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/stackdepot.h>
30 #include <linux/debugobjects.h>
31 #include <linux/kallsyms.h>
32 #include <linux/kfence.h>
33 #include <linux/memory.h>
34 #include <linux/math64.h>
35 #include <linux/fault-inject.h>
36 #include <linux/stacktrace.h>
37 #include <linux/prefetch.h>
38 #include <linux/memcontrol.h>
39 #include <linux/random.h>
40 #include <kunit/test.h>
42 #include <linux/debugfs.h>
43 #include <trace/events/kmem.h>
49 * 1. slab_mutex (Global Mutex)
51 * 3. slab_lock(page) (Only on some arches and for debugging)
55 * The role of the slab_mutex is to protect the list of all the slabs
56 * and to synchronize major metadata changes to slab cache structures.
58 * The slab_lock is only used for debugging and on arches that do not
59 * have the ability to do a cmpxchg_double. It only protects:
60 * A. page->freelist -> List of object free in a page
61 * B. page->inuse -> Number of objects in use
62 * C. page->objects -> Number of objects in page
63 * D. page->frozen -> frozen state
65 * If a slab is frozen then it is exempt from list management. It is not
66 * on any list except per cpu partial list. The processor that froze the
67 * slab is the one who can perform list operations on the page. Other
68 * processors may put objects onto the freelist but the processor that
69 * froze the slab is the only one that can retrieve the objects from the
72 * The list_lock protects the partial and full list on each node and
73 * the partial slab counter. If taken then no new slabs may be added or
74 * removed from the lists nor make the number of partial slabs be modified.
75 * (Note that the total number of slabs is an atomic value that may be
76 * modified without taking the list lock).
78 * The list_lock is a centralized lock and thus we avoid taking it as
79 * much as possible. As long as SLUB does not have to handle partial
80 * slabs, operations can continue without any centralized lock. F.e.
81 * allocating a long series of objects that fill up slabs does not require
83 * Interrupts are disabled during allocation and deallocation in order to
84 * make the slab allocator safe to use in the context of an irq. In addition
85 * interrupts are disabled to ensure that the processor does not change
86 * while handling per_cpu slabs, due to kernel preemption.
88 * SLUB assigns one slab for allocation to each processor.
89 * Allocations only occur from these slabs called cpu slabs.
91 * Slabs with free elements are kept on a partial list and during regular
92 * operations no list for full slabs is used. If an object in a full slab is
93 * freed then the slab will show up again on the partial lists.
94 * We track full slabs for debugging purposes though because otherwise we
95 * cannot scan all objects.
97 * Slabs are freed when they become empty. Teardown and setup is
98 * minimal so we rely on the page allocators per cpu caches for
99 * fast frees and allocs.
101 * page->frozen The slab is frozen and exempt from list processing.
102 * This means that the slab is dedicated to a purpose
103 * such as satisfying allocations for a specific
104 * processor. Objects may be freed in the slab while
105 * it is frozen but slab_free will then skip the usual
106 * list operations. It is up to the processor holding
107 * the slab to integrate the slab into the slab lists
108 * when the slab is no longer needed.
110 * One use of this flag is to mark slabs that are
111 * used for allocations. Then such a slab becomes a cpu
112 * slab. The cpu slab may be equipped with an additional
113 * freelist that allows lockless access to
114 * free objects in addition to the regular freelist
115 * that requires the slab lock.
117 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
118 * options set. This moves slab handling out of
119 * the fast path and disables lockless freelists.
122 #ifdef CONFIG_SLUB_DEBUG
124 #ifdef CONFIG_SLUB_DEBUG_ON
125 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
127 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
130 static inline bool __slub_debug_enabled(void)
132 return static_branch_unlikely(&slub_debug_enabled);
135 #else /* CONFIG_SLUB_DEBUG */
137 static inline bool __slub_debug_enabled(void)
142 #endif /* CONFIG_SLUB_DEBUG */
144 static inline bool kmem_cache_debug(struct kmem_cache *s)
146 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
149 void *fixup_red_left(struct kmem_cache *s, void *p)
151 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
152 p += s->red_left_pad;
157 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
159 #ifdef CONFIG_SLUB_CPU_PARTIAL
160 return !kmem_cache_debug(s);
167 * Issues still to be resolved:
169 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
171 * - Variable sizing of the per node arrays
174 /* Enable to log cmpxchg failures */
175 #undef SLUB_DEBUG_CMPXCHG
178 * Minimum number of partial slabs. These will be left on the partial
179 * lists even if they are empty. kmem_cache_shrink may reclaim them.
181 #define MIN_PARTIAL 5
184 * Maximum number of desirable partial slabs.
185 * The existence of more partial slabs makes kmem_cache_shrink
186 * sort the partial list by the number of objects in use.
188 #define MAX_PARTIAL 10
190 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
191 SLAB_POISON | SLAB_STORE_USER)
194 * These debug flags cannot use CMPXCHG because there might be consistency
195 * issues when checking or reading debug information
197 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
202 * Debugging flags that require metadata to be stored in the slab. These get
203 * disabled when slub_debug=O is used and a cache's min order increases with
206 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
209 #define OO_MASK ((1 << OO_SHIFT) - 1)
210 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
212 /* Internal SLUB flags */
214 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
215 /* Use cmpxchg_double */
216 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
219 * Tracking user of a slab.
221 #define TRACK_ADDRS_COUNT 16
223 unsigned long addr; /* Called from address */
224 #ifdef CONFIG_STACKDEPOT
225 depot_stack_handle_t handle;
227 int cpu; /* Was running on cpu */
228 int pid; /* Pid context */
229 unsigned long when; /* When did the operation occur */
232 enum track_item { TRACK_ALLOC, TRACK_FREE };
235 static int sysfs_slab_add(struct kmem_cache *);
236 static int sysfs_slab_alias(struct kmem_cache *, const char *);
238 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
239 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
243 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
244 static void debugfs_slab_add(struct kmem_cache *);
246 static inline void debugfs_slab_add(struct kmem_cache *s) { }
249 static inline void stat(const struct kmem_cache *s, enum stat_item si)
251 #ifdef CONFIG_SLUB_STATS
253 * The rmw is racy on a preemptible kernel but this is acceptable, so
254 * avoid this_cpu_add()'s irq-disable overhead.
256 raw_cpu_inc(s->cpu_slab->stat[si]);
261 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
262 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
263 * differ during memory hotplug/hotremove operations.
264 * Protected by slab_mutex.
266 static nodemask_t slab_nodes;
268 /********************************************************************
269 * Core slab cache functions
270 *******************************************************************/
273 * Returns freelist pointer (ptr). With hardening, this is obfuscated
274 * with an XOR of the address where the pointer is held and a per-cache
277 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
278 unsigned long ptr_addr)
280 #ifdef CONFIG_SLAB_FREELIST_HARDENED
282 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
283 * Normally, this doesn't cause any issues, as both set_freepointer()
284 * and get_freepointer() are called with a pointer with the same tag.
285 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
286 * example, when __free_slub() iterates over objects in a cache, it
287 * passes untagged pointers to check_object(). check_object() in turns
288 * calls get_freepointer() with an untagged pointer, which causes the
289 * freepointer to be restored incorrectly.
291 return (void *)((unsigned long)ptr ^ s->random ^
292 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
298 /* Returns the freelist pointer recorded at location ptr_addr. */
299 static inline void *freelist_dereference(const struct kmem_cache *s,
302 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
303 (unsigned long)ptr_addr);
306 static inline void *get_freepointer(struct kmem_cache *s, void *object)
308 object = kasan_reset_tag(object);
309 return freelist_dereference(s, object + s->offset);
312 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
314 prefetch(object + s->offset);
317 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
319 unsigned long freepointer_addr;
322 if (!debug_pagealloc_enabled_static())
323 return get_freepointer(s, object);
325 object = kasan_reset_tag(object);
326 freepointer_addr = (unsigned long)object + s->offset;
327 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
328 return freelist_ptr(s, p, freepointer_addr);
331 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
333 unsigned long freeptr_addr = (unsigned long)object + s->offset;
335 #ifdef CONFIG_SLAB_FREELIST_HARDENED
336 BUG_ON(object == fp); /* naive detection of double free or corruption */
339 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
340 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
343 /* Loop over all objects in a slab */
344 #define for_each_object(__p, __s, __addr, __objects) \
345 for (__p = fixup_red_left(__s, __addr); \
346 __p < (__addr) + (__objects) * (__s)->size; \
349 static inline unsigned int order_objects(unsigned int order, unsigned int size)
351 return ((unsigned int)PAGE_SIZE << order) / size;
354 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
357 struct kmem_cache_order_objects x = {
358 (order << OO_SHIFT) + order_objects(order, size)
364 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
366 return x.x >> OO_SHIFT;
369 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
371 return x.x & OO_MASK;
375 * Per slab locking using the pagelock
377 static __always_inline void slab_lock(struct page *page)
379 VM_BUG_ON_PAGE(PageTail(page), page);
380 bit_spin_lock(PG_locked, &page->flags);
383 static __always_inline void slab_unlock(struct page *page)
385 VM_BUG_ON_PAGE(PageTail(page), page);
386 __bit_spin_unlock(PG_locked, &page->flags);
389 /* Interrupts must be disabled (for the fallback code to work right) */
390 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
391 void *freelist_old, unsigned long counters_old,
392 void *freelist_new, unsigned long counters_new,
395 VM_BUG_ON(!irqs_disabled());
396 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
397 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
398 if (s->flags & __CMPXCHG_DOUBLE) {
399 if (cmpxchg_double(&page->freelist, &page->counters,
400 freelist_old, counters_old,
401 freelist_new, counters_new))
407 if (page->freelist == freelist_old &&
408 page->counters == counters_old) {
409 page->freelist = freelist_new;
410 page->counters = counters_new;
418 stat(s, CMPXCHG_DOUBLE_FAIL);
420 #ifdef SLUB_DEBUG_CMPXCHG
421 pr_info("%s %s: cmpxchg double redo ", n, s->name);
427 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
428 void *freelist_old, unsigned long counters_old,
429 void *freelist_new, unsigned long counters_new,
432 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
433 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
434 if (s->flags & __CMPXCHG_DOUBLE) {
435 if (cmpxchg_double(&page->freelist, &page->counters,
436 freelist_old, counters_old,
437 freelist_new, counters_new))
444 local_irq_save(flags);
446 if (page->freelist == freelist_old &&
447 page->counters == counters_old) {
448 page->freelist = freelist_new;
449 page->counters = counters_new;
451 local_irq_restore(flags);
455 local_irq_restore(flags);
459 stat(s, CMPXCHG_DOUBLE_FAIL);
461 #ifdef SLUB_DEBUG_CMPXCHG
462 pr_info("%s %s: cmpxchg double redo ", n, s->name);
468 #ifdef CONFIG_SLUB_DEBUG
469 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
470 static DEFINE_SPINLOCK(object_map_lock);
472 #if IS_ENABLED(CONFIG_KUNIT)
473 static bool slab_add_kunit_errors(void)
475 struct kunit_resource *resource;
477 if (likely(!current->kunit_test))
480 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
484 (*(int *)resource->data)++;
485 kunit_put_resource(resource);
489 static inline bool slab_add_kunit_errors(void) { return false; }
493 * Determine a map of object in use on a page.
495 * Node listlock must be held to guarantee that the page does
496 * not vanish from under us.
498 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
499 __acquires(&object_map_lock)
502 void *addr = page_address(page);
504 VM_BUG_ON(!irqs_disabled());
506 spin_lock(&object_map_lock);
508 bitmap_zero(object_map, page->objects);
510 for (p = page->freelist; p; p = get_freepointer(s, p))
511 set_bit(__obj_to_index(s, addr, p), object_map);
516 static void put_map(unsigned long *map) __releases(&object_map_lock)
518 VM_BUG_ON(map != object_map);
519 spin_unlock(&object_map_lock);
522 static inline unsigned int size_from_object(struct kmem_cache *s)
524 if (s->flags & SLAB_RED_ZONE)
525 return s->size - s->red_left_pad;
530 static inline void *restore_red_left(struct kmem_cache *s, void *p)
532 if (s->flags & SLAB_RED_ZONE)
533 p -= s->red_left_pad;
541 #if defined(CONFIG_SLUB_DEBUG_ON)
542 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
544 static slab_flags_t slub_debug;
547 static char *slub_debug_string;
548 static int disable_higher_order_debug;
551 * slub is about to manipulate internal object metadata. This memory lies
552 * outside the range of the allocated object, so accessing it would normally
553 * be reported by kasan as a bounds error. metadata_access_enable() is used
554 * to tell kasan that these accesses are OK.
556 static inline void metadata_access_enable(void)
558 kasan_disable_current();
561 static inline void metadata_access_disable(void)
563 kasan_enable_current();
570 /* Verify that a pointer has an address that is valid within a slab page */
571 static inline int check_valid_pointer(struct kmem_cache *s,
572 struct page *page, void *object)
579 base = page_address(page);
580 object = kasan_reset_tag(object);
581 object = restore_red_left(s, object);
582 if (object < base || object >= base + page->objects * s->size ||
583 (object - base) % s->size) {
590 static void print_section(char *level, char *text, u8 *addr,
593 metadata_access_enable();
594 print_hex_dump(level, kasan_reset_tag(text), DUMP_PREFIX_ADDRESS,
595 16, 1, addr, length, 1);
596 metadata_access_disable();
600 * See comment in calculate_sizes().
602 static inline bool freeptr_outside_object(struct kmem_cache *s)
604 return s->offset >= s->inuse;
608 * Return offset of the end of info block which is inuse + free pointer if
609 * not overlapping with object.
611 static inline unsigned int get_info_end(struct kmem_cache *s)
613 if (freeptr_outside_object(s))
614 return s->inuse + sizeof(void *);
619 static struct track *get_track(struct kmem_cache *s, void *object,
620 enum track_item alloc)
624 p = object + get_info_end(s);
626 return kasan_reset_tag(p + alloc);
629 #ifdef CONFIG_STACKDEPOT
630 static depot_stack_handle_t save_stack_depot_trace(gfp_t flags)
632 unsigned long entries[TRACK_ADDRS_COUNT];
633 depot_stack_handle_t handle;
634 unsigned int nr_entries;
636 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 4);
637 handle = stack_depot_save(entries, nr_entries, flags);
642 static void set_track(struct kmem_cache *s, void *object,
643 enum track_item alloc, unsigned long addr)
645 struct track *p = get_track(s, object, alloc);
648 #ifdef CONFIG_STACKDEPOT
649 p->handle = save_stack_depot_trace(GFP_NOWAIT);
652 p->cpu = smp_processor_id();
653 p->pid = current->pid;
656 memset(p, 0, sizeof(struct track));
660 static void init_tracking(struct kmem_cache *s, void *object)
662 if (!(s->flags & SLAB_STORE_USER))
665 set_track(s, object, TRACK_FREE, 0UL);
666 set_track(s, object, TRACK_ALLOC, 0UL);
669 static void print_track(const char *s, struct track *t, unsigned long pr_time)
674 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
675 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
676 #ifdef CONFIG_STACKDEPOT
678 depot_stack_handle_t handle;
679 unsigned long *entries;
680 unsigned int nr_entries;
682 handle = READ_ONCE(t->handle);
684 pr_err("object allocation/free stack trace missing\n");
686 nr_entries = stack_depot_fetch(handle, &entries);
687 stack_trace_print(entries, nr_entries, 0);
693 void print_tracking(struct kmem_cache *s, void *object)
695 unsigned long pr_time = jiffies;
696 if (!(s->flags & SLAB_STORE_USER))
699 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
700 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
703 static void print_page_info(struct page *page)
705 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
706 page, page->objects, page->inuse, page->freelist,
707 page->flags, &page->flags);
711 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
713 struct va_format vaf;
719 pr_err("=============================================================================\n");
720 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
721 pr_err("-----------------------------------------------------------------------------\n\n");
726 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
728 struct va_format vaf;
731 if (slab_add_kunit_errors())
737 pr_err("FIX %s: %pV\n", s->name, &vaf);
741 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
742 void **freelist, void *nextfree)
744 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
745 !check_valid_pointer(s, page, nextfree) && freelist) {
746 object_err(s, page, *freelist, "Freechain corrupt");
748 slab_fix(s, "Isolate corrupted freechain");
755 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
757 unsigned int off; /* Offset of last byte */
758 u8 *addr = page_address(page);
760 print_tracking(s, p);
762 print_page_info(page);
764 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
765 p, p - addr, get_freepointer(s, p));
767 if (s->flags & SLAB_RED_ZONE)
768 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
770 else if (p > addr + 16)
771 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
773 print_section(KERN_ERR, "Object ", p,
774 min_t(unsigned int, s->object_size, PAGE_SIZE));
775 if (s->flags & SLAB_RED_ZONE)
776 print_section(KERN_ERR, "Redzone ", p + s->object_size,
777 s->inuse - s->object_size);
779 off = get_info_end(s);
781 if (s->flags & SLAB_STORE_USER)
782 off += 2 * sizeof(struct track);
784 off += kasan_metadata_size(s);
786 if (off != size_from_object(s))
787 /* Beginning of the filler is the free pointer */
788 print_section(KERN_ERR, "Padding ", p + off,
789 size_from_object(s) - off);
794 void object_err(struct kmem_cache *s, struct page *page,
795 u8 *object, char *reason)
797 if (slab_add_kunit_errors())
800 slab_bug(s, "%s", reason);
801 print_trailer(s, page, object);
802 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
805 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
806 const char *fmt, ...)
811 if (slab_add_kunit_errors())
815 vsnprintf(buf, sizeof(buf), fmt, args);
817 slab_bug(s, "%s", buf);
818 print_page_info(page);
820 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
823 static void init_object(struct kmem_cache *s, void *object, u8 val)
825 u8 *p = kasan_reset_tag(object);
827 if (s->flags & SLAB_RED_ZONE)
828 memset(p - s->red_left_pad, val, s->red_left_pad);
830 if (s->flags & __OBJECT_POISON) {
831 memset(p, POISON_FREE, s->object_size - 1);
832 p[s->object_size - 1] = POISON_END;
835 if (s->flags & SLAB_RED_ZONE)
836 memset(p + s->object_size, val, s->inuse - s->object_size);
839 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
840 void *from, void *to)
842 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
843 memset(from, data, to - from);
846 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
847 u8 *object, char *what,
848 u8 *start, unsigned int value, unsigned int bytes)
852 u8 *addr = page_address(page);
854 metadata_access_enable();
855 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
856 metadata_access_disable();
861 while (end > fault && end[-1] == value)
864 if (slab_add_kunit_errors())
867 slab_bug(s, "%s overwritten", what);
868 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
869 fault, end - 1, fault - addr,
871 print_trailer(s, page, object);
872 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
875 restore_bytes(s, what, value, fault, end);
883 * Bytes of the object to be managed.
884 * If the freepointer may overlay the object then the free
885 * pointer is at the middle of the object.
887 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
890 * object + s->object_size
891 * Padding to reach word boundary. This is also used for Redzoning.
892 * Padding is extended by another word if Redzoning is enabled and
893 * object_size == inuse.
895 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
896 * 0xcc (RED_ACTIVE) for objects in use.
899 * Meta data starts here.
901 * A. Free pointer (if we cannot overwrite object on free)
902 * B. Tracking data for SLAB_STORE_USER
903 * C. Padding to reach required alignment boundary or at minimum
904 * one word if debugging is on to be able to detect writes
905 * before the word boundary.
907 * Padding is done using 0x5a (POISON_INUSE)
910 * Nothing is used beyond s->size.
912 * If slabcaches are merged then the object_size and inuse boundaries are mostly
913 * ignored. And therefore no slab options that rely on these boundaries
914 * may be used with merged slabcaches.
917 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
919 unsigned long off = get_info_end(s); /* The end of info */
921 if (s->flags & SLAB_STORE_USER)
922 /* We also have user information there */
923 off += 2 * sizeof(struct track);
925 off += kasan_metadata_size(s);
927 if (size_from_object(s) == off)
930 return check_bytes_and_report(s, page, p, "Object padding",
931 p + off, POISON_INUSE, size_from_object(s) - off);
934 /* Check the pad bytes at the end of a slab page */
935 static int slab_pad_check(struct kmem_cache *s, struct page *page)
944 if (!(s->flags & SLAB_POISON))
947 start = page_address(page);
948 length = page_size(page);
949 end = start + length;
950 remainder = length % s->size;
954 pad = end - remainder;
955 metadata_access_enable();
956 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
957 metadata_access_disable();
960 while (end > fault && end[-1] == POISON_INUSE)
963 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
964 fault, end - 1, fault - start);
965 print_section(KERN_ERR, "Padding ", pad, remainder);
967 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
971 static int check_object(struct kmem_cache *s, struct page *page,
972 void *object, u8 val)
975 u8 *endobject = object + s->object_size;
977 if (s->flags & SLAB_RED_ZONE) {
978 if (!check_bytes_and_report(s, page, object, "Left Redzone",
979 object - s->red_left_pad, val, s->red_left_pad))
982 if (!check_bytes_and_report(s, page, object, "Right Redzone",
983 endobject, val, s->inuse - s->object_size))
986 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
987 check_bytes_and_report(s, page, p, "Alignment padding",
988 endobject, POISON_INUSE,
989 s->inuse - s->object_size);
993 if (s->flags & SLAB_POISON) {
994 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
995 (!check_bytes_and_report(s, page, p, "Poison", p,
996 POISON_FREE, s->object_size - 1) ||
997 !check_bytes_and_report(s, page, p, "End Poison",
998 p + s->object_size - 1, POISON_END, 1)))
1001 * check_pad_bytes cleans up on its own.
1003 check_pad_bytes(s, page, p);
1006 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1008 * Object and freepointer overlap. Cannot check
1009 * freepointer while object is allocated.
1013 /* Check free pointer validity */
1014 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
1015 object_err(s, page, p, "Freepointer corrupt");
1017 * No choice but to zap it and thus lose the remainder
1018 * of the free objects in this slab. May cause
1019 * another error because the object count is now wrong.
1021 set_freepointer(s, p, NULL);
1027 static int check_slab(struct kmem_cache *s, struct page *page)
1031 VM_BUG_ON(!irqs_disabled());
1033 if (!PageSlab(page)) {
1034 slab_err(s, page, "Not a valid slab page");
1038 maxobj = order_objects(compound_order(page), s->size);
1039 if (page->objects > maxobj) {
1040 slab_err(s, page, "objects %u > max %u",
1041 page->objects, maxobj);
1044 if (page->inuse > page->objects) {
1045 slab_err(s, page, "inuse %u > max %u",
1046 page->inuse, page->objects);
1049 /* Slab_pad_check fixes things up after itself */
1050 slab_pad_check(s, page);
1055 * Determine if a certain object on a page is on the freelist. Must hold the
1056 * slab lock to guarantee that the chains are in a consistent state.
1058 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
1062 void *object = NULL;
1065 fp = page->freelist;
1066 while (fp && nr <= page->objects) {
1069 if (!check_valid_pointer(s, page, fp)) {
1071 object_err(s, page, object,
1072 "Freechain corrupt");
1073 set_freepointer(s, object, NULL);
1075 slab_err(s, page, "Freepointer corrupt");
1076 page->freelist = NULL;
1077 page->inuse = page->objects;
1078 slab_fix(s, "Freelist cleared");
1084 fp = get_freepointer(s, object);
1088 max_objects = order_objects(compound_order(page), s->size);
1089 if (max_objects > MAX_OBJS_PER_PAGE)
1090 max_objects = MAX_OBJS_PER_PAGE;
1092 if (page->objects != max_objects) {
1093 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1094 page->objects, max_objects);
1095 page->objects = max_objects;
1096 slab_fix(s, "Number of objects adjusted");
1098 if (page->inuse != page->objects - nr) {
1099 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1100 page->inuse, page->objects - nr);
1101 page->inuse = page->objects - nr;
1102 slab_fix(s, "Object count adjusted");
1104 return search == NULL;
1107 static void trace(struct kmem_cache *s, struct page *page, void *object,
1110 if (s->flags & SLAB_TRACE) {
1111 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1113 alloc ? "alloc" : "free",
1114 object, page->inuse,
1118 print_section(KERN_INFO, "Object ", (void *)object,
1126 * Tracking of fully allocated slabs for debugging purposes.
1128 static void add_full(struct kmem_cache *s,
1129 struct kmem_cache_node *n, struct page *page)
1131 if (!(s->flags & SLAB_STORE_USER))
1134 lockdep_assert_held(&n->list_lock);
1135 list_add(&page->slab_list, &n->full);
1138 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1140 if (!(s->flags & SLAB_STORE_USER))
1143 lockdep_assert_held(&n->list_lock);
1144 list_del(&page->slab_list);
1147 /* Tracking of the number of slabs for debugging purposes */
1148 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1150 struct kmem_cache_node *n = get_node(s, node);
1152 return atomic_long_read(&n->nr_slabs);
1155 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1157 return atomic_long_read(&n->nr_slabs);
1160 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1162 struct kmem_cache_node *n = get_node(s, node);
1165 * May be called early in order to allocate a slab for the
1166 * kmem_cache_node structure. Solve the chicken-egg
1167 * dilemma by deferring the increment of the count during
1168 * bootstrap (see early_kmem_cache_node_alloc).
1171 atomic_long_inc(&n->nr_slabs);
1172 atomic_long_add(objects, &n->total_objects);
1175 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1177 struct kmem_cache_node *n = get_node(s, node);
1179 atomic_long_dec(&n->nr_slabs);
1180 atomic_long_sub(objects, &n->total_objects);
1183 /* Object debug checks for alloc/free paths */
1184 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1187 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1190 init_object(s, object, SLUB_RED_INACTIVE);
1191 init_tracking(s, object);
1195 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1197 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1200 metadata_access_enable();
1201 memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page));
1202 metadata_access_disable();
1205 static inline int alloc_consistency_checks(struct kmem_cache *s,
1206 struct page *page, void *object)
1208 if (!check_slab(s, page))
1211 if (!check_valid_pointer(s, page, object)) {
1212 object_err(s, page, object, "Freelist Pointer check fails");
1216 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1222 static noinline int alloc_debug_processing(struct kmem_cache *s,
1224 void *object, unsigned long addr)
1226 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1227 if (!alloc_consistency_checks(s, page, object))
1231 /* Success perform special debug activities for allocs */
1232 if (s->flags & SLAB_STORE_USER)
1233 set_track(s, object, TRACK_ALLOC, addr);
1234 trace(s, page, object, 1);
1235 init_object(s, object, SLUB_RED_ACTIVE);
1239 if (PageSlab(page)) {
1241 * If this is a slab page then lets do the best we can
1242 * to avoid issues in the future. Marking all objects
1243 * as used avoids touching the remaining objects.
1245 slab_fix(s, "Marking all objects used");
1246 page->inuse = page->objects;
1247 page->freelist = NULL;
1252 static inline int free_consistency_checks(struct kmem_cache *s,
1253 struct page *page, void *object, unsigned long addr)
1255 if (!check_valid_pointer(s, page, object)) {
1256 slab_err(s, page, "Invalid object pointer 0x%p", object);
1260 if (on_freelist(s, page, object)) {
1261 object_err(s, page, object, "Object already free");
1265 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1268 if (unlikely(s != page->slab_cache)) {
1269 if (!PageSlab(page)) {
1270 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1272 } else if (!page->slab_cache) {
1273 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1277 object_err(s, page, object,
1278 "page slab pointer corrupt.");
1284 /* Supports checking bulk free of a constructed freelist */
1285 static noinline int free_debug_processing(
1286 struct kmem_cache *s, struct page *page,
1287 void *head, void *tail, int bulk_cnt,
1290 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1291 void *object = head;
1293 unsigned long flags;
1296 spin_lock_irqsave(&n->list_lock, flags);
1299 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1300 if (!check_slab(s, page))
1307 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1308 if (!free_consistency_checks(s, page, object, addr))
1312 if (s->flags & SLAB_STORE_USER)
1313 set_track(s, object, TRACK_FREE, addr);
1314 trace(s, page, object, 0);
1315 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1316 init_object(s, object, SLUB_RED_INACTIVE);
1318 /* Reached end of constructed freelist yet? */
1319 if (object != tail) {
1320 object = get_freepointer(s, object);
1326 if (cnt != bulk_cnt)
1327 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1331 spin_unlock_irqrestore(&n->list_lock, flags);
1333 slab_fix(s, "Object at 0x%p not freed", object);
1338 * Parse a block of slub_debug options. Blocks are delimited by ';'
1340 * @str: start of block
1341 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1342 * @slabs: return start of list of slabs, or NULL when there's no list
1343 * @init: assume this is initial parsing and not per-kmem-create parsing
1345 * returns the start of next block if there's any, or NULL
1348 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1350 bool higher_order_disable = false;
1352 /* Skip any completely empty blocks */
1353 while (*str && *str == ';')
1358 * No options but restriction on slabs. This means full
1359 * debugging for slabs matching a pattern.
1361 *flags = DEBUG_DEFAULT_FLAGS;
1366 /* Determine which debug features should be switched on */
1367 for (; *str && *str != ',' && *str != ';'; str++) {
1368 switch (tolower(*str)) {
1373 *flags |= SLAB_CONSISTENCY_CHECKS;
1376 *flags |= SLAB_RED_ZONE;
1379 *flags |= SLAB_POISON;
1382 *flags |= SLAB_STORE_USER;
1385 *flags |= SLAB_TRACE;
1388 *flags |= SLAB_FAILSLAB;
1392 * Avoid enabling debugging on caches if its minimum
1393 * order would increase as a result.
1395 higher_order_disable = true;
1399 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1408 /* Skip over the slab list */
1409 while (*str && *str != ';')
1412 /* Skip any completely empty blocks */
1413 while (*str && *str == ';')
1416 if (init && higher_order_disable)
1417 disable_higher_order_debug = 1;
1425 static int __init setup_slub_debug(char *str)
1430 bool global_slub_debug_changed = false;
1431 bool slab_list_specified = false;
1433 slub_debug = DEBUG_DEFAULT_FLAGS;
1434 if (*str++ != '=' || !*str)
1436 * No options specified. Switch on full debugging.
1442 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1446 global_slub_debug_changed = true;
1448 slab_list_specified = true;
1453 * For backwards compatibility, a single list of flags with list of
1454 * slabs means debugging is only enabled for those slabs, so the global
1455 * slub_debug should be 0. We can extended that to multiple lists as
1456 * long as there is no option specifying flags without a slab list.
1458 if (slab_list_specified) {
1459 if (!global_slub_debug_changed)
1461 slub_debug_string = saved_str;
1464 if (slub_debug != 0 || slub_debug_string)
1465 static_branch_enable(&slub_debug_enabled);
1467 static_branch_disable(&slub_debug_enabled);
1468 if ((static_branch_unlikely(&init_on_alloc) ||
1469 static_branch_unlikely(&init_on_free)) &&
1470 (slub_debug & SLAB_POISON))
1471 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1475 __setup("slub_debug", setup_slub_debug);
1478 * kmem_cache_flags - apply debugging options to the cache
1479 * @object_size: the size of an object without meta data
1480 * @flags: flags to set
1481 * @name: name of the cache
1483 * Debug option(s) are applied to @flags. In addition to the debug
1484 * option(s), if a slab name (or multiple) is specified i.e.
1485 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1486 * then only the select slabs will receive the debug option(s).
1488 slab_flags_t kmem_cache_flags(unsigned int object_size,
1489 slab_flags_t flags, const char *name)
1494 slab_flags_t block_flags;
1495 slab_flags_t slub_debug_local = slub_debug;
1498 * If the slab cache is for debugging (e.g. kmemleak) then
1499 * don't store user (stack trace) information by default,
1500 * but let the user enable it via the command line below.
1502 if (flags & SLAB_NOLEAKTRACE)
1503 slub_debug_local &= ~SLAB_STORE_USER;
1506 next_block = slub_debug_string;
1507 /* Go through all blocks of debug options, see if any matches our slab's name */
1508 while (next_block) {
1509 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1512 /* Found a block that has a slab list, search it */
1517 end = strchrnul(iter, ',');
1518 if (next_block && next_block < end)
1519 end = next_block - 1;
1521 glob = strnchr(iter, end - iter, '*');
1523 cmplen = glob - iter;
1525 cmplen = max_t(size_t, len, (end - iter));
1527 if (!strncmp(name, iter, cmplen)) {
1528 flags |= block_flags;
1532 if (!*end || *end == ';')
1538 return flags | slub_debug_local;
1540 #else /* !CONFIG_SLUB_DEBUG */
1541 static inline void setup_object_debug(struct kmem_cache *s,
1542 struct page *page, void *object) {}
1544 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1546 static inline int alloc_debug_processing(struct kmem_cache *s,
1547 struct page *page, void *object, unsigned long addr) { return 0; }
1549 static inline int free_debug_processing(
1550 struct kmem_cache *s, struct page *page,
1551 void *head, void *tail, int bulk_cnt,
1552 unsigned long addr) { return 0; }
1554 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1556 static inline int check_object(struct kmem_cache *s, struct page *page,
1557 void *object, u8 val) { return 1; }
1558 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1559 struct page *page) {}
1560 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1561 struct page *page) {}
1562 slab_flags_t kmem_cache_flags(unsigned int object_size,
1563 slab_flags_t flags, const char *name)
1567 #define slub_debug 0
1569 #define disable_higher_order_debug 0
1571 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1573 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1575 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1577 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1580 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1581 void **freelist, void *nextfree)
1585 #endif /* CONFIG_SLUB_DEBUG */
1588 * Hooks for other subsystems that check memory allocations. In a typical
1589 * production configuration these hooks all should produce no code at all.
1591 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1593 ptr = kasan_kmalloc_large(ptr, size, flags);
1594 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1595 kmemleak_alloc(ptr, size, 1, flags);
1599 static __always_inline void kfree_hook(void *x)
1602 kasan_kfree_large(x);
1605 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1608 kmemleak_free_recursive(x, s->flags);
1611 * Trouble is that we may no longer disable interrupts in the fast path
1612 * So in order to make the debug calls that expect irqs to be
1613 * disabled we need to disable interrupts temporarily.
1615 #ifdef CONFIG_LOCKDEP
1617 unsigned long flags;
1619 local_irq_save(flags);
1620 debug_check_no_locks_freed(x, s->object_size);
1621 local_irq_restore(flags);
1624 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1625 debug_check_no_obj_freed(x, s->object_size);
1627 /* Use KCSAN to help debug racy use-after-free. */
1628 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1629 __kcsan_check_access(x, s->object_size,
1630 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1633 * As memory initialization might be integrated into KASAN,
1634 * kasan_slab_free and initialization memset's must be
1635 * kept together to avoid discrepancies in behavior.
1637 * The initialization memset's clear the object and the metadata,
1638 * but don't touch the SLAB redzone.
1643 if (!kasan_has_integrated_init())
1644 memset(kasan_reset_tag(x), 0, s->object_size);
1645 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1646 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1647 s->size - s->inuse - rsize);
1649 /* KASAN might put x into memory quarantine, delaying its reuse. */
1650 return kasan_slab_free(s, x, init);
1653 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1654 void **head, void **tail)
1659 void *old_tail = *tail ? *tail : *head;
1661 if (is_kfence_address(next)) {
1662 slab_free_hook(s, next, false);
1666 /* Head and tail of the reconstructed freelist */
1672 next = get_freepointer(s, object);
1674 /* If object's reuse doesn't have to be delayed */
1675 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1676 /* Move object to the new freelist */
1677 set_freepointer(s, object, *head);
1682 } while (object != old_tail);
1687 return *head != NULL;
1690 static void *setup_object(struct kmem_cache *s, struct page *page,
1693 setup_object_debug(s, page, object);
1694 object = kasan_init_slab_obj(s, object);
1695 if (unlikely(s->ctor)) {
1696 kasan_unpoison_object_data(s, object);
1698 kasan_poison_object_data(s, object);
1704 * Slab allocation and freeing
1706 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1707 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1710 unsigned int order = oo_order(oo);
1712 if (node == NUMA_NO_NODE)
1713 page = alloc_pages(flags, order);
1715 page = __alloc_pages_node(node, flags, order);
1720 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1721 /* Pre-initialize the random sequence cache */
1722 static int init_cache_random_seq(struct kmem_cache *s)
1724 unsigned int count = oo_objects(s->oo);
1727 /* Bailout if already initialised */
1731 err = cache_random_seq_create(s, count, GFP_KERNEL);
1733 pr_err("SLUB: Unable to initialize free list for %s\n",
1738 /* Transform to an offset on the set of pages */
1739 if (s->random_seq) {
1742 for (i = 0; i < count; i++)
1743 s->random_seq[i] *= s->size;
1748 /* Initialize each random sequence freelist per cache */
1749 static void __init init_freelist_randomization(void)
1751 struct kmem_cache *s;
1753 mutex_lock(&slab_mutex);
1755 list_for_each_entry(s, &slab_caches, list)
1756 init_cache_random_seq(s);
1758 mutex_unlock(&slab_mutex);
1761 /* Get the next entry on the pre-computed freelist randomized */
1762 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1763 unsigned long *pos, void *start,
1764 unsigned long page_limit,
1765 unsigned long freelist_count)
1770 * If the target page allocation failed, the number of objects on the
1771 * page might be smaller than the usual size defined by the cache.
1774 idx = s->random_seq[*pos];
1776 if (*pos >= freelist_count)
1778 } while (unlikely(idx >= page_limit));
1780 return (char *)start + idx;
1783 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1784 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1789 unsigned long idx, pos, page_limit, freelist_count;
1791 if (page->objects < 2 || !s->random_seq)
1794 freelist_count = oo_objects(s->oo);
1795 pos = get_random_int() % freelist_count;
1797 page_limit = page->objects * s->size;
1798 start = fixup_red_left(s, page_address(page));
1800 /* First entry is used as the base of the freelist */
1801 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1803 cur = setup_object(s, page, cur);
1804 page->freelist = cur;
1806 for (idx = 1; idx < page->objects; idx++) {
1807 next = next_freelist_entry(s, page, &pos, start, page_limit,
1809 next = setup_object(s, page, next);
1810 set_freepointer(s, cur, next);
1813 set_freepointer(s, cur, NULL);
1818 static inline int init_cache_random_seq(struct kmem_cache *s)
1822 static inline void init_freelist_randomization(void) { }
1823 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1827 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1829 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1832 struct kmem_cache_order_objects oo = s->oo;
1834 void *start, *p, *next;
1838 flags &= gfp_allowed_mask;
1840 if (gfpflags_allow_blocking(flags))
1843 flags |= s->allocflags;
1846 * Let the initial higher-order allocation fail under memory pressure
1847 * so we fall-back to the minimum order allocation.
1849 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1850 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1851 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1853 page = alloc_slab_page(s, alloc_gfp, node, oo);
1854 if (unlikely(!page)) {
1858 * Allocation may have failed due to fragmentation.
1859 * Try a lower order alloc if possible
1861 page = alloc_slab_page(s, alloc_gfp, node, oo);
1862 if (unlikely(!page))
1864 stat(s, ORDER_FALLBACK);
1867 page->objects = oo_objects(oo);
1869 account_slab_page(page, oo_order(oo), s, flags);
1871 page->slab_cache = s;
1872 __SetPageSlab(page);
1873 if (page_is_pfmemalloc(page))
1874 SetPageSlabPfmemalloc(page);
1876 kasan_poison_slab(page);
1878 start = page_address(page);
1880 setup_page_debug(s, page, start);
1882 shuffle = shuffle_freelist(s, page);
1885 start = fixup_red_left(s, start);
1886 start = setup_object(s, page, start);
1887 page->freelist = start;
1888 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1890 next = setup_object(s, page, next);
1891 set_freepointer(s, p, next);
1894 set_freepointer(s, p, NULL);
1897 page->inuse = page->objects;
1901 if (gfpflags_allow_blocking(flags))
1902 local_irq_disable();
1906 inc_slabs_node(s, page_to_nid(page), page->objects);
1911 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1913 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1914 flags = kmalloc_fix_flags(flags);
1916 return allocate_slab(s,
1917 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1920 static void __free_slab(struct kmem_cache *s, struct page *page)
1922 int order = compound_order(page);
1923 int pages = 1 << order;
1925 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1928 slab_pad_check(s, page);
1929 for_each_object(p, s, page_address(page),
1931 check_object(s, page, p, SLUB_RED_INACTIVE);
1934 __ClearPageSlabPfmemalloc(page);
1935 __ClearPageSlab(page);
1936 /* In union with page->mapping where page allocator expects NULL */
1937 page->slab_cache = NULL;
1938 if (current->reclaim_state)
1939 current->reclaim_state->reclaimed_slab += pages;
1940 unaccount_slab_page(page, order, s);
1941 __free_pages(page, order);
1944 static void rcu_free_slab(struct rcu_head *h)
1946 struct page *page = container_of(h, struct page, rcu_head);
1948 __free_slab(page->slab_cache, page);
1951 static void free_slab(struct kmem_cache *s, struct page *page)
1953 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1954 call_rcu(&page->rcu_head, rcu_free_slab);
1956 __free_slab(s, page);
1959 static void discard_slab(struct kmem_cache *s, struct page *page)
1961 dec_slabs_node(s, page_to_nid(page), page->objects);
1966 * Management of partially allocated slabs.
1969 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1972 if (tail == DEACTIVATE_TO_TAIL)
1973 list_add_tail(&page->slab_list, &n->partial);
1975 list_add(&page->slab_list, &n->partial);
1978 static inline void add_partial(struct kmem_cache_node *n,
1979 struct page *page, int tail)
1981 lockdep_assert_held(&n->list_lock);
1982 __add_partial(n, page, tail);
1985 static inline void remove_partial(struct kmem_cache_node *n,
1988 lockdep_assert_held(&n->list_lock);
1989 list_del(&page->slab_list);
1994 * Remove slab from the partial list, freeze it and
1995 * return the pointer to the freelist.
1997 * Returns a list of objects or NULL if it fails.
1999 static inline void *acquire_slab(struct kmem_cache *s,
2000 struct kmem_cache_node *n, struct page *page,
2001 int mode, int *objects)
2004 unsigned long counters;
2007 lockdep_assert_held(&n->list_lock);
2010 * Zap the freelist and set the frozen bit.
2011 * The old freelist is the list of objects for the
2012 * per cpu allocation list.
2014 freelist = page->freelist;
2015 counters = page->counters;
2016 new.counters = counters;
2017 *objects = new.objects - new.inuse;
2019 new.inuse = page->objects;
2020 new.freelist = NULL;
2022 new.freelist = freelist;
2025 VM_BUG_ON(new.frozen);
2028 if (!__cmpxchg_double_slab(s, page,
2030 new.freelist, new.counters,
2034 remove_partial(n, page);
2039 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
2040 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
2043 * Try to allocate a partial slab from a specific node.
2045 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2046 struct kmem_cache_cpu *c, gfp_t flags)
2048 struct page *page, *page2;
2049 void *object = NULL;
2050 unsigned int available = 0;
2054 * Racy check. If we mistakenly see no partial slabs then we
2055 * just allocate an empty slab. If we mistakenly try to get a
2056 * partial slab and there is none available then get_partial()
2059 if (!n || !n->nr_partial)
2062 spin_lock(&n->list_lock);
2063 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
2066 if (!pfmemalloc_match(page, flags))
2069 t = acquire_slab(s, n, page, object == NULL, &objects);
2073 available += objects;
2076 stat(s, ALLOC_FROM_PARTIAL);
2079 put_cpu_partial(s, page, 0);
2080 stat(s, CPU_PARTIAL_NODE);
2082 if (!kmem_cache_has_cpu_partial(s)
2083 || available > slub_cpu_partial(s) / 2)
2087 spin_unlock(&n->list_lock);
2092 * Get a page from somewhere. Search in increasing NUMA distances.
2094 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2095 struct kmem_cache_cpu *c)
2098 struct zonelist *zonelist;
2101 enum zone_type highest_zoneidx = gfp_zone(flags);
2103 unsigned int cpuset_mems_cookie;
2106 * The defrag ratio allows a configuration of the tradeoffs between
2107 * inter node defragmentation and node local allocations. A lower
2108 * defrag_ratio increases the tendency to do local allocations
2109 * instead of attempting to obtain partial slabs from other nodes.
2111 * If the defrag_ratio is set to 0 then kmalloc() always
2112 * returns node local objects. If the ratio is higher then kmalloc()
2113 * may return off node objects because partial slabs are obtained
2114 * from other nodes and filled up.
2116 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2117 * (which makes defrag_ratio = 1000) then every (well almost)
2118 * allocation will first attempt to defrag slab caches on other nodes.
2119 * This means scanning over all nodes to look for partial slabs which
2120 * may be expensive if we do it every time we are trying to find a slab
2121 * with available objects.
2123 if (!s->remote_node_defrag_ratio ||
2124 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2128 cpuset_mems_cookie = read_mems_allowed_begin();
2129 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2130 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2131 struct kmem_cache_node *n;
2133 n = get_node(s, zone_to_nid(zone));
2135 if (n && cpuset_zone_allowed(zone, flags) &&
2136 n->nr_partial > s->min_partial) {
2137 object = get_partial_node(s, n, c, flags);
2140 * Don't check read_mems_allowed_retry()
2141 * here - if mems_allowed was updated in
2142 * parallel, that was a harmless race
2143 * between allocation and the cpuset
2150 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2151 #endif /* CONFIG_NUMA */
2156 * Get a partial page, lock it and return it.
2158 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2159 struct kmem_cache_cpu *c)
2162 int searchnode = node;
2164 if (node == NUMA_NO_NODE)
2165 searchnode = numa_mem_id();
2167 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2168 if (object || node != NUMA_NO_NODE)
2171 return get_any_partial(s, flags, c);
2174 #ifdef CONFIG_PREEMPTION
2176 * Calculate the next globally unique transaction for disambiguation
2177 * during cmpxchg. The transactions start with the cpu number and are then
2178 * incremented by CONFIG_NR_CPUS.
2180 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2183 * No preemption supported therefore also no need to check for
2189 static inline unsigned long next_tid(unsigned long tid)
2191 return tid + TID_STEP;
2194 #ifdef SLUB_DEBUG_CMPXCHG
2195 static inline unsigned int tid_to_cpu(unsigned long tid)
2197 return tid % TID_STEP;
2200 static inline unsigned long tid_to_event(unsigned long tid)
2202 return tid / TID_STEP;
2206 static inline unsigned int init_tid(int cpu)
2211 static inline void note_cmpxchg_failure(const char *n,
2212 const struct kmem_cache *s, unsigned long tid)
2214 #ifdef SLUB_DEBUG_CMPXCHG
2215 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2217 pr_info("%s %s: cmpxchg redo ", n, s->name);
2219 #ifdef CONFIG_PREEMPTION
2220 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2221 pr_warn("due to cpu change %d -> %d\n",
2222 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2225 if (tid_to_event(tid) != tid_to_event(actual_tid))
2226 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2227 tid_to_event(tid), tid_to_event(actual_tid));
2229 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2230 actual_tid, tid, next_tid(tid));
2232 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2235 static void init_kmem_cache_cpus(struct kmem_cache *s)
2239 for_each_possible_cpu(cpu)
2240 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2244 * Remove the cpu slab
2246 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2247 void *freelist, struct kmem_cache_cpu *c)
2249 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2250 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2251 int lock = 0, free_delta = 0;
2252 enum slab_modes l = M_NONE, m = M_NONE;
2253 void *nextfree, *freelist_iter, *freelist_tail;
2254 int tail = DEACTIVATE_TO_HEAD;
2258 if (page->freelist) {
2259 stat(s, DEACTIVATE_REMOTE_FREES);
2260 tail = DEACTIVATE_TO_TAIL;
2264 * Stage one: Count the objects on cpu's freelist as free_delta and
2265 * remember the last object in freelist_tail for later splicing.
2267 freelist_tail = NULL;
2268 freelist_iter = freelist;
2269 while (freelist_iter) {
2270 nextfree = get_freepointer(s, freelist_iter);
2273 * If 'nextfree' is invalid, it is possible that the object at
2274 * 'freelist_iter' is already corrupted. So isolate all objects
2275 * starting at 'freelist_iter' by skipping them.
2277 if (freelist_corrupted(s, page, &freelist_iter, nextfree))
2280 freelist_tail = freelist_iter;
2283 freelist_iter = nextfree;
2287 * Stage two: Unfreeze the page while splicing the per-cpu
2288 * freelist to the head of page's freelist.
2290 * Ensure that the page is unfrozen while the list presence
2291 * reflects the actual number of objects during unfreeze.
2293 * We setup the list membership and then perform a cmpxchg
2294 * with the count. If there is a mismatch then the page
2295 * is not unfrozen but the page is on the wrong list.
2297 * Then we restart the process which may have to remove
2298 * the page from the list that we just put it on again
2299 * because the number of objects in the slab may have
2304 old.freelist = READ_ONCE(page->freelist);
2305 old.counters = READ_ONCE(page->counters);
2306 VM_BUG_ON(!old.frozen);
2308 /* Determine target state of the slab */
2309 new.counters = old.counters;
2310 if (freelist_tail) {
2311 new.inuse -= free_delta;
2312 set_freepointer(s, freelist_tail, old.freelist);
2313 new.freelist = freelist;
2315 new.freelist = old.freelist;
2319 if (!new.inuse && n->nr_partial >= s->min_partial)
2321 else if (new.freelist) {
2326 * Taking the spinlock removes the possibility
2327 * that acquire_slab() will see a slab page that
2330 spin_lock(&n->list_lock);
2334 if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) {
2337 * This also ensures that the scanning of full
2338 * slabs from diagnostic functions will not see
2341 spin_lock(&n->list_lock);
2347 remove_partial(n, page);
2348 else if (l == M_FULL)
2349 remove_full(s, n, page);
2352 add_partial(n, page, tail);
2353 else if (m == M_FULL)
2354 add_full(s, n, page);
2358 if (!__cmpxchg_double_slab(s, page,
2359 old.freelist, old.counters,
2360 new.freelist, new.counters,
2365 spin_unlock(&n->list_lock);
2369 else if (m == M_FULL)
2370 stat(s, DEACTIVATE_FULL);
2371 else if (m == M_FREE) {
2372 stat(s, DEACTIVATE_EMPTY);
2373 discard_slab(s, page);
2382 * Unfreeze all the cpu partial slabs.
2384 * This function must be called with interrupts disabled
2385 * for the cpu using c (or some other guarantee must be there
2386 * to guarantee no concurrent accesses).
2388 static void unfreeze_partials(struct kmem_cache *s,
2389 struct kmem_cache_cpu *c)
2391 #ifdef CONFIG_SLUB_CPU_PARTIAL
2392 struct kmem_cache_node *n = NULL, *n2 = NULL;
2393 struct page *page, *discard_page = NULL;
2395 while ((page = slub_percpu_partial(c))) {
2399 slub_set_percpu_partial(c, page);
2401 n2 = get_node(s, page_to_nid(page));
2404 spin_unlock(&n->list_lock);
2407 spin_lock(&n->list_lock);
2412 old.freelist = page->freelist;
2413 old.counters = page->counters;
2414 VM_BUG_ON(!old.frozen);
2416 new.counters = old.counters;
2417 new.freelist = old.freelist;
2421 } while (!__cmpxchg_double_slab(s, page,
2422 old.freelist, old.counters,
2423 new.freelist, new.counters,
2424 "unfreezing slab"));
2426 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2427 page->next = discard_page;
2428 discard_page = page;
2430 add_partial(n, page, DEACTIVATE_TO_TAIL);
2431 stat(s, FREE_ADD_PARTIAL);
2436 spin_unlock(&n->list_lock);
2438 while (discard_page) {
2439 page = discard_page;
2440 discard_page = discard_page->next;
2442 stat(s, DEACTIVATE_EMPTY);
2443 discard_slab(s, page);
2446 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2450 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2451 * partial page slot if available.
2453 * If we did not find a slot then simply move all the partials to the
2454 * per node partial list.
2456 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2458 #ifdef CONFIG_SLUB_CPU_PARTIAL
2459 struct page *oldpage;
2467 oldpage = this_cpu_read(s->cpu_slab->partial);
2470 pobjects = oldpage->pobjects;
2471 pages = oldpage->pages;
2472 if (drain && pobjects > slub_cpu_partial(s)) {
2473 unsigned long flags;
2475 * partial array is full. Move the existing
2476 * set to the per node partial list.
2478 local_irq_save(flags);
2479 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2480 local_irq_restore(flags);
2484 stat(s, CPU_PARTIAL_DRAIN);
2489 pobjects += page->objects - page->inuse;
2491 page->pages = pages;
2492 page->pobjects = pobjects;
2493 page->next = oldpage;
2495 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2497 if (unlikely(!slub_cpu_partial(s))) {
2498 unsigned long flags;
2500 local_irq_save(flags);
2501 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2502 local_irq_restore(flags);
2505 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2508 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2510 stat(s, CPUSLAB_FLUSH);
2511 deactivate_slab(s, c->page, c->freelist, c);
2513 c->tid = next_tid(c->tid);
2519 * Called from IPI handler with interrupts disabled.
2521 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2523 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2528 unfreeze_partials(s, c);
2531 static void flush_cpu_slab(void *d)
2533 struct kmem_cache *s = d;
2535 __flush_cpu_slab(s, smp_processor_id());
2538 static bool has_cpu_slab(int cpu, void *info)
2540 struct kmem_cache *s = info;
2541 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2543 return c->page || slub_percpu_partial(c);
2546 static void flush_all(struct kmem_cache *s)
2548 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2552 * Use the cpu notifier to insure that the cpu slabs are flushed when
2555 static int slub_cpu_dead(unsigned int cpu)
2557 struct kmem_cache *s;
2558 unsigned long flags;
2560 mutex_lock(&slab_mutex);
2561 list_for_each_entry(s, &slab_caches, list) {
2562 local_irq_save(flags);
2563 __flush_cpu_slab(s, cpu);
2564 local_irq_restore(flags);
2566 mutex_unlock(&slab_mutex);
2571 * Check if the objects in a per cpu structure fit numa
2572 * locality expectations.
2574 static inline int node_match(struct page *page, int node)
2577 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2583 #ifdef CONFIG_SLUB_DEBUG
2584 static int count_free(struct page *page)
2586 return page->objects - page->inuse;
2589 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2591 return atomic_long_read(&n->total_objects);
2593 #endif /* CONFIG_SLUB_DEBUG */
2595 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2596 static unsigned long count_partial(struct kmem_cache_node *n,
2597 int (*get_count)(struct page *))
2599 unsigned long flags;
2600 unsigned long x = 0;
2603 spin_lock_irqsave(&n->list_lock, flags);
2604 list_for_each_entry(page, &n->partial, slab_list)
2605 x += get_count(page);
2606 spin_unlock_irqrestore(&n->list_lock, flags);
2609 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2611 static noinline void
2612 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2614 #ifdef CONFIG_SLUB_DEBUG
2615 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2616 DEFAULT_RATELIMIT_BURST);
2618 struct kmem_cache_node *n;
2620 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2623 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2624 nid, gfpflags, &gfpflags);
2625 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2626 s->name, s->object_size, s->size, oo_order(s->oo),
2629 if (oo_order(s->min) > get_order(s->object_size))
2630 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2633 for_each_kmem_cache_node(s, node, n) {
2634 unsigned long nr_slabs;
2635 unsigned long nr_objs;
2636 unsigned long nr_free;
2638 nr_free = count_partial(n, count_free);
2639 nr_slabs = node_nr_slabs(n);
2640 nr_objs = node_nr_objs(n);
2642 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2643 node, nr_slabs, nr_objs, nr_free);
2648 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2649 int node, struct kmem_cache_cpu **pc)
2652 struct kmem_cache_cpu *c = *pc;
2655 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2657 freelist = get_partial(s, flags, node, c);
2662 page = new_slab(s, flags, node);
2664 c = raw_cpu_ptr(s->cpu_slab);
2669 * No other reference to the page yet so we can
2670 * muck around with it freely without cmpxchg
2672 freelist = page->freelist;
2673 page->freelist = NULL;
2675 stat(s, ALLOC_SLAB);
2683 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2685 if (unlikely(PageSlabPfmemalloc(page)))
2686 return gfp_pfmemalloc_allowed(gfpflags);
2692 * Check the page->freelist of a page and either transfer the freelist to the
2693 * per cpu freelist or deactivate the page.
2695 * The page is still frozen if the return value is not NULL.
2697 * If this function returns NULL then the page has been unfrozen.
2699 * This function must be called with interrupt disabled.
2701 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2704 unsigned long counters;
2708 freelist = page->freelist;
2709 counters = page->counters;
2711 new.counters = counters;
2712 VM_BUG_ON(!new.frozen);
2714 new.inuse = page->objects;
2715 new.frozen = freelist != NULL;
2717 } while (!__cmpxchg_double_slab(s, page,
2726 * Slow path. The lockless freelist is empty or we need to perform
2729 * Processing is still very fast if new objects have been freed to the
2730 * regular freelist. In that case we simply take over the regular freelist
2731 * as the lockless freelist and zap the regular freelist.
2733 * If that is not working then we fall back to the partial lists. We take the
2734 * first element of the freelist as the object to allocate now and move the
2735 * rest of the freelist to the lockless freelist.
2737 * And if we were unable to get a new slab from the partial slab lists then
2738 * we need to allocate a new slab. This is the slowest path since it involves
2739 * a call to the page allocator and the setup of a new slab.
2741 * Version of __slab_alloc to use when we know that interrupts are
2742 * already disabled (which is the case for bulk allocation).
2744 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2745 unsigned long addr, struct kmem_cache_cpu *c)
2750 stat(s, ALLOC_SLOWPATH);
2755 * if the node is not online or has no normal memory, just
2756 * ignore the node constraint
2758 if (unlikely(node != NUMA_NO_NODE &&
2759 !node_isset(node, slab_nodes)))
2760 node = NUMA_NO_NODE;
2765 if (unlikely(!node_match(page, node))) {
2767 * same as above but node_match() being false already
2768 * implies node != NUMA_NO_NODE
2770 if (!node_isset(node, slab_nodes)) {
2771 node = NUMA_NO_NODE;
2774 stat(s, ALLOC_NODE_MISMATCH);
2775 deactivate_slab(s, page, c->freelist, c);
2781 * By rights, we should be searching for a slab page that was
2782 * PFMEMALLOC but right now, we are losing the pfmemalloc
2783 * information when the page leaves the per-cpu allocator
2785 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2786 deactivate_slab(s, page, c->freelist, c);
2790 /* must check again c->freelist in case of cpu migration or IRQ */
2791 freelist = c->freelist;
2795 freelist = get_freelist(s, page);
2799 stat(s, DEACTIVATE_BYPASS);
2803 stat(s, ALLOC_REFILL);
2807 * freelist is pointing to the list of objects to be used.
2808 * page is pointing to the page from which the objects are obtained.
2809 * That page must be frozen for per cpu allocations to work.
2811 VM_BUG_ON(!c->page->frozen);
2812 c->freelist = get_freepointer(s, freelist);
2813 c->tid = next_tid(c->tid);
2818 if (slub_percpu_partial(c)) {
2819 page = c->page = slub_percpu_partial(c);
2820 slub_set_percpu_partial(c, page);
2821 stat(s, CPU_PARTIAL_ALLOC);
2825 freelist = new_slab_objects(s, gfpflags, node, &c);
2827 if (unlikely(!freelist)) {
2828 slab_out_of_memory(s, gfpflags, node);
2833 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2836 /* Only entered in the debug case */
2837 if (kmem_cache_debug(s) &&
2838 !alloc_debug_processing(s, page, freelist, addr))
2839 goto new_slab; /* Slab failed checks. Next slab needed */
2841 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2846 * Another one that disabled interrupt and compensates for possible
2847 * cpu changes by refetching the per cpu area pointer.
2849 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2850 unsigned long addr, struct kmem_cache_cpu *c)
2853 unsigned long flags;
2855 local_irq_save(flags);
2856 #ifdef CONFIG_PREEMPTION
2858 * We may have been preempted and rescheduled on a different
2859 * cpu before disabling interrupts. Need to reload cpu area
2862 c = this_cpu_ptr(s->cpu_slab);
2865 p = ___slab_alloc(s, gfpflags, node, addr, c);
2866 local_irq_restore(flags);
2871 * If the object has been wiped upon free, make sure it's fully initialized by
2872 * zeroing out freelist pointer.
2874 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2877 if (unlikely(slab_want_init_on_free(s)) && obj)
2878 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
2883 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2884 * have the fastpath folded into their functions. So no function call
2885 * overhead for requests that can be satisfied on the fastpath.
2887 * The fastpath works by first checking if the lockless freelist can be used.
2888 * If not then __slab_alloc is called for slow processing.
2890 * Otherwise we can simply pick the next object from the lockless free list.
2892 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2893 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
2896 struct kmem_cache_cpu *c;
2899 struct obj_cgroup *objcg = NULL;
2902 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2906 object = kfence_alloc(s, orig_size, gfpflags);
2907 if (unlikely(object))
2912 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2913 * enabled. We may switch back and forth between cpus while
2914 * reading from one cpu area. That does not matter as long
2915 * as we end up on the original cpu again when doing the cmpxchg.
2917 * We should guarantee that tid and kmem_cache are retrieved on
2918 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2919 * to check if it is matched or not.
2922 tid = this_cpu_read(s->cpu_slab->tid);
2923 c = raw_cpu_ptr(s->cpu_slab);
2924 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2925 unlikely(tid != READ_ONCE(c->tid)));
2928 * Irqless object alloc/free algorithm used here depends on sequence
2929 * of fetching cpu_slab's data. tid should be fetched before anything
2930 * on c to guarantee that object and page associated with previous tid
2931 * won't be used with current tid. If we fetch tid first, object and
2932 * page could be one associated with next tid and our alloc/free
2933 * request will be failed. In this case, we will retry. So, no problem.
2938 * The transaction ids are globally unique per cpu and per operation on
2939 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2940 * occurs on the right processor and that there was no operation on the
2941 * linked list in between.
2944 object = c->freelist;
2946 if (unlikely(!object || !page || !node_match(page, node))) {
2947 object = __slab_alloc(s, gfpflags, node, addr, c);
2949 void *next_object = get_freepointer_safe(s, object);
2952 * The cmpxchg will only match if there was no additional
2953 * operation and if we are on the right processor.
2955 * The cmpxchg does the following atomically (without lock
2957 * 1. Relocate first pointer to the current per cpu area.
2958 * 2. Verify that tid and freelist have not been changed
2959 * 3. If they were not changed replace tid and freelist
2961 * Since this is without lock semantics the protection is only
2962 * against code executing on this cpu *not* from access by
2965 if (unlikely(!this_cpu_cmpxchg_double(
2966 s->cpu_slab->freelist, s->cpu_slab->tid,
2968 next_object, next_tid(tid)))) {
2970 note_cmpxchg_failure("slab_alloc", s, tid);
2973 prefetch_freepointer(s, next_object);
2974 stat(s, ALLOC_FASTPATH);
2977 maybe_wipe_obj_freeptr(s, object);
2978 init = slab_want_init_on_alloc(gfpflags, s);
2981 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
2986 static __always_inline void *slab_alloc(struct kmem_cache *s,
2987 gfp_t gfpflags, unsigned long addr, size_t orig_size)
2989 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size);
2992 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2994 void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size);
2996 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
3001 EXPORT_SYMBOL(kmem_cache_alloc);
3003 #ifdef CONFIG_TRACING
3004 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3006 void *ret = slab_alloc(s, gfpflags, _RET_IP_, size);
3007 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
3008 ret = kasan_kmalloc(s, ret, size, gfpflags);
3011 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3015 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3017 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size);
3019 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3020 s->object_size, s->size, gfpflags, node);
3024 EXPORT_SYMBOL(kmem_cache_alloc_node);
3026 #ifdef CONFIG_TRACING
3027 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3029 int node, size_t size)
3031 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size);
3033 trace_kmalloc_node(_RET_IP_, ret,
3034 size, s->size, gfpflags, node);
3036 ret = kasan_kmalloc(s, ret, size, gfpflags);
3039 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3041 #endif /* CONFIG_NUMA */
3044 * Slow path handling. This may still be called frequently since objects
3045 * have a longer lifetime than the cpu slabs in most processing loads.
3047 * So we still attempt to reduce cache line usage. Just take the slab
3048 * lock and free the item. If there is no additional partial page
3049 * handling required then we can return immediately.
3051 static void __slab_free(struct kmem_cache *s, struct page *page,
3052 void *head, void *tail, int cnt,
3059 unsigned long counters;
3060 struct kmem_cache_node *n = NULL;
3061 unsigned long flags;
3063 stat(s, FREE_SLOWPATH);
3065 if (kfence_free(head))
3068 if (kmem_cache_debug(s) &&
3069 !free_debug_processing(s, page, head, tail, cnt, addr))
3074 spin_unlock_irqrestore(&n->list_lock, flags);
3077 prior = page->freelist;
3078 counters = page->counters;
3079 set_freepointer(s, tail, prior);
3080 new.counters = counters;
3081 was_frozen = new.frozen;
3083 if ((!new.inuse || !prior) && !was_frozen) {
3085 if (kmem_cache_has_cpu_partial(s) && !prior) {
3088 * Slab was on no list before and will be
3090 * We can defer the list move and instead
3095 } else { /* Needs to be taken off a list */
3097 n = get_node(s, page_to_nid(page));
3099 * Speculatively acquire the list_lock.
3100 * If the cmpxchg does not succeed then we may
3101 * drop the list_lock without any processing.
3103 * Otherwise the list_lock will synchronize with
3104 * other processors updating the list of slabs.
3106 spin_lock_irqsave(&n->list_lock, flags);
3111 } while (!cmpxchg_double_slab(s, page,
3118 if (likely(was_frozen)) {
3120 * The list lock was not taken therefore no list
3121 * activity can be necessary.
3123 stat(s, FREE_FROZEN);
3124 } else if (new.frozen) {
3126 * If we just froze the page then put it onto the
3127 * per cpu partial list.
3129 put_cpu_partial(s, page, 1);
3130 stat(s, CPU_PARTIAL_FREE);
3136 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3140 * Objects left in the slab. If it was not on the partial list before
3143 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3144 remove_full(s, n, page);
3145 add_partial(n, page, DEACTIVATE_TO_TAIL);
3146 stat(s, FREE_ADD_PARTIAL);
3148 spin_unlock_irqrestore(&n->list_lock, flags);
3154 * Slab on the partial list.
3156 remove_partial(n, page);
3157 stat(s, FREE_REMOVE_PARTIAL);
3159 /* Slab must be on the full list */
3160 remove_full(s, n, page);
3163 spin_unlock_irqrestore(&n->list_lock, flags);
3165 discard_slab(s, page);
3169 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3170 * can perform fastpath freeing without additional function calls.
3172 * The fastpath is only possible if we are freeing to the current cpu slab
3173 * of this processor. This typically the case if we have just allocated
3176 * If fastpath is not possible then fall back to __slab_free where we deal
3177 * with all sorts of special processing.
3179 * Bulk free of a freelist with several objects (all pointing to the
3180 * same page) possible by specifying head and tail ptr, plus objects
3181 * count (cnt). Bulk free indicated by tail pointer being set.
3183 static __always_inline void do_slab_free(struct kmem_cache *s,
3184 struct page *page, void *head, void *tail,
3185 int cnt, unsigned long addr)
3187 void *tail_obj = tail ? : head;
3188 struct kmem_cache_cpu *c;
3191 memcg_slab_free_hook(s, &head, 1);
3194 * Determine the currently cpus per cpu slab.
3195 * The cpu may change afterward. However that does not matter since
3196 * data is retrieved via this pointer. If we are on the same cpu
3197 * during the cmpxchg then the free will succeed.
3200 tid = this_cpu_read(s->cpu_slab->tid);
3201 c = raw_cpu_ptr(s->cpu_slab);
3202 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3203 unlikely(tid != READ_ONCE(c->tid)));
3205 /* Same with comment on barrier() in slab_alloc_node() */
3208 if (likely(page == c->page)) {
3209 void **freelist = READ_ONCE(c->freelist);
3211 set_freepointer(s, tail_obj, freelist);
3213 if (unlikely(!this_cpu_cmpxchg_double(
3214 s->cpu_slab->freelist, s->cpu_slab->tid,
3216 head, next_tid(tid)))) {
3218 note_cmpxchg_failure("slab_free", s, tid);
3221 stat(s, FREE_FASTPATH);
3223 __slab_free(s, page, head, tail_obj, cnt, addr);
3227 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3228 void *head, void *tail, int cnt,
3232 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3233 * to remove objects, whose reuse must be delayed.
3235 if (slab_free_freelist_hook(s, &head, &tail))
3236 do_slab_free(s, page, head, tail, cnt, addr);
3239 #ifdef CONFIG_KASAN_GENERIC
3240 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3242 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3246 void kmem_cache_free(struct kmem_cache *s, void *x)
3248 s = cache_from_obj(s, x);
3251 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3252 trace_kmem_cache_free(_RET_IP_, x, s->name);
3254 EXPORT_SYMBOL(kmem_cache_free);
3256 struct detached_freelist {
3261 struct kmem_cache *s;
3265 * This function progressively scans the array with free objects (with
3266 * a limited look ahead) and extract objects belonging to the same
3267 * page. It builds a detached freelist directly within the given
3268 * page/objects. This can happen without any need for
3269 * synchronization, because the objects are owned by running process.
3270 * The freelist is build up as a single linked list in the objects.
3271 * The idea is, that this detached freelist can then be bulk
3272 * transferred to the real freelist(s), but only requiring a single
3273 * synchronization primitive. Look ahead in the array is limited due
3274 * to performance reasons.
3277 int build_detached_freelist(struct kmem_cache *s, size_t size,
3278 void **p, struct detached_freelist *df)
3280 size_t first_skipped_index = 0;
3285 /* Always re-init detached_freelist */
3290 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3291 } while (!object && size);
3296 page = virt_to_head_page(object);
3298 /* Handle kalloc'ed objects */
3299 if (unlikely(!PageSlab(page))) {
3300 BUG_ON(!PageCompound(page));
3302 __free_pages(page, compound_order(page));
3303 p[size] = NULL; /* mark object processed */
3306 /* Derive kmem_cache from object */
3307 df->s = page->slab_cache;
3309 df->s = cache_from_obj(s, object); /* Support for memcg */
3312 if (is_kfence_address(object)) {
3313 slab_free_hook(df->s, object, false);
3314 __kfence_free(object);
3315 p[size] = NULL; /* mark object processed */
3319 /* Start new detached freelist */
3321 set_freepointer(df->s, object, NULL);
3323 df->freelist = object;
3324 p[size] = NULL; /* mark object processed */
3330 continue; /* Skip processed objects */
3332 /* df->page is always set at this point */
3333 if (df->page == virt_to_head_page(object)) {
3334 /* Opportunity build freelist */
3335 set_freepointer(df->s, object, df->freelist);
3336 df->freelist = object;
3338 p[size] = NULL; /* mark object processed */
3343 /* Limit look ahead search */
3347 if (!first_skipped_index)
3348 first_skipped_index = size + 1;
3351 return first_skipped_index;
3354 /* Note that interrupts must be enabled when calling this function. */
3355 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3360 memcg_slab_free_hook(s, p, size);
3362 struct detached_freelist df;
3364 size = build_detached_freelist(s, size, p, &df);
3368 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
3369 } while (likely(size));
3371 EXPORT_SYMBOL(kmem_cache_free_bulk);
3373 /* Note that interrupts must be enabled when calling this function. */
3374 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3377 struct kmem_cache_cpu *c;
3379 struct obj_cgroup *objcg = NULL;
3381 /* memcg and kmem_cache debug support */
3382 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3386 * Drain objects in the per cpu slab, while disabling local
3387 * IRQs, which protects against PREEMPT and interrupts
3388 * handlers invoking normal fastpath.
3390 local_irq_disable();
3391 c = this_cpu_ptr(s->cpu_slab);
3393 for (i = 0; i < size; i++) {
3394 void *object = kfence_alloc(s, s->object_size, flags);
3396 if (unlikely(object)) {
3401 object = c->freelist;
3402 if (unlikely(!object)) {
3404 * We may have removed an object from c->freelist using
3405 * the fastpath in the previous iteration; in that case,
3406 * c->tid has not been bumped yet.
3407 * Since ___slab_alloc() may reenable interrupts while
3408 * allocating memory, we should bump c->tid now.
3410 c->tid = next_tid(c->tid);
3413 * Invoking slow path likely have side-effect
3414 * of re-populating per CPU c->freelist
3416 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3418 if (unlikely(!p[i]))
3421 c = this_cpu_ptr(s->cpu_slab);
3422 maybe_wipe_obj_freeptr(s, p[i]);
3424 continue; /* goto for-loop */
3426 c->freelist = get_freepointer(s, object);
3428 maybe_wipe_obj_freeptr(s, p[i]);
3430 c->tid = next_tid(c->tid);
3434 * memcg and kmem_cache debug support and memory initialization.
3435 * Done outside of the IRQ disabled fastpath loop.
3437 slab_post_alloc_hook(s, objcg, flags, size, p,
3438 slab_want_init_on_alloc(flags, s));
3442 slab_post_alloc_hook(s, objcg, flags, i, p, false);
3443 __kmem_cache_free_bulk(s, i, p);
3446 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3450 * Object placement in a slab is made very easy because we always start at
3451 * offset 0. If we tune the size of the object to the alignment then we can
3452 * get the required alignment by putting one properly sized object after
3455 * Notice that the allocation order determines the sizes of the per cpu
3456 * caches. Each processor has always one slab available for allocations.
3457 * Increasing the allocation order reduces the number of times that slabs
3458 * must be moved on and off the partial lists and is therefore a factor in
3463 * Minimum / Maximum order of slab pages. This influences locking overhead
3464 * and slab fragmentation. A higher order reduces the number of partial slabs
3465 * and increases the number of allocations possible without having to
3466 * take the list_lock.
3468 static unsigned int slub_min_order;
3469 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3470 static unsigned int slub_min_objects;
3473 * Calculate the order of allocation given an slab object size.
3475 * The order of allocation has significant impact on performance and other
3476 * system components. Generally order 0 allocations should be preferred since
3477 * order 0 does not cause fragmentation in the page allocator. Larger objects
3478 * be problematic to put into order 0 slabs because there may be too much
3479 * unused space left. We go to a higher order if more than 1/16th of the slab
3482 * In order to reach satisfactory performance we must ensure that a minimum
3483 * number of objects is in one slab. Otherwise we may generate too much
3484 * activity on the partial lists which requires taking the list_lock. This is
3485 * less a concern for large slabs though which are rarely used.
3487 * slub_max_order specifies the order where we begin to stop considering the
3488 * number of objects in a slab as critical. If we reach slub_max_order then
3489 * we try to keep the page order as low as possible. So we accept more waste
3490 * of space in favor of a small page order.
3492 * Higher order allocations also allow the placement of more objects in a
3493 * slab and thereby reduce object handling overhead. If the user has
3494 * requested a higher minimum order then we start with that one instead of
3495 * the smallest order which will fit the object.
3497 static inline unsigned int slab_order(unsigned int size,
3498 unsigned int min_objects, unsigned int max_order,
3499 unsigned int fract_leftover)
3501 unsigned int min_order = slub_min_order;
3504 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3505 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3507 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3508 order <= max_order; order++) {
3510 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3513 rem = slab_size % size;
3515 if (rem <= slab_size / fract_leftover)
3522 static inline int calculate_order(unsigned int size)
3525 unsigned int min_objects;
3526 unsigned int max_objects;
3527 unsigned int nr_cpus;
3530 * Attempt to find best configuration for a slab. This
3531 * works by first attempting to generate a layout with
3532 * the best configuration and backing off gradually.
3534 * First we increase the acceptable waste in a slab. Then
3535 * we reduce the minimum objects required in a slab.
3537 min_objects = slub_min_objects;
3540 * Some architectures will only update present cpus when
3541 * onlining them, so don't trust the number if it's just 1. But
3542 * we also don't want to use nr_cpu_ids always, as on some other
3543 * architectures, there can be many possible cpus, but never
3544 * onlined. Here we compromise between trying to avoid too high
3545 * order on systems that appear larger than they are, and too
3546 * low order on systems that appear smaller than they are.
3548 nr_cpus = num_present_cpus();
3550 nr_cpus = nr_cpu_ids;
3551 min_objects = 4 * (fls(nr_cpus) + 1);
3553 max_objects = order_objects(slub_max_order, size);
3554 min_objects = min(min_objects, max_objects);
3556 while (min_objects > 1) {
3557 unsigned int fraction;
3560 while (fraction >= 4) {
3561 order = slab_order(size, min_objects,
3562 slub_max_order, fraction);
3563 if (order <= slub_max_order)
3571 * We were unable to place multiple objects in a slab. Now
3572 * lets see if we can place a single object there.
3574 order = slab_order(size, 1, slub_max_order, 1);
3575 if (order <= slub_max_order)
3579 * Doh this slab cannot be placed using slub_max_order.
3581 order = slab_order(size, 1, MAX_ORDER, 1);
3582 if (order < MAX_ORDER)
3588 init_kmem_cache_node(struct kmem_cache_node *n)
3591 spin_lock_init(&n->list_lock);
3592 INIT_LIST_HEAD(&n->partial);
3593 #ifdef CONFIG_SLUB_DEBUG
3594 atomic_long_set(&n->nr_slabs, 0);
3595 atomic_long_set(&n->total_objects, 0);
3596 INIT_LIST_HEAD(&n->full);
3600 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3602 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3603 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3606 * Must align to double word boundary for the double cmpxchg
3607 * instructions to work; see __pcpu_double_call_return_bool().
3609 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3610 2 * sizeof(void *));
3615 init_kmem_cache_cpus(s);
3620 static struct kmem_cache *kmem_cache_node;
3623 * No kmalloc_node yet so do it by hand. We know that this is the first
3624 * slab on the node for this slabcache. There are no concurrent accesses
3627 * Note that this function only works on the kmem_cache_node
3628 * when allocating for the kmem_cache_node. This is used for bootstrapping
3629 * memory on a fresh node that has no slab structures yet.
3631 static void early_kmem_cache_node_alloc(int node)
3634 struct kmem_cache_node *n;
3636 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3638 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3641 if (page_to_nid(page) != node) {
3642 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3643 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3648 #ifdef CONFIG_SLUB_DEBUG
3649 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3650 init_tracking(kmem_cache_node, n);
3652 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3653 page->freelist = get_freepointer(kmem_cache_node, n);
3656 kmem_cache_node->node[node] = n;
3657 init_kmem_cache_node(n);
3658 inc_slabs_node(kmem_cache_node, node, page->objects);
3661 * No locks need to be taken here as it has just been
3662 * initialized and there is no concurrent access.
3664 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3667 static void free_kmem_cache_nodes(struct kmem_cache *s)
3670 struct kmem_cache_node *n;
3672 for_each_kmem_cache_node(s, node, n) {
3673 s->node[node] = NULL;
3674 kmem_cache_free(kmem_cache_node, n);
3678 void __kmem_cache_release(struct kmem_cache *s)
3680 cache_random_seq_destroy(s);
3681 free_percpu(s->cpu_slab);
3682 free_kmem_cache_nodes(s);
3685 static int init_kmem_cache_nodes(struct kmem_cache *s)
3689 for_each_node_mask(node, slab_nodes) {
3690 struct kmem_cache_node *n;
3692 if (slab_state == DOWN) {
3693 early_kmem_cache_node_alloc(node);
3696 n = kmem_cache_alloc_node(kmem_cache_node,
3700 free_kmem_cache_nodes(s);
3704 init_kmem_cache_node(n);
3710 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3712 if (min < MIN_PARTIAL)
3714 else if (min > MAX_PARTIAL)
3716 s->min_partial = min;
3719 static void set_cpu_partial(struct kmem_cache *s)
3721 #ifdef CONFIG_SLUB_CPU_PARTIAL
3723 * cpu_partial determined the maximum number of objects kept in the
3724 * per cpu partial lists of a processor.
3726 * Per cpu partial lists mainly contain slabs that just have one
3727 * object freed. If they are used for allocation then they can be
3728 * filled up again with minimal effort. The slab will never hit the
3729 * per node partial lists and therefore no locking will be required.
3731 * This setting also determines
3733 * A) The number of objects from per cpu partial slabs dumped to the
3734 * per node list when we reach the limit.
3735 * B) The number of objects in cpu partial slabs to extract from the
3736 * per node list when we run out of per cpu objects. We only fetch
3737 * 50% to keep some capacity around for frees.
3739 if (!kmem_cache_has_cpu_partial(s))
3740 slub_set_cpu_partial(s, 0);
3741 else if (s->size >= PAGE_SIZE)
3742 slub_set_cpu_partial(s, 2);
3743 else if (s->size >= 1024)
3744 slub_set_cpu_partial(s, 6);
3745 else if (s->size >= 256)
3746 slub_set_cpu_partial(s, 13);
3748 slub_set_cpu_partial(s, 30);
3753 * calculate_sizes() determines the order and the distribution of data within
3756 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3758 slab_flags_t flags = s->flags;
3759 unsigned int size = s->object_size;
3763 * Round up object size to the next word boundary. We can only
3764 * place the free pointer at word boundaries and this determines
3765 * the possible location of the free pointer.
3767 size = ALIGN(size, sizeof(void *));
3769 #ifdef CONFIG_SLUB_DEBUG
3771 * Determine if we can poison the object itself. If the user of
3772 * the slab may touch the object after free or before allocation
3773 * then we should never poison the object itself.
3775 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3777 s->flags |= __OBJECT_POISON;
3779 s->flags &= ~__OBJECT_POISON;
3783 * If we are Redzoning then check if there is some space between the
3784 * end of the object and the free pointer. If not then add an
3785 * additional word to have some bytes to store Redzone information.
3787 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3788 size += sizeof(void *);
3792 * With that we have determined the number of bytes in actual use
3793 * by the object and redzoning.
3797 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3798 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
3801 * Relocate free pointer after the object if it is not
3802 * permitted to overwrite the first word of the object on
3805 * This is the case if we do RCU, have a constructor or
3806 * destructor, are poisoning the objects, or are
3807 * redzoning an object smaller than sizeof(void *).
3809 * The assumption that s->offset >= s->inuse means free
3810 * pointer is outside of the object is used in the
3811 * freeptr_outside_object() function. If that is no
3812 * longer true, the function needs to be modified.
3815 size += sizeof(void *);
3818 * Store freelist pointer near middle of object to keep
3819 * it away from the edges of the object to avoid small
3820 * sized over/underflows from neighboring allocations.
3822 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
3825 #ifdef CONFIG_SLUB_DEBUG
3826 if (flags & SLAB_STORE_USER)
3828 * Need to store information about allocs and frees after
3831 size += 2 * sizeof(struct track);
3834 kasan_cache_create(s, &size, &s->flags);
3835 #ifdef CONFIG_SLUB_DEBUG
3836 if (flags & SLAB_RED_ZONE) {
3838 * Add some empty padding so that we can catch
3839 * overwrites from earlier objects rather than let
3840 * tracking information or the free pointer be
3841 * corrupted if a user writes before the start
3844 size += sizeof(void *);
3846 s->red_left_pad = sizeof(void *);
3847 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3848 size += s->red_left_pad;
3853 * SLUB stores one object immediately after another beginning from
3854 * offset 0. In order to align the objects we have to simply size
3855 * each object to conform to the alignment.
3857 size = ALIGN(size, s->align);
3859 s->reciprocal_size = reciprocal_value(size);
3860 if (forced_order >= 0)
3861 order = forced_order;
3863 order = calculate_order(size);
3870 s->allocflags |= __GFP_COMP;
3872 if (s->flags & SLAB_CACHE_DMA)
3873 s->allocflags |= GFP_DMA;
3875 if (s->flags & SLAB_CACHE_DMA32)
3876 s->allocflags |= GFP_DMA32;
3878 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3879 s->allocflags |= __GFP_RECLAIMABLE;
3882 * Determine the number of objects per slab
3884 s->oo = oo_make(order, size);
3885 s->min = oo_make(get_order(size), size);
3886 if (oo_objects(s->oo) > oo_objects(s->max))
3889 return !!oo_objects(s->oo);
3892 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3894 s->flags = kmem_cache_flags(s->size, flags, s->name);
3895 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3896 s->random = get_random_long();
3899 if (!calculate_sizes(s, -1))
3901 if (disable_higher_order_debug) {
3903 * Disable debugging flags that store metadata if the min slab
3906 if (get_order(s->size) > get_order(s->object_size)) {
3907 s->flags &= ~DEBUG_METADATA_FLAGS;
3909 if (!calculate_sizes(s, -1))
3914 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3915 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3916 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3917 /* Enable fast mode */
3918 s->flags |= __CMPXCHG_DOUBLE;
3922 * The larger the object size is, the more pages we want on the partial
3923 * list to avoid pounding the page allocator excessively.
3925 set_min_partial(s, ilog2(s->size) / 2);
3930 s->remote_node_defrag_ratio = 1000;
3933 /* Initialize the pre-computed randomized freelist if slab is up */
3934 if (slab_state >= UP) {
3935 if (init_cache_random_seq(s))
3939 if (!init_kmem_cache_nodes(s))
3942 if (alloc_kmem_cache_cpus(s))
3945 free_kmem_cache_nodes(s);
3950 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3953 #ifdef CONFIG_SLUB_DEBUG
3954 void *addr = page_address(page);
3958 slab_err(s, page, text, s->name);
3961 map = get_map(s, page);
3962 for_each_object(p, s, addr, page->objects) {
3964 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3965 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
3966 print_tracking(s, p);
3975 * Attempt to free all partial slabs on a node.
3976 * This is called from __kmem_cache_shutdown(). We must take list_lock
3977 * because sysfs file might still access partial list after the shutdowning.
3979 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3982 struct page *page, *h;
3984 BUG_ON(irqs_disabled());
3985 spin_lock_irq(&n->list_lock);
3986 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3988 remove_partial(n, page);
3989 list_add(&page->slab_list, &discard);
3991 list_slab_objects(s, page,
3992 "Objects remaining in %s on __kmem_cache_shutdown()");
3995 spin_unlock_irq(&n->list_lock);
3997 list_for_each_entry_safe(page, h, &discard, slab_list)
3998 discard_slab(s, page);
4001 bool __kmem_cache_empty(struct kmem_cache *s)
4004 struct kmem_cache_node *n;
4006 for_each_kmem_cache_node(s, node, n)
4007 if (n->nr_partial || slabs_node(s, node))
4013 * Release all resources used by a slab cache.
4015 int __kmem_cache_shutdown(struct kmem_cache *s)
4018 struct kmem_cache_node *n;
4021 /* Attempt to free all objects */
4022 for_each_kmem_cache_node(s, node, n) {
4024 if (n->nr_partial || slabs_node(s, node))
4030 #ifdef CONFIG_PRINTK
4031 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
4034 int __maybe_unused i;
4038 struct kmem_cache *s = page->slab_cache;
4039 struct track __maybe_unused *trackp;
4041 kpp->kp_ptr = object;
4042 kpp->kp_page = page;
4043 kpp->kp_slab_cache = s;
4044 base = page_address(page);
4045 objp0 = kasan_reset_tag(object);
4046 #ifdef CONFIG_SLUB_DEBUG
4047 objp = restore_red_left(s, objp0);
4051 objnr = obj_to_index(s, page, objp);
4052 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4053 objp = base + s->size * objnr;
4054 kpp->kp_objp = objp;
4055 if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) ||
4056 !(s->flags & SLAB_STORE_USER))
4058 #ifdef CONFIG_SLUB_DEBUG
4059 objp = fixup_red_left(s, objp);
4060 trackp = get_track(s, objp, TRACK_ALLOC);
4061 kpp->kp_ret = (void *)trackp->addr;
4062 #ifdef CONFIG_STACKDEPOT
4064 depot_stack_handle_t handle;
4065 unsigned long *entries;
4066 unsigned int nr_entries;
4068 handle = READ_ONCE(trackp->handle);
4070 nr_entries = stack_depot_fetch(handle, &entries);
4071 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4072 kpp->kp_stack[i] = (void *)entries[i];
4075 trackp = get_track(s, objp, TRACK_FREE);
4076 handle = READ_ONCE(trackp->handle);
4078 nr_entries = stack_depot_fetch(handle, &entries);
4079 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4080 kpp->kp_free_stack[i] = (void *)entries[i];
4088 /********************************************************************
4090 *******************************************************************/
4092 static int __init setup_slub_min_order(char *str)
4094 get_option(&str, (int *)&slub_min_order);
4099 __setup("slub_min_order=", setup_slub_min_order);
4101 static int __init setup_slub_max_order(char *str)
4103 get_option(&str, (int *)&slub_max_order);
4104 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4109 __setup("slub_max_order=", setup_slub_max_order);
4111 static int __init setup_slub_min_objects(char *str)
4113 get_option(&str, (int *)&slub_min_objects);
4118 __setup("slub_min_objects=", setup_slub_min_objects);
4120 void *__kmalloc(size_t size, gfp_t flags)
4122 struct kmem_cache *s;
4125 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4126 return kmalloc_large(size, flags);
4128 s = kmalloc_slab(size, flags);
4130 if (unlikely(ZERO_OR_NULL_PTR(s)))
4133 ret = slab_alloc(s, flags, _RET_IP_, size);
4135 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
4137 ret = kasan_kmalloc(s, ret, size, flags);
4141 EXPORT_SYMBOL(__kmalloc);
4144 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4148 unsigned int order = get_order(size);
4150 flags |= __GFP_COMP;
4151 page = alloc_pages_node(node, flags, order);
4153 ptr = page_address(page);
4154 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4155 PAGE_SIZE << order);
4158 return kmalloc_large_node_hook(ptr, size, flags);
4161 void *__kmalloc_node(size_t size, gfp_t flags, int node)
4163 struct kmem_cache *s;
4166 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4167 ret = kmalloc_large_node(size, flags, node);
4169 trace_kmalloc_node(_RET_IP_, ret,
4170 size, PAGE_SIZE << get_order(size),
4176 s = kmalloc_slab(size, flags);
4178 if (unlikely(ZERO_OR_NULL_PTR(s)))
4181 ret = slab_alloc_node(s, flags, node, _RET_IP_, size);
4183 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4185 ret = kasan_kmalloc(s, ret, size, flags);
4189 EXPORT_SYMBOL(__kmalloc_node);
4190 #endif /* CONFIG_NUMA */
4192 #ifdef CONFIG_HARDENED_USERCOPY
4194 * Rejects incorrectly sized objects and objects that are to be copied
4195 * to/from userspace but do not fall entirely within the containing slab
4196 * cache's usercopy region.
4198 * Returns NULL if check passes, otherwise const char * to name of cache
4199 * to indicate an error.
4201 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4204 struct kmem_cache *s;
4205 unsigned int offset;
4207 bool is_kfence = is_kfence_address(ptr);
4209 ptr = kasan_reset_tag(ptr);
4211 /* Find object and usable object size. */
4212 s = page->slab_cache;
4214 /* Reject impossible pointers. */
4215 if (ptr < page_address(page))
4216 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4219 /* Find offset within object. */
4221 offset = ptr - kfence_object_start(ptr);
4223 offset = (ptr - page_address(page)) % s->size;
4225 /* Adjust for redzone and reject if within the redzone. */
4226 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4227 if (offset < s->red_left_pad)
4228 usercopy_abort("SLUB object in left red zone",
4229 s->name, to_user, offset, n);
4230 offset -= s->red_left_pad;
4233 /* Allow address range falling entirely within usercopy region. */
4234 if (offset >= s->useroffset &&
4235 offset - s->useroffset <= s->usersize &&
4236 n <= s->useroffset - offset + s->usersize)
4240 * If the copy is still within the allocated object, produce
4241 * a warning instead of rejecting the copy. This is intended
4242 * to be a temporary method to find any missing usercopy
4245 object_size = slab_ksize(s);
4246 if (usercopy_fallback &&
4247 offset <= object_size && n <= object_size - offset) {
4248 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4252 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4254 #endif /* CONFIG_HARDENED_USERCOPY */
4256 size_t __ksize(const void *object)
4260 if (unlikely(object == ZERO_SIZE_PTR))
4263 page = virt_to_head_page(object);
4265 if (unlikely(!PageSlab(page))) {
4266 WARN_ON(!PageCompound(page));
4267 return page_size(page);
4270 return slab_ksize(page->slab_cache);
4272 EXPORT_SYMBOL(__ksize);
4274 void kfree(const void *x)
4277 void *object = (void *)x;
4279 trace_kfree(_RET_IP_, x);
4281 if (unlikely(ZERO_OR_NULL_PTR(x)))
4284 page = virt_to_head_page(x);
4285 if (unlikely(!PageSlab(page))) {
4286 unsigned int order = compound_order(page);
4288 BUG_ON(!PageCompound(page));
4290 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4291 -(PAGE_SIZE << order));
4292 __free_pages(page, order);
4295 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4297 EXPORT_SYMBOL(kfree);
4299 #define SHRINK_PROMOTE_MAX 32
4302 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4303 * up most to the head of the partial lists. New allocations will then
4304 * fill those up and thus they can be removed from the partial lists.
4306 * The slabs with the least items are placed last. This results in them
4307 * being allocated from last increasing the chance that the last objects
4308 * are freed in them.
4310 int __kmem_cache_shrink(struct kmem_cache *s)
4314 struct kmem_cache_node *n;
4317 struct list_head discard;
4318 struct list_head promote[SHRINK_PROMOTE_MAX];
4319 unsigned long flags;
4323 for_each_kmem_cache_node(s, node, n) {
4324 INIT_LIST_HEAD(&discard);
4325 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4326 INIT_LIST_HEAD(promote + i);
4328 spin_lock_irqsave(&n->list_lock, flags);
4331 * Build lists of slabs to discard or promote.
4333 * Note that concurrent frees may occur while we hold the
4334 * list_lock. page->inuse here is the upper limit.
4336 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4337 int free = page->objects - page->inuse;
4339 /* Do not reread page->inuse */
4342 /* We do not keep full slabs on the list */
4345 if (free == page->objects) {
4346 list_move(&page->slab_list, &discard);
4348 } else if (free <= SHRINK_PROMOTE_MAX)
4349 list_move(&page->slab_list, promote + free - 1);
4353 * Promote the slabs filled up most to the head of the
4356 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4357 list_splice(promote + i, &n->partial);
4359 spin_unlock_irqrestore(&n->list_lock, flags);
4361 /* Release empty slabs */
4362 list_for_each_entry_safe(page, t, &discard, slab_list)
4363 discard_slab(s, page);
4365 if (slabs_node(s, node))
4372 static int slab_mem_going_offline_callback(void *arg)
4374 struct kmem_cache *s;
4376 mutex_lock(&slab_mutex);
4377 list_for_each_entry(s, &slab_caches, list)
4378 __kmem_cache_shrink(s);
4379 mutex_unlock(&slab_mutex);
4384 static void slab_mem_offline_callback(void *arg)
4386 struct memory_notify *marg = arg;
4389 offline_node = marg->status_change_nid_normal;
4392 * If the node still has available memory. we need kmem_cache_node
4395 if (offline_node < 0)
4398 mutex_lock(&slab_mutex);
4399 node_clear(offline_node, slab_nodes);
4401 * We no longer free kmem_cache_node structures here, as it would be
4402 * racy with all get_node() users, and infeasible to protect them with
4405 mutex_unlock(&slab_mutex);
4408 static int slab_mem_going_online_callback(void *arg)
4410 struct kmem_cache_node *n;
4411 struct kmem_cache *s;
4412 struct memory_notify *marg = arg;
4413 int nid = marg->status_change_nid_normal;
4417 * If the node's memory is already available, then kmem_cache_node is
4418 * already created. Nothing to do.
4424 * We are bringing a node online. No memory is available yet. We must
4425 * allocate a kmem_cache_node structure in order to bring the node
4428 mutex_lock(&slab_mutex);
4429 list_for_each_entry(s, &slab_caches, list) {
4431 * The structure may already exist if the node was previously
4432 * onlined and offlined.
4434 if (get_node(s, nid))
4437 * XXX: kmem_cache_alloc_node will fallback to other nodes
4438 * since memory is not yet available from the node that
4441 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4446 init_kmem_cache_node(n);
4450 * Any cache created after this point will also have kmem_cache_node
4451 * initialized for the new node.
4453 node_set(nid, slab_nodes);
4455 mutex_unlock(&slab_mutex);
4459 static int slab_memory_callback(struct notifier_block *self,
4460 unsigned long action, void *arg)
4465 case MEM_GOING_ONLINE:
4466 ret = slab_mem_going_online_callback(arg);
4468 case MEM_GOING_OFFLINE:
4469 ret = slab_mem_going_offline_callback(arg);
4472 case MEM_CANCEL_ONLINE:
4473 slab_mem_offline_callback(arg);
4476 case MEM_CANCEL_OFFLINE:
4480 ret = notifier_from_errno(ret);
4486 static struct notifier_block slab_memory_callback_nb = {
4487 .notifier_call = slab_memory_callback,
4488 .priority = SLAB_CALLBACK_PRI,
4491 /********************************************************************
4492 * Basic setup of slabs
4493 *******************************************************************/
4496 * Used for early kmem_cache structures that were allocated using
4497 * the page allocator. Allocate them properly then fix up the pointers
4498 * that may be pointing to the wrong kmem_cache structure.
4501 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4504 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4505 struct kmem_cache_node *n;
4507 memcpy(s, static_cache, kmem_cache->object_size);
4510 * This runs very early, and only the boot processor is supposed to be
4511 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4514 __flush_cpu_slab(s, smp_processor_id());
4515 for_each_kmem_cache_node(s, node, n) {
4518 list_for_each_entry(p, &n->partial, slab_list)
4521 #ifdef CONFIG_SLUB_DEBUG
4522 list_for_each_entry(p, &n->full, slab_list)
4526 list_add(&s->list, &slab_caches);
4530 void __init kmem_cache_init(void)
4532 static __initdata struct kmem_cache boot_kmem_cache,
4533 boot_kmem_cache_node;
4536 if (debug_guardpage_minorder())
4539 /* Print slub debugging pointers without hashing */
4540 if (__slub_debug_enabled())
4541 no_hash_pointers_enable(NULL);
4543 kmem_cache_node = &boot_kmem_cache_node;
4544 kmem_cache = &boot_kmem_cache;
4547 * Initialize the nodemask for which we will allocate per node
4548 * structures. Here we don't need taking slab_mutex yet.
4550 for_each_node_state(node, N_NORMAL_MEMORY)
4551 node_set(node, slab_nodes);
4553 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4554 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4556 register_hotmemory_notifier(&slab_memory_callback_nb);
4558 /* Able to allocate the per node structures */
4559 slab_state = PARTIAL;
4561 create_boot_cache(kmem_cache, "kmem_cache",
4562 offsetof(struct kmem_cache, node) +
4563 nr_node_ids * sizeof(struct kmem_cache_node *),
4564 SLAB_HWCACHE_ALIGN, 0, 0);
4566 kmem_cache = bootstrap(&boot_kmem_cache);
4567 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4569 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4570 setup_kmalloc_cache_index_table();
4571 create_kmalloc_caches(0);
4573 /* Setup random freelists for each cache */
4574 init_freelist_randomization();
4576 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4579 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4581 slub_min_order, slub_max_order, slub_min_objects,
4582 nr_cpu_ids, nr_node_ids);
4585 void __init kmem_cache_init_late(void)
4590 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4591 slab_flags_t flags, void (*ctor)(void *))
4593 struct kmem_cache *s;
4595 s = find_mergeable(size, align, flags, name, ctor);
4600 * Adjust the object sizes so that we clear
4601 * the complete object on kzalloc.
4603 s->object_size = max(s->object_size, size);
4604 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4606 if (sysfs_slab_alias(s, name)) {
4615 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4619 err = kmem_cache_open(s, flags);
4623 /* Mutex is not taken during early boot */
4624 if (slab_state <= UP)
4627 err = sysfs_slab_add(s);
4629 __kmem_cache_release(s);
4631 if (s->flags & SLAB_STORE_USER)
4632 debugfs_slab_add(s);
4637 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4639 struct kmem_cache *s;
4642 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4643 return kmalloc_large(size, gfpflags);
4645 s = kmalloc_slab(size, gfpflags);
4647 if (unlikely(ZERO_OR_NULL_PTR(s)))
4650 ret = slab_alloc(s, gfpflags, caller, size);
4652 /* Honor the call site pointer we received. */
4653 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4657 EXPORT_SYMBOL(__kmalloc_track_caller);
4660 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4661 int node, unsigned long caller)
4663 struct kmem_cache *s;
4666 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4667 ret = kmalloc_large_node(size, gfpflags, node);
4669 trace_kmalloc_node(caller, ret,
4670 size, PAGE_SIZE << get_order(size),
4676 s = kmalloc_slab(size, gfpflags);
4678 if (unlikely(ZERO_OR_NULL_PTR(s)))
4681 ret = slab_alloc_node(s, gfpflags, node, caller, size);
4683 /* Honor the call site pointer we received. */
4684 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4688 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4692 static int count_inuse(struct page *page)
4697 static int count_total(struct page *page)
4699 return page->objects;
4703 #ifdef CONFIG_SLUB_DEBUG
4704 static void validate_slab(struct kmem_cache *s, struct page *page)
4707 void *addr = page_address(page);
4712 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4715 /* Now we know that a valid freelist exists */
4716 map = get_map(s, page);
4717 for_each_object(p, s, addr, page->objects) {
4718 u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
4719 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4721 if (!check_object(s, page, p, val))
4729 static int validate_slab_node(struct kmem_cache *s,
4730 struct kmem_cache_node *n)
4732 unsigned long count = 0;
4734 unsigned long flags;
4736 spin_lock_irqsave(&n->list_lock, flags);
4738 list_for_each_entry(page, &n->partial, slab_list) {
4739 validate_slab(s, page);
4742 if (count != n->nr_partial) {
4743 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4744 s->name, count, n->nr_partial);
4745 slab_add_kunit_errors();
4748 if (!(s->flags & SLAB_STORE_USER))
4751 list_for_each_entry(page, &n->full, slab_list) {
4752 validate_slab(s, page);
4755 if (count != atomic_long_read(&n->nr_slabs)) {
4756 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4757 s->name, count, atomic_long_read(&n->nr_slabs));
4758 slab_add_kunit_errors();
4762 spin_unlock_irqrestore(&n->list_lock, flags);
4766 long validate_slab_cache(struct kmem_cache *s)
4769 unsigned long count = 0;
4770 struct kmem_cache_node *n;
4773 for_each_kmem_cache_node(s, node, n)
4774 count += validate_slab_node(s, n);
4778 EXPORT_SYMBOL(validate_slab_cache);
4780 #ifdef CONFIG_DEBUG_FS
4782 * Generate lists of code addresses where slabcache objects are allocated
4787 unsigned long count;
4794 DECLARE_BITMAP(cpus, NR_CPUS);
4800 unsigned long count;
4801 struct location *loc;
4804 static struct dentry *slab_debugfs_root;
4806 static void free_loc_track(struct loc_track *t)
4809 free_pages((unsigned long)t->loc,
4810 get_order(sizeof(struct location) * t->max));
4813 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4818 order = get_order(sizeof(struct location) * max);
4820 l = (void *)__get_free_pages(flags, order);
4825 memcpy(l, t->loc, sizeof(struct location) * t->count);
4833 static int add_location(struct loc_track *t, struct kmem_cache *s,
4834 const struct track *track)
4836 long start, end, pos;
4838 unsigned long caddr;
4839 unsigned long age = jiffies - track->when;
4845 pos = start + (end - start + 1) / 2;
4848 * There is nothing at "end". If we end up there
4849 * we need to add something to before end.
4854 caddr = t->loc[pos].addr;
4855 if (track->addr == caddr) {
4861 if (age < l->min_time)
4863 if (age > l->max_time)
4866 if (track->pid < l->min_pid)
4867 l->min_pid = track->pid;
4868 if (track->pid > l->max_pid)
4869 l->max_pid = track->pid;
4871 cpumask_set_cpu(track->cpu,
4872 to_cpumask(l->cpus));
4874 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4878 if (track->addr < caddr)
4885 * Not found. Insert new tracking element.
4887 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4893 (t->count - pos) * sizeof(struct location));
4896 l->addr = track->addr;
4900 l->min_pid = track->pid;
4901 l->max_pid = track->pid;
4902 cpumask_clear(to_cpumask(l->cpus));
4903 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4904 nodes_clear(l->nodes);
4905 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4909 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4910 struct page *page, enum track_item alloc)
4912 void *addr = page_address(page);
4916 map = get_map(s, page);
4917 for_each_object(p, s, addr, page->objects)
4918 if (!test_bit(__obj_to_index(s, addr, p), map))
4919 add_location(t, s, get_track(s, p, alloc));
4922 #endif /* CONFIG_DEBUG_FS */
4923 #endif /* CONFIG_SLUB_DEBUG */
4926 enum slab_stat_type {
4927 SL_ALL, /* All slabs */
4928 SL_PARTIAL, /* Only partially allocated slabs */
4929 SL_CPU, /* Only slabs used for cpu caches */
4930 SL_OBJECTS, /* Determine allocated objects not slabs */
4931 SL_TOTAL /* Determine object capacity not slabs */
4934 #define SO_ALL (1 << SL_ALL)
4935 #define SO_PARTIAL (1 << SL_PARTIAL)
4936 #define SO_CPU (1 << SL_CPU)
4937 #define SO_OBJECTS (1 << SL_OBJECTS)
4938 #define SO_TOTAL (1 << SL_TOTAL)
4940 static ssize_t show_slab_objects(struct kmem_cache *s,
4941 char *buf, unsigned long flags)
4943 unsigned long total = 0;
4946 unsigned long *nodes;
4949 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4953 if (flags & SO_CPU) {
4956 for_each_possible_cpu(cpu) {
4957 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4962 page = READ_ONCE(c->page);
4966 node = page_to_nid(page);
4967 if (flags & SO_TOTAL)
4969 else if (flags & SO_OBJECTS)
4977 page = slub_percpu_partial_read_once(c);
4979 node = page_to_nid(page);
4980 if (flags & SO_TOTAL)
4982 else if (flags & SO_OBJECTS)
4993 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4994 * already held which will conflict with an existing lock order:
4996 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4998 * We don't really need mem_hotplug_lock (to hold off
4999 * slab_mem_going_offline_callback) here because slab's memory hot
5000 * unplug code doesn't destroy the kmem_cache->node[] data.
5003 #ifdef CONFIG_SLUB_DEBUG
5004 if (flags & SO_ALL) {
5005 struct kmem_cache_node *n;
5007 for_each_kmem_cache_node(s, node, n) {
5009 if (flags & SO_TOTAL)
5010 x = atomic_long_read(&n->total_objects);
5011 else if (flags & SO_OBJECTS)
5012 x = atomic_long_read(&n->total_objects) -
5013 count_partial(n, count_free);
5015 x = atomic_long_read(&n->nr_slabs);
5022 if (flags & SO_PARTIAL) {
5023 struct kmem_cache_node *n;
5025 for_each_kmem_cache_node(s, node, n) {
5026 if (flags & SO_TOTAL)
5027 x = count_partial(n, count_total);
5028 else if (flags & SO_OBJECTS)
5029 x = count_partial(n, count_inuse);
5037 len += sysfs_emit_at(buf, len, "%lu", total);
5039 for (node = 0; node < nr_node_ids; node++) {
5041 len += sysfs_emit_at(buf, len, " N%d=%lu",
5045 len += sysfs_emit_at(buf, len, "\n");
5051 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5052 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5054 struct slab_attribute {
5055 struct attribute attr;
5056 ssize_t (*show)(struct kmem_cache *s, char *buf);
5057 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5060 #define SLAB_ATTR_RO(_name) \
5061 static struct slab_attribute _name##_attr = \
5062 __ATTR(_name, 0400, _name##_show, NULL)
5064 #define SLAB_ATTR(_name) \
5065 static struct slab_attribute _name##_attr = \
5066 __ATTR(_name, 0600, _name##_show, _name##_store)
5068 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5070 return sysfs_emit(buf, "%u\n", s->size);
5072 SLAB_ATTR_RO(slab_size);
5074 static ssize_t align_show(struct kmem_cache *s, char *buf)
5076 return sysfs_emit(buf, "%u\n", s->align);
5078 SLAB_ATTR_RO(align);
5080 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5082 return sysfs_emit(buf, "%u\n", s->object_size);
5084 SLAB_ATTR_RO(object_size);
5086 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5088 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5090 SLAB_ATTR_RO(objs_per_slab);
5092 static ssize_t order_show(struct kmem_cache *s, char *buf)
5094 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5096 SLAB_ATTR_RO(order);
5098 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5100 return sysfs_emit(buf, "%lu\n", s->min_partial);
5103 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5109 err = kstrtoul(buf, 10, &min);
5113 set_min_partial(s, min);
5116 SLAB_ATTR(min_partial);
5118 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5120 return sysfs_emit(buf, "%u\n", slub_cpu_partial(s));
5123 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5126 unsigned int objects;
5129 err = kstrtouint(buf, 10, &objects);
5132 if (objects && !kmem_cache_has_cpu_partial(s))
5135 slub_set_cpu_partial(s, objects);
5139 SLAB_ATTR(cpu_partial);
5141 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5145 return sysfs_emit(buf, "%pS\n", s->ctor);
5149 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5151 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5153 SLAB_ATTR_RO(aliases);
5155 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5157 return show_slab_objects(s, buf, SO_PARTIAL);
5159 SLAB_ATTR_RO(partial);
5161 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5163 return show_slab_objects(s, buf, SO_CPU);
5165 SLAB_ATTR_RO(cpu_slabs);
5167 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5169 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5171 SLAB_ATTR_RO(objects);
5173 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5175 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5177 SLAB_ATTR_RO(objects_partial);
5179 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5186 for_each_online_cpu(cpu) {
5189 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5192 pages += page->pages;
5193 objects += page->pobjects;
5197 len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages);
5200 for_each_online_cpu(cpu) {
5203 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5205 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5206 cpu, page->pobjects, page->pages);
5209 len += sysfs_emit_at(buf, len, "\n");
5213 SLAB_ATTR_RO(slabs_cpu_partial);
5215 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5217 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5219 SLAB_ATTR_RO(reclaim_account);
5221 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5223 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5225 SLAB_ATTR_RO(hwcache_align);
5227 #ifdef CONFIG_ZONE_DMA
5228 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5230 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5232 SLAB_ATTR_RO(cache_dma);
5235 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5237 return sysfs_emit(buf, "%u\n", s->usersize);
5239 SLAB_ATTR_RO(usersize);
5241 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5243 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5245 SLAB_ATTR_RO(destroy_by_rcu);
5247 #ifdef CONFIG_SLUB_DEBUG
5248 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5250 return show_slab_objects(s, buf, SO_ALL);
5252 SLAB_ATTR_RO(slabs);
5254 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5256 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5258 SLAB_ATTR_RO(total_objects);
5260 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5262 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5264 SLAB_ATTR_RO(sanity_checks);
5266 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5268 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5270 SLAB_ATTR_RO(trace);
5272 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5274 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5277 SLAB_ATTR_RO(red_zone);
5279 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5281 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5284 SLAB_ATTR_RO(poison);
5286 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5288 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5291 SLAB_ATTR_RO(store_user);
5293 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5298 static ssize_t validate_store(struct kmem_cache *s,
5299 const char *buf, size_t length)
5303 if (buf[0] == '1') {
5304 ret = validate_slab_cache(s);
5310 SLAB_ATTR(validate);
5312 #endif /* CONFIG_SLUB_DEBUG */
5314 #ifdef CONFIG_FAILSLAB
5315 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5317 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5319 SLAB_ATTR_RO(failslab);
5322 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5327 static ssize_t shrink_store(struct kmem_cache *s,
5328 const char *buf, size_t length)
5331 kmem_cache_shrink(s);
5339 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5341 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5344 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5345 const char *buf, size_t length)
5350 err = kstrtouint(buf, 10, &ratio);
5356 s->remote_node_defrag_ratio = ratio * 10;
5360 SLAB_ATTR(remote_node_defrag_ratio);
5363 #ifdef CONFIG_SLUB_STATS
5364 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5366 unsigned long sum = 0;
5369 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5374 for_each_online_cpu(cpu) {
5375 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5381 len += sysfs_emit_at(buf, len, "%lu", sum);
5384 for_each_online_cpu(cpu) {
5386 len += sysfs_emit_at(buf, len, " C%d=%u",
5391 len += sysfs_emit_at(buf, len, "\n");
5396 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5400 for_each_online_cpu(cpu)
5401 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5404 #define STAT_ATTR(si, text) \
5405 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5407 return show_stat(s, buf, si); \
5409 static ssize_t text##_store(struct kmem_cache *s, \
5410 const char *buf, size_t length) \
5412 if (buf[0] != '0') \
5414 clear_stat(s, si); \
5419 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5420 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5421 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5422 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5423 STAT_ATTR(FREE_FROZEN, free_frozen);
5424 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5425 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5426 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5427 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5428 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5429 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5430 STAT_ATTR(FREE_SLAB, free_slab);
5431 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5432 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5433 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5434 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5435 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5436 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5437 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5438 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5439 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5440 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5441 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5442 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5443 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5444 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5445 #endif /* CONFIG_SLUB_STATS */
5447 static struct attribute *slab_attrs[] = {
5448 &slab_size_attr.attr,
5449 &object_size_attr.attr,
5450 &objs_per_slab_attr.attr,
5452 &min_partial_attr.attr,
5453 &cpu_partial_attr.attr,
5455 &objects_partial_attr.attr,
5457 &cpu_slabs_attr.attr,
5461 &hwcache_align_attr.attr,
5462 &reclaim_account_attr.attr,
5463 &destroy_by_rcu_attr.attr,
5465 &slabs_cpu_partial_attr.attr,
5466 #ifdef CONFIG_SLUB_DEBUG
5467 &total_objects_attr.attr,
5469 &sanity_checks_attr.attr,
5471 &red_zone_attr.attr,
5473 &store_user_attr.attr,
5474 &validate_attr.attr,
5476 #ifdef CONFIG_ZONE_DMA
5477 &cache_dma_attr.attr,
5480 &remote_node_defrag_ratio_attr.attr,
5482 #ifdef CONFIG_SLUB_STATS
5483 &alloc_fastpath_attr.attr,
5484 &alloc_slowpath_attr.attr,
5485 &free_fastpath_attr.attr,
5486 &free_slowpath_attr.attr,
5487 &free_frozen_attr.attr,
5488 &free_add_partial_attr.attr,
5489 &free_remove_partial_attr.attr,
5490 &alloc_from_partial_attr.attr,
5491 &alloc_slab_attr.attr,
5492 &alloc_refill_attr.attr,
5493 &alloc_node_mismatch_attr.attr,
5494 &free_slab_attr.attr,
5495 &cpuslab_flush_attr.attr,
5496 &deactivate_full_attr.attr,
5497 &deactivate_empty_attr.attr,
5498 &deactivate_to_head_attr.attr,
5499 &deactivate_to_tail_attr.attr,
5500 &deactivate_remote_frees_attr.attr,
5501 &deactivate_bypass_attr.attr,
5502 &order_fallback_attr.attr,
5503 &cmpxchg_double_fail_attr.attr,
5504 &cmpxchg_double_cpu_fail_attr.attr,
5505 &cpu_partial_alloc_attr.attr,
5506 &cpu_partial_free_attr.attr,
5507 &cpu_partial_node_attr.attr,
5508 &cpu_partial_drain_attr.attr,
5510 #ifdef CONFIG_FAILSLAB
5511 &failslab_attr.attr,
5513 &usersize_attr.attr,
5518 static const struct attribute_group slab_attr_group = {
5519 .attrs = slab_attrs,
5522 static ssize_t slab_attr_show(struct kobject *kobj,
5523 struct attribute *attr,
5526 struct slab_attribute *attribute;
5527 struct kmem_cache *s;
5530 attribute = to_slab_attr(attr);
5533 if (!attribute->show)
5536 err = attribute->show(s, buf);
5541 static ssize_t slab_attr_store(struct kobject *kobj,
5542 struct attribute *attr,
5543 const char *buf, size_t len)
5545 struct slab_attribute *attribute;
5546 struct kmem_cache *s;
5549 attribute = to_slab_attr(attr);
5552 if (!attribute->store)
5555 err = attribute->store(s, buf, len);
5559 static void kmem_cache_release(struct kobject *k)
5561 slab_kmem_cache_release(to_slab(k));
5564 static const struct sysfs_ops slab_sysfs_ops = {
5565 .show = slab_attr_show,
5566 .store = slab_attr_store,
5569 static struct kobj_type slab_ktype = {
5570 .sysfs_ops = &slab_sysfs_ops,
5571 .release = kmem_cache_release,
5574 static struct kset *slab_kset;
5576 static inline struct kset *cache_kset(struct kmem_cache *s)
5581 #define ID_STR_LENGTH 64
5583 /* Create a unique string id for a slab cache:
5585 * Format :[flags-]size
5587 static char *create_unique_id(struct kmem_cache *s)
5589 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5596 * First flags affecting slabcache operations. We will only
5597 * get here for aliasable slabs so we do not need to support
5598 * too many flags. The flags here must cover all flags that
5599 * are matched during merging to guarantee that the id is
5602 if (s->flags & SLAB_CACHE_DMA)
5604 if (s->flags & SLAB_CACHE_DMA32)
5606 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5608 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5610 if (s->flags & SLAB_ACCOUNT)
5614 p += sprintf(p, "%07u", s->size);
5616 BUG_ON(p > name + ID_STR_LENGTH - 1);
5620 static int sysfs_slab_add(struct kmem_cache *s)
5624 struct kset *kset = cache_kset(s);
5625 int unmergeable = slab_unmergeable(s);
5628 kobject_init(&s->kobj, &slab_ktype);
5632 if (!unmergeable && disable_higher_order_debug &&
5633 (slub_debug & DEBUG_METADATA_FLAGS))
5638 * Slabcache can never be merged so we can use the name proper.
5639 * This is typically the case for debug situations. In that
5640 * case we can catch duplicate names easily.
5642 sysfs_remove_link(&slab_kset->kobj, s->name);
5646 * Create a unique name for the slab as a target
5649 name = create_unique_id(s);
5652 s->kobj.kset = kset;
5653 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5657 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5662 /* Setup first alias */
5663 sysfs_slab_alias(s, s->name);
5670 kobject_del(&s->kobj);
5674 void sysfs_slab_unlink(struct kmem_cache *s)
5676 if (slab_state >= FULL)
5677 kobject_del(&s->kobj);
5680 void sysfs_slab_release(struct kmem_cache *s)
5682 if (slab_state >= FULL)
5683 kobject_put(&s->kobj);
5687 * Need to buffer aliases during bootup until sysfs becomes
5688 * available lest we lose that information.
5690 struct saved_alias {
5691 struct kmem_cache *s;
5693 struct saved_alias *next;
5696 static struct saved_alias *alias_list;
5698 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5700 struct saved_alias *al;
5702 if (slab_state == FULL) {
5704 * If we have a leftover link then remove it.
5706 sysfs_remove_link(&slab_kset->kobj, name);
5707 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5710 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5716 al->next = alias_list;
5721 static int __init slab_sysfs_init(void)
5723 struct kmem_cache *s;
5726 mutex_lock(&slab_mutex);
5728 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5730 mutex_unlock(&slab_mutex);
5731 pr_err("Cannot register slab subsystem.\n");
5737 list_for_each_entry(s, &slab_caches, list) {
5738 err = sysfs_slab_add(s);
5740 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5744 while (alias_list) {
5745 struct saved_alias *al = alias_list;
5747 alias_list = alias_list->next;
5748 err = sysfs_slab_alias(al->s, al->name);
5750 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5755 mutex_unlock(&slab_mutex);
5759 __initcall(slab_sysfs_init);
5760 #endif /* CONFIG_SYSFS */
5762 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
5763 static int slab_debugfs_show(struct seq_file *seq, void *v)
5767 unsigned int idx = *(unsigned int *)v;
5768 struct loc_track *t = seq->private;
5770 if (idx < t->count) {
5773 seq_printf(seq, "%7ld ", l->count);
5776 seq_printf(seq, "%pS", (void *)l->addr);
5778 seq_puts(seq, "<not-available>");
5780 if (l->sum_time != l->min_time) {
5781 seq_printf(seq, " age=%ld/%llu/%ld",
5782 l->min_time, div_u64(l->sum_time, l->count),
5785 seq_printf(seq, " age=%ld", l->min_time);
5787 if (l->min_pid != l->max_pid)
5788 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
5790 seq_printf(seq, " pid=%ld",
5793 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
5794 seq_printf(seq, " cpus=%*pbl",
5795 cpumask_pr_args(to_cpumask(l->cpus)));
5797 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
5798 seq_printf(seq, " nodes=%*pbl",
5799 nodemask_pr_args(&l->nodes));
5801 seq_puts(seq, "\n");
5804 if (!idx && !t->count)
5805 seq_puts(seq, "No data\n");
5810 static void slab_debugfs_stop(struct seq_file *seq, void *v)
5814 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
5816 struct loc_track *t = seq->private;
5820 if (*ppos <= t->count)
5826 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
5831 static const struct seq_operations slab_debugfs_sops = {
5832 .start = slab_debugfs_start,
5833 .next = slab_debugfs_next,
5834 .stop = slab_debugfs_stop,
5835 .show = slab_debugfs_show,
5838 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
5841 struct kmem_cache_node *n;
5842 enum track_item alloc;
5844 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
5845 sizeof(struct loc_track));
5846 struct kmem_cache *s = file_inode(filep)->i_private;
5848 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
5849 alloc = TRACK_ALLOC;
5853 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL))
5856 /* Push back cpu slabs */
5859 for_each_kmem_cache_node(s, node, n) {
5860 unsigned long flags;
5863 if (!atomic_long_read(&n->nr_slabs))
5866 spin_lock_irqsave(&n->list_lock, flags);
5867 list_for_each_entry(page, &n->partial, slab_list)
5868 process_slab(t, s, page, alloc);
5869 list_for_each_entry(page, &n->full, slab_list)
5870 process_slab(t, s, page, alloc);
5871 spin_unlock_irqrestore(&n->list_lock, flags);
5877 static int slab_debug_trace_release(struct inode *inode, struct file *file)
5879 struct seq_file *seq = file->private_data;
5880 struct loc_track *t = seq->private;
5883 return seq_release_private(inode, file);
5886 static const struct file_operations slab_debugfs_fops = {
5887 .open = slab_debug_trace_open,
5889 .llseek = seq_lseek,
5890 .release = slab_debug_trace_release,
5893 static void debugfs_slab_add(struct kmem_cache *s)
5895 struct dentry *slab_cache_dir;
5897 if (unlikely(!slab_debugfs_root))
5900 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
5902 debugfs_create_file("alloc_traces", 0400,
5903 slab_cache_dir, s, &slab_debugfs_fops);
5905 debugfs_create_file("free_traces", 0400,
5906 slab_cache_dir, s, &slab_debugfs_fops);
5909 void debugfs_slab_release(struct kmem_cache *s)
5911 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
5914 static int __init slab_debugfs_init(void)
5916 struct kmem_cache *s;
5918 slab_debugfs_root = debugfs_create_dir("slab", NULL);
5920 list_for_each_entry(s, &slab_caches, list)
5921 if (s->flags & SLAB_STORE_USER)
5922 debugfs_slab_add(s);
5927 __initcall(slab_debugfs_init);
5930 * The /proc/slabinfo ABI
5932 #ifdef CONFIG_SLUB_DEBUG
5933 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5935 unsigned long nr_slabs = 0;
5936 unsigned long nr_objs = 0;
5937 unsigned long nr_free = 0;
5939 struct kmem_cache_node *n;
5941 for_each_kmem_cache_node(s, node, n) {
5942 nr_slabs += node_nr_slabs(n);
5943 nr_objs += node_nr_objs(n);
5944 nr_free += count_partial(n, count_free);
5947 sinfo->active_objs = nr_objs - nr_free;
5948 sinfo->num_objs = nr_objs;
5949 sinfo->active_slabs = nr_slabs;
5950 sinfo->num_slabs = nr_slabs;
5951 sinfo->objects_per_slab = oo_objects(s->oo);
5952 sinfo->cache_order = oo_order(s->oo);
5955 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5959 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5960 size_t count, loff_t *ppos)
5964 #endif /* CONFIG_SLUB_DEBUG */