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/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
22 #include <linux/proc_fs.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37 #include <linux/random.h>
39 #include <trace/events/kmem.h>
45 * 1. slab_mutex (Global Mutex)
47 * 3. slab_lock(page) (Only on some arches and for debugging)
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects:
56 * A. page->freelist -> List of object free in a page
57 * B. page->inuse -> Number of objects in use
58 * C. page->objects -> Number of objects in page
59 * D. page->frozen -> frozen state
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list. The processor that froze the slab is the one who can
63 * perform list operations on the page. Other processors may put objects
64 * onto the freelist but the processor that froze the slab is the only
65 * one that can retrieve the objects from the page's freelist.
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * Overloading of page flags that are otherwise used for LRU management.
98 * PageActive The slab is frozen and exempt from list processing.
99 * This means that the slab is dedicated to a purpose
100 * such as satisfying allocations for a specific
101 * processor. Objects may be freed in the slab while
102 * it is frozen but slab_free will then skip the usual
103 * list operations. It is up to the processor holding
104 * the slab to integrate the slab into the slab lists
105 * when the slab is no longer needed.
107 * One use of this flag is to mark slabs that are
108 * used for allocations. Then such a slab becomes a cpu
109 * slab. The cpu slab may be equipped with an additional
110 * freelist that allows lockless access to
111 * free objects in addition to the regular freelist
112 * that requires the slab lock.
114 * PageError Slab requires special handling due to debug
115 * options set. This moves slab handling out of
116 * the fast path and disables lockless freelists.
119 static inline int kmem_cache_debug(struct kmem_cache *s)
121 #ifdef CONFIG_SLUB_DEBUG
122 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
128 void *fixup_red_left(struct kmem_cache *s, void *p)
130 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
131 p += s->red_left_pad;
136 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
138 #ifdef CONFIG_SLUB_CPU_PARTIAL
139 return !kmem_cache_debug(s);
146 * Issues still to be resolved:
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
156 /* Enable to log cmpxchg failures */
157 #undef SLUB_DEBUG_CMPXCHG
160 * Mininum number of partial slabs. These will be left on the partial
161 * lists even if they are empty. kmem_cache_shrink may reclaim them.
163 #define MIN_PARTIAL 5
166 * Maximum number of desirable partial slabs.
167 * The existence of more partial slabs makes kmem_cache_shrink
168 * sort the partial list by the number of objects in use.
170 #define MAX_PARTIAL 10
172 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_STORE_USER)
176 * These debug flags cannot use CMPXCHG because there might be consistency
177 * issues when checking or reading debug information
179 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
184 * Debugging flags that require metadata to be stored in the slab. These get
185 * disabled when slub_debug=O is used and a cache's min order increases with
188 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
191 #define OO_MASK ((1 << OO_SHIFT) - 1)
192 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
194 /* Internal SLUB flags */
196 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
197 /* Use cmpxchg_double */
198 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
201 * Tracking user of a slab.
203 #define TRACK_ADDRS_COUNT 16
205 unsigned long addr; /* Called from address */
206 #ifdef CONFIG_STACKTRACE
207 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
209 int cpu; /* Was running on cpu */
210 int pid; /* Pid context */
211 unsigned long when; /* When did the operation occur */
214 enum track_item { TRACK_ALLOC, TRACK_FREE };
217 static int sysfs_slab_add(struct kmem_cache *);
218 static int sysfs_slab_alias(struct kmem_cache *, const char *);
219 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
220 static void sysfs_slab_remove(struct kmem_cache *s);
222 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
226 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
229 static inline void stat(const struct kmem_cache *s, enum stat_item si)
231 #ifdef CONFIG_SLUB_STATS
233 * The rmw is racy on a preemptible kernel but this is acceptable, so
234 * avoid this_cpu_add()'s irq-disable overhead.
236 raw_cpu_inc(s->cpu_slab->stat[si]);
240 /********************************************************************
241 * Core slab cache functions
242 *******************************************************************/
245 * Returns freelist pointer (ptr). With hardening, this is obfuscated
246 * with an XOR of the address where the pointer is held and a per-cache
249 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
250 unsigned long ptr_addr)
252 #ifdef CONFIG_SLAB_FREELIST_HARDENED
253 return (void *)((unsigned long)ptr ^ s->random ^ swab(ptr_addr));
259 /* Returns the freelist pointer recorded at location ptr_addr. */
260 static inline void *freelist_dereference(const struct kmem_cache *s,
263 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
264 (unsigned long)ptr_addr);
267 static inline void *get_freepointer(struct kmem_cache *s, void *object)
269 return freelist_dereference(s, object + s->offset);
272 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
274 prefetch(object + s->offset);
277 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
279 unsigned long freepointer_addr;
282 if (!debug_pagealloc_enabled())
283 return get_freepointer(s, object);
285 freepointer_addr = (unsigned long)object + s->offset;
286 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
287 return freelist_ptr(s, p, freepointer_addr);
290 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
292 unsigned long freeptr_addr = (unsigned long)object + s->offset;
294 #ifdef CONFIG_SLAB_FREELIST_HARDENED
295 BUG_ON(object == fp); /* naive detection of double free or corruption */
298 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
301 /* Loop over all objects in a slab */
302 #define for_each_object(__p, __s, __addr, __objects) \
303 for (__p = fixup_red_left(__s, __addr); \
304 __p < (__addr) + (__objects) * (__s)->size; \
307 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
308 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
309 __idx <= __objects; \
310 __p += (__s)->size, __idx++)
312 /* Determine object index from a given position */
313 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
315 return (p - addr) / s->size;
318 static inline unsigned int order_objects(unsigned int order, unsigned int size)
320 return ((unsigned int)PAGE_SIZE << order) / size;
323 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
326 struct kmem_cache_order_objects x = {
327 (order << OO_SHIFT) + order_objects(order, size)
333 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
335 return x.x >> OO_SHIFT;
338 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
340 return x.x & OO_MASK;
344 * Per slab locking using the pagelock
346 static __always_inline void slab_lock(struct page *page)
348 VM_BUG_ON_PAGE(PageTail(page), page);
349 bit_spin_lock(PG_locked, &page->flags);
352 static __always_inline void slab_unlock(struct page *page)
354 VM_BUG_ON_PAGE(PageTail(page), page);
355 __bit_spin_unlock(PG_locked, &page->flags);
358 /* Interrupts must be disabled (for the fallback code to work right) */
359 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
360 void *freelist_old, unsigned long counters_old,
361 void *freelist_new, unsigned long counters_new,
364 VM_BUG_ON(!irqs_disabled());
365 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
367 if (s->flags & __CMPXCHG_DOUBLE) {
368 if (cmpxchg_double(&page->freelist, &page->counters,
369 freelist_old, counters_old,
370 freelist_new, counters_new))
376 if (page->freelist == freelist_old &&
377 page->counters == counters_old) {
378 page->freelist = freelist_new;
379 page->counters = counters_new;
387 stat(s, CMPXCHG_DOUBLE_FAIL);
389 #ifdef SLUB_DEBUG_CMPXCHG
390 pr_info("%s %s: cmpxchg double redo ", n, s->name);
396 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
397 void *freelist_old, unsigned long counters_old,
398 void *freelist_new, unsigned long counters_new,
401 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
402 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
403 if (s->flags & __CMPXCHG_DOUBLE) {
404 if (cmpxchg_double(&page->freelist, &page->counters,
405 freelist_old, counters_old,
406 freelist_new, counters_new))
413 local_irq_save(flags);
415 if (page->freelist == freelist_old &&
416 page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
420 local_irq_restore(flags);
424 local_irq_restore(flags);
428 stat(s, CMPXCHG_DOUBLE_FAIL);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 pr_info("%s %s: cmpxchg double redo ", n, s->name);
437 #ifdef CONFIG_SLUB_DEBUG
439 * Determine a map of object in use on a page.
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
444 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
447 void *addr = page_address(page);
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
453 static inline unsigned int size_from_object(struct kmem_cache *s)
455 if (s->flags & SLAB_RED_ZONE)
456 return s->size - s->red_left_pad;
461 static inline void *restore_red_left(struct kmem_cache *s, void *p)
463 if (s->flags & SLAB_RED_ZONE)
464 p -= s->red_left_pad;
472 #if defined(CONFIG_SLUB_DEBUG_ON)
473 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
475 static slab_flags_t slub_debug;
478 static char *slub_debug_slabs;
479 static int disable_higher_order_debug;
482 * slub is about to manipulate internal object metadata. This memory lies
483 * outside the range of the allocated object, so accessing it would normally
484 * be reported by kasan as a bounds error. metadata_access_enable() is used
485 * to tell kasan that these accesses are OK.
487 static inline void metadata_access_enable(void)
489 kasan_disable_current();
492 static inline void metadata_access_disable(void)
494 kasan_enable_current();
501 /* Verify that a pointer has an address that is valid within a slab page */
502 static inline int check_valid_pointer(struct kmem_cache *s,
503 struct page *page, void *object)
510 base = page_address(page);
511 object = restore_red_left(s, object);
512 if (object < base || object >= base + page->objects * s->size ||
513 (object - base) % s->size) {
520 static void print_section(char *level, char *text, u8 *addr,
523 metadata_access_enable();
524 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
526 metadata_access_disable();
529 static struct track *get_track(struct kmem_cache *s, void *object,
530 enum track_item alloc)
535 p = object + s->offset + sizeof(void *);
537 p = object + s->inuse;
542 static void set_track(struct kmem_cache *s, void *object,
543 enum track_item alloc, unsigned long addr)
545 struct track *p = get_track(s, object, alloc);
548 #ifdef CONFIG_STACKTRACE
549 struct stack_trace trace;
552 trace.nr_entries = 0;
553 trace.max_entries = TRACK_ADDRS_COUNT;
554 trace.entries = p->addrs;
556 metadata_access_enable();
557 save_stack_trace(&trace);
558 metadata_access_disable();
560 /* See rant in lockdep.c */
561 if (trace.nr_entries != 0 &&
562 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
565 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
569 p->cpu = smp_processor_id();
570 p->pid = current->pid;
573 memset(p, 0, sizeof(struct track));
576 static void init_tracking(struct kmem_cache *s, void *object)
578 if (!(s->flags & SLAB_STORE_USER))
581 set_track(s, object, TRACK_FREE, 0UL);
582 set_track(s, object, TRACK_ALLOC, 0UL);
585 static void print_track(const char *s, struct track *t, unsigned long pr_time)
590 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
591 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
592 #ifdef CONFIG_STACKTRACE
595 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
597 pr_err("\t%pS\n", (void *)t->addrs[i]);
604 static void print_tracking(struct kmem_cache *s, void *object)
606 unsigned long pr_time = jiffies;
607 if (!(s->flags & SLAB_STORE_USER))
610 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
611 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
614 static void print_page_info(struct page *page)
616 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
617 page, page->objects, page->inuse, page->freelist, page->flags);
621 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
623 struct va_format vaf;
629 pr_err("=============================================================================\n");
630 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
631 pr_err("-----------------------------------------------------------------------------\n\n");
633 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
637 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
639 struct va_format vaf;
645 pr_err("FIX %s: %pV\n", s->name, &vaf);
649 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
650 void **freelist, void *nextfree)
652 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
653 !check_valid_pointer(s, page, nextfree) && freelist) {
654 object_err(s, page, *freelist, "Freechain corrupt");
656 slab_fix(s, "Isolate corrupted freechain");
663 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
665 unsigned int off; /* Offset of last byte */
666 u8 *addr = page_address(page);
668 print_tracking(s, p);
670 print_page_info(page);
672 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
673 p, p - addr, get_freepointer(s, p));
675 if (s->flags & SLAB_RED_ZONE)
676 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
678 else if (p > addr + 16)
679 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
681 print_section(KERN_ERR, "Object ", p,
682 min_t(unsigned int, s->object_size, PAGE_SIZE));
683 if (s->flags & SLAB_RED_ZONE)
684 print_section(KERN_ERR, "Redzone ", p + s->object_size,
685 s->inuse - s->object_size);
688 off = s->offset + sizeof(void *);
692 if (s->flags & SLAB_STORE_USER)
693 off += 2 * sizeof(struct track);
695 off += kasan_metadata_size(s);
697 if (off != size_from_object(s))
698 /* Beginning of the filler is the free pointer */
699 print_section(KERN_ERR, "Padding ", p + off,
700 size_from_object(s) - off);
705 void object_err(struct kmem_cache *s, struct page *page,
706 u8 *object, char *reason)
708 slab_bug(s, "%s", reason);
709 print_trailer(s, page, object);
712 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
713 const char *fmt, ...)
719 vsnprintf(buf, sizeof(buf), fmt, args);
721 slab_bug(s, "%s", buf);
722 print_page_info(page);
726 static void init_object(struct kmem_cache *s, void *object, u8 val)
730 if (s->flags & SLAB_RED_ZONE)
731 memset(p - s->red_left_pad, val, s->red_left_pad);
733 if (s->flags & __OBJECT_POISON) {
734 memset(p, POISON_FREE, s->object_size - 1);
735 p[s->object_size - 1] = POISON_END;
738 if (s->flags & SLAB_RED_ZONE)
739 memset(p + s->object_size, val, s->inuse - s->object_size);
742 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
743 void *from, void *to)
745 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
746 memset(from, data, to - from);
749 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
750 u8 *object, char *what,
751 u8 *start, unsigned int value, unsigned int bytes)
756 metadata_access_enable();
757 fault = memchr_inv(start, value, bytes);
758 metadata_access_disable();
763 while (end > fault && end[-1] == value)
766 slab_bug(s, "%s overwritten", what);
767 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
768 fault, end - 1, fault[0], value);
769 print_trailer(s, page, object);
771 restore_bytes(s, what, value, fault, end);
779 * Bytes of the object to be managed.
780 * If the freepointer may overlay the object then the free
781 * pointer is the first word of the object.
783 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
786 * object + s->object_size
787 * Padding to reach word boundary. This is also used for Redzoning.
788 * Padding is extended by another word if Redzoning is enabled and
789 * object_size == inuse.
791 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
792 * 0xcc (RED_ACTIVE) for objects in use.
795 * Meta data starts here.
797 * A. Free pointer (if we cannot overwrite object on free)
798 * B. Tracking data for SLAB_STORE_USER
799 * C. Padding to reach required alignment boundary or at mininum
800 * one word if debugging is on to be able to detect writes
801 * before the word boundary.
803 * Padding is done using 0x5a (POISON_INUSE)
806 * Nothing is used beyond s->size.
808 * If slabcaches are merged then the object_size and inuse boundaries are mostly
809 * ignored. And therefore no slab options that rely on these boundaries
810 * may be used with merged slabcaches.
813 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
815 unsigned long off = s->inuse; /* The end of info */
818 /* Freepointer is placed after the object. */
819 off += sizeof(void *);
821 if (s->flags & SLAB_STORE_USER)
822 /* We also have user information there */
823 off += 2 * sizeof(struct track);
825 off += kasan_metadata_size(s);
827 if (size_from_object(s) == off)
830 return check_bytes_and_report(s, page, p, "Object padding",
831 p + off, POISON_INUSE, size_from_object(s) - off);
834 /* Check the pad bytes at the end of a slab page */
835 static int slab_pad_check(struct kmem_cache *s, struct page *page)
844 if (!(s->flags & SLAB_POISON))
847 start = page_address(page);
848 length = PAGE_SIZE << compound_order(page);
849 end = start + length;
850 remainder = length % s->size;
854 pad = end - remainder;
855 metadata_access_enable();
856 fault = memchr_inv(pad, POISON_INUSE, remainder);
857 metadata_access_disable();
860 while (end > fault && end[-1] == POISON_INUSE)
863 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
864 print_section(KERN_ERR, "Padding ", pad, remainder);
866 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
870 static int check_object(struct kmem_cache *s, struct page *page,
871 void *object, u8 val)
874 u8 *endobject = object + s->object_size;
876 if (s->flags & SLAB_RED_ZONE) {
877 if (!check_bytes_and_report(s, page, object, "Left Redzone",
878 object - s->red_left_pad, val, s->red_left_pad))
881 if (!check_bytes_and_report(s, page, object, "Right Redzone",
882 endobject, val, s->inuse - s->object_size))
885 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
886 check_bytes_and_report(s, page, p, "Alignment padding",
887 endobject, POISON_INUSE,
888 s->inuse - s->object_size);
892 if (s->flags & SLAB_POISON) {
893 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
894 (!check_bytes_and_report(s, page, p, "Poison", p,
895 POISON_FREE, s->object_size - 1) ||
896 !check_bytes_and_report(s, page, p, "End Poison",
897 p + s->object_size - 1, POISON_END, 1)))
900 * check_pad_bytes cleans up on its own.
902 check_pad_bytes(s, page, p);
905 if (!s->offset && val == SLUB_RED_ACTIVE)
907 * Object and freepointer overlap. Cannot check
908 * freepointer while object is allocated.
912 /* Check free pointer validity */
913 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
914 object_err(s, page, p, "Freepointer corrupt");
916 * No choice but to zap it and thus lose the remainder
917 * of the free objects in this slab. May cause
918 * another error because the object count is now wrong.
920 set_freepointer(s, p, NULL);
926 static int check_slab(struct kmem_cache *s, struct page *page)
930 VM_BUG_ON(!irqs_disabled());
932 if (!PageSlab(page)) {
933 slab_err(s, page, "Not a valid slab page");
937 maxobj = order_objects(compound_order(page), s->size);
938 if (page->objects > maxobj) {
939 slab_err(s, page, "objects %u > max %u",
940 page->objects, maxobj);
943 if (page->inuse > page->objects) {
944 slab_err(s, page, "inuse %u > max %u",
945 page->inuse, page->objects);
948 /* Slab_pad_check fixes things up after itself */
949 slab_pad_check(s, page);
954 * Determine if a certain object on a page is on the freelist. Must hold the
955 * slab lock to guarantee that the chains are in a consistent state.
957 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
965 while (fp && nr <= page->objects) {
968 if (!check_valid_pointer(s, page, fp)) {
970 object_err(s, page, object,
971 "Freechain corrupt");
972 set_freepointer(s, object, NULL);
974 slab_err(s, page, "Freepointer corrupt");
975 page->freelist = NULL;
976 page->inuse = page->objects;
977 slab_fix(s, "Freelist cleared");
983 fp = get_freepointer(s, object);
987 max_objects = order_objects(compound_order(page), s->size);
988 if (max_objects > MAX_OBJS_PER_PAGE)
989 max_objects = MAX_OBJS_PER_PAGE;
991 if (page->objects != max_objects) {
992 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
993 page->objects, max_objects);
994 page->objects = max_objects;
995 slab_fix(s, "Number of objects adjusted.");
997 if (page->inuse != page->objects - nr) {
998 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
999 page->inuse, page->objects - nr);
1000 page->inuse = page->objects - nr;
1001 slab_fix(s, "Object count adjusted.");
1003 return search == NULL;
1006 static void trace(struct kmem_cache *s, struct page *page, void *object,
1009 if (s->flags & SLAB_TRACE) {
1010 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1012 alloc ? "alloc" : "free",
1013 object, page->inuse,
1017 print_section(KERN_INFO, "Object ", (void *)object,
1025 * Tracking of fully allocated slabs for debugging purposes.
1027 static void add_full(struct kmem_cache *s,
1028 struct kmem_cache_node *n, struct page *page)
1030 if (!(s->flags & SLAB_STORE_USER))
1033 lockdep_assert_held(&n->list_lock);
1034 list_add(&page->lru, &n->full);
1037 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1039 if (!(s->flags & SLAB_STORE_USER))
1042 lockdep_assert_held(&n->list_lock);
1043 list_del(&page->lru);
1046 /* Tracking of the number of slabs for debugging purposes */
1047 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1049 struct kmem_cache_node *n = get_node(s, node);
1051 return atomic_long_read(&n->nr_slabs);
1054 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1056 return atomic_long_read(&n->nr_slabs);
1059 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1061 struct kmem_cache_node *n = get_node(s, node);
1064 * May be called early in order to allocate a slab for the
1065 * kmem_cache_node structure. Solve the chicken-egg
1066 * dilemma by deferring the increment of the count during
1067 * bootstrap (see early_kmem_cache_node_alloc).
1070 atomic_long_inc(&n->nr_slabs);
1071 atomic_long_add(objects, &n->total_objects);
1074 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1076 struct kmem_cache_node *n = get_node(s, node);
1078 atomic_long_dec(&n->nr_slabs);
1079 atomic_long_sub(objects, &n->total_objects);
1082 /* Object debug checks for alloc/free paths */
1083 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1086 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1089 init_object(s, object, SLUB_RED_INACTIVE);
1090 init_tracking(s, object);
1093 static inline int alloc_consistency_checks(struct kmem_cache *s,
1095 void *object, unsigned long addr)
1097 if (!check_slab(s, page))
1100 if (!check_valid_pointer(s, page, object)) {
1101 object_err(s, page, object, "Freelist Pointer check fails");
1105 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1111 static noinline int alloc_debug_processing(struct kmem_cache *s,
1113 void *object, unsigned long addr)
1115 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1116 if (!alloc_consistency_checks(s, page, object, addr))
1120 /* Success perform special debug activities for allocs */
1121 if (s->flags & SLAB_STORE_USER)
1122 set_track(s, object, TRACK_ALLOC, addr);
1123 trace(s, page, object, 1);
1124 init_object(s, object, SLUB_RED_ACTIVE);
1128 if (PageSlab(page)) {
1130 * If this is a slab page then lets do the best we can
1131 * to avoid issues in the future. Marking all objects
1132 * as used avoids touching the remaining objects.
1134 slab_fix(s, "Marking all objects used");
1135 page->inuse = page->objects;
1136 page->freelist = NULL;
1141 static inline int free_consistency_checks(struct kmem_cache *s,
1142 struct page *page, void *object, unsigned long addr)
1144 if (!check_valid_pointer(s, page, object)) {
1145 slab_err(s, page, "Invalid object pointer 0x%p", object);
1149 if (on_freelist(s, page, object)) {
1150 object_err(s, page, object, "Object already free");
1154 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1157 if (unlikely(s != page->slab_cache)) {
1158 if (!PageSlab(page)) {
1159 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1161 } else if (!page->slab_cache) {
1162 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1166 object_err(s, page, object,
1167 "page slab pointer corrupt.");
1173 /* Supports checking bulk free of a constructed freelist */
1174 static noinline int free_debug_processing(
1175 struct kmem_cache *s, struct page *page,
1176 void *head, void *tail, int bulk_cnt,
1179 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1180 void *object = head;
1182 unsigned long uninitialized_var(flags);
1185 spin_lock_irqsave(&n->list_lock, flags);
1188 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1189 if (!check_slab(s, page))
1196 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1197 if (!free_consistency_checks(s, page, object, addr))
1201 if (s->flags & SLAB_STORE_USER)
1202 set_track(s, object, TRACK_FREE, addr);
1203 trace(s, page, object, 0);
1204 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1205 init_object(s, object, SLUB_RED_INACTIVE);
1207 /* Reached end of constructed freelist yet? */
1208 if (object != tail) {
1209 object = get_freepointer(s, object);
1215 if (cnt != bulk_cnt)
1216 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1220 spin_unlock_irqrestore(&n->list_lock, flags);
1222 slab_fix(s, "Object at 0x%p not freed", object);
1226 static int __init setup_slub_debug(char *str)
1228 slub_debug = DEBUG_DEFAULT_FLAGS;
1229 if (*str++ != '=' || !*str)
1231 * No options specified. Switch on full debugging.
1237 * No options but restriction on slabs. This means full
1238 * debugging for slabs matching a pattern.
1245 * Switch off all debugging measures.
1250 * Determine which debug features should be switched on
1252 for (; *str && *str != ','; str++) {
1253 switch (tolower(*str)) {
1255 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1258 slub_debug |= SLAB_RED_ZONE;
1261 slub_debug |= SLAB_POISON;
1264 slub_debug |= SLAB_STORE_USER;
1267 slub_debug |= SLAB_TRACE;
1270 slub_debug |= SLAB_FAILSLAB;
1274 * Avoid enabling debugging on caches if its minimum
1275 * order would increase as a result.
1277 disable_higher_order_debug = 1;
1280 pr_err("slub_debug option '%c' unknown. skipped\n",
1287 slub_debug_slabs = str + 1;
1292 __setup("slub_debug", setup_slub_debug);
1294 slab_flags_t kmem_cache_flags(unsigned int object_size,
1295 slab_flags_t flags, const char *name,
1296 void (*ctor)(void *))
1299 * Enable debugging if selected on the kernel commandline.
1301 if (slub_debug && (!slub_debug_slabs || (name &&
1302 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1303 flags |= slub_debug;
1307 #else /* !CONFIG_SLUB_DEBUG */
1308 static inline void setup_object_debug(struct kmem_cache *s,
1309 struct page *page, void *object) {}
1311 static inline int alloc_debug_processing(struct kmem_cache *s,
1312 struct page *page, void *object, unsigned long addr) { return 0; }
1314 static inline int free_debug_processing(
1315 struct kmem_cache *s, struct page *page,
1316 void *head, void *tail, int bulk_cnt,
1317 unsigned long addr) { return 0; }
1319 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1321 static inline int check_object(struct kmem_cache *s, struct page *page,
1322 void *object, u8 val) { return 1; }
1323 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1324 struct page *page) {}
1325 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1326 struct page *page) {}
1327 slab_flags_t kmem_cache_flags(unsigned int object_size,
1328 slab_flags_t flags, const char *name,
1329 void (*ctor)(void *))
1333 #define slub_debug 0
1335 #define disable_higher_order_debug 0
1337 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1339 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1341 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1343 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1346 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1347 void **freelist, void *nextfree)
1351 #endif /* CONFIG_SLUB_DEBUG */
1354 * Hooks for other subsystems that check memory allocations. In a typical
1355 * production configuration these hooks all should produce no code at all.
1357 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1359 kmemleak_alloc(ptr, size, 1, flags);
1360 kasan_kmalloc_large(ptr, size, flags);
1363 static __always_inline void kfree_hook(void *x)
1366 kasan_kfree_large(x, _RET_IP_);
1369 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1371 kmemleak_free_recursive(x, s->flags);
1374 * Trouble is that we may no longer disable interrupts in the fast path
1375 * So in order to make the debug calls that expect irqs to be
1376 * disabled we need to disable interrupts temporarily.
1378 #ifdef CONFIG_LOCKDEP
1380 unsigned long flags;
1382 local_irq_save(flags);
1383 debug_check_no_locks_freed(x, s->object_size);
1384 local_irq_restore(flags);
1387 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1388 debug_check_no_obj_freed(x, s->object_size);
1390 /* KASAN might put x into memory quarantine, delaying its reuse */
1391 return kasan_slab_free(s, x, _RET_IP_);
1394 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1395 void **head, void **tail,
1399 * Compiler cannot detect this function can be removed if slab_free_hook()
1400 * evaluates to nothing. Thus, catch all relevant config debug options here.
1402 #if defined(CONFIG_LOCKDEP) || \
1403 defined(CONFIG_DEBUG_KMEMLEAK) || \
1404 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1405 defined(CONFIG_KASAN)
1409 void *old_tail = *tail ? *tail : *head;
1411 /* Head and tail of the reconstructed freelist */
1417 next = get_freepointer(s, object);
1418 /* If object's reuse doesn't have to be delayed */
1419 if (!slab_free_hook(s, object)) {
1420 /* Move object to the new freelist */
1421 set_freepointer(s, object, *head);
1427 * Adjust the reconstructed freelist depth
1428 * accordingly if object's reuse is delayed.
1432 } while (object != old_tail);
1437 return *head != NULL;
1443 static void setup_object(struct kmem_cache *s, struct page *page,
1446 setup_object_debug(s, page, object);
1447 kasan_init_slab_obj(s, object);
1448 if (unlikely(s->ctor)) {
1449 kasan_unpoison_object_data(s, object);
1451 kasan_poison_object_data(s, object);
1456 * Slab allocation and freeing
1458 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1459 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1462 unsigned int order = oo_order(oo);
1464 if (node == NUMA_NO_NODE)
1465 page = alloc_pages(flags, order);
1467 page = __alloc_pages_node(node, flags, order);
1469 if (page && memcg_charge_slab(page, flags, order, s)) {
1470 __free_pages(page, order);
1477 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1478 /* Pre-initialize the random sequence cache */
1479 static int init_cache_random_seq(struct kmem_cache *s)
1481 unsigned int count = oo_objects(s->oo);
1484 /* Bailout if already initialised */
1488 err = cache_random_seq_create(s, count, GFP_KERNEL);
1490 pr_err("SLUB: Unable to initialize free list for %s\n",
1495 /* Transform to an offset on the set of pages */
1496 if (s->random_seq) {
1499 for (i = 0; i < count; i++)
1500 s->random_seq[i] *= s->size;
1505 /* Initialize each random sequence freelist per cache */
1506 static void __init init_freelist_randomization(void)
1508 struct kmem_cache *s;
1510 mutex_lock(&slab_mutex);
1512 list_for_each_entry(s, &slab_caches, list)
1513 init_cache_random_seq(s);
1515 mutex_unlock(&slab_mutex);
1518 /* Get the next entry on the pre-computed freelist randomized */
1519 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1520 unsigned long *pos, void *start,
1521 unsigned long page_limit,
1522 unsigned long freelist_count)
1527 * If the target page allocation failed, the number of objects on the
1528 * page might be smaller than the usual size defined by the cache.
1531 idx = s->random_seq[*pos];
1533 if (*pos >= freelist_count)
1535 } while (unlikely(idx >= page_limit));
1537 return (char *)start + idx;
1540 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1541 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1546 unsigned long idx, pos, page_limit, freelist_count;
1548 if (page->objects < 2 || !s->random_seq)
1551 freelist_count = oo_objects(s->oo);
1552 pos = get_random_int() % freelist_count;
1554 page_limit = page->objects * s->size;
1555 start = fixup_red_left(s, page_address(page));
1557 /* First entry is used as the base of the freelist */
1558 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1560 page->freelist = cur;
1562 for (idx = 1; idx < page->objects; idx++) {
1563 setup_object(s, page, cur);
1564 next = next_freelist_entry(s, page, &pos, start, page_limit,
1566 set_freepointer(s, cur, next);
1569 setup_object(s, page, cur);
1570 set_freepointer(s, cur, NULL);
1575 static inline int init_cache_random_seq(struct kmem_cache *s)
1579 static inline void init_freelist_randomization(void) { }
1580 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1584 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1586 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1589 struct kmem_cache_order_objects oo = s->oo;
1595 flags &= gfp_allowed_mask;
1597 if (gfpflags_allow_blocking(flags))
1600 flags |= s->allocflags;
1603 * Let the initial higher-order allocation fail under memory pressure
1604 * so we fall-back to the minimum order allocation.
1606 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1607 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1608 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1610 page = alloc_slab_page(s, alloc_gfp, node, oo);
1611 if (unlikely(!page)) {
1615 * Allocation may have failed due to fragmentation.
1616 * Try a lower order alloc if possible
1618 page = alloc_slab_page(s, alloc_gfp, node, oo);
1619 if (unlikely(!page))
1621 stat(s, ORDER_FALLBACK);
1624 page->objects = oo_objects(oo);
1626 order = compound_order(page);
1627 page->slab_cache = s;
1628 __SetPageSlab(page);
1629 if (page_is_pfmemalloc(page))
1630 SetPageSlabPfmemalloc(page);
1632 start = page_address(page);
1634 if (unlikely(s->flags & SLAB_POISON))
1635 memset(start, POISON_INUSE, PAGE_SIZE << order);
1637 kasan_poison_slab(page);
1639 shuffle = shuffle_freelist(s, page);
1642 for_each_object_idx(p, idx, s, start, page->objects) {
1643 setup_object(s, page, p);
1644 if (likely(idx < page->objects))
1645 set_freepointer(s, p, p + s->size);
1647 set_freepointer(s, p, NULL);
1649 page->freelist = fixup_red_left(s, start);
1652 page->inuse = page->objects;
1656 if (gfpflags_allow_blocking(flags))
1657 local_irq_disable();
1661 mod_lruvec_page_state(page,
1662 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1663 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1666 inc_slabs_node(s, page_to_nid(page), page->objects);
1671 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1673 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1674 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1675 flags &= ~GFP_SLAB_BUG_MASK;
1676 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1677 invalid_mask, &invalid_mask, flags, &flags);
1681 return allocate_slab(s,
1682 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1685 static void __free_slab(struct kmem_cache *s, struct page *page)
1687 int order = compound_order(page);
1688 int pages = 1 << order;
1690 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1693 slab_pad_check(s, page);
1694 for_each_object(p, s, page_address(page),
1696 check_object(s, page, p, SLUB_RED_INACTIVE);
1699 mod_lruvec_page_state(page,
1700 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1701 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1704 __ClearPageSlabPfmemalloc(page);
1705 __ClearPageSlab(page);
1707 page->mapping = NULL;
1708 if (current->reclaim_state)
1709 current->reclaim_state->reclaimed_slab += pages;
1710 memcg_uncharge_slab(page, order, s);
1711 __free_pages(page, order);
1714 static void rcu_free_slab(struct rcu_head *h)
1716 struct page *page = container_of(h, struct page, rcu_head);
1718 __free_slab(page->slab_cache, page);
1721 static void free_slab(struct kmem_cache *s, struct page *page)
1723 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1724 call_rcu(&page->rcu_head, rcu_free_slab);
1726 __free_slab(s, page);
1729 static void discard_slab(struct kmem_cache *s, struct page *page)
1731 dec_slabs_node(s, page_to_nid(page), page->objects);
1736 * Management of partially allocated slabs.
1739 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1742 if (tail == DEACTIVATE_TO_TAIL)
1743 list_add_tail(&page->lru, &n->partial);
1745 list_add(&page->lru, &n->partial);
1748 static inline void add_partial(struct kmem_cache_node *n,
1749 struct page *page, int tail)
1751 lockdep_assert_held(&n->list_lock);
1752 __add_partial(n, page, tail);
1755 static inline void remove_partial(struct kmem_cache_node *n,
1758 lockdep_assert_held(&n->list_lock);
1759 list_del(&page->lru);
1764 * Remove slab from the partial list, freeze it and
1765 * return the pointer to the freelist.
1767 * Returns a list of objects or NULL if it fails.
1769 static inline void *acquire_slab(struct kmem_cache *s,
1770 struct kmem_cache_node *n, struct page *page,
1771 int mode, int *objects)
1774 unsigned long counters;
1777 lockdep_assert_held(&n->list_lock);
1780 * Zap the freelist and set the frozen bit.
1781 * The old freelist is the list of objects for the
1782 * per cpu allocation list.
1784 freelist = page->freelist;
1785 counters = page->counters;
1786 new.counters = counters;
1787 *objects = new.objects - new.inuse;
1789 new.inuse = page->objects;
1790 new.freelist = NULL;
1792 new.freelist = freelist;
1795 VM_BUG_ON(new.frozen);
1798 if (!__cmpxchg_double_slab(s, page,
1800 new.freelist, new.counters,
1804 remove_partial(n, page);
1809 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1810 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1813 * Try to allocate a partial slab from a specific node.
1815 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1816 struct kmem_cache_cpu *c, gfp_t flags)
1818 struct page *page, *page2;
1819 void *object = NULL;
1820 unsigned int available = 0;
1824 * Racy check. If we mistakenly see no partial slabs then we
1825 * just allocate an empty slab. If we mistakenly try to get a
1826 * partial slab and there is none available then get_partials()
1829 if (!n || !n->nr_partial)
1832 spin_lock(&n->list_lock);
1833 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1836 if (!pfmemalloc_match(page, flags))
1839 t = acquire_slab(s, n, page, object == NULL, &objects);
1843 available += objects;
1846 stat(s, ALLOC_FROM_PARTIAL);
1849 put_cpu_partial(s, page, 0);
1850 stat(s, CPU_PARTIAL_NODE);
1852 if (!kmem_cache_has_cpu_partial(s)
1853 || available > slub_cpu_partial(s) / 2)
1857 spin_unlock(&n->list_lock);
1862 * Get a page from somewhere. Search in increasing NUMA distances.
1864 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1865 struct kmem_cache_cpu *c)
1868 struct zonelist *zonelist;
1871 enum zone_type high_zoneidx = gfp_zone(flags);
1873 unsigned int cpuset_mems_cookie;
1876 * The defrag ratio allows a configuration of the tradeoffs between
1877 * inter node defragmentation and node local allocations. A lower
1878 * defrag_ratio increases the tendency to do local allocations
1879 * instead of attempting to obtain partial slabs from other nodes.
1881 * If the defrag_ratio is set to 0 then kmalloc() always
1882 * returns node local objects. If the ratio is higher then kmalloc()
1883 * may return off node objects because partial slabs are obtained
1884 * from other nodes and filled up.
1886 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1887 * (which makes defrag_ratio = 1000) then every (well almost)
1888 * allocation will first attempt to defrag slab caches on other nodes.
1889 * This means scanning over all nodes to look for partial slabs which
1890 * may be expensive if we do it every time we are trying to find a slab
1891 * with available objects.
1893 if (!s->remote_node_defrag_ratio ||
1894 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1898 cpuset_mems_cookie = read_mems_allowed_begin();
1899 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1900 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1901 struct kmem_cache_node *n;
1903 n = get_node(s, zone_to_nid(zone));
1905 if (n && cpuset_zone_allowed(zone, flags) &&
1906 n->nr_partial > s->min_partial) {
1907 object = get_partial_node(s, n, c, flags);
1910 * Don't check read_mems_allowed_retry()
1911 * here - if mems_allowed was updated in
1912 * parallel, that was a harmless race
1913 * between allocation and the cpuset
1920 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1926 * Get a partial page, lock it and return it.
1928 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1929 struct kmem_cache_cpu *c)
1932 int searchnode = node;
1934 if (node == NUMA_NO_NODE)
1935 searchnode = numa_mem_id();
1937 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1938 if (object || node != NUMA_NO_NODE)
1941 return get_any_partial(s, flags, c);
1944 #ifdef CONFIG_PREEMPT
1946 * Calculate the next globally unique transaction for disambiguiation
1947 * during cmpxchg. The transactions start with the cpu number and are then
1948 * incremented by CONFIG_NR_CPUS.
1950 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1953 * No preemption supported therefore also no need to check for
1959 static inline unsigned long next_tid(unsigned long tid)
1961 return tid + TID_STEP;
1964 static inline unsigned int tid_to_cpu(unsigned long tid)
1966 return tid % TID_STEP;
1969 static inline unsigned long tid_to_event(unsigned long tid)
1971 return tid / TID_STEP;
1974 static inline unsigned int init_tid(int cpu)
1979 static inline void note_cmpxchg_failure(const char *n,
1980 const struct kmem_cache *s, unsigned long tid)
1982 #ifdef SLUB_DEBUG_CMPXCHG
1983 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1985 pr_info("%s %s: cmpxchg redo ", n, s->name);
1987 #ifdef CONFIG_PREEMPT
1988 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1989 pr_warn("due to cpu change %d -> %d\n",
1990 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1993 if (tid_to_event(tid) != tid_to_event(actual_tid))
1994 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1995 tid_to_event(tid), tid_to_event(actual_tid));
1997 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1998 actual_tid, tid, next_tid(tid));
2000 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2003 static void init_kmem_cache_cpus(struct kmem_cache *s)
2007 for_each_possible_cpu(cpu)
2008 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2012 * Remove the cpu slab
2014 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2015 void *freelist, struct kmem_cache_cpu *c)
2017 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2018 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2020 enum slab_modes l = M_NONE, m = M_NONE;
2022 int tail = DEACTIVATE_TO_HEAD;
2026 if (page->freelist) {
2027 stat(s, DEACTIVATE_REMOTE_FREES);
2028 tail = DEACTIVATE_TO_TAIL;
2032 * Stage one: Free all available per cpu objects back
2033 * to the page freelist while it is still frozen. Leave the
2036 * There is no need to take the list->lock because the page
2039 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2041 unsigned long counters;
2044 * If 'nextfree' is invalid, it is possible that the object at
2045 * 'freelist' is already corrupted. So isolate all objects
2046 * starting at 'freelist'.
2048 if (freelist_corrupted(s, page, &freelist, nextfree))
2052 prior = page->freelist;
2053 counters = page->counters;
2054 set_freepointer(s, freelist, prior);
2055 new.counters = counters;
2057 VM_BUG_ON(!new.frozen);
2059 } while (!__cmpxchg_double_slab(s, page,
2061 freelist, new.counters,
2062 "drain percpu freelist"));
2064 freelist = nextfree;
2068 * Stage two: Ensure that the page is unfrozen while the
2069 * list presence reflects the actual number of objects
2072 * We setup the list membership and then perform a cmpxchg
2073 * with the count. If there is a mismatch then the page
2074 * is not unfrozen but the page is on the wrong list.
2076 * Then we restart the process which may have to remove
2077 * the page from the list that we just put it on again
2078 * because the number of objects in the slab may have
2083 old.freelist = page->freelist;
2084 old.counters = page->counters;
2085 VM_BUG_ON(!old.frozen);
2087 /* Determine target state of the slab */
2088 new.counters = old.counters;
2091 set_freepointer(s, freelist, old.freelist);
2092 new.freelist = freelist;
2094 new.freelist = old.freelist;
2098 if (!new.inuse && n->nr_partial >= s->min_partial)
2100 else if (new.freelist) {
2105 * Taking the spinlock removes the possiblity
2106 * that acquire_slab() will see a slab page that
2109 spin_lock(&n->list_lock);
2113 if (kmem_cache_debug(s) && !lock) {
2116 * This also ensures that the scanning of full
2117 * slabs from diagnostic functions will not see
2120 spin_lock(&n->list_lock);
2128 remove_partial(n, page);
2130 else if (l == M_FULL)
2132 remove_full(s, n, page);
2134 if (m == M_PARTIAL) {
2136 add_partial(n, page, tail);
2139 } else if (m == M_FULL) {
2141 stat(s, DEACTIVATE_FULL);
2142 add_full(s, n, page);
2148 if (!__cmpxchg_double_slab(s, page,
2149 old.freelist, old.counters,
2150 new.freelist, new.counters,
2155 spin_unlock(&n->list_lock);
2158 stat(s, DEACTIVATE_EMPTY);
2159 discard_slab(s, page);
2168 * Unfreeze all the cpu partial slabs.
2170 * This function must be called with interrupts disabled
2171 * for the cpu using c (or some other guarantee must be there
2172 * to guarantee no concurrent accesses).
2174 static void unfreeze_partials(struct kmem_cache *s,
2175 struct kmem_cache_cpu *c)
2177 #ifdef CONFIG_SLUB_CPU_PARTIAL
2178 struct kmem_cache_node *n = NULL, *n2 = NULL;
2179 struct page *page, *discard_page = NULL;
2181 while ((page = c->partial)) {
2185 c->partial = page->next;
2187 n2 = get_node(s, page_to_nid(page));
2190 spin_unlock(&n->list_lock);
2193 spin_lock(&n->list_lock);
2198 old.freelist = page->freelist;
2199 old.counters = page->counters;
2200 VM_BUG_ON(!old.frozen);
2202 new.counters = old.counters;
2203 new.freelist = old.freelist;
2207 } while (!__cmpxchg_double_slab(s, page,
2208 old.freelist, old.counters,
2209 new.freelist, new.counters,
2210 "unfreezing slab"));
2212 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2213 page->next = discard_page;
2214 discard_page = page;
2216 add_partial(n, page, DEACTIVATE_TO_TAIL);
2217 stat(s, FREE_ADD_PARTIAL);
2222 spin_unlock(&n->list_lock);
2224 while (discard_page) {
2225 page = discard_page;
2226 discard_page = discard_page->next;
2228 stat(s, DEACTIVATE_EMPTY);
2229 discard_slab(s, page);
2236 * Put a page that was just frozen (in __slab_free) into a partial page
2237 * slot if available.
2239 * If we did not find a slot then simply move all the partials to the
2240 * per node partial list.
2242 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2244 #ifdef CONFIG_SLUB_CPU_PARTIAL
2245 struct page *oldpage;
2253 oldpage = this_cpu_read(s->cpu_slab->partial);
2256 pobjects = oldpage->pobjects;
2257 pages = oldpage->pages;
2258 if (drain && pobjects > s->cpu_partial) {
2259 unsigned long flags;
2261 * partial array is full. Move the existing
2262 * set to the per node partial list.
2264 local_irq_save(flags);
2265 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2266 local_irq_restore(flags);
2270 stat(s, CPU_PARTIAL_DRAIN);
2275 pobjects += page->objects - page->inuse;
2277 page->pages = pages;
2278 page->pobjects = pobjects;
2279 page->next = oldpage;
2281 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2283 if (unlikely(!s->cpu_partial)) {
2284 unsigned long flags;
2286 local_irq_save(flags);
2287 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2288 local_irq_restore(flags);
2294 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2296 stat(s, CPUSLAB_FLUSH);
2297 deactivate_slab(s, c->page, c->freelist, c);
2299 c->tid = next_tid(c->tid);
2305 * Called from IPI handler with interrupts disabled.
2307 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2309 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2315 unfreeze_partials(s, c);
2319 static void flush_cpu_slab(void *d)
2321 struct kmem_cache *s = d;
2323 __flush_cpu_slab(s, smp_processor_id());
2326 static bool has_cpu_slab(int cpu, void *info)
2328 struct kmem_cache *s = info;
2329 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2331 return c->page || slub_percpu_partial(c);
2334 static void flush_all(struct kmem_cache *s)
2336 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2340 * Use the cpu notifier to insure that the cpu slabs are flushed when
2343 static int slub_cpu_dead(unsigned int cpu)
2345 struct kmem_cache *s;
2346 unsigned long flags;
2348 mutex_lock(&slab_mutex);
2349 list_for_each_entry(s, &slab_caches, list) {
2350 local_irq_save(flags);
2351 __flush_cpu_slab(s, cpu);
2352 local_irq_restore(flags);
2354 mutex_unlock(&slab_mutex);
2359 * Check if the objects in a per cpu structure fit numa
2360 * locality expectations.
2362 static inline int node_match(struct page *page, int node)
2365 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2371 #ifdef CONFIG_SLUB_DEBUG
2372 static int count_free(struct page *page)
2374 return page->objects - page->inuse;
2377 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2379 return atomic_long_read(&n->total_objects);
2381 #endif /* CONFIG_SLUB_DEBUG */
2383 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2384 static unsigned long count_partial(struct kmem_cache_node *n,
2385 int (*get_count)(struct page *))
2387 unsigned long flags;
2388 unsigned long x = 0;
2391 spin_lock_irqsave(&n->list_lock, flags);
2392 list_for_each_entry(page, &n->partial, lru)
2393 x += get_count(page);
2394 spin_unlock_irqrestore(&n->list_lock, flags);
2397 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2399 static noinline void
2400 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2402 #ifdef CONFIG_SLUB_DEBUG
2403 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2404 DEFAULT_RATELIMIT_BURST);
2406 struct kmem_cache_node *n;
2408 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2411 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2412 nid, gfpflags, &gfpflags);
2413 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2414 s->name, s->object_size, s->size, oo_order(s->oo),
2417 if (oo_order(s->min) > get_order(s->object_size))
2418 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2421 for_each_kmem_cache_node(s, node, n) {
2422 unsigned long nr_slabs;
2423 unsigned long nr_objs;
2424 unsigned long nr_free;
2426 nr_free = count_partial(n, count_free);
2427 nr_slabs = node_nr_slabs(n);
2428 nr_objs = node_nr_objs(n);
2430 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2431 node, nr_slabs, nr_objs, nr_free);
2436 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2437 int node, struct kmem_cache_cpu **pc)
2440 struct kmem_cache_cpu *c = *pc;
2443 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2445 freelist = get_partial(s, flags, node, c);
2450 page = new_slab(s, flags, node);
2452 c = raw_cpu_ptr(s->cpu_slab);
2457 * No other reference to the page yet so we can
2458 * muck around with it freely without cmpxchg
2460 freelist = page->freelist;
2461 page->freelist = NULL;
2463 stat(s, ALLOC_SLAB);
2472 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2474 if (unlikely(PageSlabPfmemalloc(page)))
2475 return gfp_pfmemalloc_allowed(gfpflags);
2481 * Check the page->freelist of a page and either transfer the freelist to the
2482 * per cpu freelist or deactivate the page.
2484 * The page is still frozen if the return value is not NULL.
2486 * If this function returns NULL then the page has been unfrozen.
2488 * This function must be called with interrupt disabled.
2490 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2493 unsigned long counters;
2497 freelist = page->freelist;
2498 counters = page->counters;
2500 new.counters = counters;
2501 VM_BUG_ON(!new.frozen);
2503 new.inuse = page->objects;
2504 new.frozen = freelist != NULL;
2506 } while (!__cmpxchg_double_slab(s, page,
2515 * Slow path. The lockless freelist is empty or we need to perform
2518 * Processing is still very fast if new objects have been freed to the
2519 * regular freelist. In that case we simply take over the regular freelist
2520 * as the lockless freelist and zap the regular freelist.
2522 * If that is not working then we fall back to the partial lists. We take the
2523 * first element of the freelist as the object to allocate now and move the
2524 * rest of the freelist to the lockless freelist.
2526 * And if we were unable to get a new slab from the partial slab lists then
2527 * we need to allocate a new slab. This is the slowest path since it involves
2528 * a call to the page allocator and the setup of a new slab.
2530 * Version of __slab_alloc to use when we know that interrupts are
2531 * already disabled (which is the case for bulk allocation).
2533 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2534 unsigned long addr, struct kmem_cache_cpu *c)
2542 * if the node is not online or has no normal memory, just
2543 * ignore the node constraint
2545 if (unlikely(node != NUMA_NO_NODE &&
2546 !node_state(node, N_NORMAL_MEMORY)))
2547 node = NUMA_NO_NODE;
2552 if (unlikely(!node_match(page, node))) {
2554 * same as above but node_match() being false already
2555 * implies node != NUMA_NO_NODE
2557 if (!node_state(node, N_NORMAL_MEMORY)) {
2558 node = NUMA_NO_NODE;
2561 stat(s, ALLOC_NODE_MISMATCH);
2562 deactivate_slab(s, page, c->freelist, c);
2568 * By rights, we should be searching for a slab page that was
2569 * PFMEMALLOC but right now, we are losing the pfmemalloc
2570 * information when the page leaves the per-cpu allocator
2572 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2573 deactivate_slab(s, page, c->freelist, c);
2577 /* must check again c->freelist in case of cpu migration or IRQ */
2578 freelist = c->freelist;
2582 freelist = get_freelist(s, page);
2586 stat(s, DEACTIVATE_BYPASS);
2590 stat(s, ALLOC_REFILL);
2594 * freelist is pointing to the list of objects to be used.
2595 * page is pointing to the page from which the objects are obtained.
2596 * That page must be frozen for per cpu allocations to work.
2598 VM_BUG_ON(!c->page->frozen);
2599 c->freelist = get_freepointer(s, freelist);
2600 c->tid = next_tid(c->tid);
2605 if (slub_percpu_partial(c)) {
2606 page = c->page = slub_percpu_partial(c);
2607 slub_set_percpu_partial(c, page);
2608 stat(s, CPU_PARTIAL_ALLOC);
2612 freelist = new_slab_objects(s, gfpflags, node, &c);
2614 if (unlikely(!freelist)) {
2615 slab_out_of_memory(s, gfpflags, node);
2620 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2623 /* Only entered in the debug case */
2624 if (kmem_cache_debug(s) &&
2625 !alloc_debug_processing(s, page, freelist, addr))
2626 goto new_slab; /* Slab failed checks. Next slab needed */
2628 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2633 * Another one that disabled interrupt and compensates for possible
2634 * cpu changes by refetching the per cpu area pointer.
2636 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2637 unsigned long addr, struct kmem_cache_cpu *c)
2640 unsigned long flags;
2642 local_irq_save(flags);
2643 #ifdef CONFIG_PREEMPT
2645 * We may have been preempted and rescheduled on a different
2646 * cpu before disabling interrupts. Need to reload cpu area
2649 c = this_cpu_ptr(s->cpu_slab);
2652 p = ___slab_alloc(s, gfpflags, node, addr, c);
2653 local_irq_restore(flags);
2658 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2659 * have the fastpath folded into their functions. So no function call
2660 * overhead for requests that can be satisfied on the fastpath.
2662 * The fastpath works by first checking if the lockless freelist can be used.
2663 * If not then __slab_alloc is called for slow processing.
2665 * Otherwise we can simply pick the next object from the lockless free list.
2667 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2668 gfp_t gfpflags, int node, unsigned long addr)
2671 struct kmem_cache_cpu *c;
2675 s = slab_pre_alloc_hook(s, gfpflags);
2680 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2681 * enabled. We may switch back and forth between cpus while
2682 * reading from one cpu area. That does not matter as long
2683 * as we end up on the original cpu again when doing the cmpxchg.
2685 * We should guarantee that tid and kmem_cache are retrieved on
2686 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2687 * to check if it is matched or not.
2690 tid = this_cpu_read(s->cpu_slab->tid);
2691 c = raw_cpu_ptr(s->cpu_slab);
2692 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2693 unlikely(tid != READ_ONCE(c->tid)));
2696 * Irqless object alloc/free algorithm used here depends on sequence
2697 * of fetching cpu_slab's data. tid should be fetched before anything
2698 * on c to guarantee that object and page associated with previous tid
2699 * won't be used with current tid. If we fetch tid first, object and
2700 * page could be one associated with next tid and our alloc/free
2701 * request will be failed. In this case, we will retry. So, no problem.
2706 * The transaction ids are globally unique per cpu and per operation on
2707 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2708 * occurs on the right processor and that there was no operation on the
2709 * linked list in between.
2712 object = c->freelist;
2714 if (unlikely(!object || !node_match(page, node))) {
2715 object = __slab_alloc(s, gfpflags, node, addr, c);
2716 stat(s, ALLOC_SLOWPATH);
2718 void *next_object = get_freepointer_safe(s, object);
2721 * The cmpxchg will only match if there was no additional
2722 * operation and if we are on the right processor.
2724 * The cmpxchg does the following atomically (without lock
2726 * 1. Relocate first pointer to the current per cpu area.
2727 * 2. Verify that tid and freelist have not been changed
2728 * 3. If they were not changed replace tid and freelist
2730 * Since this is without lock semantics the protection is only
2731 * against code executing on this cpu *not* from access by
2734 if (unlikely(!this_cpu_cmpxchg_double(
2735 s->cpu_slab->freelist, s->cpu_slab->tid,
2737 next_object, next_tid(tid)))) {
2739 note_cmpxchg_failure("slab_alloc", s, tid);
2742 prefetch_freepointer(s, next_object);
2743 stat(s, ALLOC_FASTPATH);
2746 if (unlikely(gfpflags & __GFP_ZERO) && object)
2747 memset(object, 0, s->object_size);
2749 slab_post_alloc_hook(s, gfpflags, 1, &object);
2754 static __always_inline void *slab_alloc(struct kmem_cache *s,
2755 gfp_t gfpflags, unsigned long addr)
2757 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2760 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2762 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2764 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2769 EXPORT_SYMBOL(kmem_cache_alloc);
2771 #ifdef CONFIG_TRACING
2772 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2774 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2775 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2776 kasan_kmalloc(s, ret, size, gfpflags);
2779 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2783 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2785 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2787 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2788 s->object_size, s->size, gfpflags, node);
2792 EXPORT_SYMBOL(kmem_cache_alloc_node);
2794 #ifdef CONFIG_TRACING
2795 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2797 int node, size_t size)
2799 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2801 trace_kmalloc_node(_RET_IP_, ret,
2802 size, s->size, gfpflags, node);
2804 kasan_kmalloc(s, ret, size, gfpflags);
2807 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2812 * Slow path handling. This may still be called frequently since objects
2813 * have a longer lifetime than the cpu slabs in most processing loads.
2815 * So we still attempt to reduce cache line usage. Just take the slab
2816 * lock and free the item. If there is no additional partial page
2817 * handling required then we can return immediately.
2819 static void __slab_free(struct kmem_cache *s, struct page *page,
2820 void *head, void *tail, int cnt,
2827 unsigned long counters;
2828 struct kmem_cache_node *n = NULL;
2829 unsigned long uninitialized_var(flags);
2831 stat(s, FREE_SLOWPATH);
2833 if (kmem_cache_debug(s) &&
2834 !free_debug_processing(s, page, head, tail, cnt, addr))
2839 spin_unlock_irqrestore(&n->list_lock, flags);
2842 prior = page->freelist;
2843 counters = page->counters;
2844 set_freepointer(s, tail, prior);
2845 new.counters = counters;
2846 was_frozen = new.frozen;
2848 if ((!new.inuse || !prior) && !was_frozen) {
2850 if (kmem_cache_has_cpu_partial(s) && !prior) {
2853 * Slab was on no list before and will be
2855 * We can defer the list move and instead
2860 } else { /* Needs to be taken off a list */
2862 n = get_node(s, page_to_nid(page));
2864 * Speculatively acquire the list_lock.
2865 * If the cmpxchg does not succeed then we may
2866 * drop the list_lock without any processing.
2868 * Otherwise the list_lock will synchronize with
2869 * other processors updating the list of slabs.
2871 spin_lock_irqsave(&n->list_lock, flags);
2876 } while (!cmpxchg_double_slab(s, page,
2884 * If we just froze the page then put it onto the
2885 * per cpu partial list.
2887 if (new.frozen && !was_frozen) {
2888 put_cpu_partial(s, page, 1);
2889 stat(s, CPU_PARTIAL_FREE);
2892 * The list lock was not taken therefore no list
2893 * activity can be necessary.
2896 stat(s, FREE_FROZEN);
2900 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2904 * Objects left in the slab. If it was not on the partial list before
2907 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2908 if (kmem_cache_debug(s))
2909 remove_full(s, n, page);
2910 add_partial(n, page, DEACTIVATE_TO_TAIL);
2911 stat(s, FREE_ADD_PARTIAL);
2913 spin_unlock_irqrestore(&n->list_lock, flags);
2919 * Slab on the partial list.
2921 remove_partial(n, page);
2922 stat(s, FREE_REMOVE_PARTIAL);
2924 /* Slab must be on the full list */
2925 remove_full(s, n, page);
2928 spin_unlock_irqrestore(&n->list_lock, flags);
2930 discard_slab(s, page);
2934 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2935 * can perform fastpath freeing without additional function calls.
2937 * The fastpath is only possible if we are freeing to the current cpu slab
2938 * of this processor. This typically the case if we have just allocated
2941 * If fastpath is not possible then fall back to __slab_free where we deal
2942 * with all sorts of special processing.
2944 * Bulk free of a freelist with several objects (all pointing to the
2945 * same page) possible by specifying head and tail ptr, plus objects
2946 * count (cnt). Bulk free indicated by tail pointer being set.
2948 static __always_inline void do_slab_free(struct kmem_cache *s,
2949 struct page *page, void *head, void *tail,
2950 int cnt, unsigned long addr)
2952 void *tail_obj = tail ? : head;
2953 struct kmem_cache_cpu *c;
2957 * Determine the currently cpus per cpu slab.
2958 * The cpu may change afterward. However that does not matter since
2959 * data is retrieved via this pointer. If we are on the same cpu
2960 * during the cmpxchg then the free will succeed.
2963 tid = this_cpu_read(s->cpu_slab->tid);
2964 c = raw_cpu_ptr(s->cpu_slab);
2965 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2966 unlikely(tid != READ_ONCE(c->tid)));
2968 /* Same with comment on barrier() in slab_alloc_node() */
2971 if (likely(page == c->page)) {
2972 void **freelist = READ_ONCE(c->freelist);
2974 set_freepointer(s, tail_obj, freelist);
2976 if (unlikely(!this_cpu_cmpxchg_double(
2977 s->cpu_slab->freelist, s->cpu_slab->tid,
2979 head, next_tid(tid)))) {
2981 note_cmpxchg_failure("slab_free", s, tid);
2984 stat(s, FREE_FASTPATH);
2986 __slab_free(s, page, head, tail_obj, cnt, addr);
2990 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2991 void *head, void *tail, int cnt,
2995 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2996 * to remove objects, whose reuse must be delayed.
2998 if (slab_free_freelist_hook(s, &head, &tail, &cnt))
2999 do_slab_free(s, page, head, tail, cnt, addr);
3003 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3005 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3009 void kmem_cache_free(struct kmem_cache *s, void *x)
3011 s = cache_from_obj(s, x);
3014 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3015 trace_kmem_cache_free(_RET_IP_, x);
3017 EXPORT_SYMBOL(kmem_cache_free);
3019 struct detached_freelist {
3024 struct kmem_cache *s;
3028 * This function progressively scans the array with free objects (with
3029 * a limited look ahead) and extract objects belonging to the same
3030 * page. It builds a detached freelist directly within the given
3031 * page/objects. This can happen without any need for
3032 * synchronization, because the objects are owned by running process.
3033 * The freelist is build up as a single linked list in the objects.
3034 * The idea is, that this detached freelist can then be bulk
3035 * transferred to the real freelist(s), but only requiring a single
3036 * synchronization primitive. Look ahead in the array is limited due
3037 * to performance reasons.
3040 int build_detached_freelist(struct kmem_cache *s, size_t size,
3041 void **p, struct detached_freelist *df)
3043 size_t first_skipped_index = 0;
3048 /* Always re-init detached_freelist */
3053 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3054 } while (!object && size);
3059 page = virt_to_head_page(object);
3061 /* Handle kalloc'ed objects */
3062 if (unlikely(!PageSlab(page))) {
3063 BUG_ON(!PageCompound(page));
3065 __free_pages(page, compound_order(page));
3066 p[size] = NULL; /* mark object processed */
3069 /* Derive kmem_cache from object */
3070 df->s = page->slab_cache;
3072 df->s = cache_from_obj(s, object); /* Support for memcg */
3075 /* Start new detached freelist */
3077 set_freepointer(df->s, object, NULL);
3079 df->freelist = object;
3080 p[size] = NULL; /* mark object processed */
3086 continue; /* Skip processed objects */
3088 /* df->page is always set at this point */
3089 if (df->page == virt_to_head_page(object)) {
3090 /* Opportunity build freelist */
3091 set_freepointer(df->s, object, df->freelist);
3092 df->freelist = object;
3094 p[size] = NULL; /* mark object processed */
3099 /* Limit look ahead search */
3103 if (!first_skipped_index)
3104 first_skipped_index = size + 1;
3107 return first_skipped_index;
3110 /* Note that interrupts must be enabled when calling this function. */
3111 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3117 struct detached_freelist df;
3119 size = build_detached_freelist(s, size, p, &df);
3123 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3124 } while (likely(size));
3126 EXPORT_SYMBOL(kmem_cache_free_bulk);
3128 /* Note that interrupts must be enabled when calling this function. */
3129 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3132 struct kmem_cache_cpu *c;
3135 /* memcg and kmem_cache debug support */
3136 s = slab_pre_alloc_hook(s, flags);
3140 * Drain objects in the per cpu slab, while disabling local
3141 * IRQs, which protects against PREEMPT and interrupts
3142 * handlers invoking normal fastpath.
3144 local_irq_disable();
3145 c = this_cpu_ptr(s->cpu_slab);
3147 for (i = 0; i < size; i++) {
3148 void *object = c->freelist;
3150 if (unlikely(!object)) {
3152 * We may have removed an object from c->freelist using
3153 * the fastpath in the previous iteration; in that case,
3154 * c->tid has not been bumped yet.
3155 * Since ___slab_alloc() may reenable interrupts while
3156 * allocating memory, we should bump c->tid now.
3158 c->tid = next_tid(c->tid);
3161 * Invoking slow path likely have side-effect
3162 * of re-populating per CPU c->freelist
3164 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3166 if (unlikely(!p[i]))
3169 c = this_cpu_ptr(s->cpu_slab);
3170 continue; /* goto for-loop */
3172 c->freelist = get_freepointer(s, object);
3175 c->tid = next_tid(c->tid);
3178 /* Clear memory outside IRQ disabled fastpath loop */
3179 if (unlikely(flags & __GFP_ZERO)) {
3182 for (j = 0; j < i; j++)
3183 memset(p[j], 0, s->object_size);
3186 /* memcg and kmem_cache debug support */
3187 slab_post_alloc_hook(s, flags, size, p);
3191 slab_post_alloc_hook(s, flags, i, p);
3192 __kmem_cache_free_bulk(s, i, p);
3195 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3199 * Object placement in a slab is made very easy because we always start at
3200 * offset 0. If we tune the size of the object to the alignment then we can
3201 * get the required alignment by putting one properly sized object after
3204 * Notice that the allocation order determines the sizes of the per cpu
3205 * caches. Each processor has always one slab available for allocations.
3206 * Increasing the allocation order reduces the number of times that slabs
3207 * must be moved on and off the partial lists and is therefore a factor in
3212 * Mininum / Maximum order of slab pages. This influences locking overhead
3213 * and slab fragmentation. A higher order reduces the number of partial slabs
3214 * and increases the number of allocations possible without having to
3215 * take the list_lock.
3217 static unsigned int slub_min_order;
3218 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3219 static unsigned int slub_min_objects;
3222 * Calculate the order of allocation given an slab object size.
3224 * The order of allocation has significant impact on performance and other
3225 * system components. Generally order 0 allocations should be preferred since
3226 * order 0 does not cause fragmentation in the page allocator. Larger objects
3227 * be problematic to put into order 0 slabs because there may be too much
3228 * unused space left. We go to a higher order if more than 1/16th of the slab
3231 * In order to reach satisfactory performance we must ensure that a minimum
3232 * number of objects is in one slab. Otherwise we may generate too much
3233 * activity on the partial lists which requires taking the list_lock. This is
3234 * less a concern for large slabs though which are rarely used.
3236 * slub_max_order specifies the order where we begin to stop considering the
3237 * number of objects in a slab as critical. If we reach slub_max_order then
3238 * we try to keep the page order as low as possible. So we accept more waste
3239 * of space in favor of a small page order.
3241 * Higher order allocations also allow the placement of more objects in a
3242 * slab and thereby reduce object handling overhead. If the user has
3243 * requested a higher mininum order then we start with that one instead of
3244 * the smallest order which will fit the object.
3246 static inline unsigned int slab_order(unsigned int size,
3247 unsigned int min_objects, unsigned int max_order,
3248 unsigned int fract_leftover)
3250 unsigned int min_order = slub_min_order;
3253 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3254 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3256 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3257 order <= max_order; order++) {
3259 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3262 rem = slab_size % size;
3264 if (rem <= slab_size / fract_leftover)
3271 static inline int calculate_order(unsigned int size)
3274 unsigned int min_objects;
3275 unsigned int max_objects;
3278 * Attempt to find best configuration for a slab. This
3279 * works by first attempting to generate a layout with
3280 * the best configuration and backing off gradually.
3282 * First we increase the acceptable waste in a slab. Then
3283 * we reduce the minimum objects required in a slab.
3285 min_objects = slub_min_objects;
3287 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3288 max_objects = order_objects(slub_max_order, size);
3289 min_objects = min(min_objects, max_objects);
3291 while (min_objects > 1) {
3292 unsigned int fraction;
3295 while (fraction >= 4) {
3296 order = slab_order(size, min_objects,
3297 slub_max_order, fraction);
3298 if (order <= slub_max_order)
3306 * We were unable to place multiple objects in a slab. Now
3307 * lets see if we can place a single object there.
3309 order = slab_order(size, 1, slub_max_order, 1);
3310 if (order <= slub_max_order)
3314 * Doh this slab cannot be placed using slub_max_order.
3316 order = slab_order(size, 1, MAX_ORDER, 1);
3317 if (order < MAX_ORDER)
3323 init_kmem_cache_node(struct kmem_cache_node *n)
3326 spin_lock_init(&n->list_lock);
3327 INIT_LIST_HEAD(&n->partial);
3328 #ifdef CONFIG_SLUB_DEBUG
3329 atomic_long_set(&n->nr_slabs, 0);
3330 atomic_long_set(&n->total_objects, 0);
3331 INIT_LIST_HEAD(&n->full);
3335 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3337 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3338 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3341 * Must align to double word boundary for the double cmpxchg
3342 * instructions to work; see __pcpu_double_call_return_bool().
3344 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3345 2 * sizeof(void *));
3350 init_kmem_cache_cpus(s);
3355 static struct kmem_cache *kmem_cache_node;
3358 * No kmalloc_node yet so do it by hand. We know that this is the first
3359 * slab on the node for this slabcache. There are no concurrent accesses
3362 * Note that this function only works on the kmem_cache_node
3363 * when allocating for the kmem_cache_node. This is used for bootstrapping
3364 * memory on a fresh node that has no slab structures yet.
3366 static void early_kmem_cache_node_alloc(int node)
3369 struct kmem_cache_node *n;
3371 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3373 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3376 if (page_to_nid(page) != node) {
3377 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3378 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3383 page->freelist = get_freepointer(kmem_cache_node, n);
3386 kmem_cache_node->node[node] = n;
3387 #ifdef CONFIG_SLUB_DEBUG
3388 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3389 init_tracking(kmem_cache_node, n);
3391 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3393 init_kmem_cache_node(n);
3394 inc_slabs_node(kmem_cache_node, node, page->objects);
3397 * No locks need to be taken here as it has just been
3398 * initialized and there is no concurrent access.
3400 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3403 static void free_kmem_cache_nodes(struct kmem_cache *s)
3406 struct kmem_cache_node *n;
3408 for_each_kmem_cache_node(s, node, n) {
3409 s->node[node] = NULL;
3410 kmem_cache_free(kmem_cache_node, n);
3414 void __kmem_cache_release(struct kmem_cache *s)
3416 cache_random_seq_destroy(s);
3417 free_percpu(s->cpu_slab);
3418 free_kmem_cache_nodes(s);
3421 static int init_kmem_cache_nodes(struct kmem_cache *s)
3425 for_each_node_state(node, N_NORMAL_MEMORY) {
3426 struct kmem_cache_node *n;
3428 if (slab_state == DOWN) {
3429 early_kmem_cache_node_alloc(node);
3432 n = kmem_cache_alloc_node(kmem_cache_node,
3436 free_kmem_cache_nodes(s);
3440 init_kmem_cache_node(n);
3446 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3448 if (min < MIN_PARTIAL)
3450 else if (min > MAX_PARTIAL)
3452 s->min_partial = min;
3455 static void set_cpu_partial(struct kmem_cache *s)
3457 #ifdef CONFIG_SLUB_CPU_PARTIAL
3459 * cpu_partial determined the maximum number of objects kept in the
3460 * per cpu partial lists of a processor.
3462 * Per cpu partial lists mainly contain slabs that just have one
3463 * object freed. If they are used for allocation then they can be
3464 * filled up again with minimal effort. The slab will never hit the
3465 * per node partial lists and therefore no locking will be required.
3467 * This setting also determines
3469 * A) The number of objects from per cpu partial slabs dumped to the
3470 * per node list when we reach the limit.
3471 * B) The number of objects in cpu partial slabs to extract from the
3472 * per node list when we run out of per cpu objects. We only fetch
3473 * 50% to keep some capacity around for frees.
3475 if (!kmem_cache_has_cpu_partial(s))
3477 else if (s->size >= PAGE_SIZE)
3479 else if (s->size >= 1024)
3481 else if (s->size >= 256)
3482 s->cpu_partial = 13;
3484 s->cpu_partial = 30;
3489 * calculate_sizes() determines the order and the distribution of data within
3492 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3494 slab_flags_t flags = s->flags;
3495 unsigned int size = s->object_size;
3499 * Round up object size to the next word boundary. We can only
3500 * place the free pointer at word boundaries and this determines
3501 * the possible location of the free pointer.
3503 size = ALIGN(size, sizeof(void *));
3505 #ifdef CONFIG_SLUB_DEBUG
3507 * Determine if we can poison the object itself. If the user of
3508 * the slab may touch the object after free or before allocation
3509 * then we should never poison the object itself.
3511 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3513 s->flags |= __OBJECT_POISON;
3515 s->flags &= ~__OBJECT_POISON;
3519 * If we are Redzoning then check if there is some space between the
3520 * end of the object and the free pointer. If not then add an
3521 * additional word to have some bytes to store Redzone information.
3523 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3524 size += sizeof(void *);
3528 * With that we have determined the number of bytes in actual use
3529 * by the object. This is the potential offset to the free pointer.
3533 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3536 * Relocate free pointer after the object if it is not
3537 * permitted to overwrite the first word of the object on
3540 * This is the case if we do RCU, have a constructor or
3541 * destructor or are poisoning the objects.
3544 size += sizeof(void *);
3547 #ifdef CONFIG_SLUB_DEBUG
3548 if (flags & SLAB_STORE_USER)
3550 * Need to store information about allocs and frees after
3553 size += 2 * sizeof(struct track);
3556 kasan_cache_create(s, &size, &s->flags);
3557 #ifdef CONFIG_SLUB_DEBUG
3558 if (flags & SLAB_RED_ZONE) {
3560 * Add some empty padding so that we can catch
3561 * overwrites from earlier objects rather than let
3562 * tracking information or the free pointer be
3563 * corrupted if a user writes before the start
3566 size += sizeof(void *);
3568 s->red_left_pad = sizeof(void *);
3569 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3570 size += s->red_left_pad;
3575 * SLUB stores one object immediately after another beginning from
3576 * offset 0. In order to align the objects we have to simply size
3577 * each object to conform to the alignment.
3579 size = ALIGN(size, s->align);
3581 if (forced_order >= 0)
3582 order = forced_order;
3584 order = calculate_order(size);
3591 s->allocflags |= __GFP_COMP;
3593 if (s->flags & SLAB_CACHE_DMA)
3594 s->allocflags |= GFP_DMA;
3596 if (s->flags & SLAB_CACHE_DMA32)
3597 s->allocflags |= GFP_DMA32;
3599 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3600 s->allocflags |= __GFP_RECLAIMABLE;
3603 * Determine the number of objects per slab
3605 s->oo = oo_make(order, size);
3606 s->min = oo_make(get_order(size), size);
3607 if (oo_objects(s->oo) > oo_objects(s->max))
3610 return !!oo_objects(s->oo);
3613 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3615 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3616 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3617 s->random = get_random_long();
3620 if (!calculate_sizes(s, -1))
3622 if (disable_higher_order_debug) {
3624 * Disable debugging flags that store metadata if the min slab
3627 if (get_order(s->size) > get_order(s->object_size)) {
3628 s->flags &= ~DEBUG_METADATA_FLAGS;
3630 if (!calculate_sizes(s, -1))
3635 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3636 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3637 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3638 /* Enable fast mode */
3639 s->flags |= __CMPXCHG_DOUBLE;
3643 * The larger the object size is, the more pages we want on the partial
3644 * list to avoid pounding the page allocator excessively.
3646 set_min_partial(s, ilog2(s->size) / 2);
3651 s->remote_node_defrag_ratio = 1000;
3654 /* Initialize the pre-computed randomized freelist if slab is up */
3655 if (slab_state >= UP) {
3656 if (init_cache_random_seq(s))
3660 if (!init_kmem_cache_nodes(s))
3663 if (alloc_kmem_cache_cpus(s))
3666 free_kmem_cache_nodes(s);
3668 if (flags & SLAB_PANIC)
3669 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3670 s->name, s->size, s->size,
3671 oo_order(s->oo), s->offset, (unsigned long)flags);
3675 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3678 #ifdef CONFIG_SLUB_DEBUG
3679 void *addr = page_address(page);
3681 unsigned long *map = kcalloc(BITS_TO_LONGS(page->objects),
3686 slab_err(s, page, text, s->name);
3689 get_map(s, page, map);
3690 for_each_object(p, s, addr, page->objects) {
3692 if (!test_bit(slab_index(p, s, addr), map)) {
3693 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3694 print_tracking(s, p);
3703 * Attempt to free all partial slabs on a node.
3704 * This is called from __kmem_cache_shutdown(). We must take list_lock
3705 * because sysfs file might still access partial list after the shutdowning.
3707 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3710 struct page *page, *h;
3712 BUG_ON(irqs_disabled());
3713 spin_lock_irq(&n->list_lock);
3714 list_for_each_entry_safe(page, h, &n->partial, lru) {
3716 remove_partial(n, page);
3717 list_add(&page->lru, &discard);
3719 list_slab_objects(s, page,
3720 "Objects remaining in %s on __kmem_cache_shutdown()");
3723 spin_unlock_irq(&n->list_lock);
3725 list_for_each_entry_safe(page, h, &discard, lru)
3726 discard_slab(s, page);
3729 bool __kmem_cache_empty(struct kmem_cache *s)
3732 struct kmem_cache_node *n;
3734 for_each_kmem_cache_node(s, node, n)
3735 if (n->nr_partial || slabs_node(s, node))
3741 * Release all resources used by a slab cache.
3743 int __kmem_cache_shutdown(struct kmem_cache *s)
3746 struct kmem_cache_node *n;
3749 /* Attempt to free all objects */
3750 for_each_kmem_cache_node(s, node, n) {
3752 if (n->nr_partial || slabs_node(s, node))
3755 sysfs_slab_remove(s);
3759 /********************************************************************
3761 *******************************************************************/
3763 static int __init setup_slub_min_order(char *str)
3765 get_option(&str, (int *)&slub_min_order);
3770 __setup("slub_min_order=", setup_slub_min_order);
3772 static int __init setup_slub_max_order(char *str)
3774 get_option(&str, (int *)&slub_max_order);
3775 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3780 __setup("slub_max_order=", setup_slub_max_order);
3782 static int __init setup_slub_min_objects(char *str)
3784 get_option(&str, (int *)&slub_min_objects);
3789 __setup("slub_min_objects=", setup_slub_min_objects);
3791 void *__kmalloc(size_t size, gfp_t flags)
3793 struct kmem_cache *s;
3796 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3797 return kmalloc_large(size, flags);
3799 s = kmalloc_slab(size, flags);
3801 if (unlikely(ZERO_OR_NULL_PTR(s)))
3804 ret = slab_alloc(s, flags, _RET_IP_);
3806 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3808 kasan_kmalloc(s, ret, size, flags);
3812 EXPORT_SYMBOL(__kmalloc);
3815 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3820 flags |= __GFP_COMP;
3821 page = alloc_pages_node(node, flags, get_order(size));
3823 ptr = page_address(page);
3825 kmalloc_large_node_hook(ptr, size, flags);
3829 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3831 struct kmem_cache *s;
3834 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3835 ret = kmalloc_large_node(size, flags, node);
3837 trace_kmalloc_node(_RET_IP_, ret,
3838 size, PAGE_SIZE << get_order(size),
3844 s = kmalloc_slab(size, flags);
3846 if (unlikely(ZERO_OR_NULL_PTR(s)))
3849 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3851 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3853 kasan_kmalloc(s, ret, size, flags);
3857 EXPORT_SYMBOL(__kmalloc_node);
3860 #ifdef CONFIG_HARDENED_USERCOPY
3862 * Rejects incorrectly sized objects and objects that are to be copied
3863 * to/from userspace but do not fall entirely within the containing slab
3864 * cache's usercopy region.
3866 * Returns NULL if check passes, otherwise const char * to name of cache
3867 * to indicate an error.
3869 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3872 struct kmem_cache *s;
3873 unsigned int offset;
3876 /* Find object and usable object size. */
3877 s = page->slab_cache;
3879 /* Reject impossible pointers. */
3880 if (ptr < page_address(page))
3881 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3884 /* Find offset within object. */
3885 offset = (ptr - page_address(page)) % s->size;
3887 /* Adjust for redzone and reject if within the redzone. */
3888 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3889 if (offset < s->red_left_pad)
3890 usercopy_abort("SLUB object in left red zone",
3891 s->name, to_user, offset, n);
3892 offset -= s->red_left_pad;
3895 /* Allow address range falling entirely within usercopy region. */
3896 if (offset >= s->useroffset &&
3897 offset - s->useroffset <= s->usersize &&
3898 n <= s->useroffset - offset + s->usersize)
3902 * If the copy is still within the allocated object, produce
3903 * a warning instead of rejecting the copy. This is intended
3904 * to be a temporary method to find any missing usercopy
3907 object_size = slab_ksize(s);
3908 if (usercopy_fallback &&
3909 offset <= object_size && n <= object_size - offset) {
3910 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3914 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3916 #endif /* CONFIG_HARDENED_USERCOPY */
3918 static size_t __ksize(const void *object)
3922 if (unlikely(object == ZERO_SIZE_PTR))
3925 page = virt_to_head_page(object);
3927 if (unlikely(!PageSlab(page))) {
3928 WARN_ON(!PageCompound(page));
3929 return PAGE_SIZE << compound_order(page);
3932 return slab_ksize(page->slab_cache);
3935 size_t ksize(const void *object)
3937 size_t size = __ksize(object);
3938 /* We assume that ksize callers could use whole allocated area,
3939 * so we need to unpoison this area.
3941 kasan_unpoison_shadow(object, size);
3944 EXPORT_SYMBOL(ksize);
3946 void kfree(const void *x)
3949 void *object = (void *)x;
3951 trace_kfree(_RET_IP_, x);
3953 if (unlikely(ZERO_OR_NULL_PTR(x)))
3956 page = virt_to_head_page(x);
3957 if (unlikely(!PageSlab(page))) {
3958 BUG_ON(!PageCompound(page));
3960 __free_pages(page, compound_order(page));
3963 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3965 EXPORT_SYMBOL(kfree);
3967 #define SHRINK_PROMOTE_MAX 32
3970 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3971 * up most to the head of the partial lists. New allocations will then
3972 * fill those up and thus they can be removed from the partial lists.
3974 * The slabs with the least items are placed last. This results in them
3975 * being allocated from last increasing the chance that the last objects
3976 * are freed in them.
3978 int __kmem_cache_shrink(struct kmem_cache *s)
3982 struct kmem_cache_node *n;
3985 struct list_head discard;
3986 struct list_head promote[SHRINK_PROMOTE_MAX];
3987 unsigned long flags;
3991 for_each_kmem_cache_node(s, node, n) {
3992 INIT_LIST_HEAD(&discard);
3993 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3994 INIT_LIST_HEAD(promote + i);
3996 spin_lock_irqsave(&n->list_lock, flags);
3999 * Build lists of slabs to discard or promote.
4001 * Note that concurrent frees may occur while we hold the
4002 * list_lock. page->inuse here is the upper limit.
4004 list_for_each_entry_safe(page, t, &n->partial, lru) {
4005 int free = page->objects - page->inuse;
4007 /* Do not reread page->inuse */
4010 /* We do not keep full slabs on the list */
4013 if (free == page->objects) {
4014 list_move(&page->lru, &discard);
4016 } else if (free <= SHRINK_PROMOTE_MAX)
4017 list_move(&page->lru, promote + free - 1);
4021 * Promote the slabs filled up most to the head of the
4024 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4025 list_splice(promote + i, &n->partial);
4027 spin_unlock_irqrestore(&n->list_lock, flags);
4029 /* Release empty slabs */
4030 list_for_each_entry_safe(page, t, &discard, lru)
4031 discard_slab(s, page);
4033 if (slabs_node(s, node))
4041 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
4044 * Called with all the locks held after a sched RCU grace period.
4045 * Even if @s becomes empty after shrinking, we can't know that @s
4046 * doesn't have allocations already in-flight and thus can't
4047 * destroy @s until the associated memcg is released.
4049 * However, let's remove the sysfs files for empty caches here.
4050 * Each cache has a lot of interface files which aren't
4051 * particularly useful for empty draining caches; otherwise, we can
4052 * easily end up with millions of unnecessary sysfs files on
4053 * systems which have a lot of memory and transient cgroups.
4055 if (!__kmem_cache_shrink(s))
4056 sysfs_slab_remove(s);
4059 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4062 * Disable empty slabs caching. Used to avoid pinning offline
4063 * memory cgroups by kmem pages that can be freed.
4065 slub_set_cpu_partial(s, 0);
4069 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4070 * we have to make sure the change is visible before shrinking.
4072 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4076 static int slab_mem_going_offline_callback(void *arg)
4078 struct kmem_cache *s;
4080 mutex_lock(&slab_mutex);
4081 list_for_each_entry(s, &slab_caches, list)
4082 __kmem_cache_shrink(s);
4083 mutex_unlock(&slab_mutex);
4088 static void slab_mem_offline_callback(void *arg)
4090 struct kmem_cache_node *n;
4091 struct kmem_cache *s;
4092 struct memory_notify *marg = arg;
4095 offline_node = marg->status_change_nid_normal;
4098 * If the node still has available memory. we need kmem_cache_node
4101 if (offline_node < 0)
4104 mutex_lock(&slab_mutex);
4105 list_for_each_entry(s, &slab_caches, list) {
4106 n = get_node(s, offline_node);
4109 * if n->nr_slabs > 0, slabs still exist on the node
4110 * that is going down. We were unable to free them,
4111 * and offline_pages() function shouldn't call this
4112 * callback. So, we must fail.
4114 BUG_ON(slabs_node(s, offline_node));
4116 s->node[offline_node] = NULL;
4117 kmem_cache_free(kmem_cache_node, n);
4120 mutex_unlock(&slab_mutex);
4123 static int slab_mem_going_online_callback(void *arg)
4125 struct kmem_cache_node *n;
4126 struct kmem_cache *s;
4127 struct memory_notify *marg = arg;
4128 int nid = marg->status_change_nid_normal;
4132 * If the node's memory is already available, then kmem_cache_node is
4133 * already created. Nothing to do.
4139 * We are bringing a node online. No memory is available yet. We must
4140 * allocate a kmem_cache_node structure in order to bring the node
4143 mutex_lock(&slab_mutex);
4144 list_for_each_entry(s, &slab_caches, list) {
4146 * XXX: kmem_cache_alloc_node will fallback to other nodes
4147 * since memory is not yet available from the node that
4150 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4155 init_kmem_cache_node(n);
4159 mutex_unlock(&slab_mutex);
4163 static int slab_memory_callback(struct notifier_block *self,
4164 unsigned long action, void *arg)
4169 case MEM_GOING_ONLINE:
4170 ret = slab_mem_going_online_callback(arg);
4172 case MEM_GOING_OFFLINE:
4173 ret = slab_mem_going_offline_callback(arg);
4176 case MEM_CANCEL_ONLINE:
4177 slab_mem_offline_callback(arg);
4180 case MEM_CANCEL_OFFLINE:
4184 ret = notifier_from_errno(ret);
4190 static struct notifier_block slab_memory_callback_nb = {
4191 .notifier_call = slab_memory_callback,
4192 .priority = SLAB_CALLBACK_PRI,
4195 /********************************************************************
4196 * Basic setup of slabs
4197 *******************************************************************/
4200 * Used for early kmem_cache structures that were allocated using
4201 * the page allocator. Allocate them properly then fix up the pointers
4202 * that may be pointing to the wrong kmem_cache structure.
4205 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4208 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4209 struct kmem_cache_node *n;
4211 memcpy(s, static_cache, kmem_cache->object_size);
4214 * This runs very early, and only the boot processor is supposed to be
4215 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4218 __flush_cpu_slab(s, smp_processor_id());
4219 for_each_kmem_cache_node(s, node, n) {
4222 list_for_each_entry(p, &n->partial, lru)
4225 #ifdef CONFIG_SLUB_DEBUG
4226 list_for_each_entry(p, &n->full, lru)
4230 slab_init_memcg_params(s);
4231 list_add(&s->list, &slab_caches);
4232 memcg_link_cache(s);
4236 void __init kmem_cache_init(void)
4238 static __initdata struct kmem_cache boot_kmem_cache,
4239 boot_kmem_cache_node;
4241 if (debug_guardpage_minorder())
4244 kmem_cache_node = &boot_kmem_cache_node;
4245 kmem_cache = &boot_kmem_cache;
4247 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4248 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4250 register_hotmemory_notifier(&slab_memory_callback_nb);
4252 /* Able to allocate the per node structures */
4253 slab_state = PARTIAL;
4255 create_boot_cache(kmem_cache, "kmem_cache",
4256 offsetof(struct kmem_cache, node) +
4257 nr_node_ids * sizeof(struct kmem_cache_node *),
4258 SLAB_HWCACHE_ALIGN, 0, 0);
4260 kmem_cache = bootstrap(&boot_kmem_cache);
4261 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4263 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4264 setup_kmalloc_cache_index_table();
4265 create_kmalloc_caches(0);
4267 /* Setup random freelists for each cache */
4268 init_freelist_randomization();
4270 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4273 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%d\n",
4275 slub_min_order, slub_max_order, slub_min_objects,
4276 nr_cpu_ids, nr_node_ids);
4279 void __init kmem_cache_init_late(void)
4284 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4285 slab_flags_t flags, void (*ctor)(void *))
4287 struct kmem_cache *s, *c;
4289 s = find_mergeable(size, align, flags, name, ctor);
4294 * Adjust the object sizes so that we clear
4295 * the complete object on kzalloc.
4297 s->object_size = max(s->object_size, size);
4298 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4300 for_each_memcg_cache(c, s) {
4301 c->object_size = s->object_size;
4302 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4305 if (sysfs_slab_alias(s, name)) {
4314 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4318 err = kmem_cache_open(s, flags);
4322 /* Mutex is not taken during early boot */
4323 if (slab_state <= UP)
4326 memcg_propagate_slab_attrs(s);
4327 err = sysfs_slab_add(s);
4329 __kmem_cache_release(s);
4334 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4336 struct kmem_cache *s;
4339 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4340 return kmalloc_large(size, gfpflags);
4342 s = kmalloc_slab(size, gfpflags);
4344 if (unlikely(ZERO_OR_NULL_PTR(s)))
4347 ret = slab_alloc(s, gfpflags, caller);
4349 /* Honor the call site pointer we received. */
4350 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4356 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4357 int node, unsigned long caller)
4359 struct kmem_cache *s;
4362 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4363 ret = kmalloc_large_node(size, gfpflags, node);
4365 trace_kmalloc_node(caller, ret,
4366 size, PAGE_SIZE << get_order(size),
4372 s = kmalloc_slab(size, gfpflags);
4374 if (unlikely(ZERO_OR_NULL_PTR(s)))
4377 ret = slab_alloc_node(s, gfpflags, node, caller);
4379 /* Honor the call site pointer we received. */
4380 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4387 static int count_inuse(struct page *page)
4392 static int count_total(struct page *page)
4394 return page->objects;
4398 #ifdef CONFIG_SLUB_DEBUG
4399 static int validate_slab(struct kmem_cache *s, struct page *page,
4403 void *addr = page_address(page);
4405 if (!check_slab(s, page) ||
4406 !on_freelist(s, page, NULL))
4409 /* Now we know that a valid freelist exists */
4410 bitmap_zero(map, page->objects);
4412 get_map(s, page, map);
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_INACTIVE))
4419 for_each_object(p, s, addr, page->objects)
4420 if (!test_bit(slab_index(p, s, addr), map))
4421 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4426 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4430 validate_slab(s, page, map);
4434 static int validate_slab_node(struct kmem_cache *s,
4435 struct kmem_cache_node *n, unsigned long *map)
4437 unsigned long count = 0;
4439 unsigned long flags;
4441 spin_lock_irqsave(&n->list_lock, flags);
4443 list_for_each_entry(page, &n->partial, lru) {
4444 validate_slab_slab(s, page, map);
4447 if (count != n->nr_partial)
4448 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4449 s->name, count, n->nr_partial);
4451 if (!(s->flags & SLAB_STORE_USER))
4454 list_for_each_entry(page, &n->full, lru) {
4455 validate_slab_slab(s, page, map);
4458 if (count != atomic_long_read(&n->nr_slabs))
4459 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4460 s->name, count, atomic_long_read(&n->nr_slabs));
4463 spin_unlock_irqrestore(&n->list_lock, flags);
4467 static long validate_slab_cache(struct kmem_cache *s)
4470 unsigned long count = 0;
4471 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4472 sizeof(unsigned long),
4474 struct kmem_cache_node *n;
4480 for_each_kmem_cache_node(s, node, n)
4481 count += validate_slab_node(s, n, map);
4486 * Generate lists of code addresses where slabcache objects are allocated
4491 unsigned long count;
4498 DECLARE_BITMAP(cpus, NR_CPUS);
4504 unsigned long count;
4505 struct location *loc;
4508 static void free_loc_track(struct loc_track *t)
4511 free_pages((unsigned long)t->loc,
4512 get_order(sizeof(struct location) * t->max));
4515 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4520 order = get_order(sizeof(struct location) * max);
4522 l = (void *)__get_free_pages(flags, order);
4527 memcpy(l, t->loc, sizeof(struct location) * t->count);
4535 static int add_location(struct loc_track *t, struct kmem_cache *s,
4536 const struct track *track)
4538 long start, end, pos;
4540 unsigned long caddr;
4541 unsigned long age = jiffies - track->when;
4547 pos = start + (end - start + 1) / 2;
4550 * There is nothing at "end". If we end up there
4551 * we need to add something to before end.
4556 caddr = t->loc[pos].addr;
4557 if (track->addr == caddr) {
4563 if (age < l->min_time)
4565 if (age > l->max_time)
4568 if (track->pid < l->min_pid)
4569 l->min_pid = track->pid;
4570 if (track->pid > l->max_pid)
4571 l->max_pid = track->pid;
4573 cpumask_set_cpu(track->cpu,
4574 to_cpumask(l->cpus));
4576 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4580 if (track->addr < caddr)
4587 * Not found. Insert new tracking element.
4589 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4595 (t->count - pos) * sizeof(struct location));
4598 l->addr = track->addr;
4602 l->min_pid = track->pid;
4603 l->max_pid = track->pid;
4604 cpumask_clear(to_cpumask(l->cpus));
4605 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4606 nodes_clear(l->nodes);
4607 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4611 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4612 struct page *page, enum track_item alloc,
4615 void *addr = page_address(page);
4618 bitmap_zero(map, page->objects);
4619 get_map(s, page, map);
4621 for_each_object(p, s, addr, page->objects)
4622 if (!test_bit(slab_index(p, s, addr), map))
4623 add_location(t, s, get_track(s, p, alloc));
4626 static int list_locations(struct kmem_cache *s, char *buf,
4627 enum track_item alloc)
4631 struct loc_track t = { 0, 0, NULL };
4633 unsigned long *map = kmalloc_array(BITS_TO_LONGS(oo_objects(s->max)),
4634 sizeof(unsigned long),
4636 struct kmem_cache_node *n;
4638 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4641 return sprintf(buf, "Out of memory\n");
4643 /* Push back cpu slabs */
4646 for_each_kmem_cache_node(s, node, n) {
4647 unsigned long flags;
4650 if (!atomic_long_read(&n->nr_slabs))
4653 spin_lock_irqsave(&n->list_lock, flags);
4654 list_for_each_entry(page, &n->partial, lru)
4655 process_slab(&t, s, page, alloc, map);
4656 list_for_each_entry(page, &n->full, lru)
4657 process_slab(&t, s, page, alloc, map);
4658 spin_unlock_irqrestore(&n->list_lock, flags);
4661 for (i = 0; i < t.count; i++) {
4662 struct location *l = &t.loc[i];
4664 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4666 len += sprintf(buf + len, "%7ld ", l->count);
4669 len += sprintf(buf + len, "%pS", (void *)l->addr);
4671 len += sprintf(buf + len, "<not-available>");
4673 if (l->sum_time != l->min_time) {
4674 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4676 (long)div_u64(l->sum_time, l->count),
4679 len += sprintf(buf + len, " age=%ld",
4682 if (l->min_pid != l->max_pid)
4683 len += sprintf(buf + len, " pid=%ld-%ld",
4684 l->min_pid, l->max_pid);
4686 len += sprintf(buf + len, " pid=%ld",
4689 if (num_online_cpus() > 1 &&
4690 !cpumask_empty(to_cpumask(l->cpus)) &&
4691 len < PAGE_SIZE - 60)
4692 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4694 cpumask_pr_args(to_cpumask(l->cpus)));
4696 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4697 len < PAGE_SIZE - 60)
4698 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4700 nodemask_pr_args(&l->nodes));
4702 len += sprintf(buf + len, "\n");
4708 len += sprintf(buf, "No data\n");
4713 #ifdef SLUB_RESILIENCY_TEST
4714 static void __init resiliency_test(void)
4718 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4720 pr_err("SLUB resiliency testing\n");
4721 pr_err("-----------------------\n");
4722 pr_err("A. Corruption after allocation\n");
4724 p = kzalloc(16, GFP_KERNEL);
4726 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4729 validate_slab_cache(kmalloc_caches[4]);
4731 /* Hmmm... The next two are dangerous */
4732 p = kzalloc(32, GFP_KERNEL);
4733 p[32 + sizeof(void *)] = 0x34;
4734 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4736 pr_err("If allocated object is overwritten then not detectable\n\n");
4738 validate_slab_cache(kmalloc_caches[5]);
4739 p = kzalloc(64, GFP_KERNEL);
4740 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4742 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4744 pr_err("If allocated object is overwritten then not detectable\n\n");
4745 validate_slab_cache(kmalloc_caches[6]);
4747 pr_err("\nB. Corruption after free\n");
4748 p = kzalloc(128, GFP_KERNEL);
4751 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4752 validate_slab_cache(kmalloc_caches[7]);
4754 p = kzalloc(256, GFP_KERNEL);
4757 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4758 validate_slab_cache(kmalloc_caches[8]);
4760 p = kzalloc(512, GFP_KERNEL);
4763 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4764 validate_slab_cache(kmalloc_caches[9]);
4768 static void resiliency_test(void) {};
4773 enum slab_stat_type {
4774 SL_ALL, /* All slabs */
4775 SL_PARTIAL, /* Only partially allocated slabs */
4776 SL_CPU, /* Only slabs used for cpu caches */
4777 SL_OBJECTS, /* Determine allocated objects not slabs */
4778 SL_TOTAL /* Determine object capacity not slabs */
4781 #define SO_ALL (1 << SL_ALL)
4782 #define SO_PARTIAL (1 << SL_PARTIAL)
4783 #define SO_CPU (1 << SL_CPU)
4784 #define SO_OBJECTS (1 << SL_OBJECTS)
4785 #define SO_TOTAL (1 << SL_TOTAL)
4788 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4790 static int __init setup_slub_memcg_sysfs(char *str)
4794 if (get_option(&str, &v) > 0)
4795 memcg_sysfs_enabled = v;
4800 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4803 static ssize_t show_slab_objects(struct kmem_cache *s,
4804 char *buf, unsigned long flags)
4806 unsigned long total = 0;
4809 unsigned long *nodes;
4811 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4815 if (flags & SO_CPU) {
4818 for_each_possible_cpu(cpu) {
4819 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4824 page = READ_ONCE(c->page);
4828 node = page_to_nid(page);
4829 if (flags & SO_TOTAL)
4831 else if (flags & SO_OBJECTS)
4839 page = slub_percpu_partial_read_once(c);
4841 node = page_to_nid(page);
4842 if (flags & SO_TOTAL)
4844 else if (flags & SO_OBJECTS)
4855 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4856 * already held which will conflict with an existing lock order:
4858 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4860 * We don't really need mem_hotplug_lock (to hold off
4861 * slab_mem_going_offline_callback) here because slab's memory hot
4862 * unplug code doesn't destroy the kmem_cache->node[] data.
4865 #ifdef CONFIG_SLUB_DEBUG
4866 if (flags & SO_ALL) {
4867 struct kmem_cache_node *n;
4869 for_each_kmem_cache_node(s, node, n) {
4871 if (flags & SO_TOTAL)
4872 x = atomic_long_read(&n->total_objects);
4873 else if (flags & SO_OBJECTS)
4874 x = atomic_long_read(&n->total_objects) -
4875 count_partial(n, count_free);
4877 x = atomic_long_read(&n->nr_slabs);
4884 if (flags & SO_PARTIAL) {
4885 struct kmem_cache_node *n;
4887 for_each_kmem_cache_node(s, node, n) {
4888 if (flags & SO_TOTAL)
4889 x = count_partial(n, count_total);
4890 else if (flags & SO_OBJECTS)
4891 x = count_partial(n, count_inuse);
4898 x = sprintf(buf, "%lu", total);
4900 for (node = 0; node < nr_node_ids; node++)
4902 x += sprintf(buf + x, " N%d=%lu",
4906 return x + sprintf(buf + x, "\n");
4909 #ifdef CONFIG_SLUB_DEBUG
4910 static int any_slab_objects(struct kmem_cache *s)
4913 struct kmem_cache_node *n;
4915 for_each_kmem_cache_node(s, node, n)
4916 if (atomic_long_read(&n->total_objects))
4923 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4924 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4926 struct slab_attribute {
4927 struct attribute attr;
4928 ssize_t (*show)(struct kmem_cache *s, char *buf);
4929 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4932 #define SLAB_ATTR_RO(_name) \
4933 static struct slab_attribute _name##_attr = \
4934 __ATTR(_name, 0400, _name##_show, NULL)
4936 #define SLAB_ATTR(_name) \
4937 static struct slab_attribute _name##_attr = \
4938 __ATTR(_name, 0600, _name##_show, _name##_store)
4940 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4942 return sprintf(buf, "%u\n", s->size);
4944 SLAB_ATTR_RO(slab_size);
4946 static ssize_t align_show(struct kmem_cache *s, char *buf)
4948 return sprintf(buf, "%u\n", s->align);
4950 SLAB_ATTR_RO(align);
4952 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4954 return sprintf(buf, "%u\n", s->object_size);
4956 SLAB_ATTR_RO(object_size);
4958 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4960 return sprintf(buf, "%u\n", oo_objects(s->oo));
4962 SLAB_ATTR_RO(objs_per_slab);
4964 static ssize_t order_store(struct kmem_cache *s,
4965 const char *buf, size_t length)
4970 err = kstrtouint(buf, 10, &order);
4974 if (order > slub_max_order || order < slub_min_order)
4977 calculate_sizes(s, order);
4981 static ssize_t order_show(struct kmem_cache *s, char *buf)
4983 return sprintf(buf, "%u\n", oo_order(s->oo));
4987 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4989 return sprintf(buf, "%lu\n", s->min_partial);
4992 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4998 err = kstrtoul(buf, 10, &min);
5002 set_min_partial(s, min);
5005 SLAB_ATTR(min_partial);
5007 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5009 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5012 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5015 unsigned int objects;
5018 err = kstrtouint(buf, 10, &objects);
5021 if (objects && !kmem_cache_has_cpu_partial(s))
5024 slub_set_cpu_partial(s, objects);
5028 SLAB_ATTR(cpu_partial);
5030 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5034 return sprintf(buf, "%pS\n", s->ctor);
5038 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5040 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5042 SLAB_ATTR_RO(aliases);
5044 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5046 return show_slab_objects(s, buf, SO_PARTIAL);
5048 SLAB_ATTR_RO(partial);
5050 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5052 return show_slab_objects(s, buf, SO_CPU);
5054 SLAB_ATTR_RO(cpu_slabs);
5056 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5058 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5060 SLAB_ATTR_RO(objects);
5062 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5064 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5066 SLAB_ATTR_RO(objects_partial);
5068 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5075 for_each_online_cpu(cpu) {
5078 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5081 pages += page->pages;
5082 objects += page->pobjects;
5086 len = sprintf(buf, "%d(%d)", objects, pages);
5089 for_each_online_cpu(cpu) {
5092 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5094 if (page && len < PAGE_SIZE - 20)
5095 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5096 page->pobjects, page->pages);
5099 return len + sprintf(buf + len, "\n");
5101 SLAB_ATTR_RO(slabs_cpu_partial);
5103 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5105 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5108 static ssize_t reclaim_account_store(struct kmem_cache *s,
5109 const char *buf, size_t length)
5111 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5113 s->flags |= SLAB_RECLAIM_ACCOUNT;
5116 SLAB_ATTR(reclaim_account);
5118 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5122 SLAB_ATTR_RO(hwcache_align);
5124 #ifdef CONFIG_ZONE_DMA
5125 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5127 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5129 SLAB_ATTR_RO(cache_dma);
5132 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5134 return sprintf(buf, "%u\n", s->usersize);
5136 SLAB_ATTR_RO(usersize);
5138 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5140 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5142 SLAB_ATTR_RO(destroy_by_rcu);
5144 #ifdef CONFIG_SLUB_DEBUG
5145 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5147 return show_slab_objects(s, buf, SO_ALL);
5149 SLAB_ATTR_RO(slabs);
5151 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5153 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5155 SLAB_ATTR_RO(total_objects);
5157 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5159 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5162 static ssize_t sanity_checks_store(struct kmem_cache *s,
5163 const char *buf, size_t length)
5165 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5166 if (buf[0] == '1') {
5167 s->flags &= ~__CMPXCHG_DOUBLE;
5168 s->flags |= SLAB_CONSISTENCY_CHECKS;
5172 SLAB_ATTR(sanity_checks);
5174 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5176 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5179 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5183 * Tracing a merged cache is going to give confusing results
5184 * as well as cause other issues like converting a mergeable
5185 * cache into an umergeable one.
5187 if (s->refcount > 1)
5190 s->flags &= ~SLAB_TRACE;
5191 if (buf[0] == '1') {
5192 s->flags &= ~__CMPXCHG_DOUBLE;
5193 s->flags |= SLAB_TRACE;
5199 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5201 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5204 static ssize_t red_zone_store(struct kmem_cache *s,
5205 const char *buf, size_t length)
5207 if (any_slab_objects(s))
5210 s->flags &= ~SLAB_RED_ZONE;
5211 if (buf[0] == '1') {
5212 s->flags |= SLAB_RED_ZONE;
5214 calculate_sizes(s, -1);
5217 SLAB_ATTR(red_zone);
5219 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5221 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5224 static ssize_t poison_store(struct kmem_cache *s,
5225 const char *buf, size_t length)
5227 if (any_slab_objects(s))
5230 s->flags &= ~SLAB_POISON;
5231 if (buf[0] == '1') {
5232 s->flags |= SLAB_POISON;
5234 calculate_sizes(s, -1);
5239 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5241 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5244 static ssize_t store_user_store(struct kmem_cache *s,
5245 const char *buf, size_t length)
5247 if (any_slab_objects(s))
5250 s->flags &= ~SLAB_STORE_USER;
5251 if (buf[0] == '1') {
5252 s->flags &= ~__CMPXCHG_DOUBLE;
5253 s->flags |= SLAB_STORE_USER;
5255 calculate_sizes(s, -1);
5258 SLAB_ATTR(store_user);
5260 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5265 static ssize_t validate_store(struct kmem_cache *s,
5266 const char *buf, size_t length)
5270 if (buf[0] == '1') {
5271 ret = validate_slab_cache(s);
5277 SLAB_ATTR(validate);
5279 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5281 if (!(s->flags & SLAB_STORE_USER))
5283 return list_locations(s, buf, TRACK_ALLOC);
5285 SLAB_ATTR_RO(alloc_calls);
5287 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5289 if (!(s->flags & SLAB_STORE_USER))
5291 return list_locations(s, buf, TRACK_FREE);
5293 SLAB_ATTR_RO(free_calls);
5294 #endif /* CONFIG_SLUB_DEBUG */
5296 #ifdef CONFIG_FAILSLAB
5297 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5299 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5302 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5305 if (s->refcount > 1)
5308 s->flags &= ~SLAB_FAILSLAB;
5310 s->flags |= SLAB_FAILSLAB;
5313 SLAB_ATTR(failslab);
5316 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5321 static ssize_t shrink_store(struct kmem_cache *s,
5322 const char *buf, size_t length)
5325 kmem_cache_shrink(s);
5333 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5335 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5338 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5339 const char *buf, size_t length)
5344 err = kstrtouint(buf, 10, &ratio);
5350 s->remote_node_defrag_ratio = ratio * 10;
5354 SLAB_ATTR(remote_node_defrag_ratio);
5357 #ifdef CONFIG_SLUB_STATS
5358 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5360 unsigned long sum = 0;
5363 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5368 for_each_online_cpu(cpu) {
5369 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5375 len = sprintf(buf, "%lu", sum);
5378 for_each_online_cpu(cpu) {
5379 if (data[cpu] && len < PAGE_SIZE - 20)
5380 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5384 return len + sprintf(buf + len, "\n");
5387 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5391 for_each_online_cpu(cpu)
5392 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5395 #define STAT_ATTR(si, text) \
5396 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5398 return show_stat(s, buf, si); \
5400 static ssize_t text##_store(struct kmem_cache *s, \
5401 const char *buf, size_t length) \
5403 if (buf[0] != '0') \
5405 clear_stat(s, si); \
5410 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5411 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5412 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5413 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5414 STAT_ATTR(FREE_FROZEN, free_frozen);
5415 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5416 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5417 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5418 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5419 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5420 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5421 STAT_ATTR(FREE_SLAB, free_slab);
5422 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5423 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5424 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5425 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5426 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5427 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5428 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5429 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5430 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5431 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5432 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5433 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5434 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5435 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5438 static struct attribute *slab_attrs[] = {
5439 &slab_size_attr.attr,
5440 &object_size_attr.attr,
5441 &objs_per_slab_attr.attr,
5443 &min_partial_attr.attr,
5444 &cpu_partial_attr.attr,
5446 &objects_partial_attr.attr,
5448 &cpu_slabs_attr.attr,
5452 &hwcache_align_attr.attr,
5453 &reclaim_account_attr.attr,
5454 &destroy_by_rcu_attr.attr,
5456 &slabs_cpu_partial_attr.attr,
5457 #ifdef CONFIG_SLUB_DEBUG
5458 &total_objects_attr.attr,
5460 &sanity_checks_attr.attr,
5462 &red_zone_attr.attr,
5464 &store_user_attr.attr,
5465 &validate_attr.attr,
5466 &alloc_calls_attr.attr,
5467 &free_calls_attr.attr,
5469 #ifdef CONFIG_ZONE_DMA
5470 &cache_dma_attr.attr,
5473 &remote_node_defrag_ratio_attr.attr,
5475 #ifdef CONFIG_SLUB_STATS
5476 &alloc_fastpath_attr.attr,
5477 &alloc_slowpath_attr.attr,
5478 &free_fastpath_attr.attr,
5479 &free_slowpath_attr.attr,
5480 &free_frozen_attr.attr,
5481 &free_add_partial_attr.attr,
5482 &free_remove_partial_attr.attr,
5483 &alloc_from_partial_attr.attr,
5484 &alloc_slab_attr.attr,
5485 &alloc_refill_attr.attr,
5486 &alloc_node_mismatch_attr.attr,
5487 &free_slab_attr.attr,
5488 &cpuslab_flush_attr.attr,
5489 &deactivate_full_attr.attr,
5490 &deactivate_empty_attr.attr,
5491 &deactivate_to_head_attr.attr,
5492 &deactivate_to_tail_attr.attr,
5493 &deactivate_remote_frees_attr.attr,
5494 &deactivate_bypass_attr.attr,
5495 &order_fallback_attr.attr,
5496 &cmpxchg_double_fail_attr.attr,
5497 &cmpxchg_double_cpu_fail_attr.attr,
5498 &cpu_partial_alloc_attr.attr,
5499 &cpu_partial_free_attr.attr,
5500 &cpu_partial_node_attr.attr,
5501 &cpu_partial_drain_attr.attr,
5503 #ifdef CONFIG_FAILSLAB
5504 &failslab_attr.attr,
5506 &usersize_attr.attr,
5511 static const struct attribute_group slab_attr_group = {
5512 .attrs = slab_attrs,
5515 static ssize_t slab_attr_show(struct kobject *kobj,
5516 struct attribute *attr,
5519 struct slab_attribute *attribute;
5520 struct kmem_cache *s;
5523 attribute = to_slab_attr(attr);
5526 if (!attribute->show)
5529 err = attribute->show(s, buf);
5534 static ssize_t slab_attr_store(struct kobject *kobj,
5535 struct attribute *attr,
5536 const char *buf, size_t len)
5538 struct slab_attribute *attribute;
5539 struct kmem_cache *s;
5542 attribute = to_slab_attr(attr);
5545 if (!attribute->store)
5548 err = attribute->store(s, buf, len);
5550 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5551 struct kmem_cache *c;
5553 mutex_lock(&slab_mutex);
5554 if (s->max_attr_size < len)
5555 s->max_attr_size = len;
5558 * This is a best effort propagation, so this function's return
5559 * value will be determined by the parent cache only. This is
5560 * basically because not all attributes will have a well
5561 * defined semantics for rollbacks - most of the actions will
5562 * have permanent effects.
5564 * Returning the error value of any of the children that fail
5565 * is not 100 % defined, in the sense that users seeing the
5566 * error code won't be able to know anything about the state of
5569 * Only returning the error code for the parent cache at least
5570 * has well defined semantics. The cache being written to
5571 * directly either failed or succeeded, in which case we loop
5572 * through the descendants with best-effort propagation.
5574 for_each_memcg_cache(c, s)
5575 attribute->store(c, buf, len);
5576 mutex_unlock(&slab_mutex);
5582 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5586 char *buffer = NULL;
5587 struct kmem_cache *root_cache;
5589 if (is_root_cache(s))
5592 root_cache = s->memcg_params.root_cache;
5595 * This mean this cache had no attribute written. Therefore, no point
5596 * in copying default values around
5598 if (!root_cache->max_attr_size)
5601 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5604 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5607 if (!attr || !attr->store || !attr->show)
5611 * It is really bad that we have to allocate here, so we will
5612 * do it only as a fallback. If we actually allocate, though,
5613 * we can just use the allocated buffer until the end.
5615 * Most of the slub attributes will tend to be very small in
5616 * size, but sysfs allows buffers up to a page, so they can
5617 * theoretically happen.
5621 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5622 !IS_ENABLED(CONFIG_SLUB_STATS))
5625 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5626 if (WARN_ON(!buffer))
5631 len = attr->show(root_cache, buf);
5633 attr->store(s, buf, len);
5637 free_page((unsigned long)buffer);
5641 static void kmem_cache_release(struct kobject *k)
5643 slab_kmem_cache_release(to_slab(k));
5646 static const struct sysfs_ops slab_sysfs_ops = {
5647 .show = slab_attr_show,
5648 .store = slab_attr_store,
5651 static struct kobj_type slab_ktype = {
5652 .sysfs_ops = &slab_sysfs_ops,
5653 .release = kmem_cache_release,
5656 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5658 struct kobj_type *ktype = get_ktype(kobj);
5660 if (ktype == &slab_ktype)
5665 static const struct kset_uevent_ops slab_uevent_ops = {
5666 .filter = uevent_filter,
5669 static struct kset *slab_kset;
5671 static inline struct kset *cache_kset(struct kmem_cache *s)
5674 if (!is_root_cache(s))
5675 return s->memcg_params.root_cache->memcg_kset;
5680 #define ID_STR_LENGTH 64
5682 /* Create a unique string id for a slab cache:
5684 * Format :[flags-]size
5686 static char *create_unique_id(struct kmem_cache *s)
5688 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5695 * First flags affecting slabcache operations. We will only
5696 * get here for aliasable slabs so we do not need to support
5697 * too many flags. The flags here must cover all flags that
5698 * are matched during merging to guarantee that the id is
5701 if (s->flags & SLAB_CACHE_DMA)
5703 if (s->flags & SLAB_CACHE_DMA32)
5705 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5707 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5709 if (s->flags & SLAB_ACCOUNT)
5713 p += sprintf(p, "%07u", s->size);
5715 BUG_ON(p > name + ID_STR_LENGTH - 1);
5719 static void sysfs_slab_remove_workfn(struct work_struct *work)
5721 struct kmem_cache *s =
5722 container_of(work, struct kmem_cache, kobj_remove_work);
5724 if (!s->kobj.state_in_sysfs)
5726 * For a memcg cache, this may be called during
5727 * deactivation and again on shutdown. Remove only once.
5728 * A cache is never shut down before deactivation is
5729 * complete, so no need to worry about synchronization.
5734 kset_unregister(s->memcg_kset);
5736 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5738 kobject_put(&s->kobj);
5741 static int sysfs_slab_add(struct kmem_cache *s)
5745 struct kset *kset = cache_kset(s);
5746 int unmergeable = slab_unmergeable(s);
5748 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5751 kobject_init(&s->kobj, &slab_ktype);
5755 if (!unmergeable && disable_higher_order_debug &&
5756 (slub_debug & DEBUG_METADATA_FLAGS))
5761 * Slabcache can never be merged so we can use the name proper.
5762 * This is typically the case for debug situations. In that
5763 * case we can catch duplicate names easily.
5765 sysfs_remove_link(&slab_kset->kobj, s->name);
5769 * Create a unique name for the slab as a target
5772 name = create_unique_id(s);
5775 s->kobj.kset = kset;
5776 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5780 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5785 if (is_root_cache(s) && memcg_sysfs_enabled) {
5786 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5787 if (!s->memcg_kset) {
5794 kobject_uevent(&s->kobj, KOBJ_ADD);
5796 /* Setup first alias */
5797 sysfs_slab_alias(s, s->name);
5804 kobject_del(&s->kobj);
5808 static void sysfs_slab_remove(struct kmem_cache *s)
5810 if (slab_state < FULL)
5812 * Sysfs has not been setup yet so no need to remove the
5817 kobject_get(&s->kobj);
5818 schedule_work(&s->kobj_remove_work);
5821 void sysfs_slab_unlink(struct kmem_cache *s)
5823 if (slab_state >= FULL)
5824 kobject_del(&s->kobj);
5827 void sysfs_slab_release(struct kmem_cache *s)
5829 if (slab_state >= FULL)
5830 kobject_put(&s->kobj);
5834 * Need to buffer aliases during bootup until sysfs becomes
5835 * available lest we lose that information.
5837 struct saved_alias {
5838 struct kmem_cache *s;
5840 struct saved_alias *next;
5843 static struct saved_alias *alias_list;
5845 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5847 struct saved_alias *al;
5849 if (slab_state == FULL) {
5851 * If we have a leftover link then remove it.
5853 sysfs_remove_link(&slab_kset->kobj, name);
5854 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5857 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5863 al->next = alias_list;
5868 static int __init slab_sysfs_init(void)
5870 struct kmem_cache *s;
5873 mutex_lock(&slab_mutex);
5875 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5877 mutex_unlock(&slab_mutex);
5878 pr_err("Cannot register slab subsystem.\n");
5884 list_for_each_entry(s, &slab_caches, list) {
5885 err = sysfs_slab_add(s);
5887 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5891 while (alias_list) {
5892 struct saved_alias *al = alias_list;
5894 alias_list = alias_list->next;
5895 err = sysfs_slab_alias(al->s, al->name);
5897 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5902 mutex_unlock(&slab_mutex);
5907 __initcall(slab_sysfs_init);
5908 #endif /* CONFIG_SYSFS */
5911 * The /proc/slabinfo ABI
5913 #ifdef CONFIG_SLUB_DEBUG
5914 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5916 unsigned long nr_slabs = 0;
5917 unsigned long nr_objs = 0;
5918 unsigned long nr_free = 0;
5920 struct kmem_cache_node *n;
5922 for_each_kmem_cache_node(s, node, n) {
5923 nr_slabs += node_nr_slabs(n);
5924 nr_objs += node_nr_objs(n);
5925 nr_free += count_partial(n, count_free);
5928 sinfo->active_objs = nr_objs - nr_free;
5929 sinfo->num_objs = nr_objs;
5930 sinfo->active_slabs = nr_slabs;
5931 sinfo->num_slabs = nr_slabs;
5932 sinfo->objects_per_slab = oo_objects(s->oo);
5933 sinfo->cache_order = oo_order(s->oo);
5936 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5940 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5941 size_t count, loff_t *ppos)
5945 #endif /* CONFIG_SLUB_DEBUG */