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 * Overloading of page flags that are otherwise used for LRU management.
98 * PageActive The slab is frozen and exempt from list processing.
99 * This means that the slab is dedicated to a purpose
100 * such as satisfying allocations for a specific
101 * processor. Objects may be freed in the slab while
102 * it is frozen but slab_free will then skip the usual
103 * list operations. It is up to the processor holding
104 * the slab to integrate the slab into the slab lists
105 * when the slab is no longer needed.
107 * One use of this flag is to mark slabs that are
108 * used for allocations. Then such a slab becomes a cpu
109 * slab. The cpu slab may be equipped with an additional
110 * freelist that allows lockless access to
111 * free objects in addition to the regular freelist
112 * that requires the slab lock.
114 * PageError Slab requires special handling due to debug
115 * options set. This moves slab handling out of
116 * the fast path and disables lockless freelists.
119 static inline int kmem_cache_debug(struct kmem_cache *s)
121 #ifdef CONFIG_SLUB_DEBUG
122 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
128 void *fixup_red_left(struct kmem_cache *s, void *p)
130 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
131 p += s->red_left_pad;
136 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
138 #ifdef CONFIG_SLUB_CPU_PARTIAL
139 return !kmem_cache_debug(s);
146 * Issues still to be resolved:
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
156 /* Enable to log cmpxchg failures */
157 #undef SLUB_DEBUG_CMPXCHG
160 * Mininum number of partial slabs. These will be left on the partial
161 * lists even if they are empty. kmem_cache_shrink may reclaim them.
163 #define MIN_PARTIAL 5
166 * Maximum number of desirable partial slabs.
167 * The existence of more partial slabs makes kmem_cache_shrink
168 * sort the partial list by the number of objects in use.
170 #define MAX_PARTIAL 10
172 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_STORE_USER)
176 * These debug flags cannot use CMPXCHG because there might be consistency
177 * issues when checking or reading debug information
179 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
184 * Debugging flags that require metadata to be stored in the slab. These get
185 * disabled when slub_debug=O is used and a cache's min order increases with
188 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
191 #define OO_MASK ((1 << OO_SHIFT) - 1)
192 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
194 /* Internal SLUB flags */
196 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
197 /* Use cmpxchg_double */
198 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
201 * Tracking user of a slab.
203 #define TRACK_ADDRS_COUNT 16
205 unsigned long addr; /* Called from address */
206 #ifdef CONFIG_STACKTRACE
207 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
209 int cpu; /* Was running on cpu */
210 int pid; /* Pid context */
211 unsigned long when; /* When did the operation occur */
214 enum track_item { TRACK_ALLOC, TRACK_FREE };
217 static int sysfs_slab_add(struct kmem_cache *);
218 static int sysfs_slab_alias(struct kmem_cache *, const char *);
219 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
220 static void sysfs_slab_remove(struct kmem_cache *s);
222 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
226 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
229 static inline void stat(const struct kmem_cache *s, enum stat_item si)
231 #ifdef CONFIG_SLUB_STATS
233 * The rmw is racy on a preemptible kernel but this is acceptable, so
234 * avoid this_cpu_add()'s irq-disable overhead.
236 raw_cpu_inc(s->cpu_slab->stat[si]);
240 /********************************************************************
241 * Core slab cache functions
242 *******************************************************************/
245 * Returns freelist pointer (ptr). With hardening, this is obfuscated
246 * with an XOR of the address where the pointer is held and a per-cache
249 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
250 unsigned long ptr_addr)
252 #ifdef CONFIG_SLAB_FREELIST_HARDENED
254 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
255 * Normally, this doesn't cause any issues, as both set_freepointer()
256 * and get_freepointer() are called with a pointer with the same tag.
257 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
258 * example, when __free_slub() iterates over objects in a cache, it
259 * passes untagged pointers to check_object(). check_object() in turns
260 * calls get_freepointer() with an untagged pointer, which causes the
261 * freepointer to be restored incorrectly.
263 return (void *)((unsigned long)ptr ^ s->random ^
264 (unsigned long)kasan_reset_tag((void *)ptr_addr));
270 /* Returns the freelist pointer recorded at location ptr_addr. */
271 static inline void *freelist_dereference(const struct kmem_cache *s,
274 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
275 (unsigned long)ptr_addr);
278 static inline void *get_freepointer(struct kmem_cache *s, void *object)
280 return freelist_dereference(s, object + s->offset);
283 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
285 prefetch(object + s->offset);
288 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
290 unsigned long freepointer_addr;
293 if (!debug_pagealloc_enabled())
294 return get_freepointer(s, object);
296 freepointer_addr = (unsigned long)object + s->offset;
297 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
298 return freelist_ptr(s, p, freepointer_addr);
301 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
303 unsigned long freeptr_addr = (unsigned long)object + s->offset;
305 #ifdef CONFIG_SLAB_FREELIST_HARDENED
306 BUG_ON(object == fp); /* naive detection of double free or corruption */
309 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
312 /* Loop over all objects in a slab */
313 #define for_each_object(__p, __s, __addr, __objects) \
314 for (__p = fixup_red_left(__s, __addr); \
315 __p < (__addr) + (__objects) * (__s)->size; \
318 /* Determine object index from a given position */
319 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
321 return (kasan_reset_tag(p) - addr) / s->size;
324 static inline unsigned int order_objects(unsigned int order, unsigned int size)
326 return ((unsigned int)PAGE_SIZE << order) / size;
329 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size)
339 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
341 return x.x >> OO_SHIFT;
344 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
346 return x.x & OO_MASK;
350 * Per slab locking using the pagelock
352 static __always_inline void slab_lock(struct page *page)
354 VM_BUG_ON_PAGE(PageTail(page), page);
355 bit_spin_lock(PG_locked, &page->flags);
358 static __always_inline void slab_unlock(struct page *page)
360 VM_BUG_ON_PAGE(PageTail(page), page);
361 __bit_spin_unlock(PG_locked, &page->flags);
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 page->counters = counters_new;
393 stat(s, CMPXCHG_DOUBLE_FAIL);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
402 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
419 local_irq_save(flags);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 page->counters = counters_new;
426 local_irq_restore(flags);
430 local_irq_restore(flags);
434 stat(s, CMPXCHG_DOUBLE_FAIL);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453 void *addr = page_address(page);
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
459 static inline unsigned int size_from_object(struct kmem_cache *s)
461 if (s->flags & SLAB_RED_ZONE)
462 return s->size - s->red_left_pad;
467 static inline void *restore_red_left(struct kmem_cache *s, void *p)
469 if (s->flags & SLAB_RED_ZONE)
470 p -= s->red_left_pad;
478 #if defined(CONFIG_SLUB_DEBUG_ON)
479 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
481 static slab_flags_t slub_debug;
484 static char *slub_debug_slabs;
485 static int disable_higher_order_debug;
488 * slub is about to manipulate internal object metadata. This memory lies
489 * outside the range of the allocated object, so accessing it would normally
490 * be reported by kasan as a bounds error. metadata_access_enable() is used
491 * to tell kasan that these accesses are OK.
493 static inline void metadata_access_enable(void)
495 kasan_disable_current();
498 static inline void metadata_access_disable(void)
500 kasan_enable_current();
507 /* Verify that a pointer has an address that is valid within a slab page */
508 static inline int check_valid_pointer(struct kmem_cache *s,
509 struct page *page, void *object)
516 base = page_address(page);
517 object = kasan_reset_tag(object);
518 object = restore_red_left(s, object);
519 if (object < base || object >= base + page->objects * s->size ||
520 (object - base) % s->size) {
527 static void print_section(char *level, char *text, u8 *addr,
530 metadata_access_enable();
531 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
533 metadata_access_disable();
536 static struct track *get_track(struct kmem_cache *s, void *object,
537 enum track_item alloc)
542 p = object + s->offset + sizeof(void *);
544 p = object + s->inuse;
549 static void set_track(struct kmem_cache *s, void *object,
550 enum track_item alloc, unsigned long addr)
552 struct track *p = get_track(s, object, alloc);
555 #ifdef CONFIG_STACKTRACE
556 unsigned int nr_entries;
558 metadata_access_enable();
559 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
560 metadata_access_disable();
562 if (nr_entries < TRACK_ADDRS_COUNT)
563 p->addrs[nr_entries] = 0;
566 p->cpu = smp_processor_id();
567 p->pid = current->pid;
570 memset(p, 0, sizeof(struct track));
574 static void init_tracking(struct kmem_cache *s, void *object)
576 if (!(s->flags & SLAB_STORE_USER))
579 set_track(s, object, TRACK_FREE, 0UL);
580 set_track(s, object, TRACK_ALLOC, 0UL);
583 static void print_track(const char *s, struct track *t, unsigned long pr_time)
588 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
589 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
590 #ifdef CONFIG_STACKTRACE
593 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
595 pr_err("\t%pS\n", (void *)t->addrs[i]);
602 static void print_tracking(struct kmem_cache *s, void *object)
604 unsigned long pr_time = jiffies;
605 if (!(s->flags & SLAB_STORE_USER))
608 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
609 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
612 static void print_page_info(struct page *page)
614 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
615 page, page->objects, page->inuse, page->freelist, page->flags);
619 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
621 struct va_format vaf;
627 pr_err("=============================================================================\n");
628 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
629 pr_err("-----------------------------------------------------------------------------\n\n");
631 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
635 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
637 struct va_format vaf;
643 pr_err("FIX %s: %pV\n", s->name, &vaf);
647 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
649 unsigned int off; /* Offset of last byte */
650 u8 *addr = page_address(page);
652 print_tracking(s, p);
654 print_page_info(page);
656 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
657 p, p - addr, get_freepointer(s, p));
659 if (s->flags & SLAB_RED_ZONE)
660 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
662 else if (p > addr + 16)
663 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
665 print_section(KERN_ERR, "Object ", p,
666 min_t(unsigned int, s->object_size, PAGE_SIZE));
667 if (s->flags & SLAB_RED_ZONE)
668 print_section(KERN_ERR, "Redzone ", p + s->object_size,
669 s->inuse - s->object_size);
672 off = s->offset + sizeof(void *);
676 if (s->flags & SLAB_STORE_USER)
677 off += 2 * sizeof(struct track);
679 off += kasan_metadata_size(s);
681 if (off != size_from_object(s))
682 /* Beginning of the filler is the free pointer */
683 print_section(KERN_ERR, "Padding ", p + off,
684 size_from_object(s) - off);
689 void object_err(struct kmem_cache *s, struct page *page,
690 u8 *object, char *reason)
692 slab_bug(s, "%s", reason);
693 print_trailer(s, page, object);
696 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
697 const char *fmt, ...)
703 vsnprintf(buf, sizeof(buf), fmt, args);
705 slab_bug(s, "%s", buf);
706 print_page_info(page);
710 static void init_object(struct kmem_cache *s, void *object, u8 val)
714 if (s->flags & SLAB_RED_ZONE)
715 memset(p - s->red_left_pad, val, s->red_left_pad);
717 if (s->flags & __OBJECT_POISON) {
718 memset(p, POISON_FREE, s->object_size - 1);
719 p[s->object_size - 1] = POISON_END;
722 if (s->flags & SLAB_RED_ZONE)
723 memset(p + s->object_size, val, s->inuse - s->object_size);
726 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
727 void *from, void *to)
729 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
730 memset(from, data, to - from);
733 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
734 u8 *object, char *what,
735 u8 *start, unsigned int value, unsigned int bytes)
740 metadata_access_enable();
741 fault = memchr_inv(start, value, bytes);
742 metadata_access_disable();
747 while (end > fault && end[-1] == value)
750 slab_bug(s, "%s overwritten", what);
751 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
752 fault, end - 1, fault[0], value);
753 print_trailer(s, page, object);
755 restore_bytes(s, what, value, fault, end);
763 * Bytes of the object to be managed.
764 * If the freepointer may overlay the object then the free
765 * pointer is the first word of the object.
767 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
770 * object + s->object_size
771 * Padding to reach word boundary. This is also used for Redzoning.
772 * Padding is extended by another word if Redzoning is enabled and
773 * object_size == inuse.
775 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
776 * 0xcc (RED_ACTIVE) for objects in use.
779 * Meta data starts here.
781 * A. Free pointer (if we cannot overwrite object on free)
782 * B. Tracking data for SLAB_STORE_USER
783 * C. Padding to reach required alignment boundary or at mininum
784 * one word if debugging is on to be able to detect writes
785 * before the word boundary.
787 * Padding is done using 0x5a (POISON_INUSE)
790 * Nothing is used beyond s->size.
792 * If slabcaches are merged then the object_size and inuse boundaries are mostly
793 * ignored. And therefore no slab options that rely on these boundaries
794 * may be used with merged slabcaches.
797 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
799 unsigned long off = s->inuse; /* The end of info */
802 /* Freepointer is placed after the object. */
803 off += sizeof(void *);
805 if (s->flags & SLAB_STORE_USER)
806 /* We also have user information there */
807 off += 2 * sizeof(struct track);
809 off += kasan_metadata_size(s);
811 if (size_from_object(s) == off)
814 return check_bytes_and_report(s, page, p, "Object padding",
815 p + off, POISON_INUSE, size_from_object(s) - off);
818 /* Check the pad bytes at the end of a slab page */
819 static int slab_pad_check(struct kmem_cache *s, struct page *page)
828 if (!(s->flags & SLAB_POISON))
831 start = page_address(page);
832 length = page_size(page);
833 end = start + length;
834 remainder = length % s->size;
838 pad = end - remainder;
839 metadata_access_enable();
840 fault = memchr_inv(pad, POISON_INUSE, remainder);
841 metadata_access_disable();
844 while (end > fault && end[-1] == POISON_INUSE)
847 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
848 print_section(KERN_ERR, "Padding ", pad, remainder);
850 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
854 static int check_object(struct kmem_cache *s, struct page *page,
855 void *object, u8 val)
858 u8 *endobject = object + s->object_size;
860 if (s->flags & SLAB_RED_ZONE) {
861 if (!check_bytes_and_report(s, page, object, "Redzone",
862 object - s->red_left_pad, val, s->red_left_pad))
865 if (!check_bytes_and_report(s, page, object, "Redzone",
866 endobject, val, s->inuse - s->object_size))
869 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
870 check_bytes_and_report(s, page, p, "Alignment padding",
871 endobject, POISON_INUSE,
872 s->inuse - s->object_size);
876 if (s->flags & SLAB_POISON) {
877 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
878 (!check_bytes_and_report(s, page, p, "Poison", p,
879 POISON_FREE, s->object_size - 1) ||
880 !check_bytes_and_report(s, page, p, "Poison",
881 p + s->object_size - 1, POISON_END, 1)))
884 * check_pad_bytes cleans up on its own.
886 check_pad_bytes(s, page, p);
889 if (!s->offset && val == SLUB_RED_ACTIVE)
891 * Object and freepointer overlap. Cannot check
892 * freepointer while object is allocated.
896 /* Check free pointer validity */
897 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
898 object_err(s, page, p, "Freepointer corrupt");
900 * No choice but to zap it and thus lose the remainder
901 * of the free objects in this slab. May cause
902 * another error because the object count is now wrong.
904 set_freepointer(s, p, NULL);
910 static int check_slab(struct kmem_cache *s, struct page *page)
914 VM_BUG_ON(!irqs_disabled());
916 if (!PageSlab(page)) {
917 slab_err(s, page, "Not a valid slab page");
921 maxobj = order_objects(compound_order(page), s->size);
922 if (page->objects > maxobj) {
923 slab_err(s, page, "objects %u > max %u",
924 page->objects, maxobj);
927 if (page->inuse > page->objects) {
928 slab_err(s, page, "inuse %u > max %u",
929 page->inuse, page->objects);
932 /* Slab_pad_check fixes things up after itself */
933 slab_pad_check(s, page);
938 * Determine if a certain object on a page is on the freelist. Must hold the
939 * slab lock to guarantee that the chains are in a consistent state.
941 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
949 while (fp && nr <= page->objects) {
952 if (!check_valid_pointer(s, page, fp)) {
954 object_err(s, page, object,
955 "Freechain corrupt");
956 set_freepointer(s, object, NULL);
958 slab_err(s, page, "Freepointer corrupt");
959 page->freelist = NULL;
960 page->inuse = page->objects;
961 slab_fix(s, "Freelist cleared");
967 fp = get_freepointer(s, object);
971 max_objects = order_objects(compound_order(page), s->size);
972 if (max_objects > MAX_OBJS_PER_PAGE)
973 max_objects = MAX_OBJS_PER_PAGE;
975 if (page->objects != max_objects) {
976 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
977 page->objects, max_objects);
978 page->objects = max_objects;
979 slab_fix(s, "Number of objects adjusted.");
981 if (page->inuse != page->objects - nr) {
982 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
983 page->inuse, page->objects - nr);
984 page->inuse = page->objects - nr;
985 slab_fix(s, "Object count adjusted.");
987 return search == NULL;
990 static void trace(struct kmem_cache *s, struct page *page, void *object,
993 if (s->flags & SLAB_TRACE) {
994 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
996 alloc ? "alloc" : "free",
1001 print_section(KERN_INFO, "Object ", (void *)object,
1009 * Tracking of fully allocated slabs for debugging purposes.
1011 static void add_full(struct kmem_cache *s,
1012 struct kmem_cache_node *n, struct page *page)
1014 if (!(s->flags & SLAB_STORE_USER))
1017 lockdep_assert_held(&n->list_lock);
1018 list_add(&page->slab_list, &n->full);
1021 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1023 if (!(s->flags & SLAB_STORE_USER))
1026 lockdep_assert_held(&n->list_lock);
1027 list_del(&page->slab_list);
1030 /* Tracking of the number of slabs for debugging purposes */
1031 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1033 struct kmem_cache_node *n = get_node(s, node);
1035 return atomic_long_read(&n->nr_slabs);
1038 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1040 return atomic_long_read(&n->nr_slabs);
1043 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1045 struct kmem_cache_node *n = get_node(s, node);
1048 * May be called early in order to allocate a slab for the
1049 * kmem_cache_node structure. Solve the chicken-egg
1050 * dilemma by deferring the increment of the count during
1051 * bootstrap (see early_kmem_cache_node_alloc).
1054 atomic_long_inc(&n->nr_slabs);
1055 atomic_long_add(objects, &n->total_objects);
1058 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1060 struct kmem_cache_node *n = get_node(s, node);
1062 atomic_long_dec(&n->nr_slabs);
1063 atomic_long_sub(objects, &n->total_objects);
1066 /* Object debug checks for alloc/free paths */
1067 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1070 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1073 init_object(s, object, SLUB_RED_INACTIVE);
1074 init_tracking(s, object);
1078 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1080 if (!(s->flags & SLAB_POISON))
1083 metadata_access_enable();
1084 memset(addr, POISON_INUSE, page_size(page));
1085 metadata_access_disable();
1088 static inline int alloc_consistency_checks(struct kmem_cache *s,
1089 struct page *page, void *object)
1091 if (!check_slab(s, page))
1094 if (!check_valid_pointer(s, page, object)) {
1095 object_err(s, page, object, "Freelist Pointer check fails");
1099 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1105 static noinline int alloc_debug_processing(struct kmem_cache *s,
1107 void *object, unsigned long addr)
1109 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1110 if (!alloc_consistency_checks(s, page, object))
1114 /* Success perform special debug activities for allocs */
1115 if (s->flags & SLAB_STORE_USER)
1116 set_track(s, object, TRACK_ALLOC, addr);
1117 trace(s, page, object, 1);
1118 init_object(s, object, SLUB_RED_ACTIVE);
1122 if (PageSlab(page)) {
1124 * If this is a slab page then lets do the best we can
1125 * to avoid issues in the future. Marking all objects
1126 * as used avoids touching the remaining objects.
1128 slab_fix(s, "Marking all objects used");
1129 page->inuse = page->objects;
1130 page->freelist = NULL;
1135 static inline int free_consistency_checks(struct kmem_cache *s,
1136 struct page *page, void *object, unsigned long addr)
1138 if (!check_valid_pointer(s, page, object)) {
1139 slab_err(s, page, "Invalid object pointer 0x%p", object);
1143 if (on_freelist(s, page, object)) {
1144 object_err(s, page, object, "Object already free");
1148 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1151 if (unlikely(s != page->slab_cache)) {
1152 if (!PageSlab(page)) {
1153 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1155 } else if (!page->slab_cache) {
1156 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1160 object_err(s, page, object,
1161 "page slab pointer corrupt.");
1167 /* Supports checking bulk free of a constructed freelist */
1168 static noinline int free_debug_processing(
1169 struct kmem_cache *s, struct page *page,
1170 void *head, void *tail, int bulk_cnt,
1173 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1174 void *object = head;
1176 unsigned long uninitialized_var(flags);
1179 spin_lock_irqsave(&n->list_lock, flags);
1182 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1183 if (!check_slab(s, page))
1190 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1191 if (!free_consistency_checks(s, page, object, addr))
1195 if (s->flags & SLAB_STORE_USER)
1196 set_track(s, object, TRACK_FREE, addr);
1197 trace(s, page, object, 0);
1198 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1199 init_object(s, object, SLUB_RED_INACTIVE);
1201 /* Reached end of constructed freelist yet? */
1202 if (object != tail) {
1203 object = get_freepointer(s, object);
1209 if (cnt != bulk_cnt)
1210 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1214 spin_unlock_irqrestore(&n->list_lock, flags);
1216 slab_fix(s, "Object at 0x%p not freed", object);
1220 static int __init setup_slub_debug(char *str)
1222 slub_debug = DEBUG_DEFAULT_FLAGS;
1223 if (*str++ != '=' || !*str)
1225 * No options specified. Switch on full debugging.
1231 * No options but restriction on slabs. This means full
1232 * debugging for slabs matching a pattern.
1239 * Switch off all debugging measures.
1244 * Determine which debug features should be switched on
1246 for (; *str && *str != ','; str++) {
1247 switch (tolower(*str)) {
1249 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1252 slub_debug |= SLAB_RED_ZONE;
1255 slub_debug |= SLAB_POISON;
1258 slub_debug |= SLAB_STORE_USER;
1261 slub_debug |= SLAB_TRACE;
1264 slub_debug |= SLAB_FAILSLAB;
1268 * Avoid enabling debugging on caches if its minimum
1269 * order would increase as a result.
1271 disable_higher_order_debug = 1;
1274 pr_err("slub_debug option '%c' unknown. skipped\n",
1281 slub_debug_slabs = str + 1;
1283 if ((static_branch_unlikely(&init_on_alloc) ||
1284 static_branch_unlikely(&init_on_free)) &&
1285 (slub_debug & SLAB_POISON))
1286 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1290 __setup("slub_debug", setup_slub_debug);
1293 * kmem_cache_flags - apply debugging options to the cache
1294 * @object_size: the size of an object without meta data
1295 * @flags: flags to set
1296 * @name: name of the cache
1297 * @ctor: constructor function
1299 * Debug option(s) are applied to @flags. In addition to the debug
1300 * option(s), if a slab name (or multiple) is specified i.e.
1301 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1302 * then only the select slabs will receive the debug option(s).
1304 slab_flags_t kmem_cache_flags(unsigned int object_size,
1305 slab_flags_t flags, const char *name,
1306 void (*ctor)(void *))
1311 /* If slub_debug = 0, it folds into the if conditional. */
1312 if (!slub_debug_slabs)
1313 return flags | slub_debug;
1316 iter = slub_debug_slabs;
1321 end = strchrnul(iter, ',');
1323 glob = strnchr(iter, end - iter, '*');
1325 cmplen = glob - iter;
1327 cmplen = max_t(size_t, len, (end - iter));
1329 if (!strncmp(name, iter, cmplen)) {
1330 flags |= slub_debug;
1341 #else /* !CONFIG_SLUB_DEBUG */
1342 static inline void setup_object_debug(struct kmem_cache *s,
1343 struct page *page, void *object) {}
1345 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1347 static inline int alloc_debug_processing(struct kmem_cache *s,
1348 struct page *page, void *object, unsigned long addr) { return 0; }
1350 static inline int free_debug_processing(
1351 struct kmem_cache *s, struct page *page,
1352 void *head, void *tail, int bulk_cnt,
1353 unsigned long addr) { return 0; }
1355 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1357 static inline int check_object(struct kmem_cache *s, struct page *page,
1358 void *object, u8 val) { return 1; }
1359 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1360 struct page *page) {}
1361 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1362 struct page *page) {}
1363 slab_flags_t kmem_cache_flags(unsigned int object_size,
1364 slab_flags_t flags, const char *name,
1365 void (*ctor)(void *))
1369 #define slub_debug 0
1371 #define disable_higher_order_debug 0
1373 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1375 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1377 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1379 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1382 #endif /* CONFIG_SLUB_DEBUG */
1385 * Hooks for other subsystems that check memory allocations. In a typical
1386 * production configuration these hooks all should produce no code at all.
1388 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1390 ptr = kasan_kmalloc_large(ptr, size, flags);
1391 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1392 kmemleak_alloc(ptr, size, 1, flags);
1396 static __always_inline void kfree_hook(void *x)
1399 kasan_kfree_large(x, _RET_IP_);
1402 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1404 kmemleak_free_recursive(x, s->flags);
1407 * Trouble is that we may no longer disable interrupts in the fast path
1408 * So in order to make the debug calls that expect irqs to be
1409 * disabled we need to disable interrupts temporarily.
1411 #ifdef CONFIG_LOCKDEP
1413 unsigned long flags;
1415 local_irq_save(flags);
1416 debug_check_no_locks_freed(x, s->object_size);
1417 local_irq_restore(flags);
1420 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1421 debug_check_no_obj_freed(x, s->object_size);
1423 /* KASAN might put x into memory quarantine, delaying its reuse */
1424 return kasan_slab_free(s, x, _RET_IP_);
1427 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1428 void **head, void **tail)
1433 void *old_tail = *tail ? *tail : *head;
1436 if (slab_want_init_on_free(s)) {
1441 next = get_freepointer(s, object);
1443 * Clear the object and the metadata, but don't touch
1446 memset(object, 0, s->object_size);
1447 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1449 memset((char *)object + s->inuse, 0,
1450 s->size - s->inuse - rsize);
1451 set_freepointer(s, object, p);
1453 } while (object != old_tail);
1457 * Compiler cannot detect this function can be removed if slab_free_hook()
1458 * evaluates to nothing. Thus, catch all relevant config debug options here.
1460 #if defined(CONFIG_LOCKDEP) || \
1461 defined(CONFIG_DEBUG_KMEMLEAK) || \
1462 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1463 defined(CONFIG_KASAN)
1467 /* Head and tail of the reconstructed freelist */
1473 next = get_freepointer(s, object);
1474 /* If object's reuse doesn't have to be delayed */
1475 if (!slab_free_hook(s, object)) {
1476 /* Move object to the new freelist */
1477 set_freepointer(s, object, *head);
1482 } while (object != old_tail);
1487 return *head != NULL;
1493 static void *setup_object(struct kmem_cache *s, struct page *page,
1496 setup_object_debug(s, page, object);
1497 object = kasan_init_slab_obj(s, object);
1498 if (unlikely(s->ctor)) {
1499 kasan_unpoison_object_data(s, object);
1501 kasan_poison_object_data(s, object);
1507 * Slab allocation and freeing
1509 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1510 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1513 unsigned int order = oo_order(oo);
1515 if (node == NUMA_NO_NODE)
1516 page = alloc_pages(flags, order);
1518 page = __alloc_pages_node(node, flags, order);
1520 if (page && charge_slab_page(page, flags, order, s)) {
1521 __free_pages(page, order);
1528 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1529 /* Pre-initialize the random sequence cache */
1530 static int init_cache_random_seq(struct kmem_cache *s)
1532 unsigned int count = oo_objects(s->oo);
1535 /* Bailout if already initialised */
1539 err = cache_random_seq_create(s, count, GFP_KERNEL);
1541 pr_err("SLUB: Unable to initialize free list for %s\n",
1546 /* Transform to an offset on the set of pages */
1547 if (s->random_seq) {
1550 for (i = 0; i < count; i++)
1551 s->random_seq[i] *= s->size;
1556 /* Initialize each random sequence freelist per cache */
1557 static void __init init_freelist_randomization(void)
1559 struct kmem_cache *s;
1561 mutex_lock(&slab_mutex);
1563 list_for_each_entry(s, &slab_caches, list)
1564 init_cache_random_seq(s);
1566 mutex_unlock(&slab_mutex);
1569 /* Get the next entry on the pre-computed freelist randomized */
1570 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1571 unsigned long *pos, void *start,
1572 unsigned long page_limit,
1573 unsigned long freelist_count)
1578 * If the target page allocation failed, the number of objects on the
1579 * page might be smaller than the usual size defined by the cache.
1582 idx = s->random_seq[*pos];
1584 if (*pos >= freelist_count)
1586 } while (unlikely(idx >= page_limit));
1588 return (char *)start + idx;
1591 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1592 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1597 unsigned long idx, pos, page_limit, freelist_count;
1599 if (page->objects < 2 || !s->random_seq)
1602 freelist_count = oo_objects(s->oo);
1603 pos = get_random_int() % freelist_count;
1605 page_limit = page->objects * s->size;
1606 start = fixup_red_left(s, page_address(page));
1608 /* First entry is used as the base of the freelist */
1609 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1611 cur = setup_object(s, page, cur);
1612 page->freelist = cur;
1614 for (idx = 1; idx < page->objects; idx++) {
1615 next = next_freelist_entry(s, page, &pos, start, page_limit,
1617 next = setup_object(s, page, next);
1618 set_freepointer(s, cur, next);
1621 set_freepointer(s, cur, NULL);
1626 static inline int init_cache_random_seq(struct kmem_cache *s)
1630 static inline void init_freelist_randomization(void) { }
1631 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1635 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1637 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1640 struct kmem_cache_order_objects oo = s->oo;
1642 void *start, *p, *next;
1646 flags &= gfp_allowed_mask;
1648 if (gfpflags_allow_blocking(flags))
1651 flags |= s->allocflags;
1654 * Let the initial higher-order allocation fail under memory pressure
1655 * so we fall-back to the minimum order allocation.
1657 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1658 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1659 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1661 page = alloc_slab_page(s, alloc_gfp, node, oo);
1662 if (unlikely(!page)) {
1666 * Allocation may have failed due to fragmentation.
1667 * Try a lower order alloc if possible
1669 page = alloc_slab_page(s, alloc_gfp, node, oo);
1670 if (unlikely(!page))
1672 stat(s, ORDER_FALLBACK);
1675 page->objects = oo_objects(oo);
1677 page->slab_cache = s;
1678 __SetPageSlab(page);
1679 if (page_is_pfmemalloc(page))
1680 SetPageSlabPfmemalloc(page);
1682 kasan_poison_slab(page);
1684 start = page_address(page);
1686 setup_page_debug(s, page, start);
1688 shuffle = shuffle_freelist(s, page);
1691 start = fixup_red_left(s, start);
1692 start = setup_object(s, page, start);
1693 page->freelist = start;
1694 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1696 next = setup_object(s, page, next);
1697 set_freepointer(s, p, next);
1700 set_freepointer(s, p, NULL);
1703 page->inuse = page->objects;
1707 if (gfpflags_allow_blocking(flags))
1708 local_irq_disable();
1712 inc_slabs_node(s, page_to_nid(page), page->objects);
1717 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1719 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1720 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1721 flags &= ~GFP_SLAB_BUG_MASK;
1722 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1723 invalid_mask, &invalid_mask, flags, &flags);
1727 return allocate_slab(s,
1728 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1731 static void __free_slab(struct kmem_cache *s, struct page *page)
1733 int order = compound_order(page);
1734 int pages = 1 << order;
1736 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1739 slab_pad_check(s, page);
1740 for_each_object(p, s, page_address(page),
1742 check_object(s, page, p, SLUB_RED_INACTIVE);
1745 __ClearPageSlabPfmemalloc(page);
1746 __ClearPageSlab(page);
1748 page->mapping = NULL;
1749 if (current->reclaim_state)
1750 current->reclaim_state->reclaimed_slab += pages;
1751 uncharge_slab_page(page, order, s);
1752 __free_pages(page, order);
1755 static void rcu_free_slab(struct rcu_head *h)
1757 struct page *page = container_of(h, struct page, rcu_head);
1759 __free_slab(page->slab_cache, page);
1762 static void free_slab(struct kmem_cache *s, struct page *page)
1764 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1765 call_rcu(&page->rcu_head, rcu_free_slab);
1767 __free_slab(s, page);
1770 static void discard_slab(struct kmem_cache *s, struct page *page)
1772 dec_slabs_node(s, page_to_nid(page), page->objects);
1777 * Management of partially allocated slabs.
1780 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1783 if (tail == DEACTIVATE_TO_TAIL)
1784 list_add_tail(&page->slab_list, &n->partial);
1786 list_add(&page->slab_list, &n->partial);
1789 static inline void add_partial(struct kmem_cache_node *n,
1790 struct page *page, int tail)
1792 lockdep_assert_held(&n->list_lock);
1793 __add_partial(n, page, tail);
1796 static inline void remove_partial(struct kmem_cache_node *n,
1799 lockdep_assert_held(&n->list_lock);
1800 list_del(&page->slab_list);
1805 * Remove slab from the partial list, freeze it and
1806 * return the pointer to the freelist.
1808 * Returns a list of objects or NULL if it fails.
1810 static inline void *acquire_slab(struct kmem_cache *s,
1811 struct kmem_cache_node *n, struct page *page,
1812 int mode, int *objects)
1815 unsigned long counters;
1818 lockdep_assert_held(&n->list_lock);
1821 * Zap the freelist and set the frozen bit.
1822 * The old freelist is the list of objects for the
1823 * per cpu allocation list.
1825 freelist = page->freelist;
1826 counters = page->counters;
1827 new.counters = counters;
1828 *objects = new.objects - new.inuse;
1830 new.inuse = page->objects;
1831 new.freelist = NULL;
1833 new.freelist = freelist;
1836 VM_BUG_ON(new.frozen);
1839 if (!__cmpxchg_double_slab(s, page,
1841 new.freelist, new.counters,
1845 remove_partial(n, page);
1850 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1851 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1854 * Try to allocate a partial slab from a specific node.
1856 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1857 struct kmem_cache_cpu *c, gfp_t flags)
1859 struct page *page, *page2;
1860 void *object = NULL;
1861 unsigned int available = 0;
1865 * Racy check. If we mistakenly see no partial slabs then we
1866 * just allocate an empty slab. If we mistakenly try to get a
1867 * partial slab and there is none available then get_partials()
1870 if (!n || !n->nr_partial)
1873 spin_lock(&n->list_lock);
1874 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1877 if (!pfmemalloc_match(page, flags))
1880 t = acquire_slab(s, n, page, object == NULL, &objects);
1884 available += objects;
1887 stat(s, ALLOC_FROM_PARTIAL);
1890 put_cpu_partial(s, page, 0);
1891 stat(s, CPU_PARTIAL_NODE);
1893 if (!kmem_cache_has_cpu_partial(s)
1894 || available > slub_cpu_partial(s) / 2)
1898 spin_unlock(&n->list_lock);
1903 * Get a page from somewhere. Search in increasing NUMA distances.
1905 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1906 struct kmem_cache_cpu *c)
1909 struct zonelist *zonelist;
1912 enum zone_type high_zoneidx = gfp_zone(flags);
1914 unsigned int cpuset_mems_cookie;
1917 * The defrag ratio allows a configuration of the tradeoffs between
1918 * inter node defragmentation and node local allocations. A lower
1919 * defrag_ratio increases the tendency to do local allocations
1920 * instead of attempting to obtain partial slabs from other nodes.
1922 * If the defrag_ratio is set to 0 then kmalloc() always
1923 * returns node local objects. If the ratio is higher then kmalloc()
1924 * may return off node objects because partial slabs are obtained
1925 * from other nodes and filled up.
1927 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1928 * (which makes defrag_ratio = 1000) then every (well almost)
1929 * allocation will first attempt to defrag slab caches on other nodes.
1930 * This means scanning over all nodes to look for partial slabs which
1931 * may be expensive if we do it every time we are trying to find a slab
1932 * with available objects.
1934 if (!s->remote_node_defrag_ratio ||
1935 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1939 cpuset_mems_cookie = read_mems_allowed_begin();
1940 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1941 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1942 struct kmem_cache_node *n;
1944 n = get_node(s, zone_to_nid(zone));
1946 if (n && cpuset_zone_allowed(zone, flags) &&
1947 n->nr_partial > s->min_partial) {
1948 object = get_partial_node(s, n, c, flags);
1951 * Don't check read_mems_allowed_retry()
1952 * here - if mems_allowed was updated in
1953 * parallel, that was a harmless race
1954 * between allocation and the cpuset
1961 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1962 #endif /* CONFIG_NUMA */
1967 * Get a partial page, lock it and return it.
1969 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1970 struct kmem_cache_cpu *c)
1973 int searchnode = node;
1975 if (node == NUMA_NO_NODE)
1976 searchnode = numa_mem_id();
1977 else if (!node_present_pages(node))
1978 searchnode = node_to_mem_node(node);
1980 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1981 if (object || node != NUMA_NO_NODE)
1984 return get_any_partial(s, flags, c);
1987 #ifdef CONFIG_PREEMPT
1989 * Calculate the next globally unique transaction for disambiguiation
1990 * during cmpxchg. The transactions start with the cpu number and are then
1991 * incremented by CONFIG_NR_CPUS.
1993 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1996 * No preemption supported therefore also no need to check for
2002 static inline unsigned long next_tid(unsigned long tid)
2004 return tid + TID_STEP;
2007 #ifdef SLUB_DEBUG_CMPXCHG
2008 static inline unsigned int tid_to_cpu(unsigned long tid)
2010 return tid % TID_STEP;
2013 static inline unsigned long tid_to_event(unsigned long tid)
2015 return tid / TID_STEP;
2019 static inline unsigned int init_tid(int cpu)
2024 static inline void note_cmpxchg_failure(const char *n,
2025 const struct kmem_cache *s, unsigned long tid)
2027 #ifdef SLUB_DEBUG_CMPXCHG
2028 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2030 pr_info("%s %s: cmpxchg redo ", n, s->name);
2032 #ifdef CONFIG_PREEMPT
2033 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2034 pr_warn("due to cpu change %d -> %d\n",
2035 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2038 if (tid_to_event(tid) != tid_to_event(actual_tid))
2039 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2040 tid_to_event(tid), tid_to_event(actual_tid));
2042 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2043 actual_tid, tid, next_tid(tid));
2045 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2048 static void init_kmem_cache_cpus(struct kmem_cache *s)
2052 for_each_possible_cpu(cpu)
2053 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2057 * Remove the cpu slab
2059 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2060 void *freelist, struct kmem_cache_cpu *c)
2062 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2063 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2065 enum slab_modes l = M_NONE, m = M_NONE;
2067 int tail = DEACTIVATE_TO_HEAD;
2071 if (page->freelist) {
2072 stat(s, DEACTIVATE_REMOTE_FREES);
2073 tail = DEACTIVATE_TO_TAIL;
2077 * Stage one: Free all available per cpu objects back
2078 * to the page freelist while it is still frozen. Leave the
2081 * There is no need to take the list->lock because the page
2084 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2086 unsigned long counters;
2089 prior = page->freelist;
2090 counters = page->counters;
2091 set_freepointer(s, freelist, prior);
2092 new.counters = counters;
2094 VM_BUG_ON(!new.frozen);
2096 } while (!__cmpxchg_double_slab(s, page,
2098 freelist, new.counters,
2099 "drain percpu freelist"));
2101 freelist = nextfree;
2105 * Stage two: Ensure that the page is unfrozen while the
2106 * list presence reflects the actual number of objects
2109 * We setup the list membership and then perform a cmpxchg
2110 * with the count. If there is a mismatch then the page
2111 * is not unfrozen but the page is on the wrong list.
2113 * Then we restart the process which may have to remove
2114 * the page from the list that we just put it on again
2115 * because the number of objects in the slab may have
2120 old.freelist = page->freelist;
2121 old.counters = page->counters;
2122 VM_BUG_ON(!old.frozen);
2124 /* Determine target state of the slab */
2125 new.counters = old.counters;
2128 set_freepointer(s, freelist, old.freelist);
2129 new.freelist = freelist;
2131 new.freelist = old.freelist;
2135 if (!new.inuse && n->nr_partial >= s->min_partial)
2137 else if (new.freelist) {
2142 * Taking the spinlock removes the possibility
2143 * that acquire_slab() will see a slab page that
2146 spin_lock(&n->list_lock);
2150 if (kmem_cache_debug(s) && !lock) {
2153 * This also ensures that the scanning of full
2154 * slabs from diagnostic functions will not see
2157 spin_lock(&n->list_lock);
2163 remove_partial(n, page);
2164 else if (l == M_FULL)
2165 remove_full(s, n, page);
2168 add_partial(n, page, tail);
2169 else if (m == M_FULL)
2170 add_full(s, n, page);
2174 if (!__cmpxchg_double_slab(s, page,
2175 old.freelist, old.counters,
2176 new.freelist, new.counters,
2181 spin_unlock(&n->list_lock);
2185 else if (m == M_FULL)
2186 stat(s, DEACTIVATE_FULL);
2187 else if (m == M_FREE) {
2188 stat(s, DEACTIVATE_EMPTY);
2189 discard_slab(s, page);
2198 * Unfreeze all the cpu partial slabs.
2200 * This function must be called with interrupts disabled
2201 * for the cpu using c (or some other guarantee must be there
2202 * to guarantee no concurrent accesses).
2204 static void unfreeze_partials(struct kmem_cache *s,
2205 struct kmem_cache_cpu *c)
2207 #ifdef CONFIG_SLUB_CPU_PARTIAL
2208 struct kmem_cache_node *n = NULL, *n2 = NULL;
2209 struct page *page, *discard_page = NULL;
2211 while ((page = c->partial)) {
2215 c->partial = page->next;
2217 n2 = get_node(s, page_to_nid(page));
2220 spin_unlock(&n->list_lock);
2223 spin_lock(&n->list_lock);
2228 old.freelist = page->freelist;
2229 old.counters = page->counters;
2230 VM_BUG_ON(!old.frozen);
2232 new.counters = old.counters;
2233 new.freelist = old.freelist;
2237 } while (!__cmpxchg_double_slab(s, page,
2238 old.freelist, old.counters,
2239 new.freelist, new.counters,
2240 "unfreezing slab"));
2242 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2243 page->next = discard_page;
2244 discard_page = page;
2246 add_partial(n, page, DEACTIVATE_TO_TAIL);
2247 stat(s, FREE_ADD_PARTIAL);
2252 spin_unlock(&n->list_lock);
2254 while (discard_page) {
2255 page = discard_page;
2256 discard_page = discard_page->next;
2258 stat(s, DEACTIVATE_EMPTY);
2259 discard_slab(s, page);
2262 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2266 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2267 * partial page slot if available.
2269 * If we did not find a slot then simply move all the partials to the
2270 * per node partial list.
2272 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2274 #ifdef CONFIG_SLUB_CPU_PARTIAL
2275 struct page *oldpage;
2283 oldpage = this_cpu_read(s->cpu_slab->partial);
2286 pobjects = oldpage->pobjects;
2287 pages = oldpage->pages;
2288 if (drain && pobjects > s->cpu_partial) {
2289 unsigned long flags;
2291 * partial array is full. Move the existing
2292 * set to the per node partial list.
2294 local_irq_save(flags);
2295 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2296 local_irq_restore(flags);
2300 stat(s, CPU_PARTIAL_DRAIN);
2305 pobjects += page->objects - page->inuse;
2307 page->pages = pages;
2308 page->pobjects = pobjects;
2309 page->next = oldpage;
2311 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2313 if (unlikely(!s->cpu_partial)) {
2314 unsigned long flags;
2316 local_irq_save(flags);
2317 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2318 local_irq_restore(flags);
2321 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2324 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2326 stat(s, CPUSLAB_FLUSH);
2327 deactivate_slab(s, c->page, c->freelist, c);
2329 c->tid = next_tid(c->tid);
2335 * Called from IPI handler with interrupts disabled.
2337 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2339 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2344 unfreeze_partials(s, c);
2347 static void flush_cpu_slab(void *d)
2349 struct kmem_cache *s = d;
2351 __flush_cpu_slab(s, smp_processor_id());
2354 static bool has_cpu_slab(int cpu, void *info)
2356 struct kmem_cache *s = info;
2357 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2359 return c->page || slub_percpu_partial(c);
2362 static void flush_all(struct kmem_cache *s)
2364 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2368 * Use the cpu notifier to insure that the cpu slabs are flushed when
2371 static int slub_cpu_dead(unsigned int cpu)
2373 struct kmem_cache *s;
2374 unsigned long flags;
2376 mutex_lock(&slab_mutex);
2377 list_for_each_entry(s, &slab_caches, list) {
2378 local_irq_save(flags);
2379 __flush_cpu_slab(s, cpu);
2380 local_irq_restore(flags);
2382 mutex_unlock(&slab_mutex);
2387 * Check if the objects in a per cpu structure fit numa
2388 * locality expectations.
2390 static inline int node_match(struct page *page, int node)
2393 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2399 #ifdef CONFIG_SLUB_DEBUG
2400 static int count_free(struct page *page)
2402 return page->objects - page->inuse;
2405 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2407 return atomic_long_read(&n->total_objects);
2409 #endif /* CONFIG_SLUB_DEBUG */
2411 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2412 static unsigned long count_partial(struct kmem_cache_node *n,
2413 int (*get_count)(struct page *))
2415 unsigned long flags;
2416 unsigned long x = 0;
2419 spin_lock_irqsave(&n->list_lock, flags);
2420 list_for_each_entry(page, &n->partial, slab_list)
2421 x += get_count(page);
2422 spin_unlock_irqrestore(&n->list_lock, flags);
2425 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2427 static noinline void
2428 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2430 #ifdef CONFIG_SLUB_DEBUG
2431 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2432 DEFAULT_RATELIMIT_BURST);
2434 struct kmem_cache_node *n;
2436 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2439 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2440 nid, gfpflags, &gfpflags);
2441 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2442 s->name, s->object_size, s->size, oo_order(s->oo),
2445 if (oo_order(s->min) > get_order(s->object_size))
2446 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2449 for_each_kmem_cache_node(s, node, n) {
2450 unsigned long nr_slabs;
2451 unsigned long nr_objs;
2452 unsigned long nr_free;
2454 nr_free = count_partial(n, count_free);
2455 nr_slabs = node_nr_slabs(n);
2456 nr_objs = node_nr_objs(n);
2458 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2459 node, nr_slabs, nr_objs, nr_free);
2464 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2465 int node, struct kmem_cache_cpu **pc)
2468 struct kmem_cache_cpu *c = *pc;
2471 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2473 freelist = get_partial(s, flags, node, c);
2478 page = new_slab(s, flags, node);
2480 c = raw_cpu_ptr(s->cpu_slab);
2485 * No other reference to the page yet so we can
2486 * muck around with it freely without cmpxchg
2488 freelist = page->freelist;
2489 page->freelist = NULL;
2491 stat(s, ALLOC_SLAB);
2499 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2501 if (unlikely(PageSlabPfmemalloc(page)))
2502 return gfp_pfmemalloc_allowed(gfpflags);
2508 * Check the page->freelist of a page and either transfer the freelist to the
2509 * per cpu freelist or deactivate the page.
2511 * The page is still frozen if the return value is not NULL.
2513 * If this function returns NULL then the page has been unfrozen.
2515 * This function must be called with interrupt disabled.
2517 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2520 unsigned long counters;
2524 freelist = page->freelist;
2525 counters = page->counters;
2527 new.counters = counters;
2528 VM_BUG_ON(!new.frozen);
2530 new.inuse = page->objects;
2531 new.frozen = freelist != NULL;
2533 } while (!__cmpxchg_double_slab(s, page,
2542 * Slow path. The lockless freelist is empty or we need to perform
2545 * Processing is still very fast if new objects have been freed to the
2546 * regular freelist. In that case we simply take over the regular freelist
2547 * as the lockless freelist and zap the regular freelist.
2549 * If that is not working then we fall back to the partial lists. We take the
2550 * first element of the freelist as the object to allocate now and move the
2551 * rest of the freelist to the lockless freelist.
2553 * And if we were unable to get a new slab from the partial slab lists then
2554 * we need to allocate a new slab. This is the slowest path since it involves
2555 * a call to the page allocator and the setup of a new slab.
2557 * Version of __slab_alloc to use when we know that interrupts are
2558 * already disabled (which is the case for bulk allocation).
2560 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2561 unsigned long addr, struct kmem_cache_cpu *c)
2571 if (unlikely(!node_match(page, node))) {
2572 int searchnode = node;
2574 if (node != NUMA_NO_NODE && !node_present_pages(node))
2575 searchnode = node_to_mem_node(node);
2577 if (unlikely(!node_match(page, searchnode))) {
2578 stat(s, ALLOC_NODE_MISMATCH);
2579 deactivate_slab(s, page, c->freelist, c);
2585 * By rights, we should be searching for a slab page that was
2586 * PFMEMALLOC but right now, we are losing the pfmemalloc
2587 * information when the page leaves the per-cpu allocator
2589 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2590 deactivate_slab(s, page, c->freelist, c);
2594 /* must check again c->freelist in case of cpu migration or IRQ */
2595 freelist = c->freelist;
2599 freelist = get_freelist(s, page);
2603 stat(s, DEACTIVATE_BYPASS);
2607 stat(s, ALLOC_REFILL);
2611 * freelist is pointing to the list of objects to be used.
2612 * page is pointing to the page from which the objects are obtained.
2613 * That page must be frozen for per cpu allocations to work.
2615 VM_BUG_ON(!c->page->frozen);
2616 c->freelist = get_freepointer(s, freelist);
2617 c->tid = next_tid(c->tid);
2622 if (slub_percpu_partial(c)) {
2623 page = c->page = slub_percpu_partial(c);
2624 slub_set_percpu_partial(c, page);
2625 stat(s, CPU_PARTIAL_ALLOC);
2629 freelist = new_slab_objects(s, gfpflags, node, &c);
2631 if (unlikely(!freelist)) {
2632 slab_out_of_memory(s, gfpflags, node);
2637 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2640 /* Only entered in the debug case */
2641 if (kmem_cache_debug(s) &&
2642 !alloc_debug_processing(s, page, freelist, addr))
2643 goto new_slab; /* Slab failed checks. Next slab needed */
2645 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2650 * Another one that disabled interrupt and compensates for possible
2651 * cpu changes by refetching the per cpu area pointer.
2653 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2654 unsigned long addr, struct kmem_cache_cpu *c)
2657 unsigned long flags;
2659 local_irq_save(flags);
2660 #ifdef CONFIG_PREEMPT
2662 * We may have been preempted and rescheduled on a different
2663 * cpu before disabling interrupts. Need to reload cpu area
2666 c = this_cpu_ptr(s->cpu_slab);
2669 p = ___slab_alloc(s, gfpflags, node, addr, c);
2670 local_irq_restore(flags);
2675 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2676 * have the fastpath folded into their functions. So no function call
2677 * overhead for requests that can be satisfied on the fastpath.
2679 * The fastpath works by first checking if the lockless freelist can be used.
2680 * If not then __slab_alloc is called for slow processing.
2682 * Otherwise we can simply pick the next object from the lockless free list.
2684 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2685 gfp_t gfpflags, int node, unsigned long addr)
2688 struct kmem_cache_cpu *c;
2692 s = slab_pre_alloc_hook(s, gfpflags);
2697 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2698 * enabled. We may switch back and forth between cpus while
2699 * reading from one cpu area. That does not matter as long
2700 * as we end up on the original cpu again when doing the cmpxchg.
2702 * We should guarantee that tid and kmem_cache are retrieved on
2703 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2704 * to check if it is matched or not.
2707 tid = this_cpu_read(s->cpu_slab->tid);
2708 c = raw_cpu_ptr(s->cpu_slab);
2709 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2710 unlikely(tid != READ_ONCE(c->tid)));
2713 * Irqless object alloc/free algorithm used here depends on sequence
2714 * of fetching cpu_slab's data. tid should be fetched before anything
2715 * on c to guarantee that object and page associated with previous tid
2716 * won't be used with current tid. If we fetch tid first, object and
2717 * page could be one associated with next tid and our alloc/free
2718 * request will be failed. In this case, we will retry. So, no problem.
2723 * The transaction ids are globally unique per cpu and per operation on
2724 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2725 * occurs on the right processor and that there was no operation on the
2726 * linked list in between.
2729 object = c->freelist;
2731 if (unlikely(!object || !node_match(page, node))) {
2732 object = __slab_alloc(s, gfpflags, node, addr, c);
2733 stat(s, ALLOC_SLOWPATH);
2735 void *next_object = get_freepointer_safe(s, object);
2738 * The cmpxchg will only match if there was no additional
2739 * operation and if we are on the right processor.
2741 * The cmpxchg does the following atomically (without lock
2743 * 1. Relocate first pointer to the current per cpu area.
2744 * 2. Verify that tid and freelist have not been changed
2745 * 3. If they were not changed replace tid and freelist
2747 * Since this is without lock semantics the protection is only
2748 * against code executing on this cpu *not* from access by
2751 if (unlikely(!this_cpu_cmpxchg_double(
2752 s->cpu_slab->freelist, s->cpu_slab->tid,
2754 next_object, next_tid(tid)))) {
2756 note_cmpxchg_failure("slab_alloc", s, tid);
2759 prefetch_freepointer(s, next_object);
2760 stat(s, ALLOC_FASTPATH);
2763 * If the object has been wiped upon free, make sure it's fully
2764 * initialized by zeroing out freelist pointer.
2766 if (unlikely(slab_want_init_on_free(s)) && object)
2767 memset(object + s->offset, 0, sizeof(void *));
2769 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2770 memset(object, 0, s->object_size);
2772 slab_post_alloc_hook(s, gfpflags, 1, &object);
2777 static __always_inline void *slab_alloc(struct kmem_cache *s,
2778 gfp_t gfpflags, unsigned long addr)
2780 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2783 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2785 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2787 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2792 EXPORT_SYMBOL(kmem_cache_alloc);
2794 #ifdef CONFIG_TRACING
2795 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2797 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2798 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2799 ret = kasan_kmalloc(s, ret, size, gfpflags);
2802 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2806 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2808 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2810 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2811 s->object_size, s->size, gfpflags, node);
2815 EXPORT_SYMBOL(kmem_cache_alloc_node);
2817 #ifdef CONFIG_TRACING
2818 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2820 int node, size_t size)
2822 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2824 trace_kmalloc_node(_RET_IP_, ret,
2825 size, s->size, gfpflags, node);
2827 ret = kasan_kmalloc(s, ret, size, gfpflags);
2830 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2832 #endif /* CONFIG_NUMA */
2835 * Slow path handling. This may still be called frequently since objects
2836 * have a longer lifetime than the cpu slabs in most processing loads.
2838 * So we still attempt to reduce cache line usage. Just take the slab
2839 * lock and free the item. If there is no additional partial page
2840 * handling required then we can return immediately.
2842 static void __slab_free(struct kmem_cache *s, struct page *page,
2843 void *head, void *tail, int cnt,
2850 unsigned long counters;
2851 struct kmem_cache_node *n = NULL;
2852 unsigned long uninitialized_var(flags);
2854 stat(s, FREE_SLOWPATH);
2856 if (kmem_cache_debug(s) &&
2857 !free_debug_processing(s, page, head, tail, cnt, addr))
2862 spin_unlock_irqrestore(&n->list_lock, flags);
2865 prior = page->freelist;
2866 counters = page->counters;
2867 set_freepointer(s, tail, prior);
2868 new.counters = counters;
2869 was_frozen = new.frozen;
2871 if ((!new.inuse || !prior) && !was_frozen) {
2873 if (kmem_cache_has_cpu_partial(s) && !prior) {
2876 * Slab was on no list before and will be
2878 * We can defer the list move and instead
2883 } else { /* Needs to be taken off a list */
2885 n = get_node(s, page_to_nid(page));
2887 * Speculatively acquire the list_lock.
2888 * If the cmpxchg does not succeed then we may
2889 * drop the list_lock without any processing.
2891 * Otherwise the list_lock will synchronize with
2892 * other processors updating the list of slabs.
2894 spin_lock_irqsave(&n->list_lock, flags);
2899 } while (!cmpxchg_double_slab(s, page,
2907 * If we just froze the page then put it onto the
2908 * per cpu partial list.
2910 if (new.frozen && !was_frozen) {
2911 put_cpu_partial(s, page, 1);
2912 stat(s, CPU_PARTIAL_FREE);
2915 * The list lock was not taken therefore no list
2916 * activity can be necessary.
2919 stat(s, FREE_FROZEN);
2923 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2927 * Objects left in the slab. If it was not on the partial list before
2930 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2931 remove_full(s, n, page);
2932 add_partial(n, page, DEACTIVATE_TO_TAIL);
2933 stat(s, FREE_ADD_PARTIAL);
2935 spin_unlock_irqrestore(&n->list_lock, flags);
2941 * Slab on the partial list.
2943 remove_partial(n, page);
2944 stat(s, FREE_REMOVE_PARTIAL);
2946 /* Slab must be on the full list */
2947 remove_full(s, n, page);
2950 spin_unlock_irqrestore(&n->list_lock, flags);
2952 discard_slab(s, page);
2956 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2957 * can perform fastpath freeing without additional function calls.
2959 * The fastpath is only possible if we are freeing to the current cpu slab
2960 * of this processor. This typically the case if we have just allocated
2963 * If fastpath is not possible then fall back to __slab_free where we deal
2964 * with all sorts of special processing.
2966 * Bulk free of a freelist with several objects (all pointing to the
2967 * same page) possible by specifying head and tail ptr, plus objects
2968 * count (cnt). Bulk free indicated by tail pointer being set.
2970 static __always_inline void do_slab_free(struct kmem_cache *s,
2971 struct page *page, void *head, void *tail,
2972 int cnt, unsigned long addr)
2974 void *tail_obj = tail ? : head;
2975 struct kmem_cache_cpu *c;
2979 * Determine the currently cpus per cpu slab.
2980 * The cpu may change afterward. However that does not matter since
2981 * data is retrieved via this pointer. If we are on the same cpu
2982 * during the cmpxchg then the free will succeed.
2985 tid = this_cpu_read(s->cpu_slab->tid);
2986 c = raw_cpu_ptr(s->cpu_slab);
2987 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2988 unlikely(tid != READ_ONCE(c->tid)));
2990 /* Same with comment on barrier() in slab_alloc_node() */
2993 if (likely(page == c->page)) {
2994 set_freepointer(s, tail_obj, c->freelist);
2996 if (unlikely(!this_cpu_cmpxchg_double(
2997 s->cpu_slab->freelist, s->cpu_slab->tid,
2999 head, next_tid(tid)))) {
3001 note_cmpxchg_failure("slab_free", s, tid);
3004 stat(s, FREE_FASTPATH);
3006 __slab_free(s, page, head, tail_obj, cnt, addr);
3010 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3011 void *head, void *tail, int cnt,
3015 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3016 * to remove objects, whose reuse must be delayed.
3018 if (slab_free_freelist_hook(s, &head, &tail))
3019 do_slab_free(s, page, head, tail, cnt, addr);
3022 #ifdef CONFIG_KASAN_GENERIC
3023 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3025 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3029 void kmem_cache_free(struct kmem_cache *s, void *x)
3031 s = cache_from_obj(s, x);
3034 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3035 trace_kmem_cache_free(_RET_IP_, x);
3037 EXPORT_SYMBOL(kmem_cache_free);
3039 struct detached_freelist {
3044 struct kmem_cache *s;
3048 * This function progressively scans the array with free objects (with
3049 * a limited look ahead) and extract objects belonging to the same
3050 * page. It builds a detached freelist directly within the given
3051 * page/objects. This can happen without any need for
3052 * synchronization, because the objects are owned by running process.
3053 * The freelist is build up as a single linked list in the objects.
3054 * The idea is, that this detached freelist can then be bulk
3055 * transferred to the real freelist(s), but only requiring a single
3056 * synchronization primitive. Look ahead in the array is limited due
3057 * to performance reasons.
3060 int build_detached_freelist(struct kmem_cache *s, size_t size,
3061 void **p, struct detached_freelist *df)
3063 size_t first_skipped_index = 0;
3068 /* Always re-init detached_freelist */
3073 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3074 } while (!object && size);
3079 page = virt_to_head_page(object);
3081 /* Handle kalloc'ed objects */
3082 if (unlikely(!PageSlab(page))) {
3083 BUG_ON(!PageCompound(page));
3085 __free_pages(page, compound_order(page));
3086 p[size] = NULL; /* mark object processed */
3089 /* Derive kmem_cache from object */
3090 df->s = page->slab_cache;
3092 df->s = cache_from_obj(s, object); /* Support for memcg */
3095 /* Start new detached freelist */
3097 set_freepointer(df->s, object, NULL);
3099 df->freelist = object;
3100 p[size] = NULL; /* mark object processed */
3106 continue; /* Skip processed objects */
3108 /* df->page is always set at this point */
3109 if (df->page == virt_to_head_page(object)) {
3110 /* Opportunity build freelist */
3111 set_freepointer(df->s, object, df->freelist);
3112 df->freelist = object;
3114 p[size] = NULL; /* mark object processed */
3119 /* Limit look ahead search */
3123 if (!first_skipped_index)
3124 first_skipped_index = size + 1;
3127 return first_skipped_index;
3130 /* Note that interrupts must be enabled when calling this function. */
3131 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3137 struct detached_freelist df;
3139 size = build_detached_freelist(s, size, p, &df);
3143 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3144 } while (likely(size));
3146 EXPORT_SYMBOL(kmem_cache_free_bulk);
3148 /* Note that interrupts must be enabled when calling this function. */
3149 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3152 struct kmem_cache_cpu *c;
3155 /* memcg and kmem_cache debug support */
3156 s = slab_pre_alloc_hook(s, flags);
3160 * Drain objects in the per cpu slab, while disabling local
3161 * IRQs, which protects against PREEMPT and interrupts
3162 * handlers invoking normal fastpath.
3164 local_irq_disable();
3165 c = this_cpu_ptr(s->cpu_slab);
3167 for (i = 0; i < size; i++) {
3168 void *object = c->freelist;
3170 if (unlikely(!object)) {
3172 * Invoking slow path likely have side-effect
3173 * of re-populating per CPU c->freelist
3175 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3177 if (unlikely(!p[i]))
3180 c = this_cpu_ptr(s->cpu_slab);
3181 continue; /* goto for-loop */
3183 c->freelist = get_freepointer(s, object);
3186 c->tid = next_tid(c->tid);
3189 /* Clear memory outside IRQ disabled fastpath loop */
3190 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3193 for (j = 0; j < i; j++)
3194 memset(p[j], 0, s->object_size);
3197 /* memcg and kmem_cache debug support */
3198 slab_post_alloc_hook(s, flags, size, p);
3202 slab_post_alloc_hook(s, flags, i, p);
3203 __kmem_cache_free_bulk(s, i, p);
3206 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3210 * Object placement in a slab is made very easy because we always start at
3211 * offset 0. If we tune the size of the object to the alignment then we can
3212 * get the required alignment by putting one properly sized object after
3215 * Notice that the allocation order determines the sizes of the per cpu
3216 * caches. Each processor has always one slab available for allocations.
3217 * Increasing the allocation order reduces the number of times that slabs
3218 * must be moved on and off the partial lists and is therefore a factor in
3223 * Mininum / Maximum order of slab pages. This influences locking overhead
3224 * and slab fragmentation. A higher order reduces the number of partial slabs
3225 * and increases the number of allocations possible without having to
3226 * take the list_lock.
3228 static unsigned int slub_min_order;
3229 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3230 static unsigned int slub_min_objects;
3233 * Calculate the order of allocation given an slab object size.
3235 * The order of allocation has significant impact on performance and other
3236 * system components. Generally order 0 allocations should be preferred since
3237 * order 0 does not cause fragmentation in the page allocator. Larger objects
3238 * be problematic to put into order 0 slabs because there may be too much
3239 * unused space left. We go to a higher order if more than 1/16th of the slab
3242 * In order to reach satisfactory performance we must ensure that a minimum
3243 * number of objects is in one slab. Otherwise we may generate too much
3244 * activity on the partial lists which requires taking the list_lock. This is
3245 * less a concern for large slabs though which are rarely used.
3247 * slub_max_order specifies the order where we begin to stop considering the
3248 * number of objects in a slab as critical. If we reach slub_max_order then
3249 * we try to keep the page order as low as possible. So we accept more waste
3250 * of space in favor of a small page order.
3252 * Higher order allocations also allow the placement of more objects in a
3253 * slab and thereby reduce object handling overhead. If the user has
3254 * requested a higher mininum order then we start with that one instead of
3255 * the smallest order which will fit the object.
3257 static inline unsigned int slab_order(unsigned int size,
3258 unsigned int min_objects, unsigned int max_order,
3259 unsigned int fract_leftover)
3261 unsigned int min_order = slub_min_order;
3264 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3265 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3267 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3268 order <= max_order; order++) {
3270 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3273 rem = slab_size % size;
3275 if (rem <= slab_size / fract_leftover)
3282 static inline int calculate_order(unsigned int size)
3285 unsigned int min_objects;
3286 unsigned int max_objects;
3289 * Attempt to find best configuration for a slab. This
3290 * works by first attempting to generate a layout with
3291 * the best configuration and backing off gradually.
3293 * First we increase the acceptable waste in a slab. Then
3294 * we reduce the minimum objects required in a slab.
3296 min_objects = slub_min_objects;
3298 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3299 max_objects = order_objects(slub_max_order, size);
3300 min_objects = min(min_objects, max_objects);
3302 while (min_objects > 1) {
3303 unsigned int fraction;
3306 while (fraction >= 4) {
3307 order = slab_order(size, min_objects,
3308 slub_max_order, fraction);
3309 if (order <= slub_max_order)
3317 * We were unable to place multiple objects in a slab. Now
3318 * lets see if we can place a single object there.
3320 order = slab_order(size, 1, slub_max_order, 1);
3321 if (order <= slub_max_order)
3325 * Doh this slab cannot be placed using slub_max_order.
3327 order = slab_order(size, 1, MAX_ORDER, 1);
3328 if (order < MAX_ORDER)
3334 init_kmem_cache_node(struct kmem_cache_node *n)
3337 spin_lock_init(&n->list_lock);
3338 INIT_LIST_HEAD(&n->partial);
3339 #ifdef CONFIG_SLUB_DEBUG
3340 atomic_long_set(&n->nr_slabs, 0);
3341 atomic_long_set(&n->total_objects, 0);
3342 INIT_LIST_HEAD(&n->full);
3346 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3348 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3349 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3352 * Must align to double word boundary for the double cmpxchg
3353 * instructions to work; see __pcpu_double_call_return_bool().
3355 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3356 2 * sizeof(void *));
3361 init_kmem_cache_cpus(s);
3366 static struct kmem_cache *kmem_cache_node;
3369 * No kmalloc_node yet so do it by hand. We know that this is the first
3370 * slab on the node for this slabcache. There are no concurrent accesses
3373 * Note that this function only works on the kmem_cache_node
3374 * when allocating for the kmem_cache_node. This is used for bootstrapping
3375 * memory on a fresh node that has no slab structures yet.
3377 static void early_kmem_cache_node_alloc(int node)
3380 struct kmem_cache_node *n;
3382 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3384 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3387 if (page_to_nid(page) != node) {
3388 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3389 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3394 #ifdef CONFIG_SLUB_DEBUG
3395 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3396 init_tracking(kmem_cache_node, n);
3398 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3400 page->freelist = get_freepointer(kmem_cache_node, n);
3403 kmem_cache_node->node[node] = n;
3404 init_kmem_cache_node(n);
3405 inc_slabs_node(kmem_cache_node, node, page->objects);
3408 * No locks need to be taken here as it has just been
3409 * initialized and there is no concurrent access.
3411 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3414 static void free_kmem_cache_nodes(struct kmem_cache *s)
3417 struct kmem_cache_node *n;
3419 for_each_kmem_cache_node(s, node, n) {
3420 s->node[node] = NULL;
3421 kmem_cache_free(kmem_cache_node, n);
3425 void __kmem_cache_release(struct kmem_cache *s)
3427 cache_random_seq_destroy(s);
3428 free_percpu(s->cpu_slab);
3429 free_kmem_cache_nodes(s);
3432 static int init_kmem_cache_nodes(struct kmem_cache *s)
3436 for_each_node_state(node, N_NORMAL_MEMORY) {
3437 struct kmem_cache_node *n;
3439 if (slab_state == DOWN) {
3440 early_kmem_cache_node_alloc(node);
3443 n = kmem_cache_alloc_node(kmem_cache_node,
3447 free_kmem_cache_nodes(s);
3451 init_kmem_cache_node(n);
3457 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3459 if (min < MIN_PARTIAL)
3461 else if (min > MAX_PARTIAL)
3463 s->min_partial = min;
3466 static void set_cpu_partial(struct kmem_cache *s)
3468 #ifdef CONFIG_SLUB_CPU_PARTIAL
3470 * cpu_partial determined the maximum number of objects kept in the
3471 * per cpu partial lists of a processor.
3473 * Per cpu partial lists mainly contain slabs that just have one
3474 * object freed. If they are used for allocation then they can be
3475 * filled up again with minimal effort. The slab will never hit the
3476 * per node partial lists and therefore no locking will be required.
3478 * This setting also determines
3480 * A) The number of objects from per cpu partial slabs dumped to the
3481 * per node list when we reach the limit.
3482 * B) The number of objects in cpu partial slabs to extract from the
3483 * per node list when we run out of per cpu objects. We only fetch
3484 * 50% to keep some capacity around for frees.
3486 if (!kmem_cache_has_cpu_partial(s))
3488 else if (s->size >= PAGE_SIZE)
3490 else if (s->size >= 1024)
3492 else if (s->size >= 256)
3493 s->cpu_partial = 13;
3495 s->cpu_partial = 30;
3500 * calculate_sizes() determines the order and the distribution of data within
3503 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3505 slab_flags_t flags = s->flags;
3506 unsigned int size = s->object_size;
3510 * Round up object size to the next word boundary. We can only
3511 * place the free pointer at word boundaries and this determines
3512 * the possible location of the free pointer.
3514 size = ALIGN(size, sizeof(void *));
3516 #ifdef CONFIG_SLUB_DEBUG
3518 * Determine if we can poison the object itself. If the user of
3519 * the slab may touch the object after free or before allocation
3520 * then we should never poison the object itself.
3522 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3524 s->flags |= __OBJECT_POISON;
3526 s->flags &= ~__OBJECT_POISON;
3530 * If we are Redzoning then check if there is some space between the
3531 * end of the object and the free pointer. If not then add an
3532 * additional word to have some bytes to store Redzone information.
3534 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3535 size += sizeof(void *);
3539 * With that we have determined the number of bytes in actual use
3540 * by the object. This is the potential offset to the free pointer.
3544 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3547 * Relocate free pointer after the object if it is not
3548 * permitted to overwrite the first word of the object on
3551 * This is the case if we do RCU, have a constructor or
3552 * destructor or are poisoning the objects.
3555 size += sizeof(void *);
3558 #ifdef CONFIG_SLUB_DEBUG
3559 if (flags & SLAB_STORE_USER)
3561 * Need to store information about allocs and frees after
3564 size += 2 * sizeof(struct track);
3567 kasan_cache_create(s, &size, &s->flags);
3568 #ifdef CONFIG_SLUB_DEBUG
3569 if (flags & SLAB_RED_ZONE) {
3571 * Add some empty padding so that we can catch
3572 * overwrites from earlier objects rather than let
3573 * tracking information or the free pointer be
3574 * corrupted if a user writes before the start
3577 size += sizeof(void *);
3579 s->red_left_pad = sizeof(void *);
3580 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3581 size += s->red_left_pad;
3586 * SLUB stores one object immediately after another beginning from
3587 * offset 0. In order to align the objects we have to simply size
3588 * each object to conform to the alignment.
3590 size = ALIGN(size, s->align);
3592 if (forced_order >= 0)
3593 order = forced_order;
3595 order = calculate_order(size);
3602 s->allocflags |= __GFP_COMP;
3604 if (s->flags & SLAB_CACHE_DMA)
3605 s->allocflags |= GFP_DMA;
3607 if (s->flags & SLAB_CACHE_DMA32)
3608 s->allocflags |= GFP_DMA32;
3610 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3611 s->allocflags |= __GFP_RECLAIMABLE;
3614 * Determine the number of objects per slab
3616 s->oo = oo_make(order, size);
3617 s->min = oo_make(get_order(size), size);
3618 if (oo_objects(s->oo) > oo_objects(s->max))
3621 return !!oo_objects(s->oo);
3624 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3626 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3627 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3628 s->random = get_random_long();
3631 if (!calculate_sizes(s, -1))
3633 if (disable_higher_order_debug) {
3635 * Disable debugging flags that store metadata if the min slab
3638 if (get_order(s->size) > get_order(s->object_size)) {
3639 s->flags &= ~DEBUG_METADATA_FLAGS;
3641 if (!calculate_sizes(s, -1))
3646 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3647 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3648 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3649 /* Enable fast mode */
3650 s->flags |= __CMPXCHG_DOUBLE;
3654 * The larger the object size is, the more pages we want on the partial
3655 * list to avoid pounding the page allocator excessively.
3657 set_min_partial(s, ilog2(s->size) / 2);
3662 s->remote_node_defrag_ratio = 1000;
3665 /* Initialize the pre-computed randomized freelist if slab is up */
3666 if (slab_state >= UP) {
3667 if (init_cache_random_seq(s))
3671 if (!init_kmem_cache_nodes(s))
3674 if (alloc_kmem_cache_cpus(s))
3677 free_kmem_cache_nodes(s);
3682 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3685 #ifdef CONFIG_SLUB_DEBUG
3686 void *addr = page_address(page);
3688 unsigned long *map = bitmap_zalloc(page->objects, GFP_ATOMIC);
3691 slab_err(s, page, text, s->name);
3694 get_map(s, page, map);
3695 for_each_object(p, s, addr, page->objects) {
3697 if (!test_bit(slab_index(p, s, addr), map)) {
3698 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3699 print_tracking(s, p);
3708 * Attempt to free all partial slabs on a node.
3709 * This is called from __kmem_cache_shutdown(). We must take list_lock
3710 * because sysfs file might still access partial list after the shutdowning.
3712 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3715 struct page *page, *h;
3717 BUG_ON(irqs_disabled());
3718 spin_lock_irq(&n->list_lock);
3719 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3721 remove_partial(n, page);
3722 list_add(&page->slab_list, &discard);
3724 list_slab_objects(s, page,
3725 "Objects remaining in %s on __kmem_cache_shutdown()");
3728 spin_unlock_irq(&n->list_lock);
3730 list_for_each_entry_safe(page, h, &discard, slab_list)
3731 discard_slab(s, page);
3734 bool __kmem_cache_empty(struct kmem_cache *s)
3737 struct kmem_cache_node *n;
3739 for_each_kmem_cache_node(s, node, n)
3740 if (n->nr_partial || slabs_node(s, node))
3746 * Release all resources used by a slab cache.
3748 int __kmem_cache_shutdown(struct kmem_cache *s)
3751 struct kmem_cache_node *n;
3754 /* Attempt to free all objects */
3755 for_each_kmem_cache_node(s, node, n) {
3757 if (n->nr_partial || slabs_node(s, node))
3760 sysfs_slab_remove(s);
3764 /********************************************************************
3766 *******************************************************************/
3768 static int __init setup_slub_min_order(char *str)
3770 get_option(&str, (int *)&slub_min_order);
3775 __setup("slub_min_order=", setup_slub_min_order);
3777 static int __init setup_slub_max_order(char *str)
3779 get_option(&str, (int *)&slub_max_order);
3780 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3785 __setup("slub_max_order=", setup_slub_max_order);
3787 static int __init setup_slub_min_objects(char *str)
3789 get_option(&str, (int *)&slub_min_objects);
3794 __setup("slub_min_objects=", setup_slub_min_objects);
3796 void *__kmalloc(size_t size, gfp_t flags)
3798 struct kmem_cache *s;
3801 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3802 return kmalloc_large(size, flags);
3804 s = kmalloc_slab(size, flags);
3806 if (unlikely(ZERO_OR_NULL_PTR(s)))
3809 ret = slab_alloc(s, flags, _RET_IP_);
3811 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3813 ret = kasan_kmalloc(s, ret, size, flags);
3817 EXPORT_SYMBOL(__kmalloc);
3820 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3824 unsigned int order = get_order(size);
3826 flags |= __GFP_COMP;
3827 page = alloc_pages_node(node, flags, order);
3829 ptr = page_address(page);
3830 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3834 return kmalloc_large_node_hook(ptr, size, flags);
3837 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3839 struct kmem_cache *s;
3842 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3843 ret = kmalloc_large_node(size, flags, node);
3845 trace_kmalloc_node(_RET_IP_, ret,
3846 size, PAGE_SIZE << get_order(size),
3852 s = kmalloc_slab(size, flags);
3854 if (unlikely(ZERO_OR_NULL_PTR(s)))
3857 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3859 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3861 ret = kasan_kmalloc(s, ret, size, flags);
3865 EXPORT_SYMBOL(__kmalloc_node);
3866 #endif /* CONFIG_NUMA */
3868 #ifdef CONFIG_HARDENED_USERCOPY
3870 * Rejects incorrectly sized objects and objects that are to be copied
3871 * to/from userspace but do not fall entirely within the containing slab
3872 * cache's usercopy region.
3874 * Returns NULL if check passes, otherwise const char * to name of cache
3875 * to indicate an error.
3877 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3880 struct kmem_cache *s;
3881 unsigned int offset;
3884 ptr = kasan_reset_tag(ptr);
3886 /* Find object and usable object size. */
3887 s = page->slab_cache;
3889 /* Reject impossible pointers. */
3890 if (ptr < page_address(page))
3891 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3894 /* Find offset within object. */
3895 offset = (ptr - page_address(page)) % s->size;
3897 /* Adjust for redzone and reject if within the redzone. */
3898 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3899 if (offset < s->red_left_pad)
3900 usercopy_abort("SLUB object in left red zone",
3901 s->name, to_user, offset, n);
3902 offset -= s->red_left_pad;
3905 /* Allow address range falling entirely within usercopy region. */
3906 if (offset >= s->useroffset &&
3907 offset - s->useroffset <= s->usersize &&
3908 n <= s->useroffset - offset + s->usersize)
3912 * If the copy is still within the allocated object, produce
3913 * a warning instead of rejecting the copy. This is intended
3914 * to be a temporary method to find any missing usercopy
3917 object_size = slab_ksize(s);
3918 if (usercopy_fallback &&
3919 offset <= object_size && n <= object_size - offset) {
3920 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3924 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3926 #endif /* CONFIG_HARDENED_USERCOPY */
3928 size_t __ksize(const void *object)
3932 if (unlikely(object == ZERO_SIZE_PTR))
3935 page = virt_to_head_page(object);
3937 if (unlikely(!PageSlab(page))) {
3938 WARN_ON(!PageCompound(page));
3939 return page_size(page);
3942 return slab_ksize(page->slab_cache);
3944 EXPORT_SYMBOL(__ksize);
3946 void kfree(const void *x)
3949 void *object = (void *)x;
3951 trace_kfree(_RET_IP_, x);
3953 if (unlikely(ZERO_OR_NULL_PTR(x)))
3956 page = virt_to_head_page(x);
3957 if (unlikely(!PageSlab(page))) {
3958 unsigned int order = compound_order(page);
3960 BUG_ON(!PageCompound(page));
3962 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3964 __free_pages(page, order);
3967 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3969 EXPORT_SYMBOL(kfree);
3971 #define SHRINK_PROMOTE_MAX 32
3974 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3975 * up most to the head of the partial lists. New allocations will then
3976 * fill those up and thus they can be removed from the partial lists.
3978 * The slabs with the least items are placed last. This results in them
3979 * being allocated from last increasing the chance that the last objects
3980 * are freed in them.
3982 int __kmem_cache_shrink(struct kmem_cache *s)
3986 struct kmem_cache_node *n;
3989 struct list_head discard;
3990 struct list_head promote[SHRINK_PROMOTE_MAX];
3991 unsigned long flags;
3995 for_each_kmem_cache_node(s, node, n) {
3996 INIT_LIST_HEAD(&discard);
3997 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3998 INIT_LIST_HEAD(promote + i);
4000 spin_lock_irqsave(&n->list_lock, flags);
4003 * Build lists of slabs to discard or promote.
4005 * Note that concurrent frees may occur while we hold the
4006 * list_lock. page->inuse here is the upper limit.
4008 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4009 int free = page->objects - page->inuse;
4011 /* Do not reread page->inuse */
4014 /* We do not keep full slabs on the list */
4017 if (free == page->objects) {
4018 list_move(&page->slab_list, &discard);
4020 } else if (free <= SHRINK_PROMOTE_MAX)
4021 list_move(&page->slab_list, promote + free - 1);
4025 * Promote the slabs filled up most to the head of the
4028 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4029 list_splice(promote + i, &n->partial);
4031 spin_unlock_irqrestore(&n->list_lock, flags);
4033 /* Release empty slabs */
4034 list_for_each_entry_safe(page, t, &discard, slab_list)
4035 discard_slab(s, page);
4037 if (slabs_node(s, node))
4045 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4048 * Called with all the locks held after a sched RCU grace period.
4049 * Even if @s becomes empty after shrinking, we can't know that @s
4050 * doesn't have allocations already in-flight and thus can't
4051 * destroy @s until the associated memcg is released.
4053 * However, let's remove the sysfs files for empty caches here.
4054 * Each cache has a lot of interface files which aren't
4055 * particularly useful for empty draining caches; otherwise, we can
4056 * easily end up with millions of unnecessary sysfs files on
4057 * systems which have a lot of memory and transient cgroups.
4059 if (!__kmem_cache_shrink(s))
4060 sysfs_slab_remove(s);
4063 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4066 * Disable empty slabs caching. Used to avoid pinning offline
4067 * memory cgroups by kmem pages that can be freed.
4069 slub_set_cpu_partial(s, 0);
4072 #endif /* CONFIG_MEMCG */
4074 static int slab_mem_going_offline_callback(void *arg)
4076 struct kmem_cache *s;
4078 mutex_lock(&slab_mutex);
4079 list_for_each_entry(s, &slab_caches, list)
4080 __kmem_cache_shrink(s);
4081 mutex_unlock(&slab_mutex);
4086 static void slab_mem_offline_callback(void *arg)
4088 struct kmem_cache_node *n;
4089 struct kmem_cache *s;
4090 struct memory_notify *marg = arg;
4093 offline_node = marg->status_change_nid_normal;
4096 * If the node still has available memory. we need kmem_cache_node
4099 if (offline_node < 0)
4102 mutex_lock(&slab_mutex);
4103 list_for_each_entry(s, &slab_caches, list) {
4104 n = get_node(s, offline_node);
4107 * if n->nr_slabs > 0, slabs still exist on the node
4108 * that is going down. We were unable to free them,
4109 * and offline_pages() function shouldn't call this
4110 * callback. So, we must fail.
4112 BUG_ON(slabs_node(s, offline_node));
4114 s->node[offline_node] = NULL;
4115 kmem_cache_free(kmem_cache_node, n);
4118 mutex_unlock(&slab_mutex);
4121 static int slab_mem_going_online_callback(void *arg)
4123 struct kmem_cache_node *n;
4124 struct kmem_cache *s;
4125 struct memory_notify *marg = arg;
4126 int nid = marg->status_change_nid_normal;
4130 * If the node's memory is already available, then kmem_cache_node is
4131 * already created. Nothing to do.
4137 * We are bringing a node online. No memory is available yet. We must
4138 * allocate a kmem_cache_node structure in order to bring the node
4141 mutex_lock(&slab_mutex);
4142 list_for_each_entry(s, &slab_caches, list) {
4144 * XXX: kmem_cache_alloc_node will fallback to other nodes
4145 * since memory is not yet available from the node that
4148 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4153 init_kmem_cache_node(n);
4157 mutex_unlock(&slab_mutex);
4161 static int slab_memory_callback(struct notifier_block *self,
4162 unsigned long action, void *arg)
4167 case MEM_GOING_ONLINE:
4168 ret = slab_mem_going_online_callback(arg);
4170 case MEM_GOING_OFFLINE:
4171 ret = slab_mem_going_offline_callback(arg);
4174 case MEM_CANCEL_ONLINE:
4175 slab_mem_offline_callback(arg);
4178 case MEM_CANCEL_OFFLINE:
4182 ret = notifier_from_errno(ret);
4188 static struct notifier_block slab_memory_callback_nb = {
4189 .notifier_call = slab_memory_callback,
4190 .priority = SLAB_CALLBACK_PRI,
4193 /********************************************************************
4194 * Basic setup of slabs
4195 *******************************************************************/
4198 * Used for early kmem_cache structures that were allocated using
4199 * the page allocator. Allocate them properly then fix up the pointers
4200 * that may be pointing to the wrong kmem_cache structure.
4203 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4206 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4207 struct kmem_cache_node *n;
4209 memcpy(s, static_cache, kmem_cache->object_size);
4212 * This runs very early, and only the boot processor is supposed to be
4213 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4216 __flush_cpu_slab(s, smp_processor_id());
4217 for_each_kmem_cache_node(s, node, n) {
4220 list_for_each_entry(p, &n->partial, slab_list)
4223 #ifdef CONFIG_SLUB_DEBUG
4224 list_for_each_entry(p, &n->full, slab_list)
4228 slab_init_memcg_params(s);
4229 list_add(&s->list, &slab_caches);
4230 memcg_link_cache(s, NULL);
4234 void __init kmem_cache_init(void)
4236 static __initdata struct kmem_cache boot_kmem_cache,
4237 boot_kmem_cache_node;
4239 if (debug_guardpage_minorder())
4242 kmem_cache_node = &boot_kmem_cache_node;
4243 kmem_cache = &boot_kmem_cache;
4245 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4246 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4248 register_hotmemory_notifier(&slab_memory_callback_nb);
4250 /* Able to allocate the per node structures */
4251 slab_state = PARTIAL;
4253 create_boot_cache(kmem_cache, "kmem_cache",
4254 offsetof(struct kmem_cache, node) +
4255 nr_node_ids * sizeof(struct kmem_cache_node *),
4256 SLAB_HWCACHE_ALIGN, 0, 0);
4258 kmem_cache = bootstrap(&boot_kmem_cache);
4259 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4261 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4262 setup_kmalloc_cache_index_table();
4263 create_kmalloc_caches(0);
4265 /* Setup random freelists for each cache */
4266 init_freelist_randomization();
4268 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4271 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4273 slub_min_order, slub_max_order, slub_min_objects,
4274 nr_cpu_ids, nr_node_ids);
4277 void __init kmem_cache_init_late(void)
4282 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4283 slab_flags_t flags, void (*ctor)(void *))
4285 struct kmem_cache *s, *c;
4287 s = find_mergeable(size, align, flags, name, ctor);
4292 * Adjust the object sizes so that we clear
4293 * the complete object on kzalloc.
4295 s->object_size = max(s->object_size, size);
4296 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4298 for_each_memcg_cache(c, s) {
4299 c->object_size = s->object_size;
4300 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4303 if (sysfs_slab_alias(s, name)) {
4312 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4316 err = kmem_cache_open(s, flags);
4320 /* Mutex is not taken during early boot */
4321 if (slab_state <= UP)
4324 memcg_propagate_slab_attrs(s);
4325 err = sysfs_slab_add(s);
4327 __kmem_cache_release(s);
4332 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4334 struct kmem_cache *s;
4337 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4338 return kmalloc_large(size, gfpflags);
4340 s = kmalloc_slab(size, gfpflags);
4342 if (unlikely(ZERO_OR_NULL_PTR(s)))
4345 ret = slab_alloc(s, gfpflags, caller);
4347 /* Honor the call site pointer we received. */
4348 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4354 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4355 int node, unsigned long caller)
4357 struct kmem_cache *s;
4360 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4361 ret = kmalloc_large_node(size, gfpflags, node);
4363 trace_kmalloc_node(caller, ret,
4364 size, PAGE_SIZE << get_order(size),
4370 s = kmalloc_slab(size, gfpflags);
4372 if (unlikely(ZERO_OR_NULL_PTR(s)))
4375 ret = slab_alloc_node(s, gfpflags, node, caller);
4377 /* Honor the call site pointer we received. */
4378 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4385 static int count_inuse(struct page *page)
4390 static int count_total(struct page *page)
4392 return page->objects;
4396 #ifdef CONFIG_SLUB_DEBUG
4397 static int validate_slab(struct kmem_cache *s, struct page *page,
4401 void *addr = page_address(page);
4403 if (!check_slab(s, page) ||
4404 !on_freelist(s, page, NULL))
4407 /* Now we know that a valid freelist exists */
4408 bitmap_zero(map, page->objects);
4410 get_map(s, page, map);
4411 for_each_object(p, s, addr, page->objects) {
4412 if (test_bit(slab_index(p, s, addr), map))
4413 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4417 for_each_object(p, s, addr, page->objects)
4418 if (!test_bit(slab_index(p, s, addr), map))
4419 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4424 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4428 validate_slab(s, page, map);
4432 static int validate_slab_node(struct kmem_cache *s,
4433 struct kmem_cache_node *n, unsigned long *map)
4435 unsigned long count = 0;
4437 unsigned long flags;
4439 spin_lock_irqsave(&n->list_lock, flags);
4441 list_for_each_entry(page, &n->partial, slab_list) {
4442 validate_slab_slab(s, page, map);
4445 if (count != n->nr_partial)
4446 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4447 s->name, count, n->nr_partial);
4449 if (!(s->flags & SLAB_STORE_USER))
4452 list_for_each_entry(page, &n->full, slab_list) {
4453 validate_slab_slab(s, page, map);
4456 if (count != atomic_long_read(&n->nr_slabs))
4457 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4458 s->name, count, atomic_long_read(&n->nr_slabs));
4461 spin_unlock_irqrestore(&n->list_lock, flags);
4465 static long validate_slab_cache(struct kmem_cache *s)
4468 unsigned long count = 0;
4469 struct kmem_cache_node *n;
4470 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4476 for_each_kmem_cache_node(s, node, n)
4477 count += validate_slab_node(s, n, map);
4482 * Generate lists of code addresses where slabcache objects are allocated
4487 unsigned long count;
4494 DECLARE_BITMAP(cpus, NR_CPUS);
4500 unsigned long count;
4501 struct location *loc;
4504 static void free_loc_track(struct loc_track *t)
4507 free_pages((unsigned long)t->loc,
4508 get_order(sizeof(struct location) * t->max));
4511 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4516 order = get_order(sizeof(struct location) * max);
4518 l = (void *)__get_free_pages(flags, order);
4523 memcpy(l, t->loc, sizeof(struct location) * t->count);
4531 static int add_location(struct loc_track *t, struct kmem_cache *s,
4532 const struct track *track)
4534 long start, end, pos;
4536 unsigned long caddr;
4537 unsigned long age = jiffies - track->when;
4543 pos = start + (end - start + 1) / 2;
4546 * There is nothing at "end". If we end up there
4547 * we need to add something to before end.
4552 caddr = t->loc[pos].addr;
4553 if (track->addr == caddr) {
4559 if (age < l->min_time)
4561 if (age > l->max_time)
4564 if (track->pid < l->min_pid)
4565 l->min_pid = track->pid;
4566 if (track->pid > l->max_pid)
4567 l->max_pid = track->pid;
4569 cpumask_set_cpu(track->cpu,
4570 to_cpumask(l->cpus));
4572 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4576 if (track->addr < caddr)
4583 * Not found. Insert new tracking element.
4585 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4591 (t->count - pos) * sizeof(struct location));
4594 l->addr = track->addr;
4598 l->min_pid = track->pid;
4599 l->max_pid = track->pid;
4600 cpumask_clear(to_cpumask(l->cpus));
4601 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4602 nodes_clear(l->nodes);
4603 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4607 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4608 struct page *page, enum track_item alloc,
4611 void *addr = page_address(page);
4614 bitmap_zero(map, page->objects);
4615 get_map(s, page, map);
4617 for_each_object(p, s, addr, page->objects)
4618 if (!test_bit(slab_index(p, s, addr), map))
4619 add_location(t, s, get_track(s, p, alloc));
4622 static int list_locations(struct kmem_cache *s, char *buf,
4623 enum track_item alloc)
4627 struct loc_track t = { 0, 0, NULL };
4629 struct kmem_cache_node *n;
4630 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4632 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4635 return sprintf(buf, "Out of memory\n");
4637 /* Push back cpu slabs */
4640 for_each_kmem_cache_node(s, node, n) {
4641 unsigned long flags;
4644 if (!atomic_long_read(&n->nr_slabs))
4647 spin_lock_irqsave(&n->list_lock, flags);
4648 list_for_each_entry(page, &n->partial, slab_list)
4649 process_slab(&t, s, page, alloc, map);
4650 list_for_each_entry(page, &n->full, slab_list)
4651 process_slab(&t, s, page, alloc, map);
4652 spin_unlock_irqrestore(&n->list_lock, flags);
4655 for (i = 0; i < t.count; i++) {
4656 struct location *l = &t.loc[i];
4658 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4660 len += sprintf(buf + len, "%7ld ", l->count);
4663 len += sprintf(buf + len, "%pS", (void *)l->addr);
4665 len += sprintf(buf + len, "<not-available>");
4667 if (l->sum_time != l->min_time) {
4668 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4670 (long)div_u64(l->sum_time, l->count),
4673 len += sprintf(buf + len, " age=%ld",
4676 if (l->min_pid != l->max_pid)
4677 len += sprintf(buf + len, " pid=%ld-%ld",
4678 l->min_pid, l->max_pid);
4680 len += sprintf(buf + len, " pid=%ld",
4683 if (num_online_cpus() > 1 &&
4684 !cpumask_empty(to_cpumask(l->cpus)) &&
4685 len < PAGE_SIZE - 60)
4686 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4688 cpumask_pr_args(to_cpumask(l->cpus)));
4690 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4691 len < PAGE_SIZE - 60)
4692 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4694 nodemask_pr_args(&l->nodes));
4696 len += sprintf(buf + len, "\n");
4702 len += sprintf(buf, "No data\n");
4705 #endif /* CONFIG_SLUB_DEBUG */
4707 #ifdef SLUB_RESILIENCY_TEST
4708 static void __init resiliency_test(void)
4711 int type = KMALLOC_NORMAL;
4713 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4715 pr_err("SLUB resiliency testing\n");
4716 pr_err("-----------------------\n");
4717 pr_err("A. Corruption after allocation\n");
4719 p = kzalloc(16, GFP_KERNEL);
4721 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4724 validate_slab_cache(kmalloc_caches[type][4]);
4726 /* Hmmm... The next two are dangerous */
4727 p = kzalloc(32, GFP_KERNEL);
4728 p[32 + sizeof(void *)] = 0x34;
4729 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4731 pr_err("If allocated object is overwritten then not detectable\n\n");
4733 validate_slab_cache(kmalloc_caches[type][5]);
4734 p = kzalloc(64, GFP_KERNEL);
4735 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4737 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4739 pr_err("If allocated object is overwritten then not detectable\n\n");
4740 validate_slab_cache(kmalloc_caches[type][6]);
4742 pr_err("\nB. Corruption after free\n");
4743 p = kzalloc(128, GFP_KERNEL);
4746 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4747 validate_slab_cache(kmalloc_caches[type][7]);
4749 p = kzalloc(256, GFP_KERNEL);
4752 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4753 validate_slab_cache(kmalloc_caches[type][8]);
4755 p = kzalloc(512, GFP_KERNEL);
4758 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4759 validate_slab_cache(kmalloc_caches[type][9]);
4763 static void resiliency_test(void) {};
4765 #endif /* SLUB_RESILIENCY_TEST */
4768 enum slab_stat_type {
4769 SL_ALL, /* All slabs */
4770 SL_PARTIAL, /* Only partially allocated slabs */
4771 SL_CPU, /* Only slabs used for cpu caches */
4772 SL_OBJECTS, /* Determine allocated objects not slabs */
4773 SL_TOTAL /* Determine object capacity not slabs */
4776 #define SO_ALL (1 << SL_ALL)
4777 #define SO_PARTIAL (1 << SL_PARTIAL)
4778 #define SO_CPU (1 << SL_CPU)
4779 #define SO_OBJECTS (1 << SL_OBJECTS)
4780 #define SO_TOTAL (1 << SL_TOTAL)
4783 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4785 static int __init setup_slub_memcg_sysfs(char *str)
4789 if (get_option(&str, &v) > 0)
4790 memcg_sysfs_enabled = v;
4795 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4798 static ssize_t show_slab_objects(struct kmem_cache *s,
4799 char *buf, unsigned long flags)
4801 unsigned long total = 0;
4804 unsigned long *nodes;
4806 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4810 if (flags & SO_CPU) {
4813 for_each_possible_cpu(cpu) {
4814 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4819 page = READ_ONCE(c->page);
4823 node = page_to_nid(page);
4824 if (flags & SO_TOTAL)
4826 else if (flags & SO_OBJECTS)
4834 page = slub_percpu_partial_read_once(c);
4836 node = page_to_nid(page);
4837 if (flags & SO_TOTAL)
4839 else if (flags & SO_OBJECTS)
4850 #ifdef CONFIG_SLUB_DEBUG
4851 if (flags & SO_ALL) {
4852 struct kmem_cache_node *n;
4854 for_each_kmem_cache_node(s, node, n) {
4856 if (flags & SO_TOTAL)
4857 x = atomic_long_read(&n->total_objects);
4858 else if (flags & SO_OBJECTS)
4859 x = atomic_long_read(&n->total_objects) -
4860 count_partial(n, count_free);
4862 x = atomic_long_read(&n->nr_slabs);
4869 if (flags & SO_PARTIAL) {
4870 struct kmem_cache_node *n;
4872 for_each_kmem_cache_node(s, node, n) {
4873 if (flags & SO_TOTAL)
4874 x = count_partial(n, count_total);
4875 else if (flags & SO_OBJECTS)
4876 x = count_partial(n, count_inuse);
4883 x = sprintf(buf, "%lu", total);
4885 for (node = 0; node < nr_node_ids; node++)
4887 x += sprintf(buf + x, " N%d=%lu",
4892 return x + sprintf(buf + x, "\n");
4895 #ifdef CONFIG_SLUB_DEBUG
4896 static int any_slab_objects(struct kmem_cache *s)
4899 struct kmem_cache_node *n;
4901 for_each_kmem_cache_node(s, node, n)
4902 if (atomic_long_read(&n->total_objects))
4909 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4910 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4912 struct slab_attribute {
4913 struct attribute attr;
4914 ssize_t (*show)(struct kmem_cache *s, char *buf);
4915 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4918 #define SLAB_ATTR_RO(_name) \
4919 static struct slab_attribute _name##_attr = \
4920 __ATTR(_name, 0400, _name##_show, NULL)
4922 #define SLAB_ATTR(_name) \
4923 static struct slab_attribute _name##_attr = \
4924 __ATTR(_name, 0600, _name##_show, _name##_store)
4926 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4928 return sprintf(buf, "%u\n", s->size);
4930 SLAB_ATTR_RO(slab_size);
4932 static ssize_t align_show(struct kmem_cache *s, char *buf)
4934 return sprintf(buf, "%u\n", s->align);
4936 SLAB_ATTR_RO(align);
4938 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4940 return sprintf(buf, "%u\n", s->object_size);
4942 SLAB_ATTR_RO(object_size);
4944 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4946 return sprintf(buf, "%u\n", oo_objects(s->oo));
4948 SLAB_ATTR_RO(objs_per_slab);
4950 static ssize_t order_store(struct kmem_cache *s,
4951 const char *buf, size_t length)
4956 err = kstrtouint(buf, 10, &order);
4960 if (order > slub_max_order || order < slub_min_order)
4963 calculate_sizes(s, order);
4967 static ssize_t order_show(struct kmem_cache *s, char *buf)
4969 return sprintf(buf, "%u\n", oo_order(s->oo));
4973 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4975 return sprintf(buf, "%lu\n", s->min_partial);
4978 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4984 err = kstrtoul(buf, 10, &min);
4988 set_min_partial(s, min);
4991 SLAB_ATTR(min_partial);
4993 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4995 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4998 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5001 unsigned int objects;
5004 err = kstrtouint(buf, 10, &objects);
5007 if (objects && !kmem_cache_has_cpu_partial(s))
5010 slub_set_cpu_partial(s, objects);
5014 SLAB_ATTR(cpu_partial);
5016 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5020 return sprintf(buf, "%pS\n", s->ctor);
5024 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5026 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5028 SLAB_ATTR_RO(aliases);
5030 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5032 return show_slab_objects(s, buf, SO_PARTIAL);
5034 SLAB_ATTR_RO(partial);
5036 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5038 return show_slab_objects(s, buf, SO_CPU);
5040 SLAB_ATTR_RO(cpu_slabs);
5042 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5044 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5046 SLAB_ATTR_RO(objects);
5048 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5050 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5052 SLAB_ATTR_RO(objects_partial);
5054 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5061 for_each_online_cpu(cpu) {
5064 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5067 pages += page->pages;
5068 objects += page->pobjects;
5072 len = sprintf(buf, "%d(%d)", objects, pages);
5075 for_each_online_cpu(cpu) {
5078 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5080 if (page && len < PAGE_SIZE - 20)
5081 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5082 page->pobjects, page->pages);
5085 return len + sprintf(buf + len, "\n");
5087 SLAB_ATTR_RO(slabs_cpu_partial);
5089 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5091 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5094 static ssize_t reclaim_account_store(struct kmem_cache *s,
5095 const char *buf, size_t length)
5097 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5099 s->flags |= SLAB_RECLAIM_ACCOUNT;
5102 SLAB_ATTR(reclaim_account);
5104 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5106 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5108 SLAB_ATTR_RO(hwcache_align);
5110 #ifdef CONFIG_ZONE_DMA
5111 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5113 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5115 SLAB_ATTR_RO(cache_dma);
5118 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5120 return sprintf(buf, "%u\n", s->usersize);
5122 SLAB_ATTR_RO(usersize);
5124 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5126 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5128 SLAB_ATTR_RO(destroy_by_rcu);
5130 #ifdef CONFIG_SLUB_DEBUG
5131 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5133 return show_slab_objects(s, buf, SO_ALL);
5135 SLAB_ATTR_RO(slabs);
5137 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5139 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5141 SLAB_ATTR_RO(total_objects);
5143 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5145 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5148 static ssize_t sanity_checks_store(struct kmem_cache *s,
5149 const char *buf, size_t length)
5151 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5152 if (buf[0] == '1') {
5153 s->flags &= ~__CMPXCHG_DOUBLE;
5154 s->flags |= SLAB_CONSISTENCY_CHECKS;
5158 SLAB_ATTR(sanity_checks);
5160 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5162 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5165 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5169 * Tracing a merged cache is going to give confusing results
5170 * as well as cause other issues like converting a mergeable
5171 * cache into an umergeable one.
5173 if (s->refcount > 1)
5176 s->flags &= ~SLAB_TRACE;
5177 if (buf[0] == '1') {
5178 s->flags &= ~__CMPXCHG_DOUBLE;
5179 s->flags |= SLAB_TRACE;
5185 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5187 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5190 static ssize_t red_zone_store(struct kmem_cache *s,
5191 const char *buf, size_t length)
5193 if (any_slab_objects(s))
5196 s->flags &= ~SLAB_RED_ZONE;
5197 if (buf[0] == '1') {
5198 s->flags |= SLAB_RED_ZONE;
5200 calculate_sizes(s, -1);
5203 SLAB_ATTR(red_zone);
5205 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5207 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5210 static ssize_t poison_store(struct kmem_cache *s,
5211 const char *buf, size_t length)
5213 if (any_slab_objects(s))
5216 s->flags &= ~SLAB_POISON;
5217 if (buf[0] == '1') {
5218 s->flags |= SLAB_POISON;
5220 calculate_sizes(s, -1);
5225 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5227 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5230 static ssize_t store_user_store(struct kmem_cache *s,
5231 const char *buf, size_t length)
5233 if (any_slab_objects(s))
5236 s->flags &= ~SLAB_STORE_USER;
5237 if (buf[0] == '1') {
5238 s->flags &= ~__CMPXCHG_DOUBLE;
5239 s->flags |= SLAB_STORE_USER;
5241 calculate_sizes(s, -1);
5244 SLAB_ATTR(store_user);
5246 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5251 static ssize_t validate_store(struct kmem_cache *s,
5252 const char *buf, size_t length)
5256 if (buf[0] == '1') {
5257 ret = validate_slab_cache(s);
5263 SLAB_ATTR(validate);
5265 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5267 if (!(s->flags & SLAB_STORE_USER))
5269 return list_locations(s, buf, TRACK_ALLOC);
5271 SLAB_ATTR_RO(alloc_calls);
5273 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5275 if (!(s->flags & SLAB_STORE_USER))
5277 return list_locations(s, buf, TRACK_FREE);
5279 SLAB_ATTR_RO(free_calls);
5280 #endif /* CONFIG_SLUB_DEBUG */
5282 #ifdef CONFIG_FAILSLAB
5283 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5285 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5288 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5291 if (s->refcount > 1)
5294 s->flags &= ~SLAB_FAILSLAB;
5296 s->flags |= SLAB_FAILSLAB;
5299 SLAB_ATTR(failslab);
5302 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5307 static ssize_t shrink_store(struct kmem_cache *s,
5308 const char *buf, size_t length)
5311 kmem_cache_shrink_all(s);
5319 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5321 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5324 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5325 const char *buf, size_t length)
5330 err = kstrtouint(buf, 10, &ratio);
5336 s->remote_node_defrag_ratio = ratio * 10;
5340 SLAB_ATTR(remote_node_defrag_ratio);
5343 #ifdef CONFIG_SLUB_STATS
5344 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5346 unsigned long sum = 0;
5349 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5354 for_each_online_cpu(cpu) {
5355 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5361 len = sprintf(buf, "%lu", sum);
5364 for_each_online_cpu(cpu) {
5365 if (data[cpu] && len < PAGE_SIZE - 20)
5366 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5370 return len + sprintf(buf + len, "\n");
5373 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5377 for_each_online_cpu(cpu)
5378 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5381 #define STAT_ATTR(si, text) \
5382 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5384 return show_stat(s, buf, si); \
5386 static ssize_t text##_store(struct kmem_cache *s, \
5387 const char *buf, size_t length) \
5389 if (buf[0] != '0') \
5391 clear_stat(s, si); \
5396 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5397 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5398 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5399 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5400 STAT_ATTR(FREE_FROZEN, free_frozen);
5401 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5402 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5403 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5404 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5405 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5406 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5407 STAT_ATTR(FREE_SLAB, free_slab);
5408 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5409 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5410 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5411 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5412 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5413 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5414 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5415 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5416 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5417 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5418 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5419 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5420 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5421 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5422 #endif /* CONFIG_SLUB_STATS */
5424 static struct attribute *slab_attrs[] = {
5425 &slab_size_attr.attr,
5426 &object_size_attr.attr,
5427 &objs_per_slab_attr.attr,
5429 &min_partial_attr.attr,
5430 &cpu_partial_attr.attr,
5432 &objects_partial_attr.attr,
5434 &cpu_slabs_attr.attr,
5438 &hwcache_align_attr.attr,
5439 &reclaim_account_attr.attr,
5440 &destroy_by_rcu_attr.attr,
5442 &slabs_cpu_partial_attr.attr,
5443 #ifdef CONFIG_SLUB_DEBUG
5444 &total_objects_attr.attr,
5446 &sanity_checks_attr.attr,
5448 &red_zone_attr.attr,
5450 &store_user_attr.attr,
5451 &validate_attr.attr,
5452 &alloc_calls_attr.attr,
5453 &free_calls_attr.attr,
5455 #ifdef CONFIG_ZONE_DMA
5456 &cache_dma_attr.attr,
5459 &remote_node_defrag_ratio_attr.attr,
5461 #ifdef CONFIG_SLUB_STATS
5462 &alloc_fastpath_attr.attr,
5463 &alloc_slowpath_attr.attr,
5464 &free_fastpath_attr.attr,
5465 &free_slowpath_attr.attr,
5466 &free_frozen_attr.attr,
5467 &free_add_partial_attr.attr,
5468 &free_remove_partial_attr.attr,
5469 &alloc_from_partial_attr.attr,
5470 &alloc_slab_attr.attr,
5471 &alloc_refill_attr.attr,
5472 &alloc_node_mismatch_attr.attr,
5473 &free_slab_attr.attr,
5474 &cpuslab_flush_attr.attr,
5475 &deactivate_full_attr.attr,
5476 &deactivate_empty_attr.attr,
5477 &deactivate_to_head_attr.attr,
5478 &deactivate_to_tail_attr.attr,
5479 &deactivate_remote_frees_attr.attr,
5480 &deactivate_bypass_attr.attr,
5481 &order_fallback_attr.attr,
5482 &cmpxchg_double_fail_attr.attr,
5483 &cmpxchg_double_cpu_fail_attr.attr,
5484 &cpu_partial_alloc_attr.attr,
5485 &cpu_partial_free_attr.attr,
5486 &cpu_partial_node_attr.attr,
5487 &cpu_partial_drain_attr.attr,
5489 #ifdef CONFIG_FAILSLAB
5490 &failslab_attr.attr,
5492 &usersize_attr.attr,
5497 static const struct attribute_group slab_attr_group = {
5498 .attrs = slab_attrs,
5501 static ssize_t slab_attr_show(struct kobject *kobj,
5502 struct attribute *attr,
5505 struct slab_attribute *attribute;
5506 struct kmem_cache *s;
5509 attribute = to_slab_attr(attr);
5512 if (!attribute->show)
5515 err = attribute->show(s, buf);
5520 static ssize_t slab_attr_store(struct kobject *kobj,
5521 struct attribute *attr,
5522 const char *buf, size_t len)
5524 struct slab_attribute *attribute;
5525 struct kmem_cache *s;
5528 attribute = to_slab_attr(attr);
5531 if (!attribute->store)
5534 err = attribute->store(s, buf, len);
5536 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5537 struct kmem_cache *c;
5539 mutex_lock(&slab_mutex);
5540 if (s->max_attr_size < len)
5541 s->max_attr_size = len;
5544 * This is a best effort propagation, so this function's return
5545 * value will be determined by the parent cache only. This is
5546 * basically because not all attributes will have a well
5547 * defined semantics for rollbacks - most of the actions will
5548 * have permanent effects.
5550 * Returning the error value of any of the children that fail
5551 * is not 100 % defined, in the sense that users seeing the
5552 * error code won't be able to know anything about the state of
5555 * Only returning the error code for the parent cache at least
5556 * has well defined semantics. The cache being written to
5557 * directly either failed or succeeded, in which case we loop
5558 * through the descendants with best-effort propagation.
5560 for_each_memcg_cache(c, s)
5561 attribute->store(c, buf, len);
5562 mutex_unlock(&slab_mutex);
5568 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5572 char *buffer = NULL;
5573 struct kmem_cache *root_cache;
5575 if (is_root_cache(s))
5578 root_cache = s->memcg_params.root_cache;
5581 * This mean this cache had no attribute written. Therefore, no point
5582 * in copying default values around
5584 if (!root_cache->max_attr_size)
5587 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5590 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5593 if (!attr || !attr->store || !attr->show)
5597 * It is really bad that we have to allocate here, so we will
5598 * do it only as a fallback. If we actually allocate, though,
5599 * we can just use the allocated buffer until the end.
5601 * Most of the slub attributes will tend to be very small in
5602 * size, but sysfs allows buffers up to a page, so they can
5603 * theoretically happen.
5607 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5610 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5611 if (WARN_ON(!buffer))
5616 len = attr->show(root_cache, buf);
5618 attr->store(s, buf, len);
5622 free_page((unsigned long)buffer);
5623 #endif /* CONFIG_MEMCG */
5626 static void kmem_cache_release(struct kobject *k)
5628 slab_kmem_cache_release(to_slab(k));
5631 static const struct sysfs_ops slab_sysfs_ops = {
5632 .show = slab_attr_show,
5633 .store = slab_attr_store,
5636 static struct kobj_type slab_ktype = {
5637 .sysfs_ops = &slab_sysfs_ops,
5638 .release = kmem_cache_release,
5641 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5643 struct kobj_type *ktype = get_ktype(kobj);
5645 if (ktype == &slab_ktype)
5650 static const struct kset_uevent_ops slab_uevent_ops = {
5651 .filter = uevent_filter,
5654 static struct kset *slab_kset;
5656 static inline struct kset *cache_kset(struct kmem_cache *s)
5659 if (!is_root_cache(s))
5660 return s->memcg_params.root_cache->memcg_kset;
5665 #define ID_STR_LENGTH 64
5667 /* Create a unique string id for a slab cache:
5669 * Format :[flags-]size
5671 static char *create_unique_id(struct kmem_cache *s)
5673 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5680 * First flags affecting slabcache operations. We will only
5681 * get here for aliasable slabs so we do not need to support
5682 * too many flags. The flags here must cover all flags that
5683 * are matched during merging to guarantee that the id is
5686 if (s->flags & SLAB_CACHE_DMA)
5688 if (s->flags & SLAB_CACHE_DMA32)
5690 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5692 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5694 if (s->flags & SLAB_ACCOUNT)
5698 p += sprintf(p, "%07u", s->size);
5700 BUG_ON(p > name + ID_STR_LENGTH - 1);
5704 static void sysfs_slab_remove_workfn(struct work_struct *work)
5706 struct kmem_cache *s =
5707 container_of(work, struct kmem_cache, kobj_remove_work);
5709 if (!s->kobj.state_in_sysfs)
5711 * For a memcg cache, this may be called during
5712 * deactivation and again on shutdown. Remove only once.
5713 * A cache is never shut down before deactivation is
5714 * complete, so no need to worry about synchronization.
5719 kset_unregister(s->memcg_kset);
5721 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5723 kobject_put(&s->kobj);
5726 static int sysfs_slab_add(struct kmem_cache *s)
5730 struct kset *kset = cache_kset(s);
5731 int unmergeable = slab_unmergeable(s);
5733 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5736 kobject_init(&s->kobj, &slab_ktype);
5740 if (!unmergeable && disable_higher_order_debug &&
5741 (slub_debug & DEBUG_METADATA_FLAGS))
5746 * Slabcache can never be merged so we can use the name proper.
5747 * This is typically the case for debug situations. In that
5748 * case we can catch duplicate names easily.
5750 sysfs_remove_link(&slab_kset->kobj, s->name);
5754 * Create a unique name for the slab as a target
5757 name = create_unique_id(s);
5760 s->kobj.kset = kset;
5761 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5765 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5770 if (is_root_cache(s) && memcg_sysfs_enabled) {
5771 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5772 if (!s->memcg_kset) {
5779 kobject_uevent(&s->kobj, KOBJ_ADD);
5781 /* Setup first alias */
5782 sysfs_slab_alias(s, s->name);
5789 kobject_del(&s->kobj);
5793 static void sysfs_slab_remove(struct kmem_cache *s)
5795 if (slab_state < FULL)
5797 * Sysfs has not been setup yet so no need to remove the
5802 kobject_get(&s->kobj);
5803 schedule_work(&s->kobj_remove_work);
5806 void sysfs_slab_unlink(struct kmem_cache *s)
5808 if (slab_state >= FULL)
5809 kobject_del(&s->kobj);
5812 void sysfs_slab_release(struct kmem_cache *s)
5814 if (slab_state >= FULL)
5815 kobject_put(&s->kobj);
5819 * Need to buffer aliases during bootup until sysfs becomes
5820 * available lest we lose that information.
5822 struct saved_alias {
5823 struct kmem_cache *s;
5825 struct saved_alias *next;
5828 static struct saved_alias *alias_list;
5830 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5832 struct saved_alias *al;
5834 if (slab_state == FULL) {
5836 * If we have a leftover link then remove it.
5838 sysfs_remove_link(&slab_kset->kobj, name);
5839 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5842 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5848 al->next = alias_list;
5853 static int __init slab_sysfs_init(void)
5855 struct kmem_cache *s;
5858 mutex_lock(&slab_mutex);
5860 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5862 mutex_unlock(&slab_mutex);
5863 pr_err("Cannot register slab subsystem.\n");
5869 list_for_each_entry(s, &slab_caches, list) {
5870 err = sysfs_slab_add(s);
5872 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5876 while (alias_list) {
5877 struct saved_alias *al = alias_list;
5879 alias_list = alias_list->next;
5880 err = sysfs_slab_alias(al->s, al->name);
5882 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5887 mutex_unlock(&slab_mutex);
5892 __initcall(slab_sysfs_init);
5893 #endif /* CONFIG_SYSFS */
5896 * The /proc/slabinfo ABI
5898 #ifdef CONFIG_SLUB_DEBUG
5899 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5901 unsigned long nr_slabs = 0;
5902 unsigned long nr_objs = 0;
5903 unsigned long nr_free = 0;
5905 struct kmem_cache_node *n;
5907 for_each_kmem_cache_node(s, node, n) {
5908 nr_slabs += node_nr_slabs(n);
5909 nr_objs += node_nr_objs(n);
5910 nr_free += count_partial(n, count_free);
5913 sinfo->active_objs = nr_objs - nr_free;
5914 sinfo->num_objs = nr_objs;
5915 sinfo->active_slabs = nr_slabs;
5916 sinfo->num_slabs = nr_slabs;
5917 sinfo->objects_per_slab = oo_objects(s->oo);
5918 sinfo->cache_order = oo_order(s->oo);
5921 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5925 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5926 size_t count, loff_t *ppos)
5930 #endif /* CONFIG_SLUB_DEBUG */