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
123 #ifdef CONFIG_SLUB_DEBUG_ON
124 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
126 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
128 #endif /* CONFIG_SLUB_DEBUG */
130 static inline bool kmem_cache_debug(struct kmem_cache *s)
132 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
135 void *fixup_red_left(struct kmem_cache *s, void *p)
137 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
138 p += s->red_left_pad;
143 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
145 #ifdef CONFIG_SLUB_CPU_PARTIAL
146 return !kmem_cache_debug(s);
153 * Issues still to be resolved:
155 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
157 * - Variable sizing of the per node arrays
160 /* Enable to log cmpxchg failures */
161 #undef SLUB_DEBUG_CMPXCHG
164 * Minimum number of partial slabs. These will be left on the partial
165 * lists even if they are empty. kmem_cache_shrink may reclaim them.
167 #define MIN_PARTIAL 5
170 * Maximum number of desirable partial slabs.
171 * The existence of more partial slabs makes kmem_cache_shrink
172 * sort the partial list by the number of objects in use.
174 #define MAX_PARTIAL 10
176 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
177 SLAB_POISON | SLAB_STORE_USER)
180 * These debug flags cannot use CMPXCHG because there might be consistency
181 * issues when checking or reading debug information
183 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
188 * Debugging flags that require metadata to be stored in the slab. These get
189 * disabled when slub_debug=O is used and a cache's min order increases with
192 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
195 #define OO_MASK ((1 << OO_SHIFT) - 1)
196 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
198 /* Internal SLUB flags */
200 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
201 /* Use cmpxchg_double */
202 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
205 * Tracking user of a slab.
207 #define TRACK_ADDRS_COUNT 16
209 unsigned long addr; /* Called from address */
210 #ifdef CONFIG_STACKDEPOT
211 depot_stack_handle_t handle;
213 int cpu; /* Was running on cpu */
214 int pid; /* Pid context */
215 unsigned long when; /* When did the operation occur */
218 enum track_item { TRACK_ALLOC, TRACK_FREE };
221 static int sysfs_slab_add(struct kmem_cache *);
222 static int sysfs_slab_alias(struct kmem_cache *, const char *);
224 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
225 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
229 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
230 static void debugfs_slab_add(struct kmem_cache *);
232 static inline void debugfs_slab_add(struct kmem_cache *s) { }
235 static inline void stat(const struct kmem_cache *s, enum stat_item si)
237 #ifdef CONFIG_SLUB_STATS
239 * The rmw is racy on a preemptible kernel but this is acceptable, so
240 * avoid this_cpu_add()'s irq-disable overhead.
242 raw_cpu_inc(s->cpu_slab->stat[si]);
247 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
248 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
249 * differ during memory hotplug/hotremove operations.
250 * Protected by slab_mutex.
252 static nodemask_t slab_nodes;
254 /********************************************************************
255 * Core slab cache functions
256 *******************************************************************/
259 * Returns freelist pointer (ptr). With hardening, this is obfuscated
260 * with an XOR of the address where the pointer is held and a per-cache
263 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
264 unsigned long ptr_addr)
266 #ifdef CONFIG_SLAB_FREELIST_HARDENED
268 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
269 * Normally, this doesn't cause any issues, as both set_freepointer()
270 * and get_freepointer() are called with a pointer with the same tag.
271 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
272 * example, when __free_slub() iterates over objects in a cache, it
273 * passes untagged pointers to check_object(). check_object() in turns
274 * calls get_freepointer() with an untagged pointer, which causes the
275 * freepointer to be restored incorrectly.
277 return (void *)((unsigned long)ptr ^ s->random ^
278 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
284 /* Returns the freelist pointer recorded at location ptr_addr. */
285 static inline void *freelist_dereference(const struct kmem_cache *s,
288 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
289 (unsigned long)ptr_addr);
292 static inline void *get_freepointer(struct kmem_cache *s, void *object)
294 object = kasan_reset_tag(object);
295 return freelist_dereference(s, object + s->offset);
298 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
300 prefetch(object + s->offset);
303 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
305 unsigned long freepointer_addr;
308 if (!debug_pagealloc_enabled_static())
309 return get_freepointer(s, object);
311 object = kasan_reset_tag(object);
312 freepointer_addr = (unsigned long)object + s->offset;
313 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
314 return freelist_ptr(s, p, freepointer_addr);
317 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
319 unsigned long freeptr_addr = (unsigned long)object + s->offset;
321 #ifdef CONFIG_SLAB_FREELIST_HARDENED
322 BUG_ON(object == fp); /* naive detection of double free or corruption */
325 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
326 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
329 /* Loop over all objects in a slab */
330 #define for_each_object(__p, __s, __addr, __objects) \
331 for (__p = fixup_red_left(__s, __addr); \
332 __p < (__addr) + (__objects) * (__s)->size; \
335 static inline unsigned int order_objects(unsigned int order, unsigned int size)
337 return ((unsigned int)PAGE_SIZE << order) / size;
340 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
343 struct kmem_cache_order_objects x = {
344 (order << OO_SHIFT) + order_objects(order, size)
350 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
352 return x.x >> OO_SHIFT;
355 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
357 return x.x & OO_MASK;
361 * Per slab locking using the pagelock
363 static __always_inline void slab_lock(struct page *page)
365 VM_BUG_ON_PAGE(PageTail(page), page);
366 bit_spin_lock(PG_locked, &page->flags);
369 static __always_inline void slab_unlock(struct page *page)
371 VM_BUG_ON_PAGE(PageTail(page), page);
372 __bit_spin_unlock(PG_locked, &page->flags);
375 /* Interrupts must be disabled (for the fallback code to work right) */
376 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
377 void *freelist_old, unsigned long counters_old,
378 void *freelist_new, unsigned long counters_new,
381 VM_BUG_ON(!irqs_disabled());
382 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
383 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
384 if (s->flags & __CMPXCHG_DOUBLE) {
385 if (cmpxchg_double(&page->freelist, &page->counters,
386 freelist_old, counters_old,
387 freelist_new, counters_new))
393 if (page->freelist == freelist_old &&
394 page->counters == counters_old) {
395 page->freelist = freelist_new;
396 page->counters = counters_new;
404 stat(s, CMPXCHG_DOUBLE_FAIL);
406 #ifdef SLUB_DEBUG_CMPXCHG
407 pr_info("%s %s: cmpxchg double redo ", n, s->name);
413 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
414 void *freelist_old, unsigned long counters_old,
415 void *freelist_new, unsigned long counters_new,
418 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
419 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
420 if (s->flags & __CMPXCHG_DOUBLE) {
421 if (cmpxchg_double(&page->freelist, &page->counters,
422 freelist_old, counters_old,
423 freelist_new, counters_new))
430 local_irq_save(flags);
432 if (page->freelist == freelist_old &&
433 page->counters == counters_old) {
434 page->freelist = freelist_new;
435 page->counters = counters_new;
437 local_irq_restore(flags);
441 local_irq_restore(flags);
445 stat(s, CMPXCHG_DOUBLE_FAIL);
447 #ifdef SLUB_DEBUG_CMPXCHG
448 pr_info("%s %s: cmpxchg double redo ", n, s->name);
454 #ifdef CONFIG_SLUB_DEBUG
455 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
456 static DEFINE_SPINLOCK(object_map_lock);
458 #if IS_ENABLED(CONFIG_KUNIT)
459 static bool slab_add_kunit_errors(void)
461 struct kunit_resource *resource;
463 if (likely(!current->kunit_test))
466 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
470 (*(int *)resource->data)++;
471 kunit_put_resource(resource);
475 static inline bool slab_add_kunit_errors(void) { return false; }
479 * Determine a map of object in use on a page.
481 * Node listlock must be held to guarantee that the page does
482 * not vanish from under us.
484 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
485 __acquires(&object_map_lock)
488 void *addr = page_address(page);
490 VM_BUG_ON(!irqs_disabled());
492 spin_lock(&object_map_lock);
494 bitmap_zero(object_map, page->objects);
496 for (p = page->freelist; p; p = get_freepointer(s, p))
497 set_bit(__obj_to_index(s, addr, p), object_map);
502 static void put_map(unsigned long *map) __releases(&object_map_lock)
504 VM_BUG_ON(map != object_map);
505 spin_unlock(&object_map_lock);
508 static inline unsigned int size_from_object(struct kmem_cache *s)
510 if (s->flags & SLAB_RED_ZONE)
511 return s->size - s->red_left_pad;
516 static inline void *restore_red_left(struct kmem_cache *s, void *p)
518 if (s->flags & SLAB_RED_ZONE)
519 p -= s->red_left_pad;
527 #if defined(CONFIG_SLUB_DEBUG_ON)
528 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
530 static slab_flags_t slub_debug;
533 static char *slub_debug_string;
534 static int disable_higher_order_debug;
537 * slub is about to manipulate internal object metadata. This memory lies
538 * outside the range of the allocated object, so accessing it would normally
539 * be reported by kasan as a bounds error. metadata_access_enable() is used
540 * to tell kasan that these accesses are OK.
542 static inline void metadata_access_enable(void)
544 kasan_disable_current();
547 static inline void metadata_access_disable(void)
549 kasan_enable_current();
556 /* Verify that a pointer has an address that is valid within a slab page */
557 static inline int check_valid_pointer(struct kmem_cache *s,
558 struct page *page, void *object)
565 base = page_address(page);
566 object = kasan_reset_tag(object);
567 object = restore_red_left(s, object);
568 if (object < base || object >= base + page->objects * s->size ||
569 (object - base) % s->size) {
576 static void print_section(char *level, char *text, u8 *addr,
579 metadata_access_enable();
580 print_hex_dump(level, kasan_reset_tag(text), DUMP_PREFIX_ADDRESS,
581 16, 1, addr, length, 1);
582 metadata_access_disable();
586 * See comment in calculate_sizes().
588 static inline bool freeptr_outside_object(struct kmem_cache *s)
590 return s->offset >= s->inuse;
594 * Return offset of the end of info block which is inuse + free pointer if
595 * not overlapping with object.
597 static inline unsigned int get_info_end(struct kmem_cache *s)
599 if (freeptr_outside_object(s))
600 return s->inuse + sizeof(void *);
605 static struct track *get_track(struct kmem_cache *s, void *object,
606 enum track_item alloc)
610 p = object + get_info_end(s);
612 return kasan_reset_tag(p + alloc);
615 #ifdef CONFIG_STACKDEPOT
616 static depot_stack_handle_t save_stack_depot_trace(gfp_t flags)
618 unsigned long entries[TRACK_ADDRS_COUNT];
619 depot_stack_handle_t handle;
620 unsigned int nr_entries;
622 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 4);
623 handle = stack_depot_save(entries, nr_entries, flags);
628 static void set_track(struct kmem_cache *s, void *object,
629 enum track_item alloc, unsigned long addr)
631 struct track *p = get_track(s, object, alloc);
634 #ifdef CONFIG_STACKDEPOT
635 p->handle = save_stack_depot_trace(GFP_NOWAIT);
638 p->cpu = smp_processor_id();
639 p->pid = current->pid;
642 memset(p, 0, sizeof(struct track));
646 static void init_tracking(struct kmem_cache *s, void *object)
648 if (!(s->flags & SLAB_STORE_USER))
651 set_track(s, object, TRACK_FREE, 0UL);
652 set_track(s, object, TRACK_ALLOC, 0UL);
655 static void print_track(const char *s, struct track *t, unsigned long pr_time)
660 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
661 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
662 #ifdef CONFIG_STACKDEPOT
664 depot_stack_handle_t handle;
665 unsigned long *entries;
666 unsigned int nr_entries;
668 handle = READ_ONCE(t->handle);
670 pr_err("object allocation/free stack trace missing\n");
672 nr_entries = stack_depot_fetch(handle, &entries);
673 stack_trace_print(entries, nr_entries, 0);
679 void print_tracking(struct kmem_cache *s, void *object)
681 unsigned long pr_time = jiffies;
682 if (!(s->flags & SLAB_STORE_USER))
685 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
686 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
689 static void print_page_info(struct page *page)
691 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
692 page, page->objects, page->inuse, page->freelist,
693 page->flags, &page->flags);
697 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
699 struct va_format vaf;
705 pr_err("=============================================================================\n");
706 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
707 pr_err("-----------------------------------------------------------------------------\n\n");
712 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
714 struct va_format vaf;
717 if (slab_add_kunit_errors())
723 pr_err("FIX %s: %pV\n", s->name, &vaf);
727 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
728 void **freelist, void *nextfree)
730 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
731 !check_valid_pointer(s, page, nextfree) && freelist) {
732 object_err(s, page, *freelist, "Freechain corrupt");
734 slab_fix(s, "Isolate corrupted freechain");
741 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
743 unsigned int off; /* Offset of last byte */
744 u8 *addr = page_address(page);
746 print_tracking(s, p);
748 print_page_info(page);
750 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
751 p, p - addr, get_freepointer(s, p));
753 if (s->flags & SLAB_RED_ZONE)
754 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
756 else if (p > addr + 16)
757 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
759 print_section(KERN_ERR, "Object ", p,
760 min_t(unsigned int, s->object_size, PAGE_SIZE));
761 if (s->flags & SLAB_RED_ZONE)
762 print_section(KERN_ERR, "Redzone ", p + s->object_size,
763 s->inuse - s->object_size);
765 off = get_info_end(s);
767 if (s->flags & SLAB_STORE_USER)
768 off += 2 * sizeof(struct track);
770 off += kasan_metadata_size(s);
772 if (off != size_from_object(s))
773 /* Beginning of the filler is the free pointer */
774 print_section(KERN_ERR, "Padding ", p + off,
775 size_from_object(s) - off);
780 void object_err(struct kmem_cache *s, struct page *page,
781 u8 *object, char *reason)
783 if (slab_add_kunit_errors())
786 slab_bug(s, "%s", reason);
787 print_trailer(s, page, object);
788 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
791 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
792 const char *fmt, ...)
797 if (slab_add_kunit_errors())
801 vsnprintf(buf, sizeof(buf), fmt, args);
803 slab_bug(s, "%s", buf);
804 print_page_info(page);
806 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
809 static void init_object(struct kmem_cache *s, void *object, u8 val)
811 u8 *p = kasan_reset_tag(object);
813 if (s->flags & SLAB_RED_ZONE)
814 memset(p - s->red_left_pad, val, s->red_left_pad);
816 if (s->flags & __OBJECT_POISON) {
817 memset(p, POISON_FREE, s->object_size - 1);
818 p[s->object_size - 1] = POISON_END;
821 if (s->flags & SLAB_RED_ZONE)
822 memset(p + s->object_size, val, s->inuse - s->object_size);
825 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
826 void *from, void *to)
828 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
829 memset(from, data, to - from);
832 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
833 u8 *object, char *what,
834 u8 *start, unsigned int value, unsigned int bytes)
838 u8 *addr = page_address(page);
840 metadata_access_enable();
841 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
842 metadata_access_disable();
847 while (end > fault && end[-1] == value)
850 if (slab_add_kunit_errors())
853 slab_bug(s, "%s overwritten", what);
854 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
855 fault, end - 1, fault - addr,
857 print_trailer(s, page, object);
858 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
861 restore_bytes(s, what, value, fault, end);
869 * Bytes of the object to be managed.
870 * If the freepointer may overlay the object then the free
871 * pointer is at the middle of the object.
873 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
876 * object + s->object_size
877 * Padding to reach word boundary. This is also used for Redzoning.
878 * Padding is extended by another word if Redzoning is enabled and
879 * object_size == inuse.
881 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
882 * 0xcc (RED_ACTIVE) for objects in use.
885 * Meta data starts here.
887 * A. Free pointer (if we cannot overwrite object on free)
888 * B. Tracking data for SLAB_STORE_USER
889 * C. Padding to reach required alignment boundary or at minimum
890 * one word if debugging is on to be able to detect writes
891 * before the word boundary.
893 * Padding is done using 0x5a (POISON_INUSE)
896 * Nothing is used beyond s->size.
898 * If slabcaches are merged then the object_size and inuse boundaries are mostly
899 * ignored. And therefore no slab options that rely on these boundaries
900 * may be used with merged slabcaches.
903 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
905 unsigned long off = get_info_end(s); /* The end of info */
907 if (s->flags & SLAB_STORE_USER)
908 /* We also have user information there */
909 off += 2 * sizeof(struct track);
911 off += kasan_metadata_size(s);
913 if (size_from_object(s) == off)
916 return check_bytes_and_report(s, page, p, "Object padding",
917 p + off, POISON_INUSE, size_from_object(s) - off);
920 /* Check the pad bytes at the end of a slab page */
921 static int slab_pad_check(struct kmem_cache *s, struct page *page)
930 if (!(s->flags & SLAB_POISON))
933 start = page_address(page);
934 length = page_size(page);
935 end = start + length;
936 remainder = length % s->size;
940 pad = end - remainder;
941 metadata_access_enable();
942 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
943 metadata_access_disable();
946 while (end > fault && end[-1] == POISON_INUSE)
949 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
950 fault, end - 1, fault - start);
951 print_section(KERN_ERR, "Padding ", pad, remainder);
953 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
957 static int check_object(struct kmem_cache *s, struct page *page,
958 void *object, u8 val)
961 u8 *endobject = object + s->object_size;
963 if (s->flags & SLAB_RED_ZONE) {
964 if (!check_bytes_and_report(s, page, object, "Left Redzone",
965 object - s->red_left_pad, val, s->red_left_pad))
968 if (!check_bytes_and_report(s, page, object, "Right Redzone",
969 endobject, val, s->inuse - s->object_size))
972 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
973 check_bytes_and_report(s, page, p, "Alignment padding",
974 endobject, POISON_INUSE,
975 s->inuse - s->object_size);
979 if (s->flags & SLAB_POISON) {
980 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
981 (!check_bytes_and_report(s, page, p, "Poison", p,
982 POISON_FREE, s->object_size - 1) ||
983 !check_bytes_and_report(s, page, p, "End Poison",
984 p + s->object_size - 1, POISON_END, 1)))
987 * check_pad_bytes cleans up on its own.
989 check_pad_bytes(s, page, p);
992 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
994 * Object and freepointer overlap. Cannot check
995 * freepointer while object is allocated.
999 /* Check free pointer validity */
1000 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
1001 object_err(s, page, p, "Freepointer corrupt");
1003 * No choice but to zap it and thus lose the remainder
1004 * of the free objects in this slab. May cause
1005 * another error because the object count is now wrong.
1007 set_freepointer(s, p, NULL);
1013 static int check_slab(struct kmem_cache *s, struct page *page)
1017 VM_BUG_ON(!irqs_disabled());
1019 if (!PageSlab(page)) {
1020 slab_err(s, page, "Not a valid slab page");
1024 maxobj = order_objects(compound_order(page), s->size);
1025 if (page->objects > maxobj) {
1026 slab_err(s, page, "objects %u > max %u",
1027 page->objects, maxobj);
1030 if (page->inuse > page->objects) {
1031 slab_err(s, page, "inuse %u > max %u",
1032 page->inuse, page->objects);
1035 /* Slab_pad_check fixes things up after itself */
1036 slab_pad_check(s, page);
1041 * Determine if a certain object on a page is on the freelist. Must hold the
1042 * slab lock to guarantee that the chains are in a consistent state.
1044 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
1048 void *object = NULL;
1051 fp = page->freelist;
1052 while (fp && nr <= page->objects) {
1055 if (!check_valid_pointer(s, page, fp)) {
1057 object_err(s, page, object,
1058 "Freechain corrupt");
1059 set_freepointer(s, object, NULL);
1061 slab_err(s, page, "Freepointer corrupt");
1062 page->freelist = NULL;
1063 page->inuse = page->objects;
1064 slab_fix(s, "Freelist cleared");
1070 fp = get_freepointer(s, object);
1074 max_objects = order_objects(compound_order(page), s->size);
1075 if (max_objects > MAX_OBJS_PER_PAGE)
1076 max_objects = MAX_OBJS_PER_PAGE;
1078 if (page->objects != max_objects) {
1079 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1080 page->objects, max_objects);
1081 page->objects = max_objects;
1082 slab_fix(s, "Number of objects adjusted");
1084 if (page->inuse != page->objects - nr) {
1085 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1086 page->inuse, page->objects - nr);
1087 page->inuse = page->objects - nr;
1088 slab_fix(s, "Object count adjusted");
1090 return search == NULL;
1093 static void trace(struct kmem_cache *s, struct page *page, void *object,
1096 if (s->flags & SLAB_TRACE) {
1097 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1099 alloc ? "alloc" : "free",
1100 object, page->inuse,
1104 print_section(KERN_INFO, "Object ", (void *)object,
1112 * Tracking of fully allocated slabs for debugging purposes.
1114 static void add_full(struct kmem_cache *s,
1115 struct kmem_cache_node *n, struct page *page)
1117 if (!(s->flags & SLAB_STORE_USER))
1120 lockdep_assert_held(&n->list_lock);
1121 list_add(&page->slab_list, &n->full);
1124 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1126 if (!(s->flags & SLAB_STORE_USER))
1129 lockdep_assert_held(&n->list_lock);
1130 list_del(&page->slab_list);
1133 /* Tracking of the number of slabs for debugging purposes */
1134 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1136 struct kmem_cache_node *n = get_node(s, node);
1138 return atomic_long_read(&n->nr_slabs);
1141 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1143 return atomic_long_read(&n->nr_slabs);
1146 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1148 struct kmem_cache_node *n = get_node(s, node);
1151 * May be called early in order to allocate a slab for the
1152 * kmem_cache_node structure. Solve the chicken-egg
1153 * dilemma by deferring the increment of the count during
1154 * bootstrap (see early_kmem_cache_node_alloc).
1157 atomic_long_inc(&n->nr_slabs);
1158 atomic_long_add(objects, &n->total_objects);
1161 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1163 struct kmem_cache_node *n = get_node(s, node);
1165 atomic_long_dec(&n->nr_slabs);
1166 atomic_long_sub(objects, &n->total_objects);
1169 /* Object debug checks for alloc/free paths */
1170 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1173 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1176 init_object(s, object, SLUB_RED_INACTIVE);
1177 init_tracking(s, object);
1181 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1183 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1186 metadata_access_enable();
1187 memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page));
1188 metadata_access_disable();
1191 static inline int alloc_consistency_checks(struct kmem_cache *s,
1192 struct page *page, void *object)
1194 if (!check_slab(s, page))
1197 if (!check_valid_pointer(s, page, object)) {
1198 object_err(s, page, object, "Freelist Pointer check fails");
1202 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1208 static noinline int alloc_debug_processing(struct kmem_cache *s,
1210 void *object, unsigned long addr)
1212 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1213 if (!alloc_consistency_checks(s, page, object))
1217 /* Success perform special debug activities for allocs */
1218 if (s->flags & SLAB_STORE_USER)
1219 set_track(s, object, TRACK_ALLOC, addr);
1220 trace(s, page, object, 1);
1221 init_object(s, object, SLUB_RED_ACTIVE);
1225 if (PageSlab(page)) {
1227 * If this is a slab page then lets do the best we can
1228 * to avoid issues in the future. Marking all objects
1229 * as used avoids touching the remaining objects.
1231 slab_fix(s, "Marking all objects used");
1232 page->inuse = page->objects;
1233 page->freelist = NULL;
1238 static inline int free_consistency_checks(struct kmem_cache *s,
1239 struct page *page, void *object, unsigned long addr)
1241 if (!check_valid_pointer(s, page, object)) {
1242 slab_err(s, page, "Invalid object pointer 0x%p", object);
1246 if (on_freelist(s, page, object)) {
1247 object_err(s, page, object, "Object already free");
1251 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1254 if (unlikely(s != page->slab_cache)) {
1255 if (!PageSlab(page)) {
1256 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1258 } else if (!page->slab_cache) {
1259 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1263 object_err(s, page, object,
1264 "page slab pointer corrupt.");
1270 /* Supports checking bulk free of a constructed freelist */
1271 static noinline int free_debug_processing(
1272 struct kmem_cache *s, struct page *page,
1273 void *head, void *tail, int bulk_cnt,
1276 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1277 void *object = head;
1279 unsigned long flags;
1282 spin_lock_irqsave(&n->list_lock, flags);
1285 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1286 if (!check_slab(s, page))
1293 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1294 if (!free_consistency_checks(s, page, object, addr))
1298 if (s->flags & SLAB_STORE_USER)
1299 set_track(s, object, TRACK_FREE, addr);
1300 trace(s, page, object, 0);
1301 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1302 init_object(s, object, SLUB_RED_INACTIVE);
1304 /* Reached end of constructed freelist yet? */
1305 if (object != tail) {
1306 object = get_freepointer(s, object);
1312 if (cnt != bulk_cnt)
1313 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1317 spin_unlock_irqrestore(&n->list_lock, flags);
1319 slab_fix(s, "Object at 0x%p not freed", object);
1324 * Parse a block of slub_debug options. Blocks are delimited by ';'
1326 * @str: start of block
1327 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1328 * @slabs: return start of list of slabs, or NULL when there's no list
1329 * @init: assume this is initial parsing and not per-kmem-create parsing
1331 * returns the start of next block if there's any, or NULL
1334 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1336 bool higher_order_disable = false;
1338 /* Skip any completely empty blocks */
1339 while (*str && *str == ';')
1344 * No options but restriction on slabs. This means full
1345 * debugging for slabs matching a pattern.
1347 *flags = DEBUG_DEFAULT_FLAGS;
1352 /* Determine which debug features should be switched on */
1353 for (; *str && *str != ',' && *str != ';'; str++) {
1354 switch (tolower(*str)) {
1359 *flags |= SLAB_CONSISTENCY_CHECKS;
1362 *flags |= SLAB_RED_ZONE;
1365 *flags |= SLAB_POISON;
1368 *flags |= SLAB_STORE_USER;
1371 *flags |= SLAB_TRACE;
1374 *flags |= SLAB_FAILSLAB;
1378 * Avoid enabling debugging on caches if its minimum
1379 * order would increase as a result.
1381 higher_order_disable = true;
1385 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1394 /* Skip over the slab list */
1395 while (*str && *str != ';')
1398 /* Skip any completely empty blocks */
1399 while (*str && *str == ';')
1402 if (init && higher_order_disable)
1403 disable_higher_order_debug = 1;
1411 static int __init setup_slub_debug(char *str)
1416 bool global_slub_debug_changed = false;
1417 bool slab_list_specified = false;
1419 slub_debug = DEBUG_DEFAULT_FLAGS;
1420 if (*str++ != '=' || !*str)
1422 * No options specified. Switch on full debugging.
1428 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1432 global_slub_debug_changed = true;
1434 slab_list_specified = true;
1439 * For backwards compatibility, a single list of flags with list of
1440 * slabs means debugging is only enabled for those slabs, so the global
1441 * slub_debug should be 0. We can extended that to multiple lists as
1442 * long as there is no option specifying flags without a slab list.
1444 if (slab_list_specified) {
1445 if (!global_slub_debug_changed)
1447 slub_debug_string = saved_str;
1450 if (slub_debug != 0 || slub_debug_string)
1451 static_branch_enable(&slub_debug_enabled);
1453 static_branch_disable(&slub_debug_enabled);
1454 if ((static_branch_unlikely(&init_on_alloc) ||
1455 static_branch_unlikely(&init_on_free)) &&
1456 (slub_debug & SLAB_POISON))
1457 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1461 __setup("slub_debug", setup_slub_debug);
1464 * kmem_cache_flags - apply debugging options to the cache
1465 * @object_size: the size of an object without meta data
1466 * @flags: flags to set
1467 * @name: name of the cache
1469 * Debug option(s) are applied to @flags. In addition to the debug
1470 * option(s), if a slab name (or multiple) is specified i.e.
1471 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1472 * then only the select slabs will receive the debug option(s).
1474 slab_flags_t kmem_cache_flags(unsigned int object_size,
1475 slab_flags_t flags, const char *name)
1480 slab_flags_t block_flags;
1481 slab_flags_t slub_debug_local = slub_debug;
1484 * If the slab cache is for debugging (e.g. kmemleak) then
1485 * don't store user (stack trace) information by default,
1486 * but let the user enable it via the command line below.
1488 if (flags & SLAB_NOLEAKTRACE)
1489 slub_debug_local &= ~SLAB_STORE_USER;
1492 next_block = slub_debug_string;
1493 /* Go through all blocks of debug options, see if any matches our slab's name */
1494 while (next_block) {
1495 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1498 /* Found a block that has a slab list, search it */
1503 end = strchrnul(iter, ',');
1504 if (next_block && next_block < end)
1505 end = next_block - 1;
1507 glob = strnchr(iter, end - iter, '*');
1509 cmplen = glob - iter;
1511 cmplen = max_t(size_t, len, (end - iter));
1513 if (!strncmp(name, iter, cmplen)) {
1514 flags |= block_flags;
1518 if (!*end || *end == ';')
1524 return flags | slub_debug_local;
1526 #else /* !CONFIG_SLUB_DEBUG */
1527 static inline void setup_object_debug(struct kmem_cache *s,
1528 struct page *page, void *object) {}
1530 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1532 static inline int alloc_debug_processing(struct kmem_cache *s,
1533 struct page *page, void *object, unsigned long addr) { return 0; }
1535 static inline int free_debug_processing(
1536 struct kmem_cache *s, struct page *page,
1537 void *head, void *tail, int bulk_cnt,
1538 unsigned long addr) { return 0; }
1540 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1542 static inline int check_object(struct kmem_cache *s, struct page *page,
1543 void *object, u8 val) { return 1; }
1544 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1545 struct page *page) {}
1546 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1547 struct page *page) {}
1548 slab_flags_t kmem_cache_flags(unsigned int object_size,
1549 slab_flags_t flags, const char *name)
1553 #define slub_debug 0
1555 #define disable_higher_order_debug 0
1557 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1559 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1561 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1563 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1566 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1567 void **freelist, void *nextfree)
1571 #endif /* CONFIG_SLUB_DEBUG */
1574 * Hooks for other subsystems that check memory allocations. In a typical
1575 * production configuration these hooks all should produce no code at all.
1577 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1579 ptr = kasan_kmalloc_large(ptr, size, flags);
1580 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1581 kmemleak_alloc(ptr, size, 1, flags);
1585 static __always_inline void kfree_hook(void *x)
1588 kasan_kfree_large(x);
1591 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1594 kmemleak_free_recursive(x, s->flags);
1597 * Trouble is that we may no longer disable interrupts in the fast path
1598 * So in order to make the debug calls that expect irqs to be
1599 * disabled we need to disable interrupts temporarily.
1601 #ifdef CONFIG_LOCKDEP
1603 unsigned long flags;
1605 local_irq_save(flags);
1606 debug_check_no_locks_freed(x, s->object_size);
1607 local_irq_restore(flags);
1610 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1611 debug_check_no_obj_freed(x, s->object_size);
1613 /* Use KCSAN to help debug racy use-after-free. */
1614 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1615 __kcsan_check_access(x, s->object_size,
1616 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1619 * As memory initialization might be integrated into KASAN,
1620 * kasan_slab_free and initialization memset's must be
1621 * kept together to avoid discrepancies in behavior.
1623 * The initialization memset's clear the object and the metadata,
1624 * but don't touch the SLAB redzone.
1629 if (!kasan_has_integrated_init())
1630 memset(kasan_reset_tag(x), 0, s->object_size);
1631 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1632 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1633 s->size - s->inuse - rsize);
1635 /* KASAN might put x into memory quarantine, delaying its reuse. */
1636 return kasan_slab_free(s, x, init);
1639 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1640 void **head, void **tail)
1645 void *old_tail = *tail ? *tail : *head;
1647 if (is_kfence_address(next)) {
1648 slab_free_hook(s, next, false);
1652 /* Head and tail of the reconstructed freelist */
1658 next = get_freepointer(s, object);
1660 /* If object's reuse doesn't have to be delayed */
1661 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1662 /* Move object to the new freelist */
1663 set_freepointer(s, object, *head);
1668 } while (object != old_tail);
1673 return *head != NULL;
1676 static void *setup_object(struct kmem_cache *s, struct page *page,
1679 setup_object_debug(s, page, object);
1680 object = kasan_init_slab_obj(s, object);
1681 if (unlikely(s->ctor)) {
1682 kasan_unpoison_object_data(s, object);
1684 kasan_poison_object_data(s, object);
1690 * Slab allocation and freeing
1692 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1693 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1696 unsigned int order = oo_order(oo);
1698 if (node == NUMA_NO_NODE)
1699 page = alloc_pages(flags, order);
1701 page = __alloc_pages_node(node, flags, order);
1706 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1707 /* Pre-initialize the random sequence cache */
1708 static int init_cache_random_seq(struct kmem_cache *s)
1710 unsigned int count = oo_objects(s->oo);
1713 /* Bailout if already initialised */
1717 err = cache_random_seq_create(s, count, GFP_KERNEL);
1719 pr_err("SLUB: Unable to initialize free list for %s\n",
1724 /* Transform to an offset on the set of pages */
1725 if (s->random_seq) {
1728 for (i = 0; i < count; i++)
1729 s->random_seq[i] *= s->size;
1734 /* Initialize each random sequence freelist per cache */
1735 static void __init init_freelist_randomization(void)
1737 struct kmem_cache *s;
1739 mutex_lock(&slab_mutex);
1741 list_for_each_entry(s, &slab_caches, list)
1742 init_cache_random_seq(s);
1744 mutex_unlock(&slab_mutex);
1747 /* Get the next entry on the pre-computed freelist randomized */
1748 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1749 unsigned long *pos, void *start,
1750 unsigned long page_limit,
1751 unsigned long freelist_count)
1756 * If the target page allocation failed, the number of objects on the
1757 * page might be smaller than the usual size defined by the cache.
1760 idx = s->random_seq[*pos];
1762 if (*pos >= freelist_count)
1764 } while (unlikely(idx >= page_limit));
1766 return (char *)start + idx;
1769 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1770 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1775 unsigned long idx, pos, page_limit, freelist_count;
1777 if (page->objects < 2 || !s->random_seq)
1780 freelist_count = oo_objects(s->oo);
1781 pos = get_random_int() % freelist_count;
1783 page_limit = page->objects * s->size;
1784 start = fixup_red_left(s, page_address(page));
1786 /* First entry is used as the base of the freelist */
1787 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1789 cur = setup_object(s, page, cur);
1790 page->freelist = cur;
1792 for (idx = 1; idx < page->objects; idx++) {
1793 next = next_freelist_entry(s, page, &pos, start, page_limit,
1795 next = setup_object(s, page, next);
1796 set_freepointer(s, cur, next);
1799 set_freepointer(s, cur, NULL);
1804 static inline int init_cache_random_seq(struct kmem_cache *s)
1808 static inline void init_freelist_randomization(void) { }
1809 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1813 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1815 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1818 struct kmem_cache_order_objects oo = s->oo;
1820 void *start, *p, *next;
1824 flags &= gfp_allowed_mask;
1826 if (gfpflags_allow_blocking(flags))
1829 flags |= s->allocflags;
1832 * Let the initial higher-order allocation fail under memory pressure
1833 * so we fall-back to the minimum order allocation.
1835 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1836 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1837 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1839 page = alloc_slab_page(s, alloc_gfp, node, oo);
1840 if (unlikely(!page)) {
1844 * Allocation may have failed due to fragmentation.
1845 * Try a lower order alloc if possible
1847 page = alloc_slab_page(s, alloc_gfp, node, oo);
1848 if (unlikely(!page))
1850 stat(s, ORDER_FALLBACK);
1853 page->objects = oo_objects(oo);
1855 account_slab_page(page, oo_order(oo), s, flags);
1857 page->slab_cache = s;
1858 __SetPageSlab(page);
1859 if (page_is_pfmemalloc(page))
1860 SetPageSlabPfmemalloc(page);
1862 kasan_poison_slab(page);
1864 start = page_address(page);
1866 setup_page_debug(s, page, start);
1868 shuffle = shuffle_freelist(s, page);
1871 start = fixup_red_left(s, start);
1872 start = setup_object(s, page, start);
1873 page->freelist = start;
1874 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1876 next = setup_object(s, page, next);
1877 set_freepointer(s, p, next);
1880 set_freepointer(s, p, NULL);
1883 page->inuse = page->objects;
1887 if (gfpflags_allow_blocking(flags))
1888 local_irq_disable();
1892 inc_slabs_node(s, page_to_nid(page), page->objects);
1897 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1899 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1900 flags = kmalloc_fix_flags(flags);
1902 return allocate_slab(s,
1903 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1906 static void __free_slab(struct kmem_cache *s, struct page *page)
1908 int order = compound_order(page);
1909 int pages = 1 << order;
1911 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1914 slab_pad_check(s, page);
1915 for_each_object(p, s, page_address(page),
1917 check_object(s, page, p, SLUB_RED_INACTIVE);
1920 __ClearPageSlabPfmemalloc(page);
1921 __ClearPageSlab(page);
1922 /* In union with page->mapping where page allocator expects NULL */
1923 page->slab_cache = NULL;
1924 if (current->reclaim_state)
1925 current->reclaim_state->reclaimed_slab += pages;
1926 unaccount_slab_page(page, order, s);
1927 __free_pages(page, order);
1930 static void rcu_free_slab(struct rcu_head *h)
1932 struct page *page = container_of(h, struct page, rcu_head);
1934 __free_slab(page->slab_cache, page);
1937 static void free_slab(struct kmem_cache *s, struct page *page)
1939 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1940 call_rcu(&page->rcu_head, rcu_free_slab);
1942 __free_slab(s, page);
1945 static void discard_slab(struct kmem_cache *s, struct page *page)
1947 dec_slabs_node(s, page_to_nid(page), page->objects);
1952 * Management of partially allocated slabs.
1955 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1958 if (tail == DEACTIVATE_TO_TAIL)
1959 list_add_tail(&page->slab_list, &n->partial);
1961 list_add(&page->slab_list, &n->partial);
1964 static inline void add_partial(struct kmem_cache_node *n,
1965 struct page *page, int tail)
1967 lockdep_assert_held(&n->list_lock);
1968 __add_partial(n, page, tail);
1971 static inline void remove_partial(struct kmem_cache_node *n,
1974 lockdep_assert_held(&n->list_lock);
1975 list_del(&page->slab_list);
1980 * Remove slab from the partial list, freeze it and
1981 * return the pointer to the freelist.
1983 * Returns a list of objects or NULL if it fails.
1985 static inline void *acquire_slab(struct kmem_cache *s,
1986 struct kmem_cache_node *n, struct page *page,
1987 int mode, int *objects)
1990 unsigned long counters;
1993 lockdep_assert_held(&n->list_lock);
1996 * Zap the freelist and set the frozen bit.
1997 * The old freelist is the list of objects for the
1998 * per cpu allocation list.
2000 freelist = page->freelist;
2001 counters = page->counters;
2002 new.counters = counters;
2003 *objects = new.objects - new.inuse;
2005 new.inuse = page->objects;
2006 new.freelist = NULL;
2008 new.freelist = freelist;
2011 VM_BUG_ON(new.frozen);
2014 if (!__cmpxchg_double_slab(s, page,
2016 new.freelist, new.counters,
2020 remove_partial(n, page);
2025 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
2026 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
2029 * Try to allocate a partial slab from a specific node.
2031 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2032 struct kmem_cache_cpu *c, gfp_t flags)
2034 struct page *page, *page2;
2035 void *object = NULL;
2036 unsigned int available = 0;
2040 * Racy check. If we mistakenly see no partial slabs then we
2041 * just allocate an empty slab. If we mistakenly try to get a
2042 * partial slab and there is none available then get_partial()
2045 if (!n || !n->nr_partial)
2048 spin_lock(&n->list_lock);
2049 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
2052 if (!pfmemalloc_match(page, flags))
2055 t = acquire_slab(s, n, page, object == NULL, &objects);
2059 available += objects;
2062 stat(s, ALLOC_FROM_PARTIAL);
2065 put_cpu_partial(s, page, 0);
2066 stat(s, CPU_PARTIAL_NODE);
2068 if (!kmem_cache_has_cpu_partial(s)
2069 || available > slub_cpu_partial(s) / 2)
2073 spin_unlock(&n->list_lock);
2078 * Get a page from somewhere. Search in increasing NUMA distances.
2080 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2081 struct kmem_cache_cpu *c)
2084 struct zonelist *zonelist;
2087 enum zone_type highest_zoneidx = gfp_zone(flags);
2089 unsigned int cpuset_mems_cookie;
2092 * The defrag ratio allows a configuration of the tradeoffs between
2093 * inter node defragmentation and node local allocations. A lower
2094 * defrag_ratio increases the tendency to do local allocations
2095 * instead of attempting to obtain partial slabs from other nodes.
2097 * If the defrag_ratio is set to 0 then kmalloc() always
2098 * returns node local objects. If the ratio is higher then kmalloc()
2099 * may return off node objects because partial slabs are obtained
2100 * from other nodes and filled up.
2102 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2103 * (which makes defrag_ratio = 1000) then every (well almost)
2104 * allocation will first attempt to defrag slab caches on other nodes.
2105 * This means scanning over all nodes to look for partial slabs which
2106 * may be expensive if we do it every time we are trying to find a slab
2107 * with available objects.
2109 if (!s->remote_node_defrag_ratio ||
2110 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2114 cpuset_mems_cookie = read_mems_allowed_begin();
2115 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2116 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2117 struct kmem_cache_node *n;
2119 n = get_node(s, zone_to_nid(zone));
2121 if (n && cpuset_zone_allowed(zone, flags) &&
2122 n->nr_partial > s->min_partial) {
2123 object = get_partial_node(s, n, c, flags);
2126 * Don't check read_mems_allowed_retry()
2127 * here - if mems_allowed was updated in
2128 * parallel, that was a harmless race
2129 * between allocation and the cpuset
2136 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2137 #endif /* CONFIG_NUMA */
2142 * Get a partial page, lock it and return it.
2144 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2145 struct kmem_cache_cpu *c)
2148 int searchnode = node;
2150 if (node == NUMA_NO_NODE)
2151 searchnode = numa_mem_id();
2153 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2154 if (object || node != NUMA_NO_NODE)
2157 return get_any_partial(s, flags, c);
2160 #ifdef CONFIG_PREEMPTION
2162 * Calculate the next globally unique transaction for disambiguation
2163 * during cmpxchg. The transactions start with the cpu number and are then
2164 * incremented by CONFIG_NR_CPUS.
2166 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2169 * No preemption supported therefore also no need to check for
2175 static inline unsigned long next_tid(unsigned long tid)
2177 return tid + TID_STEP;
2180 #ifdef SLUB_DEBUG_CMPXCHG
2181 static inline unsigned int tid_to_cpu(unsigned long tid)
2183 return tid % TID_STEP;
2186 static inline unsigned long tid_to_event(unsigned long tid)
2188 return tid / TID_STEP;
2192 static inline unsigned int init_tid(int cpu)
2197 static inline void note_cmpxchg_failure(const char *n,
2198 const struct kmem_cache *s, unsigned long tid)
2200 #ifdef SLUB_DEBUG_CMPXCHG
2201 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2203 pr_info("%s %s: cmpxchg redo ", n, s->name);
2205 #ifdef CONFIG_PREEMPTION
2206 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2207 pr_warn("due to cpu change %d -> %d\n",
2208 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2211 if (tid_to_event(tid) != tid_to_event(actual_tid))
2212 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2213 tid_to_event(tid), tid_to_event(actual_tid));
2215 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2216 actual_tid, tid, next_tid(tid));
2218 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2221 static void init_kmem_cache_cpus(struct kmem_cache *s)
2225 for_each_possible_cpu(cpu)
2226 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2230 * Remove the cpu slab
2232 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2233 void *freelist, struct kmem_cache_cpu *c)
2235 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2236 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2237 int lock = 0, free_delta = 0;
2238 enum slab_modes l = M_NONE, m = M_NONE;
2239 void *nextfree, *freelist_iter, *freelist_tail;
2240 int tail = DEACTIVATE_TO_HEAD;
2244 if (page->freelist) {
2245 stat(s, DEACTIVATE_REMOTE_FREES);
2246 tail = DEACTIVATE_TO_TAIL;
2250 * Stage one: Count the objects on cpu's freelist as free_delta and
2251 * remember the last object in freelist_tail for later splicing.
2253 freelist_tail = NULL;
2254 freelist_iter = freelist;
2255 while (freelist_iter) {
2256 nextfree = get_freepointer(s, freelist_iter);
2259 * If 'nextfree' is invalid, it is possible that the object at
2260 * 'freelist_iter' is already corrupted. So isolate all objects
2261 * starting at 'freelist_iter' by skipping them.
2263 if (freelist_corrupted(s, page, &freelist_iter, nextfree))
2266 freelist_tail = freelist_iter;
2269 freelist_iter = nextfree;
2273 * Stage two: Unfreeze the page while splicing the per-cpu
2274 * freelist to the head of page's freelist.
2276 * Ensure that the page is unfrozen while the list presence
2277 * reflects the actual number of objects during unfreeze.
2279 * We setup the list membership and then perform a cmpxchg
2280 * with the count. If there is a mismatch then the page
2281 * is not unfrozen but the page is on the wrong list.
2283 * Then we restart the process which may have to remove
2284 * the page from the list that we just put it on again
2285 * because the number of objects in the slab may have
2290 old.freelist = READ_ONCE(page->freelist);
2291 old.counters = READ_ONCE(page->counters);
2292 VM_BUG_ON(!old.frozen);
2294 /* Determine target state of the slab */
2295 new.counters = old.counters;
2296 if (freelist_tail) {
2297 new.inuse -= free_delta;
2298 set_freepointer(s, freelist_tail, old.freelist);
2299 new.freelist = freelist;
2301 new.freelist = old.freelist;
2305 if (!new.inuse && n->nr_partial >= s->min_partial)
2307 else if (new.freelist) {
2312 * Taking the spinlock removes the possibility
2313 * that acquire_slab() will see a slab page that
2316 spin_lock(&n->list_lock);
2320 if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) {
2323 * This also ensures that the scanning of full
2324 * slabs from diagnostic functions will not see
2327 spin_lock(&n->list_lock);
2333 remove_partial(n, page);
2334 else if (l == M_FULL)
2335 remove_full(s, n, page);
2338 add_partial(n, page, tail);
2339 else if (m == M_FULL)
2340 add_full(s, n, page);
2344 if (!__cmpxchg_double_slab(s, page,
2345 old.freelist, old.counters,
2346 new.freelist, new.counters,
2351 spin_unlock(&n->list_lock);
2355 else if (m == M_FULL)
2356 stat(s, DEACTIVATE_FULL);
2357 else if (m == M_FREE) {
2358 stat(s, DEACTIVATE_EMPTY);
2359 discard_slab(s, page);
2368 * Unfreeze all the cpu partial slabs.
2370 * This function must be called with interrupts disabled
2371 * for the cpu using c (or some other guarantee must be there
2372 * to guarantee no concurrent accesses).
2374 static void unfreeze_partials(struct kmem_cache *s,
2375 struct kmem_cache_cpu *c)
2377 #ifdef CONFIG_SLUB_CPU_PARTIAL
2378 struct kmem_cache_node *n = NULL, *n2 = NULL;
2379 struct page *page, *discard_page = NULL;
2381 while ((page = slub_percpu_partial(c))) {
2385 slub_set_percpu_partial(c, page);
2387 n2 = get_node(s, page_to_nid(page));
2390 spin_unlock(&n->list_lock);
2393 spin_lock(&n->list_lock);
2398 old.freelist = page->freelist;
2399 old.counters = page->counters;
2400 VM_BUG_ON(!old.frozen);
2402 new.counters = old.counters;
2403 new.freelist = old.freelist;
2407 } while (!__cmpxchg_double_slab(s, page,
2408 old.freelist, old.counters,
2409 new.freelist, new.counters,
2410 "unfreezing slab"));
2412 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2413 page->next = discard_page;
2414 discard_page = page;
2416 add_partial(n, page, DEACTIVATE_TO_TAIL);
2417 stat(s, FREE_ADD_PARTIAL);
2422 spin_unlock(&n->list_lock);
2424 while (discard_page) {
2425 page = discard_page;
2426 discard_page = discard_page->next;
2428 stat(s, DEACTIVATE_EMPTY);
2429 discard_slab(s, page);
2432 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2436 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2437 * partial page slot if available.
2439 * If we did not find a slot then simply move all the partials to the
2440 * per node partial list.
2442 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2444 #ifdef CONFIG_SLUB_CPU_PARTIAL
2445 struct page *oldpage;
2453 oldpage = this_cpu_read(s->cpu_slab->partial);
2456 pobjects = oldpage->pobjects;
2457 pages = oldpage->pages;
2458 if (drain && pobjects > slub_cpu_partial(s)) {
2459 unsigned long flags;
2461 * partial array is full. Move the existing
2462 * set to the per node partial list.
2464 local_irq_save(flags);
2465 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2466 local_irq_restore(flags);
2470 stat(s, CPU_PARTIAL_DRAIN);
2475 pobjects += page->objects - page->inuse;
2477 page->pages = pages;
2478 page->pobjects = pobjects;
2479 page->next = oldpage;
2481 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2483 if (unlikely(!slub_cpu_partial(s))) {
2484 unsigned long flags;
2486 local_irq_save(flags);
2487 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2488 local_irq_restore(flags);
2491 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2494 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2496 stat(s, CPUSLAB_FLUSH);
2497 deactivate_slab(s, c->page, c->freelist, c);
2499 c->tid = next_tid(c->tid);
2505 * Called from IPI handler with interrupts disabled.
2507 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2509 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2514 unfreeze_partials(s, c);
2517 static void flush_cpu_slab(void *d)
2519 struct kmem_cache *s = d;
2521 __flush_cpu_slab(s, smp_processor_id());
2524 static bool has_cpu_slab(int cpu, void *info)
2526 struct kmem_cache *s = info;
2527 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2529 return c->page || slub_percpu_partial(c);
2532 static void flush_all(struct kmem_cache *s)
2534 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2538 * Use the cpu notifier to insure that the cpu slabs are flushed when
2541 static int slub_cpu_dead(unsigned int cpu)
2543 struct kmem_cache *s;
2544 unsigned long flags;
2546 mutex_lock(&slab_mutex);
2547 list_for_each_entry(s, &slab_caches, list) {
2548 local_irq_save(flags);
2549 __flush_cpu_slab(s, cpu);
2550 local_irq_restore(flags);
2552 mutex_unlock(&slab_mutex);
2557 * Check if the objects in a per cpu structure fit numa
2558 * locality expectations.
2560 static inline int node_match(struct page *page, int node)
2563 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2569 #ifdef CONFIG_SLUB_DEBUG
2570 static int count_free(struct page *page)
2572 return page->objects - page->inuse;
2575 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2577 return atomic_long_read(&n->total_objects);
2579 #endif /* CONFIG_SLUB_DEBUG */
2581 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2582 static unsigned long count_partial(struct kmem_cache_node *n,
2583 int (*get_count)(struct page *))
2585 unsigned long flags;
2586 unsigned long x = 0;
2589 spin_lock_irqsave(&n->list_lock, flags);
2590 list_for_each_entry(page, &n->partial, slab_list)
2591 x += get_count(page);
2592 spin_unlock_irqrestore(&n->list_lock, flags);
2595 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2597 static noinline void
2598 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2600 #ifdef CONFIG_SLUB_DEBUG
2601 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2602 DEFAULT_RATELIMIT_BURST);
2604 struct kmem_cache_node *n;
2606 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2609 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2610 nid, gfpflags, &gfpflags);
2611 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2612 s->name, s->object_size, s->size, oo_order(s->oo),
2615 if (oo_order(s->min) > get_order(s->object_size))
2616 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2619 for_each_kmem_cache_node(s, node, n) {
2620 unsigned long nr_slabs;
2621 unsigned long nr_objs;
2622 unsigned long nr_free;
2624 nr_free = count_partial(n, count_free);
2625 nr_slabs = node_nr_slabs(n);
2626 nr_objs = node_nr_objs(n);
2628 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2629 node, nr_slabs, nr_objs, nr_free);
2634 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2635 int node, struct kmem_cache_cpu **pc)
2638 struct kmem_cache_cpu *c = *pc;
2641 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2643 freelist = get_partial(s, flags, node, c);
2648 page = new_slab(s, flags, node);
2650 c = raw_cpu_ptr(s->cpu_slab);
2655 * No other reference to the page yet so we can
2656 * muck around with it freely without cmpxchg
2658 freelist = page->freelist;
2659 page->freelist = NULL;
2661 stat(s, ALLOC_SLAB);
2669 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2671 if (unlikely(PageSlabPfmemalloc(page)))
2672 return gfp_pfmemalloc_allowed(gfpflags);
2678 * Check the page->freelist of a page and either transfer the freelist to the
2679 * per cpu freelist or deactivate the page.
2681 * The page is still frozen if the return value is not NULL.
2683 * If this function returns NULL then the page has been unfrozen.
2685 * This function must be called with interrupt disabled.
2687 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2690 unsigned long counters;
2694 freelist = page->freelist;
2695 counters = page->counters;
2697 new.counters = counters;
2698 VM_BUG_ON(!new.frozen);
2700 new.inuse = page->objects;
2701 new.frozen = freelist != NULL;
2703 } while (!__cmpxchg_double_slab(s, page,
2712 * Slow path. The lockless freelist is empty or we need to perform
2715 * Processing is still very fast if new objects have been freed to the
2716 * regular freelist. In that case we simply take over the regular freelist
2717 * as the lockless freelist and zap the regular freelist.
2719 * If that is not working then we fall back to the partial lists. We take the
2720 * first element of the freelist as the object to allocate now and move the
2721 * rest of the freelist to the lockless freelist.
2723 * And if we were unable to get a new slab from the partial slab lists then
2724 * we need to allocate a new slab. This is the slowest path since it involves
2725 * a call to the page allocator and the setup of a new slab.
2727 * Version of __slab_alloc to use when we know that interrupts are
2728 * already disabled (which is the case for bulk allocation).
2730 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2731 unsigned long addr, struct kmem_cache_cpu *c)
2736 stat(s, ALLOC_SLOWPATH);
2741 * if the node is not online or has no normal memory, just
2742 * ignore the node constraint
2744 if (unlikely(node != NUMA_NO_NODE &&
2745 !node_isset(node, slab_nodes)))
2746 node = NUMA_NO_NODE;
2751 if (unlikely(!node_match(page, node))) {
2753 * same as above but node_match() being false already
2754 * implies node != NUMA_NO_NODE
2756 if (!node_isset(node, slab_nodes)) {
2757 node = NUMA_NO_NODE;
2760 stat(s, ALLOC_NODE_MISMATCH);
2761 deactivate_slab(s, page, c->freelist, c);
2767 * By rights, we should be searching for a slab page that was
2768 * PFMEMALLOC but right now, we are losing the pfmemalloc
2769 * information when the page leaves the per-cpu allocator
2771 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2772 deactivate_slab(s, page, c->freelist, c);
2776 /* must check again c->freelist in case of cpu migration or IRQ */
2777 freelist = c->freelist;
2781 freelist = get_freelist(s, page);
2785 stat(s, DEACTIVATE_BYPASS);
2789 stat(s, ALLOC_REFILL);
2793 * freelist is pointing to the list of objects to be used.
2794 * page is pointing to the page from which the objects are obtained.
2795 * That page must be frozen for per cpu allocations to work.
2797 VM_BUG_ON(!c->page->frozen);
2798 c->freelist = get_freepointer(s, freelist);
2799 c->tid = next_tid(c->tid);
2804 if (slub_percpu_partial(c)) {
2805 page = c->page = slub_percpu_partial(c);
2806 slub_set_percpu_partial(c, page);
2807 stat(s, CPU_PARTIAL_ALLOC);
2811 freelist = new_slab_objects(s, gfpflags, node, &c);
2813 if (unlikely(!freelist)) {
2814 slab_out_of_memory(s, gfpflags, node);
2819 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2822 /* Only entered in the debug case */
2823 if (kmem_cache_debug(s) &&
2824 !alloc_debug_processing(s, page, freelist, addr))
2825 goto new_slab; /* Slab failed checks. Next slab needed */
2827 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2832 * Another one that disabled interrupt and compensates for possible
2833 * cpu changes by refetching the per cpu area pointer.
2835 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2836 unsigned long addr, struct kmem_cache_cpu *c)
2839 unsigned long flags;
2841 local_irq_save(flags);
2842 #ifdef CONFIG_PREEMPTION
2844 * We may have been preempted and rescheduled on a different
2845 * cpu before disabling interrupts. Need to reload cpu area
2848 c = this_cpu_ptr(s->cpu_slab);
2851 p = ___slab_alloc(s, gfpflags, node, addr, c);
2852 local_irq_restore(flags);
2857 * If the object has been wiped upon free, make sure it's fully initialized by
2858 * zeroing out freelist pointer.
2860 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2863 if (unlikely(slab_want_init_on_free(s)) && obj)
2864 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
2869 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2870 * have the fastpath folded into their functions. So no function call
2871 * overhead for requests that can be satisfied on the fastpath.
2873 * The fastpath works by first checking if the lockless freelist can be used.
2874 * If not then __slab_alloc is called for slow processing.
2876 * Otherwise we can simply pick the next object from the lockless free list.
2878 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2879 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
2882 struct kmem_cache_cpu *c;
2885 struct obj_cgroup *objcg = NULL;
2888 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2892 object = kfence_alloc(s, orig_size, gfpflags);
2893 if (unlikely(object))
2898 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2899 * enabled. We may switch back and forth between cpus while
2900 * reading from one cpu area. That does not matter as long
2901 * as we end up on the original cpu again when doing the cmpxchg.
2903 * We should guarantee that tid and kmem_cache are retrieved on
2904 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2905 * to check if it is matched or not.
2908 tid = this_cpu_read(s->cpu_slab->tid);
2909 c = raw_cpu_ptr(s->cpu_slab);
2910 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2911 unlikely(tid != READ_ONCE(c->tid)));
2914 * Irqless object alloc/free algorithm used here depends on sequence
2915 * of fetching cpu_slab's data. tid should be fetched before anything
2916 * on c to guarantee that object and page associated with previous tid
2917 * won't be used with current tid. If we fetch tid first, object and
2918 * page could be one associated with next tid and our alloc/free
2919 * request will be failed. In this case, we will retry. So, no problem.
2924 * The transaction ids are globally unique per cpu and per operation on
2925 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2926 * occurs on the right processor and that there was no operation on the
2927 * linked list in between.
2930 object = c->freelist;
2932 if (unlikely(!object || !page || !node_match(page, node))) {
2933 object = __slab_alloc(s, gfpflags, node, addr, c);
2935 void *next_object = get_freepointer_safe(s, object);
2938 * The cmpxchg will only match if there was no additional
2939 * operation and if we are on the right processor.
2941 * The cmpxchg does the following atomically (without lock
2943 * 1. Relocate first pointer to the current per cpu area.
2944 * 2. Verify that tid and freelist have not been changed
2945 * 3. If they were not changed replace tid and freelist
2947 * Since this is without lock semantics the protection is only
2948 * against code executing on this cpu *not* from access by
2951 if (unlikely(!this_cpu_cmpxchg_double(
2952 s->cpu_slab->freelist, s->cpu_slab->tid,
2954 next_object, next_tid(tid)))) {
2956 note_cmpxchg_failure("slab_alloc", s, tid);
2959 prefetch_freepointer(s, next_object);
2960 stat(s, ALLOC_FASTPATH);
2963 maybe_wipe_obj_freeptr(s, object);
2964 init = slab_want_init_on_alloc(gfpflags, s);
2967 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
2972 static __always_inline void *slab_alloc(struct kmem_cache *s,
2973 gfp_t gfpflags, unsigned long addr, size_t orig_size)
2975 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size);
2978 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2980 void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size);
2982 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2987 EXPORT_SYMBOL(kmem_cache_alloc);
2989 #ifdef CONFIG_TRACING
2990 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2992 void *ret = slab_alloc(s, gfpflags, _RET_IP_, size);
2993 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2994 ret = kasan_kmalloc(s, ret, size, gfpflags);
2997 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3001 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3003 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size);
3005 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3006 s->object_size, s->size, gfpflags, node);
3010 EXPORT_SYMBOL(kmem_cache_alloc_node);
3012 #ifdef CONFIG_TRACING
3013 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3015 int node, size_t size)
3017 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size);
3019 trace_kmalloc_node(_RET_IP_, ret,
3020 size, s->size, gfpflags, node);
3022 ret = kasan_kmalloc(s, ret, size, gfpflags);
3025 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3027 #endif /* CONFIG_NUMA */
3030 * Slow path handling. This may still be called frequently since objects
3031 * have a longer lifetime than the cpu slabs in most processing loads.
3033 * So we still attempt to reduce cache line usage. Just take the slab
3034 * lock and free the item. If there is no additional partial page
3035 * handling required then we can return immediately.
3037 static void __slab_free(struct kmem_cache *s, struct page *page,
3038 void *head, void *tail, int cnt,
3045 unsigned long counters;
3046 struct kmem_cache_node *n = NULL;
3047 unsigned long flags;
3049 stat(s, FREE_SLOWPATH);
3051 if (kfence_free(head))
3054 if (kmem_cache_debug(s) &&
3055 !free_debug_processing(s, page, head, tail, cnt, addr))
3060 spin_unlock_irqrestore(&n->list_lock, flags);
3063 prior = page->freelist;
3064 counters = page->counters;
3065 set_freepointer(s, tail, prior);
3066 new.counters = counters;
3067 was_frozen = new.frozen;
3069 if ((!new.inuse || !prior) && !was_frozen) {
3071 if (kmem_cache_has_cpu_partial(s) && !prior) {
3074 * Slab was on no list before and will be
3076 * We can defer the list move and instead
3081 } else { /* Needs to be taken off a list */
3083 n = get_node(s, page_to_nid(page));
3085 * Speculatively acquire the list_lock.
3086 * If the cmpxchg does not succeed then we may
3087 * drop the list_lock without any processing.
3089 * Otherwise the list_lock will synchronize with
3090 * other processors updating the list of slabs.
3092 spin_lock_irqsave(&n->list_lock, flags);
3097 } while (!cmpxchg_double_slab(s, page,
3104 if (likely(was_frozen)) {
3106 * The list lock was not taken therefore no list
3107 * activity can be necessary.
3109 stat(s, FREE_FROZEN);
3110 } else if (new.frozen) {
3112 * If we just froze the page then put it onto the
3113 * per cpu partial list.
3115 put_cpu_partial(s, page, 1);
3116 stat(s, CPU_PARTIAL_FREE);
3122 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3126 * Objects left in the slab. If it was not on the partial list before
3129 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3130 remove_full(s, n, page);
3131 add_partial(n, page, DEACTIVATE_TO_TAIL);
3132 stat(s, FREE_ADD_PARTIAL);
3134 spin_unlock_irqrestore(&n->list_lock, flags);
3140 * Slab on the partial list.
3142 remove_partial(n, page);
3143 stat(s, FREE_REMOVE_PARTIAL);
3145 /* Slab must be on the full list */
3146 remove_full(s, n, page);
3149 spin_unlock_irqrestore(&n->list_lock, flags);
3151 discard_slab(s, page);
3155 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3156 * can perform fastpath freeing without additional function calls.
3158 * The fastpath is only possible if we are freeing to the current cpu slab
3159 * of this processor. This typically the case if we have just allocated
3162 * If fastpath is not possible then fall back to __slab_free where we deal
3163 * with all sorts of special processing.
3165 * Bulk free of a freelist with several objects (all pointing to the
3166 * same page) possible by specifying head and tail ptr, plus objects
3167 * count (cnt). Bulk free indicated by tail pointer being set.
3169 static __always_inline void do_slab_free(struct kmem_cache *s,
3170 struct page *page, void *head, void *tail,
3171 int cnt, unsigned long addr)
3173 void *tail_obj = tail ? : head;
3174 struct kmem_cache_cpu *c;
3177 memcg_slab_free_hook(s, &head, 1);
3180 * Determine the currently cpus per cpu slab.
3181 * The cpu may change afterward. However that does not matter since
3182 * data is retrieved via this pointer. If we are on the same cpu
3183 * during the cmpxchg then the free will succeed.
3186 tid = this_cpu_read(s->cpu_slab->tid);
3187 c = raw_cpu_ptr(s->cpu_slab);
3188 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3189 unlikely(tid != READ_ONCE(c->tid)));
3191 /* Same with comment on barrier() in slab_alloc_node() */
3194 if (likely(page == c->page)) {
3195 void **freelist = READ_ONCE(c->freelist);
3197 set_freepointer(s, tail_obj, freelist);
3199 if (unlikely(!this_cpu_cmpxchg_double(
3200 s->cpu_slab->freelist, s->cpu_slab->tid,
3202 head, next_tid(tid)))) {
3204 note_cmpxchg_failure("slab_free", s, tid);
3207 stat(s, FREE_FASTPATH);
3209 __slab_free(s, page, head, tail_obj, cnt, addr);
3213 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3214 void *head, void *tail, int cnt,
3218 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3219 * to remove objects, whose reuse must be delayed.
3221 if (slab_free_freelist_hook(s, &head, &tail))
3222 do_slab_free(s, page, head, tail, cnt, addr);
3225 #ifdef CONFIG_KASAN_GENERIC
3226 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3228 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3232 void kmem_cache_free(struct kmem_cache *s, void *x)
3234 s = cache_from_obj(s, x);
3237 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3238 trace_kmem_cache_free(_RET_IP_, x, s->name);
3240 EXPORT_SYMBOL(kmem_cache_free);
3242 struct detached_freelist {
3247 struct kmem_cache *s;
3251 * This function progressively scans the array with free objects (with
3252 * a limited look ahead) and extract objects belonging to the same
3253 * page. It builds a detached freelist directly within the given
3254 * page/objects. This can happen without any need for
3255 * synchronization, because the objects are owned by running process.
3256 * The freelist is build up as a single linked list in the objects.
3257 * The idea is, that this detached freelist can then be bulk
3258 * transferred to the real freelist(s), but only requiring a single
3259 * synchronization primitive. Look ahead in the array is limited due
3260 * to performance reasons.
3263 int build_detached_freelist(struct kmem_cache *s, size_t size,
3264 void **p, struct detached_freelist *df)
3266 size_t first_skipped_index = 0;
3271 /* Always re-init detached_freelist */
3276 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3277 } while (!object && size);
3282 page = virt_to_head_page(object);
3284 /* Handle kalloc'ed objects */
3285 if (unlikely(!PageSlab(page))) {
3286 BUG_ON(!PageCompound(page));
3288 __free_pages(page, compound_order(page));
3289 p[size] = NULL; /* mark object processed */
3292 /* Derive kmem_cache from object */
3293 df->s = page->slab_cache;
3295 df->s = cache_from_obj(s, object); /* Support for memcg */
3298 if (is_kfence_address(object)) {
3299 slab_free_hook(df->s, object, false);
3300 __kfence_free(object);
3301 p[size] = NULL; /* mark object processed */
3305 /* Start new detached freelist */
3307 set_freepointer(df->s, object, NULL);
3309 df->freelist = object;
3310 p[size] = NULL; /* mark object processed */
3316 continue; /* Skip processed objects */
3318 /* df->page is always set at this point */
3319 if (df->page == virt_to_head_page(object)) {
3320 /* Opportunity build freelist */
3321 set_freepointer(df->s, object, df->freelist);
3322 df->freelist = object;
3324 p[size] = NULL; /* mark object processed */
3329 /* Limit look ahead search */
3333 if (!first_skipped_index)
3334 first_skipped_index = size + 1;
3337 return first_skipped_index;
3340 /* Note that interrupts must be enabled when calling this function. */
3341 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3346 memcg_slab_free_hook(s, p, size);
3348 struct detached_freelist df;
3350 size = build_detached_freelist(s, size, p, &df);
3354 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
3355 } while (likely(size));
3357 EXPORT_SYMBOL(kmem_cache_free_bulk);
3359 /* Note that interrupts must be enabled when calling this function. */
3360 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3363 struct kmem_cache_cpu *c;
3365 struct obj_cgroup *objcg = NULL;
3367 /* memcg and kmem_cache debug support */
3368 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3372 * Drain objects in the per cpu slab, while disabling local
3373 * IRQs, which protects against PREEMPT and interrupts
3374 * handlers invoking normal fastpath.
3376 local_irq_disable();
3377 c = this_cpu_ptr(s->cpu_slab);
3379 for (i = 0; i < size; i++) {
3380 void *object = kfence_alloc(s, s->object_size, flags);
3382 if (unlikely(object)) {
3387 object = c->freelist;
3388 if (unlikely(!object)) {
3390 * We may have removed an object from c->freelist using
3391 * the fastpath in the previous iteration; in that case,
3392 * c->tid has not been bumped yet.
3393 * Since ___slab_alloc() may reenable interrupts while
3394 * allocating memory, we should bump c->tid now.
3396 c->tid = next_tid(c->tid);
3399 * Invoking slow path likely have side-effect
3400 * of re-populating per CPU c->freelist
3402 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3404 if (unlikely(!p[i]))
3407 c = this_cpu_ptr(s->cpu_slab);
3408 maybe_wipe_obj_freeptr(s, p[i]);
3410 continue; /* goto for-loop */
3412 c->freelist = get_freepointer(s, object);
3414 maybe_wipe_obj_freeptr(s, p[i]);
3416 c->tid = next_tid(c->tid);
3420 * memcg and kmem_cache debug support and memory initialization.
3421 * Done outside of the IRQ disabled fastpath loop.
3423 slab_post_alloc_hook(s, objcg, flags, size, p,
3424 slab_want_init_on_alloc(flags, s));
3428 slab_post_alloc_hook(s, objcg, flags, i, p, false);
3429 __kmem_cache_free_bulk(s, i, p);
3432 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3436 * Object placement in a slab is made very easy because we always start at
3437 * offset 0. If we tune the size of the object to the alignment then we can
3438 * get the required alignment by putting one properly sized object after
3441 * Notice that the allocation order determines the sizes of the per cpu
3442 * caches. Each processor has always one slab available for allocations.
3443 * Increasing the allocation order reduces the number of times that slabs
3444 * must be moved on and off the partial lists and is therefore a factor in
3449 * Minimum / Maximum order of slab pages. This influences locking overhead
3450 * and slab fragmentation. A higher order reduces the number of partial slabs
3451 * and increases the number of allocations possible without having to
3452 * take the list_lock.
3454 static unsigned int slub_min_order;
3455 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3456 static unsigned int slub_min_objects;
3459 * Calculate the order of allocation given an slab object size.
3461 * The order of allocation has significant impact on performance and other
3462 * system components. Generally order 0 allocations should be preferred since
3463 * order 0 does not cause fragmentation in the page allocator. Larger objects
3464 * be problematic to put into order 0 slabs because there may be too much
3465 * unused space left. We go to a higher order if more than 1/16th of the slab
3468 * In order to reach satisfactory performance we must ensure that a minimum
3469 * number of objects is in one slab. Otherwise we may generate too much
3470 * activity on the partial lists which requires taking the list_lock. This is
3471 * less a concern for large slabs though which are rarely used.
3473 * slub_max_order specifies the order where we begin to stop considering the
3474 * number of objects in a slab as critical. If we reach slub_max_order then
3475 * we try to keep the page order as low as possible. So we accept more waste
3476 * of space in favor of a small page order.
3478 * Higher order allocations also allow the placement of more objects in a
3479 * slab and thereby reduce object handling overhead. If the user has
3480 * requested a higher minimum order then we start with that one instead of
3481 * the smallest order which will fit the object.
3483 static inline unsigned int slab_order(unsigned int size,
3484 unsigned int min_objects, unsigned int max_order,
3485 unsigned int fract_leftover)
3487 unsigned int min_order = slub_min_order;
3490 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3491 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3493 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3494 order <= max_order; order++) {
3496 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3499 rem = slab_size % size;
3501 if (rem <= slab_size / fract_leftover)
3508 static inline int calculate_order(unsigned int size)
3511 unsigned int min_objects;
3512 unsigned int max_objects;
3513 unsigned int nr_cpus;
3516 * Attempt to find best configuration for a slab. This
3517 * works by first attempting to generate a layout with
3518 * the best configuration and backing off gradually.
3520 * First we increase the acceptable waste in a slab. Then
3521 * we reduce the minimum objects required in a slab.
3523 min_objects = slub_min_objects;
3526 * Some architectures will only update present cpus when
3527 * onlining them, so don't trust the number if it's just 1. But
3528 * we also don't want to use nr_cpu_ids always, as on some other
3529 * architectures, there can be many possible cpus, but never
3530 * onlined. Here we compromise between trying to avoid too high
3531 * order on systems that appear larger than they are, and too
3532 * low order on systems that appear smaller than they are.
3534 nr_cpus = num_present_cpus();
3536 nr_cpus = nr_cpu_ids;
3537 min_objects = 4 * (fls(nr_cpus) + 1);
3539 max_objects = order_objects(slub_max_order, size);
3540 min_objects = min(min_objects, max_objects);
3542 while (min_objects > 1) {
3543 unsigned int fraction;
3546 while (fraction >= 4) {
3547 order = slab_order(size, min_objects,
3548 slub_max_order, fraction);
3549 if (order <= slub_max_order)
3557 * We were unable to place multiple objects in a slab. Now
3558 * lets see if we can place a single object there.
3560 order = slab_order(size, 1, slub_max_order, 1);
3561 if (order <= slub_max_order)
3565 * Doh this slab cannot be placed using slub_max_order.
3567 order = slab_order(size, 1, MAX_ORDER, 1);
3568 if (order < MAX_ORDER)
3574 init_kmem_cache_node(struct kmem_cache_node *n)
3577 spin_lock_init(&n->list_lock);
3578 INIT_LIST_HEAD(&n->partial);
3579 #ifdef CONFIG_SLUB_DEBUG
3580 atomic_long_set(&n->nr_slabs, 0);
3581 atomic_long_set(&n->total_objects, 0);
3582 INIT_LIST_HEAD(&n->full);
3586 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3588 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3589 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3592 * Must align to double word boundary for the double cmpxchg
3593 * instructions to work; see __pcpu_double_call_return_bool().
3595 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3596 2 * sizeof(void *));
3601 init_kmem_cache_cpus(s);
3606 static struct kmem_cache *kmem_cache_node;
3609 * No kmalloc_node yet so do it by hand. We know that this is the first
3610 * slab on the node for this slabcache. There are no concurrent accesses
3613 * Note that this function only works on the kmem_cache_node
3614 * when allocating for the kmem_cache_node. This is used for bootstrapping
3615 * memory on a fresh node that has no slab structures yet.
3617 static void early_kmem_cache_node_alloc(int node)
3620 struct kmem_cache_node *n;
3622 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3624 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3627 if (page_to_nid(page) != node) {
3628 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3629 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3634 #ifdef CONFIG_SLUB_DEBUG
3635 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3636 init_tracking(kmem_cache_node, n);
3638 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3639 page->freelist = get_freepointer(kmem_cache_node, n);
3642 kmem_cache_node->node[node] = n;
3643 init_kmem_cache_node(n);
3644 inc_slabs_node(kmem_cache_node, node, page->objects);
3647 * No locks need to be taken here as it has just been
3648 * initialized and there is no concurrent access.
3650 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3653 static void free_kmem_cache_nodes(struct kmem_cache *s)
3656 struct kmem_cache_node *n;
3658 for_each_kmem_cache_node(s, node, n) {
3659 s->node[node] = NULL;
3660 kmem_cache_free(kmem_cache_node, n);
3664 void __kmem_cache_release(struct kmem_cache *s)
3666 cache_random_seq_destroy(s);
3667 free_percpu(s->cpu_slab);
3668 free_kmem_cache_nodes(s);
3671 static int init_kmem_cache_nodes(struct kmem_cache *s)
3675 for_each_node_mask(node, slab_nodes) {
3676 struct kmem_cache_node *n;
3678 if (slab_state == DOWN) {
3679 early_kmem_cache_node_alloc(node);
3682 n = kmem_cache_alloc_node(kmem_cache_node,
3686 free_kmem_cache_nodes(s);
3690 init_kmem_cache_node(n);
3696 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3698 if (min < MIN_PARTIAL)
3700 else if (min > MAX_PARTIAL)
3702 s->min_partial = min;
3705 static void set_cpu_partial(struct kmem_cache *s)
3707 #ifdef CONFIG_SLUB_CPU_PARTIAL
3709 * cpu_partial determined the maximum number of objects kept in the
3710 * per cpu partial lists of a processor.
3712 * Per cpu partial lists mainly contain slabs that just have one
3713 * object freed. If they are used for allocation then they can be
3714 * filled up again with minimal effort. The slab will never hit the
3715 * per node partial lists and therefore no locking will be required.
3717 * This setting also determines
3719 * A) The number of objects from per cpu partial slabs dumped to the
3720 * per node list when we reach the limit.
3721 * B) The number of objects in cpu partial slabs to extract from the
3722 * per node list when we run out of per cpu objects. We only fetch
3723 * 50% to keep some capacity around for frees.
3725 if (!kmem_cache_has_cpu_partial(s))
3726 slub_set_cpu_partial(s, 0);
3727 else if (s->size >= PAGE_SIZE)
3728 slub_set_cpu_partial(s, 2);
3729 else if (s->size >= 1024)
3730 slub_set_cpu_partial(s, 6);
3731 else if (s->size >= 256)
3732 slub_set_cpu_partial(s, 13);
3734 slub_set_cpu_partial(s, 30);
3739 * calculate_sizes() determines the order and the distribution of data within
3742 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3744 slab_flags_t flags = s->flags;
3745 unsigned int size = s->object_size;
3749 * Round up object size to the next word boundary. We can only
3750 * place the free pointer at word boundaries and this determines
3751 * the possible location of the free pointer.
3753 size = ALIGN(size, sizeof(void *));
3755 #ifdef CONFIG_SLUB_DEBUG
3757 * Determine if we can poison the object itself. If the user of
3758 * the slab may touch the object after free or before allocation
3759 * then we should never poison the object itself.
3761 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3763 s->flags |= __OBJECT_POISON;
3765 s->flags &= ~__OBJECT_POISON;
3769 * If we are Redzoning then check if there is some space between the
3770 * end of the object and the free pointer. If not then add an
3771 * additional word to have some bytes to store Redzone information.
3773 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3774 size += sizeof(void *);
3778 * With that we have determined the number of bytes in actual use
3779 * by the object and redzoning.
3783 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3784 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
3787 * Relocate free pointer after the object if it is not
3788 * permitted to overwrite the first word of the object on
3791 * This is the case if we do RCU, have a constructor or
3792 * destructor, are poisoning the objects, or are
3793 * redzoning an object smaller than sizeof(void *).
3795 * The assumption that s->offset >= s->inuse means free
3796 * pointer is outside of the object is used in the
3797 * freeptr_outside_object() function. If that is no
3798 * longer true, the function needs to be modified.
3801 size += sizeof(void *);
3804 * Store freelist pointer near middle of object to keep
3805 * it away from the edges of the object to avoid small
3806 * sized over/underflows from neighboring allocations.
3808 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
3811 #ifdef CONFIG_SLUB_DEBUG
3812 if (flags & SLAB_STORE_USER)
3814 * Need to store information about allocs and frees after
3817 size += 2 * sizeof(struct track);
3820 kasan_cache_create(s, &size, &s->flags);
3821 #ifdef CONFIG_SLUB_DEBUG
3822 if (flags & SLAB_RED_ZONE) {
3824 * Add some empty padding so that we can catch
3825 * overwrites from earlier objects rather than let
3826 * tracking information or the free pointer be
3827 * corrupted if a user writes before the start
3830 size += sizeof(void *);
3832 s->red_left_pad = sizeof(void *);
3833 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3834 size += s->red_left_pad;
3839 * SLUB stores one object immediately after another beginning from
3840 * offset 0. In order to align the objects we have to simply size
3841 * each object to conform to the alignment.
3843 size = ALIGN(size, s->align);
3845 s->reciprocal_size = reciprocal_value(size);
3846 if (forced_order >= 0)
3847 order = forced_order;
3849 order = calculate_order(size);
3856 s->allocflags |= __GFP_COMP;
3858 if (s->flags & SLAB_CACHE_DMA)
3859 s->allocflags |= GFP_DMA;
3861 if (s->flags & SLAB_CACHE_DMA32)
3862 s->allocflags |= GFP_DMA32;
3864 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3865 s->allocflags |= __GFP_RECLAIMABLE;
3868 * Determine the number of objects per slab
3870 s->oo = oo_make(order, size);
3871 s->min = oo_make(get_order(size), size);
3872 if (oo_objects(s->oo) > oo_objects(s->max))
3875 return !!oo_objects(s->oo);
3878 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3880 s->flags = kmem_cache_flags(s->size, flags, s->name);
3881 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3882 s->random = get_random_long();
3885 if (!calculate_sizes(s, -1))
3887 if (disable_higher_order_debug) {
3889 * Disable debugging flags that store metadata if the min slab
3892 if (get_order(s->size) > get_order(s->object_size)) {
3893 s->flags &= ~DEBUG_METADATA_FLAGS;
3895 if (!calculate_sizes(s, -1))
3900 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3901 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3902 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3903 /* Enable fast mode */
3904 s->flags |= __CMPXCHG_DOUBLE;
3908 * The larger the object size is, the more pages we want on the partial
3909 * list to avoid pounding the page allocator excessively.
3911 set_min_partial(s, ilog2(s->size) / 2);
3916 s->remote_node_defrag_ratio = 1000;
3919 /* Initialize the pre-computed randomized freelist if slab is up */
3920 if (slab_state >= UP) {
3921 if (init_cache_random_seq(s))
3925 if (!init_kmem_cache_nodes(s))
3928 if (alloc_kmem_cache_cpus(s))
3931 free_kmem_cache_nodes(s);
3936 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3939 #ifdef CONFIG_SLUB_DEBUG
3940 void *addr = page_address(page);
3944 slab_err(s, page, text, s->name);
3947 map = get_map(s, page);
3948 for_each_object(p, s, addr, page->objects) {
3950 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3951 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
3952 print_tracking(s, p);
3961 * Attempt to free all partial slabs on a node.
3962 * This is called from __kmem_cache_shutdown(). We must take list_lock
3963 * because sysfs file might still access partial list after the shutdowning.
3965 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3968 struct page *page, *h;
3970 BUG_ON(irqs_disabled());
3971 spin_lock_irq(&n->list_lock);
3972 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3974 remove_partial(n, page);
3975 list_add(&page->slab_list, &discard);
3977 list_slab_objects(s, page,
3978 "Objects remaining in %s on __kmem_cache_shutdown()");
3981 spin_unlock_irq(&n->list_lock);
3983 list_for_each_entry_safe(page, h, &discard, slab_list)
3984 discard_slab(s, page);
3987 bool __kmem_cache_empty(struct kmem_cache *s)
3990 struct kmem_cache_node *n;
3992 for_each_kmem_cache_node(s, node, n)
3993 if (n->nr_partial || slabs_node(s, node))
3999 * Release all resources used by a slab cache.
4001 int __kmem_cache_shutdown(struct kmem_cache *s)
4004 struct kmem_cache_node *n;
4007 /* Attempt to free all objects */
4008 for_each_kmem_cache_node(s, node, n) {
4010 if (n->nr_partial || slabs_node(s, node))
4016 #ifdef CONFIG_PRINTK
4017 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
4020 int __maybe_unused i;
4024 struct kmem_cache *s = page->slab_cache;
4025 struct track __maybe_unused *trackp;
4027 kpp->kp_ptr = object;
4028 kpp->kp_page = page;
4029 kpp->kp_slab_cache = s;
4030 base = page_address(page);
4031 objp0 = kasan_reset_tag(object);
4032 #ifdef CONFIG_SLUB_DEBUG
4033 objp = restore_red_left(s, objp0);
4037 objnr = obj_to_index(s, page, objp);
4038 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4039 objp = base + s->size * objnr;
4040 kpp->kp_objp = objp;
4041 if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) ||
4042 !(s->flags & SLAB_STORE_USER))
4044 #ifdef CONFIG_SLUB_DEBUG
4045 objp = fixup_red_left(s, objp);
4046 trackp = get_track(s, objp, TRACK_ALLOC);
4047 kpp->kp_ret = (void *)trackp->addr;
4048 #ifdef CONFIG_STACKDEPOT
4050 depot_stack_handle_t handle;
4051 unsigned long *entries;
4052 unsigned int nr_entries;
4054 handle = READ_ONCE(trackp->handle);
4056 nr_entries = stack_depot_fetch(handle, &entries);
4057 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4058 kpp->kp_stack[i] = (void *)entries[i];
4061 trackp = get_track(s, objp, TRACK_FREE);
4062 handle = READ_ONCE(trackp->handle);
4064 nr_entries = stack_depot_fetch(handle, &entries);
4065 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4066 kpp->kp_free_stack[i] = (void *)entries[i];
4074 /********************************************************************
4076 *******************************************************************/
4078 static int __init setup_slub_min_order(char *str)
4080 get_option(&str, (int *)&slub_min_order);
4085 __setup("slub_min_order=", setup_slub_min_order);
4087 static int __init setup_slub_max_order(char *str)
4089 get_option(&str, (int *)&slub_max_order);
4090 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4095 __setup("slub_max_order=", setup_slub_max_order);
4097 static int __init setup_slub_min_objects(char *str)
4099 get_option(&str, (int *)&slub_min_objects);
4104 __setup("slub_min_objects=", setup_slub_min_objects);
4106 void *__kmalloc(size_t size, gfp_t flags)
4108 struct kmem_cache *s;
4111 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4112 return kmalloc_large(size, flags);
4114 s = kmalloc_slab(size, flags);
4116 if (unlikely(ZERO_OR_NULL_PTR(s)))
4119 ret = slab_alloc(s, flags, _RET_IP_, size);
4121 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
4123 ret = kasan_kmalloc(s, ret, size, flags);
4127 EXPORT_SYMBOL(__kmalloc);
4130 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4134 unsigned int order = get_order(size);
4136 flags |= __GFP_COMP;
4137 page = alloc_pages_node(node, flags, order);
4139 ptr = page_address(page);
4140 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4141 PAGE_SIZE << order);
4144 return kmalloc_large_node_hook(ptr, size, flags);
4147 void *__kmalloc_node(size_t size, gfp_t flags, int node)
4149 struct kmem_cache *s;
4152 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4153 ret = kmalloc_large_node(size, flags, node);
4155 trace_kmalloc_node(_RET_IP_, ret,
4156 size, PAGE_SIZE << get_order(size),
4162 s = kmalloc_slab(size, flags);
4164 if (unlikely(ZERO_OR_NULL_PTR(s)))
4167 ret = slab_alloc_node(s, flags, node, _RET_IP_, size);
4169 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4171 ret = kasan_kmalloc(s, ret, size, flags);
4175 EXPORT_SYMBOL(__kmalloc_node);
4176 #endif /* CONFIG_NUMA */
4178 #ifdef CONFIG_HARDENED_USERCOPY
4180 * Rejects incorrectly sized objects and objects that are to be copied
4181 * to/from userspace but do not fall entirely within the containing slab
4182 * cache's usercopy region.
4184 * Returns NULL if check passes, otherwise const char * to name of cache
4185 * to indicate an error.
4187 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4190 struct kmem_cache *s;
4191 unsigned int offset;
4193 bool is_kfence = is_kfence_address(ptr);
4195 ptr = kasan_reset_tag(ptr);
4197 /* Find object and usable object size. */
4198 s = page->slab_cache;
4200 /* Reject impossible pointers. */
4201 if (ptr < page_address(page))
4202 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4205 /* Find offset within object. */
4207 offset = ptr - kfence_object_start(ptr);
4209 offset = (ptr - page_address(page)) % s->size;
4211 /* Adjust for redzone and reject if within the redzone. */
4212 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4213 if (offset < s->red_left_pad)
4214 usercopy_abort("SLUB object in left red zone",
4215 s->name, to_user, offset, n);
4216 offset -= s->red_left_pad;
4219 /* Allow address range falling entirely within usercopy region. */
4220 if (offset >= s->useroffset &&
4221 offset - s->useroffset <= s->usersize &&
4222 n <= s->useroffset - offset + s->usersize)
4226 * If the copy is still within the allocated object, produce
4227 * a warning instead of rejecting the copy. This is intended
4228 * to be a temporary method to find any missing usercopy
4231 object_size = slab_ksize(s);
4232 if (usercopy_fallback &&
4233 offset <= object_size && n <= object_size - offset) {
4234 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4238 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4240 #endif /* CONFIG_HARDENED_USERCOPY */
4242 size_t __ksize(const void *object)
4246 if (unlikely(object == ZERO_SIZE_PTR))
4249 page = virt_to_head_page(object);
4251 if (unlikely(!PageSlab(page))) {
4252 WARN_ON(!PageCompound(page));
4253 return page_size(page);
4256 return slab_ksize(page->slab_cache);
4258 EXPORT_SYMBOL(__ksize);
4260 void kfree(const void *x)
4263 void *object = (void *)x;
4265 trace_kfree(_RET_IP_, x);
4267 if (unlikely(ZERO_OR_NULL_PTR(x)))
4270 page = virt_to_head_page(x);
4271 if (unlikely(!PageSlab(page))) {
4272 unsigned int order = compound_order(page);
4274 BUG_ON(!PageCompound(page));
4276 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4277 -(PAGE_SIZE << order));
4278 __free_pages(page, order);
4281 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4283 EXPORT_SYMBOL(kfree);
4285 #define SHRINK_PROMOTE_MAX 32
4288 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4289 * up most to the head of the partial lists. New allocations will then
4290 * fill those up and thus they can be removed from the partial lists.
4292 * The slabs with the least items are placed last. This results in them
4293 * being allocated from last increasing the chance that the last objects
4294 * are freed in them.
4296 int __kmem_cache_shrink(struct kmem_cache *s)
4300 struct kmem_cache_node *n;
4303 struct list_head discard;
4304 struct list_head promote[SHRINK_PROMOTE_MAX];
4305 unsigned long flags;
4309 for_each_kmem_cache_node(s, node, n) {
4310 INIT_LIST_HEAD(&discard);
4311 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4312 INIT_LIST_HEAD(promote + i);
4314 spin_lock_irqsave(&n->list_lock, flags);
4317 * Build lists of slabs to discard or promote.
4319 * Note that concurrent frees may occur while we hold the
4320 * list_lock. page->inuse here is the upper limit.
4322 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4323 int free = page->objects - page->inuse;
4325 /* Do not reread page->inuse */
4328 /* We do not keep full slabs on the list */
4331 if (free == page->objects) {
4332 list_move(&page->slab_list, &discard);
4334 } else if (free <= SHRINK_PROMOTE_MAX)
4335 list_move(&page->slab_list, promote + free - 1);
4339 * Promote the slabs filled up most to the head of the
4342 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4343 list_splice(promote + i, &n->partial);
4345 spin_unlock_irqrestore(&n->list_lock, flags);
4347 /* Release empty slabs */
4348 list_for_each_entry_safe(page, t, &discard, slab_list)
4349 discard_slab(s, page);
4351 if (slabs_node(s, node))
4358 static int slab_mem_going_offline_callback(void *arg)
4360 struct kmem_cache *s;
4362 mutex_lock(&slab_mutex);
4363 list_for_each_entry(s, &slab_caches, list)
4364 __kmem_cache_shrink(s);
4365 mutex_unlock(&slab_mutex);
4370 static void slab_mem_offline_callback(void *arg)
4372 struct memory_notify *marg = arg;
4375 offline_node = marg->status_change_nid_normal;
4378 * If the node still has available memory. we need kmem_cache_node
4381 if (offline_node < 0)
4384 mutex_lock(&slab_mutex);
4385 node_clear(offline_node, slab_nodes);
4387 * We no longer free kmem_cache_node structures here, as it would be
4388 * racy with all get_node() users, and infeasible to protect them with
4391 mutex_unlock(&slab_mutex);
4394 static int slab_mem_going_online_callback(void *arg)
4396 struct kmem_cache_node *n;
4397 struct kmem_cache *s;
4398 struct memory_notify *marg = arg;
4399 int nid = marg->status_change_nid_normal;
4403 * If the node's memory is already available, then kmem_cache_node is
4404 * already created. Nothing to do.
4410 * We are bringing a node online. No memory is available yet. We must
4411 * allocate a kmem_cache_node structure in order to bring the node
4414 mutex_lock(&slab_mutex);
4415 list_for_each_entry(s, &slab_caches, list) {
4417 * The structure may already exist if the node was previously
4418 * onlined and offlined.
4420 if (get_node(s, nid))
4423 * XXX: kmem_cache_alloc_node will fallback to other nodes
4424 * since memory is not yet available from the node that
4427 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4432 init_kmem_cache_node(n);
4436 * Any cache created after this point will also have kmem_cache_node
4437 * initialized for the new node.
4439 node_set(nid, slab_nodes);
4441 mutex_unlock(&slab_mutex);
4445 static int slab_memory_callback(struct notifier_block *self,
4446 unsigned long action, void *arg)
4451 case MEM_GOING_ONLINE:
4452 ret = slab_mem_going_online_callback(arg);
4454 case MEM_GOING_OFFLINE:
4455 ret = slab_mem_going_offline_callback(arg);
4458 case MEM_CANCEL_ONLINE:
4459 slab_mem_offline_callback(arg);
4462 case MEM_CANCEL_OFFLINE:
4466 ret = notifier_from_errno(ret);
4472 static struct notifier_block slab_memory_callback_nb = {
4473 .notifier_call = slab_memory_callback,
4474 .priority = SLAB_CALLBACK_PRI,
4477 /********************************************************************
4478 * Basic setup of slabs
4479 *******************************************************************/
4482 * Used for early kmem_cache structures that were allocated using
4483 * the page allocator. Allocate them properly then fix up the pointers
4484 * that may be pointing to the wrong kmem_cache structure.
4487 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4490 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4491 struct kmem_cache_node *n;
4493 memcpy(s, static_cache, kmem_cache->object_size);
4496 * This runs very early, and only the boot processor is supposed to be
4497 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4500 __flush_cpu_slab(s, smp_processor_id());
4501 for_each_kmem_cache_node(s, node, n) {
4504 list_for_each_entry(p, &n->partial, slab_list)
4507 #ifdef CONFIG_SLUB_DEBUG
4508 list_for_each_entry(p, &n->full, slab_list)
4512 list_add(&s->list, &slab_caches);
4516 void __init kmem_cache_init(void)
4518 static __initdata struct kmem_cache boot_kmem_cache,
4519 boot_kmem_cache_node;
4522 if (debug_guardpage_minorder())
4525 /* Print slub debugging pointers without hashing */
4526 if (__slub_debug_enabled())
4527 no_hash_pointers_enable(NULL);
4529 kmem_cache_node = &boot_kmem_cache_node;
4530 kmem_cache = &boot_kmem_cache;
4533 * Initialize the nodemask for which we will allocate per node
4534 * structures. Here we don't need taking slab_mutex yet.
4536 for_each_node_state(node, N_NORMAL_MEMORY)
4537 node_set(node, slab_nodes);
4539 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4540 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4542 register_hotmemory_notifier(&slab_memory_callback_nb);
4544 /* Able to allocate the per node structures */
4545 slab_state = PARTIAL;
4547 create_boot_cache(kmem_cache, "kmem_cache",
4548 offsetof(struct kmem_cache, node) +
4549 nr_node_ids * sizeof(struct kmem_cache_node *),
4550 SLAB_HWCACHE_ALIGN, 0, 0);
4552 kmem_cache = bootstrap(&boot_kmem_cache);
4553 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4555 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4556 setup_kmalloc_cache_index_table();
4557 create_kmalloc_caches(0);
4559 /* Setup random freelists for each cache */
4560 init_freelist_randomization();
4562 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4565 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4567 slub_min_order, slub_max_order, slub_min_objects,
4568 nr_cpu_ids, nr_node_ids);
4571 void __init kmem_cache_init_late(void)
4576 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4577 slab_flags_t flags, void (*ctor)(void *))
4579 struct kmem_cache *s;
4581 s = find_mergeable(size, align, flags, name, ctor);
4586 * Adjust the object sizes so that we clear
4587 * the complete object on kzalloc.
4589 s->object_size = max(s->object_size, size);
4590 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4592 if (sysfs_slab_alias(s, name)) {
4601 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4605 err = kmem_cache_open(s, flags);
4609 /* Mutex is not taken during early boot */
4610 if (slab_state <= UP)
4613 err = sysfs_slab_add(s);
4615 __kmem_cache_release(s);
4617 if (s->flags & SLAB_STORE_USER)
4618 debugfs_slab_add(s);
4623 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4625 struct kmem_cache *s;
4628 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4629 return kmalloc_large(size, gfpflags);
4631 s = kmalloc_slab(size, gfpflags);
4633 if (unlikely(ZERO_OR_NULL_PTR(s)))
4636 ret = slab_alloc(s, gfpflags, caller, size);
4638 /* Honor the call site pointer we received. */
4639 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4643 EXPORT_SYMBOL(__kmalloc_track_caller);
4646 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4647 int node, unsigned long caller)
4649 struct kmem_cache *s;
4652 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4653 ret = kmalloc_large_node(size, gfpflags, node);
4655 trace_kmalloc_node(caller, ret,
4656 size, PAGE_SIZE << get_order(size),
4662 s = kmalloc_slab(size, gfpflags);
4664 if (unlikely(ZERO_OR_NULL_PTR(s)))
4667 ret = slab_alloc_node(s, gfpflags, node, caller, size);
4669 /* Honor the call site pointer we received. */
4670 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4674 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4678 static int count_inuse(struct page *page)
4683 static int count_total(struct page *page)
4685 return page->objects;
4689 #ifdef CONFIG_SLUB_DEBUG
4690 static void validate_slab(struct kmem_cache *s, struct page *page)
4693 void *addr = page_address(page);
4698 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4701 /* Now we know that a valid freelist exists */
4702 map = get_map(s, page);
4703 for_each_object(p, s, addr, page->objects) {
4704 u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
4705 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4707 if (!check_object(s, page, p, val))
4715 static int validate_slab_node(struct kmem_cache *s,
4716 struct kmem_cache_node *n)
4718 unsigned long count = 0;
4720 unsigned long flags;
4722 spin_lock_irqsave(&n->list_lock, flags);
4724 list_for_each_entry(page, &n->partial, slab_list) {
4725 validate_slab(s, page);
4728 if (count != n->nr_partial) {
4729 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4730 s->name, count, n->nr_partial);
4731 slab_add_kunit_errors();
4734 if (!(s->flags & SLAB_STORE_USER))
4737 list_for_each_entry(page, &n->full, slab_list) {
4738 validate_slab(s, page);
4741 if (count != atomic_long_read(&n->nr_slabs)) {
4742 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4743 s->name, count, atomic_long_read(&n->nr_slabs));
4744 slab_add_kunit_errors();
4748 spin_unlock_irqrestore(&n->list_lock, flags);
4752 long validate_slab_cache(struct kmem_cache *s)
4755 unsigned long count = 0;
4756 struct kmem_cache_node *n;
4759 for_each_kmem_cache_node(s, node, n)
4760 count += validate_slab_node(s, n);
4764 EXPORT_SYMBOL(validate_slab_cache);
4766 #ifdef CONFIG_DEBUG_FS
4768 * Generate lists of code addresses where slabcache objects are allocated
4773 unsigned long count;
4780 DECLARE_BITMAP(cpus, NR_CPUS);
4786 unsigned long count;
4787 struct location *loc;
4790 static struct dentry *slab_debugfs_root;
4792 static void free_loc_track(struct loc_track *t)
4795 free_pages((unsigned long)t->loc,
4796 get_order(sizeof(struct location) * t->max));
4799 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4804 order = get_order(sizeof(struct location) * max);
4806 l = (void *)__get_free_pages(flags, order);
4811 memcpy(l, t->loc, sizeof(struct location) * t->count);
4819 static int add_location(struct loc_track *t, struct kmem_cache *s,
4820 const struct track *track)
4822 long start, end, pos;
4824 unsigned long caddr;
4825 unsigned long age = jiffies - track->when;
4831 pos = start + (end - start + 1) / 2;
4834 * There is nothing at "end". If we end up there
4835 * we need to add something to before end.
4840 caddr = t->loc[pos].addr;
4841 if (track->addr == caddr) {
4847 if (age < l->min_time)
4849 if (age > l->max_time)
4852 if (track->pid < l->min_pid)
4853 l->min_pid = track->pid;
4854 if (track->pid > l->max_pid)
4855 l->max_pid = track->pid;
4857 cpumask_set_cpu(track->cpu,
4858 to_cpumask(l->cpus));
4860 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4864 if (track->addr < caddr)
4871 * Not found. Insert new tracking element.
4873 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4879 (t->count - pos) * sizeof(struct location));
4882 l->addr = track->addr;
4886 l->min_pid = track->pid;
4887 l->max_pid = track->pid;
4888 cpumask_clear(to_cpumask(l->cpus));
4889 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4890 nodes_clear(l->nodes);
4891 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4895 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4896 struct page *page, enum track_item alloc)
4898 void *addr = page_address(page);
4902 map = get_map(s, page);
4903 for_each_object(p, s, addr, page->objects)
4904 if (!test_bit(__obj_to_index(s, addr, p), map))
4905 add_location(t, s, get_track(s, p, alloc));
4908 #endif /* CONFIG_DEBUG_FS */
4909 #endif /* CONFIG_SLUB_DEBUG */
4912 enum slab_stat_type {
4913 SL_ALL, /* All slabs */
4914 SL_PARTIAL, /* Only partially allocated slabs */
4915 SL_CPU, /* Only slabs used for cpu caches */
4916 SL_OBJECTS, /* Determine allocated objects not slabs */
4917 SL_TOTAL /* Determine object capacity not slabs */
4920 #define SO_ALL (1 << SL_ALL)
4921 #define SO_PARTIAL (1 << SL_PARTIAL)
4922 #define SO_CPU (1 << SL_CPU)
4923 #define SO_OBJECTS (1 << SL_OBJECTS)
4924 #define SO_TOTAL (1 << SL_TOTAL)
4926 static ssize_t show_slab_objects(struct kmem_cache *s,
4927 char *buf, unsigned long flags)
4929 unsigned long total = 0;
4932 unsigned long *nodes;
4935 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4939 if (flags & SO_CPU) {
4942 for_each_possible_cpu(cpu) {
4943 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4948 page = READ_ONCE(c->page);
4952 node = page_to_nid(page);
4953 if (flags & SO_TOTAL)
4955 else if (flags & SO_OBJECTS)
4963 page = slub_percpu_partial_read_once(c);
4965 node = page_to_nid(page);
4966 if (flags & SO_TOTAL)
4968 else if (flags & SO_OBJECTS)
4979 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4980 * already held which will conflict with an existing lock order:
4982 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4984 * We don't really need mem_hotplug_lock (to hold off
4985 * slab_mem_going_offline_callback) here because slab's memory hot
4986 * unplug code doesn't destroy the kmem_cache->node[] data.
4989 #ifdef CONFIG_SLUB_DEBUG
4990 if (flags & SO_ALL) {
4991 struct kmem_cache_node *n;
4993 for_each_kmem_cache_node(s, node, n) {
4995 if (flags & SO_TOTAL)
4996 x = atomic_long_read(&n->total_objects);
4997 else if (flags & SO_OBJECTS)
4998 x = atomic_long_read(&n->total_objects) -
4999 count_partial(n, count_free);
5001 x = atomic_long_read(&n->nr_slabs);
5008 if (flags & SO_PARTIAL) {
5009 struct kmem_cache_node *n;
5011 for_each_kmem_cache_node(s, node, n) {
5012 if (flags & SO_TOTAL)
5013 x = count_partial(n, count_total);
5014 else if (flags & SO_OBJECTS)
5015 x = count_partial(n, count_inuse);
5023 len += sysfs_emit_at(buf, len, "%lu", total);
5025 for (node = 0; node < nr_node_ids; node++) {
5027 len += sysfs_emit_at(buf, len, " N%d=%lu",
5031 len += sysfs_emit_at(buf, len, "\n");
5037 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5038 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5040 struct slab_attribute {
5041 struct attribute attr;
5042 ssize_t (*show)(struct kmem_cache *s, char *buf);
5043 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5046 #define SLAB_ATTR_RO(_name) \
5047 static struct slab_attribute _name##_attr = \
5048 __ATTR(_name, 0400, _name##_show, NULL)
5050 #define SLAB_ATTR(_name) \
5051 static struct slab_attribute _name##_attr = \
5052 __ATTR(_name, 0600, _name##_show, _name##_store)
5054 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5056 return sysfs_emit(buf, "%u\n", s->size);
5058 SLAB_ATTR_RO(slab_size);
5060 static ssize_t align_show(struct kmem_cache *s, char *buf)
5062 return sysfs_emit(buf, "%u\n", s->align);
5064 SLAB_ATTR_RO(align);
5066 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5068 return sysfs_emit(buf, "%u\n", s->object_size);
5070 SLAB_ATTR_RO(object_size);
5072 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5074 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5076 SLAB_ATTR_RO(objs_per_slab);
5078 static ssize_t order_show(struct kmem_cache *s, char *buf)
5080 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5082 SLAB_ATTR_RO(order);
5084 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5086 return sysfs_emit(buf, "%lu\n", s->min_partial);
5089 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5095 err = kstrtoul(buf, 10, &min);
5099 set_min_partial(s, min);
5102 SLAB_ATTR(min_partial);
5104 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5106 return sysfs_emit(buf, "%u\n", slub_cpu_partial(s));
5109 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5112 unsigned int objects;
5115 err = kstrtouint(buf, 10, &objects);
5118 if (objects && !kmem_cache_has_cpu_partial(s))
5121 slub_set_cpu_partial(s, objects);
5125 SLAB_ATTR(cpu_partial);
5127 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5131 return sysfs_emit(buf, "%pS\n", s->ctor);
5135 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5137 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5139 SLAB_ATTR_RO(aliases);
5141 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5143 return show_slab_objects(s, buf, SO_PARTIAL);
5145 SLAB_ATTR_RO(partial);
5147 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5149 return show_slab_objects(s, buf, SO_CPU);
5151 SLAB_ATTR_RO(cpu_slabs);
5153 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5155 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5157 SLAB_ATTR_RO(objects);
5159 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5161 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5163 SLAB_ATTR_RO(objects_partial);
5165 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5172 for_each_online_cpu(cpu) {
5175 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5178 pages += page->pages;
5179 objects += page->pobjects;
5183 len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages);
5186 for_each_online_cpu(cpu) {
5189 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5191 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5192 cpu, page->pobjects, page->pages);
5195 len += sysfs_emit_at(buf, len, "\n");
5199 SLAB_ATTR_RO(slabs_cpu_partial);
5201 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5203 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5205 SLAB_ATTR_RO(reclaim_account);
5207 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5209 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5211 SLAB_ATTR_RO(hwcache_align);
5213 #ifdef CONFIG_ZONE_DMA
5214 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5216 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5218 SLAB_ATTR_RO(cache_dma);
5221 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5223 return sysfs_emit(buf, "%u\n", s->usersize);
5225 SLAB_ATTR_RO(usersize);
5227 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5229 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5231 SLAB_ATTR_RO(destroy_by_rcu);
5233 #ifdef CONFIG_SLUB_DEBUG
5234 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5236 return show_slab_objects(s, buf, SO_ALL);
5238 SLAB_ATTR_RO(slabs);
5240 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5242 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5244 SLAB_ATTR_RO(total_objects);
5246 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5248 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5250 SLAB_ATTR_RO(sanity_checks);
5252 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5254 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5256 SLAB_ATTR_RO(trace);
5258 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5260 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5263 SLAB_ATTR_RO(red_zone);
5265 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5267 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5270 SLAB_ATTR_RO(poison);
5272 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5274 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5277 SLAB_ATTR_RO(store_user);
5279 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5284 static ssize_t validate_store(struct kmem_cache *s,
5285 const char *buf, size_t length)
5289 if (buf[0] == '1') {
5290 ret = validate_slab_cache(s);
5296 SLAB_ATTR(validate);
5298 #endif /* CONFIG_SLUB_DEBUG */
5300 #ifdef CONFIG_FAILSLAB
5301 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5303 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5305 SLAB_ATTR_RO(failslab);
5308 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5313 static ssize_t shrink_store(struct kmem_cache *s,
5314 const char *buf, size_t length)
5317 kmem_cache_shrink(s);
5325 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5327 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5330 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5331 const char *buf, size_t length)
5336 err = kstrtouint(buf, 10, &ratio);
5342 s->remote_node_defrag_ratio = ratio * 10;
5346 SLAB_ATTR(remote_node_defrag_ratio);
5349 #ifdef CONFIG_SLUB_STATS
5350 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5352 unsigned long sum = 0;
5355 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5360 for_each_online_cpu(cpu) {
5361 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5367 len += sysfs_emit_at(buf, len, "%lu", sum);
5370 for_each_online_cpu(cpu) {
5372 len += sysfs_emit_at(buf, len, " C%d=%u",
5377 len += sysfs_emit_at(buf, len, "\n");
5382 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5386 for_each_online_cpu(cpu)
5387 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5390 #define STAT_ATTR(si, text) \
5391 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5393 return show_stat(s, buf, si); \
5395 static ssize_t text##_store(struct kmem_cache *s, \
5396 const char *buf, size_t length) \
5398 if (buf[0] != '0') \
5400 clear_stat(s, si); \
5405 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5406 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5407 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5408 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5409 STAT_ATTR(FREE_FROZEN, free_frozen);
5410 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5411 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5412 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5413 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5414 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5415 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5416 STAT_ATTR(FREE_SLAB, free_slab);
5417 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5418 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5419 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5420 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5421 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5422 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5423 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5424 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5425 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5426 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5427 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5428 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5429 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5430 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5431 #endif /* CONFIG_SLUB_STATS */
5433 static struct attribute *slab_attrs[] = {
5434 &slab_size_attr.attr,
5435 &object_size_attr.attr,
5436 &objs_per_slab_attr.attr,
5438 &min_partial_attr.attr,
5439 &cpu_partial_attr.attr,
5441 &objects_partial_attr.attr,
5443 &cpu_slabs_attr.attr,
5447 &hwcache_align_attr.attr,
5448 &reclaim_account_attr.attr,
5449 &destroy_by_rcu_attr.attr,
5451 &slabs_cpu_partial_attr.attr,
5452 #ifdef CONFIG_SLUB_DEBUG
5453 &total_objects_attr.attr,
5455 &sanity_checks_attr.attr,
5457 &red_zone_attr.attr,
5459 &store_user_attr.attr,
5460 &validate_attr.attr,
5462 #ifdef CONFIG_ZONE_DMA
5463 &cache_dma_attr.attr,
5466 &remote_node_defrag_ratio_attr.attr,
5468 #ifdef CONFIG_SLUB_STATS
5469 &alloc_fastpath_attr.attr,
5470 &alloc_slowpath_attr.attr,
5471 &free_fastpath_attr.attr,
5472 &free_slowpath_attr.attr,
5473 &free_frozen_attr.attr,
5474 &free_add_partial_attr.attr,
5475 &free_remove_partial_attr.attr,
5476 &alloc_from_partial_attr.attr,
5477 &alloc_slab_attr.attr,
5478 &alloc_refill_attr.attr,
5479 &alloc_node_mismatch_attr.attr,
5480 &free_slab_attr.attr,
5481 &cpuslab_flush_attr.attr,
5482 &deactivate_full_attr.attr,
5483 &deactivate_empty_attr.attr,
5484 &deactivate_to_head_attr.attr,
5485 &deactivate_to_tail_attr.attr,
5486 &deactivate_remote_frees_attr.attr,
5487 &deactivate_bypass_attr.attr,
5488 &order_fallback_attr.attr,
5489 &cmpxchg_double_fail_attr.attr,
5490 &cmpxchg_double_cpu_fail_attr.attr,
5491 &cpu_partial_alloc_attr.attr,
5492 &cpu_partial_free_attr.attr,
5493 &cpu_partial_node_attr.attr,
5494 &cpu_partial_drain_attr.attr,
5496 #ifdef CONFIG_FAILSLAB
5497 &failslab_attr.attr,
5499 &usersize_attr.attr,
5504 static const struct attribute_group slab_attr_group = {
5505 .attrs = slab_attrs,
5508 static ssize_t slab_attr_show(struct kobject *kobj,
5509 struct attribute *attr,
5512 struct slab_attribute *attribute;
5513 struct kmem_cache *s;
5516 attribute = to_slab_attr(attr);
5519 if (!attribute->show)
5522 err = attribute->show(s, buf);
5527 static ssize_t slab_attr_store(struct kobject *kobj,
5528 struct attribute *attr,
5529 const char *buf, size_t len)
5531 struct slab_attribute *attribute;
5532 struct kmem_cache *s;
5535 attribute = to_slab_attr(attr);
5538 if (!attribute->store)
5541 err = attribute->store(s, buf, len);
5545 static void kmem_cache_release(struct kobject *k)
5547 slab_kmem_cache_release(to_slab(k));
5550 static const struct sysfs_ops slab_sysfs_ops = {
5551 .show = slab_attr_show,
5552 .store = slab_attr_store,
5555 static struct kobj_type slab_ktype = {
5556 .sysfs_ops = &slab_sysfs_ops,
5557 .release = kmem_cache_release,
5560 static struct kset *slab_kset;
5562 static inline struct kset *cache_kset(struct kmem_cache *s)
5567 #define ID_STR_LENGTH 64
5569 /* Create a unique string id for a slab cache:
5571 * Format :[flags-]size
5573 static char *create_unique_id(struct kmem_cache *s)
5575 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5582 * First flags affecting slabcache operations. We will only
5583 * get here for aliasable slabs so we do not need to support
5584 * too many flags. The flags here must cover all flags that
5585 * are matched during merging to guarantee that the id is
5588 if (s->flags & SLAB_CACHE_DMA)
5590 if (s->flags & SLAB_CACHE_DMA32)
5592 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5594 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5596 if (s->flags & SLAB_ACCOUNT)
5600 p += sprintf(p, "%07u", s->size);
5602 BUG_ON(p > name + ID_STR_LENGTH - 1);
5606 static int sysfs_slab_add(struct kmem_cache *s)
5610 struct kset *kset = cache_kset(s);
5611 int unmergeable = slab_unmergeable(s);
5614 kobject_init(&s->kobj, &slab_ktype);
5618 if (!unmergeable && disable_higher_order_debug &&
5619 (slub_debug & DEBUG_METADATA_FLAGS))
5624 * Slabcache can never be merged so we can use the name proper.
5625 * This is typically the case for debug situations. In that
5626 * case we can catch duplicate names easily.
5628 sysfs_remove_link(&slab_kset->kobj, s->name);
5632 * Create a unique name for the slab as a target
5635 name = create_unique_id(s);
5638 s->kobj.kset = kset;
5639 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5643 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5648 /* Setup first alias */
5649 sysfs_slab_alias(s, s->name);
5656 kobject_del(&s->kobj);
5660 void sysfs_slab_unlink(struct kmem_cache *s)
5662 if (slab_state >= FULL)
5663 kobject_del(&s->kobj);
5666 void sysfs_slab_release(struct kmem_cache *s)
5668 if (slab_state >= FULL)
5669 kobject_put(&s->kobj);
5673 * Need to buffer aliases during bootup until sysfs becomes
5674 * available lest we lose that information.
5676 struct saved_alias {
5677 struct kmem_cache *s;
5679 struct saved_alias *next;
5682 static struct saved_alias *alias_list;
5684 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5686 struct saved_alias *al;
5688 if (slab_state == FULL) {
5690 * If we have a leftover link then remove it.
5692 sysfs_remove_link(&slab_kset->kobj, name);
5693 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5696 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5702 al->next = alias_list;
5707 static int __init slab_sysfs_init(void)
5709 struct kmem_cache *s;
5712 mutex_lock(&slab_mutex);
5714 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5716 mutex_unlock(&slab_mutex);
5717 pr_err("Cannot register slab subsystem.\n");
5723 list_for_each_entry(s, &slab_caches, list) {
5724 err = sysfs_slab_add(s);
5726 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5730 while (alias_list) {
5731 struct saved_alias *al = alias_list;
5733 alias_list = alias_list->next;
5734 err = sysfs_slab_alias(al->s, al->name);
5736 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5741 mutex_unlock(&slab_mutex);
5745 __initcall(slab_sysfs_init);
5746 #endif /* CONFIG_SYSFS */
5748 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
5749 static int slab_debugfs_show(struct seq_file *seq, void *v)
5753 unsigned int idx = *(unsigned int *)v;
5754 struct loc_track *t = seq->private;
5756 if (idx < t->count) {
5759 seq_printf(seq, "%7ld ", l->count);
5762 seq_printf(seq, "%pS", (void *)l->addr);
5764 seq_puts(seq, "<not-available>");
5766 if (l->sum_time != l->min_time) {
5767 seq_printf(seq, " age=%ld/%llu/%ld",
5768 l->min_time, div_u64(l->sum_time, l->count),
5771 seq_printf(seq, " age=%ld", l->min_time);
5773 if (l->min_pid != l->max_pid)
5774 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
5776 seq_printf(seq, " pid=%ld",
5779 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
5780 seq_printf(seq, " cpus=%*pbl",
5781 cpumask_pr_args(to_cpumask(l->cpus)));
5783 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
5784 seq_printf(seq, " nodes=%*pbl",
5785 nodemask_pr_args(&l->nodes));
5787 seq_puts(seq, "\n");
5790 if (!idx && !t->count)
5791 seq_puts(seq, "No data\n");
5796 static void slab_debugfs_stop(struct seq_file *seq, void *v)
5800 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
5802 struct loc_track *t = seq->private;
5806 if (*ppos <= t->count)
5812 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
5817 static const struct seq_operations slab_debugfs_sops = {
5818 .start = slab_debugfs_start,
5819 .next = slab_debugfs_next,
5820 .stop = slab_debugfs_stop,
5821 .show = slab_debugfs_show,
5824 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
5827 struct kmem_cache_node *n;
5828 enum track_item alloc;
5830 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
5831 sizeof(struct loc_track));
5832 struct kmem_cache *s = file_inode(filep)->i_private;
5834 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
5835 alloc = TRACK_ALLOC;
5839 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL))
5842 /* Push back cpu slabs */
5845 for_each_kmem_cache_node(s, node, n) {
5846 unsigned long flags;
5849 if (!atomic_long_read(&n->nr_slabs))
5852 spin_lock_irqsave(&n->list_lock, flags);
5853 list_for_each_entry(page, &n->partial, slab_list)
5854 process_slab(t, s, page, alloc);
5855 list_for_each_entry(page, &n->full, slab_list)
5856 process_slab(t, s, page, alloc);
5857 spin_unlock_irqrestore(&n->list_lock, flags);
5863 static int slab_debug_trace_release(struct inode *inode, struct file *file)
5865 struct seq_file *seq = file->private_data;
5866 struct loc_track *t = seq->private;
5869 return seq_release_private(inode, file);
5872 static const struct file_operations slab_debugfs_fops = {
5873 .open = slab_debug_trace_open,
5875 .llseek = seq_lseek,
5876 .release = slab_debug_trace_release,
5879 static void debugfs_slab_add(struct kmem_cache *s)
5881 struct dentry *slab_cache_dir;
5883 if (unlikely(!slab_debugfs_root))
5886 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
5888 debugfs_create_file("alloc_traces", 0400,
5889 slab_cache_dir, s, &slab_debugfs_fops);
5891 debugfs_create_file("free_traces", 0400,
5892 slab_cache_dir, s, &slab_debugfs_fops);
5895 void debugfs_slab_release(struct kmem_cache *s)
5897 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
5900 static int __init slab_debugfs_init(void)
5902 struct kmem_cache *s;
5904 slab_debugfs_root = debugfs_create_dir("slab", NULL);
5906 list_for_each_entry(s, &slab_caches, list)
5907 if (s->flags & SLAB_STORE_USER)
5908 debugfs_slab_add(s);
5913 __initcall(slab_debugfs_init);
5916 * The /proc/slabinfo ABI
5918 #ifdef CONFIG_SLUB_DEBUG
5919 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5921 unsigned long nr_slabs = 0;
5922 unsigned long nr_objs = 0;
5923 unsigned long nr_free = 0;
5925 struct kmem_cache_node *n;
5927 for_each_kmem_cache_node(s, node, n) {
5928 nr_slabs += node_nr_slabs(n);
5929 nr_objs += node_nr_objs(n);
5930 nr_free += count_partial(n, count_free);
5933 sinfo->active_objs = nr_objs - nr_free;
5934 sinfo->num_objs = nr_objs;
5935 sinfo->active_slabs = nr_slabs;
5936 sinfo->num_slabs = nr_slabs;
5937 sinfo->objects_per_slab = oo_objects(s->oo);
5938 sinfo->cache_order = oo_order(s->oo);
5941 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5945 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5946 size_t count, loff_t *ppos)
5950 #endif /* CONFIG_SLUB_DEBUG */