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 operations
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/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
22 #include <linux/proc_fs.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/kfence.h>
32 #include <linux/memory.h>
33 #include <linux/math64.h>
34 #include <linux/fault-inject.h>
35 #include <linux/stacktrace.h>
36 #include <linux/prefetch.h>
37 #include <linux/memcontrol.h>
38 #include <linux/random.h>
39 #include <kunit/test.h>
41 #include <linux/debugfs.h>
42 #include <trace/events/kmem.h>
48 * 1. slab_mutex (Global Mutex)
50 * 3. slab_lock(page) (Only on some arches and for debugging)
54 * The role of the slab_mutex is to protect the list of all the slabs
55 * and to synchronize major metadata changes to slab cache structures.
57 * The slab_lock is only used for debugging and on arches that do not
58 * have the ability to do a cmpxchg_double. It only protects:
59 * A. page->freelist -> List of object free in a page
60 * B. page->inuse -> Number of objects in use
61 * C. page->objects -> Number of objects in page
62 * D. page->frozen -> frozen state
64 * If a slab is frozen then it is exempt from list management. It is not
65 * on any list except per cpu partial list. The processor that froze the
66 * slab is the one who can perform list operations on the page. Other
67 * processors may put objects onto the freelist but the processor that
68 * froze the slab is the only one that can retrieve the objects from the
71 * The list_lock protects the partial and full list on each node and
72 * the partial slab counter. If taken then no new slabs may be added or
73 * removed from the lists nor make the number of partial slabs be modified.
74 * (Note that the total number of slabs is an atomic value that may be
75 * modified without taking the list lock).
77 * The list_lock is a centralized lock and thus we avoid taking it as
78 * much as possible. As long as SLUB does not have to handle partial
79 * slabs, operations can continue without any centralized lock. F.e.
80 * allocating a long series of objects that fill up slabs does not require
82 * Interrupts are disabled during allocation and deallocation in order to
83 * make the slab allocator safe to use in the context of an irq. In addition
84 * interrupts are disabled to ensure that the processor does not change
85 * while handling per_cpu slabs, due to kernel preemption.
87 * SLUB assigns one slab for allocation to each processor.
88 * Allocations only occur from these slabs called cpu slabs.
90 * Slabs with free elements are kept on a partial list and during regular
91 * operations no list for full slabs is used. If an object in a full slab is
92 * freed then the slab will show up again on the partial lists.
93 * We track full slabs for debugging purposes though because otherwise we
94 * cannot scan all objects.
96 * Slabs are freed when they become empty. Teardown and setup is
97 * minimal so we rely on the page allocators per cpu caches for
98 * fast frees and allocs.
100 * page->frozen The slab is frozen and exempt from list processing.
101 * This means that the slab is dedicated to a purpose
102 * such as satisfying allocations for a specific
103 * processor. Objects may be freed in the slab while
104 * it is frozen but slab_free will then skip the usual
105 * list operations. It is up to the processor holding
106 * the slab to integrate the slab into the slab lists
107 * when the slab is no longer needed.
109 * One use of this flag is to mark slabs that are
110 * used for allocations. Then such a slab becomes a cpu
111 * slab. The cpu slab may be equipped with an additional
112 * freelist that allows lockless access to
113 * free objects in addition to the regular freelist
114 * that requires the slab lock.
116 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
117 * options set. This moves slab handling out of
118 * the fast path and disables lockless freelists.
121 #ifdef CONFIG_SLUB_DEBUG
122 #ifdef CONFIG_SLUB_DEBUG_ON
123 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
125 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
127 #endif /* CONFIG_SLUB_DEBUG */
129 static inline bool kmem_cache_debug(struct kmem_cache *s)
131 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
134 void *fixup_red_left(struct kmem_cache *s, void *p)
136 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
137 p += s->red_left_pad;
142 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
144 #ifdef CONFIG_SLUB_CPU_PARTIAL
145 return !kmem_cache_debug(s);
152 * Issues still to be resolved:
154 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
156 * - Variable sizing of the per node arrays
159 /* Enable to log cmpxchg failures */
160 #undef SLUB_DEBUG_CMPXCHG
163 * Minimum number of partial slabs. These will be left on the partial
164 * lists even if they are empty. kmem_cache_shrink may reclaim them.
166 #define MIN_PARTIAL 5
169 * Maximum number of desirable partial slabs.
170 * The existence of more partial slabs makes kmem_cache_shrink
171 * sort the partial list by the number of objects in use.
173 #define MAX_PARTIAL 10
175 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_STORE_USER)
179 * These debug flags cannot use CMPXCHG because there might be consistency
180 * issues when checking or reading debug information
182 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
187 * Debugging flags that require metadata to be stored in the slab. These get
188 * disabled when slub_debug=O is used and a cache's min order increases with
191 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
194 #define OO_MASK ((1 << OO_SHIFT) - 1)
195 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
197 /* Internal SLUB flags */
199 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
200 /* Use cmpxchg_double */
201 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
204 * Tracking user of a slab.
206 #define TRACK_ADDRS_COUNT 16
208 unsigned long addr; /* Called from address */
209 #ifdef CONFIG_STACKTRACE
210 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
212 int cpu; /* Was running on cpu */
213 int pid; /* Pid context */
214 unsigned long when; /* When did the operation occur */
217 enum track_item { TRACK_ALLOC, TRACK_FREE };
220 static int sysfs_slab_add(struct kmem_cache *);
221 static int sysfs_slab_alias(struct kmem_cache *, const char *);
223 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
224 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
228 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
229 static void debugfs_slab_add(struct kmem_cache *);
231 static inline void debugfs_slab_add(struct kmem_cache *s) { }
234 static inline void stat(const struct kmem_cache *s, enum stat_item si)
236 #ifdef CONFIG_SLUB_STATS
238 * The rmw is racy on a preemptible kernel but this is acceptable, so
239 * avoid this_cpu_add()'s irq-disable overhead.
241 raw_cpu_inc(s->cpu_slab->stat[si]);
246 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
247 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
248 * differ during memory hotplug/hotremove operations.
249 * Protected by slab_mutex.
251 static nodemask_t slab_nodes;
253 /********************************************************************
254 * Core slab cache functions
255 *******************************************************************/
258 * Returns freelist pointer (ptr). With hardening, this is obfuscated
259 * with an XOR of the address where the pointer is held and a per-cache
262 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
263 unsigned long ptr_addr)
265 #ifdef CONFIG_SLAB_FREELIST_HARDENED
267 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
268 * Normally, this doesn't cause any issues, as both set_freepointer()
269 * and get_freepointer() are called with a pointer with the same tag.
270 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
271 * example, when __free_slub() iterates over objects in a cache, it
272 * passes untagged pointers to check_object(). check_object() in turns
273 * calls get_freepointer() with an untagged pointer, which causes the
274 * freepointer to be restored incorrectly.
276 return (void *)((unsigned long)ptr ^ s->random ^
277 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
283 /* Returns the freelist pointer recorded at location ptr_addr. */
284 static inline void *freelist_dereference(const struct kmem_cache *s,
287 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
288 (unsigned long)ptr_addr);
291 static inline void *get_freepointer(struct kmem_cache *s, void *object)
293 object = kasan_reset_tag(object);
294 return freelist_dereference(s, object + s->offset);
297 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
299 prefetch(object + s->offset);
302 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
304 unsigned long freepointer_addr;
307 if (!debug_pagealloc_enabled_static())
308 return get_freepointer(s, object);
310 object = kasan_reset_tag(object);
311 freepointer_addr = (unsigned long)object + s->offset;
312 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
313 return freelist_ptr(s, p, freepointer_addr);
316 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
318 unsigned long freeptr_addr = (unsigned long)object + s->offset;
320 #ifdef CONFIG_SLAB_FREELIST_HARDENED
321 BUG_ON(object == fp); /* naive detection of double free or corruption */
324 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
325 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
328 /* Loop over all objects in a slab */
329 #define for_each_object(__p, __s, __addr, __objects) \
330 for (__p = fixup_red_left(__s, __addr); \
331 __p < (__addr) + (__objects) * (__s)->size; \
334 static inline unsigned int order_objects(unsigned int order, unsigned int size)
336 return ((unsigned int)PAGE_SIZE << order) / size;
339 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
342 struct kmem_cache_order_objects x = {
343 (order << OO_SHIFT) + order_objects(order, size)
349 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
351 return x.x >> OO_SHIFT;
354 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
356 return x.x & OO_MASK;
360 * Per slab locking using the pagelock
362 static __always_inline void slab_lock(struct page *page)
364 VM_BUG_ON_PAGE(PageTail(page), page);
365 bit_spin_lock(PG_locked, &page->flags);
368 static __always_inline void slab_unlock(struct page *page)
370 VM_BUG_ON_PAGE(PageTail(page), page);
371 __bit_spin_unlock(PG_locked, &page->flags);
374 /* Interrupts must be disabled (for the fallback code to work right) */
375 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
380 VM_BUG_ON(!irqs_disabled());
381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s->flags & __CMPXCHG_DOUBLE) {
384 if (cmpxchg_double(&page->freelist, &page->counters,
385 freelist_old, counters_old,
386 freelist_new, counters_new))
392 if (page->freelist == freelist_old &&
393 page->counters == counters_old) {
394 page->freelist = freelist_new;
395 page->counters = counters_new;
403 stat(s, CMPXCHG_DOUBLE_FAIL);
405 #ifdef SLUB_DEBUG_CMPXCHG
406 pr_info("%s %s: cmpxchg double redo ", n, s->name);
412 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413 void *freelist_old, unsigned long counters_old,
414 void *freelist_new, unsigned long counters_new,
417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s->flags & __CMPXCHG_DOUBLE) {
420 if (cmpxchg_double(&page->freelist, &page->counters,
421 freelist_old, counters_old,
422 freelist_new, counters_new))
429 local_irq_save(flags);
431 if (page->freelist == freelist_old &&
432 page->counters == counters_old) {
433 page->freelist = freelist_new;
434 page->counters = counters_new;
436 local_irq_restore(flags);
440 local_irq_restore(flags);
444 stat(s, CMPXCHG_DOUBLE_FAIL);
446 #ifdef SLUB_DEBUG_CMPXCHG
447 pr_info("%s %s: cmpxchg double redo ", n, s->name);
453 #ifdef CONFIG_SLUB_DEBUG
454 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
455 static DEFINE_SPINLOCK(object_map_lock);
457 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
460 void *addr = page_address(page);
463 bitmap_zero(obj_map, page->objects);
465 for (p = page->freelist; p; p = get_freepointer(s, p))
466 set_bit(__obj_to_index(s, addr, p), obj_map);
469 #if IS_ENABLED(CONFIG_KUNIT)
470 static bool slab_add_kunit_errors(void)
472 struct kunit_resource *resource;
474 if (likely(!current->kunit_test))
477 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
481 (*(int *)resource->data)++;
482 kunit_put_resource(resource);
486 static inline bool slab_add_kunit_errors(void) { return false; }
490 * Determine a map of object in use on a page.
492 * Node listlock must be held to guarantee that the page does
493 * not vanish from under us.
495 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
496 __acquires(&object_map_lock)
498 VM_BUG_ON(!irqs_disabled());
500 spin_lock(&object_map_lock);
502 __fill_map(object_map, s, page);
507 static void put_map(unsigned long *map) __releases(&object_map_lock)
509 VM_BUG_ON(map != object_map);
510 spin_unlock(&object_map_lock);
513 static inline unsigned int size_from_object(struct kmem_cache *s)
515 if (s->flags & SLAB_RED_ZONE)
516 return s->size - s->red_left_pad;
521 static inline void *restore_red_left(struct kmem_cache *s, void *p)
523 if (s->flags & SLAB_RED_ZONE)
524 p -= s->red_left_pad;
532 #if defined(CONFIG_SLUB_DEBUG_ON)
533 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
535 static slab_flags_t slub_debug;
538 static char *slub_debug_string;
539 static int disable_higher_order_debug;
542 * slub is about to manipulate internal object metadata. This memory lies
543 * outside the range of the allocated object, so accessing it would normally
544 * be reported by kasan as a bounds error. metadata_access_enable() is used
545 * to tell kasan that these accesses are OK.
547 static inline void metadata_access_enable(void)
549 kasan_disable_current();
552 static inline void metadata_access_disable(void)
554 kasan_enable_current();
561 /* Verify that a pointer has an address that is valid within a slab page */
562 static inline int check_valid_pointer(struct kmem_cache *s,
563 struct page *page, void *object)
570 base = page_address(page);
571 object = kasan_reset_tag(object);
572 object = restore_red_left(s, object);
573 if (object < base || object >= base + page->objects * s->size ||
574 (object - base) % s->size) {
581 static void print_section(char *level, char *text, u8 *addr,
584 metadata_access_enable();
585 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
586 16, 1, kasan_reset_tag((void *)addr), length, 1);
587 metadata_access_disable();
591 * See comment in calculate_sizes().
593 static inline bool freeptr_outside_object(struct kmem_cache *s)
595 return s->offset >= s->inuse;
599 * Return offset of the end of info block which is inuse + free pointer if
600 * not overlapping with object.
602 static inline unsigned int get_info_end(struct kmem_cache *s)
604 if (freeptr_outside_object(s))
605 return s->inuse + sizeof(void *);
610 static struct track *get_track(struct kmem_cache *s, void *object,
611 enum track_item alloc)
615 p = object + get_info_end(s);
617 return kasan_reset_tag(p + alloc);
620 static void set_track(struct kmem_cache *s, void *object,
621 enum track_item alloc, unsigned long addr)
623 struct track *p = get_track(s, object, alloc);
626 #ifdef CONFIG_STACKTRACE
627 unsigned int nr_entries;
629 metadata_access_enable();
630 nr_entries = stack_trace_save(kasan_reset_tag(p->addrs),
631 TRACK_ADDRS_COUNT, 3);
632 metadata_access_disable();
634 if (nr_entries < TRACK_ADDRS_COUNT)
635 p->addrs[nr_entries] = 0;
638 p->cpu = smp_processor_id();
639 p->pid = current->pid;
642 memset(p, 0, sizeof(struct track));
646 static void init_tracking(struct kmem_cache *s, void *object)
648 if (!(s->flags & SLAB_STORE_USER))
651 set_track(s, object, TRACK_FREE, 0UL);
652 set_track(s, object, TRACK_ALLOC, 0UL);
655 static void print_track(const char *s, struct track *t, unsigned long pr_time)
660 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
661 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
662 #ifdef CONFIG_STACKTRACE
665 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
667 pr_err("\t%pS\n", (void *)t->addrs[i]);
674 void print_tracking(struct kmem_cache *s, void *object)
676 unsigned long pr_time = jiffies;
677 if (!(s->flags & SLAB_STORE_USER))
680 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
681 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
684 static void print_page_info(struct page *page)
686 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
687 page, page->objects, page->inuse, page->freelist,
688 page->flags, &page->flags);
692 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
694 struct va_format vaf;
700 pr_err("=============================================================================\n");
701 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
702 pr_err("-----------------------------------------------------------------------------\n\n");
707 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
709 struct va_format vaf;
712 if (slab_add_kunit_errors())
718 pr_err("FIX %s: %pV\n", s->name, &vaf);
722 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
723 void **freelist, void *nextfree)
725 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
726 !check_valid_pointer(s, page, nextfree) && freelist) {
727 object_err(s, page, *freelist, "Freechain corrupt");
729 slab_fix(s, "Isolate corrupted freechain");
736 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
738 unsigned int off; /* Offset of last byte */
739 u8 *addr = page_address(page);
741 print_tracking(s, p);
743 print_page_info(page);
745 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
746 p, p - addr, get_freepointer(s, p));
748 if (s->flags & SLAB_RED_ZONE)
749 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
751 else if (p > addr + 16)
752 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
754 print_section(KERN_ERR, "Object ", p,
755 min_t(unsigned int, s->object_size, PAGE_SIZE));
756 if (s->flags & SLAB_RED_ZONE)
757 print_section(KERN_ERR, "Redzone ", p + s->object_size,
758 s->inuse - s->object_size);
760 off = get_info_end(s);
762 if (s->flags & SLAB_STORE_USER)
763 off += 2 * sizeof(struct track);
765 off += kasan_metadata_size(s);
767 if (off != size_from_object(s))
768 /* Beginning of the filler is the free pointer */
769 print_section(KERN_ERR, "Padding ", p + off,
770 size_from_object(s) - off);
775 void object_err(struct kmem_cache *s, struct page *page,
776 u8 *object, char *reason)
778 if (slab_add_kunit_errors())
781 slab_bug(s, "%s", reason);
782 print_trailer(s, page, object);
783 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
786 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
787 const char *fmt, ...)
792 if (slab_add_kunit_errors())
796 vsnprintf(buf, sizeof(buf), fmt, args);
798 slab_bug(s, "%s", buf);
799 print_page_info(page);
801 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
804 static void init_object(struct kmem_cache *s, void *object, u8 val)
806 u8 *p = kasan_reset_tag(object);
808 if (s->flags & SLAB_RED_ZONE)
809 memset(p - s->red_left_pad, val, s->red_left_pad);
811 if (s->flags & __OBJECT_POISON) {
812 memset(p, POISON_FREE, s->object_size - 1);
813 p[s->object_size - 1] = POISON_END;
816 if (s->flags & SLAB_RED_ZONE)
817 memset(p + s->object_size, val, s->inuse - s->object_size);
820 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
821 void *from, void *to)
823 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
824 memset(from, data, to - from);
827 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
828 u8 *object, char *what,
829 u8 *start, unsigned int value, unsigned int bytes)
833 u8 *addr = page_address(page);
835 metadata_access_enable();
836 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
837 metadata_access_disable();
842 while (end > fault && end[-1] == value)
845 if (slab_add_kunit_errors())
848 slab_bug(s, "%s overwritten", what);
849 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
850 fault, end - 1, fault - addr,
852 print_trailer(s, page, object);
853 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
856 restore_bytes(s, what, value, fault, end);
864 * Bytes of the object to be managed.
865 * If the freepointer may overlay the object then the free
866 * pointer is at the middle of the object.
868 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
871 * object + s->object_size
872 * Padding to reach word boundary. This is also used for Redzoning.
873 * Padding is extended by another word if Redzoning is enabled and
874 * object_size == inuse.
876 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
877 * 0xcc (RED_ACTIVE) for objects in use.
880 * Meta data starts here.
882 * A. Free pointer (if we cannot overwrite object on free)
883 * B. Tracking data for SLAB_STORE_USER
884 * C. Padding to reach required alignment boundary or at minimum
885 * one word if debugging is on to be able to detect writes
886 * before the word boundary.
888 * Padding is done using 0x5a (POISON_INUSE)
891 * Nothing is used beyond s->size.
893 * If slabcaches are merged then the object_size and inuse boundaries are mostly
894 * ignored. And therefore no slab options that rely on these boundaries
895 * may be used with merged slabcaches.
898 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
900 unsigned long off = get_info_end(s); /* The end of info */
902 if (s->flags & SLAB_STORE_USER)
903 /* We also have user information there */
904 off += 2 * sizeof(struct track);
906 off += kasan_metadata_size(s);
908 if (size_from_object(s) == off)
911 return check_bytes_and_report(s, page, p, "Object padding",
912 p + off, POISON_INUSE, size_from_object(s) - off);
915 /* Check the pad bytes at the end of a slab page */
916 static int slab_pad_check(struct kmem_cache *s, struct page *page)
925 if (!(s->flags & SLAB_POISON))
928 start = page_address(page);
929 length = page_size(page);
930 end = start + length;
931 remainder = length % s->size;
935 pad = end - remainder;
936 metadata_access_enable();
937 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
938 metadata_access_disable();
941 while (end > fault && end[-1] == POISON_INUSE)
944 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
945 fault, end - 1, fault - start);
946 print_section(KERN_ERR, "Padding ", pad, remainder);
948 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
952 static int check_object(struct kmem_cache *s, struct page *page,
953 void *object, u8 val)
956 u8 *endobject = object + s->object_size;
958 if (s->flags & SLAB_RED_ZONE) {
959 if (!check_bytes_and_report(s, page, object, "Left Redzone",
960 object - s->red_left_pad, val, s->red_left_pad))
963 if (!check_bytes_and_report(s, page, object, "Right Redzone",
964 endobject, val, s->inuse - s->object_size))
967 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
968 check_bytes_and_report(s, page, p, "Alignment padding",
969 endobject, POISON_INUSE,
970 s->inuse - s->object_size);
974 if (s->flags & SLAB_POISON) {
975 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
976 (!check_bytes_and_report(s, page, p, "Poison", p,
977 POISON_FREE, s->object_size - 1) ||
978 !check_bytes_and_report(s, page, p, "End Poison",
979 p + s->object_size - 1, POISON_END, 1)))
982 * check_pad_bytes cleans up on its own.
984 check_pad_bytes(s, page, p);
987 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
989 * Object and freepointer overlap. Cannot check
990 * freepointer while object is allocated.
994 /* Check free pointer validity */
995 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
996 object_err(s, page, p, "Freepointer corrupt");
998 * No choice but to zap it and thus lose the remainder
999 * of the free objects in this slab. May cause
1000 * another error because the object count is now wrong.
1002 set_freepointer(s, p, NULL);
1008 static int check_slab(struct kmem_cache *s, struct page *page)
1012 VM_BUG_ON(!irqs_disabled());
1014 if (!PageSlab(page)) {
1015 slab_err(s, page, "Not a valid slab page");
1019 maxobj = order_objects(compound_order(page), s->size);
1020 if (page->objects > maxobj) {
1021 slab_err(s, page, "objects %u > max %u",
1022 page->objects, maxobj);
1025 if (page->inuse > page->objects) {
1026 slab_err(s, page, "inuse %u > max %u",
1027 page->inuse, page->objects);
1030 /* Slab_pad_check fixes things up after itself */
1031 slab_pad_check(s, page);
1036 * Determine if a certain object on a page is on the freelist. Must hold the
1037 * slab lock to guarantee that the chains are in a consistent state.
1039 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
1043 void *object = NULL;
1046 fp = page->freelist;
1047 while (fp && nr <= page->objects) {
1050 if (!check_valid_pointer(s, page, fp)) {
1052 object_err(s, page, object,
1053 "Freechain corrupt");
1054 set_freepointer(s, object, NULL);
1056 slab_err(s, page, "Freepointer corrupt");
1057 page->freelist = NULL;
1058 page->inuse = page->objects;
1059 slab_fix(s, "Freelist cleared");
1065 fp = get_freepointer(s, object);
1069 max_objects = order_objects(compound_order(page), s->size);
1070 if (max_objects > MAX_OBJS_PER_PAGE)
1071 max_objects = MAX_OBJS_PER_PAGE;
1073 if (page->objects != max_objects) {
1074 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1075 page->objects, max_objects);
1076 page->objects = max_objects;
1077 slab_fix(s, "Number of objects adjusted");
1079 if (page->inuse != page->objects - nr) {
1080 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1081 page->inuse, page->objects - nr);
1082 page->inuse = page->objects - nr;
1083 slab_fix(s, "Object count adjusted");
1085 return search == NULL;
1088 static void trace(struct kmem_cache *s, struct page *page, void *object,
1091 if (s->flags & SLAB_TRACE) {
1092 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1094 alloc ? "alloc" : "free",
1095 object, page->inuse,
1099 print_section(KERN_INFO, "Object ", (void *)object,
1107 * Tracking of fully allocated slabs for debugging purposes.
1109 static void add_full(struct kmem_cache *s,
1110 struct kmem_cache_node *n, struct page *page)
1112 if (!(s->flags & SLAB_STORE_USER))
1115 lockdep_assert_held(&n->list_lock);
1116 list_add(&page->slab_list, &n->full);
1119 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1121 if (!(s->flags & SLAB_STORE_USER))
1124 lockdep_assert_held(&n->list_lock);
1125 list_del(&page->slab_list);
1128 /* Tracking of the number of slabs for debugging purposes */
1129 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1131 struct kmem_cache_node *n = get_node(s, node);
1133 return atomic_long_read(&n->nr_slabs);
1136 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1138 return atomic_long_read(&n->nr_slabs);
1141 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1143 struct kmem_cache_node *n = get_node(s, node);
1146 * May be called early in order to allocate a slab for the
1147 * kmem_cache_node structure. Solve the chicken-egg
1148 * dilemma by deferring the increment of the count during
1149 * bootstrap (see early_kmem_cache_node_alloc).
1152 atomic_long_inc(&n->nr_slabs);
1153 atomic_long_add(objects, &n->total_objects);
1156 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1158 struct kmem_cache_node *n = get_node(s, node);
1160 atomic_long_dec(&n->nr_slabs);
1161 atomic_long_sub(objects, &n->total_objects);
1164 /* Object debug checks for alloc/free paths */
1165 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1168 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1171 init_object(s, object, SLUB_RED_INACTIVE);
1172 init_tracking(s, object);
1176 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1178 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1181 metadata_access_enable();
1182 memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page));
1183 metadata_access_disable();
1186 static inline int alloc_consistency_checks(struct kmem_cache *s,
1187 struct page *page, void *object)
1189 if (!check_slab(s, page))
1192 if (!check_valid_pointer(s, page, object)) {
1193 object_err(s, page, object, "Freelist Pointer check fails");
1197 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1203 static noinline int alloc_debug_processing(struct kmem_cache *s,
1205 void *object, unsigned long addr)
1207 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1208 if (!alloc_consistency_checks(s, page, object))
1212 /* Success perform special debug activities for allocs */
1213 if (s->flags & SLAB_STORE_USER)
1214 set_track(s, object, TRACK_ALLOC, addr);
1215 trace(s, page, object, 1);
1216 init_object(s, object, SLUB_RED_ACTIVE);
1220 if (PageSlab(page)) {
1222 * If this is a slab page then lets do the best we can
1223 * to avoid issues in the future. Marking all objects
1224 * as used avoids touching the remaining objects.
1226 slab_fix(s, "Marking all objects used");
1227 page->inuse = page->objects;
1228 page->freelist = NULL;
1233 static inline int free_consistency_checks(struct kmem_cache *s,
1234 struct page *page, void *object, unsigned long addr)
1236 if (!check_valid_pointer(s, page, object)) {
1237 slab_err(s, page, "Invalid object pointer 0x%p", object);
1241 if (on_freelist(s, page, object)) {
1242 object_err(s, page, object, "Object already free");
1246 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1249 if (unlikely(s != page->slab_cache)) {
1250 if (!PageSlab(page)) {
1251 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1253 } else if (!page->slab_cache) {
1254 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1258 object_err(s, page, object,
1259 "page slab pointer corrupt.");
1265 /* Supports checking bulk free of a constructed freelist */
1266 static noinline int free_debug_processing(
1267 struct kmem_cache *s, struct page *page,
1268 void *head, void *tail, int bulk_cnt,
1271 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1272 void *object = head;
1274 unsigned long flags;
1277 spin_lock_irqsave(&n->list_lock, flags);
1280 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1281 if (!check_slab(s, page))
1288 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1289 if (!free_consistency_checks(s, page, object, addr))
1293 if (s->flags & SLAB_STORE_USER)
1294 set_track(s, object, TRACK_FREE, addr);
1295 trace(s, page, object, 0);
1296 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1297 init_object(s, object, SLUB_RED_INACTIVE);
1299 /* Reached end of constructed freelist yet? */
1300 if (object != tail) {
1301 object = get_freepointer(s, object);
1307 if (cnt != bulk_cnt)
1308 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1312 spin_unlock_irqrestore(&n->list_lock, flags);
1314 slab_fix(s, "Object at 0x%p not freed", object);
1319 * Parse a block of slub_debug options. Blocks are delimited by ';'
1321 * @str: start of block
1322 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1323 * @slabs: return start of list of slabs, or NULL when there's no list
1324 * @init: assume this is initial parsing and not per-kmem-create parsing
1326 * returns the start of next block if there's any, or NULL
1329 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1331 bool higher_order_disable = false;
1333 /* Skip any completely empty blocks */
1334 while (*str && *str == ';')
1339 * No options but restriction on slabs. This means full
1340 * debugging for slabs matching a pattern.
1342 *flags = DEBUG_DEFAULT_FLAGS;
1347 /* Determine which debug features should be switched on */
1348 for (; *str && *str != ',' && *str != ';'; str++) {
1349 switch (tolower(*str)) {
1354 *flags |= SLAB_CONSISTENCY_CHECKS;
1357 *flags |= SLAB_RED_ZONE;
1360 *flags |= SLAB_POISON;
1363 *flags |= SLAB_STORE_USER;
1366 *flags |= SLAB_TRACE;
1369 *flags |= SLAB_FAILSLAB;
1373 * Avoid enabling debugging on caches if its minimum
1374 * order would increase as a result.
1376 higher_order_disable = true;
1380 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1389 /* Skip over the slab list */
1390 while (*str && *str != ';')
1393 /* Skip any completely empty blocks */
1394 while (*str && *str == ';')
1397 if (init && higher_order_disable)
1398 disable_higher_order_debug = 1;
1406 static int __init setup_slub_debug(char *str)
1409 slab_flags_t global_flags;
1412 bool global_slub_debug_changed = false;
1413 bool slab_list_specified = false;
1415 global_flags = DEBUG_DEFAULT_FLAGS;
1416 if (*str++ != '=' || !*str)
1418 * No options specified. Switch on full debugging.
1424 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1427 global_flags = flags;
1428 global_slub_debug_changed = true;
1430 slab_list_specified = true;
1435 * For backwards compatibility, a single list of flags with list of
1436 * slabs means debugging is only changed for those slabs, so the global
1437 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1438 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1439 * long as there is no option specifying flags without a slab list.
1441 if (slab_list_specified) {
1442 if (!global_slub_debug_changed)
1443 global_flags = slub_debug;
1444 slub_debug_string = saved_str;
1447 slub_debug = global_flags;
1448 if (slub_debug != 0 || slub_debug_string)
1449 static_branch_enable(&slub_debug_enabled);
1451 static_branch_disable(&slub_debug_enabled);
1452 if ((static_branch_unlikely(&init_on_alloc) ||
1453 static_branch_unlikely(&init_on_free)) &&
1454 (slub_debug & SLAB_POISON))
1455 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1459 __setup("slub_debug", setup_slub_debug);
1462 * kmem_cache_flags - apply debugging options to the cache
1463 * @object_size: the size of an object without meta data
1464 * @flags: flags to set
1465 * @name: name of the cache
1467 * Debug option(s) are applied to @flags. In addition to the debug
1468 * option(s), if a slab name (or multiple) is specified i.e.
1469 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1470 * then only the select slabs will receive the debug option(s).
1472 slab_flags_t kmem_cache_flags(unsigned int object_size,
1473 slab_flags_t flags, const char *name)
1478 slab_flags_t block_flags;
1479 slab_flags_t slub_debug_local = slub_debug;
1482 * If the slab cache is for debugging (e.g. kmemleak) then
1483 * don't store user (stack trace) information by default,
1484 * but let the user enable it via the command line below.
1486 if (flags & SLAB_NOLEAKTRACE)
1487 slub_debug_local &= ~SLAB_STORE_USER;
1490 next_block = slub_debug_string;
1491 /* Go through all blocks of debug options, see if any matches our slab's name */
1492 while (next_block) {
1493 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1496 /* Found a block that has a slab list, search it */
1501 end = strchrnul(iter, ',');
1502 if (next_block && next_block < end)
1503 end = next_block - 1;
1505 glob = strnchr(iter, end - iter, '*');
1507 cmplen = glob - iter;
1509 cmplen = max_t(size_t, len, (end - iter));
1511 if (!strncmp(name, iter, cmplen)) {
1512 flags |= block_flags;
1516 if (!*end || *end == ';')
1522 return flags | slub_debug_local;
1524 #else /* !CONFIG_SLUB_DEBUG */
1525 static inline void setup_object_debug(struct kmem_cache *s,
1526 struct page *page, void *object) {}
1528 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1530 static inline int alloc_debug_processing(struct kmem_cache *s,
1531 struct page *page, void *object, unsigned long addr) { return 0; }
1533 static inline int free_debug_processing(
1534 struct kmem_cache *s, struct page *page,
1535 void *head, void *tail, int bulk_cnt,
1536 unsigned long addr) { return 0; }
1538 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1540 static inline int check_object(struct kmem_cache *s, struct page *page,
1541 void *object, u8 val) { return 1; }
1542 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1543 struct page *page) {}
1544 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1545 struct page *page) {}
1546 slab_flags_t kmem_cache_flags(unsigned int object_size,
1547 slab_flags_t flags, const char *name)
1551 #define slub_debug 0
1553 #define disable_higher_order_debug 0
1555 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1557 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1559 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1561 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1564 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1565 void **freelist, void *nextfree)
1569 #endif /* CONFIG_SLUB_DEBUG */
1572 * Hooks for other subsystems that check memory allocations. In a typical
1573 * production configuration these hooks all should produce no code at all.
1575 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1577 ptr = kasan_kmalloc_large(ptr, size, flags);
1578 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1579 kmemleak_alloc(ptr, size, 1, flags);
1583 static __always_inline void kfree_hook(void *x)
1586 kasan_kfree_large(x);
1589 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1592 kmemleak_free_recursive(x, s->flags);
1594 debug_check_no_locks_freed(x, s->object_size);
1596 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1597 debug_check_no_obj_freed(x, s->object_size);
1599 /* Use KCSAN to help debug racy use-after-free. */
1600 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1601 __kcsan_check_access(x, s->object_size,
1602 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1605 * As memory initialization might be integrated into KASAN,
1606 * kasan_slab_free and initialization memset's must be
1607 * kept together to avoid discrepancies in behavior.
1609 * The initialization memset's clear the object and the metadata,
1610 * but don't touch the SLAB redzone.
1615 if (!kasan_has_integrated_init())
1616 memset(kasan_reset_tag(x), 0, s->object_size);
1617 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1618 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1619 s->size - s->inuse - rsize);
1621 /* KASAN might put x into memory quarantine, delaying its reuse. */
1622 return kasan_slab_free(s, x, init);
1625 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1626 void **head, void **tail)
1631 void *old_tail = *tail ? *tail : *head;
1633 if (is_kfence_address(next)) {
1634 slab_free_hook(s, next, false);
1638 /* Head and tail of the reconstructed freelist */
1644 next = get_freepointer(s, object);
1646 /* If object's reuse doesn't have to be delayed */
1647 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1648 /* Move object to the new freelist */
1649 set_freepointer(s, object, *head);
1654 } while (object != old_tail);
1659 return *head != NULL;
1662 static void *setup_object(struct kmem_cache *s, struct page *page,
1665 setup_object_debug(s, page, object);
1666 object = kasan_init_slab_obj(s, object);
1667 if (unlikely(s->ctor)) {
1668 kasan_unpoison_object_data(s, object);
1670 kasan_poison_object_data(s, object);
1676 * Slab allocation and freeing
1678 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1679 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1682 unsigned int order = oo_order(oo);
1684 if (node == NUMA_NO_NODE)
1685 page = alloc_pages(flags, order);
1687 page = __alloc_pages_node(node, flags, order);
1692 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1693 /* Pre-initialize the random sequence cache */
1694 static int init_cache_random_seq(struct kmem_cache *s)
1696 unsigned int count = oo_objects(s->oo);
1699 /* Bailout if already initialised */
1703 err = cache_random_seq_create(s, count, GFP_KERNEL);
1705 pr_err("SLUB: Unable to initialize free list for %s\n",
1710 /* Transform to an offset on the set of pages */
1711 if (s->random_seq) {
1714 for (i = 0; i < count; i++)
1715 s->random_seq[i] *= s->size;
1720 /* Initialize each random sequence freelist per cache */
1721 static void __init init_freelist_randomization(void)
1723 struct kmem_cache *s;
1725 mutex_lock(&slab_mutex);
1727 list_for_each_entry(s, &slab_caches, list)
1728 init_cache_random_seq(s);
1730 mutex_unlock(&slab_mutex);
1733 /* Get the next entry on the pre-computed freelist randomized */
1734 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1735 unsigned long *pos, void *start,
1736 unsigned long page_limit,
1737 unsigned long freelist_count)
1742 * If the target page allocation failed, the number of objects on the
1743 * page might be smaller than the usual size defined by the cache.
1746 idx = s->random_seq[*pos];
1748 if (*pos >= freelist_count)
1750 } while (unlikely(idx >= page_limit));
1752 return (char *)start + idx;
1755 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1756 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1761 unsigned long idx, pos, page_limit, freelist_count;
1763 if (page->objects < 2 || !s->random_seq)
1766 freelist_count = oo_objects(s->oo);
1767 pos = get_random_int() % freelist_count;
1769 page_limit = page->objects * s->size;
1770 start = fixup_red_left(s, page_address(page));
1772 /* First entry is used as the base of the freelist */
1773 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1775 cur = setup_object(s, page, cur);
1776 page->freelist = cur;
1778 for (idx = 1; idx < page->objects; idx++) {
1779 next = next_freelist_entry(s, page, &pos, start, page_limit,
1781 next = setup_object(s, page, next);
1782 set_freepointer(s, cur, next);
1785 set_freepointer(s, cur, NULL);
1790 static inline int init_cache_random_seq(struct kmem_cache *s)
1794 static inline void init_freelist_randomization(void) { }
1795 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1799 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1801 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1804 struct kmem_cache_order_objects oo = s->oo;
1806 void *start, *p, *next;
1810 flags &= gfp_allowed_mask;
1812 flags |= s->allocflags;
1815 * Let the initial higher-order allocation fail under memory pressure
1816 * so we fall-back to the minimum order allocation.
1818 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1819 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1820 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1822 page = alloc_slab_page(s, alloc_gfp, node, oo);
1823 if (unlikely(!page)) {
1827 * Allocation may have failed due to fragmentation.
1828 * Try a lower order alloc if possible
1830 page = alloc_slab_page(s, alloc_gfp, node, oo);
1831 if (unlikely(!page))
1833 stat(s, ORDER_FALLBACK);
1836 page->objects = oo_objects(oo);
1838 account_slab_page(page, oo_order(oo), s, flags);
1840 page->slab_cache = s;
1841 __SetPageSlab(page);
1842 if (page_is_pfmemalloc(page))
1843 SetPageSlabPfmemalloc(page);
1845 kasan_poison_slab(page);
1847 start = page_address(page);
1849 setup_page_debug(s, page, start);
1851 shuffle = shuffle_freelist(s, page);
1854 start = fixup_red_left(s, start);
1855 start = setup_object(s, page, start);
1856 page->freelist = start;
1857 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1859 next = setup_object(s, page, next);
1860 set_freepointer(s, p, next);
1863 set_freepointer(s, p, NULL);
1866 page->inuse = page->objects;
1873 inc_slabs_node(s, page_to_nid(page), page->objects);
1878 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1880 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1881 flags = kmalloc_fix_flags(flags);
1883 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
1885 return allocate_slab(s,
1886 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1889 static void __free_slab(struct kmem_cache *s, struct page *page)
1891 int order = compound_order(page);
1892 int pages = 1 << order;
1894 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1897 slab_pad_check(s, page);
1898 for_each_object(p, s, page_address(page),
1900 check_object(s, page, p, SLUB_RED_INACTIVE);
1903 __ClearPageSlabPfmemalloc(page);
1904 __ClearPageSlab(page);
1905 /* In union with page->mapping where page allocator expects NULL */
1906 page->slab_cache = NULL;
1907 if (current->reclaim_state)
1908 current->reclaim_state->reclaimed_slab += pages;
1909 unaccount_slab_page(page, order, s);
1910 __free_pages(page, order);
1913 static void rcu_free_slab(struct rcu_head *h)
1915 struct page *page = container_of(h, struct page, rcu_head);
1917 __free_slab(page->slab_cache, page);
1920 static void free_slab(struct kmem_cache *s, struct page *page)
1922 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1923 call_rcu(&page->rcu_head, rcu_free_slab);
1925 __free_slab(s, page);
1928 static void discard_slab(struct kmem_cache *s, struct page *page)
1930 dec_slabs_node(s, page_to_nid(page), page->objects);
1935 * Management of partially allocated slabs.
1938 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1941 if (tail == DEACTIVATE_TO_TAIL)
1942 list_add_tail(&page->slab_list, &n->partial);
1944 list_add(&page->slab_list, &n->partial);
1947 static inline void add_partial(struct kmem_cache_node *n,
1948 struct page *page, int tail)
1950 lockdep_assert_held(&n->list_lock);
1951 __add_partial(n, page, tail);
1954 static inline void remove_partial(struct kmem_cache_node *n,
1957 lockdep_assert_held(&n->list_lock);
1958 list_del(&page->slab_list);
1963 * Remove slab from the partial list, freeze it and
1964 * return the pointer to the freelist.
1966 * Returns a list of objects or NULL if it fails.
1968 static inline void *acquire_slab(struct kmem_cache *s,
1969 struct kmem_cache_node *n, struct page *page,
1970 int mode, int *objects)
1973 unsigned long counters;
1976 lockdep_assert_held(&n->list_lock);
1979 * Zap the freelist and set the frozen bit.
1980 * The old freelist is the list of objects for the
1981 * per cpu allocation list.
1983 freelist = page->freelist;
1984 counters = page->counters;
1985 new.counters = counters;
1986 *objects = new.objects - new.inuse;
1988 new.inuse = page->objects;
1989 new.freelist = NULL;
1991 new.freelist = freelist;
1994 VM_BUG_ON(new.frozen);
1997 if (!__cmpxchg_double_slab(s, page,
1999 new.freelist, new.counters,
2003 remove_partial(n, page);
2008 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
2009 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
2012 * Try to allocate a partial slab from a specific node.
2014 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2015 struct page **ret_page, gfp_t flags)
2017 struct page *page, *page2;
2018 void *object = NULL;
2019 unsigned int available = 0;
2023 * Racy check. If we mistakenly see no partial slabs then we
2024 * just allocate an empty slab. If we mistakenly try to get a
2025 * partial slab and there is none available then get_partial()
2028 if (!n || !n->nr_partial)
2031 spin_lock(&n->list_lock);
2032 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
2035 if (!pfmemalloc_match(page, flags))
2038 t = acquire_slab(s, n, page, object == NULL, &objects);
2042 available += objects;
2045 stat(s, ALLOC_FROM_PARTIAL);
2048 put_cpu_partial(s, page, 0);
2049 stat(s, CPU_PARTIAL_NODE);
2051 if (!kmem_cache_has_cpu_partial(s)
2052 || available > slub_cpu_partial(s) / 2)
2056 spin_unlock(&n->list_lock);
2061 * Get a page from somewhere. Search in increasing NUMA distances.
2063 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2064 struct page **ret_page)
2067 struct zonelist *zonelist;
2070 enum zone_type highest_zoneidx = gfp_zone(flags);
2072 unsigned int cpuset_mems_cookie;
2075 * The defrag ratio allows a configuration of the tradeoffs between
2076 * inter node defragmentation and node local allocations. A lower
2077 * defrag_ratio increases the tendency to do local allocations
2078 * instead of attempting to obtain partial slabs from other nodes.
2080 * If the defrag_ratio is set to 0 then kmalloc() always
2081 * returns node local objects. If the ratio is higher then kmalloc()
2082 * may return off node objects because partial slabs are obtained
2083 * from other nodes and filled up.
2085 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2086 * (which makes defrag_ratio = 1000) then every (well almost)
2087 * allocation will first attempt to defrag slab caches on other nodes.
2088 * This means scanning over all nodes to look for partial slabs which
2089 * may be expensive if we do it every time we are trying to find a slab
2090 * with available objects.
2092 if (!s->remote_node_defrag_ratio ||
2093 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2097 cpuset_mems_cookie = read_mems_allowed_begin();
2098 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2099 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2100 struct kmem_cache_node *n;
2102 n = get_node(s, zone_to_nid(zone));
2104 if (n && cpuset_zone_allowed(zone, flags) &&
2105 n->nr_partial > s->min_partial) {
2106 object = get_partial_node(s, n, ret_page, flags);
2109 * Don't check read_mems_allowed_retry()
2110 * here - if mems_allowed was updated in
2111 * parallel, that was a harmless race
2112 * between allocation and the cpuset
2119 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2120 #endif /* CONFIG_NUMA */
2125 * Get a partial page, lock it and return it.
2127 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2128 struct page **ret_page)
2131 int searchnode = node;
2133 if (node == NUMA_NO_NODE)
2134 searchnode = numa_mem_id();
2136 object = get_partial_node(s, get_node(s, searchnode), ret_page, flags);
2137 if (object || node != NUMA_NO_NODE)
2140 return get_any_partial(s, flags, ret_page);
2143 #ifdef CONFIG_PREEMPTION
2145 * Calculate the next globally unique transaction for disambiguation
2146 * during cmpxchg. The transactions start with the cpu number and are then
2147 * incremented by CONFIG_NR_CPUS.
2149 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2152 * No preemption supported therefore also no need to check for
2158 static inline unsigned long next_tid(unsigned long tid)
2160 return tid + TID_STEP;
2163 #ifdef SLUB_DEBUG_CMPXCHG
2164 static inline unsigned int tid_to_cpu(unsigned long tid)
2166 return tid % TID_STEP;
2169 static inline unsigned long tid_to_event(unsigned long tid)
2171 return tid / TID_STEP;
2175 static inline unsigned int init_tid(int cpu)
2180 static inline void note_cmpxchg_failure(const char *n,
2181 const struct kmem_cache *s, unsigned long tid)
2183 #ifdef SLUB_DEBUG_CMPXCHG
2184 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2186 pr_info("%s %s: cmpxchg redo ", n, s->name);
2188 #ifdef CONFIG_PREEMPTION
2189 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2190 pr_warn("due to cpu change %d -> %d\n",
2191 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2194 if (tid_to_event(tid) != tid_to_event(actual_tid))
2195 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2196 tid_to_event(tid), tid_to_event(actual_tid));
2198 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2199 actual_tid, tid, next_tid(tid));
2201 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2204 static void init_kmem_cache_cpus(struct kmem_cache *s)
2208 for_each_possible_cpu(cpu)
2209 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2213 * Remove the cpu slab
2215 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2216 void *freelist, struct kmem_cache_cpu *c)
2218 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2219 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2220 int lock = 0, free_delta = 0;
2221 enum slab_modes l = M_NONE, m = M_NONE;
2222 void *nextfree, *freelist_iter, *freelist_tail;
2223 int tail = DEACTIVATE_TO_HEAD;
2227 if (page->freelist) {
2228 stat(s, DEACTIVATE_REMOTE_FREES);
2229 tail = DEACTIVATE_TO_TAIL;
2233 * Stage one: Count the objects on cpu's freelist as free_delta and
2234 * remember the last object in freelist_tail for later splicing.
2236 freelist_tail = NULL;
2237 freelist_iter = freelist;
2238 while (freelist_iter) {
2239 nextfree = get_freepointer(s, freelist_iter);
2242 * If 'nextfree' is invalid, it is possible that the object at
2243 * 'freelist_iter' is already corrupted. So isolate all objects
2244 * starting at 'freelist_iter' by skipping them.
2246 if (freelist_corrupted(s, page, &freelist_iter, nextfree))
2249 freelist_tail = freelist_iter;
2252 freelist_iter = nextfree;
2256 * Stage two: Unfreeze the page while splicing the per-cpu
2257 * freelist to the head of page's freelist.
2259 * Ensure that the page is unfrozen while the list presence
2260 * reflects the actual number of objects during unfreeze.
2262 * We setup the list membership and then perform a cmpxchg
2263 * with the count. If there is a mismatch then the page
2264 * is not unfrozen but the page is on the wrong list.
2266 * Then we restart the process which may have to remove
2267 * the page from the list that we just put it on again
2268 * because the number of objects in the slab may have
2273 old.freelist = READ_ONCE(page->freelist);
2274 old.counters = READ_ONCE(page->counters);
2275 VM_BUG_ON(!old.frozen);
2277 /* Determine target state of the slab */
2278 new.counters = old.counters;
2279 if (freelist_tail) {
2280 new.inuse -= free_delta;
2281 set_freepointer(s, freelist_tail, old.freelist);
2282 new.freelist = freelist;
2284 new.freelist = old.freelist;
2288 if (!new.inuse && n->nr_partial >= s->min_partial)
2290 else if (new.freelist) {
2295 * Taking the spinlock removes the possibility
2296 * that acquire_slab() will see a slab page that
2299 spin_lock(&n->list_lock);
2303 if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) {
2306 * This also ensures that the scanning of full
2307 * slabs from diagnostic functions will not see
2310 spin_lock(&n->list_lock);
2316 remove_partial(n, page);
2317 else if (l == M_FULL)
2318 remove_full(s, n, page);
2321 add_partial(n, page, tail);
2322 else if (m == M_FULL)
2323 add_full(s, n, page);
2327 if (!__cmpxchg_double_slab(s, page,
2328 old.freelist, old.counters,
2329 new.freelist, new.counters,
2334 spin_unlock(&n->list_lock);
2338 else if (m == M_FULL)
2339 stat(s, DEACTIVATE_FULL);
2340 else if (m == M_FREE) {
2341 stat(s, DEACTIVATE_EMPTY);
2342 discard_slab(s, page);
2351 * Unfreeze all the cpu partial slabs.
2353 * This function must be called with interrupts disabled
2354 * for the cpu using c (or some other guarantee must be there
2355 * to guarantee no concurrent accesses).
2357 static void unfreeze_partials(struct kmem_cache *s,
2358 struct kmem_cache_cpu *c)
2360 #ifdef CONFIG_SLUB_CPU_PARTIAL
2361 struct kmem_cache_node *n = NULL, *n2 = NULL;
2362 struct page *page, *discard_page = NULL;
2364 while ((page = slub_percpu_partial(c))) {
2368 slub_set_percpu_partial(c, page);
2370 n2 = get_node(s, page_to_nid(page));
2373 spin_unlock(&n->list_lock);
2376 spin_lock(&n->list_lock);
2381 old.freelist = page->freelist;
2382 old.counters = page->counters;
2383 VM_BUG_ON(!old.frozen);
2385 new.counters = old.counters;
2386 new.freelist = old.freelist;
2390 } while (!__cmpxchg_double_slab(s, page,
2391 old.freelist, old.counters,
2392 new.freelist, new.counters,
2393 "unfreezing slab"));
2395 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2396 page->next = discard_page;
2397 discard_page = page;
2399 add_partial(n, page, DEACTIVATE_TO_TAIL);
2400 stat(s, FREE_ADD_PARTIAL);
2405 spin_unlock(&n->list_lock);
2407 while (discard_page) {
2408 page = discard_page;
2409 discard_page = discard_page->next;
2411 stat(s, DEACTIVATE_EMPTY);
2412 discard_slab(s, page);
2415 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2419 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2420 * partial page slot if available.
2422 * If we did not find a slot then simply move all the partials to the
2423 * per node partial list.
2425 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2427 #ifdef CONFIG_SLUB_CPU_PARTIAL
2428 struct page *oldpage;
2436 oldpage = this_cpu_read(s->cpu_slab->partial);
2439 pobjects = oldpage->pobjects;
2440 pages = oldpage->pages;
2441 if (drain && pobjects > slub_cpu_partial(s)) {
2442 unsigned long flags;
2444 * partial array is full. Move the existing
2445 * set to the per node partial list.
2447 local_irq_save(flags);
2448 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2449 local_irq_restore(flags);
2453 stat(s, CPU_PARTIAL_DRAIN);
2458 pobjects += page->objects - page->inuse;
2460 page->pages = pages;
2461 page->pobjects = pobjects;
2462 page->next = oldpage;
2464 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2467 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2470 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2472 stat(s, CPUSLAB_FLUSH);
2473 deactivate_slab(s, c->page, c->freelist, c);
2475 c->tid = next_tid(c->tid);
2481 * Called from IPI handler with interrupts disabled.
2483 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2485 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2490 unfreeze_partials(s, c);
2493 static void flush_cpu_slab(void *d)
2495 struct kmem_cache *s = d;
2497 __flush_cpu_slab(s, smp_processor_id());
2500 static bool has_cpu_slab(int cpu, void *info)
2502 struct kmem_cache *s = info;
2503 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2505 return c->page || slub_percpu_partial(c);
2508 static void flush_all(struct kmem_cache *s)
2510 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2514 * Use the cpu notifier to insure that the cpu slabs are flushed when
2517 static int slub_cpu_dead(unsigned int cpu)
2519 struct kmem_cache *s;
2520 unsigned long flags;
2522 mutex_lock(&slab_mutex);
2523 list_for_each_entry(s, &slab_caches, list) {
2524 local_irq_save(flags);
2525 __flush_cpu_slab(s, cpu);
2526 local_irq_restore(flags);
2528 mutex_unlock(&slab_mutex);
2533 * Check if the objects in a per cpu structure fit numa
2534 * locality expectations.
2536 static inline int node_match(struct page *page, int node)
2539 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2545 #ifdef CONFIG_SLUB_DEBUG
2546 static int count_free(struct page *page)
2548 return page->objects - page->inuse;
2551 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2553 return atomic_long_read(&n->total_objects);
2555 #endif /* CONFIG_SLUB_DEBUG */
2557 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2558 static unsigned long count_partial(struct kmem_cache_node *n,
2559 int (*get_count)(struct page *))
2561 unsigned long flags;
2562 unsigned long x = 0;
2565 spin_lock_irqsave(&n->list_lock, flags);
2566 list_for_each_entry(page, &n->partial, slab_list)
2567 x += get_count(page);
2568 spin_unlock_irqrestore(&n->list_lock, flags);
2571 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2573 static noinline void
2574 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2576 #ifdef CONFIG_SLUB_DEBUG
2577 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2578 DEFAULT_RATELIMIT_BURST);
2580 struct kmem_cache_node *n;
2582 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2585 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2586 nid, gfpflags, &gfpflags);
2587 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2588 s->name, s->object_size, s->size, oo_order(s->oo),
2591 if (oo_order(s->min) > get_order(s->object_size))
2592 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2595 for_each_kmem_cache_node(s, node, n) {
2596 unsigned long nr_slabs;
2597 unsigned long nr_objs;
2598 unsigned long nr_free;
2600 nr_free = count_partial(n, count_free);
2601 nr_slabs = node_nr_slabs(n);
2602 nr_objs = node_nr_objs(n);
2604 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2605 node, nr_slabs, nr_objs, nr_free);
2610 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2612 if (unlikely(PageSlabPfmemalloc(page)))
2613 return gfp_pfmemalloc_allowed(gfpflags);
2619 * A variant of pfmemalloc_match() that tests page flags without asserting
2620 * PageSlab. Intended for opportunistic checks before taking a lock and
2621 * rechecking that nobody else freed the page under us.
2623 static inline bool pfmemalloc_match_unsafe(struct page *page, gfp_t gfpflags)
2625 if (unlikely(__PageSlabPfmemalloc(page)))
2626 return gfp_pfmemalloc_allowed(gfpflags);
2632 * Check the page->freelist of a page and either transfer the freelist to the
2633 * per cpu freelist or deactivate the page.
2635 * The page is still frozen if the return value is not NULL.
2637 * If this function returns NULL then the page has been unfrozen.
2639 * This function must be called with interrupt disabled.
2641 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2644 unsigned long counters;
2648 freelist = page->freelist;
2649 counters = page->counters;
2651 new.counters = counters;
2652 VM_BUG_ON(!new.frozen);
2654 new.inuse = page->objects;
2655 new.frozen = freelist != NULL;
2657 } while (!__cmpxchg_double_slab(s, page,
2666 * Slow path. The lockless freelist is empty or we need to perform
2669 * Processing is still very fast if new objects have been freed to the
2670 * regular freelist. In that case we simply take over the regular freelist
2671 * as the lockless freelist and zap the regular freelist.
2673 * If that is not working then we fall back to the partial lists. We take the
2674 * first element of the freelist as the object to allocate now and move the
2675 * rest of the freelist to the lockless freelist.
2677 * And if we were unable to get a new slab from the partial slab lists then
2678 * we need to allocate a new slab. This is the slowest path since it involves
2679 * a call to the page allocator and the setup of a new slab.
2681 * Version of __slab_alloc to use when we know that preemption is
2682 * already disabled (which is the case for bulk allocation).
2684 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2685 unsigned long addr, struct kmem_cache_cpu *c)
2689 unsigned long flags;
2691 stat(s, ALLOC_SLOWPATH);
2695 page = READ_ONCE(c->page);
2698 * if the node is not online or has no normal memory, just
2699 * ignore the node constraint
2701 if (unlikely(node != NUMA_NO_NODE &&
2702 !node_isset(node, slab_nodes)))
2703 node = NUMA_NO_NODE;
2708 if (unlikely(!node_match(page, node))) {
2710 * same as above but node_match() being false already
2711 * implies node != NUMA_NO_NODE
2713 if (!node_isset(node, slab_nodes)) {
2714 node = NUMA_NO_NODE;
2717 stat(s, ALLOC_NODE_MISMATCH);
2718 goto deactivate_slab;
2723 * By rights, we should be searching for a slab page that was
2724 * PFMEMALLOC but right now, we are losing the pfmemalloc
2725 * information when the page leaves the per-cpu allocator
2727 if (unlikely(!pfmemalloc_match_unsafe(page, gfpflags)))
2728 goto deactivate_slab;
2730 /* must check again c->page in case IRQ handler changed it */
2731 local_irq_save(flags);
2732 if (unlikely(page != c->page)) {
2733 local_irq_restore(flags);
2736 freelist = c->freelist;
2740 freelist = get_freelist(s, page);
2744 local_irq_restore(flags);
2745 stat(s, DEACTIVATE_BYPASS);
2749 stat(s, ALLOC_REFILL);
2753 lockdep_assert_irqs_disabled();
2756 * freelist is pointing to the list of objects to be used.
2757 * page is pointing to the page from which the objects are obtained.
2758 * That page must be frozen for per cpu allocations to work.
2760 VM_BUG_ON(!c->page->frozen);
2761 c->freelist = get_freepointer(s, freelist);
2762 c->tid = next_tid(c->tid);
2763 local_irq_restore(flags);
2768 local_irq_save(flags);
2769 if (page != c->page) {
2770 local_irq_restore(flags);
2773 deactivate_slab(s, page, c->freelist, c);
2774 local_irq_restore(flags);
2778 if (slub_percpu_partial(c)) {
2779 local_irq_save(flags);
2780 if (unlikely(c->page)) {
2781 local_irq_restore(flags);
2784 if (unlikely(!slub_percpu_partial(c)))
2785 goto new_objects; /* stolen by an IRQ handler */
2787 page = c->page = slub_percpu_partial(c);
2788 slub_set_percpu_partial(c, page);
2789 local_irq_restore(flags);
2790 stat(s, CPU_PARTIAL_ALLOC);
2794 local_irq_save(flags);
2795 if (unlikely(c->page)) {
2796 local_irq_restore(flags);
2802 lockdep_assert_irqs_disabled();
2804 freelist = get_partial(s, gfpflags, node, &page);
2807 goto check_new_page;
2810 local_irq_restore(flags);
2811 put_cpu_ptr(s->cpu_slab);
2812 page = new_slab(s, gfpflags, node);
2813 c = get_cpu_ptr(s->cpu_slab);
2815 if (unlikely(!page)) {
2816 slab_out_of_memory(s, gfpflags, node);
2820 local_irq_save(flags);
2825 * No other reference to the page yet so we can
2826 * muck around with it freely without cmpxchg
2828 freelist = page->freelist;
2829 page->freelist = NULL;
2831 stat(s, ALLOC_SLAB);
2836 if (kmem_cache_debug(s)) {
2837 if (!alloc_debug_processing(s, page, freelist, addr)) {
2838 /* Slab failed checks. Next slab needed */
2840 local_irq_restore(flags);
2844 * For debug case, we don't load freelist so that all
2845 * allocations go through alloc_debug_processing()
2851 if (unlikely(!pfmemalloc_match(page, gfpflags)))
2853 * For !pfmemalloc_match() case we don't load freelist so that
2854 * we don't make further mismatched allocations easier.
2862 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2863 local_irq_restore(flags);
2868 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
2869 * disabled. Compensates for possible cpu changes by refetching the per cpu area
2872 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2873 unsigned long addr, struct kmem_cache_cpu *c)
2877 #ifdef CONFIG_PREEMPT_COUNT
2879 * We may have been preempted and rescheduled on a different
2880 * cpu before disabling preemption. Need to reload cpu area
2883 c = get_cpu_ptr(s->cpu_slab);
2886 p = ___slab_alloc(s, gfpflags, node, addr, c);
2887 #ifdef CONFIG_PREEMPT_COUNT
2888 put_cpu_ptr(s->cpu_slab);
2894 * If the object has been wiped upon free, make sure it's fully initialized by
2895 * zeroing out freelist pointer.
2897 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2900 if (unlikely(slab_want_init_on_free(s)) && obj)
2901 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
2906 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2907 * have the fastpath folded into their functions. So no function call
2908 * overhead for requests that can be satisfied on the fastpath.
2910 * The fastpath works by first checking if the lockless freelist can be used.
2911 * If not then __slab_alloc is called for slow processing.
2913 * Otherwise we can simply pick the next object from the lockless free list.
2915 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2916 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
2919 struct kmem_cache_cpu *c;
2922 struct obj_cgroup *objcg = NULL;
2925 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2929 object = kfence_alloc(s, orig_size, gfpflags);
2930 if (unlikely(object))
2935 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2936 * enabled. We may switch back and forth between cpus while
2937 * reading from one cpu area. That does not matter as long
2938 * as we end up on the original cpu again when doing the cmpxchg.
2940 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
2941 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
2942 * the tid. If we are preempted and switched to another cpu between the
2943 * two reads, it's OK as the two are still associated with the same cpu
2944 * and cmpxchg later will validate the cpu.
2946 c = raw_cpu_ptr(s->cpu_slab);
2947 tid = READ_ONCE(c->tid);
2950 * Irqless object alloc/free algorithm used here depends on sequence
2951 * of fetching cpu_slab's data. tid should be fetched before anything
2952 * on c to guarantee that object and page associated with previous tid
2953 * won't be used with current tid. If we fetch tid first, object and
2954 * page could be one associated with next tid and our alloc/free
2955 * request will be failed. In this case, we will retry. So, no problem.
2960 * The transaction ids are globally unique per cpu and per operation on
2961 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2962 * occurs on the right processor and that there was no operation on the
2963 * linked list in between.
2966 object = c->freelist;
2968 if (unlikely(!object || !page || !node_match(page, node))) {
2969 object = __slab_alloc(s, gfpflags, node, addr, c);
2971 void *next_object = get_freepointer_safe(s, object);
2974 * The cmpxchg will only match if there was no additional
2975 * operation and if we are on the right processor.
2977 * The cmpxchg does the following atomically (without lock
2979 * 1. Relocate first pointer to the current per cpu area.
2980 * 2. Verify that tid and freelist have not been changed
2981 * 3. If they were not changed replace tid and freelist
2983 * Since this is without lock semantics the protection is only
2984 * against code executing on this cpu *not* from access by
2987 if (unlikely(!this_cpu_cmpxchg_double(
2988 s->cpu_slab->freelist, s->cpu_slab->tid,
2990 next_object, next_tid(tid)))) {
2992 note_cmpxchg_failure("slab_alloc", s, tid);
2995 prefetch_freepointer(s, next_object);
2996 stat(s, ALLOC_FASTPATH);
2999 maybe_wipe_obj_freeptr(s, object);
3000 init = slab_want_init_on_alloc(gfpflags, s);
3003 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
3008 static __always_inline void *slab_alloc(struct kmem_cache *s,
3009 gfp_t gfpflags, unsigned long addr, size_t orig_size)
3011 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size);
3014 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3016 void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size);
3018 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
3023 EXPORT_SYMBOL(kmem_cache_alloc);
3025 #ifdef CONFIG_TRACING
3026 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3028 void *ret = slab_alloc(s, gfpflags, _RET_IP_, size);
3029 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
3030 ret = kasan_kmalloc(s, ret, size, gfpflags);
3033 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3037 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3039 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size);
3041 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3042 s->object_size, s->size, gfpflags, node);
3046 EXPORT_SYMBOL(kmem_cache_alloc_node);
3048 #ifdef CONFIG_TRACING
3049 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3051 int node, size_t size)
3053 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size);
3055 trace_kmalloc_node(_RET_IP_, ret,
3056 size, s->size, gfpflags, node);
3058 ret = kasan_kmalloc(s, ret, size, gfpflags);
3061 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3063 #endif /* CONFIG_NUMA */
3066 * Slow path handling. This may still be called frequently since objects
3067 * have a longer lifetime than the cpu slabs in most processing loads.
3069 * So we still attempt to reduce cache line usage. Just take the slab
3070 * lock and free the item. If there is no additional partial page
3071 * handling required then we can return immediately.
3073 static void __slab_free(struct kmem_cache *s, struct page *page,
3074 void *head, void *tail, int cnt,
3081 unsigned long counters;
3082 struct kmem_cache_node *n = NULL;
3083 unsigned long flags;
3085 stat(s, FREE_SLOWPATH);
3087 if (kfence_free(head))
3090 if (kmem_cache_debug(s) &&
3091 !free_debug_processing(s, page, head, tail, cnt, addr))
3096 spin_unlock_irqrestore(&n->list_lock, flags);
3099 prior = page->freelist;
3100 counters = page->counters;
3101 set_freepointer(s, tail, prior);
3102 new.counters = counters;
3103 was_frozen = new.frozen;
3105 if ((!new.inuse || !prior) && !was_frozen) {
3107 if (kmem_cache_has_cpu_partial(s) && !prior) {
3110 * Slab was on no list before and will be
3112 * We can defer the list move and instead
3117 } else { /* Needs to be taken off a list */
3119 n = get_node(s, page_to_nid(page));
3121 * Speculatively acquire the list_lock.
3122 * If the cmpxchg does not succeed then we may
3123 * drop the list_lock without any processing.
3125 * Otherwise the list_lock will synchronize with
3126 * other processors updating the list of slabs.
3128 spin_lock_irqsave(&n->list_lock, flags);
3133 } while (!cmpxchg_double_slab(s, page,
3140 if (likely(was_frozen)) {
3142 * The list lock was not taken therefore no list
3143 * activity can be necessary.
3145 stat(s, FREE_FROZEN);
3146 } else if (new.frozen) {
3148 * If we just froze the page then put it onto the
3149 * per cpu partial list.
3151 put_cpu_partial(s, page, 1);
3152 stat(s, CPU_PARTIAL_FREE);
3158 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3162 * Objects left in the slab. If it was not on the partial list before
3165 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3166 remove_full(s, n, page);
3167 add_partial(n, page, DEACTIVATE_TO_TAIL);
3168 stat(s, FREE_ADD_PARTIAL);
3170 spin_unlock_irqrestore(&n->list_lock, flags);
3176 * Slab on the partial list.
3178 remove_partial(n, page);
3179 stat(s, FREE_REMOVE_PARTIAL);
3181 /* Slab must be on the full list */
3182 remove_full(s, n, page);
3185 spin_unlock_irqrestore(&n->list_lock, flags);
3187 discard_slab(s, page);
3191 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3192 * can perform fastpath freeing without additional function calls.
3194 * The fastpath is only possible if we are freeing to the current cpu slab
3195 * of this processor. This typically the case if we have just allocated
3198 * If fastpath is not possible then fall back to __slab_free where we deal
3199 * with all sorts of special processing.
3201 * Bulk free of a freelist with several objects (all pointing to the
3202 * same page) possible by specifying head and tail ptr, plus objects
3203 * count (cnt). Bulk free indicated by tail pointer being set.
3205 static __always_inline void do_slab_free(struct kmem_cache *s,
3206 struct page *page, void *head, void *tail,
3207 int cnt, unsigned long addr)
3209 void *tail_obj = tail ? : head;
3210 struct kmem_cache_cpu *c;
3213 memcg_slab_free_hook(s, &head, 1);
3216 * Determine the currently cpus per cpu slab.
3217 * The cpu may change afterward. However that does not matter since
3218 * data is retrieved via this pointer. If we are on the same cpu
3219 * during the cmpxchg then the free will succeed.
3221 c = raw_cpu_ptr(s->cpu_slab);
3222 tid = READ_ONCE(c->tid);
3224 /* Same with comment on barrier() in slab_alloc_node() */
3227 if (likely(page == c->page)) {
3228 void **freelist = READ_ONCE(c->freelist);
3230 set_freepointer(s, tail_obj, freelist);
3232 if (unlikely(!this_cpu_cmpxchg_double(
3233 s->cpu_slab->freelist, s->cpu_slab->tid,
3235 head, next_tid(tid)))) {
3237 note_cmpxchg_failure("slab_free", s, tid);
3240 stat(s, FREE_FASTPATH);
3242 __slab_free(s, page, head, tail_obj, cnt, addr);
3246 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3247 void *head, void *tail, int cnt,
3251 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3252 * to remove objects, whose reuse must be delayed.
3254 if (slab_free_freelist_hook(s, &head, &tail))
3255 do_slab_free(s, page, head, tail, cnt, addr);
3258 #ifdef CONFIG_KASAN_GENERIC
3259 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3261 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3265 void kmem_cache_free(struct kmem_cache *s, void *x)
3267 s = cache_from_obj(s, x);
3270 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3271 trace_kmem_cache_free(_RET_IP_, x, s->name);
3273 EXPORT_SYMBOL(kmem_cache_free);
3275 struct detached_freelist {
3280 struct kmem_cache *s;
3283 static inline void free_nonslab_page(struct page *page, void *object)
3285 unsigned int order = compound_order(page);
3287 VM_BUG_ON_PAGE(!PageCompound(page), page);
3289 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B, -(PAGE_SIZE << order));
3290 __free_pages(page, order);
3294 * This function progressively scans the array with free objects (with
3295 * a limited look ahead) and extract objects belonging to the same
3296 * page. It builds a detached freelist directly within the given
3297 * page/objects. This can happen without any need for
3298 * synchronization, because the objects are owned by running process.
3299 * The freelist is build up as a single linked list in the objects.
3300 * The idea is, that this detached freelist can then be bulk
3301 * transferred to the real freelist(s), but only requiring a single
3302 * synchronization primitive. Look ahead in the array is limited due
3303 * to performance reasons.
3306 int build_detached_freelist(struct kmem_cache *s, size_t size,
3307 void **p, struct detached_freelist *df)
3309 size_t first_skipped_index = 0;
3314 /* Always re-init detached_freelist */
3319 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3320 } while (!object && size);
3325 page = virt_to_head_page(object);
3327 /* Handle kalloc'ed objects */
3328 if (unlikely(!PageSlab(page))) {
3329 free_nonslab_page(page, object);
3330 p[size] = NULL; /* mark object processed */
3333 /* Derive kmem_cache from object */
3334 df->s = page->slab_cache;
3336 df->s = cache_from_obj(s, object); /* Support for memcg */
3339 if (is_kfence_address(object)) {
3340 slab_free_hook(df->s, object, false);
3341 __kfence_free(object);
3342 p[size] = NULL; /* mark object processed */
3346 /* Start new detached freelist */
3348 set_freepointer(df->s, object, NULL);
3350 df->freelist = object;
3351 p[size] = NULL; /* mark object processed */
3357 continue; /* Skip processed objects */
3359 /* df->page is always set at this point */
3360 if (df->page == virt_to_head_page(object)) {
3361 /* Opportunity build freelist */
3362 set_freepointer(df->s, object, df->freelist);
3363 df->freelist = object;
3365 p[size] = NULL; /* mark object processed */
3370 /* Limit look ahead search */
3374 if (!first_skipped_index)
3375 first_skipped_index = size + 1;
3378 return first_skipped_index;
3381 /* Note that interrupts must be enabled when calling this function. */
3382 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3387 memcg_slab_free_hook(s, p, size);
3389 struct detached_freelist df;
3391 size = build_detached_freelist(s, size, p, &df);
3395 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
3396 } while (likely(size));
3398 EXPORT_SYMBOL(kmem_cache_free_bulk);
3400 /* Note that interrupts must be enabled when calling this function. */
3401 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3404 struct kmem_cache_cpu *c;
3406 struct obj_cgroup *objcg = NULL;
3408 /* memcg and kmem_cache debug support */
3409 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3413 * Drain objects in the per cpu slab, while disabling local
3414 * IRQs, which protects against PREEMPT and interrupts
3415 * handlers invoking normal fastpath.
3417 c = get_cpu_ptr(s->cpu_slab);
3418 local_irq_disable();
3420 for (i = 0; i < size; i++) {
3421 void *object = kfence_alloc(s, s->object_size, flags);
3423 if (unlikely(object)) {
3428 object = c->freelist;
3429 if (unlikely(!object)) {
3431 * We may have removed an object from c->freelist using
3432 * the fastpath in the previous iteration; in that case,
3433 * c->tid has not been bumped yet.
3434 * Since ___slab_alloc() may reenable interrupts while
3435 * allocating memory, we should bump c->tid now.
3437 c->tid = next_tid(c->tid);
3442 * Invoking slow path likely have side-effect
3443 * of re-populating per CPU c->freelist
3445 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3447 if (unlikely(!p[i]))
3450 c = this_cpu_ptr(s->cpu_slab);
3451 maybe_wipe_obj_freeptr(s, p[i]);
3453 local_irq_disable();
3455 continue; /* goto for-loop */
3457 c->freelist = get_freepointer(s, object);
3459 maybe_wipe_obj_freeptr(s, p[i]);
3461 c->tid = next_tid(c->tid);
3463 put_cpu_ptr(s->cpu_slab);
3466 * memcg and kmem_cache debug support and memory initialization.
3467 * Done outside of the IRQ disabled fastpath loop.
3469 slab_post_alloc_hook(s, objcg, flags, size, p,
3470 slab_want_init_on_alloc(flags, s));
3473 put_cpu_ptr(s->cpu_slab);
3474 slab_post_alloc_hook(s, objcg, flags, i, p, false);
3475 __kmem_cache_free_bulk(s, i, p);
3478 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3482 * Object placement in a slab is made very easy because we always start at
3483 * offset 0. If we tune the size of the object to the alignment then we can
3484 * get the required alignment by putting one properly sized object after
3487 * Notice that the allocation order determines the sizes of the per cpu
3488 * caches. Each processor has always one slab available for allocations.
3489 * Increasing the allocation order reduces the number of times that slabs
3490 * must be moved on and off the partial lists and is therefore a factor in
3495 * Minimum / Maximum order of slab pages. This influences locking overhead
3496 * and slab fragmentation. A higher order reduces the number of partial slabs
3497 * and increases the number of allocations possible without having to
3498 * take the list_lock.
3500 static unsigned int slub_min_order;
3501 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3502 static unsigned int slub_min_objects;
3505 * Calculate the order of allocation given an slab object size.
3507 * The order of allocation has significant impact on performance and other
3508 * system components. Generally order 0 allocations should be preferred since
3509 * order 0 does not cause fragmentation in the page allocator. Larger objects
3510 * be problematic to put into order 0 slabs because there may be too much
3511 * unused space left. We go to a higher order if more than 1/16th of the slab
3514 * In order to reach satisfactory performance we must ensure that a minimum
3515 * number of objects is in one slab. Otherwise we may generate too much
3516 * activity on the partial lists which requires taking the list_lock. This is
3517 * less a concern for large slabs though which are rarely used.
3519 * slub_max_order specifies the order where we begin to stop considering the
3520 * number of objects in a slab as critical. If we reach slub_max_order then
3521 * we try to keep the page order as low as possible. So we accept more waste
3522 * of space in favor of a small page order.
3524 * Higher order allocations also allow the placement of more objects in a
3525 * slab and thereby reduce object handling overhead. If the user has
3526 * requested a higher minimum order then we start with that one instead of
3527 * the smallest order which will fit the object.
3529 static inline unsigned int slab_order(unsigned int size,
3530 unsigned int min_objects, unsigned int max_order,
3531 unsigned int fract_leftover)
3533 unsigned int min_order = slub_min_order;
3536 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3537 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3539 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3540 order <= max_order; order++) {
3542 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3545 rem = slab_size % size;
3547 if (rem <= slab_size / fract_leftover)
3554 static inline int calculate_order(unsigned int size)
3557 unsigned int min_objects;
3558 unsigned int max_objects;
3559 unsigned int nr_cpus;
3562 * Attempt to find best configuration for a slab. This
3563 * works by first attempting to generate a layout with
3564 * the best configuration and backing off gradually.
3566 * First we increase the acceptable waste in a slab. Then
3567 * we reduce the minimum objects required in a slab.
3569 min_objects = slub_min_objects;
3572 * Some architectures will only update present cpus when
3573 * onlining them, so don't trust the number if it's just 1. But
3574 * we also don't want to use nr_cpu_ids always, as on some other
3575 * architectures, there can be many possible cpus, but never
3576 * onlined. Here we compromise between trying to avoid too high
3577 * order on systems that appear larger than they are, and too
3578 * low order on systems that appear smaller than they are.
3580 nr_cpus = num_present_cpus();
3582 nr_cpus = nr_cpu_ids;
3583 min_objects = 4 * (fls(nr_cpus) + 1);
3585 max_objects = order_objects(slub_max_order, size);
3586 min_objects = min(min_objects, max_objects);
3588 while (min_objects > 1) {
3589 unsigned int fraction;
3592 while (fraction >= 4) {
3593 order = slab_order(size, min_objects,
3594 slub_max_order, fraction);
3595 if (order <= slub_max_order)
3603 * We were unable to place multiple objects in a slab. Now
3604 * lets see if we can place a single object there.
3606 order = slab_order(size, 1, slub_max_order, 1);
3607 if (order <= slub_max_order)
3611 * Doh this slab cannot be placed using slub_max_order.
3613 order = slab_order(size, 1, MAX_ORDER, 1);
3614 if (order < MAX_ORDER)
3620 init_kmem_cache_node(struct kmem_cache_node *n)
3623 spin_lock_init(&n->list_lock);
3624 INIT_LIST_HEAD(&n->partial);
3625 #ifdef CONFIG_SLUB_DEBUG
3626 atomic_long_set(&n->nr_slabs, 0);
3627 atomic_long_set(&n->total_objects, 0);
3628 INIT_LIST_HEAD(&n->full);
3632 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3634 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3635 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3638 * Must align to double word boundary for the double cmpxchg
3639 * instructions to work; see __pcpu_double_call_return_bool().
3641 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3642 2 * sizeof(void *));
3647 init_kmem_cache_cpus(s);
3652 static struct kmem_cache *kmem_cache_node;
3655 * No kmalloc_node yet so do it by hand. We know that this is the first
3656 * slab on the node for this slabcache. There are no concurrent accesses
3659 * Note that this function only works on the kmem_cache_node
3660 * when allocating for the kmem_cache_node. This is used for bootstrapping
3661 * memory on a fresh node that has no slab structures yet.
3663 static void early_kmem_cache_node_alloc(int node)
3666 struct kmem_cache_node *n;
3668 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3670 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3673 if (page_to_nid(page) != node) {
3674 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3675 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3680 #ifdef CONFIG_SLUB_DEBUG
3681 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3682 init_tracking(kmem_cache_node, n);
3684 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3685 page->freelist = get_freepointer(kmem_cache_node, n);
3688 kmem_cache_node->node[node] = n;
3689 init_kmem_cache_node(n);
3690 inc_slabs_node(kmem_cache_node, node, page->objects);
3693 * No locks need to be taken here as it has just been
3694 * initialized and there is no concurrent access.
3696 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3699 static void free_kmem_cache_nodes(struct kmem_cache *s)
3702 struct kmem_cache_node *n;
3704 for_each_kmem_cache_node(s, node, n) {
3705 s->node[node] = NULL;
3706 kmem_cache_free(kmem_cache_node, n);
3710 void __kmem_cache_release(struct kmem_cache *s)
3712 cache_random_seq_destroy(s);
3713 free_percpu(s->cpu_slab);
3714 free_kmem_cache_nodes(s);
3717 static int init_kmem_cache_nodes(struct kmem_cache *s)
3721 for_each_node_mask(node, slab_nodes) {
3722 struct kmem_cache_node *n;
3724 if (slab_state == DOWN) {
3725 early_kmem_cache_node_alloc(node);
3728 n = kmem_cache_alloc_node(kmem_cache_node,
3732 free_kmem_cache_nodes(s);
3736 init_kmem_cache_node(n);
3742 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3744 if (min < MIN_PARTIAL)
3746 else if (min > MAX_PARTIAL)
3748 s->min_partial = min;
3751 static void set_cpu_partial(struct kmem_cache *s)
3753 #ifdef CONFIG_SLUB_CPU_PARTIAL
3755 * cpu_partial determined the maximum number of objects kept in the
3756 * per cpu partial lists of a processor.
3758 * Per cpu partial lists mainly contain slabs that just have one
3759 * object freed. If they are used for allocation then they can be
3760 * filled up again with minimal effort. The slab will never hit the
3761 * per node partial lists and therefore no locking will be required.
3763 * This setting also determines
3765 * A) The number of objects from per cpu partial slabs dumped to the
3766 * per node list when we reach the limit.
3767 * B) The number of objects in cpu partial slabs to extract from the
3768 * per node list when we run out of per cpu objects. We only fetch
3769 * 50% to keep some capacity around for frees.
3771 if (!kmem_cache_has_cpu_partial(s))
3772 slub_set_cpu_partial(s, 0);
3773 else if (s->size >= PAGE_SIZE)
3774 slub_set_cpu_partial(s, 2);
3775 else if (s->size >= 1024)
3776 slub_set_cpu_partial(s, 6);
3777 else if (s->size >= 256)
3778 slub_set_cpu_partial(s, 13);
3780 slub_set_cpu_partial(s, 30);
3785 * calculate_sizes() determines the order and the distribution of data within
3788 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3790 slab_flags_t flags = s->flags;
3791 unsigned int size = s->object_size;
3795 * Round up object size to the next word boundary. We can only
3796 * place the free pointer at word boundaries and this determines
3797 * the possible location of the free pointer.
3799 size = ALIGN(size, sizeof(void *));
3801 #ifdef CONFIG_SLUB_DEBUG
3803 * Determine if we can poison the object itself. If the user of
3804 * the slab may touch the object after free or before allocation
3805 * then we should never poison the object itself.
3807 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3809 s->flags |= __OBJECT_POISON;
3811 s->flags &= ~__OBJECT_POISON;
3815 * If we are Redzoning then check if there is some space between the
3816 * end of the object and the free pointer. If not then add an
3817 * additional word to have some bytes to store Redzone information.
3819 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3820 size += sizeof(void *);
3824 * With that we have determined the number of bytes in actual use
3825 * by the object and redzoning.
3829 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3830 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
3833 * Relocate free pointer after the object if it is not
3834 * permitted to overwrite the first word of the object on
3837 * This is the case if we do RCU, have a constructor or
3838 * destructor, are poisoning the objects, or are
3839 * redzoning an object smaller than sizeof(void *).
3841 * The assumption that s->offset >= s->inuse means free
3842 * pointer is outside of the object is used in the
3843 * freeptr_outside_object() function. If that is no
3844 * longer true, the function needs to be modified.
3847 size += sizeof(void *);
3850 * Store freelist pointer near middle of object to keep
3851 * it away from the edges of the object to avoid small
3852 * sized over/underflows from neighboring allocations.
3854 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
3857 #ifdef CONFIG_SLUB_DEBUG
3858 if (flags & SLAB_STORE_USER)
3860 * Need to store information about allocs and frees after
3863 size += 2 * sizeof(struct track);
3866 kasan_cache_create(s, &size, &s->flags);
3867 #ifdef CONFIG_SLUB_DEBUG
3868 if (flags & SLAB_RED_ZONE) {
3870 * Add some empty padding so that we can catch
3871 * overwrites from earlier objects rather than let
3872 * tracking information or the free pointer be
3873 * corrupted if a user writes before the start
3876 size += sizeof(void *);
3878 s->red_left_pad = sizeof(void *);
3879 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3880 size += s->red_left_pad;
3885 * SLUB stores one object immediately after another beginning from
3886 * offset 0. In order to align the objects we have to simply size
3887 * each object to conform to the alignment.
3889 size = ALIGN(size, s->align);
3891 s->reciprocal_size = reciprocal_value(size);
3892 if (forced_order >= 0)
3893 order = forced_order;
3895 order = calculate_order(size);
3902 s->allocflags |= __GFP_COMP;
3904 if (s->flags & SLAB_CACHE_DMA)
3905 s->allocflags |= GFP_DMA;
3907 if (s->flags & SLAB_CACHE_DMA32)
3908 s->allocflags |= GFP_DMA32;
3910 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3911 s->allocflags |= __GFP_RECLAIMABLE;
3914 * Determine the number of objects per slab
3916 s->oo = oo_make(order, size);
3917 s->min = oo_make(get_order(size), size);
3918 if (oo_objects(s->oo) > oo_objects(s->max))
3921 return !!oo_objects(s->oo);
3924 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3926 s->flags = kmem_cache_flags(s->size, flags, s->name);
3927 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3928 s->random = get_random_long();
3931 if (!calculate_sizes(s, -1))
3933 if (disable_higher_order_debug) {
3935 * Disable debugging flags that store metadata if the min slab
3938 if (get_order(s->size) > get_order(s->object_size)) {
3939 s->flags &= ~DEBUG_METADATA_FLAGS;
3941 if (!calculate_sizes(s, -1))
3946 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3947 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3948 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3949 /* Enable fast mode */
3950 s->flags |= __CMPXCHG_DOUBLE;
3954 * The larger the object size is, the more pages we want on the partial
3955 * list to avoid pounding the page allocator excessively.
3957 set_min_partial(s, ilog2(s->size) / 2);
3962 s->remote_node_defrag_ratio = 1000;
3965 /* Initialize the pre-computed randomized freelist if slab is up */
3966 if (slab_state >= UP) {
3967 if (init_cache_random_seq(s))
3971 if (!init_kmem_cache_nodes(s))
3974 if (alloc_kmem_cache_cpus(s))
3977 free_kmem_cache_nodes(s);
3982 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3985 #ifdef CONFIG_SLUB_DEBUG
3986 void *addr = page_address(page);
3990 slab_err(s, page, text, s->name);
3993 map = get_map(s, page);
3994 for_each_object(p, s, addr, page->objects) {
3996 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3997 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
3998 print_tracking(s, p);
4007 * Attempt to free all partial slabs on a node.
4008 * This is called from __kmem_cache_shutdown(). We must take list_lock
4009 * because sysfs file might still access partial list after the shutdowning.
4011 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4014 struct page *page, *h;
4016 BUG_ON(irqs_disabled());
4017 spin_lock_irq(&n->list_lock);
4018 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
4020 remove_partial(n, page);
4021 list_add(&page->slab_list, &discard);
4023 list_slab_objects(s, page,
4024 "Objects remaining in %s on __kmem_cache_shutdown()");
4027 spin_unlock_irq(&n->list_lock);
4029 list_for_each_entry_safe(page, h, &discard, slab_list)
4030 discard_slab(s, page);
4033 bool __kmem_cache_empty(struct kmem_cache *s)
4036 struct kmem_cache_node *n;
4038 for_each_kmem_cache_node(s, node, n)
4039 if (n->nr_partial || slabs_node(s, node))
4045 * Release all resources used by a slab cache.
4047 int __kmem_cache_shutdown(struct kmem_cache *s)
4050 struct kmem_cache_node *n;
4053 /* Attempt to free all objects */
4054 for_each_kmem_cache_node(s, node, n) {
4056 if (n->nr_partial || slabs_node(s, node))
4062 #ifdef CONFIG_PRINTK
4063 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
4066 int __maybe_unused i;
4070 struct kmem_cache *s = page->slab_cache;
4071 struct track __maybe_unused *trackp;
4073 kpp->kp_ptr = object;
4074 kpp->kp_page = page;
4075 kpp->kp_slab_cache = s;
4076 base = page_address(page);
4077 objp0 = kasan_reset_tag(object);
4078 #ifdef CONFIG_SLUB_DEBUG
4079 objp = restore_red_left(s, objp0);
4083 objnr = obj_to_index(s, page, objp);
4084 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4085 objp = base + s->size * objnr;
4086 kpp->kp_objp = objp;
4087 if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) ||
4088 !(s->flags & SLAB_STORE_USER))
4090 #ifdef CONFIG_SLUB_DEBUG
4091 objp = fixup_red_left(s, objp);
4092 trackp = get_track(s, objp, TRACK_ALLOC);
4093 kpp->kp_ret = (void *)trackp->addr;
4094 #ifdef CONFIG_STACKTRACE
4095 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4096 kpp->kp_stack[i] = (void *)trackp->addrs[i];
4097 if (!kpp->kp_stack[i])
4101 trackp = get_track(s, objp, TRACK_FREE);
4102 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4103 kpp->kp_free_stack[i] = (void *)trackp->addrs[i];
4104 if (!kpp->kp_free_stack[i])
4112 /********************************************************************
4114 *******************************************************************/
4116 static int __init setup_slub_min_order(char *str)
4118 get_option(&str, (int *)&slub_min_order);
4123 __setup("slub_min_order=", setup_slub_min_order);
4125 static int __init setup_slub_max_order(char *str)
4127 get_option(&str, (int *)&slub_max_order);
4128 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4133 __setup("slub_max_order=", setup_slub_max_order);
4135 static int __init setup_slub_min_objects(char *str)
4137 get_option(&str, (int *)&slub_min_objects);
4142 __setup("slub_min_objects=", setup_slub_min_objects);
4144 void *__kmalloc(size_t size, gfp_t flags)
4146 struct kmem_cache *s;
4149 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4150 return kmalloc_large(size, flags);
4152 s = kmalloc_slab(size, flags);
4154 if (unlikely(ZERO_OR_NULL_PTR(s)))
4157 ret = slab_alloc(s, flags, _RET_IP_, size);
4159 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
4161 ret = kasan_kmalloc(s, ret, size, flags);
4165 EXPORT_SYMBOL(__kmalloc);
4168 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4172 unsigned int order = get_order(size);
4174 flags |= __GFP_COMP;
4175 page = alloc_pages_node(node, flags, order);
4177 ptr = page_address(page);
4178 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4179 PAGE_SIZE << order);
4182 return kmalloc_large_node_hook(ptr, size, flags);
4185 void *__kmalloc_node(size_t size, gfp_t flags, int node)
4187 struct kmem_cache *s;
4190 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4191 ret = kmalloc_large_node(size, flags, node);
4193 trace_kmalloc_node(_RET_IP_, ret,
4194 size, PAGE_SIZE << get_order(size),
4200 s = kmalloc_slab(size, flags);
4202 if (unlikely(ZERO_OR_NULL_PTR(s)))
4205 ret = slab_alloc_node(s, flags, node, _RET_IP_, size);
4207 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4209 ret = kasan_kmalloc(s, ret, size, flags);
4213 EXPORT_SYMBOL(__kmalloc_node);
4214 #endif /* CONFIG_NUMA */
4216 #ifdef CONFIG_HARDENED_USERCOPY
4218 * Rejects incorrectly sized objects and objects that are to be copied
4219 * to/from userspace but do not fall entirely within the containing slab
4220 * cache's usercopy region.
4222 * Returns NULL if check passes, otherwise const char * to name of cache
4223 * to indicate an error.
4225 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4228 struct kmem_cache *s;
4229 unsigned int offset;
4231 bool is_kfence = is_kfence_address(ptr);
4233 ptr = kasan_reset_tag(ptr);
4235 /* Find object and usable object size. */
4236 s = page->slab_cache;
4238 /* Reject impossible pointers. */
4239 if (ptr < page_address(page))
4240 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4243 /* Find offset within object. */
4245 offset = ptr - kfence_object_start(ptr);
4247 offset = (ptr - page_address(page)) % s->size;
4249 /* Adjust for redzone and reject if within the redzone. */
4250 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4251 if (offset < s->red_left_pad)
4252 usercopy_abort("SLUB object in left red zone",
4253 s->name, to_user, offset, n);
4254 offset -= s->red_left_pad;
4257 /* Allow address range falling entirely within usercopy region. */
4258 if (offset >= s->useroffset &&
4259 offset - s->useroffset <= s->usersize &&
4260 n <= s->useroffset - offset + s->usersize)
4264 * If the copy is still within the allocated object, produce
4265 * a warning instead of rejecting the copy. This is intended
4266 * to be a temporary method to find any missing usercopy
4269 object_size = slab_ksize(s);
4270 if (usercopy_fallback &&
4271 offset <= object_size && n <= object_size - offset) {
4272 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4276 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4278 #endif /* CONFIG_HARDENED_USERCOPY */
4280 size_t __ksize(const void *object)
4284 if (unlikely(object == ZERO_SIZE_PTR))
4287 page = virt_to_head_page(object);
4289 if (unlikely(!PageSlab(page))) {
4290 WARN_ON(!PageCompound(page));
4291 return page_size(page);
4294 return slab_ksize(page->slab_cache);
4296 EXPORT_SYMBOL(__ksize);
4298 void kfree(const void *x)
4301 void *object = (void *)x;
4303 trace_kfree(_RET_IP_, x);
4305 if (unlikely(ZERO_OR_NULL_PTR(x)))
4308 page = virt_to_head_page(x);
4309 if (unlikely(!PageSlab(page))) {
4310 free_nonslab_page(page, object);
4313 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4315 EXPORT_SYMBOL(kfree);
4317 #define SHRINK_PROMOTE_MAX 32
4320 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4321 * up most to the head of the partial lists. New allocations will then
4322 * fill those up and thus they can be removed from the partial lists.
4324 * The slabs with the least items are placed last. This results in them
4325 * being allocated from last increasing the chance that the last objects
4326 * are freed in them.
4328 int __kmem_cache_shrink(struct kmem_cache *s)
4332 struct kmem_cache_node *n;
4335 struct list_head discard;
4336 struct list_head promote[SHRINK_PROMOTE_MAX];
4337 unsigned long flags;
4341 for_each_kmem_cache_node(s, node, n) {
4342 INIT_LIST_HEAD(&discard);
4343 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4344 INIT_LIST_HEAD(promote + i);
4346 spin_lock_irqsave(&n->list_lock, flags);
4349 * Build lists of slabs to discard or promote.
4351 * Note that concurrent frees may occur while we hold the
4352 * list_lock. page->inuse here is the upper limit.
4354 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4355 int free = page->objects - page->inuse;
4357 /* Do not reread page->inuse */
4360 /* We do not keep full slabs on the list */
4363 if (free == page->objects) {
4364 list_move(&page->slab_list, &discard);
4366 } else if (free <= SHRINK_PROMOTE_MAX)
4367 list_move(&page->slab_list, promote + free - 1);
4371 * Promote the slabs filled up most to the head of the
4374 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4375 list_splice(promote + i, &n->partial);
4377 spin_unlock_irqrestore(&n->list_lock, flags);
4379 /* Release empty slabs */
4380 list_for_each_entry_safe(page, t, &discard, slab_list)
4381 discard_slab(s, page);
4383 if (slabs_node(s, node))
4390 static int slab_mem_going_offline_callback(void *arg)
4392 struct kmem_cache *s;
4394 mutex_lock(&slab_mutex);
4395 list_for_each_entry(s, &slab_caches, list)
4396 __kmem_cache_shrink(s);
4397 mutex_unlock(&slab_mutex);
4402 static void slab_mem_offline_callback(void *arg)
4404 struct memory_notify *marg = arg;
4407 offline_node = marg->status_change_nid_normal;
4410 * If the node still has available memory. we need kmem_cache_node
4413 if (offline_node < 0)
4416 mutex_lock(&slab_mutex);
4417 node_clear(offline_node, slab_nodes);
4419 * We no longer free kmem_cache_node structures here, as it would be
4420 * racy with all get_node() users, and infeasible to protect them with
4423 mutex_unlock(&slab_mutex);
4426 static int slab_mem_going_online_callback(void *arg)
4428 struct kmem_cache_node *n;
4429 struct kmem_cache *s;
4430 struct memory_notify *marg = arg;
4431 int nid = marg->status_change_nid_normal;
4435 * If the node's memory is already available, then kmem_cache_node is
4436 * already created. Nothing to do.
4442 * We are bringing a node online. No memory is available yet. We must
4443 * allocate a kmem_cache_node structure in order to bring the node
4446 mutex_lock(&slab_mutex);
4447 list_for_each_entry(s, &slab_caches, list) {
4449 * The structure may already exist if the node was previously
4450 * onlined and offlined.
4452 if (get_node(s, nid))
4455 * XXX: kmem_cache_alloc_node will fallback to other nodes
4456 * since memory is not yet available from the node that
4459 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4464 init_kmem_cache_node(n);
4468 * Any cache created after this point will also have kmem_cache_node
4469 * initialized for the new node.
4471 node_set(nid, slab_nodes);
4473 mutex_unlock(&slab_mutex);
4477 static int slab_memory_callback(struct notifier_block *self,
4478 unsigned long action, void *arg)
4483 case MEM_GOING_ONLINE:
4484 ret = slab_mem_going_online_callback(arg);
4486 case MEM_GOING_OFFLINE:
4487 ret = slab_mem_going_offline_callback(arg);
4490 case MEM_CANCEL_ONLINE:
4491 slab_mem_offline_callback(arg);
4494 case MEM_CANCEL_OFFLINE:
4498 ret = notifier_from_errno(ret);
4504 static struct notifier_block slab_memory_callback_nb = {
4505 .notifier_call = slab_memory_callback,
4506 .priority = SLAB_CALLBACK_PRI,
4509 /********************************************************************
4510 * Basic setup of slabs
4511 *******************************************************************/
4514 * Used for early kmem_cache structures that were allocated using
4515 * the page allocator. Allocate them properly then fix up the pointers
4516 * that may be pointing to the wrong kmem_cache structure.
4519 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4522 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4523 struct kmem_cache_node *n;
4525 memcpy(s, static_cache, kmem_cache->object_size);
4528 * This runs very early, and only the boot processor is supposed to be
4529 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4532 __flush_cpu_slab(s, smp_processor_id());
4533 for_each_kmem_cache_node(s, node, n) {
4536 list_for_each_entry(p, &n->partial, slab_list)
4539 #ifdef CONFIG_SLUB_DEBUG
4540 list_for_each_entry(p, &n->full, slab_list)
4544 list_add(&s->list, &slab_caches);
4548 void __init kmem_cache_init(void)
4550 static __initdata struct kmem_cache boot_kmem_cache,
4551 boot_kmem_cache_node;
4554 if (debug_guardpage_minorder())
4557 /* Print slub debugging pointers without hashing */
4558 if (__slub_debug_enabled())
4559 no_hash_pointers_enable(NULL);
4561 kmem_cache_node = &boot_kmem_cache_node;
4562 kmem_cache = &boot_kmem_cache;
4565 * Initialize the nodemask for which we will allocate per node
4566 * structures. Here we don't need taking slab_mutex yet.
4568 for_each_node_state(node, N_NORMAL_MEMORY)
4569 node_set(node, slab_nodes);
4571 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4572 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4574 register_hotmemory_notifier(&slab_memory_callback_nb);
4576 /* Able to allocate the per node structures */
4577 slab_state = PARTIAL;
4579 create_boot_cache(kmem_cache, "kmem_cache",
4580 offsetof(struct kmem_cache, node) +
4581 nr_node_ids * sizeof(struct kmem_cache_node *),
4582 SLAB_HWCACHE_ALIGN, 0, 0);
4584 kmem_cache = bootstrap(&boot_kmem_cache);
4585 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4587 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4588 setup_kmalloc_cache_index_table();
4589 create_kmalloc_caches(0);
4591 /* Setup random freelists for each cache */
4592 init_freelist_randomization();
4594 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4597 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4599 slub_min_order, slub_max_order, slub_min_objects,
4600 nr_cpu_ids, nr_node_ids);
4603 void __init kmem_cache_init_late(void)
4608 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4609 slab_flags_t flags, void (*ctor)(void *))
4611 struct kmem_cache *s;
4613 s = find_mergeable(size, align, flags, name, ctor);
4618 * Adjust the object sizes so that we clear
4619 * the complete object on kzalloc.
4621 s->object_size = max(s->object_size, size);
4622 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4624 if (sysfs_slab_alias(s, name)) {
4633 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4637 err = kmem_cache_open(s, flags);
4641 /* Mutex is not taken during early boot */
4642 if (slab_state <= UP)
4645 err = sysfs_slab_add(s);
4647 __kmem_cache_release(s);
4649 if (s->flags & SLAB_STORE_USER)
4650 debugfs_slab_add(s);
4655 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4657 struct kmem_cache *s;
4660 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4661 return kmalloc_large(size, gfpflags);
4663 s = kmalloc_slab(size, gfpflags);
4665 if (unlikely(ZERO_OR_NULL_PTR(s)))
4668 ret = slab_alloc(s, gfpflags, caller, size);
4670 /* Honor the call site pointer we received. */
4671 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4675 EXPORT_SYMBOL(__kmalloc_track_caller);
4678 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4679 int node, unsigned long caller)
4681 struct kmem_cache *s;
4684 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4685 ret = kmalloc_large_node(size, gfpflags, node);
4687 trace_kmalloc_node(caller, ret,
4688 size, PAGE_SIZE << get_order(size),
4694 s = kmalloc_slab(size, gfpflags);
4696 if (unlikely(ZERO_OR_NULL_PTR(s)))
4699 ret = slab_alloc_node(s, gfpflags, node, caller, size);
4701 /* Honor the call site pointer we received. */
4702 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4706 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4710 static int count_inuse(struct page *page)
4715 static int count_total(struct page *page)
4717 return page->objects;
4721 #ifdef CONFIG_SLUB_DEBUG
4722 static void validate_slab(struct kmem_cache *s, struct page *page,
4723 unsigned long *obj_map)
4726 void *addr = page_address(page);
4730 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4733 /* Now we know that a valid freelist exists */
4734 __fill_map(obj_map, s, page);
4735 for_each_object(p, s, addr, page->objects) {
4736 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
4737 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4739 if (!check_object(s, page, p, val))
4746 static int validate_slab_node(struct kmem_cache *s,
4747 struct kmem_cache_node *n, unsigned long *obj_map)
4749 unsigned long count = 0;
4751 unsigned long flags;
4753 spin_lock_irqsave(&n->list_lock, flags);
4755 list_for_each_entry(page, &n->partial, slab_list) {
4756 validate_slab(s, page, obj_map);
4759 if (count != n->nr_partial) {
4760 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4761 s->name, count, n->nr_partial);
4762 slab_add_kunit_errors();
4765 if (!(s->flags & SLAB_STORE_USER))
4768 list_for_each_entry(page, &n->full, slab_list) {
4769 validate_slab(s, page, obj_map);
4772 if (count != atomic_long_read(&n->nr_slabs)) {
4773 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4774 s->name, count, atomic_long_read(&n->nr_slabs));
4775 slab_add_kunit_errors();
4779 spin_unlock_irqrestore(&n->list_lock, flags);
4783 long validate_slab_cache(struct kmem_cache *s)
4786 unsigned long count = 0;
4787 struct kmem_cache_node *n;
4788 unsigned long *obj_map;
4790 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
4795 for_each_kmem_cache_node(s, node, n)
4796 count += validate_slab_node(s, n, obj_map);
4798 bitmap_free(obj_map);
4802 EXPORT_SYMBOL(validate_slab_cache);
4804 #ifdef CONFIG_DEBUG_FS
4806 * Generate lists of code addresses where slabcache objects are allocated
4811 unsigned long count;
4818 DECLARE_BITMAP(cpus, NR_CPUS);
4824 unsigned long count;
4825 struct location *loc;
4828 static struct dentry *slab_debugfs_root;
4830 static void free_loc_track(struct loc_track *t)
4833 free_pages((unsigned long)t->loc,
4834 get_order(sizeof(struct location) * t->max));
4837 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4842 order = get_order(sizeof(struct location) * max);
4844 l = (void *)__get_free_pages(flags, order);
4849 memcpy(l, t->loc, sizeof(struct location) * t->count);
4857 static int add_location(struct loc_track *t, struct kmem_cache *s,
4858 const struct track *track)
4860 long start, end, pos;
4862 unsigned long caddr;
4863 unsigned long age = jiffies - track->when;
4869 pos = start + (end - start + 1) / 2;
4872 * There is nothing at "end". If we end up there
4873 * we need to add something to before end.
4878 caddr = t->loc[pos].addr;
4879 if (track->addr == caddr) {
4885 if (age < l->min_time)
4887 if (age > l->max_time)
4890 if (track->pid < l->min_pid)
4891 l->min_pid = track->pid;
4892 if (track->pid > l->max_pid)
4893 l->max_pid = track->pid;
4895 cpumask_set_cpu(track->cpu,
4896 to_cpumask(l->cpus));
4898 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4902 if (track->addr < caddr)
4909 * Not found. Insert new tracking element.
4911 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4917 (t->count - pos) * sizeof(struct location));
4920 l->addr = track->addr;
4924 l->min_pid = track->pid;
4925 l->max_pid = track->pid;
4926 cpumask_clear(to_cpumask(l->cpus));
4927 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4928 nodes_clear(l->nodes);
4929 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4933 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4934 struct page *page, enum track_item alloc,
4935 unsigned long *obj_map)
4937 void *addr = page_address(page);
4940 __fill_map(obj_map, s, page);
4942 for_each_object(p, s, addr, page->objects)
4943 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
4944 add_location(t, s, get_track(s, p, alloc));
4946 #endif /* CONFIG_DEBUG_FS */
4947 #endif /* CONFIG_SLUB_DEBUG */
4950 enum slab_stat_type {
4951 SL_ALL, /* All slabs */
4952 SL_PARTIAL, /* Only partially allocated slabs */
4953 SL_CPU, /* Only slabs used for cpu caches */
4954 SL_OBJECTS, /* Determine allocated objects not slabs */
4955 SL_TOTAL /* Determine object capacity not slabs */
4958 #define SO_ALL (1 << SL_ALL)
4959 #define SO_PARTIAL (1 << SL_PARTIAL)
4960 #define SO_CPU (1 << SL_CPU)
4961 #define SO_OBJECTS (1 << SL_OBJECTS)
4962 #define SO_TOTAL (1 << SL_TOTAL)
4964 static ssize_t show_slab_objects(struct kmem_cache *s,
4965 char *buf, unsigned long flags)
4967 unsigned long total = 0;
4970 unsigned long *nodes;
4973 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4977 if (flags & SO_CPU) {
4980 for_each_possible_cpu(cpu) {
4981 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4986 page = READ_ONCE(c->page);
4990 node = page_to_nid(page);
4991 if (flags & SO_TOTAL)
4993 else if (flags & SO_OBJECTS)
5001 page = slub_percpu_partial_read_once(c);
5003 node = page_to_nid(page);
5004 if (flags & SO_TOTAL)
5006 else if (flags & SO_OBJECTS)
5017 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5018 * already held which will conflict with an existing lock order:
5020 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5022 * We don't really need mem_hotplug_lock (to hold off
5023 * slab_mem_going_offline_callback) here because slab's memory hot
5024 * unplug code doesn't destroy the kmem_cache->node[] data.
5027 #ifdef CONFIG_SLUB_DEBUG
5028 if (flags & SO_ALL) {
5029 struct kmem_cache_node *n;
5031 for_each_kmem_cache_node(s, node, n) {
5033 if (flags & SO_TOTAL)
5034 x = atomic_long_read(&n->total_objects);
5035 else if (flags & SO_OBJECTS)
5036 x = atomic_long_read(&n->total_objects) -
5037 count_partial(n, count_free);
5039 x = atomic_long_read(&n->nr_slabs);
5046 if (flags & SO_PARTIAL) {
5047 struct kmem_cache_node *n;
5049 for_each_kmem_cache_node(s, node, n) {
5050 if (flags & SO_TOTAL)
5051 x = count_partial(n, count_total);
5052 else if (flags & SO_OBJECTS)
5053 x = count_partial(n, count_inuse);
5061 len += sysfs_emit_at(buf, len, "%lu", total);
5063 for (node = 0; node < nr_node_ids; node++) {
5065 len += sysfs_emit_at(buf, len, " N%d=%lu",
5069 len += sysfs_emit_at(buf, len, "\n");
5075 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5076 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5078 struct slab_attribute {
5079 struct attribute attr;
5080 ssize_t (*show)(struct kmem_cache *s, char *buf);
5081 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5084 #define SLAB_ATTR_RO(_name) \
5085 static struct slab_attribute _name##_attr = \
5086 __ATTR(_name, 0400, _name##_show, NULL)
5088 #define SLAB_ATTR(_name) \
5089 static struct slab_attribute _name##_attr = \
5090 __ATTR(_name, 0600, _name##_show, _name##_store)
5092 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5094 return sysfs_emit(buf, "%u\n", s->size);
5096 SLAB_ATTR_RO(slab_size);
5098 static ssize_t align_show(struct kmem_cache *s, char *buf)
5100 return sysfs_emit(buf, "%u\n", s->align);
5102 SLAB_ATTR_RO(align);
5104 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5106 return sysfs_emit(buf, "%u\n", s->object_size);
5108 SLAB_ATTR_RO(object_size);
5110 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5112 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5114 SLAB_ATTR_RO(objs_per_slab);
5116 static ssize_t order_show(struct kmem_cache *s, char *buf)
5118 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5120 SLAB_ATTR_RO(order);
5122 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5124 return sysfs_emit(buf, "%lu\n", s->min_partial);
5127 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5133 err = kstrtoul(buf, 10, &min);
5137 set_min_partial(s, min);
5140 SLAB_ATTR(min_partial);
5142 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5144 return sysfs_emit(buf, "%u\n", slub_cpu_partial(s));
5147 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5150 unsigned int objects;
5153 err = kstrtouint(buf, 10, &objects);
5156 if (objects && !kmem_cache_has_cpu_partial(s))
5159 slub_set_cpu_partial(s, objects);
5163 SLAB_ATTR(cpu_partial);
5165 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5169 return sysfs_emit(buf, "%pS\n", s->ctor);
5173 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5175 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5177 SLAB_ATTR_RO(aliases);
5179 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5181 return show_slab_objects(s, buf, SO_PARTIAL);
5183 SLAB_ATTR_RO(partial);
5185 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5187 return show_slab_objects(s, buf, SO_CPU);
5189 SLAB_ATTR_RO(cpu_slabs);
5191 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5193 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5195 SLAB_ATTR_RO(objects);
5197 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5199 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5201 SLAB_ATTR_RO(objects_partial);
5203 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5210 for_each_online_cpu(cpu) {
5213 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5216 pages += page->pages;
5217 objects += page->pobjects;
5221 len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages);
5224 for_each_online_cpu(cpu) {
5227 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5229 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5230 cpu, page->pobjects, page->pages);
5233 len += sysfs_emit_at(buf, len, "\n");
5237 SLAB_ATTR_RO(slabs_cpu_partial);
5239 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5241 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5243 SLAB_ATTR_RO(reclaim_account);
5245 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5247 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5249 SLAB_ATTR_RO(hwcache_align);
5251 #ifdef CONFIG_ZONE_DMA
5252 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5254 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5256 SLAB_ATTR_RO(cache_dma);
5259 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5261 return sysfs_emit(buf, "%u\n", s->usersize);
5263 SLAB_ATTR_RO(usersize);
5265 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5267 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5269 SLAB_ATTR_RO(destroy_by_rcu);
5271 #ifdef CONFIG_SLUB_DEBUG
5272 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5274 return show_slab_objects(s, buf, SO_ALL);
5276 SLAB_ATTR_RO(slabs);
5278 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5280 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5282 SLAB_ATTR_RO(total_objects);
5284 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5286 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5288 SLAB_ATTR_RO(sanity_checks);
5290 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5292 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5294 SLAB_ATTR_RO(trace);
5296 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5298 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5301 SLAB_ATTR_RO(red_zone);
5303 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5305 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5308 SLAB_ATTR_RO(poison);
5310 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5312 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5315 SLAB_ATTR_RO(store_user);
5317 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5322 static ssize_t validate_store(struct kmem_cache *s,
5323 const char *buf, size_t length)
5327 if (buf[0] == '1') {
5328 ret = validate_slab_cache(s);
5334 SLAB_ATTR(validate);
5336 #endif /* CONFIG_SLUB_DEBUG */
5338 #ifdef CONFIG_FAILSLAB
5339 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5341 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5343 SLAB_ATTR_RO(failslab);
5346 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5351 static ssize_t shrink_store(struct kmem_cache *s,
5352 const char *buf, size_t length)
5355 kmem_cache_shrink(s);
5363 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5365 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5368 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5369 const char *buf, size_t length)
5374 err = kstrtouint(buf, 10, &ratio);
5380 s->remote_node_defrag_ratio = ratio * 10;
5384 SLAB_ATTR(remote_node_defrag_ratio);
5387 #ifdef CONFIG_SLUB_STATS
5388 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5390 unsigned long sum = 0;
5393 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5398 for_each_online_cpu(cpu) {
5399 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5405 len += sysfs_emit_at(buf, len, "%lu", sum);
5408 for_each_online_cpu(cpu) {
5410 len += sysfs_emit_at(buf, len, " C%d=%u",
5415 len += sysfs_emit_at(buf, len, "\n");
5420 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5424 for_each_online_cpu(cpu)
5425 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5428 #define STAT_ATTR(si, text) \
5429 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5431 return show_stat(s, buf, si); \
5433 static ssize_t text##_store(struct kmem_cache *s, \
5434 const char *buf, size_t length) \
5436 if (buf[0] != '0') \
5438 clear_stat(s, si); \
5443 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5444 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5445 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5446 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5447 STAT_ATTR(FREE_FROZEN, free_frozen);
5448 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5449 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5450 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5451 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5452 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5453 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5454 STAT_ATTR(FREE_SLAB, free_slab);
5455 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5456 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5457 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5458 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5459 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5460 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5461 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5462 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5463 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5464 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5465 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5466 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5467 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5468 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5469 #endif /* CONFIG_SLUB_STATS */
5471 static struct attribute *slab_attrs[] = {
5472 &slab_size_attr.attr,
5473 &object_size_attr.attr,
5474 &objs_per_slab_attr.attr,
5476 &min_partial_attr.attr,
5477 &cpu_partial_attr.attr,
5479 &objects_partial_attr.attr,
5481 &cpu_slabs_attr.attr,
5485 &hwcache_align_attr.attr,
5486 &reclaim_account_attr.attr,
5487 &destroy_by_rcu_attr.attr,
5489 &slabs_cpu_partial_attr.attr,
5490 #ifdef CONFIG_SLUB_DEBUG
5491 &total_objects_attr.attr,
5493 &sanity_checks_attr.attr,
5495 &red_zone_attr.attr,
5497 &store_user_attr.attr,
5498 &validate_attr.attr,
5500 #ifdef CONFIG_ZONE_DMA
5501 &cache_dma_attr.attr,
5504 &remote_node_defrag_ratio_attr.attr,
5506 #ifdef CONFIG_SLUB_STATS
5507 &alloc_fastpath_attr.attr,
5508 &alloc_slowpath_attr.attr,
5509 &free_fastpath_attr.attr,
5510 &free_slowpath_attr.attr,
5511 &free_frozen_attr.attr,
5512 &free_add_partial_attr.attr,
5513 &free_remove_partial_attr.attr,
5514 &alloc_from_partial_attr.attr,
5515 &alloc_slab_attr.attr,
5516 &alloc_refill_attr.attr,
5517 &alloc_node_mismatch_attr.attr,
5518 &free_slab_attr.attr,
5519 &cpuslab_flush_attr.attr,
5520 &deactivate_full_attr.attr,
5521 &deactivate_empty_attr.attr,
5522 &deactivate_to_head_attr.attr,
5523 &deactivate_to_tail_attr.attr,
5524 &deactivate_remote_frees_attr.attr,
5525 &deactivate_bypass_attr.attr,
5526 &order_fallback_attr.attr,
5527 &cmpxchg_double_fail_attr.attr,
5528 &cmpxchg_double_cpu_fail_attr.attr,
5529 &cpu_partial_alloc_attr.attr,
5530 &cpu_partial_free_attr.attr,
5531 &cpu_partial_node_attr.attr,
5532 &cpu_partial_drain_attr.attr,
5534 #ifdef CONFIG_FAILSLAB
5535 &failslab_attr.attr,
5537 &usersize_attr.attr,
5542 static const struct attribute_group slab_attr_group = {
5543 .attrs = slab_attrs,
5546 static ssize_t slab_attr_show(struct kobject *kobj,
5547 struct attribute *attr,
5550 struct slab_attribute *attribute;
5551 struct kmem_cache *s;
5554 attribute = to_slab_attr(attr);
5557 if (!attribute->show)
5560 err = attribute->show(s, buf);
5565 static ssize_t slab_attr_store(struct kobject *kobj,
5566 struct attribute *attr,
5567 const char *buf, size_t len)
5569 struct slab_attribute *attribute;
5570 struct kmem_cache *s;
5573 attribute = to_slab_attr(attr);
5576 if (!attribute->store)
5579 err = attribute->store(s, buf, len);
5583 static void kmem_cache_release(struct kobject *k)
5585 slab_kmem_cache_release(to_slab(k));
5588 static const struct sysfs_ops slab_sysfs_ops = {
5589 .show = slab_attr_show,
5590 .store = slab_attr_store,
5593 static struct kobj_type slab_ktype = {
5594 .sysfs_ops = &slab_sysfs_ops,
5595 .release = kmem_cache_release,
5598 static struct kset *slab_kset;
5600 static inline struct kset *cache_kset(struct kmem_cache *s)
5605 #define ID_STR_LENGTH 64
5607 /* Create a unique string id for a slab cache:
5609 * Format :[flags-]size
5611 static char *create_unique_id(struct kmem_cache *s)
5613 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5620 * First flags affecting slabcache operations. We will only
5621 * get here for aliasable slabs so we do not need to support
5622 * too many flags. The flags here must cover all flags that
5623 * are matched during merging to guarantee that the id is
5626 if (s->flags & SLAB_CACHE_DMA)
5628 if (s->flags & SLAB_CACHE_DMA32)
5630 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5632 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5634 if (s->flags & SLAB_ACCOUNT)
5638 p += sprintf(p, "%07u", s->size);
5640 BUG_ON(p > name + ID_STR_LENGTH - 1);
5644 static int sysfs_slab_add(struct kmem_cache *s)
5648 struct kset *kset = cache_kset(s);
5649 int unmergeable = slab_unmergeable(s);
5652 kobject_init(&s->kobj, &slab_ktype);
5656 if (!unmergeable && disable_higher_order_debug &&
5657 (slub_debug & DEBUG_METADATA_FLAGS))
5662 * Slabcache can never be merged so we can use the name proper.
5663 * This is typically the case for debug situations. In that
5664 * case we can catch duplicate names easily.
5666 sysfs_remove_link(&slab_kset->kobj, s->name);
5670 * Create a unique name for the slab as a target
5673 name = create_unique_id(s);
5676 s->kobj.kset = kset;
5677 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5681 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5686 /* Setup first alias */
5687 sysfs_slab_alias(s, s->name);
5694 kobject_del(&s->kobj);
5698 void sysfs_slab_unlink(struct kmem_cache *s)
5700 if (slab_state >= FULL)
5701 kobject_del(&s->kobj);
5704 void sysfs_slab_release(struct kmem_cache *s)
5706 if (slab_state >= FULL)
5707 kobject_put(&s->kobj);
5711 * Need to buffer aliases during bootup until sysfs becomes
5712 * available lest we lose that information.
5714 struct saved_alias {
5715 struct kmem_cache *s;
5717 struct saved_alias *next;
5720 static struct saved_alias *alias_list;
5722 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5724 struct saved_alias *al;
5726 if (slab_state == FULL) {
5728 * If we have a leftover link then remove it.
5730 sysfs_remove_link(&slab_kset->kobj, name);
5731 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5734 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5740 al->next = alias_list;
5745 static int __init slab_sysfs_init(void)
5747 struct kmem_cache *s;
5750 mutex_lock(&slab_mutex);
5752 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5754 mutex_unlock(&slab_mutex);
5755 pr_err("Cannot register slab subsystem.\n");
5761 list_for_each_entry(s, &slab_caches, list) {
5762 err = sysfs_slab_add(s);
5764 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5768 while (alias_list) {
5769 struct saved_alias *al = alias_list;
5771 alias_list = alias_list->next;
5772 err = sysfs_slab_alias(al->s, al->name);
5774 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5779 mutex_unlock(&slab_mutex);
5783 __initcall(slab_sysfs_init);
5784 #endif /* CONFIG_SYSFS */
5786 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
5787 static int slab_debugfs_show(struct seq_file *seq, void *v)
5791 unsigned int idx = *(unsigned int *)v;
5792 struct loc_track *t = seq->private;
5794 if (idx < t->count) {
5797 seq_printf(seq, "%7ld ", l->count);
5800 seq_printf(seq, "%pS", (void *)l->addr);
5802 seq_puts(seq, "<not-available>");
5804 if (l->sum_time != l->min_time) {
5805 seq_printf(seq, " age=%ld/%llu/%ld",
5806 l->min_time, div_u64(l->sum_time, l->count),
5809 seq_printf(seq, " age=%ld", l->min_time);
5811 if (l->min_pid != l->max_pid)
5812 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
5814 seq_printf(seq, " pid=%ld",
5817 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
5818 seq_printf(seq, " cpus=%*pbl",
5819 cpumask_pr_args(to_cpumask(l->cpus)));
5821 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
5822 seq_printf(seq, " nodes=%*pbl",
5823 nodemask_pr_args(&l->nodes));
5825 seq_puts(seq, "\n");
5828 if (!idx && !t->count)
5829 seq_puts(seq, "No data\n");
5834 static void slab_debugfs_stop(struct seq_file *seq, void *v)
5838 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
5840 struct loc_track *t = seq->private;
5844 if (*ppos <= t->count)
5850 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
5855 static const struct seq_operations slab_debugfs_sops = {
5856 .start = slab_debugfs_start,
5857 .next = slab_debugfs_next,
5858 .stop = slab_debugfs_stop,
5859 .show = slab_debugfs_show,
5862 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
5865 struct kmem_cache_node *n;
5866 enum track_item alloc;
5868 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
5869 sizeof(struct loc_track));
5870 struct kmem_cache *s = file_inode(filep)->i_private;
5871 unsigned long *obj_map;
5873 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5877 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
5878 alloc = TRACK_ALLOC;
5882 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
5883 bitmap_free(obj_map);
5887 for_each_kmem_cache_node(s, node, n) {
5888 unsigned long flags;
5891 if (!atomic_long_read(&n->nr_slabs))
5894 spin_lock_irqsave(&n->list_lock, flags);
5895 list_for_each_entry(page, &n->partial, slab_list)
5896 process_slab(t, s, page, alloc, obj_map);
5897 list_for_each_entry(page, &n->full, slab_list)
5898 process_slab(t, s, page, alloc, obj_map);
5899 spin_unlock_irqrestore(&n->list_lock, flags);
5902 bitmap_free(obj_map);
5906 static int slab_debug_trace_release(struct inode *inode, struct file *file)
5908 struct seq_file *seq = file->private_data;
5909 struct loc_track *t = seq->private;
5912 return seq_release_private(inode, file);
5915 static const struct file_operations slab_debugfs_fops = {
5916 .open = slab_debug_trace_open,
5918 .llseek = seq_lseek,
5919 .release = slab_debug_trace_release,
5922 static void debugfs_slab_add(struct kmem_cache *s)
5924 struct dentry *slab_cache_dir;
5926 if (unlikely(!slab_debugfs_root))
5929 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
5931 debugfs_create_file("alloc_traces", 0400,
5932 slab_cache_dir, s, &slab_debugfs_fops);
5934 debugfs_create_file("free_traces", 0400,
5935 slab_cache_dir, s, &slab_debugfs_fops);
5938 void debugfs_slab_release(struct kmem_cache *s)
5940 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
5943 static int __init slab_debugfs_init(void)
5945 struct kmem_cache *s;
5947 slab_debugfs_root = debugfs_create_dir("slab", NULL);
5949 list_for_each_entry(s, &slab_caches, list)
5950 if (s->flags & SLAB_STORE_USER)
5951 debugfs_slab_add(s);
5956 __initcall(slab_debugfs_init);
5959 * The /proc/slabinfo ABI
5961 #ifdef CONFIG_SLUB_DEBUG
5962 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5964 unsigned long nr_slabs = 0;
5965 unsigned long nr_objs = 0;
5966 unsigned long nr_free = 0;
5968 struct kmem_cache_node *n;
5970 for_each_kmem_cache_node(s, node, n) {
5971 nr_slabs += node_nr_slabs(n);
5972 nr_objs += node_nr_objs(n);
5973 nr_free += count_partial(n, count_free);
5976 sinfo->active_objs = nr_objs - nr_free;
5977 sinfo->num_objs = nr_objs;
5978 sinfo->active_slabs = nr_slabs;
5979 sinfo->num_slabs = nr_slabs;
5980 sinfo->objects_per_slab = oo_objects(s->oo);
5981 sinfo->cache_order = oo_order(s->oo);
5984 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5988 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5989 size_t count, loff_t *ppos)
5993 #endif /* CONFIG_SLUB_DEBUG */