e1000e: fix acquisition of SW/FW/HW semaphore for ICHx parts
[linux-2.6.git] / mm / slub.c
blobb9f1491a58a184e80769fc541d3d2958cb4919ed
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/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemtrace.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
33 * Lock order:
34 * 1. slab_lock(page)
35 * 2. slab->list_lock
37 * The slab_lock protects operations on the object of a particular
38 * slab and its metadata in the page struct. If the slab lock
39 * has been taken then no allocations nor frees can be performed
40 * on the objects in the slab nor can the slab be added or removed
41 * from the partial or full lists since this would mean modifying
42 * the page_struct of the slab.
44 * The list_lock protects the partial and full list on each node and
45 * the partial slab counter. If taken then no new slabs may be added or
46 * removed from the lists nor make the number of partial slabs be modified.
47 * (Note that the total number of slabs is an atomic value that may be
48 * modified without taking the list lock).
50 * The list_lock is a centralized lock and thus we avoid taking it as
51 * much as possible. As long as SLUB does not have to handle partial
52 * slabs, operations can continue without any centralized lock. F.e.
53 * allocating a long series of objects that fill up slabs does not require
54 * the list lock.
56 * The lock order is sometimes inverted when we are trying to get a slab
57 * off a list. We take the list_lock and then look for a page on the list
58 * to use. While we do that objects in the slabs may be freed. We can
59 * only operate on the slab if we have also taken the slab_lock. So we use
60 * a slab_trylock() on the slab. If trylock was successful then no frees
61 * can occur anymore and we can use the slab for allocations etc. If the
62 * slab_trylock() does not succeed then frees are in progress in the slab and
63 * we must stay away from it for a while since we may cause a bouncing
64 * cacheline if we try to acquire the lock. So go onto the next slab.
65 * If all pages are busy then we may allocate a new slab instead of reusing
66 * a partial slab. A new slab has noone operating on it and thus there is
67 * no danger of cacheline contention.
69 * Interrupts are disabled during allocation and deallocation in order to
70 * make the slab allocator safe to use in the context of an irq. In addition
71 * interrupts are disabled to ensure that the processor does not change
72 * while handling per_cpu slabs, due to kernel preemption.
74 * SLUB assigns one slab for allocation to each processor.
75 * Allocations only occur from these slabs called cpu slabs.
77 * Slabs with free elements are kept on a partial list and during regular
78 * operations no list for full slabs is used. If an object in a full slab is
79 * freed then the slab will show up again on the partial lists.
80 * We track full slabs for debugging purposes though because otherwise we
81 * cannot scan all objects.
83 * Slabs are freed when they become empty. Teardown and setup is
84 * minimal so we rely on the page allocators per cpu caches for
85 * fast frees and allocs.
87 * Overloading of page flags that are otherwise used for LRU management.
89 * PageActive The slab is frozen and exempt from list processing.
90 * This means that the slab is dedicated to a purpose
91 * such as satisfying allocations for a specific
92 * processor. Objects may be freed in the slab while
93 * it is frozen but slab_free will then skip the usual
94 * list operations. It is up to the processor holding
95 * the slab to integrate the slab into the slab lists
96 * when the slab is no longer needed.
98 * One use of this flag is to mark slabs that are
99 * used for allocations. Then such a slab becomes a cpu
100 * slab. The cpu slab may be equipped with an additional
101 * freelist that allows lockless access to
102 * free objects in addition to the regular freelist
103 * that requires the slab lock.
105 * PageError Slab requires special handling due to debug
106 * options set. This moves slab handling out of
107 * the fast path and disables lockless freelists.
110 #ifdef CONFIG_SLUB_DEBUG
111 #define SLABDEBUG 1
112 #else
113 #define SLABDEBUG 0
114 #endif
117 * Issues still to be resolved:
119 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
121 * - Variable sizing of the per node arrays
124 /* Enable to test recovery from slab corruption on boot */
125 #undef SLUB_RESILIENCY_TEST
128 * Mininum number of partial slabs. These will be left on the partial
129 * lists even if they are empty. kmem_cache_shrink may reclaim them.
131 #define MIN_PARTIAL 5
134 * Maximum number of desirable partial slabs.
135 * The existence of more partial slabs makes kmem_cache_shrink
136 * sort the partial list by the number of objects in the.
138 #define MAX_PARTIAL 10
140 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
141 SLAB_POISON | SLAB_STORE_USER)
144 * Set of flags that will prevent slab merging
146 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
147 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE)
149 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
150 SLAB_CACHE_DMA | SLAB_NOTRACK)
152 #ifndef ARCH_KMALLOC_MINALIGN
153 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
154 #endif
156 #ifndef ARCH_SLAB_MINALIGN
157 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
158 #endif
160 #define OO_SHIFT 16
161 #define OO_MASK ((1 << OO_SHIFT) - 1)
162 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
164 /* Internal SLUB flags */
165 #define __OBJECT_POISON 0x80000000 /* Poison object */
166 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
168 static int kmem_size = sizeof(struct kmem_cache);
170 #ifdef CONFIG_SMP
171 static struct notifier_block slab_notifier;
172 #endif
174 static enum {
175 DOWN, /* No slab functionality available */
176 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
177 UP, /* Everything works but does not show up in sysfs */
178 SYSFS /* Sysfs up */
179 } slab_state = DOWN;
181 /* A list of all slab caches on the system */
182 static DECLARE_RWSEM(slub_lock);
183 static LIST_HEAD(slab_caches);
186 * Tracking user of a slab.
188 struct track {
189 unsigned long addr; /* Called from address */
190 int cpu; /* Was running on cpu */
191 int pid; /* Pid context */
192 unsigned long when; /* When did the operation occur */
195 enum track_item { TRACK_ALLOC, TRACK_FREE };
197 #ifdef CONFIG_SLUB_DEBUG
198 static int sysfs_slab_add(struct kmem_cache *);
199 static int sysfs_slab_alias(struct kmem_cache *, const char *);
200 static void sysfs_slab_remove(struct kmem_cache *);
202 #else
203 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
204 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
205 { return 0; }
206 static inline void sysfs_slab_remove(struct kmem_cache *s)
208 kfree(s);
211 #endif
213 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
215 #ifdef CONFIG_SLUB_STATS
216 c->stat[si]++;
217 #endif
220 /********************************************************************
221 * Core slab cache functions
222 *******************************************************************/
224 int slab_is_available(void)
226 return slab_state >= UP;
229 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
231 #ifdef CONFIG_NUMA
232 return s->node[node];
233 #else
234 return &s->local_node;
235 #endif
238 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
240 #ifdef CONFIG_SMP
241 return s->cpu_slab[cpu];
242 #else
243 return &s->cpu_slab;
244 #endif
247 /* Verify that a pointer has an address that is valid within a slab page */
248 static inline int check_valid_pointer(struct kmem_cache *s,
249 struct page *page, const void *object)
251 void *base;
253 if (!object)
254 return 1;
256 base = page_address(page);
257 if (object < base || object >= base + page->objects * s->size ||
258 (object - base) % s->size) {
259 return 0;
262 return 1;
266 * Slow version of get and set free pointer.
268 * This version requires touching the cache lines of kmem_cache which
269 * we avoid to do in the fast alloc free paths. There we obtain the offset
270 * from the page struct.
272 static inline void *get_freepointer(struct kmem_cache *s, void *object)
274 return *(void **)(object + s->offset);
277 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
279 *(void **)(object + s->offset) = fp;
282 /* Loop over all objects in a slab */
283 #define for_each_object(__p, __s, __addr, __objects) \
284 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
285 __p += (__s)->size)
287 /* Scan freelist */
288 #define for_each_free_object(__p, __s, __free) \
289 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
291 /* Determine object index from a given position */
292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
294 return (p - addr) / s->size;
297 static inline struct kmem_cache_order_objects oo_make(int order,
298 unsigned long size)
300 struct kmem_cache_order_objects x = {
301 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
304 return x;
307 static inline int oo_order(struct kmem_cache_order_objects x)
309 return x.x >> OO_SHIFT;
312 static inline int oo_objects(struct kmem_cache_order_objects x)
314 return x.x & OO_MASK;
317 #ifdef CONFIG_SLUB_DEBUG
319 * Debug settings:
321 #ifdef CONFIG_SLUB_DEBUG_ON
322 static int slub_debug = DEBUG_DEFAULT_FLAGS;
323 #else
324 static int slub_debug;
325 #endif
327 static char *slub_debug_slabs;
330 * Object debugging
332 static void print_section(char *text, u8 *addr, unsigned int length)
334 int i, offset;
335 int newline = 1;
336 char ascii[17];
338 ascii[16] = 0;
340 for (i = 0; i < length; i++) {
341 if (newline) {
342 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
343 newline = 0;
345 printk(KERN_CONT " %02x", addr[i]);
346 offset = i % 16;
347 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
348 if (offset == 15) {
349 printk(KERN_CONT " %s\n", ascii);
350 newline = 1;
353 if (!newline) {
354 i %= 16;
355 while (i < 16) {
356 printk(KERN_CONT " ");
357 ascii[i] = ' ';
358 i++;
360 printk(KERN_CONT " %s\n", ascii);
364 static struct track *get_track(struct kmem_cache *s, void *object,
365 enum track_item alloc)
367 struct track *p;
369 if (s->offset)
370 p = object + s->offset + sizeof(void *);
371 else
372 p = object + s->inuse;
374 return p + alloc;
377 static void set_track(struct kmem_cache *s, void *object,
378 enum track_item alloc, unsigned long addr)
380 struct track *p = get_track(s, object, alloc);
382 if (addr) {
383 p->addr = addr;
384 p->cpu = smp_processor_id();
385 p->pid = current->pid;
386 p->when = jiffies;
387 } else
388 memset(p, 0, sizeof(struct track));
391 static void init_tracking(struct kmem_cache *s, void *object)
393 if (!(s->flags & SLAB_STORE_USER))
394 return;
396 set_track(s, object, TRACK_FREE, 0UL);
397 set_track(s, object, TRACK_ALLOC, 0UL);
400 static void print_track(const char *s, struct track *t)
402 if (!t->addr)
403 return;
405 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
406 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
409 static void print_tracking(struct kmem_cache *s, void *object)
411 if (!(s->flags & SLAB_STORE_USER))
412 return;
414 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
415 print_track("Freed", get_track(s, object, TRACK_FREE));
418 static void print_page_info(struct page *page)
420 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
421 page, page->objects, page->inuse, page->freelist, page->flags);
425 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
427 va_list args;
428 char buf[100];
430 va_start(args, fmt);
431 vsnprintf(buf, sizeof(buf), fmt, args);
432 va_end(args);
433 printk(KERN_ERR "========================================"
434 "=====================================\n");
435 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
436 printk(KERN_ERR "----------------------------------------"
437 "-------------------------------------\n\n");
440 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
442 va_list args;
443 char buf[100];
445 va_start(args, fmt);
446 vsnprintf(buf, sizeof(buf), fmt, args);
447 va_end(args);
448 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
451 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
453 unsigned int off; /* Offset of last byte */
454 u8 *addr = page_address(page);
456 print_tracking(s, p);
458 print_page_info(page);
460 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
461 p, p - addr, get_freepointer(s, p));
463 if (p > addr + 16)
464 print_section("Bytes b4", p - 16, 16);
466 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
468 if (s->flags & SLAB_RED_ZONE)
469 print_section("Redzone", p + s->objsize,
470 s->inuse - s->objsize);
472 if (s->offset)
473 off = s->offset + sizeof(void *);
474 else
475 off = s->inuse;
477 if (s->flags & SLAB_STORE_USER)
478 off += 2 * sizeof(struct track);
480 if (off != s->size)
481 /* Beginning of the filler is the free pointer */
482 print_section("Padding", p + off, s->size - off);
484 dump_stack();
487 static void object_err(struct kmem_cache *s, struct page *page,
488 u8 *object, char *reason)
490 slab_bug(s, "%s", reason);
491 print_trailer(s, page, object);
494 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
496 va_list args;
497 char buf[100];
499 va_start(args, fmt);
500 vsnprintf(buf, sizeof(buf), fmt, args);
501 va_end(args);
502 slab_bug(s, "%s", buf);
503 print_page_info(page);
504 dump_stack();
507 static void init_object(struct kmem_cache *s, void *object, int active)
509 u8 *p = object;
511 if (s->flags & __OBJECT_POISON) {
512 memset(p, POISON_FREE, s->objsize - 1);
513 p[s->objsize - 1] = POISON_END;
516 if (s->flags & SLAB_RED_ZONE)
517 memset(p + s->objsize,
518 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
519 s->inuse - s->objsize);
522 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
524 while (bytes) {
525 if (*start != (u8)value)
526 return start;
527 start++;
528 bytes--;
530 return NULL;
533 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
534 void *from, void *to)
536 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
537 memset(from, data, to - from);
540 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
541 u8 *object, char *what,
542 u8 *start, unsigned int value, unsigned int bytes)
544 u8 *fault;
545 u8 *end;
547 fault = check_bytes(start, value, bytes);
548 if (!fault)
549 return 1;
551 end = start + bytes;
552 while (end > fault && end[-1] == value)
553 end--;
555 slab_bug(s, "%s overwritten", what);
556 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
557 fault, end - 1, fault[0], value);
558 print_trailer(s, page, object);
560 restore_bytes(s, what, value, fault, end);
561 return 0;
565 * Object layout:
567 * object address
568 * Bytes of the object to be managed.
569 * If the freepointer may overlay the object then the free
570 * pointer is the first word of the object.
572 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
573 * 0xa5 (POISON_END)
575 * object + s->objsize
576 * Padding to reach word boundary. This is also used for Redzoning.
577 * Padding is extended by another word if Redzoning is enabled and
578 * objsize == inuse.
580 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
581 * 0xcc (RED_ACTIVE) for objects in use.
583 * object + s->inuse
584 * Meta data starts here.
586 * A. Free pointer (if we cannot overwrite object on free)
587 * B. Tracking data for SLAB_STORE_USER
588 * C. Padding to reach required alignment boundary or at mininum
589 * one word if debugging is on to be able to detect writes
590 * before the word boundary.
592 * Padding is done using 0x5a (POISON_INUSE)
594 * object + s->size
595 * Nothing is used beyond s->size.
597 * If slabcaches are merged then the objsize and inuse boundaries are mostly
598 * ignored. And therefore no slab options that rely on these boundaries
599 * may be used with merged slabcaches.
602 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
604 unsigned long off = s->inuse; /* The end of info */
606 if (s->offset)
607 /* Freepointer is placed after the object. */
608 off += sizeof(void *);
610 if (s->flags & SLAB_STORE_USER)
611 /* We also have user information there */
612 off += 2 * sizeof(struct track);
614 if (s->size == off)
615 return 1;
617 return check_bytes_and_report(s, page, p, "Object padding",
618 p + off, POISON_INUSE, s->size - off);
621 /* Check the pad bytes at the end of a slab page */
622 static int slab_pad_check(struct kmem_cache *s, struct page *page)
624 u8 *start;
625 u8 *fault;
626 u8 *end;
627 int length;
628 int remainder;
630 if (!(s->flags & SLAB_POISON))
631 return 1;
633 start = page_address(page);
634 length = (PAGE_SIZE << compound_order(page));
635 end = start + length;
636 remainder = length % s->size;
637 if (!remainder)
638 return 1;
640 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
641 if (!fault)
642 return 1;
643 while (end > fault && end[-1] == POISON_INUSE)
644 end--;
646 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
647 print_section("Padding", end - remainder, remainder);
649 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
650 return 0;
653 static int check_object(struct kmem_cache *s, struct page *page,
654 void *object, int active)
656 u8 *p = object;
657 u8 *endobject = object + s->objsize;
659 if (s->flags & SLAB_RED_ZONE) {
660 unsigned int red =
661 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
663 if (!check_bytes_and_report(s, page, object, "Redzone",
664 endobject, red, s->inuse - s->objsize))
665 return 0;
666 } else {
667 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
668 check_bytes_and_report(s, page, p, "Alignment padding",
669 endobject, POISON_INUSE, s->inuse - s->objsize);
673 if (s->flags & SLAB_POISON) {
674 if (!active && (s->flags & __OBJECT_POISON) &&
675 (!check_bytes_and_report(s, page, p, "Poison", p,
676 POISON_FREE, s->objsize - 1) ||
677 !check_bytes_and_report(s, page, p, "Poison",
678 p + s->objsize - 1, POISON_END, 1)))
679 return 0;
681 * check_pad_bytes cleans up on its own.
683 check_pad_bytes(s, page, p);
686 if (!s->offset && active)
688 * Object and freepointer overlap. Cannot check
689 * freepointer while object is allocated.
691 return 1;
693 /* Check free pointer validity */
694 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
695 object_err(s, page, p, "Freepointer corrupt");
697 * No choice but to zap it and thus lose the remainder
698 * of the free objects in this slab. May cause
699 * another error because the object count is now wrong.
701 set_freepointer(s, p, NULL);
702 return 0;
704 return 1;
707 static int check_slab(struct kmem_cache *s, struct page *page)
709 int maxobj;
711 VM_BUG_ON(!irqs_disabled());
713 if (!PageSlab(page)) {
714 slab_err(s, page, "Not a valid slab page");
715 return 0;
718 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
719 if (page->objects > maxobj) {
720 slab_err(s, page, "objects %u > max %u",
721 s->name, page->objects, maxobj);
722 return 0;
724 if (page->inuse > page->objects) {
725 slab_err(s, page, "inuse %u > max %u",
726 s->name, page->inuse, page->objects);
727 return 0;
729 /* Slab_pad_check fixes things up after itself */
730 slab_pad_check(s, page);
731 return 1;
735 * Determine if a certain object on a page is on the freelist. Must hold the
736 * slab lock to guarantee that the chains are in a consistent state.
738 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
740 int nr = 0;
741 void *fp = page->freelist;
742 void *object = NULL;
743 unsigned long max_objects;
745 while (fp && nr <= page->objects) {
746 if (fp == search)
747 return 1;
748 if (!check_valid_pointer(s, page, fp)) {
749 if (object) {
750 object_err(s, page, object,
751 "Freechain corrupt");
752 set_freepointer(s, object, NULL);
753 break;
754 } else {
755 slab_err(s, page, "Freepointer corrupt");
756 page->freelist = NULL;
757 page->inuse = page->objects;
758 slab_fix(s, "Freelist cleared");
759 return 0;
761 break;
763 object = fp;
764 fp = get_freepointer(s, object);
765 nr++;
768 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
769 if (max_objects > MAX_OBJS_PER_PAGE)
770 max_objects = MAX_OBJS_PER_PAGE;
772 if (page->objects != max_objects) {
773 slab_err(s, page, "Wrong number of objects. Found %d but "
774 "should be %d", page->objects, max_objects);
775 page->objects = max_objects;
776 slab_fix(s, "Number of objects adjusted.");
778 if (page->inuse != page->objects - nr) {
779 slab_err(s, page, "Wrong object count. Counter is %d but "
780 "counted were %d", page->inuse, page->objects - nr);
781 page->inuse = page->objects - nr;
782 slab_fix(s, "Object count adjusted.");
784 return search == NULL;
787 static void trace(struct kmem_cache *s, struct page *page, void *object,
788 int alloc)
790 if (s->flags & SLAB_TRACE) {
791 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
792 s->name,
793 alloc ? "alloc" : "free",
794 object, page->inuse,
795 page->freelist);
797 if (!alloc)
798 print_section("Object", (void *)object, s->objsize);
800 dump_stack();
805 * Tracking of fully allocated slabs for debugging purposes.
807 static void add_full(struct kmem_cache_node *n, struct page *page)
809 spin_lock(&n->list_lock);
810 list_add(&page->lru, &n->full);
811 spin_unlock(&n->list_lock);
814 static void remove_full(struct kmem_cache *s, struct page *page)
816 struct kmem_cache_node *n;
818 if (!(s->flags & SLAB_STORE_USER))
819 return;
821 n = get_node(s, page_to_nid(page));
823 spin_lock(&n->list_lock);
824 list_del(&page->lru);
825 spin_unlock(&n->list_lock);
828 /* Tracking of the number of slabs for debugging purposes */
829 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
831 struct kmem_cache_node *n = get_node(s, node);
833 return atomic_long_read(&n->nr_slabs);
836 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
838 return atomic_long_read(&n->nr_slabs);
841 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
843 struct kmem_cache_node *n = get_node(s, node);
846 * May be called early in order to allocate a slab for the
847 * kmem_cache_node structure. Solve the chicken-egg
848 * dilemma by deferring the increment of the count during
849 * bootstrap (see early_kmem_cache_node_alloc).
851 if (!NUMA_BUILD || n) {
852 atomic_long_inc(&n->nr_slabs);
853 atomic_long_add(objects, &n->total_objects);
856 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
858 struct kmem_cache_node *n = get_node(s, node);
860 atomic_long_dec(&n->nr_slabs);
861 atomic_long_sub(objects, &n->total_objects);
864 /* Object debug checks for alloc/free paths */
865 static void setup_object_debug(struct kmem_cache *s, struct page *page,
866 void *object)
868 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
869 return;
871 init_object(s, object, 0);
872 init_tracking(s, object);
875 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
876 void *object, unsigned long addr)
878 if (!check_slab(s, page))
879 goto bad;
881 if (!on_freelist(s, page, object)) {
882 object_err(s, page, object, "Object already allocated");
883 goto bad;
886 if (!check_valid_pointer(s, page, object)) {
887 object_err(s, page, object, "Freelist Pointer check fails");
888 goto bad;
891 if (!check_object(s, page, object, 0))
892 goto bad;
894 /* Success perform special debug activities for allocs */
895 if (s->flags & SLAB_STORE_USER)
896 set_track(s, object, TRACK_ALLOC, addr);
897 trace(s, page, object, 1);
898 init_object(s, object, 1);
899 return 1;
901 bad:
902 if (PageSlab(page)) {
904 * If this is a slab page then lets do the best we can
905 * to avoid issues in the future. Marking all objects
906 * as used avoids touching the remaining objects.
908 slab_fix(s, "Marking all objects used");
909 page->inuse = page->objects;
910 page->freelist = NULL;
912 return 0;
915 static int free_debug_processing(struct kmem_cache *s, struct page *page,
916 void *object, unsigned long addr)
918 if (!check_slab(s, page))
919 goto fail;
921 if (!check_valid_pointer(s, page, object)) {
922 slab_err(s, page, "Invalid object pointer 0x%p", object);
923 goto fail;
926 if (on_freelist(s, page, object)) {
927 object_err(s, page, object, "Object already free");
928 goto fail;
931 if (!check_object(s, page, object, 1))
932 return 0;
934 if (unlikely(s != page->slab)) {
935 if (!PageSlab(page)) {
936 slab_err(s, page, "Attempt to free object(0x%p) "
937 "outside of slab", object);
938 } else if (!page->slab) {
939 printk(KERN_ERR
940 "SLUB <none>: no slab for object 0x%p.\n",
941 object);
942 dump_stack();
943 } else
944 object_err(s, page, object,
945 "page slab pointer corrupt.");
946 goto fail;
949 /* Special debug activities for freeing objects */
950 if (!PageSlubFrozen(page) && !page->freelist)
951 remove_full(s, page);
952 if (s->flags & SLAB_STORE_USER)
953 set_track(s, object, TRACK_FREE, addr);
954 trace(s, page, object, 0);
955 init_object(s, object, 0);
956 return 1;
958 fail:
959 slab_fix(s, "Object at 0x%p not freed", object);
960 return 0;
963 static int __init setup_slub_debug(char *str)
965 slub_debug = DEBUG_DEFAULT_FLAGS;
966 if (*str++ != '=' || !*str)
968 * No options specified. Switch on full debugging.
970 goto out;
972 if (*str == ',')
974 * No options but restriction on slabs. This means full
975 * debugging for slabs matching a pattern.
977 goto check_slabs;
979 slub_debug = 0;
980 if (*str == '-')
982 * Switch off all debugging measures.
984 goto out;
987 * Determine which debug features should be switched on
989 for (; *str && *str != ','; str++) {
990 switch (tolower(*str)) {
991 case 'f':
992 slub_debug |= SLAB_DEBUG_FREE;
993 break;
994 case 'z':
995 slub_debug |= SLAB_RED_ZONE;
996 break;
997 case 'p':
998 slub_debug |= SLAB_POISON;
999 break;
1000 case 'u':
1001 slub_debug |= SLAB_STORE_USER;
1002 break;
1003 case 't':
1004 slub_debug |= SLAB_TRACE;
1005 break;
1006 default:
1007 printk(KERN_ERR "slub_debug option '%c' "
1008 "unknown. skipped\n", *str);
1012 check_slabs:
1013 if (*str == ',')
1014 slub_debug_slabs = str + 1;
1015 out:
1016 return 1;
1019 __setup("slub_debug", setup_slub_debug);
1021 static unsigned long kmem_cache_flags(unsigned long objsize,
1022 unsigned long flags, const char *name,
1023 void (*ctor)(void *))
1026 * Enable debugging if selected on the kernel commandline.
1028 if (slub_debug && (!slub_debug_slabs ||
1029 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1030 flags |= slub_debug;
1032 return flags;
1034 #else
1035 static inline void setup_object_debug(struct kmem_cache *s,
1036 struct page *page, void *object) {}
1038 static inline int alloc_debug_processing(struct kmem_cache *s,
1039 struct page *page, void *object, unsigned long addr) { return 0; }
1041 static inline int free_debug_processing(struct kmem_cache *s,
1042 struct page *page, void *object, unsigned long addr) { return 0; }
1044 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1045 { return 1; }
1046 static inline int check_object(struct kmem_cache *s, struct page *page,
1047 void *object, int active) { return 1; }
1048 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1049 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1050 unsigned long flags, const char *name,
1051 void (*ctor)(void *))
1053 return flags;
1055 #define slub_debug 0
1057 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1058 { return 0; }
1059 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1060 { return 0; }
1061 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1062 int objects) {}
1063 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1064 int objects) {}
1065 #endif
1068 * Slab allocation and freeing
1070 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1071 struct kmem_cache_order_objects oo)
1073 int order = oo_order(oo);
1075 flags |= __GFP_NOTRACK;
1077 if (node == -1)
1078 return alloc_pages(flags, order);
1079 else
1080 return alloc_pages_node(node, flags, order);
1083 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1085 struct page *page;
1086 struct kmem_cache_order_objects oo = s->oo;
1087 gfp_t alloc_gfp;
1089 flags |= s->allocflags;
1092 * Let the initial higher-order allocation fail under memory pressure
1093 * so we fall-back to the minimum order allocation.
1095 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1097 page = alloc_slab_page(alloc_gfp, node, oo);
1098 if (unlikely(!page)) {
1099 oo = s->min;
1101 * Allocation may have failed due to fragmentation.
1102 * Try a lower order alloc if possible
1104 page = alloc_slab_page(flags, node, oo);
1105 if (!page)
1106 return NULL;
1108 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1111 if (kmemcheck_enabled
1112 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS)))
1114 int pages = 1 << oo_order(oo);
1116 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1119 * Objects from caches that have a constructor don't get
1120 * cleared when they're allocated, so we need to do it here.
1122 if (s->ctor)
1123 kmemcheck_mark_uninitialized_pages(page, pages);
1124 else
1125 kmemcheck_mark_unallocated_pages(page, pages);
1128 page->objects = oo_objects(oo);
1129 mod_zone_page_state(page_zone(page),
1130 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1131 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1132 1 << oo_order(oo));
1134 return page;
1137 static void setup_object(struct kmem_cache *s, struct page *page,
1138 void *object)
1140 setup_object_debug(s, page, object);
1141 if (unlikely(s->ctor))
1142 s->ctor(object);
1145 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1147 struct page *page;
1148 void *start;
1149 void *last;
1150 void *p;
1152 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1154 page = allocate_slab(s,
1155 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1156 if (!page)
1157 goto out;
1159 inc_slabs_node(s, page_to_nid(page), page->objects);
1160 page->slab = s;
1161 page->flags |= 1 << PG_slab;
1162 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1163 SLAB_STORE_USER | SLAB_TRACE))
1164 __SetPageSlubDebug(page);
1166 start = page_address(page);
1168 if (unlikely(s->flags & SLAB_POISON))
1169 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1171 last = start;
1172 for_each_object(p, s, start, page->objects) {
1173 setup_object(s, page, last);
1174 set_freepointer(s, last, p);
1175 last = p;
1177 setup_object(s, page, last);
1178 set_freepointer(s, last, NULL);
1180 page->freelist = start;
1181 page->inuse = 0;
1182 out:
1183 return page;
1186 static void __free_slab(struct kmem_cache *s, struct page *page)
1188 int order = compound_order(page);
1189 int pages = 1 << order;
1191 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1192 void *p;
1194 slab_pad_check(s, page);
1195 for_each_object(p, s, page_address(page),
1196 page->objects)
1197 check_object(s, page, p, 0);
1198 __ClearPageSlubDebug(page);
1201 kmemcheck_free_shadow(page, compound_order(page));
1203 mod_zone_page_state(page_zone(page),
1204 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1205 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1206 -pages);
1208 __ClearPageSlab(page);
1209 reset_page_mapcount(page);
1210 if (current->reclaim_state)
1211 current->reclaim_state->reclaimed_slab += pages;
1212 __free_pages(page, order);
1215 static void rcu_free_slab(struct rcu_head *h)
1217 struct page *page;
1219 page = container_of((struct list_head *)h, struct page, lru);
1220 __free_slab(page->slab, page);
1223 static void free_slab(struct kmem_cache *s, struct page *page)
1225 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1227 * RCU free overloads the RCU head over the LRU
1229 struct rcu_head *head = (void *)&page->lru;
1231 call_rcu(head, rcu_free_slab);
1232 } else
1233 __free_slab(s, page);
1236 static void discard_slab(struct kmem_cache *s, struct page *page)
1238 dec_slabs_node(s, page_to_nid(page), page->objects);
1239 free_slab(s, page);
1243 * Per slab locking using the pagelock
1245 static __always_inline void slab_lock(struct page *page)
1247 bit_spin_lock(PG_locked, &page->flags);
1250 static __always_inline void slab_unlock(struct page *page)
1252 __bit_spin_unlock(PG_locked, &page->flags);
1255 static __always_inline int slab_trylock(struct page *page)
1257 int rc = 1;
1259 rc = bit_spin_trylock(PG_locked, &page->flags);
1260 return rc;
1264 * Management of partially allocated slabs
1266 static void add_partial(struct kmem_cache_node *n,
1267 struct page *page, int tail)
1269 spin_lock(&n->list_lock);
1270 n->nr_partial++;
1271 if (tail)
1272 list_add_tail(&page->lru, &n->partial);
1273 else
1274 list_add(&page->lru, &n->partial);
1275 spin_unlock(&n->list_lock);
1278 static void remove_partial(struct kmem_cache *s, struct page *page)
1280 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1282 spin_lock(&n->list_lock);
1283 list_del(&page->lru);
1284 n->nr_partial--;
1285 spin_unlock(&n->list_lock);
1289 * Lock slab and remove from the partial list.
1291 * Must hold list_lock.
1293 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1294 struct page *page)
1296 if (slab_trylock(page)) {
1297 list_del(&page->lru);
1298 n->nr_partial--;
1299 __SetPageSlubFrozen(page);
1300 return 1;
1302 return 0;
1306 * Try to allocate a partial slab from a specific node.
1308 static struct page *get_partial_node(struct kmem_cache_node *n)
1310 struct page *page;
1313 * Racy check. If we mistakenly see no partial slabs then we
1314 * just allocate an empty slab. If we mistakenly try to get a
1315 * partial slab and there is none available then get_partials()
1316 * will return NULL.
1318 if (!n || !n->nr_partial)
1319 return NULL;
1321 spin_lock(&n->list_lock);
1322 list_for_each_entry(page, &n->partial, lru)
1323 if (lock_and_freeze_slab(n, page))
1324 goto out;
1325 page = NULL;
1326 out:
1327 spin_unlock(&n->list_lock);
1328 return page;
1332 * Get a page from somewhere. Search in increasing NUMA distances.
1334 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1336 #ifdef CONFIG_NUMA
1337 struct zonelist *zonelist;
1338 struct zoneref *z;
1339 struct zone *zone;
1340 enum zone_type high_zoneidx = gfp_zone(flags);
1341 struct page *page;
1344 * The defrag ratio allows a configuration of the tradeoffs between
1345 * inter node defragmentation and node local allocations. A lower
1346 * defrag_ratio increases the tendency to do local allocations
1347 * instead of attempting to obtain partial slabs from other nodes.
1349 * If the defrag_ratio is set to 0 then kmalloc() always
1350 * returns node local objects. If the ratio is higher then kmalloc()
1351 * may return off node objects because partial slabs are obtained
1352 * from other nodes and filled up.
1354 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1355 * defrag_ratio = 1000) then every (well almost) allocation will
1356 * first attempt to defrag slab caches on other nodes. This means
1357 * scanning over all nodes to look for partial slabs which may be
1358 * expensive if we do it every time we are trying to find a slab
1359 * with available objects.
1361 if (!s->remote_node_defrag_ratio ||
1362 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1363 return NULL;
1365 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1366 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1367 struct kmem_cache_node *n;
1369 n = get_node(s, zone_to_nid(zone));
1371 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1372 n->nr_partial > s->min_partial) {
1373 page = get_partial_node(n);
1374 if (page)
1375 return page;
1378 #endif
1379 return NULL;
1383 * Get a partial page, lock it and return it.
1385 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1387 struct page *page;
1388 int searchnode = (node == -1) ? numa_node_id() : node;
1390 page = get_partial_node(get_node(s, searchnode));
1391 if (page || (flags & __GFP_THISNODE))
1392 return page;
1394 return get_any_partial(s, flags);
1398 * Move a page back to the lists.
1400 * Must be called with the slab lock held.
1402 * On exit the slab lock will have been dropped.
1404 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1406 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1407 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1409 __ClearPageSlubFrozen(page);
1410 if (page->inuse) {
1412 if (page->freelist) {
1413 add_partial(n, page, tail);
1414 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1415 } else {
1416 stat(c, DEACTIVATE_FULL);
1417 if (SLABDEBUG && PageSlubDebug(page) &&
1418 (s->flags & SLAB_STORE_USER))
1419 add_full(n, page);
1421 slab_unlock(page);
1422 } else {
1423 stat(c, DEACTIVATE_EMPTY);
1424 if (n->nr_partial < s->min_partial) {
1426 * Adding an empty slab to the partial slabs in order
1427 * to avoid page allocator overhead. This slab needs
1428 * to come after the other slabs with objects in
1429 * so that the others get filled first. That way the
1430 * size of the partial list stays small.
1432 * kmem_cache_shrink can reclaim any empty slabs from
1433 * the partial list.
1435 add_partial(n, page, 1);
1436 slab_unlock(page);
1437 } else {
1438 slab_unlock(page);
1439 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1440 discard_slab(s, page);
1446 * Remove the cpu slab
1448 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1450 struct page *page = c->page;
1451 int tail = 1;
1453 if (page->freelist)
1454 stat(c, DEACTIVATE_REMOTE_FREES);
1456 * Merge cpu freelist into slab freelist. Typically we get here
1457 * because both freelists are empty. So this is unlikely
1458 * to occur.
1460 while (unlikely(c->freelist)) {
1461 void **object;
1463 tail = 0; /* Hot objects. Put the slab first */
1465 /* Retrieve object from cpu_freelist */
1466 object = c->freelist;
1467 c->freelist = c->freelist[c->offset];
1469 /* And put onto the regular freelist */
1470 object[c->offset] = page->freelist;
1471 page->freelist = object;
1472 page->inuse--;
1474 c->page = NULL;
1475 unfreeze_slab(s, page, tail);
1478 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1480 stat(c, CPUSLAB_FLUSH);
1481 slab_lock(c->page);
1482 deactivate_slab(s, c);
1486 * Flush cpu slab.
1488 * Called from IPI handler with interrupts disabled.
1490 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1492 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1494 if (likely(c && c->page))
1495 flush_slab(s, c);
1498 static void flush_cpu_slab(void *d)
1500 struct kmem_cache *s = d;
1502 __flush_cpu_slab(s, smp_processor_id());
1505 static void flush_all(struct kmem_cache *s)
1507 on_each_cpu(flush_cpu_slab, s, 1);
1511 * Check if the objects in a per cpu structure fit numa
1512 * locality expectations.
1514 static inline int node_match(struct kmem_cache_cpu *c, int node)
1516 #ifdef CONFIG_NUMA
1517 if (node != -1 && c->node != node)
1518 return 0;
1519 #endif
1520 return 1;
1523 static int count_free(struct page *page)
1525 return page->objects - page->inuse;
1528 static unsigned long count_partial(struct kmem_cache_node *n,
1529 int (*get_count)(struct page *))
1531 unsigned long flags;
1532 unsigned long x = 0;
1533 struct page *page;
1535 spin_lock_irqsave(&n->list_lock, flags);
1536 list_for_each_entry(page, &n->partial, lru)
1537 x += get_count(page);
1538 spin_unlock_irqrestore(&n->list_lock, flags);
1539 return x;
1542 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1544 #ifdef CONFIG_SLUB_DEBUG
1545 return atomic_long_read(&n->total_objects);
1546 #else
1547 return 0;
1548 #endif
1551 static noinline void
1552 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1554 int node;
1556 printk(KERN_WARNING
1557 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1558 nid, gfpflags);
1559 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1560 "default order: %d, min order: %d\n", s->name, s->objsize,
1561 s->size, oo_order(s->oo), oo_order(s->min));
1563 for_each_online_node(node) {
1564 struct kmem_cache_node *n = get_node(s, node);
1565 unsigned long nr_slabs;
1566 unsigned long nr_objs;
1567 unsigned long nr_free;
1569 if (!n)
1570 continue;
1572 nr_free = count_partial(n, count_free);
1573 nr_slabs = node_nr_slabs(n);
1574 nr_objs = node_nr_objs(n);
1576 printk(KERN_WARNING
1577 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1578 node, nr_slabs, nr_objs, nr_free);
1583 * Slow path. The lockless freelist is empty or we need to perform
1584 * debugging duties.
1586 * Interrupts are disabled.
1588 * Processing is still very fast if new objects have been freed to the
1589 * regular freelist. In that case we simply take over the regular freelist
1590 * as the lockless freelist and zap the regular freelist.
1592 * If that is not working then we fall back to the partial lists. We take the
1593 * first element of the freelist as the object to allocate now and move the
1594 * rest of the freelist to the lockless freelist.
1596 * And if we were unable to get a new slab from the partial slab lists then
1597 * we need to allocate a new slab. This is the slowest path since it involves
1598 * a call to the page allocator and the setup of a new slab.
1600 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1601 unsigned long addr, struct kmem_cache_cpu *c)
1603 void **object;
1604 struct page *new;
1606 /* We handle __GFP_ZERO in the caller */
1607 gfpflags &= ~__GFP_ZERO;
1609 if (!c->page)
1610 goto new_slab;
1612 slab_lock(c->page);
1613 if (unlikely(!node_match(c, node)))
1614 goto another_slab;
1616 stat(c, ALLOC_REFILL);
1618 load_freelist:
1619 object = c->page->freelist;
1620 if (unlikely(!object))
1621 goto another_slab;
1622 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1623 goto debug;
1625 c->freelist = object[c->offset];
1626 c->page->inuse = c->page->objects;
1627 c->page->freelist = NULL;
1628 c->node = page_to_nid(c->page);
1629 unlock_out:
1630 slab_unlock(c->page);
1631 stat(c, ALLOC_SLOWPATH);
1632 return object;
1634 another_slab:
1635 deactivate_slab(s, c);
1637 new_slab:
1638 new = get_partial(s, gfpflags, node);
1639 if (new) {
1640 c->page = new;
1641 stat(c, ALLOC_FROM_PARTIAL);
1642 goto load_freelist;
1645 if (gfpflags & __GFP_WAIT)
1646 local_irq_enable();
1648 new = new_slab(s, gfpflags, node);
1650 if (gfpflags & __GFP_WAIT)
1651 local_irq_disable();
1653 if (new) {
1654 c = get_cpu_slab(s, smp_processor_id());
1655 stat(c, ALLOC_SLAB);
1656 if (c->page)
1657 flush_slab(s, c);
1658 slab_lock(new);
1659 __SetPageSlubFrozen(new);
1660 c->page = new;
1661 goto load_freelist;
1663 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1664 slab_out_of_memory(s, gfpflags, node);
1665 return NULL;
1666 debug:
1667 if (!alloc_debug_processing(s, c->page, object, addr))
1668 goto another_slab;
1670 c->page->inuse++;
1671 c->page->freelist = object[c->offset];
1672 c->node = -1;
1673 goto unlock_out;
1677 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1678 * have the fastpath folded into their functions. So no function call
1679 * overhead for requests that can be satisfied on the fastpath.
1681 * The fastpath works by first checking if the lockless freelist can be used.
1682 * If not then __slab_alloc is called for slow processing.
1684 * Otherwise we can simply pick the next object from the lockless free list.
1686 static __always_inline void *slab_alloc(struct kmem_cache *s,
1687 gfp_t gfpflags, int node, unsigned long addr)
1689 void **object;
1690 struct kmem_cache_cpu *c;
1691 unsigned long flags;
1692 unsigned int objsize;
1694 gfpflags &= gfp_allowed_mask;
1696 lockdep_trace_alloc(gfpflags);
1697 might_sleep_if(gfpflags & __GFP_WAIT);
1699 if (should_failslab(s->objsize, gfpflags))
1700 return NULL;
1702 local_irq_save(flags);
1703 c = get_cpu_slab(s, smp_processor_id());
1704 objsize = c->objsize;
1705 if (unlikely(!c->freelist || !node_match(c, node)))
1707 object = __slab_alloc(s, gfpflags, node, addr, c);
1709 else {
1710 object = c->freelist;
1711 c->freelist = object[c->offset];
1712 stat(c, ALLOC_FASTPATH);
1714 local_irq_restore(flags);
1716 if (unlikely((gfpflags & __GFP_ZERO) && object))
1717 memset(object, 0, objsize);
1719 kmemcheck_slab_alloc(s, gfpflags, object, c->objsize);
1720 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags);
1722 return object;
1725 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1727 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1729 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1731 return ret;
1733 EXPORT_SYMBOL(kmem_cache_alloc);
1735 #ifdef CONFIG_KMEMTRACE
1736 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1738 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1740 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1741 #endif
1743 #ifdef CONFIG_NUMA
1744 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1746 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1748 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1749 s->objsize, s->size, gfpflags, node);
1751 return ret;
1753 EXPORT_SYMBOL(kmem_cache_alloc_node);
1754 #endif
1756 #ifdef CONFIG_KMEMTRACE
1757 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1758 gfp_t gfpflags,
1759 int node)
1761 return slab_alloc(s, gfpflags, node, _RET_IP_);
1763 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1764 #endif
1767 * Slow patch handling. This may still be called frequently since objects
1768 * have a longer lifetime than the cpu slabs in most processing loads.
1770 * So we still attempt to reduce cache line usage. Just take the slab
1771 * lock and free the item. If there is no additional partial page
1772 * handling required then we can return immediately.
1774 static void __slab_free(struct kmem_cache *s, struct page *page,
1775 void *x, unsigned long addr, unsigned int offset)
1777 void *prior;
1778 void **object = (void *)x;
1779 struct kmem_cache_cpu *c;
1781 c = get_cpu_slab(s, raw_smp_processor_id());
1782 stat(c, FREE_SLOWPATH);
1783 slab_lock(page);
1785 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1786 goto debug;
1788 checks_ok:
1789 prior = object[offset] = page->freelist;
1790 page->freelist = object;
1791 page->inuse--;
1793 if (unlikely(PageSlubFrozen(page))) {
1794 stat(c, FREE_FROZEN);
1795 goto out_unlock;
1798 if (unlikely(!page->inuse))
1799 goto slab_empty;
1802 * Objects left in the slab. If it was not on the partial list before
1803 * then add it.
1805 if (unlikely(!prior)) {
1806 add_partial(get_node(s, page_to_nid(page)), page, 1);
1807 stat(c, FREE_ADD_PARTIAL);
1810 out_unlock:
1811 slab_unlock(page);
1812 return;
1814 slab_empty:
1815 if (prior) {
1817 * Slab still on the partial list.
1819 remove_partial(s, page);
1820 stat(c, FREE_REMOVE_PARTIAL);
1822 slab_unlock(page);
1823 stat(c, FREE_SLAB);
1824 discard_slab(s, page);
1825 return;
1827 debug:
1828 if (!free_debug_processing(s, page, x, addr))
1829 goto out_unlock;
1830 goto checks_ok;
1834 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1835 * can perform fastpath freeing without additional function calls.
1837 * The fastpath is only possible if we are freeing to the current cpu slab
1838 * of this processor. This typically the case if we have just allocated
1839 * the item before.
1841 * If fastpath is not possible then fall back to __slab_free where we deal
1842 * with all sorts of special processing.
1844 static __always_inline void slab_free(struct kmem_cache *s,
1845 struct page *page, void *x, unsigned long addr)
1847 void **object = (void *)x;
1848 struct kmem_cache_cpu *c;
1849 unsigned long flags;
1851 kmemleak_free_recursive(x, s->flags);
1852 local_irq_save(flags);
1853 c = get_cpu_slab(s, smp_processor_id());
1854 kmemcheck_slab_free(s, object, c->objsize);
1855 debug_check_no_locks_freed(object, c->objsize);
1856 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1857 debug_check_no_obj_freed(object, c->objsize);
1858 if (likely(page == c->page && c->node >= 0)) {
1859 object[c->offset] = c->freelist;
1860 c->freelist = object;
1861 stat(c, FREE_FASTPATH);
1862 } else
1863 __slab_free(s, page, x, addr, c->offset);
1865 local_irq_restore(flags);
1868 void kmem_cache_free(struct kmem_cache *s, void *x)
1870 struct page *page;
1872 page = virt_to_head_page(x);
1874 slab_free(s, page, x, _RET_IP_);
1876 trace_kmem_cache_free(_RET_IP_, x);
1878 EXPORT_SYMBOL(kmem_cache_free);
1880 /* Figure out on which slab page the object resides */
1881 static struct page *get_object_page(const void *x)
1883 struct page *page = virt_to_head_page(x);
1885 if (!PageSlab(page))
1886 return NULL;
1888 return page;
1892 * Object placement in a slab is made very easy because we always start at
1893 * offset 0. If we tune the size of the object to the alignment then we can
1894 * get the required alignment by putting one properly sized object after
1895 * another.
1897 * Notice that the allocation order determines the sizes of the per cpu
1898 * caches. Each processor has always one slab available for allocations.
1899 * Increasing the allocation order reduces the number of times that slabs
1900 * must be moved on and off the partial lists and is therefore a factor in
1901 * locking overhead.
1905 * Mininum / Maximum order of slab pages. This influences locking overhead
1906 * and slab fragmentation. A higher order reduces the number of partial slabs
1907 * and increases the number of allocations possible without having to
1908 * take the list_lock.
1910 static int slub_min_order;
1911 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1912 static int slub_min_objects;
1915 * Merge control. If this is set then no merging of slab caches will occur.
1916 * (Could be removed. This was introduced to pacify the merge skeptics.)
1918 static int slub_nomerge;
1921 * Calculate the order of allocation given an slab object size.
1923 * The order of allocation has significant impact on performance and other
1924 * system components. Generally order 0 allocations should be preferred since
1925 * order 0 does not cause fragmentation in the page allocator. Larger objects
1926 * be problematic to put into order 0 slabs because there may be too much
1927 * unused space left. We go to a higher order if more than 1/16th of the slab
1928 * would be wasted.
1930 * In order to reach satisfactory performance we must ensure that a minimum
1931 * number of objects is in one slab. Otherwise we may generate too much
1932 * activity on the partial lists which requires taking the list_lock. This is
1933 * less a concern for large slabs though which are rarely used.
1935 * slub_max_order specifies the order where we begin to stop considering the
1936 * number of objects in a slab as critical. If we reach slub_max_order then
1937 * we try to keep the page order as low as possible. So we accept more waste
1938 * of space in favor of a small page order.
1940 * Higher order allocations also allow the placement of more objects in a
1941 * slab and thereby reduce object handling overhead. If the user has
1942 * requested a higher mininum order then we start with that one instead of
1943 * the smallest order which will fit the object.
1945 static inline int slab_order(int size, int min_objects,
1946 int max_order, int fract_leftover)
1948 int order;
1949 int rem;
1950 int min_order = slub_min_order;
1952 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1953 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1955 for (order = max(min_order,
1956 fls(min_objects * size - 1) - PAGE_SHIFT);
1957 order <= max_order; order++) {
1959 unsigned long slab_size = PAGE_SIZE << order;
1961 if (slab_size < min_objects * size)
1962 continue;
1964 rem = slab_size % size;
1966 if (rem <= slab_size / fract_leftover)
1967 break;
1971 return order;
1974 static inline int calculate_order(int size)
1976 int order;
1977 int min_objects;
1978 int fraction;
1979 int max_objects;
1982 * Attempt to find best configuration for a slab. This
1983 * works by first attempting to generate a layout with
1984 * the best configuration and backing off gradually.
1986 * First we reduce the acceptable waste in a slab. Then
1987 * we reduce the minimum objects required in a slab.
1989 min_objects = slub_min_objects;
1990 if (!min_objects)
1991 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1992 max_objects = (PAGE_SIZE << slub_max_order)/size;
1993 min_objects = min(min_objects, max_objects);
1995 while (min_objects > 1) {
1996 fraction = 16;
1997 while (fraction >= 4) {
1998 order = slab_order(size, min_objects,
1999 slub_max_order, fraction);
2000 if (order <= slub_max_order)
2001 return order;
2002 fraction /= 2;
2004 min_objects --;
2008 * We were unable to place multiple objects in a slab. Now
2009 * lets see if we can place a single object there.
2011 order = slab_order(size, 1, slub_max_order, 1);
2012 if (order <= slub_max_order)
2013 return order;
2016 * Doh this slab cannot be placed using slub_max_order.
2018 order = slab_order(size, 1, MAX_ORDER, 1);
2019 if (order < MAX_ORDER)
2020 return order;
2021 return -ENOSYS;
2025 * Figure out what the alignment of the objects will be.
2027 static unsigned long calculate_alignment(unsigned long flags,
2028 unsigned long align, unsigned long size)
2031 * If the user wants hardware cache aligned objects then follow that
2032 * suggestion if the object is sufficiently large.
2034 * The hardware cache alignment cannot override the specified
2035 * alignment though. If that is greater then use it.
2037 if (flags & SLAB_HWCACHE_ALIGN) {
2038 unsigned long ralign = cache_line_size();
2039 while (size <= ralign / 2)
2040 ralign /= 2;
2041 align = max(align, ralign);
2044 if (align < ARCH_SLAB_MINALIGN)
2045 align = ARCH_SLAB_MINALIGN;
2047 return ALIGN(align, sizeof(void *));
2050 static void init_kmem_cache_cpu(struct kmem_cache *s,
2051 struct kmem_cache_cpu *c)
2053 c->page = NULL;
2054 c->freelist = NULL;
2055 c->node = 0;
2056 c->offset = s->offset / sizeof(void *);
2057 c->objsize = s->objsize;
2058 #ifdef CONFIG_SLUB_STATS
2059 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
2060 #endif
2063 static void
2064 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2066 n->nr_partial = 0;
2067 spin_lock_init(&n->list_lock);
2068 INIT_LIST_HEAD(&n->partial);
2069 #ifdef CONFIG_SLUB_DEBUG
2070 atomic_long_set(&n->nr_slabs, 0);
2071 atomic_long_set(&n->total_objects, 0);
2072 INIT_LIST_HEAD(&n->full);
2073 #endif
2076 #ifdef CONFIG_SMP
2078 * Per cpu array for per cpu structures.
2080 * The per cpu array places all kmem_cache_cpu structures from one processor
2081 * close together meaning that it becomes possible that multiple per cpu
2082 * structures are contained in one cacheline. This may be particularly
2083 * beneficial for the kmalloc caches.
2085 * A desktop system typically has around 60-80 slabs. With 100 here we are
2086 * likely able to get per cpu structures for all caches from the array defined
2087 * here. We must be able to cover all kmalloc caches during bootstrap.
2089 * If the per cpu array is exhausted then fall back to kmalloc
2090 * of individual cachelines. No sharing is possible then.
2092 #define NR_KMEM_CACHE_CPU 100
2094 static DEFINE_PER_CPU(struct kmem_cache_cpu,
2095 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2097 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2098 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
2100 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2101 int cpu, gfp_t flags)
2103 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2105 if (c)
2106 per_cpu(kmem_cache_cpu_free, cpu) =
2107 (void *)c->freelist;
2108 else {
2109 /* Table overflow: So allocate ourselves */
2110 c = kmalloc_node(
2111 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2112 flags, cpu_to_node(cpu));
2113 if (!c)
2114 return NULL;
2117 init_kmem_cache_cpu(s, c);
2118 return c;
2121 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2123 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2124 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2125 kfree(c);
2126 return;
2128 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2129 per_cpu(kmem_cache_cpu_free, cpu) = c;
2132 static void free_kmem_cache_cpus(struct kmem_cache *s)
2134 int cpu;
2136 for_each_online_cpu(cpu) {
2137 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2139 if (c) {
2140 s->cpu_slab[cpu] = NULL;
2141 free_kmem_cache_cpu(c, cpu);
2146 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2148 int cpu;
2150 for_each_online_cpu(cpu) {
2151 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2153 if (c)
2154 continue;
2156 c = alloc_kmem_cache_cpu(s, cpu, flags);
2157 if (!c) {
2158 free_kmem_cache_cpus(s);
2159 return 0;
2161 s->cpu_slab[cpu] = c;
2163 return 1;
2167 * Initialize the per cpu array.
2169 static void init_alloc_cpu_cpu(int cpu)
2171 int i;
2173 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2174 return;
2176 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2177 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2179 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2182 static void __init init_alloc_cpu(void)
2184 int cpu;
2186 for_each_online_cpu(cpu)
2187 init_alloc_cpu_cpu(cpu);
2190 #else
2191 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2192 static inline void init_alloc_cpu(void) {}
2194 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2196 init_kmem_cache_cpu(s, &s->cpu_slab);
2197 return 1;
2199 #endif
2201 #ifdef CONFIG_NUMA
2203 * No kmalloc_node yet so do it by hand. We know that this is the first
2204 * slab on the node for this slabcache. There are no concurrent accesses
2205 * possible.
2207 * Note that this function only works on the kmalloc_node_cache
2208 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2209 * memory on a fresh node that has no slab structures yet.
2211 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2213 struct page *page;
2214 struct kmem_cache_node *n;
2215 unsigned long flags;
2217 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2219 page = new_slab(kmalloc_caches, gfpflags, node);
2221 BUG_ON(!page);
2222 if (page_to_nid(page) != node) {
2223 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2224 "node %d\n", node);
2225 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2226 "in order to be able to continue\n");
2229 n = page->freelist;
2230 BUG_ON(!n);
2231 page->freelist = get_freepointer(kmalloc_caches, n);
2232 page->inuse++;
2233 kmalloc_caches->node[node] = n;
2234 #ifdef CONFIG_SLUB_DEBUG
2235 init_object(kmalloc_caches, n, 1);
2236 init_tracking(kmalloc_caches, n);
2237 #endif
2238 init_kmem_cache_node(n, kmalloc_caches);
2239 inc_slabs_node(kmalloc_caches, node, page->objects);
2242 * lockdep requires consistent irq usage for each lock
2243 * so even though there cannot be a race this early in
2244 * the boot sequence, we still disable irqs.
2246 local_irq_save(flags);
2247 add_partial(n, page, 0);
2248 local_irq_restore(flags);
2251 static void free_kmem_cache_nodes(struct kmem_cache *s)
2253 int node;
2255 for_each_node_state(node, N_NORMAL_MEMORY) {
2256 struct kmem_cache_node *n = s->node[node];
2257 if (n && n != &s->local_node)
2258 kmem_cache_free(kmalloc_caches, n);
2259 s->node[node] = NULL;
2263 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2265 int node;
2266 int local_node;
2268 if (slab_state >= UP)
2269 local_node = page_to_nid(virt_to_page(s));
2270 else
2271 local_node = 0;
2273 for_each_node_state(node, N_NORMAL_MEMORY) {
2274 struct kmem_cache_node *n;
2276 if (local_node == node)
2277 n = &s->local_node;
2278 else {
2279 if (slab_state == DOWN) {
2280 early_kmem_cache_node_alloc(gfpflags, node);
2281 continue;
2283 n = kmem_cache_alloc_node(kmalloc_caches,
2284 gfpflags, node);
2286 if (!n) {
2287 free_kmem_cache_nodes(s);
2288 return 0;
2292 s->node[node] = n;
2293 init_kmem_cache_node(n, s);
2295 return 1;
2297 #else
2298 static void free_kmem_cache_nodes(struct kmem_cache *s)
2302 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2304 init_kmem_cache_node(&s->local_node, s);
2305 return 1;
2307 #endif
2309 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2311 if (min < MIN_PARTIAL)
2312 min = MIN_PARTIAL;
2313 else if (min > MAX_PARTIAL)
2314 min = MAX_PARTIAL;
2315 s->min_partial = min;
2319 * calculate_sizes() determines the order and the distribution of data within
2320 * a slab object.
2322 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2324 unsigned long flags = s->flags;
2325 unsigned long size = s->objsize;
2326 unsigned long align = s->align;
2327 int order;
2330 * Round up object size to the next word boundary. We can only
2331 * place the free pointer at word boundaries and this determines
2332 * the possible location of the free pointer.
2334 size = ALIGN(size, sizeof(void *));
2336 #ifdef CONFIG_SLUB_DEBUG
2338 * Determine if we can poison the object itself. If the user of
2339 * the slab may touch the object after free or before allocation
2340 * then we should never poison the object itself.
2342 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2343 !s->ctor)
2344 s->flags |= __OBJECT_POISON;
2345 else
2346 s->flags &= ~__OBJECT_POISON;
2350 * If we are Redzoning then check if there is some space between the
2351 * end of the object and the free pointer. If not then add an
2352 * additional word to have some bytes to store Redzone information.
2354 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2355 size += sizeof(void *);
2356 #endif
2359 * With that we have determined the number of bytes in actual use
2360 * by the object. This is the potential offset to the free pointer.
2362 s->inuse = size;
2364 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2365 s->ctor)) {
2367 * Relocate free pointer after the object if it is not
2368 * permitted to overwrite the first word of the object on
2369 * kmem_cache_free.
2371 * This is the case if we do RCU, have a constructor or
2372 * destructor or are poisoning the objects.
2374 s->offset = size;
2375 size += sizeof(void *);
2378 #ifdef CONFIG_SLUB_DEBUG
2379 if (flags & SLAB_STORE_USER)
2381 * Need to store information about allocs and frees after
2382 * the object.
2384 size += 2 * sizeof(struct track);
2386 if (flags & SLAB_RED_ZONE)
2388 * Add some empty padding so that we can catch
2389 * overwrites from earlier objects rather than let
2390 * tracking information or the free pointer be
2391 * corrupted if a user writes before the start
2392 * of the object.
2394 size += sizeof(void *);
2395 #endif
2398 * Determine the alignment based on various parameters that the
2399 * user specified and the dynamic determination of cache line size
2400 * on bootup.
2402 align = calculate_alignment(flags, align, s->objsize);
2405 * SLUB stores one object immediately after another beginning from
2406 * offset 0. In order to align the objects we have to simply size
2407 * each object to conform to the alignment.
2409 size = ALIGN(size, align);
2410 s->size = size;
2411 if (forced_order >= 0)
2412 order = forced_order;
2413 else
2414 order = calculate_order(size);
2416 if (order < 0)
2417 return 0;
2419 s->allocflags = 0;
2420 if (order)
2421 s->allocflags |= __GFP_COMP;
2423 if (s->flags & SLAB_CACHE_DMA)
2424 s->allocflags |= SLUB_DMA;
2426 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2427 s->allocflags |= __GFP_RECLAIMABLE;
2430 * Determine the number of objects per slab
2432 s->oo = oo_make(order, size);
2433 s->min = oo_make(get_order(size), size);
2434 if (oo_objects(s->oo) > oo_objects(s->max))
2435 s->max = s->oo;
2437 return !!oo_objects(s->oo);
2441 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2442 const char *name, size_t size,
2443 size_t align, unsigned long flags,
2444 void (*ctor)(void *))
2446 memset(s, 0, kmem_size);
2447 s->name = name;
2448 s->ctor = ctor;
2449 s->objsize = size;
2450 s->align = align;
2451 s->flags = kmem_cache_flags(size, flags, name, ctor);
2453 if (!calculate_sizes(s, -1))
2454 goto error;
2457 * The larger the object size is, the more pages we want on the partial
2458 * list to avoid pounding the page allocator excessively.
2460 set_min_partial(s, ilog2(s->size));
2461 s->refcount = 1;
2462 #ifdef CONFIG_NUMA
2463 s->remote_node_defrag_ratio = 1000;
2464 #endif
2465 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2466 goto error;
2468 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2469 return 1;
2470 free_kmem_cache_nodes(s);
2471 error:
2472 if (flags & SLAB_PANIC)
2473 panic("Cannot create slab %s size=%lu realsize=%u "
2474 "order=%u offset=%u flags=%lx\n",
2475 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2476 s->offset, flags);
2477 return 0;
2481 * Check if a given pointer is valid
2483 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2485 struct page *page;
2487 page = get_object_page(object);
2489 if (!page || s != page->slab)
2490 /* No slab or wrong slab */
2491 return 0;
2493 if (!check_valid_pointer(s, page, object))
2494 return 0;
2497 * We could also check if the object is on the slabs freelist.
2498 * But this would be too expensive and it seems that the main
2499 * purpose of kmem_ptr_valid() is to check if the object belongs
2500 * to a certain slab.
2502 return 1;
2504 EXPORT_SYMBOL(kmem_ptr_validate);
2507 * Determine the size of a slab object
2509 unsigned int kmem_cache_size(struct kmem_cache *s)
2511 return s->objsize;
2513 EXPORT_SYMBOL(kmem_cache_size);
2515 const char *kmem_cache_name(struct kmem_cache *s)
2517 return s->name;
2519 EXPORT_SYMBOL(kmem_cache_name);
2521 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2522 const char *text)
2524 #ifdef CONFIG_SLUB_DEBUG
2525 void *addr = page_address(page);
2526 void *p;
2527 DECLARE_BITMAP(map, page->objects);
2529 bitmap_zero(map, page->objects);
2530 slab_err(s, page, "%s", text);
2531 slab_lock(page);
2532 for_each_free_object(p, s, page->freelist)
2533 set_bit(slab_index(p, s, addr), map);
2535 for_each_object(p, s, addr, page->objects) {
2537 if (!test_bit(slab_index(p, s, addr), map)) {
2538 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2539 p, p - addr);
2540 print_tracking(s, p);
2543 slab_unlock(page);
2544 #endif
2548 * Attempt to free all partial slabs on a node.
2550 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2552 unsigned long flags;
2553 struct page *page, *h;
2555 spin_lock_irqsave(&n->list_lock, flags);
2556 list_for_each_entry_safe(page, h, &n->partial, lru) {
2557 if (!page->inuse) {
2558 list_del(&page->lru);
2559 discard_slab(s, page);
2560 n->nr_partial--;
2561 } else {
2562 list_slab_objects(s, page,
2563 "Objects remaining on kmem_cache_close()");
2566 spin_unlock_irqrestore(&n->list_lock, flags);
2570 * Release all resources used by a slab cache.
2572 static inline int kmem_cache_close(struct kmem_cache *s)
2574 int node;
2576 flush_all(s);
2578 /* Attempt to free all objects */
2579 free_kmem_cache_cpus(s);
2580 for_each_node_state(node, N_NORMAL_MEMORY) {
2581 struct kmem_cache_node *n = get_node(s, node);
2583 free_partial(s, n);
2584 if (n->nr_partial || slabs_node(s, node))
2585 return 1;
2587 free_kmem_cache_nodes(s);
2588 return 0;
2592 * Close a cache and release the kmem_cache structure
2593 * (must be used for caches created using kmem_cache_create)
2595 void kmem_cache_destroy(struct kmem_cache *s)
2597 if (s->flags & SLAB_DESTROY_BY_RCU)
2598 rcu_barrier();
2599 down_write(&slub_lock);
2600 s->refcount--;
2601 if (!s->refcount) {
2602 list_del(&s->list);
2603 up_write(&slub_lock);
2604 if (kmem_cache_close(s)) {
2605 printk(KERN_ERR "SLUB %s: %s called for cache that "
2606 "still has objects.\n", s->name, __func__);
2607 dump_stack();
2609 sysfs_slab_remove(s);
2610 } else
2611 up_write(&slub_lock);
2613 EXPORT_SYMBOL(kmem_cache_destroy);
2615 /********************************************************************
2616 * Kmalloc subsystem
2617 *******************************************************************/
2619 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2620 EXPORT_SYMBOL(kmalloc_caches);
2622 static int __init setup_slub_min_order(char *str)
2624 get_option(&str, &slub_min_order);
2626 return 1;
2629 __setup("slub_min_order=", setup_slub_min_order);
2631 static int __init setup_slub_max_order(char *str)
2633 get_option(&str, &slub_max_order);
2634 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2636 return 1;
2639 __setup("slub_max_order=", setup_slub_max_order);
2641 static int __init setup_slub_min_objects(char *str)
2643 get_option(&str, &slub_min_objects);
2645 return 1;
2648 __setup("slub_min_objects=", setup_slub_min_objects);
2650 static int __init setup_slub_nomerge(char *str)
2652 slub_nomerge = 1;
2653 return 1;
2656 __setup("slub_nomerge", setup_slub_nomerge);
2658 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2659 const char *name, int size, gfp_t gfp_flags)
2661 unsigned int flags = 0;
2663 if (gfp_flags & SLUB_DMA)
2664 flags = SLAB_CACHE_DMA;
2667 * This function is called with IRQs disabled during early-boot on
2668 * single CPU so there's no need to take slub_lock here.
2670 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2671 flags, NULL))
2672 goto panic;
2674 list_add(&s->list, &slab_caches);
2676 if (sysfs_slab_add(s))
2677 goto panic;
2678 return s;
2680 panic:
2681 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2684 #ifdef CONFIG_ZONE_DMA
2685 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2687 static void sysfs_add_func(struct work_struct *w)
2689 struct kmem_cache *s;
2691 down_write(&slub_lock);
2692 list_for_each_entry(s, &slab_caches, list) {
2693 if (s->flags & __SYSFS_ADD_DEFERRED) {
2694 s->flags &= ~__SYSFS_ADD_DEFERRED;
2695 sysfs_slab_add(s);
2698 up_write(&slub_lock);
2701 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2703 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2705 struct kmem_cache *s;
2706 char *text;
2707 size_t realsize;
2708 unsigned long slabflags;
2710 s = kmalloc_caches_dma[index];
2711 if (s)
2712 return s;
2714 /* Dynamically create dma cache */
2715 if (flags & __GFP_WAIT)
2716 down_write(&slub_lock);
2717 else {
2718 if (!down_write_trylock(&slub_lock))
2719 goto out;
2722 if (kmalloc_caches_dma[index])
2723 goto unlock_out;
2725 realsize = kmalloc_caches[index].objsize;
2726 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2727 (unsigned int)realsize);
2728 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2731 * Must defer sysfs creation to a workqueue because we don't know
2732 * what context we are called from. Before sysfs comes up, we don't
2733 * need to do anything because our sysfs initcall will start by
2734 * adding all existing slabs to sysfs.
2736 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK;
2737 if (slab_state >= SYSFS)
2738 slabflags |= __SYSFS_ADD_DEFERRED;
2740 if (!s || !text || !kmem_cache_open(s, flags, text,
2741 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) {
2742 kfree(s);
2743 kfree(text);
2744 goto unlock_out;
2747 list_add(&s->list, &slab_caches);
2748 kmalloc_caches_dma[index] = s;
2750 if (slab_state >= SYSFS)
2751 schedule_work(&sysfs_add_work);
2753 unlock_out:
2754 up_write(&slub_lock);
2755 out:
2756 return kmalloc_caches_dma[index];
2758 #endif
2761 * Conversion table for small slabs sizes / 8 to the index in the
2762 * kmalloc array. This is necessary for slabs < 192 since we have non power
2763 * of two cache sizes there. The size of larger slabs can be determined using
2764 * fls.
2766 static s8 size_index[24] = {
2767 3, /* 8 */
2768 4, /* 16 */
2769 5, /* 24 */
2770 5, /* 32 */
2771 6, /* 40 */
2772 6, /* 48 */
2773 6, /* 56 */
2774 6, /* 64 */
2775 1, /* 72 */
2776 1, /* 80 */
2777 1, /* 88 */
2778 1, /* 96 */
2779 7, /* 104 */
2780 7, /* 112 */
2781 7, /* 120 */
2782 7, /* 128 */
2783 2, /* 136 */
2784 2, /* 144 */
2785 2, /* 152 */
2786 2, /* 160 */
2787 2, /* 168 */
2788 2, /* 176 */
2789 2, /* 184 */
2790 2 /* 192 */
2793 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2795 int index;
2797 if (size <= 192) {
2798 if (!size)
2799 return ZERO_SIZE_PTR;
2801 index = size_index[(size - 1) / 8];
2802 } else
2803 index = fls(size - 1);
2805 #ifdef CONFIG_ZONE_DMA
2806 if (unlikely((flags & SLUB_DMA)))
2807 return dma_kmalloc_cache(index, flags);
2809 #endif
2810 return &kmalloc_caches[index];
2813 void *__kmalloc(size_t size, gfp_t flags)
2815 struct kmem_cache *s;
2816 void *ret;
2818 if (unlikely(size > SLUB_MAX_SIZE))
2819 return kmalloc_large(size, flags);
2821 s = get_slab(size, flags);
2823 if (unlikely(ZERO_OR_NULL_PTR(s)))
2824 return s;
2826 ret = slab_alloc(s, flags, -1, _RET_IP_);
2828 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2830 return ret;
2832 EXPORT_SYMBOL(__kmalloc);
2834 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2836 struct page *page;
2837 void *ptr = NULL;
2839 flags |= __GFP_COMP | __GFP_NOTRACK;
2840 page = alloc_pages_node(node, flags, get_order(size));
2841 if (page)
2842 ptr = page_address(page);
2844 kmemleak_alloc(ptr, size, 1, flags);
2845 return ptr;
2848 #ifdef CONFIG_NUMA
2849 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2851 struct kmem_cache *s;
2852 void *ret;
2854 if (unlikely(size > SLUB_MAX_SIZE)) {
2855 ret = kmalloc_large_node(size, flags, node);
2857 trace_kmalloc_node(_RET_IP_, ret,
2858 size, PAGE_SIZE << get_order(size),
2859 flags, node);
2861 return ret;
2864 s = get_slab(size, flags);
2866 if (unlikely(ZERO_OR_NULL_PTR(s)))
2867 return s;
2869 ret = slab_alloc(s, flags, node, _RET_IP_);
2871 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2873 return ret;
2875 EXPORT_SYMBOL(__kmalloc_node);
2876 #endif
2878 size_t ksize(const void *object)
2880 struct page *page;
2881 struct kmem_cache *s;
2883 if (unlikely(object == ZERO_SIZE_PTR))
2884 return 0;
2886 page = virt_to_head_page(object);
2888 if (unlikely(!PageSlab(page))) {
2889 WARN_ON(!PageCompound(page));
2890 return PAGE_SIZE << compound_order(page);
2892 s = page->slab;
2894 #ifdef CONFIG_SLUB_DEBUG
2896 * Debugging requires use of the padding between object
2897 * and whatever may come after it.
2899 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2900 return s->objsize;
2902 #endif
2904 * If we have the need to store the freelist pointer
2905 * back there or track user information then we can
2906 * only use the space before that information.
2908 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2909 return s->inuse;
2911 * Else we can use all the padding etc for the allocation
2913 return s->size;
2915 EXPORT_SYMBOL(ksize);
2917 void kfree(const void *x)
2919 struct page *page;
2920 void *object = (void *)x;
2922 trace_kfree(_RET_IP_, x);
2924 if (unlikely(ZERO_OR_NULL_PTR(x)))
2925 return;
2927 page = virt_to_head_page(x);
2928 if (unlikely(!PageSlab(page))) {
2929 BUG_ON(!PageCompound(page));
2930 kmemleak_free(x);
2931 put_page(page);
2932 return;
2934 slab_free(page->slab, page, object, _RET_IP_);
2936 EXPORT_SYMBOL(kfree);
2939 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2940 * the remaining slabs by the number of items in use. The slabs with the
2941 * most items in use come first. New allocations will then fill those up
2942 * and thus they can be removed from the partial lists.
2944 * The slabs with the least items are placed last. This results in them
2945 * being allocated from last increasing the chance that the last objects
2946 * are freed in them.
2948 int kmem_cache_shrink(struct kmem_cache *s)
2950 int node;
2951 int i;
2952 struct kmem_cache_node *n;
2953 struct page *page;
2954 struct page *t;
2955 int objects = oo_objects(s->max);
2956 struct list_head *slabs_by_inuse =
2957 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2958 unsigned long flags;
2960 if (!slabs_by_inuse)
2961 return -ENOMEM;
2963 flush_all(s);
2964 for_each_node_state(node, N_NORMAL_MEMORY) {
2965 n = get_node(s, node);
2967 if (!n->nr_partial)
2968 continue;
2970 for (i = 0; i < objects; i++)
2971 INIT_LIST_HEAD(slabs_by_inuse + i);
2973 spin_lock_irqsave(&n->list_lock, flags);
2976 * Build lists indexed by the items in use in each slab.
2978 * Note that concurrent frees may occur while we hold the
2979 * list_lock. page->inuse here is the upper limit.
2981 list_for_each_entry_safe(page, t, &n->partial, lru) {
2982 if (!page->inuse && slab_trylock(page)) {
2984 * Must hold slab lock here because slab_free
2985 * may have freed the last object and be
2986 * waiting to release the slab.
2988 list_del(&page->lru);
2989 n->nr_partial--;
2990 slab_unlock(page);
2991 discard_slab(s, page);
2992 } else {
2993 list_move(&page->lru,
2994 slabs_by_inuse + page->inuse);
2999 * Rebuild the partial list with the slabs filled up most
3000 * first and the least used slabs at the end.
3002 for (i = objects - 1; i >= 0; i--)
3003 list_splice(slabs_by_inuse + i, n->partial.prev);
3005 spin_unlock_irqrestore(&n->list_lock, flags);
3008 kfree(slabs_by_inuse);
3009 return 0;
3011 EXPORT_SYMBOL(kmem_cache_shrink);
3013 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
3014 static int slab_mem_going_offline_callback(void *arg)
3016 struct kmem_cache *s;
3018 down_read(&slub_lock);
3019 list_for_each_entry(s, &slab_caches, list)
3020 kmem_cache_shrink(s);
3021 up_read(&slub_lock);
3023 return 0;
3026 static void slab_mem_offline_callback(void *arg)
3028 struct kmem_cache_node *n;
3029 struct kmem_cache *s;
3030 struct memory_notify *marg = arg;
3031 int offline_node;
3033 offline_node = marg->status_change_nid;
3036 * If the node still has available memory. we need kmem_cache_node
3037 * for it yet.
3039 if (offline_node < 0)
3040 return;
3042 down_read(&slub_lock);
3043 list_for_each_entry(s, &slab_caches, list) {
3044 n = get_node(s, offline_node);
3045 if (n) {
3047 * if n->nr_slabs > 0, slabs still exist on the node
3048 * that is going down. We were unable to free them,
3049 * and offline_pages() function shoudn't call this
3050 * callback. So, we must fail.
3052 BUG_ON(slabs_node(s, offline_node));
3054 s->node[offline_node] = NULL;
3055 kmem_cache_free(kmalloc_caches, n);
3058 up_read(&slub_lock);
3061 static int slab_mem_going_online_callback(void *arg)
3063 struct kmem_cache_node *n;
3064 struct kmem_cache *s;
3065 struct memory_notify *marg = arg;
3066 int nid = marg->status_change_nid;
3067 int ret = 0;
3070 * If the node's memory is already available, then kmem_cache_node is
3071 * already created. Nothing to do.
3073 if (nid < 0)
3074 return 0;
3077 * We are bringing a node online. No memory is available yet. We must
3078 * allocate a kmem_cache_node structure in order to bring the node
3079 * online.
3081 down_read(&slub_lock);
3082 list_for_each_entry(s, &slab_caches, list) {
3084 * XXX: kmem_cache_alloc_node will fallback to other nodes
3085 * since memory is not yet available from the node that
3086 * is brought up.
3088 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3089 if (!n) {
3090 ret = -ENOMEM;
3091 goto out;
3093 init_kmem_cache_node(n, s);
3094 s->node[nid] = n;
3096 out:
3097 up_read(&slub_lock);
3098 return ret;
3101 static int slab_memory_callback(struct notifier_block *self,
3102 unsigned long action, void *arg)
3104 int ret = 0;
3106 switch (action) {
3107 case MEM_GOING_ONLINE:
3108 ret = slab_mem_going_online_callback(arg);
3109 break;
3110 case MEM_GOING_OFFLINE:
3111 ret = slab_mem_going_offline_callback(arg);
3112 break;
3113 case MEM_OFFLINE:
3114 case MEM_CANCEL_ONLINE:
3115 slab_mem_offline_callback(arg);
3116 break;
3117 case MEM_ONLINE:
3118 case MEM_CANCEL_OFFLINE:
3119 break;
3121 if (ret)
3122 ret = notifier_from_errno(ret);
3123 else
3124 ret = NOTIFY_OK;
3125 return ret;
3128 #endif /* CONFIG_MEMORY_HOTPLUG */
3130 /********************************************************************
3131 * Basic setup of slabs
3132 *******************************************************************/
3134 void __init kmem_cache_init(void)
3136 int i;
3137 int caches = 0;
3139 init_alloc_cpu();
3141 #ifdef CONFIG_NUMA
3143 * Must first have the slab cache available for the allocations of the
3144 * struct kmem_cache_node's. There is special bootstrap code in
3145 * kmem_cache_open for slab_state == DOWN.
3147 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3148 sizeof(struct kmem_cache_node), GFP_NOWAIT);
3149 kmalloc_caches[0].refcount = -1;
3150 caches++;
3152 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3153 #endif
3155 /* Able to allocate the per node structures */
3156 slab_state = PARTIAL;
3158 /* Caches that are not of the two-to-the-power-of size */
3159 if (KMALLOC_MIN_SIZE <= 64) {
3160 create_kmalloc_cache(&kmalloc_caches[1],
3161 "kmalloc-96", 96, GFP_NOWAIT);
3162 caches++;
3163 create_kmalloc_cache(&kmalloc_caches[2],
3164 "kmalloc-192", 192, GFP_NOWAIT);
3165 caches++;
3168 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3169 create_kmalloc_cache(&kmalloc_caches[i],
3170 "kmalloc", 1 << i, GFP_NOWAIT);
3171 caches++;
3176 * Patch up the size_index table if we have strange large alignment
3177 * requirements for the kmalloc array. This is only the case for
3178 * MIPS it seems. The standard arches will not generate any code here.
3180 * Largest permitted alignment is 256 bytes due to the way we
3181 * handle the index determination for the smaller caches.
3183 * Make sure that nothing crazy happens if someone starts tinkering
3184 * around with ARCH_KMALLOC_MINALIGN
3186 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3187 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3189 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3190 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3192 if (KMALLOC_MIN_SIZE == 128) {
3194 * The 192 byte sized cache is not used if the alignment
3195 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3196 * instead.
3198 for (i = 128 + 8; i <= 192; i += 8)
3199 size_index[(i - 1) / 8] = 8;
3202 slab_state = UP;
3204 /* Provide the correct kmalloc names now that the caches are up */
3205 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3206 kmalloc_caches[i]. name =
3207 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3209 #ifdef CONFIG_SMP
3210 register_cpu_notifier(&slab_notifier);
3211 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3212 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3213 #else
3214 kmem_size = sizeof(struct kmem_cache);
3215 #endif
3217 printk(KERN_INFO
3218 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3219 " CPUs=%d, Nodes=%d\n",
3220 caches, cache_line_size(),
3221 slub_min_order, slub_max_order, slub_min_objects,
3222 nr_cpu_ids, nr_node_ids);
3225 void __init kmem_cache_init_late(void)
3230 * Find a mergeable slab cache
3232 static int slab_unmergeable(struct kmem_cache *s)
3234 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3235 return 1;
3237 if (s->ctor)
3238 return 1;
3241 * We may have set a slab to be unmergeable during bootstrap.
3243 if (s->refcount < 0)
3244 return 1;
3246 return 0;
3249 static struct kmem_cache *find_mergeable(size_t size,
3250 size_t align, unsigned long flags, const char *name,
3251 void (*ctor)(void *))
3253 struct kmem_cache *s;
3255 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3256 return NULL;
3258 if (ctor)
3259 return NULL;
3261 size = ALIGN(size, sizeof(void *));
3262 align = calculate_alignment(flags, align, size);
3263 size = ALIGN(size, align);
3264 flags = kmem_cache_flags(size, flags, name, NULL);
3266 list_for_each_entry(s, &slab_caches, list) {
3267 if (slab_unmergeable(s))
3268 continue;
3270 if (size > s->size)
3271 continue;
3273 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3274 continue;
3276 * Check if alignment is compatible.
3277 * Courtesy of Adrian Drzewiecki
3279 if ((s->size & ~(align - 1)) != s->size)
3280 continue;
3282 if (s->size - size >= sizeof(void *))
3283 continue;
3285 return s;
3287 return NULL;
3290 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3291 size_t align, unsigned long flags, void (*ctor)(void *))
3293 struct kmem_cache *s;
3295 down_write(&slub_lock);
3296 s = find_mergeable(size, align, flags, name, ctor);
3297 if (s) {
3298 int cpu;
3300 s->refcount++;
3302 * Adjust the object sizes so that we clear
3303 * the complete object on kzalloc.
3305 s->objsize = max(s->objsize, (int)size);
3308 * And then we need to update the object size in the
3309 * per cpu structures
3311 for_each_online_cpu(cpu)
3312 get_cpu_slab(s, cpu)->objsize = s->objsize;
3314 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3315 up_write(&slub_lock);
3317 if (sysfs_slab_alias(s, name)) {
3318 down_write(&slub_lock);
3319 s->refcount--;
3320 up_write(&slub_lock);
3321 goto err;
3323 return s;
3326 s = kmalloc(kmem_size, GFP_KERNEL);
3327 if (s) {
3328 if (kmem_cache_open(s, GFP_KERNEL, name,
3329 size, align, flags, ctor)) {
3330 list_add(&s->list, &slab_caches);
3331 up_write(&slub_lock);
3332 if (sysfs_slab_add(s)) {
3333 down_write(&slub_lock);
3334 list_del(&s->list);
3335 up_write(&slub_lock);
3336 kfree(s);
3337 goto err;
3339 return s;
3341 kfree(s);
3343 up_write(&slub_lock);
3345 err:
3346 if (flags & SLAB_PANIC)
3347 panic("Cannot create slabcache %s\n", name);
3348 else
3349 s = NULL;
3350 return s;
3352 EXPORT_SYMBOL(kmem_cache_create);
3354 #ifdef CONFIG_SMP
3356 * Use the cpu notifier to insure that the cpu slabs are flushed when
3357 * necessary.
3359 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3360 unsigned long action, void *hcpu)
3362 long cpu = (long)hcpu;
3363 struct kmem_cache *s;
3364 unsigned long flags;
3366 switch (action) {
3367 case CPU_UP_PREPARE:
3368 case CPU_UP_PREPARE_FROZEN:
3369 init_alloc_cpu_cpu(cpu);
3370 down_read(&slub_lock);
3371 list_for_each_entry(s, &slab_caches, list)
3372 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3373 GFP_KERNEL);
3374 up_read(&slub_lock);
3375 break;
3377 case CPU_UP_CANCELED:
3378 case CPU_UP_CANCELED_FROZEN:
3379 case CPU_DEAD:
3380 case CPU_DEAD_FROZEN:
3381 down_read(&slub_lock);
3382 list_for_each_entry(s, &slab_caches, list) {
3383 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3385 local_irq_save(flags);
3386 __flush_cpu_slab(s, cpu);
3387 local_irq_restore(flags);
3388 free_kmem_cache_cpu(c, cpu);
3389 s->cpu_slab[cpu] = NULL;
3391 up_read(&slub_lock);
3392 break;
3393 default:
3394 break;
3396 return NOTIFY_OK;
3399 static struct notifier_block __cpuinitdata slab_notifier = {
3400 .notifier_call = slab_cpuup_callback
3403 #endif
3405 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3407 struct kmem_cache *s;
3408 void *ret;
3410 if (unlikely(size > SLUB_MAX_SIZE))
3411 return kmalloc_large(size, gfpflags);
3413 s = get_slab(size, gfpflags);
3415 if (unlikely(ZERO_OR_NULL_PTR(s)))
3416 return s;
3418 ret = slab_alloc(s, gfpflags, -1, caller);
3420 /* Honor the call site pointer we recieved. */
3421 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3423 return ret;
3426 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3427 int node, unsigned long caller)
3429 struct kmem_cache *s;
3430 void *ret;
3432 if (unlikely(size > SLUB_MAX_SIZE))
3433 return kmalloc_large_node(size, gfpflags, node);
3435 s = get_slab(size, gfpflags);
3437 if (unlikely(ZERO_OR_NULL_PTR(s)))
3438 return s;
3440 ret = slab_alloc(s, gfpflags, node, caller);
3442 /* Honor the call site pointer we recieved. */
3443 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3445 return ret;
3448 #ifdef CONFIG_SLUB_DEBUG
3449 static int count_inuse(struct page *page)
3451 return page->inuse;
3454 static int count_total(struct page *page)
3456 return page->objects;
3459 static int validate_slab(struct kmem_cache *s, struct page *page,
3460 unsigned long *map)
3462 void *p;
3463 void *addr = page_address(page);
3465 if (!check_slab(s, page) ||
3466 !on_freelist(s, page, NULL))
3467 return 0;
3469 /* Now we know that a valid freelist exists */
3470 bitmap_zero(map, page->objects);
3472 for_each_free_object(p, s, page->freelist) {
3473 set_bit(slab_index(p, s, addr), map);
3474 if (!check_object(s, page, p, 0))
3475 return 0;
3478 for_each_object(p, s, addr, page->objects)
3479 if (!test_bit(slab_index(p, s, addr), map))
3480 if (!check_object(s, page, p, 1))
3481 return 0;
3482 return 1;
3485 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3486 unsigned long *map)
3488 if (slab_trylock(page)) {
3489 validate_slab(s, page, map);
3490 slab_unlock(page);
3491 } else
3492 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3493 s->name, page);
3495 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3496 if (!PageSlubDebug(page))
3497 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3498 "on slab 0x%p\n", s->name, page);
3499 } else {
3500 if (PageSlubDebug(page))
3501 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3502 "slab 0x%p\n", s->name, page);
3506 static int validate_slab_node(struct kmem_cache *s,
3507 struct kmem_cache_node *n, unsigned long *map)
3509 unsigned long count = 0;
3510 struct page *page;
3511 unsigned long flags;
3513 spin_lock_irqsave(&n->list_lock, flags);
3515 list_for_each_entry(page, &n->partial, lru) {
3516 validate_slab_slab(s, page, map);
3517 count++;
3519 if (count != n->nr_partial)
3520 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3521 "counter=%ld\n", s->name, count, n->nr_partial);
3523 if (!(s->flags & SLAB_STORE_USER))
3524 goto out;
3526 list_for_each_entry(page, &n->full, lru) {
3527 validate_slab_slab(s, page, map);
3528 count++;
3530 if (count != atomic_long_read(&n->nr_slabs))
3531 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3532 "counter=%ld\n", s->name, count,
3533 atomic_long_read(&n->nr_slabs));
3535 out:
3536 spin_unlock_irqrestore(&n->list_lock, flags);
3537 return count;
3540 static long validate_slab_cache(struct kmem_cache *s)
3542 int node;
3543 unsigned long count = 0;
3544 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3545 sizeof(unsigned long), GFP_KERNEL);
3547 if (!map)
3548 return -ENOMEM;
3550 flush_all(s);
3551 for_each_node_state(node, N_NORMAL_MEMORY) {
3552 struct kmem_cache_node *n = get_node(s, node);
3554 count += validate_slab_node(s, n, map);
3556 kfree(map);
3557 return count;
3560 #ifdef SLUB_RESILIENCY_TEST
3561 static void resiliency_test(void)
3563 u8 *p;
3565 printk(KERN_ERR "SLUB resiliency testing\n");
3566 printk(KERN_ERR "-----------------------\n");
3567 printk(KERN_ERR "A. Corruption after allocation\n");
3569 p = kzalloc(16, GFP_KERNEL);
3570 p[16] = 0x12;
3571 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3572 " 0x12->0x%p\n\n", p + 16);
3574 validate_slab_cache(kmalloc_caches + 4);
3576 /* Hmmm... The next two are dangerous */
3577 p = kzalloc(32, GFP_KERNEL);
3578 p[32 + sizeof(void *)] = 0x34;
3579 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3580 " 0x34 -> -0x%p\n", p);
3581 printk(KERN_ERR
3582 "If allocated object is overwritten then not detectable\n\n");
3584 validate_slab_cache(kmalloc_caches + 5);
3585 p = kzalloc(64, GFP_KERNEL);
3586 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3587 *p = 0x56;
3588 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3590 printk(KERN_ERR
3591 "If allocated object is overwritten then not detectable\n\n");
3592 validate_slab_cache(kmalloc_caches + 6);
3594 printk(KERN_ERR "\nB. Corruption after free\n");
3595 p = kzalloc(128, GFP_KERNEL);
3596 kfree(p);
3597 *p = 0x78;
3598 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3599 validate_slab_cache(kmalloc_caches + 7);
3601 p = kzalloc(256, GFP_KERNEL);
3602 kfree(p);
3603 p[50] = 0x9a;
3604 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3606 validate_slab_cache(kmalloc_caches + 8);
3608 p = kzalloc(512, GFP_KERNEL);
3609 kfree(p);
3610 p[512] = 0xab;
3611 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3612 validate_slab_cache(kmalloc_caches + 9);
3614 #else
3615 static void resiliency_test(void) {};
3616 #endif
3619 * Generate lists of code addresses where slabcache objects are allocated
3620 * and freed.
3623 struct location {
3624 unsigned long count;
3625 unsigned long addr;
3626 long long sum_time;
3627 long min_time;
3628 long max_time;
3629 long min_pid;
3630 long max_pid;
3631 DECLARE_BITMAP(cpus, NR_CPUS);
3632 nodemask_t nodes;
3635 struct loc_track {
3636 unsigned long max;
3637 unsigned long count;
3638 struct location *loc;
3641 static void free_loc_track(struct loc_track *t)
3643 if (t->max)
3644 free_pages((unsigned long)t->loc,
3645 get_order(sizeof(struct location) * t->max));
3648 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3650 struct location *l;
3651 int order;
3653 order = get_order(sizeof(struct location) * max);
3655 l = (void *)__get_free_pages(flags, order);
3656 if (!l)
3657 return 0;
3659 if (t->count) {
3660 memcpy(l, t->loc, sizeof(struct location) * t->count);
3661 free_loc_track(t);
3663 t->max = max;
3664 t->loc = l;
3665 return 1;
3668 static int add_location(struct loc_track *t, struct kmem_cache *s,
3669 const struct track *track)
3671 long start, end, pos;
3672 struct location *l;
3673 unsigned long caddr;
3674 unsigned long age = jiffies - track->when;
3676 start = -1;
3677 end = t->count;
3679 for ( ; ; ) {
3680 pos = start + (end - start + 1) / 2;
3683 * There is nothing at "end". If we end up there
3684 * we need to add something to before end.
3686 if (pos == end)
3687 break;
3689 caddr = t->loc[pos].addr;
3690 if (track->addr == caddr) {
3692 l = &t->loc[pos];
3693 l->count++;
3694 if (track->when) {
3695 l->sum_time += age;
3696 if (age < l->min_time)
3697 l->min_time = age;
3698 if (age > l->max_time)
3699 l->max_time = age;
3701 if (track->pid < l->min_pid)
3702 l->min_pid = track->pid;
3703 if (track->pid > l->max_pid)
3704 l->max_pid = track->pid;
3706 cpumask_set_cpu(track->cpu,
3707 to_cpumask(l->cpus));
3709 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3710 return 1;
3713 if (track->addr < caddr)
3714 end = pos;
3715 else
3716 start = pos;
3720 * Not found. Insert new tracking element.
3722 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3723 return 0;
3725 l = t->loc + pos;
3726 if (pos < t->count)
3727 memmove(l + 1, l,
3728 (t->count - pos) * sizeof(struct location));
3729 t->count++;
3730 l->count = 1;
3731 l->addr = track->addr;
3732 l->sum_time = age;
3733 l->min_time = age;
3734 l->max_time = age;
3735 l->min_pid = track->pid;
3736 l->max_pid = track->pid;
3737 cpumask_clear(to_cpumask(l->cpus));
3738 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3739 nodes_clear(l->nodes);
3740 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3741 return 1;
3744 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3745 struct page *page, enum track_item alloc)
3747 void *addr = page_address(page);
3748 DECLARE_BITMAP(map, page->objects);
3749 void *p;
3751 bitmap_zero(map, page->objects);
3752 for_each_free_object(p, s, page->freelist)
3753 set_bit(slab_index(p, s, addr), map);
3755 for_each_object(p, s, addr, page->objects)
3756 if (!test_bit(slab_index(p, s, addr), map))
3757 add_location(t, s, get_track(s, p, alloc));
3760 static int list_locations(struct kmem_cache *s, char *buf,
3761 enum track_item alloc)
3763 int len = 0;
3764 unsigned long i;
3765 struct loc_track t = { 0, 0, NULL };
3766 int node;
3768 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3769 GFP_TEMPORARY))
3770 return sprintf(buf, "Out of memory\n");
3772 /* Push back cpu slabs */
3773 flush_all(s);
3775 for_each_node_state(node, N_NORMAL_MEMORY) {
3776 struct kmem_cache_node *n = get_node(s, node);
3777 unsigned long flags;
3778 struct page *page;
3780 if (!atomic_long_read(&n->nr_slabs))
3781 continue;
3783 spin_lock_irqsave(&n->list_lock, flags);
3784 list_for_each_entry(page, &n->partial, lru)
3785 process_slab(&t, s, page, alloc);
3786 list_for_each_entry(page, &n->full, lru)
3787 process_slab(&t, s, page, alloc);
3788 spin_unlock_irqrestore(&n->list_lock, flags);
3791 for (i = 0; i < t.count; i++) {
3792 struct location *l = &t.loc[i];
3794 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3795 break;
3796 len += sprintf(buf + len, "%7ld ", l->count);
3798 if (l->addr)
3799 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3800 else
3801 len += sprintf(buf + len, "<not-available>");
3803 if (l->sum_time != l->min_time) {
3804 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3805 l->min_time,
3806 (long)div_u64(l->sum_time, l->count),
3807 l->max_time);
3808 } else
3809 len += sprintf(buf + len, " age=%ld",
3810 l->min_time);
3812 if (l->min_pid != l->max_pid)
3813 len += sprintf(buf + len, " pid=%ld-%ld",
3814 l->min_pid, l->max_pid);
3815 else
3816 len += sprintf(buf + len, " pid=%ld",
3817 l->min_pid);
3819 if (num_online_cpus() > 1 &&
3820 !cpumask_empty(to_cpumask(l->cpus)) &&
3821 len < PAGE_SIZE - 60) {
3822 len += sprintf(buf + len, " cpus=");
3823 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3824 to_cpumask(l->cpus));
3827 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3828 len < PAGE_SIZE - 60) {
3829 len += sprintf(buf + len, " nodes=");
3830 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3831 l->nodes);
3834 len += sprintf(buf + len, "\n");
3837 free_loc_track(&t);
3838 if (!t.count)
3839 len += sprintf(buf, "No data\n");
3840 return len;
3843 enum slab_stat_type {
3844 SL_ALL, /* All slabs */
3845 SL_PARTIAL, /* Only partially allocated slabs */
3846 SL_CPU, /* Only slabs used for cpu caches */
3847 SL_OBJECTS, /* Determine allocated objects not slabs */
3848 SL_TOTAL /* Determine object capacity not slabs */
3851 #define SO_ALL (1 << SL_ALL)
3852 #define SO_PARTIAL (1 << SL_PARTIAL)
3853 #define SO_CPU (1 << SL_CPU)
3854 #define SO_OBJECTS (1 << SL_OBJECTS)
3855 #define SO_TOTAL (1 << SL_TOTAL)
3857 static ssize_t show_slab_objects(struct kmem_cache *s,
3858 char *buf, unsigned long flags)
3860 unsigned long total = 0;
3861 int node;
3862 int x;
3863 unsigned long *nodes;
3864 unsigned long *per_cpu;
3866 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3867 if (!nodes)
3868 return -ENOMEM;
3869 per_cpu = nodes + nr_node_ids;
3871 if (flags & SO_CPU) {
3872 int cpu;
3874 for_each_possible_cpu(cpu) {
3875 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3877 if (!c || c->node < 0)
3878 continue;
3880 if (c->page) {
3881 if (flags & SO_TOTAL)
3882 x = c->page->objects;
3883 else if (flags & SO_OBJECTS)
3884 x = c->page->inuse;
3885 else
3886 x = 1;
3888 total += x;
3889 nodes[c->node] += x;
3891 per_cpu[c->node]++;
3895 if (flags & SO_ALL) {
3896 for_each_node_state(node, N_NORMAL_MEMORY) {
3897 struct kmem_cache_node *n = get_node(s, node);
3899 if (flags & SO_TOTAL)
3900 x = atomic_long_read(&n->total_objects);
3901 else if (flags & SO_OBJECTS)
3902 x = atomic_long_read(&n->total_objects) -
3903 count_partial(n, count_free);
3905 else
3906 x = atomic_long_read(&n->nr_slabs);
3907 total += x;
3908 nodes[node] += x;
3911 } else if (flags & SO_PARTIAL) {
3912 for_each_node_state(node, N_NORMAL_MEMORY) {
3913 struct kmem_cache_node *n = get_node(s, node);
3915 if (flags & SO_TOTAL)
3916 x = count_partial(n, count_total);
3917 else if (flags & SO_OBJECTS)
3918 x = count_partial(n, count_inuse);
3919 else
3920 x = n->nr_partial;
3921 total += x;
3922 nodes[node] += x;
3925 x = sprintf(buf, "%lu", total);
3926 #ifdef CONFIG_NUMA
3927 for_each_node_state(node, N_NORMAL_MEMORY)
3928 if (nodes[node])
3929 x += sprintf(buf + x, " N%d=%lu",
3930 node, nodes[node]);
3931 #endif
3932 kfree(nodes);
3933 return x + sprintf(buf + x, "\n");
3936 static int any_slab_objects(struct kmem_cache *s)
3938 int node;
3940 for_each_online_node(node) {
3941 struct kmem_cache_node *n = get_node(s, node);
3943 if (!n)
3944 continue;
3946 if (atomic_long_read(&n->total_objects))
3947 return 1;
3949 return 0;
3952 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3953 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3955 struct slab_attribute {
3956 struct attribute attr;
3957 ssize_t (*show)(struct kmem_cache *s, char *buf);
3958 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3961 #define SLAB_ATTR_RO(_name) \
3962 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3964 #define SLAB_ATTR(_name) \
3965 static struct slab_attribute _name##_attr = \
3966 __ATTR(_name, 0644, _name##_show, _name##_store)
3968 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3970 return sprintf(buf, "%d\n", s->size);
3972 SLAB_ATTR_RO(slab_size);
3974 static ssize_t align_show(struct kmem_cache *s, char *buf)
3976 return sprintf(buf, "%d\n", s->align);
3978 SLAB_ATTR_RO(align);
3980 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3982 return sprintf(buf, "%d\n", s->objsize);
3984 SLAB_ATTR_RO(object_size);
3986 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3988 return sprintf(buf, "%d\n", oo_objects(s->oo));
3990 SLAB_ATTR_RO(objs_per_slab);
3992 static ssize_t order_store(struct kmem_cache *s,
3993 const char *buf, size_t length)
3995 unsigned long order;
3996 int err;
3998 err = strict_strtoul(buf, 10, &order);
3999 if (err)
4000 return err;
4002 if (order > slub_max_order || order < slub_min_order)
4003 return -EINVAL;
4005 calculate_sizes(s, order);
4006 return length;
4009 static ssize_t order_show(struct kmem_cache *s, char *buf)
4011 return sprintf(buf, "%d\n", oo_order(s->oo));
4013 SLAB_ATTR(order);
4015 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4017 return sprintf(buf, "%lu\n", s->min_partial);
4020 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4021 size_t length)
4023 unsigned long min;
4024 int err;
4026 err = strict_strtoul(buf, 10, &min);
4027 if (err)
4028 return err;
4030 set_min_partial(s, min);
4031 return length;
4033 SLAB_ATTR(min_partial);
4035 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4037 if (s->ctor) {
4038 int n = sprint_symbol(buf, (unsigned long)s->ctor);
4040 return n + sprintf(buf + n, "\n");
4042 return 0;
4044 SLAB_ATTR_RO(ctor);
4046 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4048 return sprintf(buf, "%d\n", s->refcount - 1);
4050 SLAB_ATTR_RO(aliases);
4052 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4054 return show_slab_objects(s, buf, SO_ALL);
4056 SLAB_ATTR_RO(slabs);
4058 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4060 return show_slab_objects(s, buf, SO_PARTIAL);
4062 SLAB_ATTR_RO(partial);
4064 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4066 return show_slab_objects(s, buf, SO_CPU);
4068 SLAB_ATTR_RO(cpu_slabs);
4070 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4072 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4074 SLAB_ATTR_RO(objects);
4076 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4078 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4080 SLAB_ATTR_RO(objects_partial);
4082 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4084 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4086 SLAB_ATTR_RO(total_objects);
4088 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4090 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4093 static ssize_t sanity_checks_store(struct kmem_cache *s,
4094 const char *buf, size_t length)
4096 s->flags &= ~SLAB_DEBUG_FREE;
4097 if (buf[0] == '1')
4098 s->flags |= SLAB_DEBUG_FREE;
4099 return length;
4101 SLAB_ATTR(sanity_checks);
4103 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4105 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4108 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4109 size_t length)
4111 s->flags &= ~SLAB_TRACE;
4112 if (buf[0] == '1')
4113 s->flags |= SLAB_TRACE;
4114 return length;
4116 SLAB_ATTR(trace);
4118 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4123 static ssize_t reclaim_account_store(struct kmem_cache *s,
4124 const char *buf, size_t length)
4126 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4127 if (buf[0] == '1')
4128 s->flags |= SLAB_RECLAIM_ACCOUNT;
4129 return length;
4131 SLAB_ATTR(reclaim_account);
4133 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4135 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4137 SLAB_ATTR_RO(hwcache_align);
4139 #ifdef CONFIG_ZONE_DMA
4140 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4142 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4144 SLAB_ATTR_RO(cache_dma);
4145 #endif
4147 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4149 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4151 SLAB_ATTR_RO(destroy_by_rcu);
4153 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4155 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4158 static ssize_t red_zone_store(struct kmem_cache *s,
4159 const char *buf, size_t length)
4161 if (any_slab_objects(s))
4162 return -EBUSY;
4164 s->flags &= ~SLAB_RED_ZONE;
4165 if (buf[0] == '1')
4166 s->flags |= SLAB_RED_ZONE;
4167 calculate_sizes(s, -1);
4168 return length;
4170 SLAB_ATTR(red_zone);
4172 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4174 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4177 static ssize_t poison_store(struct kmem_cache *s,
4178 const char *buf, size_t length)
4180 if (any_slab_objects(s))
4181 return -EBUSY;
4183 s->flags &= ~SLAB_POISON;
4184 if (buf[0] == '1')
4185 s->flags |= SLAB_POISON;
4186 calculate_sizes(s, -1);
4187 return length;
4189 SLAB_ATTR(poison);
4191 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4193 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4196 static ssize_t store_user_store(struct kmem_cache *s,
4197 const char *buf, size_t length)
4199 if (any_slab_objects(s))
4200 return -EBUSY;
4202 s->flags &= ~SLAB_STORE_USER;
4203 if (buf[0] == '1')
4204 s->flags |= SLAB_STORE_USER;
4205 calculate_sizes(s, -1);
4206 return length;
4208 SLAB_ATTR(store_user);
4210 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4212 return 0;
4215 static ssize_t validate_store(struct kmem_cache *s,
4216 const char *buf, size_t length)
4218 int ret = -EINVAL;
4220 if (buf[0] == '1') {
4221 ret = validate_slab_cache(s);
4222 if (ret >= 0)
4223 ret = length;
4225 return ret;
4227 SLAB_ATTR(validate);
4229 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4231 return 0;
4234 static ssize_t shrink_store(struct kmem_cache *s,
4235 const char *buf, size_t length)
4237 if (buf[0] == '1') {
4238 int rc = kmem_cache_shrink(s);
4240 if (rc)
4241 return rc;
4242 } else
4243 return -EINVAL;
4244 return length;
4246 SLAB_ATTR(shrink);
4248 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4250 if (!(s->flags & SLAB_STORE_USER))
4251 return -ENOSYS;
4252 return list_locations(s, buf, TRACK_ALLOC);
4254 SLAB_ATTR_RO(alloc_calls);
4256 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4258 if (!(s->flags & SLAB_STORE_USER))
4259 return -ENOSYS;
4260 return list_locations(s, buf, TRACK_FREE);
4262 SLAB_ATTR_RO(free_calls);
4264 #ifdef CONFIG_NUMA
4265 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4267 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4270 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4271 const char *buf, size_t length)
4273 unsigned long ratio;
4274 int err;
4276 err = strict_strtoul(buf, 10, &ratio);
4277 if (err)
4278 return err;
4280 if (ratio <= 100)
4281 s->remote_node_defrag_ratio = ratio * 10;
4283 return length;
4285 SLAB_ATTR(remote_node_defrag_ratio);
4286 #endif
4288 #ifdef CONFIG_SLUB_STATS
4289 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4291 unsigned long sum = 0;
4292 int cpu;
4293 int len;
4294 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4296 if (!data)
4297 return -ENOMEM;
4299 for_each_online_cpu(cpu) {
4300 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4302 data[cpu] = x;
4303 sum += x;
4306 len = sprintf(buf, "%lu", sum);
4308 #ifdef CONFIG_SMP
4309 for_each_online_cpu(cpu) {
4310 if (data[cpu] && len < PAGE_SIZE - 20)
4311 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4313 #endif
4314 kfree(data);
4315 return len + sprintf(buf + len, "\n");
4318 #define STAT_ATTR(si, text) \
4319 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4321 return show_stat(s, buf, si); \
4323 SLAB_ATTR_RO(text); \
4325 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4326 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4327 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4328 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4329 STAT_ATTR(FREE_FROZEN, free_frozen);
4330 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4331 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4332 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4333 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4334 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4335 STAT_ATTR(FREE_SLAB, free_slab);
4336 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4337 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4338 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4339 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4340 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4341 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4342 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4343 #endif
4345 static struct attribute *slab_attrs[] = {
4346 &slab_size_attr.attr,
4347 &object_size_attr.attr,
4348 &objs_per_slab_attr.attr,
4349 &order_attr.attr,
4350 &min_partial_attr.attr,
4351 &objects_attr.attr,
4352 &objects_partial_attr.attr,
4353 &total_objects_attr.attr,
4354 &slabs_attr.attr,
4355 &partial_attr.attr,
4356 &cpu_slabs_attr.attr,
4357 &ctor_attr.attr,
4358 &aliases_attr.attr,
4359 &align_attr.attr,
4360 &sanity_checks_attr.attr,
4361 &trace_attr.attr,
4362 &hwcache_align_attr.attr,
4363 &reclaim_account_attr.attr,
4364 &destroy_by_rcu_attr.attr,
4365 &red_zone_attr.attr,
4366 &poison_attr.attr,
4367 &store_user_attr.attr,
4368 &validate_attr.attr,
4369 &shrink_attr.attr,
4370 &alloc_calls_attr.attr,
4371 &free_calls_attr.attr,
4372 #ifdef CONFIG_ZONE_DMA
4373 &cache_dma_attr.attr,
4374 #endif
4375 #ifdef CONFIG_NUMA
4376 &remote_node_defrag_ratio_attr.attr,
4377 #endif
4378 #ifdef CONFIG_SLUB_STATS
4379 &alloc_fastpath_attr.attr,
4380 &alloc_slowpath_attr.attr,
4381 &free_fastpath_attr.attr,
4382 &free_slowpath_attr.attr,
4383 &free_frozen_attr.attr,
4384 &free_add_partial_attr.attr,
4385 &free_remove_partial_attr.attr,
4386 &alloc_from_partial_attr.attr,
4387 &alloc_slab_attr.attr,
4388 &alloc_refill_attr.attr,
4389 &free_slab_attr.attr,
4390 &cpuslab_flush_attr.attr,
4391 &deactivate_full_attr.attr,
4392 &deactivate_empty_attr.attr,
4393 &deactivate_to_head_attr.attr,
4394 &deactivate_to_tail_attr.attr,
4395 &deactivate_remote_frees_attr.attr,
4396 &order_fallback_attr.attr,
4397 #endif
4398 NULL
4401 static struct attribute_group slab_attr_group = {
4402 .attrs = slab_attrs,
4405 static ssize_t slab_attr_show(struct kobject *kobj,
4406 struct attribute *attr,
4407 char *buf)
4409 struct slab_attribute *attribute;
4410 struct kmem_cache *s;
4411 int err;
4413 attribute = to_slab_attr(attr);
4414 s = to_slab(kobj);
4416 if (!attribute->show)
4417 return -EIO;
4419 err = attribute->show(s, buf);
4421 return err;
4424 static ssize_t slab_attr_store(struct kobject *kobj,
4425 struct attribute *attr,
4426 const char *buf, size_t len)
4428 struct slab_attribute *attribute;
4429 struct kmem_cache *s;
4430 int err;
4432 attribute = to_slab_attr(attr);
4433 s = to_slab(kobj);
4435 if (!attribute->store)
4436 return -EIO;
4438 err = attribute->store(s, buf, len);
4440 return err;
4443 static void kmem_cache_release(struct kobject *kobj)
4445 struct kmem_cache *s = to_slab(kobj);
4447 kfree(s);
4450 static struct sysfs_ops slab_sysfs_ops = {
4451 .show = slab_attr_show,
4452 .store = slab_attr_store,
4455 static struct kobj_type slab_ktype = {
4456 .sysfs_ops = &slab_sysfs_ops,
4457 .release = kmem_cache_release
4460 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4462 struct kobj_type *ktype = get_ktype(kobj);
4464 if (ktype == &slab_ktype)
4465 return 1;
4466 return 0;
4469 static struct kset_uevent_ops slab_uevent_ops = {
4470 .filter = uevent_filter,
4473 static struct kset *slab_kset;
4475 #define ID_STR_LENGTH 64
4477 /* Create a unique string id for a slab cache:
4479 * Format :[flags-]size
4481 static char *create_unique_id(struct kmem_cache *s)
4483 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4484 char *p = name;
4486 BUG_ON(!name);
4488 *p++ = ':';
4490 * First flags affecting slabcache operations. We will only
4491 * get here for aliasable slabs so we do not need to support
4492 * too many flags. The flags here must cover all flags that
4493 * are matched during merging to guarantee that the id is
4494 * unique.
4496 if (s->flags & SLAB_CACHE_DMA)
4497 *p++ = 'd';
4498 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4499 *p++ = 'a';
4500 if (s->flags & SLAB_DEBUG_FREE)
4501 *p++ = 'F';
4502 if (!(s->flags & SLAB_NOTRACK))
4503 *p++ = 't';
4504 if (p != name + 1)
4505 *p++ = '-';
4506 p += sprintf(p, "%07d", s->size);
4507 BUG_ON(p > name + ID_STR_LENGTH - 1);
4508 return name;
4511 static int sysfs_slab_add(struct kmem_cache *s)
4513 int err;
4514 const char *name;
4515 int unmergeable;
4517 if (slab_state < SYSFS)
4518 /* Defer until later */
4519 return 0;
4521 unmergeable = slab_unmergeable(s);
4522 if (unmergeable) {
4524 * Slabcache can never be merged so we can use the name proper.
4525 * This is typically the case for debug situations. In that
4526 * case we can catch duplicate names easily.
4528 sysfs_remove_link(&slab_kset->kobj, s->name);
4529 name = s->name;
4530 } else {
4532 * Create a unique name for the slab as a target
4533 * for the symlinks.
4535 name = create_unique_id(s);
4538 s->kobj.kset = slab_kset;
4539 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4540 if (err) {
4541 kobject_put(&s->kobj);
4542 return err;
4545 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4546 if (err)
4547 return err;
4548 kobject_uevent(&s->kobj, KOBJ_ADD);
4549 if (!unmergeable) {
4550 /* Setup first alias */
4551 sysfs_slab_alias(s, s->name);
4552 kfree(name);
4554 return 0;
4557 static void sysfs_slab_remove(struct kmem_cache *s)
4559 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4560 kobject_del(&s->kobj);
4561 kobject_put(&s->kobj);
4565 * Need to buffer aliases during bootup until sysfs becomes
4566 * available lest we lose that information.
4568 struct saved_alias {
4569 struct kmem_cache *s;
4570 const char *name;
4571 struct saved_alias *next;
4574 static struct saved_alias *alias_list;
4576 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4578 struct saved_alias *al;
4580 if (slab_state == SYSFS) {
4582 * If we have a leftover link then remove it.
4584 sysfs_remove_link(&slab_kset->kobj, name);
4585 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4588 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4589 if (!al)
4590 return -ENOMEM;
4592 al->s = s;
4593 al->name = name;
4594 al->next = alias_list;
4595 alias_list = al;
4596 return 0;
4599 static int __init slab_sysfs_init(void)
4601 struct kmem_cache *s;
4602 int err;
4604 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4605 if (!slab_kset) {
4606 printk(KERN_ERR "Cannot register slab subsystem.\n");
4607 return -ENOSYS;
4610 slab_state = SYSFS;
4612 list_for_each_entry(s, &slab_caches, list) {
4613 err = sysfs_slab_add(s);
4614 if (err)
4615 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4616 " to sysfs\n", s->name);
4619 while (alias_list) {
4620 struct saved_alias *al = alias_list;
4622 alias_list = alias_list->next;
4623 err = sysfs_slab_alias(al->s, al->name);
4624 if (err)
4625 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4626 " %s to sysfs\n", s->name);
4627 kfree(al);
4630 resiliency_test();
4631 return 0;
4634 __initcall(slab_sysfs_init);
4635 #endif
4638 * The /proc/slabinfo ABI
4640 #ifdef CONFIG_SLABINFO
4641 static void print_slabinfo_header(struct seq_file *m)
4643 seq_puts(m, "slabinfo - version: 2.1\n");
4644 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4645 "<objperslab> <pagesperslab>");
4646 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4647 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4648 seq_putc(m, '\n');
4651 static void *s_start(struct seq_file *m, loff_t *pos)
4653 loff_t n = *pos;
4655 down_read(&slub_lock);
4656 if (!n)
4657 print_slabinfo_header(m);
4659 return seq_list_start(&slab_caches, *pos);
4662 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4664 return seq_list_next(p, &slab_caches, pos);
4667 static void s_stop(struct seq_file *m, void *p)
4669 up_read(&slub_lock);
4672 static int s_show(struct seq_file *m, void *p)
4674 unsigned long nr_partials = 0;
4675 unsigned long nr_slabs = 0;
4676 unsigned long nr_inuse = 0;
4677 unsigned long nr_objs = 0;
4678 unsigned long nr_free = 0;
4679 struct kmem_cache *s;
4680 int node;
4682 s = list_entry(p, struct kmem_cache, list);
4684 for_each_online_node(node) {
4685 struct kmem_cache_node *n = get_node(s, node);
4687 if (!n)
4688 continue;
4690 nr_partials += n->nr_partial;
4691 nr_slabs += atomic_long_read(&n->nr_slabs);
4692 nr_objs += atomic_long_read(&n->total_objects);
4693 nr_free += count_partial(n, count_free);
4696 nr_inuse = nr_objs - nr_free;
4698 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4699 nr_objs, s->size, oo_objects(s->oo),
4700 (1 << oo_order(s->oo)));
4701 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4702 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4703 0UL);
4704 seq_putc(m, '\n');
4705 return 0;
4708 static const struct seq_operations slabinfo_op = {
4709 .start = s_start,
4710 .next = s_next,
4711 .stop = s_stop,
4712 .show = s_show,
4715 static int slabinfo_open(struct inode *inode, struct file *file)
4717 return seq_open(file, &slabinfo_op);
4720 static const struct file_operations proc_slabinfo_operations = {
4721 .open = slabinfo_open,
4722 .read = seq_read,
4723 .llseek = seq_lseek,
4724 .release = seq_release,
4727 static int __init slab_proc_init(void)
4729 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4730 return 0;
4732 module_init(slab_proc_init);
4733 #endif /* CONFIG_SLABINFO */