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 if (!PageSlab(page)) {
1013 slab_err(s, page, "Not a valid slab page");
1017 maxobj = order_objects(compound_order(page), s->size);
1018 if (page->objects > maxobj) {
1019 slab_err(s, page, "objects %u > max %u",
1020 page->objects, maxobj);
1023 if (page->inuse > page->objects) {
1024 slab_err(s, page, "inuse %u > max %u",
1025 page->inuse, page->objects);
1028 /* Slab_pad_check fixes things up after itself */
1029 slab_pad_check(s, page);
1034 * Determine if a certain object on a page is on the freelist. Must hold the
1035 * slab lock to guarantee that the chains are in a consistent state.
1037 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
1041 void *object = NULL;
1044 fp = page->freelist;
1045 while (fp && nr <= page->objects) {
1048 if (!check_valid_pointer(s, page, fp)) {
1050 object_err(s, page, object,
1051 "Freechain corrupt");
1052 set_freepointer(s, object, NULL);
1054 slab_err(s, page, "Freepointer corrupt");
1055 page->freelist = NULL;
1056 page->inuse = page->objects;
1057 slab_fix(s, "Freelist cleared");
1063 fp = get_freepointer(s, object);
1067 max_objects = order_objects(compound_order(page), s->size);
1068 if (max_objects > MAX_OBJS_PER_PAGE)
1069 max_objects = MAX_OBJS_PER_PAGE;
1071 if (page->objects != max_objects) {
1072 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1073 page->objects, max_objects);
1074 page->objects = max_objects;
1075 slab_fix(s, "Number of objects adjusted");
1077 if (page->inuse != page->objects - nr) {
1078 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1079 page->inuse, page->objects - nr);
1080 page->inuse = page->objects - nr;
1081 slab_fix(s, "Object count adjusted");
1083 return search == NULL;
1086 static void trace(struct kmem_cache *s, struct page *page, void *object,
1089 if (s->flags & SLAB_TRACE) {
1090 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1092 alloc ? "alloc" : "free",
1093 object, page->inuse,
1097 print_section(KERN_INFO, "Object ", (void *)object,
1105 * Tracking of fully allocated slabs for debugging purposes.
1107 static void add_full(struct kmem_cache *s,
1108 struct kmem_cache_node *n, struct page *page)
1110 if (!(s->flags & SLAB_STORE_USER))
1113 lockdep_assert_held(&n->list_lock);
1114 list_add(&page->slab_list, &n->full);
1117 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1119 if (!(s->flags & SLAB_STORE_USER))
1122 lockdep_assert_held(&n->list_lock);
1123 list_del(&page->slab_list);
1126 /* Tracking of the number of slabs for debugging purposes */
1127 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1129 struct kmem_cache_node *n = get_node(s, node);
1131 return atomic_long_read(&n->nr_slabs);
1134 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1136 return atomic_long_read(&n->nr_slabs);
1139 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1141 struct kmem_cache_node *n = get_node(s, node);
1144 * May be called early in order to allocate a slab for the
1145 * kmem_cache_node structure. Solve the chicken-egg
1146 * dilemma by deferring the increment of the count during
1147 * bootstrap (see early_kmem_cache_node_alloc).
1150 atomic_long_inc(&n->nr_slabs);
1151 atomic_long_add(objects, &n->total_objects);
1154 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1156 struct kmem_cache_node *n = get_node(s, node);
1158 atomic_long_dec(&n->nr_slabs);
1159 atomic_long_sub(objects, &n->total_objects);
1162 /* Object debug checks for alloc/free paths */
1163 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1166 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1169 init_object(s, object, SLUB_RED_INACTIVE);
1170 init_tracking(s, object);
1174 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1176 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1179 metadata_access_enable();
1180 memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page));
1181 metadata_access_disable();
1184 static inline int alloc_consistency_checks(struct kmem_cache *s,
1185 struct page *page, void *object)
1187 if (!check_slab(s, page))
1190 if (!check_valid_pointer(s, page, object)) {
1191 object_err(s, page, object, "Freelist Pointer check fails");
1195 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1201 static noinline int alloc_debug_processing(struct kmem_cache *s,
1203 void *object, unsigned long addr)
1205 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1206 if (!alloc_consistency_checks(s, page, object))
1210 /* Success perform special debug activities for allocs */
1211 if (s->flags & SLAB_STORE_USER)
1212 set_track(s, object, TRACK_ALLOC, addr);
1213 trace(s, page, object, 1);
1214 init_object(s, object, SLUB_RED_ACTIVE);
1218 if (PageSlab(page)) {
1220 * If this is a slab page then lets do the best we can
1221 * to avoid issues in the future. Marking all objects
1222 * as used avoids touching the remaining objects.
1224 slab_fix(s, "Marking all objects used");
1225 page->inuse = page->objects;
1226 page->freelist = NULL;
1231 static inline int free_consistency_checks(struct kmem_cache *s,
1232 struct page *page, void *object, unsigned long addr)
1234 if (!check_valid_pointer(s, page, object)) {
1235 slab_err(s, page, "Invalid object pointer 0x%p", object);
1239 if (on_freelist(s, page, object)) {
1240 object_err(s, page, object, "Object already free");
1244 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1247 if (unlikely(s != page->slab_cache)) {
1248 if (!PageSlab(page)) {
1249 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1251 } else if (!page->slab_cache) {
1252 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1256 object_err(s, page, object,
1257 "page slab pointer corrupt.");
1263 /* Supports checking bulk free of a constructed freelist */
1264 static noinline int free_debug_processing(
1265 struct kmem_cache *s, struct page *page,
1266 void *head, void *tail, int bulk_cnt,
1269 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1270 void *object = head;
1272 unsigned long flags;
1275 spin_lock_irqsave(&n->list_lock, flags);
1278 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1279 if (!check_slab(s, page))
1286 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1287 if (!free_consistency_checks(s, page, object, addr))
1291 if (s->flags & SLAB_STORE_USER)
1292 set_track(s, object, TRACK_FREE, addr);
1293 trace(s, page, object, 0);
1294 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1295 init_object(s, object, SLUB_RED_INACTIVE);
1297 /* Reached end of constructed freelist yet? */
1298 if (object != tail) {
1299 object = get_freepointer(s, object);
1305 if (cnt != bulk_cnt)
1306 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1310 spin_unlock_irqrestore(&n->list_lock, flags);
1312 slab_fix(s, "Object at 0x%p not freed", object);
1317 * Parse a block of slub_debug options. Blocks are delimited by ';'
1319 * @str: start of block
1320 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1321 * @slabs: return start of list of slabs, or NULL when there's no list
1322 * @init: assume this is initial parsing and not per-kmem-create parsing
1324 * returns the start of next block if there's any, or NULL
1327 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1329 bool higher_order_disable = false;
1331 /* Skip any completely empty blocks */
1332 while (*str && *str == ';')
1337 * No options but restriction on slabs. This means full
1338 * debugging for slabs matching a pattern.
1340 *flags = DEBUG_DEFAULT_FLAGS;
1345 /* Determine which debug features should be switched on */
1346 for (; *str && *str != ',' && *str != ';'; str++) {
1347 switch (tolower(*str)) {
1352 *flags |= SLAB_CONSISTENCY_CHECKS;
1355 *flags |= SLAB_RED_ZONE;
1358 *flags |= SLAB_POISON;
1361 *flags |= SLAB_STORE_USER;
1364 *flags |= SLAB_TRACE;
1367 *flags |= SLAB_FAILSLAB;
1371 * Avoid enabling debugging on caches if its minimum
1372 * order would increase as a result.
1374 higher_order_disable = true;
1378 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1387 /* Skip over the slab list */
1388 while (*str && *str != ';')
1391 /* Skip any completely empty blocks */
1392 while (*str && *str == ';')
1395 if (init && higher_order_disable)
1396 disable_higher_order_debug = 1;
1404 static int __init setup_slub_debug(char *str)
1407 slab_flags_t global_flags;
1410 bool global_slub_debug_changed = false;
1411 bool slab_list_specified = false;
1413 global_flags = DEBUG_DEFAULT_FLAGS;
1414 if (*str++ != '=' || !*str)
1416 * No options specified. Switch on full debugging.
1422 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1425 global_flags = flags;
1426 global_slub_debug_changed = true;
1428 slab_list_specified = true;
1433 * For backwards compatibility, a single list of flags with list of
1434 * slabs means debugging is only changed for those slabs, so the global
1435 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1436 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1437 * long as there is no option specifying flags without a slab list.
1439 if (slab_list_specified) {
1440 if (!global_slub_debug_changed)
1441 global_flags = slub_debug;
1442 slub_debug_string = saved_str;
1445 slub_debug = global_flags;
1446 if (slub_debug != 0 || slub_debug_string)
1447 static_branch_enable(&slub_debug_enabled);
1449 static_branch_disable(&slub_debug_enabled);
1450 if ((static_branch_unlikely(&init_on_alloc) ||
1451 static_branch_unlikely(&init_on_free)) &&
1452 (slub_debug & SLAB_POISON))
1453 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1457 __setup("slub_debug", setup_slub_debug);
1460 * kmem_cache_flags - apply debugging options to the cache
1461 * @object_size: the size of an object without meta data
1462 * @flags: flags to set
1463 * @name: name of the cache
1465 * Debug option(s) are applied to @flags. In addition to the debug
1466 * option(s), if a slab name (or multiple) is specified i.e.
1467 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1468 * then only the select slabs will receive the debug option(s).
1470 slab_flags_t kmem_cache_flags(unsigned int object_size,
1471 slab_flags_t flags, const char *name)
1476 slab_flags_t block_flags;
1477 slab_flags_t slub_debug_local = slub_debug;
1480 * If the slab cache is for debugging (e.g. kmemleak) then
1481 * don't store user (stack trace) information by default,
1482 * but let the user enable it via the command line below.
1484 if (flags & SLAB_NOLEAKTRACE)
1485 slub_debug_local &= ~SLAB_STORE_USER;
1488 next_block = slub_debug_string;
1489 /* Go through all blocks of debug options, see if any matches our slab's name */
1490 while (next_block) {
1491 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1494 /* Found a block that has a slab list, search it */
1499 end = strchrnul(iter, ',');
1500 if (next_block && next_block < end)
1501 end = next_block - 1;
1503 glob = strnchr(iter, end - iter, '*');
1505 cmplen = glob - iter;
1507 cmplen = max_t(size_t, len, (end - iter));
1509 if (!strncmp(name, iter, cmplen)) {
1510 flags |= block_flags;
1514 if (!*end || *end == ';')
1520 return flags | slub_debug_local;
1522 #else /* !CONFIG_SLUB_DEBUG */
1523 static inline void setup_object_debug(struct kmem_cache *s,
1524 struct page *page, void *object) {}
1526 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1528 static inline int alloc_debug_processing(struct kmem_cache *s,
1529 struct page *page, void *object, unsigned long addr) { return 0; }
1531 static inline int free_debug_processing(
1532 struct kmem_cache *s, struct page *page,
1533 void *head, void *tail, int bulk_cnt,
1534 unsigned long addr) { return 0; }
1536 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1538 static inline int check_object(struct kmem_cache *s, struct page *page,
1539 void *object, u8 val) { return 1; }
1540 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1541 struct page *page) {}
1542 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1543 struct page *page) {}
1544 slab_flags_t kmem_cache_flags(unsigned int object_size,
1545 slab_flags_t flags, const char *name)
1549 #define slub_debug 0
1551 #define disable_higher_order_debug 0
1553 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1555 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1557 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1559 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1562 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1563 void **freelist, void *nextfree)
1567 #endif /* CONFIG_SLUB_DEBUG */
1570 * Hooks for other subsystems that check memory allocations. In a typical
1571 * production configuration these hooks all should produce no code at all.
1573 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1575 ptr = kasan_kmalloc_large(ptr, size, flags);
1576 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1577 kmemleak_alloc(ptr, size, 1, flags);
1581 static __always_inline void kfree_hook(void *x)
1584 kasan_kfree_large(x);
1587 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1590 kmemleak_free_recursive(x, s->flags);
1592 debug_check_no_locks_freed(x, s->object_size);
1594 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1595 debug_check_no_obj_freed(x, s->object_size);
1597 /* Use KCSAN to help debug racy use-after-free. */
1598 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1599 __kcsan_check_access(x, s->object_size,
1600 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1603 * As memory initialization might be integrated into KASAN,
1604 * kasan_slab_free and initialization memset's must be
1605 * kept together to avoid discrepancies in behavior.
1607 * The initialization memset's clear the object and the metadata,
1608 * but don't touch the SLAB redzone.
1613 if (!kasan_has_integrated_init())
1614 memset(kasan_reset_tag(x), 0, s->object_size);
1615 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1616 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1617 s->size - s->inuse - rsize);
1619 /* KASAN might put x into memory quarantine, delaying its reuse. */
1620 return kasan_slab_free(s, x, init);
1623 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1624 void **head, void **tail)
1629 void *old_tail = *tail ? *tail : *head;
1631 if (is_kfence_address(next)) {
1632 slab_free_hook(s, next, false);
1636 /* Head and tail of the reconstructed freelist */
1642 next = get_freepointer(s, object);
1644 /* If object's reuse doesn't have to be delayed */
1645 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1646 /* Move object to the new freelist */
1647 set_freepointer(s, object, *head);
1652 } while (object != old_tail);
1657 return *head != NULL;
1660 static void *setup_object(struct kmem_cache *s, struct page *page,
1663 setup_object_debug(s, page, object);
1664 object = kasan_init_slab_obj(s, object);
1665 if (unlikely(s->ctor)) {
1666 kasan_unpoison_object_data(s, object);
1668 kasan_poison_object_data(s, object);
1674 * Slab allocation and freeing
1676 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1677 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1680 unsigned int order = oo_order(oo);
1682 if (node == NUMA_NO_NODE)
1683 page = alloc_pages(flags, order);
1685 page = __alloc_pages_node(node, flags, order);
1690 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1691 /* Pre-initialize the random sequence cache */
1692 static int init_cache_random_seq(struct kmem_cache *s)
1694 unsigned int count = oo_objects(s->oo);
1697 /* Bailout if already initialised */
1701 err = cache_random_seq_create(s, count, GFP_KERNEL);
1703 pr_err("SLUB: Unable to initialize free list for %s\n",
1708 /* Transform to an offset on the set of pages */
1709 if (s->random_seq) {
1712 for (i = 0; i < count; i++)
1713 s->random_seq[i] *= s->size;
1718 /* Initialize each random sequence freelist per cache */
1719 static void __init init_freelist_randomization(void)
1721 struct kmem_cache *s;
1723 mutex_lock(&slab_mutex);
1725 list_for_each_entry(s, &slab_caches, list)
1726 init_cache_random_seq(s);
1728 mutex_unlock(&slab_mutex);
1731 /* Get the next entry on the pre-computed freelist randomized */
1732 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1733 unsigned long *pos, void *start,
1734 unsigned long page_limit,
1735 unsigned long freelist_count)
1740 * If the target page allocation failed, the number of objects on the
1741 * page might be smaller than the usual size defined by the cache.
1744 idx = s->random_seq[*pos];
1746 if (*pos >= freelist_count)
1748 } while (unlikely(idx >= page_limit));
1750 return (char *)start + idx;
1753 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1754 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1759 unsigned long idx, pos, page_limit, freelist_count;
1761 if (page->objects < 2 || !s->random_seq)
1764 freelist_count = oo_objects(s->oo);
1765 pos = get_random_int() % freelist_count;
1767 page_limit = page->objects * s->size;
1768 start = fixup_red_left(s, page_address(page));
1770 /* First entry is used as the base of the freelist */
1771 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1773 cur = setup_object(s, page, cur);
1774 page->freelist = cur;
1776 for (idx = 1; idx < page->objects; idx++) {
1777 next = next_freelist_entry(s, page, &pos, start, page_limit,
1779 next = setup_object(s, page, next);
1780 set_freepointer(s, cur, next);
1783 set_freepointer(s, cur, NULL);
1788 static inline int init_cache_random_seq(struct kmem_cache *s)
1792 static inline void init_freelist_randomization(void) { }
1793 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1797 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1799 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1802 struct kmem_cache_order_objects oo = s->oo;
1804 void *start, *p, *next;
1808 flags &= gfp_allowed_mask;
1810 flags |= s->allocflags;
1813 * Let the initial higher-order allocation fail under memory pressure
1814 * so we fall-back to the minimum order allocation.
1816 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1817 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1818 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1820 page = alloc_slab_page(s, alloc_gfp, node, oo);
1821 if (unlikely(!page)) {
1825 * Allocation may have failed due to fragmentation.
1826 * Try a lower order alloc if possible
1828 page = alloc_slab_page(s, alloc_gfp, node, oo);
1829 if (unlikely(!page))
1831 stat(s, ORDER_FALLBACK);
1834 page->objects = oo_objects(oo);
1836 account_slab_page(page, oo_order(oo), s, flags);
1838 page->slab_cache = s;
1839 __SetPageSlab(page);
1840 if (page_is_pfmemalloc(page))
1841 SetPageSlabPfmemalloc(page);
1843 kasan_poison_slab(page);
1845 start = page_address(page);
1847 setup_page_debug(s, page, start);
1849 shuffle = shuffle_freelist(s, page);
1852 start = fixup_red_left(s, start);
1853 start = setup_object(s, page, start);
1854 page->freelist = start;
1855 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1857 next = setup_object(s, page, next);
1858 set_freepointer(s, p, next);
1861 set_freepointer(s, p, NULL);
1864 page->inuse = page->objects;
1871 inc_slabs_node(s, page_to_nid(page), page->objects);
1876 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1878 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1879 flags = kmalloc_fix_flags(flags);
1881 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
1883 return allocate_slab(s,
1884 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1887 static void __free_slab(struct kmem_cache *s, struct page *page)
1889 int order = compound_order(page);
1890 int pages = 1 << order;
1892 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1895 slab_pad_check(s, page);
1896 for_each_object(p, s, page_address(page),
1898 check_object(s, page, p, SLUB_RED_INACTIVE);
1901 __ClearPageSlabPfmemalloc(page);
1902 __ClearPageSlab(page);
1903 /* In union with page->mapping where page allocator expects NULL */
1904 page->slab_cache = NULL;
1905 if (current->reclaim_state)
1906 current->reclaim_state->reclaimed_slab += pages;
1907 unaccount_slab_page(page, order, s);
1908 __free_pages(page, order);
1911 static void rcu_free_slab(struct rcu_head *h)
1913 struct page *page = container_of(h, struct page, rcu_head);
1915 __free_slab(page->slab_cache, page);
1918 static void free_slab(struct kmem_cache *s, struct page *page)
1920 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1921 call_rcu(&page->rcu_head, rcu_free_slab);
1923 __free_slab(s, page);
1926 static void discard_slab(struct kmem_cache *s, struct page *page)
1928 dec_slabs_node(s, page_to_nid(page), page->objects);
1933 * Management of partially allocated slabs.
1936 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1939 if (tail == DEACTIVATE_TO_TAIL)
1940 list_add_tail(&page->slab_list, &n->partial);
1942 list_add(&page->slab_list, &n->partial);
1945 static inline void add_partial(struct kmem_cache_node *n,
1946 struct page *page, int tail)
1948 lockdep_assert_held(&n->list_lock);
1949 __add_partial(n, page, tail);
1952 static inline void remove_partial(struct kmem_cache_node *n,
1955 lockdep_assert_held(&n->list_lock);
1956 list_del(&page->slab_list);
1961 * Remove slab from the partial list, freeze it and
1962 * return the pointer to the freelist.
1964 * Returns a list of objects or NULL if it fails.
1966 static inline void *acquire_slab(struct kmem_cache *s,
1967 struct kmem_cache_node *n, struct page *page,
1968 int mode, int *objects)
1971 unsigned long counters;
1974 lockdep_assert_held(&n->list_lock);
1977 * Zap the freelist and set the frozen bit.
1978 * The old freelist is the list of objects for the
1979 * per cpu allocation list.
1981 freelist = page->freelist;
1982 counters = page->counters;
1983 new.counters = counters;
1984 *objects = new.objects - new.inuse;
1986 new.inuse = page->objects;
1987 new.freelist = NULL;
1989 new.freelist = freelist;
1992 VM_BUG_ON(new.frozen);
1995 if (!__cmpxchg_double_slab(s, page,
1997 new.freelist, new.counters,
2001 remove_partial(n, page);
2006 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
2007 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
2010 * Try to allocate a partial slab from a specific node.
2012 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2013 struct page **ret_page, gfp_t gfpflags)
2015 struct page *page, *page2;
2016 void *object = NULL;
2017 unsigned int available = 0;
2018 unsigned long flags;
2022 * Racy check. If we mistakenly see no partial slabs then we
2023 * just allocate an empty slab. If we mistakenly try to get a
2024 * partial slab and there is none available then get_partial()
2027 if (!n || !n->nr_partial)
2030 spin_lock_irqsave(&n->list_lock, flags);
2031 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
2034 if (!pfmemalloc_match(page, gfpflags))
2037 t = acquire_slab(s, n, page, object == NULL, &objects);
2041 available += objects;
2044 stat(s, ALLOC_FROM_PARTIAL);
2047 put_cpu_partial(s, page, 0);
2048 stat(s, CPU_PARTIAL_NODE);
2050 if (!kmem_cache_has_cpu_partial(s)
2051 || available > slub_cpu_partial(s) / 2)
2055 spin_unlock_irqrestore(&n->list_lock, flags);
2060 * Get a page from somewhere. Search in increasing NUMA distances.
2062 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2063 struct page **ret_page)
2066 struct zonelist *zonelist;
2069 enum zone_type highest_zoneidx = gfp_zone(flags);
2071 unsigned int cpuset_mems_cookie;
2074 * The defrag ratio allows a configuration of the tradeoffs between
2075 * inter node defragmentation and node local allocations. A lower
2076 * defrag_ratio increases the tendency to do local allocations
2077 * instead of attempting to obtain partial slabs from other nodes.
2079 * If the defrag_ratio is set to 0 then kmalloc() always
2080 * returns node local objects. If the ratio is higher then kmalloc()
2081 * may return off node objects because partial slabs are obtained
2082 * from other nodes and filled up.
2084 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2085 * (which makes defrag_ratio = 1000) then every (well almost)
2086 * allocation will first attempt to defrag slab caches on other nodes.
2087 * This means scanning over all nodes to look for partial slabs which
2088 * may be expensive if we do it every time we are trying to find a slab
2089 * with available objects.
2091 if (!s->remote_node_defrag_ratio ||
2092 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2096 cpuset_mems_cookie = read_mems_allowed_begin();
2097 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2098 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2099 struct kmem_cache_node *n;
2101 n = get_node(s, zone_to_nid(zone));
2103 if (n && cpuset_zone_allowed(zone, flags) &&
2104 n->nr_partial > s->min_partial) {
2105 object = get_partial_node(s, n, ret_page, flags);
2108 * Don't check read_mems_allowed_retry()
2109 * here - if mems_allowed was updated in
2110 * parallel, that was a harmless race
2111 * between allocation and the cpuset
2118 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2119 #endif /* CONFIG_NUMA */
2124 * Get a partial page, lock it and return it.
2126 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2127 struct page **ret_page)
2130 int searchnode = node;
2132 if (node == NUMA_NO_NODE)
2133 searchnode = numa_mem_id();
2135 object = get_partial_node(s, get_node(s, searchnode), ret_page, flags);
2136 if (object || node != NUMA_NO_NODE)
2139 return get_any_partial(s, flags, ret_page);
2142 #ifdef CONFIG_PREEMPTION
2144 * Calculate the next globally unique transaction for disambiguation
2145 * during cmpxchg. The transactions start with the cpu number and are then
2146 * incremented by CONFIG_NR_CPUS.
2148 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2151 * No preemption supported therefore also no need to check for
2157 static inline unsigned long next_tid(unsigned long tid)
2159 return tid + TID_STEP;
2162 #ifdef SLUB_DEBUG_CMPXCHG
2163 static inline unsigned int tid_to_cpu(unsigned long tid)
2165 return tid % TID_STEP;
2168 static inline unsigned long tid_to_event(unsigned long tid)
2170 return tid / TID_STEP;
2174 static inline unsigned int init_tid(int cpu)
2179 static inline void note_cmpxchg_failure(const char *n,
2180 const struct kmem_cache *s, unsigned long tid)
2182 #ifdef SLUB_DEBUG_CMPXCHG
2183 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2185 pr_info("%s %s: cmpxchg redo ", n, s->name);
2187 #ifdef CONFIG_PREEMPTION
2188 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2189 pr_warn("due to cpu change %d -> %d\n",
2190 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2193 if (tid_to_event(tid) != tid_to_event(actual_tid))
2194 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2195 tid_to_event(tid), tid_to_event(actual_tid));
2197 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2198 actual_tid, tid, next_tid(tid));
2200 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2203 static void init_kmem_cache_cpus(struct kmem_cache *s)
2207 for_each_possible_cpu(cpu)
2208 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2212 * Finishes removing the cpu slab. Merges cpu's freelist with page's freelist,
2213 * unfreezes the slabs and puts it on the proper list.
2214 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2217 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2220 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2221 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2222 int lock = 0, free_delta = 0;
2223 enum slab_modes l = M_NONE, m = M_NONE;
2224 void *nextfree, *freelist_iter, *freelist_tail;
2225 int tail = DEACTIVATE_TO_HEAD;
2226 unsigned long flags = 0;
2230 if (page->freelist) {
2231 stat(s, DEACTIVATE_REMOTE_FREES);
2232 tail = DEACTIVATE_TO_TAIL;
2236 * Stage one: Count the objects on cpu's freelist as free_delta and
2237 * remember the last object in freelist_tail for later splicing.
2239 freelist_tail = NULL;
2240 freelist_iter = freelist;
2241 while (freelist_iter) {
2242 nextfree = get_freepointer(s, freelist_iter);
2245 * If 'nextfree' is invalid, it is possible that the object at
2246 * 'freelist_iter' is already corrupted. So isolate all objects
2247 * starting at 'freelist_iter' by skipping them.
2249 if (freelist_corrupted(s, page, &freelist_iter, nextfree))
2252 freelist_tail = freelist_iter;
2255 freelist_iter = nextfree;
2259 * Stage two: Unfreeze the page while splicing the per-cpu
2260 * freelist to the head of page's freelist.
2262 * Ensure that the page is unfrozen while the list presence
2263 * reflects the actual number of objects during unfreeze.
2265 * We setup the list membership and then perform a cmpxchg
2266 * with the count. If there is a mismatch then the page
2267 * is not unfrozen but the page is on the wrong list.
2269 * Then we restart the process which may have to remove
2270 * the page from the list that we just put it on again
2271 * because the number of objects in the slab may have
2276 old.freelist = READ_ONCE(page->freelist);
2277 old.counters = READ_ONCE(page->counters);
2278 VM_BUG_ON(!old.frozen);
2280 /* Determine target state of the slab */
2281 new.counters = old.counters;
2282 if (freelist_tail) {
2283 new.inuse -= free_delta;
2284 set_freepointer(s, freelist_tail, old.freelist);
2285 new.freelist = freelist;
2287 new.freelist = old.freelist;
2291 if (!new.inuse && n->nr_partial >= s->min_partial)
2293 else if (new.freelist) {
2298 * Taking the spinlock removes the possibility
2299 * that acquire_slab() will see a slab page that
2302 spin_lock_irqsave(&n->list_lock, flags);
2306 if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) {
2309 * This also ensures that the scanning of full
2310 * slabs from diagnostic functions will not see
2313 spin_lock_irqsave(&n->list_lock, flags);
2319 remove_partial(n, page);
2320 else if (l == M_FULL)
2321 remove_full(s, n, page);
2324 add_partial(n, page, tail);
2325 else if (m == M_FULL)
2326 add_full(s, n, page);
2330 if (!cmpxchg_double_slab(s, page,
2331 old.freelist, old.counters,
2332 new.freelist, new.counters,
2337 spin_unlock_irqrestore(&n->list_lock, flags);
2341 else if (m == M_FULL)
2342 stat(s, DEACTIVATE_FULL);
2343 else if (m == M_FREE) {
2344 stat(s, DEACTIVATE_EMPTY);
2345 discard_slab(s, page);
2350 #ifdef CONFIG_SLUB_CPU_PARTIAL
2351 static void __unfreeze_partials(struct kmem_cache *s, struct page *partial_page)
2353 struct kmem_cache_node *n = NULL, *n2 = NULL;
2354 struct page *page, *discard_page = NULL;
2355 unsigned long flags = 0;
2357 while (partial_page) {
2361 page = partial_page;
2362 partial_page = page->next;
2364 n2 = get_node(s, page_to_nid(page));
2367 spin_unlock_irqrestore(&n->list_lock, flags);
2370 spin_lock_irqsave(&n->list_lock, flags);
2375 old.freelist = page->freelist;
2376 old.counters = page->counters;
2377 VM_BUG_ON(!old.frozen);
2379 new.counters = old.counters;
2380 new.freelist = old.freelist;
2384 } while (!__cmpxchg_double_slab(s, page,
2385 old.freelist, old.counters,
2386 new.freelist, new.counters,
2387 "unfreezing slab"));
2389 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2390 page->next = discard_page;
2391 discard_page = page;
2393 add_partial(n, page, DEACTIVATE_TO_TAIL);
2394 stat(s, FREE_ADD_PARTIAL);
2399 spin_unlock_irqrestore(&n->list_lock, flags);
2401 while (discard_page) {
2402 page = discard_page;
2403 discard_page = discard_page->next;
2405 stat(s, DEACTIVATE_EMPTY);
2406 discard_slab(s, page);
2412 * Unfreeze all the cpu partial slabs.
2414 static void unfreeze_partials(struct kmem_cache *s)
2416 struct page *partial_page;
2417 unsigned long flags;
2419 local_irq_save(flags);
2420 partial_page = this_cpu_read(s->cpu_slab->partial);
2421 this_cpu_write(s->cpu_slab->partial, NULL);
2422 local_irq_restore(flags);
2425 __unfreeze_partials(s, partial_page);
2428 static void unfreeze_partials_cpu(struct kmem_cache *s,
2429 struct kmem_cache_cpu *c)
2431 struct page *partial_page;
2433 partial_page = slub_percpu_partial(c);
2437 __unfreeze_partials(s, partial_page);
2440 #else /* CONFIG_SLUB_CPU_PARTIAL */
2442 static inline void unfreeze_partials(struct kmem_cache *s) { }
2443 static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2444 struct kmem_cache_cpu *c) { }
2446 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2449 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2450 * partial page slot if available.
2452 * If we did not find a slot then simply move all the partials to the
2453 * per node partial list.
2455 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2457 #ifdef CONFIG_SLUB_CPU_PARTIAL
2458 struct page *oldpage;
2466 oldpage = this_cpu_read(s->cpu_slab->partial);
2469 pobjects = oldpage->pobjects;
2470 pages = oldpage->pages;
2471 if (drain && pobjects > slub_cpu_partial(s)) {
2473 * partial array is full. Move the existing
2474 * set to the per node partial list.
2476 unfreeze_partials(s);
2480 stat(s, CPU_PARTIAL_DRAIN);
2485 pobjects += page->objects - page->inuse;
2487 page->pages = pages;
2488 page->pobjects = pobjects;
2489 page->next = oldpage;
2491 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2494 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2497 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2499 unsigned long flags;
2503 local_irq_save(flags);
2506 freelist = c->freelist;
2510 c->tid = next_tid(c->tid);
2512 local_irq_restore(flags);
2515 deactivate_slab(s, page, freelist);
2516 stat(s, CPUSLAB_FLUSH);
2520 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2522 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2523 void *freelist = c->freelist;
2524 struct page *page = c->page;
2528 c->tid = next_tid(c->tid);
2531 deactivate_slab(s, page, freelist);
2532 stat(s, CPUSLAB_FLUSH);
2535 unfreeze_partials_cpu(s, c);
2538 struct slub_flush_work {
2539 struct work_struct work;
2540 struct kmem_cache *s;
2547 * Called from CPU work handler with migration disabled.
2549 static void flush_cpu_slab(struct work_struct *w)
2551 struct kmem_cache *s;
2552 struct kmem_cache_cpu *c;
2553 struct slub_flush_work *sfw;
2555 sfw = container_of(w, struct slub_flush_work, work);
2558 c = this_cpu_ptr(s->cpu_slab);
2563 unfreeze_partials(s);
2566 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2568 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2570 return c->page || slub_percpu_partial(c);
2573 static DEFINE_MUTEX(flush_lock);
2574 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2576 static void flush_all_cpus_locked(struct kmem_cache *s)
2578 struct slub_flush_work *sfw;
2581 lockdep_assert_cpus_held();
2582 mutex_lock(&flush_lock);
2584 for_each_online_cpu(cpu) {
2585 sfw = &per_cpu(slub_flush, cpu);
2586 if (!has_cpu_slab(cpu, s)) {
2590 INIT_WORK(&sfw->work, flush_cpu_slab);
2593 schedule_work_on(cpu, &sfw->work);
2596 for_each_online_cpu(cpu) {
2597 sfw = &per_cpu(slub_flush, cpu);
2600 flush_work(&sfw->work);
2603 mutex_unlock(&flush_lock);
2606 static void flush_all(struct kmem_cache *s)
2609 flush_all_cpus_locked(s);
2614 * Use the cpu notifier to insure that the cpu slabs are flushed when
2617 static int slub_cpu_dead(unsigned int cpu)
2619 struct kmem_cache *s;
2621 mutex_lock(&slab_mutex);
2622 list_for_each_entry(s, &slab_caches, list)
2623 __flush_cpu_slab(s, cpu);
2624 mutex_unlock(&slab_mutex);
2629 * Check if the objects in a per cpu structure fit numa
2630 * locality expectations.
2632 static inline int node_match(struct page *page, int node)
2635 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2641 #ifdef CONFIG_SLUB_DEBUG
2642 static int count_free(struct page *page)
2644 return page->objects - page->inuse;
2647 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2649 return atomic_long_read(&n->total_objects);
2651 #endif /* CONFIG_SLUB_DEBUG */
2653 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2654 static unsigned long count_partial(struct kmem_cache_node *n,
2655 int (*get_count)(struct page *))
2657 unsigned long flags;
2658 unsigned long x = 0;
2661 spin_lock_irqsave(&n->list_lock, flags);
2662 list_for_each_entry(page, &n->partial, slab_list)
2663 x += get_count(page);
2664 spin_unlock_irqrestore(&n->list_lock, flags);
2667 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2669 static noinline void
2670 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2672 #ifdef CONFIG_SLUB_DEBUG
2673 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2674 DEFAULT_RATELIMIT_BURST);
2676 struct kmem_cache_node *n;
2678 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2681 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2682 nid, gfpflags, &gfpflags);
2683 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2684 s->name, s->object_size, s->size, oo_order(s->oo),
2687 if (oo_order(s->min) > get_order(s->object_size))
2688 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2691 for_each_kmem_cache_node(s, node, n) {
2692 unsigned long nr_slabs;
2693 unsigned long nr_objs;
2694 unsigned long nr_free;
2696 nr_free = count_partial(n, count_free);
2697 nr_slabs = node_nr_slabs(n);
2698 nr_objs = node_nr_objs(n);
2700 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2701 node, nr_slabs, nr_objs, nr_free);
2706 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2708 if (unlikely(PageSlabPfmemalloc(page)))
2709 return gfp_pfmemalloc_allowed(gfpflags);
2715 * A variant of pfmemalloc_match() that tests page flags without asserting
2716 * PageSlab. Intended for opportunistic checks before taking a lock and
2717 * rechecking that nobody else freed the page under us.
2719 static inline bool pfmemalloc_match_unsafe(struct page *page, gfp_t gfpflags)
2721 if (unlikely(__PageSlabPfmemalloc(page)))
2722 return gfp_pfmemalloc_allowed(gfpflags);
2728 * Check the page->freelist of a page and either transfer the freelist to the
2729 * per cpu freelist or deactivate the page.
2731 * The page is still frozen if the return value is not NULL.
2733 * If this function returns NULL then the page has been unfrozen.
2735 * This function must be called with interrupt disabled.
2737 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2740 unsigned long counters;
2744 freelist = page->freelist;
2745 counters = page->counters;
2747 new.counters = counters;
2748 VM_BUG_ON(!new.frozen);
2750 new.inuse = page->objects;
2751 new.frozen = freelist != NULL;
2753 } while (!__cmpxchg_double_slab(s, page,
2762 * Slow path. The lockless freelist is empty or we need to perform
2765 * Processing is still very fast if new objects have been freed to the
2766 * regular freelist. In that case we simply take over the regular freelist
2767 * as the lockless freelist and zap the regular freelist.
2769 * If that is not working then we fall back to the partial lists. We take the
2770 * first element of the freelist as the object to allocate now and move the
2771 * rest of the freelist to the lockless freelist.
2773 * And if we were unable to get a new slab from the partial slab lists then
2774 * we need to allocate a new slab. This is the slowest path since it involves
2775 * a call to the page allocator and the setup of a new slab.
2777 * Version of __slab_alloc to use when we know that preemption is
2778 * already disabled (which is the case for bulk allocation).
2780 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2781 unsigned long addr, struct kmem_cache_cpu *c)
2785 unsigned long flags;
2787 stat(s, ALLOC_SLOWPATH);
2791 page = READ_ONCE(c->page);
2794 * if the node is not online or has no normal memory, just
2795 * ignore the node constraint
2797 if (unlikely(node != NUMA_NO_NODE &&
2798 !node_isset(node, slab_nodes)))
2799 node = NUMA_NO_NODE;
2804 if (unlikely(!node_match(page, node))) {
2806 * same as above but node_match() being false already
2807 * implies node != NUMA_NO_NODE
2809 if (!node_isset(node, slab_nodes)) {
2810 node = NUMA_NO_NODE;
2813 stat(s, ALLOC_NODE_MISMATCH);
2814 goto deactivate_slab;
2819 * By rights, we should be searching for a slab page that was
2820 * PFMEMALLOC but right now, we are losing the pfmemalloc
2821 * information when the page leaves the per-cpu allocator
2823 if (unlikely(!pfmemalloc_match_unsafe(page, gfpflags)))
2824 goto deactivate_slab;
2826 /* must check again c->page in case IRQ handler changed it */
2827 local_irq_save(flags);
2828 if (unlikely(page != c->page)) {
2829 local_irq_restore(flags);
2832 freelist = c->freelist;
2836 freelist = get_freelist(s, page);
2840 local_irq_restore(flags);
2841 stat(s, DEACTIVATE_BYPASS);
2845 stat(s, ALLOC_REFILL);
2849 lockdep_assert_irqs_disabled();
2852 * freelist is pointing to the list of objects to be used.
2853 * page is pointing to the page from which the objects are obtained.
2854 * That page must be frozen for per cpu allocations to work.
2856 VM_BUG_ON(!c->page->frozen);
2857 c->freelist = get_freepointer(s, freelist);
2858 c->tid = next_tid(c->tid);
2859 local_irq_restore(flags);
2864 local_irq_save(flags);
2865 if (page != c->page) {
2866 local_irq_restore(flags);
2869 freelist = c->freelist;
2872 local_irq_restore(flags);
2873 deactivate_slab(s, page, freelist);
2877 if (slub_percpu_partial(c)) {
2878 local_irq_save(flags);
2879 if (unlikely(c->page)) {
2880 local_irq_restore(flags);
2883 if (unlikely(!slub_percpu_partial(c))) {
2884 local_irq_restore(flags);
2885 goto new_objects; /* stolen by an IRQ handler */
2888 page = c->page = slub_percpu_partial(c);
2889 slub_set_percpu_partial(c, page);
2890 local_irq_restore(flags);
2891 stat(s, CPU_PARTIAL_ALLOC);
2897 freelist = get_partial(s, gfpflags, node, &page);
2899 goto check_new_page;
2901 put_cpu_ptr(s->cpu_slab);
2902 page = new_slab(s, gfpflags, node);
2903 c = get_cpu_ptr(s->cpu_slab);
2905 if (unlikely(!page)) {
2906 slab_out_of_memory(s, gfpflags, node);
2911 * No other reference to the page yet so we can
2912 * muck around with it freely without cmpxchg
2914 freelist = page->freelist;
2915 page->freelist = NULL;
2917 stat(s, ALLOC_SLAB);
2921 if (kmem_cache_debug(s)) {
2922 if (!alloc_debug_processing(s, page, freelist, addr)) {
2923 /* Slab failed checks. Next slab needed */
2927 * For debug case, we don't load freelist so that all
2928 * allocations go through alloc_debug_processing()
2934 if (unlikely(!pfmemalloc_match(page, gfpflags)))
2936 * For !pfmemalloc_match() case we don't load freelist so that
2937 * we don't make further mismatched allocations easier.
2943 local_irq_save(flags);
2944 if (unlikely(c->page)) {
2945 void *flush_freelist = c->freelist;
2946 struct page *flush_page = c->page;
2950 c->tid = next_tid(c->tid);
2952 local_irq_restore(flags);
2954 deactivate_slab(s, flush_page, flush_freelist);
2956 stat(s, CPUSLAB_FLUSH);
2958 goto retry_load_page;
2966 deactivate_slab(s, page, get_freepointer(s, freelist));
2971 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
2972 * disabled. Compensates for possible cpu changes by refetching the per cpu area
2975 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2976 unsigned long addr, struct kmem_cache_cpu *c)
2980 #ifdef CONFIG_PREEMPT_COUNT
2982 * We may have been preempted and rescheduled on a different
2983 * cpu before disabling preemption. Need to reload cpu area
2986 c = get_cpu_ptr(s->cpu_slab);
2989 p = ___slab_alloc(s, gfpflags, node, addr, c);
2990 #ifdef CONFIG_PREEMPT_COUNT
2991 put_cpu_ptr(s->cpu_slab);
2997 * If the object has been wiped upon free, make sure it's fully initialized by
2998 * zeroing out freelist pointer.
3000 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3003 if (unlikely(slab_want_init_on_free(s)) && obj)
3004 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3009 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3010 * have the fastpath folded into their functions. So no function call
3011 * overhead for requests that can be satisfied on the fastpath.
3013 * The fastpath works by first checking if the lockless freelist can be used.
3014 * If not then __slab_alloc is called for slow processing.
3016 * Otherwise we can simply pick the next object from the lockless free list.
3018 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
3019 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3022 struct kmem_cache_cpu *c;
3025 struct obj_cgroup *objcg = NULL;
3028 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
3032 object = kfence_alloc(s, orig_size, gfpflags);
3033 if (unlikely(object))
3038 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3039 * enabled. We may switch back and forth between cpus while
3040 * reading from one cpu area. That does not matter as long
3041 * as we end up on the original cpu again when doing the cmpxchg.
3043 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3044 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3045 * the tid. If we are preempted and switched to another cpu between the
3046 * two reads, it's OK as the two are still associated with the same cpu
3047 * and cmpxchg later will validate the cpu.
3049 c = raw_cpu_ptr(s->cpu_slab);
3050 tid = READ_ONCE(c->tid);
3053 * Irqless object alloc/free algorithm used here depends on sequence
3054 * of fetching cpu_slab's data. tid should be fetched before anything
3055 * on c to guarantee that object and page associated with previous tid
3056 * won't be used with current tid. If we fetch tid first, object and
3057 * page could be one associated with next tid and our alloc/free
3058 * request will be failed. In this case, we will retry. So, no problem.
3063 * The transaction ids are globally unique per cpu and per operation on
3064 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3065 * occurs on the right processor and that there was no operation on the
3066 * linked list in between.
3069 object = c->freelist;
3071 if (unlikely(!object || !page || !node_match(page, node))) {
3072 object = __slab_alloc(s, gfpflags, node, addr, c);
3074 void *next_object = get_freepointer_safe(s, object);
3077 * The cmpxchg will only match if there was no additional
3078 * operation and if we are on the right processor.
3080 * The cmpxchg does the following atomically (without lock
3082 * 1. Relocate first pointer to the current per cpu area.
3083 * 2. Verify that tid and freelist have not been changed
3084 * 3. If they were not changed replace tid and freelist
3086 * Since this is without lock semantics the protection is only
3087 * against code executing on this cpu *not* from access by
3090 if (unlikely(!this_cpu_cmpxchg_double(
3091 s->cpu_slab->freelist, s->cpu_slab->tid,
3093 next_object, next_tid(tid)))) {
3095 note_cmpxchg_failure("slab_alloc", s, tid);
3098 prefetch_freepointer(s, next_object);
3099 stat(s, ALLOC_FASTPATH);
3102 maybe_wipe_obj_freeptr(s, object);
3103 init = slab_want_init_on_alloc(gfpflags, s);
3106 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
3111 static __always_inline void *slab_alloc(struct kmem_cache *s,
3112 gfp_t gfpflags, unsigned long addr, size_t orig_size)
3114 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size);
3117 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3119 void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size);
3121 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
3126 EXPORT_SYMBOL(kmem_cache_alloc);
3128 #ifdef CONFIG_TRACING
3129 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3131 void *ret = slab_alloc(s, gfpflags, _RET_IP_, size);
3132 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
3133 ret = kasan_kmalloc(s, ret, size, gfpflags);
3136 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3140 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3142 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size);
3144 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3145 s->object_size, s->size, gfpflags, node);
3149 EXPORT_SYMBOL(kmem_cache_alloc_node);
3151 #ifdef CONFIG_TRACING
3152 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3154 int node, size_t size)
3156 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size);
3158 trace_kmalloc_node(_RET_IP_, ret,
3159 size, s->size, gfpflags, node);
3161 ret = kasan_kmalloc(s, ret, size, gfpflags);
3164 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3166 #endif /* CONFIG_NUMA */
3169 * Slow path handling. This may still be called frequently since objects
3170 * have a longer lifetime than the cpu slabs in most processing loads.
3172 * So we still attempt to reduce cache line usage. Just take the slab
3173 * lock and free the item. If there is no additional partial page
3174 * handling required then we can return immediately.
3176 static void __slab_free(struct kmem_cache *s, struct page *page,
3177 void *head, void *tail, int cnt,
3184 unsigned long counters;
3185 struct kmem_cache_node *n = NULL;
3186 unsigned long flags;
3188 stat(s, FREE_SLOWPATH);
3190 if (kfence_free(head))
3193 if (kmem_cache_debug(s) &&
3194 !free_debug_processing(s, page, head, tail, cnt, addr))
3199 spin_unlock_irqrestore(&n->list_lock, flags);
3202 prior = page->freelist;
3203 counters = page->counters;
3204 set_freepointer(s, tail, prior);
3205 new.counters = counters;
3206 was_frozen = new.frozen;
3208 if ((!new.inuse || !prior) && !was_frozen) {
3210 if (kmem_cache_has_cpu_partial(s) && !prior) {
3213 * Slab was on no list before and will be
3215 * We can defer the list move and instead
3220 } else { /* Needs to be taken off a list */
3222 n = get_node(s, page_to_nid(page));
3224 * Speculatively acquire the list_lock.
3225 * If the cmpxchg does not succeed then we may
3226 * drop the list_lock without any processing.
3228 * Otherwise the list_lock will synchronize with
3229 * other processors updating the list of slabs.
3231 spin_lock_irqsave(&n->list_lock, flags);
3236 } while (!cmpxchg_double_slab(s, page,
3243 if (likely(was_frozen)) {
3245 * The list lock was not taken therefore no list
3246 * activity can be necessary.
3248 stat(s, FREE_FROZEN);
3249 } else if (new.frozen) {
3251 * If we just froze the page then put it onto the
3252 * per cpu partial list.
3254 put_cpu_partial(s, page, 1);
3255 stat(s, CPU_PARTIAL_FREE);
3261 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3265 * Objects left in the slab. If it was not on the partial list before
3268 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3269 remove_full(s, n, page);
3270 add_partial(n, page, DEACTIVATE_TO_TAIL);
3271 stat(s, FREE_ADD_PARTIAL);
3273 spin_unlock_irqrestore(&n->list_lock, flags);
3279 * Slab on the partial list.
3281 remove_partial(n, page);
3282 stat(s, FREE_REMOVE_PARTIAL);
3284 /* Slab must be on the full list */
3285 remove_full(s, n, page);
3288 spin_unlock_irqrestore(&n->list_lock, flags);
3290 discard_slab(s, page);
3294 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3295 * can perform fastpath freeing without additional function calls.
3297 * The fastpath is only possible if we are freeing to the current cpu slab
3298 * of this processor. This typically the case if we have just allocated
3301 * If fastpath is not possible then fall back to __slab_free where we deal
3302 * with all sorts of special processing.
3304 * Bulk free of a freelist with several objects (all pointing to the
3305 * same page) possible by specifying head and tail ptr, plus objects
3306 * count (cnt). Bulk free indicated by tail pointer being set.
3308 static __always_inline void do_slab_free(struct kmem_cache *s,
3309 struct page *page, void *head, void *tail,
3310 int cnt, unsigned long addr)
3312 void *tail_obj = tail ? : head;
3313 struct kmem_cache_cpu *c;
3316 memcg_slab_free_hook(s, &head, 1);
3319 * Determine the currently cpus per cpu slab.
3320 * The cpu may change afterward. However that does not matter since
3321 * data is retrieved via this pointer. If we are on the same cpu
3322 * during the cmpxchg then the free will succeed.
3324 c = raw_cpu_ptr(s->cpu_slab);
3325 tid = READ_ONCE(c->tid);
3327 /* Same with comment on barrier() in slab_alloc_node() */
3330 if (likely(page == c->page)) {
3331 void **freelist = READ_ONCE(c->freelist);
3333 set_freepointer(s, tail_obj, freelist);
3335 if (unlikely(!this_cpu_cmpxchg_double(
3336 s->cpu_slab->freelist, s->cpu_slab->tid,
3338 head, next_tid(tid)))) {
3340 note_cmpxchg_failure("slab_free", s, tid);
3343 stat(s, FREE_FASTPATH);
3345 __slab_free(s, page, head, tail_obj, cnt, addr);
3349 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3350 void *head, void *tail, int cnt,
3354 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3355 * to remove objects, whose reuse must be delayed.
3357 if (slab_free_freelist_hook(s, &head, &tail))
3358 do_slab_free(s, page, head, tail, cnt, addr);
3361 #ifdef CONFIG_KASAN_GENERIC
3362 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3364 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3368 void kmem_cache_free(struct kmem_cache *s, void *x)
3370 s = cache_from_obj(s, x);
3373 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3374 trace_kmem_cache_free(_RET_IP_, x, s->name);
3376 EXPORT_SYMBOL(kmem_cache_free);
3378 struct detached_freelist {
3383 struct kmem_cache *s;
3386 static inline void free_nonslab_page(struct page *page, void *object)
3388 unsigned int order = compound_order(page);
3390 VM_BUG_ON_PAGE(!PageCompound(page), page);
3392 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B, -(PAGE_SIZE << order));
3393 __free_pages(page, order);
3397 * This function progressively scans the array with free objects (with
3398 * a limited look ahead) and extract objects belonging to the same
3399 * page. It builds a detached freelist directly within the given
3400 * page/objects. This can happen without any need for
3401 * synchronization, because the objects are owned by running process.
3402 * The freelist is build up as a single linked list in the objects.
3403 * The idea is, that this detached freelist can then be bulk
3404 * transferred to the real freelist(s), but only requiring a single
3405 * synchronization primitive. Look ahead in the array is limited due
3406 * to performance reasons.
3409 int build_detached_freelist(struct kmem_cache *s, size_t size,
3410 void **p, struct detached_freelist *df)
3412 size_t first_skipped_index = 0;
3417 /* Always re-init detached_freelist */
3422 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3423 } while (!object && size);
3428 page = virt_to_head_page(object);
3430 /* Handle kalloc'ed objects */
3431 if (unlikely(!PageSlab(page))) {
3432 free_nonslab_page(page, object);
3433 p[size] = NULL; /* mark object processed */
3436 /* Derive kmem_cache from object */
3437 df->s = page->slab_cache;
3439 df->s = cache_from_obj(s, object); /* Support for memcg */
3442 if (is_kfence_address(object)) {
3443 slab_free_hook(df->s, object, false);
3444 __kfence_free(object);
3445 p[size] = NULL; /* mark object processed */
3449 /* Start new detached freelist */
3451 set_freepointer(df->s, object, NULL);
3453 df->freelist = object;
3454 p[size] = NULL; /* mark object processed */
3460 continue; /* Skip processed objects */
3462 /* df->page is always set at this point */
3463 if (df->page == virt_to_head_page(object)) {
3464 /* Opportunity build freelist */
3465 set_freepointer(df->s, object, df->freelist);
3466 df->freelist = object;
3468 p[size] = NULL; /* mark object processed */
3473 /* Limit look ahead search */
3477 if (!first_skipped_index)
3478 first_skipped_index = size + 1;
3481 return first_skipped_index;
3484 /* Note that interrupts must be enabled when calling this function. */
3485 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3490 memcg_slab_free_hook(s, p, size);
3492 struct detached_freelist df;
3494 size = build_detached_freelist(s, size, p, &df);
3498 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
3499 } while (likely(size));
3501 EXPORT_SYMBOL(kmem_cache_free_bulk);
3503 /* Note that interrupts must be enabled when calling this function. */
3504 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3507 struct kmem_cache_cpu *c;
3509 struct obj_cgroup *objcg = NULL;
3511 /* memcg and kmem_cache debug support */
3512 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3516 * Drain objects in the per cpu slab, while disabling local
3517 * IRQs, which protects against PREEMPT and interrupts
3518 * handlers invoking normal fastpath.
3520 c = get_cpu_ptr(s->cpu_slab);
3521 local_irq_disable();
3523 for (i = 0; i < size; i++) {
3524 void *object = kfence_alloc(s, s->object_size, flags);
3526 if (unlikely(object)) {
3531 object = c->freelist;
3532 if (unlikely(!object)) {
3534 * We may have removed an object from c->freelist using
3535 * the fastpath in the previous iteration; in that case,
3536 * c->tid has not been bumped yet.
3537 * Since ___slab_alloc() may reenable interrupts while
3538 * allocating memory, we should bump c->tid now.
3540 c->tid = next_tid(c->tid);
3545 * Invoking slow path likely have side-effect
3546 * of re-populating per CPU c->freelist
3548 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3550 if (unlikely(!p[i]))
3553 c = this_cpu_ptr(s->cpu_slab);
3554 maybe_wipe_obj_freeptr(s, p[i]);
3556 local_irq_disable();
3558 continue; /* goto for-loop */
3560 c->freelist = get_freepointer(s, object);
3562 maybe_wipe_obj_freeptr(s, p[i]);
3564 c->tid = next_tid(c->tid);
3566 put_cpu_ptr(s->cpu_slab);
3569 * memcg and kmem_cache debug support and memory initialization.
3570 * Done outside of the IRQ disabled fastpath loop.
3572 slab_post_alloc_hook(s, objcg, flags, size, p,
3573 slab_want_init_on_alloc(flags, s));
3576 put_cpu_ptr(s->cpu_slab);
3577 slab_post_alloc_hook(s, objcg, flags, i, p, false);
3578 __kmem_cache_free_bulk(s, i, p);
3581 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3585 * Object placement in a slab is made very easy because we always start at
3586 * offset 0. If we tune the size of the object to the alignment then we can
3587 * get the required alignment by putting one properly sized object after
3590 * Notice that the allocation order determines the sizes of the per cpu
3591 * caches. Each processor has always one slab available for allocations.
3592 * Increasing the allocation order reduces the number of times that slabs
3593 * must be moved on and off the partial lists and is therefore a factor in
3598 * Minimum / Maximum order of slab pages. This influences locking overhead
3599 * and slab fragmentation. A higher order reduces the number of partial slabs
3600 * and increases the number of allocations possible without having to
3601 * take the list_lock.
3603 static unsigned int slub_min_order;
3604 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3605 static unsigned int slub_min_objects;
3608 * Calculate the order of allocation given an slab object size.
3610 * The order of allocation has significant impact on performance and other
3611 * system components. Generally order 0 allocations should be preferred since
3612 * order 0 does not cause fragmentation in the page allocator. Larger objects
3613 * be problematic to put into order 0 slabs because there may be too much
3614 * unused space left. We go to a higher order if more than 1/16th of the slab
3617 * In order to reach satisfactory performance we must ensure that a minimum
3618 * number of objects is in one slab. Otherwise we may generate too much
3619 * activity on the partial lists which requires taking the list_lock. This is
3620 * less a concern for large slabs though which are rarely used.
3622 * slub_max_order specifies the order where we begin to stop considering the
3623 * number of objects in a slab as critical. If we reach slub_max_order then
3624 * we try to keep the page order as low as possible. So we accept more waste
3625 * of space in favor of a small page order.
3627 * Higher order allocations also allow the placement of more objects in a
3628 * slab and thereby reduce object handling overhead. If the user has
3629 * requested a higher minimum order then we start with that one instead of
3630 * the smallest order which will fit the object.
3632 static inline unsigned int slab_order(unsigned int size,
3633 unsigned int min_objects, unsigned int max_order,
3634 unsigned int fract_leftover)
3636 unsigned int min_order = slub_min_order;
3639 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3640 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3642 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3643 order <= max_order; order++) {
3645 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3648 rem = slab_size % size;
3650 if (rem <= slab_size / fract_leftover)
3657 static inline int calculate_order(unsigned int size)
3660 unsigned int min_objects;
3661 unsigned int max_objects;
3662 unsigned int nr_cpus;
3665 * Attempt to find best configuration for a slab. This
3666 * works by first attempting to generate a layout with
3667 * the best configuration and backing off gradually.
3669 * First we increase the acceptable waste in a slab. Then
3670 * we reduce the minimum objects required in a slab.
3672 min_objects = slub_min_objects;
3675 * Some architectures will only update present cpus when
3676 * onlining them, so don't trust the number if it's just 1. But
3677 * we also don't want to use nr_cpu_ids always, as on some other
3678 * architectures, there can be many possible cpus, but never
3679 * onlined. Here we compromise between trying to avoid too high
3680 * order on systems that appear larger than they are, and too
3681 * low order on systems that appear smaller than they are.
3683 nr_cpus = num_present_cpus();
3685 nr_cpus = nr_cpu_ids;
3686 min_objects = 4 * (fls(nr_cpus) + 1);
3688 max_objects = order_objects(slub_max_order, size);
3689 min_objects = min(min_objects, max_objects);
3691 while (min_objects > 1) {
3692 unsigned int fraction;
3695 while (fraction >= 4) {
3696 order = slab_order(size, min_objects,
3697 slub_max_order, fraction);
3698 if (order <= slub_max_order)
3706 * We were unable to place multiple objects in a slab. Now
3707 * lets see if we can place a single object there.
3709 order = slab_order(size, 1, slub_max_order, 1);
3710 if (order <= slub_max_order)
3714 * Doh this slab cannot be placed using slub_max_order.
3716 order = slab_order(size, 1, MAX_ORDER, 1);
3717 if (order < MAX_ORDER)
3723 init_kmem_cache_node(struct kmem_cache_node *n)
3726 spin_lock_init(&n->list_lock);
3727 INIT_LIST_HEAD(&n->partial);
3728 #ifdef CONFIG_SLUB_DEBUG
3729 atomic_long_set(&n->nr_slabs, 0);
3730 atomic_long_set(&n->total_objects, 0);
3731 INIT_LIST_HEAD(&n->full);
3735 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3737 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3738 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3741 * Must align to double word boundary for the double cmpxchg
3742 * instructions to work; see __pcpu_double_call_return_bool().
3744 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3745 2 * sizeof(void *));
3750 init_kmem_cache_cpus(s);
3755 static struct kmem_cache *kmem_cache_node;
3758 * No kmalloc_node yet so do it by hand. We know that this is the first
3759 * slab on the node for this slabcache. There are no concurrent accesses
3762 * Note that this function only works on the kmem_cache_node
3763 * when allocating for the kmem_cache_node. This is used for bootstrapping
3764 * memory on a fresh node that has no slab structures yet.
3766 static void early_kmem_cache_node_alloc(int node)
3769 struct kmem_cache_node *n;
3771 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3773 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3776 if (page_to_nid(page) != node) {
3777 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3778 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3783 #ifdef CONFIG_SLUB_DEBUG
3784 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3785 init_tracking(kmem_cache_node, n);
3787 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3788 page->freelist = get_freepointer(kmem_cache_node, n);
3791 kmem_cache_node->node[node] = n;
3792 init_kmem_cache_node(n);
3793 inc_slabs_node(kmem_cache_node, node, page->objects);
3796 * No locks need to be taken here as it has just been
3797 * initialized and there is no concurrent access.
3799 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3802 static void free_kmem_cache_nodes(struct kmem_cache *s)
3805 struct kmem_cache_node *n;
3807 for_each_kmem_cache_node(s, node, n) {
3808 s->node[node] = NULL;
3809 kmem_cache_free(kmem_cache_node, n);
3813 void __kmem_cache_release(struct kmem_cache *s)
3815 cache_random_seq_destroy(s);
3816 free_percpu(s->cpu_slab);
3817 free_kmem_cache_nodes(s);
3820 static int init_kmem_cache_nodes(struct kmem_cache *s)
3824 for_each_node_mask(node, slab_nodes) {
3825 struct kmem_cache_node *n;
3827 if (slab_state == DOWN) {
3828 early_kmem_cache_node_alloc(node);
3831 n = kmem_cache_alloc_node(kmem_cache_node,
3835 free_kmem_cache_nodes(s);
3839 init_kmem_cache_node(n);
3845 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3847 if (min < MIN_PARTIAL)
3849 else if (min > MAX_PARTIAL)
3851 s->min_partial = min;
3854 static void set_cpu_partial(struct kmem_cache *s)
3856 #ifdef CONFIG_SLUB_CPU_PARTIAL
3858 * cpu_partial determined the maximum number of objects kept in the
3859 * per cpu partial lists of a processor.
3861 * Per cpu partial lists mainly contain slabs that just have one
3862 * object freed. If they are used for allocation then they can be
3863 * filled up again with minimal effort. The slab will never hit the
3864 * per node partial lists and therefore no locking will be required.
3866 * This setting also determines
3868 * A) The number of objects from per cpu partial slabs dumped to the
3869 * per node list when we reach the limit.
3870 * B) The number of objects in cpu partial slabs to extract from the
3871 * per node list when we run out of per cpu objects. We only fetch
3872 * 50% to keep some capacity around for frees.
3874 if (!kmem_cache_has_cpu_partial(s))
3875 slub_set_cpu_partial(s, 0);
3876 else if (s->size >= PAGE_SIZE)
3877 slub_set_cpu_partial(s, 2);
3878 else if (s->size >= 1024)
3879 slub_set_cpu_partial(s, 6);
3880 else if (s->size >= 256)
3881 slub_set_cpu_partial(s, 13);
3883 slub_set_cpu_partial(s, 30);
3888 * calculate_sizes() determines the order and the distribution of data within
3891 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3893 slab_flags_t flags = s->flags;
3894 unsigned int size = s->object_size;
3898 * Round up object size to the next word boundary. We can only
3899 * place the free pointer at word boundaries and this determines
3900 * the possible location of the free pointer.
3902 size = ALIGN(size, sizeof(void *));
3904 #ifdef CONFIG_SLUB_DEBUG
3906 * Determine if we can poison the object itself. If the user of
3907 * the slab may touch the object after free or before allocation
3908 * then we should never poison the object itself.
3910 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3912 s->flags |= __OBJECT_POISON;
3914 s->flags &= ~__OBJECT_POISON;
3918 * If we are Redzoning then check if there is some space between the
3919 * end of the object and the free pointer. If not then add an
3920 * additional word to have some bytes to store Redzone information.
3922 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3923 size += sizeof(void *);
3927 * With that we have determined the number of bytes in actual use
3928 * by the object and redzoning.
3932 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3933 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
3936 * Relocate free pointer after the object if it is not
3937 * permitted to overwrite the first word of the object on
3940 * This is the case if we do RCU, have a constructor or
3941 * destructor, are poisoning the objects, or are
3942 * redzoning an object smaller than sizeof(void *).
3944 * The assumption that s->offset >= s->inuse means free
3945 * pointer is outside of the object is used in the
3946 * freeptr_outside_object() function. If that is no
3947 * longer true, the function needs to be modified.
3950 size += sizeof(void *);
3953 * Store freelist pointer near middle of object to keep
3954 * it away from the edges of the object to avoid small
3955 * sized over/underflows from neighboring allocations.
3957 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
3960 #ifdef CONFIG_SLUB_DEBUG
3961 if (flags & SLAB_STORE_USER)
3963 * Need to store information about allocs and frees after
3966 size += 2 * sizeof(struct track);
3969 kasan_cache_create(s, &size, &s->flags);
3970 #ifdef CONFIG_SLUB_DEBUG
3971 if (flags & SLAB_RED_ZONE) {
3973 * Add some empty padding so that we can catch
3974 * overwrites from earlier objects rather than let
3975 * tracking information or the free pointer be
3976 * corrupted if a user writes before the start
3979 size += sizeof(void *);
3981 s->red_left_pad = sizeof(void *);
3982 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3983 size += s->red_left_pad;
3988 * SLUB stores one object immediately after another beginning from
3989 * offset 0. In order to align the objects we have to simply size
3990 * each object to conform to the alignment.
3992 size = ALIGN(size, s->align);
3994 s->reciprocal_size = reciprocal_value(size);
3995 if (forced_order >= 0)
3996 order = forced_order;
3998 order = calculate_order(size);
4005 s->allocflags |= __GFP_COMP;
4007 if (s->flags & SLAB_CACHE_DMA)
4008 s->allocflags |= GFP_DMA;
4010 if (s->flags & SLAB_CACHE_DMA32)
4011 s->allocflags |= GFP_DMA32;
4013 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4014 s->allocflags |= __GFP_RECLAIMABLE;
4017 * Determine the number of objects per slab
4019 s->oo = oo_make(order, size);
4020 s->min = oo_make(get_order(size), size);
4021 if (oo_objects(s->oo) > oo_objects(s->max))
4024 return !!oo_objects(s->oo);
4027 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4029 s->flags = kmem_cache_flags(s->size, flags, s->name);
4030 #ifdef CONFIG_SLAB_FREELIST_HARDENED
4031 s->random = get_random_long();
4034 if (!calculate_sizes(s, -1))
4036 if (disable_higher_order_debug) {
4038 * Disable debugging flags that store metadata if the min slab
4041 if (get_order(s->size) > get_order(s->object_size)) {
4042 s->flags &= ~DEBUG_METADATA_FLAGS;
4044 if (!calculate_sizes(s, -1))
4049 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4050 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4051 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
4052 /* Enable fast mode */
4053 s->flags |= __CMPXCHG_DOUBLE;
4057 * The larger the object size is, the more pages we want on the partial
4058 * list to avoid pounding the page allocator excessively.
4060 set_min_partial(s, ilog2(s->size) / 2);
4065 s->remote_node_defrag_ratio = 1000;
4068 /* Initialize the pre-computed randomized freelist if slab is up */
4069 if (slab_state >= UP) {
4070 if (init_cache_random_seq(s))
4074 if (!init_kmem_cache_nodes(s))
4077 if (alloc_kmem_cache_cpus(s))
4080 free_kmem_cache_nodes(s);
4085 static void list_slab_objects(struct kmem_cache *s, struct page *page,
4088 #ifdef CONFIG_SLUB_DEBUG
4089 void *addr = page_address(page);
4093 slab_err(s, page, text, s->name);
4096 map = get_map(s, page);
4097 for_each_object(p, s, addr, page->objects) {
4099 if (!test_bit(__obj_to_index(s, addr, p), map)) {
4100 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4101 print_tracking(s, p);
4110 * Attempt to free all partial slabs on a node.
4111 * This is called from __kmem_cache_shutdown(). We must take list_lock
4112 * because sysfs file might still access partial list after the shutdowning.
4114 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4117 struct page *page, *h;
4119 BUG_ON(irqs_disabled());
4120 spin_lock_irq(&n->list_lock);
4121 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
4123 remove_partial(n, page);
4124 list_add(&page->slab_list, &discard);
4126 list_slab_objects(s, page,
4127 "Objects remaining in %s on __kmem_cache_shutdown()");
4130 spin_unlock_irq(&n->list_lock);
4132 list_for_each_entry_safe(page, h, &discard, slab_list)
4133 discard_slab(s, page);
4136 bool __kmem_cache_empty(struct kmem_cache *s)
4139 struct kmem_cache_node *n;
4141 for_each_kmem_cache_node(s, node, n)
4142 if (n->nr_partial || slabs_node(s, node))
4148 * Release all resources used by a slab cache.
4150 int __kmem_cache_shutdown(struct kmem_cache *s)
4153 struct kmem_cache_node *n;
4155 flush_all_cpus_locked(s);
4156 /* Attempt to free all objects */
4157 for_each_kmem_cache_node(s, node, n) {
4159 if (n->nr_partial || slabs_node(s, node))
4165 #ifdef CONFIG_PRINTK
4166 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
4169 int __maybe_unused i;
4173 struct kmem_cache *s = page->slab_cache;
4174 struct track __maybe_unused *trackp;
4176 kpp->kp_ptr = object;
4177 kpp->kp_page = page;
4178 kpp->kp_slab_cache = s;
4179 base = page_address(page);
4180 objp0 = kasan_reset_tag(object);
4181 #ifdef CONFIG_SLUB_DEBUG
4182 objp = restore_red_left(s, objp0);
4186 objnr = obj_to_index(s, page, objp);
4187 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4188 objp = base + s->size * objnr;
4189 kpp->kp_objp = objp;
4190 if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) ||
4191 !(s->flags & SLAB_STORE_USER))
4193 #ifdef CONFIG_SLUB_DEBUG
4194 objp = fixup_red_left(s, objp);
4195 trackp = get_track(s, objp, TRACK_ALLOC);
4196 kpp->kp_ret = (void *)trackp->addr;
4197 #ifdef CONFIG_STACKTRACE
4198 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4199 kpp->kp_stack[i] = (void *)trackp->addrs[i];
4200 if (!kpp->kp_stack[i])
4204 trackp = get_track(s, objp, TRACK_FREE);
4205 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4206 kpp->kp_free_stack[i] = (void *)trackp->addrs[i];
4207 if (!kpp->kp_free_stack[i])
4215 /********************************************************************
4217 *******************************************************************/
4219 static int __init setup_slub_min_order(char *str)
4221 get_option(&str, (int *)&slub_min_order);
4226 __setup("slub_min_order=", setup_slub_min_order);
4228 static int __init setup_slub_max_order(char *str)
4230 get_option(&str, (int *)&slub_max_order);
4231 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4236 __setup("slub_max_order=", setup_slub_max_order);
4238 static int __init setup_slub_min_objects(char *str)
4240 get_option(&str, (int *)&slub_min_objects);
4245 __setup("slub_min_objects=", setup_slub_min_objects);
4247 void *__kmalloc(size_t size, gfp_t flags)
4249 struct kmem_cache *s;
4252 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4253 return kmalloc_large(size, flags);
4255 s = kmalloc_slab(size, flags);
4257 if (unlikely(ZERO_OR_NULL_PTR(s)))
4260 ret = slab_alloc(s, flags, _RET_IP_, size);
4262 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
4264 ret = kasan_kmalloc(s, ret, size, flags);
4268 EXPORT_SYMBOL(__kmalloc);
4271 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4275 unsigned int order = get_order(size);
4277 flags |= __GFP_COMP;
4278 page = alloc_pages_node(node, flags, order);
4280 ptr = page_address(page);
4281 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4282 PAGE_SIZE << order);
4285 return kmalloc_large_node_hook(ptr, size, flags);
4288 void *__kmalloc_node(size_t size, gfp_t flags, int node)
4290 struct kmem_cache *s;
4293 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4294 ret = kmalloc_large_node(size, flags, node);
4296 trace_kmalloc_node(_RET_IP_, ret,
4297 size, PAGE_SIZE << get_order(size),
4303 s = kmalloc_slab(size, flags);
4305 if (unlikely(ZERO_OR_NULL_PTR(s)))
4308 ret = slab_alloc_node(s, flags, node, _RET_IP_, size);
4310 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4312 ret = kasan_kmalloc(s, ret, size, flags);
4316 EXPORT_SYMBOL(__kmalloc_node);
4317 #endif /* CONFIG_NUMA */
4319 #ifdef CONFIG_HARDENED_USERCOPY
4321 * Rejects incorrectly sized objects and objects that are to be copied
4322 * to/from userspace but do not fall entirely within the containing slab
4323 * cache's usercopy region.
4325 * Returns NULL if check passes, otherwise const char * to name of cache
4326 * to indicate an error.
4328 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4331 struct kmem_cache *s;
4332 unsigned int offset;
4334 bool is_kfence = is_kfence_address(ptr);
4336 ptr = kasan_reset_tag(ptr);
4338 /* Find object and usable object size. */
4339 s = page->slab_cache;
4341 /* Reject impossible pointers. */
4342 if (ptr < page_address(page))
4343 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4346 /* Find offset within object. */
4348 offset = ptr - kfence_object_start(ptr);
4350 offset = (ptr - page_address(page)) % s->size;
4352 /* Adjust for redzone and reject if within the redzone. */
4353 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4354 if (offset < s->red_left_pad)
4355 usercopy_abort("SLUB object in left red zone",
4356 s->name, to_user, offset, n);
4357 offset -= s->red_left_pad;
4360 /* Allow address range falling entirely within usercopy region. */
4361 if (offset >= s->useroffset &&
4362 offset - s->useroffset <= s->usersize &&
4363 n <= s->useroffset - offset + s->usersize)
4367 * If the copy is still within the allocated object, produce
4368 * a warning instead of rejecting the copy. This is intended
4369 * to be a temporary method to find any missing usercopy
4372 object_size = slab_ksize(s);
4373 if (usercopy_fallback &&
4374 offset <= object_size && n <= object_size - offset) {
4375 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4379 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4381 #endif /* CONFIG_HARDENED_USERCOPY */
4383 size_t __ksize(const void *object)
4387 if (unlikely(object == ZERO_SIZE_PTR))
4390 page = virt_to_head_page(object);
4392 if (unlikely(!PageSlab(page))) {
4393 WARN_ON(!PageCompound(page));
4394 return page_size(page);
4397 return slab_ksize(page->slab_cache);
4399 EXPORT_SYMBOL(__ksize);
4401 void kfree(const void *x)
4404 void *object = (void *)x;
4406 trace_kfree(_RET_IP_, x);
4408 if (unlikely(ZERO_OR_NULL_PTR(x)))
4411 page = virt_to_head_page(x);
4412 if (unlikely(!PageSlab(page))) {
4413 free_nonslab_page(page, object);
4416 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4418 EXPORT_SYMBOL(kfree);
4420 #define SHRINK_PROMOTE_MAX 32
4423 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4424 * up most to the head of the partial lists. New allocations will then
4425 * fill those up and thus they can be removed from the partial lists.
4427 * The slabs with the least items are placed last. This results in them
4428 * being allocated from last increasing the chance that the last objects
4429 * are freed in them.
4431 static int __kmem_cache_do_shrink(struct kmem_cache *s)
4435 struct kmem_cache_node *n;
4438 struct list_head discard;
4439 struct list_head promote[SHRINK_PROMOTE_MAX];
4440 unsigned long flags;
4443 for_each_kmem_cache_node(s, node, n) {
4444 INIT_LIST_HEAD(&discard);
4445 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4446 INIT_LIST_HEAD(promote + i);
4448 spin_lock_irqsave(&n->list_lock, flags);
4451 * Build lists of slabs to discard or promote.
4453 * Note that concurrent frees may occur while we hold the
4454 * list_lock. page->inuse here is the upper limit.
4456 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4457 int free = page->objects - page->inuse;
4459 /* Do not reread page->inuse */
4462 /* We do not keep full slabs on the list */
4465 if (free == page->objects) {
4466 list_move(&page->slab_list, &discard);
4468 } else if (free <= SHRINK_PROMOTE_MAX)
4469 list_move(&page->slab_list, promote + free - 1);
4473 * Promote the slabs filled up most to the head of the
4476 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4477 list_splice(promote + i, &n->partial);
4479 spin_unlock_irqrestore(&n->list_lock, flags);
4481 /* Release empty slabs */
4482 list_for_each_entry_safe(page, t, &discard, slab_list)
4483 discard_slab(s, page);
4485 if (slabs_node(s, node))
4492 int __kmem_cache_shrink(struct kmem_cache *s)
4495 return __kmem_cache_do_shrink(s);
4498 static int slab_mem_going_offline_callback(void *arg)
4500 struct kmem_cache *s;
4502 mutex_lock(&slab_mutex);
4503 list_for_each_entry(s, &slab_caches, list) {
4504 flush_all_cpus_locked(s);
4505 __kmem_cache_do_shrink(s);
4507 mutex_unlock(&slab_mutex);
4512 static void slab_mem_offline_callback(void *arg)
4514 struct memory_notify *marg = arg;
4517 offline_node = marg->status_change_nid_normal;
4520 * If the node still has available memory. we need kmem_cache_node
4523 if (offline_node < 0)
4526 mutex_lock(&slab_mutex);
4527 node_clear(offline_node, slab_nodes);
4529 * We no longer free kmem_cache_node structures here, as it would be
4530 * racy with all get_node() users, and infeasible to protect them with
4533 mutex_unlock(&slab_mutex);
4536 static int slab_mem_going_online_callback(void *arg)
4538 struct kmem_cache_node *n;
4539 struct kmem_cache *s;
4540 struct memory_notify *marg = arg;
4541 int nid = marg->status_change_nid_normal;
4545 * If the node's memory is already available, then kmem_cache_node is
4546 * already created. Nothing to do.
4552 * We are bringing a node online. No memory is available yet. We must
4553 * allocate a kmem_cache_node structure in order to bring the node
4556 mutex_lock(&slab_mutex);
4557 list_for_each_entry(s, &slab_caches, list) {
4559 * The structure may already exist if the node was previously
4560 * onlined and offlined.
4562 if (get_node(s, nid))
4565 * XXX: kmem_cache_alloc_node will fallback to other nodes
4566 * since memory is not yet available from the node that
4569 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4574 init_kmem_cache_node(n);
4578 * Any cache created after this point will also have kmem_cache_node
4579 * initialized for the new node.
4581 node_set(nid, slab_nodes);
4583 mutex_unlock(&slab_mutex);
4587 static int slab_memory_callback(struct notifier_block *self,
4588 unsigned long action, void *arg)
4593 case MEM_GOING_ONLINE:
4594 ret = slab_mem_going_online_callback(arg);
4596 case MEM_GOING_OFFLINE:
4597 ret = slab_mem_going_offline_callback(arg);
4600 case MEM_CANCEL_ONLINE:
4601 slab_mem_offline_callback(arg);
4604 case MEM_CANCEL_OFFLINE:
4608 ret = notifier_from_errno(ret);
4614 static struct notifier_block slab_memory_callback_nb = {
4615 .notifier_call = slab_memory_callback,
4616 .priority = SLAB_CALLBACK_PRI,
4619 /********************************************************************
4620 * Basic setup of slabs
4621 *******************************************************************/
4624 * Used for early kmem_cache structures that were allocated using
4625 * the page allocator. Allocate them properly then fix up the pointers
4626 * that may be pointing to the wrong kmem_cache structure.
4629 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4632 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4633 struct kmem_cache_node *n;
4635 memcpy(s, static_cache, kmem_cache->object_size);
4638 * This runs very early, and only the boot processor is supposed to be
4639 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4642 __flush_cpu_slab(s, smp_processor_id());
4643 for_each_kmem_cache_node(s, node, n) {
4646 list_for_each_entry(p, &n->partial, slab_list)
4649 #ifdef CONFIG_SLUB_DEBUG
4650 list_for_each_entry(p, &n->full, slab_list)
4654 list_add(&s->list, &slab_caches);
4658 void __init kmem_cache_init(void)
4660 static __initdata struct kmem_cache boot_kmem_cache,
4661 boot_kmem_cache_node;
4664 if (debug_guardpage_minorder())
4667 /* Print slub debugging pointers without hashing */
4668 if (__slub_debug_enabled())
4669 no_hash_pointers_enable(NULL);
4671 kmem_cache_node = &boot_kmem_cache_node;
4672 kmem_cache = &boot_kmem_cache;
4675 * Initialize the nodemask for which we will allocate per node
4676 * structures. Here we don't need taking slab_mutex yet.
4678 for_each_node_state(node, N_NORMAL_MEMORY)
4679 node_set(node, slab_nodes);
4681 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4682 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4684 register_hotmemory_notifier(&slab_memory_callback_nb);
4686 /* Able to allocate the per node structures */
4687 slab_state = PARTIAL;
4689 create_boot_cache(kmem_cache, "kmem_cache",
4690 offsetof(struct kmem_cache, node) +
4691 nr_node_ids * sizeof(struct kmem_cache_node *),
4692 SLAB_HWCACHE_ALIGN, 0, 0);
4694 kmem_cache = bootstrap(&boot_kmem_cache);
4695 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4697 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4698 setup_kmalloc_cache_index_table();
4699 create_kmalloc_caches(0);
4701 /* Setup random freelists for each cache */
4702 init_freelist_randomization();
4704 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4707 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4709 slub_min_order, slub_max_order, slub_min_objects,
4710 nr_cpu_ids, nr_node_ids);
4713 void __init kmem_cache_init_late(void)
4718 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4719 slab_flags_t flags, void (*ctor)(void *))
4721 struct kmem_cache *s;
4723 s = find_mergeable(size, align, flags, name, ctor);
4728 * Adjust the object sizes so that we clear
4729 * the complete object on kzalloc.
4731 s->object_size = max(s->object_size, size);
4732 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4734 if (sysfs_slab_alias(s, name)) {
4743 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4747 err = kmem_cache_open(s, flags);
4751 /* Mutex is not taken during early boot */
4752 if (slab_state <= UP)
4755 err = sysfs_slab_add(s);
4757 __kmem_cache_release(s);
4759 if (s->flags & SLAB_STORE_USER)
4760 debugfs_slab_add(s);
4765 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4767 struct kmem_cache *s;
4770 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4771 return kmalloc_large(size, gfpflags);
4773 s = kmalloc_slab(size, gfpflags);
4775 if (unlikely(ZERO_OR_NULL_PTR(s)))
4778 ret = slab_alloc(s, gfpflags, caller, size);
4780 /* Honor the call site pointer we received. */
4781 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4785 EXPORT_SYMBOL(__kmalloc_track_caller);
4788 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4789 int node, unsigned long caller)
4791 struct kmem_cache *s;
4794 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4795 ret = kmalloc_large_node(size, gfpflags, node);
4797 trace_kmalloc_node(caller, ret,
4798 size, PAGE_SIZE << get_order(size),
4804 s = kmalloc_slab(size, gfpflags);
4806 if (unlikely(ZERO_OR_NULL_PTR(s)))
4809 ret = slab_alloc_node(s, gfpflags, node, caller, size);
4811 /* Honor the call site pointer we received. */
4812 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4816 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4820 static int count_inuse(struct page *page)
4825 static int count_total(struct page *page)
4827 return page->objects;
4831 #ifdef CONFIG_SLUB_DEBUG
4832 static void validate_slab(struct kmem_cache *s, struct page *page,
4833 unsigned long *obj_map)
4836 void *addr = page_address(page);
4840 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4843 /* Now we know that a valid freelist exists */
4844 __fill_map(obj_map, s, page);
4845 for_each_object(p, s, addr, page->objects) {
4846 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
4847 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4849 if (!check_object(s, page, p, val))
4856 static int validate_slab_node(struct kmem_cache *s,
4857 struct kmem_cache_node *n, unsigned long *obj_map)
4859 unsigned long count = 0;
4861 unsigned long flags;
4863 spin_lock_irqsave(&n->list_lock, flags);
4865 list_for_each_entry(page, &n->partial, slab_list) {
4866 validate_slab(s, page, obj_map);
4869 if (count != n->nr_partial) {
4870 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4871 s->name, count, n->nr_partial);
4872 slab_add_kunit_errors();
4875 if (!(s->flags & SLAB_STORE_USER))
4878 list_for_each_entry(page, &n->full, slab_list) {
4879 validate_slab(s, page, obj_map);
4882 if (count != atomic_long_read(&n->nr_slabs)) {
4883 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4884 s->name, count, atomic_long_read(&n->nr_slabs));
4885 slab_add_kunit_errors();
4889 spin_unlock_irqrestore(&n->list_lock, flags);
4893 long validate_slab_cache(struct kmem_cache *s)
4896 unsigned long count = 0;
4897 struct kmem_cache_node *n;
4898 unsigned long *obj_map;
4900 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
4905 for_each_kmem_cache_node(s, node, n)
4906 count += validate_slab_node(s, n, obj_map);
4908 bitmap_free(obj_map);
4912 EXPORT_SYMBOL(validate_slab_cache);
4914 #ifdef CONFIG_DEBUG_FS
4916 * Generate lists of code addresses where slabcache objects are allocated
4921 unsigned long count;
4928 DECLARE_BITMAP(cpus, NR_CPUS);
4934 unsigned long count;
4935 struct location *loc;
4938 static struct dentry *slab_debugfs_root;
4940 static void free_loc_track(struct loc_track *t)
4943 free_pages((unsigned long)t->loc,
4944 get_order(sizeof(struct location) * t->max));
4947 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4952 order = get_order(sizeof(struct location) * max);
4954 l = (void *)__get_free_pages(flags, order);
4959 memcpy(l, t->loc, sizeof(struct location) * t->count);
4967 static int add_location(struct loc_track *t, struct kmem_cache *s,
4968 const struct track *track)
4970 long start, end, pos;
4972 unsigned long caddr;
4973 unsigned long age = jiffies - track->when;
4979 pos = start + (end - start + 1) / 2;
4982 * There is nothing at "end". If we end up there
4983 * we need to add something to before end.
4988 caddr = t->loc[pos].addr;
4989 if (track->addr == caddr) {
4995 if (age < l->min_time)
4997 if (age > l->max_time)
5000 if (track->pid < l->min_pid)
5001 l->min_pid = track->pid;
5002 if (track->pid > l->max_pid)
5003 l->max_pid = track->pid;
5005 cpumask_set_cpu(track->cpu,
5006 to_cpumask(l->cpus));
5008 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5012 if (track->addr < caddr)
5019 * Not found. Insert new tracking element.
5021 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5027 (t->count - pos) * sizeof(struct location));
5030 l->addr = track->addr;
5034 l->min_pid = track->pid;
5035 l->max_pid = track->pid;
5036 cpumask_clear(to_cpumask(l->cpus));
5037 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5038 nodes_clear(l->nodes);
5039 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5043 static void process_slab(struct loc_track *t, struct kmem_cache *s,
5044 struct page *page, enum track_item alloc,
5045 unsigned long *obj_map)
5047 void *addr = page_address(page);
5050 __fill_map(obj_map, s, page);
5052 for_each_object(p, s, addr, page->objects)
5053 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5054 add_location(t, s, get_track(s, p, alloc));
5056 #endif /* CONFIG_DEBUG_FS */
5057 #endif /* CONFIG_SLUB_DEBUG */
5060 enum slab_stat_type {
5061 SL_ALL, /* All slabs */
5062 SL_PARTIAL, /* Only partially allocated slabs */
5063 SL_CPU, /* Only slabs used for cpu caches */
5064 SL_OBJECTS, /* Determine allocated objects not slabs */
5065 SL_TOTAL /* Determine object capacity not slabs */
5068 #define SO_ALL (1 << SL_ALL)
5069 #define SO_PARTIAL (1 << SL_PARTIAL)
5070 #define SO_CPU (1 << SL_CPU)
5071 #define SO_OBJECTS (1 << SL_OBJECTS)
5072 #define SO_TOTAL (1 << SL_TOTAL)
5074 static ssize_t show_slab_objects(struct kmem_cache *s,
5075 char *buf, unsigned long flags)
5077 unsigned long total = 0;
5080 unsigned long *nodes;
5083 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5087 if (flags & SO_CPU) {
5090 for_each_possible_cpu(cpu) {
5091 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5096 page = READ_ONCE(c->page);
5100 node = page_to_nid(page);
5101 if (flags & SO_TOTAL)
5103 else if (flags & SO_OBJECTS)
5111 page = slub_percpu_partial_read_once(c);
5113 node = page_to_nid(page);
5114 if (flags & SO_TOTAL)
5116 else if (flags & SO_OBJECTS)
5127 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5128 * already held which will conflict with an existing lock order:
5130 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5132 * We don't really need mem_hotplug_lock (to hold off
5133 * slab_mem_going_offline_callback) here because slab's memory hot
5134 * unplug code doesn't destroy the kmem_cache->node[] data.
5137 #ifdef CONFIG_SLUB_DEBUG
5138 if (flags & SO_ALL) {
5139 struct kmem_cache_node *n;
5141 for_each_kmem_cache_node(s, node, n) {
5143 if (flags & SO_TOTAL)
5144 x = atomic_long_read(&n->total_objects);
5145 else if (flags & SO_OBJECTS)
5146 x = atomic_long_read(&n->total_objects) -
5147 count_partial(n, count_free);
5149 x = atomic_long_read(&n->nr_slabs);
5156 if (flags & SO_PARTIAL) {
5157 struct kmem_cache_node *n;
5159 for_each_kmem_cache_node(s, node, n) {
5160 if (flags & SO_TOTAL)
5161 x = count_partial(n, count_total);
5162 else if (flags & SO_OBJECTS)
5163 x = count_partial(n, count_inuse);
5171 len += sysfs_emit_at(buf, len, "%lu", total);
5173 for (node = 0; node < nr_node_ids; node++) {
5175 len += sysfs_emit_at(buf, len, " N%d=%lu",
5179 len += sysfs_emit_at(buf, len, "\n");
5185 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5186 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5188 struct slab_attribute {
5189 struct attribute attr;
5190 ssize_t (*show)(struct kmem_cache *s, char *buf);
5191 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5194 #define SLAB_ATTR_RO(_name) \
5195 static struct slab_attribute _name##_attr = \
5196 __ATTR(_name, 0400, _name##_show, NULL)
5198 #define SLAB_ATTR(_name) \
5199 static struct slab_attribute _name##_attr = \
5200 __ATTR(_name, 0600, _name##_show, _name##_store)
5202 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5204 return sysfs_emit(buf, "%u\n", s->size);
5206 SLAB_ATTR_RO(slab_size);
5208 static ssize_t align_show(struct kmem_cache *s, char *buf)
5210 return sysfs_emit(buf, "%u\n", s->align);
5212 SLAB_ATTR_RO(align);
5214 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5216 return sysfs_emit(buf, "%u\n", s->object_size);
5218 SLAB_ATTR_RO(object_size);
5220 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5222 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5224 SLAB_ATTR_RO(objs_per_slab);
5226 static ssize_t order_show(struct kmem_cache *s, char *buf)
5228 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5230 SLAB_ATTR_RO(order);
5232 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5234 return sysfs_emit(buf, "%lu\n", s->min_partial);
5237 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5243 err = kstrtoul(buf, 10, &min);
5247 set_min_partial(s, min);
5250 SLAB_ATTR(min_partial);
5252 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5254 return sysfs_emit(buf, "%u\n", slub_cpu_partial(s));
5257 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5260 unsigned int objects;
5263 err = kstrtouint(buf, 10, &objects);
5266 if (objects && !kmem_cache_has_cpu_partial(s))
5269 slub_set_cpu_partial(s, objects);
5273 SLAB_ATTR(cpu_partial);
5275 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5279 return sysfs_emit(buf, "%pS\n", s->ctor);
5283 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5285 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5287 SLAB_ATTR_RO(aliases);
5289 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5291 return show_slab_objects(s, buf, SO_PARTIAL);
5293 SLAB_ATTR_RO(partial);
5295 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5297 return show_slab_objects(s, buf, SO_CPU);
5299 SLAB_ATTR_RO(cpu_slabs);
5301 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5303 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5305 SLAB_ATTR_RO(objects);
5307 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5309 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5311 SLAB_ATTR_RO(objects_partial);
5313 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5320 for_each_online_cpu(cpu) {
5323 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5326 pages += page->pages;
5327 objects += page->pobjects;
5331 len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages);
5334 for_each_online_cpu(cpu) {
5337 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5339 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5340 cpu, page->pobjects, page->pages);
5343 len += sysfs_emit_at(buf, len, "\n");
5347 SLAB_ATTR_RO(slabs_cpu_partial);
5349 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5351 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5353 SLAB_ATTR_RO(reclaim_account);
5355 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5357 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5359 SLAB_ATTR_RO(hwcache_align);
5361 #ifdef CONFIG_ZONE_DMA
5362 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5364 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5366 SLAB_ATTR_RO(cache_dma);
5369 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5371 return sysfs_emit(buf, "%u\n", s->usersize);
5373 SLAB_ATTR_RO(usersize);
5375 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5377 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5379 SLAB_ATTR_RO(destroy_by_rcu);
5381 #ifdef CONFIG_SLUB_DEBUG
5382 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5384 return show_slab_objects(s, buf, SO_ALL);
5386 SLAB_ATTR_RO(slabs);
5388 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5390 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5392 SLAB_ATTR_RO(total_objects);
5394 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5396 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5398 SLAB_ATTR_RO(sanity_checks);
5400 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5402 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5404 SLAB_ATTR_RO(trace);
5406 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5408 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5411 SLAB_ATTR_RO(red_zone);
5413 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5415 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5418 SLAB_ATTR_RO(poison);
5420 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5422 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5425 SLAB_ATTR_RO(store_user);
5427 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5432 static ssize_t validate_store(struct kmem_cache *s,
5433 const char *buf, size_t length)
5437 if (buf[0] == '1') {
5438 ret = validate_slab_cache(s);
5444 SLAB_ATTR(validate);
5446 #endif /* CONFIG_SLUB_DEBUG */
5448 #ifdef CONFIG_FAILSLAB
5449 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5451 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5453 SLAB_ATTR_RO(failslab);
5456 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5461 static ssize_t shrink_store(struct kmem_cache *s,
5462 const char *buf, size_t length)
5465 kmem_cache_shrink(s);
5473 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5475 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5478 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5479 const char *buf, size_t length)
5484 err = kstrtouint(buf, 10, &ratio);
5490 s->remote_node_defrag_ratio = ratio * 10;
5494 SLAB_ATTR(remote_node_defrag_ratio);
5497 #ifdef CONFIG_SLUB_STATS
5498 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5500 unsigned long sum = 0;
5503 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5508 for_each_online_cpu(cpu) {
5509 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5515 len += sysfs_emit_at(buf, len, "%lu", sum);
5518 for_each_online_cpu(cpu) {
5520 len += sysfs_emit_at(buf, len, " C%d=%u",
5525 len += sysfs_emit_at(buf, len, "\n");
5530 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5534 for_each_online_cpu(cpu)
5535 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5538 #define STAT_ATTR(si, text) \
5539 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5541 return show_stat(s, buf, si); \
5543 static ssize_t text##_store(struct kmem_cache *s, \
5544 const char *buf, size_t length) \
5546 if (buf[0] != '0') \
5548 clear_stat(s, si); \
5553 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5554 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5555 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5556 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5557 STAT_ATTR(FREE_FROZEN, free_frozen);
5558 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5559 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5560 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5561 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5562 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5563 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5564 STAT_ATTR(FREE_SLAB, free_slab);
5565 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5566 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5567 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5568 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5569 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5570 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5571 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5572 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5573 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5574 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5575 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5576 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5577 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5578 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5579 #endif /* CONFIG_SLUB_STATS */
5581 static struct attribute *slab_attrs[] = {
5582 &slab_size_attr.attr,
5583 &object_size_attr.attr,
5584 &objs_per_slab_attr.attr,
5586 &min_partial_attr.attr,
5587 &cpu_partial_attr.attr,
5589 &objects_partial_attr.attr,
5591 &cpu_slabs_attr.attr,
5595 &hwcache_align_attr.attr,
5596 &reclaim_account_attr.attr,
5597 &destroy_by_rcu_attr.attr,
5599 &slabs_cpu_partial_attr.attr,
5600 #ifdef CONFIG_SLUB_DEBUG
5601 &total_objects_attr.attr,
5603 &sanity_checks_attr.attr,
5605 &red_zone_attr.attr,
5607 &store_user_attr.attr,
5608 &validate_attr.attr,
5610 #ifdef CONFIG_ZONE_DMA
5611 &cache_dma_attr.attr,
5614 &remote_node_defrag_ratio_attr.attr,
5616 #ifdef CONFIG_SLUB_STATS
5617 &alloc_fastpath_attr.attr,
5618 &alloc_slowpath_attr.attr,
5619 &free_fastpath_attr.attr,
5620 &free_slowpath_attr.attr,
5621 &free_frozen_attr.attr,
5622 &free_add_partial_attr.attr,
5623 &free_remove_partial_attr.attr,
5624 &alloc_from_partial_attr.attr,
5625 &alloc_slab_attr.attr,
5626 &alloc_refill_attr.attr,
5627 &alloc_node_mismatch_attr.attr,
5628 &free_slab_attr.attr,
5629 &cpuslab_flush_attr.attr,
5630 &deactivate_full_attr.attr,
5631 &deactivate_empty_attr.attr,
5632 &deactivate_to_head_attr.attr,
5633 &deactivate_to_tail_attr.attr,
5634 &deactivate_remote_frees_attr.attr,
5635 &deactivate_bypass_attr.attr,
5636 &order_fallback_attr.attr,
5637 &cmpxchg_double_fail_attr.attr,
5638 &cmpxchg_double_cpu_fail_attr.attr,
5639 &cpu_partial_alloc_attr.attr,
5640 &cpu_partial_free_attr.attr,
5641 &cpu_partial_node_attr.attr,
5642 &cpu_partial_drain_attr.attr,
5644 #ifdef CONFIG_FAILSLAB
5645 &failslab_attr.attr,
5647 &usersize_attr.attr,
5652 static const struct attribute_group slab_attr_group = {
5653 .attrs = slab_attrs,
5656 static ssize_t slab_attr_show(struct kobject *kobj,
5657 struct attribute *attr,
5660 struct slab_attribute *attribute;
5661 struct kmem_cache *s;
5664 attribute = to_slab_attr(attr);
5667 if (!attribute->show)
5670 err = attribute->show(s, buf);
5675 static ssize_t slab_attr_store(struct kobject *kobj,
5676 struct attribute *attr,
5677 const char *buf, size_t len)
5679 struct slab_attribute *attribute;
5680 struct kmem_cache *s;
5683 attribute = to_slab_attr(attr);
5686 if (!attribute->store)
5689 err = attribute->store(s, buf, len);
5693 static void kmem_cache_release(struct kobject *k)
5695 slab_kmem_cache_release(to_slab(k));
5698 static const struct sysfs_ops slab_sysfs_ops = {
5699 .show = slab_attr_show,
5700 .store = slab_attr_store,
5703 static struct kobj_type slab_ktype = {
5704 .sysfs_ops = &slab_sysfs_ops,
5705 .release = kmem_cache_release,
5708 static struct kset *slab_kset;
5710 static inline struct kset *cache_kset(struct kmem_cache *s)
5715 #define ID_STR_LENGTH 64
5717 /* Create a unique string id for a slab cache:
5719 * Format :[flags-]size
5721 static char *create_unique_id(struct kmem_cache *s)
5723 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5730 * First flags affecting slabcache operations. We will only
5731 * get here for aliasable slabs so we do not need to support
5732 * too many flags. The flags here must cover all flags that
5733 * are matched during merging to guarantee that the id is
5736 if (s->flags & SLAB_CACHE_DMA)
5738 if (s->flags & SLAB_CACHE_DMA32)
5740 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5742 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5744 if (s->flags & SLAB_ACCOUNT)
5748 p += sprintf(p, "%07u", s->size);
5750 BUG_ON(p > name + ID_STR_LENGTH - 1);
5754 static int sysfs_slab_add(struct kmem_cache *s)
5758 struct kset *kset = cache_kset(s);
5759 int unmergeable = slab_unmergeable(s);
5762 kobject_init(&s->kobj, &slab_ktype);
5766 if (!unmergeable && disable_higher_order_debug &&
5767 (slub_debug & DEBUG_METADATA_FLAGS))
5772 * Slabcache can never be merged so we can use the name proper.
5773 * This is typically the case for debug situations. In that
5774 * case we can catch duplicate names easily.
5776 sysfs_remove_link(&slab_kset->kobj, s->name);
5780 * Create a unique name for the slab as a target
5783 name = create_unique_id(s);
5786 s->kobj.kset = kset;
5787 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5791 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5796 /* Setup first alias */
5797 sysfs_slab_alias(s, s->name);
5804 kobject_del(&s->kobj);
5808 void sysfs_slab_unlink(struct kmem_cache *s)
5810 if (slab_state >= FULL)
5811 kobject_del(&s->kobj);
5814 void sysfs_slab_release(struct kmem_cache *s)
5816 if (slab_state >= FULL)
5817 kobject_put(&s->kobj);
5821 * Need to buffer aliases during bootup until sysfs becomes
5822 * available lest we lose that information.
5824 struct saved_alias {
5825 struct kmem_cache *s;
5827 struct saved_alias *next;
5830 static struct saved_alias *alias_list;
5832 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5834 struct saved_alias *al;
5836 if (slab_state == FULL) {
5838 * If we have a leftover link then remove it.
5840 sysfs_remove_link(&slab_kset->kobj, name);
5841 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5844 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5850 al->next = alias_list;
5855 static int __init slab_sysfs_init(void)
5857 struct kmem_cache *s;
5860 mutex_lock(&slab_mutex);
5862 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5864 mutex_unlock(&slab_mutex);
5865 pr_err("Cannot register slab subsystem.\n");
5871 list_for_each_entry(s, &slab_caches, list) {
5872 err = sysfs_slab_add(s);
5874 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5878 while (alias_list) {
5879 struct saved_alias *al = alias_list;
5881 alias_list = alias_list->next;
5882 err = sysfs_slab_alias(al->s, al->name);
5884 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5889 mutex_unlock(&slab_mutex);
5893 __initcall(slab_sysfs_init);
5894 #endif /* CONFIG_SYSFS */
5896 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
5897 static int slab_debugfs_show(struct seq_file *seq, void *v)
5901 unsigned int idx = *(unsigned int *)v;
5902 struct loc_track *t = seq->private;
5904 if (idx < t->count) {
5907 seq_printf(seq, "%7ld ", l->count);
5910 seq_printf(seq, "%pS", (void *)l->addr);
5912 seq_puts(seq, "<not-available>");
5914 if (l->sum_time != l->min_time) {
5915 seq_printf(seq, " age=%ld/%llu/%ld",
5916 l->min_time, div_u64(l->sum_time, l->count),
5919 seq_printf(seq, " age=%ld", l->min_time);
5921 if (l->min_pid != l->max_pid)
5922 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
5924 seq_printf(seq, " pid=%ld",
5927 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
5928 seq_printf(seq, " cpus=%*pbl",
5929 cpumask_pr_args(to_cpumask(l->cpus)));
5931 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
5932 seq_printf(seq, " nodes=%*pbl",
5933 nodemask_pr_args(&l->nodes));
5935 seq_puts(seq, "\n");
5938 if (!idx && !t->count)
5939 seq_puts(seq, "No data\n");
5944 static void slab_debugfs_stop(struct seq_file *seq, void *v)
5948 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
5950 struct loc_track *t = seq->private;
5954 if (*ppos <= t->count)
5960 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
5965 static const struct seq_operations slab_debugfs_sops = {
5966 .start = slab_debugfs_start,
5967 .next = slab_debugfs_next,
5968 .stop = slab_debugfs_stop,
5969 .show = slab_debugfs_show,
5972 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
5975 struct kmem_cache_node *n;
5976 enum track_item alloc;
5978 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
5979 sizeof(struct loc_track));
5980 struct kmem_cache *s = file_inode(filep)->i_private;
5981 unsigned long *obj_map;
5983 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5987 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
5988 alloc = TRACK_ALLOC;
5992 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
5993 bitmap_free(obj_map);
5997 for_each_kmem_cache_node(s, node, n) {
5998 unsigned long flags;
6001 if (!atomic_long_read(&n->nr_slabs))
6004 spin_lock_irqsave(&n->list_lock, flags);
6005 list_for_each_entry(page, &n->partial, slab_list)
6006 process_slab(t, s, page, alloc, obj_map);
6007 list_for_each_entry(page, &n->full, slab_list)
6008 process_slab(t, s, page, alloc, obj_map);
6009 spin_unlock_irqrestore(&n->list_lock, flags);
6012 bitmap_free(obj_map);
6016 static int slab_debug_trace_release(struct inode *inode, struct file *file)
6018 struct seq_file *seq = file->private_data;
6019 struct loc_track *t = seq->private;
6022 return seq_release_private(inode, file);
6025 static const struct file_operations slab_debugfs_fops = {
6026 .open = slab_debug_trace_open,
6028 .llseek = seq_lseek,
6029 .release = slab_debug_trace_release,
6032 static void debugfs_slab_add(struct kmem_cache *s)
6034 struct dentry *slab_cache_dir;
6036 if (unlikely(!slab_debugfs_root))
6039 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6041 debugfs_create_file("alloc_traces", 0400,
6042 slab_cache_dir, s, &slab_debugfs_fops);
6044 debugfs_create_file("free_traces", 0400,
6045 slab_cache_dir, s, &slab_debugfs_fops);
6048 void debugfs_slab_release(struct kmem_cache *s)
6050 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
6053 static int __init slab_debugfs_init(void)
6055 struct kmem_cache *s;
6057 slab_debugfs_root = debugfs_create_dir("slab", NULL);
6059 list_for_each_entry(s, &slab_caches, list)
6060 if (s->flags & SLAB_STORE_USER)
6061 debugfs_slab_add(s);
6066 __initcall(slab_debugfs_init);
6069 * The /proc/slabinfo ABI
6071 #ifdef CONFIG_SLUB_DEBUG
6072 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6074 unsigned long nr_slabs = 0;
6075 unsigned long nr_objs = 0;
6076 unsigned long nr_free = 0;
6078 struct kmem_cache_node *n;
6080 for_each_kmem_cache_node(s, node, n) {
6081 nr_slabs += node_nr_slabs(n);
6082 nr_objs += node_nr_objs(n);
6083 nr_free += count_partial(n, count_free);
6086 sinfo->active_objs = nr_objs - nr_free;
6087 sinfo->num_objs = nr_objs;
6088 sinfo->active_slabs = nr_slabs;
6089 sinfo->num_slabs = nr_slabs;
6090 sinfo->objects_per_slab = oo_objects(s->oo);
6091 sinfo->cache_order = oo_order(s->oo);
6094 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6098 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6099 size_t count, loff_t *ppos)
6103 #endif /* CONFIG_SLUB_DEBUG */