1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * page->frozen The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 #ifdef CONFIG_SLUB_DEBUG
118 #ifdef CONFIG_SLUB_DEBUG_ON
119 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
121 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
125 static inline bool kmem_cache_debug(struct kmem_cache *s)
127 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
130 void *fixup_red_left(struct kmem_cache *s, void *p)
132 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
133 p += s->red_left_pad;
138 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
140 #ifdef CONFIG_SLUB_CPU_PARTIAL
141 return !kmem_cache_debug(s);
148 * Issues still to be resolved:
150 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
152 * - Variable sizing of the per node arrays
155 /* Enable to test recovery from slab corruption on boot */
156 #undef SLUB_RESILIENCY_TEST
158 /* Enable to log cmpxchg failures */
159 #undef SLUB_DEBUG_CMPXCHG
162 * Mininum number of partial slabs. These will be left on the partial
163 * lists even if they are empty. kmem_cache_shrink may reclaim them.
165 #define MIN_PARTIAL 5
168 * Maximum number of desirable partial slabs.
169 * The existence of more partial slabs makes kmem_cache_shrink
170 * sort the partial list by the number of objects in use.
172 #define MAX_PARTIAL 10
174 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_STORE_USER)
178 * These debug flags cannot use CMPXCHG because there might be consistency
179 * issues when checking or reading debug information
181 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
186 * Debugging flags that require metadata to be stored in the slab. These get
187 * disabled when slub_debug=O is used and a cache's min order increases with
190 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
193 #define OO_MASK ((1 << OO_SHIFT) - 1)
194 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
196 /* Internal SLUB flags */
198 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
199 /* Use cmpxchg_double */
200 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
203 * Tracking user of a slab.
205 #define TRACK_ADDRS_COUNT 16
207 unsigned long addr; /* Called from address */
208 #ifdef CONFIG_STACKTRACE
209 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
211 int cpu; /* Was running on cpu */
212 int pid; /* Pid context */
213 unsigned long when; /* When did the operation occur */
216 enum track_item { TRACK_ALLOC, TRACK_FREE };
219 static int sysfs_slab_add(struct kmem_cache *);
220 static int sysfs_slab_alias(struct kmem_cache *, const char *);
221 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
222 static void sysfs_slab_remove(struct kmem_cache *s);
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)
227 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
228 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
231 static inline void stat(const struct kmem_cache *s, enum stat_item si)
233 #ifdef CONFIG_SLUB_STATS
235 * The rmw is racy on a preemptible kernel but this is acceptable, so
236 * avoid this_cpu_add()'s irq-disable overhead.
238 raw_cpu_inc(s->cpu_slab->stat[si]);
242 /********************************************************************
243 * Core slab cache functions
244 *******************************************************************/
247 * Returns freelist pointer (ptr). With hardening, this is obfuscated
248 * with an XOR of the address where the pointer is held and a per-cache
251 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
252 unsigned long ptr_addr)
254 #ifdef CONFIG_SLAB_FREELIST_HARDENED
256 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
257 * Normally, this doesn't cause any issues, as both set_freepointer()
258 * and get_freepointer() are called with a pointer with the same tag.
259 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
260 * example, when __free_slub() iterates over objects in a cache, it
261 * passes untagged pointers to check_object(). check_object() in turns
262 * calls get_freepointer() with an untagged pointer, which causes the
263 * freepointer to be restored incorrectly.
265 return (void *)((unsigned long)ptr ^ s->random ^
266 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
272 /* Returns the freelist pointer recorded at location ptr_addr. */
273 static inline void *freelist_dereference(const struct kmem_cache *s,
276 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
277 (unsigned long)ptr_addr);
280 static inline void *get_freepointer(struct kmem_cache *s, void *object)
282 return freelist_dereference(s, object + s->offset);
285 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
287 prefetch(object + s->offset);
290 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
292 unsigned long freepointer_addr;
295 if (!debug_pagealloc_enabled_static())
296 return get_freepointer(s, object);
298 freepointer_addr = (unsigned long)object + s->offset;
299 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
300 return freelist_ptr(s, p, freepointer_addr);
303 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
305 unsigned long freeptr_addr = (unsigned long)object + s->offset;
307 #ifdef CONFIG_SLAB_FREELIST_HARDENED
308 BUG_ON(object == fp); /* naive detection of double free or corruption */
311 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
314 /* Loop over all objects in a slab */
315 #define for_each_object(__p, __s, __addr, __objects) \
316 for (__p = fixup_red_left(__s, __addr); \
317 __p < (__addr) + (__objects) * (__s)->size; \
320 /* Determine object index from a given position */
321 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
323 return (kasan_reset_tag(p) - addr) / s->size;
326 static inline unsigned int order_objects(unsigned int order, unsigned int size)
328 return ((unsigned int)PAGE_SIZE << order) / size;
331 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
334 struct kmem_cache_order_objects x = {
335 (order << OO_SHIFT) + order_objects(order, size)
341 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
343 return x.x >> OO_SHIFT;
346 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
348 return x.x & OO_MASK;
352 * Per slab locking using the pagelock
354 static __always_inline void slab_lock(struct page *page)
356 VM_BUG_ON_PAGE(PageTail(page), page);
357 bit_spin_lock(PG_locked, &page->flags);
360 static __always_inline void slab_unlock(struct page *page)
362 VM_BUG_ON_PAGE(PageTail(page), page);
363 __bit_spin_unlock(PG_locked, &page->flags);
366 /* Interrupts must be disabled (for the fallback code to work right) */
367 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
368 void *freelist_old, unsigned long counters_old,
369 void *freelist_new, unsigned long counters_new,
372 VM_BUG_ON(!irqs_disabled());
373 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
374 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
375 if (s->flags & __CMPXCHG_DOUBLE) {
376 if (cmpxchg_double(&page->freelist, &page->counters,
377 freelist_old, counters_old,
378 freelist_new, counters_new))
384 if (page->freelist == freelist_old &&
385 page->counters == counters_old) {
386 page->freelist = freelist_new;
387 page->counters = counters_new;
395 stat(s, CMPXCHG_DOUBLE_FAIL);
397 #ifdef SLUB_DEBUG_CMPXCHG
398 pr_info("%s %s: cmpxchg double redo ", n, s->name);
404 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
405 void *freelist_old, unsigned long counters_old,
406 void *freelist_new, unsigned long counters_new,
409 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
410 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
411 if (s->flags & __CMPXCHG_DOUBLE) {
412 if (cmpxchg_double(&page->freelist, &page->counters,
413 freelist_old, counters_old,
414 freelist_new, counters_new))
421 local_irq_save(flags);
423 if (page->freelist == freelist_old &&
424 page->counters == counters_old) {
425 page->freelist = freelist_new;
426 page->counters = counters_new;
428 local_irq_restore(flags);
432 local_irq_restore(flags);
436 stat(s, CMPXCHG_DOUBLE_FAIL);
438 #ifdef SLUB_DEBUG_CMPXCHG
439 pr_info("%s %s: cmpxchg double redo ", n, s->name);
445 #ifdef CONFIG_SLUB_DEBUG
446 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
447 static DEFINE_SPINLOCK(object_map_lock);
450 * Determine a map of object in use on a page.
452 * Node listlock must be held to guarantee that the page does
453 * not vanish from under us.
455 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
456 __acquires(&object_map_lock)
459 void *addr = page_address(page);
461 VM_BUG_ON(!irqs_disabled());
463 spin_lock(&object_map_lock);
465 bitmap_zero(object_map, page->objects);
467 for (p = page->freelist; p; p = get_freepointer(s, p))
468 set_bit(slab_index(p, s, addr), object_map);
473 static void put_map(unsigned long *map) __releases(&object_map_lock)
475 VM_BUG_ON(map != object_map);
476 spin_unlock(&object_map_lock);
479 static inline unsigned int size_from_object(struct kmem_cache *s)
481 if (s->flags & SLAB_RED_ZONE)
482 return s->size - s->red_left_pad;
487 static inline void *restore_red_left(struct kmem_cache *s, void *p)
489 if (s->flags & SLAB_RED_ZONE)
490 p -= s->red_left_pad;
498 #if defined(CONFIG_SLUB_DEBUG_ON)
499 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
501 static slab_flags_t slub_debug;
504 static char *slub_debug_string;
505 static int disable_higher_order_debug;
508 * slub is about to manipulate internal object metadata. This memory lies
509 * outside the range of the allocated object, so accessing it would normally
510 * be reported by kasan as a bounds error. metadata_access_enable() is used
511 * to tell kasan that these accesses are OK.
513 static inline void metadata_access_enable(void)
515 kasan_disable_current();
518 static inline void metadata_access_disable(void)
520 kasan_enable_current();
527 /* Verify that a pointer has an address that is valid within a slab page */
528 static inline int check_valid_pointer(struct kmem_cache *s,
529 struct page *page, void *object)
536 base = page_address(page);
537 object = kasan_reset_tag(object);
538 object = restore_red_left(s, object);
539 if (object < base || object >= base + page->objects * s->size ||
540 (object - base) % s->size) {
547 static void print_section(char *level, char *text, u8 *addr,
550 metadata_access_enable();
551 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
553 metadata_access_disable();
557 * See comment in calculate_sizes().
559 static inline bool freeptr_outside_object(struct kmem_cache *s)
561 return s->offset >= s->inuse;
565 * Return offset of the end of info block which is inuse + free pointer if
566 * not overlapping with object.
568 static inline unsigned int get_info_end(struct kmem_cache *s)
570 if (freeptr_outside_object(s))
571 return s->inuse + sizeof(void *);
576 static struct track *get_track(struct kmem_cache *s, void *object,
577 enum track_item alloc)
581 p = object + get_info_end(s);
586 static void set_track(struct kmem_cache *s, void *object,
587 enum track_item alloc, unsigned long addr)
589 struct track *p = get_track(s, object, alloc);
592 #ifdef CONFIG_STACKTRACE
593 unsigned int nr_entries;
595 metadata_access_enable();
596 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
597 metadata_access_disable();
599 if (nr_entries < TRACK_ADDRS_COUNT)
600 p->addrs[nr_entries] = 0;
603 p->cpu = smp_processor_id();
604 p->pid = current->pid;
607 memset(p, 0, sizeof(struct track));
611 static void init_tracking(struct kmem_cache *s, void *object)
613 if (!(s->flags & SLAB_STORE_USER))
616 set_track(s, object, TRACK_FREE, 0UL);
617 set_track(s, object, TRACK_ALLOC, 0UL);
620 static void print_track(const char *s, struct track *t, unsigned long pr_time)
625 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
626 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
627 #ifdef CONFIG_STACKTRACE
630 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
632 pr_err("\t%pS\n", (void *)t->addrs[i]);
639 void print_tracking(struct kmem_cache *s, void *object)
641 unsigned long pr_time = jiffies;
642 if (!(s->flags & SLAB_STORE_USER))
645 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
646 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
649 static void print_page_info(struct page *page)
651 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
652 page, page->objects, page->inuse, page->freelist, page->flags);
656 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
658 struct va_format vaf;
664 pr_err("=============================================================================\n");
665 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
666 pr_err("-----------------------------------------------------------------------------\n\n");
668 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
672 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
674 struct va_format vaf;
680 pr_err("FIX %s: %pV\n", s->name, &vaf);
684 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
685 void *freelist, void *nextfree)
687 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
688 !check_valid_pointer(s, page, nextfree)) {
689 object_err(s, page, freelist, "Freechain corrupt");
691 slab_fix(s, "Isolate corrupted freechain");
698 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
700 unsigned int off; /* Offset of last byte */
701 u8 *addr = page_address(page);
703 print_tracking(s, p);
705 print_page_info(page);
707 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
708 p, p - addr, get_freepointer(s, p));
710 if (s->flags & SLAB_RED_ZONE)
711 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
713 else if (p > addr + 16)
714 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
716 print_section(KERN_ERR, "Object ", p,
717 min_t(unsigned int, s->object_size, PAGE_SIZE));
718 if (s->flags & SLAB_RED_ZONE)
719 print_section(KERN_ERR, "Redzone ", p + s->object_size,
720 s->inuse - s->object_size);
722 off = get_info_end(s);
724 if (s->flags & SLAB_STORE_USER)
725 off += 2 * sizeof(struct track);
727 off += kasan_metadata_size(s);
729 if (off != size_from_object(s))
730 /* Beginning of the filler is the free pointer */
731 print_section(KERN_ERR, "Padding ", p + off,
732 size_from_object(s) - off);
737 void object_err(struct kmem_cache *s, struct page *page,
738 u8 *object, char *reason)
740 slab_bug(s, "%s", reason);
741 print_trailer(s, page, object);
744 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
745 const char *fmt, ...)
751 vsnprintf(buf, sizeof(buf), fmt, args);
753 slab_bug(s, "%s", buf);
754 print_page_info(page);
758 static void init_object(struct kmem_cache *s, void *object, u8 val)
762 if (s->flags & SLAB_RED_ZONE)
763 memset(p - s->red_left_pad, val, s->red_left_pad);
765 if (s->flags & __OBJECT_POISON) {
766 memset(p, POISON_FREE, s->object_size - 1);
767 p[s->object_size - 1] = POISON_END;
770 if (s->flags & SLAB_RED_ZONE)
771 memset(p + s->object_size, val, s->inuse - s->object_size);
774 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
775 void *from, void *to)
777 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
778 memset(from, data, to - from);
781 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
782 u8 *object, char *what,
783 u8 *start, unsigned int value, unsigned int bytes)
787 u8 *addr = page_address(page);
789 metadata_access_enable();
790 fault = memchr_inv(start, value, bytes);
791 metadata_access_disable();
796 while (end > fault && end[-1] == value)
799 slab_bug(s, "%s overwritten", what);
800 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
801 fault, end - 1, fault - addr,
803 print_trailer(s, page, object);
805 restore_bytes(s, what, value, fault, end);
813 * Bytes of the object to be managed.
814 * If the freepointer may overlay the object then the free
815 * pointer is at the middle of the object.
817 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
820 * object + s->object_size
821 * Padding to reach word boundary. This is also used for Redzoning.
822 * Padding is extended by another word if Redzoning is enabled and
823 * object_size == inuse.
825 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
826 * 0xcc (RED_ACTIVE) for objects in use.
829 * Meta data starts here.
831 * A. Free pointer (if we cannot overwrite object on free)
832 * B. Tracking data for SLAB_STORE_USER
833 * C. Padding to reach required alignment boundary or at mininum
834 * one word if debugging is on to be able to detect writes
835 * before the word boundary.
837 * Padding is done using 0x5a (POISON_INUSE)
840 * Nothing is used beyond s->size.
842 * If slabcaches are merged then the object_size and inuse boundaries are mostly
843 * ignored. And therefore no slab options that rely on these boundaries
844 * may be used with merged slabcaches.
847 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
849 unsigned long off = get_info_end(s); /* The end of info */
851 if (s->flags & SLAB_STORE_USER)
852 /* We also have user information there */
853 off += 2 * sizeof(struct track);
855 off += kasan_metadata_size(s);
857 if (size_from_object(s) == off)
860 return check_bytes_and_report(s, page, p, "Object padding",
861 p + off, POISON_INUSE, size_from_object(s) - off);
864 /* Check the pad bytes at the end of a slab page */
865 static int slab_pad_check(struct kmem_cache *s, struct page *page)
874 if (!(s->flags & SLAB_POISON))
877 start = page_address(page);
878 length = page_size(page);
879 end = start + length;
880 remainder = length % s->size;
884 pad = end - remainder;
885 metadata_access_enable();
886 fault = memchr_inv(pad, POISON_INUSE, remainder);
887 metadata_access_disable();
890 while (end > fault && end[-1] == POISON_INUSE)
893 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
894 fault, end - 1, fault - start);
895 print_section(KERN_ERR, "Padding ", pad, remainder);
897 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
901 static int check_object(struct kmem_cache *s, struct page *page,
902 void *object, u8 val)
905 u8 *endobject = object + s->object_size;
907 if (s->flags & SLAB_RED_ZONE) {
908 if (!check_bytes_and_report(s, page, object, "Redzone",
909 object - s->red_left_pad, val, s->red_left_pad))
912 if (!check_bytes_and_report(s, page, object, "Redzone",
913 endobject, val, s->inuse - s->object_size))
916 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
917 check_bytes_and_report(s, page, p, "Alignment padding",
918 endobject, POISON_INUSE,
919 s->inuse - s->object_size);
923 if (s->flags & SLAB_POISON) {
924 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
925 (!check_bytes_and_report(s, page, p, "Poison", p,
926 POISON_FREE, s->object_size - 1) ||
927 !check_bytes_and_report(s, page, p, "Poison",
928 p + s->object_size - 1, POISON_END, 1)))
931 * check_pad_bytes cleans up on its own.
933 check_pad_bytes(s, page, p);
936 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
938 * Object and freepointer overlap. Cannot check
939 * freepointer while object is allocated.
943 /* Check free pointer validity */
944 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
945 object_err(s, page, p, "Freepointer corrupt");
947 * No choice but to zap it and thus lose the remainder
948 * of the free objects in this slab. May cause
949 * another error because the object count is now wrong.
951 set_freepointer(s, p, NULL);
957 static int check_slab(struct kmem_cache *s, struct page *page)
961 VM_BUG_ON(!irqs_disabled());
963 if (!PageSlab(page)) {
964 slab_err(s, page, "Not a valid slab page");
968 maxobj = order_objects(compound_order(page), s->size);
969 if (page->objects > maxobj) {
970 slab_err(s, page, "objects %u > max %u",
971 page->objects, maxobj);
974 if (page->inuse > page->objects) {
975 slab_err(s, page, "inuse %u > max %u",
976 page->inuse, page->objects);
979 /* Slab_pad_check fixes things up after itself */
980 slab_pad_check(s, page);
985 * Determine if a certain object on a page is on the freelist. Must hold the
986 * slab lock to guarantee that the chains are in a consistent state.
988 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
996 while (fp && nr <= page->objects) {
999 if (!check_valid_pointer(s, page, fp)) {
1001 object_err(s, page, object,
1002 "Freechain corrupt");
1003 set_freepointer(s, object, NULL);
1005 slab_err(s, page, "Freepointer corrupt");
1006 page->freelist = NULL;
1007 page->inuse = page->objects;
1008 slab_fix(s, "Freelist cleared");
1014 fp = get_freepointer(s, object);
1018 max_objects = order_objects(compound_order(page), s->size);
1019 if (max_objects > MAX_OBJS_PER_PAGE)
1020 max_objects = MAX_OBJS_PER_PAGE;
1022 if (page->objects != max_objects) {
1023 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1024 page->objects, max_objects);
1025 page->objects = max_objects;
1026 slab_fix(s, "Number of objects adjusted.");
1028 if (page->inuse != page->objects - nr) {
1029 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1030 page->inuse, page->objects - nr);
1031 page->inuse = page->objects - nr;
1032 slab_fix(s, "Object count adjusted.");
1034 return search == NULL;
1037 static void trace(struct kmem_cache *s, struct page *page, void *object,
1040 if (s->flags & SLAB_TRACE) {
1041 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1043 alloc ? "alloc" : "free",
1044 object, page->inuse,
1048 print_section(KERN_INFO, "Object ", (void *)object,
1056 * Tracking of fully allocated slabs for debugging purposes.
1058 static void add_full(struct kmem_cache *s,
1059 struct kmem_cache_node *n, struct page *page)
1061 if (!(s->flags & SLAB_STORE_USER))
1064 lockdep_assert_held(&n->list_lock);
1065 list_add(&page->slab_list, &n->full);
1068 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1070 if (!(s->flags & SLAB_STORE_USER))
1073 lockdep_assert_held(&n->list_lock);
1074 list_del(&page->slab_list);
1077 /* Tracking of the number of slabs for debugging purposes */
1078 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1080 struct kmem_cache_node *n = get_node(s, node);
1082 return atomic_long_read(&n->nr_slabs);
1085 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1087 return atomic_long_read(&n->nr_slabs);
1090 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1092 struct kmem_cache_node *n = get_node(s, node);
1095 * May be called early in order to allocate a slab for the
1096 * kmem_cache_node structure. Solve the chicken-egg
1097 * dilemma by deferring the increment of the count during
1098 * bootstrap (see early_kmem_cache_node_alloc).
1101 atomic_long_inc(&n->nr_slabs);
1102 atomic_long_add(objects, &n->total_objects);
1105 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1107 struct kmem_cache_node *n = get_node(s, node);
1109 atomic_long_dec(&n->nr_slabs);
1110 atomic_long_sub(objects, &n->total_objects);
1113 /* Object debug checks for alloc/free paths */
1114 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1117 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1120 init_object(s, object, SLUB_RED_INACTIVE);
1121 init_tracking(s, object);
1125 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1127 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1130 metadata_access_enable();
1131 memset(addr, POISON_INUSE, page_size(page));
1132 metadata_access_disable();
1135 static inline int alloc_consistency_checks(struct kmem_cache *s,
1136 struct page *page, void *object)
1138 if (!check_slab(s, page))
1141 if (!check_valid_pointer(s, page, object)) {
1142 object_err(s, page, object, "Freelist Pointer check fails");
1146 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1152 static noinline int alloc_debug_processing(struct kmem_cache *s,
1154 void *object, unsigned long addr)
1156 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1157 if (!alloc_consistency_checks(s, page, object))
1161 /* Success perform special debug activities for allocs */
1162 if (s->flags & SLAB_STORE_USER)
1163 set_track(s, object, TRACK_ALLOC, addr);
1164 trace(s, page, object, 1);
1165 init_object(s, object, SLUB_RED_ACTIVE);
1169 if (PageSlab(page)) {
1171 * If this is a slab page then lets do the best we can
1172 * to avoid issues in the future. Marking all objects
1173 * as used avoids touching the remaining objects.
1175 slab_fix(s, "Marking all objects used");
1176 page->inuse = page->objects;
1177 page->freelist = NULL;
1182 static inline int free_consistency_checks(struct kmem_cache *s,
1183 struct page *page, void *object, unsigned long addr)
1185 if (!check_valid_pointer(s, page, object)) {
1186 slab_err(s, page, "Invalid object pointer 0x%p", object);
1190 if (on_freelist(s, page, object)) {
1191 object_err(s, page, object, "Object already free");
1195 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1198 if (unlikely(s != page->slab_cache)) {
1199 if (!PageSlab(page)) {
1200 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1202 } else if (!page->slab_cache) {
1203 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1207 object_err(s, page, object,
1208 "page slab pointer corrupt.");
1214 /* Supports checking bulk free of a constructed freelist */
1215 static noinline int free_debug_processing(
1216 struct kmem_cache *s, struct page *page,
1217 void *head, void *tail, int bulk_cnt,
1220 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1221 void *object = head;
1223 unsigned long flags;
1226 spin_lock_irqsave(&n->list_lock, flags);
1229 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1230 if (!check_slab(s, page))
1237 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1238 if (!free_consistency_checks(s, page, object, addr))
1242 if (s->flags & SLAB_STORE_USER)
1243 set_track(s, object, TRACK_FREE, addr);
1244 trace(s, page, object, 0);
1245 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1246 init_object(s, object, SLUB_RED_INACTIVE);
1248 /* Reached end of constructed freelist yet? */
1249 if (object != tail) {
1250 object = get_freepointer(s, object);
1256 if (cnt != bulk_cnt)
1257 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1261 spin_unlock_irqrestore(&n->list_lock, flags);
1263 slab_fix(s, "Object at 0x%p not freed", object);
1268 * Parse a block of slub_debug options. Blocks are delimited by ';'
1270 * @str: start of block
1271 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1272 * @slabs: return start of list of slabs, or NULL when there's no list
1273 * @init: assume this is initial parsing and not per-kmem-create parsing
1275 * returns the start of next block if there's any, or NULL
1278 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1280 bool higher_order_disable = false;
1282 /* Skip any completely empty blocks */
1283 while (*str && *str == ';')
1288 * No options but restriction on slabs. This means full
1289 * debugging for slabs matching a pattern.
1291 *flags = DEBUG_DEFAULT_FLAGS;
1296 /* Determine which debug features should be switched on */
1297 for (; *str && *str != ',' && *str != ';'; str++) {
1298 switch (tolower(*str)) {
1303 *flags |= SLAB_CONSISTENCY_CHECKS;
1306 *flags |= SLAB_RED_ZONE;
1309 *flags |= SLAB_POISON;
1312 *flags |= SLAB_STORE_USER;
1315 *flags |= SLAB_TRACE;
1318 *flags |= SLAB_FAILSLAB;
1322 * Avoid enabling debugging on caches if its minimum
1323 * order would increase as a result.
1325 higher_order_disable = true;
1329 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1338 /* Skip over the slab list */
1339 while (*str && *str != ';')
1342 /* Skip any completely empty blocks */
1343 while (*str && *str == ';')
1346 if (init && higher_order_disable)
1347 disable_higher_order_debug = 1;
1355 static int __init setup_slub_debug(char *str)
1360 bool global_slub_debug_changed = false;
1361 bool slab_list_specified = false;
1363 slub_debug = DEBUG_DEFAULT_FLAGS;
1364 if (*str++ != '=' || !*str)
1366 * No options specified. Switch on full debugging.
1372 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1376 global_slub_debug_changed = true;
1378 slab_list_specified = true;
1383 * For backwards compatibility, a single list of flags with list of
1384 * slabs means debugging is only enabled for those slabs, so the global
1385 * slub_debug should be 0. We can extended that to multiple lists as
1386 * long as there is no option specifying flags without a slab list.
1388 if (slab_list_specified) {
1389 if (!global_slub_debug_changed)
1391 slub_debug_string = saved_str;
1394 if (slub_debug != 0 || slub_debug_string)
1395 static_branch_enable(&slub_debug_enabled);
1396 if ((static_branch_unlikely(&init_on_alloc) ||
1397 static_branch_unlikely(&init_on_free)) &&
1398 (slub_debug & SLAB_POISON))
1399 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1403 __setup("slub_debug", setup_slub_debug);
1406 * kmem_cache_flags - apply debugging options to the cache
1407 * @object_size: the size of an object without meta data
1408 * @flags: flags to set
1409 * @name: name of the cache
1410 * @ctor: constructor function
1412 * Debug option(s) are applied to @flags. In addition to the debug
1413 * option(s), if a slab name (or multiple) is specified i.e.
1414 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1415 * then only the select slabs will receive the debug option(s).
1417 slab_flags_t kmem_cache_flags(unsigned int object_size,
1418 slab_flags_t flags, const char *name,
1419 void (*ctor)(void *))
1424 slab_flags_t block_flags;
1426 /* If slub_debug = 0, it folds into the if conditional. */
1427 if (!slub_debug_string)
1428 return flags | slub_debug;
1431 next_block = slub_debug_string;
1432 /* Go through all blocks of debug options, see if any matches our slab's name */
1433 while (next_block) {
1434 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1437 /* Found a block that has a slab list, search it */
1442 end = strchrnul(iter, ',');
1443 if (next_block && next_block < end)
1444 end = next_block - 1;
1446 glob = strnchr(iter, end - iter, '*');
1448 cmplen = glob - iter;
1450 cmplen = max_t(size_t, len, (end - iter));
1452 if (!strncmp(name, iter, cmplen)) {
1453 flags |= block_flags;
1457 if (!*end || *end == ';')
1465 #else /* !CONFIG_SLUB_DEBUG */
1466 static inline void setup_object_debug(struct kmem_cache *s,
1467 struct page *page, void *object) {}
1469 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1471 static inline int alloc_debug_processing(struct kmem_cache *s,
1472 struct page *page, void *object, unsigned long addr) { return 0; }
1474 static inline int free_debug_processing(
1475 struct kmem_cache *s, struct page *page,
1476 void *head, void *tail, int bulk_cnt,
1477 unsigned long addr) { return 0; }
1479 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1481 static inline int check_object(struct kmem_cache *s, struct page *page,
1482 void *object, u8 val) { return 1; }
1483 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1484 struct page *page) {}
1485 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1486 struct page *page) {}
1487 slab_flags_t kmem_cache_flags(unsigned int object_size,
1488 slab_flags_t flags, const char *name,
1489 void (*ctor)(void *))
1493 #define slub_debug 0
1495 #define disable_higher_order_debug 0
1497 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1499 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1501 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1503 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1506 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1507 void *freelist, void *nextfree)
1511 #endif /* CONFIG_SLUB_DEBUG */
1514 * Hooks for other subsystems that check memory allocations. In a typical
1515 * production configuration these hooks all should produce no code at all.
1517 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1519 ptr = kasan_kmalloc_large(ptr, size, flags);
1520 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1521 kmemleak_alloc(ptr, size, 1, flags);
1525 static __always_inline void kfree_hook(void *x)
1528 kasan_kfree_large(x, _RET_IP_);
1531 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1533 kmemleak_free_recursive(x, s->flags);
1536 * Trouble is that we may no longer disable interrupts in the fast path
1537 * So in order to make the debug calls that expect irqs to be
1538 * disabled we need to disable interrupts temporarily.
1540 #ifdef CONFIG_LOCKDEP
1542 unsigned long flags;
1544 local_irq_save(flags);
1545 debug_check_no_locks_freed(x, s->object_size);
1546 local_irq_restore(flags);
1549 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1550 debug_check_no_obj_freed(x, s->object_size);
1552 /* Use KCSAN to help debug racy use-after-free. */
1553 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1554 __kcsan_check_access(x, s->object_size,
1555 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1557 /* KASAN might put x into memory quarantine, delaying its reuse */
1558 return kasan_slab_free(s, x, _RET_IP_);
1561 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1562 void **head, void **tail)
1567 void *old_tail = *tail ? *tail : *head;
1570 /* Head and tail of the reconstructed freelist */
1576 next = get_freepointer(s, object);
1578 if (slab_want_init_on_free(s)) {
1580 * Clear the object and the metadata, but don't touch
1583 memset(object, 0, s->object_size);
1584 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1586 memset((char *)object + s->inuse, 0,
1587 s->size - s->inuse - rsize);
1590 /* If object's reuse doesn't have to be delayed */
1591 if (!slab_free_hook(s, object)) {
1592 /* Move object to the new freelist */
1593 set_freepointer(s, object, *head);
1598 } while (object != old_tail);
1603 return *head != NULL;
1606 static void *setup_object(struct kmem_cache *s, struct page *page,
1609 setup_object_debug(s, page, object);
1610 object = kasan_init_slab_obj(s, object);
1611 if (unlikely(s->ctor)) {
1612 kasan_unpoison_object_data(s, object);
1614 kasan_poison_object_data(s, object);
1620 * Slab allocation and freeing
1622 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1623 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1626 unsigned int order = oo_order(oo);
1628 if (node == NUMA_NO_NODE)
1629 page = alloc_pages(flags, order);
1631 page = __alloc_pages_node(node, flags, order);
1633 if (page && charge_slab_page(page, flags, order, s)) {
1634 __free_pages(page, order);
1641 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1642 /* Pre-initialize the random sequence cache */
1643 static int init_cache_random_seq(struct kmem_cache *s)
1645 unsigned int count = oo_objects(s->oo);
1648 /* Bailout if already initialised */
1652 err = cache_random_seq_create(s, count, GFP_KERNEL);
1654 pr_err("SLUB: Unable to initialize free list for %s\n",
1659 /* Transform to an offset on the set of pages */
1660 if (s->random_seq) {
1663 for (i = 0; i < count; i++)
1664 s->random_seq[i] *= s->size;
1669 /* Initialize each random sequence freelist per cache */
1670 static void __init init_freelist_randomization(void)
1672 struct kmem_cache *s;
1674 mutex_lock(&slab_mutex);
1676 list_for_each_entry(s, &slab_caches, list)
1677 init_cache_random_seq(s);
1679 mutex_unlock(&slab_mutex);
1682 /* Get the next entry on the pre-computed freelist randomized */
1683 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1684 unsigned long *pos, void *start,
1685 unsigned long page_limit,
1686 unsigned long freelist_count)
1691 * If the target page allocation failed, the number of objects on the
1692 * page might be smaller than the usual size defined by the cache.
1695 idx = s->random_seq[*pos];
1697 if (*pos >= freelist_count)
1699 } while (unlikely(idx >= page_limit));
1701 return (char *)start + idx;
1704 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1705 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1710 unsigned long idx, pos, page_limit, freelist_count;
1712 if (page->objects < 2 || !s->random_seq)
1715 freelist_count = oo_objects(s->oo);
1716 pos = get_random_int() % freelist_count;
1718 page_limit = page->objects * s->size;
1719 start = fixup_red_left(s, page_address(page));
1721 /* First entry is used as the base of the freelist */
1722 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1724 cur = setup_object(s, page, cur);
1725 page->freelist = cur;
1727 for (idx = 1; idx < page->objects; idx++) {
1728 next = next_freelist_entry(s, page, &pos, start, page_limit,
1730 next = setup_object(s, page, next);
1731 set_freepointer(s, cur, next);
1734 set_freepointer(s, cur, NULL);
1739 static inline int init_cache_random_seq(struct kmem_cache *s)
1743 static inline void init_freelist_randomization(void) { }
1744 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1748 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1750 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1753 struct kmem_cache_order_objects oo = s->oo;
1755 void *start, *p, *next;
1759 flags &= gfp_allowed_mask;
1761 if (gfpflags_allow_blocking(flags))
1764 flags |= s->allocflags;
1767 * Let the initial higher-order allocation fail under memory pressure
1768 * so we fall-back to the minimum order allocation.
1770 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1771 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1772 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1774 page = alloc_slab_page(s, alloc_gfp, node, oo);
1775 if (unlikely(!page)) {
1779 * Allocation may have failed due to fragmentation.
1780 * Try a lower order alloc if possible
1782 page = alloc_slab_page(s, alloc_gfp, node, oo);
1783 if (unlikely(!page))
1785 stat(s, ORDER_FALLBACK);
1788 page->objects = oo_objects(oo);
1790 page->slab_cache = s;
1791 __SetPageSlab(page);
1792 if (page_is_pfmemalloc(page))
1793 SetPageSlabPfmemalloc(page);
1795 kasan_poison_slab(page);
1797 start = page_address(page);
1799 setup_page_debug(s, page, start);
1801 shuffle = shuffle_freelist(s, page);
1804 start = fixup_red_left(s, start);
1805 start = setup_object(s, page, start);
1806 page->freelist = start;
1807 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1809 next = setup_object(s, page, next);
1810 set_freepointer(s, p, next);
1813 set_freepointer(s, p, NULL);
1816 page->inuse = page->objects;
1820 if (gfpflags_allow_blocking(flags))
1821 local_irq_disable();
1825 inc_slabs_node(s, page_to_nid(page), page->objects);
1830 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1832 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1833 flags = kmalloc_fix_flags(flags);
1835 return allocate_slab(s,
1836 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1839 static void __free_slab(struct kmem_cache *s, struct page *page)
1841 int order = compound_order(page);
1842 int pages = 1 << order;
1844 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1847 slab_pad_check(s, page);
1848 for_each_object(p, s, page_address(page),
1850 check_object(s, page, p, SLUB_RED_INACTIVE);
1853 __ClearPageSlabPfmemalloc(page);
1854 __ClearPageSlab(page);
1856 page->mapping = NULL;
1857 if (current->reclaim_state)
1858 current->reclaim_state->reclaimed_slab += pages;
1859 uncharge_slab_page(page, order, s);
1860 __free_pages(page, order);
1863 static void rcu_free_slab(struct rcu_head *h)
1865 struct page *page = container_of(h, struct page, rcu_head);
1867 __free_slab(page->slab_cache, page);
1870 static void free_slab(struct kmem_cache *s, struct page *page)
1872 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1873 call_rcu(&page->rcu_head, rcu_free_slab);
1875 __free_slab(s, page);
1878 static void discard_slab(struct kmem_cache *s, struct page *page)
1880 dec_slabs_node(s, page_to_nid(page), page->objects);
1885 * Management of partially allocated slabs.
1888 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1891 if (tail == DEACTIVATE_TO_TAIL)
1892 list_add_tail(&page->slab_list, &n->partial);
1894 list_add(&page->slab_list, &n->partial);
1897 static inline void add_partial(struct kmem_cache_node *n,
1898 struct page *page, int tail)
1900 lockdep_assert_held(&n->list_lock);
1901 __add_partial(n, page, tail);
1904 static inline void remove_partial(struct kmem_cache_node *n,
1907 lockdep_assert_held(&n->list_lock);
1908 list_del(&page->slab_list);
1913 * Remove slab from the partial list, freeze it and
1914 * return the pointer to the freelist.
1916 * Returns a list of objects or NULL if it fails.
1918 static inline void *acquire_slab(struct kmem_cache *s,
1919 struct kmem_cache_node *n, struct page *page,
1920 int mode, int *objects)
1923 unsigned long counters;
1926 lockdep_assert_held(&n->list_lock);
1929 * Zap the freelist and set the frozen bit.
1930 * The old freelist is the list of objects for the
1931 * per cpu allocation list.
1933 freelist = page->freelist;
1934 counters = page->counters;
1935 new.counters = counters;
1936 *objects = new.objects - new.inuse;
1938 new.inuse = page->objects;
1939 new.freelist = NULL;
1941 new.freelist = freelist;
1944 VM_BUG_ON(new.frozen);
1947 if (!__cmpxchg_double_slab(s, page,
1949 new.freelist, new.counters,
1953 remove_partial(n, page);
1958 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1959 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1962 * Try to allocate a partial slab from a specific node.
1964 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1965 struct kmem_cache_cpu *c, gfp_t flags)
1967 struct page *page, *page2;
1968 void *object = NULL;
1969 unsigned int available = 0;
1973 * Racy check. If we mistakenly see no partial slabs then we
1974 * just allocate an empty slab. If we mistakenly try to get a
1975 * partial slab and there is none available then get_partials()
1978 if (!n || !n->nr_partial)
1981 spin_lock(&n->list_lock);
1982 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1985 if (!pfmemalloc_match(page, flags))
1988 t = acquire_slab(s, n, page, object == NULL, &objects);
1992 available += objects;
1995 stat(s, ALLOC_FROM_PARTIAL);
1998 put_cpu_partial(s, page, 0);
1999 stat(s, CPU_PARTIAL_NODE);
2001 if (!kmem_cache_has_cpu_partial(s)
2002 || available > slub_cpu_partial(s) / 2)
2006 spin_unlock(&n->list_lock);
2011 * Get a page from somewhere. Search in increasing NUMA distances.
2013 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2014 struct kmem_cache_cpu *c)
2017 struct zonelist *zonelist;
2020 enum zone_type highest_zoneidx = gfp_zone(flags);
2022 unsigned int cpuset_mems_cookie;
2025 * The defrag ratio allows a configuration of the tradeoffs between
2026 * inter node defragmentation and node local allocations. A lower
2027 * defrag_ratio increases the tendency to do local allocations
2028 * instead of attempting to obtain partial slabs from other nodes.
2030 * If the defrag_ratio is set to 0 then kmalloc() always
2031 * returns node local objects. If the ratio is higher then kmalloc()
2032 * may return off node objects because partial slabs are obtained
2033 * from other nodes and filled up.
2035 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2036 * (which makes defrag_ratio = 1000) then every (well almost)
2037 * allocation will first attempt to defrag slab caches on other nodes.
2038 * This means scanning over all nodes to look for partial slabs which
2039 * may be expensive if we do it every time we are trying to find a slab
2040 * with available objects.
2042 if (!s->remote_node_defrag_ratio ||
2043 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2047 cpuset_mems_cookie = read_mems_allowed_begin();
2048 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2049 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2050 struct kmem_cache_node *n;
2052 n = get_node(s, zone_to_nid(zone));
2054 if (n && cpuset_zone_allowed(zone, flags) &&
2055 n->nr_partial > s->min_partial) {
2056 object = get_partial_node(s, n, c, flags);
2059 * Don't check read_mems_allowed_retry()
2060 * here - if mems_allowed was updated in
2061 * parallel, that was a harmless race
2062 * between allocation and the cpuset
2069 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2070 #endif /* CONFIG_NUMA */
2075 * Get a partial page, lock it and return it.
2077 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2078 struct kmem_cache_cpu *c)
2081 int searchnode = node;
2083 if (node == NUMA_NO_NODE)
2084 searchnode = numa_mem_id();
2086 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2087 if (object || node != NUMA_NO_NODE)
2090 return get_any_partial(s, flags, c);
2093 #ifdef CONFIG_PREEMPTION
2095 * Calculate the next globally unique transaction for disambiguation
2096 * during cmpxchg. The transactions start with the cpu number and are then
2097 * incremented by CONFIG_NR_CPUS.
2099 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2102 * No preemption supported therefore also no need to check for
2108 static inline unsigned long next_tid(unsigned long tid)
2110 return tid + TID_STEP;
2113 #ifdef SLUB_DEBUG_CMPXCHG
2114 static inline unsigned int tid_to_cpu(unsigned long tid)
2116 return tid % TID_STEP;
2119 static inline unsigned long tid_to_event(unsigned long tid)
2121 return tid / TID_STEP;
2125 static inline unsigned int init_tid(int cpu)
2130 static inline void note_cmpxchg_failure(const char *n,
2131 const struct kmem_cache *s, unsigned long tid)
2133 #ifdef SLUB_DEBUG_CMPXCHG
2134 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2136 pr_info("%s %s: cmpxchg redo ", n, s->name);
2138 #ifdef CONFIG_PREEMPTION
2139 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2140 pr_warn("due to cpu change %d -> %d\n",
2141 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2144 if (tid_to_event(tid) != tid_to_event(actual_tid))
2145 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2146 tid_to_event(tid), tid_to_event(actual_tid));
2148 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2149 actual_tid, tid, next_tid(tid));
2151 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2154 static void init_kmem_cache_cpus(struct kmem_cache *s)
2158 for_each_possible_cpu(cpu)
2159 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2163 * Remove the cpu slab
2165 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2166 void *freelist, struct kmem_cache_cpu *c)
2168 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2169 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2171 enum slab_modes l = M_NONE, m = M_NONE;
2173 int tail = DEACTIVATE_TO_HEAD;
2177 if (page->freelist) {
2178 stat(s, DEACTIVATE_REMOTE_FREES);
2179 tail = DEACTIVATE_TO_TAIL;
2183 * Stage one: Free all available per cpu objects back
2184 * to the page freelist while it is still frozen. Leave the
2187 * There is no need to take the list->lock because the page
2190 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2192 unsigned long counters;
2195 * If 'nextfree' is invalid, it is possible that the object at
2196 * 'freelist' is already corrupted. So isolate all objects
2197 * starting at 'freelist'.
2199 if (freelist_corrupted(s, page, freelist, nextfree))
2203 prior = page->freelist;
2204 counters = page->counters;
2205 set_freepointer(s, freelist, prior);
2206 new.counters = counters;
2208 VM_BUG_ON(!new.frozen);
2210 } while (!__cmpxchg_double_slab(s, page,
2212 freelist, new.counters,
2213 "drain percpu freelist"));
2215 freelist = nextfree;
2219 * Stage two: Ensure that the page is unfrozen while the
2220 * list presence reflects the actual number of objects
2223 * We setup the list membership and then perform a cmpxchg
2224 * with the count. If there is a mismatch then the page
2225 * is not unfrozen but the page is on the wrong list.
2227 * Then we restart the process which may have to remove
2228 * the page from the list that we just put it on again
2229 * because the number of objects in the slab may have
2234 old.freelist = page->freelist;
2235 old.counters = page->counters;
2236 VM_BUG_ON(!old.frozen);
2238 /* Determine target state of the slab */
2239 new.counters = old.counters;
2242 set_freepointer(s, freelist, old.freelist);
2243 new.freelist = freelist;
2245 new.freelist = old.freelist;
2249 if (!new.inuse && n->nr_partial >= s->min_partial)
2251 else if (new.freelist) {
2256 * Taking the spinlock removes the possibility
2257 * that acquire_slab() will see a slab page that
2260 spin_lock(&n->list_lock);
2264 if (kmem_cache_debug(s) && !lock) {
2267 * This also ensures that the scanning of full
2268 * slabs from diagnostic functions will not see
2271 spin_lock(&n->list_lock);
2277 remove_partial(n, page);
2278 else if (l == M_FULL)
2279 remove_full(s, n, page);
2282 add_partial(n, page, tail);
2283 else if (m == M_FULL)
2284 add_full(s, n, page);
2288 if (!__cmpxchg_double_slab(s, page,
2289 old.freelist, old.counters,
2290 new.freelist, new.counters,
2295 spin_unlock(&n->list_lock);
2299 else if (m == M_FULL)
2300 stat(s, DEACTIVATE_FULL);
2301 else if (m == M_FREE) {
2302 stat(s, DEACTIVATE_EMPTY);
2303 discard_slab(s, page);
2312 * Unfreeze all the cpu partial slabs.
2314 * This function must be called with interrupts disabled
2315 * for the cpu using c (or some other guarantee must be there
2316 * to guarantee no concurrent accesses).
2318 static void unfreeze_partials(struct kmem_cache *s,
2319 struct kmem_cache_cpu *c)
2321 #ifdef CONFIG_SLUB_CPU_PARTIAL
2322 struct kmem_cache_node *n = NULL, *n2 = NULL;
2323 struct page *page, *discard_page = NULL;
2325 while ((page = slub_percpu_partial(c))) {
2329 slub_set_percpu_partial(c, page);
2331 n2 = get_node(s, page_to_nid(page));
2334 spin_unlock(&n->list_lock);
2337 spin_lock(&n->list_lock);
2342 old.freelist = page->freelist;
2343 old.counters = page->counters;
2344 VM_BUG_ON(!old.frozen);
2346 new.counters = old.counters;
2347 new.freelist = old.freelist;
2351 } while (!__cmpxchg_double_slab(s, page,
2352 old.freelist, old.counters,
2353 new.freelist, new.counters,
2354 "unfreezing slab"));
2356 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2357 page->next = discard_page;
2358 discard_page = page;
2360 add_partial(n, page, DEACTIVATE_TO_TAIL);
2361 stat(s, FREE_ADD_PARTIAL);
2366 spin_unlock(&n->list_lock);
2368 while (discard_page) {
2369 page = discard_page;
2370 discard_page = discard_page->next;
2372 stat(s, DEACTIVATE_EMPTY);
2373 discard_slab(s, page);
2376 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2380 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2381 * partial page slot if available.
2383 * If we did not find a slot then simply move all the partials to the
2384 * per node partial list.
2386 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2388 #ifdef CONFIG_SLUB_CPU_PARTIAL
2389 struct page *oldpage;
2397 oldpage = this_cpu_read(s->cpu_slab->partial);
2400 pobjects = oldpage->pobjects;
2401 pages = oldpage->pages;
2402 if (drain && pobjects > slub_cpu_partial(s)) {
2403 unsigned long flags;
2405 * partial array is full. Move the existing
2406 * set to the per node partial list.
2408 local_irq_save(flags);
2409 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2410 local_irq_restore(flags);
2414 stat(s, CPU_PARTIAL_DRAIN);
2419 pobjects += page->objects - page->inuse;
2421 page->pages = pages;
2422 page->pobjects = pobjects;
2423 page->next = oldpage;
2425 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2427 if (unlikely(!slub_cpu_partial(s))) {
2428 unsigned long flags;
2430 local_irq_save(flags);
2431 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2432 local_irq_restore(flags);
2435 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2438 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2440 stat(s, CPUSLAB_FLUSH);
2441 deactivate_slab(s, c->page, c->freelist, c);
2443 c->tid = next_tid(c->tid);
2449 * Called from IPI handler with interrupts disabled.
2451 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2453 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2458 unfreeze_partials(s, c);
2461 static void flush_cpu_slab(void *d)
2463 struct kmem_cache *s = d;
2465 __flush_cpu_slab(s, smp_processor_id());
2468 static bool has_cpu_slab(int cpu, void *info)
2470 struct kmem_cache *s = info;
2471 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2473 return c->page || slub_percpu_partial(c);
2476 static void flush_all(struct kmem_cache *s)
2478 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2482 * Use the cpu notifier to insure that the cpu slabs are flushed when
2485 static int slub_cpu_dead(unsigned int cpu)
2487 struct kmem_cache *s;
2488 unsigned long flags;
2490 mutex_lock(&slab_mutex);
2491 list_for_each_entry(s, &slab_caches, list) {
2492 local_irq_save(flags);
2493 __flush_cpu_slab(s, cpu);
2494 local_irq_restore(flags);
2496 mutex_unlock(&slab_mutex);
2501 * Check if the objects in a per cpu structure fit numa
2502 * locality expectations.
2504 static inline int node_match(struct page *page, int node)
2507 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2513 #ifdef CONFIG_SLUB_DEBUG
2514 static int count_free(struct page *page)
2516 return page->objects - page->inuse;
2519 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2521 return atomic_long_read(&n->total_objects);
2523 #endif /* CONFIG_SLUB_DEBUG */
2525 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2526 static unsigned long count_partial(struct kmem_cache_node *n,
2527 int (*get_count)(struct page *))
2529 unsigned long flags;
2530 unsigned long x = 0;
2533 spin_lock_irqsave(&n->list_lock, flags);
2534 list_for_each_entry(page, &n->partial, slab_list)
2535 x += get_count(page);
2536 spin_unlock_irqrestore(&n->list_lock, flags);
2539 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2541 static noinline void
2542 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2544 #ifdef CONFIG_SLUB_DEBUG
2545 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2546 DEFAULT_RATELIMIT_BURST);
2548 struct kmem_cache_node *n;
2550 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2553 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2554 nid, gfpflags, &gfpflags);
2555 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2556 s->name, s->object_size, s->size, oo_order(s->oo),
2559 if (oo_order(s->min) > get_order(s->object_size))
2560 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2563 for_each_kmem_cache_node(s, node, n) {
2564 unsigned long nr_slabs;
2565 unsigned long nr_objs;
2566 unsigned long nr_free;
2568 nr_free = count_partial(n, count_free);
2569 nr_slabs = node_nr_slabs(n);
2570 nr_objs = node_nr_objs(n);
2572 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2573 node, nr_slabs, nr_objs, nr_free);
2578 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2579 int node, struct kmem_cache_cpu **pc)
2582 struct kmem_cache_cpu *c = *pc;
2585 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2587 freelist = get_partial(s, flags, node, c);
2592 page = new_slab(s, flags, node);
2594 c = raw_cpu_ptr(s->cpu_slab);
2599 * No other reference to the page yet so we can
2600 * muck around with it freely without cmpxchg
2602 freelist = page->freelist;
2603 page->freelist = NULL;
2605 stat(s, ALLOC_SLAB);
2613 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2615 if (unlikely(PageSlabPfmemalloc(page)))
2616 return gfp_pfmemalloc_allowed(gfpflags);
2622 * Check the page->freelist of a page and either transfer the freelist to the
2623 * per cpu freelist or deactivate the page.
2625 * The page is still frozen if the return value is not NULL.
2627 * If this function returns NULL then the page has been unfrozen.
2629 * This function must be called with interrupt disabled.
2631 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2634 unsigned long counters;
2638 freelist = page->freelist;
2639 counters = page->counters;
2641 new.counters = counters;
2642 VM_BUG_ON(!new.frozen);
2644 new.inuse = page->objects;
2645 new.frozen = freelist != NULL;
2647 } while (!__cmpxchg_double_slab(s, page,
2656 * Slow path. The lockless freelist is empty or we need to perform
2659 * Processing is still very fast if new objects have been freed to the
2660 * regular freelist. In that case we simply take over the regular freelist
2661 * as the lockless freelist and zap the regular freelist.
2663 * If that is not working then we fall back to the partial lists. We take the
2664 * first element of the freelist as the object to allocate now and move the
2665 * rest of the freelist to the lockless freelist.
2667 * And if we were unable to get a new slab from the partial slab lists then
2668 * we need to allocate a new slab. This is the slowest path since it involves
2669 * a call to the page allocator and the setup of a new slab.
2671 * Version of __slab_alloc to use when we know that interrupts are
2672 * already disabled (which is the case for bulk allocation).
2674 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2675 unsigned long addr, struct kmem_cache_cpu *c)
2683 * if the node is not online or has no normal memory, just
2684 * ignore the node constraint
2686 if (unlikely(node != NUMA_NO_NODE &&
2687 !node_state(node, N_NORMAL_MEMORY)))
2688 node = NUMA_NO_NODE;
2693 if (unlikely(!node_match(page, node))) {
2695 * same as above but node_match() being false already
2696 * implies node != NUMA_NO_NODE
2698 if (!node_state(node, N_NORMAL_MEMORY)) {
2699 node = NUMA_NO_NODE;
2702 stat(s, ALLOC_NODE_MISMATCH);
2703 deactivate_slab(s, page, c->freelist, c);
2709 * By rights, we should be searching for a slab page that was
2710 * PFMEMALLOC but right now, we are losing the pfmemalloc
2711 * information when the page leaves the per-cpu allocator
2713 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2714 deactivate_slab(s, page, c->freelist, c);
2718 /* must check again c->freelist in case of cpu migration or IRQ */
2719 freelist = c->freelist;
2723 freelist = get_freelist(s, page);
2727 stat(s, DEACTIVATE_BYPASS);
2731 stat(s, ALLOC_REFILL);
2735 * freelist is pointing to the list of objects to be used.
2736 * page is pointing to the page from which the objects are obtained.
2737 * That page must be frozen for per cpu allocations to work.
2739 VM_BUG_ON(!c->page->frozen);
2740 c->freelist = get_freepointer(s, freelist);
2741 c->tid = next_tid(c->tid);
2746 if (slub_percpu_partial(c)) {
2747 page = c->page = slub_percpu_partial(c);
2748 slub_set_percpu_partial(c, page);
2749 stat(s, CPU_PARTIAL_ALLOC);
2753 freelist = new_slab_objects(s, gfpflags, node, &c);
2755 if (unlikely(!freelist)) {
2756 slab_out_of_memory(s, gfpflags, node);
2761 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2764 /* Only entered in the debug case */
2765 if (kmem_cache_debug(s) &&
2766 !alloc_debug_processing(s, page, freelist, addr))
2767 goto new_slab; /* Slab failed checks. Next slab needed */
2769 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2774 * Another one that disabled interrupt and compensates for possible
2775 * cpu changes by refetching the per cpu area pointer.
2777 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2778 unsigned long addr, struct kmem_cache_cpu *c)
2781 unsigned long flags;
2783 local_irq_save(flags);
2784 #ifdef CONFIG_PREEMPTION
2786 * We may have been preempted and rescheduled on a different
2787 * cpu before disabling interrupts. Need to reload cpu area
2790 c = this_cpu_ptr(s->cpu_slab);
2793 p = ___slab_alloc(s, gfpflags, node, addr, c);
2794 local_irq_restore(flags);
2799 * If the object has been wiped upon free, make sure it's fully initialized by
2800 * zeroing out freelist pointer.
2802 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2805 if (unlikely(slab_want_init_on_free(s)) && obj)
2806 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2810 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2811 * have the fastpath folded into their functions. So no function call
2812 * overhead for requests that can be satisfied on the fastpath.
2814 * The fastpath works by first checking if the lockless freelist can be used.
2815 * If not then __slab_alloc is called for slow processing.
2817 * Otherwise we can simply pick the next object from the lockless free list.
2819 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2820 gfp_t gfpflags, int node, unsigned long addr)
2823 struct kmem_cache_cpu *c;
2827 s = slab_pre_alloc_hook(s, gfpflags);
2832 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2833 * enabled. We may switch back and forth between cpus while
2834 * reading from one cpu area. That does not matter as long
2835 * as we end up on the original cpu again when doing the cmpxchg.
2837 * We should guarantee that tid and kmem_cache are retrieved on
2838 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2839 * to check if it is matched or not.
2842 tid = this_cpu_read(s->cpu_slab->tid);
2843 c = raw_cpu_ptr(s->cpu_slab);
2844 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2845 unlikely(tid != READ_ONCE(c->tid)));
2848 * Irqless object alloc/free algorithm used here depends on sequence
2849 * of fetching cpu_slab's data. tid should be fetched before anything
2850 * on c to guarantee that object and page associated with previous tid
2851 * won't be used with current tid. If we fetch tid first, object and
2852 * page could be one associated with next tid and our alloc/free
2853 * request will be failed. In this case, we will retry. So, no problem.
2858 * The transaction ids are globally unique per cpu and per operation on
2859 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2860 * occurs on the right processor and that there was no operation on the
2861 * linked list in between.
2864 object = c->freelist;
2866 if (unlikely(!object || !node_match(page, node))) {
2867 object = __slab_alloc(s, gfpflags, node, addr, c);
2868 stat(s, ALLOC_SLOWPATH);
2870 void *next_object = get_freepointer_safe(s, object);
2873 * The cmpxchg will only match if there was no additional
2874 * operation and if we are on the right processor.
2876 * The cmpxchg does the following atomically (without lock
2878 * 1. Relocate first pointer to the current per cpu area.
2879 * 2. Verify that tid and freelist have not been changed
2880 * 3. If they were not changed replace tid and freelist
2882 * Since this is without lock semantics the protection is only
2883 * against code executing on this cpu *not* from access by
2886 if (unlikely(!this_cpu_cmpxchg_double(
2887 s->cpu_slab->freelist, s->cpu_slab->tid,
2889 next_object, next_tid(tid)))) {
2891 note_cmpxchg_failure("slab_alloc", s, tid);
2894 prefetch_freepointer(s, next_object);
2895 stat(s, ALLOC_FASTPATH);
2898 maybe_wipe_obj_freeptr(s, object);
2900 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2901 memset(object, 0, s->object_size);
2903 slab_post_alloc_hook(s, gfpflags, 1, &object);
2908 static __always_inline void *slab_alloc(struct kmem_cache *s,
2909 gfp_t gfpflags, unsigned long addr)
2911 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2914 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2916 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2918 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2923 EXPORT_SYMBOL(kmem_cache_alloc);
2925 #ifdef CONFIG_TRACING
2926 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2928 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2929 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2930 ret = kasan_kmalloc(s, ret, size, gfpflags);
2933 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2937 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2939 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2941 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2942 s->object_size, s->size, gfpflags, node);
2946 EXPORT_SYMBOL(kmem_cache_alloc_node);
2948 #ifdef CONFIG_TRACING
2949 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2951 int node, size_t size)
2953 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2955 trace_kmalloc_node(_RET_IP_, ret,
2956 size, s->size, gfpflags, node);
2958 ret = kasan_kmalloc(s, ret, size, gfpflags);
2961 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2963 #endif /* CONFIG_NUMA */
2966 * Slow path handling. This may still be called frequently since objects
2967 * have a longer lifetime than the cpu slabs in most processing loads.
2969 * So we still attempt to reduce cache line usage. Just take the slab
2970 * lock and free the item. If there is no additional partial page
2971 * handling required then we can return immediately.
2973 static void __slab_free(struct kmem_cache *s, struct page *page,
2974 void *head, void *tail, int cnt,
2981 unsigned long counters;
2982 struct kmem_cache_node *n = NULL;
2983 unsigned long flags;
2985 stat(s, FREE_SLOWPATH);
2987 if (kmem_cache_debug(s) &&
2988 !free_debug_processing(s, page, head, tail, cnt, addr))
2993 spin_unlock_irqrestore(&n->list_lock, flags);
2996 prior = page->freelist;
2997 counters = page->counters;
2998 set_freepointer(s, tail, prior);
2999 new.counters = counters;
3000 was_frozen = new.frozen;
3002 if ((!new.inuse || !prior) && !was_frozen) {
3004 if (kmem_cache_has_cpu_partial(s) && !prior) {
3007 * Slab was on no list before and will be
3009 * We can defer the list move and instead
3014 } else { /* Needs to be taken off a list */
3016 n = get_node(s, page_to_nid(page));
3018 * Speculatively acquire the list_lock.
3019 * If the cmpxchg does not succeed then we may
3020 * drop the list_lock without any processing.
3022 * Otherwise the list_lock will synchronize with
3023 * other processors updating the list of slabs.
3025 spin_lock_irqsave(&n->list_lock, flags);
3030 } while (!cmpxchg_double_slab(s, page,
3038 * If we just froze the page then put it onto the
3039 * per cpu partial list.
3041 if (new.frozen && !was_frozen) {
3042 put_cpu_partial(s, page, 1);
3043 stat(s, CPU_PARTIAL_FREE);
3046 * The list lock was not taken therefore no list
3047 * activity can be necessary.
3050 stat(s, FREE_FROZEN);
3054 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3058 * Objects left in the slab. If it was not on the partial list before
3061 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3062 remove_full(s, n, page);
3063 add_partial(n, page, DEACTIVATE_TO_TAIL);
3064 stat(s, FREE_ADD_PARTIAL);
3066 spin_unlock_irqrestore(&n->list_lock, flags);
3072 * Slab on the partial list.
3074 remove_partial(n, page);
3075 stat(s, FREE_REMOVE_PARTIAL);
3077 /* Slab must be on the full list */
3078 remove_full(s, n, page);
3081 spin_unlock_irqrestore(&n->list_lock, flags);
3083 discard_slab(s, page);
3087 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3088 * can perform fastpath freeing without additional function calls.
3090 * The fastpath is only possible if we are freeing to the current cpu slab
3091 * of this processor. This typically the case if we have just allocated
3094 * If fastpath is not possible then fall back to __slab_free where we deal
3095 * with all sorts of special processing.
3097 * Bulk free of a freelist with several objects (all pointing to the
3098 * same page) possible by specifying head and tail ptr, plus objects
3099 * count (cnt). Bulk free indicated by tail pointer being set.
3101 static __always_inline void do_slab_free(struct kmem_cache *s,
3102 struct page *page, void *head, void *tail,
3103 int cnt, unsigned long addr)
3105 void *tail_obj = tail ? : head;
3106 struct kmem_cache_cpu *c;
3110 * Determine the currently cpus per cpu slab.
3111 * The cpu may change afterward. However that does not matter since
3112 * data is retrieved via this pointer. If we are on the same cpu
3113 * during the cmpxchg then the free will succeed.
3116 tid = this_cpu_read(s->cpu_slab->tid);
3117 c = raw_cpu_ptr(s->cpu_slab);
3118 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3119 unlikely(tid != READ_ONCE(c->tid)));
3121 /* Same with comment on barrier() in slab_alloc_node() */
3124 if (likely(page == c->page)) {
3125 void **freelist = READ_ONCE(c->freelist);
3127 set_freepointer(s, tail_obj, freelist);
3129 if (unlikely(!this_cpu_cmpxchg_double(
3130 s->cpu_slab->freelist, s->cpu_slab->tid,
3132 head, next_tid(tid)))) {
3134 note_cmpxchg_failure("slab_free", s, tid);
3137 stat(s, FREE_FASTPATH);
3139 __slab_free(s, page, head, tail_obj, cnt, addr);
3143 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3144 void *head, void *tail, int cnt,
3148 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3149 * to remove objects, whose reuse must be delayed.
3151 if (slab_free_freelist_hook(s, &head, &tail))
3152 do_slab_free(s, page, head, tail, cnt, addr);
3155 #ifdef CONFIG_KASAN_GENERIC
3156 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3158 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3162 void kmem_cache_free(struct kmem_cache *s, void *x)
3164 s = cache_from_obj(s, x);
3167 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3168 trace_kmem_cache_free(_RET_IP_, x);
3170 EXPORT_SYMBOL(kmem_cache_free);
3172 struct detached_freelist {
3177 struct kmem_cache *s;
3181 * This function progressively scans the array with free objects (with
3182 * a limited look ahead) and extract objects belonging to the same
3183 * page. It builds a detached freelist directly within the given
3184 * page/objects. This can happen without any need for
3185 * synchronization, because the objects are owned by running process.
3186 * The freelist is build up as a single linked list in the objects.
3187 * The idea is, that this detached freelist can then be bulk
3188 * transferred to the real freelist(s), but only requiring a single
3189 * synchronization primitive. Look ahead in the array is limited due
3190 * to performance reasons.
3193 int build_detached_freelist(struct kmem_cache *s, size_t size,
3194 void **p, struct detached_freelist *df)
3196 size_t first_skipped_index = 0;
3201 /* Always re-init detached_freelist */
3206 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3207 } while (!object && size);
3212 page = virt_to_head_page(object);
3214 /* Handle kalloc'ed objects */
3215 if (unlikely(!PageSlab(page))) {
3216 BUG_ON(!PageCompound(page));
3218 __free_pages(page, compound_order(page));
3219 p[size] = NULL; /* mark object processed */
3222 /* Derive kmem_cache from object */
3223 df->s = page->slab_cache;
3225 df->s = cache_from_obj(s, object); /* Support for memcg */
3228 /* Start new detached freelist */
3230 set_freepointer(df->s, object, NULL);
3232 df->freelist = object;
3233 p[size] = NULL; /* mark object processed */
3239 continue; /* Skip processed objects */
3241 /* df->page is always set at this point */
3242 if (df->page == virt_to_head_page(object)) {
3243 /* Opportunity build freelist */
3244 set_freepointer(df->s, object, df->freelist);
3245 df->freelist = object;
3247 p[size] = NULL; /* mark object processed */
3252 /* Limit look ahead search */
3256 if (!first_skipped_index)
3257 first_skipped_index = size + 1;
3260 return first_skipped_index;
3263 /* Note that interrupts must be enabled when calling this function. */
3264 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3270 struct detached_freelist df;
3272 size = build_detached_freelist(s, size, p, &df);
3276 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3277 } while (likely(size));
3279 EXPORT_SYMBOL(kmem_cache_free_bulk);
3281 /* Note that interrupts must be enabled when calling this function. */
3282 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3285 struct kmem_cache_cpu *c;
3288 /* memcg and kmem_cache debug support */
3289 s = slab_pre_alloc_hook(s, flags);
3293 * Drain objects in the per cpu slab, while disabling local
3294 * IRQs, which protects against PREEMPT and interrupts
3295 * handlers invoking normal fastpath.
3297 local_irq_disable();
3298 c = this_cpu_ptr(s->cpu_slab);
3300 for (i = 0; i < size; i++) {
3301 void *object = c->freelist;
3303 if (unlikely(!object)) {
3305 * We may have removed an object from c->freelist using
3306 * the fastpath in the previous iteration; in that case,
3307 * c->tid has not been bumped yet.
3308 * Since ___slab_alloc() may reenable interrupts while
3309 * allocating memory, we should bump c->tid now.
3311 c->tid = next_tid(c->tid);
3314 * Invoking slow path likely have side-effect
3315 * of re-populating per CPU c->freelist
3317 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3319 if (unlikely(!p[i]))
3322 c = this_cpu_ptr(s->cpu_slab);
3323 maybe_wipe_obj_freeptr(s, p[i]);
3325 continue; /* goto for-loop */
3327 c->freelist = get_freepointer(s, object);
3329 maybe_wipe_obj_freeptr(s, p[i]);
3331 c->tid = next_tid(c->tid);
3334 /* Clear memory outside IRQ disabled fastpath loop */
3335 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3338 for (j = 0; j < i; j++)
3339 memset(p[j], 0, s->object_size);
3342 /* memcg and kmem_cache debug support */
3343 slab_post_alloc_hook(s, flags, size, p);
3347 slab_post_alloc_hook(s, flags, i, p);
3348 __kmem_cache_free_bulk(s, i, p);
3351 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3355 * Object placement in a slab is made very easy because we always start at
3356 * offset 0. If we tune the size of the object to the alignment then we can
3357 * get the required alignment by putting one properly sized object after
3360 * Notice that the allocation order determines the sizes of the per cpu
3361 * caches. Each processor has always one slab available for allocations.
3362 * Increasing the allocation order reduces the number of times that slabs
3363 * must be moved on and off the partial lists and is therefore a factor in
3368 * Mininum / Maximum order of slab pages. This influences locking overhead
3369 * and slab fragmentation. A higher order reduces the number of partial slabs
3370 * and increases the number of allocations possible without having to
3371 * take the list_lock.
3373 static unsigned int slub_min_order;
3374 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3375 static unsigned int slub_min_objects;
3378 * Calculate the order of allocation given an slab object size.
3380 * The order of allocation has significant impact on performance and other
3381 * system components. Generally order 0 allocations should be preferred since
3382 * order 0 does not cause fragmentation in the page allocator. Larger objects
3383 * be problematic to put into order 0 slabs because there may be too much
3384 * unused space left. We go to a higher order if more than 1/16th of the slab
3387 * In order to reach satisfactory performance we must ensure that a minimum
3388 * number of objects is in one slab. Otherwise we may generate too much
3389 * activity on the partial lists which requires taking the list_lock. This is
3390 * less a concern for large slabs though which are rarely used.
3392 * slub_max_order specifies the order where we begin to stop considering the
3393 * number of objects in a slab as critical. If we reach slub_max_order then
3394 * we try to keep the page order as low as possible. So we accept more waste
3395 * of space in favor of a small page order.
3397 * Higher order allocations also allow the placement of more objects in a
3398 * slab and thereby reduce object handling overhead. If the user has
3399 * requested a higher mininum order then we start with that one instead of
3400 * the smallest order which will fit the object.
3402 static inline unsigned int slab_order(unsigned int size,
3403 unsigned int min_objects, unsigned int max_order,
3404 unsigned int fract_leftover)
3406 unsigned int min_order = slub_min_order;
3409 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3410 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3412 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3413 order <= max_order; order++) {
3415 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3418 rem = slab_size % size;
3420 if (rem <= slab_size / fract_leftover)
3427 static inline int calculate_order(unsigned int size)
3430 unsigned int min_objects;
3431 unsigned int max_objects;
3434 * Attempt to find best configuration for a slab. This
3435 * works by first attempting to generate a layout with
3436 * the best configuration and backing off gradually.
3438 * First we increase the acceptable waste in a slab. Then
3439 * we reduce the minimum objects required in a slab.
3441 min_objects = slub_min_objects;
3443 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3444 max_objects = order_objects(slub_max_order, size);
3445 min_objects = min(min_objects, max_objects);
3447 while (min_objects > 1) {
3448 unsigned int fraction;
3451 while (fraction >= 4) {
3452 order = slab_order(size, min_objects,
3453 slub_max_order, fraction);
3454 if (order <= slub_max_order)
3462 * We were unable to place multiple objects in a slab. Now
3463 * lets see if we can place a single object there.
3465 order = slab_order(size, 1, slub_max_order, 1);
3466 if (order <= slub_max_order)
3470 * Doh this slab cannot be placed using slub_max_order.
3472 order = slab_order(size, 1, MAX_ORDER, 1);
3473 if (order < MAX_ORDER)
3479 init_kmem_cache_node(struct kmem_cache_node *n)
3482 spin_lock_init(&n->list_lock);
3483 INIT_LIST_HEAD(&n->partial);
3484 #ifdef CONFIG_SLUB_DEBUG
3485 atomic_long_set(&n->nr_slabs, 0);
3486 atomic_long_set(&n->total_objects, 0);
3487 INIT_LIST_HEAD(&n->full);
3491 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3493 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3494 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3497 * Must align to double word boundary for the double cmpxchg
3498 * instructions to work; see __pcpu_double_call_return_bool().
3500 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3501 2 * sizeof(void *));
3506 init_kmem_cache_cpus(s);
3511 static struct kmem_cache *kmem_cache_node;
3514 * No kmalloc_node yet so do it by hand. We know that this is the first
3515 * slab on the node for this slabcache. There are no concurrent accesses
3518 * Note that this function only works on the kmem_cache_node
3519 * when allocating for the kmem_cache_node. This is used for bootstrapping
3520 * memory on a fresh node that has no slab structures yet.
3522 static void early_kmem_cache_node_alloc(int node)
3525 struct kmem_cache_node *n;
3527 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3529 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3532 if (page_to_nid(page) != node) {
3533 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3534 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3539 #ifdef CONFIG_SLUB_DEBUG
3540 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3541 init_tracking(kmem_cache_node, n);
3543 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3545 page->freelist = get_freepointer(kmem_cache_node, n);
3548 kmem_cache_node->node[node] = n;
3549 init_kmem_cache_node(n);
3550 inc_slabs_node(kmem_cache_node, node, page->objects);
3553 * No locks need to be taken here as it has just been
3554 * initialized and there is no concurrent access.
3556 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3559 static void free_kmem_cache_nodes(struct kmem_cache *s)
3562 struct kmem_cache_node *n;
3564 for_each_kmem_cache_node(s, node, n) {
3565 s->node[node] = NULL;
3566 kmem_cache_free(kmem_cache_node, n);
3570 void __kmem_cache_release(struct kmem_cache *s)
3572 cache_random_seq_destroy(s);
3573 free_percpu(s->cpu_slab);
3574 free_kmem_cache_nodes(s);
3577 static int init_kmem_cache_nodes(struct kmem_cache *s)
3581 for_each_node_state(node, N_NORMAL_MEMORY) {
3582 struct kmem_cache_node *n;
3584 if (slab_state == DOWN) {
3585 early_kmem_cache_node_alloc(node);
3588 n = kmem_cache_alloc_node(kmem_cache_node,
3592 free_kmem_cache_nodes(s);
3596 init_kmem_cache_node(n);
3602 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3604 if (min < MIN_PARTIAL)
3606 else if (min > MAX_PARTIAL)
3608 s->min_partial = min;
3611 static void set_cpu_partial(struct kmem_cache *s)
3613 #ifdef CONFIG_SLUB_CPU_PARTIAL
3615 * cpu_partial determined the maximum number of objects kept in the
3616 * per cpu partial lists of a processor.
3618 * Per cpu partial lists mainly contain slabs that just have one
3619 * object freed. If they are used for allocation then they can be
3620 * filled up again with minimal effort. The slab will never hit the
3621 * per node partial lists and therefore no locking will be required.
3623 * This setting also determines
3625 * A) The number of objects from per cpu partial slabs dumped to the
3626 * per node list when we reach the limit.
3627 * B) The number of objects in cpu partial slabs to extract from the
3628 * per node list when we run out of per cpu objects. We only fetch
3629 * 50% to keep some capacity around for frees.
3631 if (!kmem_cache_has_cpu_partial(s))
3632 slub_set_cpu_partial(s, 0);
3633 else if (s->size >= PAGE_SIZE)
3634 slub_set_cpu_partial(s, 2);
3635 else if (s->size >= 1024)
3636 slub_set_cpu_partial(s, 6);
3637 else if (s->size >= 256)
3638 slub_set_cpu_partial(s, 13);
3640 slub_set_cpu_partial(s, 30);
3645 * calculate_sizes() determines the order and the distribution of data within
3648 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3650 slab_flags_t flags = s->flags;
3651 unsigned int size = s->object_size;
3652 unsigned int freepointer_area;
3656 * Round up object size to the next word boundary. We can only
3657 * place the free pointer at word boundaries and this determines
3658 * the possible location of the free pointer.
3660 size = ALIGN(size, sizeof(void *));
3662 * This is the area of the object where a freepointer can be
3663 * safely written. If redzoning adds more to the inuse size, we
3664 * can't use that portion for writing the freepointer, so
3665 * s->offset must be limited within this for the general case.
3667 freepointer_area = size;
3669 #ifdef CONFIG_SLUB_DEBUG
3671 * Determine if we can poison the object itself. If the user of
3672 * the slab may touch the object after free or before allocation
3673 * then we should never poison the object itself.
3675 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3677 s->flags |= __OBJECT_POISON;
3679 s->flags &= ~__OBJECT_POISON;
3683 * If we are Redzoning then check if there is some space between the
3684 * end of the object and the free pointer. If not then add an
3685 * additional word to have some bytes to store Redzone information.
3687 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3688 size += sizeof(void *);
3692 * With that we have determined the number of bytes in actual use
3693 * by the object. This is the potential offset to the free pointer.
3697 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3700 * Relocate free pointer after the object if it is not
3701 * permitted to overwrite the first word of the object on
3704 * This is the case if we do RCU, have a constructor or
3705 * destructor or are poisoning the objects.
3707 * The assumption that s->offset >= s->inuse means free
3708 * pointer is outside of the object is used in the
3709 * freeptr_outside_object() function. If that is no
3710 * longer true, the function needs to be modified.
3713 size += sizeof(void *);
3714 } else if (freepointer_area > sizeof(void *)) {
3716 * Store freelist pointer near middle of object to keep
3717 * it away from the edges of the object to avoid small
3718 * sized over/underflows from neighboring allocations.
3720 s->offset = ALIGN(freepointer_area / 2, sizeof(void *));
3723 #ifdef CONFIG_SLUB_DEBUG
3724 if (flags & SLAB_STORE_USER)
3726 * Need to store information about allocs and frees after
3729 size += 2 * sizeof(struct track);
3732 kasan_cache_create(s, &size, &s->flags);
3733 #ifdef CONFIG_SLUB_DEBUG
3734 if (flags & SLAB_RED_ZONE) {
3736 * Add some empty padding so that we can catch
3737 * overwrites from earlier objects rather than let
3738 * tracking information or the free pointer be
3739 * corrupted if a user writes before the start
3742 size += sizeof(void *);
3744 s->red_left_pad = sizeof(void *);
3745 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3746 size += s->red_left_pad;
3751 * SLUB stores one object immediately after another beginning from
3752 * offset 0. In order to align the objects we have to simply size
3753 * each object to conform to the alignment.
3755 size = ALIGN(size, s->align);
3757 if (forced_order >= 0)
3758 order = forced_order;
3760 order = calculate_order(size);
3767 s->allocflags |= __GFP_COMP;
3769 if (s->flags & SLAB_CACHE_DMA)
3770 s->allocflags |= GFP_DMA;
3772 if (s->flags & SLAB_CACHE_DMA32)
3773 s->allocflags |= GFP_DMA32;
3775 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3776 s->allocflags |= __GFP_RECLAIMABLE;
3779 * Determine the number of objects per slab
3781 s->oo = oo_make(order, size);
3782 s->min = oo_make(get_order(size), size);
3783 if (oo_objects(s->oo) > oo_objects(s->max))
3786 return !!oo_objects(s->oo);
3789 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3791 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3792 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3793 s->random = get_random_long();
3796 if (!calculate_sizes(s, -1))
3798 if (disable_higher_order_debug) {
3800 * Disable debugging flags that store metadata if the min slab
3803 if (get_order(s->size) > get_order(s->object_size)) {
3804 s->flags &= ~DEBUG_METADATA_FLAGS;
3806 if (!calculate_sizes(s, -1))
3811 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3812 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3813 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3814 /* Enable fast mode */
3815 s->flags |= __CMPXCHG_DOUBLE;
3819 * The larger the object size is, the more pages we want on the partial
3820 * list to avoid pounding the page allocator excessively.
3822 set_min_partial(s, ilog2(s->size) / 2);
3827 s->remote_node_defrag_ratio = 1000;
3830 /* Initialize the pre-computed randomized freelist if slab is up */
3831 if (slab_state >= UP) {
3832 if (init_cache_random_seq(s))
3836 if (!init_kmem_cache_nodes(s))
3839 if (alloc_kmem_cache_cpus(s))
3842 free_kmem_cache_nodes(s);
3847 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3850 #ifdef CONFIG_SLUB_DEBUG
3851 void *addr = page_address(page);
3855 slab_err(s, page, text, s->name);
3858 map = get_map(s, page);
3859 for_each_object(p, s, addr, page->objects) {
3861 if (!test_bit(slab_index(p, s, addr), map)) {
3862 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3863 print_tracking(s, p);
3872 * Attempt to free all partial slabs on a node.
3873 * This is called from __kmem_cache_shutdown(). We must take list_lock
3874 * because sysfs file might still access partial list after the shutdowning.
3876 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3879 struct page *page, *h;
3881 BUG_ON(irqs_disabled());
3882 spin_lock_irq(&n->list_lock);
3883 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3885 remove_partial(n, page);
3886 list_add(&page->slab_list, &discard);
3888 list_slab_objects(s, page,
3889 "Objects remaining in %s on __kmem_cache_shutdown()");
3892 spin_unlock_irq(&n->list_lock);
3894 list_for_each_entry_safe(page, h, &discard, slab_list)
3895 discard_slab(s, page);
3898 bool __kmem_cache_empty(struct kmem_cache *s)
3901 struct kmem_cache_node *n;
3903 for_each_kmem_cache_node(s, node, n)
3904 if (n->nr_partial || slabs_node(s, node))
3910 * Release all resources used by a slab cache.
3912 int __kmem_cache_shutdown(struct kmem_cache *s)
3915 struct kmem_cache_node *n;
3918 /* Attempt to free all objects */
3919 for_each_kmem_cache_node(s, node, n) {
3921 if (n->nr_partial || slabs_node(s, node))
3924 sysfs_slab_remove(s);
3928 /********************************************************************
3930 *******************************************************************/
3932 static int __init setup_slub_min_order(char *str)
3934 get_option(&str, (int *)&slub_min_order);
3939 __setup("slub_min_order=", setup_slub_min_order);
3941 static int __init setup_slub_max_order(char *str)
3943 get_option(&str, (int *)&slub_max_order);
3944 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3949 __setup("slub_max_order=", setup_slub_max_order);
3951 static int __init setup_slub_min_objects(char *str)
3953 get_option(&str, (int *)&slub_min_objects);
3958 __setup("slub_min_objects=", setup_slub_min_objects);
3960 void *__kmalloc(size_t size, gfp_t flags)
3962 struct kmem_cache *s;
3965 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3966 return kmalloc_large(size, flags);
3968 s = kmalloc_slab(size, flags);
3970 if (unlikely(ZERO_OR_NULL_PTR(s)))
3973 ret = slab_alloc(s, flags, _RET_IP_);
3975 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3977 ret = kasan_kmalloc(s, ret, size, flags);
3981 EXPORT_SYMBOL(__kmalloc);
3984 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3988 unsigned int order = get_order(size);
3990 flags |= __GFP_COMP;
3991 page = alloc_pages_node(node, flags, order);
3993 ptr = page_address(page);
3994 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
3995 PAGE_SIZE << order);
3998 return kmalloc_large_node_hook(ptr, size, flags);
4001 void *__kmalloc_node(size_t size, gfp_t flags, int node)
4003 struct kmem_cache *s;
4006 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4007 ret = kmalloc_large_node(size, flags, node);
4009 trace_kmalloc_node(_RET_IP_, ret,
4010 size, PAGE_SIZE << get_order(size),
4016 s = kmalloc_slab(size, flags);
4018 if (unlikely(ZERO_OR_NULL_PTR(s)))
4021 ret = slab_alloc_node(s, flags, node, _RET_IP_);
4023 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4025 ret = kasan_kmalloc(s, ret, size, flags);
4029 EXPORT_SYMBOL(__kmalloc_node);
4030 #endif /* CONFIG_NUMA */
4032 #ifdef CONFIG_HARDENED_USERCOPY
4034 * Rejects incorrectly sized objects and objects that are to be copied
4035 * to/from userspace but do not fall entirely within the containing slab
4036 * cache's usercopy region.
4038 * Returns NULL if check passes, otherwise const char * to name of cache
4039 * to indicate an error.
4041 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4044 struct kmem_cache *s;
4045 unsigned int offset;
4048 ptr = kasan_reset_tag(ptr);
4050 /* Find object and usable object size. */
4051 s = page->slab_cache;
4053 /* Reject impossible pointers. */
4054 if (ptr < page_address(page))
4055 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4058 /* Find offset within object. */
4059 offset = (ptr - page_address(page)) % s->size;
4061 /* Adjust for redzone and reject if within the redzone. */
4062 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4063 if (offset < s->red_left_pad)
4064 usercopy_abort("SLUB object in left red zone",
4065 s->name, to_user, offset, n);
4066 offset -= s->red_left_pad;
4069 /* Allow address range falling entirely within usercopy region. */
4070 if (offset >= s->useroffset &&
4071 offset - s->useroffset <= s->usersize &&
4072 n <= s->useroffset - offset + s->usersize)
4076 * If the copy is still within the allocated object, produce
4077 * a warning instead of rejecting the copy. This is intended
4078 * to be a temporary method to find any missing usercopy
4081 object_size = slab_ksize(s);
4082 if (usercopy_fallback &&
4083 offset <= object_size && n <= object_size - offset) {
4084 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4088 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4090 #endif /* CONFIG_HARDENED_USERCOPY */
4092 size_t __ksize(const void *object)
4096 if (unlikely(object == ZERO_SIZE_PTR))
4099 page = virt_to_head_page(object);
4101 if (unlikely(!PageSlab(page))) {
4102 WARN_ON(!PageCompound(page));
4103 return page_size(page);
4106 return slab_ksize(page->slab_cache);
4108 EXPORT_SYMBOL(__ksize);
4110 void kfree(const void *x)
4113 void *object = (void *)x;
4115 trace_kfree(_RET_IP_, x);
4117 if (unlikely(ZERO_OR_NULL_PTR(x)))
4120 page = virt_to_head_page(x);
4121 if (unlikely(!PageSlab(page))) {
4122 unsigned int order = compound_order(page);
4124 BUG_ON(!PageCompound(page));
4126 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
4127 -(PAGE_SIZE << order));
4128 __free_pages(page, order);
4131 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4133 EXPORT_SYMBOL(kfree);
4135 #define SHRINK_PROMOTE_MAX 32
4138 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4139 * up most to the head of the partial lists. New allocations will then
4140 * fill those up and thus they can be removed from the partial lists.
4142 * The slabs with the least items are placed last. This results in them
4143 * being allocated from last increasing the chance that the last objects
4144 * are freed in them.
4146 int __kmem_cache_shrink(struct kmem_cache *s)
4150 struct kmem_cache_node *n;
4153 struct list_head discard;
4154 struct list_head promote[SHRINK_PROMOTE_MAX];
4155 unsigned long flags;
4159 for_each_kmem_cache_node(s, node, n) {
4160 INIT_LIST_HEAD(&discard);
4161 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4162 INIT_LIST_HEAD(promote + i);
4164 spin_lock_irqsave(&n->list_lock, flags);
4167 * Build lists of slabs to discard or promote.
4169 * Note that concurrent frees may occur while we hold the
4170 * list_lock. page->inuse here is the upper limit.
4172 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4173 int free = page->objects - page->inuse;
4175 /* Do not reread page->inuse */
4178 /* We do not keep full slabs on the list */
4181 if (free == page->objects) {
4182 list_move(&page->slab_list, &discard);
4184 } else if (free <= SHRINK_PROMOTE_MAX)
4185 list_move(&page->slab_list, promote + free - 1);
4189 * Promote the slabs filled up most to the head of the
4192 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4193 list_splice(promote + i, &n->partial);
4195 spin_unlock_irqrestore(&n->list_lock, flags);
4197 /* Release empty slabs */
4198 list_for_each_entry_safe(page, t, &discard, slab_list)
4199 discard_slab(s, page);
4201 if (slabs_node(s, node))
4209 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4212 * Called with all the locks held after a sched RCU grace period.
4213 * Even if @s becomes empty after shrinking, we can't know that @s
4214 * doesn't have allocations already in-flight and thus can't
4215 * destroy @s until the associated memcg is released.
4217 * However, let's remove the sysfs files for empty caches here.
4218 * Each cache has a lot of interface files which aren't
4219 * particularly useful for empty draining caches; otherwise, we can
4220 * easily end up with millions of unnecessary sysfs files on
4221 * systems which have a lot of memory and transient cgroups.
4223 if (!__kmem_cache_shrink(s))
4224 sysfs_slab_remove(s);
4227 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4230 * Disable empty slabs caching. Used to avoid pinning offline
4231 * memory cgroups by kmem pages that can be freed.
4233 slub_set_cpu_partial(s, 0);
4236 #endif /* CONFIG_MEMCG */
4238 static int slab_mem_going_offline_callback(void *arg)
4240 struct kmem_cache *s;
4242 mutex_lock(&slab_mutex);
4243 list_for_each_entry(s, &slab_caches, list)
4244 __kmem_cache_shrink(s);
4245 mutex_unlock(&slab_mutex);
4250 static void slab_mem_offline_callback(void *arg)
4252 struct kmem_cache_node *n;
4253 struct kmem_cache *s;
4254 struct memory_notify *marg = arg;
4257 offline_node = marg->status_change_nid_normal;
4260 * If the node still has available memory. we need kmem_cache_node
4263 if (offline_node < 0)
4266 mutex_lock(&slab_mutex);
4267 list_for_each_entry(s, &slab_caches, list) {
4268 n = get_node(s, offline_node);
4271 * if n->nr_slabs > 0, slabs still exist on the node
4272 * that is going down. We were unable to free them,
4273 * and offline_pages() function shouldn't call this
4274 * callback. So, we must fail.
4276 BUG_ON(slabs_node(s, offline_node));
4278 s->node[offline_node] = NULL;
4279 kmem_cache_free(kmem_cache_node, n);
4282 mutex_unlock(&slab_mutex);
4285 static int slab_mem_going_online_callback(void *arg)
4287 struct kmem_cache_node *n;
4288 struct kmem_cache *s;
4289 struct memory_notify *marg = arg;
4290 int nid = marg->status_change_nid_normal;
4294 * If the node's memory is already available, then kmem_cache_node is
4295 * already created. Nothing to do.
4301 * We are bringing a node online. No memory is available yet. We must
4302 * allocate a kmem_cache_node structure in order to bring the node
4305 mutex_lock(&slab_mutex);
4306 list_for_each_entry(s, &slab_caches, list) {
4308 * XXX: kmem_cache_alloc_node will fallback to other nodes
4309 * since memory is not yet available from the node that
4312 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4317 init_kmem_cache_node(n);
4321 mutex_unlock(&slab_mutex);
4325 static int slab_memory_callback(struct notifier_block *self,
4326 unsigned long action, void *arg)
4331 case MEM_GOING_ONLINE:
4332 ret = slab_mem_going_online_callback(arg);
4334 case MEM_GOING_OFFLINE:
4335 ret = slab_mem_going_offline_callback(arg);
4338 case MEM_CANCEL_ONLINE:
4339 slab_mem_offline_callback(arg);
4342 case MEM_CANCEL_OFFLINE:
4346 ret = notifier_from_errno(ret);
4352 static struct notifier_block slab_memory_callback_nb = {
4353 .notifier_call = slab_memory_callback,
4354 .priority = SLAB_CALLBACK_PRI,
4357 /********************************************************************
4358 * Basic setup of slabs
4359 *******************************************************************/
4362 * Used for early kmem_cache structures that were allocated using
4363 * the page allocator. Allocate them properly then fix up the pointers
4364 * that may be pointing to the wrong kmem_cache structure.
4367 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4370 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4371 struct kmem_cache_node *n;
4373 memcpy(s, static_cache, kmem_cache->object_size);
4376 * This runs very early, and only the boot processor is supposed to be
4377 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4380 __flush_cpu_slab(s, smp_processor_id());
4381 for_each_kmem_cache_node(s, node, n) {
4384 list_for_each_entry(p, &n->partial, slab_list)
4387 #ifdef CONFIG_SLUB_DEBUG
4388 list_for_each_entry(p, &n->full, slab_list)
4392 slab_init_memcg_params(s);
4393 list_add(&s->list, &slab_caches);
4394 memcg_link_cache(s, NULL);
4398 void __init kmem_cache_init(void)
4400 static __initdata struct kmem_cache boot_kmem_cache,
4401 boot_kmem_cache_node;
4403 if (debug_guardpage_minorder())
4406 kmem_cache_node = &boot_kmem_cache_node;
4407 kmem_cache = &boot_kmem_cache;
4409 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4410 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4412 register_hotmemory_notifier(&slab_memory_callback_nb);
4414 /* Able to allocate the per node structures */
4415 slab_state = PARTIAL;
4417 create_boot_cache(kmem_cache, "kmem_cache",
4418 offsetof(struct kmem_cache, node) +
4419 nr_node_ids * sizeof(struct kmem_cache_node *),
4420 SLAB_HWCACHE_ALIGN, 0, 0);
4422 kmem_cache = bootstrap(&boot_kmem_cache);
4423 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4425 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4426 setup_kmalloc_cache_index_table();
4427 create_kmalloc_caches(0);
4429 /* Setup random freelists for each cache */
4430 init_freelist_randomization();
4432 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4435 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4437 slub_min_order, slub_max_order, slub_min_objects,
4438 nr_cpu_ids, nr_node_ids);
4441 void __init kmem_cache_init_late(void)
4446 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4447 slab_flags_t flags, void (*ctor)(void *))
4449 struct kmem_cache *s, *c;
4451 s = find_mergeable(size, align, flags, name, ctor);
4456 * Adjust the object sizes so that we clear
4457 * the complete object on kzalloc.
4459 s->object_size = max(s->object_size, size);
4460 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4462 for_each_memcg_cache(c, s) {
4463 c->object_size = s->object_size;
4464 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4467 if (sysfs_slab_alias(s, name)) {
4476 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4480 err = kmem_cache_open(s, flags);
4484 /* Mutex is not taken during early boot */
4485 if (slab_state <= UP)
4488 memcg_propagate_slab_attrs(s);
4489 err = sysfs_slab_add(s);
4491 __kmem_cache_release(s);
4496 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4498 struct kmem_cache *s;
4501 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4502 return kmalloc_large(size, gfpflags);
4504 s = kmalloc_slab(size, gfpflags);
4506 if (unlikely(ZERO_OR_NULL_PTR(s)))
4509 ret = slab_alloc(s, gfpflags, caller);
4511 /* Honor the call site pointer we received. */
4512 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4516 EXPORT_SYMBOL(__kmalloc_track_caller);
4519 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4520 int node, unsigned long caller)
4522 struct kmem_cache *s;
4525 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4526 ret = kmalloc_large_node(size, gfpflags, node);
4528 trace_kmalloc_node(caller, ret,
4529 size, PAGE_SIZE << get_order(size),
4535 s = kmalloc_slab(size, gfpflags);
4537 if (unlikely(ZERO_OR_NULL_PTR(s)))
4540 ret = slab_alloc_node(s, gfpflags, node, caller);
4542 /* Honor the call site pointer we received. */
4543 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4547 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4551 static int count_inuse(struct page *page)
4556 static int count_total(struct page *page)
4558 return page->objects;
4562 #ifdef CONFIG_SLUB_DEBUG
4563 static void validate_slab(struct kmem_cache *s, struct page *page)
4566 void *addr = page_address(page);
4571 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4574 /* Now we know that a valid freelist exists */
4575 map = get_map(s, page);
4576 for_each_object(p, s, addr, page->objects) {
4577 u8 val = test_bit(slab_index(p, s, addr), map) ?
4578 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4580 if (!check_object(s, page, p, val))
4588 static int validate_slab_node(struct kmem_cache *s,
4589 struct kmem_cache_node *n)
4591 unsigned long count = 0;
4593 unsigned long flags;
4595 spin_lock_irqsave(&n->list_lock, flags);
4597 list_for_each_entry(page, &n->partial, slab_list) {
4598 validate_slab(s, page);
4601 if (count != n->nr_partial)
4602 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4603 s->name, count, n->nr_partial);
4605 if (!(s->flags & SLAB_STORE_USER))
4608 list_for_each_entry(page, &n->full, slab_list) {
4609 validate_slab(s, page);
4612 if (count != atomic_long_read(&n->nr_slabs))
4613 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4614 s->name, count, atomic_long_read(&n->nr_slabs));
4617 spin_unlock_irqrestore(&n->list_lock, flags);
4621 static long validate_slab_cache(struct kmem_cache *s)
4624 unsigned long count = 0;
4625 struct kmem_cache_node *n;
4628 for_each_kmem_cache_node(s, node, n)
4629 count += validate_slab_node(s, n);
4634 * Generate lists of code addresses where slabcache objects are allocated
4639 unsigned long count;
4646 DECLARE_BITMAP(cpus, NR_CPUS);
4652 unsigned long count;
4653 struct location *loc;
4656 static void free_loc_track(struct loc_track *t)
4659 free_pages((unsigned long)t->loc,
4660 get_order(sizeof(struct location) * t->max));
4663 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4668 order = get_order(sizeof(struct location) * max);
4670 l = (void *)__get_free_pages(flags, order);
4675 memcpy(l, t->loc, sizeof(struct location) * t->count);
4683 static int add_location(struct loc_track *t, struct kmem_cache *s,
4684 const struct track *track)
4686 long start, end, pos;
4688 unsigned long caddr;
4689 unsigned long age = jiffies - track->when;
4695 pos = start + (end - start + 1) / 2;
4698 * There is nothing at "end". If we end up there
4699 * we need to add something to before end.
4704 caddr = t->loc[pos].addr;
4705 if (track->addr == caddr) {
4711 if (age < l->min_time)
4713 if (age > l->max_time)
4716 if (track->pid < l->min_pid)
4717 l->min_pid = track->pid;
4718 if (track->pid > l->max_pid)
4719 l->max_pid = track->pid;
4721 cpumask_set_cpu(track->cpu,
4722 to_cpumask(l->cpus));
4724 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4728 if (track->addr < caddr)
4735 * Not found. Insert new tracking element.
4737 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4743 (t->count - pos) * sizeof(struct location));
4746 l->addr = track->addr;
4750 l->min_pid = track->pid;
4751 l->max_pid = track->pid;
4752 cpumask_clear(to_cpumask(l->cpus));
4753 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4754 nodes_clear(l->nodes);
4755 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4759 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4760 struct page *page, enum track_item alloc)
4762 void *addr = page_address(page);
4766 map = get_map(s, page);
4767 for_each_object(p, s, addr, page->objects)
4768 if (!test_bit(slab_index(p, s, addr), map))
4769 add_location(t, s, get_track(s, p, alloc));
4773 static int list_locations(struct kmem_cache *s, char *buf,
4774 enum track_item alloc)
4778 struct loc_track t = { 0, 0, NULL };
4780 struct kmem_cache_node *n;
4782 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4784 return sprintf(buf, "Out of memory\n");
4786 /* Push back cpu slabs */
4789 for_each_kmem_cache_node(s, node, n) {
4790 unsigned long flags;
4793 if (!atomic_long_read(&n->nr_slabs))
4796 spin_lock_irqsave(&n->list_lock, flags);
4797 list_for_each_entry(page, &n->partial, slab_list)
4798 process_slab(&t, s, page, alloc);
4799 list_for_each_entry(page, &n->full, slab_list)
4800 process_slab(&t, s, page, alloc);
4801 spin_unlock_irqrestore(&n->list_lock, flags);
4804 for (i = 0; i < t.count; i++) {
4805 struct location *l = &t.loc[i];
4807 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4809 len += sprintf(buf + len, "%7ld ", l->count);
4812 len += sprintf(buf + len, "%pS", (void *)l->addr);
4814 len += sprintf(buf + len, "<not-available>");
4816 if (l->sum_time != l->min_time) {
4817 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4819 (long)div_u64(l->sum_time, l->count),
4822 len += sprintf(buf + len, " age=%ld",
4825 if (l->min_pid != l->max_pid)
4826 len += sprintf(buf + len, " pid=%ld-%ld",
4827 l->min_pid, l->max_pid);
4829 len += sprintf(buf + len, " pid=%ld",
4832 if (num_online_cpus() > 1 &&
4833 !cpumask_empty(to_cpumask(l->cpus)) &&
4834 len < PAGE_SIZE - 60)
4835 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4837 cpumask_pr_args(to_cpumask(l->cpus)));
4839 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4840 len < PAGE_SIZE - 60)
4841 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4843 nodemask_pr_args(&l->nodes));
4845 len += sprintf(buf + len, "\n");
4850 len += sprintf(buf, "No data\n");
4853 #endif /* CONFIG_SLUB_DEBUG */
4855 #ifdef SLUB_RESILIENCY_TEST
4856 static void __init resiliency_test(void)
4859 int type = KMALLOC_NORMAL;
4861 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4863 pr_err("SLUB resiliency testing\n");
4864 pr_err("-----------------------\n");
4865 pr_err("A. Corruption after allocation\n");
4867 p = kzalloc(16, GFP_KERNEL);
4869 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4872 validate_slab_cache(kmalloc_caches[type][4]);
4874 /* Hmmm... The next two are dangerous */
4875 p = kzalloc(32, GFP_KERNEL);
4876 p[32 + sizeof(void *)] = 0x34;
4877 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4879 pr_err("If allocated object is overwritten then not detectable\n\n");
4881 validate_slab_cache(kmalloc_caches[type][5]);
4882 p = kzalloc(64, GFP_KERNEL);
4883 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4885 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4887 pr_err("If allocated object is overwritten then not detectable\n\n");
4888 validate_slab_cache(kmalloc_caches[type][6]);
4890 pr_err("\nB. Corruption after free\n");
4891 p = kzalloc(128, GFP_KERNEL);
4894 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4895 validate_slab_cache(kmalloc_caches[type][7]);
4897 p = kzalloc(256, GFP_KERNEL);
4900 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4901 validate_slab_cache(kmalloc_caches[type][8]);
4903 p = kzalloc(512, GFP_KERNEL);
4906 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4907 validate_slab_cache(kmalloc_caches[type][9]);
4911 static void resiliency_test(void) {};
4913 #endif /* SLUB_RESILIENCY_TEST */
4916 enum slab_stat_type {
4917 SL_ALL, /* All slabs */
4918 SL_PARTIAL, /* Only partially allocated slabs */
4919 SL_CPU, /* Only slabs used for cpu caches */
4920 SL_OBJECTS, /* Determine allocated objects not slabs */
4921 SL_TOTAL /* Determine object capacity not slabs */
4924 #define SO_ALL (1 << SL_ALL)
4925 #define SO_PARTIAL (1 << SL_PARTIAL)
4926 #define SO_CPU (1 << SL_CPU)
4927 #define SO_OBJECTS (1 << SL_OBJECTS)
4928 #define SO_TOTAL (1 << SL_TOTAL)
4931 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4933 static int __init setup_slub_memcg_sysfs(char *str)
4937 if (get_option(&str, &v) > 0)
4938 memcg_sysfs_enabled = v;
4943 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4946 static ssize_t show_slab_objects(struct kmem_cache *s,
4947 char *buf, unsigned long flags)
4949 unsigned long total = 0;
4952 unsigned long *nodes;
4954 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4958 if (flags & SO_CPU) {
4961 for_each_possible_cpu(cpu) {
4962 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4967 page = READ_ONCE(c->page);
4971 node = page_to_nid(page);
4972 if (flags & SO_TOTAL)
4974 else if (flags & SO_OBJECTS)
4982 page = slub_percpu_partial_read_once(c);
4984 node = page_to_nid(page);
4985 if (flags & SO_TOTAL)
4987 else if (flags & SO_OBJECTS)
4998 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4999 * already held which will conflict with an existing lock order:
5001 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5003 * We don't really need mem_hotplug_lock (to hold off
5004 * slab_mem_going_offline_callback) here because slab's memory hot
5005 * unplug code doesn't destroy the kmem_cache->node[] data.
5008 #ifdef CONFIG_SLUB_DEBUG
5009 if (flags & SO_ALL) {
5010 struct kmem_cache_node *n;
5012 for_each_kmem_cache_node(s, node, n) {
5014 if (flags & SO_TOTAL)
5015 x = atomic_long_read(&n->total_objects);
5016 else if (flags & SO_OBJECTS)
5017 x = atomic_long_read(&n->total_objects) -
5018 count_partial(n, count_free);
5020 x = atomic_long_read(&n->nr_slabs);
5027 if (flags & SO_PARTIAL) {
5028 struct kmem_cache_node *n;
5030 for_each_kmem_cache_node(s, node, n) {
5031 if (flags & SO_TOTAL)
5032 x = count_partial(n, count_total);
5033 else if (flags & SO_OBJECTS)
5034 x = count_partial(n, count_inuse);
5041 x = sprintf(buf, "%lu", total);
5043 for (node = 0; node < nr_node_ids; node++)
5045 x += sprintf(buf + x, " N%d=%lu",
5049 return x + sprintf(buf + x, "\n");
5052 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5053 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5055 struct slab_attribute {
5056 struct attribute attr;
5057 ssize_t (*show)(struct kmem_cache *s, char *buf);
5058 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5061 #define SLAB_ATTR_RO(_name) \
5062 static struct slab_attribute _name##_attr = \
5063 __ATTR(_name, 0400, _name##_show, NULL)
5065 #define SLAB_ATTR(_name) \
5066 static struct slab_attribute _name##_attr = \
5067 __ATTR(_name, 0600, _name##_show, _name##_store)
5069 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5071 return sprintf(buf, "%u\n", s->size);
5073 SLAB_ATTR_RO(slab_size);
5075 static ssize_t align_show(struct kmem_cache *s, char *buf)
5077 return sprintf(buf, "%u\n", s->align);
5079 SLAB_ATTR_RO(align);
5081 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5083 return sprintf(buf, "%u\n", s->object_size);
5085 SLAB_ATTR_RO(object_size);
5087 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5089 return sprintf(buf, "%u\n", oo_objects(s->oo));
5091 SLAB_ATTR_RO(objs_per_slab);
5093 static ssize_t order_show(struct kmem_cache *s, char *buf)
5095 return sprintf(buf, "%u\n", oo_order(s->oo));
5097 SLAB_ATTR_RO(order);
5099 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5101 return sprintf(buf, "%lu\n", s->min_partial);
5104 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5110 err = kstrtoul(buf, 10, &min);
5114 set_min_partial(s, min);
5117 SLAB_ATTR(min_partial);
5119 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5121 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5124 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5127 unsigned int objects;
5130 err = kstrtouint(buf, 10, &objects);
5133 if (objects && !kmem_cache_has_cpu_partial(s))
5136 slub_set_cpu_partial(s, objects);
5140 SLAB_ATTR(cpu_partial);
5142 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5146 return sprintf(buf, "%pS\n", s->ctor);
5150 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5152 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5154 SLAB_ATTR_RO(aliases);
5156 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5158 return show_slab_objects(s, buf, SO_PARTIAL);
5160 SLAB_ATTR_RO(partial);
5162 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5164 return show_slab_objects(s, buf, SO_CPU);
5166 SLAB_ATTR_RO(cpu_slabs);
5168 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5170 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5172 SLAB_ATTR_RO(objects);
5174 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5176 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5178 SLAB_ATTR_RO(objects_partial);
5180 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5187 for_each_online_cpu(cpu) {
5190 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5193 pages += page->pages;
5194 objects += page->pobjects;
5198 len = sprintf(buf, "%d(%d)", objects, pages);
5201 for_each_online_cpu(cpu) {
5204 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5206 if (page && len < PAGE_SIZE - 20)
5207 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5208 page->pobjects, page->pages);
5211 return len + sprintf(buf + len, "\n");
5213 SLAB_ATTR_RO(slabs_cpu_partial);
5215 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5217 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5219 SLAB_ATTR_RO(reclaim_account);
5221 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5223 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5225 SLAB_ATTR_RO(hwcache_align);
5227 #ifdef CONFIG_ZONE_DMA
5228 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5230 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5232 SLAB_ATTR_RO(cache_dma);
5235 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5237 return sprintf(buf, "%u\n", s->usersize);
5239 SLAB_ATTR_RO(usersize);
5241 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5243 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5245 SLAB_ATTR_RO(destroy_by_rcu);
5247 #ifdef CONFIG_SLUB_DEBUG
5248 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5250 return show_slab_objects(s, buf, SO_ALL);
5252 SLAB_ATTR_RO(slabs);
5254 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5256 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5258 SLAB_ATTR_RO(total_objects);
5260 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5262 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5264 SLAB_ATTR_RO(sanity_checks);
5266 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5268 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5270 SLAB_ATTR_RO(trace);
5272 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5274 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5277 SLAB_ATTR_RO(red_zone);
5279 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5281 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5284 SLAB_ATTR_RO(poison);
5286 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5288 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5291 SLAB_ATTR_RO(store_user);
5293 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5298 static ssize_t validate_store(struct kmem_cache *s,
5299 const char *buf, size_t length)
5303 if (buf[0] == '1') {
5304 ret = validate_slab_cache(s);
5310 SLAB_ATTR(validate);
5312 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5314 if (!(s->flags & SLAB_STORE_USER))
5316 return list_locations(s, buf, TRACK_ALLOC);
5318 SLAB_ATTR_RO(alloc_calls);
5320 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5322 if (!(s->flags & SLAB_STORE_USER))
5324 return list_locations(s, buf, TRACK_FREE);
5326 SLAB_ATTR_RO(free_calls);
5327 #endif /* CONFIG_SLUB_DEBUG */
5329 #ifdef CONFIG_FAILSLAB
5330 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5332 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5334 SLAB_ATTR_RO(failslab);
5337 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5342 static ssize_t shrink_store(struct kmem_cache *s,
5343 const char *buf, size_t length)
5346 kmem_cache_shrink_all(s);
5354 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5356 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5359 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5360 const char *buf, size_t length)
5365 err = kstrtouint(buf, 10, &ratio);
5371 s->remote_node_defrag_ratio = ratio * 10;
5375 SLAB_ATTR(remote_node_defrag_ratio);
5378 #ifdef CONFIG_SLUB_STATS
5379 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5381 unsigned long sum = 0;
5384 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5389 for_each_online_cpu(cpu) {
5390 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5396 len = sprintf(buf, "%lu", sum);
5399 for_each_online_cpu(cpu) {
5400 if (data[cpu] && len < PAGE_SIZE - 20)
5401 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5405 return len + sprintf(buf + len, "\n");
5408 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5412 for_each_online_cpu(cpu)
5413 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5416 #define STAT_ATTR(si, text) \
5417 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5419 return show_stat(s, buf, si); \
5421 static ssize_t text##_store(struct kmem_cache *s, \
5422 const char *buf, size_t length) \
5424 if (buf[0] != '0') \
5426 clear_stat(s, si); \
5431 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5432 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5433 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5434 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5435 STAT_ATTR(FREE_FROZEN, free_frozen);
5436 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5437 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5438 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5439 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5440 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5441 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5442 STAT_ATTR(FREE_SLAB, free_slab);
5443 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5444 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5445 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5446 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5447 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5448 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5449 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5450 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5451 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5452 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5453 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5454 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5455 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5456 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5457 #endif /* CONFIG_SLUB_STATS */
5459 static struct attribute *slab_attrs[] = {
5460 &slab_size_attr.attr,
5461 &object_size_attr.attr,
5462 &objs_per_slab_attr.attr,
5464 &min_partial_attr.attr,
5465 &cpu_partial_attr.attr,
5467 &objects_partial_attr.attr,
5469 &cpu_slabs_attr.attr,
5473 &hwcache_align_attr.attr,
5474 &reclaim_account_attr.attr,
5475 &destroy_by_rcu_attr.attr,
5477 &slabs_cpu_partial_attr.attr,
5478 #ifdef CONFIG_SLUB_DEBUG
5479 &total_objects_attr.attr,
5481 &sanity_checks_attr.attr,
5483 &red_zone_attr.attr,
5485 &store_user_attr.attr,
5486 &validate_attr.attr,
5487 &alloc_calls_attr.attr,
5488 &free_calls_attr.attr,
5490 #ifdef CONFIG_ZONE_DMA
5491 &cache_dma_attr.attr,
5494 &remote_node_defrag_ratio_attr.attr,
5496 #ifdef CONFIG_SLUB_STATS
5497 &alloc_fastpath_attr.attr,
5498 &alloc_slowpath_attr.attr,
5499 &free_fastpath_attr.attr,
5500 &free_slowpath_attr.attr,
5501 &free_frozen_attr.attr,
5502 &free_add_partial_attr.attr,
5503 &free_remove_partial_attr.attr,
5504 &alloc_from_partial_attr.attr,
5505 &alloc_slab_attr.attr,
5506 &alloc_refill_attr.attr,
5507 &alloc_node_mismatch_attr.attr,
5508 &free_slab_attr.attr,
5509 &cpuslab_flush_attr.attr,
5510 &deactivate_full_attr.attr,
5511 &deactivate_empty_attr.attr,
5512 &deactivate_to_head_attr.attr,
5513 &deactivate_to_tail_attr.attr,
5514 &deactivate_remote_frees_attr.attr,
5515 &deactivate_bypass_attr.attr,
5516 &order_fallback_attr.attr,
5517 &cmpxchg_double_fail_attr.attr,
5518 &cmpxchg_double_cpu_fail_attr.attr,
5519 &cpu_partial_alloc_attr.attr,
5520 &cpu_partial_free_attr.attr,
5521 &cpu_partial_node_attr.attr,
5522 &cpu_partial_drain_attr.attr,
5524 #ifdef CONFIG_FAILSLAB
5525 &failslab_attr.attr,
5527 &usersize_attr.attr,
5532 static const struct attribute_group slab_attr_group = {
5533 .attrs = slab_attrs,
5536 static ssize_t slab_attr_show(struct kobject *kobj,
5537 struct attribute *attr,
5540 struct slab_attribute *attribute;
5541 struct kmem_cache *s;
5544 attribute = to_slab_attr(attr);
5547 if (!attribute->show)
5550 err = attribute->show(s, buf);
5555 static ssize_t slab_attr_store(struct kobject *kobj,
5556 struct attribute *attr,
5557 const char *buf, size_t len)
5559 struct slab_attribute *attribute;
5560 struct kmem_cache *s;
5563 attribute = to_slab_attr(attr);
5566 if (!attribute->store)
5569 err = attribute->store(s, buf, len);
5571 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5572 struct kmem_cache *c;
5574 mutex_lock(&slab_mutex);
5575 if (s->max_attr_size < len)
5576 s->max_attr_size = len;
5579 * This is a best effort propagation, so this function's return
5580 * value will be determined by the parent cache only. This is
5581 * basically because not all attributes will have a well
5582 * defined semantics for rollbacks - most of the actions will
5583 * have permanent effects.
5585 * Returning the error value of any of the children that fail
5586 * is not 100 % defined, in the sense that users seeing the
5587 * error code won't be able to know anything about the state of
5590 * Only returning the error code for the parent cache at least
5591 * has well defined semantics. The cache being written to
5592 * directly either failed or succeeded, in which case we loop
5593 * through the descendants with best-effort propagation.
5595 for_each_memcg_cache(c, s)
5596 attribute->store(c, buf, len);
5597 mutex_unlock(&slab_mutex);
5603 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5607 char *buffer = NULL;
5608 struct kmem_cache *root_cache;
5610 if (is_root_cache(s))
5613 root_cache = s->memcg_params.root_cache;
5616 * This mean this cache had no attribute written. Therefore, no point
5617 * in copying default values around
5619 if (!root_cache->max_attr_size)
5622 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5625 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5628 if (!attr || !attr->store || !attr->show)
5632 * It is really bad that we have to allocate here, so we will
5633 * do it only as a fallback. If we actually allocate, though,
5634 * we can just use the allocated buffer until the end.
5636 * Most of the slub attributes will tend to be very small in
5637 * size, but sysfs allows buffers up to a page, so they can
5638 * theoretically happen.
5642 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5643 !IS_ENABLED(CONFIG_SLUB_STATS))
5646 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5647 if (WARN_ON(!buffer))
5652 len = attr->show(root_cache, buf);
5654 attr->store(s, buf, len);
5658 free_page((unsigned long)buffer);
5659 #endif /* CONFIG_MEMCG */
5662 static void kmem_cache_release(struct kobject *k)
5664 slab_kmem_cache_release(to_slab(k));
5667 static const struct sysfs_ops slab_sysfs_ops = {
5668 .show = slab_attr_show,
5669 .store = slab_attr_store,
5672 static struct kobj_type slab_ktype = {
5673 .sysfs_ops = &slab_sysfs_ops,
5674 .release = kmem_cache_release,
5677 static struct kset *slab_kset;
5679 static inline struct kset *cache_kset(struct kmem_cache *s)
5682 if (!is_root_cache(s))
5683 return s->memcg_params.root_cache->memcg_kset;
5688 #define ID_STR_LENGTH 64
5690 /* Create a unique string id for a slab cache:
5692 * Format :[flags-]size
5694 static char *create_unique_id(struct kmem_cache *s)
5696 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5703 * First flags affecting slabcache operations. We will only
5704 * get here for aliasable slabs so we do not need to support
5705 * too many flags. The flags here must cover all flags that
5706 * are matched during merging to guarantee that the id is
5709 if (s->flags & SLAB_CACHE_DMA)
5711 if (s->flags & SLAB_CACHE_DMA32)
5713 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5715 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5717 if (s->flags & SLAB_ACCOUNT)
5721 p += sprintf(p, "%07u", s->size);
5723 BUG_ON(p > name + ID_STR_LENGTH - 1);
5727 static void sysfs_slab_remove_workfn(struct work_struct *work)
5729 struct kmem_cache *s =
5730 container_of(work, struct kmem_cache, kobj_remove_work);
5732 if (!s->kobj.state_in_sysfs)
5734 * For a memcg cache, this may be called during
5735 * deactivation and again on shutdown. Remove only once.
5736 * A cache is never shut down before deactivation is
5737 * complete, so no need to worry about synchronization.
5742 kset_unregister(s->memcg_kset);
5745 kobject_put(&s->kobj);
5748 static int sysfs_slab_add(struct kmem_cache *s)
5752 struct kset *kset = cache_kset(s);
5753 int unmergeable = slab_unmergeable(s);
5755 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5758 kobject_init(&s->kobj, &slab_ktype);
5762 if (!unmergeable && disable_higher_order_debug &&
5763 (slub_debug & DEBUG_METADATA_FLAGS))
5768 * Slabcache can never be merged so we can use the name proper.
5769 * This is typically the case for debug situations. In that
5770 * case we can catch duplicate names easily.
5772 sysfs_remove_link(&slab_kset->kobj, s->name);
5776 * Create a unique name for the slab as a target
5779 name = create_unique_id(s);
5782 s->kobj.kset = kset;
5783 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5785 kobject_put(&s->kobj);
5789 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5794 if (is_root_cache(s) && memcg_sysfs_enabled) {
5795 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5796 if (!s->memcg_kset) {
5804 /* Setup first alias */
5805 sysfs_slab_alias(s, s->name);
5812 kobject_del(&s->kobj);
5816 static void sysfs_slab_remove(struct kmem_cache *s)
5818 if (slab_state < FULL)
5820 * Sysfs has not been setup yet so no need to remove the
5825 kobject_get(&s->kobj);
5826 schedule_work(&s->kobj_remove_work);
5829 void sysfs_slab_unlink(struct kmem_cache *s)
5831 if (slab_state >= FULL)
5832 kobject_del(&s->kobj);
5835 void sysfs_slab_release(struct kmem_cache *s)
5837 if (slab_state >= FULL)
5838 kobject_put(&s->kobj);
5842 * Need to buffer aliases during bootup until sysfs becomes
5843 * available lest we lose that information.
5845 struct saved_alias {
5846 struct kmem_cache *s;
5848 struct saved_alias *next;
5851 static struct saved_alias *alias_list;
5853 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5855 struct saved_alias *al;
5857 if (slab_state == FULL) {
5859 * If we have a leftover link then remove it.
5861 sysfs_remove_link(&slab_kset->kobj, name);
5862 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5865 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5871 al->next = alias_list;
5876 static int __init slab_sysfs_init(void)
5878 struct kmem_cache *s;
5881 mutex_lock(&slab_mutex);
5883 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5885 mutex_unlock(&slab_mutex);
5886 pr_err("Cannot register slab subsystem.\n");
5892 list_for_each_entry(s, &slab_caches, list) {
5893 err = sysfs_slab_add(s);
5895 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5899 while (alias_list) {
5900 struct saved_alias *al = alias_list;
5902 alias_list = alias_list->next;
5903 err = sysfs_slab_alias(al->s, al->name);
5905 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5910 mutex_unlock(&slab_mutex);
5915 __initcall(slab_sysfs_init);
5916 #endif /* CONFIG_SYSFS */
5919 * The /proc/slabinfo ABI
5921 #ifdef CONFIG_SLUB_DEBUG
5922 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5924 unsigned long nr_slabs = 0;
5925 unsigned long nr_objs = 0;
5926 unsigned long nr_free = 0;
5928 struct kmem_cache_node *n;
5930 for_each_kmem_cache_node(s, node, n) {
5931 nr_slabs += node_nr_slabs(n);
5932 nr_objs += node_nr_objs(n);
5933 nr_free += count_partial(n, count_free);
5936 sinfo->active_objs = nr_objs - nr_free;
5937 sinfo->num_objs = nr_objs;
5938 sinfo->active_slabs = nr_slabs;
5939 sinfo->num_slabs = nr_slabs;
5940 sinfo->objects_per_slab = oo_objects(s->oo);
5941 sinfo->cache_order = oo_order(s->oo);
5944 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5948 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5949 size_t count, loff_t *ppos)
5953 #endif /* CONFIG_SLUB_DEBUG */