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 << compound_order(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);
1077 static void setup_page_debug(struct kmem_cache *s, void *addr, int order)
1079 if (!(s->flags & SLAB_POISON))
1082 metadata_access_enable();
1083 memset(addr, POISON_INUSE, PAGE_SIZE << order);
1084 metadata_access_disable();
1087 static inline int alloc_consistency_checks(struct kmem_cache *s,
1088 struct page *page, void *object)
1090 if (!check_slab(s, page))
1093 if (!check_valid_pointer(s, page, object)) {
1094 object_err(s, page, object, "Freelist Pointer check fails");
1098 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1104 static noinline int alloc_debug_processing(struct kmem_cache *s,
1106 void *object, unsigned long addr)
1108 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1109 if (!alloc_consistency_checks(s, page, object))
1113 /* Success perform special debug activities for allocs */
1114 if (s->flags & SLAB_STORE_USER)
1115 set_track(s, object, TRACK_ALLOC, addr);
1116 trace(s, page, object, 1);
1117 init_object(s, object, SLUB_RED_ACTIVE);
1121 if (PageSlab(page)) {
1123 * If this is a slab page then lets do the best we can
1124 * to avoid issues in the future. Marking all objects
1125 * as used avoids touching the remaining objects.
1127 slab_fix(s, "Marking all objects used");
1128 page->inuse = page->objects;
1129 page->freelist = NULL;
1134 static inline int free_consistency_checks(struct kmem_cache *s,
1135 struct page *page, void *object, unsigned long addr)
1137 if (!check_valid_pointer(s, page, object)) {
1138 slab_err(s, page, "Invalid object pointer 0x%p", object);
1142 if (on_freelist(s, page, object)) {
1143 object_err(s, page, object, "Object already free");
1147 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1150 if (unlikely(s != page->slab_cache)) {
1151 if (!PageSlab(page)) {
1152 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1154 } else if (!page->slab_cache) {
1155 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1159 object_err(s, page, object,
1160 "page slab pointer corrupt.");
1166 /* Supports checking bulk free of a constructed freelist */
1167 static noinline int free_debug_processing(
1168 struct kmem_cache *s, struct page *page,
1169 void *head, void *tail, int bulk_cnt,
1172 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1173 void *object = head;
1175 unsigned long uninitialized_var(flags);
1178 spin_lock_irqsave(&n->list_lock, flags);
1181 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1182 if (!check_slab(s, page))
1189 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1190 if (!free_consistency_checks(s, page, object, addr))
1194 if (s->flags & SLAB_STORE_USER)
1195 set_track(s, object, TRACK_FREE, addr);
1196 trace(s, page, object, 0);
1197 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1198 init_object(s, object, SLUB_RED_INACTIVE);
1200 /* Reached end of constructed freelist yet? */
1201 if (object != tail) {
1202 object = get_freepointer(s, object);
1208 if (cnt != bulk_cnt)
1209 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1213 spin_unlock_irqrestore(&n->list_lock, flags);
1215 slab_fix(s, "Object at 0x%p not freed", object);
1219 static int __init setup_slub_debug(char *str)
1221 slub_debug = DEBUG_DEFAULT_FLAGS;
1222 if (*str++ != '=' || !*str)
1224 * No options specified. Switch on full debugging.
1230 * No options but restriction on slabs. This means full
1231 * debugging for slabs matching a pattern.
1238 * Switch off all debugging measures.
1243 * Determine which debug features should be switched on
1245 for (; *str && *str != ','; str++) {
1246 switch (tolower(*str)) {
1248 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1251 slub_debug |= SLAB_RED_ZONE;
1254 slub_debug |= SLAB_POISON;
1257 slub_debug |= SLAB_STORE_USER;
1260 slub_debug |= SLAB_TRACE;
1263 slub_debug |= SLAB_FAILSLAB;
1267 * Avoid enabling debugging on caches if its minimum
1268 * order would increase as a result.
1270 disable_higher_order_debug = 1;
1273 pr_err("slub_debug option '%c' unknown. skipped\n",
1280 slub_debug_slabs = str + 1;
1282 if ((static_branch_unlikely(&init_on_alloc) ||
1283 static_branch_unlikely(&init_on_free)) &&
1284 (slub_debug & SLAB_POISON))
1285 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1289 __setup("slub_debug", setup_slub_debug);
1292 * kmem_cache_flags - apply debugging options to the cache
1293 * @object_size: the size of an object without meta data
1294 * @flags: flags to set
1295 * @name: name of the cache
1296 * @ctor: constructor function
1298 * Debug option(s) are applied to @flags. In addition to the debug
1299 * option(s), if a slab name (or multiple) is specified i.e.
1300 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1301 * then only the select slabs will receive the debug option(s).
1303 slab_flags_t kmem_cache_flags(unsigned int object_size,
1304 slab_flags_t flags, const char *name,
1305 void (*ctor)(void *))
1310 /* If slub_debug = 0, it folds into the if conditional. */
1311 if (!slub_debug_slabs)
1312 return flags | slub_debug;
1315 iter = slub_debug_slabs;
1320 end = strchrnul(iter, ',');
1322 glob = strnchr(iter, end - iter, '*');
1324 cmplen = glob - iter;
1326 cmplen = max_t(size_t, len, (end - iter));
1328 if (!strncmp(name, iter, cmplen)) {
1329 flags |= slub_debug;
1340 #else /* !CONFIG_SLUB_DEBUG */
1341 static inline void setup_object_debug(struct kmem_cache *s,
1342 struct page *page, void *object) {}
1343 static inline void setup_page_debug(struct kmem_cache *s,
1344 void *addr, int order) {}
1346 static inline int alloc_debug_processing(struct kmem_cache *s,
1347 struct page *page, void *object, unsigned long addr) { return 0; }
1349 static inline int free_debug_processing(
1350 struct kmem_cache *s, struct page *page,
1351 void *head, void *tail, int bulk_cnt,
1352 unsigned long addr) { return 0; }
1354 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1356 static inline int check_object(struct kmem_cache *s, struct page *page,
1357 void *object, u8 val) { return 1; }
1358 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1359 struct page *page) {}
1360 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1361 struct page *page) {}
1362 slab_flags_t kmem_cache_flags(unsigned int object_size,
1363 slab_flags_t flags, const char *name,
1364 void (*ctor)(void *))
1368 #define slub_debug 0
1370 #define disable_higher_order_debug 0
1372 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1374 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1376 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1378 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1381 #endif /* CONFIG_SLUB_DEBUG */
1384 * Hooks for other subsystems that check memory allocations. In a typical
1385 * production configuration these hooks all should produce no code at all.
1387 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1389 ptr = kasan_kmalloc_large(ptr, size, flags);
1390 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1391 kmemleak_alloc(ptr, size, 1, flags);
1395 static __always_inline void kfree_hook(void *x)
1398 kasan_kfree_large(x, _RET_IP_);
1401 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1403 kmemleak_free_recursive(x, s->flags);
1406 * Trouble is that we may no longer disable interrupts in the fast path
1407 * So in order to make the debug calls that expect irqs to be
1408 * disabled we need to disable interrupts temporarily.
1410 #ifdef CONFIG_LOCKDEP
1412 unsigned long flags;
1414 local_irq_save(flags);
1415 debug_check_no_locks_freed(x, s->object_size);
1416 local_irq_restore(flags);
1419 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1420 debug_check_no_obj_freed(x, s->object_size);
1422 /* KASAN might put x into memory quarantine, delaying its reuse */
1423 return kasan_slab_free(s, x, _RET_IP_);
1426 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1427 void **head, void **tail)
1432 void *old_tail = *tail ? *tail : *head;
1435 if (slab_want_init_on_free(s)) {
1440 next = get_freepointer(s, object);
1442 * Clear the object and the metadata, but don't touch
1445 memset(object, 0, s->object_size);
1446 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1448 memset((char *)object + s->inuse, 0,
1449 s->size - s->inuse - rsize);
1450 set_freepointer(s, object, p);
1452 } while (object != old_tail);
1456 * Compiler cannot detect this function can be removed if slab_free_hook()
1457 * evaluates to nothing. Thus, catch all relevant config debug options here.
1459 #if defined(CONFIG_LOCKDEP) || \
1460 defined(CONFIG_DEBUG_KMEMLEAK) || \
1461 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1462 defined(CONFIG_KASAN)
1466 /* Head and tail of the reconstructed freelist */
1472 next = get_freepointer(s, object);
1473 /* If object's reuse doesn't have to be delayed */
1474 if (!slab_free_hook(s, object)) {
1475 /* Move object to the new freelist */
1476 set_freepointer(s, object, *head);
1481 } while (object != old_tail);
1486 return *head != NULL;
1492 static void *setup_object(struct kmem_cache *s, struct page *page,
1495 setup_object_debug(s, page, object);
1496 object = kasan_init_slab_obj(s, object);
1497 if (unlikely(s->ctor)) {
1498 kasan_unpoison_object_data(s, object);
1500 kasan_poison_object_data(s, object);
1506 * Slab allocation and freeing
1508 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1509 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1512 unsigned int order = oo_order(oo);
1514 if (node == NUMA_NO_NODE)
1515 page = alloc_pages(flags, order);
1517 page = __alloc_pages_node(node, flags, order);
1519 if (page && charge_slab_page(page, flags, order, s)) {
1520 __free_pages(page, order);
1527 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1528 /* Pre-initialize the random sequence cache */
1529 static int init_cache_random_seq(struct kmem_cache *s)
1531 unsigned int count = oo_objects(s->oo);
1534 /* Bailout if already initialised */
1538 err = cache_random_seq_create(s, count, GFP_KERNEL);
1540 pr_err("SLUB: Unable to initialize free list for %s\n",
1545 /* Transform to an offset on the set of pages */
1546 if (s->random_seq) {
1549 for (i = 0; i < count; i++)
1550 s->random_seq[i] *= s->size;
1555 /* Initialize each random sequence freelist per cache */
1556 static void __init init_freelist_randomization(void)
1558 struct kmem_cache *s;
1560 mutex_lock(&slab_mutex);
1562 list_for_each_entry(s, &slab_caches, list)
1563 init_cache_random_seq(s);
1565 mutex_unlock(&slab_mutex);
1568 /* Get the next entry on the pre-computed freelist randomized */
1569 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1570 unsigned long *pos, void *start,
1571 unsigned long page_limit,
1572 unsigned long freelist_count)
1577 * If the target page allocation failed, the number of objects on the
1578 * page might be smaller than the usual size defined by the cache.
1581 idx = s->random_seq[*pos];
1583 if (*pos >= freelist_count)
1585 } while (unlikely(idx >= page_limit));
1587 return (char *)start + idx;
1590 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1591 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1596 unsigned long idx, pos, page_limit, freelist_count;
1598 if (page->objects < 2 || !s->random_seq)
1601 freelist_count = oo_objects(s->oo);
1602 pos = get_random_int() % freelist_count;
1604 page_limit = page->objects * s->size;
1605 start = fixup_red_left(s, page_address(page));
1607 /* First entry is used as the base of the freelist */
1608 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1610 cur = setup_object(s, page, cur);
1611 page->freelist = cur;
1613 for (idx = 1; idx < page->objects; idx++) {
1614 next = next_freelist_entry(s, page, &pos, start, page_limit,
1616 next = setup_object(s, page, next);
1617 set_freepointer(s, cur, next);
1620 set_freepointer(s, cur, NULL);
1625 static inline int init_cache_random_seq(struct kmem_cache *s)
1629 static inline void init_freelist_randomization(void) { }
1630 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1634 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1636 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1639 struct kmem_cache_order_objects oo = s->oo;
1641 void *start, *p, *next;
1645 flags &= gfp_allowed_mask;
1647 if (gfpflags_allow_blocking(flags))
1650 flags |= s->allocflags;
1653 * Let the initial higher-order allocation fail under memory pressure
1654 * so we fall-back to the minimum order allocation.
1656 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1657 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1658 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1660 page = alloc_slab_page(s, alloc_gfp, node, oo);
1661 if (unlikely(!page)) {
1665 * Allocation may have failed due to fragmentation.
1666 * Try a lower order alloc if possible
1668 page = alloc_slab_page(s, alloc_gfp, node, oo);
1669 if (unlikely(!page))
1671 stat(s, ORDER_FALLBACK);
1674 page->objects = oo_objects(oo);
1676 order = compound_order(page);
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, start, order);
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 static inline unsigned int tid_to_cpu(unsigned long tid)
2009 return tid % TID_STEP;
2012 static inline unsigned long tid_to_event(unsigned long tid)
2014 return tid / TID_STEP;
2017 static inline unsigned int init_tid(int cpu)
2022 static inline void note_cmpxchg_failure(const char *n,
2023 const struct kmem_cache *s, unsigned long tid)
2025 #ifdef SLUB_DEBUG_CMPXCHG
2026 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2028 pr_info("%s %s: cmpxchg redo ", n, s->name);
2030 #ifdef CONFIG_PREEMPT
2031 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2032 pr_warn("due to cpu change %d -> %d\n",
2033 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2036 if (tid_to_event(tid) != tid_to_event(actual_tid))
2037 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2038 tid_to_event(tid), tid_to_event(actual_tid));
2040 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2041 actual_tid, tid, next_tid(tid));
2043 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2046 static void init_kmem_cache_cpus(struct kmem_cache *s)
2050 for_each_possible_cpu(cpu)
2051 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2055 * Remove the cpu slab
2057 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2058 void *freelist, struct kmem_cache_cpu *c)
2060 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2061 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2063 enum slab_modes l = M_NONE, m = M_NONE;
2065 int tail = DEACTIVATE_TO_HEAD;
2069 if (page->freelist) {
2070 stat(s, DEACTIVATE_REMOTE_FREES);
2071 tail = DEACTIVATE_TO_TAIL;
2075 * Stage one: Free all available per cpu objects back
2076 * to the page freelist while it is still frozen. Leave the
2079 * There is no need to take the list->lock because the page
2082 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2084 unsigned long counters;
2087 prior = page->freelist;
2088 counters = page->counters;
2089 set_freepointer(s, freelist, prior);
2090 new.counters = counters;
2092 VM_BUG_ON(!new.frozen);
2094 } while (!__cmpxchg_double_slab(s, page,
2096 freelist, new.counters,
2097 "drain percpu freelist"));
2099 freelist = nextfree;
2103 * Stage two: Ensure that the page is unfrozen while the
2104 * list presence reflects the actual number of objects
2107 * We setup the list membership and then perform a cmpxchg
2108 * with the count. If there is a mismatch then the page
2109 * is not unfrozen but the page is on the wrong list.
2111 * Then we restart the process which may have to remove
2112 * the page from the list that we just put it on again
2113 * because the number of objects in the slab may have
2118 old.freelist = page->freelist;
2119 old.counters = page->counters;
2120 VM_BUG_ON(!old.frozen);
2122 /* Determine target state of the slab */
2123 new.counters = old.counters;
2126 set_freepointer(s, freelist, old.freelist);
2127 new.freelist = freelist;
2129 new.freelist = old.freelist;
2133 if (!new.inuse && n->nr_partial >= s->min_partial)
2135 else if (new.freelist) {
2140 * Taking the spinlock removes the possibility
2141 * that acquire_slab() will see a slab page that
2144 spin_lock(&n->list_lock);
2148 if (kmem_cache_debug(s) && !lock) {
2151 * This also ensures that the scanning of full
2152 * slabs from diagnostic functions will not see
2155 spin_lock(&n->list_lock);
2161 remove_partial(n, page);
2162 else if (l == M_FULL)
2163 remove_full(s, n, page);
2166 add_partial(n, page, tail);
2167 else if (m == M_FULL)
2168 add_full(s, n, page);
2172 if (!__cmpxchg_double_slab(s, page,
2173 old.freelist, old.counters,
2174 new.freelist, new.counters,
2179 spin_unlock(&n->list_lock);
2183 else if (m == M_FULL)
2184 stat(s, DEACTIVATE_FULL);
2185 else if (m == M_FREE) {
2186 stat(s, DEACTIVATE_EMPTY);
2187 discard_slab(s, page);
2196 * Unfreeze all the cpu partial slabs.
2198 * This function must be called with interrupts disabled
2199 * for the cpu using c (or some other guarantee must be there
2200 * to guarantee no concurrent accesses).
2202 static void unfreeze_partials(struct kmem_cache *s,
2203 struct kmem_cache_cpu *c)
2205 #ifdef CONFIG_SLUB_CPU_PARTIAL
2206 struct kmem_cache_node *n = NULL, *n2 = NULL;
2207 struct page *page, *discard_page = NULL;
2209 while ((page = c->partial)) {
2213 c->partial = page->next;
2215 n2 = get_node(s, page_to_nid(page));
2218 spin_unlock(&n->list_lock);
2221 spin_lock(&n->list_lock);
2226 old.freelist = page->freelist;
2227 old.counters = page->counters;
2228 VM_BUG_ON(!old.frozen);
2230 new.counters = old.counters;
2231 new.freelist = old.freelist;
2235 } while (!__cmpxchg_double_slab(s, page,
2236 old.freelist, old.counters,
2237 new.freelist, new.counters,
2238 "unfreezing slab"));
2240 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2241 page->next = discard_page;
2242 discard_page = page;
2244 add_partial(n, page, DEACTIVATE_TO_TAIL);
2245 stat(s, FREE_ADD_PARTIAL);
2250 spin_unlock(&n->list_lock);
2252 while (discard_page) {
2253 page = discard_page;
2254 discard_page = discard_page->next;
2256 stat(s, DEACTIVATE_EMPTY);
2257 discard_slab(s, page);
2260 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2264 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2265 * partial page slot if available.
2267 * If we did not find a slot then simply move all the partials to the
2268 * per node partial list.
2270 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2272 #ifdef CONFIG_SLUB_CPU_PARTIAL
2273 struct page *oldpage;
2281 oldpage = this_cpu_read(s->cpu_slab->partial);
2284 pobjects = oldpage->pobjects;
2285 pages = oldpage->pages;
2286 if (drain && pobjects > s->cpu_partial) {
2287 unsigned long flags;
2289 * partial array is full. Move the existing
2290 * set to the per node partial list.
2292 local_irq_save(flags);
2293 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2294 local_irq_restore(flags);
2298 stat(s, CPU_PARTIAL_DRAIN);
2303 pobjects += page->objects - page->inuse;
2305 page->pages = pages;
2306 page->pobjects = pobjects;
2307 page->next = oldpage;
2309 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2311 if (unlikely(!s->cpu_partial)) {
2312 unsigned long flags;
2314 local_irq_save(flags);
2315 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2316 local_irq_restore(flags);
2319 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2322 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2324 stat(s, CPUSLAB_FLUSH);
2325 deactivate_slab(s, c->page, c->freelist, c);
2327 c->tid = next_tid(c->tid);
2333 * Called from IPI handler with interrupts disabled.
2335 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2337 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2342 unfreeze_partials(s, c);
2345 static void flush_cpu_slab(void *d)
2347 struct kmem_cache *s = d;
2349 __flush_cpu_slab(s, smp_processor_id());
2352 static bool has_cpu_slab(int cpu, void *info)
2354 struct kmem_cache *s = info;
2355 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2357 return c->page || slub_percpu_partial(c);
2360 static void flush_all(struct kmem_cache *s)
2362 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2366 * Use the cpu notifier to insure that the cpu slabs are flushed when
2369 static int slub_cpu_dead(unsigned int cpu)
2371 struct kmem_cache *s;
2372 unsigned long flags;
2374 mutex_lock(&slab_mutex);
2375 list_for_each_entry(s, &slab_caches, list) {
2376 local_irq_save(flags);
2377 __flush_cpu_slab(s, cpu);
2378 local_irq_restore(flags);
2380 mutex_unlock(&slab_mutex);
2385 * Check if the objects in a per cpu structure fit numa
2386 * locality expectations.
2388 static inline int node_match(struct page *page, int node)
2391 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2397 #ifdef CONFIG_SLUB_DEBUG
2398 static int count_free(struct page *page)
2400 return page->objects - page->inuse;
2403 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2405 return atomic_long_read(&n->total_objects);
2407 #endif /* CONFIG_SLUB_DEBUG */
2409 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2410 static unsigned long count_partial(struct kmem_cache_node *n,
2411 int (*get_count)(struct page *))
2413 unsigned long flags;
2414 unsigned long x = 0;
2417 spin_lock_irqsave(&n->list_lock, flags);
2418 list_for_each_entry(page, &n->partial, slab_list)
2419 x += get_count(page);
2420 spin_unlock_irqrestore(&n->list_lock, flags);
2423 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2425 static noinline void
2426 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2428 #ifdef CONFIG_SLUB_DEBUG
2429 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2430 DEFAULT_RATELIMIT_BURST);
2432 struct kmem_cache_node *n;
2434 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2437 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2438 nid, gfpflags, &gfpflags);
2439 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2440 s->name, s->object_size, s->size, oo_order(s->oo),
2443 if (oo_order(s->min) > get_order(s->object_size))
2444 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2447 for_each_kmem_cache_node(s, node, n) {
2448 unsigned long nr_slabs;
2449 unsigned long nr_objs;
2450 unsigned long nr_free;
2452 nr_free = count_partial(n, count_free);
2453 nr_slabs = node_nr_slabs(n);
2454 nr_objs = node_nr_objs(n);
2456 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2457 node, nr_slabs, nr_objs, nr_free);
2462 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2463 int node, struct kmem_cache_cpu **pc)
2466 struct kmem_cache_cpu *c = *pc;
2469 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2471 freelist = get_partial(s, flags, node, c);
2476 page = new_slab(s, flags, node);
2478 c = raw_cpu_ptr(s->cpu_slab);
2483 * No other reference to the page yet so we can
2484 * muck around with it freely without cmpxchg
2486 freelist = page->freelist;
2487 page->freelist = NULL;
2489 stat(s, ALLOC_SLAB);
2497 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2499 if (unlikely(PageSlabPfmemalloc(page)))
2500 return gfp_pfmemalloc_allowed(gfpflags);
2506 * Check the page->freelist of a page and either transfer the freelist to the
2507 * per cpu freelist or deactivate the page.
2509 * The page is still frozen if the return value is not NULL.
2511 * If this function returns NULL then the page has been unfrozen.
2513 * This function must be called with interrupt disabled.
2515 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2518 unsigned long counters;
2522 freelist = page->freelist;
2523 counters = page->counters;
2525 new.counters = counters;
2526 VM_BUG_ON(!new.frozen);
2528 new.inuse = page->objects;
2529 new.frozen = freelist != NULL;
2531 } while (!__cmpxchg_double_slab(s, page,
2540 * Slow path. The lockless freelist is empty or we need to perform
2543 * Processing is still very fast if new objects have been freed to the
2544 * regular freelist. In that case we simply take over the regular freelist
2545 * as the lockless freelist and zap the regular freelist.
2547 * If that is not working then we fall back to the partial lists. We take the
2548 * first element of the freelist as the object to allocate now and move the
2549 * rest of the freelist to the lockless freelist.
2551 * And if we were unable to get a new slab from the partial slab lists then
2552 * we need to allocate a new slab. This is the slowest path since it involves
2553 * a call to the page allocator and the setup of a new slab.
2555 * Version of __slab_alloc to use when we know that interrupts are
2556 * already disabled (which is the case for bulk allocation).
2558 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2559 unsigned long addr, struct kmem_cache_cpu *c)
2569 if (unlikely(!node_match(page, node))) {
2570 int searchnode = node;
2572 if (node != NUMA_NO_NODE && !node_present_pages(node))
2573 searchnode = node_to_mem_node(node);
2575 if (unlikely(!node_match(page, searchnode))) {
2576 stat(s, ALLOC_NODE_MISMATCH);
2577 deactivate_slab(s, page, c->freelist, c);
2583 * By rights, we should be searching for a slab page that was
2584 * PFMEMALLOC but right now, we are losing the pfmemalloc
2585 * information when the page leaves the per-cpu allocator
2587 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2588 deactivate_slab(s, page, c->freelist, c);
2592 /* must check again c->freelist in case of cpu migration or IRQ */
2593 freelist = c->freelist;
2597 freelist = get_freelist(s, page);
2601 stat(s, DEACTIVATE_BYPASS);
2605 stat(s, ALLOC_REFILL);
2609 * freelist is pointing to the list of objects to be used.
2610 * page is pointing to the page from which the objects are obtained.
2611 * That page must be frozen for per cpu allocations to work.
2613 VM_BUG_ON(!c->page->frozen);
2614 c->freelist = get_freepointer(s, freelist);
2615 c->tid = next_tid(c->tid);
2620 if (slub_percpu_partial(c)) {
2621 page = c->page = slub_percpu_partial(c);
2622 slub_set_percpu_partial(c, page);
2623 stat(s, CPU_PARTIAL_ALLOC);
2627 freelist = new_slab_objects(s, gfpflags, node, &c);
2629 if (unlikely(!freelist)) {
2630 slab_out_of_memory(s, gfpflags, node);
2635 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2638 /* Only entered in the debug case */
2639 if (kmem_cache_debug(s) &&
2640 !alloc_debug_processing(s, page, freelist, addr))
2641 goto new_slab; /* Slab failed checks. Next slab needed */
2643 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2648 * Another one that disabled interrupt and compensates for possible
2649 * cpu changes by refetching the per cpu area pointer.
2651 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2652 unsigned long addr, struct kmem_cache_cpu *c)
2655 unsigned long flags;
2657 local_irq_save(flags);
2658 #ifdef CONFIG_PREEMPT
2660 * We may have been preempted and rescheduled on a different
2661 * cpu before disabling interrupts. Need to reload cpu area
2664 c = this_cpu_ptr(s->cpu_slab);
2667 p = ___slab_alloc(s, gfpflags, node, addr, c);
2668 local_irq_restore(flags);
2673 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2674 * have the fastpath folded into their functions. So no function call
2675 * overhead for requests that can be satisfied on the fastpath.
2677 * The fastpath works by first checking if the lockless freelist can be used.
2678 * If not then __slab_alloc is called for slow processing.
2680 * Otherwise we can simply pick the next object from the lockless free list.
2682 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2683 gfp_t gfpflags, int node, unsigned long addr)
2686 struct kmem_cache_cpu *c;
2690 s = slab_pre_alloc_hook(s, gfpflags);
2695 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2696 * enabled. We may switch back and forth between cpus while
2697 * reading from one cpu area. That does not matter as long
2698 * as we end up on the original cpu again when doing the cmpxchg.
2700 * We should guarantee that tid and kmem_cache are retrieved on
2701 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2702 * to check if it is matched or not.
2705 tid = this_cpu_read(s->cpu_slab->tid);
2706 c = raw_cpu_ptr(s->cpu_slab);
2707 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2708 unlikely(tid != READ_ONCE(c->tid)));
2711 * Irqless object alloc/free algorithm used here depends on sequence
2712 * of fetching cpu_slab's data. tid should be fetched before anything
2713 * on c to guarantee that object and page associated with previous tid
2714 * won't be used with current tid. If we fetch tid first, object and
2715 * page could be one associated with next tid and our alloc/free
2716 * request will be failed. In this case, we will retry. So, no problem.
2721 * The transaction ids are globally unique per cpu and per operation on
2722 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2723 * occurs on the right processor and that there was no operation on the
2724 * linked list in between.
2727 object = c->freelist;
2729 if (unlikely(!object || !node_match(page, node))) {
2730 object = __slab_alloc(s, gfpflags, node, addr, c);
2731 stat(s, ALLOC_SLOWPATH);
2733 void *next_object = get_freepointer_safe(s, object);
2736 * The cmpxchg will only match if there was no additional
2737 * operation and if we are on the right processor.
2739 * The cmpxchg does the following atomically (without lock
2741 * 1. Relocate first pointer to the current per cpu area.
2742 * 2. Verify that tid and freelist have not been changed
2743 * 3. If they were not changed replace tid and freelist
2745 * Since this is without lock semantics the protection is only
2746 * against code executing on this cpu *not* from access by
2749 if (unlikely(!this_cpu_cmpxchg_double(
2750 s->cpu_slab->freelist, s->cpu_slab->tid,
2752 next_object, next_tid(tid)))) {
2754 note_cmpxchg_failure("slab_alloc", s, tid);
2757 prefetch_freepointer(s, next_object);
2758 stat(s, ALLOC_FASTPATH);
2761 * If the object has been wiped upon free, make sure it's fully
2762 * initialized by zeroing out freelist pointer.
2764 if (unlikely(slab_want_init_on_free(s)) && object)
2765 memset(object + s->offset, 0, sizeof(void *));
2767 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2768 memset(object, 0, s->object_size);
2770 slab_post_alloc_hook(s, gfpflags, 1, &object);
2775 static __always_inline void *slab_alloc(struct kmem_cache *s,
2776 gfp_t gfpflags, unsigned long addr)
2778 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2781 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2783 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2785 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2790 EXPORT_SYMBOL(kmem_cache_alloc);
2792 #ifdef CONFIG_TRACING
2793 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2795 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2796 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2797 ret = kasan_kmalloc(s, ret, size, gfpflags);
2800 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2804 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2806 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2808 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2809 s->object_size, s->size, gfpflags, node);
2813 EXPORT_SYMBOL(kmem_cache_alloc_node);
2815 #ifdef CONFIG_TRACING
2816 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2818 int node, size_t size)
2820 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2822 trace_kmalloc_node(_RET_IP_, ret,
2823 size, s->size, gfpflags, node);
2825 ret = kasan_kmalloc(s, ret, size, gfpflags);
2828 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2830 #endif /* CONFIG_NUMA */
2833 * Slow path handling. This may still be called frequently since objects
2834 * have a longer lifetime than the cpu slabs in most processing loads.
2836 * So we still attempt to reduce cache line usage. Just take the slab
2837 * lock and free the item. If there is no additional partial page
2838 * handling required then we can return immediately.
2840 static void __slab_free(struct kmem_cache *s, struct page *page,
2841 void *head, void *tail, int cnt,
2848 unsigned long counters;
2849 struct kmem_cache_node *n = NULL;
2850 unsigned long uninitialized_var(flags);
2852 stat(s, FREE_SLOWPATH);
2854 if (kmem_cache_debug(s) &&
2855 !free_debug_processing(s, page, head, tail, cnt, addr))
2860 spin_unlock_irqrestore(&n->list_lock, flags);
2863 prior = page->freelist;
2864 counters = page->counters;
2865 set_freepointer(s, tail, prior);
2866 new.counters = counters;
2867 was_frozen = new.frozen;
2869 if ((!new.inuse || !prior) && !was_frozen) {
2871 if (kmem_cache_has_cpu_partial(s) && !prior) {
2874 * Slab was on no list before and will be
2876 * We can defer the list move and instead
2881 } else { /* Needs to be taken off a list */
2883 n = get_node(s, page_to_nid(page));
2885 * Speculatively acquire the list_lock.
2886 * If the cmpxchg does not succeed then we may
2887 * drop the list_lock without any processing.
2889 * Otherwise the list_lock will synchronize with
2890 * other processors updating the list of slabs.
2892 spin_lock_irqsave(&n->list_lock, flags);
2897 } while (!cmpxchg_double_slab(s, page,
2905 * If we just froze the page then put it onto the
2906 * per cpu partial list.
2908 if (new.frozen && !was_frozen) {
2909 put_cpu_partial(s, page, 1);
2910 stat(s, CPU_PARTIAL_FREE);
2913 * The list lock was not taken therefore no list
2914 * activity can be necessary.
2917 stat(s, FREE_FROZEN);
2921 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2925 * Objects left in the slab. If it was not on the partial list before
2928 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2929 remove_full(s, n, page);
2930 add_partial(n, page, DEACTIVATE_TO_TAIL);
2931 stat(s, FREE_ADD_PARTIAL);
2933 spin_unlock_irqrestore(&n->list_lock, flags);
2939 * Slab on the partial list.
2941 remove_partial(n, page);
2942 stat(s, FREE_REMOVE_PARTIAL);
2944 /* Slab must be on the full list */
2945 remove_full(s, n, page);
2948 spin_unlock_irqrestore(&n->list_lock, flags);
2950 discard_slab(s, page);
2954 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2955 * can perform fastpath freeing without additional function calls.
2957 * The fastpath is only possible if we are freeing to the current cpu slab
2958 * of this processor. This typically the case if we have just allocated
2961 * If fastpath is not possible then fall back to __slab_free where we deal
2962 * with all sorts of special processing.
2964 * Bulk free of a freelist with several objects (all pointing to the
2965 * same page) possible by specifying head and tail ptr, plus objects
2966 * count (cnt). Bulk free indicated by tail pointer being set.
2968 static __always_inline void do_slab_free(struct kmem_cache *s,
2969 struct page *page, void *head, void *tail,
2970 int cnt, unsigned long addr)
2972 void *tail_obj = tail ? : head;
2973 struct kmem_cache_cpu *c;
2977 * Determine the currently cpus per cpu slab.
2978 * The cpu may change afterward. However that does not matter since
2979 * data is retrieved via this pointer. If we are on the same cpu
2980 * during the cmpxchg then the free will succeed.
2983 tid = this_cpu_read(s->cpu_slab->tid);
2984 c = raw_cpu_ptr(s->cpu_slab);
2985 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2986 unlikely(tid != READ_ONCE(c->tid)));
2988 /* Same with comment on barrier() in slab_alloc_node() */
2991 if (likely(page == c->page)) {
2992 set_freepointer(s, tail_obj, c->freelist);
2994 if (unlikely(!this_cpu_cmpxchg_double(
2995 s->cpu_slab->freelist, s->cpu_slab->tid,
2997 head, next_tid(tid)))) {
2999 note_cmpxchg_failure("slab_free", s, tid);
3002 stat(s, FREE_FASTPATH);
3004 __slab_free(s, page, head, tail_obj, cnt, addr);
3008 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3009 void *head, void *tail, int cnt,
3013 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3014 * to remove objects, whose reuse must be delayed.
3016 if (slab_free_freelist_hook(s, &head, &tail))
3017 do_slab_free(s, page, head, tail, cnt, addr);
3020 #ifdef CONFIG_KASAN_GENERIC
3021 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3023 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3027 void kmem_cache_free(struct kmem_cache *s, void *x)
3029 s = cache_from_obj(s, x);
3032 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3033 trace_kmem_cache_free(_RET_IP_, x);
3035 EXPORT_SYMBOL(kmem_cache_free);
3037 struct detached_freelist {
3042 struct kmem_cache *s;
3046 * This function progressively scans the array with free objects (with
3047 * a limited look ahead) and extract objects belonging to the same
3048 * page. It builds a detached freelist directly within the given
3049 * page/objects. This can happen without any need for
3050 * synchronization, because the objects are owned by running process.
3051 * The freelist is build up as a single linked list in the objects.
3052 * The idea is, that this detached freelist can then be bulk
3053 * transferred to the real freelist(s), but only requiring a single
3054 * synchronization primitive. Look ahead in the array is limited due
3055 * to performance reasons.
3058 int build_detached_freelist(struct kmem_cache *s, size_t size,
3059 void **p, struct detached_freelist *df)
3061 size_t first_skipped_index = 0;
3066 /* Always re-init detached_freelist */
3071 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3072 } while (!object && size);
3077 page = virt_to_head_page(object);
3079 /* Handle kalloc'ed objects */
3080 if (unlikely(!PageSlab(page))) {
3081 BUG_ON(!PageCompound(page));
3083 __free_pages(page, compound_order(page));
3084 p[size] = NULL; /* mark object processed */
3087 /* Derive kmem_cache from object */
3088 df->s = page->slab_cache;
3090 df->s = cache_from_obj(s, object); /* Support for memcg */
3093 /* Start new detached freelist */
3095 set_freepointer(df->s, object, NULL);
3097 df->freelist = object;
3098 p[size] = NULL; /* mark object processed */
3104 continue; /* Skip processed objects */
3106 /* df->page is always set at this point */
3107 if (df->page == virt_to_head_page(object)) {
3108 /* Opportunity build freelist */
3109 set_freepointer(df->s, object, df->freelist);
3110 df->freelist = object;
3112 p[size] = NULL; /* mark object processed */
3117 /* Limit look ahead search */
3121 if (!first_skipped_index)
3122 first_skipped_index = size + 1;
3125 return first_skipped_index;
3128 /* Note that interrupts must be enabled when calling this function. */
3129 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3135 struct detached_freelist df;
3137 size = build_detached_freelist(s, size, p, &df);
3141 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3142 } while (likely(size));
3144 EXPORT_SYMBOL(kmem_cache_free_bulk);
3146 /* Note that interrupts must be enabled when calling this function. */
3147 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3150 struct kmem_cache_cpu *c;
3153 /* memcg and kmem_cache debug support */
3154 s = slab_pre_alloc_hook(s, flags);
3158 * Drain objects in the per cpu slab, while disabling local
3159 * IRQs, which protects against PREEMPT and interrupts
3160 * handlers invoking normal fastpath.
3162 local_irq_disable();
3163 c = this_cpu_ptr(s->cpu_slab);
3165 for (i = 0; i < size; i++) {
3166 void *object = c->freelist;
3168 if (unlikely(!object)) {
3170 * Invoking slow path likely have side-effect
3171 * of re-populating per CPU c->freelist
3173 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3175 if (unlikely(!p[i]))
3178 c = this_cpu_ptr(s->cpu_slab);
3179 continue; /* goto for-loop */
3181 c->freelist = get_freepointer(s, object);
3184 c->tid = next_tid(c->tid);
3187 /* Clear memory outside IRQ disabled fastpath loop */
3188 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3191 for (j = 0; j < i; j++)
3192 memset(p[j], 0, s->object_size);
3195 /* memcg and kmem_cache debug support */
3196 slab_post_alloc_hook(s, flags, size, p);
3200 slab_post_alloc_hook(s, flags, i, p);
3201 __kmem_cache_free_bulk(s, i, p);
3204 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3208 * Object placement in a slab is made very easy because we always start at
3209 * offset 0. If we tune the size of the object to the alignment then we can
3210 * get the required alignment by putting one properly sized object after
3213 * Notice that the allocation order determines the sizes of the per cpu
3214 * caches. Each processor has always one slab available for allocations.
3215 * Increasing the allocation order reduces the number of times that slabs
3216 * must be moved on and off the partial lists and is therefore a factor in
3221 * Mininum / Maximum order of slab pages. This influences locking overhead
3222 * and slab fragmentation. A higher order reduces the number of partial slabs
3223 * and increases the number of allocations possible without having to
3224 * take the list_lock.
3226 static unsigned int slub_min_order;
3227 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3228 static unsigned int slub_min_objects;
3231 * Calculate the order of allocation given an slab object size.
3233 * The order of allocation has significant impact on performance and other
3234 * system components. Generally order 0 allocations should be preferred since
3235 * order 0 does not cause fragmentation in the page allocator. Larger objects
3236 * be problematic to put into order 0 slabs because there may be too much
3237 * unused space left. We go to a higher order if more than 1/16th of the slab
3240 * In order to reach satisfactory performance we must ensure that a minimum
3241 * number of objects is in one slab. Otherwise we may generate too much
3242 * activity on the partial lists which requires taking the list_lock. This is
3243 * less a concern for large slabs though which are rarely used.
3245 * slub_max_order specifies the order where we begin to stop considering the
3246 * number of objects in a slab as critical. If we reach slub_max_order then
3247 * we try to keep the page order as low as possible. So we accept more waste
3248 * of space in favor of a small page order.
3250 * Higher order allocations also allow the placement of more objects in a
3251 * slab and thereby reduce object handling overhead. If the user has
3252 * requested a higher mininum order then we start with that one instead of
3253 * the smallest order which will fit the object.
3255 static inline unsigned int slab_order(unsigned int size,
3256 unsigned int min_objects, unsigned int max_order,
3257 unsigned int fract_leftover)
3259 unsigned int min_order = slub_min_order;
3262 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3263 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3265 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3266 order <= max_order; order++) {
3268 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3271 rem = slab_size % size;
3273 if (rem <= slab_size / fract_leftover)
3280 static inline int calculate_order(unsigned int size)
3283 unsigned int min_objects;
3284 unsigned int max_objects;
3287 * Attempt to find best configuration for a slab. This
3288 * works by first attempting to generate a layout with
3289 * the best configuration and backing off gradually.
3291 * First we increase the acceptable waste in a slab. Then
3292 * we reduce the minimum objects required in a slab.
3294 min_objects = slub_min_objects;
3296 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3297 max_objects = order_objects(slub_max_order, size);
3298 min_objects = min(min_objects, max_objects);
3300 while (min_objects > 1) {
3301 unsigned int fraction;
3304 while (fraction >= 4) {
3305 order = slab_order(size, min_objects,
3306 slub_max_order, fraction);
3307 if (order <= slub_max_order)
3315 * We were unable to place multiple objects in a slab. Now
3316 * lets see if we can place a single object there.
3318 order = slab_order(size, 1, slub_max_order, 1);
3319 if (order <= slub_max_order)
3323 * Doh this slab cannot be placed using slub_max_order.
3325 order = slab_order(size, 1, MAX_ORDER, 1);
3326 if (order < MAX_ORDER)
3332 init_kmem_cache_node(struct kmem_cache_node *n)
3335 spin_lock_init(&n->list_lock);
3336 INIT_LIST_HEAD(&n->partial);
3337 #ifdef CONFIG_SLUB_DEBUG
3338 atomic_long_set(&n->nr_slabs, 0);
3339 atomic_long_set(&n->total_objects, 0);
3340 INIT_LIST_HEAD(&n->full);
3344 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3346 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3347 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3350 * Must align to double word boundary for the double cmpxchg
3351 * instructions to work; see __pcpu_double_call_return_bool().
3353 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3354 2 * sizeof(void *));
3359 init_kmem_cache_cpus(s);
3364 static struct kmem_cache *kmem_cache_node;
3367 * No kmalloc_node yet so do it by hand. We know that this is the first
3368 * slab on the node for this slabcache. There are no concurrent accesses
3371 * Note that this function only works on the kmem_cache_node
3372 * when allocating for the kmem_cache_node. This is used for bootstrapping
3373 * memory on a fresh node that has no slab structures yet.
3375 static void early_kmem_cache_node_alloc(int node)
3378 struct kmem_cache_node *n;
3380 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3382 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3385 if (page_to_nid(page) != node) {
3386 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3387 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3392 #ifdef CONFIG_SLUB_DEBUG
3393 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3394 init_tracking(kmem_cache_node, n);
3396 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3398 page->freelist = get_freepointer(kmem_cache_node, n);
3401 kmem_cache_node->node[node] = n;
3402 init_kmem_cache_node(n);
3403 inc_slabs_node(kmem_cache_node, node, page->objects);
3406 * No locks need to be taken here as it has just been
3407 * initialized and there is no concurrent access.
3409 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3412 static void free_kmem_cache_nodes(struct kmem_cache *s)
3415 struct kmem_cache_node *n;
3417 for_each_kmem_cache_node(s, node, n) {
3418 s->node[node] = NULL;
3419 kmem_cache_free(kmem_cache_node, n);
3423 void __kmem_cache_release(struct kmem_cache *s)
3425 cache_random_seq_destroy(s);
3426 free_percpu(s->cpu_slab);
3427 free_kmem_cache_nodes(s);
3430 static int init_kmem_cache_nodes(struct kmem_cache *s)
3434 for_each_node_state(node, N_NORMAL_MEMORY) {
3435 struct kmem_cache_node *n;
3437 if (slab_state == DOWN) {
3438 early_kmem_cache_node_alloc(node);
3441 n = kmem_cache_alloc_node(kmem_cache_node,
3445 free_kmem_cache_nodes(s);
3449 init_kmem_cache_node(n);
3455 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3457 if (min < MIN_PARTIAL)
3459 else if (min > MAX_PARTIAL)
3461 s->min_partial = min;
3464 static void set_cpu_partial(struct kmem_cache *s)
3466 #ifdef CONFIG_SLUB_CPU_PARTIAL
3468 * cpu_partial determined the maximum number of objects kept in the
3469 * per cpu partial lists of a processor.
3471 * Per cpu partial lists mainly contain slabs that just have one
3472 * object freed. If they are used for allocation then they can be
3473 * filled up again with minimal effort. The slab will never hit the
3474 * per node partial lists and therefore no locking will be required.
3476 * This setting also determines
3478 * A) The number of objects from per cpu partial slabs dumped to the
3479 * per node list when we reach the limit.
3480 * B) The number of objects in cpu partial slabs to extract from the
3481 * per node list when we run out of per cpu objects. We only fetch
3482 * 50% to keep some capacity around for frees.
3484 if (!kmem_cache_has_cpu_partial(s))
3486 else if (s->size >= PAGE_SIZE)
3488 else if (s->size >= 1024)
3490 else if (s->size >= 256)
3491 s->cpu_partial = 13;
3493 s->cpu_partial = 30;
3498 * calculate_sizes() determines the order and the distribution of data within
3501 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3503 slab_flags_t flags = s->flags;
3504 unsigned int size = s->object_size;
3508 * Round up object size to the next word boundary. We can only
3509 * place the free pointer at word boundaries and this determines
3510 * the possible location of the free pointer.
3512 size = ALIGN(size, sizeof(void *));
3514 #ifdef CONFIG_SLUB_DEBUG
3516 * Determine if we can poison the object itself. If the user of
3517 * the slab may touch the object after free or before allocation
3518 * then we should never poison the object itself.
3520 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3522 s->flags |= __OBJECT_POISON;
3524 s->flags &= ~__OBJECT_POISON;
3528 * If we are Redzoning then check if there is some space between the
3529 * end of the object and the free pointer. If not then add an
3530 * additional word to have some bytes to store Redzone information.
3532 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3533 size += sizeof(void *);
3537 * With that we have determined the number of bytes in actual use
3538 * by the object. This is the potential offset to the free pointer.
3542 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3545 * Relocate free pointer after the object if it is not
3546 * permitted to overwrite the first word of the object on
3549 * This is the case if we do RCU, have a constructor or
3550 * destructor or are poisoning the objects.
3553 size += sizeof(void *);
3556 #ifdef CONFIG_SLUB_DEBUG
3557 if (flags & SLAB_STORE_USER)
3559 * Need to store information about allocs and frees after
3562 size += 2 * sizeof(struct track);
3565 kasan_cache_create(s, &size, &s->flags);
3566 #ifdef CONFIG_SLUB_DEBUG
3567 if (flags & SLAB_RED_ZONE) {
3569 * Add some empty padding so that we can catch
3570 * overwrites from earlier objects rather than let
3571 * tracking information or the free pointer be
3572 * corrupted if a user writes before the start
3575 size += sizeof(void *);
3577 s->red_left_pad = sizeof(void *);
3578 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3579 size += s->red_left_pad;
3584 * SLUB stores one object immediately after another beginning from
3585 * offset 0. In order to align the objects we have to simply size
3586 * each object to conform to the alignment.
3588 size = ALIGN(size, s->align);
3590 if (forced_order >= 0)
3591 order = forced_order;
3593 order = calculate_order(size);
3600 s->allocflags |= __GFP_COMP;
3602 if (s->flags & SLAB_CACHE_DMA)
3603 s->allocflags |= GFP_DMA;
3605 if (s->flags & SLAB_CACHE_DMA32)
3606 s->allocflags |= GFP_DMA32;
3608 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3609 s->allocflags |= __GFP_RECLAIMABLE;
3612 * Determine the number of objects per slab
3614 s->oo = oo_make(order, size);
3615 s->min = oo_make(get_order(size), size);
3616 if (oo_objects(s->oo) > oo_objects(s->max))
3619 return !!oo_objects(s->oo);
3622 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3624 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3625 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3626 s->random = get_random_long();
3629 if (!calculate_sizes(s, -1))
3631 if (disable_higher_order_debug) {
3633 * Disable debugging flags that store metadata if the min slab
3636 if (get_order(s->size) > get_order(s->object_size)) {
3637 s->flags &= ~DEBUG_METADATA_FLAGS;
3639 if (!calculate_sizes(s, -1))
3644 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3645 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3646 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3647 /* Enable fast mode */
3648 s->flags |= __CMPXCHG_DOUBLE;
3652 * The larger the object size is, the more pages we want on the partial
3653 * list to avoid pounding the page allocator excessively.
3655 set_min_partial(s, ilog2(s->size) / 2);
3660 s->remote_node_defrag_ratio = 1000;
3663 /* Initialize the pre-computed randomized freelist if slab is up */
3664 if (slab_state >= UP) {
3665 if (init_cache_random_seq(s))
3669 if (!init_kmem_cache_nodes(s))
3672 if (alloc_kmem_cache_cpus(s))
3675 free_kmem_cache_nodes(s);
3680 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3683 #ifdef CONFIG_SLUB_DEBUG
3684 void *addr = page_address(page);
3686 unsigned long *map = bitmap_zalloc(page->objects, GFP_ATOMIC);
3689 slab_err(s, page, text, s->name);
3692 get_map(s, page, map);
3693 for_each_object(p, s, addr, page->objects) {
3695 if (!test_bit(slab_index(p, s, addr), map)) {
3696 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3697 print_tracking(s, p);
3706 * Attempt to free all partial slabs on a node.
3707 * This is called from __kmem_cache_shutdown(). We must take list_lock
3708 * because sysfs file might still access partial list after the shutdowning.
3710 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3713 struct page *page, *h;
3715 BUG_ON(irqs_disabled());
3716 spin_lock_irq(&n->list_lock);
3717 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3719 remove_partial(n, page);
3720 list_add(&page->slab_list, &discard);
3722 list_slab_objects(s, page,
3723 "Objects remaining in %s on __kmem_cache_shutdown()");
3726 spin_unlock_irq(&n->list_lock);
3728 list_for_each_entry_safe(page, h, &discard, slab_list)
3729 discard_slab(s, page);
3732 bool __kmem_cache_empty(struct kmem_cache *s)
3735 struct kmem_cache_node *n;
3737 for_each_kmem_cache_node(s, node, n)
3738 if (n->nr_partial || slabs_node(s, node))
3744 * Release all resources used by a slab cache.
3746 int __kmem_cache_shutdown(struct kmem_cache *s)
3749 struct kmem_cache_node *n;
3752 /* Attempt to free all objects */
3753 for_each_kmem_cache_node(s, node, n) {
3755 if (n->nr_partial || slabs_node(s, node))
3758 sysfs_slab_remove(s);
3762 /********************************************************************
3764 *******************************************************************/
3766 static int __init setup_slub_min_order(char *str)
3768 get_option(&str, (int *)&slub_min_order);
3773 __setup("slub_min_order=", setup_slub_min_order);
3775 static int __init setup_slub_max_order(char *str)
3777 get_option(&str, (int *)&slub_max_order);
3778 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3783 __setup("slub_max_order=", setup_slub_max_order);
3785 static int __init setup_slub_min_objects(char *str)
3787 get_option(&str, (int *)&slub_min_objects);
3792 __setup("slub_min_objects=", setup_slub_min_objects);
3794 void *__kmalloc(size_t size, gfp_t flags)
3796 struct kmem_cache *s;
3799 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3800 return kmalloc_large(size, flags);
3802 s = kmalloc_slab(size, flags);
3804 if (unlikely(ZERO_OR_NULL_PTR(s)))
3807 ret = slab_alloc(s, flags, _RET_IP_);
3809 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3811 ret = kasan_kmalloc(s, ret, size, flags);
3815 EXPORT_SYMBOL(__kmalloc);
3818 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3823 flags |= __GFP_COMP;
3824 page = alloc_pages_node(node, flags, get_order(size));
3826 ptr = page_address(page);
3828 return kmalloc_large_node_hook(ptr, size, flags);
3831 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3833 struct kmem_cache *s;
3836 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3837 ret = kmalloc_large_node(size, flags, node);
3839 trace_kmalloc_node(_RET_IP_, ret,
3840 size, PAGE_SIZE << get_order(size),
3846 s = kmalloc_slab(size, flags);
3848 if (unlikely(ZERO_OR_NULL_PTR(s)))
3851 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3853 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3855 ret = kasan_kmalloc(s, ret, size, flags);
3859 EXPORT_SYMBOL(__kmalloc_node);
3860 #endif /* CONFIG_NUMA */
3862 #ifdef CONFIG_HARDENED_USERCOPY
3864 * Rejects incorrectly sized objects and objects that are to be copied
3865 * to/from userspace but do not fall entirely within the containing slab
3866 * cache's usercopy region.
3868 * Returns NULL if check passes, otherwise const char * to name of cache
3869 * to indicate an error.
3871 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3874 struct kmem_cache *s;
3875 unsigned int offset;
3878 ptr = kasan_reset_tag(ptr);
3880 /* Find object and usable object size. */
3881 s = page->slab_cache;
3883 /* Reject impossible pointers. */
3884 if (ptr < page_address(page))
3885 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3888 /* Find offset within object. */
3889 offset = (ptr - page_address(page)) % s->size;
3891 /* Adjust for redzone and reject if within the redzone. */
3892 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3893 if (offset < s->red_left_pad)
3894 usercopy_abort("SLUB object in left red zone",
3895 s->name, to_user, offset, n);
3896 offset -= s->red_left_pad;
3899 /* Allow address range falling entirely within usercopy region. */
3900 if (offset >= s->useroffset &&
3901 offset - s->useroffset <= s->usersize &&
3902 n <= s->useroffset - offset + s->usersize)
3906 * If the copy is still within the allocated object, produce
3907 * a warning instead of rejecting the copy. This is intended
3908 * to be a temporary method to find any missing usercopy
3911 object_size = slab_ksize(s);
3912 if (usercopy_fallback &&
3913 offset <= object_size && n <= object_size - offset) {
3914 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3918 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3920 #endif /* CONFIG_HARDENED_USERCOPY */
3922 size_t __ksize(const void *object)
3926 if (unlikely(object == ZERO_SIZE_PTR))
3929 page = virt_to_head_page(object);
3931 if (unlikely(!PageSlab(page))) {
3932 WARN_ON(!PageCompound(page));
3933 return PAGE_SIZE << compound_order(page);
3936 return slab_ksize(page->slab_cache);
3938 EXPORT_SYMBOL(__ksize);
3940 void kfree(const void *x)
3943 void *object = (void *)x;
3945 trace_kfree(_RET_IP_, x);
3947 if (unlikely(ZERO_OR_NULL_PTR(x)))
3950 page = virt_to_head_page(x);
3951 if (unlikely(!PageSlab(page))) {
3952 BUG_ON(!PageCompound(page));
3954 __free_pages(page, compound_order(page));
3957 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3959 EXPORT_SYMBOL(kfree);
3961 #define SHRINK_PROMOTE_MAX 32
3964 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3965 * up most to the head of the partial lists. New allocations will then
3966 * fill those up and thus they can be removed from the partial lists.
3968 * The slabs with the least items are placed last. This results in them
3969 * being allocated from last increasing the chance that the last objects
3970 * are freed in them.
3972 int __kmem_cache_shrink(struct kmem_cache *s)
3976 struct kmem_cache_node *n;
3979 struct list_head discard;
3980 struct list_head promote[SHRINK_PROMOTE_MAX];
3981 unsigned long flags;
3985 for_each_kmem_cache_node(s, node, n) {
3986 INIT_LIST_HEAD(&discard);
3987 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3988 INIT_LIST_HEAD(promote + i);
3990 spin_lock_irqsave(&n->list_lock, flags);
3993 * Build lists of slabs to discard or promote.
3995 * Note that concurrent frees may occur while we hold the
3996 * list_lock. page->inuse here is the upper limit.
3998 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
3999 int free = page->objects - page->inuse;
4001 /* Do not reread page->inuse */
4004 /* We do not keep full slabs on the list */
4007 if (free == page->objects) {
4008 list_move(&page->slab_list, &discard);
4010 } else if (free <= SHRINK_PROMOTE_MAX)
4011 list_move(&page->slab_list, promote + free - 1);
4015 * Promote the slabs filled up most to the head of the
4018 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4019 list_splice(promote + i, &n->partial);
4021 spin_unlock_irqrestore(&n->list_lock, flags);
4023 /* Release empty slabs */
4024 list_for_each_entry_safe(page, t, &discard, slab_list)
4025 discard_slab(s, page);
4027 if (slabs_node(s, node))
4035 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4038 * Called with all the locks held after a sched RCU grace period.
4039 * Even if @s becomes empty after shrinking, we can't know that @s
4040 * doesn't have allocations already in-flight and thus can't
4041 * destroy @s until the associated memcg is released.
4043 * However, let's remove the sysfs files for empty caches here.
4044 * Each cache has a lot of interface files which aren't
4045 * particularly useful for empty draining caches; otherwise, we can
4046 * easily end up with millions of unnecessary sysfs files on
4047 * systems which have a lot of memory and transient cgroups.
4049 if (!__kmem_cache_shrink(s))
4050 sysfs_slab_remove(s);
4053 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4056 * Disable empty slabs caching. Used to avoid pinning offline
4057 * memory cgroups by kmem pages that can be freed.
4059 slub_set_cpu_partial(s, 0);
4062 #endif /* CONFIG_MEMCG */
4064 static int slab_mem_going_offline_callback(void *arg)
4066 struct kmem_cache *s;
4068 mutex_lock(&slab_mutex);
4069 list_for_each_entry(s, &slab_caches, list)
4070 __kmem_cache_shrink(s);
4071 mutex_unlock(&slab_mutex);
4076 static void slab_mem_offline_callback(void *arg)
4078 struct kmem_cache_node *n;
4079 struct kmem_cache *s;
4080 struct memory_notify *marg = arg;
4083 offline_node = marg->status_change_nid_normal;
4086 * If the node still has available memory. we need kmem_cache_node
4089 if (offline_node < 0)
4092 mutex_lock(&slab_mutex);
4093 list_for_each_entry(s, &slab_caches, list) {
4094 n = get_node(s, offline_node);
4097 * if n->nr_slabs > 0, slabs still exist on the node
4098 * that is going down. We were unable to free them,
4099 * and offline_pages() function shouldn't call this
4100 * callback. So, we must fail.
4102 BUG_ON(slabs_node(s, offline_node));
4104 s->node[offline_node] = NULL;
4105 kmem_cache_free(kmem_cache_node, n);
4108 mutex_unlock(&slab_mutex);
4111 static int slab_mem_going_online_callback(void *arg)
4113 struct kmem_cache_node *n;
4114 struct kmem_cache *s;
4115 struct memory_notify *marg = arg;
4116 int nid = marg->status_change_nid_normal;
4120 * If the node's memory is already available, then kmem_cache_node is
4121 * already created. Nothing to do.
4127 * We are bringing a node online. No memory is available yet. We must
4128 * allocate a kmem_cache_node structure in order to bring the node
4131 mutex_lock(&slab_mutex);
4132 list_for_each_entry(s, &slab_caches, list) {
4134 * XXX: kmem_cache_alloc_node will fallback to other nodes
4135 * since memory is not yet available from the node that
4138 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4143 init_kmem_cache_node(n);
4147 mutex_unlock(&slab_mutex);
4151 static int slab_memory_callback(struct notifier_block *self,
4152 unsigned long action, void *arg)
4157 case MEM_GOING_ONLINE:
4158 ret = slab_mem_going_online_callback(arg);
4160 case MEM_GOING_OFFLINE:
4161 ret = slab_mem_going_offline_callback(arg);
4164 case MEM_CANCEL_ONLINE:
4165 slab_mem_offline_callback(arg);
4168 case MEM_CANCEL_OFFLINE:
4172 ret = notifier_from_errno(ret);
4178 static struct notifier_block slab_memory_callback_nb = {
4179 .notifier_call = slab_memory_callback,
4180 .priority = SLAB_CALLBACK_PRI,
4183 /********************************************************************
4184 * Basic setup of slabs
4185 *******************************************************************/
4188 * Used for early kmem_cache structures that were allocated using
4189 * the page allocator. Allocate them properly then fix up the pointers
4190 * that may be pointing to the wrong kmem_cache structure.
4193 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4196 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4197 struct kmem_cache_node *n;
4199 memcpy(s, static_cache, kmem_cache->object_size);
4202 * This runs very early, and only the boot processor is supposed to be
4203 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4206 __flush_cpu_slab(s, smp_processor_id());
4207 for_each_kmem_cache_node(s, node, n) {
4210 list_for_each_entry(p, &n->partial, slab_list)
4213 #ifdef CONFIG_SLUB_DEBUG
4214 list_for_each_entry(p, &n->full, slab_list)
4218 slab_init_memcg_params(s);
4219 list_add(&s->list, &slab_caches);
4220 memcg_link_cache(s, NULL);
4224 void __init kmem_cache_init(void)
4226 static __initdata struct kmem_cache boot_kmem_cache,
4227 boot_kmem_cache_node;
4229 if (debug_guardpage_minorder())
4232 kmem_cache_node = &boot_kmem_cache_node;
4233 kmem_cache = &boot_kmem_cache;
4235 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4236 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4238 register_hotmemory_notifier(&slab_memory_callback_nb);
4240 /* Able to allocate the per node structures */
4241 slab_state = PARTIAL;
4243 create_boot_cache(kmem_cache, "kmem_cache",
4244 offsetof(struct kmem_cache, node) +
4245 nr_node_ids * sizeof(struct kmem_cache_node *),
4246 SLAB_HWCACHE_ALIGN, 0, 0);
4248 kmem_cache = bootstrap(&boot_kmem_cache);
4249 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4251 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4252 setup_kmalloc_cache_index_table();
4253 create_kmalloc_caches(0);
4255 /* Setup random freelists for each cache */
4256 init_freelist_randomization();
4258 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4261 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4263 slub_min_order, slub_max_order, slub_min_objects,
4264 nr_cpu_ids, nr_node_ids);
4267 void __init kmem_cache_init_late(void)
4272 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4273 slab_flags_t flags, void (*ctor)(void *))
4275 struct kmem_cache *s, *c;
4277 s = find_mergeable(size, align, flags, name, ctor);
4282 * Adjust the object sizes so that we clear
4283 * the complete object on kzalloc.
4285 s->object_size = max(s->object_size, size);
4286 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4288 for_each_memcg_cache(c, s) {
4289 c->object_size = s->object_size;
4290 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4293 if (sysfs_slab_alias(s, name)) {
4302 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4306 err = kmem_cache_open(s, flags);
4310 /* Mutex is not taken during early boot */
4311 if (slab_state <= UP)
4314 memcg_propagate_slab_attrs(s);
4315 err = sysfs_slab_add(s);
4317 __kmem_cache_release(s);
4322 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4324 struct kmem_cache *s;
4327 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4328 return kmalloc_large(size, gfpflags);
4330 s = kmalloc_slab(size, gfpflags);
4332 if (unlikely(ZERO_OR_NULL_PTR(s)))
4335 ret = slab_alloc(s, gfpflags, caller);
4337 /* Honor the call site pointer we received. */
4338 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4344 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4345 int node, unsigned long caller)
4347 struct kmem_cache *s;
4350 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4351 ret = kmalloc_large_node(size, gfpflags, node);
4353 trace_kmalloc_node(caller, ret,
4354 size, PAGE_SIZE << get_order(size),
4360 s = kmalloc_slab(size, gfpflags);
4362 if (unlikely(ZERO_OR_NULL_PTR(s)))
4365 ret = slab_alloc_node(s, gfpflags, node, caller);
4367 /* Honor the call site pointer we received. */
4368 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4375 static int count_inuse(struct page *page)
4380 static int count_total(struct page *page)
4382 return page->objects;
4386 #ifdef CONFIG_SLUB_DEBUG
4387 static int validate_slab(struct kmem_cache *s, struct page *page,
4391 void *addr = page_address(page);
4393 if (!check_slab(s, page) ||
4394 !on_freelist(s, page, NULL))
4397 /* Now we know that a valid freelist exists */
4398 bitmap_zero(map, page->objects);
4400 get_map(s, page, map);
4401 for_each_object(p, s, addr, page->objects) {
4402 if (test_bit(slab_index(p, s, addr), map))
4403 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4407 for_each_object(p, s, addr, page->objects)
4408 if (!test_bit(slab_index(p, s, addr), map))
4409 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4414 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4418 validate_slab(s, page, map);
4422 static int validate_slab_node(struct kmem_cache *s,
4423 struct kmem_cache_node *n, unsigned long *map)
4425 unsigned long count = 0;
4427 unsigned long flags;
4429 spin_lock_irqsave(&n->list_lock, flags);
4431 list_for_each_entry(page, &n->partial, slab_list) {
4432 validate_slab_slab(s, page, map);
4435 if (count != n->nr_partial)
4436 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4437 s->name, count, n->nr_partial);
4439 if (!(s->flags & SLAB_STORE_USER))
4442 list_for_each_entry(page, &n->full, slab_list) {
4443 validate_slab_slab(s, page, map);
4446 if (count != atomic_long_read(&n->nr_slabs))
4447 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4448 s->name, count, atomic_long_read(&n->nr_slabs));
4451 spin_unlock_irqrestore(&n->list_lock, flags);
4455 static long validate_slab_cache(struct kmem_cache *s)
4458 unsigned long count = 0;
4459 struct kmem_cache_node *n;
4460 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4466 for_each_kmem_cache_node(s, node, n)
4467 count += validate_slab_node(s, n, map);
4472 * Generate lists of code addresses where slabcache objects are allocated
4477 unsigned long count;
4484 DECLARE_BITMAP(cpus, NR_CPUS);
4490 unsigned long count;
4491 struct location *loc;
4494 static void free_loc_track(struct loc_track *t)
4497 free_pages((unsigned long)t->loc,
4498 get_order(sizeof(struct location) * t->max));
4501 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4506 order = get_order(sizeof(struct location) * max);
4508 l = (void *)__get_free_pages(flags, order);
4513 memcpy(l, t->loc, sizeof(struct location) * t->count);
4521 static int add_location(struct loc_track *t, struct kmem_cache *s,
4522 const struct track *track)
4524 long start, end, pos;
4526 unsigned long caddr;
4527 unsigned long age = jiffies - track->when;
4533 pos = start + (end - start + 1) / 2;
4536 * There is nothing at "end". If we end up there
4537 * we need to add something to before end.
4542 caddr = t->loc[pos].addr;
4543 if (track->addr == caddr) {
4549 if (age < l->min_time)
4551 if (age > l->max_time)
4554 if (track->pid < l->min_pid)
4555 l->min_pid = track->pid;
4556 if (track->pid > l->max_pid)
4557 l->max_pid = track->pid;
4559 cpumask_set_cpu(track->cpu,
4560 to_cpumask(l->cpus));
4562 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4566 if (track->addr < caddr)
4573 * Not found. Insert new tracking element.
4575 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4581 (t->count - pos) * sizeof(struct location));
4584 l->addr = track->addr;
4588 l->min_pid = track->pid;
4589 l->max_pid = track->pid;
4590 cpumask_clear(to_cpumask(l->cpus));
4591 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4592 nodes_clear(l->nodes);
4593 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4597 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4598 struct page *page, enum track_item alloc,
4601 void *addr = page_address(page);
4604 bitmap_zero(map, page->objects);
4605 get_map(s, page, map);
4607 for_each_object(p, s, addr, page->objects)
4608 if (!test_bit(slab_index(p, s, addr), map))
4609 add_location(t, s, get_track(s, p, alloc));
4612 static int list_locations(struct kmem_cache *s, char *buf,
4613 enum track_item alloc)
4617 struct loc_track t = { 0, 0, NULL };
4619 struct kmem_cache_node *n;
4620 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4622 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4625 return sprintf(buf, "Out of memory\n");
4627 /* Push back cpu slabs */
4630 for_each_kmem_cache_node(s, node, n) {
4631 unsigned long flags;
4634 if (!atomic_long_read(&n->nr_slabs))
4637 spin_lock_irqsave(&n->list_lock, flags);
4638 list_for_each_entry(page, &n->partial, slab_list)
4639 process_slab(&t, s, page, alloc, map);
4640 list_for_each_entry(page, &n->full, slab_list)
4641 process_slab(&t, s, page, alloc, map);
4642 spin_unlock_irqrestore(&n->list_lock, flags);
4645 for (i = 0; i < t.count; i++) {
4646 struct location *l = &t.loc[i];
4648 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4650 len += sprintf(buf + len, "%7ld ", l->count);
4653 len += sprintf(buf + len, "%pS", (void *)l->addr);
4655 len += sprintf(buf + len, "<not-available>");
4657 if (l->sum_time != l->min_time) {
4658 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4660 (long)div_u64(l->sum_time, l->count),
4663 len += sprintf(buf + len, " age=%ld",
4666 if (l->min_pid != l->max_pid)
4667 len += sprintf(buf + len, " pid=%ld-%ld",
4668 l->min_pid, l->max_pid);
4670 len += sprintf(buf + len, " pid=%ld",
4673 if (num_online_cpus() > 1 &&
4674 !cpumask_empty(to_cpumask(l->cpus)) &&
4675 len < PAGE_SIZE - 60)
4676 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4678 cpumask_pr_args(to_cpumask(l->cpus)));
4680 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4681 len < PAGE_SIZE - 60)
4682 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4684 nodemask_pr_args(&l->nodes));
4686 len += sprintf(buf + len, "\n");
4692 len += sprintf(buf, "No data\n");
4695 #endif /* CONFIG_SLUB_DEBUG */
4697 #ifdef SLUB_RESILIENCY_TEST
4698 static void __init resiliency_test(void)
4701 int type = KMALLOC_NORMAL;
4703 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4705 pr_err("SLUB resiliency testing\n");
4706 pr_err("-----------------------\n");
4707 pr_err("A. Corruption after allocation\n");
4709 p = kzalloc(16, GFP_KERNEL);
4711 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4714 validate_slab_cache(kmalloc_caches[type][4]);
4716 /* Hmmm... The next two are dangerous */
4717 p = kzalloc(32, GFP_KERNEL);
4718 p[32 + sizeof(void *)] = 0x34;
4719 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4721 pr_err("If allocated object is overwritten then not detectable\n\n");
4723 validate_slab_cache(kmalloc_caches[type][5]);
4724 p = kzalloc(64, GFP_KERNEL);
4725 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4727 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4729 pr_err("If allocated object is overwritten then not detectable\n\n");
4730 validate_slab_cache(kmalloc_caches[type][6]);
4732 pr_err("\nB. Corruption after free\n");
4733 p = kzalloc(128, GFP_KERNEL);
4736 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4737 validate_slab_cache(kmalloc_caches[type][7]);
4739 p = kzalloc(256, GFP_KERNEL);
4742 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4743 validate_slab_cache(kmalloc_caches[type][8]);
4745 p = kzalloc(512, GFP_KERNEL);
4748 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4749 validate_slab_cache(kmalloc_caches[type][9]);
4753 static void resiliency_test(void) {};
4755 #endif /* SLUB_RESILIENCY_TEST */
4758 enum slab_stat_type {
4759 SL_ALL, /* All slabs */
4760 SL_PARTIAL, /* Only partially allocated slabs */
4761 SL_CPU, /* Only slabs used for cpu caches */
4762 SL_OBJECTS, /* Determine allocated objects not slabs */
4763 SL_TOTAL /* Determine object capacity not slabs */
4766 #define SO_ALL (1 << SL_ALL)
4767 #define SO_PARTIAL (1 << SL_PARTIAL)
4768 #define SO_CPU (1 << SL_CPU)
4769 #define SO_OBJECTS (1 << SL_OBJECTS)
4770 #define SO_TOTAL (1 << SL_TOTAL)
4773 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4775 static int __init setup_slub_memcg_sysfs(char *str)
4779 if (get_option(&str, &v) > 0)
4780 memcg_sysfs_enabled = v;
4785 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4788 static ssize_t show_slab_objects(struct kmem_cache *s,
4789 char *buf, unsigned long flags)
4791 unsigned long total = 0;
4794 unsigned long *nodes;
4796 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4800 if (flags & SO_CPU) {
4803 for_each_possible_cpu(cpu) {
4804 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4809 page = READ_ONCE(c->page);
4813 node = page_to_nid(page);
4814 if (flags & SO_TOTAL)
4816 else if (flags & SO_OBJECTS)
4824 page = slub_percpu_partial_read_once(c);
4826 node = page_to_nid(page);
4827 if (flags & SO_TOTAL)
4829 else if (flags & SO_OBJECTS)
4840 #ifdef CONFIG_SLUB_DEBUG
4841 if (flags & SO_ALL) {
4842 struct kmem_cache_node *n;
4844 for_each_kmem_cache_node(s, node, n) {
4846 if (flags & SO_TOTAL)
4847 x = atomic_long_read(&n->total_objects);
4848 else if (flags & SO_OBJECTS)
4849 x = atomic_long_read(&n->total_objects) -
4850 count_partial(n, count_free);
4852 x = atomic_long_read(&n->nr_slabs);
4859 if (flags & SO_PARTIAL) {
4860 struct kmem_cache_node *n;
4862 for_each_kmem_cache_node(s, node, n) {
4863 if (flags & SO_TOTAL)
4864 x = count_partial(n, count_total);
4865 else if (flags & SO_OBJECTS)
4866 x = count_partial(n, count_inuse);
4873 x = sprintf(buf, "%lu", total);
4875 for (node = 0; node < nr_node_ids; node++)
4877 x += sprintf(buf + x, " N%d=%lu",
4882 return x + sprintf(buf + x, "\n");
4885 #ifdef CONFIG_SLUB_DEBUG
4886 static int any_slab_objects(struct kmem_cache *s)
4889 struct kmem_cache_node *n;
4891 for_each_kmem_cache_node(s, node, n)
4892 if (atomic_long_read(&n->total_objects))
4899 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4900 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4902 struct slab_attribute {
4903 struct attribute attr;
4904 ssize_t (*show)(struct kmem_cache *s, char *buf);
4905 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4908 #define SLAB_ATTR_RO(_name) \
4909 static struct slab_attribute _name##_attr = \
4910 __ATTR(_name, 0400, _name##_show, NULL)
4912 #define SLAB_ATTR(_name) \
4913 static struct slab_attribute _name##_attr = \
4914 __ATTR(_name, 0600, _name##_show, _name##_store)
4916 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4918 return sprintf(buf, "%u\n", s->size);
4920 SLAB_ATTR_RO(slab_size);
4922 static ssize_t align_show(struct kmem_cache *s, char *buf)
4924 return sprintf(buf, "%u\n", s->align);
4926 SLAB_ATTR_RO(align);
4928 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4930 return sprintf(buf, "%u\n", s->object_size);
4932 SLAB_ATTR_RO(object_size);
4934 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4936 return sprintf(buf, "%u\n", oo_objects(s->oo));
4938 SLAB_ATTR_RO(objs_per_slab);
4940 static ssize_t order_store(struct kmem_cache *s,
4941 const char *buf, size_t length)
4946 err = kstrtouint(buf, 10, &order);
4950 if (order > slub_max_order || order < slub_min_order)
4953 calculate_sizes(s, order);
4957 static ssize_t order_show(struct kmem_cache *s, char *buf)
4959 return sprintf(buf, "%u\n", oo_order(s->oo));
4963 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4965 return sprintf(buf, "%lu\n", s->min_partial);
4968 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4974 err = kstrtoul(buf, 10, &min);
4978 set_min_partial(s, min);
4981 SLAB_ATTR(min_partial);
4983 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4985 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4988 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4991 unsigned int objects;
4994 err = kstrtouint(buf, 10, &objects);
4997 if (objects && !kmem_cache_has_cpu_partial(s))
5000 slub_set_cpu_partial(s, objects);
5004 SLAB_ATTR(cpu_partial);
5006 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5010 return sprintf(buf, "%pS\n", s->ctor);
5014 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5016 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5018 SLAB_ATTR_RO(aliases);
5020 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5022 return show_slab_objects(s, buf, SO_PARTIAL);
5024 SLAB_ATTR_RO(partial);
5026 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5028 return show_slab_objects(s, buf, SO_CPU);
5030 SLAB_ATTR_RO(cpu_slabs);
5032 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5034 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5036 SLAB_ATTR_RO(objects);
5038 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5040 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5042 SLAB_ATTR_RO(objects_partial);
5044 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5051 for_each_online_cpu(cpu) {
5054 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5057 pages += page->pages;
5058 objects += page->pobjects;
5062 len = sprintf(buf, "%d(%d)", objects, pages);
5065 for_each_online_cpu(cpu) {
5068 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5070 if (page && len < PAGE_SIZE - 20)
5071 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5072 page->pobjects, page->pages);
5075 return len + sprintf(buf + len, "\n");
5077 SLAB_ATTR_RO(slabs_cpu_partial);
5079 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5081 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5084 static ssize_t reclaim_account_store(struct kmem_cache *s,
5085 const char *buf, size_t length)
5087 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5089 s->flags |= SLAB_RECLAIM_ACCOUNT;
5092 SLAB_ATTR(reclaim_account);
5094 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5096 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5098 SLAB_ATTR_RO(hwcache_align);
5100 #ifdef CONFIG_ZONE_DMA
5101 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5103 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5105 SLAB_ATTR_RO(cache_dma);
5108 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5110 return sprintf(buf, "%u\n", s->usersize);
5112 SLAB_ATTR_RO(usersize);
5114 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5116 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5118 SLAB_ATTR_RO(destroy_by_rcu);
5120 #ifdef CONFIG_SLUB_DEBUG
5121 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5123 return show_slab_objects(s, buf, SO_ALL);
5125 SLAB_ATTR_RO(slabs);
5127 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5129 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5131 SLAB_ATTR_RO(total_objects);
5133 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5135 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5138 static ssize_t sanity_checks_store(struct kmem_cache *s,
5139 const char *buf, size_t length)
5141 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5142 if (buf[0] == '1') {
5143 s->flags &= ~__CMPXCHG_DOUBLE;
5144 s->flags |= SLAB_CONSISTENCY_CHECKS;
5148 SLAB_ATTR(sanity_checks);
5150 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5152 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5155 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5159 * Tracing a merged cache is going to give confusing results
5160 * as well as cause other issues like converting a mergeable
5161 * cache into an umergeable one.
5163 if (s->refcount > 1)
5166 s->flags &= ~SLAB_TRACE;
5167 if (buf[0] == '1') {
5168 s->flags &= ~__CMPXCHG_DOUBLE;
5169 s->flags |= SLAB_TRACE;
5175 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5177 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5180 static ssize_t red_zone_store(struct kmem_cache *s,
5181 const char *buf, size_t length)
5183 if (any_slab_objects(s))
5186 s->flags &= ~SLAB_RED_ZONE;
5187 if (buf[0] == '1') {
5188 s->flags |= SLAB_RED_ZONE;
5190 calculate_sizes(s, -1);
5193 SLAB_ATTR(red_zone);
5195 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5197 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5200 static ssize_t poison_store(struct kmem_cache *s,
5201 const char *buf, size_t length)
5203 if (any_slab_objects(s))
5206 s->flags &= ~SLAB_POISON;
5207 if (buf[0] == '1') {
5208 s->flags |= SLAB_POISON;
5210 calculate_sizes(s, -1);
5215 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5217 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5220 static ssize_t store_user_store(struct kmem_cache *s,
5221 const char *buf, size_t length)
5223 if (any_slab_objects(s))
5226 s->flags &= ~SLAB_STORE_USER;
5227 if (buf[0] == '1') {
5228 s->flags &= ~__CMPXCHG_DOUBLE;
5229 s->flags |= SLAB_STORE_USER;
5231 calculate_sizes(s, -1);
5234 SLAB_ATTR(store_user);
5236 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5241 static ssize_t validate_store(struct kmem_cache *s,
5242 const char *buf, size_t length)
5246 if (buf[0] == '1') {
5247 ret = validate_slab_cache(s);
5253 SLAB_ATTR(validate);
5255 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5257 if (!(s->flags & SLAB_STORE_USER))
5259 return list_locations(s, buf, TRACK_ALLOC);
5261 SLAB_ATTR_RO(alloc_calls);
5263 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5265 if (!(s->flags & SLAB_STORE_USER))
5267 return list_locations(s, buf, TRACK_FREE);
5269 SLAB_ATTR_RO(free_calls);
5270 #endif /* CONFIG_SLUB_DEBUG */
5272 #ifdef CONFIG_FAILSLAB
5273 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5275 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5278 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5281 if (s->refcount > 1)
5284 s->flags &= ~SLAB_FAILSLAB;
5286 s->flags |= SLAB_FAILSLAB;
5289 SLAB_ATTR(failslab);
5292 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5297 static ssize_t shrink_store(struct kmem_cache *s,
5298 const char *buf, size_t length)
5301 kmem_cache_shrink(s);
5309 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5311 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5314 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5315 const char *buf, size_t length)
5320 err = kstrtouint(buf, 10, &ratio);
5326 s->remote_node_defrag_ratio = ratio * 10;
5330 SLAB_ATTR(remote_node_defrag_ratio);
5333 #ifdef CONFIG_SLUB_STATS
5334 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5336 unsigned long sum = 0;
5339 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5344 for_each_online_cpu(cpu) {
5345 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5351 len = sprintf(buf, "%lu", sum);
5354 for_each_online_cpu(cpu) {
5355 if (data[cpu] && len < PAGE_SIZE - 20)
5356 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5360 return len + sprintf(buf + len, "\n");
5363 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5367 for_each_online_cpu(cpu)
5368 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5371 #define STAT_ATTR(si, text) \
5372 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5374 return show_stat(s, buf, si); \
5376 static ssize_t text##_store(struct kmem_cache *s, \
5377 const char *buf, size_t length) \
5379 if (buf[0] != '0') \
5381 clear_stat(s, si); \
5386 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5387 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5388 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5389 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5390 STAT_ATTR(FREE_FROZEN, free_frozen);
5391 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5392 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5393 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5394 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5395 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5396 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5397 STAT_ATTR(FREE_SLAB, free_slab);
5398 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5399 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5400 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5401 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5402 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5403 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5404 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5405 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5406 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5407 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5408 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5409 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5410 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5411 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5412 #endif /* CONFIG_SLUB_STATS */
5414 static struct attribute *slab_attrs[] = {
5415 &slab_size_attr.attr,
5416 &object_size_attr.attr,
5417 &objs_per_slab_attr.attr,
5419 &min_partial_attr.attr,
5420 &cpu_partial_attr.attr,
5422 &objects_partial_attr.attr,
5424 &cpu_slabs_attr.attr,
5428 &hwcache_align_attr.attr,
5429 &reclaim_account_attr.attr,
5430 &destroy_by_rcu_attr.attr,
5432 &slabs_cpu_partial_attr.attr,
5433 #ifdef CONFIG_SLUB_DEBUG
5434 &total_objects_attr.attr,
5436 &sanity_checks_attr.attr,
5438 &red_zone_attr.attr,
5440 &store_user_attr.attr,
5441 &validate_attr.attr,
5442 &alloc_calls_attr.attr,
5443 &free_calls_attr.attr,
5445 #ifdef CONFIG_ZONE_DMA
5446 &cache_dma_attr.attr,
5449 &remote_node_defrag_ratio_attr.attr,
5451 #ifdef CONFIG_SLUB_STATS
5452 &alloc_fastpath_attr.attr,
5453 &alloc_slowpath_attr.attr,
5454 &free_fastpath_attr.attr,
5455 &free_slowpath_attr.attr,
5456 &free_frozen_attr.attr,
5457 &free_add_partial_attr.attr,
5458 &free_remove_partial_attr.attr,
5459 &alloc_from_partial_attr.attr,
5460 &alloc_slab_attr.attr,
5461 &alloc_refill_attr.attr,
5462 &alloc_node_mismatch_attr.attr,
5463 &free_slab_attr.attr,
5464 &cpuslab_flush_attr.attr,
5465 &deactivate_full_attr.attr,
5466 &deactivate_empty_attr.attr,
5467 &deactivate_to_head_attr.attr,
5468 &deactivate_to_tail_attr.attr,
5469 &deactivate_remote_frees_attr.attr,
5470 &deactivate_bypass_attr.attr,
5471 &order_fallback_attr.attr,
5472 &cmpxchg_double_fail_attr.attr,
5473 &cmpxchg_double_cpu_fail_attr.attr,
5474 &cpu_partial_alloc_attr.attr,
5475 &cpu_partial_free_attr.attr,
5476 &cpu_partial_node_attr.attr,
5477 &cpu_partial_drain_attr.attr,
5479 #ifdef CONFIG_FAILSLAB
5480 &failslab_attr.attr,
5482 &usersize_attr.attr,
5487 static const struct attribute_group slab_attr_group = {
5488 .attrs = slab_attrs,
5491 static ssize_t slab_attr_show(struct kobject *kobj,
5492 struct attribute *attr,
5495 struct slab_attribute *attribute;
5496 struct kmem_cache *s;
5499 attribute = to_slab_attr(attr);
5502 if (!attribute->show)
5505 err = attribute->show(s, buf);
5510 static ssize_t slab_attr_store(struct kobject *kobj,
5511 struct attribute *attr,
5512 const char *buf, size_t len)
5514 struct slab_attribute *attribute;
5515 struct kmem_cache *s;
5518 attribute = to_slab_attr(attr);
5521 if (!attribute->store)
5524 err = attribute->store(s, buf, len);
5526 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5527 struct kmem_cache *c;
5529 mutex_lock(&slab_mutex);
5530 if (s->max_attr_size < len)
5531 s->max_attr_size = len;
5534 * This is a best effort propagation, so this function's return
5535 * value will be determined by the parent cache only. This is
5536 * basically because not all attributes will have a well
5537 * defined semantics for rollbacks - most of the actions will
5538 * have permanent effects.
5540 * Returning the error value of any of the children that fail
5541 * is not 100 % defined, in the sense that users seeing the
5542 * error code won't be able to know anything about the state of
5545 * Only returning the error code for the parent cache at least
5546 * has well defined semantics. The cache being written to
5547 * directly either failed or succeeded, in which case we loop
5548 * through the descendants with best-effort propagation.
5550 for_each_memcg_cache(c, s)
5551 attribute->store(c, buf, len);
5552 mutex_unlock(&slab_mutex);
5558 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5562 char *buffer = NULL;
5563 struct kmem_cache *root_cache;
5565 if (is_root_cache(s))
5568 root_cache = s->memcg_params.root_cache;
5571 * This mean this cache had no attribute written. Therefore, no point
5572 * in copying default values around
5574 if (!root_cache->max_attr_size)
5577 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5580 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5583 if (!attr || !attr->store || !attr->show)
5587 * It is really bad that we have to allocate here, so we will
5588 * do it only as a fallback. If we actually allocate, though,
5589 * we can just use the allocated buffer until the end.
5591 * Most of the slub attributes will tend to be very small in
5592 * size, but sysfs allows buffers up to a page, so they can
5593 * theoretically happen.
5597 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5600 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5601 if (WARN_ON(!buffer))
5606 len = attr->show(root_cache, buf);
5608 attr->store(s, buf, len);
5612 free_page((unsigned long)buffer);
5613 #endif /* CONFIG_MEMCG */
5616 static void kmem_cache_release(struct kobject *k)
5618 slab_kmem_cache_release(to_slab(k));
5621 static const struct sysfs_ops slab_sysfs_ops = {
5622 .show = slab_attr_show,
5623 .store = slab_attr_store,
5626 static struct kobj_type slab_ktype = {
5627 .sysfs_ops = &slab_sysfs_ops,
5628 .release = kmem_cache_release,
5631 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5633 struct kobj_type *ktype = get_ktype(kobj);
5635 if (ktype == &slab_ktype)
5640 static const struct kset_uevent_ops slab_uevent_ops = {
5641 .filter = uevent_filter,
5644 static struct kset *slab_kset;
5646 static inline struct kset *cache_kset(struct kmem_cache *s)
5649 if (!is_root_cache(s))
5650 return s->memcg_params.root_cache->memcg_kset;
5655 #define ID_STR_LENGTH 64
5657 /* Create a unique string id for a slab cache:
5659 * Format :[flags-]size
5661 static char *create_unique_id(struct kmem_cache *s)
5663 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5670 * First flags affecting slabcache operations. We will only
5671 * get here for aliasable slabs so we do not need to support
5672 * too many flags. The flags here must cover all flags that
5673 * are matched during merging to guarantee that the id is
5676 if (s->flags & SLAB_CACHE_DMA)
5678 if (s->flags & SLAB_CACHE_DMA32)
5680 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5682 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5684 if (s->flags & SLAB_ACCOUNT)
5688 p += sprintf(p, "%07u", s->size);
5690 BUG_ON(p > name + ID_STR_LENGTH - 1);
5694 static void sysfs_slab_remove_workfn(struct work_struct *work)
5696 struct kmem_cache *s =
5697 container_of(work, struct kmem_cache, kobj_remove_work);
5699 if (!s->kobj.state_in_sysfs)
5701 * For a memcg cache, this may be called during
5702 * deactivation and again on shutdown. Remove only once.
5703 * A cache is never shut down before deactivation is
5704 * complete, so no need to worry about synchronization.
5709 kset_unregister(s->memcg_kset);
5711 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5713 kobject_put(&s->kobj);
5716 static int sysfs_slab_add(struct kmem_cache *s)
5720 struct kset *kset = cache_kset(s);
5721 int unmergeable = slab_unmergeable(s);
5723 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5726 kobject_init(&s->kobj, &slab_ktype);
5730 if (!unmergeable && disable_higher_order_debug &&
5731 (slub_debug & DEBUG_METADATA_FLAGS))
5736 * Slabcache can never be merged so we can use the name proper.
5737 * This is typically the case for debug situations. In that
5738 * case we can catch duplicate names easily.
5740 sysfs_remove_link(&slab_kset->kobj, s->name);
5744 * Create a unique name for the slab as a target
5747 name = create_unique_id(s);
5750 s->kobj.kset = kset;
5751 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5755 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5760 if (is_root_cache(s) && memcg_sysfs_enabled) {
5761 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5762 if (!s->memcg_kset) {
5769 kobject_uevent(&s->kobj, KOBJ_ADD);
5771 /* Setup first alias */
5772 sysfs_slab_alias(s, s->name);
5779 kobject_del(&s->kobj);
5783 static void sysfs_slab_remove(struct kmem_cache *s)
5785 if (slab_state < FULL)
5787 * Sysfs has not been setup yet so no need to remove the
5792 kobject_get(&s->kobj);
5793 schedule_work(&s->kobj_remove_work);
5796 void sysfs_slab_unlink(struct kmem_cache *s)
5798 if (slab_state >= FULL)
5799 kobject_del(&s->kobj);
5802 void sysfs_slab_release(struct kmem_cache *s)
5804 if (slab_state >= FULL)
5805 kobject_put(&s->kobj);
5809 * Need to buffer aliases during bootup until sysfs becomes
5810 * available lest we lose that information.
5812 struct saved_alias {
5813 struct kmem_cache *s;
5815 struct saved_alias *next;
5818 static struct saved_alias *alias_list;
5820 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5822 struct saved_alias *al;
5824 if (slab_state == FULL) {
5826 * If we have a leftover link then remove it.
5828 sysfs_remove_link(&slab_kset->kobj, name);
5829 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5832 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5838 al->next = alias_list;
5843 static int __init slab_sysfs_init(void)
5845 struct kmem_cache *s;
5848 mutex_lock(&slab_mutex);
5850 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5852 mutex_unlock(&slab_mutex);
5853 pr_err("Cannot register slab subsystem.\n");
5859 list_for_each_entry(s, &slab_caches, list) {
5860 err = sysfs_slab_add(s);
5862 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5866 while (alias_list) {
5867 struct saved_alias *al = alias_list;
5869 alias_list = alias_list->next;
5870 err = sysfs_slab_alias(al->s, al->name);
5872 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5877 mutex_unlock(&slab_mutex);
5882 __initcall(slab_sysfs_init);
5883 #endif /* CONFIG_SYSFS */
5886 * The /proc/slabinfo ABI
5888 #ifdef CONFIG_SLUB_DEBUG
5889 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5891 unsigned long nr_slabs = 0;
5892 unsigned long nr_objs = 0;
5893 unsigned long nr_free = 0;
5895 struct kmem_cache_node *n;
5897 for_each_kmem_cache_node(s, node, n) {
5898 nr_slabs += node_nr_slabs(n);
5899 nr_objs += node_nr_objs(n);
5900 nr_free += count_partial(n, count_free);
5903 sinfo->active_objs = nr_objs - nr_free;
5904 sinfo->num_objs = nr_objs;
5905 sinfo->active_slabs = nr_slabs;
5906 sinfo->num_slabs = nr_slabs;
5907 sinfo->objects_per_slab = oo_objects(s->oo);
5908 sinfo->cache_order = oo_order(s->oo);
5911 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5915 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5916 size_t count, loff_t *ppos)
5920 #endif /* CONFIG_SLUB_DEBUG */