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,
1097 void *object, unsigned long addr)
1099 if (!check_slab(s, page))
1102 if (!check_valid_pointer(s, page, object)) {
1103 object_err(s, page, object, "Freelist Pointer check fails");
1107 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1113 static noinline int alloc_debug_processing(struct kmem_cache *s,
1115 void *object, unsigned long addr)
1117 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1118 if (!alloc_consistency_checks(s, page, object, addr))
1122 /* Success perform special debug activities for allocs */
1123 if (s->flags & SLAB_STORE_USER)
1124 set_track(s, object, TRACK_ALLOC, addr);
1125 trace(s, page, object, 1);
1126 init_object(s, object, SLUB_RED_ACTIVE);
1130 if (PageSlab(page)) {
1132 * If this is a slab page then lets do the best we can
1133 * to avoid issues in the future. Marking all objects
1134 * as used avoids touching the remaining objects.
1136 slab_fix(s, "Marking all objects used");
1137 page->inuse = page->objects;
1138 page->freelist = NULL;
1143 static inline int free_consistency_checks(struct kmem_cache *s,
1144 struct page *page, void *object, unsigned long addr)
1146 if (!check_valid_pointer(s, page, object)) {
1147 slab_err(s, page, "Invalid object pointer 0x%p", object);
1151 if (on_freelist(s, page, object)) {
1152 object_err(s, page, object, "Object already free");
1156 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1159 if (unlikely(s != page->slab_cache)) {
1160 if (!PageSlab(page)) {
1161 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1163 } else if (!page->slab_cache) {
1164 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1168 object_err(s, page, object,
1169 "page slab pointer corrupt.");
1175 /* Supports checking bulk free of a constructed freelist */
1176 static noinline int free_debug_processing(
1177 struct kmem_cache *s, struct page *page,
1178 void *head, void *tail, int bulk_cnt,
1181 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1182 void *object = head;
1184 unsigned long uninitialized_var(flags);
1187 spin_lock_irqsave(&n->list_lock, flags);
1190 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1191 if (!check_slab(s, page))
1198 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1199 if (!free_consistency_checks(s, page, object, addr))
1203 if (s->flags & SLAB_STORE_USER)
1204 set_track(s, object, TRACK_FREE, addr);
1205 trace(s, page, object, 0);
1206 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1207 init_object(s, object, SLUB_RED_INACTIVE);
1209 /* Reached end of constructed freelist yet? */
1210 if (object != tail) {
1211 object = get_freepointer(s, object);
1217 if (cnt != bulk_cnt)
1218 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1222 spin_unlock_irqrestore(&n->list_lock, flags);
1224 slab_fix(s, "Object at 0x%p not freed", object);
1228 static int __init setup_slub_debug(char *str)
1230 slub_debug = DEBUG_DEFAULT_FLAGS;
1231 if (*str++ != '=' || !*str)
1233 * No options specified. Switch on full debugging.
1239 * No options but restriction on slabs. This means full
1240 * debugging for slabs matching a pattern.
1247 * Switch off all debugging measures.
1252 * Determine which debug features should be switched on
1254 for (; *str && *str != ','; str++) {
1255 switch (tolower(*str)) {
1257 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1260 slub_debug |= SLAB_RED_ZONE;
1263 slub_debug |= SLAB_POISON;
1266 slub_debug |= SLAB_STORE_USER;
1269 slub_debug |= SLAB_TRACE;
1272 slub_debug |= SLAB_FAILSLAB;
1276 * Avoid enabling debugging on caches if its minimum
1277 * order would increase as a result.
1279 disable_higher_order_debug = 1;
1282 pr_err("slub_debug option '%c' unknown. skipped\n",
1289 slub_debug_slabs = str + 1;
1294 __setup("slub_debug", setup_slub_debug);
1297 * kmem_cache_flags - apply debugging options to the cache
1298 * @object_size: the size of an object without meta data
1299 * @flags: flags to set
1300 * @name: name of the cache
1301 * @ctor: constructor function
1303 * Debug option(s) are applied to @flags. In addition to the debug
1304 * option(s), if a slab name (or multiple) is specified i.e.
1305 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1306 * then only the select slabs will receive the debug option(s).
1308 slab_flags_t kmem_cache_flags(unsigned int object_size,
1309 slab_flags_t flags, const char *name,
1310 void (*ctor)(void *))
1315 /* If slub_debug = 0, it folds into the if conditional. */
1316 if (!slub_debug_slabs)
1317 return flags | slub_debug;
1320 iter = slub_debug_slabs;
1325 end = strchr(iter, ',');
1327 end = iter + strlen(iter);
1329 glob = strnchr(iter, end - iter, '*');
1331 cmplen = glob - iter;
1333 cmplen = max_t(size_t, len, (end - iter));
1335 if (!strncmp(name, iter, cmplen)) {
1336 flags |= slub_debug;
1347 #else /* !CONFIG_SLUB_DEBUG */
1348 static inline void setup_object_debug(struct kmem_cache *s,
1349 struct page *page, void *object) {}
1350 static inline void setup_page_debug(struct kmem_cache *s,
1351 void *addr, int order) {}
1353 static inline int alloc_debug_processing(struct kmem_cache *s,
1354 struct page *page, void *object, unsigned long addr) { return 0; }
1356 static inline int free_debug_processing(
1357 struct kmem_cache *s, struct page *page,
1358 void *head, void *tail, int bulk_cnt,
1359 unsigned long addr) { return 0; }
1361 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1363 static inline int check_object(struct kmem_cache *s, struct page *page,
1364 void *object, u8 val) { return 1; }
1365 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1366 struct page *page) {}
1367 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1368 struct page *page) {}
1369 slab_flags_t kmem_cache_flags(unsigned int object_size,
1370 slab_flags_t flags, const char *name,
1371 void (*ctor)(void *))
1375 #define slub_debug 0
1377 #define disable_higher_order_debug 0
1379 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1381 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1383 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1385 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1388 #endif /* CONFIG_SLUB_DEBUG */
1391 * Hooks for other subsystems that check memory allocations. In a typical
1392 * production configuration these hooks all should produce no code at all.
1394 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1396 ptr = kasan_kmalloc_large(ptr, size, flags);
1397 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1398 kmemleak_alloc(ptr, size, 1, flags);
1402 static __always_inline void kfree_hook(void *x)
1405 kasan_kfree_large(x, _RET_IP_);
1408 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1410 kmemleak_free_recursive(x, s->flags);
1413 * Trouble is that we may no longer disable interrupts in the fast path
1414 * So in order to make the debug calls that expect irqs to be
1415 * disabled we need to disable interrupts temporarily.
1417 #ifdef CONFIG_LOCKDEP
1419 unsigned long flags;
1421 local_irq_save(flags);
1422 debug_check_no_locks_freed(x, s->object_size);
1423 local_irq_restore(flags);
1426 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1427 debug_check_no_obj_freed(x, s->object_size);
1429 /* KASAN might put x into memory quarantine, delaying its reuse */
1430 return kasan_slab_free(s, x, _RET_IP_);
1433 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1434 void **head, void **tail)
1437 * Compiler cannot detect this function can be removed if slab_free_hook()
1438 * evaluates to nothing. Thus, catch all relevant config debug options here.
1440 #if defined(CONFIG_LOCKDEP) || \
1441 defined(CONFIG_DEBUG_KMEMLEAK) || \
1442 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1443 defined(CONFIG_KASAN)
1447 void *old_tail = *tail ? *tail : *head;
1449 /* Head and tail of the reconstructed freelist */
1455 next = get_freepointer(s, object);
1456 /* If object's reuse doesn't have to be delayed */
1457 if (!slab_free_hook(s, object)) {
1458 /* Move object to the new freelist */
1459 set_freepointer(s, object, *head);
1464 } while (object != old_tail);
1469 return *head != NULL;
1475 static void *setup_object(struct kmem_cache *s, struct page *page,
1478 setup_object_debug(s, page, object);
1479 object = kasan_init_slab_obj(s, object);
1480 if (unlikely(s->ctor)) {
1481 kasan_unpoison_object_data(s, object);
1483 kasan_poison_object_data(s, object);
1489 * Slab allocation and freeing
1491 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1492 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1495 unsigned int order = oo_order(oo);
1497 if (node == NUMA_NO_NODE)
1498 page = alloc_pages(flags, order);
1500 page = __alloc_pages_node(node, flags, order);
1502 if (page && memcg_charge_slab(page, flags, order, s)) {
1503 __free_pages(page, order);
1510 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1511 /* Pre-initialize the random sequence cache */
1512 static int init_cache_random_seq(struct kmem_cache *s)
1514 unsigned int count = oo_objects(s->oo);
1517 /* Bailout if already initialised */
1521 err = cache_random_seq_create(s, count, GFP_KERNEL);
1523 pr_err("SLUB: Unable to initialize free list for %s\n",
1528 /* Transform to an offset on the set of pages */
1529 if (s->random_seq) {
1532 for (i = 0; i < count; i++)
1533 s->random_seq[i] *= s->size;
1538 /* Initialize each random sequence freelist per cache */
1539 static void __init init_freelist_randomization(void)
1541 struct kmem_cache *s;
1543 mutex_lock(&slab_mutex);
1545 list_for_each_entry(s, &slab_caches, list)
1546 init_cache_random_seq(s);
1548 mutex_unlock(&slab_mutex);
1551 /* Get the next entry on the pre-computed freelist randomized */
1552 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1553 unsigned long *pos, void *start,
1554 unsigned long page_limit,
1555 unsigned long freelist_count)
1560 * If the target page allocation failed, the number of objects on the
1561 * page might be smaller than the usual size defined by the cache.
1564 idx = s->random_seq[*pos];
1566 if (*pos >= freelist_count)
1568 } while (unlikely(idx >= page_limit));
1570 return (char *)start + idx;
1573 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1574 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1579 unsigned long idx, pos, page_limit, freelist_count;
1581 if (page->objects < 2 || !s->random_seq)
1584 freelist_count = oo_objects(s->oo);
1585 pos = get_random_int() % freelist_count;
1587 page_limit = page->objects * s->size;
1588 start = fixup_red_left(s, page_address(page));
1590 /* First entry is used as the base of the freelist */
1591 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1593 cur = setup_object(s, page, cur);
1594 page->freelist = cur;
1596 for (idx = 1; idx < page->objects; idx++) {
1597 next = next_freelist_entry(s, page, &pos, start, page_limit,
1599 next = setup_object(s, page, next);
1600 set_freepointer(s, cur, next);
1603 set_freepointer(s, cur, NULL);
1608 static inline int init_cache_random_seq(struct kmem_cache *s)
1612 static inline void init_freelist_randomization(void) { }
1613 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1617 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1619 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1622 struct kmem_cache_order_objects oo = s->oo;
1624 void *start, *p, *next;
1628 flags &= gfp_allowed_mask;
1630 if (gfpflags_allow_blocking(flags))
1633 flags |= s->allocflags;
1636 * Let the initial higher-order allocation fail under memory pressure
1637 * so we fall-back to the minimum order allocation.
1639 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1640 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1641 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1643 page = alloc_slab_page(s, alloc_gfp, node, oo);
1644 if (unlikely(!page)) {
1648 * Allocation may have failed due to fragmentation.
1649 * Try a lower order alloc if possible
1651 page = alloc_slab_page(s, alloc_gfp, node, oo);
1652 if (unlikely(!page))
1654 stat(s, ORDER_FALLBACK);
1657 page->objects = oo_objects(oo);
1659 order = compound_order(page);
1660 page->slab_cache = s;
1661 __SetPageSlab(page);
1662 if (page_is_pfmemalloc(page))
1663 SetPageSlabPfmemalloc(page);
1665 kasan_poison_slab(page);
1667 start = page_address(page);
1669 setup_page_debug(s, start, order);
1671 shuffle = shuffle_freelist(s, page);
1674 start = fixup_red_left(s, start);
1675 start = setup_object(s, page, start);
1676 page->freelist = start;
1677 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1679 next = setup_object(s, page, next);
1680 set_freepointer(s, p, next);
1683 set_freepointer(s, p, NULL);
1686 page->inuse = page->objects;
1690 if (gfpflags_allow_blocking(flags))
1691 local_irq_disable();
1695 mod_lruvec_page_state(page,
1696 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1697 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1700 inc_slabs_node(s, page_to_nid(page), page->objects);
1705 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1707 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1708 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1709 flags &= ~GFP_SLAB_BUG_MASK;
1710 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1711 invalid_mask, &invalid_mask, flags, &flags);
1715 return allocate_slab(s,
1716 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1719 static void __free_slab(struct kmem_cache *s, struct page *page)
1721 int order = compound_order(page);
1722 int pages = 1 << order;
1724 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1727 slab_pad_check(s, page);
1728 for_each_object(p, s, page_address(page),
1730 check_object(s, page, p, SLUB_RED_INACTIVE);
1733 mod_lruvec_page_state(page,
1734 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1735 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1738 __ClearPageSlabPfmemalloc(page);
1739 __ClearPageSlab(page);
1741 page->mapping = NULL;
1742 if (current->reclaim_state)
1743 current->reclaim_state->reclaimed_slab += pages;
1744 memcg_uncharge_slab(page, order, s);
1745 __free_pages(page, order);
1748 static void rcu_free_slab(struct rcu_head *h)
1750 struct page *page = container_of(h, struct page, rcu_head);
1752 __free_slab(page->slab_cache, page);
1755 static void free_slab(struct kmem_cache *s, struct page *page)
1757 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1758 call_rcu(&page->rcu_head, rcu_free_slab);
1760 __free_slab(s, page);
1763 static void discard_slab(struct kmem_cache *s, struct page *page)
1765 dec_slabs_node(s, page_to_nid(page), page->objects);
1770 * Management of partially allocated slabs.
1773 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1776 if (tail == DEACTIVATE_TO_TAIL)
1777 list_add_tail(&page->lru, &n->partial);
1779 list_add(&page->lru, &n->partial);
1782 static inline void add_partial(struct kmem_cache_node *n,
1783 struct page *page, int tail)
1785 lockdep_assert_held(&n->list_lock);
1786 __add_partial(n, page, tail);
1789 static inline void remove_partial(struct kmem_cache_node *n,
1792 lockdep_assert_held(&n->list_lock);
1793 list_del(&page->lru);
1798 * Remove slab from the partial list, freeze it and
1799 * return the pointer to the freelist.
1801 * Returns a list of objects or NULL if it fails.
1803 static inline void *acquire_slab(struct kmem_cache *s,
1804 struct kmem_cache_node *n, struct page *page,
1805 int mode, int *objects)
1808 unsigned long counters;
1811 lockdep_assert_held(&n->list_lock);
1814 * Zap the freelist and set the frozen bit.
1815 * The old freelist is the list of objects for the
1816 * per cpu allocation list.
1818 freelist = page->freelist;
1819 counters = page->counters;
1820 new.counters = counters;
1821 *objects = new.objects - new.inuse;
1823 new.inuse = page->objects;
1824 new.freelist = NULL;
1826 new.freelist = freelist;
1829 VM_BUG_ON(new.frozen);
1832 if (!__cmpxchg_double_slab(s, page,
1834 new.freelist, new.counters,
1838 remove_partial(n, page);
1843 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1844 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1847 * Try to allocate a partial slab from a specific node.
1849 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1850 struct kmem_cache_cpu *c, gfp_t flags)
1852 struct page *page, *page2;
1853 void *object = NULL;
1854 unsigned int available = 0;
1858 * Racy check. If we mistakenly see no partial slabs then we
1859 * just allocate an empty slab. If we mistakenly try to get a
1860 * partial slab and there is none available then get_partials()
1863 if (!n || !n->nr_partial)
1866 spin_lock(&n->list_lock);
1867 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1870 if (!pfmemalloc_match(page, flags))
1873 t = acquire_slab(s, n, page, object == NULL, &objects);
1877 available += objects;
1880 stat(s, ALLOC_FROM_PARTIAL);
1883 put_cpu_partial(s, page, 0);
1884 stat(s, CPU_PARTIAL_NODE);
1886 if (!kmem_cache_has_cpu_partial(s)
1887 || available > slub_cpu_partial(s) / 2)
1891 spin_unlock(&n->list_lock);
1896 * Get a page from somewhere. Search in increasing NUMA distances.
1898 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1899 struct kmem_cache_cpu *c)
1902 struct zonelist *zonelist;
1905 enum zone_type high_zoneidx = gfp_zone(flags);
1907 unsigned int cpuset_mems_cookie;
1910 * The defrag ratio allows a configuration of the tradeoffs between
1911 * inter node defragmentation and node local allocations. A lower
1912 * defrag_ratio increases the tendency to do local allocations
1913 * instead of attempting to obtain partial slabs from other nodes.
1915 * If the defrag_ratio is set to 0 then kmalloc() always
1916 * returns node local objects. If the ratio is higher then kmalloc()
1917 * may return off node objects because partial slabs are obtained
1918 * from other nodes and filled up.
1920 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1921 * (which makes defrag_ratio = 1000) then every (well almost)
1922 * allocation will first attempt to defrag slab caches on other nodes.
1923 * This means scanning over all nodes to look for partial slabs which
1924 * may be expensive if we do it every time we are trying to find a slab
1925 * with available objects.
1927 if (!s->remote_node_defrag_ratio ||
1928 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1932 cpuset_mems_cookie = read_mems_allowed_begin();
1933 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1934 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1935 struct kmem_cache_node *n;
1937 n = get_node(s, zone_to_nid(zone));
1939 if (n && cpuset_zone_allowed(zone, flags) &&
1940 n->nr_partial > s->min_partial) {
1941 object = get_partial_node(s, n, c, flags);
1944 * Don't check read_mems_allowed_retry()
1945 * here - if mems_allowed was updated in
1946 * parallel, that was a harmless race
1947 * between allocation and the cpuset
1954 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1960 * Get a partial page, lock it and return it.
1962 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1963 struct kmem_cache_cpu *c)
1966 int searchnode = node;
1968 if (node == NUMA_NO_NODE)
1969 searchnode = numa_mem_id();
1970 else if (!node_present_pages(node))
1971 searchnode = node_to_mem_node(node);
1973 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1974 if (object || node != NUMA_NO_NODE)
1977 return get_any_partial(s, flags, c);
1980 #ifdef CONFIG_PREEMPT
1982 * Calculate the next globally unique transaction for disambiguiation
1983 * during cmpxchg. The transactions start with the cpu number and are then
1984 * incremented by CONFIG_NR_CPUS.
1986 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1989 * No preemption supported therefore also no need to check for
1995 static inline unsigned long next_tid(unsigned long tid)
1997 return tid + TID_STEP;
2000 static inline unsigned int tid_to_cpu(unsigned long tid)
2002 return tid % TID_STEP;
2005 static inline unsigned long tid_to_event(unsigned long tid)
2007 return tid / TID_STEP;
2010 static inline unsigned int init_tid(int cpu)
2015 static inline void note_cmpxchg_failure(const char *n,
2016 const struct kmem_cache *s, unsigned long tid)
2018 #ifdef SLUB_DEBUG_CMPXCHG
2019 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2021 pr_info("%s %s: cmpxchg redo ", n, s->name);
2023 #ifdef CONFIG_PREEMPT
2024 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2025 pr_warn("due to cpu change %d -> %d\n",
2026 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2029 if (tid_to_event(tid) != tid_to_event(actual_tid))
2030 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2031 tid_to_event(tid), tid_to_event(actual_tid));
2033 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2034 actual_tid, tid, next_tid(tid));
2036 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2039 static void init_kmem_cache_cpus(struct kmem_cache *s)
2043 for_each_possible_cpu(cpu)
2044 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2048 * Remove the cpu slab
2050 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2051 void *freelist, struct kmem_cache_cpu *c)
2053 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2054 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2056 enum slab_modes l = M_NONE, m = M_NONE;
2058 int tail = DEACTIVATE_TO_HEAD;
2062 if (page->freelist) {
2063 stat(s, DEACTIVATE_REMOTE_FREES);
2064 tail = DEACTIVATE_TO_TAIL;
2068 * Stage one: Free all available per cpu objects back
2069 * to the page freelist while it is still frozen. Leave the
2072 * There is no need to take the list->lock because the page
2075 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2077 unsigned long counters;
2080 prior = page->freelist;
2081 counters = page->counters;
2082 set_freepointer(s, freelist, prior);
2083 new.counters = counters;
2085 VM_BUG_ON(!new.frozen);
2087 } while (!__cmpxchg_double_slab(s, page,
2089 freelist, new.counters,
2090 "drain percpu freelist"));
2092 freelist = nextfree;
2096 * Stage two: Ensure that the page is unfrozen while the
2097 * list presence reflects the actual number of objects
2100 * We setup the list membership and then perform a cmpxchg
2101 * with the count. If there is a mismatch then the page
2102 * is not unfrozen but the page is on the wrong list.
2104 * Then we restart the process which may have to remove
2105 * the page from the list that we just put it on again
2106 * because the number of objects in the slab may have
2111 old.freelist = page->freelist;
2112 old.counters = page->counters;
2113 VM_BUG_ON(!old.frozen);
2115 /* Determine target state of the slab */
2116 new.counters = old.counters;
2119 set_freepointer(s, freelist, old.freelist);
2120 new.freelist = freelist;
2122 new.freelist = old.freelist;
2126 if (!new.inuse && n->nr_partial >= s->min_partial)
2128 else if (new.freelist) {
2133 * Taking the spinlock removes the possiblity
2134 * that acquire_slab() will see a slab page that
2137 spin_lock(&n->list_lock);
2141 if (kmem_cache_debug(s) && !lock) {
2144 * This also ensures that the scanning of full
2145 * slabs from diagnostic functions will not see
2148 spin_lock(&n->list_lock);
2154 remove_partial(n, page);
2155 else if (l == M_FULL)
2156 remove_full(s, n, page);
2159 add_partial(n, page, tail);
2160 else if (m == M_FULL)
2161 add_full(s, n, page);
2165 if (!__cmpxchg_double_slab(s, page,
2166 old.freelist, old.counters,
2167 new.freelist, new.counters,
2172 spin_unlock(&n->list_lock);
2176 else if (m == M_FULL)
2177 stat(s, DEACTIVATE_FULL);
2178 else if (m == M_FREE) {
2179 stat(s, DEACTIVATE_EMPTY);
2180 discard_slab(s, page);
2189 * Unfreeze all the cpu partial slabs.
2191 * This function must be called with interrupts disabled
2192 * for the cpu using c (or some other guarantee must be there
2193 * to guarantee no concurrent accesses).
2195 static void unfreeze_partials(struct kmem_cache *s,
2196 struct kmem_cache_cpu *c)
2198 #ifdef CONFIG_SLUB_CPU_PARTIAL
2199 struct kmem_cache_node *n = NULL, *n2 = NULL;
2200 struct page *page, *discard_page = NULL;
2202 while ((page = c->partial)) {
2206 c->partial = page->next;
2208 n2 = get_node(s, page_to_nid(page));
2211 spin_unlock(&n->list_lock);
2214 spin_lock(&n->list_lock);
2219 old.freelist = page->freelist;
2220 old.counters = page->counters;
2221 VM_BUG_ON(!old.frozen);
2223 new.counters = old.counters;
2224 new.freelist = old.freelist;
2228 } while (!__cmpxchg_double_slab(s, page,
2229 old.freelist, old.counters,
2230 new.freelist, new.counters,
2231 "unfreezing slab"));
2233 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2234 page->next = discard_page;
2235 discard_page = page;
2237 add_partial(n, page, DEACTIVATE_TO_TAIL);
2238 stat(s, FREE_ADD_PARTIAL);
2243 spin_unlock(&n->list_lock);
2245 while (discard_page) {
2246 page = discard_page;
2247 discard_page = discard_page->next;
2249 stat(s, DEACTIVATE_EMPTY);
2250 discard_slab(s, page);
2257 * Put a page that was just frozen (in __slab_free) into a partial page
2258 * slot if available.
2260 * If we did not find a slot then simply move all the partials to the
2261 * per node partial list.
2263 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2265 #ifdef CONFIG_SLUB_CPU_PARTIAL
2266 struct page *oldpage;
2274 oldpage = this_cpu_read(s->cpu_slab->partial);
2277 pobjects = oldpage->pobjects;
2278 pages = oldpage->pages;
2279 if (drain && pobjects > s->cpu_partial) {
2280 unsigned long flags;
2282 * partial array is full. Move the existing
2283 * set to the per node partial list.
2285 local_irq_save(flags);
2286 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2287 local_irq_restore(flags);
2291 stat(s, CPU_PARTIAL_DRAIN);
2296 pobjects += page->objects - page->inuse;
2298 page->pages = pages;
2299 page->pobjects = pobjects;
2300 page->next = oldpage;
2302 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2304 if (unlikely(!s->cpu_partial)) {
2305 unsigned long flags;
2307 local_irq_save(flags);
2308 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2309 local_irq_restore(flags);
2315 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2317 stat(s, CPUSLAB_FLUSH);
2318 deactivate_slab(s, c->page, c->freelist, c);
2320 c->tid = next_tid(c->tid);
2326 * Called from IPI handler with interrupts disabled.
2328 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2330 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2335 unfreeze_partials(s, c);
2338 static void flush_cpu_slab(void *d)
2340 struct kmem_cache *s = d;
2342 __flush_cpu_slab(s, smp_processor_id());
2345 static bool has_cpu_slab(int cpu, void *info)
2347 struct kmem_cache *s = info;
2348 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2350 return c->page || slub_percpu_partial(c);
2353 static void flush_all(struct kmem_cache *s)
2355 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2359 * Use the cpu notifier to insure that the cpu slabs are flushed when
2362 static int slub_cpu_dead(unsigned int cpu)
2364 struct kmem_cache *s;
2365 unsigned long flags;
2367 mutex_lock(&slab_mutex);
2368 list_for_each_entry(s, &slab_caches, list) {
2369 local_irq_save(flags);
2370 __flush_cpu_slab(s, cpu);
2371 local_irq_restore(flags);
2373 mutex_unlock(&slab_mutex);
2378 * Check if the objects in a per cpu structure fit numa
2379 * locality expectations.
2381 static inline int node_match(struct page *page, int node)
2384 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2390 #ifdef CONFIG_SLUB_DEBUG
2391 static int count_free(struct page *page)
2393 return page->objects - page->inuse;
2396 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2398 return atomic_long_read(&n->total_objects);
2400 #endif /* CONFIG_SLUB_DEBUG */
2402 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2403 static unsigned long count_partial(struct kmem_cache_node *n,
2404 int (*get_count)(struct page *))
2406 unsigned long flags;
2407 unsigned long x = 0;
2410 spin_lock_irqsave(&n->list_lock, flags);
2411 list_for_each_entry(page, &n->partial, lru)
2412 x += get_count(page);
2413 spin_unlock_irqrestore(&n->list_lock, flags);
2416 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2418 static noinline void
2419 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2421 #ifdef CONFIG_SLUB_DEBUG
2422 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2423 DEFAULT_RATELIMIT_BURST);
2425 struct kmem_cache_node *n;
2427 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2430 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2431 nid, gfpflags, &gfpflags);
2432 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2433 s->name, s->object_size, s->size, oo_order(s->oo),
2436 if (oo_order(s->min) > get_order(s->object_size))
2437 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2440 for_each_kmem_cache_node(s, node, n) {
2441 unsigned long nr_slabs;
2442 unsigned long nr_objs;
2443 unsigned long nr_free;
2445 nr_free = count_partial(n, count_free);
2446 nr_slabs = node_nr_slabs(n);
2447 nr_objs = node_nr_objs(n);
2449 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2450 node, nr_slabs, nr_objs, nr_free);
2455 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2456 int node, struct kmem_cache_cpu **pc)
2459 struct kmem_cache_cpu *c = *pc;
2462 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2464 freelist = get_partial(s, flags, node, c);
2469 page = new_slab(s, flags, node);
2471 c = raw_cpu_ptr(s->cpu_slab);
2476 * No other reference to the page yet so we can
2477 * muck around with it freely without cmpxchg
2479 freelist = page->freelist;
2480 page->freelist = NULL;
2482 stat(s, ALLOC_SLAB);
2491 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2493 if (unlikely(PageSlabPfmemalloc(page)))
2494 return gfp_pfmemalloc_allowed(gfpflags);
2500 * Check the page->freelist of a page and either transfer the freelist to the
2501 * per cpu freelist or deactivate the page.
2503 * The page is still frozen if the return value is not NULL.
2505 * If this function returns NULL then the page has been unfrozen.
2507 * This function must be called with interrupt disabled.
2509 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2512 unsigned long counters;
2516 freelist = page->freelist;
2517 counters = page->counters;
2519 new.counters = counters;
2520 VM_BUG_ON(!new.frozen);
2522 new.inuse = page->objects;
2523 new.frozen = freelist != NULL;
2525 } while (!__cmpxchg_double_slab(s, page,
2534 * Slow path. The lockless freelist is empty or we need to perform
2537 * Processing is still very fast if new objects have been freed to the
2538 * regular freelist. In that case we simply take over the regular freelist
2539 * as the lockless freelist and zap the regular freelist.
2541 * If that is not working then we fall back to the partial lists. We take the
2542 * first element of the freelist as the object to allocate now and move the
2543 * rest of the freelist to the lockless freelist.
2545 * And if we were unable to get a new slab from the partial slab lists then
2546 * we need to allocate a new slab. This is the slowest path since it involves
2547 * a call to the page allocator and the setup of a new slab.
2549 * Version of __slab_alloc to use when we know that interrupts are
2550 * already disabled (which is the case for bulk allocation).
2552 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2553 unsigned long addr, struct kmem_cache_cpu *c)
2563 if (unlikely(!node_match(page, node))) {
2564 int searchnode = node;
2566 if (node != NUMA_NO_NODE && !node_present_pages(node))
2567 searchnode = node_to_mem_node(node);
2569 if (unlikely(!node_match(page, searchnode))) {
2570 stat(s, ALLOC_NODE_MISMATCH);
2571 deactivate_slab(s, page, c->freelist, c);
2577 * By rights, we should be searching for a slab page that was
2578 * PFMEMALLOC but right now, we are losing the pfmemalloc
2579 * information when the page leaves the per-cpu allocator
2581 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2582 deactivate_slab(s, page, c->freelist, c);
2586 /* must check again c->freelist in case of cpu migration or IRQ */
2587 freelist = c->freelist;
2591 freelist = get_freelist(s, page);
2595 stat(s, DEACTIVATE_BYPASS);
2599 stat(s, ALLOC_REFILL);
2603 * freelist is pointing to the list of objects to be used.
2604 * page is pointing to the page from which the objects are obtained.
2605 * That page must be frozen for per cpu allocations to work.
2607 VM_BUG_ON(!c->page->frozen);
2608 c->freelist = get_freepointer(s, freelist);
2609 c->tid = next_tid(c->tid);
2614 if (slub_percpu_partial(c)) {
2615 page = c->page = slub_percpu_partial(c);
2616 slub_set_percpu_partial(c, page);
2617 stat(s, CPU_PARTIAL_ALLOC);
2621 freelist = new_slab_objects(s, gfpflags, node, &c);
2623 if (unlikely(!freelist)) {
2624 slab_out_of_memory(s, gfpflags, node);
2629 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2632 /* Only entered in the debug case */
2633 if (kmem_cache_debug(s) &&
2634 !alloc_debug_processing(s, page, freelist, addr))
2635 goto new_slab; /* Slab failed checks. Next slab needed */
2637 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2642 * Another one that disabled interrupt and compensates for possible
2643 * cpu changes by refetching the per cpu area pointer.
2645 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2646 unsigned long addr, struct kmem_cache_cpu *c)
2649 unsigned long flags;
2651 local_irq_save(flags);
2652 #ifdef CONFIG_PREEMPT
2654 * We may have been preempted and rescheduled on a different
2655 * cpu before disabling interrupts. Need to reload cpu area
2658 c = this_cpu_ptr(s->cpu_slab);
2661 p = ___slab_alloc(s, gfpflags, node, addr, c);
2662 local_irq_restore(flags);
2667 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2668 * have the fastpath folded into their functions. So no function call
2669 * overhead for requests that can be satisfied on the fastpath.
2671 * The fastpath works by first checking if the lockless freelist can be used.
2672 * If not then __slab_alloc is called for slow processing.
2674 * Otherwise we can simply pick the next object from the lockless free list.
2676 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2677 gfp_t gfpflags, int node, unsigned long addr)
2680 struct kmem_cache_cpu *c;
2684 s = slab_pre_alloc_hook(s, gfpflags);
2689 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2690 * enabled. We may switch back and forth between cpus while
2691 * reading from one cpu area. That does not matter as long
2692 * as we end up on the original cpu again when doing the cmpxchg.
2694 * We should guarantee that tid and kmem_cache are retrieved on
2695 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2696 * to check if it is matched or not.
2699 tid = this_cpu_read(s->cpu_slab->tid);
2700 c = raw_cpu_ptr(s->cpu_slab);
2701 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2702 unlikely(tid != READ_ONCE(c->tid)));
2705 * Irqless object alloc/free algorithm used here depends on sequence
2706 * of fetching cpu_slab's data. tid should be fetched before anything
2707 * on c to guarantee that object and page associated with previous tid
2708 * won't be used with current tid. If we fetch tid first, object and
2709 * page could be one associated with next tid and our alloc/free
2710 * request will be failed. In this case, we will retry. So, no problem.
2715 * The transaction ids are globally unique per cpu and per operation on
2716 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2717 * occurs on the right processor and that there was no operation on the
2718 * linked list in between.
2721 object = c->freelist;
2723 if (unlikely(!object || !node_match(page, node))) {
2724 object = __slab_alloc(s, gfpflags, node, addr, c);
2725 stat(s, ALLOC_SLOWPATH);
2727 void *next_object = get_freepointer_safe(s, object);
2730 * The cmpxchg will only match if there was no additional
2731 * operation and if we are on the right processor.
2733 * The cmpxchg does the following atomically (without lock
2735 * 1. Relocate first pointer to the current per cpu area.
2736 * 2. Verify that tid and freelist have not been changed
2737 * 3. If they were not changed replace tid and freelist
2739 * Since this is without lock semantics the protection is only
2740 * against code executing on this cpu *not* from access by
2743 if (unlikely(!this_cpu_cmpxchg_double(
2744 s->cpu_slab->freelist, s->cpu_slab->tid,
2746 next_object, next_tid(tid)))) {
2748 note_cmpxchg_failure("slab_alloc", s, tid);
2751 prefetch_freepointer(s, next_object);
2752 stat(s, ALLOC_FASTPATH);
2755 if (unlikely(gfpflags & __GFP_ZERO) && object)
2756 memset(object, 0, s->object_size);
2758 slab_post_alloc_hook(s, gfpflags, 1, &object);
2763 static __always_inline void *slab_alloc(struct kmem_cache *s,
2764 gfp_t gfpflags, unsigned long addr)
2766 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2769 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2771 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2773 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2778 EXPORT_SYMBOL(kmem_cache_alloc);
2780 #ifdef CONFIG_TRACING
2781 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2783 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2784 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2785 ret = kasan_kmalloc(s, ret, size, gfpflags);
2788 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2792 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2794 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2796 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2797 s->object_size, s->size, gfpflags, node);
2801 EXPORT_SYMBOL(kmem_cache_alloc_node);
2803 #ifdef CONFIG_TRACING
2804 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2806 int node, size_t size)
2808 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2810 trace_kmalloc_node(_RET_IP_, ret,
2811 size, s->size, gfpflags, node);
2813 ret = kasan_kmalloc(s, ret, size, gfpflags);
2816 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2821 * Slow path handling. This may still be called frequently since objects
2822 * have a longer lifetime than the cpu slabs in most processing loads.
2824 * So we still attempt to reduce cache line usage. Just take the slab
2825 * lock and free the item. If there is no additional partial page
2826 * handling required then we can return immediately.
2828 static void __slab_free(struct kmem_cache *s, struct page *page,
2829 void *head, void *tail, int cnt,
2836 unsigned long counters;
2837 struct kmem_cache_node *n = NULL;
2838 unsigned long uninitialized_var(flags);
2840 stat(s, FREE_SLOWPATH);
2842 if (kmem_cache_debug(s) &&
2843 !free_debug_processing(s, page, head, tail, cnt, addr))
2848 spin_unlock_irqrestore(&n->list_lock, flags);
2851 prior = page->freelist;
2852 counters = page->counters;
2853 set_freepointer(s, tail, prior);
2854 new.counters = counters;
2855 was_frozen = new.frozen;
2857 if ((!new.inuse || !prior) && !was_frozen) {
2859 if (kmem_cache_has_cpu_partial(s) && !prior) {
2862 * Slab was on no list before and will be
2864 * We can defer the list move and instead
2869 } else { /* Needs to be taken off a list */
2871 n = get_node(s, page_to_nid(page));
2873 * Speculatively acquire the list_lock.
2874 * If the cmpxchg does not succeed then we may
2875 * drop the list_lock without any processing.
2877 * Otherwise the list_lock will synchronize with
2878 * other processors updating the list of slabs.
2880 spin_lock_irqsave(&n->list_lock, flags);
2885 } while (!cmpxchg_double_slab(s, page,
2893 * If we just froze the page then put it onto the
2894 * per cpu partial list.
2896 if (new.frozen && !was_frozen) {
2897 put_cpu_partial(s, page, 1);
2898 stat(s, CPU_PARTIAL_FREE);
2901 * The list lock was not taken therefore no list
2902 * activity can be necessary.
2905 stat(s, FREE_FROZEN);
2909 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2913 * Objects left in the slab. If it was not on the partial list before
2916 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2917 if (kmem_cache_debug(s))
2918 remove_full(s, n, page);
2919 add_partial(n, page, DEACTIVATE_TO_TAIL);
2920 stat(s, FREE_ADD_PARTIAL);
2922 spin_unlock_irqrestore(&n->list_lock, flags);
2928 * Slab on the partial list.
2930 remove_partial(n, page);
2931 stat(s, FREE_REMOVE_PARTIAL);
2933 /* Slab must be on the full list */
2934 remove_full(s, n, page);
2937 spin_unlock_irqrestore(&n->list_lock, flags);
2939 discard_slab(s, page);
2943 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2944 * can perform fastpath freeing without additional function calls.
2946 * The fastpath is only possible if we are freeing to the current cpu slab
2947 * of this processor. This typically the case if we have just allocated
2950 * If fastpath is not possible then fall back to __slab_free where we deal
2951 * with all sorts of special processing.
2953 * Bulk free of a freelist with several objects (all pointing to the
2954 * same page) possible by specifying head and tail ptr, plus objects
2955 * count (cnt). Bulk free indicated by tail pointer being set.
2957 static __always_inline void do_slab_free(struct kmem_cache *s,
2958 struct page *page, void *head, void *tail,
2959 int cnt, unsigned long addr)
2961 void *tail_obj = tail ? : head;
2962 struct kmem_cache_cpu *c;
2966 * Determine the currently cpus per cpu slab.
2967 * The cpu may change afterward. However that does not matter since
2968 * data is retrieved via this pointer. If we are on the same cpu
2969 * during the cmpxchg then the free will succeed.
2972 tid = this_cpu_read(s->cpu_slab->tid);
2973 c = raw_cpu_ptr(s->cpu_slab);
2974 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2975 unlikely(tid != READ_ONCE(c->tid)));
2977 /* Same with comment on barrier() in slab_alloc_node() */
2980 if (likely(page == c->page)) {
2981 set_freepointer(s, tail_obj, c->freelist);
2983 if (unlikely(!this_cpu_cmpxchg_double(
2984 s->cpu_slab->freelist, s->cpu_slab->tid,
2986 head, next_tid(tid)))) {
2988 note_cmpxchg_failure("slab_free", s, tid);
2991 stat(s, FREE_FASTPATH);
2993 __slab_free(s, page, head, tail_obj, cnt, addr);
2997 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2998 void *head, void *tail, int cnt,
3002 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3003 * to remove objects, whose reuse must be delayed.
3005 if (slab_free_freelist_hook(s, &head, &tail))
3006 do_slab_free(s, page, head, tail, cnt, addr);
3009 #ifdef CONFIG_KASAN_GENERIC
3010 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3012 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3016 void kmem_cache_free(struct kmem_cache *s, void *x)
3018 s = cache_from_obj(s, x);
3021 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3022 trace_kmem_cache_free(_RET_IP_, x);
3024 EXPORT_SYMBOL(kmem_cache_free);
3026 struct detached_freelist {
3031 struct kmem_cache *s;
3035 * This function progressively scans the array with free objects (with
3036 * a limited look ahead) and extract objects belonging to the same
3037 * page. It builds a detached freelist directly within the given
3038 * page/objects. This can happen without any need for
3039 * synchronization, because the objects are owned by running process.
3040 * The freelist is build up as a single linked list in the objects.
3041 * The idea is, that this detached freelist can then be bulk
3042 * transferred to the real freelist(s), but only requiring a single
3043 * synchronization primitive. Look ahead in the array is limited due
3044 * to performance reasons.
3047 int build_detached_freelist(struct kmem_cache *s, size_t size,
3048 void **p, struct detached_freelist *df)
3050 size_t first_skipped_index = 0;
3055 /* Always re-init detached_freelist */
3060 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3061 } while (!object && size);
3066 page = virt_to_head_page(object);
3068 /* Handle kalloc'ed objects */
3069 if (unlikely(!PageSlab(page))) {
3070 BUG_ON(!PageCompound(page));
3072 __free_pages(page, compound_order(page));
3073 p[size] = NULL; /* mark object processed */
3076 /* Derive kmem_cache from object */
3077 df->s = page->slab_cache;
3079 df->s = cache_from_obj(s, object); /* Support for memcg */
3082 /* Start new detached freelist */
3084 set_freepointer(df->s, object, NULL);
3086 df->freelist = object;
3087 p[size] = NULL; /* mark object processed */
3093 continue; /* Skip processed objects */
3095 /* df->page is always set at this point */
3096 if (df->page == virt_to_head_page(object)) {
3097 /* Opportunity build freelist */
3098 set_freepointer(df->s, object, df->freelist);
3099 df->freelist = object;
3101 p[size] = NULL; /* mark object processed */
3106 /* Limit look ahead search */
3110 if (!first_skipped_index)
3111 first_skipped_index = size + 1;
3114 return first_skipped_index;
3117 /* Note that interrupts must be enabled when calling this function. */
3118 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3124 struct detached_freelist df;
3126 size = build_detached_freelist(s, size, p, &df);
3130 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3131 } while (likely(size));
3133 EXPORT_SYMBOL(kmem_cache_free_bulk);
3135 /* Note that interrupts must be enabled when calling this function. */
3136 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3139 struct kmem_cache_cpu *c;
3142 /* memcg and kmem_cache debug support */
3143 s = slab_pre_alloc_hook(s, flags);
3147 * Drain objects in the per cpu slab, while disabling local
3148 * IRQs, which protects against PREEMPT and interrupts
3149 * handlers invoking normal fastpath.
3151 local_irq_disable();
3152 c = this_cpu_ptr(s->cpu_slab);
3154 for (i = 0; i < size; i++) {
3155 void *object = c->freelist;
3157 if (unlikely(!object)) {
3159 * Invoking slow path likely have side-effect
3160 * of re-populating per CPU c->freelist
3162 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3164 if (unlikely(!p[i]))
3167 c = this_cpu_ptr(s->cpu_slab);
3168 continue; /* goto for-loop */
3170 c->freelist = get_freepointer(s, object);
3173 c->tid = next_tid(c->tid);
3176 /* Clear memory outside IRQ disabled fastpath loop */
3177 if (unlikely(flags & __GFP_ZERO)) {
3180 for (j = 0; j < i; j++)
3181 memset(p[j], 0, s->object_size);
3184 /* memcg and kmem_cache debug support */
3185 slab_post_alloc_hook(s, flags, size, p);
3189 slab_post_alloc_hook(s, flags, i, p);
3190 __kmem_cache_free_bulk(s, i, p);
3193 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3197 * Object placement in a slab is made very easy because we always start at
3198 * offset 0. If we tune the size of the object to the alignment then we can
3199 * get the required alignment by putting one properly sized object after
3202 * Notice that the allocation order determines the sizes of the per cpu
3203 * caches. Each processor has always one slab available for allocations.
3204 * Increasing the allocation order reduces the number of times that slabs
3205 * must be moved on and off the partial lists and is therefore a factor in
3210 * Mininum / Maximum order of slab pages. This influences locking overhead
3211 * and slab fragmentation. A higher order reduces the number of partial slabs
3212 * and increases the number of allocations possible without having to
3213 * take the list_lock.
3215 static unsigned int slub_min_order;
3216 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3217 static unsigned int slub_min_objects;
3220 * Calculate the order of allocation given an slab object size.
3222 * The order of allocation has significant impact on performance and other
3223 * system components. Generally order 0 allocations should be preferred since
3224 * order 0 does not cause fragmentation in the page allocator. Larger objects
3225 * be problematic to put into order 0 slabs because there may be too much
3226 * unused space left. We go to a higher order if more than 1/16th of the slab
3229 * In order to reach satisfactory performance we must ensure that a minimum
3230 * number of objects is in one slab. Otherwise we may generate too much
3231 * activity on the partial lists which requires taking the list_lock. This is
3232 * less a concern for large slabs though which are rarely used.
3234 * slub_max_order specifies the order where we begin to stop considering the
3235 * number of objects in a slab as critical. If we reach slub_max_order then
3236 * we try to keep the page order as low as possible. So we accept more waste
3237 * of space in favor of a small page order.
3239 * Higher order allocations also allow the placement of more objects in a
3240 * slab and thereby reduce object handling overhead. If the user has
3241 * requested a higher mininum order then we start with that one instead of
3242 * the smallest order which will fit the object.
3244 static inline unsigned int slab_order(unsigned int size,
3245 unsigned int min_objects, unsigned int max_order,
3246 unsigned int fract_leftover)
3248 unsigned int min_order = slub_min_order;
3251 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3252 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3254 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3255 order <= max_order; order++) {
3257 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3260 rem = slab_size % size;
3262 if (rem <= slab_size / fract_leftover)
3269 static inline int calculate_order(unsigned int size)
3272 unsigned int min_objects;
3273 unsigned int max_objects;
3276 * Attempt to find best configuration for a slab. This
3277 * works by first attempting to generate a layout with
3278 * the best configuration and backing off gradually.
3280 * First we increase the acceptable waste in a slab. Then
3281 * we reduce the minimum objects required in a slab.
3283 min_objects = slub_min_objects;
3285 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3286 max_objects = order_objects(slub_max_order, size);
3287 min_objects = min(min_objects, max_objects);
3289 while (min_objects > 1) {
3290 unsigned int fraction;
3293 while (fraction >= 4) {
3294 order = slab_order(size, min_objects,
3295 slub_max_order, fraction);
3296 if (order <= slub_max_order)
3304 * We were unable to place multiple objects in a slab. Now
3305 * lets see if we can place a single object there.
3307 order = slab_order(size, 1, slub_max_order, 1);
3308 if (order <= slub_max_order)
3312 * Doh this slab cannot be placed using slub_max_order.
3314 order = slab_order(size, 1, MAX_ORDER, 1);
3315 if (order < MAX_ORDER)
3321 init_kmem_cache_node(struct kmem_cache_node *n)
3324 spin_lock_init(&n->list_lock);
3325 INIT_LIST_HEAD(&n->partial);
3326 #ifdef CONFIG_SLUB_DEBUG
3327 atomic_long_set(&n->nr_slabs, 0);
3328 atomic_long_set(&n->total_objects, 0);
3329 INIT_LIST_HEAD(&n->full);
3333 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3335 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3336 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3339 * Must align to double word boundary for the double cmpxchg
3340 * instructions to work; see __pcpu_double_call_return_bool().
3342 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3343 2 * sizeof(void *));
3348 init_kmem_cache_cpus(s);
3353 static struct kmem_cache *kmem_cache_node;
3356 * No kmalloc_node yet so do it by hand. We know that this is the first
3357 * slab on the node for this slabcache. There are no concurrent accesses
3360 * Note that this function only works on the kmem_cache_node
3361 * when allocating for the kmem_cache_node. This is used for bootstrapping
3362 * memory on a fresh node that has no slab structures yet.
3364 static void early_kmem_cache_node_alloc(int node)
3367 struct kmem_cache_node *n;
3369 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3371 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3374 if (page_to_nid(page) != node) {
3375 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3376 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3381 #ifdef CONFIG_SLUB_DEBUG
3382 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3383 init_tracking(kmem_cache_node, n);
3385 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3387 page->freelist = get_freepointer(kmem_cache_node, n);
3390 kmem_cache_node->node[node] = n;
3391 init_kmem_cache_node(n);
3392 inc_slabs_node(kmem_cache_node, node, page->objects);
3395 * No locks need to be taken here as it has just been
3396 * initialized and there is no concurrent access.
3398 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3401 static void free_kmem_cache_nodes(struct kmem_cache *s)
3404 struct kmem_cache_node *n;
3406 for_each_kmem_cache_node(s, node, n) {
3407 s->node[node] = NULL;
3408 kmem_cache_free(kmem_cache_node, n);
3412 void __kmem_cache_release(struct kmem_cache *s)
3414 cache_random_seq_destroy(s);
3415 free_percpu(s->cpu_slab);
3416 free_kmem_cache_nodes(s);
3419 static int init_kmem_cache_nodes(struct kmem_cache *s)
3423 for_each_node_state(node, N_NORMAL_MEMORY) {
3424 struct kmem_cache_node *n;
3426 if (slab_state == DOWN) {
3427 early_kmem_cache_node_alloc(node);
3430 n = kmem_cache_alloc_node(kmem_cache_node,
3434 free_kmem_cache_nodes(s);
3438 init_kmem_cache_node(n);
3444 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3446 if (min < MIN_PARTIAL)
3448 else if (min > MAX_PARTIAL)
3450 s->min_partial = min;
3453 static void set_cpu_partial(struct kmem_cache *s)
3455 #ifdef CONFIG_SLUB_CPU_PARTIAL
3457 * cpu_partial determined the maximum number of objects kept in the
3458 * per cpu partial lists of a processor.
3460 * Per cpu partial lists mainly contain slabs that just have one
3461 * object freed. If they are used for allocation then they can be
3462 * filled up again with minimal effort. The slab will never hit the
3463 * per node partial lists and therefore no locking will be required.
3465 * This setting also determines
3467 * A) The number of objects from per cpu partial slabs dumped to the
3468 * per node list when we reach the limit.
3469 * B) The number of objects in cpu partial slabs to extract from the
3470 * per node list when we run out of per cpu objects. We only fetch
3471 * 50% to keep some capacity around for frees.
3473 if (!kmem_cache_has_cpu_partial(s))
3475 else if (s->size >= PAGE_SIZE)
3477 else if (s->size >= 1024)
3479 else if (s->size >= 256)
3480 s->cpu_partial = 13;
3482 s->cpu_partial = 30;
3487 * calculate_sizes() determines the order and the distribution of data within
3490 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3492 slab_flags_t flags = s->flags;
3493 unsigned int size = s->object_size;
3497 * Round up object size to the next word boundary. We can only
3498 * place the free pointer at word boundaries and this determines
3499 * the possible location of the free pointer.
3501 size = ALIGN(size, sizeof(void *));
3503 #ifdef CONFIG_SLUB_DEBUG
3505 * Determine if we can poison the object itself. If the user of
3506 * the slab may touch the object after free or before allocation
3507 * then we should never poison the object itself.
3509 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3511 s->flags |= __OBJECT_POISON;
3513 s->flags &= ~__OBJECT_POISON;
3517 * If we are Redzoning then check if there is some space between the
3518 * end of the object and the free pointer. If not then add an
3519 * additional word to have some bytes to store Redzone information.
3521 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3522 size += sizeof(void *);
3526 * With that we have determined the number of bytes in actual use
3527 * by the object. This is the potential offset to the free pointer.
3531 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3534 * Relocate free pointer after the object if it is not
3535 * permitted to overwrite the first word of the object on
3538 * This is the case if we do RCU, have a constructor or
3539 * destructor or are poisoning the objects.
3542 size += sizeof(void *);
3545 #ifdef CONFIG_SLUB_DEBUG
3546 if (flags & SLAB_STORE_USER)
3548 * Need to store information about allocs and frees after
3551 size += 2 * sizeof(struct track);
3554 kasan_cache_create(s, &size, &s->flags);
3555 #ifdef CONFIG_SLUB_DEBUG
3556 if (flags & SLAB_RED_ZONE) {
3558 * Add some empty padding so that we can catch
3559 * overwrites from earlier objects rather than let
3560 * tracking information or the free pointer be
3561 * corrupted if a user writes before the start
3564 size += sizeof(void *);
3566 s->red_left_pad = sizeof(void *);
3567 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3568 size += s->red_left_pad;
3573 * SLUB stores one object immediately after another beginning from
3574 * offset 0. In order to align the objects we have to simply size
3575 * each object to conform to the alignment.
3577 size = ALIGN(size, s->align);
3579 if (forced_order >= 0)
3580 order = forced_order;
3582 order = calculate_order(size);
3589 s->allocflags |= __GFP_COMP;
3591 if (s->flags & SLAB_CACHE_DMA)
3592 s->allocflags |= GFP_DMA;
3594 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3595 s->allocflags |= __GFP_RECLAIMABLE;
3598 * Determine the number of objects per slab
3600 s->oo = oo_make(order, size);
3601 s->min = oo_make(get_order(size), size);
3602 if (oo_objects(s->oo) > oo_objects(s->max))
3605 return !!oo_objects(s->oo);
3608 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3610 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3611 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3612 s->random = get_random_long();
3615 if (!calculate_sizes(s, -1))
3617 if (disable_higher_order_debug) {
3619 * Disable debugging flags that store metadata if the min slab
3622 if (get_order(s->size) > get_order(s->object_size)) {
3623 s->flags &= ~DEBUG_METADATA_FLAGS;
3625 if (!calculate_sizes(s, -1))
3630 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3631 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3632 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3633 /* Enable fast mode */
3634 s->flags |= __CMPXCHG_DOUBLE;
3638 * The larger the object size is, the more pages we want on the partial
3639 * list to avoid pounding the page allocator excessively.
3641 set_min_partial(s, ilog2(s->size) / 2);
3646 s->remote_node_defrag_ratio = 1000;
3649 /* Initialize the pre-computed randomized freelist if slab is up */
3650 if (slab_state >= UP) {
3651 if (init_cache_random_seq(s))
3655 if (!init_kmem_cache_nodes(s))
3658 if (alloc_kmem_cache_cpus(s))
3661 free_kmem_cache_nodes(s);
3663 if (flags & SLAB_PANIC)
3664 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3665 s->name, s->size, s->size,
3666 oo_order(s->oo), s->offset, (unsigned long)flags);
3670 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3673 #ifdef CONFIG_SLUB_DEBUG
3674 void *addr = page_address(page);
3676 unsigned long *map = bitmap_zalloc(page->objects, GFP_ATOMIC);
3679 slab_err(s, page, text, s->name);
3682 get_map(s, page, map);
3683 for_each_object(p, s, addr, page->objects) {
3685 if (!test_bit(slab_index(p, s, addr), map)) {
3686 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3687 print_tracking(s, p);
3696 * Attempt to free all partial slabs on a node.
3697 * This is called from __kmem_cache_shutdown(). We must take list_lock
3698 * because sysfs file might still access partial list after the shutdowning.
3700 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3703 struct page *page, *h;
3705 BUG_ON(irqs_disabled());
3706 spin_lock_irq(&n->list_lock);
3707 list_for_each_entry_safe(page, h, &n->partial, lru) {
3709 remove_partial(n, page);
3710 list_add(&page->lru, &discard);
3712 list_slab_objects(s, page,
3713 "Objects remaining in %s on __kmem_cache_shutdown()");
3716 spin_unlock_irq(&n->list_lock);
3718 list_for_each_entry_safe(page, h, &discard, lru)
3719 discard_slab(s, page);
3722 bool __kmem_cache_empty(struct kmem_cache *s)
3725 struct kmem_cache_node *n;
3727 for_each_kmem_cache_node(s, node, n)
3728 if (n->nr_partial || slabs_node(s, node))
3734 * Release all resources used by a slab cache.
3736 int __kmem_cache_shutdown(struct kmem_cache *s)
3739 struct kmem_cache_node *n;
3742 /* Attempt to free all objects */
3743 for_each_kmem_cache_node(s, node, n) {
3745 if (n->nr_partial || slabs_node(s, node))
3748 sysfs_slab_remove(s);
3752 /********************************************************************
3754 *******************************************************************/
3756 static int __init setup_slub_min_order(char *str)
3758 get_option(&str, (int *)&slub_min_order);
3763 __setup("slub_min_order=", setup_slub_min_order);
3765 static int __init setup_slub_max_order(char *str)
3767 get_option(&str, (int *)&slub_max_order);
3768 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3773 __setup("slub_max_order=", setup_slub_max_order);
3775 static int __init setup_slub_min_objects(char *str)
3777 get_option(&str, (int *)&slub_min_objects);
3782 __setup("slub_min_objects=", setup_slub_min_objects);
3784 void *__kmalloc(size_t size, gfp_t flags)
3786 struct kmem_cache *s;
3789 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3790 return kmalloc_large(size, flags);
3792 s = kmalloc_slab(size, flags);
3794 if (unlikely(ZERO_OR_NULL_PTR(s)))
3797 ret = slab_alloc(s, flags, _RET_IP_);
3799 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3801 ret = kasan_kmalloc(s, ret, size, flags);
3805 EXPORT_SYMBOL(__kmalloc);
3808 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3813 flags |= __GFP_COMP;
3814 page = alloc_pages_node(node, flags, get_order(size));
3816 ptr = page_address(page);
3818 return kmalloc_large_node_hook(ptr, size, flags);
3821 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3823 struct kmem_cache *s;
3826 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3827 ret = kmalloc_large_node(size, flags, node);
3829 trace_kmalloc_node(_RET_IP_, ret,
3830 size, PAGE_SIZE << get_order(size),
3836 s = kmalloc_slab(size, flags);
3838 if (unlikely(ZERO_OR_NULL_PTR(s)))
3841 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3843 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3845 ret = kasan_kmalloc(s, ret, size, flags);
3849 EXPORT_SYMBOL(__kmalloc_node);
3852 #ifdef CONFIG_HARDENED_USERCOPY
3854 * Rejects incorrectly sized objects and objects that are to be copied
3855 * to/from userspace but do not fall entirely within the containing slab
3856 * cache's usercopy region.
3858 * Returns NULL if check passes, otherwise const char * to name of cache
3859 * to indicate an error.
3861 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3864 struct kmem_cache *s;
3865 unsigned int offset;
3868 ptr = kasan_reset_tag(ptr);
3870 /* Find object and usable object size. */
3871 s = page->slab_cache;
3873 /* Reject impossible pointers. */
3874 if (ptr < page_address(page))
3875 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3878 /* Find offset within object. */
3879 offset = (ptr - page_address(page)) % s->size;
3881 /* Adjust for redzone and reject if within the redzone. */
3882 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3883 if (offset < s->red_left_pad)
3884 usercopy_abort("SLUB object in left red zone",
3885 s->name, to_user, offset, n);
3886 offset -= s->red_left_pad;
3889 /* Allow address range falling entirely within usercopy region. */
3890 if (offset >= s->useroffset &&
3891 offset - s->useroffset <= s->usersize &&
3892 n <= s->useroffset - offset + s->usersize)
3896 * If the copy is still within the allocated object, produce
3897 * a warning instead of rejecting the copy. This is intended
3898 * to be a temporary method to find any missing usercopy
3901 object_size = slab_ksize(s);
3902 if (usercopy_fallback &&
3903 offset <= object_size && n <= object_size - offset) {
3904 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3908 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3910 #endif /* CONFIG_HARDENED_USERCOPY */
3912 static size_t __ksize(const void *object)
3916 if (unlikely(object == ZERO_SIZE_PTR))
3919 page = virt_to_head_page(object);
3921 if (unlikely(!PageSlab(page))) {
3922 WARN_ON(!PageCompound(page));
3923 return PAGE_SIZE << compound_order(page);
3926 return slab_ksize(page->slab_cache);
3929 size_t ksize(const void *object)
3931 size_t size = __ksize(object);
3932 /* We assume that ksize callers could use whole allocated area,
3933 * so we need to unpoison this area.
3935 kasan_unpoison_shadow(object, size);
3938 EXPORT_SYMBOL(ksize);
3940 void kfree(const void *x)
3943 void *object = (void *)x;
3945 trace_kfree(_RET_IP_, x);
3947 if (unlikely(ZERO_OR_NULL_PTR(x)))
3950 page = virt_to_head_page(x);
3951 if (unlikely(!PageSlab(page))) {
3952 BUG_ON(!PageCompound(page));
3954 __free_pages(page, compound_order(page));
3957 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3959 EXPORT_SYMBOL(kfree);
3961 #define SHRINK_PROMOTE_MAX 32
3964 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3965 * up most to the head of the partial lists. New allocations will then
3966 * fill those up and thus they can be removed from the partial lists.
3968 * The slabs with the least items are placed last. This results in them
3969 * being allocated from last increasing the chance that the last objects
3970 * are freed in them.
3972 int __kmem_cache_shrink(struct kmem_cache *s)
3976 struct kmem_cache_node *n;
3979 struct list_head discard;
3980 struct list_head promote[SHRINK_PROMOTE_MAX];
3981 unsigned long flags;
3985 for_each_kmem_cache_node(s, node, n) {
3986 INIT_LIST_HEAD(&discard);
3987 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3988 INIT_LIST_HEAD(promote + i);
3990 spin_lock_irqsave(&n->list_lock, flags);
3993 * Build lists of slabs to discard or promote.
3995 * Note that concurrent frees may occur while we hold the
3996 * list_lock. page->inuse here is the upper limit.
3998 list_for_each_entry_safe(page, t, &n->partial, lru) {
3999 int free = page->objects - page->inuse;
4001 /* Do not reread page->inuse */
4004 /* We do not keep full slabs on the list */
4007 if (free == page->objects) {
4008 list_move(&page->lru, &discard);
4010 } else if (free <= SHRINK_PROMOTE_MAX)
4011 list_move(&page->lru, promote + free - 1);
4015 * Promote the slabs filled up most to the head of the
4018 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4019 list_splice(promote + i, &n->partial);
4021 spin_unlock_irqrestore(&n->list_lock, flags);
4023 /* Release empty slabs */
4024 list_for_each_entry_safe(page, t, &discard, lru)
4025 discard_slab(s, page);
4027 if (slabs_node(s, node))
4035 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
4038 * Called with all the locks held after a sched RCU grace period.
4039 * Even if @s becomes empty after shrinking, we can't know that @s
4040 * doesn't have allocations already in-flight and thus can't
4041 * destroy @s until the associated memcg is released.
4043 * However, let's remove the sysfs files for empty caches here.
4044 * Each cache has a lot of interface files which aren't
4045 * particularly useful for empty draining caches; otherwise, we can
4046 * easily end up with millions of unnecessary sysfs files on
4047 * systems which have a lot of memory and transient cgroups.
4049 if (!__kmem_cache_shrink(s))
4050 sysfs_slab_remove(s);
4053 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4056 * Disable empty slabs caching. Used to avoid pinning offline
4057 * memory cgroups by kmem pages that can be freed.
4059 slub_set_cpu_partial(s, 0);
4063 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4064 * we have to make sure the change is visible before shrinking.
4066 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4070 static int slab_mem_going_offline_callback(void *arg)
4072 struct kmem_cache *s;
4074 mutex_lock(&slab_mutex);
4075 list_for_each_entry(s, &slab_caches, list)
4076 __kmem_cache_shrink(s);
4077 mutex_unlock(&slab_mutex);
4082 static void slab_mem_offline_callback(void *arg)
4084 struct kmem_cache_node *n;
4085 struct kmem_cache *s;
4086 struct memory_notify *marg = arg;
4089 offline_node = marg->status_change_nid_normal;
4092 * If the node still has available memory. we need kmem_cache_node
4095 if (offline_node < 0)
4098 mutex_lock(&slab_mutex);
4099 list_for_each_entry(s, &slab_caches, list) {
4100 n = get_node(s, offline_node);
4103 * if n->nr_slabs > 0, slabs still exist on the node
4104 * that is going down. We were unable to free them,
4105 * and offline_pages() function shouldn't call this
4106 * callback. So, we must fail.
4108 BUG_ON(slabs_node(s, offline_node));
4110 s->node[offline_node] = NULL;
4111 kmem_cache_free(kmem_cache_node, n);
4114 mutex_unlock(&slab_mutex);
4117 static int slab_mem_going_online_callback(void *arg)
4119 struct kmem_cache_node *n;
4120 struct kmem_cache *s;
4121 struct memory_notify *marg = arg;
4122 int nid = marg->status_change_nid_normal;
4126 * If the node's memory is already available, then kmem_cache_node is
4127 * already created. Nothing to do.
4133 * We are bringing a node online. No memory is available yet. We must
4134 * allocate a kmem_cache_node structure in order to bring the node
4137 mutex_lock(&slab_mutex);
4138 list_for_each_entry(s, &slab_caches, list) {
4140 * XXX: kmem_cache_alloc_node will fallback to other nodes
4141 * since memory is not yet available from the node that
4144 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4149 init_kmem_cache_node(n);
4153 mutex_unlock(&slab_mutex);
4157 static int slab_memory_callback(struct notifier_block *self,
4158 unsigned long action, void *arg)
4163 case MEM_GOING_ONLINE:
4164 ret = slab_mem_going_online_callback(arg);
4166 case MEM_GOING_OFFLINE:
4167 ret = slab_mem_going_offline_callback(arg);
4170 case MEM_CANCEL_ONLINE:
4171 slab_mem_offline_callback(arg);
4174 case MEM_CANCEL_OFFLINE:
4178 ret = notifier_from_errno(ret);
4184 static struct notifier_block slab_memory_callback_nb = {
4185 .notifier_call = slab_memory_callback,
4186 .priority = SLAB_CALLBACK_PRI,
4189 /********************************************************************
4190 * Basic setup of slabs
4191 *******************************************************************/
4194 * Used for early kmem_cache structures that were allocated using
4195 * the page allocator. Allocate them properly then fix up the pointers
4196 * that may be pointing to the wrong kmem_cache structure.
4199 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4202 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4203 struct kmem_cache_node *n;
4205 memcpy(s, static_cache, kmem_cache->object_size);
4208 * This runs very early, and only the boot processor is supposed to be
4209 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4212 __flush_cpu_slab(s, smp_processor_id());
4213 for_each_kmem_cache_node(s, node, n) {
4216 list_for_each_entry(p, &n->partial, lru)
4219 #ifdef CONFIG_SLUB_DEBUG
4220 list_for_each_entry(p, &n->full, lru)
4224 slab_init_memcg_params(s);
4225 list_add(&s->list, &slab_caches);
4226 memcg_link_cache(s);
4230 void __init kmem_cache_init(void)
4232 static __initdata struct kmem_cache boot_kmem_cache,
4233 boot_kmem_cache_node;
4235 if (debug_guardpage_minorder())
4238 kmem_cache_node = &boot_kmem_cache_node;
4239 kmem_cache = &boot_kmem_cache;
4241 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4242 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4244 register_hotmemory_notifier(&slab_memory_callback_nb);
4246 /* Able to allocate the per node structures */
4247 slab_state = PARTIAL;
4249 create_boot_cache(kmem_cache, "kmem_cache",
4250 offsetof(struct kmem_cache, node) +
4251 nr_node_ids * sizeof(struct kmem_cache_node *),
4252 SLAB_HWCACHE_ALIGN, 0, 0);
4254 kmem_cache = bootstrap(&boot_kmem_cache);
4255 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4257 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4258 setup_kmalloc_cache_index_table();
4259 create_kmalloc_caches(0);
4261 /* Setup random freelists for each cache */
4262 init_freelist_randomization();
4264 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4267 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%d\n",
4269 slub_min_order, slub_max_order, slub_min_objects,
4270 nr_cpu_ids, nr_node_ids);
4273 void __init kmem_cache_init_late(void)
4278 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4279 slab_flags_t flags, void (*ctor)(void *))
4281 struct kmem_cache *s, *c;
4283 s = find_mergeable(size, align, flags, name, ctor);
4288 * Adjust the object sizes so that we clear
4289 * the complete object on kzalloc.
4291 s->object_size = max(s->object_size, size);
4292 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4294 for_each_memcg_cache(c, s) {
4295 c->object_size = s->object_size;
4296 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4299 if (sysfs_slab_alias(s, name)) {
4308 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4312 err = kmem_cache_open(s, flags);
4316 /* Mutex is not taken during early boot */
4317 if (slab_state <= UP)
4320 memcg_propagate_slab_attrs(s);
4321 err = sysfs_slab_add(s);
4323 __kmem_cache_release(s);
4328 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4330 struct kmem_cache *s;
4333 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4334 return kmalloc_large(size, gfpflags);
4336 s = kmalloc_slab(size, gfpflags);
4338 if (unlikely(ZERO_OR_NULL_PTR(s)))
4341 ret = slab_alloc(s, gfpflags, caller);
4343 /* Honor the call site pointer we received. */
4344 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4350 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4351 int node, unsigned long caller)
4353 struct kmem_cache *s;
4356 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4357 ret = kmalloc_large_node(size, gfpflags, node);
4359 trace_kmalloc_node(caller, ret,
4360 size, PAGE_SIZE << get_order(size),
4366 s = kmalloc_slab(size, gfpflags);
4368 if (unlikely(ZERO_OR_NULL_PTR(s)))
4371 ret = slab_alloc_node(s, gfpflags, node, caller);
4373 /* Honor the call site pointer we received. */
4374 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4381 static int count_inuse(struct page *page)
4386 static int count_total(struct page *page)
4388 return page->objects;
4392 #ifdef CONFIG_SLUB_DEBUG
4393 static int validate_slab(struct kmem_cache *s, struct page *page,
4397 void *addr = page_address(page);
4399 if (!check_slab(s, page) ||
4400 !on_freelist(s, page, NULL))
4403 /* Now we know that a valid freelist exists */
4404 bitmap_zero(map, page->objects);
4406 get_map(s, page, map);
4407 for_each_object(p, s, addr, page->objects) {
4408 if (test_bit(slab_index(p, s, addr), map))
4409 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4413 for_each_object(p, s, addr, page->objects)
4414 if (!test_bit(slab_index(p, s, addr), map))
4415 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4420 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4424 validate_slab(s, page, map);
4428 static int validate_slab_node(struct kmem_cache *s,
4429 struct kmem_cache_node *n, unsigned long *map)
4431 unsigned long count = 0;
4433 unsigned long flags;
4435 spin_lock_irqsave(&n->list_lock, flags);
4437 list_for_each_entry(page, &n->partial, lru) {
4438 validate_slab_slab(s, page, map);
4441 if (count != n->nr_partial)
4442 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4443 s->name, count, n->nr_partial);
4445 if (!(s->flags & SLAB_STORE_USER))
4448 list_for_each_entry(page, &n->full, lru) {
4449 validate_slab_slab(s, page, map);
4452 if (count != atomic_long_read(&n->nr_slabs))
4453 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4454 s->name, count, atomic_long_read(&n->nr_slabs));
4457 spin_unlock_irqrestore(&n->list_lock, flags);
4461 static long validate_slab_cache(struct kmem_cache *s)
4464 unsigned long count = 0;
4465 struct kmem_cache_node *n;
4466 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4472 for_each_kmem_cache_node(s, node, n)
4473 count += validate_slab_node(s, n, map);
4478 * Generate lists of code addresses where slabcache objects are allocated
4483 unsigned long count;
4490 DECLARE_BITMAP(cpus, NR_CPUS);
4496 unsigned long count;
4497 struct location *loc;
4500 static void free_loc_track(struct loc_track *t)
4503 free_pages((unsigned long)t->loc,
4504 get_order(sizeof(struct location) * t->max));
4507 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4512 order = get_order(sizeof(struct location) * max);
4514 l = (void *)__get_free_pages(flags, order);
4519 memcpy(l, t->loc, sizeof(struct location) * t->count);
4527 static int add_location(struct loc_track *t, struct kmem_cache *s,
4528 const struct track *track)
4530 long start, end, pos;
4532 unsigned long caddr;
4533 unsigned long age = jiffies - track->when;
4539 pos = start + (end - start + 1) / 2;
4542 * There is nothing at "end". If we end up there
4543 * we need to add something to before end.
4548 caddr = t->loc[pos].addr;
4549 if (track->addr == caddr) {
4555 if (age < l->min_time)
4557 if (age > l->max_time)
4560 if (track->pid < l->min_pid)
4561 l->min_pid = track->pid;
4562 if (track->pid > l->max_pid)
4563 l->max_pid = track->pid;
4565 cpumask_set_cpu(track->cpu,
4566 to_cpumask(l->cpus));
4568 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4572 if (track->addr < caddr)
4579 * Not found. Insert new tracking element.
4581 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4587 (t->count - pos) * sizeof(struct location));
4590 l->addr = track->addr;
4594 l->min_pid = track->pid;
4595 l->max_pid = track->pid;
4596 cpumask_clear(to_cpumask(l->cpus));
4597 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4598 nodes_clear(l->nodes);
4599 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4603 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4604 struct page *page, enum track_item alloc,
4607 void *addr = page_address(page);
4610 bitmap_zero(map, page->objects);
4611 get_map(s, page, map);
4613 for_each_object(p, s, addr, page->objects)
4614 if (!test_bit(slab_index(p, s, addr), map))
4615 add_location(t, s, get_track(s, p, alloc));
4618 static int list_locations(struct kmem_cache *s, char *buf,
4619 enum track_item alloc)
4623 struct loc_track t = { 0, 0, NULL };
4625 struct kmem_cache_node *n;
4626 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4628 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4631 return sprintf(buf, "Out of memory\n");
4633 /* Push back cpu slabs */
4636 for_each_kmem_cache_node(s, node, n) {
4637 unsigned long flags;
4640 if (!atomic_long_read(&n->nr_slabs))
4643 spin_lock_irqsave(&n->list_lock, flags);
4644 list_for_each_entry(page, &n->partial, lru)
4645 process_slab(&t, s, page, alloc, map);
4646 list_for_each_entry(page, &n->full, lru)
4647 process_slab(&t, s, page, alloc, map);
4648 spin_unlock_irqrestore(&n->list_lock, flags);
4651 for (i = 0; i < t.count; i++) {
4652 struct location *l = &t.loc[i];
4654 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4656 len += sprintf(buf + len, "%7ld ", l->count);
4659 len += sprintf(buf + len, "%pS", (void *)l->addr);
4661 len += sprintf(buf + len, "<not-available>");
4663 if (l->sum_time != l->min_time) {
4664 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4666 (long)div_u64(l->sum_time, l->count),
4669 len += sprintf(buf + len, " age=%ld",
4672 if (l->min_pid != l->max_pid)
4673 len += sprintf(buf + len, " pid=%ld-%ld",
4674 l->min_pid, l->max_pid);
4676 len += sprintf(buf + len, " pid=%ld",
4679 if (num_online_cpus() > 1 &&
4680 !cpumask_empty(to_cpumask(l->cpus)) &&
4681 len < PAGE_SIZE - 60)
4682 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4684 cpumask_pr_args(to_cpumask(l->cpus)));
4686 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4687 len < PAGE_SIZE - 60)
4688 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4690 nodemask_pr_args(&l->nodes));
4692 len += sprintf(buf + len, "\n");
4698 len += sprintf(buf, "No data\n");
4703 #ifdef SLUB_RESILIENCY_TEST
4704 static void __init resiliency_test(void)
4707 int type = KMALLOC_NORMAL;
4709 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4711 pr_err("SLUB resiliency testing\n");
4712 pr_err("-----------------------\n");
4713 pr_err("A. Corruption after allocation\n");
4715 p = kzalloc(16, GFP_KERNEL);
4717 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4720 validate_slab_cache(kmalloc_caches[type][4]);
4722 /* Hmmm... The next two are dangerous */
4723 p = kzalloc(32, GFP_KERNEL);
4724 p[32 + sizeof(void *)] = 0x34;
4725 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4727 pr_err("If allocated object is overwritten then not detectable\n\n");
4729 validate_slab_cache(kmalloc_caches[type][5]);
4730 p = kzalloc(64, GFP_KERNEL);
4731 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4733 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4735 pr_err("If allocated object is overwritten then not detectable\n\n");
4736 validate_slab_cache(kmalloc_caches[type][6]);
4738 pr_err("\nB. Corruption after free\n");
4739 p = kzalloc(128, GFP_KERNEL);
4742 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4743 validate_slab_cache(kmalloc_caches[type][7]);
4745 p = kzalloc(256, GFP_KERNEL);
4748 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4749 validate_slab_cache(kmalloc_caches[type][8]);
4751 p = kzalloc(512, GFP_KERNEL);
4754 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4755 validate_slab_cache(kmalloc_caches[type][9]);
4759 static void resiliency_test(void) {};
4764 enum slab_stat_type {
4765 SL_ALL, /* All slabs */
4766 SL_PARTIAL, /* Only partially allocated slabs */
4767 SL_CPU, /* Only slabs used for cpu caches */
4768 SL_OBJECTS, /* Determine allocated objects not slabs */
4769 SL_TOTAL /* Determine object capacity not slabs */
4772 #define SO_ALL (1 << SL_ALL)
4773 #define SO_PARTIAL (1 << SL_PARTIAL)
4774 #define SO_CPU (1 << SL_CPU)
4775 #define SO_OBJECTS (1 << SL_OBJECTS)
4776 #define SO_TOTAL (1 << SL_TOTAL)
4779 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4781 static int __init setup_slub_memcg_sysfs(char *str)
4785 if (get_option(&str, &v) > 0)
4786 memcg_sysfs_enabled = v;
4791 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4794 static ssize_t show_slab_objects(struct kmem_cache *s,
4795 char *buf, unsigned long flags)
4797 unsigned long total = 0;
4800 unsigned long *nodes;
4802 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4806 if (flags & SO_CPU) {
4809 for_each_possible_cpu(cpu) {
4810 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4815 page = READ_ONCE(c->page);
4819 node = page_to_nid(page);
4820 if (flags & SO_TOTAL)
4822 else if (flags & SO_OBJECTS)
4830 page = slub_percpu_partial_read_once(c);
4832 node = page_to_nid(page);
4833 if (flags & SO_TOTAL)
4835 else if (flags & SO_OBJECTS)
4846 #ifdef CONFIG_SLUB_DEBUG
4847 if (flags & SO_ALL) {
4848 struct kmem_cache_node *n;
4850 for_each_kmem_cache_node(s, node, n) {
4852 if (flags & SO_TOTAL)
4853 x = atomic_long_read(&n->total_objects);
4854 else if (flags & SO_OBJECTS)
4855 x = atomic_long_read(&n->total_objects) -
4856 count_partial(n, count_free);
4858 x = atomic_long_read(&n->nr_slabs);
4865 if (flags & SO_PARTIAL) {
4866 struct kmem_cache_node *n;
4868 for_each_kmem_cache_node(s, node, n) {
4869 if (flags & SO_TOTAL)
4870 x = count_partial(n, count_total);
4871 else if (flags & SO_OBJECTS)
4872 x = count_partial(n, count_inuse);
4879 x = sprintf(buf, "%lu", total);
4881 for (node = 0; node < nr_node_ids; node++)
4883 x += sprintf(buf + x, " N%d=%lu",
4888 return x + sprintf(buf + x, "\n");
4891 #ifdef CONFIG_SLUB_DEBUG
4892 static int any_slab_objects(struct kmem_cache *s)
4895 struct kmem_cache_node *n;
4897 for_each_kmem_cache_node(s, node, n)
4898 if (atomic_long_read(&n->total_objects))
4905 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4906 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4908 struct slab_attribute {
4909 struct attribute attr;
4910 ssize_t (*show)(struct kmem_cache *s, char *buf);
4911 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4914 #define SLAB_ATTR_RO(_name) \
4915 static struct slab_attribute _name##_attr = \
4916 __ATTR(_name, 0400, _name##_show, NULL)
4918 #define SLAB_ATTR(_name) \
4919 static struct slab_attribute _name##_attr = \
4920 __ATTR(_name, 0600, _name##_show, _name##_store)
4922 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4924 return sprintf(buf, "%u\n", s->size);
4926 SLAB_ATTR_RO(slab_size);
4928 static ssize_t align_show(struct kmem_cache *s, char *buf)
4930 return sprintf(buf, "%u\n", s->align);
4932 SLAB_ATTR_RO(align);
4934 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4936 return sprintf(buf, "%u\n", s->object_size);
4938 SLAB_ATTR_RO(object_size);
4940 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4942 return sprintf(buf, "%u\n", oo_objects(s->oo));
4944 SLAB_ATTR_RO(objs_per_slab);
4946 static ssize_t order_store(struct kmem_cache *s,
4947 const char *buf, size_t length)
4952 err = kstrtouint(buf, 10, &order);
4956 if (order > slub_max_order || order < slub_min_order)
4959 calculate_sizes(s, order);
4963 static ssize_t order_show(struct kmem_cache *s, char *buf)
4965 return sprintf(buf, "%u\n", oo_order(s->oo));
4969 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4971 return sprintf(buf, "%lu\n", s->min_partial);
4974 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4980 err = kstrtoul(buf, 10, &min);
4984 set_min_partial(s, min);
4987 SLAB_ATTR(min_partial);
4989 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4991 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4994 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4997 unsigned int objects;
5000 err = kstrtouint(buf, 10, &objects);
5003 if (objects && !kmem_cache_has_cpu_partial(s))
5006 slub_set_cpu_partial(s, objects);
5010 SLAB_ATTR(cpu_partial);
5012 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5016 return sprintf(buf, "%pS\n", s->ctor);
5020 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5022 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5024 SLAB_ATTR_RO(aliases);
5026 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5028 return show_slab_objects(s, buf, SO_PARTIAL);
5030 SLAB_ATTR_RO(partial);
5032 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5034 return show_slab_objects(s, buf, SO_CPU);
5036 SLAB_ATTR_RO(cpu_slabs);
5038 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5040 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5042 SLAB_ATTR_RO(objects);
5044 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5046 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5048 SLAB_ATTR_RO(objects_partial);
5050 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5057 for_each_online_cpu(cpu) {
5060 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5063 pages += page->pages;
5064 objects += page->pobjects;
5068 len = sprintf(buf, "%d(%d)", objects, pages);
5071 for_each_online_cpu(cpu) {
5074 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5076 if (page && len < PAGE_SIZE - 20)
5077 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5078 page->pobjects, page->pages);
5081 return len + sprintf(buf + len, "\n");
5083 SLAB_ATTR_RO(slabs_cpu_partial);
5085 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5087 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5090 static ssize_t reclaim_account_store(struct kmem_cache *s,
5091 const char *buf, size_t length)
5093 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5095 s->flags |= SLAB_RECLAIM_ACCOUNT;
5098 SLAB_ATTR(reclaim_account);
5100 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5102 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5104 SLAB_ATTR_RO(hwcache_align);
5106 #ifdef CONFIG_ZONE_DMA
5107 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5109 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5111 SLAB_ATTR_RO(cache_dma);
5114 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5116 return sprintf(buf, "%u\n", s->usersize);
5118 SLAB_ATTR_RO(usersize);
5120 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5122 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5124 SLAB_ATTR_RO(destroy_by_rcu);
5126 #ifdef CONFIG_SLUB_DEBUG
5127 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5129 return show_slab_objects(s, buf, SO_ALL);
5131 SLAB_ATTR_RO(slabs);
5133 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5135 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5137 SLAB_ATTR_RO(total_objects);
5139 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5141 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5144 static ssize_t sanity_checks_store(struct kmem_cache *s,
5145 const char *buf, size_t length)
5147 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5148 if (buf[0] == '1') {
5149 s->flags &= ~__CMPXCHG_DOUBLE;
5150 s->flags |= SLAB_CONSISTENCY_CHECKS;
5154 SLAB_ATTR(sanity_checks);
5156 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5158 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5161 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5165 * Tracing a merged cache is going to give confusing results
5166 * as well as cause other issues like converting a mergeable
5167 * cache into an umergeable one.
5169 if (s->refcount > 1)
5172 s->flags &= ~SLAB_TRACE;
5173 if (buf[0] == '1') {
5174 s->flags &= ~__CMPXCHG_DOUBLE;
5175 s->flags |= SLAB_TRACE;
5181 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5183 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5186 static ssize_t red_zone_store(struct kmem_cache *s,
5187 const char *buf, size_t length)
5189 if (any_slab_objects(s))
5192 s->flags &= ~SLAB_RED_ZONE;
5193 if (buf[0] == '1') {
5194 s->flags |= SLAB_RED_ZONE;
5196 calculate_sizes(s, -1);
5199 SLAB_ATTR(red_zone);
5201 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5203 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5206 static ssize_t poison_store(struct kmem_cache *s,
5207 const char *buf, size_t length)
5209 if (any_slab_objects(s))
5212 s->flags &= ~SLAB_POISON;
5213 if (buf[0] == '1') {
5214 s->flags |= SLAB_POISON;
5216 calculate_sizes(s, -1);
5221 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5223 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5226 static ssize_t store_user_store(struct kmem_cache *s,
5227 const char *buf, size_t length)
5229 if (any_slab_objects(s))
5232 s->flags &= ~SLAB_STORE_USER;
5233 if (buf[0] == '1') {
5234 s->flags &= ~__CMPXCHG_DOUBLE;
5235 s->flags |= SLAB_STORE_USER;
5237 calculate_sizes(s, -1);
5240 SLAB_ATTR(store_user);
5242 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5247 static ssize_t validate_store(struct kmem_cache *s,
5248 const char *buf, size_t length)
5252 if (buf[0] == '1') {
5253 ret = validate_slab_cache(s);
5259 SLAB_ATTR(validate);
5261 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5263 if (!(s->flags & SLAB_STORE_USER))
5265 return list_locations(s, buf, TRACK_ALLOC);
5267 SLAB_ATTR_RO(alloc_calls);
5269 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5271 if (!(s->flags & SLAB_STORE_USER))
5273 return list_locations(s, buf, TRACK_FREE);
5275 SLAB_ATTR_RO(free_calls);
5276 #endif /* CONFIG_SLUB_DEBUG */
5278 #ifdef CONFIG_FAILSLAB
5279 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5281 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5284 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5287 if (s->refcount > 1)
5290 s->flags &= ~SLAB_FAILSLAB;
5292 s->flags |= SLAB_FAILSLAB;
5295 SLAB_ATTR(failslab);
5298 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5303 static ssize_t shrink_store(struct kmem_cache *s,
5304 const char *buf, size_t length)
5307 kmem_cache_shrink(s);
5315 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5317 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5320 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5321 const char *buf, size_t length)
5326 err = kstrtouint(buf, 10, &ratio);
5332 s->remote_node_defrag_ratio = ratio * 10;
5336 SLAB_ATTR(remote_node_defrag_ratio);
5339 #ifdef CONFIG_SLUB_STATS
5340 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5342 unsigned long sum = 0;
5345 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5350 for_each_online_cpu(cpu) {
5351 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5357 len = sprintf(buf, "%lu", sum);
5360 for_each_online_cpu(cpu) {
5361 if (data[cpu] && len < PAGE_SIZE - 20)
5362 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5366 return len + sprintf(buf + len, "\n");
5369 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5373 for_each_online_cpu(cpu)
5374 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5377 #define STAT_ATTR(si, text) \
5378 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5380 return show_stat(s, buf, si); \
5382 static ssize_t text##_store(struct kmem_cache *s, \
5383 const char *buf, size_t length) \
5385 if (buf[0] != '0') \
5387 clear_stat(s, si); \
5392 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5393 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5394 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5395 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5396 STAT_ATTR(FREE_FROZEN, free_frozen);
5397 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5398 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5399 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5400 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5401 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5402 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5403 STAT_ATTR(FREE_SLAB, free_slab);
5404 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5405 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5406 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5407 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5408 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5409 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5410 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5411 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5412 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5413 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5414 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5415 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5416 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5417 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5420 static struct attribute *slab_attrs[] = {
5421 &slab_size_attr.attr,
5422 &object_size_attr.attr,
5423 &objs_per_slab_attr.attr,
5425 &min_partial_attr.attr,
5426 &cpu_partial_attr.attr,
5428 &objects_partial_attr.attr,
5430 &cpu_slabs_attr.attr,
5434 &hwcache_align_attr.attr,
5435 &reclaim_account_attr.attr,
5436 &destroy_by_rcu_attr.attr,
5438 &slabs_cpu_partial_attr.attr,
5439 #ifdef CONFIG_SLUB_DEBUG
5440 &total_objects_attr.attr,
5442 &sanity_checks_attr.attr,
5444 &red_zone_attr.attr,
5446 &store_user_attr.attr,
5447 &validate_attr.attr,
5448 &alloc_calls_attr.attr,
5449 &free_calls_attr.attr,
5451 #ifdef CONFIG_ZONE_DMA
5452 &cache_dma_attr.attr,
5455 &remote_node_defrag_ratio_attr.attr,
5457 #ifdef CONFIG_SLUB_STATS
5458 &alloc_fastpath_attr.attr,
5459 &alloc_slowpath_attr.attr,
5460 &free_fastpath_attr.attr,
5461 &free_slowpath_attr.attr,
5462 &free_frozen_attr.attr,
5463 &free_add_partial_attr.attr,
5464 &free_remove_partial_attr.attr,
5465 &alloc_from_partial_attr.attr,
5466 &alloc_slab_attr.attr,
5467 &alloc_refill_attr.attr,
5468 &alloc_node_mismatch_attr.attr,
5469 &free_slab_attr.attr,
5470 &cpuslab_flush_attr.attr,
5471 &deactivate_full_attr.attr,
5472 &deactivate_empty_attr.attr,
5473 &deactivate_to_head_attr.attr,
5474 &deactivate_to_tail_attr.attr,
5475 &deactivate_remote_frees_attr.attr,
5476 &deactivate_bypass_attr.attr,
5477 &order_fallback_attr.attr,
5478 &cmpxchg_double_fail_attr.attr,
5479 &cmpxchg_double_cpu_fail_attr.attr,
5480 &cpu_partial_alloc_attr.attr,
5481 &cpu_partial_free_attr.attr,
5482 &cpu_partial_node_attr.attr,
5483 &cpu_partial_drain_attr.attr,
5485 #ifdef CONFIG_FAILSLAB
5486 &failslab_attr.attr,
5488 &usersize_attr.attr,
5493 static const struct attribute_group slab_attr_group = {
5494 .attrs = slab_attrs,
5497 static ssize_t slab_attr_show(struct kobject *kobj,
5498 struct attribute *attr,
5501 struct slab_attribute *attribute;
5502 struct kmem_cache *s;
5505 attribute = to_slab_attr(attr);
5508 if (!attribute->show)
5511 err = attribute->show(s, buf);
5516 static ssize_t slab_attr_store(struct kobject *kobj,
5517 struct attribute *attr,
5518 const char *buf, size_t len)
5520 struct slab_attribute *attribute;
5521 struct kmem_cache *s;
5524 attribute = to_slab_attr(attr);
5527 if (!attribute->store)
5530 err = attribute->store(s, buf, len);
5532 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5533 struct kmem_cache *c;
5535 mutex_lock(&slab_mutex);
5536 if (s->max_attr_size < len)
5537 s->max_attr_size = len;
5540 * This is a best effort propagation, so this function's return
5541 * value will be determined by the parent cache only. This is
5542 * basically because not all attributes will have a well
5543 * defined semantics for rollbacks - most of the actions will
5544 * have permanent effects.
5546 * Returning the error value of any of the children that fail
5547 * is not 100 % defined, in the sense that users seeing the
5548 * error code won't be able to know anything about the state of
5551 * Only returning the error code for the parent cache at least
5552 * has well defined semantics. The cache being written to
5553 * directly either failed or succeeded, in which case we loop
5554 * through the descendants with best-effort propagation.
5556 for_each_memcg_cache(c, s)
5557 attribute->store(c, buf, len);
5558 mutex_unlock(&slab_mutex);
5564 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5568 char *buffer = NULL;
5569 struct kmem_cache *root_cache;
5571 if (is_root_cache(s))
5574 root_cache = s->memcg_params.root_cache;
5577 * This mean this cache had no attribute written. Therefore, no point
5578 * in copying default values around
5580 if (!root_cache->max_attr_size)
5583 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5586 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5589 if (!attr || !attr->store || !attr->show)
5593 * It is really bad that we have to allocate here, so we will
5594 * do it only as a fallback. If we actually allocate, though,
5595 * we can just use the allocated buffer until the end.
5597 * Most of the slub attributes will tend to be very small in
5598 * size, but sysfs allows buffers up to a page, so they can
5599 * theoretically happen.
5603 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5606 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5607 if (WARN_ON(!buffer))
5612 len = attr->show(root_cache, buf);
5614 attr->store(s, buf, len);
5618 free_page((unsigned long)buffer);
5622 static void kmem_cache_release(struct kobject *k)
5624 slab_kmem_cache_release(to_slab(k));
5627 static const struct sysfs_ops slab_sysfs_ops = {
5628 .show = slab_attr_show,
5629 .store = slab_attr_store,
5632 static struct kobj_type slab_ktype = {
5633 .sysfs_ops = &slab_sysfs_ops,
5634 .release = kmem_cache_release,
5637 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5639 struct kobj_type *ktype = get_ktype(kobj);
5641 if (ktype == &slab_ktype)
5646 static const struct kset_uevent_ops slab_uevent_ops = {
5647 .filter = uevent_filter,
5650 static struct kset *slab_kset;
5652 static inline struct kset *cache_kset(struct kmem_cache *s)
5655 if (!is_root_cache(s))
5656 return s->memcg_params.root_cache->memcg_kset;
5661 #define ID_STR_LENGTH 64
5663 /* Create a unique string id for a slab cache:
5665 * Format :[flags-]size
5667 static char *create_unique_id(struct kmem_cache *s)
5669 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5676 * First flags affecting slabcache operations. We will only
5677 * get here for aliasable slabs so we do not need to support
5678 * too many flags. The flags here must cover all flags that
5679 * are matched during merging to guarantee that the id is
5682 if (s->flags & SLAB_CACHE_DMA)
5684 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5686 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5688 if (s->flags & SLAB_ACCOUNT)
5692 p += sprintf(p, "%07u", s->size);
5694 BUG_ON(p > name + ID_STR_LENGTH - 1);
5698 static void sysfs_slab_remove_workfn(struct work_struct *work)
5700 struct kmem_cache *s =
5701 container_of(work, struct kmem_cache, kobj_remove_work);
5703 if (!s->kobj.state_in_sysfs)
5705 * For a memcg cache, this may be called during
5706 * deactivation and again on shutdown. Remove only once.
5707 * A cache is never shut down before deactivation is
5708 * complete, so no need to worry about synchronization.
5713 kset_unregister(s->memcg_kset);
5715 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5717 kobject_put(&s->kobj);
5720 static int sysfs_slab_add(struct kmem_cache *s)
5724 struct kset *kset = cache_kset(s);
5725 int unmergeable = slab_unmergeable(s);
5727 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5730 kobject_init(&s->kobj, &slab_ktype);
5734 if (!unmergeable && disable_higher_order_debug &&
5735 (slub_debug & DEBUG_METADATA_FLAGS))
5740 * Slabcache can never be merged so we can use the name proper.
5741 * This is typically the case for debug situations. In that
5742 * case we can catch duplicate names easily.
5744 sysfs_remove_link(&slab_kset->kobj, s->name);
5748 * Create a unique name for the slab as a target
5751 name = create_unique_id(s);
5754 s->kobj.kset = kset;
5755 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5759 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5764 if (is_root_cache(s) && memcg_sysfs_enabled) {
5765 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5766 if (!s->memcg_kset) {
5773 kobject_uevent(&s->kobj, KOBJ_ADD);
5775 /* Setup first alias */
5776 sysfs_slab_alias(s, s->name);
5783 kobject_del(&s->kobj);
5787 static void sysfs_slab_remove(struct kmem_cache *s)
5789 if (slab_state < FULL)
5791 * Sysfs has not been setup yet so no need to remove the
5796 kobject_get(&s->kobj);
5797 schedule_work(&s->kobj_remove_work);
5800 void sysfs_slab_unlink(struct kmem_cache *s)
5802 if (slab_state >= FULL)
5803 kobject_del(&s->kobj);
5806 void sysfs_slab_release(struct kmem_cache *s)
5808 if (slab_state >= FULL)
5809 kobject_put(&s->kobj);
5813 * Need to buffer aliases during bootup until sysfs becomes
5814 * available lest we lose that information.
5816 struct saved_alias {
5817 struct kmem_cache *s;
5819 struct saved_alias *next;
5822 static struct saved_alias *alias_list;
5824 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5826 struct saved_alias *al;
5828 if (slab_state == FULL) {
5830 * If we have a leftover link then remove it.
5832 sysfs_remove_link(&slab_kset->kobj, name);
5833 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5836 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5842 al->next = alias_list;
5847 static int __init slab_sysfs_init(void)
5849 struct kmem_cache *s;
5852 mutex_lock(&slab_mutex);
5854 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5856 mutex_unlock(&slab_mutex);
5857 pr_err("Cannot register slab subsystem.\n");
5863 list_for_each_entry(s, &slab_caches, list) {
5864 err = sysfs_slab_add(s);
5866 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5870 while (alias_list) {
5871 struct saved_alias *al = alias_list;
5873 alias_list = alias_list->next;
5874 err = sysfs_slab_alias(al->s, al->name);
5876 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5881 mutex_unlock(&slab_mutex);
5886 __initcall(slab_sysfs_init);
5887 #endif /* CONFIG_SYSFS */
5890 * The /proc/slabinfo ABI
5892 #ifdef CONFIG_SLUB_DEBUG
5893 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5895 unsigned long nr_slabs = 0;
5896 unsigned long nr_objs = 0;
5897 unsigned long nr_free = 0;
5899 struct kmem_cache_node *n;
5901 for_each_kmem_cache_node(s, node, n) {
5902 nr_slabs += node_nr_slabs(n);
5903 nr_objs += node_nr_objs(n);
5904 nr_free += count_partial(n, count_free);
5907 sinfo->active_objs = nr_objs - nr_free;
5908 sinfo->num_objs = nr_objs;
5909 sinfo->active_slabs = nr_slabs;
5910 sinfo->num_slabs = nr_slabs;
5911 sinfo->objects_per_slab = oo_objects(s->oo);
5912 sinfo->cache_order = oo_order(s->oo);
5915 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5919 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5920 size_t count, loff_t *ppos)
5924 #endif /* CONFIG_SLUB_DEBUG */