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. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 void *fixup_red_left(struct kmem_cache *s, void *p)
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
137 #ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
145 * Issues still to be resolved:
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 * - Variable sizing of the per node arrays
152 /* Enable to test recovery from slab corruption on boot */
153 #undef SLUB_RESILIENCY_TEST
155 /* Enable to log cmpxchg failures */
156 #undef SLUB_DEBUG_CMPXCHG
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
162 #define MIN_PARTIAL 5
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
169 #define MAX_PARTIAL 10
171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
190 #define OO_MASK ((1 << OO_SHIFT) - 1)
191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
193 /* Internal SLUB flags */
195 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
196 /* Use cmpxchg_double */
197 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
200 * Tracking user of a slab.
202 #define TRACK_ADDRS_COUNT 16
204 unsigned long addr; /* Called from address */
205 #ifdef CONFIG_STACKTRACE
206 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
208 int cpu; /* Was running on cpu */
209 int pid; /* Pid context */
210 unsigned long when; /* When did the operation occur */
213 enum track_item { TRACK_ALLOC, TRACK_FREE };
216 static int sysfs_slab_add(struct kmem_cache *);
217 static int sysfs_slab_alias(struct kmem_cache *, const char *);
218 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
219 static void sysfs_slab_remove(struct kmem_cache *s);
221 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
222 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
225 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s->cpu_slab->stat[si]);
239 /********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
244 * Returns freelist pointer (ptr). With hardening, this is obfuscated
245 * with an XOR of the address where the pointer is held and a per-cache
248 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
249 unsigned long ptr_addr)
251 #ifdef CONFIG_SLAB_FREELIST_HARDENED
253 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
254 * Normally, this doesn't cause any issues, as both set_freepointer()
255 * and get_freepointer() are called with a pointer with the same tag.
256 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
257 * example, when __free_slub() iterates over objects in a cache, it
258 * passes untagged pointers to check_object(). check_object() in turns
259 * calls get_freepointer() with an untagged pointer, which causes the
260 * freepointer to be restored incorrectly.
262 return (void *)((unsigned long)ptr ^ s->random ^
263 (unsigned long)kasan_reset_tag((void *)ptr_addr));
269 /* Returns the freelist pointer recorded at location ptr_addr. */
270 static inline void *freelist_dereference(const struct kmem_cache *s,
273 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
274 (unsigned long)ptr_addr);
277 static inline void *get_freepointer(struct kmem_cache *s, void *object)
279 return freelist_dereference(s, object + s->offset);
282 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
284 prefetch(object + s->offset);
287 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
289 unsigned long freepointer_addr;
292 if (!debug_pagealloc_enabled())
293 return get_freepointer(s, object);
295 freepointer_addr = (unsigned long)object + s->offset;
296 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
297 return freelist_ptr(s, p, freepointer_addr);
300 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
302 unsigned long freeptr_addr = (unsigned long)object + s->offset;
304 #ifdef CONFIG_SLAB_FREELIST_HARDENED
305 BUG_ON(object == fp); /* naive detection of double free or corruption */
308 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
311 /* Loop over all objects in a slab */
312 #define for_each_object(__p, __s, __addr, __objects) \
313 for (__p = fixup_red_left(__s, __addr); \
314 __p < (__addr) + (__objects) * (__s)->size; \
317 /* Determine object index from a given position */
318 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
320 return (kasan_reset_tag(p) - addr) / s->size;
323 static inline unsigned int order_objects(unsigned int order, unsigned int size)
325 return ((unsigned int)PAGE_SIZE << order) / size;
328 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
331 struct kmem_cache_order_objects x = {
332 (order << OO_SHIFT) + order_objects(order, size)
338 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
340 return x.x >> OO_SHIFT;
343 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
345 return x.x & OO_MASK;
349 * Per slab locking using the pagelock
351 static __always_inline void slab_lock(struct page *page)
353 VM_BUG_ON_PAGE(PageTail(page), page);
354 bit_spin_lock(PG_locked, &page->flags);
357 static __always_inline void slab_unlock(struct page *page)
359 VM_BUG_ON_PAGE(PageTail(page), page);
360 __bit_spin_unlock(PG_locked, &page->flags);
363 /* Interrupts must be disabled (for the fallback code to work right) */
364 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
365 void *freelist_old, unsigned long counters_old,
366 void *freelist_new, unsigned long counters_new,
369 VM_BUG_ON(!irqs_disabled());
370 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
371 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
372 if (s->flags & __CMPXCHG_DOUBLE) {
373 if (cmpxchg_double(&page->freelist, &page->counters,
374 freelist_old, counters_old,
375 freelist_new, counters_new))
381 if (page->freelist == freelist_old &&
382 page->counters == counters_old) {
383 page->freelist = freelist_new;
384 page->counters = counters_new;
392 stat(s, CMPXCHG_DOUBLE_FAIL);
394 #ifdef SLUB_DEBUG_CMPXCHG
395 pr_info("%s %s: cmpxchg double redo ", n, s->name);
401 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
402 void *freelist_old, unsigned long counters_old,
403 void *freelist_new, unsigned long counters_new,
406 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
407 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
408 if (s->flags & __CMPXCHG_DOUBLE) {
409 if (cmpxchg_double(&page->freelist, &page->counters,
410 freelist_old, counters_old,
411 freelist_new, counters_new))
418 local_irq_save(flags);
420 if (page->freelist == freelist_old &&
421 page->counters == counters_old) {
422 page->freelist = freelist_new;
423 page->counters = counters_new;
425 local_irq_restore(flags);
429 local_irq_restore(flags);
433 stat(s, CMPXCHG_DOUBLE_FAIL);
435 #ifdef SLUB_DEBUG_CMPXCHG
436 pr_info("%s %s: cmpxchg double redo ", n, s->name);
442 #ifdef CONFIG_SLUB_DEBUG
444 * Determine a map of object in use on a page.
446 * Node listlock must be held to guarantee that the page does
447 * not vanish from under us.
449 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
452 void *addr = page_address(page);
454 for (p = page->freelist; p; p = get_freepointer(s, p))
455 set_bit(slab_index(p, s, addr), map);
458 static inline unsigned int size_from_object(struct kmem_cache *s)
460 if (s->flags & SLAB_RED_ZONE)
461 return s->size - s->red_left_pad;
466 static inline void *restore_red_left(struct kmem_cache *s, void *p)
468 if (s->flags & SLAB_RED_ZONE)
469 p -= s->red_left_pad;
477 #if defined(CONFIG_SLUB_DEBUG_ON)
478 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
480 static slab_flags_t slub_debug;
483 static char *slub_debug_slabs;
484 static int disable_higher_order_debug;
487 * slub is about to manipulate internal object metadata. This memory lies
488 * outside the range of the allocated object, so accessing it would normally
489 * be reported by kasan as a bounds error. metadata_access_enable() is used
490 * to tell kasan that these accesses are OK.
492 static inline void metadata_access_enable(void)
494 kasan_disable_current();
497 static inline void metadata_access_disable(void)
499 kasan_enable_current();
506 /* Verify that a pointer has an address that is valid within a slab page */
507 static inline int check_valid_pointer(struct kmem_cache *s,
508 struct page *page, void *object)
515 base = page_address(page);
516 object = kasan_reset_tag(object);
517 object = restore_red_left(s, object);
518 if (object < base || object >= base + page->objects * s->size ||
519 (object - base) % s->size) {
526 static void print_section(char *level, char *text, u8 *addr,
529 metadata_access_enable();
530 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
532 metadata_access_disable();
535 static struct track *get_track(struct kmem_cache *s, void *object,
536 enum track_item alloc)
541 p = object + s->offset + sizeof(void *);
543 p = object + s->inuse;
548 static void set_track(struct kmem_cache *s, void *object,
549 enum track_item alloc, unsigned long addr)
551 struct track *p = get_track(s, object, alloc);
554 #ifdef CONFIG_STACKTRACE
555 struct stack_trace trace;
558 trace.nr_entries = 0;
559 trace.max_entries = TRACK_ADDRS_COUNT;
560 trace.entries = p->addrs;
562 metadata_access_enable();
563 save_stack_trace(&trace);
564 metadata_access_disable();
566 /* See rant in lockdep.c */
567 if (trace.nr_entries != 0 &&
568 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
571 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
575 p->cpu = smp_processor_id();
576 p->pid = current->pid;
579 memset(p, 0, sizeof(struct track));
582 static void init_tracking(struct kmem_cache *s, void *object)
584 if (!(s->flags & SLAB_STORE_USER))
587 set_track(s, object, TRACK_FREE, 0UL);
588 set_track(s, object, TRACK_ALLOC, 0UL);
591 static void print_track(const char *s, struct track *t, unsigned long pr_time)
596 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
597 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
598 #ifdef CONFIG_STACKTRACE
601 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
603 pr_err("\t%pS\n", (void *)t->addrs[i]);
610 static void print_tracking(struct kmem_cache *s, void *object)
612 unsigned long pr_time = jiffies;
613 if (!(s->flags & SLAB_STORE_USER))
616 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
617 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
620 static void print_page_info(struct page *page)
622 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
623 page, page->objects, page->inuse, page->freelist, page->flags);
627 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
629 struct va_format vaf;
635 pr_err("=============================================================================\n");
636 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
637 pr_err("-----------------------------------------------------------------------------\n\n");
639 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
643 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
645 struct va_format vaf;
651 pr_err("FIX %s: %pV\n", s->name, &vaf);
655 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
657 unsigned int off; /* Offset of last byte */
658 u8 *addr = page_address(page);
660 print_tracking(s, p);
662 print_page_info(page);
664 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
665 p, p - addr, get_freepointer(s, p));
667 if (s->flags & SLAB_RED_ZONE)
668 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
670 else if (p > addr + 16)
671 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
673 print_section(KERN_ERR, "Object ", p,
674 min_t(unsigned int, s->object_size, PAGE_SIZE));
675 if (s->flags & SLAB_RED_ZONE)
676 print_section(KERN_ERR, "Redzone ", p + s->object_size,
677 s->inuse - s->object_size);
680 off = s->offset + sizeof(void *);
684 if (s->flags & SLAB_STORE_USER)
685 off += 2 * sizeof(struct track);
687 off += kasan_metadata_size(s);
689 if (off != size_from_object(s))
690 /* Beginning of the filler is the free pointer */
691 print_section(KERN_ERR, "Padding ", p + off,
692 size_from_object(s) - off);
697 void object_err(struct kmem_cache *s, struct page *page,
698 u8 *object, char *reason)
700 slab_bug(s, "%s", reason);
701 print_trailer(s, page, object);
704 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
705 const char *fmt, ...)
711 vsnprintf(buf, sizeof(buf), fmt, args);
713 slab_bug(s, "%s", buf);
714 print_page_info(page);
718 static void init_object(struct kmem_cache *s, void *object, u8 val)
722 if (s->flags & SLAB_RED_ZONE)
723 memset(p - s->red_left_pad, val, s->red_left_pad);
725 if (s->flags & __OBJECT_POISON) {
726 memset(p, POISON_FREE, s->object_size - 1);
727 p[s->object_size - 1] = POISON_END;
730 if (s->flags & SLAB_RED_ZONE)
731 memset(p + s->object_size, val, s->inuse - s->object_size);
734 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
735 void *from, void *to)
737 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
738 memset(from, data, to - from);
741 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
742 u8 *object, char *what,
743 u8 *start, unsigned int value, unsigned int bytes)
748 metadata_access_enable();
749 fault = memchr_inv(start, value, bytes);
750 metadata_access_disable();
755 while (end > fault && end[-1] == value)
758 slab_bug(s, "%s overwritten", what);
759 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
760 fault, end - 1, fault[0], value);
761 print_trailer(s, page, object);
763 restore_bytes(s, what, value, fault, end);
771 * Bytes of the object to be managed.
772 * If the freepointer may overlay the object then the free
773 * pointer is the first word of the object.
775 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
778 * object + s->object_size
779 * Padding to reach word boundary. This is also used for Redzoning.
780 * Padding is extended by another word if Redzoning is enabled and
781 * object_size == inuse.
783 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
784 * 0xcc (RED_ACTIVE) for objects in use.
787 * Meta data starts here.
789 * A. Free pointer (if we cannot overwrite object on free)
790 * B. Tracking data for SLAB_STORE_USER
791 * C. Padding to reach required alignment boundary or at mininum
792 * one word if debugging is on to be able to detect writes
793 * before the word boundary.
795 * Padding is done using 0x5a (POISON_INUSE)
798 * Nothing is used beyond s->size.
800 * If slabcaches are merged then the object_size and inuse boundaries are mostly
801 * ignored. And therefore no slab options that rely on these boundaries
802 * may be used with merged slabcaches.
805 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
807 unsigned long off = s->inuse; /* The end of info */
810 /* Freepointer is placed after the object. */
811 off += sizeof(void *);
813 if (s->flags & SLAB_STORE_USER)
814 /* We also have user information there */
815 off += 2 * sizeof(struct track);
817 off += kasan_metadata_size(s);
819 if (size_from_object(s) == off)
822 return check_bytes_and_report(s, page, p, "Object padding",
823 p + off, POISON_INUSE, size_from_object(s) - off);
826 /* Check the pad bytes at the end of a slab page */
827 static int slab_pad_check(struct kmem_cache *s, struct page *page)
836 if (!(s->flags & SLAB_POISON))
839 start = page_address(page);
840 length = PAGE_SIZE << compound_order(page);
841 end = start + length;
842 remainder = length % s->size;
846 pad = end - remainder;
847 metadata_access_enable();
848 fault = memchr_inv(pad, POISON_INUSE, remainder);
849 metadata_access_disable();
852 while (end > fault && end[-1] == POISON_INUSE)
855 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
856 print_section(KERN_ERR, "Padding ", pad, remainder);
858 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
862 static int check_object(struct kmem_cache *s, struct page *page,
863 void *object, u8 val)
866 u8 *endobject = object + s->object_size;
868 if (s->flags & SLAB_RED_ZONE) {
869 if (!check_bytes_and_report(s, page, object, "Redzone",
870 object - s->red_left_pad, val, s->red_left_pad))
873 if (!check_bytes_and_report(s, page, object, "Redzone",
874 endobject, val, s->inuse - s->object_size))
877 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
878 check_bytes_and_report(s, page, p, "Alignment padding",
879 endobject, POISON_INUSE,
880 s->inuse - s->object_size);
884 if (s->flags & SLAB_POISON) {
885 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
886 (!check_bytes_and_report(s, page, p, "Poison", p,
887 POISON_FREE, s->object_size - 1) ||
888 !check_bytes_and_report(s, page, p, "Poison",
889 p + s->object_size - 1, POISON_END, 1)))
892 * check_pad_bytes cleans up on its own.
894 check_pad_bytes(s, page, p);
897 if (!s->offset && val == SLUB_RED_ACTIVE)
899 * Object and freepointer overlap. Cannot check
900 * freepointer while object is allocated.
904 /* Check free pointer validity */
905 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
906 object_err(s, page, p, "Freepointer corrupt");
908 * No choice but to zap it and thus lose the remainder
909 * of the free objects in this slab. May cause
910 * another error because the object count is now wrong.
912 set_freepointer(s, p, NULL);
918 static int check_slab(struct kmem_cache *s, struct page *page)
922 VM_BUG_ON(!irqs_disabled());
924 if (!PageSlab(page)) {
925 slab_err(s, page, "Not a valid slab page");
929 maxobj = order_objects(compound_order(page), s->size);
930 if (page->objects > maxobj) {
931 slab_err(s, page, "objects %u > max %u",
932 page->objects, maxobj);
935 if (page->inuse > page->objects) {
936 slab_err(s, page, "inuse %u > max %u",
937 page->inuse, page->objects);
940 /* Slab_pad_check fixes things up after itself */
941 slab_pad_check(s, page);
946 * Determine if a certain object on a page is on the freelist. Must hold the
947 * slab lock to guarantee that the chains are in a consistent state.
949 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
957 while (fp && nr <= page->objects) {
960 if (!check_valid_pointer(s, page, fp)) {
962 object_err(s, page, object,
963 "Freechain corrupt");
964 set_freepointer(s, object, NULL);
966 slab_err(s, page, "Freepointer corrupt");
967 page->freelist = NULL;
968 page->inuse = page->objects;
969 slab_fix(s, "Freelist cleared");
975 fp = get_freepointer(s, object);
979 max_objects = order_objects(compound_order(page), s->size);
980 if (max_objects > MAX_OBJS_PER_PAGE)
981 max_objects = MAX_OBJS_PER_PAGE;
983 if (page->objects != max_objects) {
984 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
985 page->objects, max_objects);
986 page->objects = max_objects;
987 slab_fix(s, "Number of objects adjusted.");
989 if (page->inuse != page->objects - nr) {
990 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
991 page->inuse, page->objects - nr);
992 page->inuse = page->objects - nr;
993 slab_fix(s, "Object count adjusted.");
995 return search == NULL;
998 static void trace(struct kmem_cache *s, struct page *page, void *object,
1001 if (s->flags & SLAB_TRACE) {
1002 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1004 alloc ? "alloc" : "free",
1005 object, page->inuse,
1009 print_section(KERN_INFO, "Object ", (void *)object,
1017 * Tracking of fully allocated slabs for debugging purposes.
1019 static void add_full(struct kmem_cache *s,
1020 struct kmem_cache_node *n, struct page *page)
1022 if (!(s->flags & SLAB_STORE_USER))
1025 lockdep_assert_held(&n->list_lock);
1026 list_add(&page->lru, &n->full);
1029 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1031 if (!(s->flags & SLAB_STORE_USER))
1034 lockdep_assert_held(&n->list_lock);
1035 list_del(&page->lru);
1038 /* Tracking of the number of slabs for debugging purposes */
1039 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1041 struct kmem_cache_node *n = get_node(s, node);
1043 return atomic_long_read(&n->nr_slabs);
1046 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1048 return atomic_long_read(&n->nr_slabs);
1051 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1053 struct kmem_cache_node *n = get_node(s, node);
1056 * May be called early in order to allocate a slab for the
1057 * kmem_cache_node structure. Solve the chicken-egg
1058 * dilemma by deferring the increment of the count during
1059 * bootstrap (see early_kmem_cache_node_alloc).
1062 atomic_long_inc(&n->nr_slabs);
1063 atomic_long_add(objects, &n->total_objects);
1066 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1068 struct kmem_cache_node *n = get_node(s, node);
1070 atomic_long_dec(&n->nr_slabs);
1071 atomic_long_sub(objects, &n->total_objects);
1074 /* Object debug checks for alloc/free paths */
1075 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1078 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1081 init_object(s, object, SLUB_RED_INACTIVE);
1082 init_tracking(s, object);
1085 static void setup_page_debug(struct kmem_cache *s, void *addr, int order)
1087 if (!(s->flags & SLAB_POISON))
1090 metadata_access_enable();
1091 memset(addr, POISON_INUSE, PAGE_SIZE << order);
1092 metadata_access_disable();
1095 static inline int alloc_consistency_checks(struct kmem_cache *s,
1096 struct page *page, void *object)
1098 if (!check_slab(s, page))
1101 if (!check_valid_pointer(s, page, object)) {
1102 object_err(s, page, object, "Freelist Pointer check fails");
1106 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1112 static noinline int alloc_debug_processing(struct kmem_cache *s,
1114 void *object, unsigned long addr)
1116 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1117 if (!alloc_consistency_checks(s, page, object))
1121 /* Success perform special debug activities for allocs */
1122 if (s->flags & SLAB_STORE_USER)
1123 set_track(s, object, TRACK_ALLOC, addr);
1124 trace(s, page, object, 1);
1125 init_object(s, object, SLUB_RED_ACTIVE);
1129 if (PageSlab(page)) {
1131 * If this is a slab page then lets do the best we can
1132 * to avoid issues in the future. Marking all objects
1133 * as used avoids touching the remaining objects.
1135 slab_fix(s, "Marking all objects used");
1136 page->inuse = page->objects;
1137 page->freelist = NULL;
1142 static inline int free_consistency_checks(struct kmem_cache *s,
1143 struct page *page, void *object, unsigned long addr)
1145 if (!check_valid_pointer(s, page, object)) {
1146 slab_err(s, page, "Invalid object pointer 0x%p", object);
1150 if (on_freelist(s, page, object)) {
1151 object_err(s, page, object, "Object already free");
1155 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1158 if (unlikely(s != page->slab_cache)) {
1159 if (!PageSlab(page)) {
1160 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1162 } else if (!page->slab_cache) {
1163 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1167 object_err(s, page, object,
1168 "page slab pointer corrupt.");
1174 /* Supports checking bulk free of a constructed freelist */
1175 static noinline int free_debug_processing(
1176 struct kmem_cache *s, struct page *page,
1177 void *head, void *tail, int bulk_cnt,
1180 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1181 void *object = head;
1183 unsigned long uninitialized_var(flags);
1186 spin_lock_irqsave(&n->list_lock, flags);
1189 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1190 if (!check_slab(s, page))
1197 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1198 if (!free_consistency_checks(s, page, object, addr))
1202 if (s->flags & SLAB_STORE_USER)
1203 set_track(s, object, TRACK_FREE, addr);
1204 trace(s, page, object, 0);
1205 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1206 init_object(s, object, SLUB_RED_INACTIVE);
1208 /* Reached end of constructed freelist yet? */
1209 if (object != tail) {
1210 object = get_freepointer(s, object);
1216 if (cnt != bulk_cnt)
1217 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1221 spin_unlock_irqrestore(&n->list_lock, flags);
1223 slab_fix(s, "Object at 0x%p not freed", object);
1227 static int __init setup_slub_debug(char *str)
1229 slub_debug = DEBUG_DEFAULT_FLAGS;
1230 if (*str++ != '=' || !*str)
1232 * No options specified. Switch on full debugging.
1238 * No options but restriction on slabs. This means full
1239 * debugging for slabs matching a pattern.
1246 * Switch off all debugging measures.
1251 * Determine which debug features should be switched on
1253 for (; *str && *str != ','; str++) {
1254 switch (tolower(*str)) {
1256 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1259 slub_debug |= SLAB_RED_ZONE;
1262 slub_debug |= SLAB_POISON;
1265 slub_debug |= SLAB_STORE_USER;
1268 slub_debug |= SLAB_TRACE;
1271 slub_debug |= SLAB_FAILSLAB;
1275 * Avoid enabling debugging on caches if its minimum
1276 * order would increase as a result.
1278 disable_higher_order_debug = 1;
1281 pr_err("slub_debug option '%c' unknown. skipped\n",
1288 slub_debug_slabs = str + 1;
1293 __setup("slub_debug", setup_slub_debug);
1296 * kmem_cache_flags - apply debugging options to the cache
1297 * @object_size: the size of an object without meta data
1298 * @flags: flags to set
1299 * @name: name of the cache
1300 * @ctor: constructor function
1302 * Debug option(s) are applied to @flags. In addition to the debug
1303 * option(s), if a slab name (or multiple) is specified i.e.
1304 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1305 * then only the select slabs will receive the debug option(s).
1307 slab_flags_t kmem_cache_flags(unsigned int object_size,
1308 slab_flags_t flags, const char *name,
1309 void (*ctor)(void *))
1314 /* If slub_debug = 0, it folds into the if conditional. */
1315 if (!slub_debug_slabs)
1316 return flags | slub_debug;
1319 iter = slub_debug_slabs;
1324 end = strchr(iter, ',');
1326 end = iter + strlen(iter);
1328 glob = strnchr(iter, end - iter, '*');
1330 cmplen = glob - iter;
1332 cmplen = max_t(size_t, len, (end - iter));
1334 if (!strncmp(name, iter, cmplen)) {
1335 flags |= slub_debug;
1346 #else /* !CONFIG_SLUB_DEBUG */
1347 static inline void setup_object_debug(struct kmem_cache *s,
1348 struct page *page, void *object) {}
1349 static inline void setup_page_debug(struct kmem_cache *s,
1350 void *addr, int order) {}
1352 static inline int alloc_debug_processing(struct kmem_cache *s,
1353 struct page *page, void *object, unsigned long addr) { return 0; }
1355 static inline int free_debug_processing(
1356 struct kmem_cache *s, struct page *page,
1357 void *head, void *tail, int bulk_cnt,
1358 unsigned long addr) { return 0; }
1360 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1362 static inline int check_object(struct kmem_cache *s, struct page *page,
1363 void *object, u8 val) { return 1; }
1364 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1365 struct page *page) {}
1366 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1367 struct page *page) {}
1368 slab_flags_t kmem_cache_flags(unsigned int object_size,
1369 slab_flags_t flags, const char *name,
1370 void (*ctor)(void *))
1374 #define slub_debug 0
1376 #define disable_higher_order_debug 0
1378 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1380 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1382 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1384 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1387 #endif /* CONFIG_SLUB_DEBUG */
1390 * Hooks for other subsystems that check memory allocations. In a typical
1391 * production configuration these hooks all should produce no code at all.
1393 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1395 ptr = kasan_kmalloc_large(ptr, size, flags);
1396 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1397 kmemleak_alloc(ptr, size, 1, flags);
1401 static __always_inline void kfree_hook(void *x)
1404 kasan_kfree_large(x, _RET_IP_);
1407 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1409 kmemleak_free_recursive(x, s->flags);
1412 * Trouble is that we may no longer disable interrupts in the fast path
1413 * So in order to make the debug calls that expect irqs to be
1414 * disabled we need to disable interrupts temporarily.
1416 #ifdef CONFIG_LOCKDEP
1418 unsigned long flags;
1420 local_irq_save(flags);
1421 debug_check_no_locks_freed(x, s->object_size);
1422 local_irq_restore(flags);
1425 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1426 debug_check_no_obj_freed(x, s->object_size);
1428 /* KASAN might put x into memory quarantine, delaying its reuse */
1429 return kasan_slab_free(s, x, _RET_IP_);
1432 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1433 void **head, void **tail)
1436 * Compiler cannot detect this function can be removed if slab_free_hook()
1437 * evaluates to nothing. Thus, catch all relevant config debug options here.
1439 #if defined(CONFIG_LOCKDEP) || \
1440 defined(CONFIG_DEBUG_KMEMLEAK) || \
1441 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1442 defined(CONFIG_KASAN)
1446 void *old_tail = *tail ? *tail : *head;
1448 /* Head and tail of the reconstructed freelist */
1454 next = get_freepointer(s, object);
1455 /* If object's reuse doesn't have to be delayed */
1456 if (!slab_free_hook(s, object)) {
1457 /* Move object to the new freelist */
1458 set_freepointer(s, object, *head);
1463 } while (object != old_tail);
1468 return *head != NULL;
1474 static void *setup_object(struct kmem_cache *s, struct page *page,
1477 setup_object_debug(s, page, object);
1478 object = kasan_init_slab_obj(s, object);
1479 if (unlikely(s->ctor)) {
1480 kasan_unpoison_object_data(s, object);
1482 kasan_poison_object_data(s, object);
1488 * Slab allocation and freeing
1490 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1491 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1494 unsigned int order = oo_order(oo);
1496 if (node == NUMA_NO_NODE)
1497 page = alloc_pages(flags, order);
1499 page = __alloc_pages_node(node, flags, order);
1501 if (page && memcg_charge_slab(page, flags, order, s)) {
1502 __free_pages(page, order);
1509 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1510 /* Pre-initialize the random sequence cache */
1511 static int init_cache_random_seq(struct kmem_cache *s)
1513 unsigned int count = oo_objects(s->oo);
1516 /* Bailout if already initialised */
1520 err = cache_random_seq_create(s, count, GFP_KERNEL);
1522 pr_err("SLUB: Unable to initialize free list for %s\n",
1527 /* Transform to an offset on the set of pages */
1528 if (s->random_seq) {
1531 for (i = 0; i < count; i++)
1532 s->random_seq[i] *= s->size;
1537 /* Initialize each random sequence freelist per cache */
1538 static void __init init_freelist_randomization(void)
1540 struct kmem_cache *s;
1542 mutex_lock(&slab_mutex);
1544 list_for_each_entry(s, &slab_caches, list)
1545 init_cache_random_seq(s);
1547 mutex_unlock(&slab_mutex);
1550 /* Get the next entry on the pre-computed freelist randomized */
1551 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1552 unsigned long *pos, void *start,
1553 unsigned long page_limit,
1554 unsigned long freelist_count)
1559 * If the target page allocation failed, the number of objects on the
1560 * page might be smaller than the usual size defined by the cache.
1563 idx = s->random_seq[*pos];
1565 if (*pos >= freelist_count)
1567 } while (unlikely(idx >= page_limit));
1569 return (char *)start + idx;
1572 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1573 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1578 unsigned long idx, pos, page_limit, freelist_count;
1580 if (page->objects < 2 || !s->random_seq)
1583 freelist_count = oo_objects(s->oo);
1584 pos = get_random_int() % freelist_count;
1586 page_limit = page->objects * s->size;
1587 start = fixup_red_left(s, page_address(page));
1589 /* First entry is used as the base of the freelist */
1590 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1592 cur = setup_object(s, page, cur);
1593 page->freelist = cur;
1595 for (idx = 1; idx < page->objects; idx++) {
1596 next = next_freelist_entry(s, page, &pos, start, page_limit,
1598 next = setup_object(s, page, next);
1599 set_freepointer(s, cur, next);
1602 set_freepointer(s, cur, NULL);
1607 static inline int init_cache_random_seq(struct kmem_cache *s)
1611 static inline void init_freelist_randomization(void) { }
1612 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1616 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1618 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1621 struct kmem_cache_order_objects oo = s->oo;
1623 void *start, *p, *next;
1627 flags &= gfp_allowed_mask;
1629 if (gfpflags_allow_blocking(flags))
1632 flags |= s->allocflags;
1635 * Let the initial higher-order allocation fail under memory pressure
1636 * so we fall-back to the minimum order allocation.
1638 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1639 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1640 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1642 page = alloc_slab_page(s, alloc_gfp, node, oo);
1643 if (unlikely(!page)) {
1647 * Allocation may have failed due to fragmentation.
1648 * Try a lower order alloc if possible
1650 page = alloc_slab_page(s, alloc_gfp, node, oo);
1651 if (unlikely(!page))
1653 stat(s, ORDER_FALLBACK);
1656 page->objects = oo_objects(oo);
1658 order = compound_order(page);
1659 page->slab_cache = s;
1660 __SetPageSlab(page);
1661 if (page_is_pfmemalloc(page))
1662 SetPageSlabPfmemalloc(page);
1664 kasan_poison_slab(page);
1666 start = page_address(page);
1668 setup_page_debug(s, start, order);
1670 shuffle = shuffle_freelist(s, page);
1673 start = fixup_red_left(s, start);
1674 start = setup_object(s, page, start);
1675 page->freelist = start;
1676 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1678 next = setup_object(s, page, next);
1679 set_freepointer(s, p, next);
1682 set_freepointer(s, p, NULL);
1685 page->inuse = page->objects;
1689 if (gfpflags_allow_blocking(flags))
1690 local_irq_disable();
1694 mod_lruvec_page_state(page,
1695 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1696 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1699 inc_slabs_node(s, page_to_nid(page), page->objects);
1704 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1706 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1707 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1708 flags &= ~GFP_SLAB_BUG_MASK;
1709 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1710 invalid_mask, &invalid_mask, flags, &flags);
1714 return allocate_slab(s,
1715 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1718 static void __free_slab(struct kmem_cache *s, struct page *page)
1720 int order = compound_order(page);
1721 int pages = 1 << order;
1723 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1726 slab_pad_check(s, page);
1727 for_each_object(p, s, page_address(page),
1729 check_object(s, page, p, SLUB_RED_INACTIVE);
1732 mod_lruvec_page_state(page,
1733 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1734 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1737 __ClearPageSlabPfmemalloc(page);
1738 __ClearPageSlab(page);
1740 page->mapping = NULL;
1741 if (current->reclaim_state)
1742 current->reclaim_state->reclaimed_slab += pages;
1743 memcg_uncharge_slab(page, order, s);
1744 __free_pages(page, order);
1747 static void rcu_free_slab(struct rcu_head *h)
1749 struct page *page = container_of(h, struct page, rcu_head);
1751 __free_slab(page->slab_cache, page);
1754 static void free_slab(struct kmem_cache *s, struct page *page)
1756 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1757 call_rcu(&page->rcu_head, rcu_free_slab);
1759 __free_slab(s, page);
1762 static void discard_slab(struct kmem_cache *s, struct page *page)
1764 dec_slabs_node(s, page_to_nid(page), page->objects);
1769 * Management of partially allocated slabs.
1772 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1775 if (tail == DEACTIVATE_TO_TAIL)
1776 list_add_tail(&page->lru, &n->partial);
1778 list_add(&page->lru, &n->partial);
1781 static inline void add_partial(struct kmem_cache_node *n,
1782 struct page *page, int tail)
1784 lockdep_assert_held(&n->list_lock);
1785 __add_partial(n, page, tail);
1788 static inline void remove_partial(struct kmem_cache_node *n,
1791 lockdep_assert_held(&n->list_lock);
1792 list_del(&page->lru);
1797 * Remove slab from the partial list, freeze it and
1798 * return the pointer to the freelist.
1800 * Returns a list of objects or NULL if it fails.
1802 static inline void *acquire_slab(struct kmem_cache *s,
1803 struct kmem_cache_node *n, struct page *page,
1804 int mode, int *objects)
1807 unsigned long counters;
1810 lockdep_assert_held(&n->list_lock);
1813 * Zap the freelist and set the frozen bit.
1814 * The old freelist is the list of objects for the
1815 * per cpu allocation list.
1817 freelist = page->freelist;
1818 counters = page->counters;
1819 new.counters = counters;
1820 *objects = new.objects - new.inuse;
1822 new.inuse = page->objects;
1823 new.freelist = NULL;
1825 new.freelist = freelist;
1828 VM_BUG_ON(new.frozen);
1831 if (!__cmpxchg_double_slab(s, page,
1833 new.freelist, new.counters,
1837 remove_partial(n, page);
1842 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1843 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1846 * Try to allocate a partial slab from a specific node.
1848 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1849 struct kmem_cache_cpu *c, gfp_t flags)
1851 struct page *page, *page2;
1852 void *object = NULL;
1853 unsigned int available = 0;
1857 * Racy check. If we mistakenly see no partial slabs then we
1858 * just allocate an empty slab. If we mistakenly try to get a
1859 * partial slab and there is none available then get_partials()
1862 if (!n || !n->nr_partial)
1865 spin_lock(&n->list_lock);
1866 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1869 if (!pfmemalloc_match(page, flags))
1872 t = acquire_slab(s, n, page, object == NULL, &objects);
1876 available += objects;
1879 stat(s, ALLOC_FROM_PARTIAL);
1882 put_cpu_partial(s, page, 0);
1883 stat(s, CPU_PARTIAL_NODE);
1885 if (!kmem_cache_has_cpu_partial(s)
1886 || available > slub_cpu_partial(s) / 2)
1890 spin_unlock(&n->list_lock);
1895 * Get a page from somewhere. Search in increasing NUMA distances.
1897 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1898 struct kmem_cache_cpu *c)
1901 struct zonelist *zonelist;
1904 enum zone_type high_zoneidx = gfp_zone(flags);
1906 unsigned int cpuset_mems_cookie;
1909 * The defrag ratio allows a configuration of the tradeoffs between
1910 * inter node defragmentation and node local allocations. A lower
1911 * defrag_ratio increases the tendency to do local allocations
1912 * instead of attempting to obtain partial slabs from other nodes.
1914 * If the defrag_ratio is set to 0 then kmalloc() always
1915 * returns node local objects. If the ratio is higher then kmalloc()
1916 * may return off node objects because partial slabs are obtained
1917 * from other nodes and filled up.
1919 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1920 * (which makes defrag_ratio = 1000) then every (well almost)
1921 * allocation will first attempt to defrag slab caches on other nodes.
1922 * This means scanning over all nodes to look for partial slabs which
1923 * may be expensive if we do it every time we are trying to find a slab
1924 * with available objects.
1926 if (!s->remote_node_defrag_ratio ||
1927 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1931 cpuset_mems_cookie = read_mems_allowed_begin();
1932 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1933 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1934 struct kmem_cache_node *n;
1936 n = get_node(s, zone_to_nid(zone));
1938 if (n && cpuset_zone_allowed(zone, flags) &&
1939 n->nr_partial > s->min_partial) {
1940 object = get_partial_node(s, n, c, flags);
1943 * Don't check read_mems_allowed_retry()
1944 * here - if mems_allowed was updated in
1945 * parallel, that was a harmless race
1946 * between allocation and the cpuset
1953 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1959 * Get a partial page, lock it and return it.
1961 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1962 struct kmem_cache_cpu *c)
1965 int searchnode = node;
1967 if (node == NUMA_NO_NODE)
1968 searchnode = numa_mem_id();
1969 else if (!node_present_pages(node))
1970 searchnode = node_to_mem_node(node);
1972 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1973 if (object || node != NUMA_NO_NODE)
1976 return get_any_partial(s, flags, c);
1979 #ifdef CONFIG_PREEMPT
1981 * Calculate the next globally unique transaction for disambiguiation
1982 * during cmpxchg. The transactions start with the cpu number and are then
1983 * incremented by CONFIG_NR_CPUS.
1985 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1988 * No preemption supported therefore also no need to check for
1994 static inline unsigned long next_tid(unsigned long tid)
1996 return tid + TID_STEP;
1999 static inline unsigned int tid_to_cpu(unsigned long tid)
2001 return tid % TID_STEP;
2004 static inline unsigned long tid_to_event(unsigned long tid)
2006 return tid / TID_STEP;
2009 static inline unsigned int init_tid(int cpu)
2014 static inline void note_cmpxchg_failure(const char *n,
2015 const struct kmem_cache *s, unsigned long tid)
2017 #ifdef SLUB_DEBUG_CMPXCHG
2018 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2020 pr_info("%s %s: cmpxchg redo ", n, s->name);
2022 #ifdef CONFIG_PREEMPT
2023 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2024 pr_warn("due to cpu change %d -> %d\n",
2025 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2028 if (tid_to_event(tid) != tid_to_event(actual_tid))
2029 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2030 tid_to_event(tid), tid_to_event(actual_tid));
2032 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2033 actual_tid, tid, next_tid(tid));
2035 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2038 static void init_kmem_cache_cpus(struct kmem_cache *s)
2042 for_each_possible_cpu(cpu)
2043 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2047 * Remove the cpu slab
2049 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2050 void *freelist, struct kmem_cache_cpu *c)
2052 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2053 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2055 enum slab_modes l = M_NONE, m = M_NONE;
2057 int tail = DEACTIVATE_TO_HEAD;
2061 if (page->freelist) {
2062 stat(s, DEACTIVATE_REMOTE_FREES);
2063 tail = DEACTIVATE_TO_TAIL;
2067 * Stage one: Free all available per cpu objects back
2068 * to the page freelist while it is still frozen. Leave the
2071 * There is no need to take the list->lock because the page
2074 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2076 unsigned long counters;
2079 prior = page->freelist;
2080 counters = page->counters;
2081 set_freepointer(s, freelist, prior);
2082 new.counters = counters;
2084 VM_BUG_ON(!new.frozen);
2086 } while (!__cmpxchg_double_slab(s, page,
2088 freelist, new.counters,
2089 "drain percpu freelist"));
2091 freelist = nextfree;
2095 * Stage two: Ensure that the page is unfrozen while the
2096 * list presence reflects the actual number of objects
2099 * We setup the list membership and then perform a cmpxchg
2100 * with the count. If there is a mismatch then the page
2101 * is not unfrozen but the page is on the wrong list.
2103 * Then we restart the process which may have to remove
2104 * the page from the list that we just put it on again
2105 * because the number of objects in the slab may have
2110 old.freelist = page->freelist;
2111 old.counters = page->counters;
2112 VM_BUG_ON(!old.frozen);
2114 /* Determine target state of the slab */
2115 new.counters = old.counters;
2118 set_freepointer(s, freelist, old.freelist);
2119 new.freelist = freelist;
2121 new.freelist = old.freelist;
2125 if (!new.inuse && n->nr_partial >= s->min_partial)
2127 else if (new.freelist) {
2132 * Taking the spinlock removes the possibility
2133 * that acquire_slab() will see a slab page that
2136 spin_lock(&n->list_lock);
2140 if (kmem_cache_debug(s) && !lock) {
2143 * This also ensures that the scanning of full
2144 * slabs from diagnostic functions will not see
2147 spin_lock(&n->list_lock);
2153 remove_partial(n, page);
2154 else if (l == M_FULL)
2155 remove_full(s, n, page);
2158 add_partial(n, page, tail);
2159 else if (m == M_FULL)
2160 add_full(s, n, page);
2164 if (!__cmpxchg_double_slab(s, page,
2165 old.freelist, old.counters,
2166 new.freelist, new.counters,
2171 spin_unlock(&n->list_lock);
2175 else if (m == M_FULL)
2176 stat(s, DEACTIVATE_FULL);
2177 else if (m == M_FREE) {
2178 stat(s, DEACTIVATE_EMPTY);
2179 discard_slab(s, page);
2188 * Unfreeze all the cpu partial slabs.
2190 * This function must be called with interrupts disabled
2191 * for the cpu using c (or some other guarantee must be there
2192 * to guarantee no concurrent accesses).
2194 static void unfreeze_partials(struct kmem_cache *s,
2195 struct kmem_cache_cpu *c)
2197 #ifdef CONFIG_SLUB_CPU_PARTIAL
2198 struct kmem_cache_node *n = NULL, *n2 = NULL;
2199 struct page *page, *discard_page = NULL;
2201 while ((page = c->partial)) {
2205 c->partial = page->next;
2207 n2 = get_node(s, page_to_nid(page));
2210 spin_unlock(&n->list_lock);
2213 spin_lock(&n->list_lock);
2218 old.freelist = page->freelist;
2219 old.counters = page->counters;
2220 VM_BUG_ON(!old.frozen);
2222 new.counters = old.counters;
2223 new.freelist = old.freelist;
2227 } while (!__cmpxchg_double_slab(s, page,
2228 old.freelist, old.counters,
2229 new.freelist, new.counters,
2230 "unfreezing slab"));
2232 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2233 page->next = discard_page;
2234 discard_page = page;
2236 add_partial(n, page, DEACTIVATE_TO_TAIL);
2237 stat(s, FREE_ADD_PARTIAL);
2242 spin_unlock(&n->list_lock);
2244 while (discard_page) {
2245 page = discard_page;
2246 discard_page = discard_page->next;
2248 stat(s, DEACTIVATE_EMPTY);
2249 discard_slab(s, page);
2256 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2257 * partial page slot if available.
2259 * If we did not find a slot then simply move all the partials to the
2260 * per node partial list.
2262 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2264 #ifdef CONFIG_SLUB_CPU_PARTIAL
2265 struct page *oldpage;
2273 oldpage = this_cpu_read(s->cpu_slab->partial);
2276 pobjects = oldpage->pobjects;
2277 pages = oldpage->pages;
2278 if (drain && pobjects > s->cpu_partial) {
2279 unsigned long flags;
2281 * partial array is full. Move the existing
2282 * set to the per node partial list.
2284 local_irq_save(flags);
2285 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2286 local_irq_restore(flags);
2290 stat(s, CPU_PARTIAL_DRAIN);
2295 pobjects += page->objects - page->inuse;
2297 page->pages = pages;
2298 page->pobjects = pobjects;
2299 page->next = oldpage;
2301 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2303 if (unlikely(!s->cpu_partial)) {
2304 unsigned long flags;
2306 local_irq_save(flags);
2307 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2308 local_irq_restore(flags);
2314 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2316 stat(s, CPUSLAB_FLUSH);
2317 deactivate_slab(s, c->page, c->freelist, c);
2319 c->tid = next_tid(c->tid);
2325 * Called from IPI handler with interrupts disabled.
2327 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2329 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2334 unfreeze_partials(s, c);
2337 static void flush_cpu_slab(void *d)
2339 struct kmem_cache *s = d;
2341 __flush_cpu_slab(s, smp_processor_id());
2344 static bool has_cpu_slab(int cpu, void *info)
2346 struct kmem_cache *s = info;
2347 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2349 return c->page || slub_percpu_partial(c);
2352 static void flush_all(struct kmem_cache *s)
2354 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2358 * Use the cpu notifier to insure that the cpu slabs are flushed when
2361 static int slub_cpu_dead(unsigned int cpu)
2363 struct kmem_cache *s;
2364 unsigned long flags;
2366 mutex_lock(&slab_mutex);
2367 list_for_each_entry(s, &slab_caches, list) {
2368 local_irq_save(flags);
2369 __flush_cpu_slab(s, cpu);
2370 local_irq_restore(flags);
2372 mutex_unlock(&slab_mutex);
2377 * Check if the objects in a per cpu structure fit numa
2378 * locality expectations.
2380 static inline int node_match(struct page *page, int node)
2383 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2389 #ifdef CONFIG_SLUB_DEBUG
2390 static int count_free(struct page *page)
2392 return page->objects - page->inuse;
2395 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2397 return atomic_long_read(&n->total_objects);
2399 #endif /* CONFIG_SLUB_DEBUG */
2401 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2402 static unsigned long count_partial(struct kmem_cache_node *n,
2403 int (*get_count)(struct page *))
2405 unsigned long flags;
2406 unsigned long x = 0;
2409 spin_lock_irqsave(&n->list_lock, flags);
2410 list_for_each_entry(page, &n->partial, lru)
2411 x += get_count(page);
2412 spin_unlock_irqrestore(&n->list_lock, flags);
2415 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2417 static noinline void
2418 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2420 #ifdef CONFIG_SLUB_DEBUG
2421 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2422 DEFAULT_RATELIMIT_BURST);
2424 struct kmem_cache_node *n;
2426 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2429 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2430 nid, gfpflags, &gfpflags);
2431 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2432 s->name, s->object_size, s->size, oo_order(s->oo),
2435 if (oo_order(s->min) > get_order(s->object_size))
2436 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2439 for_each_kmem_cache_node(s, node, n) {
2440 unsigned long nr_slabs;
2441 unsigned long nr_objs;
2442 unsigned long nr_free;
2444 nr_free = count_partial(n, count_free);
2445 nr_slabs = node_nr_slabs(n);
2446 nr_objs = node_nr_objs(n);
2448 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2449 node, nr_slabs, nr_objs, nr_free);
2454 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2455 int node, struct kmem_cache_cpu **pc)
2458 struct kmem_cache_cpu *c = *pc;
2461 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2463 freelist = get_partial(s, flags, node, c);
2468 page = new_slab(s, flags, node);
2470 c = raw_cpu_ptr(s->cpu_slab);
2475 * No other reference to the page yet so we can
2476 * muck around with it freely without cmpxchg
2478 freelist = page->freelist;
2479 page->freelist = NULL;
2481 stat(s, ALLOC_SLAB);
2489 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2491 if (unlikely(PageSlabPfmemalloc(page)))
2492 return gfp_pfmemalloc_allowed(gfpflags);
2498 * Check the page->freelist of a page and either transfer the freelist to the
2499 * per cpu freelist or deactivate the page.
2501 * The page is still frozen if the return value is not NULL.
2503 * If this function returns NULL then the page has been unfrozen.
2505 * This function must be called with interrupt disabled.
2507 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2510 unsigned long counters;
2514 freelist = page->freelist;
2515 counters = page->counters;
2517 new.counters = counters;
2518 VM_BUG_ON(!new.frozen);
2520 new.inuse = page->objects;
2521 new.frozen = freelist != NULL;
2523 } while (!__cmpxchg_double_slab(s, page,
2532 * Slow path. The lockless freelist is empty or we need to perform
2535 * Processing is still very fast if new objects have been freed to the
2536 * regular freelist. In that case we simply take over the regular freelist
2537 * as the lockless freelist and zap the regular freelist.
2539 * If that is not working then we fall back to the partial lists. We take the
2540 * first element of the freelist as the object to allocate now and move the
2541 * rest of the freelist to the lockless freelist.
2543 * And if we were unable to get a new slab from the partial slab lists then
2544 * we need to allocate a new slab. This is the slowest path since it involves
2545 * a call to the page allocator and the setup of a new slab.
2547 * Version of __slab_alloc to use when we know that interrupts are
2548 * already disabled (which is the case for bulk allocation).
2550 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2551 unsigned long addr, struct kmem_cache_cpu *c)
2561 if (unlikely(!node_match(page, node))) {
2562 int searchnode = node;
2564 if (node != NUMA_NO_NODE && !node_present_pages(node))
2565 searchnode = node_to_mem_node(node);
2567 if (unlikely(!node_match(page, searchnode))) {
2568 stat(s, ALLOC_NODE_MISMATCH);
2569 deactivate_slab(s, page, c->freelist, c);
2575 * By rights, we should be searching for a slab page that was
2576 * PFMEMALLOC but right now, we are losing the pfmemalloc
2577 * information when the page leaves the per-cpu allocator
2579 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2580 deactivate_slab(s, page, c->freelist, c);
2584 /* must check again c->freelist in case of cpu migration or IRQ */
2585 freelist = c->freelist;
2589 freelist = get_freelist(s, page);
2593 stat(s, DEACTIVATE_BYPASS);
2597 stat(s, ALLOC_REFILL);
2601 * freelist is pointing to the list of objects to be used.
2602 * page is pointing to the page from which the objects are obtained.
2603 * That page must be frozen for per cpu allocations to work.
2605 VM_BUG_ON(!c->page->frozen);
2606 c->freelist = get_freepointer(s, freelist);
2607 c->tid = next_tid(c->tid);
2612 if (slub_percpu_partial(c)) {
2613 page = c->page = slub_percpu_partial(c);
2614 slub_set_percpu_partial(c, page);
2615 stat(s, CPU_PARTIAL_ALLOC);
2619 freelist = new_slab_objects(s, gfpflags, node, &c);
2621 if (unlikely(!freelist)) {
2622 slab_out_of_memory(s, gfpflags, node);
2627 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2630 /* Only entered in the debug case */
2631 if (kmem_cache_debug(s) &&
2632 !alloc_debug_processing(s, page, freelist, addr))
2633 goto new_slab; /* Slab failed checks. Next slab needed */
2635 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2640 * Another one that disabled interrupt and compensates for possible
2641 * cpu changes by refetching the per cpu area pointer.
2643 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2644 unsigned long addr, struct kmem_cache_cpu *c)
2647 unsigned long flags;
2649 local_irq_save(flags);
2650 #ifdef CONFIG_PREEMPT
2652 * We may have been preempted and rescheduled on a different
2653 * cpu before disabling interrupts. Need to reload cpu area
2656 c = this_cpu_ptr(s->cpu_slab);
2659 p = ___slab_alloc(s, gfpflags, node, addr, c);
2660 local_irq_restore(flags);
2665 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2666 * have the fastpath folded into their functions. So no function call
2667 * overhead for requests that can be satisfied on the fastpath.
2669 * The fastpath works by first checking if the lockless freelist can be used.
2670 * If not then __slab_alloc is called for slow processing.
2672 * Otherwise we can simply pick the next object from the lockless free list.
2674 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2675 gfp_t gfpflags, int node, unsigned long addr)
2678 struct kmem_cache_cpu *c;
2682 s = slab_pre_alloc_hook(s, gfpflags);
2687 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2688 * enabled. We may switch back and forth between cpus while
2689 * reading from one cpu area. That does not matter as long
2690 * as we end up on the original cpu again when doing the cmpxchg.
2692 * We should guarantee that tid and kmem_cache are retrieved on
2693 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2694 * to check if it is matched or not.
2697 tid = this_cpu_read(s->cpu_slab->tid);
2698 c = raw_cpu_ptr(s->cpu_slab);
2699 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2700 unlikely(tid != READ_ONCE(c->tid)));
2703 * Irqless object alloc/free algorithm used here depends on sequence
2704 * of fetching cpu_slab's data. tid should be fetched before anything
2705 * on c to guarantee that object and page associated with previous tid
2706 * won't be used with current tid. If we fetch tid first, object and
2707 * page could be one associated with next tid and our alloc/free
2708 * request will be failed. In this case, we will retry. So, no problem.
2713 * The transaction ids are globally unique per cpu and per operation on
2714 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2715 * occurs on the right processor and that there was no operation on the
2716 * linked list in between.
2719 object = c->freelist;
2721 if (unlikely(!object || !node_match(page, node))) {
2722 object = __slab_alloc(s, gfpflags, node, addr, c);
2723 stat(s, ALLOC_SLOWPATH);
2725 void *next_object = get_freepointer_safe(s, object);
2728 * The cmpxchg will only match if there was no additional
2729 * operation and if we are on the right processor.
2731 * The cmpxchg does the following atomically (without lock
2733 * 1. Relocate first pointer to the current per cpu area.
2734 * 2. Verify that tid and freelist have not been changed
2735 * 3. If they were not changed replace tid and freelist
2737 * Since this is without lock semantics the protection is only
2738 * against code executing on this cpu *not* from access by
2741 if (unlikely(!this_cpu_cmpxchg_double(
2742 s->cpu_slab->freelist, s->cpu_slab->tid,
2744 next_object, next_tid(tid)))) {
2746 note_cmpxchg_failure("slab_alloc", s, tid);
2749 prefetch_freepointer(s, next_object);
2750 stat(s, ALLOC_FASTPATH);
2753 if (unlikely(gfpflags & __GFP_ZERO) && object)
2754 memset(object, 0, s->object_size);
2756 slab_post_alloc_hook(s, gfpflags, 1, &object);
2761 static __always_inline void *slab_alloc(struct kmem_cache *s,
2762 gfp_t gfpflags, unsigned long addr)
2764 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2767 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2769 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2771 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2776 EXPORT_SYMBOL(kmem_cache_alloc);
2778 #ifdef CONFIG_TRACING
2779 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2781 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2782 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2783 ret = kasan_kmalloc(s, ret, size, gfpflags);
2786 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2790 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2792 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2794 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2795 s->object_size, s->size, gfpflags, node);
2799 EXPORT_SYMBOL(kmem_cache_alloc_node);
2801 #ifdef CONFIG_TRACING
2802 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2804 int node, size_t size)
2806 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2808 trace_kmalloc_node(_RET_IP_, ret,
2809 size, s->size, gfpflags, node);
2811 ret = kasan_kmalloc(s, ret, size, gfpflags);
2814 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2819 * Slow path handling. This may still be called frequently since objects
2820 * have a longer lifetime than the cpu slabs in most processing loads.
2822 * So we still attempt to reduce cache line usage. Just take the slab
2823 * lock and free the item. If there is no additional partial page
2824 * handling required then we can return immediately.
2826 static void __slab_free(struct kmem_cache *s, struct page *page,
2827 void *head, void *tail, int cnt,
2834 unsigned long counters;
2835 struct kmem_cache_node *n = NULL;
2836 unsigned long uninitialized_var(flags);
2838 stat(s, FREE_SLOWPATH);
2840 if (kmem_cache_debug(s) &&
2841 !free_debug_processing(s, page, head, tail, cnt, addr))
2846 spin_unlock_irqrestore(&n->list_lock, flags);
2849 prior = page->freelist;
2850 counters = page->counters;
2851 set_freepointer(s, tail, prior);
2852 new.counters = counters;
2853 was_frozen = new.frozen;
2855 if ((!new.inuse || !prior) && !was_frozen) {
2857 if (kmem_cache_has_cpu_partial(s) && !prior) {
2860 * Slab was on no list before and will be
2862 * We can defer the list move and instead
2867 } else { /* Needs to be taken off a list */
2869 n = get_node(s, page_to_nid(page));
2871 * Speculatively acquire the list_lock.
2872 * If the cmpxchg does not succeed then we may
2873 * drop the list_lock without any processing.
2875 * Otherwise the list_lock will synchronize with
2876 * other processors updating the list of slabs.
2878 spin_lock_irqsave(&n->list_lock, flags);
2883 } while (!cmpxchg_double_slab(s, page,
2891 * If we just froze the page then put it onto the
2892 * per cpu partial list.
2894 if (new.frozen && !was_frozen) {
2895 put_cpu_partial(s, page, 1);
2896 stat(s, CPU_PARTIAL_FREE);
2899 * The list lock was not taken therefore no list
2900 * activity can be necessary.
2903 stat(s, FREE_FROZEN);
2907 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2911 * Objects left in the slab. If it was not on the partial list before
2914 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2915 if (kmem_cache_debug(s))
2916 remove_full(s, n, page);
2917 add_partial(n, page, DEACTIVATE_TO_TAIL);
2918 stat(s, FREE_ADD_PARTIAL);
2920 spin_unlock_irqrestore(&n->list_lock, flags);
2926 * Slab on the partial list.
2928 remove_partial(n, page);
2929 stat(s, FREE_REMOVE_PARTIAL);
2931 /* Slab must be on the full list */
2932 remove_full(s, n, page);
2935 spin_unlock_irqrestore(&n->list_lock, flags);
2937 discard_slab(s, page);
2941 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2942 * can perform fastpath freeing without additional function calls.
2944 * The fastpath is only possible if we are freeing to the current cpu slab
2945 * of this processor. This typically the case if we have just allocated
2948 * If fastpath is not possible then fall back to __slab_free where we deal
2949 * with all sorts of special processing.
2951 * Bulk free of a freelist with several objects (all pointing to the
2952 * same page) possible by specifying head and tail ptr, plus objects
2953 * count (cnt). Bulk free indicated by tail pointer being set.
2955 static __always_inline void do_slab_free(struct kmem_cache *s,
2956 struct page *page, void *head, void *tail,
2957 int cnt, unsigned long addr)
2959 void *tail_obj = tail ? : head;
2960 struct kmem_cache_cpu *c;
2964 * Determine the currently cpus per cpu slab.
2965 * The cpu may change afterward. However that does not matter since
2966 * data is retrieved via this pointer. If we are on the same cpu
2967 * during the cmpxchg then the free will succeed.
2970 tid = this_cpu_read(s->cpu_slab->tid);
2971 c = raw_cpu_ptr(s->cpu_slab);
2972 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2973 unlikely(tid != READ_ONCE(c->tid)));
2975 /* Same with comment on barrier() in slab_alloc_node() */
2978 if (likely(page == c->page)) {
2979 set_freepointer(s, tail_obj, c->freelist);
2981 if (unlikely(!this_cpu_cmpxchg_double(
2982 s->cpu_slab->freelist, s->cpu_slab->tid,
2984 head, next_tid(tid)))) {
2986 note_cmpxchg_failure("slab_free", s, tid);
2989 stat(s, FREE_FASTPATH);
2991 __slab_free(s, page, head, tail_obj, cnt, addr);
2995 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2996 void *head, void *tail, int cnt,
3000 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3001 * to remove objects, whose reuse must be delayed.
3003 if (slab_free_freelist_hook(s, &head, &tail))
3004 do_slab_free(s, page, head, tail, cnt, addr);
3007 #ifdef CONFIG_KASAN_GENERIC
3008 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3010 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3014 void kmem_cache_free(struct kmem_cache *s, void *x)
3016 s = cache_from_obj(s, x);
3019 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3020 trace_kmem_cache_free(_RET_IP_, x);
3022 EXPORT_SYMBOL(kmem_cache_free);
3024 struct detached_freelist {
3029 struct kmem_cache *s;
3033 * This function progressively scans the array with free objects (with
3034 * a limited look ahead) and extract objects belonging to the same
3035 * page. It builds a detached freelist directly within the given
3036 * page/objects. This can happen without any need for
3037 * synchronization, because the objects are owned by running process.
3038 * The freelist is build up as a single linked list in the objects.
3039 * The idea is, that this detached freelist can then be bulk
3040 * transferred to the real freelist(s), but only requiring a single
3041 * synchronization primitive. Look ahead in the array is limited due
3042 * to performance reasons.
3045 int build_detached_freelist(struct kmem_cache *s, size_t size,
3046 void **p, struct detached_freelist *df)
3048 size_t first_skipped_index = 0;
3053 /* Always re-init detached_freelist */
3058 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3059 } while (!object && size);
3064 page = virt_to_head_page(object);
3066 /* Handle kalloc'ed objects */
3067 if (unlikely(!PageSlab(page))) {
3068 BUG_ON(!PageCompound(page));
3070 __free_pages(page, compound_order(page));
3071 p[size] = NULL; /* mark object processed */
3074 /* Derive kmem_cache from object */
3075 df->s = page->slab_cache;
3077 df->s = cache_from_obj(s, object); /* Support for memcg */
3080 /* Start new detached freelist */
3082 set_freepointer(df->s, object, NULL);
3084 df->freelist = object;
3085 p[size] = NULL; /* mark object processed */
3091 continue; /* Skip processed objects */
3093 /* df->page is always set at this point */
3094 if (df->page == virt_to_head_page(object)) {
3095 /* Opportunity build freelist */
3096 set_freepointer(df->s, object, df->freelist);
3097 df->freelist = object;
3099 p[size] = NULL; /* mark object processed */
3104 /* Limit look ahead search */
3108 if (!first_skipped_index)
3109 first_skipped_index = size + 1;
3112 return first_skipped_index;
3115 /* Note that interrupts must be enabled when calling this function. */
3116 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3122 struct detached_freelist df;
3124 size = build_detached_freelist(s, size, p, &df);
3128 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3129 } while (likely(size));
3131 EXPORT_SYMBOL(kmem_cache_free_bulk);
3133 /* Note that interrupts must be enabled when calling this function. */
3134 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3137 struct kmem_cache_cpu *c;
3140 /* memcg and kmem_cache debug support */
3141 s = slab_pre_alloc_hook(s, flags);
3145 * Drain objects in the per cpu slab, while disabling local
3146 * IRQs, which protects against PREEMPT and interrupts
3147 * handlers invoking normal fastpath.
3149 local_irq_disable();
3150 c = this_cpu_ptr(s->cpu_slab);
3152 for (i = 0; i < size; i++) {
3153 void *object = c->freelist;
3155 if (unlikely(!object)) {
3157 * Invoking slow path likely have side-effect
3158 * of re-populating per CPU c->freelist
3160 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3162 if (unlikely(!p[i]))
3165 c = this_cpu_ptr(s->cpu_slab);
3166 continue; /* goto for-loop */
3168 c->freelist = get_freepointer(s, object);
3171 c->tid = next_tid(c->tid);
3174 /* Clear memory outside IRQ disabled fastpath loop */
3175 if (unlikely(flags & __GFP_ZERO)) {
3178 for (j = 0; j < i; j++)
3179 memset(p[j], 0, s->object_size);
3182 /* memcg and kmem_cache debug support */
3183 slab_post_alloc_hook(s, flags, size, p);
3187 slab_post_alloc_hook(s, flags, i, p);
3188 __kmem_cache_free_bulk(s, i, p);
3191 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3195 * Object placement in a slab is made very easy because we always start at
3196 * offset 0. If we tune the size of the object to the alignment then we can
3197 * get the required alignment by putting one properly sized object after
3200 * Notice that the allocation order determines the sizes of the per cpu
3201 * caches. Each processor has always one slab available for allocations.
3202 * Increasing the allocation order reduces the number of times that slabs
3203 * must be moved on and off the partial lists and is therefore a factor in
3208 * Mininum / Maximum order of slab pages. This influences locking overhead
3209 * and slab fragmentation. A higher order reduces the number of partial slabs
3210 * and increases the number of allocations possible without having to
3211 * take the list_lock.
3213 static unsigned int slub_min_order;
3214 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3215 static unsigned int slub_min_objects;
3218 * Calculate the order of allocation given an slab object size.
3220 * The order of allocation has significant impact on performance and other
3221 * system components. Generally order 0 allocations should be preferred since
3222 * order 0 does not cause fragmentation in the page allocator. Larger objects
3223 * be problematic to put into order 0 slabs because there may be too much
3224 * unused space left. We go to a higher order if more than 1/16th of the slab
3227 * In order to reach satisfactory performance we must ensure that a minimum
3228 * number of objects is in one slab. Otherwise we may generate too much
3229 * activity on the partial lists which requires taking the list_lock. This is
3230 * less a concern for large slabs though which are rarely used.
3232 * slub_max_order specifies the order where we begin to stop considering the
3233 * number of objects in a slab as critical. If we reach slub_max_order then
3234 * we try to keep the page order as low as possible. So we accept more waste
3235 * of space in favor of a small page order.
3237 * Higher order allocations also allow the placement of more objects in a
3238 * slab and thereby reduce object handling overhead. If the user has
3239 * requested a higher mininum order then we start with that one instead of
3240 * the smallest order which will fit the object.
3242 static inline unsigned int slab_order(unsigned int size,
3243 unsigned int min_objects, unsigned int max_order,
3244 unsigned int fract_leftover)
3246 unsigned int min_order = slub_min_order;
3249 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3250 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3252 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3253 order <= max_order; order++) {
3255 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3258 rem = slab_size % size;
3260 if (rem <= slab_size / fract_leftover)
3267 static inline int calculate_order(unsigned int size)
3270 unsigned int min_objects;
3271 unsigned int max_objects;
3274 * Attempt to find best configuration for a slab. This
3275 * works by first attempting to generate a layout with
3276 * the best configuration and backing off gradually.
3278 * First we increase the acceptable waste in a slab. Then
3279 * we reduce the minimum objects required in a slab.
3281 min_objects = slub_min_objects;
3283 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3284 max_objects = order_objects(slub_max_order, size);
3285 min_objects = min(min_objects, max_objects);
3287 while (min_objects > 1) {
3288 unsigned int fraction;
3291 while (fraction >= 4) {
3292 order = slab_order(size, min_objects,
3293 slub_max_order, fraction);
3294 if (order <= slub_max_order)
3302 * We were unable to place multiple objects in a slab. Now
3303 * lets see if we can place a single object there.
3305 order = slab_order(size, 1, slub_max_order, 1);
3306 if (order <= slub_max_order)
3310 * Doh this slab cannot be placed using slub_max_order.
3312 order = slab_order(size, 1, MAX_ORDER, 1);
3313 if (order < MAX_ORDER)
3319 init_kmem_cache_node(struct kmem_cache_node *n)
3322 spin_lock_init(&n->list_lock);
3323 INIT_LIST_HEAD(&n->partial);
3324 #ifdef CONFIG_SLUB_DEBUG
3325 atomic_long_set(&n->nr_slabs, 0);
3326 atomic_long_set(&n->total_objects, 0);
3327 INIT_LIST_HEAD(&n->full);
3331 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3333 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3334 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3337 * Must align to double word boundary for the double cmpxchg
3338 * instructions to work; see __pcpu_double_call_return_bool().
3340 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3341 2 * sizeof(void *));
3346 init_kmem_cache_cpus(s);
3351 static struct kmem_cache *kmem_cache_node;
3354 * No kmalloc_node yet so do it by hand. We know that this is the first
3355 * slab on the node for this slabcache. There are no concurrent accesses
3358 * Note that this function only works on the kmem_cache_node
3359 * when allocating for the kmem_cache_node. This is used for bootstrapping
3360 * memory on a fresh node that has no slab structures yet.
3362 static void early_kmem_cache_node_alloc(int node)
3365 struct kmem_cache_node *n;
3367 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3369 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3372 if (page_to_nid(page) != node) {
3373 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3374 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3379 #ifdef CONFIG_SLUB_DEBUG
3380 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3381 init_tracking(kmem_cache_node, n);
3383 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3385 page->freelist = get_freepointer(kmem_cache_node, n);
3388 kmem_cache_node->node[node] = n;
3389 init_kmem_cache_node(n);
3390 inc_slabs_node(kmem_cache_node, node, page->objects);
3393 * No locks need to be taken here as it has just been
3394 * initialized and there is no concurrent access.
3396 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3399 static void free_kmem_cache_nodes(struct kmem_cache *s)
3402 struct kmem_cache_node *n;
3404 for_each_kmem_cache_node(s, node, n) {
3405 s->node[node] = NULL;
3406 kmem_cache_free(kmem_cache_node, n);
3410 void __kmem_cache_release(struct kmem_cache *s)
3412 cache_random_seq_destroy(s);
3413 free_percpu(s->cpu_slab);
3414 free_kmem_cache_nodes(s);
3417 static int init_kmem_cache_nodes(struct kmem_cache *s)
3421 for_each_node_state(node, N_NORMAL_MEMORY) {
3422 struct kmem_cache_node *n;
3424 if (slab_state == DOWN) {
3425 early_kmem_cache_node_alloc(node);
3428 n = kmem_cache_alloc_node(kmem_cache_node,
3432 free_kmem_cache_nodes(s);
3436 init_kmem_cache_node(n);
3442 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3444 if (min < MIN_PARTIAL)
3446 else if (min > MAX_PARTIAL)
3448 s->min_partial = min;
3451 static void set_cpu_partial(struct kmem_cache *s)
3453 #ifdef CONFIG_SLUB_CPU_PARTIAL
3455 * cpu_partial determined the maximum number of objects kept in the
3456 * per cpu partial lists of a processor.
3458 * Per cpu partial lists mainly contain slabs that just have one
3459 * object freed. If they are used for allocation then they can be
3460 * filled up again with minimal effort. The slab will never hit the
3461 * per node partial lists and therefore no locking will be required.
3463 * This setting also determines
3465 * A) The number of objects from per cpu partial slabs dumped to the
3466 * per node list when we reach the limit.
3467 * B) The number of objects in cpu partial slabs to extract from the
3468 * per node list when we run out of per cpu objects. We only fetch
3469 * 50% to keep some capacity around for frees.
3471 if (!kmem_cache_has_cpu_partial(s))
3473 else if (s->size >= PAGE_SIZE)
3475 else if (s->size >= 1024)
3477 else if (s->size >= 256)
3478 s->cpu_partial = 13;
3480 s->cpu_partial = 30;
3485 * calculate_sizes() determines the order and the distribution of data within
3488 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3490 slab_flags_t flags = s->flags;
3491 unsigned int size = s->object_size;
3495 * Round up object size to the next word boundary. We can only
3496 * place the free pointer at word boundaries and this determines
3497 * the possible location of the free pointer.
3499 size = ALIGN(size, sizeof(void *));
3501 #ifdef CONFIG_SLUB_DEBUG
3503 * Determine if we can poison the object itself. If the user of
3504 * the slab may touch the object after free or before allocation
3505 * then we should never poison the object itself.
3507 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3509 s->flags |= __OBJECT_POISON;
3511 s->flags &= ~__OBJECT_POISON;
3515 * If we are Redzoning then check if there is some space between the
3516 * end of the object and the free pointer. If not then add an
3517 * additional word to have some bytes to store Redzone information.
3519 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3520 size += sizeof(void *);
3524 * With that we have determined the number of bytes in actual use
3525 * by the object. This is the potential offset to the free pointer.
3529 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3532 * Relocate free pointer after the object if it is not
3533 * permitted to overwrite the first word of the object on
3536 * This is the case if we do RCU, have a constructor or
3537 * destructor or are poisoning the objects.
3540 size += sizeof(void *);
3543 #ifdef CONFIG_SLUB_DEBUG
3544 if (flags & SLAB_STORE_USER)
3546 * Need to store information about allocs and frees after
3549 size += 2 * sizeof(struct track);
3552 kasan_cache_create(s, &size, &s->flags);
3553 #ifdef CONFIG_SLUB_DEBUG
3554 if (flags & SLAB_RED_ZONE) {
3556 * Add some empty padding so that we can catch
3557 * overwrites from earlier objects rather than let
3558 * tracking information or the free pointer be
3559 * corrupted if a user writes before the start
3562 size += sizeof(void *);
3564 s->red_left_pad = sizeof(void *);
3565 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3566 size += s->red_left_pad;
3571 * SLUB stores one object immediately after another beginning from
3572 * offset 0. In order to align the objects we have to simply size
3573 * each object to conform to the alignment.
3575 size = ALIGN(size, s->align);
3577 if (forced_order >= 0)
3578 order = forced_order;
3580 order = calculate_order(size);
3587 s->allocflags |= __GFP_COMP;
3589 if (s->flags & SLAB_CACHE_DMA)
3590 s->allocflags |= GFP_DMA;
3592 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3593 s->allocflags |= __GFP_RECLAIMABLE;
3596 * Determine the number of objects per slab
3598 s->oo = oo_make(order, size);
3599 s->min = oo_make(get_order(size), size);
3600 if (oo_objects(s->oo) > oo_objects(s->max))
3603 return !!oo_objects(s->oo);
3606 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3608 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3609 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3610 s->random = get_random_long();
3613 if (!calculate_sizes(s, -1))
3615 if (disable_higher_order_debug) {
3617 * Disable debugging flags that store metadata if the min slab
3620 if (get_order(s->size) > get_order(s->object_size)) {
3621 s->flags &= ~DEBUG_METADATA_FLAGS;
3623 if (!calculate_sizes(s, -1))
3628 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3629 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3630 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3631 /* Enable fast mode */
3632 s->flags |= __CMPXCHG_DOUBLE;
3636 * The larger the object size is, the more pages we want on the partial
3637 * list to avoid pounding the page allocator excessively.
3639 set_min_partial(s, ilog2(s->size) / 2);
3644 s->remote_node_defrag_ratio = 1000;
3647 /* Initialize the pre-computed randomized freelist if slab is up */
3648 if (slab_state >= UP) {
3649 if (init_cache_random_seq(s))
3653 if (!init_kmem_cache_nodes(s))
3656 if (alloc_kmem_cache_cpus(s))
3659 free_kmem_cache_nodes(s);
3661 if (flags & SLAB_PANIC)
3662 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3663 s->name, s->size, s->size,
3664 oo_order(s->oo), s->offset, (unsigned long)flags);
3668 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3671 #ifdef CONFIG_SLUB_DEBUG
3672 void *addr = page_address(page);
3674 unsigned long *map = bitmap_zalloc(page->objects, GFP_ATOMIC);
3677 slab_err(s, page, text, s->name);
3680 get_map(s, page, map);
3681 for_each_object(p, s, addr, page->objects) {
3683 if (!test_bit(slab_index(p, s, addr), map)) {
3684 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3685 print_tracking(s, p);
3694 * Attempt to free all partial slabs on a node.
3695 * This is called from __kmem_cache_shutdown(). We must take list_lock
3696 * because sysfs file might still access partial list after the shutdowning.
3698 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3701 struct page *page, *h;
3703 BUG_ON(irqs_disabled());
3704 spin_lock_irq(&n->list_lock);
3705 list_for_each_entry_safe(page, h, &n->partial, lru) {
3707 remove_partial(n, page);
3708 list_add(&page->lru, &discard);
3710 list_slab_objects(s, page,
3711 "Objects remaining in %s on __kmem_cache_shutdown()");
3714 spin_unlock_irq(&n->list_lock);
3716 list_for_each_entry_safe(page, h, &discard, lru)
3717 discard_slab(s, page);
3720 bool __kmem_cache_empty(struct kmem_cache *s)
3723 struct kmem_cache_node *n;
3725 for_each_kmem_cache_node(s, node, n)
3726 if (n->nr_partial || slabs_node(s, node))
3732 * Release all resources used by a slab cache.
3734 int __kmem_cache_shutdown(struct kmem_cache *s)
3737 struct kmem_cache_node *n;
3740 /* Attempt to free all objects */
3741 for_each_kmem_cache_node(s, node, n) {
3743 if (n->nr_partial || slabs_node(s, node))
3746 sysfs_slab_remove(s);
3750 /********************************************************************
3752 *******************************************************************/
3754 static int __init setup_slub_min_order(char *str)
3756 get_option(&str, (int *)&slub_min_order);
3761 __setup("slub_min_order=", setup_slub_min_order);
3763 static int __init setup_slub_max_order(char *str)
3765 get_option(&str, (int *)&slub_max_order);
3766 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3771 __setup("slub_max_order=", setup_slub_max_order);
3773 static int __init setup_slub_min_objects(char *str)
3775 get_option(&str, (int *)&slub_min_objects);
3780 __setup("slub_min_objects=", setup_slub_min_objects);
3782 void *__kmalloc(size_t size, gfp_t flags)
3784 struct kmem_cache *s;
3787 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3788 return kmalloc_large(size, flags);
3790 s = kmalloc_slab(size, flags);
3792 if (unlikely(ZERO_OR_NULL_PTR(s)))
3795 ret = slab_alloc(s, flags, _RET_IP_);
3797 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3799 ret = kasan_kmalloc(s, ret, size, flags);
3803 EXPORT_SYMBOL(__kmalloc);
3806 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3811 flags |= __GFP_COMP;
3812 page = alloc_pages_node(node, flags, get_order(size));
3814 ptr = page_address(page);
3816 return kmalloc_large_node_hook(ptr, size, flags);
3819 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3821 struct kmem_cache *s;
3824 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3825 ret = kmalloc_large_node(size, flags, node);
3827 trace_kmalloc_node(_RET_IP_, ret,
3828 size, PAGE_SIZE << get_order(size),
3834 s = kmalloc_slab(size, flags);
3836 if (unlikely(ZERO_OR_NULL_PTR(s)))
3839 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3841 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3843 ret = kasan_kmalloc(s, ret, size, flags);
3847 EXPORT_SYMBOL(__kmalloc_node);
3850 #ifdef CONFIG_HARDENED_USERCOPY
3852 * Rejects incorrectly sized objects and objects that are to be copied
3853 * to/from userspace but do not fall entirely within the containing slab
3854 * cache's usercopy region.
3856 * Returns NULL if check passes, otherwise const char * to name of cache
3857 * to indicate an error.
3859 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3862 struct kmem_cache *s;
3863 unsigned int offset;
3866 ptr = kasan_reset_tag(ptr);
3868 /* Find object and usable object size. */
3869 s = page->slab_cache;
3871 /* Reject impossible pointers. */
3872 if (ptr < page_address(page))
3873 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3876 /* Find offset within object. */
3877 offset = (ptr - page_address(page)) % s->size;
3879 /* Adjust for redzone and reject if within the redzone. */
3880 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3881 if (offset < s->red_left_pad)
3882 usercopy_abort("SLUB object in left red zone",
3883 s->name, to_user, offset, n);
3884 offset -= s->red_left_pad;
3887 /* Allow address range falling entirely within usercopy region. */
3888 if (offset >= s->useroffset &&
3889 offset - s->useroffset <= s->usersize &&
3890 n <= s->useroffset - offset + s->usersize)
3894 * If the copy is still within the allocated object, produce
3895 * a warning instead of rejecting the copy. This is intended
3896 * to be a temporary method to find any missing usercopy
3899 object_size = slab_ksize(s);
3900 if (usercopy_fallback &&
3901 offset <= object_size && n <= object_size - offset) {
3902 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3906 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3908 #endif /* CONFIG_HARDENED_USERCOPY */
3910 static size_t __ksize(const void *object)
3914 if (unlikely(object == ZERO_SIZE_PTR))
3917 page = virt_to_head_page(object);
3919 if (unlikely(!PageSlab(page))) {
3920 WARN_ON(!PageCompound(page));
3921 return PAGE_SIZE << compound_order(page);
3924 return slab_ksize(page->slab_cache);
3927 size_t ksize(const void *object)
3929 size_t size = __ksize(object);
3930 /* We assume that ksize callers could use whole allocated area,
3931 * so we need to unpoison this area.
3933 kasan_unpoison_shadow(object, size);
3936 EXPORT_SYMBOL(ksize);
3938 void kfree(const void *x)
3941 void *object = (void *)x;
3943 trace_kfree(_RET_IP_, x);
3945 if (unlikely(ZERO_OR_NULL_PTR(x)))
3948 page = virt_to_head_page(x);
3949 if (unlikely(!PageSlab(page))) {
3950 BUG_ON(!PageCompound(page));
3952 __free_pages(page, compound_order(page));
3955 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3957 EXPORT_SYMBOL(kfree);
3959 #define SHRINK_PROMOTE_MAX 32
3962 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3963 * up most to the head of the partial lists. New allocations will then
3964 * fill those up and thus they can be removed from the partial lists.
3966 * The slabs with the least items are placed last. This results in them
3967 * being allocated from last increasing the chance that the last objects
3968 * are freed in them.
3970 int __kmem_cache_shrink(struct kmem_cache *s)
3974 struct kmem_cache_node *n;
3977 struct list_head discard;
3978 struct list_head promote[SHRINK_PROMOTE_MAX];
3979 unsigned long flags;
3983 for_each_kmem_cache_node(s, node, n) {
3984 INIT_LIST_HEAD(&discard);
3985 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3986 INIT_LIST_HEAD(promote + i);
3988 spin_lock_irqsave(&n->list_lock, flags);
3991 * Build lists of slabs to discard or promote.
3993 * Note that concurrent frees may occur while we hold the
3994 * list_lock. page->inuse here is the upper limit.
3996 list_for_each_entry_safe(page, t, &n->partial, lru) {
3997 int free = page->objects - page->inuse;
3999 /* Do not reread page->inuse */
4002 /* We do not keep full slabs on the list */
4005 if (free == page->objects) {
4006 list_move(&page->lru, &discard);
4008 } else if (free <= SHRINK_PROMOTE_MAX)
4009 list_move(&page->lru, promote + free - 1);
4013 * Promote the slabs filled up most to the head of the
4016 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4017 list_splice(promote + i, &n->partial);
4019 spin_unlock_irqrestore(&n->list_lock, flags);
4021 /* Release empty slabs */
4022 list_for_each_entry_safe(page, t, &discard, lru)
4023 discard_slab(s, page);
4025 if (slabs_node(s, node))
4033 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
4036 * Called with all the locks held after a sched RCU grace period.
4037 * Even if @s becomes empty after shrinking, we can't know that @s
4038 * doesn't have allocations already in-flight and thus can't
4039 * destroy @s until the associated memcg is released.
4041 * However, let's remove the sysfs files for empty caches here.
4042 * Each cache has a lot of interface files which aren't
4043 * particularly useful for empty draining caches; otherwise, we can
4044 * easily end up with millions of unnecessary sysfs files on
4045 * systems which have a lot of memory and transient cgroups.
4047 if (!__kmem_cache_shrink(s))
4048 sysfs_slab_remove(s);
4051 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4054 * Disable empty slabs caching. Used to avoid pinning offline
4055 * memory cgroups by kmem pages that can be freed.
4057 slub_set_cpu_partial(s, 0);
4061 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4062 * we have to make sure the change is visible before shrinking.
4064 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4068 static int slab_mem_going_offline_callback(void *arg)
4070 struct kmem_cache *s;
4072 mutex_lock(&slab_mutex);
4073 list_for_each_entry(s, &slab_caches, list)
4074 __kmem_cache_shrink(s);
4075 mutex_unlock(&slab_mutex);
4080 static void slab_mem_offline_callback(void *arg)
4082 struct kmem_cache_node *n;
4083 struct kmem_cache *s;
4084 struct memory_notify *marg = arg;
4087 offline_node = marg->status_change_nid_normal;
4090 * If the node still has available memory. we need kmem_cache_node
4093 if (offline_node < 0)
4096 mutex_lock(&slab_mutex);
4097 list_for_each_entry(s, &slab_caches, list) {
4098 n = get_node(s, offline_node);
4101 * if n->nr_slabs > 0, slabs still exist on the node
4102 * that is going down. We were unable to free them,
4103 * and offline_pages() function shouldn't call this
4104 * callback. So, we must fail.
4106 BUG_ON(slabs_node(s, offline_node));
4108 s->node[offline_node] = NULL;
4109 kmem_cache_free(kmem_cache_node, n);
4112 mutex_unlock(&slab_mutex);
4115 static int slab_mem_going_online_callback(void *arg)
4117 struct kmem_cache_node *n;
4118 struct kmem_cache *s;
4119 struct memory_notify *marg = arg;
4120 int nid = marg->status_change_nid_normal;
4124 * If the node's memory is already available, then kmem_cache_node is
4125 * already created. Nothing to do.
4131 * We are bringing a node online. No memory is available yet. We must
4132 * allocate a kmem_cache_node structure in order to bring the node
4135 mutex_lock(&slab_mutex);
4136 list_for_each_entry(s, &slab_caches, list) {
4138 * XXX: kmem_cache_alloc_node will fallback to other nodes
4139 * since memory is not yet available from the node that
4142 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4147 init_kmem_cache_node(n);
4151 mutex_unlock(&slab_mutex);
4155 static int slab_memory_callback(struct notifier_block *self,
4156 unsigned long action, void *arg)
4161 case MEM_GOING_ONLINE:
4162 ret = slab_mem_going_online_callback(arg);
4164 case MEM_GOING_OFFLINE:
4165 ret = slab_mem_going_offline_callback(arg);
4168 case MEM_CANCEL_ONLINE:
4169 slab_mem_offline_callback(arg);
4172 case MEM_CANCEL_OFFLINE:
4176 ret = notifier_from_errno(ret);
4182 static struct notifier_block slab_memory_callback_nb = {
4183 .notifier_call = slab_memory_callback,
4184 .priority = SLAB_CALLBACK_PRI,
4187 /********************************************************************
4188 * Basic setup of slabs
4189 *******************************************************************/
4192 * Used for early kmem_cache structures that were allocated using
4193 * the page allocator. Allocate them properly then fix up the pointers
4194 * that may be pointing to the wrong kmem_cache structure.
4197 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4200 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4201 struct kmem_cache_node *n;
4203 memcpy(s, static_cache, kmem_cache->object_size);
4206 * This runs very early, and only the boot processor is supposed to be
4207 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4210 __flush_cpu_slab(s, smp_processor_id());
4211 for_each_kmem_cache_node(s, node, n) {
4214 list_for_each_entry(p, &n->partial, lru)
4217 #ifdef CONFIG_SLUB_DEBUG
4218 list_for_each_entry(p, &n->full, lru)
4222 slab_init_memcg_params(s);
4223 list_add(&s->list, &slab_caches);
4224 memcg_link_cache(s);
4228 void __init kmem_cache_init(void)
4230 static __initdata struct kmem_cache boot_kmem_cache,
4231 boot_kmem_cache_node;
4233 if (debug_guardpage_minorder())
4236 kmem_cache_node = &boot_kmem_cache_node;
4237 kmem_cache = &boot_kmem_cache;
4239 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4240 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4242 register_hotmemory_notifier(&slab_memory_callback_nb);
4244 /* Able to allocate the per node structures */
4245 slab_state = PARTIAL;
4247 create_boot_cache(kmem_cache, "kmem_cache",
4248 offsetof(struct kmem_cache, node) +
4249 nr_node_ids * sizeof(struct kmem_cache_node *),
4250 SLAB_HWCACHE_ALIGN, 0, 0);
4252 kmem_cache = bootstrap(&boot_kmem_cache);
4253 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4255 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4256 setup_kmalloc_cache_index_table();
4257 create_kmalloc_caches(0);
4259 /* Setup random freelists for each cache */
4260 init_freelist_randomization();
4262 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4265 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4267 slub_min_order, slub_max_order, slub_min_objects,
4268 nr_cpu_ids, nr_node_ids);
4271 void __init kmem_cache_init_late(void)
4276 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4277 slab_flags_t flags, void (*ctor)(void *))
4279 struct kmem_cache *s, *c;
4281 s = find_mergeable(size, align, flags, name, ctor);
4286 * Adjust the object sizes so that we clear
4287 * the complete object on kzalloc.
4289 s->object_size = max(s->object_size, size);
4290 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4292 for_each_memcg_cache(c, s) {
4293 c->object_size = s->object_size;
4294 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4297 if (sysfs_slab_alias(s, name)) {
4306 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4310 err = kmem_cache_open(s, flags);
4314 /* Mutex is not taken during early boot */
4315 if (slab_state <= UP)
4318 memcg_propagate_slab_attrs(s);
4319 err = sysfs_slab_add(s);
4321 __kmem_cache_release(s);
4326 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4328 struct kmem_cache *s;
4331 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4332 return kmalloc_large(size, gfpflags);
4334 s = kmalloc_slab(size, gfpflags);
4336 if (unlikely(ZERO_OR_NULL_PTR(s)))
4339 ret = slab_alloc(s, gfpflags, caller);
4341 /* Honor the call site pointer we received. */
4342 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4348 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4349 int node, unsigned long caller)
4351 struct kmem_cache *s;
4354 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4355 ret = kmalloc_large_node(size, gfpflags, node);
4357 trace_kmalloc_node(caller, ret,
4358 size, PAGE_SIZE << get_order(size),
4364 s = kmalloc_slab(size, gfpflags);
4366 if (unlikely(ZERO_OR_NULL_PTR(s)))
4369 ret = slab_alloc_node(s, gfpflags, node, caller);
4371 /* Honor the call site pointer we received. */
4372 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4379 static int count_inuse(struct page *page)
4384 static int count_total(struct page *page)
4386 return page->objects;
4390 #ifdef CONFIG_SLUB_DEBUG
4391 static int validate_slab(struct kmem_cache *s, struct page *page,
4395 void *addr = page_address(page);
4397 if (!check_slab(s, page) ||
4398 !on_freelist(s, page, NULL))
4401 /* Now we know that a valid freelist exists */
4402 bitmap_zero(map, page->objects);
4404 get_map(s, page, map);
4405 for_each_object(p, s, addr, page->objects) {
4406 if (test_bit(slab_index(p, s, addr), map))
4407 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4411 for_each_object(p, s, addr, page->objects)
4412 if (!test_bit(slab_index(p, s, addr), map))
4413 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4418 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4422 validate_slab(s, page, map);
4426 static int validate_slab_node(struct kmem_cache *s,
4427 struct kmem_cache_node *n, unsigned long *map)
4429 unsigned long count = 0;
4431 unsigned long flags;
4433 spin_lock_irqsave(&n->list_lock, flags);
4435 list_for_each_entry(page, &n->partial, lru) {
4436 validate_slab_slab(s, page, map);
4439 if (count != n->nr_partial)
4440 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4441 s->name, count, n->nr_partial);
4443 if (!(s->flags & SLAB_STORE_USER))
4446 list_for_each_entry(page, &n->full, lru) {
4447 validate_slab_slab(s, page, map);
4450 if (count != atomic_long_read(&n->nr_slabs))
4451 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4452 s->name, count, atomic_long_read(&n->nr_slabs));
4455 spin_unlock_irqrestore(&n->list_lock, flags);
4459 static long validate_slab_cache(struct kmem_cache *s)
4462 unsigned long count = 0;
4463 struct kmem_cache_node *n;
4464 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4470 for_each_kmem_cache_node(s, node, n)
4471 count += validate_slab_node(s, n, map);
4476 * Generate lists of code addresses where slabcache objects are allocated
4481 unsigned long count;
4488 DECLARE_BITMAP(cpus, NR_CPUS);
4494 unsigned long count;
4495 struct location *loc;
4498 static void free_loc_track(struct loc_track *t)
4501 free_pages((unsigned long)t->loc,
4502 get_order(sizeof(struct location) * t->max));
4505 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4510 order = get_order(sizeof(struct location) * max);
4512 l = (void *)__get_free_pages(flags, order);
4517 memcpy(l, t->loc, sizeof(struct location) * t->count);
4525 static int add_location(struct loc_track *t, struct kmem_cache *s,
4526 const struct track *track)
4528 long start, end, pos;
4530 unsigned long caddr;
4531 unsigned long age = jiffies - track->when;
4537 pos = start + (end - start + 1) / 2;
4540 * There is nothing at "end". If we end up there
4541 * we need to add something to before end.
4546 caddr = t->loc[pos].addr;
4547 if (track->addr == caddr) {
4553 if (age < l->min_time)
4555 if (age > l->max_time)
4558 if (track->pid < l->min_pid)
4559 l->min_pid = track->pid;
4560 if (track->pid > l->max_pid)
4561 l->max_pid = track->pid;
4563 cpumask_set_cpu(track->cpu,
4564 to_cpumask(l->cpus));
4566 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4570 if (track->addr < caddr)
4577 * Not found. Insert new tracking element.
4579 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4585 (t->count - pos) * sizeof(struct location));
4588 l->addr = track->addr;
4592 l->min_pid = track->pid;
4593 l->max_pid = track->pid;
4594 cpumask_clear(to_cpumask(l->cpus));
4595 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4596 nodes_clear(l->nodes);
4597 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4601 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4602 struct page *page, enum track_item alloc,
4605 void *addr = page_address(page);
4608 bitmap_zero(map, page->objects);
4609 get_map(s, page, map);
4611 for_each_object(p, s, addr, page->objects)
4612 if (!test_bit(slab_index(p, s, addr), map))
4613 add_location(t, s, get_track(s, p, alloc));
4616 static int list_locations(struct kmem_cache *s, char *buf,
4617 enum track_item alloc)
4621 struct loc_track t = { 0, 0, NULL };
4623 struct kmem_cache_node *n;
4624 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4626 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4629 return sprintf(buf, "Out of memory\n");
4631 /* Push back cpu slabs */
4634 for_each_kmem_cache_node(s, node, n) {
4635 unsigned long flags;
4638 if (!atomic_long_read(&n->nr_slabs))
4641 spin_lock_irqsave(&n->list_lock, flags);
4642 list_for_each_entry(page, &n->partial, lru)
4643 process_slab(&t, s, page, alloc, map);
4644 list_for_each_entry(page, &n->full, lru)
4645 process_slab(&t, s, page, alloc, map);
4646 spin_unlock_irqrestore(&n->list_lock, flags);
4649 for (i = 0; i < t.count; i++) {
4650 struct location *l = &t.loc[i];
4652 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4654 len += sprintf(buf + len, "%7ld ", l->count);
4657 len += sprintf(buf + len, "%pS", (void *)l->addr);
4659 len += sprintf(buf + len, "<not-available>");
4661 if (l->sum_time != l->min_time) {
4662 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4664 (long)div_u64(l->sum_time, l->count),
4667 len += sprintf(buf + len, " age=%ld",
4670 if (l->min_pid != l->max_pid)
4671 len += sprintf(buf + len, " pid=%ld-%ld",
4672 l->min_pid, l->max_pid);
4674 len += sprintf(buf + len, " pid=%ld",
4677 if (num_online_cpus() > 1 &&
4678 !cpumask_empty(to_cpumask(l->cpus)) &&
4679 len < PAGE_SIZE - 60)
4680 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4682 cpumask_pr_args(to_cpumask(l->cpus)));
4684 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4685 len < PAGE_SIZE - 60)
4686 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4688 nodemask_pr_args(&l->nodes));
4690 len += sprintf(buf + len, "\n");
4696 len += sprintf(buf, "No data\n");
4701 #ifdef SLUB_RESILIENCY_TEST
4702 static void __init resiliency_test(void)
4705 int type = KMALLOC_NORMAL;
4707 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4709 pr_err("SLUB resiliency testing\n");
4710 pr_err("-----------------------\n");
4711 pr_err("A. Corruption after allocation\n");
4713 p = kzalloc(16, GFP_KERNEL);
4715 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4718 validate_slab_cache(kmalloc_caches[type][4]);
4720 /* Hmmm... The next two are dangerous */
4721 p = kzalloc(32, GFP_KERNEL);
4722 p[32 + sizeof(void *)] = 0x34;
4723 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4725 pr_err("If allocated object is overwritten then not detectable\n\n");
4727 validate_slab_cache(kmalloc_caches[type][5]);
4728 p = kzalloc(64, GFP_KERNEL);
4729 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4731 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4733 pr_err("If allocated object is overwritten then not detectable\n\n");
4734 validate_slab_cache(kmalloc_caches[type][6]);
4736 pr_err("\nB. Corruption after free\n");
4737 p = kzalloc(128, GFP_KERNEL);
4740 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4741 validate_slab_cache(kmalloc_caches[type][7]);
4743 p = kzalloc(256, GFP_KERNEL);
4746 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4747 validate_slab_cache(kmalloc_caches[type][8]);
4749 p = kzalloc(512, GFP_KERNEL);
4752 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4753 validate_slab_cache(kmalloc_caches[type][9]);
4757 static void resiliency_test(void) {};
4762 enum slab_stat_type {
4763 SL_ALL, /* All slabs */
4764 SL_PARTIAL, /* Only partially allocated slabs */
4765 SL_CPU, /* Only slabs used for cpu caches */
4766 SL_OBJECTS, /* Determine allocated objects not slabs */
4767 SL_TOTAL /* Determine object capacity not slabs */
4770 #define SO_ALL (1 << SL_ALL)
4771 #define SO_PARTIAL (1 << SL_PARTIAL)
4772 #define SO_CPU (1 << SL_CPU)
4773 #define SO_OBJECTS (1 << SL_OBJECTS)
4774 #define SO_TOTAL (1 << SL_TOTAL)
4777 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4779 static int __init setup_slub_memcg_sysfs(char *str)
4783 if (get_option(&str, &v) > 0)
4784 memcg_sysfs_enabled = v;
4789 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4792 static ssize_t show_slab_objects(struct kmem_cache *s,
4793 char *buf, unsigned long flags)
4795 unsigned long total = 0;
4798 unsigned long *nodes;
4800 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4804 if (flags & SO_CPU) {
4807 for_each_possible_cpu(cpu) {
4808 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4813 page = READ_ONCE(c->page);
4817 node = page_to_nid(page);
4818 if (flags & SO_TOTAL)
4820 else if (flags & SO_OBJECTS)
4828 page = slub_percpu_partial_read_once(c);
4830 node = page_to_nid(page);
4831 if (flags & SO_TOTAL)
4833 else if (flags & SO_OBJECTS)
4844 #ifdef CONFIG_SLUB_DEBUG
4845 if (flags & SO_ALL) {
4846 struct kmem_cache_node *n;
4848 for_each_kmem_cache_node(s, node, n) {
4850 if (flags & SO_TOTAL)
4851 x = atomic_long_read(&n->total_objects);
4852 else if (flags & SO_OBJECTS)
4853 x = atomic_long_read(&n->total_objects) -
4854 count_partial(n, count_free);
4856 x = atomic_long_read(&n->nr_slabs);
4863 if (flags & SO_PARTIAL) {
4864 struct kmem_cache_node *n;
4866 for_each_kmem_cache_node(s, node, n) {
4867 if (flags & SO_TOTAL)
4868 x = count_partial(n, count_total);
4869 else if (flags & SO_OBJECTS)
4870 x = count_partial(n, count_inuse);
4877 x = sprintf(buf, "%lu", total);
4879 for (node = 0; node < nr_node_ids; node++)
4881 x += sprintf(buf + x, " N%d=%lu",
4886 return x + sprintf(buf + x, "\n");
4889 #ifdef CONFIG_SLUB_DEBUG
4890 static int any_slab_objects(struct kmem_cache *s)
4893 struct kmem_cache_node *n;
4895 for_each_kmem_cache_node(s, node, n)
4896 if (atomic_long_read(&n->total_objects))
4903 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4904 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4906 struct slab_attribute {
4907 struct attribute attr;
4908 ssize_t (*show)(struct kmem_cache *s, char *buf);
4909 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4912 #define SLAB_ATTR_RO(_name) \
4913 static struct slab_attribute _name##_attr = \
4914 __ATTR(_name, 0400, _name##_show, NULL)
4916 #define SLAB_ATTR(_name) \
4917 static struct slab_attribute _name##_attr = \
4918 __ATTR(_name, 0600, _name##_show, _name##_store)
4920 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4922 return sprintf(buf, "%u\n", s->size);
4924 SLAB_ATTR_RO(slab_size);
4926 static ssize_t align_show(struct kmem_cache *s, char *buf)
4928 return sprintf(buf, "%u\n", s->align);
4930 SLAB_ATTR_RO(align);
4932 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4934 return sprintf(buf, "%u\n", s->object_size);
4936 SLAB_ATTR_RO(object_size);
4938 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4940 return sprintf(buf, "%u\n", oo_objects(s->oo));
4942 SLAB_ATTR_RO(objs_per_slab);
4944 static ssize_t order_store(struct kmem_cache *s,
4945 const char *buf, size_t length)
4950 err = kstrtouint(buf, 10, &order);
4954 if (order > slub_max_order || order < slub_min_order)
4957 calculate_sizes(s, order);
4961 static ssize_t order_show(struct kmem_cache *s, char *buf)
4963 return sprintf(buf, "%u\n", oo_order(s->oo));
4967 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4969 return sprintf(buf, "%lu\n", s->min_partial);
4972 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4978 err = kstrtoul(buf, 10, &min);
4982 set_min_partial(s, min);
4985 SLAB_ATTR(min_partial);
4987 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4989 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4992 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4995 unsigned int objects;
4998 err = kstrtouint(buf, 10, &objects);
5001 if (objects && !kmem_cache_has_cpu_partial(s))
5004 slub_set_cpu_partial(s, objects);
5008 SLAB_ATTR(cpu_partial);
5010 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5014 return sprintf(buf, "%pS\n", s->ctor);
5018 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5020 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5022 SLAB_ATTR_RO(aliases);
5024 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5026 return show_slab_objects(s, buf, SO_PARTIAL);
5028 SLAB_ATTR_RO(partial);
5030 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5032 return show_slab_objects(s, buf, SO_CPU);
5034 SLAB_ATTR_RO(cpu_slabs);
5036 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5038 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5040 SLAB_ATTR_RO(objects);
5042 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5044 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5046 SLAB_ATTR_RO(objects_partial);
5048 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5055 for_each_online_cpu(cpu) {
5058 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5061 pages += page->pages;
5062 objects += page->pobjects;
5066 len = sprintf(buf, "%d(%d)", objects, pages);
5069 for_each_online_cpu(cpu) {
5072 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5074 if (page && len < PAGE_SIZE - 20)
5075 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5076 page->pobjects, page->pages);
5079 return len + sprintf(buf + len, "\n");
5081 SLAB_ATTR_RO(slabs_cpu_partial);
5083 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5085 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5088 static ssize_t reclaim_account_store(struct kmem_cache *s,
5089 const char *buf, size_t length)
5091 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5093 s->flags |= SLAB_RECLAIM_ACCOUNT;
5096 SLAB_ATTR(reclaim_account);
5098 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5100 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5102 SLAB_ATTR_RO(hwcache_align);
5104 #ifdef CONFIG_ZONE_DMA
5105 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5107 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5109 SLAB_ATTR_RO(cache_dma);
5112 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5114 return sprintf(buf, "%u\n", s->usersize);
5116 SLAB_ATTR_RO(usersize);
5118 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5122 SLAB_ATTR_RO(destroy_by_rcu);
5124 #ifdef CONFIG_SLUB_DEBUG
5125 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5127 return show_slab_objects(s, buf, SO_ALL);
5129 SLAB_ATTR_RO(slabs);
5131 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5133 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5135 SLAB_ATTR_RO(total_objects);
5137 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5139 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5142 static ssize_t sanity_checks_store(struct kmem_cache *s,
5143 const char *buf, size_t length)
5145 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5146 if (buf[0] == '1') {
5147 s->flags &= ~__CMPXCHG_DOUBLE;
5148 s->flags |= SLAB_CONSISTENCY_CHECKS;
5152 SLAB_ATTR(sanity_checks);
5154 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5156 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5159 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5163 * Tracing a merged cache is going to give confusing results
5164 * as well as cause other issues like converting a mergeable
5165 * cache into an umergeable one.
5167 if (s->refcount > 1)
5170 s->flags &= ~SLAB_TRACE;
5171 if (buf[0] == '1') {
5172 s->flags &= ~__CMPXCHG_DOUBLE;
5173 s->flags |= SLAB_TRACE;
5179 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5181 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5184 static ssize_t red_zone_store(struct kmem_cache *s,
5185 const char *buf, size_t length)
5187 if (any_slab_objects(s))
5190 s->flags &= ~SLAB_RED_ZONE;
5191 if (buf[0] == '1') {
5192 s->flags |= SLAB_RED_ZONE;
5194 calculate_sizes(s, -1);
5197 SLAB_ATTR(red_zone);
5199 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5201 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5204 static ssize_t poison_store(struct kmem_cache *s,
5205 const char *buf, size_t length)
5207 if (any_slab_objects(s))
5210 s->flags &= ~SLAB_POISON;
5211 if (buf[0] == '1') {
5212 s->flags |= SLAB_POISON;
5214 calculate_sizes(s, -1);
5219 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5221 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5224 static ssize_t store_user_store(struct kmem_cache *s,
5225 const char *buf, size_t length)
5227 if (any_slab_objects(s))
5230 s->flags &= ~SLAB_STORE_USER;
5231 if (buf[0] == '1') {
5232 s->flags &= ~__CMPXCHG_DOUBLE;
5233 s->flags |= SLAB_STORE_USER;
5235 calculate_sizes(s, -1);
5238 SLAB_ATTR(store_user);
5240 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5245 static ssize_t validate_store(struct kmem_cache *s,
5246 const char *buf, size_t length)
5250 if (buf[0] == '1') {
5251 ret = validate_slab_cache(s);
5257 SLAB_ATTR(validate);
5259 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5261 if (!(s->flags & SLAB_STORE_USER))
5263 return list_locations(s, buf, TRACK_ALLOC);
5265 SLAB_ATTR_RO(alloc_calls);
5267 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5269 if (!(s->flags & SLAB_STORE_USER))
5271 return list_locations(s, buf, TRACK_FREE);
5273 SLAB_ATTR_RO(free_calls);
5274 #endif /* CONFIG_SLUB_DEBUG */
5276 #ifdef CONFIG_FAILSLAB
5277 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5279 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5282 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5285 if (s->refcount > 1)
5288 s->flags &= ~SLAB_FAILSLAB;
5290 s->flags |= SLAB_FAILSLAB;
5293 SLAB_ATTR(failslab);
5296 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5301 static ssize_t shrink_store(struct kmem_cache *s,
5302 const char *buf, size_t length)
5305 kmem_cache_shrink(s);
5313 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5315 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5318 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5319 const char *buf, size_t length)
5324 err = kstrtouint(buf, 10, &ratio);
5330 s->remote_node_defrag_ratio = ratio * 10;
5334 SLAB_ATTR(remote_node_defrag_ratio);
5337 #ifdef CONFIG_SLUB_STATS
5338 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5340 unsigned long sum = 0;
5343 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5348 for_each_online_cpu(cpu) {
5349 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5355 len = sprintf(buf, "%lu", sum);
5358 for_each_online_cpu(cpu) {
5359 if (data[cpu] && len < PAGE_SIZE - 20)
5360 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5364 return len + sprintf(buf + len, "\n");
5367 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5371 for_each_online_cpu(cpu)
5372 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5375 #define STAT_ATTR(si, text) \
5376 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5378 return show_stat(s, buf, si); \
5380 static ssize_t text##_store(struct kmem_cache *s, \
5381 const char *buf, size_t length) \
5383 if (buf[0] != '0') \
5385 clear_stat(s, si); \
5390 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5391 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5392 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5393 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5394 STAT_ATTR(FREE_FROZEN, free_frozen);
5395 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5396 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5397 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5398 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5399 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5400 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5401 STAT_ATTR(FREE_SLAB, free_slab);
5402 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5403 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5404 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5405 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5406 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5407 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5408 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5409 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5410 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5411 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5412 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5413 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5414 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5415 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5418 static struct attribute *slab_attrs[] = {
5419 &slab_size_attr.attr,
5420 &object_size_attr.attr,
5421 &objs_per_slab_attr.attr,
5423 &min_partial_attr.attr,
5424 &cpu_partial_attr.attr,
5426 &objects_partial_attr.attr,
5428 &cpu_slabs_attr.attr,
5432 &hwcache_align_attr.attr,
5433 &reclaim_account_attr.attr,
5434 &destroy_by_rcu_attr.attr,
5436 &slabs_cpu_partial_attr.attr,
5437 #ifdef CONFIG_SLUB_DEBUG
5438 &total_objects_attr.attr,
5440 &sanity_checks_attr.attr,
5442 &red_zone_attr.attr,
5444 &store_user_attr.attr,
5445 &validate_attr.attr,
5446 &alloc_calls_attr.attr,
5447 &free_calls_attr.attr,
5449 #ifdef CONFIG_ZONE_DMA
5450 &cache_dma_attr.attr,
5453 &remote_node_defrag_ratio_attr.attr,
5455 #ifdef CONFIG_SLUB_STATS
5456 &alloc_fastpath_attr.attr,
5457 &alloc_slowpath_attr.attr,
5458 &free_fastpath_attr.attr,
5459 &free_slowpath_attr.attr,
5460 &free_frozen_attr.attr,
5461 &free_add_partial_attr.attr,
5462 &free_remove_partial_attr.attr,
5463 &alloc_from_partial_attr.attr,
5464 &alloc_slab_attr.attr,
5465 &alloc_refill_attr.attr,
5466 &alloc_node_mismatch_attr.attr,
5467 &free_slab_attr.attr,
5468 &cpuslab_flush_attr.attr,
5469 &deactivate_full_attr.attr,
5470 &deactivate_empty_attr.attr,
5471 &deactivate_to_head_attr.attr,
5472 &deactivate_to_tail_attr.attr,
5473 &deactivate_remote_frees_attr.attr,
5474 &deactivate_bypass_attr.attr,
5475 &order_fallback_attr.attr,
5476 &cmpxchg_double_fail_attr.attr,
5477 &cmpxchg_double_cpu_fail_attr.attr,
5478 &cpu_partial_alloc_attr.attr,
5479 &cpu_partial_free_attr.attr,
5480 &cpu_partial_node_attr.attr,
5481 &cpu_partial_drain_attr.attr,
5483 #ifdef CONFIG_FAILSLAB
5484 &failslab_attr.attr,
5486 &usersize_attr.attr,
5491 static const struct attribute_group slab_attr_group = {
5492 .attrs = slab_attrs,
5495 static ssize_t slab_attr_show(struct kobject *kobj,
5496 struct attribute *attr,
5499 struct slab_attribute *attribute;
5500 struct kmem_cache *s;
5503 attribute = to_slab_attr(attr);
5506 if (!attribute->show)
5509 err = attribute->show(s, buf);
5514 static ssize_t slab_attr_store(struct kobject *kobj,
5515 struct attribute *attr,
5516 const char *buf, size_t len)
5518 struct slab_attribute *attribute;
5519 struct kmem_cache *s;
5522 attribute = to_slab_attr(attr);
5525 if (!attribute->store)
5528 err = attribute->store(s, buf, len);
5530 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5531 struct kmem_cache *c;
5533 mutex_lock(&slab_mutex);
5534 if (s->max_attr_size < len)
5535 s->max_attr_size = len;
5538 * This is a best effort propagation, so this function's return
5539 * value will be determined by the parent cache only. This is
5540 * basically because not all attributes will have a well
5541 * defined semantics for rollbacks - most of the actions will
5542 * have permanent effects.
5544 * Returning the error value of any of the children that fail
5545 * is not 100 % defined, in the sense that users seeing the
5546 * error code won't be able to know anything about the state of
5549 * Only returning the error code for the parent cache at least
5550 * has well defined semantics. The cache being written to
5551 * directly either failed or succeeded, in which case we loop
5552 * through the descendants with best-effort propagation.
5554 for_each_memcg_cache(c, s)
5555 attribute->store(c, buf, len);
5556 mutex_unlock(&slab_mutex);
5562 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5566 char *buffer = NULL;
5567 struct kmem_cache *root_cache;
5569 if (is_root_cache(s))
5572 root_cache = s->memcg_params.root_cache;
5575 * This mean this cache had no attribute written. Therefore, no point
5576 * in copying default values around
5578 if (!root_cache->max_attr_size)
5581 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5584 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5587 if (!attr || !attr->store || !attr->show)
5591 * It is really bad that we have to allocate here, so we will
5592 * do it only as a fallback. If we actually allocate, though,
5593 * we can just use the allocated buffer until the end.
5595 * Most of the slub attributes will tend to be very small in
5596 * size, but sysfs allows buffers up to a page, so they can
5597 * theoretically happen.
5601 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5604 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5605 if (WARN_ON(!buffer))
5610 len = attr->show(root_cache, buf);
5612 attr->store(s, buf, len);
5616 free_page((unsigned long)buffer);
5620 static void kmem_cache_release(struct kobject *k)
5622 slab_kmem_cache_release(to_slab(k));
5625 static const struct sysfs_ops slab_sysfs_ops = {
5626 .show = slab_attr_show,
5627 .store = slab_attr_store,
5630 static struct kobj_type slab_ktype = {
5631 .sysfs_ops = &slab_sysfs_ops,
5632 .release = kmem_cache_release,
5635 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5637 struct kobj_type *ktype = get_ktype(kobj);
5639 if (ktype == &slab_ktype)
5644 static const struct kset_uevent_ops slab_uevent_ops = {
5645 .filter = uevent_filter,
5648 static struct kset *slab_kset;
5650 static inline struct kset *cache_kset(struct kmem_cache *s)
5653 if (!is_root_cache(s))
5654 return s->memcg_params.root_cache->memcg_kset;
5659 #define ID_STR_LENGTH 64
5661 /* Create a unique string id for a slab cache:
5663 * Format :[flags-]size
5665 static char *create_unique_id(struct kmem_cache *s)
5667 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5674 * First flags affecting slabcache operations. We will only
5675 * get here for aliasable slabs so we do not need to support
5676 * too many flags. The flags here must cover all flags that
5677 * are matched during merging to guarantee that the id is
5680 if (s->flags & SLAB_CACHE_DMA)
5682 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5684 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5686 if (s->flags & SLAB_ACCOUNT)
5690 p += sprintf(p, "%07u", s->size);
5692 BUG_ON(p > name + ID_STR_LENGTH - 1);
5696 static void sysfs_slab_remove_workfn(struct work_struct *work)
5698 struct kmem_cache *s =
5699 container_of(work, struct kmem_cache, kobj_remove_work);
5701 if (!s->kobj.state_in_sysfs)
5703 * For a memcg cache, this may be called during
5704 * deactivation and again on shutdown. Remove only once.
5705 * A cache is never shut down before deactivation is
5706 * complete, so no need to worry about synchronization.
5711 kset_unregister(s->memcg_kset);
5713 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5715 kobject_put(&s->kobj);
5718 static int sysfs_slab_add(struct kmem_cache *s)
5722 struct kset *kset = cache_kset(s);
5723 int unmergeable = slab_unmergeable(s);
5725 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5728 kobject_init(&s->kobj, &slab_ktype);
5732 if (!unmergeable && disable_higher_order_debug &&
5733 (slub_debug & DEBUG_METADATA_FLAGS))
5738 * Slabcache can never be merged so we can use the name proper.
5739 * This is typically the case for debug situations. In that
5740 * case we can catch duplicate names easily.
5742 sysfs_remove_link(&slab_kset->kobj, s->name);
5746 * Create a unique name for the slab as a target
5749 name = create_unique_id(s);
5752 s->kobj.kset = kset;
5753 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5757 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5762 if (is_root_cache(s) && memcg_sysfs_enabled) {
5763 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5764 if (!s->memcg_kset) {
5771 kobject_uevent(&s->kobj, KOBJ_ADD);
5773 /* Setup first alias */
5774 sysfs_slab_alias(s, s->name);
5781 kobject_del(&s->kobj);
5785 static void sysfs_slab_remove(struct kmem_cache *s)
5787 if (slab_state < FULL)
5789 * Sysfs has not been setup yet so no need to remove the
5794 kobject_get(&s->kobj);
5795 schedule_work(&s->kobj_remove_work);
5798 void sysfs_slab_unlink(struct kmem_cache *s)
5800 if (slab_state >= FULL)
5801 kobject_del(&s->kobj);
5804 void sysfs_slab_release(struct kmem_cache *s)
5806 if (slab_state >= FULL)
5807 kobject_put(&s->kobj);
5811 * Need to buffer aliases during bootup until sysfs becomes
5812 * available lest we lose that information.
5814 struct saved_alias {
5815 struct kmem_cache *s;
5817 struct saved_alias *next;
5820 static struct saved_alias *alias_list;
5822 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5824 struct saved_alias *al;
5826 if (slab_state == FULL) {
5828 * If we have a leftover link then remove it.
5830 sysfs_remove_link(&slab_kset->kobj, name);
5831 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5834 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5840 al->next = alias_list;
5845 static int __init slab_sysfs_init(void)
5847 struct kmem_cache *s;
5850 mutex_lock(&slab_mutex);
5852 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5854 mutex_unlock(&slab_mutex);
5855 pr_err("Cannot register slab subsystem.\n");
5861 list_for_each_entry(s, &slab_caches, list) {
5862 err = sysfs_slab_add(s);
5864 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5868 while (alias_list) {
5869 struct saved_alias *al = alias_list;
5871 alias_list = alias_list->next;
5872 err = sysfs_slab_alias(al->s, al->name);
5874 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5879 mutex_unlock(&slab_mutex);
5884 __initcall(slab_sysfs_init);
5885 #endif /* CONFIG_SYSFS */
5888 * The /proc/slabinfo ABI
5890 #ifdef CONFIG_SLUB_DEBUG
5891 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5893 unsigned long nr_slabs = 0;
5894 unsigned long nr_objs = 0;
5895 unsigned long nr_free = 0;
5897 struct kmem_cache_node *n;
5899 for_each_kmem_cache_node(s, node, n) {
5900 nr_slabs += node_nr_slabs(n);
5901 nr_objs += node_nr_objs(n);
5902 nr_free += count_partial(n, count_free);
5905 sinfo->active_objs = nr_objs - nr_free;
5906 sinfo->num_objs = nr_objs;
5907 sinfo->active_slabs = nr_slabs;
5908 sinfo->num_slabs = nr_slabs;
5909 sinfo->objects_per_slab = oo_objects(s->oo);
5910 sinfo->cache_order = oo_order(s->oo);
5913 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5917 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5918 size_t count, loff_t *ppos)
5922 #endif /* CONFIG_SLUB_DEBUG */