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 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 void *fixup_red_left(struct kmem_cache *s, void *p)
128 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
129 p += s->red_left_pad;
134 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
136 #ifdef CONFIG_SLUB_CPU_PARTIAL
137 return !kmem_cache_debug(s);
144 * Issues still to be resolved:
146 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 * - Variable sizing of the per node arrays
151 /* Enable to test recovery from slab corruption on boot */
152 #undef SLUB_RESILIENCY_TEST
154 /* Enable to log cmpxchg failures */
155 #undef SLUB_DEBUG_CMPXCHG
158 * Mininum number of partial slabs. These will be left on the partial
159 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 #define MIN_PARTIAL 5
164 * Maximum number of desirable partial slabs.
165 * The existence of more partial slabs makes kmem_cache_shrink
166 * sort the partial list by the number of objects in use.
168 #define MAX_PARTIAL 10
170 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_STORE_USER)
174 * These debug flags cannot use CMPXCHG because there might be consistency
175 * issues when checking or reading debug information
177 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
182 * Debugging flags that require metadata to be stored in the slab. These get
183 * disabled when slub_debug=O is used and a cache's min order increases with
186 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
189 #define OO_MASK ((1 << OO_SHIFT) - 1)
190 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
192 /* Internal SLUB flags */
194 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
195 /* Use cmpxchg_double */
196 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
199 * Tracking user of a slab.
201 #define TRACK_ADDRS_COUNT 16
203 unsigned long addr; /* Called from address */
204 #ifdef CONFIG_STACKTRACE
205 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
207 int cpu; /* Was running on cpu */
208 int pid; /* Pid context */
209 unsigned long when; /* When did the operation occur */
212 enum track_item { TRACK_ALLOC, TRACK_FREE };
215 static int sysfs_slab_add(struct kmem_cache *);
216 static int sysfs_slab_alias(struct kmem_cache *, const char *);
217 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
218 static void sysfs_slab_remove(struct kmem_cache *s);
220 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
221 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
223 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
224 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
227 static inline void stat(const struct kmem_cache *s, enum stat_item si)
229 #ifdef CONFIG_SLUB_STATS
231 * The rmw is racy on a preemptible kernel but this is acceptable, so
232 * avoid this_cpu_add()'s irq-disable overhead.
234 raw_cpu_inc(s->cpu_slab->stat[si]);
238 /********************************************************************
239 * Core slab cache functions
240 *******************************************************************/
243 * Returns freelist pointer (ptr). With hardening, this is obfuscated
244 * with an XOR of the address where the pointer is held and a per-cache
247 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
248 unsigned long ptr_addr)
250 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
253 * Normally, this doesn't cause any issues, as both set_freepointer()
254 * and get_freepointer() are called with a pointer with the same tag.
255 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256 * example, when __free_slub() iterates over objects in a cache, it
257 * passes untagged pointers to check_object(). check_object() in turns
258 * calls get_freepointer() with an untagged pointer, which causes the
259 * freepointer to be restored incorrectly.
261 return (void *)((unsigned long)ptr ^ s->random ^
262 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
268 /* Returns the freelist pointer recorded at location ptr_addr. */
269 static inline void *freelist_dereference(const struct kmem_cache *s,
272 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
273 (unsigned long)ptr_addr);
276 static inline void *get_freepointer(struct kmem_cache *s, void *object)
278 return freelist_dereference(s, object + s->offset);
281 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
283 prefetch(object + s->offset);
286 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
288 unsigned long freepointer_addr;
291 if (!debug_pagealloc_enabled_static())
292 return get_freepointer(s, object);
294 freepointer_addr = (unsigned long)object + s->offset;
295 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
296 return freelist_ptr(s, p, freepointer_addr);
299 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
301 unsigned long freeptr_addr = (unsigned long)object + s->offset;
303 #ifdef CONFIG_SLAB_FREELIST_HARDENED
304 BUG_ON(object == fp); /* naive detection of double free or corruption */
307 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
310 /* Loop over all objects in a slab */
311 #define for_each_object(__p, __s, __addr, __objects) \
312 for (__p = fixup_red_left(__s, __addr); \
313 __p < (__addr) + (__objects) * (__s)->size; \
316 /* Determine object index from a given position */
317 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
319 return (kasan_reset_tag(p) - addr) / s->size;
322 static inline unsigned int order_objects(unsigned int order, unsigned int size)
324 return ((unsigned int)PAGE_SIZE << order) / size;
327 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
330 struct kmem_cache_order_objects x = {
331 (order << OO_SHIFT) + order_objects(order, size)
337 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
339 return x.x >> OO_SHIFT;
342 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
344 return x.x & OO_MASK;
348 * Per slab locking using the pagelock
350 static __always_inline void slab_lock(struct page *page)
352 VM_BUG_ON_PAGE(PageTail(page), page);
353 bit_spin_lock(PG_locked, &page->flags);
356 static __always_inline void slab_unlock(struct page *page)
358 VM_BUG_ON_PAGE(PageTail(page), page);
359 __bit_spin_unlock(PG_locked, &page->flags);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
368 VM_BUG_ON(!irqs_disabled());
369 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
370 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
371 if (s->flags & __CMPXCHG_DOUBLE) {
372 if (cmpxchg_double(&page->freelist, &page->counters,
373 freelist_old, counters_old,
374 freelist_new, counters_new))
380 if (page->freelist == freelist_old &&
381 page->counters == counters_old) {
382 page->freelist = freelist_new;
383 page->counters = counters_new;
391 stat(s, CMPXCHG_DOUBLE_FAIL);
393 #ifdef SLUB_DEBUG_CMPXCHG
394 pr_info("%s %s: cmpxchg double redo ", n, s->name);
400 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
401 void *freelist_old, unsigned long counters_old,
402 void *freelist_new, unsigned long counters_new,
405 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
406 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
407 if (s->flags & __CMPXCHG_DOUBLE) {
408 if (cmpxchg_double(&page->freelist, &page->counters,
409 freelist_old, counters_old,
410 freelist_new, counters_new))
417 local_irq_save(flags);
419 if (page->freelist == freelist_old &&
420 page->counters == counters_old) {
421 page->freelist = freelist_new;
422 page->counters = counters_new;
424 local_irq_restore(flags);
428 local_irq_restore(flags);
432 stat(s, CMPXCHG_DOUBLE_FAIL);
434 #ifdef SLUB_DEBUG_CMPXCHG
435 pr_info("%s %s: cmpxchg double redo ", n, s->name);
441 #ifdef CONFIG_SLUB_DEBUG
442 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
443 static DEFINE_SPINLOCK(object_map_lock);
446 * Determine a map of object in use on a page.
448 * Node listlock must be held to guarantee that the page does
449 * not vanish from under us.
451 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
452 __acquires(&object_map_lock)
455 void *addr = page_address(page);
457 VM_BUG_ON(!irqs_disabled());
459 spin_lock(&object_map_lock);
461 bitmap_zero(object_map, page->objects);
463 for (p = page->freelist; p; p = get_freepointer(s, p))
464 set_bit(slab_index(p, s, addr), object_map);
469 static void put_map(unsigned long *map) __releases(&object_map_lock)
471 VM_BUG_ON(map != object_map);
472 lockdep_assert_held(&object_map_lock);
474 spin_unlock(&object_map_lock);
477 static inline unsigned int size_from_object(struct kmem_cache *s)
479 if (s->flags & SLAB_RED_ZONE)
480 return s->size - s->red_left_pad;
485 static inline void *restore_red_left(struct kmem_cache *s, void *p)
487 if (s->flags & SLAB_RED_ZONE)
488 p -= s->red_left_pad;
496 #if defined(CONFIG_SLUB_DEBUG_ON)
497 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
499 static slab_flags_t slub_debug;
502 static char *slub_debug_slabs;
503 static int disable_higher_order_debug;
506 * slub is about to manipulate internal object metadata. This memory lies
507 * outside the range of the allocated object, so accessing it would normally
508 * be reported by kasan as a bounds error. metadata_access_enable() is used
509 * to tell kasan that these accesses are OK.
511 static inline void metadata_access_enable(void)
513 kasan_disable_current();
516 static inline void metadata_access_disable(void)
518 kasan_enable_current();
525 /* Verify that a pointer has an address that is valid within a slab page */
526 static inline int check_valid_pointer(struct kmem_cache *s,
527 struct page *page, void *object)
534 base = page_address(page);
535 object = kasan_reset_tag(object);
536 object = restore_red_left(s, object);
537 if (object < base || object >= base + page->objects * s->size ||
538 (object - base) % s->size) {
545 static void print_section(char *level, char *text, u8 *addr,
548 metadata_access_enable();
549 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
551 metadata_access_disable();
555 * See comment in calculate_sizes().
557 static inline bool freeptr_outside_object(struct kmem_cache *s)
559 return s->offset >= s->inuse;
563 * Return offset of the end of info block which is inuse + free pointer if
564 * not overlapping with object.
566 static inline unsigned int get_info_end(struct kmem_cache *s)
568 if (freeptr_outside_object(s))
569 return s->inuse + sizeof(void *);
574 static struct track *get_track(struct kmem_cache *s, void *object,
575 enum track_item alloc)
579 p = object + get_info_end(s);
584 static void set_track(struct kmem_cache *s, void *object,
585 enum track_item alloc, unsigned long addr)
587 struct track *p = get_track(s, object, alloc);
590 #ifdef CONFIG_STACKTRACE
591 unsigned int nr_entries;
593 metadata_access_enable();
594 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
595 metadata_access_disable();
597 if (nr_entries < TRACK_ADDRS_COUNT)
598 p->addrs[nr_entries] = 0;
601 p->cpu = smp_processor_id();
602 p->pid = current->pid;
605 memset(p, 0, sizeof(struct track));
609 static void init_tracking(struct kmem_cache *s, void *object)
611 if (!(s->flags & SLAB_STORE_USER))
614 set_track(s, object, TRACK_FREE, 0UL);
615 set_track(s, object, TRACK_ALLOC, 0UL);
618 static void print_track(const char *s, struct track *t, unsigned long pr_time)
623 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
624 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
625 #ifdef CONFIG_STACKTRACE
628 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
630 pr_err("\t%pS\n", (void *)t->addrs[i]);
637 static void print_tracking(struct kmem_cache *s, void *object)
639 unsigned long pr_time = jiffies;
640 if (!(s->flags & SLAB_STORE_USER))
643 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
644 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
647 static void print_page_info(struct page *page)
649 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
650 page, page->objects, page->inuse, page->freelist, page->flags);
654 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
656 struct va_format vaf;
662 pr_err("=============================================================================\n");
663 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
664 pr_err("-----------------------------------------------------------------------------\n\n");
666 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
670 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
672 struct va_format vaf;
678 pr_err("FIX %s: %pV\n", s->name, &vaf);
682 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
683 void *freelist, void *nextfree)
685 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
686 !check_valid_pointer(s, page, nextfree)) {
687 object_err(s, page, freelist, "Freechain corrupt");
689 slab_fix(s, "Isolate corrupted freechain");
696 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
698 unsigned int off; /* Offset of last byte */
699 u8 *addr = page_address(page);
701 print_tracking(s, p);
703 print_page_info(page);
705 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
706 p, p - addr, get_freepointer(s, p));
708 if (s->flags & SLAB_RED_ZONE)
709 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
711 else if (p > addr + 16)
712 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
714 print_section(KERN_ERR, "Object ", p,
715 min_t(unsigned int, s->object_size, PAGE_SIZE));
716 if (s->flags & SLAB_RED_ZONE)
717 print_section(KERN_ERR, "Redzone ", p + s->object_size,
718 s->inuse - s->object_size);
720 off = get_info_end(s);
722 if (s->flags & SLAB_STORE_USER)
723 off += 2 * sizeof(struct track);
725 off += kasan_metadata_size(s);
727 if (off != size_from_object(s))
728 /* Beginning of the filler is the free pointer */
729 print_section(KERN_ERR, "Padding ", p + off,
730 size_from_object(s) - off);
735 void object_err(struct kmem_cache *s, struct page *page,
736 u8 *object, char *reason)
738 slab_bug(s, "%s", reason);
739 print_trailer(s, page, object);
742 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
743 const char *fmt, ...)
749 vsnprintf(buf, sizeof(buf), fmt, args);
751 slab_bug(s, "%s", buf);
752 print_page_info(page);
756 static void init_object(struct kmem_cache *s, void *object, u8 val)
760 if (s->flags & SLAB_RED_ZONE)
761 memset(p - s->red_left_pad, val, s->red_left_pad);
763 if (s->flags & __OBJECT_POISON) {
764 memset(p, POISON_FREE, s->object_size - 1);
765 p[s->object_size - 1] = POISON_END;
768 if (s->flags & SLAB_RED_ZONE)
769 memset(p + s->object_size, val, s->inuse - s->object_size);
772 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
773 void *from, void *to)
775 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
776 memset(from, data, to - from);
779 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
780 u8 *object, char *what,
781 u8 *start, unsigned int value, unsigned int bytes)
785 u8 *addr = page_address(page);
787 metadata_access_enable();
788 fault = memchr_inv(start, value, bytes);
789 metadata_access_disable();
794 while (end > fault && end[-1] == value)
797 slab_bug(s, "%s overwritten", what);
798 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
799 fault, end - 1, fault - addr,
801 print_trailer(s, page, object);
803 restore_bytes(s, what, value, fault, end);
811 * Bytes of the object to be managed.
812 * If the freepointer may overlay the object then the free
813 * pointer is at the middle of the object.
815 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
818 * object + s->object_size
819 * Padding to reach word boundary. This is also used for Redzoning.
820 * Padding is extended by another word if Redzoning is enabled and
821 * object_size == inuse.
823 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
824 * 0xcc (RED_ACTIVE) for objects in use.
827 * Meta data starts here.
829 * A. Free pointer (if we cannot overwrite object on free)
830 * B. Tracking data for SLAB_STORE_USER
831 * C. Padding to reach required alignment boundary or at mininum
832 * one word if debugging is on to be able to detect writes
833 * before the word boundary.
835 * Padding is done using 0x5a (POISON_INUSE)
838 * Nothing is used beyond s->size.
840 * If slabcaches are merged then the object_size and inuse boundaries are mostly
841 * ignored. And therefore no slab options that rely on these boundaries
842 * may be used with merged slabcaches.
845 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
847 unsigned long off = get_info_end(s); /* The end of info */
849 if (s->flags & SLAB_STORE_USER)
850 /* We also have user information there */
851 off += 2 * sizeof(struct track);
853 off += kasan_metadata_size(s);
855 if (size_from_object(s) == off)
858 return check_bytes_and_report(s, page, p, "Object padding",
859 p + off, POISON_INUSE, size_from_object(s) - off);
862 /* Check the pad bytes at the end of a slab page */
863 static int slab_pad_check(struct kmem_cache *s, struct page *page)
872 if (!(s->flags & SLAB_POISON))
875 start = page_address(page);
876 length = page_size(page);
877 end = start + length;
878 remainder = length % s->size;
882 pad = end - remainder;
883 metadata_access_enable();
884 fault = memchr_inv(pad, POISON_INUSE, remainder);
885 metadata_access_disable();
888 while (end > fault && end[-1] == POISON_INUSE)
891 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
892 fault, end - 1, fault - start);
893 print_section(KERN_ERR, "Padding ", pad, remainder);
895 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
899 static int check_object(struct kmem_cache *s, struct page *page,
900 void *object, u8 val)
903 u8 *endobject = object + s->object_size;
905 if (s->flags & SLAB_RED_ZONE) {
906 if (!check_bytes_and_report(s, page, object, "Redzone",
907 object - s->red_left_pad, val, s->red_left_pad))
910 if (!check_bytes_and_report(s, page, object, "Redzone",
911 endobject, val, s->inuse - s->object_size))
914 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
915 check_bytes_and_report(s, page, p, "Alignment padding",
916 endobject, POISON_INUSE,
917 s->inuse - s->object_size);
921 if (s->flags & SLAB_POISON) {
922 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
923 (!check_bytes_and_report(s, page, p, "Poison", p,
924 POISON_FREE, s->object_size - 1) ||
925 !check_bytes_and_report(s, page, p, "Poison",
926 p + s->object_size - 1, POISON_END, 1)))
929 * check_pad_bytes cleans up on its own.
931 check_pad_bytes(s, page, p);
934 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
936 * Object and freepointer overlap. Cannot check
937 * freepointer while object is allocated.
941 /* Check free pointer validity */
942 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
943 object_err(s, page, p, "Freepointer corrupt");
945 * No choice but to zap it and thus lose the remainder
946 * of the free objects in this slab. May cause
947 * another error because the object count is now wrong.
949 set_freepointer(s, p, NULL);
955 static int check_slab(struct kmem_cache *s, struct page *page)
959 VM_BUG_ON(!irqs_disabled());
961 if (!PageSlab(page)) {
962 slab_err(s, page, "Not a valid slab page");
966 maxobj = order_objects(compound_order(page), s->size);
967 if (page->objects > maxobj) {
968 slab_err(s, page, "objects %u > max %u",
969 page->objects, maxobj);
972 if (page->inuse > page->objects) {
973 slab_err(s, page, "inuse %u > max %u",
974 page->inuse, page->objects);
977 /* Slab_pad_check fixes things up after itself */
978 slab_pad_check(s, page);
983 * Determine if a certain object on a page is on the freelist. Must hold the
984 * slab lock to guarantee that the chains are in a consistent state.
986 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
994 while (fp && nr <= page->objects) {
997 if (!check_valid_pointer(s, page, fp)) {
999 object_err(s, page, object,
1000 "Freechain corrupt");
1001 set_freepointer(s, object, NULL);
1003 slab_err(s, page, "Freepointer corrupt");
1004 page->freelist = NULL;
1005 page->inuse = page->objects;
1006 slab_fix(s, "Freelist cleared");
1012 fp = get_freepointer(s, object);
1016 max_objects = order_objects(compound_order(page), s->size);
1017 if (max_objects > MAX_OBJS_PER_PAGE)
1018 max_objects = MAX_OBJS_PER_PAGE;
1020 if (page->objects != max_objects) {
1021 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1022 page->objects, max_objects);
1023 page->objects = max_objects;
1024 slab_fix(s, "Number of objects adjusted.");
1026 if (page->inuse != page->objects - nr) {
1027 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1028 page->inuse, page->objects - nr);
1029 page->inuse = page->objects - nr;
1030 slab_fix(s, "Object count adjusted.");
1032 return search == NULL;
1035 static void trace(struct kmem_cache *s, struct page *page, void *object,
1038 if (s->flags & SLAB_TRACE) {
1039 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1041 alloc ? "alloc" : "free",
1042 object, page->inuse,
1046 print_section(KERN_INFO, "Object ", (void *)object,
1054 * Tracking of fully allocated slabs for debugging purposes.
1056 static void add_full(struct kmem_cache *s,
1057 struct kmem_cache_node *n, struct page *page)
1059 if (!(s->flags & SLAB_STORE_USER))
1062 lockdep_assert_held(&n->list_lock);
1063 list_add(&page->slab_list, &n->full);
1066 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1068 if (!(s->flags & SLAB_STORE_USER))
1071 lockdep_assert_held(&n->list_lock);
1072 list_del(&page->slab_list);
1075 /* Tracking of the number of slabs for debugging purposes */
1076 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1078 struct kmem_cache_node *n = get_node(s, node);
1080 return atomic_long_read(&n->nr_slabs);
1083 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1085 return atomic_long_read(&n->nr_slabs);
1088 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1090 struct kmem_cache_node *n = get_node(s, node);
1093 * May be called early in order to allocate a slab for the
1094 * kmem_cache_node structure. Solve the chicken-egg
1095 * dilemma by deferring the increment of the count during
1096 * bootstrap (see early_kmem_cache_node_alloc).
1099 atomic_long_inc(&n->nr_slabs);
1100 atomic_long_add(objects, &n->total_objects);
1103 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1105 struct kmem_cache_node *n = get_node(s, node);
1107 atomic_long_dec(&n->nr_slabs);
1108 atomic_long_sub(objects, &n->total_objects);
1111 /* Object debug checks for alloc/free paths */
1112 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1115 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1118 init_object(s, object, SLUB_RED_INACTIVE);
1119 init_tracking(s, object);
1123 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1125 if (!(s->flags & SLAB_POISON))
1128 metadata_access_enable();
1129 memset(addr, POISON_INUSE, page_size(page));
1130 metadata_access_disable();
1133 static inline int alloc_consistency_checks(struct kmem_cache *s,
1134 struct page *page, void *object)
1136 if (!check_slab(s, page))
1139 if (!check_valid_pointer(s, page, object)) {
1140 object_err(s, page, object, "Freelist Pointer check fails");
1144 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1150 static noinline int alloc_debug_processing(struct kmem_cache *s,
1152 void *object, unsigned long addr)
1154 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1155 if (!alloc_consistency_checks(s, page, object))
1159 /* Success perform special debug activities for allocs */
1160 if (s->flags & SLAB_STORE_USER)
1161 set_track(s, object, TRACK_ALLOC, addr);
1162 trace(s, page, object, 1);
1163 init_object(s, object, SLUB_RED_ACTIVE);
1167 if (PageSlab(page)) {
1169 * If this is a slab page then lets do the best we can
1170 * to avoid issues in the future. Marking all objects
1171 * as used avoids touching the remaining objects.
1173 slab_fix(s, "Marking all objects used");
1174 page->inuse = page->objects;
1175 page->freelist = NULL;
1180 static inline int free_consistency_checks(struct kmem_cache *s,
1181 struct page *page, void *object, unsigned long addr)
1183 if (!check_valid_pointer(s, page, object)) {
1184 slab_err(s, page, "Invalid object pointer 0x%p", object);
1188 if (on_freelist(s, page, object)) {
1189 object_err(s, page, object, "Object already free");
1193 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1196 if (unlikely(s != page->slab_cache)) {
1197 if (!PageSlab(page)) {
1198 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1200 } else if (!page->slab_cache) {
1201 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1205 object_err(s, page, object,
1206 "page slab pointer corrupt.");
1212 /* Supports checking bulk free of a constructed freelist */
1213 static noinline int free_debug_processing(
1214 struct kmem_cache *s, struct page *page,
1215 void *head, void *tail, int bulk_cnt,
1218 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1219 void *object = head;
1221 unsigned long uninitialized_var(flags);
1224 spin_lock_irqsave(&n->list_lock, flags);
1227 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1228 if (!check_slab(s, page))
1235 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1236 if (!free_consistency_checks(s, page, object, addr))
1240 if (s->flags & SLAB_STORE_USER)
1241 set_track(s, object, TRACK_FREE, addr);
1242 trace(s, page, object, 0);
1243 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1244 init_object(s, object, SLUB_RED_INACTIVE);
1246 /* Reached end of constructed freelist yet? */
1247 if (object != tail) {
1248 object = get_freepointer(s, object);
1254 if (cnt != bulk_cnt)
1255 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1259 spin_unlock_irqrestore(&n->list_lock, flags);
1261 slab_fix(s, "Object at 0x%p not freed", object);
1265 static int __init setup_slub_debug(char *str)
1267 slub_debug = DEBUG_DEFAULT_FLAGS;
1268 if (*str++ != '=' || !*str)
1270 * No options specified. Switch on full debugging.
1276 * No options but restriction on slabs. This means full
1277 * debugging for slabs matching a pattern.
1284 * Switch off all debugging measures.
1289 * Determine which debug features should be switched on
1291 for (; *str && *str != ','; str++) {
1292 switch (tolower(*str)) {
1294 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1297 slub_debug |= SLAB_RED_ZONE;
1300 slub_debug |= SLAB_POISON;
1303 slub_debug |= SLAB_STORE_USER;
1306 slub_debug |= SLAB_TRACE;
1309 slub_debug |= SLAB_FAILSLAB;
1313 * Avoid enabling debugging on caches if its minimum
1314 * order would increase as a result.
1316 disable_higher_order_debug = 1;
1319 pr_err("slub_debug option '%c' unknown. skipped\n",
1326 slub_debug_slabs = str + 1;
1328 if ((static_branch_unlikely(&init_on_alloc) ||
1329 static_branch_unlikely(&init_on_free)) &&
1330 (slub_debug & SLAB_POISON))
1331 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1335 __setup("slub_debug", setup_slub_debug);
1338 * kmem_cache_flags - apply debugging options to the cache
1339 * @object_size: the size of an object without meta data
1340 * @flags: flags to set
1341 * @name: name of the cache
1342 * @ctor: constructor function
1344 * Debug option(s) are applied to @flags. In addition to the debug
1345 * option(s), if a slab name (or multiple) is specified i.e.
1346 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1347 * then only the select slabs will receive the debug option(s).
1349 slab_flags_t kmem_cache_flags(unsigned int object_size,
1350 slab_flags_t flags, const char *name,
1351 void (*ctor)(void *))
1356 /* If slub_debug = 0, it folds into the if conditional. */
1357 if (!slub_debug_slabs)
1358 return flags | slub_debug;
1361 iter = slub_debug_slabs;
1366 end = strchrnul(iter, ',');
1368 glob = strnchr(iter, end - iter, '*');
1370 cmplen = glob - iter;
1372 cmplen = max_t(size_t, len, (end - iter));
1374 if (!strncmp(name, iter, cmplen)) {
1375 flags |= slub_debug;
1386 #else /* !CONFIG_SLUB_DEBUG */
1387 static inline void setup_object_debug(struct kmem_cache *s,
1388 struct page *page, void *object) {}
1390 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1392 static inline int alloc_debug_processing(struct kmem_cache *s,
1393 struct page *page, void *object, unsigned long addr) { return 0; }
1395 static inline int free_debug_processing(
1396 struct kmem_cache *s, struct page *page,
1397 void *head, void *tail, int bulk_cnt,
1398 unsigned long addr) { return 0; }
1400 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1402 static inline int check_object(struct kmem_cache *s, struct page *page,
1403 void *object, u8 val) { return 1; }
1404 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1405 struct page *page) {}
1406 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1407 struct page *page) {}
1408 slab_flags_t kmem_cache_flags(unsigned int object_size,
1409 slab_flags_t flags, const char *name,
1410 void (*ctor)(void *))
1414 #define slub_debug 0
1416 #define disable_higher_order_debug 0
1418 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1420 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1422 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1424 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1427 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1428 void *freelist, void *nextfree)
1432 #endif /* CONFIG_SLUB_DEBUG */
1435 * Hooks for other subsystems that check memory allocations. In a typical
1436 * production configuration these hooks all should produce no code at all.
1438 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1440 ptr = kasan_kmalloc_large(ptr, size, flags);
1441 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1442 kmemleak_alloc(ptr, size, 1, flags);
1446 static __always_inline void kfree_hook(void *x)
1449 kasan_kfree_large(x, _RET_IP_);
1452 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1454 kmemleak_free_recursive(x, s->flags);
1457 * Trouble is that we may no longer disable interrupts in the fast path
1458 * So in order to make the debug calls that expect irqs to be
1459 * disabled we need to disable interrupts temporarily.
1461 #ifdef CONFIG_LOCKDEP
1463 unsigned long flags;
1465 local_irq_save(flags);
1466 debug_check_no_locks_freed(x, s->object_size);
1467 local_irq_restore(flags);
1470 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1471 debug_check_no_obj_freed(x, s->object_size);
1473 /* KASAN might put x into memory quarantine, delaying its reuse */
1474 return kasan_slab_free(s, x, _RET_IP_);
1477 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1478 void **head, void **tail)
1483 void *old_tail = *tail ? *tail : *head;
1486 /* Head and tail of the reconstructed freelist */
1492 next = get_freepointer(s, object);
1494 if (slab_want_init_on_free(s)) {
1496 * Clear the object and the metadata, but don't touch
1499 memset(object, 0, s->object_size);
1500 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1502 memset((char *)object + s->inuse, 0,
1503 s->size - s->inuse - rsize);
1506 /* If object's reuse doesn't have to be delayed */
1507 if (!slab_free_hook(s, object)) {
1508 /* Move object to the new freelist */
1509 set_freepointer(s, object, *head);
1514 } while (object != old_tail);
1519 return *head != NULL;
1522 static void *setup_object(struct kmem_cache *s, struct page *page,
1525 setup_object_debug(s, page, object);
1526 object = kasan_init_slab_obj(s, object);
1527 if (unlikely(s->ctor)) {
1528 kasan_unpoison_object_data(s, object);
1530 kasan_poison_object_data(s, object);
1536 * Slab allocation and freeing
1538 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1539 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1542 unsigned int order = oo_order(oo);
1544 if (node == NUMA_NO_NODE)
1545 page = alloc_pages(flags, order);
1547 page = __alloc_pages_node(node, flags, order);
1549 if (page && charge_slab_page(page, flags, order, s)) {
1550 __free_pages(page, order);
1557 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1558 /* Pre-initialize the random sequence cache */
1559 static int init_cache_random_seq(struct kmem_cache *s)
1561 unsigned int count = oo_objects(s->oo);
1564 /* Bailout if already initialised */
1568 err = cache_random_seq_create(s, count, GFP_KERNEL);
1570 pr_err("SLUB: Unable to initialize free list for %s\n",
1575 /* Transform to an offset on the set of pages */
1576 if (s->random_seq) {
1579 for (i = 0; i < count; i++)
1580 s->random_seq[i] *= s->size;
1585 /* Initialize each random sequence freelist per cache */
1586 static void __init init_freelist_randomization(void)
1588 struct kmem_cache *s;
1590 mutex_lock(&slab_mutex);
1592 list_for_each_entry(s, &slab_caches, list)
1593 init_cache_random_seq(s);
1595 mutex_unlock(&slab_mutex);
1598 /* Get the next entry on the pre-computed freelist randomized */
1599 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1600 unsigned long *pos, void *start,
1601 unsigned long page_limit,
1602 unsigned long freelist_count)
1607 * If the target page allocation failed, the number of objects on the
1608 * page might be smaller than the usual size defined by the cache.
1611 idx = s->random_seq[*pos];
1613 if (*pos >= freelist_count)
1615 } while (unlikely(idx >= page_limit));
1617 return (char *)start + idx;
1620 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1621 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1626 unsigned long idx, pos, page_limit, freelist_count;
1628 if (page->objects < 2 || !s->random_seq)
1631 freelist_count = oo_objects(s->oo);
1632 pos = get_random_int() % freelist_count;
1634 page_limit = page->objects * s->size;
1635 start = fixup_red_left(s, page_address(page));
1637 /* First entry is used as the base of the freelist */
1638 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1640 cur = setup_object(s, page, cur);
1641 page->freelist = cur;
1643 for (idx = 1; idx < page->objects; idx++) {
1644 next = next_freelist_entry(s, page, &pos, start, page_limit,
1646 next = setup_object(s, page, next);
1647 set_freepointer(s, cur, next);
1650 set_freepointer(s, cur, NULL);
1655 static inline int init_cache_random_seq(struct kmem_cache *s)
1659 static inline void init_freelist_randomization(void) { }
1660 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1664 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1666 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1669 struct kmem_cache_order_objects oo = s->oo;
1671 void *start, *p, *next;
1675 flags &= gfp_allowed_mask;
1677 if (gfpflags_allow_blocking(flags))
1680 flags |= s->allocflags;
1683 * Let the initial higher-order allocation fail under memory pressure
1684 * so we fall-back to the minimum order allocation.
1686 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1687 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1688 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1690 page = alloc_slab_page(s, alloc_gfp, node, oo);
1691 if (unlikely(!page)) {
1695 * Allocation may have failed due to fragmentation.
1696 * Try a lower order alloc if possible
1698 page = alloc_slab_page(s, alloc_gfp, node, oo);
1699 if (unlikely(!page))
1701 stat(s, ORDER_FALLBACK);
1704 page->objects = oo_objects(oo);
1706 page->slab_cache = s;
1707 __SetPageSlab(page);
1708 if (page_is_pfmemalloc(page))
1709 SetPageSlabPfmemalloc(page);
1711 kasan_poison_slab(page);
1713 start = page_address(page);
1715 setup_page_debug(s, page, start);
1717 shuffle = shuffle_freelist(s, page);
1720 start = fixup_red_left(s, start);
1721 start = setup_object(s, page, start);
1722 page->freelist = start;
1723 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1725 next = setup_object(s, page, next);
1726 set_freepointer(s, p, next);
1729 set_freepointer(s, p, NULL);
1732 page->inuse = page->objects;
1736 if (gfpflags_allow_blocking(flags))
1737 local_irq_disable();
1741 inc_slabs_node(s, page_to_nid(page), page->objects);
1746 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1748 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1749 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1750 flags &= ~GFP_SLAB_BUG_MASK;
1751 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1752 invalid_mask, &invalid_mask, flags, &flags);
1756 return allocate_slab(s,
1757 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1760 static void __free_slab(struct kmem_cache *s, struct page *page)
1762 int order = compound_order(page);
1763 int pages = 1 << order;
1765 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1768 slab_pad_check(s, page);
1769 for_each_object(p, s, page_address(page),
1771 check_object(s, page, p, SLUB_RED_INACTIVE);
1774 __ClearPageSlabPfmemalloc(page);
1775 __ClearPageSlab(page);
1777 page->mapping = NULL;
1778 if (current->reclaim_state)
1779 current->reclaim_state->reclaimed_slab += pages;
1780 uncharge_slab_page(page, order, s);
1781 __free_pages(page, order);
1784 static void rcu_free_slab(struct rcu_head *h)
1786 struct page *page = container_of(h, struct page, rcu_head);
1788 __free_slab(page->slab_cache, page);
1791 static void free_slab(struct kmem_cache *s, struct page *page)
1793 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1794 call_rcu(&page->rcu_head, rcu_free_slab);
1796 __free_slab(s, page);
1799 static void discard_slab(struct kmem_cache *s, struct page *page)
1801 dec_slabs_node(s, page_to_nid(page), page->objects);
1806 * Management of partially allocated slabs.
1809 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1812 if (tail == DEACTIVATE_TO_TAIL)
1813 list_add_tail(&page->slab_list, &n->partial);
1815 list_add(&page->slab_list, &n->partial);
1818 static inline void add_partial(struct kmem_cache_node *n,
1819 struct page *page, int tail)
1821 lockdep_assert_held(&n->list_lock);
1822 __add_partial(n, page, tail);
1825 static inline void remove_partial(struct kmem_cache_node *n,
1828 lockdep_assert_held(&n->list_lock);
1829 list_del(&page->slab_list);
1834 * Remove slab from the partial list, freeze it and
1835 * return the pointer to the freelist.
1837 * Returns a list of objects or NULL if it fails.
1839 static inline void *acquire_slab(struct kmem_cache *s,
1840 struct kmem_cache_node *n, struct page *page,
1841 int mode, int *objects)
1844 unsigned long counters;
1847 lockdep_assert_held(&n->list_lock);
1850 * Zap the freelist and set the frozen bit.
1851 * The old freelist is the list of objects for the
1852 * per cpu allocation list.
1854 freelist = page->freelist;
1855 counters = page->counters;
1856 new.counters = counters;
1857 *objects = new.objects - new.inuse;
1859 new.inuse = page->objects;
1860 new.freelist = NULL;
1862 new.freelist = freelist;
1865 VM_BUG_ON(new.frozen);
1868 if (!__cmpxchg_double_slab(s, page,
1870 new.freelist, new.counters,
1874 remove_partial(n, page);
1879 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1880 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1883 * Try to allocate a partial slab from a specific node.
1885 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1886 struct kmem_cache_cpu *c, gfp_t flags)
1888 struct page *page, *page2;
1889 void *object = NULL;
1890 unsigned int available = 0;
1894 * Racy check. If we mistakenly see no partial slabs then we
1895 * just allocate an empty slab. If we mistakenly try to get a
1896 * partial slab and there is none available then get_partials()
1899 if (!n || !n->nr_partial)
1902 spin_lock(&n->list_lock);
1903 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1906 if (!pfmemalloc_match(page, flags))
1909 t = acquire_slab(s, n, page, object == NULL, &objects);
1913 available += objects;
1916 stat(s, ALLOC_FROM_PARTIAL);
1919 put_cpu_partial(s, page, 0);
1920 stat(s, CPU_PARTIAL_NODE);
1922 if (!kmem_cache_has_cpu_partial(s)
1923 || available > slub_cpu_partial(s) / 2)
1927 spin_unlock(&n->list_lock);
1932 * Get a page from somewhere. Search in increasing NUMA distances.
1934 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1935 struct kmem_cache_cpu *c)
1938 struct zonelist *zonelist;
1941 enum zone_type highest_zoneidx = gfp_zone(flags);
1943 unsigned int cpuset_mems_cookie;
1946 * The defrag ratio allows a configuration of the tradeoffs between
1947 * inter node defragmentation and node local allocations. A lower
1948 * defrag_ratio increases the tendency to do local allocations
1949 * instead of attempting to obtain partial slabs from other nodes.
1951 * If the defrag_ratio is set to 0 then kmalloc() always
1952 * returns node local objects. If the ratio is higher then kmalloc()
1953 * may return off node objects because partial slabs are obtained
1954 * from other nodes and filled up.
1956 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1957 * (which makes defrag_ratio = 1000) then every (well almost)
1958 * allocation will first attempt to defrag slab caches on other nodes.
1959 * This means scanning over all nodes to look for partial slabs which
1960 * may be expensive if we do it every time we are trying to find a slab
1961 * with available objects.
1963 if (!s->remote_node_defrag_ratio ||
1964 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1968 cpuset_mems_cookie = read_mems_allowed_begin();
1969 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1970 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
1971 struct kmem_cache_node *n;
1973 n = get_node(s, zone_to_nid(zone));
1975 if (n && cpuset_zone_allowed(zone, flags) &&
1976 n->nr_partial > s->min_partial) {
1977 object = get_partial_node(s, n, c, flags);
1980 * Don't check read_mems_allowed_retry()
1981 * here - if mems_allowed was updated in
1982 * parallel, that was a harmless race
1983 * between allocation and the cpuset
1990 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1991 #endif /* CONFIG_NUMA */
1996 * Get a partial page, lock it and return it.
1998 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1999 struct kmem_cache_cpu *c)
2002 int searchnode = node;
2004 if (node == NUMA_NO_NODE)
2005 searchnode = numa_mem_id();
2007 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2008 if (object || node != NUMA_NO_NODE)
2011 return get_any_partial(s, flags, c);
2014 #ifdef CONFIG_PREEMPTION
2016 * Calculate the next globally unique transaction for disambiguation
2017 * during cmpxchg. The transactions start with the cpu number and are then
2018 * incremented by CONFIG_NR_CPUS.
2020 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2023 * No preemption supported therefore also no need to check for
2029 static inline unsigned long next_tid(unsigned long tid)
2031 return tid + TID_STEP;
2034 #ifdef SLUB_DEBUG_CMPXCHG
2035 static inline unsigned int tid_to_cpu(unsigned long tid)
2037 return tid % TID_STEP;
2040 static inline unsigned long tid_to_event(unsigned long tid)
2042 return tid / TID_STEP;
2046 static inline unsigned int init_tid(int cpu)
2051 static inline void note_cmpxchg_failure(const char *n,
2052 const struct kmem_cache *s, unsigned long tid)
2054 #ifdef SLUB_DEBUG_CMPXCHG
2055 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2057 pr_info("%s %s: cmpxchg redo ", n, s->name);
2059 #ifdef CONFIG_PREEMPTION
2060 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2061 pr_warn("due to cpu change %d -> %d\n",
2062 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2065 if (tid_to_event(tid) != tid_to_event(actual_tid))
2066 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2067 tid_to_event(tid), tid_to_event(actual_tid));
2069 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2070 actual_tid, tid, next_tid(tid));
2072 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2075 static void init_kmem_cache_cpus(struct kmem_cache *s)
2079 for_each_possible_cpu(cpu)
2080 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2084 * Remove the cpu slab
2086 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2087 void *freelist, struct kmem_cache_cpu *c)
2089 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2090 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2092 enum slab_modes l = M_NONE, m = M_NONE;
2094 int tail = DEACTIVATE_TO_HEAD;
2098 if (page->freelist) {
2099 stat(s, DEACTIVATE_REMOTE_FREES);
2100 tail = DEACTIVATE_TO_TAIL;
2104 * Stage one: Free all available per cpu objects back
2105 * to the page freelist while it is still frozen. Leave the
2108 * There is no need to take the list->lock because the page
2111 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2113 unsigned long counters;
2116 * If 'nextfree' is invalid, it is possible that the object at
2117 * 'freelist' is already corrupted. So isolate all objects
2118 * starting at 'freelist'.
2120 if (freelist_corrupted(s, page, freelist, nextfree))
2124 prior = page->freelist;
2125 counters = page->counters;
2126 set_freepointer(s, freelist, prior);
2127 new.counters = counters;
2129 VM_BUG_ON(!new.frozen);
2131 } while (!__cmpxchg_double_slab(s, page,
2133 freelist, new.counters,
2134 "drain percpu freelist"));
2136 freelist = nextfree;
2140 * Stage two: Ensure that the page is unfrozen while the
2141 * list presence reflects the actual number of objects
2144 * We setup the list membership and then perform a cmpxchg
2145 * with the count. If there is a mismatch then the page
2146 * is not unfrozen but the page is on the wrong list.
2148 * Then we restart the process which may have to remove
2149 * the page from the list that we just put it on again
2150 * because the number of objects in the slab may have
2155 old.freelist = page->freelist;
2156 old.counters = page->counters;
2157 VM_BUG_ON(!old.frozen);
2159 /* Determine target state of the slab */
2160 new.counters = old.counters;
2163 set_freepointer(s, freelist, old.freelist);
2164 new.freelist = freelist;
2166 new.freelist = old.freelist;
2170 if (!new.inuse && n->nr_partial >= s->min_partial)
2172 else if (new.freelist) {
2177 * Taking the spinlock removes the possibility
2178 * that acquire_slab() will see a slab page that
2181 spin_lock(&n->list_lock);
2185 if (kmem_cache_debug(s) && !lock) {
2188 * This also ensures that the scanning of full
2189 * slabs from diagnostic functions will not see
2192 spin_lock(&n->list_lock);
2198 remove_partial(n, page);
2199 else if (l == M_FULL)
2200 remove_full(s, n, page);
2203 add_partial(n, page, tail);
2204 else if (m == M_FULL)
2205 add_full(s, n, page);
2209 if (!__cmpxchg_double_slab(s, page,
2210 old.freelist, old.counters,
2211 new.freelist, new.counters,
2216 spin_unlock(&n->list_lock);
2220 else if (m == M_FULL)
2221 stat(s, DEACTIVATE_FULL);
2222 else if (m == M_FREE) {
2223 stat(s, DEACTIVATE_EMPTY);
2224 discard_slab(s, page);
2233 * Unfreeze all the cpu partial slabs.
2235 * This function must be called with interrupts disabled
2236 * for the cpu using c (or some other guarantee must be there
2237 * to guarantee no concurrent accesses).
2239 static void unfreeze_partials(struct kmem_cache *s,
2240 struct kmem_cache_cpu *c)
2242 #ifdef CONFIG_SLUB_CPU_PARTIAL
2243 struct kmem_cache_node *n = NULL, *n2 = NULL;
2244 struct page *page, *discard_page = NULL;
2246 while ((page = slub_percpu_partial(c))) {
2250 slub_set_percpu_partial(c, page);
2252 n2 = get_node(s, page_to_nid(page));
2255 spin_unlock(&n->list_lock);
2258 spin_lock(&n->list_lock);
2263 old.freelist = page->freelist;
2264 old.counters = page->counters;
2265 VM_BUG_ON(!old.frozen);
2267 new.counters = old.counters;
2268 new.freelist = old.freelist;
2272 } while (!__cmpxchg_double_slab(s, page,
2273 old.freelist, old.counters,
2274 new.freelist, new.counters,
2275 "unfreezing slab"));
2277 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2278 page->next = discard_page;
2279 discard_page = page;
2281 add_partial(n, page, DEACTIVATE_TO_TAIL);
2282 stat(s, FREE_ADD_PARTIAL);
2287 spin_unlock(&n->list_lock);
2289 while (discard_page) {
2290 page = discard_page;
2291 discard_page = discard_page->next;
2293 stat(s, DEACTIVATE_EMPTY);
2294 discard_slab(s, page);
2297 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2301 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2302 * partial page slot if available.
2304 * If we did not find a slot then simply move all the partials to the
2305 * per node partial list.
2307 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2309 #ifdef CONFIG_SLUB_CPU_PARTIAL
2310 struct page *oldpage;
2318 oldpage = this_cpu_read(s->cpu_slab->partial);
2321 pobjects = oldpage->pobjects;
2322 pages = oldpage->pages;
2323 if (drain && pobjects > slub_cpu_partial(s)) {
2324 unsigned long flags;
2326 * partial array is full. Move the existing
2327 * set to the per node partial list.
2329 local_irq_save(flags);
2330 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2331 local_irq_restore(flags);
2335 stat(s, CPU_PARTIAL_DRAIN);
2340 pobjects += page->objects - page->inuse;
2342 page->pages = pages;
2343 page->pobjects = pobjects;
2344 page->next = oldpage;
2346 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2348 if (unlikely(!slub_cpu_partial(s))) {
2349 unsigned long flags;
2351 local_irq_save(flags);
2352 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2353 local_irq_restore(flags);
2356 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2359 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2361 stat(s, CPUSLAB_FLUSH);
2362 deactivate_slab(s, c->page, c->freelist, c);
2364 c->tid = next_tid(c->tid);
2370 * Called from IPI handler with interrupts disabled.
2372 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2374 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2379 unfreeze_partials(s, c);
2382 static void flush_cpu_slab(void *d)
2384 struct kmem_cache *s = d;
2386 __flush_cpu_slab(s, smp_processor_id());
2389 static bool has_cpu_slab(int cpu, void *info)
2391 struct kmem_cache *s = info;
2392 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2394 return c->page || slub_percpu_partial(c);
2397 static void flush_all(struct kmem_cache *s)
2399 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2403 * Use the cpu notifier to insure that the cpu slabs are flushed when
2406 static int slub_cpu_dead(unsigned int cpu)
2408 struct kmem_cache *s;
2409 unsigned long flags;
2411 mutex_lock(&slab_mutex);
2412 list_for_each_entry(s, &slab_caches, list) {
2413 local_irq_save(flags);
2414 __flush_cpu_slab(s, cpu);
2415 local_irq_restore(flags);
2417 mutex_unlock(&slab_mutex);
2422 * Check if the objects in a per cpu structure fit numa
2423 * locality expectations.
2425 static inline int node_match(struct page *page, int node)
2428 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2434 #ifdef CONFIG_SLUB_DEBUG
2435 static int count_free(struct page *page)
2437 return page->objects - page->inuse;
2440 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2442 return atomic_long_read(&n->total_objects);
2444 #endif /* CONFIG_SLUB_DEBUG */
2446 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2447 static unsigned long count_partial(struct kmem_cache_node *n,
2448 int (*get_count)(struct page *))
2450 unsigned long flags;
2451 unsigned long x = 0;
2454 spin_lock_irqsave(&n->list_lock, flags);
2455 list_for_each_entry(page, &n->partial, slab_list)
2456 x += get_count(page);
2457 spin_unlock_irqrestore(&n->list_lock, flags);
2460 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2462 static noinline void
2463 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2465 #ifdef CONFIG_SLUB_DEBUG
2466 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2467 DEFAULT_RATELIMIT_BURST);
2469 struct kmem_cache_node *n;
2471 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2474 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2475 nid, gfpflags, &gfpflags);
2476 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2477 s->name, s->object_size, s->size, oo_order(s->oo),
2480 if (oo_order(s->min) > get_order(s->object_size))
2481 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2484 for_each_kmem_cache_node(s, node, n) {
2485 unsigned long nr_slabs;
2486 unsigned long nr_objs;
2487 unsigned long nr_free;
2489 nr_free = count_partial(n, count_free);
2490 nr_slabs = node_nr_slabs(n);
2491 nr_objs = node_nr_objs(n);
2493 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2494 node, nr_slabs, nr_objs, nr_free);
2499 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2500 int node, struct kmem_cache_cpu **pc)
2503 struct kmem_cache_cpu *c = *pc;
2506 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2508 freelist = get_partial(s, flags, node, c);
2513 page = new_slab(s, flags, node);
2515 c = raw_cpu_ptr(s->cpu_slab);
2520 * No other reference to the page yet so we can
2521 * muck around with it freely without cmpxchg
2523 freelist = page->freelist;
2524 page->freelist = NULL;
2526 stat(s, ALLOC_SLAB);
2534 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2536 if (unlikely(PageSlabPfmemalloc(page)))
2537 return gfp_pfmemalloc_allowed(gfpflags);
2543 * Check the page->freelist of a page and either transfer the freelist to the
2544 * per cpu freelist or deactivate the page.
2546 * The page is still frozen if the return value is not NULL.
2548 * If this function returns NULL then the page has been unfrozen.
2550 * This function must be called with interrupt disabled.
2552 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2555 unsigned long counters;
2559 freelist = page->freelist;
2560 counters = page->counters;
2562 new.counters = counters;
2563 VM_BUG_ON(!new.frozen);
2565 new.inuse = page->objects;
2566 new.frozen = freelist != NULL;
2568 } while (!__cmpxchg_double_slab(s, page,
2577 * Slow path. The lockless freelist is empty or we need to perform
2580 * Processing is still very fast if new objects have been freed to the
2581 * regular freelist. In that case we simply take over the regular freelist
2582 * as the lockless freelist and zap the regular freelist.
2584 * If that is not working then we fall back to the partial lists. We take the
2585 * first element of the freelist as the object to allocate now and move the
2586 * rest of the freelist to the lockless freelist.
2588 * And if we were unable to get a new slab from the partial slab lists then
2589 * we need to allocate a new slab. This is the slowest path since it involves
2590 * a call to the page allocator and the setup of a new slab.
2592 * Version of __slab_alloc to use when we know that interrupts are
2593 * already disabled (which is the case for bulk allocation).
2595 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2596 unsigned long addr, struct kmem_cache_cpu *c)
2604 * if the node is not online or has no normal memory, just
2605 * ignore the node constraint
2607 if (unlikely(node != NUMA_NO_NODE &&
2608 !node_state(node, N_NORMAL_MEMORY)))
2609 node = NUMA_NO_NODE;
2614 if (unlikely(!node_match(page, node))) {
2616 * same as above but node_match() being false already
2617 * implies node != NUMA_NO_NODE
2619 if (!node_state(node, N_NORMAL_MEMORY)) {
2620 node = NUMA_NO_NODE;
2623 stat(s, ALLOC_NODE_MISMATCH);
2624 deactivate_slab(s, page, c->freelist, c);
2630 * By rights, we should be searching for a slab page that was
2631 * PFMEMALLOC but right now, we are losing the pfmemalloc
2632 * information when the page leaves the per-cpu allocator
2634 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2635 deactivate_slab(s, page, c->freelist, c);
2639 /* must check again c->freelist in case of cpu migration or IRQ */
2640 freelist = c->freelist;
2644 freelist = get_freelist(s, page);
2648 stat(s, DEACTIVATE_BYPASS);
2652 stat(s, ALLOC_REFILL);
2656 * freelist is pointing to the list of objects to be used.
2657 * page is pointing to the page from which the objects are obtained.
2658 * That page must be frozen for per cpu allocations to work.
2660 VM_BUG_ON(!c->page->frozen);
2661 c->freelist = get_freepointer(s, freelist);
2662 c->tid = next_tid(c->tid);
2667 if (slub_percpu_partial(c)) {
2668 page = c->page = slub_percpu_partial(c);
2669 slub_set_percpu_partial(c, page);
2670 stat(s, CPU_PARTIAL_ALLOC);
2674 freelist = new_slab_objects(s, gfpflags, node, &c);
2676 if (unlikely(!freelist)) {
2677 slab_out_of_memory(s, gfpflags, node);
2682 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2685 /* Only entered in the debug case */
2686 if (kmem_cache_debug(s) &&
2687 !alloc_debug_processing(s, page, freelist, addr))
2688 goto new_slab; /* Slab failed checks. Next slab needed */
2690 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2695 * Another one that disabled interrupt and compensates for possible
2696 * cpu changes by refetching the per cpu area pointer.
2698 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2699 unsigned long addr, struct kmem_cache_cpu *c)
2702 unsigned long flags;
2704 local_irq_save(flags);
2705 #ifdef CONFIG_PREEMPTION
2707 * We may have been preempted and rescheduled on a different
2708 * cpu before disabling interrupts. Need to reload cpu area
2711 c = this_cpu_ptr(s->cpu_slab);
2714 p = ___slab_alloc(s, gfpflags, node, addr, c);
2715 local_irq_restore(flags);
2720 * If the object has been wiped upon free, make sure it's fully initialized by
2721 * zeroing out freelist pointer.
2723 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2726 if (unlikely(slab_want_init_on_free(s)) && obj)
2727 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2731 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2732 * have the fastpath folded into their functions. So no function call
2733 * overhead for requests that can be satisfied on the fastpath.
2735 * The fastpath works by first checking if the lockless freelist can be used.
2736 * If not then __slab_alloc is called for slow processing.
2738 * Otherwise we can simply pick the next object from the lockless free list.
2740 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2741 gfp_t gfpflags, int node, unsigned long addr)
2744 struct kmem_cache_cpu *c;
2748 s = slab_pre_alloc_hook(s, gfpflags);
2753 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2754 * enabled. We may switch back and forth between cpus while
2755 * reading from one cpu area. That does not matter as long
2756 * as we end up on the original cpu again when doing the cmpxchg.
2758 * We should guarantee that tid and kmem_cache are retrieved on
2759 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2760 * to check if it is matched or not.
2763 tid = this_cpu_read(s->cpu_slab->tid);
2764 c = raw_cpu_ptr(s->cpu_slab);
2765 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2766 unlikely(tid != READ_ONCE(c->tid)));
2769 * Irqless object alloc/free algorithm used here depends on sequence
2770 * of fetching cpu_slab's data. tid should be fetched before anything
2771 * on c to guarantee that object and page associated with previous tid
2772 * won't be used with current tid. If we fetch tid first, object and
2773 * page could be one associated with next tid and our alloc/free
2774 * request will be failed. In this case, we will retry. So, no problem.
2779 * The transaction ids are globally unique per cpu and per operation on
2780 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2781 * occurs on the right processor and that there was no operation on the
2782 * linked list in between.
2785 object = c->freelist;
2787 if (unlikely(!object || !node_match(page, node))) {
2788 object = __slab_alloc(s, gfpflags, node, addr, c);
2789 stat(s, ALLOC_SLOWPATH);
2791 void *next_object = get_freepointer_safe(s, object);
2794 * The cmpxchg will only match if there was no additional
2795 * operation and if we are on the right processor.
2797 * The cmpxchg does the following atomically (without lock
2799 * 1. Relocate first pointer to the current per cpu area.
2800 * 2. Verify that tid and freelist have not been changed
2801 * 3. If they were not changed replace tid and freelist
2803 * Since this is without lock semantics the protection is only
2804 * against code executing on this cpu *not* from access by
2807 if (unlikely(!this_cpu_cmpxchg_double(
2808 s->cpu_slab->freelist, s->cpu_slab->tid,
2810 next_object, next_tid(tid)))) {
2812 note_cmpxchg_failure("slab_alloc", s, tid);
2815 prefetch_freepointer(s, next_object);
2816 stat(s, ALLOC_FASTPATH);
2819 maybe_wipe_obj_freeptr(s, object);
2821 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2822 memset(object, 0, s->object_size);
2824 slab_post_alloc_hook(s, gfpflags, 1, &object);
2829 static __always_inline void *slab_alloc(struct kmem_cache *s,
2830 gfp_t gfpflags, unsigned long addr)
2832 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2835 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2837 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2839 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2844 EXPORT_SYMBOL(kmem_cache_alloc);
2846 #ifdef CONFIG_TRACING
2847 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2849 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2850 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2851 ret = kasan_kmalloc(s, ret, size, gfpflags);
2854 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2858 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2860 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2862 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2863 s->object_size, s->size, gfpflags, node);
2867 EXPORT_SYMBOL(kmem_cache_alloc_node);
2869 #ifdef CONFIG_TRACING
2870 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2872 int node, size_t size)
2874 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2876 trace_kmalloc_node(_RET_IP_, ret,
2877 size, s->size, gfpflags, node);
2879 ret = kasan_kmalloc(s, ret, size, gfpflags);
2882 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2884 #endif /* CONFIG_NUMA */
2887 * Slow path handling. This may still be called frequently since objects
2888 * have a longer lifetime than the cpu slabs in most processing loads.
2890 * So we still attempt to reduce cache line usage. Just take the slab
2891 * lock and free the item. If there is no additional partial page
2892 * handling required then we can return immediately.
2894 static void __slab_free(struct kmem_cache *s, struct page *page,
2895 void *head, void *tail, int cnt,
2902 unsigned long counters;
2903 struct kmem_cache_node *n = NULL;
2904 unsigned long uninitialized_var(flags);
2906 stat(s, FREE_SLOWPATH);
2908 if (kmem_cache_debug(s) &&
2909 !free_debug_processing(s, page, head, tail, cnt, addr))
2914 spin_unlock_irqrestore(&n->list_lock, flags);
2917 prior = page->freelist;
2918 counters = page->counters;
2919 set_freepointer(s, tail, prior);
2920 new.counters = counters;
2921 was_frozen = new.frozen;
2923 if ((!new.inuse || !prior) && !was_frozen) {
2925 if (kmem_cache_has_cpu_partial(s) && !prior) {
2928 * Slab was on no list before and will be
2930 * We can defer the list move and instead
2935 } else { /* Needs to be taken off a list */
2937 n = get_node(s, page_to_nid(page));
2939 * Speculatively acquire the list_lock.
2940 * If the cmpxchg does not succeed then we may
2941 * drop the list_lock without any processing.
2943 * Otherwise the list_lock will synchronize with
2944 * other processors updating the list of slabs.
2946 spin_lock_irqsave(&n->list_lock, flags);
2951 } while (!cmpxchg_double_slab(s, page,
2959 * If we just froze the page then put it onto the
2960 * per cpu partial list.
2962 if (new.frozen && !was_frozen) {
2963 put_cpu_partial(s, page, 1);
2964 stat(s, CPU_PARTIAL_FREE);
2967 * The list lock was not taken therefore no list
2968 * activity can be necessary.
2971 stat(s, FREE_FROZEN);
2975 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2979 * Objects left in the slab. If it was not on the partial list before
2982 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2983 remove_full(s, n, page);
2984 add_partial(n, page, DEACTIVATE_TO_TAIL);
2985 stat(s, FREE_ADD_PARTIAL);
2987 spin_unlock_irqrestore(&n->list_lock, flags);
2993 * Slab on the partial list.
2995 remove_partial(n, page);
2996 stat(s, FREE_REMOVE_PARTIAL);
2998 /* Slab must be on the full list */
2999 remove_full(s, n, page);
3002 spin_unlock_irqrestore(&n->list_lock, flags);
3004 discard_slab(s, page);
3008 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3009 * can perform fastpath freeing without additional function calls.
3011 * The fastpath is only possible if we are freeing to the current cpu slab
3012 * of this processor. This typically the case if we have just allocated
3015 * If fastpath is not possible then fall back to __slab_free where we deal
3016 * with all sorts of special processing.
3018 * Bulk free of a freelist with several objects (all pointing to the
3019 * same page) possible by specifying head and tail ptr, plus objects
3020 * count (cnt). Bulk free indicated by tail pointer being set.
3022 static __always_inline void do_slab_free(struct kmem_cache *s,
3023 struct page *page, void *head, void *tail,
3024 int cnt, unsigned long addr)
3026 void *tail_obj = tail ? : head;
3027 struct kmem_cache_cpu *c;
3031 * Determine the currently cpus per cpu slab.
3032 * The cpu may change afterward. However that does not matter since
3033 * data is retrieved via this pointer. If we are on the same cpu
3034 * during the cmpxchg then the free will succeed.
3037 tid = this_cpu_read(s->cpu_slab->tid);
3038 c = raw_cpu_ptr(s->cpu_slab);
3039 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3040 unlikely(tid != READ_ONCE(c->tid)));
3042 /* Same with comment on barrier() in slab_alloc_node() */
3045 if (likely(page == c->page)) {
3046 void **freelist = READ_ONCE(c->freelist);
3048 set_freepointer(s, tail_obj, freelist);
3050 if (unlikely(!this_cpu_cmpxchg_double(
3051 s->cpu_slab->freelist, s->cpu_slab->tid,
3053 head, next_tid(tid)))) {
3055 note_cmpxchg_failure("slab_free", s, tid);
3058 stat(s, FREE_FASTPATH);
3060 __slab_free(s, page, head, tail_obj, cnt, addr);
3064 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3065 void *head, void *tail, int cnt,
3069 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3070 * to remove objects, whose reuse must be delayed.
3072 if (slab_free_freelist_hook(s, &head, &tail))
3073 do_slab_free(s, page, head, tail, cnt, addr);
3076 #ifdef CONFIG_KASAN_GENERIC
3077 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3079 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3083 void kmem_cache_free(struct kmem_cache *s, void *x)
3085 s = cache_from_obj(s, x);
3088 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3089 trace_kmem_cache_free(_RET_IP_, x);
3091 EXPORT_SYMBOL(kmem_cache_free);
3093 struct detached_freelist {
3098 struct kmem_cache *s;
3102 * This function progressively scans the array with free objects (with
3103 * a limited look ahead) and extract objects belonging to the same
3104 * page. It builds a detached freelist directly within the given
3105 * page/objects. This can happen without any need for
3106 * synchronization, because the objects are owned by running process.
3107 * The freelist is build up as a single linked list in the objects.
3108 * The idea is, that this detached freelist can then be bulk
3109 * transferred to the real freelist(s), but only requiring a single
3110 * synchronization primitive. Look ahead in the array is limited due
3111 * to performance reasons.
3114 int build_detached_freelist(struct kmem_cache *s, size_t size,
3115 void **p, struct detached_freelist *df)
3117 size_t first_skipped_index = 0;
3122 /* Always re-init detached_freelist */
3127 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3128 } while (!object && size);
3133 page = virt_to_head_page(object);
3135 /* Handle kalloc'ed objects */
3136 if (unlikely(!PageSlab(page))) {
3137 BUG_ON(!PageCompound(page));
3139 __free_pages(page, compound_order(page));
3140 p[size] = NULL; /* mark object processed */
3143 /* Derive kmem_cache from object */
3144 df->s = page->slab_cache;
3146 df->s = cache_from_obj(s, object); /* Support for memcg */
3149 /* Start new detached freelist */
3151 set_freepointer(df->s, object, NULL);
3153 df->freelist = object;
3154 p[size] = NULL; /* mark object processed */
3160 continue; /* Skip processed objects */
3162 /* df->page is always set at this point */
3163 if (df->page == virt_to_head_page(object)) {
3164 /* Opportunity build freelist */
3165 set_freepointer(df->s, object, df->freelist);
3166 df->freelist = object;
3168 p[size] = NULL; /* mark object processed */
3173 /* Limit look ahead search */
3177 if (!first_skipped_index)
3178 first_skipped_index = size + 1;
3181 return first_skipped_index;
3184 /* Note that interrupts must be enabled when calling this function. */
3185 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3191 struct detached_freelist df;
3193 size = build_detached_freelist(s, size, p, &df);
3197 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3198 } while (likely(size));
3200 EXPORT_SYMBOL(kmem_cache_free_bulk);
3202 /* Note that interrupts must be enabled when calling this function. */
3203 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3206 struct kmem_cache_cpu *c;
3209 /* memcg and kmem_cache debug support */
3210 s = slab_pre_alloc_hook(s, flags);
3214 * Drain objects in the per cpu slab, while disabling local
3215 * IRQs, which protects against PREEMPT and interrupts
3216 * handlers invoking normal fastpath.
3218 local_irq_disable();
3219 c = this_cpu_ptr(s->cpu_slab);
3221 for (i = 0; i < size; i++) {
3222 void *object = c->freelist;
3224 if (unlikely(!object)) {
3226 * We may have removed an object from c->freelist using
3227 * the fastpath in the previous iteration; in that case,
3228 * c->tid has not been bumped yet.
3229 * Since ___slab_alloc() may reenable interrupts while
3230 * allocating memory, we should bump c->tid now.
3232 c->tid = next_tid(c->tid);
3235 * Invoking slow path likely have side-effect
3236 * of re-populating per CPU c->freelist
3238 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3240 if (unlikely(!p[i]))
3243 c = this_cpu_ptr(s->cpu_slab);
3244 maybe_wipe_obj_freeptr(s, p[i]);
3246 continue; /* goto for-loop */
3248 c->freelist = get_freepointer(s, object);
3250 maybe_wipe_obj_freeptr(s, p[i]);
3252 c->tid = next_tid(c->tid);
3255 /* Clear memory outside IRQ disabled fastpath loop */
3256 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3259 for (j = 0; j < i; j++)
3260 memset(p[j], 0, s->object_size);
3263 /* memcg and kmem_cache debug support */
3264 slab_post_alloc_hook(s, flags, size, p);
3268 slab_post_alloc_hook(s, flags, i, p);
3269 __kmem_cache_free_bulk(s, i, p);
3272 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3276 * Object placement in a slab is made very easy because we always start at
3277 * offset 0. If we tune the size of the object to the alignment then we can
3278 * get the required alignment by putting one properly sized object after
3281 * Notice that the allocation order determines the sizes of the per cpu
3282 * caches. Each processor has always one slab available for allocations.
3283 * Increasing the allocation order reduces the number of times that slabs
3284 * must be moved on and off the partial lists and is therefore a factor in
3289 * Mininum / Maximum order of slab pages. This influences locking overhead
3290 * and slab fragmentation. A higher order reduces the number of partial slabs
3291 * and increases the number of allocations possible without having to
3292 * take the list_lock.
3294 static unsigned int slub_min_order;
3295 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3296 static unsigned int slub_min_objects;
3299 * Calculate the order of allocation given an slab object size.
3301 * The order of allocation has significant impact on performance and other
3302 * system components. Generally order 0 allocations should be preferred since
3303 * order 0 does not cause fragmentation in the page allocator. Larger objects
3304 * be problematic to put into order 0 slabs because there may be too much
3305 * unused space left. We go to a higher order if more than 1/16th of the slab
3308 * In order to reach satisfactory performance we must ensure that a minimum
3309 * number of objects is in one slab. Otherwise we may generate too much
3310 * activity on the partial lists which requires taking the list_lock. This is
3311 * less a concern for large slabs though which are rarely used.
3313 * slub_max_order specifies the order where we begin to stop considering the
3314 * number of objects in a slab as critical. If we reach slub_max_order then
3315 * we try to keep the page order as low as possible. So we accept more waste
3316 * of space in favor of a small page order.
3318 * Higher order allocations also allow the placement of more objects in a
3319 * slab and thereby reduce object handling overhead. If the user has
3320 * requested a higher mininum order then we start with that one instead of
3321 * the smallest order which will fit the object.
3323 static inline unsigned int slab_order(unsigned int size,
3324 unsigned int min_objects, unsigned int max_order,
3325 unsigned int fract_leftover)
3327 unsigned int min_order = slub_min_order;
3330 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3331 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3333 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3334 order <= max_order; order++) {
3336 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3339 rem = slab_size % size;
3341 if (rem <= slab_size / fract_leftover)
3348 static inline int calculate_order(unsigned int size)
3351 unsigned int min_objects;
3352 unsigned int max_objects;
3355 * Attempt to find best configuration for a slab. This
3356 * works by first attempting to generate a layout with
3357 * the best configuration and backing off gradually.
3359 * First we increase the acceptable waste in a slab. Then
3360 * we reduce the minimum objects required in a slab.
3362 min_objects = slub_min_objects;
3364 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3365 max_objects = order_objects(slub_max_order, size);
3366 min_objects = min(min_objects, max_objects);
3368 while (min_objects > 1) {
3369 unsigned int fraction;
3372 while (fraction >= 4) {
3373 order = slab_order(size, min_objects,
3374 slub_max_order, fraction);
3375 if (order <= slub_max_order)
3383 * We were unable to place multiple objects in a slab. Now
3384 * lets see if we can place a single object there.
3386 order = slab_order(size, 1, slub_max_order, 1);
3387 if (order <= slub_max_order)
3391 * Doh this slab cannot be placed using slub_max_order.
3393 order = slab_order(size, 1, MAX_ORDER, 1);
3394 if (order < MAX_ORDER)
3400 init_kmem_cache_node(struct kmem_cache_node *n)
3403 spin_lock_init(&n->list_lock);
3404 INIT_LIST_HEAD(&n->partial);
3405 #ifdef CONFIG_SLUB_DEBUG
3406 atomic_long_set(&n->nr_slabs, 0);
3407 atomic_long_set(&n->total_objects, 0);
3408 INIT_LIST_HEAD(&n->full);
3412 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3414 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3415 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3418 * Must align to double word boundary for the double cmpxchg
3419 * instructions to work; see __pcpu_double_call_return_bool().
3421 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3422 2 * sizeof(void *));
3427 init_kmem_cache_cpus(s);
3432 static struct kmem_cache *kmem_cache_node;
3435 * No kmalloc_node yet so do it by hand. We know that this is the first
3436 * slab on the node for this slabcache. There are no concurrent accesses
3439 * Note that this function only works on the kmem_cache_node
3440 * when allocating for the kmem_cache_node. This is used for bootstrapping
3441 * memory on a fresh node that has no slab structures yet.
3443 static void early_kmem_cache_node_alloc(int node)
3446 struct kmem_cache_node *n;
3448 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3450 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3453 if (page_to_nid(page) != node) {
3454 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3455 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3460 #ifdef CONFIG_SLUB_DEBUG
3461 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3462 init_tracking(kmem_cache_node, n);
3464 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3466 page->freelist = get_freepointer(kmem_cache_node, n);
3469 kmem_cache_node->node[node] = n;
3470 init_kmem_cache_node(n);
3471 inc_slabs_node(kmem_cache_node, node, page->objects);
3474 * No locks need to be taken here as it has just been
3475 * initialized and there is no concurrent access.
3477 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3480 static void free_kmem_cache_nodes(struct kmem_cache *s)
3483 struct kmem_cache_node *n;
3485 for_each_kmem_cache_node(s, node, n) {
3486 s->node[node] = NULL;
3487 kmem_cache_free(kmem_cache_node, n);
3491 void __kmem_cache_release(struct kmem_cache *s)
3493 cache_random_seq_destroy(s);
3494 free_percpu(s->cpu_slab);
3495 free_kmem_cache_nodes(s);
3498 static int init_kmem_cache_nodes(struct kmem_cache *s)
3502 for_each_node_state(node, N_NORMAL_MEMORY) {
3503 struct kmem_cache_node *n;
3505 if (slab_state == DOWN) {
3506 early_kmem_cache_node_alloc(node);
3509 n = kmem_cache_alloc_node(kmem_cache_node,
3513 free_kmem_cache_nodes(s);
3517 init_kmem_cache_node(n);
3523 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3525 if (min < MIN_PARTIAL)
3527 else if (min > MAX_PARTIAL)
3529 s->min_partial = min;
3532 static void set_cpu_partial(struct kmem_cache *s)
3534 #ifdef CONFIG_SLUB_CPU_PARTIAL
3536 * cpu_partial determined the maximum number of objects kept in the
3537 * per cpu partial lists of a processor.
3539 * Per cpu partial lists mainly contain slabs that just have one
3540 * object freed. If they are used for allocation then they can be
3541 * filled up again with minimal effort. The slab will never hit the
3542 * per node partial lists and therefore no locking will be required.
3544 * This setting also determines
3546 * A) The number of objects from per cpu partial slabs dumped to the
3547 * per node list when we reach the limit.
3548 * B) The number of objects in cpu partial slabs to extract from the
3549 * per node list when we run out of per cpu objects. We only fetch
3550 * 50% to keep some capacity around for frees.
3552 if (!kmem_cache_has_cpu_partial(s))
3553 slub_set_cpu_partial(s, 0);
3554 else if (s->size >= PAGE_SIZE)
3555 slub_set_cpu_partial(s, 2);
3556 else if (s->size >= 1024)
3557 slub_set_cpu_partial(s, 6);
3558 else if (s->size >= 256)
3559 slub_set_cpu_partial(s, 13);
3561 slub_set_cpu_partial(s, 30);
3566 * calculate_sizes() determines the order and the distribution of data within
3569 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3571 slab_flags_t flags = s->flags;
3572 unsigned int size = s->object_size;
3573 unsigned int freepointer_area;
3577 * Round up object size to the next word boundary. We can only
3578 * place the free pointer at word boundaries and this determines
3579 * the possible location of the free pointer.
3581 size = ALIGN(size, sizeof(void *));
3583 * This is the area of the object where a freepointer can be
3584 * safely written. If redzoning adds more to the inuse size, we
3585 * can't use that portion for writing the freepointer, so
3586 * s->offset must be limited within this for the general case.
3588 freepointer_area = size;
3590 #ifdef CONFIG_SLUB_DEBUG
3592 * Determine if we can poison the object itself. If the user of
3593 * the slab may touch the object after free or before allocation
3594 * then we should never poison the object itself.
3596 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3598 s->flags |= __OBJECT_POISON;
3600 s->flags &= ~__OBJECT_POISON;
3604 * If we are Redzoning then check if there is some space between the
3605 * end of the object and the free pointer. If not then add an
3606 * additional word to have some bytes to store Redzone information.
3608 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3609 size += sizeof(void *);
3613 * With that we have determined the number of bytes in actual use
3614 * by the object. This is the potential offset to the free pointer.
3618 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3621 * Relocate free pointer after the object if it is not
3622 * permitted to overwrite the first word of the object on
3625 * This is the case if we do RCU, have a constructor or
3626 * destructor or are poisoning the objects.
3628 * The assumption that s->offset >= s->inuse means free
3629 * pointer is outside of the object is used in the
3630 * freeptr_outside_object() function. If that is no
3631 * longer true, the function needs to be modified.
3634 size += sizeof(void *);
3635 } else if (freepointer_area > sizeof(void *)) {
3637 * Store freelist pointer near middle of object to keep
3638 * it away from the edges of the object to avoid small
3639 * sized over/underflows from neighboring allocations.
3641 s->offset = ALIGN(freepointer_area / 2, sizeof(void *));
3644 #ifdef CONFIG_SLUB_DEBUG
3645 if (flags & SLAB_STORE_USER)
3647 * Need to store information about allocs and frees after
3650 size += 2 * sizeof(struct track);
3653 kasan_cache_create(s, &size, &s->flags);
3654 #ifdef CONFIG_SLUB_DEBUG
3655 if (flags & SLAB_RED_ZONE) {
3657 * Add some empty padding so that we can catch
3658 * overwrites from earlier objects rather than let
3659 * tracking information or the free pointer be
3660 * corrupted if a user writes before the start
3663 size += sizeof(void *);
3665 s->red_left_pad = sizeof(void *);
3666 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3667 size += s->red_left_pad;
3672 * SLUB stores one object immediately after another beginning from
3673 * offset 0. In order to align the objects we have to simply size
3674 * each object to conform to the alignment.
3676 size = ALIGN(size, s->align);
3678 if (forced_order >= 0)
3679 order = forced_order;
3681 order = calculate_order(size);
3688 s->allocflags |= __GFP_COMP;
3690 if (s->flags & SLAB_CACHE_DMA)
3691 s->allocflags |= GFP_DMA;
3693 if (s->flags & SLAB_CACHE_DMA32)
3694 s->allocflags |= GFP_DMA32;
3696 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3697 s->allocflags |= __GFP_RECLAIMABLE;
3700 * Determine the number of objects per slab
3702 s->oo = oo_make(order, size);
3703 s->min = oo_make(get_order(size), size);
3704 if (oo_objects(s->oo) > oo_objects(s->max))
3707 return !!oo_objects(s->oo);
3710 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3712 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3713 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3714 s->random = get_random_long();
3717 if (!calculate_sizes(s, -1))
3719 if (disable_higher_order_debug) {
3721 * Disable debugging flags that store metadata if the min slab
3724 if (get_order(s->size) > get_order(s->object_size)) {
3725 s->flags &= ~DEBUG_METADATA_FLAGS;
3727 if (!calculate_sizes(s, -1))
3732 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3733 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3734 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3735 /* Enable fast mode */
3736 s->flags |= __CMPXCHG_DOUBLE;
3740 * The larger the object size is, the more pages we want on the partial
3741 * list to avoid pounding the page allocator excessively.
3743 set_min_partial(s, ilog2(s->size) / 2);
3748 s->remote_node_defrag_ratio = 1000;
3751 /* Initialize the pre-computed randomized freelist if slab is up */
3752 if (slab_state >= UP) {
3753 if (init_cache_random_seq(s))
3757 if (!init_kmem_cache_nodes(s))
3760 if (alloc_kmem_cache_cpus(s))
3763 free_kmem_cache_nodes(s);
3768 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3771 #ifdef CONFIG_SLUB_DEBUG
3772 void *addr = page_address(page);
3776 slab_err(s, page, text, s->name);
3779 map = get_map(s, page);
3780 for_each_object(p, s, addr, page->objects) {
3782 if (!test_bit(slab_index(p, s, addr), map)) {
3783 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3784 print_tracking(s, p);
3793 * Attempt to free all partial slabs on a node.
3794 * This is called from __kmem_cache_shutdown(). We must take list_lock
3795 * because sysfs file might still access partial list after the shutdowning.
3797 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3800 struct page *page, *h;
3802 BUG_ON(irqs_disabled());
3803 spin_lock_irq(&n->list_lock);
3804 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3806 remove_partial(n, page);
3807 list_add(&page->slab_list, &discard);
3809 list_slab_objects(s, page,
3810 "Objects remaining in %s on __kmem_cache_shutdown()");
3813 spin_unlock_irq(&n->list_lock);
3815 list_for_each_entry_safe(page, h, &discard, slab_list)
3816 discard_slab(s, page);
3819 bool __kmem_cache_empty(struct kmem_cache *s)
3822 struct kmem_cache_node *n;
3824 for_each_kmem_cache_node(s, node, n)
3825 if (n->nr_partial || slabs_node(s, node))
3831 * Release all resources used by a slab cache.
3833 int __kmem_cache_shutdown(struct kmem_cache *s)
3836 struct kmem_cache_node *n;
3839 /* Attempt to free all objects */
3840 for_each_kmem_cache_node(s, node, n) {
3842 if (n->nr_partial || slabs_node(s, node))
3845 sysfs_slab_remove(s);
3849 /********************************************************************
3851 *******************************************************************/
3853 static int __init setup_slub_min_order(char *str)
3855 get_option(&str, (int *)&slub_min_order);
3860 __setup("slub_min_order=", setup_slub_min_order);
3862 static int __init setup_slub_max_order(char *str)
3864 get_option(&str, (int *)&slub_max_order);
3865 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3870 __setup("slub_max_order=", setup_slub_max_order);
3872 static int __init setup_slub_min_objects(char *str)
3874 get_option(&str, (int *)&slub_min_objects);
3879 __setup("slub_min_objects=", setup_slub_min_objects);
3881 void *__kmalloc(size_t size, gfp_t flags)
3883 struct kmem_cache *s;
3886 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3887 return kmalloc_large(size, flags);
3889 s = kmalloc_slab(size, flags);
3891 if (unlikely(ZERO_OR_NULL_PTR(s)))
3894 ret = slab_alloc(s, flags, _RET_IP_);
3896 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3898 ret = kasan_kmalloc(s, ret, size, flags);
3902 EXPORT_SYMBOL(__kmalloc);
3905 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3909 unsigned int order = get_order(size);
3911 flags |= __GFP_COMP;
3912 page = alloc_pages_node(node, flags, order);
3914 ptr = page_address(page);
3915 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3919 return kmalloc_large_node_hook(ptr, size, flags);
3922 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3924 struct kmem_cache *s;
3927 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3928 ret = kmalloc_large_node(size, flags, node);
3930 trace_kmalloc_node(_RET_IP_, ret,
3931 size, PAGE_SIZE << get_order(size),
3937 s = kmalloc_slab(size, flags);
3939 if (unlikely(ZERO_OR_NULL_PTR(s)))
3942 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3944 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3946 ret = kasan_kmalloc(s, ret, size, flags);
3950 EXPORT_SYMBOL(__kmalloc_node);
3951 #endif /* CONFIG_NUMA */
3953 #ifdef CONFIG_HARDENED_USERCOPY
3955 * Rejects incorrectly sized objects and objects that are to be copied
3956 * to/from userspace but do not fall entirely within the containing slab
3957 * cache's usercopy region.
3959 * Returns NULL if check passes, otherwise const char * to name of cache
3960 * to indicate an error.
3962 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3965 struct kmem_cache *s;
3966 unsigned int offset;
3969 ptr = kasan_reset_tag(ptr);
3971 /* Find object and usable object size. */
3972 s = page->slab_cache;
3974 /* Reject impossible pointers. */
3975 if (ptr < page_address(page))
3976 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3979 /* Find offset within object. */
3980 offset = (ptr - page_address(page)) % s->size;
3982 /* Adjust for redzone and reject if within the redzone. */
3983 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3984 if (offset < s->red_left_pad)
3985 usercopy_abort("SLUB object in left red zone",
3986 s->name, to_user, offset, n);
3987 offset -= s->red_left_pad;
3990 /* Allow address range falling entirely within usercopy region. */
3991 if (offset >= s->useroffset &&
3992 offset - s->useroffset <= s->usersize &&
3993 n <= s->useroffset - offset + s->usersize)
3997 * If the copy is still within the allocated object, produce
3998 * a warning instead of rejecting the copy. This is intended
3999 * to be a temporary method to find any missing usercopy
4002 object_size = slab_ksize(s);
4003 if (usercopy_fallback &&
4004 offset <= object_size && n <= object_size - offset) {
4005 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4009 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4011 #endif /* CONFIG_HARDENED_USERCOPY */
4013 size_t __ksize(const void *object)
4017 if (unlikely(object == ZERO_SIZE_PTR))
4020 page = virt_to_head_page(object);
4022 if (unlikely(!PageSlab(page))) {
4023 WARN_ON(!PageCompound(page));
4024 return page_size(page);
4027 return slab_ksize(page->slab_cache);
4029 EXPORT_SYMBOL(__ksize);
4031 void kfree(const void *x)
4034 void *object = (void *)x;
4036 trace_kfree(_RET_IP_, x);
4038 if (unlikely(ZERO_OR_NULL_PTR(x)))
4041 page = virt_to_head_page(x);
4042 if (unlikely(!PageSlab(page))) {
4043 unsigned int order = compound_order(page);
4045 BUG_ON(!PageCompound(page));
4047 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
4049 __free_pages(page, order);
4052 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4054 EXPORT_SYMBOL(kfree);
4056 #define SHRINK_PROMOTE_MAX 32
4059 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4060 * up most to the head of the partial lists. New allocations will then
4061 * fill those up and thus they can be removed from the partial lists.
4063 * The slabs with the least items are placed last. This results in them
4064 * being allocated from last increasing the chance that the last objects
4065 * are freed in them.
4067 int __kmem_cache_shrink(struct kmem_cache *s)
4071 struct kmem_cache_node *n;
4074 struct list_head discard;
4075 struct list_head promote[SHRINK_PROMOTE_MAX];
4076 unsigned long flags;
4080 for_each_kmem_cache_node(s, node, n) {
4081 INIT_LIST_HEAD(&discard);
4082 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4083 INIT_LIST_HEAD(promote + i);
4085 spin_lock_irqsave(&n->list_lock, flags);
4088 * Build lists of slabs to discard or promote.
4090 * Note that concurrent frees may occur while we hold the
4091 * list_lock. page->inuse here is the upper limit.
4093 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4094 int free = page->objects - page->inuse;
4096 /* Do not reread page->inuse */
4099 /* We do not keep full slabs on the list */
4102 if (free == page->objects) {
4103 list_move(&page->slab_list, &discard);
4105 } else if (free <= SHRINK_PROMOTE_MAX)
4106 list_move(&page->slab_list, promote + free - 1);
4110 * Promote the slabs filled up most to the head of the
4113 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4114 list_splice(promote + i, &n->partial);
4116 spin_unlock_irqrestore(&n->list_lock, flags);
4118 /* Release empty slabs */
4119 list_for_each_entry_safe(page, t, &discard, slab_list)
4120 discard_slab(s, page);
4122 if (slabs_node(s, node))
4130 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4133 * Called with all the locks held after a sched RCU grace period.
4134 * Even if @s becomes empty after shrinking, we can't know that @s
4135 * doesn't have allocations already in-flight and thus can't
4136 * destroy @s until the associated memcg is released.
4138 * However, let's remove the sysfs files for empty caches here.
4139 * Each cache has a lot of interface files which aren't
4140 * particularly useful for empty draining caches; otherwise, we can
4141 * easily end up with millions of unnecessary sysfs files on
4142 * systems which have a lot of memory and transient cgroups.
4144 if (!__kmem_cache_shrink(s))
4145 sysfs_slab_remove(s);
4148 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4151 * Disable empty slabs caching. Used to avoid pinning offline
4152 * memory cgroups by kmem pages that can be freed.
4154 slub_set_cpu_partial(s, 0);
4157 #endif /* CONFIG_MEMCG */
4159 static int slab_mem_going_offline_callback(void *arg)
4161 struct kmem_cache *s;
4163 mutex_lock(&slab_mutex);
4164 list_for_each_entry(s, &slab_caches, list)
4165 __kmem_cache_shrink(s);
4166 mutex_unlock(&slab_mutex);
4171 static void slab_mem_offline_callback(void *arg)
4173 struct kmem_cache_node *n;
4174 struct kmem_cache *s;
4175 struct memory_notify *marg = arg;
4178 offline_node = marg->status_change_nid_normal;
4181 * If the node still has available memory. we need kmem_cache_node
4184 if (offline_node < 0)
4187 mutex_lock(&slab_mutex);
4188 list_for_each_entry(s, &slab_caches, list) {
4189 n = get_node(s, offline_node);
4192 * if n->nr_slabs > 0, slabs still exist on the node
4193 * that is going down. We were unable to free them,
4194 * and offline_pages() function shouldn't call this
4195 * callback. So, we must fail.
4197 BUG_ON(slabs_node(s, offline_node));
4199 s->node[offline_node] = NULL;
4200 kmem_cache_free(kmem_cache_node, n);
4203 mutex_unlock(&slab_mutex);
4206 static int slab_mem_going_online_callback(void *arg)
4208 struct kmem_cache_node *n;
4209 struct kmem_cache *s;
4210 struct memory_notify *marg = arg;
4211 int nid = marg->status_change_nid_normal;
4215 * If the node's memory is already available, then kmem_cache_node is
4216 * already created. Nothing to do.
4222 * We are bringing a node online. No memory is available yet. We must
4223 * allocate a kmem_cache_node structure in order to bring the node
4226 mutex_lock(&slab_mutex);
4227 list_for_each_entry(s, &slab_caches, list) {
4229 * XXX: kmem_cache_alloc_node will fallback to other nodes
4230 * since memory is not yet available from the node that
4233 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4238 init_kmem_cache_node(n);
4242 mutex_unlock(&slab_mutex);
4246 static int slab_memory_callback(struct notifier_block *self,
4247 unsigned long action, void *arg)
4252 case MEM_GOING_ONLINE:
4253 ret = slab_mem_going_online_callback(arg);
4255 case MEM_GOING_OFFLINE:
4256 ret = slab_mem_going_offline_callback(arg);
4259 case MEM_CANCEL_ONLINE:
4260 slab_mem_offline_callback(arg);
4263 case MEM_CANCEL_OFFLINE:
4267 ret = notifier_from_errno(ret);
4273 static struct notifier_block slab_memory_callback_nb = {
4274 .notifier_call = slab_memory_callback,
4275 .priority = SLAB_CALLBACK_PRI,
4278 /********************************************************************
4279 * Basic setup of slabs
4280 *******************************************************************/
4283 * Used for early kmem_cache structures that were allocated using
4284 * the page allocator. Allocate them properly then fix up the pointers
4285 * that may be pointing to the wrong kmem_cache structure.
4288 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4291 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4292 struct kmem_cache_node *n;
4294 memcpy(s, static_cache, kmem_cache->object_size);
4297 * This runs very early, and only the boot processor is supposed to be
4298 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4301 __flush_cpu_slab(s, smp_processor_id());
4302 for_each_kmem_cache_node(s, node, n) {
4305 list_for_each_entry(p, &n->partial, slab_list)
4308 #ifdef CONFIG_SLUB_DEBUG
4309 list_for_each_entry(p, &n->full, slab_list)
4313 slab_init_memcg_params(s);
4314 list_add(&s->list, &slab_caches);
4315 memcg_link_cache(s, NULL);
4319 void __init kmem_cache_init(void)
4321 static __initdata struct kmem_cache boot_kmem_cache,
4322 boot_kmem_cache_node;
4324 if (debug_guardpage_minorder())
4327 kmem_cache_node = &boot_kmem_cache_node;
4328 kmem_cache = &boot_kmem_cache;
4330 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4331 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4333 register_hotmemory_notifier(&slab_memory_callback_nb);
4335 /* Able to allocate the per node structures */
4336 slab_state = PARTIAL;
4338 create_boot_cache(kmem_cache, "kmem_cache",
4339 offsetof(struct kmem_cache, node) +
4340 nr_node_ids * sizeof(struct kmem_cache_node *),
4341 SLAB_HWCACHE_ALIGN, 0, 0);
4343 kmem_cache = bootstrap(&boot_kmem_cache);
4344 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4346 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4347 setup_kmalloc_cache_index_table();
4348 create_kmalloc_caches(0);
4350 /* Setup random freelists for each cache */
4351 init_freelist_randomization();
4353 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4356 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4358 slub_min_order, slub_max_order, slub_min_objects,
4359 nr_cpu_ids, nr_node_ids);
4362 void __init kmem_cache_init_late(void)
4367 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4368 slab_flags_t flags, void (*ctor)(void *))
4370 struct kmem_cache *s, *c;
4372 s = find_mergeable(size, align, flags, name, ctor);
4377 * Adjust the object sizes so that we clear
4378 * the complete object on kzalloc.
4380 s->object_size = max(s->object_size, size);
4381 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4383 for_each_memcg_cache(c, s) {
4384 c->object_size = s->object_size;
4385 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4388 if (sysfs_slab_alias(s, name)) {
4397 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4401 err = kmem_cache_open(s, flags);
4405 /* Mutex is not taken during early boot */
4406 if (slab_state <= UP)
4409 memcg_propagate_slab_attrs(s);
4410 err = sysfs_slab_add(s);
4412 __kmem_cache_release(s);
4417 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4419 struct kmem_cache *s;
4422 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4423 return kmalloc_large(size, gfpflags);
4425 s = kmalloc_slab(size, gfpflags);
4427 if (unlikely(ZERO_OR_NULL_PTR(s)))
4430 ret = slab_alloc(s, gfpflags, caller);
4432 /* Honor the call site pointer we received. */
4433 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4437 EXPORT_SYMBOL(__kmalloc_track_caller);
4440 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4441 int node, unsigned long caller)
4443 struct kmem_cache *s;
4446 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4447 ret = kmalloc_large_node(size, gfpflags, node);
4449 trace_kmalloc_node(caller, ret,
4450 size, PAGE_SIZE << get_order(size),
4456 s = kmalloc_slab(size, gfpflags);
4458 if (unlikely(ZERO_OR_NULL_PTR(s)))
4461 ret = slab_alloc_node(s, gfpflags, node, caller);
4463 /* Honor the call site pointer we received. */
4464 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4468 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4472 static int count_inuse(struct page *page)
4477 static int count_total(struct page *page)
4479 return page->objects;
4483 #ifdef CONFIG_SLUB_DEBUG
4484 static void validate_slab(struct kmem_cache *s, struct page *page)
4487 void *addr = page_address(page);
4492 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4495 /* Now we know that a valid freelist exists */
4496 map = get_map(s, page);
4497 for_each_object(p, s, addr, page->objects) {
4498 u8 val = test_bit(slab_index(p, s, addr), map) ?
4499 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4501 if (!check_object(s, page, p, val))
4509 static int validate_slab_node(struct kmem_cache *s,
4510 struct kmem_cache_node *n)
4512 unsigned long count = 0;
4514 unsigned long flags;
4516 spin_lock_irqsave(&n->list_lock, flags);
4518 list_for_each_entry(page, &n->partial, slab_list) {
4519 validate_slab(s, page);
4522 if (count != n->nr_partial)
4523 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4524 s->name, count, n->nr_partial);
4526 if (!(s->flags & SLAB_STORE_USER))
4529 list_for_each_entry(page, &n->full, slab_list) {
4530 validate_slab(s, page);
4533 if (count != atomic_long_read(&n->nr_slabs))
4534 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4535 s->name, count, atomic_long_read(&n->nr_slabs));
4538 spin_unlock_irqrestore(&n->list_lock, flags);
4542 static long validate_slab_cache(struct kmem_cache *s)
4545 unsigned long count = 0;
4546 struct kmem_cache_node *n;
4549 for_each_kmem_cache_node(s, node, n)
4550 count += validate_slab_node(s, n);
4555 * Generate lists of code addresses where slabcache objects are allocated
4560 unsigned long count;
4567 DECLARE_BITMAP(cpus, NR_CPUS);
4573 unsigned long count;
4574 struct location *loc;
4577 static void free_loc_track(struct loc_track *t)
4580 free_pages((unsigned long)t->loc,
4581 get_order(sizeof(struct location) * t->max));
4584 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4589 order = get_order(sizeof(struct location) * max);
4591 l = (void *)__get_free_pages(flags, order);
4596 memcpy(l, t->loc, sizeof(struct location) * t->count);
4604 static int add_location(struct loc_track *t, struct kmem_cache *s,
4605 const struct track *track)
4607 long start, end, pos;
4609 unsigned long caddr;
4610 unsigned long age = jiffies - track->when;
4616 pos = start + (end - start + 1) / 2;
4619 * There is nothing at "end". If we end up there
4620 * we need to add something to before end.
4625 caddr = t->loc[pos].addr;
4626 if (track->addr == caddr) {
4632 if (age < l->min_time)
4634 if (age > l->max_time)
4637 if (track->pid < l->min_pid)
4638 l->min_pid = track->pid;
4639 if (track->pid > l->max_pid)
4640 l->max_pid = track->pid;
4642 cpumask_set_cpu(track->cpu,
4643 to_cpumask(l->cpus));
4645 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4649 if (track->addr < caddr)
4656 * Not found. Insert new tracking element.
4658 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4664 (t->count - pos) * sizeof(struct location));
4667 l->addr = track->addr;
4671 l->min_pid = track->pid;
4672 l->max_pid = track->pid;
4673 cpumask_clear(to_cpumask(l->cpus));
4674 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4675 nodes_clear(l->nodes);
4676 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4680 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4681 struct page *page, enum track_item alloc)
4683 void *addr = page_address(page);
4687 map = get_map(s, page);
4688 for_each_object(p, s, addr, page->objects)
4689 if (!test_bit(slab_index(p, s, addr), map))
4690 add_location(t, s, get_track(s, p, alloc));
4694 static int list_locations(struct kmem_cache *s, char *buf,
4695 enum track_item alloc)
4699 struct loc_track t = { 0, 0, NULL };
4701 struct kmem_cache_node *n;
4703 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4705 return sprintf(buf, "Out of memory\n");
4707 /* Push back cpu slabs */
4710 for_each_kmem_cache_node(s, node, n) {
4711 unsigned long flags;
4714 if (!atomic_long_read(&n->nr_slabs))
4717 spin_lock_irqsave(&n->list_lock, flags);
4718 list_for_each_entry(page, &n->partial, slab_list)
4719 process_slab(&t, s, page, alloc);
4720 list_for_each_entry(page, &n->full, slab_list)
4721 process_slab(&t, s, page, alloc);
4722 spin_unlock_irqrestore(&n->list_lock, flags);
4725 for (i = 0; i < t.count; i++) {
4726 struct location *l = &t.loc[i];
4728 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4730 len += sprintf(buf + len, "%7ld ", l->count);
4733 len += sprintf(buf + len, "%pS", (void *)l->addr);
4735 len += sprintf(buf + len, "<not-available>");
4737 if (l->sum_time != l->min_time) {
4738 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4740 (long)div_u64(l->sum_time, l->count),
4743 len += sprintf(buf + len, " age=%ld",
4746 if (l->min_pid != l->max_pid)
4747 len += sprintf(buf + len, " pid=%ld-%ld",
4748 l->min_pid, l->max_pid);
4750 len += sprintf(buf + len, " pid=%ld",
4753 if (num_online_cpus() > 1 &&
4754 !cpumask_empty(to_cpumask(l->cpus)) &&
4755 len < PAGE_SIZE - 60)
4756 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4758 cpumask_pr_args(to_cpumask(l->cpus)));
4760 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4761 len < PAGE_SIZE - 60)
4762 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4764 nodemask_pr_args(&l->nodes));
4766 len += sprintf(buf + len, "\n");
4771 len += sprintf(buf, "No data\n");
4774 #endif /* CONFIG_SLUB_DEBUG */
4776 #ifdef SLUB_RESILIENCY_TEST
4777 static void __init resiliency_test(void)
4780 int type = KMALLOC_NORMAL;
4782 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4784 pr_err("SLUB resiliency testing\n");
4785 pr_err("-----------------------\n");
4786 pr_err("A. Corruption after allocation\n");
4788 p = kzalloc(16, GFP_KERNEL);
4790 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4793 validate_slab_cache(kmalloc_caches[type][4]);
4795 /* Hmmm... The next two are dangerous */
4796 p = kzalloc(32, GFP_KERNEL);
4797 p[32 + sizeof(void *)] = 0x34;
4798 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4800 pr_err("If allocated object is overwritten then not detectable\n\n");
4802 validate_slab_cache(kmalloc_caches[type][5]);
4803 p = kzalloc(64, GFP_KERNEL);
4804 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4806 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4808 pr_err("If allocated object is overwritten then not detectable\n\n");
4809 validate_slab_cache(kmalloc_caches[type][6]);
4811 pr_err("\nB. Corruption after free\n");
4812 p = kzalloc(128, GFP_KERNEL);
4815 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4816 validate_slab_cache(kmalloc_caches[type][7]);
4818 p = kzalloc(256, GFP_KERNEL);
4821 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4822 validate_slab_cache(kmalloc_caches[type][8]);
4824 p = kzalloc(512, GFP_KERNEL);
4827 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4828 validate_slab_cache(kmalloc_caches[type][9]);
4832 static void resiliency_test(void) {};
4834 #endif /* SLUB_RESILIENCY_TEST */
4837 enum slab_stat_type {
4838 SL_ALL, /* All slabs */
4839 SL_PARTIAL, /* Only partially allocated slabs */
4840 SL_CPU, /* Only slabs used for cpu caches */
4841 SL_OBJECTS, /* Determine allocated objects not slabs */
4842 SL_TOTAL /* Determine object capacity not slabs */
4845 #define SO_ALL (1 << SL_ALL)
4846 #define SO_PARTIAL (1 << SL_PARTIAL)
4847 #define SO_CPU (1 << SL_CPU)
4848 #define SO_OBJECTS (1 << SL_OBJECTS)
4849 #define SO_TOTAL (1 << SL_TOTAL)
4852 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4854 static int __init setup_slub_memcg_sysfs(char *str)
4858 if (get_option(&str, &v) > 0)
4859 memcg_sysfs_enabled = v;
4864 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4867 static ssize_t show_slab_objects(struct kmem_cache *s,
4868 char *buf, unsigned long flags)
4870 unsigned long total = 0;
4873 unsigned long *nodes;
4875 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4879 if (flags & SO_CPU) {
4882 for_each_possible_cpu(cpu) {
4883 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4888 page = READ_ONCE(c->page);
4892 node = page_to_nid(page);
4893 if (flags & SO_TOTAL)
4895 else if (flags & SO_OBJECTS)
4903 page = slub_percpu_partial_read_once(c);
4905 node = page_to_nid(page);
4906 if (flags & SO_TOTAL)
4908 else if (flags & SO_OBJECTS)
4919 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4920 * already held which will conflict with an existing lock order:
4922 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4924 * We don't really need mem_hotplug_lock (to hold off
4925 * slab_mem_going_offline_callback) here because slab's memory hot
4926 * unplug code doesn't destroy the kmem_cache->node[] data.
4929 #ifdef CONFIG_SLUB_DEBUG
4930 if (flags & SO_ALL) {
4931 struct kmem_cache_node *n;
4933 for_each_kmem_cache_node(s, node, n) {
4935 if (flags & SO_TOTAL)
4936 x = atomic_long_read(&n->total_objects);
4937 else if (flags & SO_OBJECTS)
4938 x = atomic_long_read(&n->total_objects) -
4939 count_partial(n, count_free);
4941 x = atomic_long_read(&n->nr_slabs);
4948 if (flags & SO_PARTIAL) {
4949 struct kmem_cache_node *n;
4951 for_each_kmem_cache_node(s, node, n) {
4952 if (flags & SO_TOTAL)
4953 x = count_partial(n, count_total);
4954 else if (flags & SO_OBJECTS)
4955 x = count_partial(n, count_inuse);
4962 x = sprintf(buf, "%lu", total);
4964 for (node = 0; node < nr_node_ids; node++)
4966 x += sprintf(buf + x, " N%d=%lu",
4970 return x + sprintf(buf + x, "\n");
4973 #ifdef CONFIG_SLUB_DEBUG
4974 static int any_slab_objects(struct kmem_cache *s)
4977 struct kmem_cache_node *n;
4979 for_each_kmem_cache_node(s, node, n)
4980 if (atomic_long_read(&n->total_objects))
4987 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4988 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4990 struct slab_attribute {
4991 struct attribute attr;
4992 ssize_t (*show)(struct kmem_cache *s, char *buf);
4993 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4996 #define SLAB_ATTR_RO(_name) \
4997 static struct slab_attribute _name##_attr = \
4998 __ATTR(_name, 0400, _name##_show, NULL)
5000 #define SLAB_ATTR(_name) \
5001 static struct slab_attribute _name##_attr = \
5002 __ATTR(_name, 0600, _name##_show, _name##_store)
5004 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5006 return sprintf(buf, "%u\n", s->size);
5008 SLAB_ATTR_RO(slab_size);
5010 static ssize_t align_show(struct kmem_cache *s, char *buf)
5012 return sprintf(buf, "%u\n", s->align);
5014 SLAB_ATTR_RO(align);
5016 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5018 return sprintf(buf, "%u\n", s->object_size);
5020 SLAB_ATTR_RO(object_size);
5022 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5024 return sprintf(buf, "%u\n", oo_objects(s->oo));
5026 SLAB_ATTR_RO(objs_per_slab);
5028 static ssize_t order_store(struct kmem_cache *s,
5029 const char *buf, size_t length)
5034 err = kstrtouint(buf, 10, &order);
5038 if (order > slub_max_order || order < slub_min_order)
5041 calculate_sizes(s, order);
5045 static ssize_t order_show(struct kmem_cache *s, char *buf)
5047 return sprintf(buf, "%u\n", oo_order(s->oo));
5051 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5053 return sprintf(buf, "%lu\n", s->min_partial);
5056 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5062 err = kstrtoul(buf, 10, &min);
5066 set_min_partial(s, min);
5069 SLAB_ATTR(min_partial);
5071 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5073 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5076 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5079 unsigned int objects;
5082 err = kstrtouint(buf, 10, &objects);
5085 if (objects && !kmem_cache_has_cpu_partial(s))
5088 slub_set_cpu_partial(s, objects);
5092 SLAB_ATTR(cpu_partial);
5094 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5098 return sprintf(buf, "%pS\n", s->ctor);
5102 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5104 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5106 SLAB_ATTR_RO(aliases);
5108 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5110 return show_slab_objects(s, buf, SO_PARTIAL);
5112 SLAB_ATTR_RO(partial);
5114 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5116 return show_slab_objects(s, buf, SO_CPU);
5118 SLAB_ATTR_RO(cpu_slabs);
5120 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5122 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5124 SLAB_ATTR_RO(objects);
5126 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5128 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5130 SLAB_ATTR_RO(objects_partial);
5132 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5139 for_each_online_cpu(cpu) {
5142 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5145 pages += page->pages;
5146 objects += page->pobjects;
5150 len = sprintf(buf, "%d(%d)", objects, pages);
5153 for_each_online_cpu(cpu) {
5156 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5158 if (page && len < PAGE_SIZE - 20)
5159 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5160 page->pobjects, page->pages);
5163 return len + sprintf(buf + len, "\n");
5165 SLAB_ATTR_RO(slabs_cpu_partial);
5167 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5169 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5172 static ssize_t reclaim_account_store(struct kmem_cache *s,
5173 const char *buf, size_t length)
5175 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5177 s->flags |= SLAB_RECLAIM_ACCOUNT;
5180 SLAB_ATTR(reclaim_account);
5182 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5184 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5186 SLAB_ATTR_RO(hwcache_align);
5188 #ifdef CONFIG_ZONE_DMA
5189 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5191 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5193 SLAB_ATTR_RO(cache_dma);
5196 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5198 return sprintf(buf, "%u\n", s->usersize);
5200 SLAB_ATTR_RO(usersize);
5202 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5204 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5206 SLAB_ATTR_RO(destroy_by_rcu);
5208 #ifdef CONFIG_SLUB_DEBUG
5209 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5211 return show_slab_objects(s, buf, SO_ALL);
5213 SLAB_ATTR_RO(slabs);
5215 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5217 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5219 SLAB_ATTR_RO(total_objects);
5221 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5223 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5226 static ssize_t sanity_checks_store(struct kmem_cache *s,
5227 const char *buf, size_t length)
5229 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5230 if (buf[0] == '1') {
5231 s->flags &= ~__CMPXCHG_DOUBLE;
5232 s->flags |= SLAB_CONSISTENCY_CHECKS;
5236 SLAB_ATTR(sanity_checks);
5238 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5240 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5243 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5247 * Tracing a merged cache is going to give confusing results
5248 * as well as cause other issues like converting a mergeable
5249 * cache into an umergeable one.
5251 if (s->refcount > 1)
5254 s->flags &= ~SLAB_TRACE;
5255 if (buf[0] == '1') {
5256 s->flags &= ~__CMPXCHG_DOUBLE;
5257 s->flags |= SLAB_TRACE;
5263 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5265 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5268 static ssize_t red_zone_store(struct kmem_cache *s,
5269 const char *buf, size_t length)
5271 if (any_slab_objects(s))
5274 s->flags &= ~SLAB_RED_ZONE;
5275 if (buf[0] == '1') {
5276 s->flags |= SLAB_RED_ZONE;
5278 calculate_sizes(s, -1);
5281 SLAB_ATTR(red_zone);
5283 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5285 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5288 static ssize_t poison_store(struct kmem_cache *s,
5289 const char *buf, size_t length)
5291 if (any_slab_objects(s))
5294 s->flags &= ~SLAB_POISON;
5295 if (buf[0] == '1') {
5296 s->flags |= SLAB_POISON;
5298 calculate_sizes(s, -1);
5303 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5305 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5308 static ssize_t store_user_store(struct kmem_cache *s,
5309 const char *buf, size_t length)
5311 if (any_slab_objects(s))
5314 s->flags &= ~SLAB_STORE_USER;
5315 if (buf[0] == '1') {
5316 s->flags &= ~__CMPXCHG_DOUBLE;
5317 s->flags |= SLAB_STORE_USER;
5319 calculate_sizes(s, -1);
5322 SLAB_ATTR(store_user);
5324 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5329 static ssize_t validate_store(struct kmem_cache *s,
5330 const char *buf, size_t length)
5334 if (buf[0] == '1') {
5335 ret = validate_slab_cache(s);
5341 SLAB_ATTR(validate);
5343 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5345 if (!(s->flags & SLAB_STORE_USER))
5347 return list_locations(s, buf, TRACK_ALLOC);
5349 SLAB_ATTR_RO(alloc_calls);
5351 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5353 if (!(s->flags & SLAB_STORE_USER))
5355 return list_locations(s, buf, TRACK_FREE);
5357 SLAB_ATTR_RO(free_calls);
5358 #endif /* CONFIG_SLUB_DEBUG */
5360 #ifdef CONFIG_FAILSLAB
5361 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5363 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5366 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5369 if (s->refcount > 1)
5372 s->flags &= ~SLAB_FAILSLAB;
5374 s->flags |= SLAB_FAILSLAB;
5377 SLAB_ATTR(failslab);
5380 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5385 static ssize_t shrink_store(struct kmem_cache *s,
5386 const char *buf, size_t length)
5389 kmem_cache_shrink_all(s);
5397 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5399 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5402 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5403 const char *buf, size_t length)
5408 err = kstrtouint(buf, 10, &ratio);
5414 s->remote_node_defrag_ratio = ratio * 10;
5418 SLAB_ATTR(remote_node_defrag_ratio);
5421 #ifdef CONFIG_SLUB_STATS
5422 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5424 unsigned long sum = 0;
5427 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5432 for_each_online_cpu(cpu) {
5433 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5439 len = sprintf(buf, "%lu", sum);
5442 for_each_online_cpu(cpu) {
5443 if (data[cpu] && len < PAGE_SIZE - 20)
5444 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5448 return len + sprintf(buf + len, "\n");
5451 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5455 for_each_online_cpu(cpu)
5456 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5459 #define STAT_ATTR(si, text) \
5460 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5462 return show_stat(s, buf, si); \
5464 static ssize_t text##_store(struct kmem_cache *s, \
5465 const char *buf, size_t length) \
5467 if (buf[0] != '0') \
5469 clear_stat(s, si); \
5474 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5475 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5476 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5477 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5478 STAT_ATTR(FREE_FROZEN, free_frozen);
5479 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5480 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5481 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5482 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5483 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5484 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5485 STAT_ATTR(FREE_SLAB, free_slab);
5486 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5487 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5488 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5489 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5490 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5491 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5492 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5493 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5494 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5495 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5496 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5497 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5498 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5499 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5500 #endif /* CONFIG_SLUB_STATS */
5502 static struct attribute *slab_attrs[] = {
5503 &slab_size_attr.attr,
5504 &object_size_attr.attr,
5505 &objs_per_slab_attr.attr,
5507 &min_partial_attr.attr,
5508 &cpu_partial_attr.attr,
5510 &objects_partial_attr.attr,
5512 &cpu_slabs_attr.attr,
5516 &hwcache_align_attr.attr,
5517 &reclaim_account_attr.attr,
5518 &destroy_by_rcu_attr.attr,
5520 &slabs_cpu_partial_attr.attr,
5521 #ifdef CONFIG_SLUB_DEBUG
5522 &total_objects_attr.attr,
5524 &sanity_checks_attr.attr,
5526 &red_zone_attr.attr,
5528 &store_user_attr.attr,
5529 &validate_attr.attr,
5530 &alloc_calls_attr.attr,
5531 &free_calls_attr.attr,
5533 #ifdef CONFIG_ZONE_DMA
5534 &cache_dma_attr.attr,
5537 &remote_node_defrag_ratio_attr.attr,
5539 #ifdef CONFIG_SLUB_STATS
5540 &alloc_fastpath_attr.attr,
5541 &alloc_slowpath_attr.attr,
5542 &free_fastpath_attr.attr,
5543 &free_slowpath_attr.attr,
5544 &free_frozen_attr.attr,
5545 &free_add_partial_attr.attr,
5546 &free_remove_partial_attr.attr,
5547 &alloc_from_partial_attr.attr,
5548 &alloc_slab_attr.attr,
5549 &alloc_refill_attr.attr,
5550 &alloc_node_mismatch_attr.attr,
5551 &free_slab_attr.attr,
5552 &cpuslab_flush_attr.attr,
5553 &deactivate_full_attr.attr,
5554 &deactivate_empty_attr.attr,
5555 &deactivate_to_head_attr.attr,
5556 &deactivate_to_tail_attr.attr,
5557 &deactivate_remote_frees_attr.attr,
5558 &deactivate_bypass_attr.attr,
5559 &order_fallback_attr.attr,
5560 &cmpxchg_double_fail_attr.attr,
5561 &cmpxchg_double_cpu_fail_attr.attr,
5562 &cpu_partial_alloc_attr.attr,
5563 &cpu_partial_free_attr.attr,
5564 &cpu_partial_node_attr.attr,
5565 &cpu_partial_drain_attr.attr,
5567 #ifdef CONFIG_FAILSLAB
5568 &failslab_attr.attr,
5570 &usersize_attr.attr,
5575 static const struct attribute_group slab_attr_group = {
5576 .attrs = slab_attrs,
5579 static ssize_t slab_attr_show(struct kobject *kobj,
5580 struct attribute *attr,
5583 struct slab_attribute *attribute;
5584 struct kmem_cache *s;
5587 attribute = to_slab_attr(attr);
5590 if (!attribute->show)
5593 err = attribute->show(s, buf);
5598 static ssize_t slab_attr_store(struct kobject *kobj,
5599 struct attribute *attr,
5600 const char *buf, size_t len)
5602 struct slab_attribute *attribute;
5603 struct kmem_cache *s;
5606 attribute = to_slab_attr(attr);
5609 if (!attribute->store)
5612 err = attribute->store(s, buf, len);
5614 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5615 struct kmem_cache *c;
5617 mutex_lock(&slab_mutex);
5618 if (s->max_attr_size < len)
5619 s->max_attr_size = len;
5622 * This is a best effort propagation, so this function's return
5623 * value will be determined by the parent cache only. This is
5624 * basically because not all attributes will have a well
5625 * defined semantics for rollbacks - most of the actions will
5626 * have permanent effects.
5628 * Returning the error value of any of the children that fail
5629 * is not 100 % defined, in the sense that users seeing the
5630 * error code won't be able to know anything about the state of
5633 * Only returning the error code for the parent cache at least
5634 * has well defined semantics. The cache being written to
5635 * directly either failed or succeeded, in which case we loop
5636 * through the descendants with best-effort propagation.
5638 for_each_memcg_cache(c, s)
5639 attribute->store(c, buf, len);
5640 mutex_unlock(&slab_mutex);
5646 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5650 char *buffer = NULL;
5651 struct kmem_cache *root_cache;
5653 if (is_root_cache(s))
5656 root_cache = s->memcg_params.root_cache;
5659 * This mean this cache had no attribute written. Therefore, no point
5660 * in copying default values around
5662 if (!root_cache->max_attr_size)
5665 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5668 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5671 if (!attr || !attr->store || !attr->show)
5675 * It is really bad that we have to allocate here, so we will
5676 * do it only as a fallback. If we actually allocate, though,
5677 * we can just use the allocated buffer until the end.
5679 * Most of the slub attributes will tend to be very small in
5680 * size, but sysfs allows buffers up to a page, so they can
5681 * theoretically happen.
5685 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5686 !IS_ENABLED(CONFIG_SLUB_STATS))
5689 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5690 if (WARN_ON(!buffer))
5695 len = attr->show(root_cache, buf);
5697 attr->store(s, buf, len);
5701 free_page((unsigned long)buffer);
5702 #endif /* CONFIG_MEMCG */
5705 static void kmem_cache_release(struct kobject *k)
5707 slab_kmem_cache_release(to_slab(k));
5710 static const struct sysfs_ops slab_sysfs_ops = {
5711 .show = slab_attr_show,
5712 .store = slab_attr_store,
5715 static struct kobj_type slab_ktype = {
5716 .sysfs_ops = &slab_sysfs_ops,
5717 .release = kmem_cache_release,
5720 static struct kset *slab_kset;
5722 static inline struct kset *cache_kset(struct kmem_cache *s)
5725 if (!is_root_cache(s))
5726 return s->memcg_params.root_cache->memcg_kset;
5731 #define ID_STR_LENGTH 64
5733 /* Create a unique string id for a slab cache:
5735 * Format :[flags-]size
5737 static char *create_unique_id(struct kmem_cache *s)
5739 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5746 * First flags affecting slabcache operations. We will only
5747 * get here for aliasable slabs so we do not need to support
5748 * too many flags. The flags here must cover all flags that
5749 * are matched during merging to guarantee that the id is
5752 if (s->flags & SLAB_CACHE_DMA)
5754 if (s->flags & SLAB_CACHE_DMA32)
5756 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5758 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5760 if (s->flags & SLAB_ACCOUNT)
5764 p += sprintf(p, "%07u", s->size);
5766 BUG_ON(p > name + ID_STR_LENGTH - 1);
5770 static void sysfs_slab_remove_workfn(struct work_struct *work)
5772 struct kmem_cache *s =
5773 container_of(work, struct kmem_cache, kobj_remove_work);
5775 if (!s->kobj.state_in_sysfs)
5777 * For a memcg cache, this may be called during
5778 * deactivation and again on shutdown. Remove only once.
5779 * A cache is never shut down before deactivation is
5780 * complete, so no need to worry about synchronization.
5785 kset_unregister(s->memcg_kset);
5788 kobject_put(&s->kobj);
5791 static int sysfs_slab_add(struct kmem_cache *s)
5795 struct kset *kset = cache_kset(s);
5796 int unmergeable = slab_unmergeable(s);
5798 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5801 kobject_init(&s->kobj, &slab_ktype);
5805 if (!unmergeable && disable_higher_order_debug &&
5806 (slub_debug & DEBUG_METADATA_FLAGS))
5811 * Slabcache can never be merged so we can use the name proper.
5812 * This is typically the case for debug situations. In that
5813 * case we can catch duplicate names easily.
5815 sysfs_remove_link(&slab_kset->kobj, s->name);
5819 * Create a unique name for the slab as a target
5822 name = create_unique_id(s);
5825 s->kobj.kset = kset;
5826 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5828 kobject_put(&s->kobj);
5832 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5837 if (is_root_cache(s) && memcg_sysfs_enabled) {
5838 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5839 if (!s->memcg_kset) {
5847 /* Setup first alias */
5848 sysfs_slab_alias(s, s->name);
5855 kobject_del(&s->kobj);
5859 static void sysfs_slab_remove(struct kmem_cache *s)
5861 if (slab_state < FULL)
5863 * Sysfs has not been setup yet so no need to remove the
5868 kobject_get(&s->kobj);
5869 schedule_work(&s->kobj_remove_work);
5872 void sysfs_slab_unlink(struct kmem_cache *s)
5874 if (slab_state >= FULL)
5875 kobject_del(&s->kobj);
5878 void sysfs_slab_release(struct kmem_cache *s)
5880 if (slab_state >= FULL)
5881 kobject_put(&s->kobj);
5885 * Need to buffer aliases during bootup until sysfs becomes
5886 * available lest we lose that information.
5888 struct saved_alias {
5889 struct kmem_cache *s;
5891 struct saved_alias *next;
5894 static struct saved_alias *alias_list;
5896 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5898 struct saved_alias *al;
5900 if (slab_state == FULL) {
5902 * If we have a leftover link then remove it.
5904 sysfs_remove_link(&slab_kset->kobj, name);
5905 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5908 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5914 al->next = alias_list;
5919 static int __init slab_sysfs_init(void)
5921 struct kmem_cache *s;
5924 mutex_lock(&slab_mutex);
5926 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5928 mutex_unlock(&slab_mutex);
5929 pr_err("Cannot register slab subsystem.\n");
5935 list_for_each_entry(s, &slab_caches, list) {
5936 err = sysfs_slab_add(s);
5938 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5942 while (alias_list) {
5943 struct saved_alias *al = alias_list;
5945 alias_list = alias_list->next;
5946 err = sysfs_slab_alias(al->s, al->name);
5948 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5953 mutex_unlock(&slab_mutex);
5958 __initcall(slab_sysfs_init);
5959 #endif /* CONFIG_SYSFS */
5962 * The /proc/slabinfo ABI
5964 #ifdef CONFIG_SLUB_DEBUG
5965 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5967 unsigned long nr_slabs = 0;
5968 unsigned long nr_objs = 0;
5969 unsigned long nr_free = 0;
5971 struct kmem_cache_node *n;
5973 for_each_kmem_cache_node(s, node, n) {
5974 nr_slabs += node_nr_slabs(n);
5975 nr_objs += node_nr_objs(n);
5976 nr_free += count_partial(n, count_free);
5979 sinfo->active_objs = nr_objs - nr_free;
5980 sinfo->num_objs = nr_objs;
5981 sinfo->active_slabs = nr_slabs;
5982 sinfo->num_slabs = nr_slabs;
5983 sinfo->objects_per_slab = oo_objects(s->oo);
5984 sinfo->cache_order = oo_order(s->oo);
5987 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5991 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5992 size_t count, loff_t *ppos)
5996 #endif /* CONFIG_SLUB_DEBUG */