytes to store Redzone information.
3562 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3563 size += sizeof(void *);
3567 * With that we have determined the number of bytes in actual use
3568 * by the object. This is the potential offset to the free pointer.
3572 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3575 * Relocate free pointer after the object if it is not
3576 * permitted to overwrite the first word of the object on
3579 * This is the case if we do RCU, have a constructor or
3580 * destructor or are poisoning the objects.
3583 size += sizeof(void *);
3584 } else if (size > sizeof(void *)) {
3586 * Store freelist pointer near middle of object to keep
3587 * it away from the edges of the object to avoid small
3588 * sized over/underflows from neighboring allocations.
3590 s->offset = ALIGN(size / 2, sizeof(void *));
3593 #ifdef CONFIG_SLUB_DEBUG
3594 if (flags & SLAB_STORE_USER)
3596 * Need to store information about allocs and frees after
3599 size += 2 * sizeof(struct track);
3602 kasan_cache_create(s, &size, &s->flags);
3603 #ifdef CONFIG_SLUB_DEBUG
3604 if (flags & SLAB_RED_ZONE) {
3606 * Add some empty padding so that we can catch
3607 * overwrites from earlier objects rather than let
3608 * tracking information or the free pointer be
3609 * corrupted if a user writes before the start
3612 size += sizeof(void *);
3614 s->red_left_pad = sizeof(void *);
3615 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3616 size += s->red_left_pad;
3621 * SLUB stores one object immediately after another beginning from
3622 * offset 0. In order to align the objects we have to simply size
3623 * each object to conform to the alignment.
3625 size = ALIGN(size, s->align);
3627 if (forced_order >= 0)
3628 order = forced_order;
3630 order = calculate_order(size);
3637 s->allocflags |= __GFP_COMP;
3639 if (s->flags & SLAB_CACHE_DMA)
3640 s->allocflags |= GFP_DMA;
3642 if (s->flags & SLAB_CACHE_DMA32)
3643 s->allocflags |= GFP_DMA32;
3645 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3646 s->allocflags |= __GFP_RECLAIMABLE;
3649 * Determine the number of objects per slab
3651 s->oo = oo_make(order, size);
3652 s->min = oo_make(get_order(size), size);
3653 if (oo_objects(s->oo) > oo_objects(s->max))
3656 return !!oo_objects(s->oo);
3659 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3661 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3662 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3663 s->random = get_random_long();
3666 if (!calculate_sizes(s, -1))
3668 if (disable_higher_order_debug) {
3670 * Disable debugging flags that store metadata if the min slab
3673 if (get_order(s->size) > get_order(s->object_size)) {
3674 s->flags &= ~DEBUG_METADATA_FLAGS;
3676 if (!calculate_sizes(s, -1))
3681 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3682 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3683 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3684 /* Enable fast mode */
3685 s->flags |= __CMPXCHG_DOUBLE;
3689 * The larger the object size is, the more pages we want on the partial
3690 * list to avoid pounding the page allocator excessively.
3692 set_min_partial(s, ilog2(s->size) / 2);
3697 s->remote_node_defrag_ratio = 1000;
3700 /* Initialize the pre-computed randomized freelist if slab is up */
3701 if (slab_state >= UP) {
3702 if (init_cache_random_seq(s))
3706 if (!init_kmem_cache_nodes(s))
3709 if (alloc_kmem_cache_cpus(s))
3712 free_kmem_cache_nodes(s);
3717 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3720 #ifdef CONFIG_SLUB_DEBUG
3721 void *addr = page_address(page);
3725 slab_err(s, page, text, s->name);
3728 map = get_map(s, page);
3729 for_each_object(p, s, addr, page->objects) {
3731 if (!test_bit(slab_index(p, s, addr), map)) {
3732 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3733 print_tracking(s, p);
3743 * Attempt to free all partial slabs on a node.
3744 * This is called from __kmem_cache_shutdown(). We must take list_lock
3745 * because sysfs file might still access partial list after the shutdowning.
3747 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3750 struct page *page, *h;
3752 BUG_ON(irqs_disabled());
3753 spin_lock_irq(&n->list_lock);
3754 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3756 remove_partial(n, page);
3757 list_add(&page->slab_list, &discard);
3759 list_slab_objects(s, page,
3760 "Objects remaining in %s on __kmem_cache_shutdown()");
3763 spin_unlock_irq(&n->list_lock);
3765 list_for_each_entry_safe(page, h, &discard, slab_list)
3766 discard_slab(s, page);
3769 bool __kmem_cache_empty(struct kmem_cache *s)
3772 struct kmem_cache_node *n;
3774 for_each_kmem_cache_node(s, node, n)
3775 if (n->nr_partial || slabs_node(s, node))
3781 * Release all resources used by a slab cache.
3783 int __kmem_cache_shutdown(struct kmem_cache *s)
3786 struct kmem_cache_node *n;
3789 /* Attempt to free all objects */
3790 for_each_kmem_cache_node(s, node, n) {
3792 if (n->nr_partial || slabs_node(s, node))
3795 sysfs_slab_remove(s);
3799 /********************************************************************
3801 *******************************************************************/
3803 static int __init setup_slub_min_order(char *str)
3805 get_option(&str, (int *)&slub_min_order);
3810 __setup("slub_min_order=", setup_slub_min_order);
3812 static int __init setup_slub_max_order(char *str)
3814 get_option(&str, (int *)&slub_max_order);
3815 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3820 __setup("slub_max_order=", setup_slub_max_order);
3822 static int __init setup_slub_min_objects(char *str)
3824 get_option(&str, (int *)&slub_min_objects);
3829 __setup("slub_min_objects=", setup_slub_min_objects);
3831 void *__kmalloc(size_t size, gfp_t flags)
3833 struct kmem_cache *s;
3836 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3837 return kmalloc_large(size, flags);
3839 s = kmalloc_slab(size, flags);
3841 if (unlikely(ZERO_OR_NULL_PTR(s)))
3844 ret = slab_alloc(s, flags, _RET_IP_);
3846 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3848 ret = kasan_kmalloc(s, ret, size, flags);
3852 EXPORT_SYMBOL(__kmalloc);
3855 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3859 unsigned int order = get_order(size);
3861 flags |= __GFP_COMP;
3862 page = alloc_pages_node(node, flags, order);
3864 ptr = page_address(page);
3865 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3869 return kmalloc_large_node_hook(ptr, size, flags);
3872 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3874 struct kmem_cache *s;
3877 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3878 ret = kmalloc_large_node(size, flags, node);
3880 trace_kmalloc_node(_RET_IP_, ret,
3881 size, PAGE_SIZE << get_order(size),
3887 s = kmalloc_slab(size, flags);
3889 if (unlikely(ZERO_OR_NULL_PTR(s)))
3892 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3894 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3896 ret = kasan_kmalloc(s, ret, size, flags);
3900 EXPORT_SYMBOL(__kmalloc_node);
3901 #endif /* CONFIG_NUMA */
3903 #ifdef CONFIG_HARDENED_USERCOPY
3905 * Rejects incorrectly sized objects and objects that are to be copied
3906 * to/from userspace but do not fall entirely within the containing slab
3907 * cache's usercopy region.
3909 * Returns NULL if check passes, otherwise const char * to name of cache
3910 * to indicate an error.
3912 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3915 struct kmem_cache *s;
3916 unsigned int offset;
3919 ptr = kasan_reset_tag(ptr);
3921 /* Find object and usable object size. */
3922 s = page->slab_cache;
3924 /* Reject impossible pointers. */
3925 if (ptr < page_address(page))
3926 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3929 /* Find offset within object. */
3930 offset = (ptr - page_address(page)) % s->size;
3932 /* Adjust for redzone and reject if within the redzone. */
3933 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3934 if (offset < s->red_left_pad)
3935 usercopy_abort("SLUB object in left red zone",
3936 s->name, to_user, offset, n);
3937 offset -= s->red_left_pad;
3940 /* Allow address range falling entirely within usercopy region. */
3941 if (offset >= s->useroffset &&
3942 offset - s->useroffset <= s->usersize &&
3943 n <= s->useroffset - offset + s->usersize)
3947 * If the copy is still within the allocated object, produce
3948 * a warning instead of rejecting the copy. This is intended
3949 * to be a temporary method to find any missing usercopy
3952 object_size = slab_ksize(s);
3953 if (usercopy_fallback &&
3954 offset <= object_size && n <= object_size - offset) {
3955 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3959 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3961 #endif /* CONFIG_HARDENED_USERCOPY */
3963 size_t __ksize(const void *object)
3967 if (unlikely(object == ZERO_SIZE_PTR))
3970 page = virt_to_head_page(object);
3972 if (unlikely(!PageSlab(page))) {
3973 WARN_ON(!PageCompound(page));
3974 return page_size(page);
3977 return slab_ksize(page->slab_cache);
3979 EXPORT_SYMBOL(__ksize);
3981 void kfree(const void *x)
3984 void *object = (void *)x;
3986 trace_kfree(_RET_IP_, x);
3988 if (unlikely(ZERO_OR_NULL_PTR(x)))
3991 page = virt_to_head_page(x);
3992 if (unlikely(!PageSlab(page))) {
3993 unsigned int order = compound_order(page);
3995 BUG_ON(!PageCompound(page));
3997 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3999 __free_pages(page, order);
4002 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4004 EXPORT_SYMBOL(kfree);
4006 #define SHRINK_PROMOTE_MAX 32
4009 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4010 * up most to the head of the partial lists. New allocations will then
4011 * fill those up and thus they can be removed from the partial lists.
4013 * The slabs with the least items are placed last. This results in them
4014 * being allocated from last increasing the chance that the last objects
4015 * are freed in them.
4017 int __kmem_cache_shrink(struct kmem_cache *s)
4021 struct kmem_cache_node *n;
4024 struct list_head discard;
4025 struct list_head promote[SHRINK_PROMOTE_MAX];
4026 unsigned long flags;
4030 for_each_kmem_cache_node(s, node, n) {
4031 INIT_LIST_HEAD(&discard);
4032 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4033 INIT_LIST_HEAD(promote + i);
4035 spin_lock_irqsave(&n->list_lock, flags);
4038 * Build lists of slabs to discard or promote.
4040 * Note that concurrent frees may occur while we hold the
4041 * list_lock. page->inuse here is the upper limit.
4043 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4044 int free = page->objects - page->inuse;
4046 /* Do not reread page->inuse */
4049 /* We do not keep full slabs on the list */
4052 if (free == page->objects) {
4053 list_move(&page->slab_list, &discard);
4055 } else if (free <= SHRINK_PROMOTE_MAX)
4056 list_move(&page->slab_list, promote + free - 1);
4060 * Promote the slabs filled up most to the head of the
4063 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4064 list_splice(promote + i, &n->partial);
4066 spin_unlock_irqrestore(&n->list_lock, flags);
4068 /* Release empty slabs */
4069 list_for_each_entry_safe(page, t, &discard, slab_list)
4070 discard_slab(s, page);
4072 if (slabs_node(s, node))
4080 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4083 * Called with all the locks held after a sched RCU grace period.
4084 * Even if @s becomes empty after shrinking, we can't know that @s
4085 * doesn't have allocations already in-flight and thus can't
4086 * destroy @s until the associated memcg is released.
4088 * However, let's remove the sysfs files for empty caches here.
4089 * Each cache has a lot of interface files which aren't
4090 * particularly useful for empty draining caches; otherwise, we can
4091 * easily end up with millions of unnecessary sysfs files on
4092 * systems which have a lot of memory and transient cgroups.
4094 if (!__kmem_cache_shrink(s))
4095 sysfs_slab_remove(s);
4098 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4101 * Disable empty slabs caching. Used to avoid pinning offline
4102 * memory cgroups by kmem pages that can be freed.
4104 slub_set_cpu_partial(s, 0);
4107 #endif /* CONFIG_MEMCG */
4109 static int slab_mem_going_offline_callback(void *arg)
4111 struct kmem_cache *s;
4113 mutex_lock(&slab_mutex);
4114 list_for_each_entry(s, &slab_caches, list)
4115 __kmem_cache_shrink(s);
4116 mutex_unlock(&slab_mutex);
4121 static void slab_mem_offline_callback(void *arg)
4123 struct kmem_cache_node *n;
4124 struct kmem_cache *s;
4125 struct memory_notify *marg = arg;
4128 offline_node = marg->status_change_nid_normal;
4131 * If the node still has available memory. we need kmem_cache_node
4134 if (offline_node < 0)
4137 mutex_lock(&slab_mutex);
4138 list_for_each_entry(s, &slab_caches, list) {
4139 n = get_node(s, offline_node);
4142 * if n->nr_slabs > 0, slabs still exist on the node
4143 * that is going down. We were unable to free them,
4144 * and offline_pages() function shouldn't call this
4145 * callback. So, we must fail.
4147 BUG_ON(slabs_node(s, offline_node));
4149 s->node[offline_node] = NULL;
4150 kmem_cache_free(kmem_cache_node, n);
4153 mutex_unlock(&slab_mutex);
4156 static int slab_mem_going_online_callback(void *arg)
4158 struct kmem_cache_node *n;
4159 struct kmem_cache *s;
4160 struct memory_notify *marg = arg;
4161 int nid = marg->status_change_nid_normal;
4165 * If the node's memory is already available, then kmem_cache_node is
4166 * already created. Nothing to do.
4172 * We are bringing a node online. No memory is available yet. We must
4173 * allocate a kmem_cache_node structure in order to bring the node
4176 mutex_lock(&slab_mutex);
4177 list_for_each_entry(s, &slab_caches, list) {
4179 * XXX: kmem_cache_alloc_node will fallback to other nodes
4180 * since memory is not yet available from the node that
4183 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4188 init_kmem_cache_node(n);
4192 mutex_unlock(&slab_mutex);
4196 static int slab_memory_callback(struct notifier_block *self,
4197 unsigned long action, void *arg)
4202 case MEM_GOING_ONLINE:
4203 ret = slab_mem_going_online_callback(arg);
4205 case MEM_GOING_OFFLINE:
4206 ret = slab_mem_going_offline_callback(arg);
4209 case MEM_CANCEL_ONLINE:
4210 slab_mem_offline_callback(arg);
4213 case MEM_CANCEL_OFFLINE:
4217 ret = notifier_from_errno(ret);
4223 static struct notifier_block slab_memory_callback_nb = {
4224 .notifier_call = slab_memory_callback,
4225 .priority = SLAB_CALLBACK_PRI,
4228 /********************************************************************
4229 * Basic setup of slabs
4230 *******************************************************************/
4233 * Used for early kmem_cache structures that were allocated using
4234 * the page allocator. Allocate them properly then fix up the pointers
4235 * that may be pointing to the wrong kmem_cache structure.
4238 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4241 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4242 struct kmem_cache_node *n;
4244 memcpy(s, static_cache, kmem_cache->object_size);
4247 * This runs very early, and only the boot processor is supposed to be
4248 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4251 __flush_cpu_slab(s, smp_processor_id());
4252 for_each_kmem_cache_node(s, node, n) {
4255 list_for_each_entry(p, &n->partial, slab_list)
4258 #ifdef CONFIG_SLUB_DEBUG
4259 list_for_each_entry(p, &n->full, slab_list)
4263 slab_init_memcg_params(s);
4264 list_add(&s->list, &slab_caches);
4265 memcg_link_cache(s, NULL);
4269 void __init kmem_cache_init(void)
4271 static __initdata struct kmem_cache boot_kmem_cache,
4272 boot_kmem_cache_node;
4274 if (debug_guardpage_minorder())
4277 kmem_cache_node = &boot_kmem_cache_node;
4278 kmem_cache = &boot_kmem_cache;
4280 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4281 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4283 register_hotmemory_notifier(&slab_memory_callback_nb);
4285 /* Able to allocate the per node structures */
4286 slab_state = PARTIAL;
4288 create_boot_cache(kmem_cache, "kmem_cache",
4289 offsetof(struct kmem_cache, node) +
4290 nr_node_ids * sizeof(struct kmem_cache_node *),
4291 SLAB_HWCACHE_ALIGN, 0, 0);
4293 kmem_cache = bootstrap(&boot_kmem_cache);
4294 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4296 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4297 setup_kmalloc_cache_index_table();
4298 create_kmalloc_caches(0);
4300 /* Setup random freelists for each cache */
4301 init_freelist_randomization();
4303 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4306 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4308 slub_min_order, slub_max_order, slub_min_objects,
4309 nr_cpu_ids, nr_node_ids);
4312 void __init kmem_cache_init_late(void)
4317 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4318 slab_flags_t flags, void (*ctor)(void *))
4320 struct kmem_cache *s, *c;
4322 s = find_mergeable(size, align, flags, name, ctor);
4327 * Adjust the object sizes so that we clear
4328 * the complete object on kzalloc.
4330 s->object_size = max(s->object_size, size);
4331 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4333 for_each_memcg_cache(c, s) {
4334 c->object_size = s->object_size;
4335 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4338 if (sysfs_slab_alias(s, name)) {
4347 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4351 err = kmem_cache_open(s, flags);
4355 /* Mutex is not taken during early boot */
4356 if (slab_state <= UP)
4359 memcg_propagate_slab_attrs(s);
4360 err = sysfs_slab_add(s);
4362 __kmem_cache_release(s);
4367 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4369 struct kmem_cache *s;
4372 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4373 return kmalloc_large(size, gfpflags);
4375 s = kmalloc_slab(size, gfpflags);
4377 if (unlikely(ZERO_OR_NULL_PTR(s)))
4380 ret = slab_alloc(s, gfpflags, caller);
4382 /* Honor the call site pointer we received. */
4383 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4389 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4390 int node, unsigned long caller)
4392 struct kmem_cache *s;
4395 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4396 ret = kmalloc_large_node(size, gfpflags, node);
4398 trace_kmalloc_node(caller, ret,
4399 size, PAGE_SIZE << get_order(size),
4405 s = kmalloc_slab(size, gfpflags);
4407 if (unlikely(ZERO_OR_NULL_PTR(s)))
4410 ret = slab_alloc_node(s, gfpflags, node, caller);
4412 /* Honor the call site pointer we received. */
4413 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4420 static int count_inuse(struct page *page)
4425 static int count_total(struct page *page)
4427 return page->objects;
4431 #ifdef CONFIG_SLUB_DEBUG
4432 static void validate_slab(struct kmem_cache *s, struct page *page)
4435 void *addr = page_address(page);
4440 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4443 /* Now we know that a valid freelist exists */
4444 map = get_map(s, page);
4445 for_each_object(p, s, addr, page->objects) {
4446 u8 val = test_bit(slab_index(p, s, addr), map) ?
4447 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4449 if (!check_object(s, page, p, val))
4457 static int validate_slab_node(struct kmem_cache *s,
4458 struct kmem_cache_node *n)
4460 unsigned long count = 0;
4462 unsigned long flags;
4464 spin_lock_irqsave(&n->list_lock, flags);
4466 list_for_each_entry(page, &n->partial, slab_list) {
4467 validate_slab(s, page);
4470 if (count != n->nr_partial)
4471 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4472 s->name, count, n->nr_partial);
4474 if (!(s->flags & SLAB_STORE_USER))
4477 list_for_each_entry(page, &n->full, slab_list) {
4478 validate_slab(s, page);
4481 if (count != atomic_long_read(&n->nr_slabs))
4482 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4483 s->name, count, atomic_long_read(&n->nr_slabs));
4486 spin_unlock_irqrestore(&n->list_lock, flags);
4490 static long validate_slab_cache(struct kmem_cache *s)
4493 unsigned long count = 0;
4494 struct kmem_cache_node *n;
4497 for_each_kmem_cache_node(s, node, n)
4498 count += validate_slab_node(s, n);
4503 * Generate lists of code addresses where slabcache objects are allocated
4508 unsigned long count;
4515 DECLARE_BITMAP(cpus, NR_CPUS);
4521 unsigned long count;
4522 struct location *loc;
4525 static void free_loc_track(struct loc_track *t)
4528 free_pages((unsigned long)t->loc,
4529 get_order(sizeof(struct location) * t->max));
4532 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4537 order = get_order(sizeof(struct location) * max);
4539 l = (void *)__get_free_pages(flags, order);
4544 memcpy(l, t->loc, sizeof(struct location) * t->count);
4552 static int add_location(struct loc_track *t, struct kmem_cache *s,
4553 const struct track *track)
4555 long start, end, pos;
4557 unsigned long caddr;
4558 unsigned long age = jiffies - track->when;
4564 pos = start + (end - start + 1) / 2;
4567 * There is nothing at "end". If we end up there
4568 * we need to add something to before end.
4573 caddr = t->loc[pos].addr;
4574 if (track->addr == caddr) {
4580 if (age < l->min_time)
4582 if (age > l->max_time)
4585 if (track->pid < l->min_pid)
4586 l->min_pid = track->pid;
4587 if (track->pid > l->max_pid)
4588 l->max_pid = track->pid;
4590 cpumask_set_cpu(track->cpu,
4591 to_cpumask(l->cpus));
4593 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4597 if (track->addr < caddr)
4604 * Not found. Insert new tracking element.
4606 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4612 (t->count - pos) * sizeof(struct location));
4615 l->addr = track->addr;
4619 l->min_pid = track->pid;
4620 l->max_pid = track->pid;
4621 cpumask_clear(to_cpumask(l->cpus));
4622 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4623 nodes_clear(l->nodes);
4624 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4628 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4629 struct page *page, enum track_item alloc)
4631 void *addr = page_address(page);
4635 map = get_map(s, page);
4636 for_each_object(p, s, addr, page->objects)
4637 if (!test_bit(slab_index(p, s, addr), map))
4638 add_location(t, s, get_track(s, p, alloc));
4642 static int list_locations(struct kmem_cache *s, char *buf,
4643 enum track_item alloc)
4647 struct loc_track t = { 0, 0, NULL };
4649 struct kmem_cache_node *n;
4651 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4653 return sprintf(buf, "Out of memory\n");
4655 /* Push back cpu slabs */
4658 for_each_kmem_cache_node(s, node, n) {
4659 unsigned long flags;
4662 if (!atomic_long_read(&n->nr_slabs))
4665 spin_lock_irqsave(&n->list_lock, flags);
4666 list_for_each_entry(page, &n->partial, slab_list)
4667 process_slab(&t, s, page, alloc);
4668 list_for_each_entry(page, &n->full, slab_list)
4669 process_slab(&t, s, page, alloc);
4670 spin_unlock_irqrestore(&n->list_lock, flags);
4673 for (i = 0; i < t.count; i++) {
4674 struct location *l = &t.loc[i];
4676 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4678 len += sprintf(buf + len, "%7ld ", l->count);
4681 len += sprintf(buf + len, "%pS", (void *)l->addr);
4683 len += sprintf(buf + len, "<not-available>");
4685 if (l->sum_time != l->min_time) {
4686 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4688 (long)div_u64(l->sum_time, l->count),
4691 len += sprintf(buf + len, " age=%ld",
4694 if (l->min_pid != l->max_pid)
4695 len += sprintf(buf + len, " pid=%ld-%ld",
4696 l->min_pid, l->max_pid);
4698 len += sprintf(buf + len, " pid=%ld",
4701 if (num_online_cpus() > 1 &&
4702 !cpumask_empty(to_cpumask(l->cpus)) &&
4703 len < PAGE_SIZE - 60)
4704 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4706 cpumask_pr_args(to_cpumask(l->cpus)));
4708 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4709 len < PAGE_SIZE - 60)
4710 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4712 nodemask_pr_args(&l->nodes));
4714 len += sprintf(buf + len, "\n");
4719 len += sprintf(buf, "No data\n");
4722 #endif /* CONFIG_SLUB_DEBUG */
4724 #ifdef SLUB_RESILIENCY_TEST
4725 static void __init resiliency_test(void)
4728 int type = KMALLOC_NORMAL;
4730 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4732 pr_err("SLUB resiliency testing\n");
4733 pr_err("-----------------------\n");
4734 pr_err("A. Corruption after allocation\n");
4736 p = kzalloc(16, GFP_KERNEL);
4738 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4741 validate_slab_cache(kmalloc_caches[type][4]);
4743 /* Hmmm... The next two are dangerous */
4744 p = kzalloc(32, GFP_KERNEL);
4745 p[32 + sizeof(void *)] = 0x34;
4746 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4748 pr_err("If allocated object is overwritten then not detectable\n\n");
4750 validate_slab_cache(kmalloc_caches[type][5]);
4751 p = kzalloc(64, GFP_KERNEL);
4752 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4754 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4756 pr_err("If allocated object is overwritten then not detectable\n\n");
4757 validate_slab_cache(kmalloc_caches[type][6]);
4759 pr_err("\nB. Corruption after free\n");
4760 p = kzalloc(128, GFP_KERNEL);
4763 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4764 validate_slab_cache(kmalloc_caches[type][7]);
4766 p = kzalloc(256, GFP_KERNEL);
4769 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4770 validate_slab_cache(kmalloc_caches[type][8]);
4772 p = kzalloc(512, GFP_KERNEL);
4775 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4776 validate_slab_cache(kmalloc_caches[type][9]);
4780 static void resiliency_test(void) {};
4782 #endif /* SLUB_RESILIENCY_TEST */
4785 enum slab_stat_type {
4786 SL_ALL, /* All slabs */
4787 SL_PARTIAL, /* Only partially allocated slabs */
4788 SL_CPU, /* Only slabs used for cpu caches */
4789 SL_OBJECTS, /* Determine allocated objects not slabs */
4790 SL_TOTAL /* Determine object capacity not slabs */
4793 #define SO_ALL (1 << SL_ALL)
4794 #define SO_PARTIAL (1 << SL_PARTIAL)
4795 #define SO_CPU (1 << SL_CPU)
4796 #define SO_OBJECTS (1 << SL_OBJECTS)
4797 #define SO_TOTAL (1 << SL_TOTAL)
4800 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4802 static int __init setup_slub_memcg_sysfs(char *str)
4806 if (get_option(&str, &v) > 0)
4807 memcg_sysfs_enabled = v;
4812 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4815 static ssize_t show_slab_objects(struct kmem_cache *s,
4816 char *buf, unsigned long flags)
4818 unsigned long total = 0;
4821 unsigned long *nodes;
4823 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4827 if (flags & SO_CPU) {
4830 for_each_possible_cpu(cpu) {
4831 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4836 page = READ_ONCE(c->page);
4840 node = page_to_nid(page);
4841 if (flags & SO_TOTAL)
4843 else if (flags & SO_OBJECTS)
4851 page = slub_percpu_partial_read_once(c);
4853 node = page_to_nid(page);
4854 if (flags & SO_TOTAL)
4856 else if (flags & SO_OBJECTS)
4867 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4868 * already held which will conflict with an existing lock order:
4870 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4872 * We don't really need mem_hotplug_lock (to hold off
4873 * slab_mem_going_offline_callback) here because slab's memory hot
4874 * unplug code doesn't destroy the kmem_cache->node[] data.
4877 #ifdef CONFIG_SLUB_DEBUG
4878 if (flags & SO_ALL) {
4879 struct kmem_cache_node *n;
4881 for_each_kmem_cache_node(s, node, n) {
4883 if (flags & SO_TOTAL)
4884 x = atomic_long_read(&n->total_objects);
4885 else if (flags & SO_OBJECTS)
4886 x = atomic_long_read(&n->total_objects) -
4887 count_partial(n, count_free);
4889 x = atomic_long_read(&n->nr_slabs);
4896 if (flags & SO_PARTIAL) {
4897 struct kmem_cache_node *n;
4899 for_each_kmem_cache_node(s, node, n) {
4900 if (flags & SO_TOTAL)
4901 x = count_partial(n, count_total);
4902 else if (flags & SO_OBJECTS)
4903 x = count_partial(n, count_inuse);
4910 x = sprintf(buf, "%lu", total);
4912 for (node = 0; node < nr_node_ids; node++)
4914 x += sprintf(buf + x, " N%d=%lu",
4918 return x + sprintf(buf + x, "\n");
4921 #ifdef CONFIG_SLUB_DEBUG
4922 static int any_slab_objects(struct kmem_cache *s)
4925 struct kmem_cache_node *n;
4927 for_each_kmem_cache_node(s, node, n)
4928 if (atomic_long_read(&n->total_objects))
4935 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4936 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4938 struct slab_attribute {
4939 struct attribute attr;
4940 ssize_t (*show)(struct kmem_cache *s, char *buf);
4941 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4944 #define SLAB_ATTR_RO(_name) \
4945 static struct slab_attribute _name##_attr = \
4946 __ATTR(_name, 0400, _name##_show, NULL)
4948 #define SLAB_ATTR(_name) \
4949 static struct slab_attribute _name##_attr = \
4950 __ATTR(_name, 0600, _name##_show, _name##_store)
4952 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4954 return sprintf(buf, "%u\n", s->size);
4956 SLAB_ATTR_RO(slab_size);
4958 static ssize_t align_show(struct kmem_cache *s, char *buf)
4960 return sprintf(buf, "%u\n", s->align);
4962 SLAB_ATTR_RO(align);
4964 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4966 return sprintf(buf, "%u\n", s->object_size);
4968 SLAB_ATTR_RO(object_size);
4970 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4972 return sprintf(buf, "%u\n", oo_objects(s->oo));
4974 SLAB_ATTR_RO(objs_per_slab);
4976 static ssize_t order_store(struct kmem_cache *s,
4977 const char *buf, size_t length)
4982 err = kstrtouint(buf, 10, &order);
4986 if (order > slub_max_order || order < slub_min_order)
4989 calculate_sizes(s, order);
4993 static ssize_t order_show(struct kmem_cache *s, char *buf)
4995 return sprintf(buf, "%u\n", oo_order(s->oo));
4999 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5001 return sprintf(buf, "%lu\n", s->min_partial);
5004 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5010 err = kstrtoul(buf, 10, &min);
5014 set_min_partial(s, min);
5017 SLAB_ATTR(min_partial);
5019 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5021 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5024 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5027 unsigned int objects;
5030 err = kstrtouint(buf, 10, &objects);
5033 if (objects && !kmem_cache_has_cpu_partial(s))
5036 slub_set_cpu_partial(s, objects);
5040 SLAB_ATTR(cpu_partial);
5042 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5046 return sprintf(buf, "%pS\n", s->ctor);
5050 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5052 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5054 SLAB_ATTR_RO(aliases);
5056 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5058 return show_slab_objects(s, buf, SO_PARTIAL);
5060 SLAB_ATTR_RO(partial);
5062 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5064 return show_slab_objects(s, buf, SO_CPU);
5066 SLAB_ATTR_RO(cpu_slabs);
5068 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5070 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5072 SLAB_ATTR_RO(objects);
5074 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5076 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5078 SLAB_ATTR_RO(objects_partial);
5080 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5087 for_each_online_cpu(cpu) {
5090 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5093 pages += page->pages;
5094 objects += page->pobjects;
5098 len = sprintf(buf, "%d(%d)", objects, pages);
5101 for_each_online_cpu(cpu) {
5104 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5106 if (page && len < PAGE_SIZE - 20)
5107 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5108 page->pobjects, page->pages);
5111 return len + sprintf(buf + len, "\n");
5113 SLAB_ATTR_RO(slabs_cpu_partial);
5115 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5117 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5120 static ssize_t reclaim_account_store(struct kmem_cache *s,
5121 const char *buf, size_t length)
5123 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5125 s->flags |= SLAB_RECLAIM_ACCOUNT;
5128 SLAB_ATTR(reclaim_account);
5130 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5132 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5134 SLAB_ATTR_RO(hwcache_align);
5136 #ifdef CONFIG_ZONE_DMA
5137 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5139 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5141 SLAB_ATTR_RO(cache_dma);
5144 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5146 return sprintf(buf, "%u\n", s->usersize);
5148 SLAB_ATTR_RO(usersize);
5150 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5152 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5154 SLAB_ATTR_RO(destroy_by_rcu);
5156 #ifdef CONFIG_SLUB_DEBUG
5157 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5159 return show_slab_objects(s, buf, SO_ALL);
5161 SLAB_ATTR_RO(slabs);
5163 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5165 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5167 SLAB_ATTR_RO(total_objects);
5169 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5171 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5174 static ssize_t sanity_checks_store(struct kmem_cache *s,
5175 const char *buf, size_t length)
5177 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5178 if (buf[0] == '1') {
5179 s->flags &= ~__CMPXCHG_DOUBLE;
5180 s->flags |= SLAB_CONSISTENCY_CHECKS;
5184 SLAB_ATTR(sanity_checks);
5186 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5188 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5191 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5195 * Tracing a merged cache is going to give confusing results
5196 * as well as cause other issues like converting a mergeable
5197 * cache into an umergeable one.
5199 if (s->refcount > 1)
5202 s->flags &= ~SLAB_TRACE;
5203 if (buf[0] == '1') {
5204 s->flags &= ~__CMPXCHG_DOUBLE;
5205 s->flags |= SLAB_TRACE;
5211 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5213 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5216 static ssize_t red_zone_store(struct kmem_cache *s,
5217 const char *buf, size_t length)
5219 if (any_slab_objects(s))
5222 s->flags &= ~SLAB_RED_ZONE;
5223 if (buf[0] == '1') {
5224 s->flags |= SLAB_RED_ZONE;
5226 calculate_sizes(s, -1);
5229 SLAB_ATTR(red_zone);
5231 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5233 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5236 static ssize_t poison_store(struct kmem_cache *s,
5237 const char *buf, size_t length)
5239 if (any_slab_objects(s))
5242 s->flags &= ~SLAB_POISON;
5243 if (buf[0] == '1') {
5244 s->flags |= SLAB_POISON;
5246 calculate_sizes(s, -1);
5251 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5253 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5256 static ssize_t store_user_store(struct kmem_cache *s,
5257 const char *buf, size_t length)
5259 if (any_slab_objects(s))
5262 s->flags &= ~SLAB_STORE_USER;
5263 if (buf[0] == '1') {
5264 s->flags &= ~__CMPXCHG_DOUBLE;
5265 s->flags |= SLAB_STORE_USER;
5267 calculate_sizes(s, -1);
5270 SLAB_ATTR(store_user);
5272 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5277 static ssize_t validate_store(struct kmem_cache *s,
5278 const char *buf, size_t length)
5282 if (buf[0] == '1') {
5283 ret = validate_slab_cache(s);
5289 SLAB_ATTR(validate);
5291 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5293 if (!(s->flags & SLAB_STORE_USER))
5295 return list_locations(s, buf, TRACK_ALLOC);
5297 SLAB_ATTR_RO(alloc_calls);
5299 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5301 if (!(s->flags & SLAB_STORE_USER))
5303 return list_locations(s, buf, TRACK_FREE);
5305 SLAB_ATTR_RO(free_calls);
5306 #endif /* CONFIG_SLUB_DEBUG */
5308 #ifdef CONFIG_FAILSLAB
5309 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5311 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5314 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5317 if (s->refcount > 1)
5320 s->flags &= ~SLAB_FAILSLAB;
5322 s->flags |= SLAB_FAILSLAB;
5325 SLAB_ATTR(failslab);
5328 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5333 static ssize_t shrink_store(struct kmem_cache *s,
5334 const char *buf, size_t length)
5337 kmem_cache_shrink_all(s);
5345 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5347 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5350 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5351 const char *buf, size_t length)
5356 err = kstrtouint(buf, 10, &ratio);
5362 s->remote_node_defrag_ratio = ratio * 10;
5366 SLAB_ATTR(remote_node_defrag_ratio);
5369 #ifdef CONFIG_SLUB_STATS
5370 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5372 unsigned long sum = 0;
5375 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5380 for_each_online_cpu(cpu) {
5381 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5387 len = sprintf(buf, "%lu", sum);
5390 for_each_online_cpu(cpu) {
5391 if (data[cpu] && len < PAGE_SIZE - 20)
5392 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5396 return len + sprintf(buf + len, "\n");
5399 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5403 for_each_online_cpu(cpu)
5404 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5407 #define STAT_ATTR(si, text) \
5408 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5410 return show_stat(s, buf, si); \
5412 static ssize_t text##_store(struct kmem_cache *s, \
5413 const char *buf, size_t length) \
5415 if (buf[0] != '0') \
5417 clear_stat(s, si); \
5422 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5423 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5424 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5425 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5426 STAT_ATTR(FREE_FROZEN, free_frozen);
5427 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5428 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5429 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5430 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5431 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5432 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5433 STAT_ATTR(FREE_SLAB, free_slab);
5434 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5435 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5436 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5437 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5438 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5439 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5440 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5441 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5442 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5443 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5444 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5445 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5446 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5447 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5448 #endif /* CONFIG_SLUB_STATS */
5450 static struct attribute *slab_attrs[] = {
5451 &slab_size_attr.attr,
5452 &object_size_attr.attr,
5453 &objs_per_slab_attr.attr,
5455 &min_partial_attr.attr,
5456 &cpu_partial_attr.attr,
5458 &objects_partial_attr.attr,
5460 &cpu_slabs_attr.attr,
5464 &hwcache_align_attr.attr,
5465 &reclaim_account_attr.attr,
5466 &destroy_by_rcu_attr.attr,
5468 &slabs_cpu_partial_attr.attr,
5469 #ifdef CONFIG_SLUB_DEBUG
5470 &total_objects_attr.attr,
5472 &sanity_checks_attr.attr,
5474 &red_zone_attr.attr,
5476 &store_user_attr.attr,
5477 &validate_attr.attr,
5478 &alloc_calls_attr.attr,
5479 &free_calls_attr.attr,
5481 #ifdef CONFIG_ZONE_DMA
5482 &cache_dma_attr.attr,
5485 &remote_node_defrag_ratio_attr.attr,
5487 #ifdef CONFIG_SLUB_STATS
5488 &alloc_fastpath_attr.attr,
5489 &alloc_slowpath_attr.attr,
5490 &free_fastpath_attr.attr,
5491 &free_slowpath_attr.attr,
5492 &free_frozen_attr.attr,
5493 &free_add_partial_attr.attr,
5494 &free_remove_partial_attr.attr,
5495 &alloc_from_partial_attr.attr,
5496 &alloc_slab_attr.attr,
5497 &alloc_refill_attr.attr,
5498 &alloc_node_mismatch_attr.attr,
5499 &free_slab_attr.attr,
5500 &cpuslab_flush_attr.attr,
5501 &deactivate_full_attr.attr,
5502 &deactivate_empty_attr.attr,
5503 &deactivate_to_head_attr.attr,
5504 &deactivate_to_tail_attr.attr,
5505 &deactivate_remote_frees_attr.attr,
5506 &deactivate_bypass_attr.attr,
5507 &order_fallback_attr.attr,
5508 &cmpxchg_double_fail_attr.attr,
5509 &cmpxchg_double_cpu_fail_attr.attr,
5510 &cpu_partial_alloc_attr.attr,
5511 &cpu_partial_free_attr.attr,
5512 &cpu_partial_node_attr.attr,
5513 &cpu_partial_drain_attr.attr,
5515 #ifdef CONFIG_FAILSLAB
5516 &failslab_attr.attr,
5518 &usersize_attr.attr,
5523 static const struct attribute_group slab_attr_group = {
5524 .attrs = slab_attrs,
5527 static ssize_t slab_attr_show(struct kobject *kobj,
5528 struct attribute *attr,
5531 struct slab_attribute *attribute;
5532 struct kmem_cache *s;
5535 attribute = to_slab_attr(attr);
5538 if (!attribute->show)
5541 err = attribute->show(s, buf);
5546 static ssize_t slab_attr_store(struct kobject *kobj,
5547 struct attribute *attr,
5548 const char *buf, size_t len)
5550 struct slab_attribute *attribute;
5551 struct kmem_cache *s;
5554 attribute = to_slab_attr(attr);
5557 if (!attribute->store)
5560 err = attribute->store(s, buf, len);
5562 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5563 struct kmem_cache *c;
5565 mutex_lock(&slab_mutex);
5566 if (s->max_attr_size < len)
5567 s->max_attr_size = len;
5570 * This is a best effort propagation, so this function's return
5571 * value will be determined by the parent cache only. This is
5572 * basically because not all attributes will have a well
5573 * defined semantics for rollbacks - most of the actions will
5574 * have permanent effects.
5576 * Returning the error value of any of the children that fail
5577 * is not 100 % defined, in the sense that users seeing the
5578 * error code won't be able to know anything about the state of
5581 * Only returning the error code for the parent cache at least
5582 * has well defined semantics. The cache being written to
5583 * directly either failed or succeeded, in which case we loop
5584 * through the descendants with best-effort propagation.
5586 for_each_memcg_cache(c, s)
5587 attribute->store(c, buf, len);
5588 mutex_unlock(&slab_mutex);
5594 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5598 char *buffer = NULL;
5599 struct kmem_cache *root_cache;
5601 if (is_root_cache(s))
5604 root_cache = s->memcg_params.root_cache;
5607 * This mean this cache had no attribute written. Therefore, no point
5608 * in copying default values around
5610 if (!root_cache->max_attr_size)
5613 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5616 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5619 if (!attr || !attr->store || !attr->show)
5623 * It is really bad that we have to allocate here, so we will
5624 * do it only as a fallback. If we actually allocate, though,
5625 * we can just use the allocated buffer until the end.
5627 * Most of the slub attributes will tend to be very small in
5628 * size, but sysfs allows buffers up to a page, so they can
5629 * theoretically happen.
5633 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5636 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5637 if (WARN_ON(!buffer))
5642 len = attr->show(root_cache, buf);
5644 attr->store(s, buf, len);
5648 free_page((unsigned long)buffer);
5649 #endif /* CONFIG_MEMCG */
5652 static void kmem_cache_release(struct kobject *k)
5654 slab_kmem_cache_release(to_slab(k));
5657 static const struct sysfs_ops slab_sysfs_ops = {
5658 .show = slab_attr_show,
5659 .store = slab_attr_store,
5662 static struct kobj_type slab_ktype = {
5663 .sysfs_ops = &slab_sysfs_ops,
5664 .release = kmem_cache_release,
5667 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5669 struct kobj_type *ktype = get_ktype(kobj);
5671 if (ktype == &slab_ktype)
5676 static const struct kset_uevent_ops slab_uevent_ops = {
5677 .filter = uevent_filter,
5680 static struct kset *slab_kset;
5682 static inline struct kset *cache_kset(struct kmem_cache *s)
5685 if (!is_root_cache(s))
5686 return s->memcg_params.root_cache->memcg_kset;
5691 #define ID_STR_LENGTH 64
5693 /* Create a unique string id for a slab cache:
5695 * Format :[flags-]size
5697 static char *create_unique_id(struct kmem_cache *s)
5699 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5706 * First flags affecting slabcache operations. We will only
5707 * get here for aliasable slabs so we do not need to support
5708 * too many flags. The flags here must cover all flags that
5709 * are matched during merging to guarantee that the id is
5712 if (s->flags & SLAB_CACHE_DMA)
5714 if (s->flags & SLAB_CACHE_DMA32)
5716 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5718 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5720 if (s->flags & SLAB_ACCOUNT)
5724 p += sprintf(p, "%07u", s->size);
5726 BUG_ON(p > name + ID_STR_LENGTH - 1);
5730 static void sysfs_slab_remove_workfn(struct work_struct *work)
5732 struct kmem_cache *s =
5733 container_of(work, struct kmem_cache, kobj_remove_work);
5735 if (!s->kobj.state_in_sysfs)
5737 * For a memcg cache, this may be called during
5738 * deactivation and again on shutdown. Remove only once.
5739 * A cache is never shut down before deactivation is
5740 * complete, so no need to worry about synchronization.
5745 kset_unregister(s->memcg_kset);
5747 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5749 kobject_put(&s->kobj);
5752 static int sysfs_slab_add(struct kmem_cache *s)
5756 struct kset *kset = cache_kset(s);
5757 int unmergeable = slab_unmergeable(s);
5759 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5762 kobject_init(&s->kobj, &slab_ktype);
5766 if (!unmergeable && disable_higher_order_debug &&
5767 (slub_debug & DEBUG_METADATA_FLAGS))
5772 * Slabcache can never be merged so we can use the name proper.
5773 * This is typically the case for debug situations. In that
5774 * case we can catch duplicate names easily.
5776 sysfs_remove_link(&slab_kset->kobj, s->name);
5780 * Create a unique name for the slab as a target
5783 name = create_unique_id(s);
5786 s->kobj.kset = kset;
5787 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5791 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5796 if (is_root_cache(s) && memcg_sysfs_enabled) {
5797 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5798 if (!s->memcg_kset) {
5805 kobject_uevent(&s->kobj, KOBJ_ADD);
5807 /* Setup first alias */
5808 sysfs_slab_alias(s, s->name);
5815 kobject_del(&s->kobj);
5819 static void sysfs_slab_remove(struct kmem_cache *s)
5821 if (slab_state < FULL)
5823 * Sysfs has not been setup yet so no need to remove the
5828 kobject_get(&s->kobj);
5829 schedule_work(&s->kobj_remove_work);
5832 void sysfs_slab_unlink(struct kmem_cache *s)
5834 if (slab_state >= FULL)
5835 kobject_del(&s->kobj);
5838 void sysfs_slab_release(struct kmem_cache *s)
5840 if (slab_state >= FULL)
5841 kobject_put(&s->kobj);
5845 * Need to buffer aliases during bootup until sysfs becomes
5846 * available lest we lose that information.
5848 struct saved_alias {
5849 struct kmem_cache *s;
5851 struct saved_alias *next;
5854 static struct saved_alias *alias_list;
5856 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5858 struct saved_alias *al;
5860 if (slab_state == FULL) {
5862 * If we have a leftover link then remove it.
5864 sysfs_remove_link(&slab_kset->kobj, name);
5865 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5868 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5874 al->next = alias_list;
5879 static int __init slab_sysfs_init(void)
5881 struct kmem_cache *s;
5884 mutex_lock(&slab_mutex);
5886 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5888 mutex_unlock(&slab_mutex);
5889 pr_err("Cannot register slab subsystem.\n");
5895 list_for_each_entry(s, &slab_caches, list) {
5896 err = sysfs_slab_add(s);
5898 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5902 while (alias_list) {
5903 struct saved_alias *al = alias_list;
5905 alias_list = alias_list->next;
5906 err = sysfs_slab_alias(al->s, al->name);
5908 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5913 mutex_unlock(&slab_mutex);
5918 __initcall(slab_sysfs_init);
5919 #endif /* CONFIG_SYSFS */
5922 * The /proc/slabinfo ABI
5924 #ifdef CONFIG_SLUB_DEBUG
5925 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5927 unsigned long nr_slabs = 0;
5928 unsigned long nr_objs = 0;
5929 unsigned long nr_free = 0;
5931 struct kmem_cache_node *n;
5933 for_each_kmem_cache_node(s, node, n) {
5934 nr_slabs += node_nr_slabs(n);
5935 nr_objs += node_nr_objs(n);
5936 nr_free += count_partial(n, count_free);
5939 sinfo->active_objs = nr_objs - nr_free;
5940 sinfo->num_objs = nr_objs;
5941 sinfo->active_slabs = nr_slabs;
5942 sinfo->num_slabs = nr_slabs;
5943 sinfo->objects_per_slab = oo_objects(s->oo);
5944 sinfo->cache_order = oo_order(s->oo);
5947 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5951 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5952 size_t count, loff_t *ppos)
5956 #endif /* CONFIG_SLUB_DEBUG */