2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
136 * Issues still to be resolved:
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 * - Variable sizing of the per node arrays
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 #define MIN_PARTIAL 5
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in use.
160 #define MAX_PARTIAL 10
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
173 * Set of flags that will prevent slab merging
175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
183 #define OO_MASK ((1 << OO_SHIFT) - 1)
184 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
186 /* Internal SLUB flags */
187 #define __OBJECT_POISON 0x80000000UL /* Poison object */
188 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
191 static struct notifier_block slab_notifier;
195 * Tracking user of a slab.
197 #define TRACK_ADDRS_COUNT 16
199 unsigned long addr; /* Called from address */
200 #ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
208 enum track_item { TRACK_ALLOC, TRACK_FREE };
211 static int sysfs_slab_add(struct kmem_cache *);
212 static int sysfs_slab_alias(struct kmem_cache *, const char *);
213 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
215 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
216 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
218 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
221 static inline void stat(const struct kmem_cache *s, enum stat_item si)
223 #ifdef CONFIG_SLUB_STATS
225 * The rmw is racy on a preemptible kernel but this is acceptable, so
226 * avoid this_cpu_add()'s irq-disable overhead.
228 raw_cpu_inc(s->cpu_slab->stat[si]);
232 /********************************************************************
233 * Core slab cache functions
234 *******************************************************************/
236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
238 return s->node[node];
241 /* Verify that a pointer has an address that is valid within a slab page */
242 static inline int check_valid_pointer(struct kmem_cache *s,
243 struct page *page, const void *object)
250 base = page_address(page);
251 if (object < base || object >= base + page->objects * s->size ||
252 (object - base) % s->size) {
259 static inline void *get_freepointer(struct kmem_cache *s, void *object)
261 return *(void **)(object + s->offset);
264 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
266 prefetch(object + s->offset);
269 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
273 #ifdef CONFIG_DEBUG_PAGEALLOC
274 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
276 p = get_freepointer(s, object);
281 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
283 *(void **)(object + s->offset) = fp;
286 /* Loop over all objects in a slab */
287 #define for_each_object(__p, __s, __addr, __objects) \
288 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
291 /* Determine object index from a given position */
292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
294 return (p - addr) / s->size;
297 static inline size_t slab_ksize(const struct kmem_cache *s)
299 #ifdef CONFIG_SLUB_DEBUG
301 * Debugging requires use of the padding between object
302 * and whatever may come after it.
304 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
305 return s->object_size;
309 * If we have the need to store the freelist pointer
310 * back there or track user information then we can
311 * only use the space before that information.
313 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
316 * Else we can use all the padding etc for the allocation
321 static inline int order_objects(int order, unsigned long size, int reserved)
323 return ((PAGE_SIZE << order) - reserved) / size;
326 static inline struct kmem_cache_order_objects oo_make(int order,
327 unsigned long size, int reserved)
329 struct kmem_cache_order_objects x = {
330 (order << OO_SHIFT) + order_objects(order, size, reserved)
336 static inline int oo_order(struct kmem_cache_order_objects x)
338 return x.x >> OO_SHIFT;
341 static inline int oo_objects(struct kmem_cache_order_objects x)
343 return x.x & OO_MASK;
347 * Per slab locking using the pagelock
349 static __always_inline void slab_lock(struct page *page)
351 bit_spin_lock(PG_locked, &page->flags);
354 static __always_inline void slab_unlock(struct page *page)
356 __bit_spin_unlock(PG_locked, &page->flags);
359 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
362 tmp.counters = counters_new;
364 * page->counters can cover frozen/inuse/objects as well
365 * as page->_count. If we assign to ->counters directly
366 * we run the risk of losing updates to page->_count, so
367 * be careful and only assign to the fields we need.
369 page->frozen = tmp.frozen;
370 page->inuse = tmp.inuse;
371 page->objects = tmp.objects;
374 /* Interrupts must be disabled (for the fallback code to work right) */
375 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
380 VM_BUG_ON(!irqs_disabled());
381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s->flags & __CMPXCHG_DOUBLE) {
384 if (cmpxchg_double(&page->freelist, &page->counters,
385 freelist_old, counters_old,
386 freelist_new, counters_new))
392 if (page->freelist == freelist_old &&
393 page->counters == counters_old) {
394 page->freelist = freelist_new;
395 set_page_slub_counters(page, counters_new);
403 stat(s, CMPXCHG_DOUBLE_FAIL);
405 #ifdef SLUB_DEBUG_CMPXCHG
406 pr_info("%s %s: cmpxchg double redo ", n, s->name);
412 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413 void *freelist_old, unsigned long counters_old,
414 void *freelist_new, unsigned long counters_new,
417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s->flags & __CMPXCHG_DOUBLE) {
420 if (cmpxchg_double(&page->freelist, &page->counters,
421 freelist_old, counters_old,
422 freelist_new, counters_new))
429 local_irq_save(flags);
431 if (page->freelist == freelist_old &&
432 page->counters == counters_old) {
433 page->freelist = freelist_new;
434 set_page_slub_counters(page, counters_new);
436 local_irq_restore(flags);
440 local_irq_restore(flags);
444 stat(s, CMPXCHG_DOUBLE_FAIL);
446 #ifdef SLUB_DEBUG_CMPXCHG
447 pr_info("%s %s: cmpxchg double redo ", n, s->name);
453 #ifdef CONFIG_SLUB_DEBUG
455 * Determine a map of object in use on a page.
457 * Node listlock must be held to guarantee that the page does
458 * not vanish from under us.
460 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
463 void *addr = page_address(page);
465 for (p = page->freelist; p; p = get_freepointer(s, p))
466 set_bit(slab_index(p, s, addr), map);
472 #ifdef CONFIG_SLUB_DEBUG_ON
473 static int slub_debug = DEBUG_DEFAULT_FLAGS;
475 static int slub_debug;
478 static char *slub_debug_slabs;
479 static int disable_higher_order_debug;
484 static void print_section(char *text, u8 *addr, unsigned int length)
486 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
490 static struct track *get_track(struct kmem_cache *s, void *object,
491 enum track_item alloc)
496 p = object + s->offset + sizeof(void *);
498 p = object + s->inuse;
503 static void set_track(struct kmem_cache *s, void *object,
504 enum track_item alloc, unsigned long addr)
506 struct track *p = get_track(s, object, alloc);
509 #ifdef CONFIG_STACKTRACE
510 struct stack_trace trace;
513 trace.nr_entries = 0;
514 trace.max_entries = TRACK_ADDRS_COUNT;
515 trace.entries = p->addrs;
517 save_stack_trace(&trace);
519 /* See rant in lockdep.c */
520 if (trace.nr_entries != 0 &&
521 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
524 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
528 p->cpu = smp_processor_id();
529 p->pid = current->pid;
532 memset(p, 0, sizeof(struct track));
535 static void init_tracking(struct kmem_cache *s, void *object)
537 if (!(s->flags & SLAB_STORE_USER))
540 set_track(s, object, TRACK_FREE, 0UL);
541 set_track(s, object, TRACK_ALLOC, 0UL);
544 static void print_track(const char *s, struct track *t)
549 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
550 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
551 #ifdef CONFIG_STACKTRACE
554 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
556 pr_err("\t%pS\n", (void *)t->addrs[i]);
563 static void print_tracking(struct kmem_cache *s, void *object)
565 if (!(s->flags & SLAB_STORE_USER))
568 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
569 print_track("Freed", get_track(s, object, TRACK_FREE));
572 static void print_page_info(struct page *page)
574 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
575 page, page->objects, page->inuse, page->freelist, page->flags);
579 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
581 struct va_format vaf;
587 pr_err("=============================================================================\n");
588 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
589 pr_err("-----------------------------------------------------------------------------\n\n");
591 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
595 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
597 struct va_format vaf;
603 pr_err("FIX %s: %pV\n", s->name, &vaf);
607 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
609 unsigned int off; /* Offset of last byte */
610 u8 *addr = page_address(page);
612 print_tracking(s, p);
614 print_page_info(page);
616 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
617 p, p - addr, get_freepointer(s, p));
620 print_section("Bytes b4 ", p - 16, 16);
622 print_section("Object ", p, min_t(unsigned long, s->object_size,
624 if (s->flags & SLAB_RED_ZONE)
625 print_section("Redzone ", p + s->object_size,
626 s->inuse - s->object_size);
629 off = s->offset + sizeof(void *);
633 if (s->flags & SLAB_STORE_USER)
634 off += 2 * sizeof(struct track);
637 /* Beginning of the filler is the free pointer */
638 print_section("Padding ", p + off, s->size - off);
643 static void object_err(struct kmem_cache *s, struct page *page,
644 u8 *object, char *reason)
646 slab_bug(s, "%s", reason);
647 print_trailer(s, page, object);
650 static void slab_err(struct kmem_cache *s, struct page *page,
651 const char *fmt, ...)
657 vsnprintf(buf, sizeof(buf), fmt, args);
659 slab_bug(s, "%s", buf);
660 print_page_info(page);
664 static void init_object(struct kmem_cache *s, void *object, u8 val)
668 if (s->flags & __OBJECT_POISON) {
669 memset(p, POISON_FREE, s->object_size - 1);
670 p[s->object_size - 1] = POISON_END;
673 if (s->flags & SLAB_RED_ZONE)
674 memset(p + s->object_size, val, s->inuse - s->object_size);
677 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
678 void *from, void *to)
680 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
681 memset(from, data, to - from);
684 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
685 u8 *object, char *what,
686 u8 *start, unsigned int value, unsigned int bytes)
691 fault = memchr_inv(start, value, bytes);
696 while (end > fault && end[-1] == value)
699 slab_bug(s, "%s overwritten", what);
700 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
701 fault, end - 1, fault[0], value);
702 print_trailer(s, page, object);
704 restore_bytes(s, what, value, fault, end);
712 * Bytes of the object to be managed.
713 * If the freepointer may overlay the object then the free
714 * pointer is the first word of the object.
716 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
719 * object + s->object_size
720 * Padding to reach word boundary. This is also used for Redzoning.
721 * Padding is extended by another word if Redzoning is enabled and
722 * object_size == inuse.
724 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
725 * 0xcc (RED_ACTIVE) for objects in use.
728 * Meta data starts here.
730 * A. Free pointer (if we cannot overwrite object on free)
731 * B. Tracking data for SLAB_STORE_USER
732 * C. Padding to reach required alignment boundary or at mininum
733 * one word if debugging is on to be able to detect writes
734 * before the word boundary.
736 * Padding is done using 0x5a (POISON_INUSE)
739 * Nothing is used beyond s->size.
741 * If slabcaches are merged then the object_size and inuse boundaries are mostly
742 * ignored. And therefore no slab options that rely on these boundaries
743 * may be used with merged slabcaches.
746 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
748 unsigned long off = s->inuse; /* The end of info */
751 /* Freepointer is placed after the object. */
752 off += sizeof(void *);
754 if (s->flags & SLAB_STORE_USER)
755 /* We also have user information there */
756 off += 2 * sizeof(struct track);
761 return check_bytes_and_report(s, page, p, "Object padding",
762 p + off, POISON_INUSE, s->size - off);
765 /* Check the pad bytes at the end of a slab page */
766 static int slab_pad_check(struct kmem_cache *s, struct page *page)
774 if (!(s->flags & SLAB_POISON))
777 start = page_address(page);
778 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
779 end = start + length;
780 remainder = length % s->size;
784 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
787 while (end > fault && end[-1] == POISON_INUSE)
790 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
791 print_section("Padding ", end - remainder, remainder);
793 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
797 static int check_object(struct kmem_cache *s, struct page *page,
798 void *object, u8 val)
801 u8 *endobject = object + s->object_size;
803 if (s->flags & SLAB_RED_ZONE) {
804 if (!check_bytes_and_report(s, page, object, "Redzone",
805 endobject, val, s->inuse - s->object_size))
808 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
809 check_bytes_and_report(s, page, p, "Alignment padding",
810 endobject, POISON_INUSE,
811 s->inuse - s->object_size);
815 if (s->flags & SLAB_POISON) {
816 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
817 (!check_bytes_and_report(s, page, p, "Poison", p,
818 POISON_FREE, s->object_size - 1) ||
819 !check_bytes_and_report(s, page, p, "Poison",
820 p + s->object_size - 1, POISON_END, 1)))
823 * check_pad_bytes cleans up on its own.
825 check_pad_bytes(s, page, p);
828 if (!s->offset && val == SLUB_RED_ACTIVE)
830 * Object and freepointer overlap. Cannot check
831 * freepointer while object is allocated.
835 /* Check free pointer validity */
836 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
837 object_err(s, page, p, "Freepointer corrupt");
839 * No choice but to zap it and thus lose the remainder
840 * of the free objects in this slab. May cause
841 * another error because the object count is now wrong.
843 set_freepointer(s, p, NULL);
849 static int check_slab(struct kmem_cache *s, struct page *page)
853 VM_BUG_ON(!irqs_disabled());
855 if (!PageSlab(page)) {
856 slab_err(s, page, "Not a valid slab page");
860 maxobj = order_objects(compound_order(page), s->size, s->reserved);
861 if (page->objects > maxobj) {
862 slab_err(s, page, "objects %u > max %u",
863 s->name, page->objects, maxobj);
866 if (page->inuse > page->objects) {
867 slab_err(s, page, "inuse %u > max %u",
868 s->name, page->inuse, page->objects);
871 /* Slab_pad_check fixes things up after itself */
872 slab_pad_check(s, page);
877 * Determine if a certain object on a page is on the freelist. Must hold the
878 * slab lock to guarantee that the chains are in a consistent state.
880 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
885 unsigned long max_objects;
888 while (fp && nr <= page->objects) {
891 if (!check_valid_pointer(s, page, fp)) {
893 object_err(s, page, object,
894 "Freechain corrupt");
895 set_freepointer(s, object, NULL);
897 slab_err(s, page, "Freepointer corrupt");
898 page->freelist = NULL;
899 page->inuse = page->objects;
900 slab_fix(s, "Freelist cleared");
906 fp = get_freepointer(s, object);
910 max_objects = order_objects(compound_order(page), s->size, s->reserved);
911 if (max_objects > MAX_OBJS_PER_PAGE)
912 max_objects = MAX_OBJS_PER_PAGE;
914 if (page->objects != max_objects) {
915 slab_err(s, page, "Wrong number of objects. Found %d but "
916 "should be %d", page->objects, max_objects);
917 page->objects = max_objects;
918 slab_fix(s, "Number of objects adjusted.");
920 if (page->inuse != page->objects - nr) {
921 slab_err(s, page, "Wrong object count. Counter is %d but "
922 "counted were %d", page->inuse, page->objects - nr);
923 page->inuse = page->objects - nr;
924 slab_fix(s, "Object count adjusted.");
926 return search == NULL;
929 static void trace(struct kmem_cache *s, struct page *page, void *object,
932 if (s->flags & SLAB_TRACE) {
933 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
935 alloc ? "alloc" : "free",
940 print_section("Object ", (void *)object,
948 * Hooks for other subsystems that check memory allocations. In a typical
949 * production configuration these hooks all should produce no code at all.
951 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
953 kmemleak_alloc(ptr, size, 1, flags);
956 static inline void kfree_hook(const void *x)
961 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
963 flags &= gfp_allowed_mask;
964 lockdep_trace_alloc(flags);
965 might_sleep_if(flags & __GFP_WAIT);
967 return should_failslab(s->object_size, flags, s->flags);
970 static inline void slab_post_alloc_hook(struct kmem_cache *s,
971 gfp_t flags, void *object)
973 flags &= gfp_allowed_mask;
974 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
975 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
978 static inline void slab_free_hook(struct kmem_cache *s, void *x)
980 kmemleak_free_recursive(x, s->flags);
983 * Trouble is that we may no longer disable interrupts in the fast path
984 * So in order to make the debug calls that expect irqs to be
985 * disabled we need to disable interrupts temporarily.
987 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
991 local_irq_save(flags);
992 kmemcheck_slab_free(s, x, s->object_size);
993 debug_check_no_locks_freed(x, s->object_size);
994 local_irq_restore(flags);
997 if (!(s->flags & SLAB_DEBUG_OBJECTS))
998 debug_check_no_obj_freed(x, s->object_size);
1002 * Tracking of fully allocated slabs for debugging purposes.
1004 static void add_full(struct kmem_cache *s,
1005 struct kmem_cache_node *n, struct page *page)
1007 if (!(s->flags & SLAB_STORE_USER))
1010 lockdep_assert_held(&n->list_lock);
1011 list_add(&page->lru, &n->full);
1014 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1016 if (!(s->flags & SLAB_STORE_USER))
1019 lockdep_assert_held(&n->list_lock);
1020 list_del(&page->lru);
1023 /* Tracking of the number of slabs for debugging purposes */
1024 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1026 struct kmem_cache_node *n = get_node(s, node);
1028 return atomic_long_read(&n->nr_slabs);
1031 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1033 return atomic_long_read(&n->nr_slabs);
1036 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1038 struct kmem_cache_node *n = get_node(s, node);
1041 * May be called early in order to allocate a slab for the
1042 * kmem_cache_node structure. Solve the chicken-egg
1043 * dilemma by deferring the increment of the count during
1044 * bootstrap (see early_kmem_cache_node_alloc).
1047 atomic_long_inc(&n->nr_slabs);
1048 atomic_long_add(objects, &n->total_objects);
1051 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1053 struct kmem_cache_node *n = get_node(s, node);
1055 atomic_long_dec(&n->nr_slabs);
1056 atomic_long_sub(objects, &n->total_objects);
1059 /* Object debug checks for alloc/free paths */
1060 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1063 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1066 init_object(s, object, SLUB_RED_INACTIVE);
1067 init_tracking(s, object);
1070 static noinline int alloc_debug_processing(struct kmem_cache *s,
1072 void *object, unsigned long addr)
1074 if (!check_slab(s, page))
1077 if (!check_valid_pointer(s, page, object)) {
1078 object_err(s, page, object, "Freelist Pointer check fails");
1082 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1085 /* Success perform special debug activities for allocs */
1086 if (s->flags & SLAB_STORE_USER)
1087 set_track(s, object, TRACK_ALLOC, addr);
1088 trace(s, page, object, 1);
1089 init_object(s, object, SLUB_RED_ACTIVE);
1093 if (PageSlab(page)) {
1095 * If this is a slab page then lets do the best we can
1096 * to avoid issues in the future. Marking all objects
1097 * as used avoids touching the remaining objects.
1099 slab_fix(s, "Marking all objects used");
1100 page->inuse = page->objects;
1101 page->freelist = NULL;
1106 static noinline struct kmem_cache_node *free_debug_processing(
1107 struct kmem_cache *s, struct page *page, void *object,
1108 unsigned long addr, unsigned long *flags)
1110 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1112 spin_lock_irqsave(&n->list_lock, *flags);
1115 if (!check_slab(s, page))
1118 if (!check_valid_pointer(s, page, object)) {
1119 slab_err(s, page, "Invalid object pointer 0x%p", object);
1123 if (on_freelist(s, page, object)) {
1124 object_err(s, page, object, "Object already free");
1128 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1131 if (unlikely(s != page->slab_cache)) {
1132 if (!PageSlab(page)) {
1133 slab_err(s, page, "Attempt to free object(0x%p) "
1134 "outside of slab", object);
1135 } else if (!page->slab_cache) {
1136 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1140 object_err(s, page, object,
1141 "page slab pointer corrupt.");
1145 if (s->flags & SLAB_STORE_USER)
1146 set_track(s, object, TRACK_FREE, addr);
1147 trace(s, page, object, 0);
1148 init_object(s, object, SLUB_RED_INACTIVE);
1152 * Keep node_lock to preserve integrity
1153 * until the object is actually freed
1159 spin_unlock_irqrestore(&n->list_lock, *flags);
1160 slab_fix(s, "Object at 0x%p not freed", object);
1164 static int __init setup_slub_debug(char *str)
1166 slub_debug = DEBUG_DEFAULT_FLAGS;
1167 if (*str++ != '=' || !*str)
1169 * No options specified. Switch on full debugging.
1175 * No options but restriction on slabs. This means full
1176 * debugging for slabs matching a pattern.
1180 if (tolower(*str) == 'o') {
1182 * Avoid enabling debugging on caches if its minimum order
1183 * would increase as a result.
1185 disable_higher_order_debug = 1;
1192 * Switch off all debugging measures.
1197 * Determine which debug features should be switched on
1199 for (; *str && *str != ','; str++) {
1200 switch (tolower(*str)) {
1202 slub_debug |= SLAB_DEBUG_FREE;
1205 slub_debug |= SLAB_RED_ZONE;
1208 slub_debug |= SLAB_POISON;
1211 slub_debug |= SLAB_STORE_USER;
1214 slub_debug |= SLAB_TRACE;
1217 slub_debug |= SLAB_FAILSLAB;
1220 pr_err("slub_debug option '%c' unknown. skipped\n",
1227 slub_debug_slabs = str + 1;
1232 __setup("slub_debug", setup_slub_debug);
1234 static unsigned long kmem_cache_flags(unsigned long object_size,
1235 unsigned long flags, const char *name,
1236 void (*ctor)(void *))
1239 * Enable debugging if selected on the kernel commandline.
1241 if (slub_debug && (!slub_debug_slabs || (name &&
1242 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1243 flags |= slub_debug;
1248 static inline void setup_object_debug(struct kmem_cache *s,
1249 struct page *page, void *object) {}
1251 static inline int alloc_debug_processing(struct kmem_cache *s,
1252 struct page *page, void *object, unsigned long addr) { return 0; }
1254 static inline struct kmem_cache_node *free_debug_processing(
1255 struct kmem_cache *s, struct page *page, void *object,
1256 unsigned long addr, unsigned long *flags) { return NULL; }
1258 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1260 static inline int check_object(struct kmem_cache *s, struct page *page,
1261 void *object, u8 val) { return 1; }
1262 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1263 struct page *page) {}
1264 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1265 struct page *page) {}
1266 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1267 unsigned long flags, const char *name,
1268 void (*ctor)(void *))
1272 #define slub_debug 0
1274 #define disable_higher_order_debug 0
1276 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1278 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1280 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1282 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1285 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1287 kmemleak_alloc(ptr, size, 1, flags);
1290 static inline void kfree_hook(const void *x)
1295 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1298 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1301 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1302 flags & gfp_allowed_mask);
1305 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1307 kmemleak_free_recursive(x, s->flags);
1310 #endif /* CONFIG_SLUB_DEBUG */
1313 * Slab allocation and freeing
1315 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1316 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1319 int order = oo_order(oo);
1321 flags |= __GFP_NOTRACK;
1323 if (memcg_charge_slab(s, flags, order))
1326 if (node == NUMA_NO_NODE)
1327 page = alloc_pages(flags, order);
1329 page = alloc_pages_exact_node(node, flags, order);
1332 memcg_uncharge_slab(s, order);
1337 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1340 struct kmem_cache_order_objects oo = s->oo;
1343 flags &= gfp_allowed_mask;
1345 if (flags & __GFP_WAIT)
1348 flags |= s->allocflags;
1351 * Let the initial higher-order allocation fail under memory pressure
1352 * so we fall-back to the minimum order allocation.
1354 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1356 page = alloc_slab_page(s, alloc_gfp, node, oo);
1357 if (unlikely(!page)) {
1361 * Allocation may have failed due to fragmentation.
1362 * Try a lower order alloc if possible
1364 page = alloc_slab_page(s, alloc_gfp, node, oo);
1367 stat(s, ORDER_FALLBACK);
1370 if (kmemcheck_enabled && page
1371 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1372 int pages = 1 << oo_order(oo);
1374 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1377 * Objects from caches that have a constructor don't get
1378 * cleared when they're allocated, so we need to do it here.
1381 kmemcheck_mark_uninitialized_pages(page, pages);
1383 kmemcheck_mark_unallocated_pages(page, pages);
1386 if (flags & __GFP_WAIT)
1387 local_irq_disable();
1391 page->objects = oo_objects(oo);
1392 mod_zone_page_state(page_zone(page),
1393 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1394 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1400 static void setup_object(struct kmem_cache *s, struct page *page,
1403 setup_object_debug(s, page, object);
1404 if (unlikely(s->ctor))
1408 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1416 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1418 page = allocate_slab(s,
1419 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1423 order = compound_order(page);
1424 inc_slabs_node(s, page_to_nid(page), page->objects);
1425 memcg_bind_pages(s, order);
1426 page->slab_cache = s;
1427 __SetPageSlab(page);
1428 if (page->pfmemalloc)
1429 SetPageSlabPfmemalloc(page);
1431 start = page_address(page);
1433 if (unlikely(s->flags & SLAB_POISON))
1434 memset(start, POISON_INUSE, PAGE_SIZE << order);
1437 for_each_object(p, s, start, page->objects) {
1438 setup_object(s, page, last);
1439 set_freepointer(s, last, p);
1442 setup_object(s, page, last);
1443 set_freepointer(s, last, NULL);
1445 page->freelist = start;
1446 page->inuse = page->objects;
1452 static void __free_slab(struct kmem_cache *s, struct page *page)
1454 int order = compound_order(page);
1455 int pages = 1 << order;
1457 if (kmem_cache_debug(s)) {
1460 slab_pad_check(s, page);
1461 for_each_object(p, s, page_address(page),
1463 check_object(s, page, p, SLUB_RED_INACTIVE);
1466 kmemcheck_free_shadow(page, compound_order(page));
1468 mod_zone_page_state(page_zone(page),
1469 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1470 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1473 __ClearPageSlabPfmemalloc(page);
1474 __ClearPageSlab(page);
1476 memcg_release_pages(s, order);
1477 page_mapcount_reset(page);
1478 if (current->reclaim_state)
1479 current->reclaim_state->reclaimed_slab += pages;
1480 __free_pages(page, order);
1481 memcg_uncharge_slab(s, order);
1484 #define need_reserve_slab_rcu \
1485 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1487 static void rcu_free_slab(struct rcu_head *h)
1491 if (need_reserve_slab_rcu)
1492 page = virt_to_head_page(h);
1494 page = container_of((struct list_head *)h, struct page, lru);
1496 __free_slab(page->slab_cache, page);
1499 static void free_slab(struct kmem_cache *s, struct page *page)
1501 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1502 struct rcu_head *head;
1504 if (need_reserve_slab_rcu) {
1505 int order = compound_order(page);
1506 int offset = (PAGE_SIZE << order) - s->reserved;
1508 VM_BUG_ON(s->reserved != sizeof(*head));
1509 head = page_address(page) + offset;
1512 * RCU free overloads the RCU head over the LRU
1514 head = (void *)&page->lru;
1517 call_rcu(head, rcu_free_slab);
1519 __free_slab(s, page);
1522 static void discard_slab(struct kmem_cache *s, struct page *page)
1524 dec_slabs_node(s, page_to_nid(page), page->objects);
1529 * Management of partially allocated slabs.
1532 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1535 if (tail == DEACTIVATE_TO_TAIL)
1536 list_add_tail(&page->lru, &n->partial);
1538 list_add(&page->lru, &n->partial);
1541 static inline void add_partial(struct kmem_cache_node *n,
1542 struct page *page, int tail)
1544 lockdep_assert_held(&n->list_lock);
1545 __add_partial(n, page, tail);
1549 __remove_partial(struct kmem_cache_node *n, struct page *page)
1551 list_del(&page->lru);
1555 static inline void remove_partial(struct kmem_cache_node *n,
1558 lockdep_assert_held(&n->list_lock);
1559 __remove_partial(n, page);
1563 * Remove slab from the partial list, freeze it and
1564 * return the pointer to the freelist.
1566 * Returns a list of objects or NULL if it fails.
1568 static inline void *acquire_slab(struct kmem_cache *s,
1569 struct kmem_cache_node *n, struct page *page,
1570 int mode, int *objects)
1573 unsigned long counters;
1576 lockdep_assert_held(&n->list_lock);
1579 * Zap the freelist and set the frozen bit.
1580 * The old freelist is the list of objects for the
1581 * per cpu allocation list.
1583 freelist = page->freelist;
1584 counters = page->counters;
1585 new.counters = counters;
1586 *objects = new.objects - new.inuse;
1588 new.inuse = page->objects;
1589 new.freelist = NULL;
1591 new.freelist = freelist;
1594 VM_BUG_ON(new.frozen);
1597 if (!__cmpxchg_double_slab(s, page,
1599 new.freelist, new.counters,
1603 remove_partial(n, page);
1608 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1609 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1612 * Try to allocate a partial slab from a specific node.
1614 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1615 struct kmem_cache_cpu *c, gfp_t flags)
1617 struct page *page, *page2;
1618 void *object = NULL;
1623 * Racy check. If we mistakenly see no partial slabs then we
1624 * just allocate an empty slab. If we mistakenly try to get a
1625 * partial slab and there is none available then get_partials()
1628 if (!n || !n->nr_partial)
1631 spin_lock(&n->list_lock);
1632 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1635 if (!pfmemalloc_match(page, flags))
1638 t = acquire_slab(s, n, page, object == NULL, &objects);
1642 available += objects;
1645 stat(s, ALLOC_FROM_PARTIAL);
1648 put_cpu_partial(s, page, 0);
1649 stat(s, CPU_PARTIAL_NODE);
1651 if (!kmem_cache_has_cpu_partial(s)
1652 || available > s->cpu_partial / 2)
1656 spin_unlock(&n->list_lock);
1661 * Get a page from somewhere. Search in increasing NUMA distances.
1663 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1664 struct kmem_cache_cpu *c)
1667 struct zonelist *zonelist;
1670 enum zone_type high_zoneidx = gfp_zone(flags);
1672 unsigned int cpuset_mems_cookie;
1675 * The defrag ratio allows a configuration of the tradeoffs between
1676 * inter node defragmentation and node local allocations. A lower
1677 * defrag_ratio increases the tendency to do local allocations
1678 * instead of attempting to obtain partial slabs from other nodes.
1680 * If the defrag_ratio is set to 0 then kmalloc() always
1681 * returns node local objects. If the ratio is higher then kmalloc()
1682 * may return off node objects because partial slabs are obtained
1683 * from other nodes and filled up.
1685 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1686 * defrag_ratio = 1000) then every (well almost) allocation will
1687 * first attempt to defrag slab caches on other nodes. This means
1688 * scanning over all nodes to look for partial slabs which may be
1689 * expensive if we do it every time we are trying to find a slab
1690 * with available objects.
1692 if (!s->remote_node_defrag_ratio ||
1693 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1697 cpuset_mems_cookie = read_mems_allowed_begin();
1698 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1699 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1700 struct kmem_cache_node *n;
1702 n = get_node(s, zone_to_nid(zone));
1704 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1705 n->nr_partial > s->min_partial) {
1706 object = get_partial_node(s, n, c, flags);
1709 * Don't check read_mems_allowed_retry()
1710 * here - if mems_allowed was updated in
1711 * parallel, that was a harmless race
1712 * between allocation and the cpuset
1719 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1725 * Get a partial page, lock it and return it.
1727 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1728 struct kmem_cache_cpu *c)
1731 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1733 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1734 if (object || node != NUMA_NO_NODE)
1737 return get_any_partial(s, flags, c);
1740 #ifdef CONFIG_PREEMPT
1742 * Calculate the next globally unique transaction for disambiguiation
1743 * during cmpxchg. The transactions start with the cpu number and are then
1744 * incremented by CONFIG_NR_CPUS.
1746 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1749 * No preemption supported therefore also no need to check for
1755 static inline unsigned long next_tid(unsigned long tid)
1757 return tid + TID_STEP;
1760 static inline unsigned int tid_to_cpu(unsigned long tid)
1762 return tid % TID_STEP;
1765 static inline unsigned long tid_to_event(unsigned long tid)
1767 return tid / TID_STEP;
1770 static inline unsigned int init_tid(int cpu)
1775 static inline void note_cmpxchg_failure(const char *n,
1776 const struct kmem_cache *s, unsigned long tid)
1778 #ifdef SLUB_DEBUG_CMPXCHG
1779 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1781 pr_info("%s %s: cmpxchg redo ", n, s->name);
1783 #ifdef CONFIG_PREEMPT
1784 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1785 pr_warn("due to cpu change %d -> %d\n",
1786 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1789 if (tid_to_event(tid) != tid_to_event(actual_tid))
1790 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1791 tid_to_event(tid), tid_to_event(actual_tid));
1793 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1794 actual_tid, tid, next_tid(tid));
1796 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1799 static void init_kmem_cache_cpus(struct kmem_cache *s)
1803 for_each_possible_cpu(cpu)
1804 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1808 * Remove the cpu slab
1810 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1813 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1814 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1816 enum slab_modes l = M_NONE, m = M_NONE;
1818 int tail = DEACTIVATE_TO_HEAD;
1822 if (page->freelist) {
1823 stat(s, DEACTIVATE_REMOTE_FREES);
1824 tail = DEACTIVATE_TO_TAIL;
1828 * Stage one: Free all available per cpu objects back
1829 * to the page freelist while it is still frozen. Leave the
1832 * There is no need to take the list->lock because the page
1835 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1837 unsigned long counters;
1840 prior = page->freelist;
1841 counters = page->counters;
1842 set_freepointer(s, freelist, prior);
1843 new.counters = counters;
1845 VM_BUG_ON(!new.frozen);
1847 } while (!__cmpxchg_double_slab(s, page,
1849 freelist, new.counters,
1850 "drain percpu freelist"));
1852 freelist = nextfree;
1856 * Stage two: Ensure that the page is unfrozen while the
1857 * list presence reflects the actual number of objects
1860 * We setup the list membership and then perform a cmpxchg
1861 * with the count. If there is a mismatch then the page
1862 * is not unfrozen but the page is on the wrong list.
1864 * Then we restart the process which may have to remove
1865 * the page from the list that we just put it on again
1866 * because the number of objects in the slab may have
1871 old.freelist = page->freelist;
1872 old.counters = page->counters;
1873 VM_BUG_ON(!old.frozen);
1875 /* Determine target state of the slab */
1876 new.counters = old.counters;
1879 set_freepointer(s, freelist, old.freelist);
1880 new.freelist = freelist;
1882 new.freelist = old.freelist;
1886 if (!new.inuse && n->nr_partial > s->min_partial)
1888 else if (new.freelist) {
1893 * Taking the spinlock removes the possiblity
1894 * that acquire_slab() will see a slab page that
1897 spin_lock(&n->list_lock);
1901 if (kmem_cache_debug(s) && !lock) {
1904 * This also ensures that the scanning of full
1905 * slabs from diagnostic functions will not see
1908 spin_lock(&n->list_lock);
1916 remove_partial(n, page);
1918 else if (l == M_FULL)
1920 remove_full(s, n, page);
1922 if (m == M_PARTIAL) {
1924 add_partial(n, page, tail);
1927 } else if (m == M_FULL) {
1929 stat(s, DEACTIVATE_FULL);
1930 add_full(s, n, page);
1936 if (!__cmpxchg_double_slab(s, page,
1937 old.freelist, old.counters,
1938 new.freelist, new.counters,
1943 spin_unlock(&n->list_lock);
1946 stat(s, DEACTIVATE_EMPTY);
1947 discard_slab(s, page);
1953 * Unfreeze all the cpu partial slabs.
1955 * This function must be called with interrupts disabled
1956 * for the cpu using c (or some other guarantee must be there
1957 * to guarantee no concurrent accesses).
1959 static void unfreeze_partials(struct kmem_cache *s,
1960 struct kmem_cache_cpu *c)
1962 #ifdef CONFIG_SLUB_CPU_PARTIAL
1963 struct kmem_cache_node *n = NULL, *n2 = NULL;
1964 struct page *page, *discard_page = NULL;
1966 while ((page = c->partial)) {
1970 c->partial = page->next;
1972 n2 = get_node(s, page_to_nid(page));
1975 spin_unlock(&n->list_lock);
1978 spin_lock(&n->list_lock);
1983 old.freelist = page->freelist;
1984 old.counters = page->counters;
1985 VM_BUG_ON(!old.frozen);
1987 new.counters = old.counters;
1988 new.freelist = old.freelist;
1992 } while (!__cmpxchg_double_slab(s, page,
1993 old.freelist, old.counters,
1994 new.freelist, new.counters,
1995 "unfreezing slab"));
1997 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1998 page->next = discard_page;
1999 discard_page = page;
2001 add_partial(n, page, DEACTIVATE_TO_TAIL);
2002 stat(s, FREE_ADD_PARTIAL);
2007 spin_unlock(&n->list_lock);
2009 while (discard_page) {
2010 page = discard_page;
2011 discard_page = discard_page->next;
2013 stat(s, DEACTIVATE_EMPTY);
2014 discard_slab(s, page);
2021 * Put a page that was just frozen (in __slab_free) into a partial page
2022 * slot if available. This is done without interrupts disabled and without
2023 * preemption disabled. The cmpxchg is racy and may put the partial page
2024 * onto a random cpus partial slot.
2026 * If we did not find a slot then simply move all the partials to the
2027 * per node partial list.
2029 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2031 #ifdef CONFIG_SLUB_CPU_PARTIAL
2032 struct page *oldpage;
2039 oldpage = this_cpu_read(s->cpu_slab->partial);
2042 pobjects = oldpage->pobjects;
2043 pages = oldpage->pages;
2044 if (drain && pobjects > s->cpu_partial) {
2045 unsigned long flags;
2047 * partial array is full. Move the existing
2048 * set to the per node partial list.
2050 local_irq_save(flags);
2051 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2052 local_irq_restore(flags);
2056 stat(s, CPU_PARTIAL_DRAIN);
2061 pobjects += page->objects - page->inuse;
2063 page->pages = pages;
2064 page->pobjects = pobjects;
2065 page->next = oldpage;
2067 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2072 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2074 stat(s, CPUSLAB_FLUSH);
2075 deactivate_slab(s, c->page, c->freelist);
2077 c->tid = next_tid(c->tid);
2085 * Called from IPI handler with interrupts disabled.
2087 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2089 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2095 unfreeze_partials(s, c);
2099 static void flush_cpu_slab(void *d)
2101 struct kmem_cache *s = d;
2103 __flush_cpu_slab(s, smp_processor_id());
2106 static bool has_cpu_slab(int cpu, void *info)
2108 struct kmem_cache *s = info;
2109 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2111 return c->page || c->partial;
2114 static void flush_all(struct kmem_cache *s)
2116 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2120 * Check if the objects in a per cpu structure fit numa
2121 * locality expectations.
2123 static inline int node_match(struct page *page, int node)
2126 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2132 #ifdef CONFIG_SLUB_DEBUG
2133 static int count_free(struct page *page)
2135 return page->objects - page->inuse;
2138 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2140 return atomic_long_read(&n->total_objects);
2142 #endif /* CONFIG_SLUB_DEBUG */
2144 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2145 static unsigned long count_partial(struct kmem_cache_node *n,
2146 int (*get_count)(struct page *))
2148 unsigned long flags;
2149 unsigned long x = 0;
2152 spin_lock_irqsave(&n->list_lock, flags);
2153 list_for_each_entry(page, &n->partial, lru)
2154 x += get_count(page);
2155 spin_unlock_irqrestore(&n->list_lock, flags);
2158 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2160 static noinline void
2161 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2163 #ifdef CONFIG_SLUB_DEBUG
2164 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2165 DEFAULT_RATELIMIT_BURST);
2168 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2171 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2173 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2174 s->name, s->object_size, s->size, oo_order(s->oo),
2177 if (oo_order(s->min) > get_order(s->object_size))
2178 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2181 for_each_online_node(node) {
2182 struct kmem_cache_node *n = get_node(s, node);
2183 unsigned long nr_slabs;
2184 unsigned long nr_objs;
2185 unsigned long nr_free;
2190 nr_free = count_partial(n, count_free);
2191 nr_slabs = node_nr_slabs(n);
2192 nr_objs = node_nr_objs(n);
2194 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2195 node, nr_slabs, nr_objs, nr_free);
2200 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2201 int node, struct kmem_cache_cpu **pc)
2204 struct kmem_cache_cpu *c = *pc;
2207 freelist = get_partial(s, flags, node, c);
2212 page = new_slab(s, flags, node);
2214 c = __this_cpu_ptr(s->cpu_slab);
2219 * No other reference to the page yet so we can
2220 * muck around with it freely without cmpxchg
2222 freelist = page->freelist;
2223 page->freelist = NULL;
2225 stat(s, ALLOC_SLAB);
2234 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2236 if (unlikely(PageSlabPfmemalloc(page)))
2237 return gfp_pfmemalloc_allowed(gfpflags);
2243 * Check the page->freelist of a page and either transfer the freelist to the
2244 * per cpu freelist or deactivate the page.
2246 * The page is still frozen if the return value is not NULL.
2248 * If this function returns NULL then the page has been unfrozen.
2250 * This function must be called with interrupt disabled.
2252 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2255 unsigned long counters;
2259 freelist = page->freelist;
2260 counters = page->counters;
2262 new.counters = counters;
2263 VM_BUG_ON(!new.frozen);
2265 new.inuse = page->objects;
2266 new.frozen = freelist != NULL;
2268 } while (!__cmpxchg_double_slab(s, page,
2277 * Slow path. The lockless freelist is empty or we need to perform
2280 * Processing is still very fast if new objects have been freed to the
2281 * regular freelist. In that case we simply take over the regular freelist
2282 * as the lockless freelist and zap the regular freelist.
2284 * If that is not working then we fall back to the partial lists. We take the
2285 * first element of the freelist as the object to allocate now and move the
2286 * rest of the freelist to the lockless freelist.
2288 * And if we were unable to get a new slab from the partial slab lists then
2289 * we need to allocate a new slab. This is the slowest path since it involves
2290 * a call to the page allocator and the setup of a new slab.
2292 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2293 unsigned long addr, struct kmem_cache_cpu *c)
2297 unsigned long flags;
2299 local_irq_save(flags);
2300 #ifdef CONFIG_PREEMPT
2302 * We may have been preempted and rescheduled on a different
2303 * cpu before disabling interrupts. Need to reload cpu area
2306 c = this_cpu_ptr(s->cpu_slab);
2314 if (unlikely(!node_match(page, node))) {
2315 stat(s, ALLOC_NODE_MISMATCH);
2316 deactivate_slab(s, page, c->freelist);
2323 * By rights, we should be searching for a slab page that was
2324 * PFMEMALLOC but right now, we are losing the pfmemalloc
2325 * information when the page leaves the per-cpu allocator
2327 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2328 deactivate_slab(s, page, c->freelist);
2334 /* must check again c->freelist in case of cpu migration or IRQ */
2335 freelist = c->freelist;
2339 freelist = get_freelist(s, page);
2343 stat(s, DEACTIVATE_BYPASS);
2347 stat(s, ALLOC_REFILL);
2351 * freelist is pointing to the list of objects to be used.
2352 * page is pointing to the page from which the objects are obtained.
2353 * That page must be frozen for per cpu allocations to work.
2355 VM_BUG_ON(!c->page->frozen);
2356 c->freelist = get_freepointer(s, freelist);
2357 c->tid = next_tid(c->tid);
2358 local_irq_restore(flags);
2364 page = c->page = c->partial;
2365 c->partial = page->next;
2366 stat(s, CPU_PARTIAL_ALLOC);
2371 freelist = new_slab_objects(s, gfpflags, node, &c);
2373 if (unlikely(!freelist)) {
2374 slab_out_of_memory(s, gfpflags, node);
2375 local_irq_restore(flags);
2380 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2383 /* Only entered in the debug case */
2384 if (kmem_cache_debug(s) &&
2385 !alloc_debug_processing(s, page, freelist, addr))
2386 goto new_slab; /* Slab failed checks. Next slab needed */
2388 deactivate_slab(s, page, get_freepointer(s, freelist));
2391 local_irq_restore(flags);
2396 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2397 * have the fastpath folded into their functions. So no function call
2398 * overhead for requests that can be satisfied on the fastpath.
2400 * The fastpath works by first checking if the lockless freelist can be used.
2401 * If not then __slab_alloc is called for slow processing.
2403 * Otherwise we can simply pick the next object from the lockless free list.
2405 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2406 gfp_t gfpflags, int node, unsigned long addr)
2409 struct kmem_cache_cpu *c;
2413 if (slab_pre_alloc_hook(s, gfpflags))
2416 s = memcg_kmem_get_cache(s, gfpflags);
2419 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2420 * enabled. We may switch back and forth between cpus while
2421 * reading from one cpu area. That does not matter as long
2422 * as we end up on the original cpu again when doing the cmpxchg.
2424 * Preemption is disabled for the retrieval of the tid because that
2425 * must occur from the current processor. We cannot allow rescheduling
2426 * on a different processor between the determination of the pointer
2427 * and the retrieval of the tid.
2430 c = __this_cpu_ptr(s->cpu_slab);
2433 * The transaction ids are globally unique per cpu and per operation on
2434 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2435 * occurs on the right processor and that there was no operation on the
2436 * linked list in between.
2441 object = c->freelist;
2443 if (unlikely(!object || !node_match(page, node))) {
2444 object = __slab_alloc(s, gfpflags, node, addr, c);
2445 stat(s, ALLOC_SLOWPATH);
2447 void *next_object = get_freepointer_safe(s, object);
2450 * The cmpxchg will only match if there was no additional
2451 * operation and if we are on the right processor.
2453 * The cmpxchg does the following atomically (without lock
2455 * 1. Relocate first pointer to the current per cpu area.
2456 * 2. Verify that tid and freelist have not been changed
2457 * 3. If they were not changed replace tid and freelist
2459 * Since this is without lock semantics the protection is only
2460 * against code executing on this cpu *not* from access by
2463 if (unlikely(!this_cpu_cmpxchg_double(
2464 s->cpu_slab->freelist, s->cpu_slab->tid,
2466 next_object, next_tid(tid)))) {
2468 note_cmpxchg_failure("slab_alloc", s, tid);
2471 prefetch_freepointer(s, next_object);
2472 stat(s, ALLOC_FASTPATH);
2475 if (unlikely(gfpflags & __GFP_ZERO) && object)
2476 memset(object, 0, s->object_size);
2478 slab_post_alloc_hook(s, gfpflags, object);
2483 static __always_inline void *slab_alloc(struct kmem_cache *s,
2484 gfp_t gfpflags, unsigned long addr)
2486 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2489 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2491 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2493 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2498 EXPORT_SYMBOL(kmem_cache_alloc);
2500 #ifdef CONFIG_TRACING
2501 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2503 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2504 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2507 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2511 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2513 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2515 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2516 s->object_size, s->size, gfpflags, node);
2520 EXPORT_SYMBOL(kmem_cache_alloc_node);
2522 #ifdef CONFIG_TRACING
2523 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2525 int node, size_t size)
2527 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2529 trace_kmalloc_node(_RET_IP_, ret,
2530 size, s->size, gfpflags, node);
2533 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2538 * Slow patch handling. This may still be called frequently since objects
2539 * have a longer lifetime than the cpu slabs in most processing loads.
2541 * So we still attempt to reduce cache line usage. Just take the slab
2542 * lock and free the item. If there is no additional partial page
2543 * handling required then we can return immediately.
2545 static void __slab_free(struct kmem_cache *s, struct page *page,
2546 void *x, unsigned long addr)
2549 void **object = (void *)x;
2552 unsigned long counters;
2553 struct kmem_cache_node *n = NULL;
2554 unsigned long uninitialized_var(flags);
2556 stat(s, FREE_SLOWPATH);
2558 if (kmem_cache_debug(s) &&
2559 !(n = free_debug_processing(s, page, x, addr, &flags)))
2564 spin_unlock_irqrestore(&n->list_lock, flags);
2567 prior = page->freelist;
2568 counters = page->counters;
2569 set_freepointer(s, object, prior);
2570 new.counters = counters;
2571 was_frozen = new.frozen;
2573 if ((!new.inuse || !prior) && !was_frozen) {
2575 if (kmem_cache_has_cpu_partial(s) && !prior) {
2578 * Slab was on no list before and will be
2580 * We can defer the list move and instead
2585 } else { /* Needs to be taken off a list */
2587 n = get_node(s, page_to_nid(page));
2589 * Speculatively acquire the list_lock.
2590 * If the cmpxchg does not succeed then we may
2591 * drop the list_lock without any processing.
2593 * Otherwise the list_lock will synchronize with
2594 * other processors updating the list of slabs.
2596 spin_lock_irqsave(&n->list_lock, flags);
2601 } while (!cmpxchg_double_slab(s, page,
2603 object, new.counters,
2609 * If we just froze the page then put it onto the
2610 * per cpu partial list.
2612 if (new.frozen && !was_frozen) {
2613 put_cpu_partial(s, page, 1);
2614 stat(s, CPU_PARTIAL_FREE);
2617 * The list lock was not taken therefore no list
2618 * activity can be necessary.
2621 stat(s, FREE_FROZEN);
2625 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2629 * Objects left in the slab. If it was not on the partial list before
2632 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2633 if (kmem_cache_debug(s))
2634 remove_full(s, n, page);
2635 add_partial(n, page, DEACTIVATE_TO_TAIL);
2636 stat(s, FREE_ADD_PARTIAL);
2638 spin_unlock_irqrestore(&n->list_lock, flags);
2644 * Slab on the partial list.
2646 remove_partial(n, page);
2647 stat(s, FREE_REMOVE_PARTIAL);
2649 /* Slab must be on the full list */
2650 remove_full(s, n, page);
2653 spin_unlock_irqrestore(&n->list_lock, flags);
2655 discard_slab(s, page);
2659 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2660 * can perform fastpath freeing without additional function calls.
2662 * The fastpath is only possible if we are freeing to the current cpu slab
2663 * of this processor. This typically the case if we have just allocated
2666 * If fastpath is not possible then fall back to __slab_free where we deal
2667 * with all sorts of special processing.
2669 static __always_inline void slab_free(struct kmem_cache *s,
2670 struct page *page, void *x, unsigned long addr)
2672 void **object = (void *)x;
2673 struct kmem_cache_cpu *c;
2676 slab_free_hook(s, x);
2680 * Determine the currently cpus per cpu slab.
2681 * The cpu may change afterward. However that does not matter since
2682 * data is retrieved via this pointer. If we are on the same cpu
2683 * during the cmpxchg then the free will succedd.
2686 c = __this_cpu_ptr(s->cpu_slab);
2691 if (likely(page == c->page)) {
2692 set_freepointer(s, object, c->freelist);
2694 if (unlikely(!this_cpu_cmpxchg_double(
2695 s->cpu_slab->freelist, s->cpu_slab->tid,
2697 object, next_tid(tid)))) {
2699 note_cmpxchg_failure("slab_free", s, tid);
2702 stat(s, FREE_FASTPATH);
2704 __slab_free(s, page, x, addr);
2708 void kmem_cache_free(struct kmem_cache *s, void *x)
2710 s = cache_from_obj(s, x);
2713 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2714 trace_kmem_cache_free(_RET_IP_, x);
2716 EXPORT_SYMBOL(kmem_cache_free);
2719 * Object placement in a slab is made very easy because we always start at
2720 * offset 0. If we tune the size of the object to the alignment then we can
2721 * get the required alignment by putting one properly sized object after
2724 * Notice that the allocation order determines the sizes of the per cpu
2725 * caches. Each processor has always one slab available for allocations.
2726 * Increasing the allocation order reduces the number of times that slabs
2727 * must be moved on and off the partial lists and is therefore a factor in
2732 * Mininum / Maximum order of slab pages. This influences locking overhead
2733 * and slab fragmentation. A higher order reduces the number of partial slabs
2734 * and increases the number of allocations possible without having to
2735 * take the list_lock.
2737 static int slub_min_order;
2738 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2739 static int slub_min_objects;
2742 * Merge control. If this is set then no merging of slab caches will occur.
2743 * (Could be removed. This was introduced to pacify the merge skeptics.)
2745 static int slub_nomerge;
2748 * Calculate the order of allocation given an slab object size.
2750 * The order of allocation has significant impact on performance and other
2751 * system components. Generally order 0 allocations should be preferred since
2752 * order 0 does not cause fragmentation in the page allocator. Larger objects
2753 * be problematic to put into order 0 slabs because there may be too much
2754 * unused space left. We go to a higher order if more than 1/16th of the slab
2757 * In order to reach satisfactory performance we must ensure that a minimum
2758 * number of objects is in one slab. Otherwise we may generate too much
2759 * activity on the partial lists which requires taking the list_lock. This is
2760 * less a concern for large slabs though which are rarely used.
2762 * slub_max_order specifies the order where we begin to stop considering the
2763 * number of objects in a slab as critical. If we reach slub_max_order then
2764 * we try to keep the page order as low as possible. So we accept more waste
2765 * of space in favor of a small page order.
2767 * Higher order allocations also allow the placement of more objects in a
2768 * slab and thereby reduce object handling overhead. If the user has
2769 * requested a higher mininum order then we start with that one instead of
2770 * the smallest order which will fit the object.
2772 static inline int slab_order(int size, int min_objects,
2773 int max_order, int fract_leftover, int reserved)
2777 int min_order = slub_min_order;
2779 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2780 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2782 for (order = max(min_order,
2783 fls(min_objects * size - 1) - PAGE_SHIFT);
2784 order <= max_order; order++) {
2786 unsigned long slab_size = PAGE_SIZE << order;
2788 if (slab_size < min_objects * size + reserved)
2791 rem = (slab_size - reserved) % size;
2793 if (rem <= slab_size / fract_leftover)
2801 static inline int calculate_order(int size, int reserved)
2809 * Attempt to find best configuration for a slab. This
2810 * works by first attempting to generate a layout with
2811 * the best configuration and backing off gradually.
2813 * First we reduce the acceptable waste in a slab. Then
2814 * we reduce the minimum objects required in a slab.
2816 min_objects = slub_min_objects;
2818 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2819 max_objects = order_objects(slub_max_order, size, reserved);
2820 min_objects = min(min_objects, max_objects);
2822 while (min_objects > 1) {
2824 while (fraction >= 4) {
2825 order = slab_order(size, min_objects,
2826 slub_max_order, fraction, reserved);
2827 if (order <= slub_max_order)
2835 * We were unable to place multiple objects in a slab. Now
2836 * lets see if we can place a single object there.
2838 order = slab_order(size, 1, slub_max_order, 1, reserved);
2839 if (order <= slub_max_order)
2843 * Doh this slab cannot be placed using slub_max_order.
2845 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2846 if (order < MAX_ORDER)
2852 init_kmem_cache_node(struct kmem_cache_node *n)
2855 spin_lock_init(&n->list_lock);
2856 INIT_LIST_HEAD(&n->partial);
2857 #ifdef CONFIG_SLUB_DEBUG
2858 atomic_long_set(&n->nr_slabs, 0);
2859 atomic_long_set(&n->total_objects, 0);
2860 INIT_LIST_HEAD(&n->full);
2864 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2866 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2867 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2870 * Must align to double word boundary for the double cmpxchg
2871 * instructions to work; see __pcpu_double_call_return_bool().
2873 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2874 2 * sizeof(void *));
2879 init_kmem_cache_cpus(s);
2884 static struct kmem_cache *kmem_cache_node;
2887 * No kmalloc_node yet so do it by hand. We know that this is the first
2888 * slab on the node for this slabcache. There are no concurrent accesses
2891 * Note that this function only works on the kmem_cache_node
2892 * when allocating for the kmem_cache_node. This is used for bootstrapping
2893 * memory on a fresh node that has no slab structures yet.
2895 static void early_kmem_cache_node_alloc(int node)
2898 struct kmem_cache_node *n;
2900 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2902 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2905 if (page_to_nid(page) != node) {
2906 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2907 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2912 page->freelist = get_freepointer(kmem_cache_node, n);
2915 kmem_cache_node->node[node] = n;
2916 #ifdef CONFIG_SLUB_DEBUG
2917 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2918 init_tracking(kmem_cache_node, n);
2920 init_kmem_cache_node(n);
2921 inc_slabs_node(kmem_cache_node, node, page->objects);
2924 * No locks need to be taken here as it has just been
2925 * initialized and there is no concurrent access.
2927 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2930 static void free_kmem_cache_nodes(struct kmem_cache *s)
2934 for_each_node_state(node, N_NORMAL_MEMORY) {
2935 struct kmem_cache_node *n = s->node[node];
2938 kmem_cache_free(kmem_cache_node, n);
2940 s->node[node] = NULL;
2944 static int init_kmem_cache_nodes(struct kmem_cache *s)
2948 for_each_node_state(node, N_NORMAL_MEMORY) {
2949 struct kmem_cache_node *n;
2951 if (slab_state == DOWN) {
2952 early_kmem_cache_node_alloc(node);
2955 n = kmem_cache_alloc_node(kmem_cache_node,
2959 free_kmem_cache_nodes(s);
2964 init_kmem_cache_node(n);
2969 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2971 if (min < MIN_PARTIAL)
2973 else if (min > MAX_PARTIAL)
2975 s->min_partial = min;
2979 * calculate_sizes() determines the order and the distribution of data within
2982 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2984 unsigned long flags = s->flags;
2985 unsigned long size = s->object_size;
2989 * Round up object size to the next word boundary. We can only
2990 * place the free pointer at word boundaries and this determines
2991 * the possible location of the free pointer.
2993 size = ALIGN(size, sizeof(void *));
2995 #ifdef CONFIG_SLUB_DEBUG
2997 * Determine if we can poison the object itself. If the user of
2998 * the slab may touch the object after free or before allocation
2999 * then we should never poison the object itself.
3001 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3003 s->flags |= __OBJECT_POISON;
3005 s->flags &= ~__OBJECT_POISON;
3009 * If we are Redzoning then check if there is some space between the
3010 * end of the object and the free pointer. If not then add an
3011 * additional word to have some bytes to store Redzone information.
3013 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3014 size += sizeof(void *);
3018 * With that we have determined the number of bytes in actual use
3019 * by the object. This is the potential offset to the free pointer.
3023 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3026 * Relocate free pointer after the object if it is not
3027 * permitted to overwrite the first word of the object on
3030 * This is the case if we do RCU, have a constructor or
3031 * destructor or are poisoning the objects.
3034 size += sizeof(void *);
3037 #ifdef CONFIG_SLUB_DEBUG
3038 if (flags & SLAB_STORE_USER)
3040 * Need to store information about allocs and frees after
3043 size += 2 * sizeof(struct track);
3045 if (flags & SLAB_RED_ZONE)
3047 * Add some empty padding so that we can catch
3048 * overwrites from earlier objects rather than let
3049 * tracking information or the free pointer be
3050 * corrupted if a user writes before the start
3053 size += sizeof(void *);
3057 * SLUB stores one object immediately after another beginning from
3058 * offset 0. In order to align the objects we have to simply size
3059 * each object to conform to the alignment.
3061 size = ALIGN(size, s->align);
3063 if (forced_order >= 0)
3064 order = forced_order;
3066 order = calculate_order(size, s->reserved);
3073 s->allocflags |= __GFP_COMP;
3075 if (s->flags & SLAB_CACHE_DMA)
3076 s->allocflags |= GFP_DMA;
3078 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3079 s->allocflags |= __GFP_RECLAIMABLE;
3082 * Determine the number of objects per slab
3084 s->oo = oo_make(order, size, s->reserved);
3085 s->min = oo_make(get_order(size), size, s->reserved);
3086 if (oo_objects(s->oo) > oo_objects(s->max))
3089 return !!oo_objects(s->oo);
3092 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3094 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3097 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3098 s->reserved = sizeof(struct rcu_head);
3100 if (!calculate_sizes(s, -1))
3102 if (disable_higher_order_debug) {
3104 * Disable debugging flags that store metadata if the min slab
3107 if (get_order(s->size) > get_order(s->object_size)) {
3108 s->flags &= ~DEBUG_METADATA_FLAGS;
3110 if (!calculate_sizes(s, -1))
3115 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3116 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3117 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3118 /* Enable fast mode */
3119 s->flags |= __CMPXCHG_DOUBLE;
3123 * The larger the object size is, the more pages we want on the partial
3124 * list to avoid pounding the page allocator excessively.
3126 set_min_partial(s, ilog2(s->size) / 2);
3129 * cpu_partial determined the maximum number of objects kept in the
3130 * per cpu partial lists of a processor.
3132 * Per cpu partial lists mainly contain slabs that just have one
3133 * object freed. If they are used for allocation then they can be
3134 * filled up again with minimal effort. The slab will never hit the
3135 * per node partial lists and therefore no locking will be required.
3137 * This setting also determines
3139 * A) The number of objects from per cpu partial slabs dumped to the
3140 * per node list when we reach the limit.
3141 * B) The number of objects in cpu partial slabs to extract from the
3142 * per node list when we run out of per cpu objects. We only fetch
3143 * 50% to keep some capacity around for frees.
3145 if (!kmem_cache_has_cpu_partial(s))
3147 else if (s->size >= PAGE_SIZE)
3149 else if (s->size >= 1024)
3151 else if (s->size >= 256)
3152 s->cpu_partial = 13;
3154 s->cpu_partial = 30;
3157 s->remote_node_defrag_ratio = 1000;
3159 if (!init_kmem_cache_nodes(s))
3162 if (alloc_kmem_cache_cpus(s))
3165 free_kmem_cache_nodes(s);
3167 if (flags & SLAB_PANIC)
3168 panic("Cannot create slab %s size=%lu realsize=%u "
3169 "order=%u offset=%u flags=%lx\n",
3170 s->name, (unsigned long)s->size, s->size,
3171 oo_order(s->oo), s->offset, flags);
3175 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3178 #ifdef CONFIG_SLUB_DEBUG
3179 void *addr = page_address(page);
3181 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3182 sizeof(long), GFP_ATOMIC);
3185 slab_err(s, page, text, s->name);
3188 get_map(s, page, map);
3189 for_each_object(p, s, addr, page->objects) {
3191 if (!test_bit(slab_index(p, s, addr), map)) {
3192 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3193 print_tracking(s, p);
3202 * Attempt to free all partial slabs on a node.
3203 * This is called from kmem_cache_close(). We must be the last thread
3204 * using the cache and therefore we do not need to lock anymore.
3206 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3208 struct page *page, *h;
3210 list_for_each_entry_safe(page, h, &n->partial, lru) {
3212 __remove_partial(n, page);
3213 discard_slab(s, page);
3215 list_slab_objects(s, page,
3216 "Objects remaining in %s on kmem_cache_close()");
3222 * Release all resources used by a slab cache.
3224 static inline int kmem_cache_close(struct kmem_cache *s)
3229 /* Attempt to free all objects */
3230 for_each_node_state(node, N_NORMAL_MEMORY) {
3231 struct kmem_cache_node *n = get_node(s, node);
3234 if (n->nr_partial || slabs_node(s, node))
3237 free_percpu(s->cpu_slab);
3238 free_kmem_cache_nodes(s);
3242 int __kmem_cache_shutdown(struct kmem_cache *s)
3244 return kmem_cache_close(s);
3247 /********************************************************************
3249 *******************************************************************/
3251 static int __init setup_slub_min_order(char *str)
3253 get_option(&str, &slub_min_order);
3258 __setup("slub_min_order=", setup_slub_min_order);
3260 static int __init setup_slub_max_order(char *str)
3262 get_option(&str, &slub_max_order);
3263 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3268 __setup("slub_max_order=", setup_slub_max_order);
3270 static int __init setup_slub_min_objects(char *str)
3272 get_option(&str, &slub_min_objects);
3277 __setup("slub_min_objects=", setup_slub_min_objects);
3279 static int __init setup_slub_nomerge(char *str)
3285 __setup("slub_nomerge", setup_slub_nomerge);
3287 void *__kmalloc(size_t size, gfp_t flags)
3289 struct kmem_cache *s;
3292 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3293 return kmalloc_large(size, flags);
3295 s = kmalloc_slab(size, flags);
3297 if (unlikely(ZERO_OR_NULL_PTR(s)))
3300 ret = slab_alloc(s, flags, _RET_IP_);
3302 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3306 EXPORT_SYMBOL(__kmalloc);
3309 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3314 flags |= __GFP_COMP | __GFP_NOTRACK;
3315 page = alloc_kmem_pages_node(node, flags, get_order(size));
3317 ptr = page_address(page);
3319 kmalloc_large_node_hook(ptr, size, flags);
3323 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3325 struct kmem_cache *s;
3328 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3329 ret = kmalloc_large_node(size, flags, node);
3331 trace_kmalloc_node(_RET_IP_, ret,
3332 size, PAGE_SIZE << get_order(size),
3338 s = kmalloc_slab(size, flags);
3340 if (unlikely(ZERO_OR_NULL_PTR(s)))
3343 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3345 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3349 EXPORT_SYMBOL(__kmalloc_node);
3352 size_t ksize(const void *object)
3356 if (unlikely(object == ZERO_SIZE_PTR))
3359 page = virt_to_head_page(object);
3361 if (unlikely(!PageSlab(page))) {
3362 WARN_ON(!PageCompound(page));
3363 return PAGE_SIZE << compound_order(page);
3366 return slab_ksize(page->slab_cache);
3368 EXPORT_SYMBOL(ksize);
3370 void kfree(const void *x)
3373 void *object = (void *)x;
3375 trace_kfree(_RET_IP_, x);
3377 if (unlikely(ZERO_OR_NULL_PTR(x)))
3380 page = virt_to_head_page(x);
3381 if (unlikely(!PageSlab(page))) {
3382 BUG_ON(!PageCompound(page));
3384 __free_kmem_pages(page, compound_order(page));
3387 slab_free(page->slab_cache, page, object, _RET_IP_);
3389 EXPORT_SYMBOL(kfree);
3392 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3393 * the remaining slabs by the number of items in use. The slabs with the
3394 * most items in use come first. New allocations will then fill those up
3395 * and thus they can be removed from the partial lists.
3397 * The slabs with the least items are placed last. This results in them
3398 * being allocated from last increasing the chance that the last objects
3399 * are freed in them.
3401 int kmem_cache_shrink(struct kmem_cache *s)
3405 struct kmem_cache_node *n;
3408 int objects = oo_objects(s->max);
3409 struct list_head *slabs_by_inuse =
3410 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3411 unsigned long flags;
3413 if (!slabs_by_inuse)
3417 for_each_node_state(node, N_NORMAL_MEMORY) {
3418 n = get_node(s, node);
3423 for (i = 0; i < objects; i++)
3424 INIT_LIST_HEAD(slabs_by_inuse + i);
3426 spin_lock_irqsave(&n->list_lock, flags);
3429 * Build lists indexed by the items in use in each slab.
3431 * Note that concurrent frees may occur while we hold the
3432 * list_lock. page->inuse here is the upper limit.
3434 list_for_each_entry_safe(page, t, &n->partial, lru) {
3435 list_move(&page->lru, slabs_by_inuse + page->inuse);
3441 * Rebuild the partial list with the slabs filled up most
3442 * first and the least used slabs at the end.
3444 for (i = objects - 1; i > 0; i--)
3445 list_splice(slabs_by_inuse + i, n->partial.prev);
3447 spin_unlock_irqrestore(&n->list_lock, flags);
3449 /* Release empty slabs */
3450 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3451 discard_slab(s, page);
3454 kfree(slabs_by_inuse);
3457 EXPORT_SYMBOL(kmem_cache_shrink);
3459 static int slab_mem_going_offline_callback(void *arg)
3461 struct kmem_cache *s;
3463 mutex_lock(&slab_mutex);
3464 list_for_each_entry(s, &slab_caches, list)
3465 kmem_cache_shrink(s);
3466 mutex_unlock(&slab_mutex);
3471 static void slab_mem_offline_callback(void *arg)
3473 struct kmem_cache_node *n;
3474 struct kmem_cache *s;
3475 struct memory_notify *marg = arg;
3478 offline_node = marg->status_change_nid_normal;
3481 * If the node still has available memory. we need kmem_cache_node
3484 if (offline_node < 0)
3487 mutex_lock(&slab_mutex);
3488 list_for_each_entry(s, &slab_caches, list) {
3489 n = get_node(s, offline_node);
3492 * if n->nr_slabs > 0, slabs still exist on the node
3493 * that is going down. We were unable to free them,
3494 * and offline_pages() function shouldn't call this
3495 * callback. So, we must fail.
3497 BUG_ON(slabs_node(s, offline_node));
3499 s->node[offline_node] = NULL;
3500 kmem_cache_free(kmem_cache_node, n);
3503 mutex_unlock(&slab_mutex);
3506 static int slab_mem_going_online_callback(void *arg)
3508 struct kmem_cache_node *n;
3509 struct kmem_cache *s;
3510 struct memory_notify *marg = arg;
3511 int nid = marg->status_change_nid_normal;
3515 * If the node's memory is already available, then kmem_cache_node is
3516 * already created. Nothing to do.
3522 * We are bringing a node online. No memory is available yet. We must
3523 * allocate a kmem_cache_node structure in order to bring the node
3526 mutex_lock(&slab_mutex);
3527 list_for_each_entry(s, &slab_caches, list) {
3529 * XXX: kmem_cache_alloc_node will fallback to other nodes
3530 * since memory is not yet available from the node that
3533 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3538 init_kmem_cache_node(n);
3542 mutex_unlock(&slab_mutex);
3546 static int slab_memory_callback(struct notifier_block *self,
3547 unsigned long action, void *arg)
3552 case MEM_GOING_ONLINE:
3553 ret = slab_mem_going_online_callback(arg);
3555 case MEM_GOING_OFFLINE:
3556 ret = slab_mem_going_offline_callback(arg);
3559 case MEM_CANCEL_ONLINE:
3560 slab_mem_offline_callback(arg);
3563 case MEM_CANCEL_OFFLINE:
3567 ret = notifier_from_errno(ret);
3573 static struct notifier_block slab_memory_callback_nb = {
3574 .notifier_call = slab_memory_callback,
3575 .priority = SLAB_CALLBACK_PRI,
3578 /********************************************************************
3579 * Basic setup of slabs
3580 *******************************************************************/
3583 * Used for early kmem_cache structures that were allocated using
3584 * the page allocator. Allocate them properly then fix up the pointers
3585 * that may be pointing to the wrong kmem_cache structure.
3588 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3591 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3593 memcpy(s, static_cache, kmem_cache->object_size);
3596 * This runs very early, and only the boot processor is supposed to be
3597 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3600 __flush_cpu_slab(s, smp_processor_id());
3601 for_each_node_state(node, N_NORMAL_MEMORY) {
3602 struct kmem_cache_node *n = get_node(s, node);
3606 list_for_each_entry(p, &n->partial, lru)
3609 #ifdef CONFIG_SLUB_DEBUG
3610 list_for_each_entry(p, &n->full, lru)
3615 list_add(&s->list, &slab_caches);
3619 void __init kmem_cache_init(void)
3621 static __initdata struct kmem_cache boot_kmem_cache,
3622 boot_kmem_cache_node;
3624 if (debug_guardpage_minorder())
3627 kmem_cache_node = &boot_kmem_cache_node;
3628 kmem_cache = &boot_kmem_cache;
3630 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3631 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3633 register_hotmemory_notifier(&slab_memory_callback_nb);
3635 /* Able to allocate the per node structures */
3636 slab_state = PARTIAL;
3638 create_boot_cache(kmem_cache, "kmem_cache",
3639 offsetof(struct kmem_cache, node) +
3640 nr_node_ids * sizeof(struct kmem_cache_node *),
3641 SLAB_HWCACHE_ALIGN);
3643 kmem_cache = bootstrap(&boot_kmem_cache);
3646 * Allocate kmem_cache_node properly from the kmem_cache slab.
3647 * kmem_cache_node is separately allocated so no need to
3648 * update any list pointers.
3650 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3652 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3653 create_kmalloc_caches(0);
3656 register_cpu_notifier(&slab_notifier);
3659 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3661 slub_min_order, slub_max_order, slub_min_objects,
3662 nr_cpu_ids, nr_node_ids);
3665 void __init kmem_cache_init_late(void)
3670 * Find a mergeable slab cache
3672 static int slab_unmergeable(struct kmem_cache *s)
3674 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3677 if (!is_root_cache(s))
3684 * We may have set a slab to be unmergeable during bootstrap.
3686 if (s->refcount < 0)
3692 static struct kmem_cache *find_mergeable(size_t size, size_t align,
3693 unsigned long flags, const char *name, void (*ctor)(void *))
3695 struct kmem_cache *s;
3697 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3703 size = ALIGN(size, sizeof(void *));
3704 align = calculate_alignment(flags, align, size);
3705 size = ALIGN(size, align);
3706 flags = kmem_cache_flags(size, flags, name, NULL);
3708 list_for_each_entry(s, &slab_caches, list) {
3709 if (slab_unmergeable(s))
3715 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3718 * Check if alignment is compatible.
3719 * Courtesy of Adrian Drzewiecki
3721 if ((s->size & ~(align - 1)) != s->size)
3724 if (s->size - size >= sizeof(void *))
3733 __kmem_cache_alias(const char *name, size_t size, size_t align,
3734 unsigned long flags, void (*ctor)(void *))
3736 struct kmem_cache *s;
3738 s = find_mergeable(size, align, flags, name, ctor);
3741 struct kmem_cache *c;
3746 * Adjust the object sizes so that we clear
3747 * the complete object on kzalloc.
3749 s->object_size = max(s->object_size, (int)size);
3750 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3752 for_each_memcg_cache_index(i) {
3753 c = cache_from_memcg_idx(s, i);
3756 c->object_size = s->object_size;
3757 c->inuse = max_t(int, c->inuse,
3758 ALIGN(size, sizeof(void *)));
3761 if (sysfs_slab_alias(s, name)) {
3770 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3774 err = kmem_cache_open(s, flags);
3778 /* Mutex is not taken during early boot */
3779 if (slab_state <= UP)
3782 memcg_propagate_slab_attrs(s);
3783 err = sysfs_slab_add(s);
3785 kmem_cache_close(s);
3792 * Use the cpu notifier to insure that the cpu slabs are flushed when
3795 static int slab_cpuup_callback(struct notifier_block *nfb,
3796 unsigned long action, void *hcpu)
3798 long cpu = (long)hcpu;
3799 struct kmem_cache *s;
3800 unsigned long flags;
3803 case CPU_UP_CANCELED:
3804 case CPU_UP_CANCELED_FROZEN:
3806 case CPU_DEAD_FROZEN:
3807 mutex_lock(&slab_mutex);
3808 list_for_each_entry(s, &slab_caches, list) {
3809 local_irq_save(flags);
3810 __flush_cpu_slab(s, cpu);
3811 local_irq_restore(flags);
3813 mutex_unlock(&slab_mutex);
3821 static struct notifier_block slab_notifier = {
3822 .notifier_call = slab_cpuup_callback
3827 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3829 struct kmem_cache *s;
3832 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3833 return kmalloc_large(size, gfpflags);
3835 s = kmalloc_slab(size, gfpflags);
3837 if (unlikely(ZERO_OR_NULL_PTR(s)))
3840 ret = slab_alloc(s, gfpflags, caller);
3842 /* Honor the call site pointer we received. */
3843 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3849 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3850 int node, unsigned long caller)
3852 struct kmem_cache *s;
3855 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3856 ret = kmalloc_large_node(size, gfpflags, node);
3858 trace_kmalloc_node(caller, ret,
3859 size, PAGE_SIZE << get_order(size),
3865 s = kmalloc_slab(size, gfpflags);
3867 if (unlikely(ZERO_OR_NULL_PTR(s)))
3870 ret = slab_alloc_node(s, gfpflags, node, caller);
3872 /* Honor the call site pointer we received. */
3873 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3880 static int count_inuse(struct page *page)
3885 static int count_total(struct page *page)
3887 return page->objects;
3891 #ifdef CONFIG_SLUB_DEBUG
3892 static int validate_slab(struct kmem_cache *s, struct page *page,
3896 void *addr = page_address(page);
3898 if (!check_slab(s, page) ||
3899 !on_freelist(s, page, NULL))
3902 /* Now we know that a valid freelist exists */
3903 bitmap_zero(map, page->objects);
3905 get_map(s, page, map);
3906 for_each_object(p, s, addr, page->objects) {
3907 if (test_bit(slab_index(p, s, addr), map))
3908 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3912 for_each_object(p, s, addr, page->objects)
3913 if (!test_bit(slab_index(p, s, addr), map))
3914 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3919 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3923 validate_slab(s, page, map);
3927 static int validate_slab_node(struct kmem_cache *s,
3928 struct kmem_cache_node *n, unsigned long *map)
3930 unsigned long count = 0;
3932 unsigned long flags;
3934 spin_lock_irqsave(&n->list_lock, flags);
3936 list_for_each_entry(page, &n->partial, lru) {
3937 validate_slab_slab(s, page, map);
3940 if (count != n->nr_partial)
3941 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3942 s->name, count, n->nr_partial);
3944 if (!(s->flags & SLAB_STORE_USER))
3947 list_for_each_entry(page, &n->full, lru) {
3948 validate_slab_slab(s, page, map);
3951 if (count != atomic_long_read(&n->nr_slabs))
3952 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3953 s->name, count, atomic_long_read(&n->nr_slabs));
3956 spin_unlock_irqrestore(&n->list_lock, flags);
3960 static long validate_slab_cache(struct kmem_cache *s)
3963 unsigned long count = 0;
3964 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3965 sizeof(unsigned long), GFP_KERNEL);
3971 for_each_node_state(node, N_NORMAL_MEMORY) {
3972 struct kmem_cache_node *n = get_node(s, node);
3974 count += validate_slab_node(s, n, map);
3980 * Generate lists of code addresses where slabcache objects are allocated
3985 unsigned long count;
3992 DECLARE_BITMAP(cpus, NR_CPUS);
3998 unsigned long count;
3999 struct location *loc;
4002 static void free_loc_track(struct loc_track *t)
4005 free_pages((unsigned long)t->loc,
4006 get_order(sizeof(struct location) * t->max));
4009 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4014 order = get_order(sizeof(struct location) * max);
4016 l = (void *)__get_free_pages(flags, order);
4021 memcpy(l, t->loc, sizeof(struct location) * t->count);
4029 static int add_location(struct loc_track *t, struct kmem_cache *s,
4030 const struct track *track)
4032 long start, end, pos;
4034 unsigned long caddr;
4035 unsigned long age = jiffies - track->when;
4041 pos = start + (end - start + 1) / 2;
4044 * There is nothing at "end". If we end up there
4045 * we need to add something to before end.
4050 caddr = t->loc[pos].addr;
4051 if (track->addr == caddr) {
4057 if (age < l->min_time)
4059 if (age > l->max_time)
4062 if (track->pid < l->min_pid)
4063 l->min_pid = track->pid;
4064 if (track->pid > l->max_pid)
4065 l->max_pid = track->pid;
4067 cpumask_set_cpu(track->cpu,
4068 to_cpumask(l->cpus));
4070 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4074 if (track->addr < caddr)
4081 * Not found. Insert new tracking element.
4083 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4089 (t->count - pos) * sizeof(struct location));
4092 l->addr = track->addr;
4096 l->min_pid = track->pid;
4097 l->max_pid = track->pid;
4098 cpumask_clear(to_cpumask(l->cpus));
4099 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4100 nodes_clear(l->nodes);
4101 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4105 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4106 struct page *page, enum track_item alloc,
4109 void *addr = page_address(page);
4112 bitmap_zero(map, page->objects);
4113 get_map(s, page, map);
4115 for_each_object(p, s, addr, page->objects)
4116 if (!test_bit(slab_index(p, s, addr), map))
4117 add_location(t, s, get_track(s, p, alloc));
4120 static int list_locations(struct kmem_cache *s, char *buf,
4121 enum track_item alloc)
4125 struct loc_track t = { 0, 0, NULL };
4127 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4128 sizeof(unsigned long), GFP_KERNEL);
4130 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4133 return sprintf(buf, "Out of memory\n");
4135 /* Push back cpu slabs */
4138 for_each_node_state(node, N_NORMAL_MEMORY) {
4139 struct kmem_cache_node *n = get_node(s, node);
4140 unsigned long flags;
4143 if (!atomic_long_read(&n->nr_slabs))
4146 spin_lock_irqsave(&n->list_lock, flags);
4147 list_for_each_entry(page, &n->partial, lru)
4148 process_slab(&t, s, page, alloc, map);
4149 list_for_each_entry(page, &n->full, lru)
4150 process_slab(&t, s, page, alloc, map);
4151 spin_unlock_irqrestore(&n->list_lock, flags);
4154 for (i = 0; i < t.count; i++) {
4155 struct location *l = &t.loc[i];
4157 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4159 len += sprintf(buf + len, "%7ld ", l->count);
4162 len += sprintf(buf + len, "%pS", (void *)l->addr);
4164 len += sprintf(buf + len, "<not-available>");
4166 if (l->sum_time != l->min_time) {
4167 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4169 (long)div_u64(l->sum_time, l->count),
4172 len += sprintf(buf + len, " age=%ld",
4175 if (l->min_pid != l->max_pid)
4176 len += sprintf(buf + len, " pid=%ld-%ld",
4177 l->min_pid, l->max_pid);
4179 len += sprintf(buf + len, " pid=%ld",
4182 if (num_online_cpus() > 1 &&
4183 !cpumask_empty(to_cpumask(l->cpus)) &&
4184 len < PAGE_SIZE - 60) {
4185 len += sprintf(buf + len, " cpus=");
4186 len += cpulist_scnprintf(buf + len,
4187 PAGE_SIZE - len - 50,
4188 to_cpumask(l->cpus));
4191 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4192 len < PAGE_SIZE - 60) {
4193 len += sprintf(buf + len, " nodes=");
4194 len += nodelist_scnprintf(buf + len,
4195 PAGE_SIZE - len - 50,
4199 len += sprintf(buf + len, "\n");
4205 len += sprintf(buf, "No data\n");
4210 #ifdef SLUB_RESILIENCY_TEST
4211 static void resiliency_test(void)
4215 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4217 pr_err("SLUB resiliency testing\n");
4218 pr_err("-----------------------\n");
4219 pr_err("A. Corruption after allocation\n");
4221 p = kzalloc(16, GFP_KERNEL);
4223 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4226 validate_slab_cache(kmalloc_caches[4]);
4228 /* Hmmm... The next two are dangerous */
4229 p = kzalloc(32, GFP_KERNEL);
4230 p[32 + sizeof(void *)] = 0x34;
4231 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4233 pr_err("If allocated object is overwritten then not detectable\n\n");
4235 validate_slab_cache(kmalloc_caches[5]);
4236 p = kzalloc(64, GFP_KERNEL);
4237 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4239 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4241 pr_err("If allocated object is overwritten then not detectable\n\n");
4242 validate_slab_cache(kmalloc_caches[6]);
4244 pr_err("\nB. Corruption after free\n");
4245 p = kzalloc(128, GFP_KERNEL);
4248 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4249 validate_slab_cache(kmalloc_caches[7]);
4251 p = kzalloc(256, GFP_KERNEL);
4254 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4255 validate_slab_cache(kmalloc_caches[8]);
4257 p = kzalloc(512, GFP_KERNEL);
4260 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4261 validate_slab_cache(kmalloc_caches[9]);
4265 static void resiliency_test(void) {};
4270 enum slab_stat_type {
4271 SL_ALL, /* All slabs */
4272 SL_PARTIAL, /* Only partially allocated slabs */
4273 SL_CPU, /* Only slabs used for cpu caches */
4274 SL_OBJECTS, /* Determine allocated objects not slabs */
4275 SL_TOTAL /* Determine object capacity not slabs */
4278 #define SO_ALL (1 << SL_ALL)
4279 #define SO_PARTIAL (1 << SL_PARTIAL)
4280 #define SO_CPU (1 << SL_CPU)
4281 #define SO_OBJECTS (1 << SL_OBJECTS)
4282 #define SO_TOTAL (1 << SL_TOTAL)
4284 static ssize_t show_slab_objects(struct kmem_cache *s,
4285 char *buf, unsigned long flags)
4287 unsigned long total = 0;
4290 unsigned long *nodes;
4292 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4296 if (flags & SO_CPU) {
4299 for_each_possible_cpu(cpu) {
4300 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4305 page = ACCESS_ONCE(c->page);
4309 node = page_to_nid(page);
4310 if (flags & SO_TOTAL)
4312 else if (flags & SO_OBJECTS)
4320 page = ACCESS_ONCE(c->partial);
4322 node = page_to_nid(page);
4323 if (flags & SO_TOTAL)
4325 else if (flags & SO_OBJECTS)
4336 #ifdef CONFIG_SLUB_DEBUG
4337 if (flags & SO_ALL) {
4338 for_each_node_state(node, N_NORMAL_MEMORY) {
4339 struct kmem_cache_node *n = get_node(s, node);
4341 if (flags & SO_TOTAL)
4342 x = atomic_long_read(&n->total_objects);
4343 else if (flags & SO_OBJECTS)
4344 x = atomic_long_read(&n->total_objects) -
4345 count_partial(n, count_free);
4347 x = atomic_long_read(&n->nr_slabs);
4354 if (flags & SO_PARTIAL) {
4355 for_each_node_state(node, N_NORMAL_MEMORY) {
4356 struct kmem_cache_node *n = get_node(s, node);
4358 if (flags & SO_TOTAL)
4359 x = count_partial(n, count_total);
4360 else if (flags & SO_OBJECTS)
4361 x = count_partial(n, count_inuse);
4368 x = sprintf(buf, "%lu", total);
4370 for_each_node_state(node, N_NORMAL_MEMORY)
4372 x += sprintf(buf + x, " N%d=%lu",
4377 return x + sprintf(buf + x, "\n");
4380 #ifdef CONFIG_SLUB_DEBUG
4381 static int any_slab_objects(struct kmem_cache *s)
4385 for_each_online_node(node) {
4386 struct kmem_cache_node *n = get_node(s, node);
4391 if (atomic_long_read(&n->total_objects))
4398 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4399 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4401 struct slab_attribute {
4402 struct attribute attr;
4403 ssize_t (*show)(struct kmem_cache *s, char *buf);
4404 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4407 #define SLAB_ATTR_RO(_name) \
4408 static struct slab_attribute _name##_attr = \
4409 __ATTR(_name, 0400, _name##_show, NULL)
4411 #define SLAB_ATTR(_name) \
4412 static struct slab_attribute _name##_attr = \
4413 __ATTR(_name, 0600, _name##_show, _name##_store)
4415 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4417 return sprintf(buf, "%d\n", s->size);
4419 SLAB_ATTR_RO(slab_size);
4421 static ssize_t align_show(struct kmem_cache *s, char *buf)
4423 return sprintf(buf, "%d\n", s->align);
4425 SLAB_ATTR_RO(align);
4427 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4429 return sprintf(buf, "%d\n", s->object_size);
4431 SLAB_ATTR_RO(object_size);
4433 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4435 return sprintf(buf, "%d\n", oo_objects(s->oo));
4437 SLAB_ATTR_RO(objs_per_slab);
4439 static ssize_t order_store(struct kmem_cache *s,
4440 const char *buf, size_t length)
4442 unsigned long order;
4445 err = kstrtoul(buf, 10, &order);
4449 if (order > slub_max_order || order < slub_min_order)
4452 calculate_sizes(s, order);
4456 static ssize_t order_show(struct kmem_cache *s, char *buf)
4458 return sprintf(buf, "%d\n", oo_order(s->oo));
4462 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4464 return sprintf(buf, "%lu\n", s->min_partial);
4467 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4473 err = kstrtoul(buf, 10, &min);
4477 set_min_partial(s, min);
4480 SLAB_ATTR(min_partial);
4482 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4484 return sprintf(buf, "%u\n", s->cpu_partial);
4487 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4490 unsigned long objects;
4493 err = kstrtoul(buf, 10, &objects);
4496 if (objects && !kmem_cache_has_cpu_partial(s))
4499 s->cpu_partial = objects;
4503 SLAB_ATTR(cpu_partial);
4505 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4509 return sprintf(buf, "%pS\n", s->ctor);
4513 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4515 return sprintf(buf, "%d\n", s->refcount - 1);
4517 SLAB_ATTR_RO(aliases);
4519 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4521 return show_slab_objects(s, buf, SO_PARTIAL);
4523 SLAB_ATTR_RO(partial);
4525 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4527 return show_slab_objects(s, buf, SO_CPU);
4529 SLAB_ATTR_RO(cpu_slabs);
4531 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4533 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4535 SLAB_ATTR_RO(objects);
4537 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4539 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4541 SLAB_ATTR_RO(objects_partial);
4543 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4550 for_each_online_cpu(cpu) {
4551 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4554 pages += page->pages;
4555 objects += page->pobjects;
4559 len = sprintf(buf, "%d(%d)", objects, pages);
4562 for_each_online_cpu(cpu) {
4563 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4565 if (page && len < PAGE_SIZE - 20)
4566 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4567 page->pobjects, page->pages);
4570 return len + sprintf(buf + len, "\n");
4572 SLAB_ATTR_RO(slabs_cpu_partial);
4574 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4579 static ssize_t reclaim_account_store(struct kmem_cache *s,
4580 const char *buf, size_t length)
4582 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4584 s->flags |= SLAB_RECLAIM_ACCOUNT;
4587 SLAB_ATTR(reclaim_account);
4589 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4591 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4593 SLAB_ATTR_RO(hwcache_align);
4595 #ifdef CONFIG_ZONE_DMA
4596 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4598 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4600 SLAB_ATTR_RO(cache_dma);
4603 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4605 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4607 SLAB_ATTR_RO(destroy_by_rcu);
4609 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4611 return sprintf(buf, "%d\n", s->reserved);
4613 SLAB_ATTR_RO(reserved);
4615 #ifdef CONFIG_SLUB_DEBUG
4616 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4618 return show_slab_objects(s, buf, SO_ALL);
4620 SLAB_ATTR_RO(slabs);
4622 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4624 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4626 SLAB_ATTR_RO(total_objects);
4628 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4630 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4633 static ssize_t sanity_checks_store(struct kmem_cache *s,
4634 const char *buf, size_t length)
4636 s->flags &= ~SLAB_DEBUG_FREE;
4637 if (buf[0] == '1') {
4638 s->flags &= ~__CMPXCHG_DOUBLE;
4639 s->flags |= SLAB_DEBUG_FREE;
4643 SLAB_ATTR(sanity_checks);
4645 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4647 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4650 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4653 s->flags &= ~SLAB_TRACE;
4654 if (buf[0] == '1') {
4655 s->flags &= ~__CMPXCHG_DOUBLE;
4656 s->flags |= SLAB_TRACE;
4662 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4664 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4667 static ssize_t red_zone_store(struct kmem_cache *s,
4668 const char *buf, size_t length)
4670 if (any_slab_objects(s))
4673 s->flags &= ~SLAB_RED_ZONE;
4674 if (buf[0] == '1') {
4675 s->flags &= ~__CMPXCHG_DOUBLE;
4676 s->flags |= SLAB_RED_ZONE;
4678 calculate_sizes(s, -1);
4681 SLAB_ATTR(red_zone);
4683 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4685 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4688 static ssize_t poison_store(struct kmem_cache *s,
4689 const char *buf, size_t length)
4691 if (any_slab_objects(s))
4694 s->flags &= ~SLAB_POISON;
4695 if (buf[0] == '1') {
4696 s->flags &= ~__CMPXCHG_DOUBLE;
4697 s->flags |= SLAB_POISON;
4699 calculate_sizes(s, -1);
4704 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4706 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4709 static ssize_t store_user_store(struct kmem_cache *s,
4710 const char *buf, size_t length)
4712 if (any_slab_objects(s))
4715 s->flags &= ~SLAB_STORE_USER;
4716 if (buf[0] == '1') {
4717 s->flags &= ~__CMPXCHG_DOUBLE;
4718 s->flags |= SLAB_STORE_USER;
4720 calculate_sizes(s, -1);
4723 SLAB_ATTR(store_user);
4725 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4730 static ssize_t validate_store(struct kmem_cache *s,
4731 const char *buf, size_t length)
4735 if (buf[0] == '1') {
4736 ret = validate_slab_cache(s);
4742 SLAB_ATTR(validate);
4744 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4746 if (!(s->flags & SLAB_STORE_USER))
4748 return list_locations(s, buf, TRACK_ALLOC);
4750 SLAB_ATTR_RO(alloc_calls);
4752 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4754 if (!(s->flags & SLAB_STORE_USER))
4756 return list_locations(s, buf, TRACK_FREE);
4758 SLAB_ATTR_RO(free_calls);
4759 #endif /* CONFIG_SLUB_DEBUG */
4761 #ifdef CONFIG_FAILSLAB
4762 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4764 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4767 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4770 s->flags &= ~SLAB_FAILSLAB;
4772 s->flags |= SLAB_FAILSLAB;
4775 SLAB_ATTR(failslab);
4778 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4783 static ssize_t shrink_store(struct kmem_cache *s,
4784 const char *buf, size_t length)
4786 if (buf[0] == '1') {
4787 int rc = kmem_cache_shrink(s);
4798 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4800 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4803 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4804 const char *buf, size_t length)
4806 unsigned long ratio;
4809 err = kstrtoul(buf, 10, &ratio);
4814 s->remote_node_defrag_ratio = ratio * 10;
4818 SLAB_ATTR(remote_node_defrag_ratio);
4821 #ifdef CONFIG_SLUB_STATS
4822 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4824 unsigned long sum = 0;
4827 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4832 for_each_online_cpu(cpu) {
4833 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4839 len = sprintf(buf, "%lu", sum);
4842 for_each_online_cpu(cpu) {
4843 if (data[cpu] && len < PAGE_SIZE - 20)
4844 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4848 return len + sprintf(buf + len, "\n");
4851 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4855 for_each_online_cpu(cpu)
4856 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4859 #define STAT_ATTR(si, text) \
4860 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4862 return show_stat(s, buf, si); \
4864 static ssize_t text##_store(struct kmem_cache *s, \
4865 const char *buf, size_t length) \
4867 if (buf[0] != '0') \
4869 clear_stat(s, si); \
4874 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4875 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4876 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4877 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4878 STAT_ATTR(FREE_FROZEN, free_frozen);
4879 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4880 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4881 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4882 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4883 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4884 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4885 STAT_ATTR(FREE_SLAB, free_slab);
4886 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4887 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4888 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4889 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4890 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4891 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4892 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4893 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4894 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4895 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4896 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4897 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4898 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4899 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4902 static struct attribute *slab_attrs[] = {
4903 &slab_size_attr.attr,
4904 &object_size_attr.attr,
4905 &objs_per_slab_attr.attr,
4907 &min_partial_attr.attr,
4908 &cpu_partial_attr.attr,
4910 &objects_partial_attr.attr,
4912 &cpu_slabs_attr.attr,
4916 &hwcache_align_attr.attr,
4917 &reclaim_account_attr.attr,
4918 &destroy_by_rcu_attr.attr,
4920 &reserved_attr.attr,
4921 &slabs_cpu_partial_attr.attr,
4922 #ifdef CONFIG_SLUB_DEBUG
4923 &total_objects_attr.attr,
4925 &sanity_checks_attr.attr,
4927 &red_zone_attr.attr,
4929 &store_user_attr.attr,
4930 &validate_attr.attr,
4931 &alloc_calls_attr.attr,
4932 &free_calls_attr.attr,
4934 #ifdef CONFIG_ZONE_DMA
4935 &cache_dma_attr.attr,
4938 &remote_node_defrag_ratio_attr.attr,
4940 #ifdef CONFIG_SLUB_STATS
4941 &alloc_fastpath_attr.attr,
4942 &alloc_slowpath_attr.attr,
4943 &free_fastpath_attr.attr,
4944 &free_slowpath_attr.attr,
4945 &free_frozen_attr.attr,
4946 &free_add_partial_attr.attr,
4947 &free_remove_partial_attr.attr,
4948 &alloc_from_partial_attr.attr,
4949 &alloc_slab_attr.attr,
4950 &alloc_refill_attr.attr,
4951 &alloc_node_mismatch_attr.attr,
4952 &free_slab_attr.attr,
4953 &cpuslab_flush_attr.attr,
4954 &deactivate_full_attr.attr,
4955 &deactivate_empty_attr.attr,
4956 &deactivate_to_head_attr.attr,
4957 &deactivate_to_tail_attr.attr,
4958 &deactivate_remote_frees_attr.attr,
4959 &deactivate_bypass_attr.attr,
4960 &order_fallback_attr.attr,
4961 &cmpxchg_double_fail_attr.attr,
4962 &cmpxchg_double_cpu_fail_attr.attr,
4963 &cpu_partial_alloc_attr.attr,
4964 &cpu_partial_free_attr.attr,
4965 &cpu_partial_node_attr.attr,
4966 &cpu_partial_drain_attr.attr,
4968 #ifdef CONFIG_FAILSLAB
4969 &failslab_attr.attr,
4975 static struct attribute_group slab_attr_group = {
4976 .attrs = slab_attrs,
4979 static ssize_t slab_attr_show(struct kobject *kobj,
4980 struct attribute *attr,
4983 struct slab_attribute *attribute;
4984 struct kmem_cache *s;
4987 attribute = to_slab_attr(attr);
4990 if (!attribute->show)
4993 err = attribute->show(s, buf);
4998 static ssize_t slab_attr_store(struct kobject *kobj,
4999 struct attribute *attr,
5000 const char *buf, size_t len)
5002 struct slab_attribute *attribute;
5003 struct kmem_cache *s;
5006 attribute = to_slab_attr(attr);
5009 if (!attribute->store)
5012 err = attribute->store(s, buf, len);
5013 #ifdef CONFIG_MEMCG_KMEM
5014 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5017 mutex_lock(&slab_mutex);
5018 if (s->max_attr_size < len)
5019 s->max_attr_size = len;
5022 * This is a best effort propagation, so this function's return
5023 * value will be determined by the parent cache only. This is
5024 * basically because not all attributes will have a well
5025 * defined semantics for rollbacks - most of the actions will
5026 * have permanent effects.
5028 * Returning the error value of any of the children that fail
5029 * is not 100 % defined, in the sense that users seeing the
5030 * error code won't be able to know anything about the state of
5033 * Only returning the error code for the parent cache at least
5034 * has well defined semantics. The cache being written to
5035 * directly either failed or succeeded, in which case we loop
5036 * through the descendants with best-effort propagation.
5038 for_each_memcg_cache_index(i) {
5039 struct kmem_cache *c = cache_from_memcg_idx(s, i);
5041 attribute->store(c, buf, len);
5043 mutex_unlock(&slab_mutex);
5049 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5051 #ifdef CONFIG_MEMCG_KMEM
5053 char *buffer = NULL;
5054 struct kmem_cache *root_cache;
5056 if (is_root_cache(s))
5059 root_cache = s->memcg_params->root_cache;
5062 * This mean this cache had no attribute written. Therefore, no point
5063 * in copying default values around
5065 if (!root_cache->max_attr_size)
5068 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5071 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5073 if (!attr || !attr->store || !attr->show)
5077 * It is really bad that we have to allocate here, so we will
5078 * do it only as a fallback. If we actually allocate, though,
5079 * we can just use the allocated buffer until the end.
5081 * Most of the slub attributes will tend to be very small in
5082 * size, but sysfs allows buffers up to a page, so they can
5083 * theoretically happen.
5087 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5090 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5091 if (WARN_ON(!buffer))
5096 attr->show(root_cache, buf);
5097 attr->store(s, buf, strlen(buf));
5101 free_page((unsigned long)buffer);
5105 static void kmem_cache_release(struct kobject *k)
5107 slab_kmem_cache_release(to_slab(k));
5110 static const struct sysfs_ops slab_sysfs_ops = {
5111 .show = slab_attr_show,
5112 .store = slab_attr_store,
5115 static struct kobj_type slab_ktype = {
5116 .sysfs_ops = &slab_sysfs_ops,
5117 .release = kmem_cache_release,
5120 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5122 struct kobj_type *ktype = get_ktype(kobj);
5124 if (ktype == &slab_ktype)
5129 static const struct kset_uevent_ops slab_uevent_ops = {
5130 .filter = uevent_filter,
5133 static struct kset *slab_kset;
5135 static inline struct kset *cache_kset(struct kmem_cache *s)
5137 #ifdef CONFIG_MEMCG_KMEM
5138 if (!is_root_cache(s))
5139 return s->memcg_params->root_cache->memcg_kset;
5144 #define ID_STR_LENGTH 64
5146 /* Create a unique string id for a slab cache:
5148 * Format :[flags-]size
5150 static char *create_unique_id(struct kmem_cache *s)
5152 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5159 * First flags affecting slabcache operations. We will only
5160 * get here for aliasable slabs so we do not need to support
5161 * too many flags. The flags here must cover all flags that
5162 * are matched during merging to guarantee that the id is
5165 if (s->flags & SLAB_CACHE_DMA)
5167 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5169 if (s->flags & SLAB_DEBUG_FREE)
5171 if (!(s->flags & SLAB_NOTRACK))
5175 p += sprintf(p, "%07d", s->size);
5177 #ifdef CONFIG_MEMCG_KMEM
5178 if (!is_root_cache(s))
5179 p += sprintf(p, "-%08d",
5180 memcg_cache_id(s->memcg_params->memcg));
5183 BUG_ON(p > name + ID_STR_LENGTH - 1);
5187 static int sysfs_slab_add(struct kmem_cache *s)
5191 int unmergeable = slab_unmergeable(s);
5195 * Slabcache can never be merged so we can use the name proper.
5196 * This is typically the case for debug situations. In that
5197 * case we can catch duplicate names easily.
5199 sysfs_remove_link(&slab_kset->kobj, s->name);
5203 * Create a unique name for the slab as a target
5206 name = create_unique_id(s);
5209 s->kobj.kset = cache_kset(s);
5210 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5214 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5218 #ifdef CONFIG_MEMCG_KMEM
5219 if (is_root_cache(s)) {
5220 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5221 if (!s->memcg_kset) {
5228 kobject_uevent(&s->kobj, KOBJ_ADD);
5230 /* Setup first alias */
5231 sysfs_slab_alias(s, s->name);
5238 kobject_del(&s->kobj);
5240 kobject_put(&s->kobj);
5244 void sysfs_slab_remove(struct kmem_cache *s)
5246 if (slab_state < FULL)
5248 * Sysfs has not been setup yet so no need to remove the
5253 #ifdef CONFIG_MEMCG_KMEM
5254 kset_unregister(s->memcg_kset);
5256 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5257 kobject_del(&s->kobj);
5258 kobject_put(&s->kobj);
5262 * Need to buffer aliases during bootup until sysfs becomes
5263 * available lest we lose that information.
5265 struct saved_alias {
5266 struct kmem_cache *s;
5268 struct saved_alias *next;
5271 static struct saved_alias *alias_list;
5273 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5275 struct saved_alias *al;
5277 if (slab_state == FULL) {
5279 * If we have a leftover link then remove it.
5281 sysfs_remove_link(&slab_kset->kobj, name);
5282 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5285 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5291 al->next = alias_list;
5296 static int __init slab_sysfs_init(void)
5298 struct kmem_cache *s;
5301 mutex_lock(&slab_mutex);
5303 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5305 mutex_unlock(&slab_mutex);
5306 pr_err("Cannot register slab subsystem.\n");
5312 list_for_each_entry(s, &slab_caches, list) {
5313 err = sysfs_slab_add(s);
5315 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5319 while (alias_list) {
5320 struct saved_alias *al = alias_list;
5322 alias_list = alias_list->next;
5323 err = sysfs_slab_alias(al->s, al->name);
5325 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5330 mutex_unlock(&slab_mutex);
5335 __initcall(slab_sysfs_init);
5336 #endif /* CONFIG_SYSFS */
5339 * The /proc/slabinfo ABI
5341 #ifdef CONFIG_SLABINFO
5342 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5344 unsigned long nr_slabs = 0;
5345 unsigned long nr_objs = 0;
5346 unsigned long nr_free = 0;
5349 for_each_online_node(node) {
5350 struct kmem_cache_node *n = get_node(s, node);
5355 nr_slabs += node_nr_slabs(n);
5356 nr_objs += node_nr_objs(n);
5357 nr_free += count_partial(n, count_free);
5360 sinfo->active_objs = nr_objs - nr_free;
5361 sinfo->num_objs = nr_objs;
5362 sinfo->active_slabs = nr_slabs;
5363 sinfo->num_slabs = nr_slabs;
5364 sinfo->objects_per_slab = oo_objects(s->oo);
5365 sinfo->cache_order = oo_order(s->oo);
5368 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5372 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5373 size_t count, loff_t *ppos)
5377 #endif /* CONFIG_SLABINFO */