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 #define OO_MASK ((1 << OO_SHIFT) - 1)
174 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
176 /* Internal SLUB flags */
177 #define __OBJECT_POISON 0x80000000UL /* Poison object */
178 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
181 static struct notifier_block slab_notifier;
185 * Tracking user of a slab.
187 #define TRACK_ADDRS_COUNT 16
189 unsigned long addr; /* Called from address */
190 #ifdef CONFIG_STACKTRACE
191 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
193 int cpu; /* Was running on cpu */
194 int pid; /* Pid context */
195 unsigned long when; /* When did the operation occur */
198 enum track_item { TRACK_ALLOC, TRACK_FREE };
201 static int sysfs_slab_add(struct kmem_cache *);
202 static int sysfs_slab_alias(struct kmem_cache *, const char *);
203 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
205 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
206 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
208 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
211 static inline void stat(const struct kmem_cache *s, enum stat_item si)
213 #ifdef CONFIG_SLUB_STATS
215 * The rmw is racy on a preemptible kernel but this is acceptable, so
216 * avoid this_cpu_add()'s irq-disable overhead.
218 raw_cpu_inc(s->cpu_slab->stat[si]);
222 /********************************************************************
223 * Core slab cache functions
224 *******************************************************************/
226 /* Verify that a pointer has an address that is valid within a slab page */
227 static inline int check_valid_pointer(struct kmem_cache *s,
228 struct page *page, const void *object)
235 base = page_address(page);
236 if (object < base || object >= base + page->objects * s->size ||
237 (object - base) % s->size) {
244 static inline void *get_freepointer(struct kmem_cache *s, void *object)
246 return *(void **)(object + s->offset);
249 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
251 prefetch(object + s->offset);
254 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
258 #ifdef CONFIG_DEBUG_PAGEALLOC
259 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
261 p = get_freepointer(s, object);
266 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
268 *(void **)(object + s->offset) = fp;
271 /* Loop over all objects in a slab */
272 #define for_each_object(__p, __s, __addr, __objects) \
273 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
276 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
277 for (__p = (__addr), __idx = 1; __idx <= __objects;\
278 __p += (__s)->size, __idx++)
280 /* Determine object index from a given position */
281 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
283 return (p - addr) / s->size;
286 static inline size_t slab_ksize(const struct kmem_cache *s)
288 #ifdef CONFIG_SLUB_DEBUG
290 * Debugging requires use of the padding between object
291 * and whatever may come after it.
293 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
294 return s->object_size;
298 * If we have the need to store the freelist pointer
299 * back there or track user information then we can
300 * only use the space before that information.
302 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
305 * Else we can use all the padding etc for the allocation
310 static inline int order_objects(int order, unsigned long size, int reserved)
312 return ((PAGE_SIZE << order) - reserved) / size;
315 static inline struct kmem_cache_order_objects oo_make(int order,
316 unsigned long size, int reserved)
318 struct kmem_cache_order_objects x = {
319 (order << OO_SHIFT) + order_objects(order, size, reserved)
325 static inline int oo_order(struct kmem_cache_order_objects x)
327 return x.x >> OO_SHIFT;
330 static inline int oo_objects(struct kmem_cache_order_objects x)
332 return x.x & OO_MASK;
336 * Per slab locking using the pagelock
338 static __always_inline void slab_lock(struct page *page)
340 bit_spin_lock(PG_locked, &page->flags);
343 static __always_inline void slab_unlock(struct page *page)
345 __bit_spin_unlock(PG_locked, &page->flags);
348 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
351 tmp.counters = counters_new;
353 * page->counters can cover frozen/inuse/objects as well
354 * as page->_count. If we assign to ->counters directly
355 * we run the risk of losing updates to page->_count, so
356 * be careful and only assign to the fields we need.
358 page->frozen = tmp.frozen;
359 page->inuse = tmp.inuse;
360 page->objects = tmp.objects;
363 /* Interrupts must be disabled (for the fallback code to work right) */
364 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
365 void *freelist_old, unsigned long counters_old,
366 void *freelist_new, unsigned long counters_new,
369 VM_BUG_ON(!irqs_disabled());
370 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
371 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
372 if (s->flags & __CMPXCHG_DOUBLE) {
373 if (cmpxchg_double(&page->freelist, &page->counters,
374 freelist_old, counters_old,
375 freelist_new, counters_new))
381 if (page->freelist == freelist_old &&
382 page->counters == counters_old) {
383 page->freelist = freelist_new;
384 set_page_slub_counters(page, counters_new);
392 stat(s, CMPXCHG_DOUBLE_FAIL);
394 #ifdef SLUB_DEBUG_CMPXCHG
395 pr_info("%s %s: cmpxchg double redo ", n, s->name);
401 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
402 void *freelist_old, unsigned long counters_old,
403 void *freelist_new, unsigned long counters_new,
406 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
407 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
408 if (s->flags & __CMPXCHG_DOUBLE) {
409 if (cmpxchg_double(&page->freelist, &page->counters,
410 freelist_old, counters_old,
411 freelist_new, counters_new))
418 local_irq_save(flags);
420 if (page->freelist == freelist_old &&
421 page->counters == counters_old) {
422 page->freelist = freelist_new;
423 set_page_slub_counters(page, counters_new);
425 local_irq_restore(flags);
429 local_irq_restore(flags);
433 stat(s, CMPXCHG_DOUBLE_FAIL);
435 #ifdef SLUB_DEBUG_CMPXCHG
436 pr_info("%s %s: cmpxchg double redo ", n, s->name);
442 #ifdef CONFIG_SLUB_DEBUG
444 * Determine a map of object in use on a page.
446 * Node listlock must be held to guarantee that the page does
447 * not vanish from under us.
449 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
452 void *addr = page_address(page);
454 for (p = page->freelist; p; p = get_freepointer(s, p))
455 set_bit(slab_index(p, s, addr), map);
461 #ifdef CONFIG_SLUB_DEBUG_ON
462 static int slub_debug = DEBUG_DEFAULT_FLAGS;
464 static int slub_debug;
467 static char *slub_debug_slabs;
468 static int disable_higher_order_debug;
473 static void print_section(char *text, u8 *addr, unsigned int length)
475 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
479 static struct track *get_track(struct kmem_cache *s, void *object,
480 enum track_item alloc)
485 p = object + s->offset + sizeof(void *);
487 p = object + s->inuse;
492 static void set_track(struct kmem_cache *s, void *object,
493 enum track_item alloc, unsigned long addr)
495 struct track *p = get_track(s, object, alloc);
498 #ifdef CONFIG_STACKTRACE
499 struct stack_trace trace;
502 trace.nr_entries = 0;
503 trace.max_entries = TRACK_ADDRS_COUNT;
504 trace.entries = p->addrs;
506 save_stack_trace(&trace);
508 /* See rant in lockdep.c */
509 if (trace.nr_entries != 0 &&
510 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
513 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
517 p->cpu = smp_processor_id();
518 p->pid = current->pid;
521 memset(p, 0, sizeof(struct track));
524 static void init_tracking(struct kmem_cache *s, void *object)
526 if (!(s->flags & SLAB_STORE_USER))
529 set_track(s, object, TRACK_FREE, 0UL);
530 set_track(s, object, TRACK_ALLOC, 0UL);
533 static void print_track(const char *s, struct track *t)
538 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
539 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
540 #ifdef CONFIG_STACKTRACE
543 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
545 pr_err("\t%pS\n", (void *)t->addrs[i]);
552 static void print_tracking(struct kmem_cache *s, void *object)
554 if (!(s->flags & SLAB_STORE_USER))
557 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
558 print_track("Freed", get_track(s, object, TRACK_FREE));
561 static void print_page_info(struct page *page)
563 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
564 page, page->objects, page->inuse, page->freelist, page->flags);
568 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
570 struct va_format vaf;
576 pr_err("=============================================================================\n");
577 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
578 pr_err("-----------------------------------------------------------------------------\n\n");
580 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
584 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
586 struct va_format vaf;
592 pr_err("FIX %s: %pV\n", s->name, &vaf);
596 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
598 unsigned int off; /* Offset of last byte */
599 u8 *addr = page_address(page);
601 print_tracking(s, p);
603 print_page_info(page);
605 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
606 p, p - addr, get_freepointer(s, p));
609 print_section("Bytes b4 ", p - 16, 16);
611 print_section("Object ", p, min_t(unsigned long, s->object_size,
613 if (s->flags & SLAB_RED_ZONE)
614 print_section("Redzone ", p + s->object_size,
615 s->inuse - s->object_size);
618 off = s->offset + sizeof(void *);
622 if (s->flags & SLAB_STORE_USER)
623 off += 2 * sizeof(struct track);
626 /* Beginning of the filler is the free pointer */
627 print_section("Padding ", p + off, s->size - off);
632 static void object_err(struct kmem_cache *s, struct page *page,
633 u8 *object, char *reason)
635 slab_bug(s, "%s", reason);
636 print_trailer(s, page, object);
639 static void slab_err(struct kmem_cache *s, struct page *page,
640 const char *fmt, ...)
646 vsnprintf(buf, sizeof(buf), fmt, args);
648 slab_bug(s, "%s", buf);
649 print_page_info(page);
653 static void init_object(struct kmem_cache *s, void *object, u8 val)
657 if (s->flags & __OBJECT_POISON) {
658 memset(p, POISON_FREE, s->object_size - 1);
659 p[s->object_size - 1] = POISON_END;
662 if (s->flags & SLAB_RED_ZONE)
663 memset(p + s->object_size, val, s->inuse - s->object_size);
666 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667 void *from, void *to)
669 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670 memset(from, data, to - from);
673 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674 u8 *object, char *what,
675 u8 *start, unsigned int value, unsigned int bytes)
680 fault = memchr_inv(start, value, bytes);
685 while (end > fault && end[-1] == value)
688 slab_bug(s, "%s overwritten", what);
689 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690 fault, end - 1, fault[0], value);
691 print_trailer(s, page, object);
693 restore_bytes(s, what, value, fault, end);
701 * Bytes of the object to be managed.
702 * If the freepointer may overlay the object then the free
703 * pointer is the first word of the object.
705 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
708 * object + s->object_size
709 * Padding to reach word boundary. This is also used for Redzoning.
710 * Padding is extended by another word if Redzoning is enabled and
711 * object_size == inuse.
713 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714 * 0xcc (RED_ACTIVE) for objects in use.
717 * Meta data starts here.
719 * A. Free pointer (if we cannot overwrite object on free)
720 * B. Tracking data for SLAB_STORE_USER
721 * C. Padding to reach required alignment boundary or at mininum
722 * one word if debugging is on to be able to detect writes
723 * before the word boundary.
725 * Padding is done using 0x5a (POISON_INUSE)
728 * Nothing is used beyond s->size.
730 * If slabcaches are merged then the object_size and inuse boundaries are mostly
731 * ignored. And therefore no slab options that rely on these boundaries
732 * may be used with merged slabcaches.
735 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
737 unsigned long off = s->inuse; /* The end of info */
740 /* Freepointer is placed after the object. */
741 off += sizeof(void *);
743 if (s->flags & SLAB_STORE_USER)
744 /* We also have user information there */
745 off += 2 * sizeof(struct track);
750 return check_bytes_and_report(s, page, p, "Object padding",
751 p + off, POISON_INUSE, s->size - off);
754 /* Check the pad bytes at the end of a slab page */
755 static int slab_pad_check(struct kmem_cache *s, struct page *page)
763 if (!(s->flags & SLAB_POISON))
766 start = page_address(page);
767 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768 end = start + length;
769 remainder = length % s->size;
773 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
776 while (end > fault && end[-1] == POISON_INUSE)
779 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780 print_section("Padding ", end - remainder, remainder);
782 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
786 static int check_object(struct kmem_cache *s, struct page *page,
787 void *object, u8 val)
790 u8 *endobject = object + s->object_size;
792 if (s->flags & SLAB_RED_ZONE) {
793 if (!check_bytes_and_report(s, page, object, "Redzone",
794 endobject, val, s->inuse - s->object_size))
797 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
798 check_bytes_and_report(s, page, p, "Alignment padding",
799 endobject, POISON_INUSE,
800 s->inuse - s->object_size);
804 if (s->flags & SLAB_POISON) {
805 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
806 (!check_bytes_and_report(s, page, p, "Poison", p,
807 POISON_FREE, s->object_size - 1) ||
808 !check_bytes_and_report(s, page, p, "Poison",
809 p + s->object_size - 1, POISON_END, 1)))
812 * check_pad_bytes cleans up on its own.
814 check_pad_bytes(s, page, p);
817 if (!s->offset && val == SLUB_RED_ACTIVE)
819 * Object and freepointer overlap. Cannot check
820 * freepointer while object is allocated.
824 /* Check free pointer validity */
825 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
826 object_err(s, page, p, "Freepointer corrupt");
828 * No choice but to zap it and thus lose the remainder
829 * of the free objects in this slab. May cause
830 * another error because the object count is now wrong.
832 set_freepointer(s, p, NULL);
838 static int check_slab(struct kmem_cache *s, struct page *page)
842 VM_BUG_ON(!irqs_disabled());
844 if (!PageSlab(page)) {
845 slab_err(s, page, "Not a valid slab page");
849 maxobj = order_objects(compound_order(page), s->size, s->reserved);
850 if (page->objects > maxobj) {
851 slab_err(s, page, "objects %u > max %u",
852 page->objects, maxobj);
855 if (page->inuse > page->objects) {
856 slab_err(s, page, "inuse %u > max %u",
857 page->inuse, page->objects);
860 /* Slab_pad_check fixes things up after itself */
861 slab_pad_check(s, page);
866 * Determine if a certain object on a page is on the freelist. Must hold the
867 * slab lock to guarantee that the chains are in a consistent state.
869 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
877 while (fp && nr <= page->objects) {
880 if (!check_valid_pointer(s, page, fp)) {
882 object_err(s, page, object,
883 "Freechain corrupt");
884 set_freepointer(s, object, NULL);
886 slab_err(s, page, "Freepointer corrupt");
887 page->freelist = NULL;
888 page->inuse = page->objects;
889 slab_fix(s, "Freelist cleared");
895 fp = get_freepointer(s, object);
899 max_objects = order_objects(compound_order(page), s->size, s->reserved);
900 if (max_objects > MAX_OBJS_PER_PAGE)
901 max_objects = MAX_OBJS_PER_PAGE;
903 if (page->objects != max_objects) {
904 slab_err(s, page, "Wrong number of objects. Found %d but "
905 "should be %d", page->objects, max_objects);
906 page->objects = max_objects;
907 slab_fix(s, "Number of objects adjusted.");
909 if (page->inuse != page->objects - nr) {
910 slab_err(s, page, "Wrong object count. Counter is %d but "
911 "counted were %d", page->inuse, page->objects - nr);
912 page->inuse = page->objects - nr;
913 slab_fix(s, "Object count adjusted.");
915 return search == NULL;
918 static void trace(struct kmem_cache *s, struct page *page, void *object,
921 if (s->flags & SLAB_TRACE) {
922 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
924 alloc ? "alloc" : "free",
929 print_section("Object ", (void *)object,
937 * Tracking of fully allocated slabs for debugging purposes.
939 static void add_full(struct kmem_cache *s,
940 struct kmem_cache_node *n, struct page *page)
942 if (!(s->flags & SLAB_STORE_USER))
945 lockdep_assert_held(&n->list_lock);
946 list_add(&page->lru, &n->full);
949 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
951 if (!(s->flags & SLAB_STORE_USER))
954 lockdep_assert_held(&n->list_lock);
955 list_del(&page->lru);
958 /* Tracking of the number of slabs for debugging purposes */
959 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
961 struct kmem_cache_node *n = get_node(s, node);
963 return atomic_long_read(&n->nr_slabs);
966 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
968 return atomic_long_read(&n->nr_slabs);
971 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
973 struct kmem_cache_node *n = get_node(s, node);
976 * May be called early in order to allocate a slab for the
977 * kmem_cache_node structure. Solve the chicken-egg
978 * dilemma by deferring the increment of the count during
979 * bootstrap (see early_kmem_cache_node_alloc).
982 atomic_long_inc(&n->nr_slabs);
983 atomic_long_add(objects, &n->total_objects);
986 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
988 struct kmem_cache_node *n = get_node(s, node);
990 atomic_long_dec(&n->nr_slabs);
991 atomic_long_sub(objects, &n->total_objects);
994 /* Object debug checks for alloc/free paths */
995 static void setup_object_debug(struct kmem_cache *s, struct page *page,
998 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1001 init_object(s, object, SLUB_RED_INACTIVE);
1002 init_tracking(s, object);
1005 static noinline int alloc_debug_processing(struct kmem_cache *s,
1007 void *object, unsigned long addr)
1009 if (!check_slab(s, page))
1012 if (!check_valid_pointer(s, page, object)) {
1013 object_err(s, page, object, "Freelist Pointer check fails");
1017 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1020 /* Success perform special debug activities for allocs */
1021 if (s->flags & SLAB_STORE_USER)
1022 set_track(s, object, TRACK_ALLOC, addr);
1023 trace(s, page, object, 1);
1024 init_object(s, object, SLUB_RED_ACTIVE);
1028 if (PageSlab(page)) {
1030 * If this is a slab page then lets do the best we can
1031 * to avoid issues in the future. Marking all objects
1032 * as used avoids touching the remaining objects.
1034 slab_fix(s, "Marking all objects used");
1035 page->inuse = page->objects;
1036 page->freelist = NULL;
1041 static noinline struct kmem_cache_node *free_debug_processing(
1042 struct kmem_cache *s, struct page *page, void *object,
1043 unsigned long addr, unsigned long *flags)
1045 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1047 spin_lock_irqsave(&n->list_lock, *flags);
1050 if (!check_slab(s, page))
1053 if (!check_valid_pointer(s, page, object)) {
1054 slab_err(s, page, "Invalid object pointer 0x%p", object);
1058 if (on_freelist(s, page, object)) {
1059 object_err(s, page, object, "Object already free");
1063 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1066 if (unlikely(s != page->slab_cache)) {
1067 if (!PageSlab(page)) {
1068 slab_err(s, page, "Attempt to free object(0x%p) "
1069 "outside of slab", object);
1070 } else if (!page->slab_cache) {
1071 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1075 object_err(s, page, object,
1076 "page slab pointer corrupt.");
1080 if (s->flags & SLAB_STORE_USER)
1081 set_track(s, object, TRACK_FREE, addr);
1082 trace(s, page, object, 0);
1083 init_object(s, object, SLUB_RED_INACTIVE);
1087 * Keep node_lock to preserve integrity
1088 * until the object is actually freed
1094 spin_unlock_irqrestore(&n->list_lock, *flags);
1095 slab_fix(s, "Object at 0x%p not freed", object);
1099 static int __init setup_slub_debug(char *str)
1101 slub_debug = DEBUG_DEFAULT_FLAGS;
1102 if (*str++ != '=' || !*str)
1104 * No options specified. Switch on full debugging.
1110 * No options but restriction on slabs. This means full
1111 * debugging for slabs matching a pattern.
1115 if (tolower(*str) == 'o') {
1117 * Avoid enabling debugging on caches if its minimum order
1118 * would increase as a result.
1120 disable_higher_order_debug = 1;
1127 * Switch off all debugging measures.
1132 * Determine which debug features should be switched on
1134 for (; *str && *str != ','; str++) {
1135 switch (tolower(*str)) {
1137 slub_debug |= SLAB_DEBUG_FREE;
1140 slub_debug |= SLAB_RED_ZONE;
1143 slub_debug |= SLAB_POISON;
1146 slub_debug |= SLAB_STORE_USER;
1149 slub_debug |= SLAB_TRACE;
1152 slub_debug |= SLAB_FAILSLAB;
1155 pr_err("slub_debug option '%c' unknown. skipped\n",
1162 slub_debug_slabs = str + 1;
1167 __setup("slub_debug", setup_slub_debug);
1169 unsigned long kmem_cache_flags(unsigned long object_size,
1170 unsigned long flags, const char *name,
1171 void (*ctor)(void *))
1174 * Enable debugging if selected on the kernel commandline.
1176 if (slub_debug && (!slub_debug_slabs || (name &&
1177 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1178 flags |= slub_debug;
1183 static inline void setup_object_debug(struct kmem_cache *s,
1184 struct page *page, void *object) {}
1186 static inline int alloc_debug_processing(struct kmem_cache *s,
1187 struct page *page, void *object, unsigned long addr) { return 0; }
1189 static inline struct kmem_cache_node *free_debug_processing(
1190 struct kmem_cache *s, struct page *page, void *object,
1191 unsigned long addr, unsigned long *flags) { return NULL; }
1193 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1195 static inline int check_object(struct kmem_cache *s, struct page *page,
1196 void *object, u8 val) { return 1; }
1197 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1198 struct page *page) {}
1199 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1200 struct page *page) {}
1201 unsigned long kmem_cache_flags(unsigned long object_size,
1202 unsigned long flags, const char *name,
1203 void (*ctor)(void *))
1207 #define slub_debug 0
1209 #define disable_higher_order_debug 0
1211 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1213 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1215 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1217 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1220 #endif /* CONFIG_SLUB_DEBUG */
1223 * Hooks for other subsystems that check memory allocations. In a typical
1224 * production configuration these hooks all should produce no code at all.
1226 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1228 kmemleak_alloc(ptr, size, 1, flags);
1231 static inline void kfree_hook(const void *x)
1236 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1238 flags &= gfp_allowed_mask;
1239 lockdep_trace_alloc(flags);
1240 might_sleep_if(flags & __GFP_WAIT);
1242 return should_failslab(s->object_size, flags, s->flags);
1245 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1246 gfp_t flags, void *object)
1248 flags &= gfp_allowed_mask;
1249 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1250 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1253 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1255 kmemleak_free_recursive(x, s->flags);
1258 * Trouble is that we may no longer disable interrupts in the fast path
1259 * So in order to make the debug calls that expect irqs to be
1260 * disabled we need to disable interrupts temporarily.
1262 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1264 unsigned long flags;
1266 local_irq_save(flags);
1267 kmemcheck_slab_free(s, x, s->object_size);
1268 debug_check_no_locks_freed(x, s->object_size);
1269 local_irq_restore(flags);
1272 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1273 debug_check_no_obj_freed(x, s->object_size);
1277 * Slab allocation and freeing
1279 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1280 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1283 int order = oo_order(oo);
1285 flags |= __GFP_NOTRACK;
1287 if (memcg_charge_slab(s, flags, order))
1290 if (node == NUMA_NO_NODE)
1291 page = alloc_pages(flags, order);
1293 page = alloc_pages_exact_node(node, flags, order);
1296 memcg_uncharge_slab(s, order);
1301 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1304 struct kmem_cache_order_objects oo = s->oo;
1307 flags &= gfp_allowed_mask;
1309 if (flags & __GFP_WAIT)
1312 flags |= s->allocflags;
1315 * Let the initial higher-order allocation fail under memory pressure
1316 * so we fall-back to the minimum order allocation.
1318 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1320 page = alloc_slab_page(s, alloc_gfp, node, oo);
1321 if (unlikely(!page)) {
1325 * Allocation may have failed due to fragmentation.
1326 * Try a lower order alloc if possible
1328 page = alloc_slab_page(s, alloc_gfp, node, oo);
1331 stat(s, ORDER_FALLBACK);
1334 if (kmemcheck_enabled && page
1335 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1336 int pages = 1 << oo_order(oo);
1338 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1341 * Objects from caches that have a constructor don't get
1342 * cleared when they're allocated, so we need to do it here.
1345 kmemcheck_mark_uninitialized_pages(page, pages);
1347 kmemcheck_mark_unallocated_pages(page, pages);
1350 if (flags & __GFP_WAIT)
1351 local_irq_disable();
1355 page->objects = oo_objects(oo);
1356 mod_zone_page_state(page_zone(page),
1357 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1358 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1364 static void setup_object(struct kmem_cache *s, struct page *page,
1367 setup_object_debug(s, page, object);
1368 if (unlikely(s->ctor))
1372 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1380 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1381 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1385 page = allocate_slab(s,
1386 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1390 order = compound_order(page);
1391 inc_slabs_node(s, page_to_nid(page), page->objects);
1392 page->slab_cache = s;
1393 __SetPageSlab(page);
1394 if (page->pfmemalloc)
1395 SetPageSlabPfmemalloc(page);
1397 start = page_address(page);
1399 if (unlikely(s->flags & SLAB_POISON))
1400 memset(start, POISON_INUSE, PAGE_SIZE << order);
1402 for_each_object_idx(p, idx, s, start, page->objects) {
1403 setup_object(s, page, p);
1404 if (likely(idx < page->objects))
1405 set_freepointer(s, p, p + s->size);
1407 set_freepointer(s, p, NULL);
1410 page->freelist = start;
1411 page->inuse = page->objects;
1417 static void __free_slab(struct kmem_cache *s, struct page *page)
1419 int order = compound_order(page);
1420 int pages = 1 << order;
1422 if (kmem_cache_debug(s)) {
1425 slab_pad_check(s, page);
1426 for_each_object(p, s, page_address(page),
1428 check_object(s, page, p, SLUB_RED_INACTIVE);
1431 kmemcheck_free_shadow(page, compound_order(page));
1433 mod_zone_page_state(page_zone(page),
1434 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1435 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1438 __ClearPageSlabPfmemalloc(page);
1439 __ClearPageSlab(page);
1441 page_mapcount_reset(page);
1442 if (current->reclaim_state)
1443 current->reclaim_state->reclaimed_slab += pages;
1444 __free_pages(page, order);
1445 memcg_uncharge_slab(s, order);
1448 #define need_reserve_slab_rcu \
1449 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1451 static void rcu_free_slab(struct rcu_head *h)
1455 if (need_reserve_slab_rcu)
1456 page = virt_to_head_page(h);
1458 page = container_of((struct list_head *)h, struct page, lru);
1460 __free_slab(page->slab_cache, page);
1463 static void free_slab(struct kmem_cache *s, struct page *page)
1465 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1466 struct rcu_head *head;
1468 if (need_reserve_slab_rcu) {
1469 int order = compound_order(page);
1470 int offset = (PAGE_SIZE << order) - s->reserved;
1472 VM_BUG_ON(s->reserved != sizeof(*head));
1473 head = page_address(page) + offset;
1476 * RCU free overloads the RCU head over the LRU
1478 head = (void *)&page->lru;
1481 call_rcu(head, rcu_free_slab);
1483 __free_slab(s, page);
1486 static void discard_slab(struct kmem_cache *s, struct page *page)
1488 dec_slabs_node(s, page_to_nid(page), page->objects);
1493 * Management of partially allocated slabs.
1496 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1499 if (tail == DEACTIVATE_TO_TAIL)
1500 list_add_tail(&page->lru, &n->partial);
1502 list_add(&page->lru, &n->partial);
1505 static inline void add_partial(struct kmem_cache_node *n,
1506 struct page *page, int tail)
1508 lockdep_assert_held(&n->list_lock);
1509 __add_partial(n, page, tail);
1513 __remove_partial(struct kmem_cache_node *n, struct page *page)
1515 list_del(&page->lru);
1519 static inline void remove_partial(struct kmem_cache_node *n,
1522 lockdep_assert_held(&n->list_lock);
1523 __remove_partial(n, page);
1527 * Remove slab from the partial list, freeze it and
1528 * return the pointer to the freelist.
1530 * Returns a list of objects or NULL if it fails.
1532 static inline void *acquire_slab(struct kmem_cache *s,
1533 struct kmem_cache_node *n, struct page *page,
1534 int mode, int *objects)
1537 unsigned long counters;
1540 lockdep_assert_held(&n->list_lock);
1543 * Zap the freelist and set the frozen bit.
1544 * The old freelist is the list of objects for the
1545 * per cpu allocation list.
1547 freelist = page->freelist;
1548 counters = page->counters;
1549 new.counters = counters;
1550 *objects = new.objects - new.inuse;
1552 new.inuse = page->objects;
1553 new.freelist = NULL;
1555 new.freelist = freelist;
1558 VM_BUG_ON(new.frozen);
1561 if (!__cmpxchg_double_slab(s, page,
1563 new.freelist, new.counters,
1567 remove_partial(n, page);
1572 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1573 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1576 * Try to allocate a partial slab from a specific node.
1578 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1579 struct kmem_cache_cpu *c, gfp_t flags)
1581 struct page *page, *page2;
1582 void *object = NULL;
1587 * Racy check. If we mistakenly see no partial slabs then we
1588 * just allocate an empty slab. If we mistakenly try to get a
1589 * partial slab and there is none available then get_partials()
1592 if (!n || !n->nr_partial)
1595 spin_lock(&n->list_lock);
1596 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1599 if (!pfmemalloc_match(page, flags))
1602 t = acquire_slab(s, n, page, object == NULL, &objects);
1606 available += objects;
1609 stat(s, ALLOC_FROM_PARTIAL);
1612 put_cpu_partial(s, page, 0);
1613 stat(s, CPU_PARTIAL_NODE);
1615 if (!kmem_cache_has_cpu_partial(s)
1616 || available > s->cpu_partial / 2)
1620 spin_unlock(&n->list_lock);
1625 * Get a page from somewhere. Search in increasing NUMA distances.
1627 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1628 struct kmem_cache_cpu *c)
1631 struct zonelist *zonelist;
1634 enum zone_type high_zoneidx = gfp_zone(flags);
1636 unsigned int cpuset_mems_cookie;
1639 * The defrag ratio allows a configuration of the tradeoffs between
1640 * inter node defragmentation and node local allocations. A lower
1641 * defrag_ratio increases the tendency to do local allocations
1642 * instead of attempting to obtain partial slabs from other nodes.
1644 * If the defrag_ratio is set to 0 then kmalloc() always
1645 * returns node local objects. If the ratio is higher then kmalloc()
1646 * may return off node objects because partial slabs are obtained
1647 * from other nodes and filled up.
1649 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1650 * defrag_ratio = 1000) then every (well almost) allocation will
1651 * first attempt to defrag slab caches on other nodes. This means
1652 * scanning over all nodes to look for partial slabs which may be
1653 * expensive if we do it every time we are trying to find a slab
1654 * with available objects.
1656 if (!s->remote_node_defrag_ratio ||
1657 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1661 cpuset_mems_cookie = read_mems_allowed_begin();
1662 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1663 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1664 struct kmem_cache_node *n;
1666 n = get_node(s, zone_to_nid(zone));
1668 if (n && cpuset_zone_allowed(zone,
1669 flags | __GFP_HARDWALL) &&
1670 n->nr_partial > s->min_partial) {
1671 object = get_partial_node(s, n, c, flags);
1674 * Don't check read_mems_allowed_retry()
1675 * here - if mems_allowed was updated in
1676 * parallel, that was a harmless race
1677 * between allocation and the cpuset
1684 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1690 * Get a partial page, lock it and return it.
1692 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1693 struct kmem_cache_cpu *c)
1696 int searchnode = node;
1698 if (node == NUMA_NO_NODE)
1699 searchnode = numa_mem_id();
1700 else if (!node_present_pages(node))
1701 searchnode = node_to_mem_node(node);
1703 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1704 if (object || node != NUMA_NO_NODE)
1707 return get_any_partial(s, flags, c);
1710 #ifdef CONFIG_PREEMPT
1712 * Calculate the next globally unique transaction for disambiguiation
1713 * during cmpxchg. The transactions start with the cpu number and are then
1714 * incremented by CONFIG_NR_CPUS.
1716 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1719 * No preemption supported therefore also no need to check for
1725 static inline unsigned long next_tid(unsigned long tid)
1727 return tid + TID_STEP;
1730 static inline unsigned int tid_to_cpu(unsigned long tid)
1732 return tid % TID_STEP;
1735 static inline unsigned long tid_to_event(unsigned long tid)
1737 return tid / TID_STEP;
1740 static inline unsigned int init_tid(int cpu)
1745 static inline void note_cmpxchg_failure(const char *n,
1746 const struct kmem_cache *s, unsigned long tid)
1748 #ifdef SLUB_DEBUG_CMPXCHG
1749 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1751 pr_info("%s %s: cmpxchg redo ", n, s->name);
1753 #ifdef CONFIG_PREEMPT
1754 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1755 pr_warn("due to cpu change %d -> %d\n",
1756 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1759 if (tid_to_event(tid) != tid_to_event(actual_tid))
1760 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1761 tid_to_event(tid), tid_to_event(actual_tid));
1763 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1764 actual_tid, tid, next_tid(tid));
1766 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1769 static void init_kmem_cache_cpus(struct kmem_cache *s)
1773 for_each_possible_cpu(cpu)
1774 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1778 * Remove the cpu slab
1780 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1783 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1784 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1786 enum slab_modes l = M_NONE, m = M_NONE;
1788 int tail = DEACTIVATE_TO_HEAD;
1792 if (page->freelist) {
1793 stat(s, DEACTIVATE_REMOTE_FREES);
1794 tail = DEACTIVATE_TO_TAIL;
1798 * Stage one: Free all available per cpu objects back
1799 * to the page freelist while it is still frozen. Leave the
1802 * There is no need to take the list->lock because the page
1805 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1807 unsigned long counters;
1810 prior = page->freelist;
1811 counters = page->counters;
1812 set_freepointer(s, freelist, prior);
1813 new.counters = counters;
1815 VM_BUG_ON(!new.frozen);
1817 } while (!__cmpxchg_double_slab(s, page,
1819 freelist, new.counters,
1820 "drain percpu freelist"));
1822 freelist = nextfree;
1826 * Stage two: Ensure that the page is unfrozen while the
1827 * list presence reflects the actual number of objects
1830 * We setup the list membership and then perform a cmpxchg
1831 * with the count. If there is a mismatch then the page
1832 * is not unfrozen but the page is on the wrong list.
1834 * Then we restart the process which may have to remove
1835 * the page from the list that we just put it on again
1836 * because the number of objects in the slab may have
1841 old.freelist = page->freelist;
1842 old.counters = page->counters;
1843 VM_BUG_ON(!old.frozen);
1845 /* Determine target state of the slab */
1846 new.counters = old.counters;
1849 set_freepointer(s, freelist, old.freelist);
1850 new.freelist = freelist;
1852 new.freelist = old.freelist;
1856 if (!new.inuse && n->nr_partial >= s->min_partial)
1858 else if (new.freelist) {
1863 * Taking the spinlock removes the possiblity
1864 * that acquire_slab() will see a slab page that
1867 spin_lock(&n->list_lock);
1871 if (kmem_cache_debug(s) && !lock) {
1874 * This also ensures that the scanning of full
1875 * slabs from diagnostic functions will not see
1878 spin_lock(&n->list_lock);
1886 remove_partial(n, page);
1888 else if (l == M_FULL)
1890 remove_full(s, n, page);
1892 if (m == M_PARTIAL) {
1894 add_partial(n, page, tail);
1897 } else if (m == M_FULL) {
1899 stat(s, DEACTIVATE_FULL);
1900 add_full(s, n, page);
1906 if (!__cmpxchg_double_slab(s, page,
1907 old.freelist, old.counters,
1908 new.freelist, new.counters,
1913 spin_unlock(&n->list_lock);
1916 stat(s, DEACTIVATE_EMPTY);
1917 discard_slab(s, page);
1923 * Unfreeze all the cpu partial slabs.
1925 * This function must be called with interrupts disabled
1926 * for the cpu using c (or some other guarantee must be there
1927 * to guarantee no concurrent accesses).
1929 static void unfreeze_partials(struct kmem_cache *s,
1930 struct kmem_cache_cpu *c)
1932 #ifdef CONFIG_SLUB_CPU_PARTIAL
1933 struct kmem_cache_node *n = NULL, *n2 = NULL;
1934 struct page *page, *discard_page = NULL;
1936 while ((page = c->partial)) {
1940 c->partial = page->next;
1942 n2 = get_node(s, page_to_nid(page));
1945 spin_unlock(&n->list_lock);
1948 spin_lock(&n->list_lock);
1953 old.freelist = page->freelist;
1954 old.counters = page->counters;
1955 VM_BUG_ON(!old.frozen);
1957 new.counters = old.counters;
1958 new.freelist = old.freelist;
1962 } while (!__cmpxchg_double_slab(s, page,
1963 old.freelist, old.counters,
1964 new.freelist, new.counters,
1965 "unfreezing slab"));
1967 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1968 page->next = discard_page;
1969 discard_page = page;
1971 add_partial(n, page, DEACTIVATE_TO_TAIL);
1972 stat(s, FREE_ADD_PARTIAL);
1977 spin_unlock(&n->list_lock);
1979 while (discard_page) {
1980 page = discard_page;
1981 discard_page = discard_page->next;
1983 stat(s, DEACTIVATE_EMPTY);
1984 discard_slab(s, page);
1991 * Put a page that was just frozen (in __slab_free) into a partial page
1992 * slot if available. This is done without interrupts disabled and without
1993 * preemption disabled. The cmpxchg is racy and may put the partial page
1994 * onto a random cpus partial slot.
1996 * If we did not find a slot then simply move all the partials to the
1997 * per node partial list.
1999 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2001 #ifdef CONFIG_SLUB_CPU_PARTIAL
2002 struct page *oldpage;
2009 oldpage = this_cpu_read(s->cpu_slab->partial);
2012 pobjects = oldpage->pobjects;
2013 pages = oldpage->pages;
2014 if (drain && pobjects > s->cpu_partial) {
2015 unsigned long flags;
2017 * partial array is full. Move the existing
2018 * set to the per node partial list.
2020 local_irq_save(flags);
2021 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2022 local_irq_restore(flags);
2026 stat(s, CPU_PARTIAL_DRAIN);
2031 pobjects += page->objects - page->inuse;
2033 page->pages = pages;
2034 page->pobjects = pobjects;
2035 page->next = oldpage;
2037 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2042 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2044 stat(s, CPUSLAB_FLUSH);
2045 deactivate_slab(s, c->page, c->freelist);
2047 c->tid = next_tid(c->tid);
2055 * Called from IPI handler with interrupts disabled.
2057 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2059 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2065 unfreeze_partials(s, c);
2069 static void flush_cpu_slab(void *d)
2071 struct kmem_cache *s = d;
2073 __flush_cpu_slab(s, smp_processor_id());
2076 static bool has_cpu_slab(int cpu, void *info)
2078 struct kmem_cache *s = info;
2079 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2081 return c->page || c->partial;
2084 static void flush_all(struct kmem_cache *s)
2086 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2090 * Check if the objects in a per cpu structure fit numa
2091 * locality expectations.
2093 static inline int node_match(struct page *page, int node)
2096 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2102 #ifdef CONFIG_SLUB_DEBUG
2103 static int count_free(struct page *page)
2105 return page->objects - page->inuse;
2108 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2110 return atomic_long_read(&n->total_objects);
2112 #endif /* CONFIG_SLUB_DEBUG */
2114 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2115 static unsigned long count_partial(struct kmem_cache_node *n,
2116 int (*get_count)(struct page *))
2118 unsigned long flags;
2119 unsigned long x = 0;
2122 spin_lock_irqsave(&n->list_lock, flags);
2123 list_for_each_entry(page, &n->partial, lru)
2124 x += get_count(page);
2125 spin_unlock_irqrestore(&n->list_lock, flags);
2128 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2130 static noinline void
2131 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2133 #ifdef CONFIG_SLUB_DEBUG
2134 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2135 DEFAULT_RATELIMIT_BURST);
2137 struct kmem_cache_node *n;
2139 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2142 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2144 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2145 s->name, s->object_size, s->size, oo_order(s->oo),
2148 if (oo_order(s->min) > get_order(s->object_size))
2149 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2152 for_each_kmem_cache_node(s, node, n) {
2153 unsigned long nr_slabs;
2154 unsigned long nr_objs;
2155 unsigned long nr_free;
2157 nr_free = count_partial(n, count_free);
2158 nr_slabs = node_nr_slabs(n);
2159 nr_objs = node_nr_objs(n);
2161 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2162 node, nr_slabs, nr_objs, nr_free);
2167 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2168 int node, struct kmem_cache_cpu **pc)
2171 struct kmem_cache_cpu *c = *pc;
2174 freelist = get_partial(s, flags, node, c);
2179 page = new_slab(s, flags, node);
2181 c = raw_cpu_ptr(s->cpu_slab);
2186 * No other reference to the page yet so we can
2187 * muck around with it freely without cmpxchg
2189 freelist = page->freelist;
2190 page->freelist = NULL;
2192 stat(s, ALLOC_SLAB);
2201 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2203 if (unlikely(PageSlabPfmemalloc(page)))
2204 return gfp_pfmemalloc_allowed(gfpflags);
2210 * Check the page->freelist of a page and either transfer the freelist to the
2211 * per cpu freelist or deactivate the page.
2213 * The page is still frozen if the return value is not NULL.
2215 * If this function returns NULL then the page has been unfrozen.
2217 * This function must be called with interrupt disabled.
2219 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2222 unsigned long counters;
2226 freelist = page->freelist;
2227 counters = page->counters;
2229 new.counters = counters;
2230 VM_BUG_ON(!new.frozen);
2232 new.inuse = page->objects;
2233 new.frozen = freelist != NULL;
2235 } while (!__cmpxchg_double_slab(s, page,
2244 * Slow path. The lockless freelist is empty or we need to perform
2247 * Processing is still very fast if new objects have been freed to the
2248 * regular freelist. In that case we simply take over the regular freelist
2249 * as the lockless freelist and zap the regular freelist.
2251 * If that is not working then we fall back to the partial lists. We take the
2252 * first element of the freelist as the object to allocate now and move the
2253 * rest of the freelist to the lockless freelist.
2255 * And if we were unable to get a new slab from the partial slab lists then
2256 * we need to allocate a new slab. This is the slowest path since it involves
2257 * a call to the page allocator and the setup of a new slab.
2259 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2260 unsigned long addr, struct kmem_cache_cpu *c)
2264 unsigned long flags;
2266 local_irq_save(flags);
2267 #ifdef CONFIG_PREEMPT
2269 * We may have been preempted and rescheduled on a different
2270 * cpu before disabling interrupts. Need to reload cpu area
2273 c = this_cpu_ptr(s->cpu_slab);
2281 if (unlikely(!node_match(page, node))) {
2282 int searchnode = node;
2284 if (node != NUMA_NO_NODE && !node_present_pages(node))
2285 searchnode = node_to_mem_node(node);
2287 if (unlikely(!node_match(page, searchnode))) {
2288 stat(s, ALLOC_NODE_MISMATCH);
2289 deactivate_slab(s, page, c->freelist);
2297 * By rights, we should be searching for a slab page that was
2298 * PFMEMALLOC but right now, we are losing the pfmemalloc
2299 * information when the page leaves the per-cpu allocator
2301 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2302 deactivate_slab(s, page, c->freelist);
2308 /* must check again c->freelist in case of cpu migration or IRQ */
2309 freelist = c->freelist;
2313 freelist = get_freelist(s, page);
2317 stat(s, DEACTIVATE_BYPASS);
2321 stat(s, ALLOC_REFILL);
2325 * freelist is pointing to the list of objects to be used.
2326 * page is pointing to the page from which the objects are obtained.
2327 * That page must be frozen for per cpu allocations to work.
2329 VM_BUG_ON(!c->page->frozen);
2330 c->freelist = get_freepointer(s, freelist);
2331 c->tid = next_tid(c->tid);
2332 local_irq_restore(flags);
2338 page = c->page = c->partial;
2339 c->partial = page->next;
2340 stat(s, CPU_PARTIAL_ALLOC);
2345 freelist = new_slab_objects(s, gfpflags, node, &c);
2347 if (unlikely(!freelist)) {
2348 slab_out_of_memory(s, gfpflags, node);
2349 local_irq_restore(flags);
2354 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2357 /* Only entered in the debug case */
2358 if (kmem_cache_debug(s) &&
2359 !alloc_debug_processing(s, page, freelist, addr))
2360 goto new_slab; /* Slab failed checks. Next slab needed */
2362 deactivate_slab(s, page, get_freepointer(s, freelist));
2365 local_irq_restore(flags);
2370 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2371 * have the fastpath folded into their functions. So no function call
2372 * overhead for requests that can be satisfied on the fastpath.
2374 * The fastpath works by first checking if the lockless freelist can be used.
2375 * If not then __slab_alloc is called for slow processing.
2377 * Otherwise we can simply pick the next object from the lockless free list.
2379 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2380 gfp_t gfpflags, int node, unsigned long addr)
2383 struct kmem_cache_cpu *c;
2387 if (slab_pre_alloc_hook(s, gfpflags))
2390 s = memcg_kmem_get_cache(s, gfpflags);
2393 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2394 * enabled. We may switch back and forth between cpus while
2395 * reading from one cpu area. That does not matter as long
2396 * as we end up on the original cpu again when doing the cmpxchg.
2398 * Preemption is disabled for the retrieval of the tid because that
2399 * must occur from the current processor. We cannot allow rescheduling
2400 * on a different processor between the determination of the pointer
2401 * and the retrieval of the tid.
2404 c = this_cpu_ptr(s->cpu_slab);
2407 * The transaction ids are globally unique per cpu and per operation on
2408 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2409 * occurs on the right processor and that there was no operation on the
2410 * linked list in between.
2415 object = c->freelist;
2417 if (unlikely(!object || !node_match(page, node))) {
2418 object = __slab_alloc(s, gfpflags, node, addr, c);
2419 stat(s, ALLOC_SLOWPATH);
2421 void *next_object = get_freepointer_safe(s, object);
2424 * The cmpxchg will only match if there was no additional
2425 * operation and if we are on the right processor.
2427 * The cmpxchg does the following atomically (without lock
2429 * 1. Relocate first pointer to the current per cpu area.
2430 * 2. Verify that tid and freelist have not been changed
2431 * 3. If they were not changed replace tid and freelist
2433 * Since this is without lock semantics the protection is only
2434 * against code executing on this cpu *not* from access by
2437 if (unlikely(!this_cpu_cmpxchg_double(
2438 s->cpu_slab->freelist, s->cpu_slab->tid,
2440 next_object, next_tid(tid)))) {
2442 note_cmpxchg_failure("slab_alloc", s, tid);
2445 prefetch_freepointer(s, next_object);
2446 stat(s, ALLOC_FASTPATH);
2449 if (unlikely(gfpflags & __GFP_ZERO) && object)
2450 memset(object, 0, s->object_size);
2452 slab_post_alloc_hook(s, gfpflags, object);
2457 static __always_inline void *slab_alloc(struct kmem_cache *s,
2458 gfp_t gfpflags, unsigned long addr)
2460 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2463 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2465 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2467 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2472 EXPORT_SYMBOL(kmem_cache_alloc);
2474 #ifdef CONFIG_TRACING
2475 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2477 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2478 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2481 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2485 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2487 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2489 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2490 s->object_size, s->size, gfpflags, node);
2494 EXPORT_SYMBOL(kmem_cache_alloc_node);
2496 #ifdef CONFIG_TRACING
2497 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2499 int node, size_t size)
2501 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2503 trace_kmalloc_node(_RET_IP_, ret,
2504 size, s->size, gfpflags, node);
2507 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2512 * Slow patch handling. This may still be called frequently since objects
2513 * have a longer lifetime than the cpu slabs in most processing loads.
2515 * So we still attempt to reduce cache line usage. Just take the slab
2516 * lock and free the item. If there is no additional partial page
2517 * handling required then we can return immediately.
2519 static void __slab_free(struct kmem_cache *s, struct page *page,
2520 void *x, unsigned long addr)
2523 void **object = (void *)x;
2526 unsigned long counters;
2527 struct kmem_cache_node *n = NULL;
2528 unsigned long uninitialized_var(flags);
2530 stat(s, FREE_SLOWPATH);
2532 if (kmem_cache_debug(s) &&
2533 !(n = free_debug_processing(s, page, x, addr, &flags)))
2538 spin_unlock_irqrestore(&n->list_lock, flags);
2541 prior = page->freelist;
2542 counters = page->counters;
2543 set_freepointer(s, object, prior);
2544 new.counters = counters;
2545 was_frozen = new.frozen;
2547 if ((!new.inuse || !prior) && !was_frozen) {
2549 if (kmem_cache_has_cpu_partial(s) && !prior) {
2552 * Slab was on no list before and will be
2554 * We can defer the list move and instead
2559 } else { /* Needs to be taken off a list */
2561 n = get_node(s, page_to_nid(page));
2563 * Speculatively acquire the list_lock.
2564 * If the cmpxchg does not succeed then we may
2565 * drop the list_lock without any processing.
2567 * Otherwise the list_lock will synchronize with
2568 * other processors updating the list of slabs.
2570 spin_lock_irqsave(&n->list_lock, flags);
2575 } while (!cmpxchg_double_slab(s, page,
2577 object, new.counters,
2583 * If we just froze the page then put it onto the
2584 * per cpu partial list.
2586 if (new.frozen && !was_frozen) {
2587 put_cpu_partial(s, page, 1);
2588 stat(s, CPU_PARTIAL_FREE);
2591 * The list lock was not taken therefore no list
2592 * activity can be necessary.
2595 stat(s, FREE_FROZEN);
2599 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2603 * Objects left in the slab. If it was not on the partial list before
2606 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2607 if (kmem_cache_debug(s))
2608 remove_full(s, n, page);
2609 add_partial(n, page, DEACTIVATE_TO_TAIL);
2610 stat(s, FREE_ADD_PARTIAL);
2612 spin_unlock_irqrestore(&n->list_lock, flags);
2618 * Slab on the partial list.
2620 remove_partial(n, page);
2621 stat(s, FREE_REMOVE_PARTIAL);
2623 /* Slab must be on the full list */
2624 remove_full(s, n, page);
2627 spin_unlock_irqrestore(&n->list_lock, flags);
2629 discard_slab(s, page);
2633 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2634 * can perform fastpath freeing without additional function calls.
2636 * The fastpath is only possible if we are freeing to the current cpu slab
2637 * of this processor. This typically the case if we have just allocated
2640 * If fastpath is not possible then fall back to __slab_free where we deal
2641 * with all sorts of special processing.
2643 static __always_inline void slab_free(struct kmem_cache *s,
2644 struct page *page, void *x, unsigned long addr)
2646 void **object = (void *)x;
2647 struct kmem_cache_cpu *c;
2650 slab_free_hook(s, x);
2654 * Determine the currently cpus per cpu slab.
2655 * The cpu may change afterward. However that does not matter since
2656 * data is retrieved via this pointer. If we are on the same cpu
2657 * during the cmpxchg then the free will succedd.
2660 c = this_cpu_ptr(s->cpu_slab);
2665 if (likely(page == c->page)) {
2666 set_freepointer(s, object, c->freelist);
2668 if (unlikely(!this_cpu_cmpxchg_double(
2669 s->cpu_slab->freelist, s->cpu_slab->tid,
2671 object, next_tid(tid)))) {
2673 note_cmpxchg_failure("slab_free", s, tid);
2676 stat(s, FREE_FASTPATH);
2678 __slab_free(s, page, x, addr);
2682 void kmem_cache_free(struct kmem_cache *s, void *x)
2684 s = cache_from_obj(s, x);
2687 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2688 trace_kmem_cache_free(_RET_IP_, x);
2690 EXPORT_SYMBOL(kmem_cache_free);
2693 * Object placement in a slab is made very easy because we always start at
2694 * offset 0. If we tune the size of the object to the alignment then we can
2695 * get the required alignment by putting one properly sized object after
2698 * Notice that the allocation order determines the sizes of the per cpu
2699 * caches. Each processor has always one slab available for allocations.
2700 * Increasing the allocation order reduces the number of times that slabs
2701 * must be moved on and off the partial lists and is therefore a factor in
2706 * Mininum / Maximum order of slab pages. This influences locking overhead
2707 * and slab fragmentation. A higher order reduces the number of partial slabs
2708 * and increases the number of allocations possible without having to
2709 * take the list_lock.
2711 static int slub_min_order;
2712 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2713 static int slub_min_objects;
2716 * Calculate the order of allocation given an slab object size.
2718 * The order of allocation has significant impact on performance and other
2719 * system components. Generally order 0 allocations should be preferred since
2720 * order 0 does not cause fragmentation in the page allocator. Larger objects
2721 * be problematic to put into order 0 slabs because there may be too much
2722 * unused space left. We go to a higher order if more than 1/16th of the slab
2725 * In order to reach satisfactory performance we must ensure that a minimum
2726 * number of objects is in one slab. Otherwise we may generate too much
2727 * activity on the partial lists which requires taking the list_lock. This is
2728 * less a concern for large slabs though which are rarely used.
2730 * slub_max_order specifies the order where we begin to stop considering the
2731 * number of objects in a slab as critical. If we reach slub_max_order then
2732 * we try to keep the page order as low as possible. So we accept more waste
2733 * of space in favor of a small page order.
2735 * Higher order allocations also allow the placement of more objects in a
2736 * slab and thereby reduce object handling overhead. If the user has
2737 * requested a higher mininum order then we start with that one instead of
2738 * the smallest order which will fit the object.
2740 static inline int slab_order(int size, int min_objects,
2741 int max_order, int fract_leftover, int reserved)
2745 int min_order = slub_min_order;
2747 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2748 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2750 for (order = max(min_order,
2751 fls(min_objects * size - 1) - PAGE_SHIFT);
2752 order <= max_order; order++) {
2754 unsigned long slab_size = PAGE_SIZE << order;
2756 if (slab_size < min_objects * size + reserved)
2759 rem = (slab_size - reserved) % size;
2761 if (rem <= slab_size / fract_leftover)
2769 static inline int calculate_order(int size, int reserved)
2777 * Attempt to find best configuration for a slab. This
2778 * works by first attempting to generate a layout with
2779 * the best configuration and backing off gradually.
2781 * First we reduce the acceptable waste in a slab. Then
2782 * we reduce the minimum objects required in a slab.
2784 min_objects = slub_min_objects;
2786 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2787 max_objects = order_objects(slub_max_order, size, reserved);
2788 min_objects = min(min_objects, max_objects);
2790 while (min_objects > 1) {
2792 while (fraction >= 4) {
2793 order = slab_order(size, min_objects,
2794 slub_max_order, fraction, reserved);
2795 if (order <= slub_max_order)
2803 * We were unable to place multiple objects in a slab. Now
2804 * lets see if we can place a single object there.
2806 order = slab_order(size, 1, slub_max_order, 1, reserved);
2807 if (order <= slub_max_order)
2811 * Doh this slab cannot be placed using slub_max_order.
2813 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2814 if (order < MAX_ORDER)
2820 init_kmem_cache_node(struct kmem_cache_node *n)
2823 spin_lock_init(&n->list_lock);
2824 INIT_LIST_HEAD(&n->partial);
2825 #ifdef CONFIG_SLUB_DEBUG
2826 atomic_long_set(&n->nr_slabs, 0);
2827 atomic_long_set(&n->total_objects, 0);
2828 INIT_LIST_HEAD(&n->full);
2832 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2834 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2835 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2838 * Must align to double word boundary for the double cmpxchg
2839 * instructions to work; see __pcpu_double_call_return_bool().
2841 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2842 2 * sizeof(void *));
2847 init_kmem_cache_cpus(s);
2852 static struct kmem_cache *kmem_cache_node;
2855 * No kmalloc_node yet so do it by hand. We know that this is the first
2856 * slab on the node for this slabcache. There are no concurrent accesses
2859 * Note that this function only works on the kmem_cache_node
2860 * when allocating for the kmem_cache_node. This is used for bootstrapping
2861 * memory on a fresh node that has no slab structures yet.
2863 static void early_kmem_cache_node_alloc(int node)
2866 struct kmem_cache_node *n;
2868 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2870 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2873 if (page_to_nid(page) != node) {
2874 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2875 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2880 page->freelist = get_freepointer(kmem_cache_node, n);
2883 kmem_cache_node->node[node] = n;
2884 #ifdef CONFIG_SLUB_DEBUG
2885 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2886 init_tracking(kmem_cache_node, n);
2888 init_kmem_cache_node(n);
2889 inc_slabs_node(kmem_cache_node, node, page->objects);
2892 * No locks need to be taken here as it has just been
2893 * initialized and there is no concurrent access.
2895 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2898 static void free_kmem_cache_nodes(struct kmem_cache *s)
2901 struct kmem_cache_node *n;
2903 for_each_kmem_cache_node(s, node, n) {
2904 kmem_cache_free(kmem_cache_node, n);
2905 s->node[node] = NULL;
2909 static int init_kmem_cache_nodes(struct kmem_cache *s)
2913 for_each_node_state(node, N_NORMAL_MEMORY) {
2914 struct kmem_cache_node *n;
2916 if (slab_state == DOWN) {
2917 early_kmem_cache_node_alloc(node);
2920 n = kmem_cache_alloc_node(kmem_cache_node,
2924 free_kmem_cache_nodes(s);
2929 init_kmem_cache_node(n);
2934 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2936 if (min < MIN_PARTIAL)
2938 else if (min > MAX_PARTIAL)
2940 s->min_partial = min;
2944 * calculate_sizes() determines the order and the distribution of data within
2947 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2949 unsigned long flags = s->flags;
2950 unsigned long size = s->object_size;
2954 * Round up object size to the next word boundary. We can only
2955 * place the free pointer at word boundaries and this determines
2956 * the possible location of the free pointer.
2958 size = ALIGN(size, sizeof(void *));
2960 #ifdef CONFIG_SLUB_DEBUG
2962 * Determine if we can poison the object itself. If the user of
2963 * the slab may touch the object after free or before allocation
2964 * then we should never poison the object itself.
2966 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2968 s->flags |= __OBJECT_POISON;
2970 s->flags &= ~__OBJECT_POISON;
2974 * If we are Redzoning then check if there is some space between the
2975 * end of the object and the free pointer. If not then add an
2976 * additional word to have some bytes to store Redzone information.
2978 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2979 size += sizeof(void *);
2983 * With that we have determined the number of bytes in actual use
2984 * by the object. This is the potential offset to the free pointer.
2988 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2991 * Relocate free pointer after the object if it is not
2992 * permitted to overwrite the first word of the object on
2995 * This is the case if we do RCU, have a constructor or
2996 * destructor or are poisoning the objects.
2999 size += sizeof(void *);
3002 #ifdef CONFIG_SLUB_DEBUG
3003 if (flags & SLAB_STORE_USER)
3005 * Need to store information about allocs and frees after
3008 size += 2 * sizeof(struct track);
3010 if (flags & SLAB_RED_ZONE)
3012 * Add some empty padding so that we can catch
3013 * overwrites from earlier objects rather than let
3014 * tracking information or the free pointer be
3015 * corrupted if a user writes before the start
3018 size += sizeof(void *);
3022 * SLUB stores one object immediately after another beginning from
3023 * offset 0. In order to align the objects we have to simply size
3024 * each object to conform to the alignment.
3026 size = ALIGN(size, s->align);
3028 if (forced_order >= 0)
3029 order = forced_order;
3031 order = calculate_order(size, s->reserved);
3038 s->allocflags |= __GFP_COMP;
3040 if (s->flags & SLAB_CACHE_DMA)
3041 s->allocflags |= GFP_DMA;
3043 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3044 s->allocflags |= __GFP_RECLAIMABLE;
3047 * Determine the number of objects per slab
3049 s->oo = oo_make(order, size, s->reserved);
3050 s->min = oo_make(get_order(size), size, s->reserved);
3051 if (oo_objects(s->oo) > oo_objects(s->max))
3054 return !!oo_objects(s->oo);
3057 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3059 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3062 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3063 s->reserved = sizeof(struct rcu_head);
3065 if (!calculate_sizes(s, -1))
3067 if (disable_higher_order_debug) {
3069 * Disable debugging flags that store metadata if the min slab
3072 if (get_order(s->size) > get_order(s->object_size)) {
3073 s->flags &= ~DEBUG_METADATA_FLAGS;
3075 if (!calculate_sizes(s, -1))
3080 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3081 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3082 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3083 /* Enable fast mode */
3084 s->flags |= __CMPXCHG_DOUBLE;
3088 * The larger the object size is, the more pages we want on the partial
3089 * list to avoid pounding the page allocator excessively.
3091 set_min_partial(s, ilog2(s->size) / 2);
3094 * cpu_partial determined the maximum number of objects kept in the
3095 * per cpu partial lists of a processor.
3097 * Per cpu partial lists mainly contain slabs that just have one
3098 * object freed. If they are used for allocation then they can be
3099 * filled up again with minimal effort. The slab will never hit the
3100 * per node partial lists and therefore no locking will be required.
3102 * This setting also determines
3104 * A) The number of objects from per cpu partial slabs dumped to the
3105 * per node list when we reach the limit.
3106 * B) The number of objects in cpu partial slabs to extract from the
3107 * per node list when we run out of per cpu objects. We only fetch
3108 * 50% to keep some capacity around for frees.
3110 if (!kmem_cache_has_cpu_partial(s))
3112 else if (s->size >= PAGE_SIZE)
3114 else if (s->size >= 1024)
3116 else if (s->size >= 256)
3117 s->cpu_partial = 13;
3119 s->cpu_partial = 30;
3122 s->remote_node_defrag_ratio = 1000;
3124 if (!init_kmem_cache_nodes(s))
3127 if (alloc_kmem_cache_cpus(s))
3130 free_kmem_cache_nodes(s);
3132 if (flags & SLAB_PANIC)
3133 panic("Cannot create slab %s size=%lu realsize=%u "
3134 "order=%u offset=%u flags=%lx\n",
3135 s->name, (unsigned long)s->size, s->size,
3136 oo_order(s->oo), s->offset, flags);
3140 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3143 #ifdef CONFIG_SLUB_DEBUG
3144 void *addr = page_address(page);
3146 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3147 sizeof(long), GFP_ATOMIC);
3150 slab_err(s, page, text, s->name);
3153 get_map(s, page, map);
3154 for_each_object(p, s, addr, page->objects) {
3156 if (!test_bit(slab_index(p, s, addr), map)) {
3157 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3158 print_tracking(s, p);
3167 * Attempt to free all partial slabs on a node.
3168 * This is called from kmem_cache_close(). We must be the last thread
3169 * using the cache and therefore we do not need to lock anymore.
3171 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3173 struct page *page, *h;
3175 list_for_each_entry_safe(page, h, &n->partial, lru) {
3177 __remove_partial(n, page);
3178 discard_slab(s, page);
3180 list_slab_objects(s, page,
3181 "Objects remaining in %s on kmem_cache_close()");
3187 * Release all resources used by a slab cache.
3189 static inline int kmem_cache_close(struct kmem_cache *s)
3192 struct kmem_cache_node *n;
3195 /* Attempt to free all objects */
3196 for_each_kmem_cache_node(s, node, n) {
3198 if (n->nr_partial || slabs_node(s, node))
3201 free_percpu(s->cpu_slab);
3202 free_kmem_cache_nodes(s);
3206 int __kmem_cache_shutdown(struct kmem_cache *s)
3208 return kmem_cache_close(s);
3211 /********************************************************************
3213 *******************************************************************/
3215 static int __init setup_slub_min_order(char *str)
3217 get_option(&str, &slub_min_order);
3222 __setup("slub_min_order=", setup_slub_min_order);
3224 static int __init setup_slub_max_order(char *str)
3226 get_option(&str, &slub_max_order);
3227 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3232 __setup("slub_max_order=", setup_slub_max_order);
3234 static int __init setup_slub_min_objects(char *str)
3236 get_option(&str, &slub_min_objects);
3241 __setup("slub_min_objects=", setup_slub_min_objects);
3243 void *__kmalloc(size_t size, gfp_t flags)
3245 struct kmem_cache *s;
3248 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3249 return kmalloc_large(size, flags);
3251 s = kmalloc_slab(size, flags);
3253 if (unlikely(ZERO_OR_NULL_PTR(s)))
3256 ret = slab_alloc(s, flags, _RET_IP_);
3258 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3262 EXPORT_SYMBOL(__kmalloc);
3265 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3270 flags |= __GFP_COMP | __GFP_NOTRACK;
3271 page = alloc_kmem_pages_node(node, flags, get_order(size));
3273 ptr = page_address(page);
3275 kmalloc_large_node_hook(ptr, size, flags);
3279 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3281 struct kmem_cache *s;
3284 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3285 ret = kmalloc_large_node(size, flags, node);
3287 trace_kmalloc_node(_RET_IP_, ret,
3288 size, PAGE_SIZE << get_order(size),
3294 s = kmalloc_slab(size, flags);
3296 if (unlikely(ZERO_OR_NULL_PTR(s)))
3299 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3301 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3305 EXPORT_SYMBOL(__kmalloc_node);
3308 size_t ksize(const void *object)
3312 if (unlikely(object == ZERO_SIZE_PTR))
3315 page = virt_to_head_page(object);
3317 if (unlikely(!PageSlab(page))) {
3318 WARN_ON(!PageCompound(page));
3319 return PAGE_SIZE << compound_order(page);
3322 return slab_ksize(page->slab_cache);
3324 EXPORT_SYMBOL(ksize);
3326 void kfree(const void *x)
3329 void *object = (void *)x;
3331 trace_kfree(_RET_IP_, x);
3333 if (unlikely(ZERO_OR_NULL_PTR(x)))
3336 page = virt_to_head_page(x);
3337 if (unlikely(!PageSlab(page))) {
3338 BUG_ON(!PageCompound(page));
3340 __free_kmem_pages(page, compound_order(page));
3343 slab_free(page->slab_cache, page, object, _RET_IP_);
3345 EXPORT_SYMBOL(kfree);
3348 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3349 * the remaining slabs by the number of items in use. The slabs with the
3350 * most items in use come first. New allocations will then fill those up
3351 * and thus they can be removed from the partial lists.
3353 * The slabs with the least items are placed last. This results in them
3354 * being allocated from last increasing the chance that the last objects
3355 * are freed in them.
3357 int __kmem_cache_shrink(struct kmem_cache *s)
3361 struct kmem_cache_node *n;
3364 int objects = oo_objects(s->max);
3365 struct list_head *slabs_by_inuse =
3366 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3367 unsigned long flags;
3369 if (!slabs_by_inuse)
3373 for_each_kmem_cache_node(s, node, n) {
3377 for (i = 0; i < objects; i++)
3378 INIT_LIST_HEAD(slabs_by_inuse + i);
3380 spin_lock_irqsave(&n->list_lock, flags);
3383 * Build lists indexed by the items in use in each slab.
3385 * Note that concurrent frees may occur while we hold the
3386 * list_lock. page->inuse here is the upper limit.
3388 list_for_each_entry_safe(page, t, &n->partial, lru) {
3389 list_move(&page->lru, slabs_by_inuse + page->inuse);
3395 * Rebuild the partial list with the slabs filled up most
3396 * first and the least used slabs at the end.
3398 for (i = objects - 1; i > 0; i--)
3399 list_splice(slabs_by_inuse + i, n->partial.prev);
3401 spin_unlock_irqrestore(&n->list_lock, flags);
3403 /* Release empty slabs */
3404 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3405 discard_slab(s, page);
3408 kfree(slabs_by_inuse);
3412 static int slab_mem_going_offline_callback(void *arg)
3414 struct kmem_cache *s;
3416 mutex_lock(&slab_mutex);
3417 list_for_each_entry(s, &slab_caches, list)
3418 __kmem_cache_shrink(s);
3419 mutex_unlock(&slab_mutex);
3424 static void slab_mem_offline_callback(void *arg)
3426 struct kmem_cache_node *n;
3427 struct kmem_cache *s;
3428 struct memory_notify *marg = arg;
3431 offline_node = marg->status_change_nid_normal;
3434 * If the node still has available memory. we need kmem_cache_node
3437 if (offline_node < 0)
3440 mutex_lock(&slab_mutex);
3441 list_for_each_entry(s, &slab_caches, list) {
3442 n = get_node(s, offline_node);
3445 * if n->nr_slabs > 0, slabs still exist on the node
3446 * that is going down. We were unable to free them,
3447 * and offline_pages() function shouldn't call this
3448 * callback. So, we must fail.
3450 BUG_ON(slabs_node(s, offline_node));
3452 s->node[offline_node] = NULL;
3453 kmem_cache_free(kmem_cache_node, n);
3456 mutex_unlock(&slab_mutex);
3459 static int slab_mem_going_online_callback(void *arg)
3461 struct kmem_cache_node *n;
3462 struct kmem_cache *s;
3463 struct memory_notify *marg = arg;
3464 int nid = marg->status_change_nid_normal;
3468 * If the node's memory is already available, then kmem_cache_node is
3469 * already created. Nothing to do.
3475 * We are bringing a node online. No memory is available yet. We must
3476 * allocate a kmem_cache_node structure in order to bring the node
3479 mutex_lock(&slab_mutex);
3480 list_for_each_entry(s, &slab_caches, list) {
3482 * XXX: kmem_cache_alloc_node will fallback to other nodes
3483 * since memory is not yet available from the node that
3486 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3491 init_kmem_cache_node(n);
3495 mutex_unlock(&slab_mutex);
3499 static int slab_memory_callback(struct notifier_block *self,
3500 unsigned long action, void *arg)
3505 case MEM_GOING_ONLINE:
3506 ret = slab_mem_going_online_callback(arg);
3508 case MEM_GOING_OFFLINE:
3509 ret = slab_mem_going_offline_callback(arg);
3512 case MEM_CANCEL_ONLINE:
3513 slab_mem_offline_callback(arg);
3516 case MEM_CANCEL_OFFLINE:
3520 ret = notifier_from_errno(ret);
3526 static struct notifier_block slab_memory_callback_nb = {
3527 .notifier_call = slab_memory_callback,
3528 .priority = SLAB_CALLBACK_PRI,
3531 /********************************************************************
3532 * Basic setup of slabs
3533 *******************************************************************/
3536 * Used for early kmem_cache structures that were allocated using
3537 * the page allocator. Allocate them properly then fix up the pointers
3538 * that may be pointing to the wrong kmem_cache structure.
3541 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3544 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3545 struct kmem_cache_node *n;
3547 memcpy(s, static_cache, kmem_cache->object_size);
3550 * This runs very early, and only the boot processor is supposed to be
3551 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3554 __flush_cpu_slab(s, smp_processor_id());
3555 for_each_kmem_cache_node(s, node, n) {
3558 list_for_each_entry(p, &n->partial, lru)
3561 #ifdef CONFIG_SLUB_DEBUG
3562 list_for_each_entry(p, &n->full, lru)
3566 list_add(&s->list, &slab_caches);
3570 void __init kmem_cache_init(void)
3572 static __initdata struct kmem_cache boot_kmem_cache,
3573 boot_kmem_cache_node;
3575 if (debug_guardpage_minorder())
3578 kmem_cache_node = &boot_kmem_cache_node;
3579 kmem_cache = &boot_kmem_cache;
3581 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3582 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3584 register_hotmemory_notifier(&slab_memory_callback_nb);
3586 /* Able to allocate the per node structures */
3587 slab_state = PARTIAL;
3589 create_boot_cache(kmem_cache, "kmem_cache",
3590 offsetof(struct kmem_cache, node) +
3591 nr_node_ids * sizeof(struct kmem_cache_node *),
3592 SLAB_HWCACHE_ALIGN);
3594 kmem_cache = bootstrap(&boot_kmem_cache);
3597 * Allocate kmem_cache_node properly from the kmem_cache slab.
3598 * kmem_cache_node is separately allocated so no need to
3599 * update any list pointers.
3601 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3603 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3604 create_kmalloc_caches(0);
3607 register_cpu_notifier(&slab_notifier);
3610 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3612 slub_min_order, slub_max_order, slub_min_objects,
3613 nr_cpu_ids, nr_node_ids);
3616 void __init kmem_cache_init_late(void)
3621 __kmem_cache_alias(const char *name, size_t size, size_t align,
3622 unsigned long flags, void (*ctor)(void *))
3624 struct kmem_cache *s;
3626 s = find_mergeable(size, align, flags, name, ctor);
3629 struct kmem_cache *c;
3634 * Adjust the object sizes so that we clear
3635 * the complete object on kzalloc.
3637 s->object_size = max(s->object_size, (int)size);
3638 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3640 for_each_memcg_cache_index(i) {
3641 c = cache_from_memcg_idx(s, i);
3644 c->object_size = s->object_size;
3645 c->inuse = max_t(int, c->inuse,
3646 ALIGN(size, sizeof(void *)));
3649 if (sysfs_slab_alias(s, name)) {
3658 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3662 err = kmem_cache_open(s, flags);
3666 /* Mutex is not taken during early boot */
3667 if (slab_state <= UP)
3670 memcg_propagate_slab_attrs(s);
3671 err = sysfs_slab_add(s);
3673 kmem_cache_close(s);
3680 * Use the cpu notifier to insure that the cpu slabs are flushed when
3683 static int slab_cpuup_callback(struct notifier_block *nfb,
3684 unsigned long action, void *hcpu)
3686 long cpu = (long)hcpu;
3687 struct kmem_cache *s;
3688 unsigned long flags;
3691 case CPU_UP_CANCELED:
3692 case CPU_UP_CANCELED_FROZEN:
3694 case CPU_DEAD_FROZEN:
3695 mutex_lock(&slab_mutex);
3696 list_for_each_entry(s, &slab_caches, list) {
3697 local_irq_save(flags);
3698 __flush_cpu_slab(s, cpu);
3699 local_irq_restore(flags);
3701 mutex_unlock(&slab_mutex);
3709 static struct notifier_block slab_notifier = {
3710 .notifier_call = slab_cpuup_callback
3715 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3717 struct kmem_cache *s;
3720 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3721 return kmalloc_large(size, gfpflags);
3723 s = kmalloc_slab(size, gfpflags);
3725 if (unlikely(ZERO_OR_NULL_PTR(s)))
3728 ret = slab_alloc(s, gfpflags, caller);
3730 /* Honor the call site pointer we received. */
3731 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3737 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3738 int node, unsigned long caller)
3740 struct kmem_cache *s;
3743 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3744 ret = kmalloc_large_node(size, gfpflags, node);
3746 trace_kmalloc_node(caller, ret,
3747 size, PAGE_SIZE << get_order(size),
3753 s = kmalloc_slab(size, gfpflags);
3755 if (unlikely(ZERO_OR_NULL_PTR(s)))
3758 ret = slab_alloc_node(s, gfpflags, node, caller);
3760 /* Honor the call site pointer we received. */
3761 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3768 static int count_inuse(struct page *page)
3773 static int count_total(struct page *page)
3775 return page->objects;
3779 #ifdef CONFIG_SLUB_DEBUG
3780 static int validate_slab(struct kmem_cache *s, struct page *page,
3784 void *addr = page_address(page);
3786 if (!check_slab(s, page) ||
3787 !on_freelist(s, page, NULL))
3790 /* Now we know that a valid freelist exists */
3791 bitmap_zero(map, page->objects);
3793 get_map(s, page, map);
3794 for_each_object(p, s, addr, page->objects) {
3795 if (test_bit(slab_index(p, s, addr), map))
3796 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3800 for_each_object(p, s, addr, page->objects)
3801 if (!test_bit(slab_index(p, s, addr), map))
3802 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3807 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3811 validate_slab(s, page, map);
3815 static int validate_slab_node(struct kmem_cache *s,
3816 struct kmem_cache_node *n, unsigned long *map)
3818 unsigned long count = 0;
3820 unsigned long flags;
3822 spin_lock_irqsave(&n->list_lock, flags);
3824 list_for_each_entry(page, &n->partial, lru) {
3825 validate_slab_slab(s, page, map);
3828 if (count != n->nr_partial)
3829 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3830 s->name, count, n->nr_partial);
3832 if (!(s->flags & SLAB_STORE_USER))
3835 list_for_each_entry(page, &n->full, lru) {
3836 validate_slab_slab(s, page, map);
3839 if (count != atomic_long_read(&n->nr_slabs))
3840 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3841 s->name, count, atomic_long_read(&n->nr_slabs));
3844 spin_unlock_irqrestore(&n->list_lock, flags);
3848 static long validate_slab_cache(struct kmem_cache *s)
3851 unsigned long count = 0;
3852 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3853 sizeof(unsigned long), GFP_KERNEL);
3854 struct kmem_cache_node *n;
3860 for_each_kmem_cache_node(s, node, n)
3861 count += validate_slab_node(s, n, map);
3866 * Generate lists of code addresses where slabcache objects are allocated
3871 unsigned long count;
3878 DECLARE_BITMAP(cpus, NR_CPUS);
3884 unsigned long count;
3885 struct location *loc;
3888 static void free_loc_track(struct loc_track *t)
3891 free_pages((unsigned long)t->loc,
3892 get_order(sizeof(struct location) * t->max));
3895 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3900 order = get_order(sizeof(struct location) * max);
3902 l = (void *)__get_free_pages(flags, order);
3907 memcpy(l, t->loc, sizeof(struct location) * t->count);
3915 static int add_location(struct loc_track *t, struct kmem_cache *s,
3916 const struct track *track)
3918 long start, end, pos;
3920 unsigned long caddr;
3921 unsigned long age = jiffies - track->when;
3927 pos = start + (end - start + 1) / 2;
3930 * There is nothing at "end". If we end up there
3931 * we need to add something to before end.
3936 caddr = t->loc[pos].addr;
3937 if (track->addr == caddr) {
3943 if (age < l->min_time)
3945 if (age > l->max_time)
3948 if (track->pid < l->min_pid)
3949 l->min_pid = track->pid;
3950 if (track->pid > l->max_pid)
3951 l->max_pid = track->pid;
3953 cpumask_set_cpu(track->cpu,
3954 to_cpumask(l->cpus));
3956 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3960 if (track->addr < caddr)
3967 * Not found. Insert new tracking element.
3969 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3975 (t->count - pos) * sizeof(struct location));
3978 l->addr = track->addr;
3982 l->min_pid = track->pid;
3983 l->max_pid = track->pid;
3984 cpumask_clear(to_cpumask(l->cpus));
3985 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3986 nodes_clear(l->nodes);
3987 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3991 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3992 struct page *page, enum track_item alloc,
3995 void *addr = page_address(page);
3998 bitmap_zero(map, page->objects);
3999 get_map(s, page, map);
4001 for_each_object(p, s, addr, page->objects)
4002 if (!test_bit(slab_index(p, s, addr), map))
4003 add_location(t, s, get_track(s, p, alloc));
4006 static int list_locations(struct kmem_cache *s, char *buf,
4007 enum track_item alloc)
4011 struct loc_track t = { 0, 0, NULL };
4013 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4014 sizeof(unsigned long), GFP_KERNEL);
4015 struct kmem_cache_node *n;
4017 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4020 return sprintf(buf, "Out of memory\n");
4022 /* Push back cpu slabs */
4025 for_each_kmem_cache_node(s, node, n) {
4026 unsigned long flags;
4029 if (!atomic_long_read(&n->nr_slabs))
4032 spin_lock_irqsave(&n->list_lock, flags);
4033 list_for_each_entry(page, &n->partial, lru)
4034 process_slab(&t, s, page, alloc, map);
4035 list_for_each_entry(page, &n->full, lru)
4036 process_slab(&t, s, page, alloc, map);
4037 spin_unlock_irqrestore(&n->list_lock, flags);
4040 for (i = 0; i < t.count; i++) {
4041 struct location *l = &t.loc[i];
4043 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4045 len += sprintf(buf + len, "%7ld ", l->count);
4048 len += sprintf(buf + len, "%pS", (void *)l->addr);
4050 len += sprintf(buf + len, "<not-available>");
4052 if (l->sum_time != l->min_time) {
4053 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4055 (long)div_u64(l->sum_time, l->count),
4058 len += sprintf(buf + len, " age=%ld",
4061 if (l->min_pid != l->max_pid)
4062 len += sprintf(buf + len, " pid=%ld-%ld",
4063 l->min_pid, l->max_pid);
4065 len += sprintf(buf + len, " pid=%ld",
4068 if (num_online_cpus() > 1 &&
4069 !cpumask_empty(to_cpumask(l->cpus)) &&
4070 len < PAGE_SIZE - 60) {
4071 len += sprintf(buf + len, " cpus=");
4072 len += cpulist_scnprintf(buf + len,
4073 PAGE_SIZE - len - 50,
4074 to_cpumask(l->cpus));
4077 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4078 len < PAGE_SIZE - 60) {
4079 len += sprintf(buf + len, " nodes=");
4080 len += nodelist_scnprintf(buf + len,
4081 PAGE_SIZE - len - 50,
4085 len += sprintf(buf + len, "\n");
4091 len += sprintf(buf, "No data\n");
4096 #ifdef SLUB_RESILIENCY_TEST
4097 static void __init resiliency_test(void)
4101 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4103 pr_err("SLUB resiliency testing\n");
4104 pr_err("-----------------------\n");
4105 pr_err("A. Corruption after allocation\n");
4107 p = kzalloc(16, GFP_KERNEL);
4109 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4112 validate_slab_cache(kmalloc_caches[4]);
4114 /* Hmmm... The next two are dangerous */
4115 p = kzalloc(32, GFP_KERNEL);
4116 p[32 + sizeof(void *)] = 0x34;
4117 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4119 pr_err("If allocated object is overwritten then not detectable\n\n");
4121 validate_slab_cache(kmalloc_caches[5]);
4122 p = kzalloc(64, GFP_KERNEL);
4123 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4125 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4127 pr_err("If allocated object is overwritten then not detectable\n\n");
4128 validate_slab_cache(kmalloc_caches[6]);
4130 pr_err("\nB. Corruption after free\n");
4131 p = kzalloc(128, GFP_KERNEL);
4134 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4135 validate_slab_cache(kmalloc_caches[7]);
4137 p = kzalloc(256, GFP_KERNEL);
4140 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4141 validate_slab_cache(kmalloc_caches[8]);
4143 p = kzalloc(512, GFP_KERNEL);
4146 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4147 validate_slab_cache(kmalloc_caches[9]);
4151 static void resiliency_test(void) {};
4156 enum slab_stat_type {
4157 SL_ALL, /* All slabs */
4158 SL_PARTIAL, /* Only partially allocated slabs */
4159 SL_CPU, /* Only slabs used for cpu caches */
4160 SL_OBJECTS, /* Determine allocated objects not slabs */
4161 SL_TOTAL /* Determine object capacity not slabs */
4164 #define SO_ALL (1 << SL_ALL)
4165 #define SO_PARTIAL (1 << SL_PARTIAL)
4166 #define SO_CPU (1 << SL_CPU)
4167 #define SO_OBJECTS (1 << SL_OBJECTS)
4168 #define SO_TOTAL (1 << SL_TOTAL)
4170 static ssize_t show_slab_objects(struct kmem_cache *s,
4171 char *buf, unsigned long flags)
4173 unsigned long total = 0;
4176 unsigned long *nodes;
4178 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4182 if (flags & SO_CPU) {
4185 for_each_possible_cpu(cpu) {
4186 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4191 page = ACCESS_ONCE(c->page);
4195 node = page_to_nid(page);
4196 if (flags & SO_TOTAL)
4198 else if (flags & SO_OBJECTS)
4206 page = ACCESS_ONCE(c->partial);
4208 node = page_to_nid(page);
4209 if (flags & SO_TOTAL)
4211 else if (flags & SO_OBJECTS)
4222 #ifdef CONFIG_SLUB_DEBUG
4223 if (flags & SO_ALL) {
4224 struct kmem_cache_node *n;
4226 for_each_kmem_cache_node(s, node, n) {
4228 if (flags & SO_TOTAL)
4229 x = atomic_long_read(&n->total_objects);
4230 else if (flags & SO_OBJECTS)
4231 x = atomic_long_read(&n->total_objects) -
4232 count_partial(n, count_free);
4234 x = atomic_long_read(&n->nr_slabs);
4241 if (flags & SO_PARTIAL) {
4242 struct kmem_cache_node *n;
4244 for_each_kmem_cache_node(s, node, n) {
4245 if (flags & SO_TOTAL)
4246 x = count_partial(n, count_total);
4247 else if (flags & SO_OBJECTS)
4248 x = count_partial(n, count_inuse);
4255 x = sprintf(buf, "%lu", total);
4257 for (node = 0; node < nr_node_ids; node++)
4259 x += sprintf(buf + x, " N%d=%lu",
4264 return x + sprintf(buf + x, "\n");
4267 #ifdef CONFIG_SLUB_DEBUG
4268 static int any_slab_objects(struct kmem_cache *s)
4271 struct kmem_cache_node *n;
4273 for_each_kmem_cache_node(s, node, n)
4274 if (atomic_long_read(&n->total_objects))
4281 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4282 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4284 struct slab_attribute {
4285 struct attribute attr;
4286 ssize_t (*show)(struct kmem_cache *s, char *buf);
4287 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4290 #define SLAB_ATTR_RO(_name) \
4291 static struct slab_attribute _name##_attr = \
4292 __ATTR(_name, 0400, _name##_show, NULL)
4294 #define SLAB_ATTR(_name) \
4295 static struct slab_attribute _name##_attr = \
4296 __ATTR(_name, 0600, _name##_show, _name##_store)
4298 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4300 return sprintf(buf, "%d\n", s->size);
4302 SLAB_ATTR_RO(slab_size);
4304 static ssize_t align_show(struct kmem_cache *s, char *buf)
4306 return sprintf(buf, "%d\n", s->align);
4308 SLAB_ATTR_RO(align);
4310 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4312 return sprintf(buf, "%d\n", s->object_size);
4314 SLAB_ATTR_RO(object_size);
4316 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4318 return sprintf(buf, "%d\n", oo_objects(s->oo));
4320 SLAB_ATTR_RO(objs_per_slab);
4322 static ssize_t order_store(struct kmem_cache *s,
4323 const char *buf, size_t length)
4325 unsigned long order;
4328 err = kstrtoul(buf, 10, &order);
4332 if (order > slub_max_order || order < slub_min_order)
4335 calculate_sizes(s, order);
4339 static ssize_t order_show(struct kmem_cache *s, char *buf)
4341 return sprintf(buf, "%d\n", oo_order(s->oo));
4345 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4347 return sprintf(buf, "%lu\n", s->min_partial);
4350 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4356 err = kstrtoul(buf, 10, &min);
4360 set_min_partial(s, min);
4363 SLAB_ATTR(min_partial);
4365 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4367 return sprintf(buf, "%u\n", s->cpu_partial);
4370 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4373 unsigned long objects;
4376 err = kstrtoul(buf, 10, &objects);
4379 if (objects && !kmem_cache_has_cpu_partial(s))
4382 s->cpu_partial = objects;
4386 SLAB_ATTR(cpu_partial);
4388 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4392 return sprintf(buf, "%pS\n", s->ctor);
4396 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4398 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4400 SLAB_ATTR_RO(aliases);
4402 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4404 return show_slab_objects(s, buf, SO_PARTIAL);
4406 SLAB_ATTR_RO(partial);
4408 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4410 return show_slab_objects(s, buf, SO_CPU);
4412 SLAB_ATTR_RO(cpu_slabs);
4414 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4416 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4418 SLAB_ATTR_RO(objects);
4420 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4422 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4424 SLAB_ATTR_RO(objects_partial);
4426 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4433 for_each_online_cpu(cpu) {
4434 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4437 pages += page->pages;
4438 objects += page->pobjects;
4442 len = sprintf(buf, "%d(%d)", objects, pages);
4445 for_each_online_cpu(cpu) {
4446 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4448 if (page && len < PAGE_SIZE - 20)
4449 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4450 page->pobjects, page->pages);
4453 return len + sprintf(buf + len, "\n");
4455 SLAB_ATTR_RO(slabs_cpu_partial);
4457 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4459 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4462 static ssize_t reclaim_account_store(struct kmem_cache *s,
4463 const char *buf, size_t length)
4465 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4467 s->flags |= SLAB_RECLAIM_ACCOUNT;
4470 SLAB_ATTR(reclaim_account);
4472 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4474 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4476 SLAB_ATTR_RO(hwcache_align);
4478 #ifdef CONFIG_ZONE_DMA
4479 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4481 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4483 SLAB_ATTR_RO(cache_dma);
4486 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4488 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4490 SLAB_ATTR_RO(destroy_by_rcu);
4492 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4494 return sprintf(buf, "%d\n", s->reserved);
4496 SLAB_ATTR_RO(reserved);
4498 #ifdef CONFIG_SLUB_DEBUG
4499 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4501 return show_slab_objects(s, buf, SO_ALL);
4503 SLAB_ATTR_RO(slabs);
4505 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4507 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4509 SLAB_ATTR_RO(total_objects);
4511 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4513 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4516 static ssize_t sanity_checks_store(struct kmem_cache *s,
4517 const char *buf, size_t length)
4519 s->flags &= ~SLAB_DEBUG_FREE;
4520 if (buf[0] == '1') {
4521 s->flags &= ~__CMPXCHG_DOUBLE;
4522 s->flags |= SLAB_DEBUG_FREE;
4526 SLAB_ATTR(sanity_checks);
4528 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4530 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4533 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4537 * Tracing a merged cache is going to give confusing results
4538 * as well as cause other issues like converting a mergeable
4539 * cache into an umergeable one.
4541 if (s->refcount > 1)
4544 s->flags &= ~SLAB_TRACE;
4545 if (buf[0] == '1') {
4546 s->flags &= ~__CMPXCHG_DOUBLE;
4547 s->flags |= SLAB_TRACE;
4553 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4555 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4558 static ssize_t red_zone_store(struct kmem_cache *s,
4559 const char *buf, size_t length)
4561 if (any_slab_objects(s))
4564 s->flags &= ~SLAB_RED_ZONE;
4565 if (buf[0] == '1') {
4566 s->flags &= ~__CMPXCHG_DOUBLE;
4567 s->flags |= SLAB_RED_ZONE;
4569 calculate_sizes(s, -1);
4572 SLAB_ATTR(red_zone);
4574 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4579 static ssize_t poison_store(struct kmem_cache *s,
4580 const char *buf, size_t length)
4582 if (any_slab_objects(s))
4585 s->flags &= ~SLAB_POISON;
4586 if (buf[0] == '1') {
4587 s->flags &= ~__CMPXCHG_DOUBLE;
4588 s->flags |= SLAB_POISON;
4590 calculate_sizes(s, -1);
4595 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4597 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4600 static ssize_t store_user_store(struct kmem_cache *s,
4601 const char *buf, size_t length)
4603 if (any_slab_objects(s))
4606 s->flags &= ~SLAB_STORE_USER;
4607 if (buf[0] == '1') {
4608 s->flags &= ~__CMPXCHG_DOUBLE;
4609 s->flags |= SLAB_STORE_USER;
4611 calculate_sizes(s, -1);
4614 SLAB_ATTR(store_user);
4616 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4621 static ssize_t validate_store(struct kmem_cache *s,
4622 const char *buf, size_t length)
4626 if (buf[0] == '1') {
4627 ret = validate_slab_cache(s);
4633 SLAB_ATTR(validate);
4635 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4637 if (!(s->flags & SLAB_STORE_USER))
4639 return list_locations(s, buf, TRACK_ALLOC);
4641 SLAB_ATTR_RO(alloc_calls);
4643 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4645 if (!(s->flags & SLAB_STORE_USER))
4647 return list_locations(s, buf, TRACK_FREE);
4649 SLAB_ATTR_RO(free_calls);
4650 #endif /* CONFIG_SLUB_DEBUG */
4652 #ifdef CONFIG_FAILSLAB
4653 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4655 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4658 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4661 if (s->refcount > 1)
4664 s->flags &= ~SLAB_FAILSLAB;
4666 s->flags |= SLAB_FAILSLAB;
4669 SLAB_ATTR(failslab);
4672 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4677 static ssize_t shrink_store(struct kmem_cache *s,
4678 const char *buf, size_t length)
4680 if (buf[0] == '1') {
4681 int rc = kmem_cache_shrink(s);
4692 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4694 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4697 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4698 const char *buf, size_t length)
4700 unsigned long ratio;
4703 err = kstrtoul(buf, 10, &ratio);
4708 s->remote_node_defrag_ratio = ratio * 10;
4712 SLAB_ATTR(remote_node_defrag_ratio);
4715 #ifdef CONFIG_SLUB_STATS
4716 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4718 unsigned long sum = 0;
4721 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4726 for_each_online_cpu(cpu) {
4727 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4733 len = sprintf(buf, "%lu", sum);
4736 for_each_online_cpu(cpu) {
4737 if (data[cpu] && len < PAGE_SIZE - 20)
4738 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4742 return len + sprintf(buf + len, "\n");
4745 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4749 for_each_online_cpu(cpu)
4750 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4753 #define STAT_ATTR(si, text) \
4754 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4756 return show_stat(s, buf, si); \
4758 static ssize_t text##_store(struct kmem_cache *s, \
4759 const char *buf, size_t length) \
4761 if (buf[0] != '0') \
4763 clear_stat(s, si); \
4768 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4769 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4770 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4771 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4772 STAT_ATTR(FREE_FROZEN, free_frozen);
4773 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4774 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4775 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4776 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4777 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4778 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4779 STAT_ATTR(FREE_SLAB, free_slab);
4780 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4781 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4782 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4783 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4784 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4785 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4786 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4787 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4788 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4789 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4790 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4791 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4792 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4793 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4796 static struct attribute *slab_attrs[] = {
4797 &slab_size_attr.attr,
4798 &object_size_attr.attr,
4799 &objs_per_slab_attr.attr,
4801 &min_partial_attr.attr,
4802 &cpu_partial_attr.attr,
4804 &objects_partial_attr.attr,
4806 &cpu_slabs_attr.attr,
4810 &hwcache_align_attr.attr,
4811 &reclaim_account_attr.attr,
4812 &destroy_by_rcu_attr.attr,
4814 &reserved_attr.attr,
4815 &slabs_cpu_partial_attr.attr,
4816 #ifdef CONFIG_SLUB_DEBUG
4817 &total_objects_attr.attr,
4819 &sanity_checks_attr.attr,
4821 &red_zone_attr.attr,
4823 &store_user_attr.attr,
4824 &validate_attr.attr,
4825 &alloc_calls_attr.attr,
4826 &free_calls_attr.attr,
4828 #ifdef CONFIG_ZONE_DMA
4829 &cache_dma_attr.attr,
4832 &remote_node_defrag_ratio_attr.attr,
4834 #ifdef CONFIG_SLUB_STATS
4835 &alloc_fastpath_attr.attr,
4836 &alloc_slowpath_attr.attr,
4837 &free_fastpath_attr.attr,
4838 &free_slowpath_attr.attr,
4839 &free_frozen_attr.attr,
4840 &free_add_partial_attr.attr,
4841 &free_remove_partial_attr.attr,
4842 &alloc_from_partial_attr.attr,
4843 &alloc_slab_attr.attr,
4844 &alloc_refill_attr.attr,
4845 &alloc_node_mismatch_attr.attr,
4846 &free_slab_attr.attr,
4847 &cpuslab_flush_attr.attr,
4848 &deactivate_full_attr.attr,
4849 &deactivate_empty_attr.attr,
4850 &deactivate_to_head_attr.attr,
4851 &deactivate_to_tail_attr.attr,
4852 &deactivate_remote_frees_attr.attr,
4853 &deactivate_bypass_attr.attr,
4854 &order_fallback_attr.attr,
4855 &cmpxchg_double_fail_attr.attr,
4856 &cmpxchg_double_cpu_fail_attr.attr,
4857 &cpu_partial_alloc_attr.attr,
4858 &cpu_partial_free_attr.attr,
4859 &cpu_partial_node_attr.attr,
4860 &cpu_partial_drain_attr.attr,
4862 #ifdef CONFIG_FAILSLAB
4863 &failslab_attr.attr,
4869 static struct attribute_group slab_attr_group = {
4870 .attrs = slab_attrs,
4873 static ssize_t slab_attr_show(struct kobject *kobj,
4874 struct attribute *attr,
4877 struct slab_attribute *attribute;
4878 struct kmem_cache *s;
4881 attribute = to_slab_attr(attr);
4884 if (!attribute->show)
4887 err = attribute->show(s, buf);
4892 static ssize_t slab_attr_store(struct kobject *kobj,
4893 struct attribute *attr,
4894 const char *buf, size_t len)
4896 struct slab_attribute *attribute;
4897 struct kmem_cache *s;
4900 attribute = to_slab_attr(attr);
4903 if (!attribute->store)
4906 err = attribute->store(s, buf, len);
4907 #ifdef CONFIG_MEMCG_KMEM
4908 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4911 mutex_lock(&slab_mutex);
4912 if (s->max_attr_size < len)
4913 s->max_attr_size = len;
4916 * This is a best effort propagation, so this function's return
4917 * value will be determined by the parent cache only. This is
4918 * basically because not all attributes will have a well
4919 * defined semantics for rollbacks - most of the actions will
4920 * have permanent effects.
4922 * Returning the error value of any of the children that fail
4923 * is not 100 % defined, in the sense that users seeing the
4924 * error code won't be able to know anything about the state of
4927 * Only returning the error code for the parent cache at least
4928 * has well defined semantics. The cache being written to
4929 * directly either failed or succeeded, in which case we loop
4930 * through the descendants with best-effort propagation.
4932 for_each_memcg_cache_index(i) {
4933 struct kmem_cache *c = cache_from_memcg_idx(s, i);
4935 attribute->store(c, buf, len);
4937 mutex_unlock(&slab_mutex);
4943 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
4945 #ifdef CONFIG_MEMCG_KMEM
4947 char *buffer = NULL;
4948 struct kmem_cache *root_cache;
4950 if (is_root_cache(s))
4953 root_cache = s->memcg_params->root_cache;
4956 * This mean this cache had no attribute written. Therefore, no point
4957 * in copying default values around
4959 if (!root_cache->max_attr_size)
4962 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
4965 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
4967 if (!attr || !attr->store || !attr->show)
4971 * It is really bad that we have to allocate here, so we will
4972 * do it only as a fallback. If we actually allocate, though,
4973 * we can just use the allocated buffer until the end.
4975 * Most of the slub attributes will tend to be very small in
4976 * size, but sysfs allows buffers up to a page, so they can
4977 * theoretically happen.
4981 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
4984 buffer = (char *) get_zeroed_page(GFP_KERNEL);
4985 if (WARN_ON(!buffer))
4990 attr->show(root_cache, buf);
4991 attr->store(s, buf, strlen(buf));
4995 free_page((unsigned long)buffer);
4999 static void kmem_cache_release(struct kobject *k)
5001 slab_kmem_cache_release(to_slab(k));
5004 static const struct sysfs_ops slab_sysfs_ops = {
5005 .show = slab_attr_show,
5006 .store = slab_attr_store,
5009 static struct kobj_type slab_ktype = {
5010 .sysfs_ops = &slab_sysfs_ops,
5011 .release = kmem_cache_release,
5014 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5016 struct kobj_type *ktype = get_ktype(kobj);
5018 if (ktype == &slab_ktype)
5023 static const struct kset_uevent_ops slab_uevent_ops = {
5024 .filter = uevent_filter,
5027 static struct kset *slab_kset;
5029 static inline struct kset *cache_kset(struct kmem_cache *s)
5031 #ifdef CONFIG_MEMCG_KMEM
5032 if (!is_root_cache(s))
5033 return s->memcg_params->root_cache->memcg_kset;
5038 #define ID_STR_LENGTH 64
5040 /* Create a unique string id for a slab cache:
5042 * Format :[flags-]size
5044 static char *create_unique_id(struct kmem_cache *s)
5046 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5053 * First flags affecting slabcache operations. We will only
5054 * get here for aliasable slabs so we do not need to support
5055 * too many flags. The flags here must cover all flags that
5056 * are matched during merging to guarantee that the id is
5059 if (s->flags & SLAB_CACHE_DMA)
5061 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5063 if (s->flags & SLAB_DEBUG_FREE)
5065 if (!(s->flags & SLAB_NOTRACK))
5069 p += sprintf(p, "%07d", s->size);
5071 BUG_ON(p > name + ID_STR_LENGTH - 1);
5075 static int sysfs_slab_add(struct kmem_cache *s)
5079 int unmergeable = slab_unmergeable(s);
5083 * Slabcache can never be merged so we can use the name proper.
5084 * This is typically the case for debug situations. In that
5085 * case we can catch duplicate names easily.
5087 sysfs_remove_link(&slab_kset->kobj, s->name);
5091 * Create a unique name for the slab as a target
5094 name = create_unique_id(s);
5097 s->kobj.kset = cache_kset(s);
5098 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5102 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5106 #ifdef CONFIG_MEMCG_KMEM
5107 if (is_root_cache(s)) {
5108 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5109 if (!s->memcg_kset) {
5116 kobject_uevent(&s->kobj, KOBJ_ADD);
5118 /* Setup first alias */
5119 sysfs_slab_alias(s, s->name);
5126 kobject_del(&s->kobj);
5128 kobject_put(&s->kobj);
5132 void sysfs_slab_remove(struct kmem_cache *s)
5134 if (slab_state < FULL)
5136 * Sysfs has not been setup yet so no need to remove the
5141 #ifdef CONFIG_MEMCG_KMEM
5142 kset_unregister(s->memcg_kset);
5144 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5145 kobject_del(&s->kobj);
5146 kobject_put(&s->kobj);
5150 * Need to buffer aliases during bootup until sysfs becomes
5151 * available lest we lose that information.
5153 struct saved_alias {
5154 struct kmem_cache *s;
5156 struct saved_alias *next;
5159 static struct saved_alias *alias_list;
5161 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5163 struct saved_alias *al;
5165 if (slab_state == FULL) {
5167 * If we have a leftover link then remove it.
5169 sysfs_remove_link(&slab_kset->kobj, name);
5170 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5173 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5179 al->next = alias_list;
5184 static int __init slab_sysfs_init(void)
5186 struct kmem_cache *s;
5189 mutex_lock(&slab_mutex);
5191 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5193 mutex_unlock(&slab_mutex);
5194 pr_err("Cannot register slab subsystem.\n");
5200 list_for_each_entry(s, &slab_caches, list) {
5201 err = sysfs_slab_add(s);
5203 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5207 while (alias_list) {
5208 struct saved_alias *al = alias_list;
5210 alias_list = alias_list->next;
5211 err = sysfs_slab_alias(al->s, al->name);
5213 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5218 mutex_unlock(&slab_mutex);
5223 __initcall(slab_sysfs_init);
5224 #endif /* CONFIG_SYSFS */
5227 * The /proc/slabinfo ABI
5229 #ifdef CONFIG_SLABINFO
5230 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5232 unsigned long nr_slabs = 0;
5233 unsigned long nr_objs = 0;
5234 unsigned long nr_free = 0;
5236 struct kmem_cache_node *n;
5238 for_each_kmem_cache_node(s, node, n) {
5239 nr_slabs += node_nr_slabs(n);
5240 nr_objs += node_nr_objs(n);
5241 nr_free += count_partial(n, count_free);
5244 sinfo->active_objs = nr_objs - nr_free;
5245 sinfo->num_objs = nr_objs;
5246 sinfo->active_slabs = nr_slabs;
5247 sinfo->num_slabs = nr_slabs;
5248 sinfo->objects_per_slab = oo_objects(s->oo);
5249 sinfo->cache_order = oo_order(s->oo);
5252 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5256 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5257 size_t count, loff_t *ppos)
5261 #endif /* CONFIG_SLABINFO */