kmemtrace: use tracepoints
[linux-2.6/mini2440.git] / mm / slub.c
bloba98078bf738b209c328a43f18764bb056849db70
1 /*
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 and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 */
11 #include <linux/mm.h>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/proc_fs.h>
18 #include <linux/seq_file.h>
19 #include <trace/kmemtrace.h>
20 #include <linux/cpu.h>
21 #include <linux/cpuset.h>
22 #include <linux/mempolicy.h>
23 #include <linux/ctype.h>
24 #include <linux/debugobjects.h>
25 #include <linux/kallsyms.h>
26 #include <linux/memory.h>
27 #include <linux/math64.h>
28 #include <linux/fault-inject.h>
31 * Lock order:
32 * 1. slab_lock(page)
33 * 2. slab->list_lock
35 * The slab_lock protects operations on the object of a particular
36 * slab and its metadata in the page struct. If the slab lock
37 * has been taken then no allocations nor frees can be performed
38 * on the objects in the slab nor can the slab be added or removed
39 * from the partial or full lists since this would mean modifying
40 * the page_struct of the slab.
42 * The list_lock protects the partial and full list on each node and
43 * the partial slab counter. If taken then no new slabs may be added or
44 * removed from the lists nor make the number of partial slabs be modified.
45 * (Note that the total number of slabs is an atomic value that may be
46 * modified without taking the list lock).
48 * The list_lock is a centralized lock and thus we avoid taking it as
49 * much as possible. As long as SLUB does not have to handle partial
50 * slabs, operations can continue without any centralized lock. F.e.
51 * allocating a long series of objects that fill up slabs does not require
52 * the list lock.
54 * The lock order is sometimes inverted when we are trying to get a slab
55 * off a list. We take the list_lock and then look for a page on the list
56 * to use. While we do that objects in the slabs may be freed. We can
57 * only operate on the slab if we have also taken the slab_lock. So we use
58 * a slab_trylock() on the slab. If trylock was successful then no frees
59 * can occur anymore and we can use the slab for allocations etc. If the
60 * slab_trylock() does not succeed then frees are in progress in the slab and
61 * we must stay away from it for a while since we may cause a bouncing
62 * cacheline if we try to acquire the lock. So go onto the next slab.
63 * If all pages are busy then we may allocate a new slab instead of reusing
64 * a partial slab. A new slab has noone operating on it and thus there is
65 * no danger of cacheline contention.
67 * Interrupts are disabled during allocation and deallocation in order to
68 * make the slab allocator safe to use in the context of an irq. In addition
69 * interrupts are disabled to ensure that the processor does not change
70 * while handling per_cpu slabs, due to kernel preemption.
72 * SLUB assigns one slab for allocation to each processor.
73 * Allocations only occur from these slabs called cpu slabs.
75 * Slabs with free elements are kept on a partial list and during regular
76 * operations no list for full slabs is used. If an object in a full slab is
77 * freed then the slab will show up again on the partial lists.
78 * We track full slabs for debugging purposes though because otherwise we
79 * cannot scan all objects.
81 * Slabs are freed when they become empty. Teardown and setup is
82 * minimal so we rely on the page allocators per cpu caches for
83 * fast frees and allocs.
85 * Overloading of page flags that are otherwise used for LRU management.
87 * PageActive The slab is frozen and exempt from list processing.
88 * This means that the slab is dedicated to a purpose
89 * such as satisfying allocations for a specific
90 * processor. Objects may be freed in the slab while
91 * it is frozen but slab_free will then skip the usual
92 * list operations. It is up to the processor holding
93 * the slab to integrate the slab into the slab lists
94 * when the slab is no longer needed.
96 * One use of this flag is to mark slabs that are
97 * used for allocations. Then such a slab becomes a cpu
98 * slab. The cpu slab may be equipped with an additional
99 * freelist that allows lockless access to
100 * free objects in addition to the regular freelist
101 * that requires the slab lock.
103 * PageError Slab requires special handling due to debug
104 * options set. This moves slab handling out of
105 * the fast path and disables lockless freelists.
108 #ifdef CONFIG_SLUB_DEBUG
109 #define SLABDEBUG 1
110 #else
111 #define SLABDEBUG 0
112 #endif
115 * Issues still to be resolved:
117 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
119 * - Variable sizing of the per node arrays
122 /* Enable to test recovery from slab corruption on boot */
123 #undef SLUB_RESILIENCY_TEST
126 * Mininum number of partial slabs. These will be left on the partial
127 * lists even if they are empty. kmem_cache_shrink may reclaim them.
129 #define MIN_PARTIAL 5
132 * Maximum number of desirable partial slabs.
133 * The existence of more partial slabs makes kmem_cache_shrink
134 * sort the partial list by the number of objects in the.
136 #define MAX_PARTIAL 10
138 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
139 SLAB_POISON | SLAB_STORE_USER)
142 * Set of flags that will prevent slab merging
144 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
145 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
147 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
148 SLAB_CACHE_DMA)
150 #ifndef ARCH_KMALLOC_MINALIGN
151 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
152 #endif
154 #ifndef ARCH_SLAB_MINALIGN
155 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
156 #endif
158 #define OO_SHIFT 16
159 #define OO_MASK ((1 << OO_SHIFT) - 1)
160 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
162 /* Internal SLUB flags */
163 #define __OBJECT_POISON 0x80000000 /* Poison object */
164 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
166 static int kmem_size = sizeof(struct kmem_cache);
168 #ifdef CONFIG_SMP
169 static struct notifier_block slab_notifier;
170 #endif
172 static enum {
173 DOWN, /* No slab functionality available */
174 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
175 UP, /* Everything works but does not show up in sysfs */
176 SYSFS /* Sysfs up */
177 } slab_state = DOWN;
179 /* A list of all slab caches on the system */
180 static DECLARE_RWSEM(slub_lock);
181 static LIST_HEAD(slab_caches);
184 * Tracking user of a slab.
186 struct track {
187 unsigned long addr; /* Called from address */
188 int cpu; /* Was running on cpu */
189 int pid; /* Pid context */
190 unsigned long when; /* When did the operation occur */
193 enum track_item { TRACK_ALLOC, TRACK_FREE };
195 #ifdef CONFIG_SLUB_DEBUG
196 static int sysfs_slab_add(struct kmem_cache *);
197 static int sysfs_slab_alias(struct kmem_cache *, const char *);
198 static void sysfs_slab_remove(struct kmem_cache *);
200 #else
201 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
202 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
203 { return 0; }
204 static inline void sysfs_slab_remove(struct kmem_cache *s)
206 kfree(s);
209 #endif
211 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
213 #ifdef CONFIG_SLUB_STATS
214 c->stat[si]++;
215 #endif
218 /********************************************************************
219 * Core slab cache functions
220 *******************************************************************/
222 int slab_is_available(void)
224 return slab_state >= UP;
227 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
229 #ifdef CONFIG_NUMA
230 return s->node[node];
231 #else
232 return &s->local_node;
233 #endif
236 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
238 #ifdef CONFIG_SMP
239 return s->cpu_slab[cpu];
240 #else
241 return &s->cpu_slab;
242 #endif
245 /* Verify that a pointer has an address that is valid within a slab page */
246 static inline int check_valid_pointer(struct kmem_cache *s,
247 struct page *page, const void *object)
249 void *base;
251 if (!object)
252 return 1;
254 base = page_address(page);
255 if (object < base || object >= base + page->objects * s->size ||
256 (object - base) % s->size) {
257 return 0;
260 return 1;
264 * Slow version of get and set free pointer.
266 * This version requires touching the cache lines of kmem_cache which
267 * we avoid to do in the fast alloc free paths. There we obtain the offset
268 * from the page struct.
270 static inline void *get_freepointer(struct kmem_cache *s, void *object)
272 return *(void **)(object + s->offset);
275 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
277 *(void **)(object + s->offset) = fp;
280 /* Loop over all objects in a slab */
281 #define for_each_object(__p, __s, __addr, __objects) \
282 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
283 __p += (__s)->size)
285 /* Scan freelist */
286 #define for_each_free_object(__p, __s, __free) \
287 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
289 /* Determine object index from a given position */
290 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292 return (p - addr) / s->size;
295 static inline struct kmem_cache_order_objects oo_make(int order,
296 unsigned long size)
298 struct kmem_cache_order_objects x = {
299 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
302 return x;
305 static inline int oo_order(struct kmem_cache_order_objects x)
307 return x.x >> OO_SHIFT;
310 static inline int oo_objects(struct kmem_cache_order_objects x)
312 return x.x & OO_MASK;
315 #ifdef CONFIG_SLUB_DEBUG
317 * Debug settings:
319 #ifdef CONFIG_SLUB_DEBUG_ON
320 static int slub_debug = DEBUG_DEFAULT_FLAGS;
321 #else
322 static int slub_debug;
323 #endif
325 static char *slub_debug_slabs;
328 * Object debugging
330 static void print_section(char *text, u8 *addr, unsigned int length)
332 int i, offset;
333 int newline = 1;
334 char ascii[17];
336 ascii[16] = 0;
338 for (i = 0; i < length; i++) {
339 if (newline) {
340 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
341 newline = 0;
343 printk(KERN_CONT " %02x", addr[i]);
344 offset = i % 16;
345 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
346 if (offset == 15) {
347 printk(KERN_CONT " %s\n", ascii);
348 newline = 1;
351 if (!newline) {
352 i %= 16;
353 while (i < 16) {
354 printk(KERN_CONT " ");
355 ascii[i] = ' ';
356 i++;
358 printk(KERN_CONT " %s\n", ascii);
362 static struct track *get_track(struct kmem_cache *s, void *object,
363 enum track_item alloc)
365 struct track *p;
367 if (s->offset)
368 p = object + s->offset + sizeof(void *);
369 else
370 p = object + s->inuse;
372 return p + alloc;
375 static void set_track(struct kmem_cache *s, void *object,
376 enum track_item alloc, unsigned long addr)
378 struct track *p = get_track(s, object, alloc);
380 if (addr) {
381 p->addr = addr;
382 p->cpu = smp_processor_id();
383 p->pid = current->pid;
384 p->when = jiffies;
385 } else
386 memset(p, 0, sizeof(struct track));
389 static void init_tracking(struct kmem_cache *s, void *object)
391 if (!(s->flags & SLAB_STORE_USER))
392 return;
394 set_track(s, object, TRACK_FREE, 0UL);
395 set_track(s, object, TRACK_ALLOC, 0UL);
398 static void print_track(const char *s, struct track *t)
400 if (!t->addr)
401 return;
403 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
404 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
407 static void print_tracking(struct kmem_cache *s, void *object)
409 if (!(s->flags & SLAB_STORE_USER))
410 return;
412 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
413 print_track("Freed", get_track(s, object, TRACK_FREE));
416 static void print_page_info(struct page *page)
418 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
419 page, page->objects, page->inuse, page->freelist, page->flags);
423 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
425 va_list args;
426 char buf[100];
428 va_start(args, fmt);
429 vsnprintf(buf, sizeof(buf), fmt, args);
430 va_end(args);
431 printk(KERN_ERR "========================================"
432 "=====================================\n");
433 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
434 printk(KERN_ERR "----------------------------------------"
435 "-------------------------------------\n\n");
438 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
440 va_list args;
441 char buf[100];
443 va_start(args, fmt);
444 vsnprintf(buf, sizeof(buf), fmt, args);
445 va_end(args);
446 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
449 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
451 unsigned int off; /* Offset of last byte */
452 u8 *addr = page_address(page);
454 print_tracking(s, p);
456 print_page_info(page);
458 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
459 p, p - addr, get_freepointer(s, p));
461 if (p > addr + 16)
462 print_section("Bytes b4", p - 16, 16);
464 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
466 if (s->flags & SLAB_RED_ZONE)
467 print_section("Redzone", p + s->objsize,
468 s->inuse - s->objsize);
470 if (s->offset)
471 off = s->offset + sizeof(void *);
472 else
473 off = s->inuse;
475 if (s->flags & SLAB_STORE_USER)
476 off += 2 * sizeof(struct track);
478 if (off != s->size)
479 /* Beginning of the filler is the free pointer */
480 print_section("Padding", p + off, s->size - off);
482 dump_stack();
485 static void object_err(struct kmem_cache *s, struct page *page,
486 u8 *object, char *reason)
488 slab_bug(s, "%s", reason);
489 print_trailer(s, page, object);
492 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
494 va_list args;
495 char buf[100];
497 va_start(args, fmt);
498 vsnprintf(buf, sizeof(buf), fmt, args);
499 va_end(args);
500 slab_bug(s, "%s", buf);
501 print_page_info(page);
502 dump_stack();
505 static void init_object(struct kmem_cache *s, void *object, int active)
507 u8 *p = object;
509 if (s->flags & __OBJECT_POISON) {
510 memset(p, POISON_FREE, s->objsize - 1);
511 p[s->objsize - 1] = POISON_END;
514 if (s->flags & SLAB_RED_ZONE)
515 memset(p + s->objsize,
516 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
517 s->inuse - s->objsize);
520 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
522 while (bytes) {
523 if (*start != (u8)value)
524 return start;
525 start++;
526 bytes--;
528 return NULL;
531 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
532 void *from, void *to)
534 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
535 memset(from, data, to - from);
538 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
539 u8 *object, char *what,
540 u8 *start, unsigned int value, unsigned int bytes)
542 u8 *fault;
543 u8 *end;
545 fault = check_bytes(start, value, bytes);
546 if (!fault)
547 return 1;
549 end = start + bytes;
550 while (end > fault && end[-1] == value)
551 end--;
553 slab_bug(s, "%s overwritten", what);
554 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
555 fault, end - 1, fault[0], value);
556 print_trailer(s, page, object);
558 restore_bytes(s, what, value, fault, end);
559 return 0;
563 * Object layout:
565 * object address
566 * Bytes of the object to be managed.
567 * If the freepointer may overlay the object then the free
568 * pointer is the first word of the object.
570 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
571 * 0xa5 (POISON_END)
573 * object + s->objsize
574 * Padding to reach word boundary. This is also used for Redzoning.
575 * Padding is extended by another word if Redzoning is enabled and
576 * objsize == inuse.
578 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
579 * 0xcc (RED_ACTIVE) for objects in use.
581 * object + s->inuse
582 * Meta data starts here.
584 * A. Free pointer (if we cannot overwrite object on free)
585 * B. Tracking data for SLAB_STORE_USER
586 * C. Padding to reach required alignment boundary or at mininum
587 * one word if debugging is on to be able to detect writes
588 * before the word boundary.
590 * Padding is done using 0x5a (POISON_INUSE)
592 * object + s->size
593 * Nothing is used beyond s->size.
595 * If slabcaches are merged then the objsize and inuse boundaries are mostly
596 * ignored. And therefore no slab options that rely on these boundaries
597 * may be used with merged slabcaches.
600 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
602 unsigned long off = s->inuse; /* The end of info */
604 if (s->offset)
605 /* Freepointer is placed after the object. */
606 off += sizeof(void *);
608 if (s->flags & SLAB_STORE_USER)
609 /* We also have user information there */
610 off += 2 * sizeof(struct track);
612 if (s->size == off)
613 return 1;
615 return check_bytes_and_report(s, page, p, "Object padding",
616 p + off, POISON_INUSE, s->size - off);
619 /* Check the pad bytes at the end of a slab page */
620 static int slab_pad_check(struct kmem_cache *s, struct page *page)
622 u8 *start;
623 u8 *fault;
624 u8 *end;
625 int length;
626 int remainder;
628 if (!(s->flags & SLAB_POISON))
629 return 1;
631 start = page_address(page);
632 length = (PAGE_SIZE << compound_order(page));
633 end = start + length;
634 remainder = length % s->size;
635 if (!remainder)
636 return 1;
638 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
639 if (!fault)
640 return 1;
641 while (end > fault && end[-1] == POISON_INUSE)
642 end--;
644 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
645 print_section("Padding", end - remainder, remainder);
647 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
648 return 0;
651 static int check_object(struct kmem_cache *s, struct page *page,
652 void *object, int active)
654 u8 *p = object;
655 u8 *endobject = object + s->objsize;
657 if (s->flags & SLAB_RED_ZONE) {
658 unsigned int red =
659 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
661 if (!check_bytes_and_report(s, page, object, "Redzone",
662 endobject, red, s->inuse - s->objsize))
663 return 0;
664 } else {
665 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
666 check_bytes_and_report(s, page, p, "Alignment padding",
667 endobject, POISON_INUSE, s->inuse - s->objsize);
671 if (s->flags & SLAB_POISON) {
672 if (!active && (s->flags & __OBJECT_POISON) &&
673 (!check_bytes_and_report(s, page, p, "Poison", p,
674 POISON_FREE, s->objsize - 1) ||
675 !check_bytes_and_report(s, page, p, "Poison",
676 p + s->objsize - 1, POISON_END, 1)))
677 return 0;
679 * check_pad_bytes cleans up on its own.
681 check_pad_bytes(s, page, p);
684 if (!s->offset && active)
686 * Object and freepointer overlap. Cannot check
687 * freepointer while object is allocated.
689 return 1;
691 /* Check free pointer validity */
692 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
693 object_err(s, page, p, "Freepointer corrupt");
695 * No choice but to zap it and thus lose the remainder
696 * of the free objects in this slab. May cause
697 * another error because the object count is now wrong.
699 set_freepointer(s, p, NULL);
700 return 0;
702 return 1;
705 static int check_slab(struct kmem_cache *s, struct page *page)
707 int maxobj;
709 VM_BUG_ON(!irqs_disabled());
711 if (!PageSlab(page)) {
712 slab_err(s, page, "Not a valid slab page");
713 return 0;
716 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
717 if (page->objects > maxobj) {
718 slab_err(s, page, "objects %u > max %u",
719 s->name, page->objects, maxobj);
720 return 0;
722 if (page->inuse > page->objects) {
723 slab_err(s, page, "inuse %u > max %u",
724 s->name, page->inuse, page->objects);
725 return 0;
727 /* Slab_pad_check fixes things up after itself */
728 slab_pad_check(s, page);
729 return 1;
733 * Determine if a certain object on a page is on the freelist. Must hold the
734 * slab lock to guarantee that the chains are in a consistent state.
736 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
738 int nr = 0;
739 void *fp = page->freelist;
740 void *object = NULL;
741 unsigned long max_objects;
743 while (fp && nr <= page->objects) {
744 if (fp == search)
745 return 1;
746 if (!check_valid_pointer(s, page, fp)) {
747 if (object) {
748 object_err(s, page, object,
749 "Freechain corrupt");
750 set_freepointer(s, object, NULL);
751 break;
752 } else {
753 slab_err(s, page, "Freepointer corrupt");
754 page->freelist = NULL;
755 page->inuse = page->objects;
756 slab_fix(s, "Freelist cleared");
757 return 0;
759 break;
761 object = fp;
762 fp = get_freepointer(s, object);
763 nr++;
766 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
767 if (max_objects > MAX_OBJS_PER_PAGE)
768 max_objects = MAX_OBJS_PER_PAGE;
770 if (page->objects != max_objects) {
771 slab_err(s, page, "Wrong number of objects. Found %d but "
772 "should be %d", page->objects, max_objects);
773 page->objects = max_objects;
774 slab_fix(s, "Number of objects adjusted.");
776 if (page->inuse != page->objects - nr) {
777 slab_err(s, page, "Wrong object count. Counter is %d but "
778 "counted were %d", page->inuse, page->objects - nr);
779 page->inuse = page->objects - nr;
780 slab_fix(s, "Object count adjusted.");
782 return search == NULL;
785 static void trace(struct kmem_cache *s, struct page *page, void *object,
786 int alloc)
788 if (s->flags & SLAB_TRACE) {
789 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
790 s->name,
791 alloc ? "alloc" : "free",
792 object, page->inuse,
793 page->freelist);
795 if (!alloc)
796 print_section("Object", (void *)object, s->objsize);
798 dump_stack();
803 * Tracking of fully allocated slabs for debugging purposes.
805 static void add_full(struct kmem_cache_node *n, struct page *page)
807 spin_lock(&n->list_lock);
808 list_add(&page->lru, &n->full);
809 spin_unlock(&n->list_lock);
812 static void remove_full(struct kmem_cache *s, struct page *page)
814 struct kmem_cache_node *n;
816 if (!(s->flags & SLAB_STORE_USER))
817 return;
819 n = get_node(s, page_to_nid(page));
821 spin_lock(&n->list_lock);
822 list_del(&page->lru);
823 spin_unlock(&n->list_lock);
826 /* Tracking of the number of slabs for debugging purposes */
827 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
829 struct kmem_cache_node *n = get_node(s, node);
831 return atomic_long_read(&n->nr_slabs);
834 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
836 struct kmem_cache_node *n = get_node(s, node);
839 * May be called early in order to allocate a slab for the
840 * kmem_cache_node structure. Solve the chicken-egg
841 * dilemma by deferring the increment of the count during
842 * bootstrap (see early_kmem_cache_node_alloc).
844 if (!NUMA_BUILD || n) {
845 atomic_long_inc(&n->nr_slabs);
846 atomic_long_add(objects, &n->total_objects);
849 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
851 struct kmem_cache_node *n = get_node(s, node);
853 atomic_long_dec(&n->nr_slabs);
854 atomic_long_sub(objects, &n->total_objects);
857 /* Object debug checks for alloc/free paths */
858 static void setup_object_debug(struct kmem_cache *s, struct page *page,
859 void *object)
861 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
862 return;
864 init_object(s, object, 0);
865 init_tracking(s, object);
868 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
869 void *object, unsigned long addr)
871 if (!check_slab(s, page))
872 goto bad;
874 if (!on_freelist(s, page, object)) {
875 object_err(s, page, object, "Object already allocated");
876 goto bad;
879 if (!check_valid_pointer(s, page, object)) {
880 object_err(s, page, object, "Freelist Pointer check fails");
881 goto bad;
884 if (!check_object(s, page, object, 0))
885 goto bad;
887 /* Success perform special debug activities for allocs */
888 if (s->flags & SLAB_STORE_USER)
889 set_track(s, object, TRACK_ALLOC, addr);
890 trace(s, page, object, 1);
891 init_object(s, object, 1);
892 return 1;
894 bad:
895 if (PageSlab(page)) {
897 * If this is a slab page then lets do the best we can
898 * to avoid issues in the future. Marking all objects
899 * as used avoids touching the remaining objects.
901 slab_fix(s, "Marking all objects used");
902 page->inuse = page->objects;
903 page->freelist = NULL;
905 return 0;
908 static int free_debug_processing(struct kmem_cache *s, struct page *page,
909 void *object, unsigned long addr)
911 if (!check_slab(s, page))
912 goto fail;
914 if (!check_valid_pointer(s, page, object)) {
915 slab_err(s, page, "Invalid object pointer 0x%p", object);
916 goto fail;
919 if (on_freelist(s, page, object)) {
920 object_err(s, page, object, "Object already free");
921 goto fail;
924 if (!check_object(s, page, object, 1))
925 return 0;
927 if (unlikely(s != page->slab)) {
928 if (!PageSlab(page)) {
929 slab_err(s, page, "Attempt to free object(0x%p) "
930 "outside of slab", object);
931 } else if (!page->slab) {
932 printk(KERN_ERR
933 "SLUB <none>: no slab for object 0x%p.\n",
934 object);
935 dump_stack();
936 } else
937 object_err(s, page, object,
938 "page slab pointer corrupt.");
939 goto fail;
942 /* Special debug activities for freeing objects */
943 if (!PageSlubFrozen(page) && !page->freelist)
944 remove_full(s, page);
945 if (s->flags & SLAB_STORE_USER)
946 set_track(s, object, TRACK_FREE, addr);
947 trace(s, page, object, 0);
948 init_object(s, object, 0);
949 return 1;
951 fail:
952 slab_fix(s, "Object at 0x%p not freed", object);
953 return 0;
956 static int __init setup_slub_debug(char *str)
958 slub_debug = DEBUG_DEFAULT_FLAGS;
959 if (*str++ != '=' || !*str)
961 * No options specified. Switch on full debugging.
963 goto out;
965 if (*str == ',')
967 * No options but restriction on slabs. This means full
968 * debugging for slabs matching a pattern.
970 goto check_slabs;
972 slub_debug = 0;
973 if (*str == '-')
975 * Switch off all debugging measures.
977 goto out;
980 * Determine which debug features should be switched on
982 for (; *str && *str != ','; str++) {
983 switch (tolower(*str)) {
984 case 'f':
985 slub_debug |= SLAB_DEBUG_FREE;
986 break;
987 case 'z':
988 slub_debug |= SLAB_RED_ZONE;
989 break;
990 case 'p':
991 slub_debug |= SLAB_POISON;
992 break;
993 case 'u':
994 slub_debug |= SLAB_STORE_USER;
995 break;
996 case 't':
997 slub_debug |= SLAB_TRACE;
998 break;
999 default:
1000 printk(KERN_ERR "slub_debug option '%c' "
1001 "unknown. skipped\n", *str);
1005 check_slabs:
1006 if (*str == ',')
1007 slub_debug_slabs = str + 1;
1008 out:
1009 return 1;
1012 __setup("slub_debug", setup_slub_debug);
1014 static unsigned long kmem_cache_flags(unsigned long objsize,
1015 unsigned long flags, const char *name,
1016 void (*ctor)(void *))
1019 * Enable debugging if selected on the kernel commandline.
1021 if (slub_debug && (!slub_debug_slabs ||
1022 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1023 flags |= slub_debug;
1025 return flags;
1027 #else
1028 static inline void setup_object_debug(struct kmem_cache *s,
1029 struct page *page, void *object) {}
1031 static inline int alloc_debug_processing(struct kmem_cache *s,
1032 struct page *page, void *object, unsigned long addr) { return 0; }
1034 static inline int free_debug_processing(struct kmem_cache *s,
1035 struct page *page, void *object, unsigned long addr) { return 0; }
1037 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1038 { return 1; }
1039 static inline int check_object(struct kmem_cache *s, struct page *page,
1040 void *object, int active) { return 1; }
1041 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1042 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1043 unsigned long flags, const char *name,
1044 void (*ctor)(void *))
1046 return flags;
1048 #define slub_debug 0
1050 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1051 { return 0; }
1052 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1053 int objects) {}
1054 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1055 int objects) {}
1056 #endif
1059 * Slab allocation and freeing
1061 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1062 struct kmem_cache_order_objects oo)
1064 int order = oo_order(oo);
1066 if (node == -1)
1067 return alloc_pages(flags, order);
1068 else
1069 return alloc_pages_node(node, flags, order);
1072 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1074 struct page *page;
1075 struct kmem_cache_order_objects oo = s->oo;
1077 flags |= s->allocflags;
1079 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1080 oo);
1081 if (unlikely(!page)) {
1082 oo = s->min;
1084 * Allocation may have failed due to fragmentation.
1085 * Try a lower order alloc if possible
1087 page = alloc_slab_page(flags, node, oo);
1088 if (!page)
1089 return NULL;
1091 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1093 page->objects = oo_objects(oo);
1094 mod_zone_page_state(page_zone(page),
1095 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1096 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1097 1 << oo_order(oo));
1099 return page;
1102 static void setup_object(struct kmem_cache *s, struct page *page,
1103 void *object)
1105 setup_object_debug(s, page, object);
1106 if (unlikely(s->ctor))
1107 s->ctor(object);
1110 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1112 struct page *page;
1113 void *start;
1114 void *last;
1115 void *p;
1117 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1119 page = allocate_slab(s,
1120 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1121 if (!page)
1122 goto out;
1124 inc_slabs_node(s, page_to_nid(page), page->objects);
1125 page->slab = s;
1126 page->flags |= 1 << PG_slab;
1127 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1128 SLAB_STORE_USER | SLAB_TRACE))
1129 __SetPageSlubDebug(page);
1131 start = page_address(page);
1133 if (unlikely(s->flags & SLAB_POISON))
1134 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1136 last = start;
1137 for_each_object(p, s, start, page->objects) {
1138 setup_object(s, page, last);
1139 set_freepointer(s, last, p);
1140 last = p;
1142 setup_object(s, page, last);
1143 set_freepointer(s, last, NULL);
1145 page->freelist = start;
1146 page->inuse = 0;
1147 out:
1148 return page;
1151 static void __free_slab(struct kmem_cache *s, struct page *page)
1153 int order = compound_order(page);
1154 int pages = 1 << order;
1156 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1157 void *p;
1159 slab_pad_check(s, page);
1160 for_each_object(p, s, page_address(page),
1161 page->objects)
1162 check_object(s, page, p, 0);
1163 __ClearPageSlubDebug(page);
1166 mod_zone_page_state(page_zone(page),
1167 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1168 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1169 -pages);
1171 __ClearPageSlab(page);
1172 reset_page_mapcount(page);
1173 __free_pages(page, order);
1176 static void rcu_free_slab(struct rcu_head *h)
1178 struct page *page;
1180 page = container_of((struct list_head *)h, struct page, lru);
1181 __free_slab(page->slab, page);
1184 static void free_slab(struct kmem_cache *s, struct page *page)
1186 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1188 * RCU free overloads the RCU head over the LRU
1190 struct rcu_head *head = (void *)&page->lru;
1192 call_rcu(head, rcu_free_slab);
1193 } else
1194 __free_slab(s, page);
1197 static void discard_slab(struct kmem_cache *s, struct page *page)
1199 dec_slabs_node(s, page_to_nid(page), page->objects);
1200 free_slab(s, page);
1204 * Per slab locking using the pagelock
1206 static __always_inline void slab_lock(struct page *page)
1208 bit_spin_lock(PG_locked, &page->flags);
1211 static __always_inline void slab_unlock(struct page *page)
1213 __bit_spin_unlock(PG_locked, &page->flags);
1216 static __always_inline int slab_trylock(struct page *page)
1218 int rc = 1;
1220 rc = bit_spin_trylock(PG_locked, &page->flags);
1221 return rc;
1225 * Management of partially allocated slabs
1227 static void add_partial(struct kmem_cache_node *n,
1228 struct page *page, int tail)
1230 spin_lock(&n->list_lock);
1231 n->nr_partial++;
1232 if (tail)
1233 list_add_tail(&page->lru, &n->partial);
1234 else
1235 list_add(&page->lru, &n->partial);
1236 spin_unlock(&n->list_lock);
1239 static void remove_partial(struct kmem_cache *s, struct page *page)
1241 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1243 spin_lock(&n->list_lock);
1244 list_del(&page->lru);
1245 n->nr_partial--;
1246 spin_unlock(&n->list_lock);
1250 * Lock slab and remove from the partial list.
1252 * Must hold list_lock.
1254 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1255 struct page *page)
1257 if (slab_trylock(page)) {
1258 list_del(&page->lru);
1259 n->nr_partial--;
1260 __SetPageSlubFrozen(page);
1261 return 1;
1263 return 0;
1267 * Try to allocate a partial slab from a specific node.
1269 static struct page *get_partial_node(struct kmem_cache_node *n)
1271 struct page *page;
1274 * Racy check. If we mistakenly see no partial slabs then we
1275 * just allocate an empty slab. If we mistakenly try to get a
1276 * partial slab and there is none available then get_partials()
1277 * will return NULL.
1279 if (!n || !n->nr_partial)
1280 return NULL;
1282 spin_lock(&n->list_lock);
1283 list_for_each_entry(page, &n->partial, lru)
1284 if (lock_and_freeze_slab(n, page))
1285 goto out;
1286 page = NULL;
1287 out:
1288 spin_unlock(&n->list_lock);
1289 return page;
1293 * Get a page from somewhere. Search in increasing NUMA distances.
1295 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1297 #ifdef CONFIG_NUMA
1298 struct zonelist *zonelist;
1299 struct zoneref *z;
1300 struct zone *zone;
1301 enum zone_type high_zoneidx = gfp_zone(flags);
1302 struct page *page;
1305 * The defrag ratio allows a configuration of the tradeoffs between
1306 * inter node defragmentation and node local allocations. A lower
1307 * defrag_ratio increases the tendency to do local allocations
1308 * instead of attempting to obtain partial slabs from other nodes.
1310 * If the defrag_ratio is set to 0 then kmalloc() always
1311 * returns node local objects. If the ratio is higher then kmalloc()
1312 * may return off node objects because partial slabs are obtained
1313 * from other nodes and filled up.
1315 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1316 * defrag_ratio = 1000) then every (well almost) allocation will
1317 * first attempt to defrag slab caches on other nodes. This means
1318 * scanning over all nodes to look for partial slabs which may be
1319 * expensive if we do it every time we are trying to find a slab
1320 * with available objects.
1322 if (!s->remote_node_defrag_ratio ||
1323 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1324 return NULL;
1326 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1327 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1328 struct kmem_cache_node *n;
1330 n = get_node(s, zone_to_nid(zone));
1332 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1333 n->nr_partial > s->min_partial) {
1334 page = get_partial_node(n);
1335 if (page)
1336 return page;
1339 #endif
1340 return NULL;
1344 * Get a partial page, lock it and return it.
1346 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1348 struct page *page;
1349 int searchnode = (node == -1) ? numa_node_id() : node;
1351 page = get_partial_node(get_node(s, searchnode));
1352 if (page || (flags & __GFP_THISNODE))
1353 return page;
1355 return get_any_partial(s, flags);
1359 * Move a page back to the lists.
1361 * Must be called with the slab lock held.
1363 * On exit the slab lock will have been dropped.
1365 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1367 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1368 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1370 __ClearPageSlubFrozen(page);
1371 if (page->inuse) {
1373 if (page->freelist) {
1374 add_partial(n, page, tail);
1375 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1376 } else {
1377 stat(c, DEACTIVATE_FULL);
1378 if (SLABDEBUG && PageSlubDebug(page) &&
1379 (s->flags & SLAB_STORE_USER))
1380 add_full(n, page);
1382 slab_unlock(page);
1383 } else {
1384 stat(c, DEACTIVATE_EMPTY);
1385 if (n->nr_partial < s->min_partial) {
1387 * Adding an empty slab to the partial slabs in order
1388 * to avoid page allocator overhead. This slab needs
1389 * to come after the other slabs with objects in
1390 * so that the others get filled first. That way the
1391 * size of the partial list stays small.
1393 * kmem_cache_shrink can reclaim any empty slabs from
1394 * the partial list.
1396 add_partial(n, page, 1);
1397 slab_unlock(page);
1398 } else {
1399 slab_unlock(page);
1400 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1401 discard_slab(s, page);
1407 * Remove the cpu slab
1409 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1411 struct page *page = c->page;
1412 int tail = 1;
1414 if (page->freelist)
1415 stat(c, DEACTIVATE_REMOTE_FREES);
1417 * Merge cpu freelist into slab freelist. Typically we get here
1418 * because both freelists are empty. So this is unlikely
1419 * to occur.
1421 while (unlikely(c->freelist)) {
1422 void **object;
1424 tail = 0; /* Hot objects. Put the slab first */
1426 /* Retrieve object from cpu_freelist */
1427 object = c->freelist;
1428 c->freelist = c->freelist[c->offset];
1430 /* And put onto the regular freelist */
1431 object[c->offset] = page->freelist;
1432 page->freelist = object;
1433 page->inuse--;
1435 c->page = NULL;
1436 unfreeze_slab(s, page, tail);
1439 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1441 stat(c, CPUSLAB_FLUSH);
1442 slab_lock(c->page);
1443 deactivate_slab(s, c);
1447 * Flush cpu slab.
1449 * Called from IPI handler with interrupts disabled.
1451 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1453 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1455 if (likely(c && c->page))
1456 flush_slab(s, c);
1459 static void flush_cpu_slab(void *d)
1461 struct kmem_cache *s = d;
1463 __flush_cpu_slab(s, smp_processor_id());
1466 static void flush_all(struct kmem_cache *s)
1468 on_each_cpu(flush_cpu_slab, s, 1);
1472 * Check if the objects in a per cpu structure fit numa
1473 * locality expectations.
1475 static inline int node_match(struct kmem_cache_cpu *c, int node)
1477 #ifdef CONFIG_NUMA
1478 if (node != -1 && c->node != node)
1479 return 0;
1480 #endif
1481 return 1;
1485 * Slow path. The lockless freelist is empty or we need to perform
1486 * debugging duties.
1488 * Interrupts are disabled.
1490 * Processing is still very fast if new objects have been freed to the
1491 * regular freelist. In that case we simply take over the regular freelist
1492 * as the lockless freelist and zap the regular freelist.
1494 * If that is not working then we fall back to the partial lists. We take the
1495 * first element of the freelist as the object to allocate now and move the
1496 * rest of the freelist to the lockless freelist.
1498 * And if we were unable to get a new slab from the partial slab lists then
1499 * we need to allocate a new slab. This is the slowest path since it involves
1500 * a call to the page allocator and the setup of a new slab.
1502 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1503 unsigned long addr, struct kmem_cache_cpu *c)
1505 void **object;
1506 struct page *new;
1508 /* We handle __GFP_ZERO in the caller */
1509 gfpflags &= ~__GFP_ZERO;
1511 if (!c->page)
1512 goto new_slab;
1514 slab_lock(c->page);
1515 if (unlikely(!node_match(c, node)))
1516 goto another_slab;
1518 stat(c, ALLOC_REFILL);
1520 load_freelist:
1521 object = c->page->freelist;
1522 if (unlikely(!object))
1523 goto another_slab;
1524 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1525 goto debug;
1527 c->freelist = object[c->offset];
1528 c->page->inuse = c->page->objects;
1529 c->page->freelist = NULL;
1530 c->node = page_to_nid(c->page);
1531 unlock_out:
1532 slab_unlock(c->page);
1533 stat(c, ALLOC_SLOWPATH);
1534 return object;
1536 another_slab:
1537 deactivate_slab(s, c);
1539 new_slab:
1540 new = get_partial(s, gfpflags, node);
1541 if (new) {
1542 c->page = new;
1543 stat(c, ALLOC_FROM_PARTIAL);
1544 goto load_freelist;
1547 if (gfpflags & __GFP_WAIT)
1548 local_irq_enable();
1550 new = new_slab(s, gfpflags, node);
1552 if (gfpflags & __GFP_WAIT)
1553 local_irq_disable();
1555 if (new) {
1556 c = get_cpu_slab(s, smp_processor_id());
1557 stat(c, ALLOC_SLAB);
1558 if (c->page)
1559 flush_slab(s, c);
1560 slab_lock(new);
1561 __SetPageSlubFrozen(new);
1562 c->page = new;
1563 goto load_freelist;
1565 return NULL;
1566 debug:
1567 if (!alloc_debug_processing(s, c->page, object, addr))
1568 goto another_slab;
1570 c->page->inuse++;
1571 c->page->freelist = object[c->offset];
1572 c->node = -1;
1573 goto unlock_out;
1577 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1578 * have the fastpath folded into their functions. So no function call
1579 * overhead for requests that can be satisfied on the fastpath.
1581 * The fastpath works by first checking if the lockless freelist can be used.
1582 * If not then __slab_alloc is called for slow processing.
1584 * Otherwise we can simply pick the next object from the lockless free list.
1586 static __always_inline void *slab_alloc(struct kmem_cache *s,
1587 gfp_t gfpflags, int node, unsigned long addr)
1589 void **object;
1590 struct kmem_cache_cpu *c;
1591 unsigned long flags;
1592 unsigned int objsize;
1594 lockdep_trace_alloc(gfpflags);
1595 might_sleep_if(gfpflags & __GFP_WAIT);
1597 if (should_failslab(s->objsize, gfpflags))
1598 return NULL;
1600 local_irq_save(flags);
1601 c = get_cpu_slab(s, smp_processor_id());
1602 objsize = c->objsize;
1603 if (unlikely(!c->freelist || !node_match(c, node)))
1605 object = __slab_alloc(s, gfpflags, node, addr, c);
1607 else {
1608 object = c->freelist;
1609 c->freelist = object[c->offset];
1610 stat(c, ALLOC_FASTPATH);
1612 local_irq_restore(flags);
1614 if (unlikely((gfpflags & __GFP_ZERO) && object))
1615 memset(object, 0, objsize);
1617 return object;
1620 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1622 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1624 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1626 return ret;
1628 EXPORT_SYMBOL(kmem_cache_alloc);
1630 #ifdef CONFIG_KMEMTRACE
1631 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1633 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1635 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1636 #endif
1638 #ifdef CONFIG_NUMA
1639 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1641 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1643 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1644 s->objsize, s->size, gfpflags, node);
1646 return ret;
1648 EXPORT_SYMBOL(kmem_cache_alloc_node);
1649 #endif
1651 #ifdef CONFIG_KMEMTRACE
1652 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1653 gfp_t gfpflags,
1654 int node)
1656 return slab_alloc(s, gfpflags, node, _RET_IP_);
1658 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1659 #endif
1662 * Slow patch handling. This may still be called frequently since objects
1663 * have a longer lifetime than the cpu slabs in most processing loads.
1665 * So we still attempt to reduce cache line usage. Just take the slab
1666 * lock and free the item. If there is no additional partial page
1667 * handling required then we can return immediately.
1669 static void __slab_free(struct kmem_cache *s, struct page *page,
1670 void *x, unsigned long addr, unsigned int offset)
1672 void *prior;
1673 void **object = (void *)x;
1674 struct kmem_cache_cpu *c;
1676 c = get_cpu_slab(s, raw_smp_processor_id());
1677 stat(c, FREE_SLOWPATH);
1678 slab_lock(page);
1680 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1681 goto debug;
1683 checks_ok:
1684 prior = object[offset] = page->freelist;
1685 page->freelist = object;
1686 page->inuse--;
1688 if (unlikely(PageSlubFrozen(page))) {
1689 stat(c, FREE_FROZEN);
1690 goto out_unlock;
1693 if (unlikely(!page->inuse))
1694 goto slab_empty;
1697 * Objects left in the slab. If it was not on the partial list before
1698 * then add it.
1700 if (unlikely(!prior)) {
1701 add_partial(get_node(s, page_to_nid(page)), page, 1);
1702 stat(c, FREE_ADD_PARTIAL);
1705 out_unlock:
1706 slab_unlock(page);
1707 return;
1709 slab_empty:
1710 if (prior) {
1712 * Slab still on the partial list.
1714 remove_partial(s, page);
1715 stat(c, FREE_REMOVE_PARTIAL);
1717 slab_unlock(page);
1718 stat(c, FREE_SLAB);
1719 discard_slab(s, page);
1720 return;
1722 debug:
1723 if (!free_debug_processing(s, page, x, addr))
1724 goto out_unlock;
1725 goto checks_ok;
1729 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1730 * can perform fastpath freeing without additional function calls.
1732 * The fastpath is only possible if we are freeing to the current cpu slab
1733 * of this processor. This typically the case if we have just allocated
1734 * the item before.
1736 * If fastpath is not possible then fall back to __slab_free where we deal
1737 * with all sorts of special processing.
1739 static __always_inline void slab_free(struct kmem_cache *s,
1740 struct page *page, void *x, unsigned long addr)
1742 void **object = (void *)x;
1743 struct kmem_cache_cpu *c;
1744 unsigned long flags;
1746 local_irq_save(flags);
1747 c = get_cpu_slab(s, smp_processor_id());
1748 debug_check_no_locks_freed(object, c->objsize);
1749 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1750 debug_check_no_obj_freed(object, c->objsize);
1751 if (likely(page == c->page && c->node >= 0)) {
1752 object[c->offset] = c->freelist;
1753 c->freelist = object;
1754 stat(c, FREE_FASTPATH);
1755 } else
1756 __slab_free(s, page, x, addr, c->offset);
1758 local_irq_restore(flags);
1761 void kmem_cache_free(struct kmem_cache *s, void *x)
1763 struct page *page;
1765 page = virt_to_head_page(x);
1767 slab_free(s, page, x, _RET_IP_);
1769 trace_kmem_cache_free(_RET_IP_, x);
1771 EXPORT_SYMBOL(kmem_cache_free);
1773 /* Figure out on which slab page the object resides */
1774 static struct page *get_object_page(const void *x)
1776 struct page *page = virt_to_head_page(x);
1778 if (!PageSlab(page))
1779 return NULL;
1781 return page;
1785 * Object placement in a slab is made very easy because we always start at
1786 * offset 0. If we tune the size of the object to the alignment then we can
1787 * get the required alignment by putting one properly sized object after
1788 * another.
1790 * Notice that the allocation order determines the sizes of the per cpu
1791 * caches. Each processor has always one slab available for allocations.
1792 * Increasing the allocation order reduces the number of times that slabs
1793 * must be moved on and off the partial lists and is therefore a factor in
1794 * locking overhead.
1798 * Mininum / Maximum order of slab pages. This influences locking overhead
1799 * and slab fragmentation. A higher order reduces the number of partial slabs
1800 * and increases the number of allocations possible without having to
1801 * take the list_lock.
1803 static int slub_min_order;
1804 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1805 static int slub_min_objects;
1808 * Merge control. If this is set then no merging of slab caches will occur.
1809 * (Could be removed. This was introduced to pacify the merge skeptics.)
1811 static int slub_nomerge;
1814 * Calculate the order of allocation given an slab object size.
1816 * The order of allocation has significant impact on performance and other
1817 * system components. Generally order 0 allocations should be preferred since
1818 * order 0 does not cause fragmentation in the page allocator. Larger objects
1819 * be problematic to put into order 0 slabs because there may be too much
1820 * unused space left. We go to a higher order if more than 1/16th of the slab
1821 * would be wasted.
1823 * In order to reach satisfactory performance we must ensure that a minimum
1824 * number of objects is in one slab. Otherwise we may generate too much
1825 * activity on the partial lists which requires taking the list_lock. This is
1826 * less a concern for large slabs though which are rarely used.
1828 * slub_max_order specifies the order where we begin to stop considering the
1829 * number of objects in a slab as critical. If we reach slub_max_order then
1830 * we try to keep the page order as low as possible. So we accept more waste
1831 * of space in favor of a small page order.
1833 * Higher order allocations also allow the placement of more objects in a
1834 * slab and thereby reduce object handling overhead. If the user has
1835 * requested a higher mininum order then we start with that one instead of
1836 * the smallest order which will fit the object.
1838 static inline int slab_order(int size, int min_objects,
1839 int max_order, int fract_leftover)
1841 int order;
1842 int rem;
1843 int min_order = slub_min_order;
1845 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1846 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1848 for (order = max(min_order,
1849 fls(min_objects * size - 1) - PAGE_SHIFT);
1850 order <= max_order; order++) {
1852 unsigned long slab_size = PAGE_SIZE << order;
1854 if (slab_size < min_objects * size)
1855 continue;
1857 rem = slab_size % size;
1859 if (rem <= slab_size / fract_leftover)
1860 break;
1864 return order;
1867 static inline int calculate_order(int size)
1869 int order;
1870 int min_objects;
1871 int fraction;
1872 int max_objects;
1875 * Attempt to find best configuration for a slab. This
1876 * works by first attempting to generate a layout with
1877 * the best configuration and backing off gradually.
1879 * First we reduce the acceptable waste in a slab. Then
1880 * we reduce the minimum objects required in a slab.
1882 min_objects = slub_min_objects;
1883 if (!min_objects)
1884 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1885 max_objects = (PAGE_SIZE << slub_max_order)/size;
1886 min_objects = min(min_objects, max_objects);
1888 while (min_objects > 1) {
1889 fraction = 16;
1890 while (fraction >= 4) {
1891 order = slab_order(size, min_objects,
1892 slub_max_order, fraction);
1893 if (order <= slub_max_order)
1894 return order;
1895 fraction /= 2;
1897 min_objects --;
1901 * We were unable to place multiple objects in a slab. Now
1902 * lets see if we can place a single object there.
1904 order = slab_order(size, 1, slub_max_order, 1);
1905 if (order <= slub_max_order)
1906 return order;
1909 * Doh this slab cannot be placed using slub_max_order.
1911 order = slab_order(size, 1, MAX_ORDER, 1);
1912 if (order <= MAX_ORDER)
1913 return order;
1914 return -ENOSYS;
1918 * Figure out what the alignment of the objects will be.
1920 static unsigned long calculate_alignment(unsigned long flags,
1921 unsigned long align, unsigned long size)
1924 * If the user wants hardware cache aligned objects then follow that
1925 * suggestion if the object is sufficiently large.
1927 * The hardware cache alignment cannot override the specified
1928 * alignment though. If that is greater then use it.
1930 if (flags & SLAB_HWCACHE_ALIGN) {
1931 unsigned long ralign = cache_line_size();
1932 while (size <= ralign / 2)
1933 ralign /= 2;
1934 align = max(align, ralign);
1937 if (align < ARCH_SLAB_MINALIGN)
1938 align = ARCH_SLAB_MINALIGN;
1940 return ALIGN(align, sizeof(void *));
1943 static void init_kmem_cache_cpu(struct kmem_cache *s,
1944 struct kmem_cache_cpu *c)
1946 c->page = NULL;
1947 c->freelist = NULL;
1948 c->node = 0;
1949 c->offset = s->offset / sizeof(void *);
1950 c->objsize = s->objsize;
1951 #ifdef CONFIG_SLUB_STATS
1952 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1953 #endif
1956 static void
1957 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1959 n->nr_partial = 0;
1960 spin_lock_init(&n->list_lock);
1961 INIT_LIST_HEAD(&n->partial);
1962 #ifdef CONFIG_SLUB_DEBUG
1963 atomic_long_set(&n->nr_slabs, 0);
1964 atomic_long_set(&n->total_objects, 0);
1965 INIT_LIST_HEAD(&n->full);
1966 #endif
1969 #ifdef CONFIG_SMP
1971 * Per cpu array for per cpu structures.
1973 * The per cpu array places all kmem_cache_cpu structures from one processor
1974 * close together meaning that it becomes possible that multiple per cpu
1975 * structures are contained in one cacheline. This may be particularly
1976 * beneficial for the kmalloc caches.
1978 * A desktop system typically has around 60-80 slabs. With 100 here we are
1979 * likely able to get per cpu structures for all caches from the array defined
1980 * here. We must be able to cover all kmalloc caches during bootstrap.
1982 * If the per cpu array is exhausted then fall back to kmalloc
1983 * of individual cachelines. No sharing is possible then.
1985 #define NR_KMEM_CACHE_CPU 100
1987 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1988 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1990 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1991 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
1993 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1994 int cpu, gfp_t flags)
1996 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1998 if (c)
1999 per_cpu(kmem_cache_cpu_free, cpu) =
2000 (void *)c->freelist;
2001 else {
2002 /* Table overflow: So allocate ourselves */
2003 c = kmalloc_node(
2004 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2005 flags, cpu_to_node(cpu));
2006 if (!c)
2007 return NULL;
2010 init_kmem_cache_cpu(s, c);
2011 return c;
2014 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2016 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2017 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2018 kfree(c);
2019 return;
2021 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2022 per_cpu(kmem_cache_cpu_free, cpu) = c;
2025 static void free_kmem_cache_cpus(struct kmem_cache *s)
2027 int cpu;
2029 for_each_online_cpu(cpu) {
2030 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2032 if (c) {
2033 s->cpu_slab[cpu] = NULL;
2034 free_kmem_cache_cpu(c, cpu);
2039 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2041 int cpu;
2043 for_each_online_cpu(cpu) {
2044 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2046 if (c)
2047 continue;
2049 c = alloc_kmem_cache_cpu(s, cpu, flags);
2050 if (!c) {
2051 free_kmem_cache_cpus(s);
2052 return 0;
2054 s->cpu_slab[cpu] = c;
2056 return 1;
2060 * Initialize the per cpu array.
2062 static void init_alloc_cpu_cpu(int cpu)
2064 int i;
2066 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2067 return;
2069 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2070 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2072 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2075 static void __init init_alloc_cpu(void)
2077 int cpu;
2079 for_each_online_cpu(cpu)
2080 init_alloc_cpu_cpu(cpu);
2083 #else
2084 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2085 static inline void init_alloc_cpu(void) {}
2087 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2089 init_kmem_cache_cpu(s, &s->cpu_slab);
2090 return 1;
2092 #endif
2094 #ifdef CONFIG_NUMA
2096 * No kmalloc_node yet so do it by hand. We know that this is the first
2097 * slab on the node for this slabcache. There are no concurrent accesses
2098 * possible.
2100 * Note that this function only works on the kmalloc_node_cache
2101 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2102 * memory on a fresh node that has no slab structures yet.
2104 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2106 struct page *page;
2107 struct kmem_cache_node *n;
2108 unsigned long flags;
2110 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2112 page = new_slab(kmalloc_caches, gfpflags, node);
2114 BUG_ON(!page);
2115 if (page_to_nid(page) != node) {
2116 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2117 "node %d\n", node);
2118 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2119 "in order to be able to continue\n");
2122 n = page->freelist;
2123 BUG_ON(!n);
2124 page->freelist = get_freepointer(kmalloc_caches, n);
2125 page->inuse++;
2126 kmalloc_caches->node[node] = n;
2127 #ifdef CONFIG_SLUB_DEBUG
2128 init_object(kmalloc_caches, n, 1);
2129 init_tracking(kmalloc_caches, n);
2130 #endif
2131 init_kmem_cache_node(n, kmalloc_caches);
2132 inc_slabs_node(kmalloc_caches, node, page->objects);
2135 * lockdep requires consistent irq usage for each lock
2136 * so even though there cannot be a race this early in
2137 * the boot sequence, we still disable irqs.
2139 local_irq_save(flags);
2140 add_partial(n, page, 0);
2141 local_irq_restore(flags);
2144 static void free_kmem_cache_nodes(struct kmem_cache *s)
2146 int node;
2148 for_each_node_state(node, N_NORMAL_MEMORY) {
2149 struct kmem_cache_node *n = s->node[node];
2150 if (n && n != &s->local_node)
2151 kmem_cache_free(kmalloc_caches, n);
2152 s->node[node] = NULL;
2156 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2158 int node;
2159 int local_node;
2161 if (slab_state >= UP)
2162 local_node = page_to_nid(virt_to_page(s));
2163 else
2164 local_node = 0;
2166 for_each_node_state(node, N_NORMAL_MEMORY) {
2167 struct kmem_cache_node *n;
2169 if (local_node == node)
2170 n = &s->local_node;
2171 else {
2172 if (slab_state == DOWN) {
2173 early_kmem_cache_node_alloc(gfpflags, node);
2174 continue;
2176 n = kmem_cache_alloc_node(kmalloc_caches,
2177 gfpflags, node);
2179 if (!n) {
2180 free_kmem_cache_nodes(s);
2181 return 0;
2185 s->node[node] = n;
2186 init_kmem_cache_node(n, s);
2188 return 1;
2190 #else
2191 static void free_kmem_cache_nodes(struct kmem_cache *s)
2195 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2197 init_kmem_cache_node(&s->local_node, s);
2198 return 1;
2200 #endif
2202 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2204 if (min < MIN_PARTIAL)
2205 min = MIN_PARTIAL;
2206 else if (min > MAX_PARTIAL)
2207 min = MAX_PARTIAL;
2208 s->min_partial = min;
2212 * calculate_sizes() determines the order and the distribution of data within
2213 * a slab object.
2215 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2217 unsigned long flags = s->flags;
2218 unsigned long size = s->objsize;
2219 unsigned long align = s->align;
2220 int order;
2223 * Round up object size to the next word boundary. We can only
2224 * place the free pointer at word boundaries and this determines
2225 * the possible location of the free pointer.
2227 size = ALIGN(size, sizeof(void *));
2229 #ifdef CONFIG_SLUB_DEBUG
2231 * Determine if we can poison the object itself. If the user of
2232 * the slab may touch the object after free or before allocation
2233 * then we should never poison the object itself.
2235 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2236 !s->ctor)
2237 s->flags |= __OBJECT_POISON;
2238 else
2239 s->flags &= ~__OBJECT_POISON;
2243 * If we are Redzoning then check if there is some space between the
2244 * end of the object and the free pointer. If not then add an
2245 * additional word to have some bytes to store Redzone information.
2247 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2248 size += sizeof(void *);
2249 #endif
2252 * With that we have determined the number of bytes in actual use
2253 * by the object. This is the potential offset to the free pointer.
2255 s->inuse = size;
2257 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2258 s->ctor)) {
2260 * Relocate free pointer after the object if it is not
2261 * permitted to overwrite the first word of the object on
2262 * kmem_cache_free.
2264 * This is the case if we do RCU, have a constructor or
2265 * destructor or are poisoning the objects.
2267 s->offset = size;
2268 size += sizeof(void *);
2271 #ifdef CONFIG_SLUB_DEBUG
2272 if (flags & SLAB_STORE_USER)
2274 * Need to store information about allocs and frees after
2275 * the object.
2277 size += 2 * sizeof(struct track);
2279 if (flags & SLAB_RED_ZONE)
2281 * Add some empty padding so that we can catch
2282 * overwrites from earlier objects rather than let
2283 * tracking information or the free pointer be
2284 * corrupted if a user writes before the start
2285 * of the object.
2287 size += sizeof(void *);
2288 #endif
2291 * Determine the alignment based on various parameters that the
2292 * user specified and the dynamic determination of cache line size
2293 * on bootup.
2295 align = calculate_alignment(flags, align, s->objsize);
2298 * SLUB stores one object immediately after another beginning from
2299 * offset 0. In order to align the objects we have to simply size
2300 * each object to conform to the alignment.
2302 size = ALIGN(size, align);
2303 s->size = size;
2304 if (forced_order >= 0)
2305 order = forced_order;
2306 else
2307 order = calculate_order(size);
2309 if (order < 0)
2310 return 0;
2312 s->allocflags = 0;
2313 if (order)
2314 s->allocflags |= __GFP_COMP;
2316 if (s->flags & SLAB_CACHE_DMA)
2317 s->allocflags |= SLUB_DMA;
2319 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2320 s->allocflags |= __GFP_RECLAIMABLE;
2323 * Determine the number of objects per slab
2325 s->oo = oo_make(order, size);
2326 s->min = oo_make(get_order(size), size);
2327 if (oo_objects(s->oo) > oo_objects(s->max))
2328 s->max = s->oo;
2330 return !!oo_objects(s->oo);
2334 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2335 const char *name, size_t size,
2336 size_t align, unsigned long flags,
2337 void (*ctor)(void *))
2339 memset(s, 0, kmem_size);
2340 s->name = name;
2341 s->ctor = ctor;
2342 s->objsize = size;
2343 s->align = align;
2344 s->flags = kmem_cache_flags(size, flags, name, ctor);
2346 if (!calculate_sizes(s, -1))
2347 goto error;
2350 * The larger the object size is, the more pages we want on the partial
2351 * list to avoid pounding the page allocator excessively.
2353 set_min_partial(s, ilog2(s->size));
2354 s->refcount = 1;
2355 #ifdef CONFIG_NUMA
2356 s->remote_node_defrag_ratio = 1000;
2357 #endif
2358 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2359 goto error;
2361 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2362 return 1;
2363 free_kmem_cache_nodes(s);
2364 error:
2365 if (flags & SLAB_PANIC)
2366 panic("Cannot create slab %s size=%lu realsize=%u "
2367 "order=%u offset=%u flags=%lx\n",
2368 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2369 s->offset, flags);
2370 return 0;
2374 * Check if a given pointer is valid
2376 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2378 struct page *page;
2380 page = get_object_page(object);
2382 if (!page || s != page->slab)
2383 /* No slab or wrong slab */
2384 return 0;
2386 if (!check_valid_pointer(s, page, object))
2387 return 0;
2390 * We could also check if the object is on the slabs freelist.
2391 * But this would be too expensive and it seems that the main
2392 * purpose of kmem_ptr_valid() is to check if the object belongs
2393 * to a certain slab.
2395 return 1;
2397 EXPORT_SYMBOL(kmem_ptr_validate);
2400 * Determine the size of a slab object
2402 unsigned int kmem_cache_size(struct kmem_cache *s)
2404 return s->objsize;
2406 EXPORT_SYMBOL(kmem_cache_size);
2408 const char *kmem_cache_name(struct kmem_cache *s)
2410 return s->name;
2412 EXPORT_SYMBOL(kmem_cache_name);
2414 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2415 const char *text)
2417 #ifdef CONFIG_SLUB_DEBUG
2418 void *addr = page_address(page);
2419 void *p;
2420 DECLARE_BITMAP(map, page->objects);
2422 bitmap_zero(map, page->objects);
2423 slab_err(s, page, "%s", text);
2424 slab_lock(page);
2425 for_each_free_object(p, s, page->freelist)
2426 set_bit(slab_index(p, s, addr), map);
2428 for_each_object(p, s, addr, page->objects) {
2430 if (!test_bit(slab_index(p, s, addr), map)) {
2431 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2432 p, p - addr);
2433 print_tracking(s, p);
2436 slab_unlock(page);
2437 #endif
2441 * Attempt to free all partial slabs on a node.
2443 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2445 unsigned long flags;
2446 struct page *page, *h;
2448 spin_lock_irqsave(&n->list_lock, flags);
2449 list_for_each_entry_safe(page, h, &n->partial, lru) {
2450 if (!page->inuse) {
2451 list_del(&page->lru);
2452 discard_slab(s, page);
2453 n->nr_partial--;
2454 } else {
2455 list_slab_objects(s, page,
2456 "Objects remaining on kmem_cache_close()");
2459 spin_unlock_irqrestore(&n->list_lock, flags);
2463 * Release all resources used by a slab cache.
2465 static inline int kmem_cache_close(struct kmem_cache *s)
2467 int node;
2469 flush_all(s);
2471 /* Attempt to free all objects */
2472 free_kmem_cache_cpus(s);
2473 for_each_node_state(node, N_NORMAL_MEMORY) {
2474 struct kmem_cache_node *n = get_node(s, node);
2476 free_partial(s, n);
2477 if (n->nr_partial || slabs_node(s, node))
2478 return 1;
2480 free_kmem_cache_nodes(s);
2481 return 0;
2485 * Close a cache and release the kmem_cache structure
2486 * (must be used for caches created using kmem_cache_create)
2488 void kmem_cache_destroy(struct kmem_cache *s)
2490 down_write(&slub_lock);
2491 s->refcount--;
2492 if (!s->refcount) {
2493 list_del(&s->list);
2494 up_write(&slub_lock);
2495 if (kmem_cache_close(s)) {
2496 printk(KERN_ERR "SLUB %s: %s called for cache that "
2497 "still has objects.\n", s->name, __func__);
2498 dump_stack();
2500 sysfs_slab_remove(s);
2501 } else
2502 up_write(&slub_lock);
2504 EXPORT_SYMBOL(kmem_cache_destroy);
2506 /********************************************************************
2507 * Kmalloc subsystem
2508 *******************************************************************/
2510 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2511 EXPORT_SYMBOL(kmalloc_caches);
2513 static int __init setup_slub_min_order(char *str)
2515 get_option(&str, &slub_min_order);
2517 return 1;
2520 __setup("slub_min_order=", setup_slub_min_order);
2522 static int __init setup_slub_max_order(char *str)
2524 get_option(&str, &slub_max_order);
2526 return 1;
2529 __setup("slub_max_order=", setup_slub_max_order);
2531 static int __init setup_slub_min_objects(char *str)
2533 get_option(&str, &slub_min_objects);
2535 return 1;
2538 __setup("slub_min_objects=", setup_slub_min_objects);
2540 static int __init setup_slub_nomerge(char *str)
2542 slub_nomerge = 1;
2543 return 1;
2546 __setup("slub_nomerge", setup_slub_nomerge);
2548 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2549 const char *name, int size, gfp_t gfp_flags)
2551 unsigned int flags = 0;
2553 if (gfp_flags & SLUB_DMA)
2554 flags = SLAB_CACHE_DMA;
2556 down_write(&slub_lock);
2557 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2558 flags, NULL))
2559 goto panic;
2561 list_add(&s->list, &slab_caches);
2562 up_write(&slub_lock);
2563 if (sysfs_slab_add(s))
2564 goto panic;
2565 return s;
2567 panic:
2568 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2571 #ifdef CONFIG_ZONE_DMA
2572 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2574 static void sysfs_add_func(struct work_struct *w)
2576 struct kmem_cache *s;
2578 down_write(&slub_lock);
2579 list_for_each_entry(s, &slab_caches, list) {
2580 if (s->flags & __SYSFS_ADD_DEFERRED) {
2581 s->flags &= ~__SYSFS_ADD_DEFERRED;
2582 sysfs_slab_add(s);
2585 up_write(&slub_lock);
2588 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2590 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2592 struct kmem_cache *s;
2593 char *text;
2594 size_t realsize;
2596 s = kmalloc_caches_dma[index];
2597 if (s)
2598 return s;
2600 /* Dynamically create dma cache */
2601 if (flags & __GFP_WAIT)
2602 down_write(&slub_lock);
2603 else {
2604 if (!down_write_trylock(&slub_lock))
2605 goto out;
2608 if (kmalloc_caches_dma[index])
2609 goto unlock_out;
2611 realsize = kmalloc_caches[index].objsize;
2612 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2613 (unsigned int)realsize);
2614 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2616 if (!s || !text || !kmem_cache_open(s, flags, text,
2617 realsize, ARCH_KMALLOC_MINALIGN,
2618 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2619 kfree(s);
2620 kfree(text);
2621 goto unlock_out;
2624 list_add(&s->list, &slab_caches);
2625 kmalloc_caches_dma[index] = s;
2627 schedule_work(&sysfs_add_work);
2629 unlock_out:
2630 up_write(&slub_lock);
2631 out:
2632 return kmalloc_caches_dma[index];
2634 #endif
2637 * Conversion table for small slabs sizes / 8 to the index in the
2638 * kmalloc array. This is necessary for slabs < 192 since we have non power
2639 * of two cache sizes there. The size of larger slabs can be determined using
2640 * fls.
2642 static s8 size_index[24] = {
2643 3, /* 8 */
2644 4, /* 16 */
2645 5, /* 24 */
2646 5, /* 32 */
2647 6, /* 40 */
2648 6, /* 48 */
2649 6, /* 56 */
2650 6, /* 64 */
2651 1, /* 72 */
2652 1, /* 80 */
2653 1, /* 88 */
2654 1, /* 96 */
2655 7, /* 104 */
2656 7, /* 112 */
2657 7, /* 120 */
2658 7, /* 128 */
2659 2, /* 136 */
2660 2, /* 144 */
2661 2, /* 152 */
2662 2, /* 160 */
2663 2, /* 168 */
2664 2, /* 176 */
2665 2, /* 184 */
2666 2 /* 192 */
2669 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2671 int index;
2673 if (size <= 192) {
2674 if (!size)
2675 return ZERO_SIZE_PTR;
2677 index = size_index[(size - 1) / 8];
2678 } else
2679 index = fls(size - 1);
2681 #ifdef CONFIG_ZONE_DMA
2682 if (unlikely((flags & SLUB_DMA)))
2683 return dma_kmalloc_cache(index, flags);
2685 #endif
2686 return &kmalloc_caches[index];
2689 void *__kmalloc(size_t size, gfp_t flags)
2691 struct kmem_cache *s;
2692 void *ret;
2694 if (unlikely(size > SLUB_MAX_SIZE))
2695 return kmalloc_large(size, flags);
2697 s = get_slab(size, flags);
2699 if (unlikely(ZERO_OR_NULL_PTR(s)))
2700 return s;
2702 ret = slab_alloc(s, flags, -1, _RET_IP_);
2704 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2706 return ret;
2708 EXPORT_SYMBOL(__kmalloc);
2710 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2712 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2713 get_order(size));
2715 if (page)
2716 return page_address(page);
2717 else
2718 return NULL;
2721 #ifdef CONFIG_NUMA
2722 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2724 struct kmem_cache *s;
2725 void *ret;
2727 if (unlikely(size > SLUB_MAX_SIZE)) {
2728 ret = kmalloc_large_node(size, flags, node);
2730 trace_kmalloc_node(_RET_IP_, ret,
2731 size, PAGE_SIZE << get_order(size),
2732 flags, node);
2734 return ret;
2737 s = get_slab(size, flags);
2739 if (unlikely(ZERO_OR_NULL_PTR(s)))
2740 return s;
2742 ret = slab_alloc(s, flags, node, _RET_IP_);
2744 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2746 return ret;
2748 EXPORT_SYMBOL(__kmalloc_node);
2749 #endif
2751 size_t ksize(const void *object)
2753 struct page *page;
2754 struct kmem_cache *s;
2756 if (unlikely(object == ZERO_SIZE_PTR))
2757 return 0;
2759 page = virt_to_head_page(object);
2761 if (unlikely(!PageSlab(page))) {
2762 WARN_ON(!PageCompound(page));
2763 return PAGE_SIZE << compound_order(page);
2765 s = page->slab;
2767 #ifdef CONFIG_SLUB_DEBUG
2769 * Debugging requires use of the padding between object
2770 * and whatever may come after it.
2772 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2773 return s->objsize;
2775 #endif
2777 * If we have the need to store the freelist pointer
2778 * back there or track user information then we can
2779 * only use the space before that information.
2781 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2782 return s->inuse;
2784 * Else we can use all the padding etc for the allocation
2786 return s->size;
2788 EXPORT_SYMBOL(ksize);
2790 void kfree(const void *x)
2792 struct page *page;
2793 void *object = (void *)x;
2795 if (unlikely(ZERO_OR_NULL_PTR(x)))
2796 return;
2798 page = virt_to_head_page(x);
2799 if (unlikely(!PageSlab(page))) {
2800 BUG_ON(!PageCompound(page));
2801 put_page(page);
2802 return;
2804 slab_free(page->slab, page, object, _RET_IP_);
2806 trace_kfree(_RET_IP_, x);
2808 EXPORT_SYMBOL(kfree);
2811 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2812 * the remaining slabs by the number of items in use. The slabs with the
2813 * most items in use come first. New allocations will then fill those up
2814 * and thus they can be removed from the partial lists.
2816 * The slabs with the least items are placed last. This results in them
2817 * being allocated from last increasing the chance that the last objects
2818 * are freed in them.
2820 int kmem_cache_shrink(struct kmem_cache *s)
2822 int node;
2823 int i;
2824 struct kmem_cache_node *n;
2825 struct page *page;
2826 struct page *t;
2827 int objects = oo_objects(s->max);
2828 struct list_head *slabs_by_inuse =
2829 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2830 unsigned long flags;
2832 if (!slabs_by_inuse)
2833 return -ENOMEM;
2835 flush_all(s);
2836 for_each_node_state(node, N_NORMAL_MEMORY) {
2837 n = get_node(s, node);
2839 if (!n->nr_partial)
2840 continue;
2842 for (i = 0; i < objects; i++)
2843 INIT_LIST_HEAD(slabs_by_inuse + i);
2845 spin_lock_irqsave(&n->list_lock, flags);
2848 * Build lists indexed by the items in use in each slab.
2850 * Note that concurrent frees may occur while we hold the
2851 * list_lock. page->inuse here is the upper limit.
2853 list_for_each_entry_safe(page, t, &n->partial, lru) {
2854 if (!page->inuse && slab_trylock(page)) {
2856 * Must hold slab lock here because slab_free
2857 * may have freed the last object and be
2858 * waiting to release the slab.
2860 list_del(&page->lru);
2861 n->nr_partial--;
2862 slab_unlock(page);
2863 discard_slab(s, page);
2864 } else {
2865 list_move(&page->lru,
2866 slabs_by_inuse + page->inuse);
2871 * Rebuild the partial list with the slabs filled up most
2872 * first and the least used slabs at the end.
2874 for (i = objects - 1; i >= 0; i--)
2875 list_splice(slabs_by_inuse + i, n->partial.prev);
2877 spin_unlock_irqrestore(&n->list_lock, flags);
2880 kfree(slabs_by_inuse);
2881 return 0;
2883 EXPORT_SYMBOL(kmem_cache_shrink);
2885 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2886 static int slab_mem_going_offline_callback(void *arg)
2888 struct kmem_cache *s;
2890 down_read(&slub_lock);
2891 list_for_each_entry(s, &slab_caches, list)
2892 kmem_cache_shrink(s);
2893 up_read(&slub_lock);
2895 return 0;
2898 static void slab_mem_offline_callback(void *arg)
2900 struct kmem_cache_node *n;
2901 struct kmem_cache *s;
2902 struct memory_notify *marg = arg;
2903 int offline_node;
2905 offline_node = marg->status_change_nid;
2908 * If the node still has available memory. we need kmem_cache_node
2909 * for it yet.
2911 if (offline_node < 0)
2912 return;
2914 down_read(&slub_lock);
2915 list_for_each_entry(s, &slab_caches, list) {
2916 n = get_node(s, offline_node);
2917 if (n) {
2919 * if n->nr_slabs > 0, slabs still exist on the node
2920 * that is going down. We were unable to free them,
2921 * and offline_pages() function shoudn't call this
2922 * callback. So, we must fail.
2924 BUG_ON(slabs_node(s, offline_node));
2926 s->node[offline_node] = NULL;
2927 kmem_cache_free(kmalloc_caches, n);
2930 up_read(&slub_lock);
2933 static int slab_mem_going_online_callback(void *arg)
2935 struct kmem_cache_node *n;
2936 struct kmem_cache *s;
2937 struct memory_notify *marg = arg;
2938 int nid = marg->status_change_nid;
2939 int ret = 0;
2942 * If the node's memory is already available, then kmem_cache_node is
2943 * already created. Nothing to do.
2945 if (nid < 0)
2946 return 0;
2949 * We are bringing a node online. No memory is available yet. We must
2950 * allocate a kmem_cache_node structure in order to bring the node
2951 * online.
2953 down_read(&slub_lock);
2954 list_for_each_entry(s, &slab_caches, list) {
2956 * XXX: kmem_cache_alloc_node will fallback to other nodes
2957 * since memory is not yet available from the node that
2958 * is brought up.
2960 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2961 if (!n) {
2962 ret = -ENOMEM;
2963 goto out;
2965 init_kmem_cache_node(n, s);
2966 s->node[nid] = n;
2968 out:
2969 up_read(&slub_lock);
2970 return ret;
2973 static int slab_memory_callback(struct notifier_block *self,
2974 unsigned long action, void *arg)
2976 int ret = 0;
2978 switch (action) {
2979 case MEM_GOING_ONLINE:
2980 ret = slab_mem_going_online_callback(arg);
2981 break;
2982 case MEM_GOING_OFFLINE:
2983 ret = slab_mem_going_offline_callback(arg);
2984 break;
2985 case MEM_OFFLINE:
2986 case MEM_CANCEL_ONLINE:
2987 slab_mem_offline_callback(arg);
2988 break;
2989 case MEM_ONLINE:
2990 case MEM_CANCEL_OFFLINE:
2991 break;
2993 if (ret)
2994 ret = notifier_from_errno(ret);
2995 else
2996 ret = NOTIFY_OK;
2997 return ret;
3000 #endif /* CONFIG_MEMORY_HOTPLUG */
3002 /********************************************************************
3003 * Basic setup of slabs
3004 *******************************************************************/
3006 void __init kmem_cache_init(void)
3008 int i;
3009 int caches = 0;
3011 init_alloc_cpu();
3013 #ifdef CONFIG_NUMA
3015 * Must first have the slab cache available for the allocations of the
3016 * struct kmem_cache_node's. There is special bootstrap code in
3017 * kmem_cache_open for slab_state == DOWN.
3019 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3020 sizeof(struct kmem_cache_node), GFP_KERNEL);
3021 kmalloc_caches[0].refcount = -1;
3022 caches++;
3024 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3025 #endif
3027 /* Able to allocate the per node structures */
3028 slab_state = PARTIAL;
3030 /* Caches that are not of the two-to-the-power-of size */
3031 if (KMALLOC_MIN_SIZE <= 64) {
3032 create_kmalloc_cache(&kmalloc_caches[1],
3033 "kmalloc-96", 96, GFP_KERNEL);
3034 caches++;
3035 create_kmalloc_cache(&kmalloc_caches[2],
3036 "kmalloc-192", 192, GFP_KERNEL);
3037 caches++;
3040 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3041 create_kmalloc_cache(&kmalloc_caches[i],
3042 "kmalloc", 1 << i, GFP_KERNEL);
3043 caches++;
3048 * Patch up the size_index table if we have strange large alignment
3049 * requirements for the kmalloc array. This is only the case for
3050 * MIPS it seems. The standard arches will not generate any code here.
3052 * Largest permitted alignment is 256 bytes due to the way we
3053 * handle the index determination for the smaller caches.
3055 * Make sure that nothing crazy happens if someone starts tinkering
3056 * around with ARCH_KMALLOC_MINALIGN
3058 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3059 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3061 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3062 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3064 if (KMALLOC_MIN_SIZE == 128) {
3066 * The 192 byte sized cache is not used if the alignment
3067 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3068 * instead.
3070 for (i = 128 + 8; i <= 192; i += 8)
3071 size_index[(i - 1) / 8] = 8;
3074 slab_state = UP;
3076 /* Provide the correct kmalloc names now that the caches are up */
3077 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3078 kmalloc_caches[i]. name =
3079 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3081 #ifdef CONFIG_SMP
3082 register_cpu_notifier(&slab_notifier);
3083 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3084 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3085 #else
3086 kmem_size = sizeof(struct kmem_cache);
3087 #endif
3089 printk(KERN_INFO
3090 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3091 " CPUs=%d, Nodes=%d\n",
3092 caches, cache_line_size(),
3093 slub_min_order, slub_max_order, slub_min_objects,
3094 nr_cpu_ids, nr_node_ids);
3098 * Find a mergeable slab cache
3100 static int slab_unmergeable(struct kmem_cache *s)
3102 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3103 return 1;
3105 if (s->ctor)
3106 return 1;
3109 * We may have set a slab to be unmergeable during bootstrap.
3111 if (s->refcount < 0)
3112 return 1;
3114 return 0;
3117 static struct kmem_cache *find_mergeable(size_t size,
3118 size_t align, unsigned long flags, const char *name,
3119 void (*ctor)(void *))
3121 struct kmem_cache *s;
3123 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3124 return NULL;
3126 if (ctor)
3127 return NULL;
3129 size = ALIGN(size, sizeof(void *));
3130 align = calculate_alignment(flags, align, size);
3131 size = ALIGN(size, align);
3132 flags = kmem_cache_flags(size, flags, name, NULL);
3134 list_for_each_entry(s, &slab_caches, list) {
3135 if (slab_unmergeable(s))
3136 continue;
3138 if (size > s->size)
3139 continue;
3141 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3142 continue;
3144 * Check if alignment is compatible.
3145 * Courtesy of Adrian Drzewiecki
3147 if ((s->size & ~(align - 1)) != s->size)
3148 continue;
3150 if (s->size - size >= sizeof(void *))
3151 continue;
3153 return s;
3155 return NULL;
3158 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3159 size_t align, unsigned long flags, void (*ctor)(void *))
3161 struct kmem_cache *s;
3163 down_write(&slub_lock);
3164 s = find_mergeable(size, align, flags, name, ctor);
3165 if (s) {
3166 int cpu;
3168 s->refcount++;
3170 * Adjust the object sizes so that we clear
3171 * the complete object on kzalloc.
3173 s->objsize = max(s->objsize, (int)size);
3176 * And then we need to update the object size in the
3177 * per cpu structures
3179 for_each_online_cpu(cpu)
3180 get_cpu_slab(s, cpu)->objsize = s->objsize;
3182 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3183 up_write(&slub_lock);
3185 if (sysfs_slab_alias(s, name)) {
3186 down_write(&slub_lock);
3187 s->refcount--;
3188 up_write(&slub_lock);
3189 goto err;
3191 return s;
3194 s = kmalloc(kmem_size, GFP_KERNEL);
3195 if (s) {
3196 if (kmem_cache_open(s, GFP_KERNEL, name,
3197 size, align, flags, ctor)) {
3198 list_add(&s->list, &slab_caches);
3199 up_write(&slub_lock);
3200 if (sysfs_slab_add(s)) {
3201 down_write(&slub_lock);
3202 list_del(&s->list);
3203 up_write(&slub_lock);
3204 kfree(s);
3205 goto err;
3207 return s;
3209 kfree(s);
3211 up_write(&slub_lock);
3213 err:
3214 if (flags & SLAB_PANIC)
3215 panic("Cannot create slabcache %s\n", name);
3216 else
3217 s = NULL;
3218 return s;
3220 EXPORT_SYMBOL(kmem_cache_create);
3222 #ifdef CONFIG_SMP
3224 * Use the cpu notifier to insure that the cpu slabs are flushed when
3225 * necessary.
3227 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3228 unsigned long action, void *hcpu)
3230 long cpu = (long)hcpu;
3231 struct kmem_cache *s;
3232 unsigned long flags;
3234 switch (action) {
3235 case CPU_UP_PREPARE:
3236 case CPU_UP_PREPARE_FROZEN:
3237 init_alloc_cpu_cpu(cpu);
3238 down_read(&slub_lock);
3239 list_for_each_entry(s, &slab_caches, list)
3240 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3241 GFP_KERNEL);
3242 up_read(&slub_lock);
3243 break;
3245 case CPU_UP_CANCELED:
3246 case CPU_UP_CANCELED_FROZEN:
3247 case CPU_DEAD:
3248 case CPU_DEAD_FROZEN:
3249 down_read(&slub_lock);
3250 list_for_each_entry(s, &slab_caches, list) {
3251 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3253 local_irq_save(flags);
3254 __flush_cpu_slab(s, cpu);
3255 local_irq_restore(flags);
3256 free_kmem_cache_cpu(c, cpu);
3257 s->cpu_slab[cpu] = NULL;
3259 up_read(&slub_lock);
3260 break;
3261 default:
3262 break;
3264 return NOTIFY_OK;
3267 static struct notifier_block __cpuinitdata slab_notifier = {
3268 .notifier_call = slab_cpuup_callback
3271 #endif
3273 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3275 struct kmem_cache *s;
3276 void *ret;
3278 if (unlikely(size > SLUB_MAX_SIZE))
3279 return kmalloc_large(size, gfpflags);
3281 s = get_slab(size, gfpflags);
3283 if (unlikely(ZERO_OR_NULL_PTR(s)))
3284 return s;
3286 ret = slab_alloc(s, gfpflags, -1, caller);
3288 /* Honor the call site pointer we recieved. */
3289 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3291 return ret;
3294 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3295 int node, unsigned long caller)
3297 struct kmem_cache *s;
3298 void *ret;
3300 if (unlikely(size > SLUB_MAX_SIZE))
3301 return kmalloc_large_node(size, gfpflags, node);
3303 s = get_slab(size, gfpflags);
3305 if (unlikely(ZERO_OR_NULL_PTR(s)))
3306 return s;
3308 ret = slab_alloc(s, gfpflags, node, caller);
3310 /* Honor the call site pointer we recieved. */
3311 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3313 return ret;
3316 #ifdef CONFIG_SLUB_DEBUG
3317 static unsigned long count_partial(struct kmem_cache_node *n,
3318 int (*get_count)(struct page *))
3320 unsigned long flags;
3321 unsigned long x = 0;
3322 struct page *page;
3324 spin_lock_irqsave(&n->list_lock, flags);
3325 list_for_each_entry(page, &n->partial, lru)
3326 x += get_count(page);
3327 spin_unlock_irqrestore(&n->list_lock, flags);
3328 return x;
3331 static int count_inuse(struct page *page)
3333 return page->inuse;
3336 static int count_total(struct page *page)
3338 return page->objects;
3341 static int count_free(struct page *page)
3343 return page->objects - page->inuse;
3346 static int validate_slab(struct kmem_cache *s, struct page *page,
3347 unsigned long *map)
3349 void *p;
3350 void *addr = page_address(page);
3352 if (!check_slab(s, page) ||
3353 !on_freelist(s, page, NULL))
3354 return 0;
3356 /* Now we know that a valid freelist exists */
3357 bitmap_zero(map, page->objects);
3359 for_each_free_object(p, s, page->freelist) {
3360 set_bit(slab_index(p, s, addr), map);
3361 if (!check_object(s, page, p, 0))
3362 return 0;
3365 for_each_object(p, s, addr, page->objects)
3366 if (!test_bit(slab_index(p, s, addr), map))
3367 if (!check_object(s, page, p, 1))
3368 return 0;
3369 return 1;
3372 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3373 unsigned long *map)
3375 if (slab_trylock(page)) {
3376 validate_slab(s, page, map);
3377 slab_unlock(page);
3378 } else
3379 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3380 s->name, page);
3382 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3383 if (!PageSlubDebug(page))
3384 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3385 "on slab 0x%p\n", s->name, page);
3386 } else {
3387 if (PageSlubDebug(page))
3388 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3389 "slab 0x%p\n", s->name, page);
3393 static int validate_slab_node(struct kmem_cache *s,
3394 struct kmem_cache_node *n, unsigned long *map)
3396 unsigned long count = 0;
3397 struct page *page;
3398 unsigned long flags;
3400 spin_lock_irqsave(&n->list_lock, flags);
3402 list_for_each_entry(page, &n->partial, lru) {
3403 validate_slab_slab(s, page, map);
3404 count++;
3406 if (count != n->nr_partial)
3407 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3408 "counter=%ld\n", s->name, count, n->nr_partial);
3410 if (!(s->flags & SLAB_STORE_USER))
3411 goto out;
3413 list_for_each_entry(page, &n->full, lru) {
3414 validate_slab_slab(s, page, map);
3415 count++;
3417 if (count != atomic_long_read(&n->nr_slabs))
3418 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3419 "counter=%ld\n", s->name, count,
3420 atomic_long_read(&n->nr_slabs));
3422 out:
3423 spin_unlock_irqrestore(&n->list_lock, flags);
3424 return count;
3427 static long validate_slab_cache(struct kmem_cache *s)
3429 int node;
3430 unsigned long count = 0;
3431 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3432 sizeof(unsigned long), GFP_KERNEL);
3434 if (!map)
3435 return -ENOMEM;
3437 flush_all(s);
3438 for_each_node_state(node, N_NORMAL_MEMORY) {
3439 struct kmem_cache_node *n = get_node(s, node);
3441 count += validate_slab_node(s, n, map);
3443 kfree(map);
3444 return count;
3447 #ifdef SLUB_RESILIENCY_TEST
3448 static void resiliency_test(void)
3450 u8 *p;
3452 printk(KERN_ERR "SLUB resiliency testing\n");
3453 printk(KERN_ERR "-----------------------\n");
3454 printk(KERN_ERR "A. Corruption after allocation\n");
3456 p = kzalloc(16, GFP_KERNEL);
3457 p[16] = 0x12;
3458 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3459 " 0x12->0x%p\n\n", p + 16);
3461 validate_slab_cache(kmalloc_caches + 4);
3463 /* Hmmm... The next two are dangerous */
3464 p = kzalloc(32, GFP_KERNEL);
3465 p[32 + sizeof(void *)] = 0x34;
3466 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3467 " 0x34 -> -0x%p\n", p);
3468 printk(KERN_ERR
3469 "If allocated object is overwritten then not detectable\n\n");
3471 validate_slab_cache(kmalloc_caches + 5);
3472 p = kzalloc(64, GFP_KERNEL);
3473 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3474 *p = 0x56;
3475 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3477 printk(KERN_ERR
3478 "If allocated object is overwritten then not detectable\n\n");
3479 validate_slab_cache(kmalloc_caches + 6);
3481 printk(KERN_ERR "\nB. Corruption after free\n");
3482 p = kzalloc(128, GFP_KERNEL);
3483 kfree(p);
3484 *p = 0x78;
3485 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3486 validate_slab_cache(kmalloc_caches + 7);
3488 p = kzalloc(256, GFP_KERNEL);
3489 kfree(p);
3490 p[50] = 0x9a;
3491 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3493 validate_slab_cache(kmalloc_caches + 8);
3495 p = kzalloc(512, GFP_KERNEL);
3496 kfree(p);
3497 p[512] = 0xab;
3498 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3499 validate_slab_cache(kmalloc_caches + 9);
3501 #else
3502 static void resiliency_test(void) {};
3503 #endif
3506 * Generate lists of code addresses where slabcache objects are allocated
3507 * and freed.
3510 struct location {
3511 unsigned long count;
3512 unsigned long addr;
3513 long long sum_time;
3514 long min_time;
3515 long max_time;
3516 long min_pid;
3517 long max_pid;
3518 DECLARE_BITMAP(cpus, NR_CPUS);
3519 nodemask_t nodes;
3522 struct loc_track {
3523 unsigned long max;
3524 unsigned long count;
3525 struct location *loc;
3528 static void free_loc_track(struct loc_track *t)
3530 if (t->max)
3531 free_pages((unsigned long)t->loc,
3532 get_order(sizeof(struct location) * t->max));
3535 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3537 struct location *l;
3538 int order;
3540 order = get_order(sizeof(struct location) * max);
3542 l = (void *)__get_free_pages(flags, order);
3543 if (!l)
3544 return 0;
3546 if (t->count) {
3547 memcpy(l, t->loc, sizeof(struct location) * t->count);
3548 free_loc_track(t);
3550 t->max = max;
3551 t->loc = l;
3552 return 1;
3555 static int add_location(struct loc_track *t, struct kmem_cache *s,
3556 const struct track *track)
3558 long start, end, pos;
3559 struct location *l;
3560 unsigned long caddr;
3561 unsigned long age = jiffies - track->when;
3563 start = -1;
3564 end = t->count;
3566 for ( ; ; ) {
3567 pos = start + (end - start + 1) / 2;
3570 * There is nothing at "end". If we end up there
3571 * we need to add something to before end.
3573 if (pos == end)
3574 break;
3576 caddr = t->loc[pos].addr;
3577 if (track->addr == caddr) {
3579 l = &t->loc[pos];
3580 l->count++;
3581 if (track->when) {
3582 l->sum_time += age;
3583 if (age < l->min_time)
3584 l->min_time = age;
3585 if (age > l->max_time)
3586 l->max_time = age;
3588 if (track->pid < l->min_pid)
3589 l->min_pid = track->pid;
3590 if (track->pid > l->max_pid)
3591 l->max_pid = track->pid;
3593 cpumask_set_cpu(track->cpu,
3594 to_cpumask(l->cpus));
3596 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3597 return 1;
3600 if (track->addr < caddr)
3601 end = pos;
3602 else
3603 start = pos;
3607 * Not found. Insert new tracking element.
3609 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3610 return 0;
3612 l = t->loc + pos;
3613 if (pos < t->count)
3614 memmove(l + 1, l,
3615 (t->count - pos) * sizeof(struct location));
3616 t->count++;
3617 l->count = 1;
3618 l->addr = track->addr;
3619 l->sum_time = age;
3620 l->min_time = age;
3621 l->max_time = age;
3622 l->min_pid = track->pid;
3623 l->max_pid = track->pid;
3624 cpumask_clear(to_cpumask(l->cpus));
3625 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3626 nodes_clear(l->nodes);
3627 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3628 return 1;
3631 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3632 struct page *page, enum track_item alloc)
3634 void *addr = page_address(page);
3635 DECLARE_BITMAP(map, page->objects);
3636 void *p;
3638 bitmap_zero(map, page->objects);
3639 for_each_free_object(p, s, page->freelist)
3640 set_bit(slab_index(p, s, addr), map);
3642 for_each_object(p, s, addr, page->objects)
3643 if (!test_bit(slab_index(p, s, addr), map))
3644 add_location(t, s, get_track(s, p, alloc));
3647 static int list_locations(struct kmem_cache *s, char *buf,
3648 enum track_item alloc)
3650 int len = 0;
3651 unsigned long i;
3652 struct loc_track t = { 0, 0, NULL };
3653 int node;
3655 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3656 GFP_TEMPORARY))
3657 return sprintf(buf, "Out of memory\n");
3659 /* Push back cpu slabs */
3660 flush_all(s);
3662 for_each_node_state(node, N_NORMAL_MEMORY) {
3663 struct kmem_cache_node *n = get_node(s, node);
3664 unsigned long flags;
3665 struct page *page;
3667 if (!atomic_long_read(&n->nr_slabs))
3668 continue;
3670 spin_lock_irqsave(&n->list_lock, flags);
3671 list_for_each_entry(page, &n->partial, lru)
3672 process_slab(&t, s, page, alloc);
3673 list_for_each_entry(page, &n->full, lru)
3674 process_slab(&t, s, page, alloc);
3675 spin_unlock_irqrestore(&n->list_lock, flags);
3678 for (i = 0; i < t.count; i++) {
3679 struct location *l = &t.loc[i];
3681 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3682 break;
3683 len += sprintf(buf + len, "%7ld ", l->count);
3685 if (l->addr)
3686 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3687 else
3688 len += sprintf(buf + len, "<not-available>");
3690 if (l->sum_time != l->min_time) {
3691 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3692 l->min_time,
3693 (long)div_u64(l->sum_time, l->count),
3694 l->max_time);
3695 } else
3696 len += sprintf(buf + len, " age=%ld",
3697 l->min_time);
3699 if (l->min_pid != l->max_pid)
3700 len += sprintf(buf + len, " pid=%ld-%ld",
3701 l->min_pid, l->max_pid);
3702 else
3703 len += sprintf(buf + len, " pid=%ld",
3704 l->min_pid);
3706 if (num_online_cpus() > 1 &&
3707 !cpumask_empty(to_cpumask(l->cpus)) &&
3708 len < PAGE_SIZE - 60) {
3709 len += sprintf(buf + len, " cpus=");
3710 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3711 to_cpumask(l->cpus));
3714 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3715 len < PAGE_SIZE - 60) {
3716 len += sprintf(buf + len, " nodes=");
3717 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3718 l->nodes);
3721 len += sprintf(buf + len, "\n");
3724 free_loc_track(&t);
3725 if (!t.count)
3726 len += sprintf(buf, "No data\n");
3727 return len;
3730 enum slab_stat_type {
3731 SL_ALL, /* All slabs */
3732 SL_PARTIAL, /* Only partially allocated slabs */
3733 SL_CPU, /* Only slabs used for cpu caches */
3734 SL_OBJECTS, /* Determine allocated objects not slabs */
3735 SL_TOTAL /* Determine object capacity not slabs */
3738 #define SO_ALL (1 << SL_ALL)
3739 #define SO_PARTIAL (1 << SL_PARTIAL)
3740 #define SO_CPU (1 << SL_CPU)
3741 #define SO_OBJECTS (1 << SL_OBJECTS)
3742 #define SO_TOTAL (1 << SL_TOTAL)
3744 static ssize_t show_slab_objects(struct kmem_cache *s,
3745 char *buf, unsigned long flags)
3747 unsigned long total = 0;
3748 int node;
3749 int x;
3750 unsigned long *nodes;
3751 unsigned long *per_cpu;
3753 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3754 if (!nodes)
3755 return -ENOMEM;
3756 per_cpu = nodes + nr_node_ids;
3758 if (flags & SO_CPU) {
3759 int cpu;
3761 for_each_possible_cpu(cpu) {
3762 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3764 if (!c || c->node < 0)
3765 continue;
3767 if (c->page) {
3768 if (flags & SO_TOTAL)
3769 x = c->page->objects;
3770 else if (flags & SO_OBJECTS)
3771 x = c->page->inuse;
3772 else
3773 x = 1;
3775 total += x;
3776 nodes[c->node] += x;
3778 per_cpu[c->node]++;
3782 if (flags & SO_ALL) {
3783 for_each_node_state(node, N_NORMAL_MEMORY) {
3784 struct kmem_cache_node *n = get_node(s, node);
3786 if (flags & SO_TOTAL)
3787 x = atomic_long_read(&n->total_objects);
3788 else if (flags & SO_OBJECTS)
3789 x = atomic_long_read(&n->total_objects) -
3790 count_partial(n, count_free);
3792 else
3793 x = atomic_long_read(&n->nr_slabs);
3794 total += x;
3795 nodes[node] += x;
3798 } else if (flags & SO_PARTIAL) {
3799 for_each_node_state(node, N_NORMAL_MEMORY) {
3800 struct kmem_cache_node *n = get_node(s, node);
3802 if (flags & SO_TOTAL)
3803 x = count_partial(n, count_total);
3804 else if (flags & SO_OBJECTS)
3805 x = count_partial(n, count_inuse);
3806 else
3807 x = n->nr_partial;
3808 total += x;
3809 nodes[node] += x;
3812 x = sprintf(buf, "%lu", total);
3813 #ifdef CONFIG_NUMA
3814 for_each_node_state(node, N_NORMAL_MEMORY)
3815 if (nodes[node])
3816 x += sprintf(buf + x, " N%d=%lu",
3817 node, nodes[node]);
3818 #endif
3819 kfree(nodes);
3820 return x + sprintf(buf + x, "\n");
3823 static int any_slab_objects(struct kmem_cache *s)
3825 int node;
3827 for_each_online_node(node) {
3828 struct kmem_cache_node *n = get_node(s, node);
3830 if (!n)
3831 continue;
3833 if (atomic_long_read(&n->total_objects))
3834 return 1;
3836 return 0;
3839 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3840 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3842 struct slab_attribute {
3843 struct attribute attr;
3844 ssize_t (*show)(struct kmem_cache *s, char *buf);
3845 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3848 #define SLAB_ATTR_RO(_name) \
3849 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3851 #define SLAB_ATTR(_name) \
3852 static struct slab_attribute _name##_attr = \
3853 __ATTR(_name, 0644, _name##_show, _name##_store)
3855 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3857 return sprintf(buf, "%d\n", s->size);
3859 SLAB_ATTR_RO(slab_size);
3861 static ssize_t align_show(struct kmem_cache *s, char *buf)
3863 return sprintf(buf, "%d\n", s->align);
3865 SLAB_ATTR_RO(align);
3867 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3869 return sprintf(buf, "%d\n", s->objsize);
3871 SLAB_ATTR_RO(object_size);
3873 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3875 return sprintf(buf, "%d\n", oo_objects(s->oo));
3877 SLAB_ATTR_RO(objs_per_slab);
3879 static ssize_t order_store(struct kmem_cache *s,
3880 const char *buf, size_t length)
3882 unsigned long order;
3883 int err;
3885 err = strict_strtoul(buf, 10, &order);
3886 if (err)
3887 return err;
3889 if (order > slub_max_order || order < slub_min_order)
3890 return -EINVAL;
3892 calculate_sizes(s, order);
3893 return length;
3896 static ssize_t order_show(struct kmem_cache *s, char *buf)
3898 return sprintf(buf, "%d\n", oo_order(s->oo));
3900 SLAB_ATTR(order);
3902 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3904 return sprintf(buf, "%lu\n", s->min_partial);
3907 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3908 size_t length)
3910 unsigned long min;
3911 int err;
3913 err = strict_strtoul(buf, 10, &min);
3914 if (err)
3915 return err;
3917 set_min_partial(s, min);
3918 return length;
3920 SLAB_ATTR(min_partial);
3922 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3924 if (s->ctor) {
3925 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3927 return n + sprintf(buf + n, "\n");
3929 return 0;
3931 SLAB_ATTR_RO(ctor);
3933 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3935 return sprintf(buf, "%d\n", s->refcount - 1);
3937 SLAB_ATTR_RO(aliases);
3939 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3941 return show_slab_objects(s, buf, SO_ALL);
3943 SLAB_ATTR_RO(slabs);
3945 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3947 return show_slab_objects(s, buf, SO_PARTIAL);
3949 SLAB_ATTR_RO(partial);
3951 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3953 return show_slab_objects(s, buf, SO_CPU);
3955 SLAB_ATTR_RO(cpu_slabs);
3957 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3959 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3961 SLAB_ATTR_RO(objects);
3963 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3965 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3967 SLAB_ATTR_RO(objects_partial);
3969 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3971 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3973 SLAB_ATTR_RO(total_objects);
3975 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3977 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3980 static ssize_t sanity_checks_store(struct kmem_cache *s,
3981 const char *buf, size_t length)
3983 s->flags &= ~SLAB_DEBUG_FREE;
3984 if (buf[0] == '1')
3985 s->flags |= SLAB_DEBUG_FREE;
3986 return length;
3988 SLAB_ATTR(sanity_checks);
3990 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3992 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3995 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3996 size_t length)
3998 s->flags &= ~SLAB_TRACE;
3999 if (buf[0] == '1')
4000 s->flags |= SLAB_TRACE;
4001 return length;
4003 SLAB_ATTR(trace);
4005 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4007 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4010 static ssize_t reclaim_account_store(struct kmem_cache *s,
4011 const char *buf, size_t length)
4013 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4014 if (buf[0] == '1')
4015 s->flags |= SLAB_RECLAIM_ACCOUNT;
4016 return length;
4018 SLAB_ATTR(reclaim_account);
4020 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4022 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4024 SLAB_ATTR_RO(hwcache_align);
4026 #ifdef CONFIG_ZONE_DMA
4027 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4029 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4031 SLAB_ATTR_RO(cache_dma);
4032 #endif
4034 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4036 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4038 SLAB_ATTR_RO(destroy_by_rcu);
4040 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4042 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4045 static ssize_t red_zone_store(struct kmem_cache *s,
4046 const char *buf, size_t length)
4048 if (any_slab_objects(s))
4049 return -EBUSY;
4051 s->flags &= ~SLAB_RED_ZONE;
4052 if (buf[0] == '1')
4053 s->flags |= SLAB_RED_ZONE;
4054 calculate_sizes(s, -1);
4055 return length;
4057 SLAB_ATTR(red_zone);
4059 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4061 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4064 static ssize_t poison_store(struct kmem_cache *s,
4065 const char *buf, size_t length)
4067 if (any_slab_objects(s))
4068 return -EBUSY;
4070 s->flags &= ~SLAB_POISON;
4071 if (buf[0] == '1')
4072 s->flags |= SLAB_POISON;
4073 calculate_sizes(s, -1);
4074 return length;
4076 SLAB_ATTR(poison);
4078 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4080 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4083 static ssize_t store_user_store(struct kmem_cache *s,
4084 const char *buf, size_t length)
4086 if (any_slab_objects(s))
4087 return -EBUSY;
4089 s->flags &= ~SLAB_STORE_USER;
4090 if (buf[0] == '1')
4091 s->flags |= SLAB_STORE_USER;
4092 calculate_sizes(s, -1);
4093 return length;
4095 SLAB_ATTR(store_user);
4097 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4099 return 0;
4102 static ssize_t validate_store(struct kmem_cache *s,
4103 const char *buf, size_t length)
4105 int ret = -EINVAL;
4107 if (buf[0] == '1') {
4108 ret = validate_slab_cache(s);
4109 if (ret >= 0)
4110 ret = length;
4112 return ret;
4114 SLAB_ATTR(validate);
4116 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4118 return 0;
4121 static ssize_t shrink_store(struct kmem_cache *s,
4122 const char *buf, size_t length)
4124 if (buf[0] == '1') {
4125 int rc = kmem_cache_shrink(s);
4127 if (rc)
4128 return rc;
4129 } else
4130 return -EINVAL;
4131 return length;
4133 SLAB_ATTR(shrink);
4135 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4137 if (!(s->flags & SLAB_STORE_USER))
4138 return -ENOSYS;
4139 return list_locations(s, buf, TRACK_ALLOC);
4141 SLAB_ATTR_RO(alloc_calls);
4143 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4145 if (!(s->flags & SLAB_STORE_USER))
4146 return -ENOSYS;
4147 return list_locations(s, buf, TRACK_FREE);
4149 SLAB_ATTR_RO(free_calls);
4151 #ifdef CONFIG_NUMA
4152 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4154 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4157 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4158 const char *buf, size_t length)
4160 unsigned long ratio;
4161 int err;
4163 err = strict_strtoul(buf, 10, &ratio);
4164 if (err)
4165 return err;
4167 if (ratio <= 100)
4168 s->remote_node_defrag_ratio = ratio * 10;
4170 return length;
4172 SLAB_ATTR(remote_node_defrag_ratio);
4173 #endif
4175 #ifdef CONFIG_SLUB_STATS
4176 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4178 unsigned long sum = 0;
4179 int cpu;
4180 int len;
4181 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4183 if (!data)
4184 return -ENOMEM;
4186 for_each_online_cpu(cpu) {
4187 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4189 data[cpu] = x;
4190 sum += x;
4193 len = sprintf(buf, "%lu", sum);
4195 #ifdef CONFIG_SMP
4196 for_each_online_cpu(cpu) {
4197 if (data[cpu] && len < PAGE_SIZE - 20)
4198 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4200 #endif
4201 kfree(data);
4202 return len + sprintf(buf + len, "\n");
4205 #define STAT_ATTR(si, text) \
4206 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4208 return show_stat(s, buf, si); \
4210 SLAB_ATTR_RO(text); \
4212 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4213 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4214 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4215 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4216 STAT_ATTR(FREE_FROZEN, free_frozen);
4217 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4218 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4219 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4220 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4221 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4222 STAT_ATTR(FREE_SLAB, free_slab);
4223 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4224 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4225 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4226 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4227 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4228 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4229 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4230 #endif
4232 static struct attribute *slab_attrs[] = {
4233 &slab_size_attr.attr,
4234 &object_size_attr.attr,
4235 &objs_per_slab_attr.attr,
4236 &order_attr.attr,
4237 &min_partial_attr.attr,
4238 &objects_attr.attr,
4239 &objects_partial_attr.attr,
4240 &total_objects_attr.attr,
4241 &slabs_attr.attr,
4242 &partial_attr.attr,
4243 &cpu_slabs_attr.attr,
4244 &ctor_attr.attr,
4245 &aliases_attr.attr,
4246 &align_attr.attr,
4247 &sanity_checks_attr.attr,
4248 &trace_attr.attr,
4249 &hwcache_align_attr.attr,
4250 &reclaim_account_attr.attr,
4251 &destroy_by_rcu_attr.attr,
4252 &red_zone_attr.attr,
4253 &poison_attr.attr,
4254 &store_user_attr.attr,
4255 &validate_attr.attr,
4256 &shrink_attr.attr,
4257 &alloc_calls_attr.attr,
4258 &free_calls_attr.attr,
4259 #ifdef CONFIG_ZONE_DMA
4260 &cache_dma_attr.attr,
4261 #endif
4262 #ifdef CONFIG_NUMA
4263 &remote_node_defrag_ratio_attr.attr,
4264 #endif
4265 #ifdef CONFIG_SLUB_STATS
4266 &alloc_fastpath_attr.attr,
4267 &alloc_slowpath_attr.attr,
4268 &free_fastpath_attr.attr,
4269 &free_slowpath_attr.attr,
4270 &free_frozen_attr.attr,
4271 &free_add_partial_attr.attr,
4272 &free_remove_partial_attr.attr,
4273 &alloc_from_partial_attr.attr,
4274 &alloc_slab_attr.attr,
4275 &alloc_refill_attr.attr,
4276 &free_slab_attr.attr,
4277 &cpuslab_flush_attr.attr,
4278 &deactivate_full_attr.attr,
4279 &deactivate_empty_attr.attr,
4280 &deactivate_to_head_attr.attr,
4281 &deactivate_to_tail_attr.attr,
4282 &deactivate_remote_frees_attr.attr,
4283 &order_fallback_attr.attr,
4284 #endif
4285 NULL
4288 static struct attribute_group slab_attr_group = {
4289 .attrs = slab_attrs,
4292 static ssize_t slab_attr_show(struct kobject *kobj,
4293 struct attribute *attr,
4294 char *buf)
4296 struct slab_attribute *attribute;
4297 struct kmem_cache *s;
4298 int err;
4300 attribute = to_slab_attr(attr);
4301 s = to_slab(kobj);
4303 if (!attribute->show)
4304 return -EIO;
4306 err = attribute->show(s, buf);
4308 return err;
4311 static ssize_t slab_attr_store(struct kobject *kobj,
4312 struct attribute *attr,
4313 const char *buf, size_t len)
4315 struct slab_attribute *attribute;
4316 struct kmem_cache *s;
4317 int err;
4319 attribute = to_slab_attr(attr);
4320 s = to_slab(kobj);
4322 if (!attribute->store)
4323 return -EIO;
4325 err = attribute->store(s, buf, len);
4327 return err;
4330 static void kmem_cache_release(struct kobject *kobj)
4332 struct kmem_cache *s = to_slab(kobj);
4334 kfree(s);
4337 static struct sysfs_ops slab_sysfs_ops = {
4338 .show = slab_attr_show,
4339 .store = slab_attr_store,
4342 static struct kobj_type slab_ktype = {
4343 .sysfs_ops = &slab_sysfs_ops,
4344 .release = kmem_cache_release
4347 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4349 struct kobj_type *ktype = get_ktype(kobj);
4351 if (ktype == &slab_ktype)
4352 return 1;
4353 return 0;
4356 static struct kset_uevent_ops slab_uevent_ops = {
4357 .filter = uevent_filter,
4360 static struct kset *slab_kset;
4362 #define ID_STR_LENGTH 64
4364 /* Create a unique string id for a slab cache:
4366 * Format :[flags-]size
4368 static char *create_unique_id(struct kmem_cache *s)
4370 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4371 char *p = name;
4373 BUG_ON(!name);
4375 *p++ = ':';
4377 * First flags affecting slabcache operations. We will only
4378 * get here for aliasable slabs so we do not need to support
4379 * too many flags. The flags here must cover all flags that
4380 * are matched during merging to guarantee that the id is
4381 * unique.
4383 if (s->flags & SLAB_CACHE_DMA)
4384 *p++ = 'd';
4385 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4386 *p++ = 'a';
4387 if (s->flags & SLAB_DEBUG_FREE)
4388 *p++ = 'F';
4389 if (p != name + 1)
4390 *p++ = '-';
4391 p += sprintf(p, "%07d", s->size);
4392 BUG_ON(p > name + ID_STR_LENGTH - 1);
4393 return name;
4396 static int sysfs_slab_add(struct kmem_cache *s)
4398 int err;
4399 const char *name;
4400 int unmergeable;
4402 if (slab_state < SYSFS)
4403 /* Defer until later */
4404 return 0;
4406 unmergeable = slab_unmergeable(s);
4407 if (unmergeable) {
4409 * Slabcache can never be merged so we can use the name proper.
4410 * This is typically the case for debug situations. In that
4411 * case we can catch duplicate names easily.
4413 sysfs_remove_link(&slab_kset->kobj, s->name);
4414 name = s->name;
4415 } else {
4417 * Create a unique name for the slab as a target
4418 * for the symlinks.
4420 name = create_unique_id(s);
4423 s->kobj.kset = slab_kset;
4424 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4425 if (err) {
4426 kobject_put(&s->kobj);
4427 return err;
4430 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4431 if (err)
4432 return err;
4433 kobject_uevent(&s->kobj, KOBJ_ADD);
4434 if (!unmergeable) {
4435 /* Setup first alias */
4436 sysfs_slab_alias(s, s->name);
4437 kfree(name);
4439 return 0;
4442 static void sysfs_slab_remove(struct kmem_cache *s)
4444 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4445 kobject_del(&s->kobj);
4446 kobject_put(&s->kobj);
4450 * Need to buffer aliases during bootup until sysfs becomes
4451 * available lest we lose that information.
4453 struct saved_alias {
4454 struct kmem_cache *s;
4455 const char *name;
4456 struct saved_alias *next;
4459 static struct saved_alias *alias_list;
4461 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4463 struct saved_alias *al;
4465 if (slab_state == SYSFS) {
4467 * If we have a leftover link then remove it.
4469 sysfs_remove_link(&slab_kset->kobj, name);
4470 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4473 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4474 if (!al)
4475 return -ENOMEM;
4477 al->s = s;
4478 al->name = name;
4479 al->next = alias_list;
4480 alias_list = al;
4481 return 0;
4484 static int __init slab_sysfs_init(void)
4486 struct kmem_cache *s;
4487 int err;
4489 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4490 if (!slab_kset) {
4491 printk(KERN_ERR "Cannot register slab subsystem.\n");
4492 return -ENOSYS;
4495 slab_state = SYSFS;
4497 list_for_each_entry(s, &slab_caches, list) {
4498 err = sysfs_slab_add(s);
4499 if (err)
4500 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4501 " to sysfs\n", s->name);
4504 while (alias_list) {
4505 struct saved_alias *al = alias_list;
4507 alias_list = alias_list->next;
4508 err = sysfs_slab_alias(al->s, al->name);
4509 if (err)
4510 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4511 " %s to sysfs\n", s->name);
4512 kfree(al);
4515 resiliency_test();
4516 return 0;
4519 __initcall(slab_sysfs_init);
4520 #endif
4523 * The /proc/slabinfo ABI
4525 #ifdef CONFIG_SLABINFO
4526 static void print_slabinfo_header(struct seq_file *m)
4528 seq_puts(m, "slabinfo - version: 2.1\n");
4529 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4530 "<objperslab> <pagesperslab>");
4531 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4532 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4533 seq_putc(m, '\n');
4536 static void *s_start(struct seq_file *m, loff_t *pos)
4538 loff_t n = *pos;
4540 down_read(&slub_lock);
4541 if (!n)
4542 print_slabinfo_header(m);
4544 return seq_list_start(&slab_caches, *pos);
4547 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4549 return seq_list_next(p, &slab_caches, pos);
4552 static void s_stop(struct seq_file *m, void *p)
4554 up_read(&slub_lock);
4557 static int s_show(struct seq_file *m, void *p)
4559 unsigned long nr_partials = 0;
4560 unsigned long nr_slabs = 0;
4561 unsigned long nr_inuse = 0;
4562 unsigned long nr_objs = 0;
4563 unsigned long nr_free = 0;
4564 struct kmem_cache *s;
4565 int node;
4567 s = list_entry(p, struct kmem_cache, list);
4569 for_each_online_node(node) {
4570 struct kmem_cache_node *n = get_node(s, node);
4572 if (!n)
4573 continue;
4575 nr_partials += n->nr_partial;
4576 nr_slabs += atomic_long_read(&n->nr_slabs);
4577 nr_objs += atomic_long_read(&n->total_objects);
4578 nr_free += count_partial(n, count_free);
4581 nr_inuse = nr_objs - nr_free;
4583 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4584 nr_objs, s->size, oo_objects(s->oo),
4585 (1 << oo_order(s->oo)));
4586 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4587 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4588 0UL);
4589 seq_putc(m, '\n');
4590 return 0;
4593 static const struct seq_operations slabinfo_op = {
4594 .start = s_start,
4595 .next = s_next,
4596 .stop = s_stop,
4597 .show = s_show,
4600 static int slabinfo_open(struct inode *inode, struct file *file)
4602 return seq_open(file, &slabinfo_op);
4605 static const struct file_operations proc_slabinfo_operations = {
4606 .open = slabinfo_open,
4607 .read = seq_read,
4608 .llseek = seq_lseek,
4609 .release = seq_release,
4612 static int __init slab_proc_init(void)
4614 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4615 return 0;
4617 module_init(slab_proc_init);
4618 #endif /* CONFIG_SLABINFO */