blob: 4996fc7195528ba931dda8c042da44b2501b083e [file] [log] [blame]
Christoph Lameter81819f02007-05-06 14:49:36 -07001/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
7 *
Christoph Lametercde53532008-07-04 09:59:22 -07008 * (C) 2007 SGI, Christoph Lameter
Christoph Lameter81819f02007-05-06 14:49:36 -07009 */
10
11#include <linux/mm.h>
Nick Piggin1eb5ac62009-05-05 19:13:44 +100012#include <linux/swap.h> /* struct reclaim_state */
Christoph Lameter81819f02007-05-06 14:49:36 -070013#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>
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +040018#include <linux/proc_fs.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070019#include <linux/seq_file.h>
Zhaolei02af61b2009-04-10 14:26:18 +080020#include <linux/kmemtrace.h>
Vegard Nossum5a896d92008-04-04 00:54:48 +020021#include <linux/kmemcheck.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070022#include <linux/cpu.h>
23#include <linux/cpuset.h>
24#include <linux/mempolicy.h>
25#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070026#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070027#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070028#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070029#include <linux/math64.h>
Akinobu Mita773ff602008-12-23 19:37:01 +090030#include <linux/fault-inject.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070031
32/*
33 * Lock order:
34 * 1. slab_lock(page)
35 * 2. slab->list_lock
36 *
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.
43 *
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).
49 *
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.
55 *
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.
68 *
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.
73 *
74 * SLUB assigns one slab for allocation to each processor.
75 * Allocations only occur from these slabs called cpu slabs.
76 *
Christoph Lameter672bba32007-05-09 02:32:39 -070077 * 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
Christoph Lameter81819f02007-05-06 14:49:36 -070079 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070080 * We track full slabs for debugging purposes though because otherwise we
81 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070082 *
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.
86 *
87 * Overloading of page flags that are otherwise used for LRU management.
88 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070089 * 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.
97 *
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
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700101 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -0700102 * free objects in addition to the regular freelist
103 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -0700104 *
105 * PageError Slab requires special handling due to debug
106 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700107 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700108 */
109
Christoph Lameter5577bd82007-05-16 22:10:56 -0700110#ifdef CONFIG_SLUB_DEBUG
Andy Whitcroft8a380822008-07-23 21:27:18 -0700111#define SLABDEBUG 1
Christoph Lameter5577bd82007-05-16 22:10:56 -0700112#else
113#define SLABDEBUG 0
114#endif
115
Christoph Lameter81819f02007-05-06 14:49:36 -0700116/*
117 * Issues still to be resolved:
118 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700119 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
120 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700121 * - Variable sizing of the per node arrays
122 */
123
124/* Enable to test recovery from slab corruption on boot */
125#undef SLUB_RESILIENCY_TEST
126
Christoph Lameter81819f02007-05-06 14:49:36 -0700127/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700128 * 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.
130 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800131#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700132
Christoph Lameter2086d262007-05-06 14:49:46 -0700133/*
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.
137 */
138#define MAX_PARTIAL 10
139
Christoph Lameter81819f02007-05-06 14:49:36 -0700140#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
141 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700142
Christoph Lameter81819f02007-05-06 14:49:36 -0700143/*
David Rientjes3de47212009-07-27 18:30:35 -0700144 * Debugging flags that require metadata to be stored in the slab. These get
145 * disabled when slub_debug=O is used and a cache's min order increases with
146 * metadata.
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700147 */
David Rientjes3de47212009-07-27 18:30:35 -0700148#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700149
150/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700151 * Set of flags that will prevent slab merging
152 */
153#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
Catalin Marinas06f22f12009-06-11 13:23:18 +0100154 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700155
156#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
Vegard Nossum5a896d92008-04-04 00:54:48 +0200157 SLAB_CACHE_DMA | SLAB_NOTRACK)
Christoph Lameter81819f02007-05-06 14:49:36 -0700158
159#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700160#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700161#endif
162
163#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700164#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700165#endif
166
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400167#define OO_SHIFT 16
168#define OO_MASK ((1 << OO_SHIFT) - 1)
169#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
170
Christoph Lameter81819f02007-05-06 14:49:36 -0700171/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700172#define __OBJECT_POISON 0x80000000 /* Poison object */
173#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700174
175static int kmem_size = sizeof(struct kmem_cache);
176
177#ifdef CONFIG_SMP
178static struct notifier_block slab_notifier;
179#endif
180
181static enum {
182 DOWN, /* No slab functionality available */
183 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700184 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700185 SYSFS /* Sysfs up */
186} slab_state = DOWN;
187
188/* A list of all slab caches on the system */
189static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700190static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700191
Christoph Lameter02cbc872007-05-09 02:32:43 -0700192/*
193 * Tracking user of a slab.
194 */
195struct track {
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300196 unsigned long addr; /* Called from address */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700197 int cpu; /* Was running on cpu */
198 int pid; /* Pid context */
199 unsigned long when; /* When did the operation occur */
200};
201
202enum track_item { TRACK_ALLOC, TRACK_FREE };
203
Christoph Lameterf6acb632008-04-29 16:16:06 -0700204#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -0700205static int sysfs_slab_add(struct kmem_cache *);
206static int sysfs_slab_alias(struct kmem_cache *, const char *);
207static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800208
Christoph Lameter81819f02007-05-06 14:49:36 -0700209#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700210static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
211static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
212 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800213static inline void sysfs_slab_remove(struct kmem_cache *s)
214{
215 kfree(s);
216}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800217
Christoph Lameter81819f02007-05-06 14:49:36 -0700218#endif
219
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800220static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
221{
222#ifdef CONFIG_SLUB_STATS
223 c->stat[si]++;
224#endif
225}
226
Christoph Lameter81819f02007-05-06 14:49:36 -0700227/********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
230
231int slab_is_available(void)
232{
233 return slab_state >= UP;
234}
235
236static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
237{
238#ifdef CONFIG_NUMA
239 return s->node[node];
240#else
241 return &s->local_node;
242#endif
243}
244
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700245static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
246{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700247#ifdef CONFIG_SMP
248 return s->cpu_slab[cpu];
249#else
250 return &s->cpu_slab;
251#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700252}
253
Christoph Lameter6446faa2008-02-15 23:45:26 -0800254/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700255static inline int check_valid_pointer(struct kmem_cache *s,
256 struct page *page, const void *object)
257{
258 void *base;
259
Christoph Lametera973e9d2008-03-01 13:40:44 -0800260 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700261 return 1;
262
Christoph Lametera973e9d2008-03-01 13:40:44 -0800263 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300264 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700265 (object - base) % s->size) {
266 return 0;
267 }
268
269 return 1;
270}
271
Christoph Lameter81819f02007-05-06 14:49:36 -0700272/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700273 * Slow version of get and set free pointer.
274 *
275 * This version requires touching the cache lines of kmem_cache which
276 * we avoid to do in the fast alloc free paths. There we obtain the offset
277 * from the page struct.
278 */
279static inline void *get_freepointer(struct kmem_cache *s, void *object)
280{
281 return *(void **)(object + s->offset);
282}
283
284static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
285{
286 *(void **)(object + s->offset) = fp;
287}
288
289/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300290#define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700292 __p += (__s)->size)
293
294/* Scan freelist */
295#define for_each_free_object(__p, __s, __free) \
Christoph Lametera973e9d2008-03-01 13:40:44 -0800296 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700297
298/* Determine object index from a given position */
299static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
300{
301 return (p - addr) / s->size;
302}
303
Christoph Lameter834f3d12008-04-14 19:11:31 +0300304static inline struct kmem_cache_order_objects oo_make(int order,
305 unsigned long size)
306{
307 struct kmem_cache_order_objects x = {
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400308 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
Christoph Lameter834f3d12008-04-14 19:11:31 +0300309 };
310
311 return x;
312}
313
314static inline int oo_order(struct kmem_cache_order_objects x)
315{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400316 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300317}
318
319static inline int oo_objects(struct kmem_cache_order_objects x)
320{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400321 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300322}
323
Christoph Lameter41ecc552007-05-09 02:32:44 -0700324#ifdef CONFIG_SLUB_DEBUG
325/*
326 * Debug settings:
327 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700328#ifdef CONFIG_SLUB_DEBUG_ON
329static int slub_debug = DEBUG_DEFAULT_FLAGS;
330#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700331static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700332#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700333
334static char *slub_debug_slabs;
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700335static int disable_higher_order_debug;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700336
Christoph Lameter7656c722007-05-09 02:32:40 -0700337/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700338 * Object debugging
339 */
340static void print_section(char *text, u8 *addr, unsigned int length)
341{
342 int i, offset;
343 int newline = 1;
344 char ascii[17];
345
346 ascii[16] = 0;
347
348 for (i = 0; i < length; i++) {
349 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700350 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700351 newline = 0;
352 }
Pekka Enberg06428782008-01-07 23:20:27 -0800353 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700354 offset = i % 16;
355 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
356 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800357 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700358 newline = 1;
359 }
360 }
361 if (!newline) {
362 i %= 16;
363 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800364 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700365 ascii[i] = ' ';
366 i++;
367 }
Pekka Enberg06428782008-01-07 23:20:27 -0800368 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700369 }
370}
371
Christoph Lameter81819f02007-05-06 14:49:36 -0700372static struct track *get_track(struct kmem_cache *s, void *object,
373 enum track_item alloc)
374{
375 struct track *p;
376
377 if (s->offset)
378 p = object + s->offset + sizeof(void *);
379 else
380 p = object + s->inuse;
381
382 return p + alloc;
383}
384
385static void set_track(struct kmem_cache *s, void *object,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300386 enum track_item alloc, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700387{
Akinobu Mita1a00df42009-03-07 00:36:21 +0900388 struct track *p = get_track(s, object, alloc);
Christoph Lameter81819f02007-05-06 14:49:36 -0700389
Christoph Lameter81819f02007-05-06 14:49:36 -0700390 if (addr) {
391 p->addr = addr;
392 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400393 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700394 p->when = jiffies;
395 } else
396 memset(p, 0, sizeof(struct track));
397}
398
Christoph Lameter81819f02007-05-06 14:49:36 -0700399static void init_tracking(struct kmem_cache *s, void *object)
400{
Christoph Lameter24922682007-07-17 04:03:18 -0700401 if (!(s->flags & SLAB_STORE_USER))
402 return;
403
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300404 set_track(s, object, TRACK_FREE, 0UL);
405 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700406}
407
408static void print_track(const char *s, struct track *t)
409{
410 if (!t->addr)
411 return;
412
Linus Torvalds7daf7052008-07-14 12:12:53 -0700413 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300414 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700415}
416
Christoph Lameter24922682007-07-17 04:03:18 -0700417static void print_tracking(struct kmem_cache *s, void *object)
418{
419 if (!(s->flags & SLAB_STORE_USER))
420 return;
421
422 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
423 print_track("Freed", get_track(s, object, TRACK_FREE));
424}
425
426static void print_page_info(struct page *page)
427{
Christoph Lameter39b26462008-04-14 19:11:30 +0300428 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
429 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700430
431}
432
433static void slab_bug(struct kmem_cache *s, char *fmt, ...)
434{
435 va_list args;
436 char buf[100];
437
438 va_start(args, fmt);
439 vsnprintf(buf, sizeof(buf), fmt, args);
440 va_end(args);
441 printk(KERN_ERR "========================================"
442 "=====================================\n");
443 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
444 printk(KERN_ERR "----------------------------------------"
445 "-------------------------------------\n\n");
446}
447
448static void slab_fix(struct kmem_cache *s, char *fmt, ...)
449{
450 va_list args;
451 char buf[100];
452
453 va_start(args, fmt);
454 vsnprintf(buf, sizeof(buf), fmt, args);
455 va_end(args);
456 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
457}
458
459static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700460{
461 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800462 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700463
464 print_tracking(s, p);
465
466 print_page_info(page);
467
468 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
469 p, p - addr, get_freepointer(s, p));
470
471 if (p > addr + 16)
472 print_section("Bytes b4", p - 16, 16);
473
Pekka Enberg0ebd6522008-07-19 14:17:22 +0300474 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700475
476 if (s->flags & SLAB_RED_ZONE)
477 print_section("Redzone", p + s->objsize,
478 s->inuse - s->objsize);
479
Christoph Lameter81819f02007-05-06 14:49:36 -0700480 if (s->offset)
481 off = s->offset + sizeof(void *);
482 else
483 off = s->inuse;
484
Christoph Lameter24922682007-07-17 04:03:18 -0700485 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700486 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700487
488 if (off != s->size)
489 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700490 print_section("Padding", p + off, s->size - off);
491
492 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700493}
494
495static void object_err(struct kmem_cache *s, struct page *page,
496 u8 *object, char *reason)
497{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700498 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700499 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700500}
501
Christoph Lameter24922682007-07-17 04:03:18 -0700502static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700503{
504 va_list args;
505 char buf[100];
506
Christoph Lameter24922682007-07-17 04:03:18 -0700507 va_start(args, fmt);
508 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700509 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700510 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700511 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700512 dump_stack();
513}
514
515static void init_object(struct kmem_cache *s, void *object, int active)
516{
517 u8 *p = object;
518
519 if (s->flags & __OBJECT_POISON) {
520 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800521 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700522 }
523
524 if (s->flags & SLAB_RED_ZONE)
525 memset(p + s->objsize,
526 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
527 s->inuse - s->objsize);
528}
529
Christoph Lameter24922682007-07-17 04:03:18 -0700530static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700531{
532 while (bytes) {
533 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700534 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700535 start++;
536 bytes--;
537 }
Christoph Lameter24922682007-07-17 04:03:18 -0700538 return NULL;
539}
540
541static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
542 void *from, void *to)
543{
544 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
545 memset(from, data, to - from);
546}
547
548static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
549 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800550 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700551{
552 u8 *fault;
553 u8 *end;
554
555 fault = check_bytes(start, value, bytes);
556 if (!fault)
557 return 1;
558
559 end = start + bytes;
560 while (end > fault && end[-1] == value)
561 end--;
562
563 slab_bug(s, "%s overwritten", what);
564 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
565 fault, end - 1, fault[0], value);
566 print_trailer(s, page, object);
567
568 restore_bytes(s, what, value, fault, end);
569 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700570}
571
Christoph Lameter81819f02007-05-06 14:49:36 -0700572/*
573 * Object layout:
574 *
575 * object address
576 * Bytes of the object to be managed.
577 * If the freepointer may overlay the object then the free
578 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700579 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700580 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
581 * 0xa5 (POISON_END)
582 *
583 * object + s->objsize
584 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700585 * Padding is extended by another word if Redzoning is enabled and
586 * objsize == inuse.
587 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700588 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
589 * 0xcc (RED_ACTIVE) for objects in use.
590 *
591 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700592 * Meta data starts here.
593 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700594 * A. Free pointer (if we cannot overwrite object on free)
595 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700596 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800597 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700598 * before the word boundary.
599 *
600 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700601 *
602 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700603 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700604 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700605 * If slabcaches are merged then the objsize and inuse boundaries are mostly
606 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700607 * may be used with merged slabcaches.
608 */
609
Christoph Lameter81819f02007-05-06 14:49:36 -0700610static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
611{
612 unsigned long off = s->inuse; /* The end of info */
613
614 if (s->offset)
615 /* Freepointer is placed after the object. */
616 off += sizeof(void *);
617
618 if (s->flags & SLAB_STORE_USER)
619 /* We also have user information there */
620 off += 2 * sizeof(struct track);
621
622 if (s->size == off)
623 return 1;
624
Christoph Lameter24922682007-07-17 04:03:18 -0700625 return check_bytes_and_report(s, page, p, "Object padding",
626 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700627}
628
Christoph Lameter39b26462008-04-14 19:11:30 +0300629/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700630static int slab_pad_check(struct kmem_cache *s, struct page *page)
631{
Christoph Lameter24922682007-07-17 04:03:18 -0700632 u8 *start;
633 u8 *fault;
634 u8 *end;
635 int length;
636 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700637
638 if (!(s->flags & SLAB_POISON))
639 return 1;
640
Christoph Lametera973e9d2008-03-01 13:40:44 -0800641 start = page_address(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +0300642 length = (PAGE_SIZE << compound_order(page));
Christoph Lameter39b26462008-04-14 19:11:30 +0300643 end = start + length;
644 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700645 if (!remainder)
646 return 1;
647
Christoph Lameter39b26462008-04-14 19:11:30 +0300648 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700649 if (!fault)
650 return 1;
651 while (end > fault && end[-1] == POISON_INUSE)
652 end--;
653
654 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Christoph Lameter39b26462008-04-14 19:11:30 +0300655 print_section("Padding", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700656
Eric Dumazet8a3d2712009-09-03 16:08:06 +0200657 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
Christoph Lameter24922682007-07-17 04:03:18 -0700658 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700659}
660
661static int check_object(struct kmem_cache *s, struct page *page,
662 void *object, int active)
663{
664 u8 *p = object;
665 u8 *endobject = object + s->objsize;
666
667 if (s->flags & SLAB_RED_ZONE) {
668 unsigned int red =
669 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
670
Christoph Lameter24922682007-07-17 04:03:18 -0700671 if (!check_bytes_and_report(s, page, object, "Redzone",
672 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700673 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700674 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800675 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
676 check_bytes_and_report(s, page, p, "Alignment padding",
677 endobject, POISON_INUSE, s->inuse - s->objsize);
678 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700679 }
680
681 if (s->flags & SLAB_POISON) {
682 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700683 (!check_bytes_and_report(s, page, p, "Poison", p,
684 POISON_FREE, s->objsize - 1) ||
685 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800686 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700687 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700688 /*
689 * check_pad_bytes cleans up on its own.
690 */
691 check_pad_bytes(s, page, p);
692 }
693
694 if (!s->offset && active)
695 /*
696 * Object and freepointer overlap. Cannot check
697 * freepointer while object is allocated.
698 */
699 return 1;
700
701 /* Check free pointer validity */
702 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
703 object_err(s, page, p, "Freepointer corrupt");
704 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100705 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700706 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700707 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700708 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800709 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700710 return 0;
711 }
712 return 1;
713}
714
715static int check_slab(struct kmem_cache *s, struct page *page)
716{
Christoph Lameter39b26462008-04-14 19:11:30 +0300717 int maxobj;
718
Christoph Lameter81819f02007-05-06 14:49:36 -0700719 VM_BUG_ON(!irqs_disabled());
720
721 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700722 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700723 return 0;
724 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300725
726 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
727 if (page->objects > maxobj) {
728 slab_err(s, page, "objects %u > max %u",
729 s->name, page->objects, maxobj);
730 return 0;
731 }
732 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700733 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300734 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700735 return 0;
736 }
737 /* Slab_pad_check fixes things up after itself */
738 slab_pad_check(s, page);
739 return 1;
740}
741
742/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700743 * Determine if a certain object on a page is on the freelist. Must hold the
744 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700745 */
746static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
747{
748 int nr = 0;
749 void *fp = page->freelist;
750 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300751 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700752
Christoph Lameter39b26462008-04-14 19:11:30 +0300753 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700754 if (fp == search)
755 return 1;
756 if (!check_valid_pointer(s, page, fp)) {
757 if (object) {
758 object_err(s, page, object,
759 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800760 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700761 break;
762 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700763 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800764 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300765 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700766 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700767 return 0;
768 }
769 break;
770 }
771 object = fp;
772 fp = get_freepointer(s, object);
773 nr++;
774 }
775
Christoph Lameter224a88b2008-04-14 19:11:31 +0300776 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400777 if (max_objects > MAX_OBJS_PER_PAGE)
778 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300779
780 if (page->objects != max_objects) {
781 slab_err(s, page, "Wrong number of objects. Found %d but "
782 "should be %d", page->objects, max_objects);
783 page->objects = max_objects;
784 slab_fix(s, "Number of objects adjusted.");
785 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300786 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700787 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300788 "counted were %d", page->inuse, page->objects - nr);
789 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700790 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700791 }
792 return search == NULL;
793}
794
Christoph Lameter0121c6192008-04-29 16:11:12 -0700795static void trace(struct kmem_cache *s, struct page *page, void *object,
796 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700797{
798 if (s->flags & SLAB_TRACE) {
799 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
800 s->name,
801 alloc ? "alloc" : "free",
802 object, page->inuse,
803 page->freelist);
804
805 if (!alloc)
806 print_section("Object", (void *)object, s->objsize);
807
808 dump_stack();
809 }
810}
811
Christoph Lameter643b1132007-05-06 14:49:42 -0700812/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700813 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700814 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700815static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700816{
Christoph Lameter643b1132007-05-06 14:49:42 -0700817 spin_lock(&n->list_lock);
818 list_add(&page->lru, &n->full);
819 spin_unlock(&n->list_lock);
820}
821
822static void remove_full(struct kmem_cache *s, struct page *page)
823{
824 struct kmem_cache_node *n;
825
826 if (!(s->flags & SLAB_STORE_USER))
827 return;
828
829 n = get_node(s, page_to_nid(page));
830
831 spin_lock(&n->list_lock);
832 list_del(&page->lru);
833 spin_unlock(&n->list_lock);
834}
835
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300836/* Tracking of the number of slabs for debugging purposes */
837static inline unsigned long slabs_node(struct kmem_cache *s, int node)
838{
839 struct kmem_cache_node *n = get_node(s, node);
840
841 return atomic_long_read(&n->nr_slabs);
842}
843
Alexander Beregalov26c02cf2009-06-11 14:08:48 +0400844static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
845{
846 return atomic_long_read(&n->nr_slabs);
847}
848
Christoph Lameter205ab992008-04-14 19:11:40 +0300849static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300850{
851 struct kmem_cache_node *n = get_node(s, node);
852
853 /*
854 * May be called early in order to allocate a slab for the
855 * kmem_cache_node structure. Solve the chicken-egg
856 * dilemma by deferring the increment of the count during
857 * bootstrap (see early_kmem_cache_node_alloc).
858 */
Christoph Lameter205ab992008-04-14 19:11:40 +0300859 if (!NUMA_BUILD || n) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300860 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300861 atomic_long_add(objects, &n->total_objects);
862 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300863}
Christoph Lameter205ab992008-04-14 19:11:40 +0300864static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300865{
866 struct kmem_cache_node *n = get_node(s, node);
867
868 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300869 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300870}
871
872/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -0700873static void setup_object_debug(struct kmem_cache *s, struct page *page,
874 void *object)
875{
876 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
877 return;
878
879 init_object(s, object, 0);
880 init_tracking(s, object);
881}
882
883static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300884 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700885{
886 if (!check_slab(s, page))
887 goto bad;
888
Christoph Lameterd692ef62008-02-15 23:45:24 -0800889 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700890 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700891 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700892 }
893
894 if (!check_valid_pointer(s, page, object)) {
895 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700896 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700897 }
898
Christoph Lameterd692ef62008-02-15 23:45:24 -0800899 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700900 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700901
Christoph Lameter3ec09742007-05-16 22:11:00 -0700902 /* Success perform special debug activities for allocs */
903 if (s->flags & SLAB_STORE_USER)
904 set_track(s, object, TRACK_ALLOC, addr);
905 trace(s, page, object, 1);
906 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700907 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700908
Christoph Lameter81819f02007-05-06 14:49:36 -0700909bad:
910 if (PageSlab(page)) {
911 /*
912 * If this is a slab page then lets do the best we can
913 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700914 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700915 */
Christoph Lameter24922682007-07-17 04:03:18 -0700916 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +0300917 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800918 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700919 }
920 return 0;
921}
922
Christoph Lameter3ec09742007-05-16 22:11:00 -0700923static int free_debug_processing(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300924 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700925{
926 if (!check_slab(s, page))
927 goto fail;
928
929 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700930 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700931 goto fail;
932 }
933
934 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700935 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700936 goto fail;
937 }
938
939 if (!check_object(s, page, object, 1))
940 return 0;
941
942 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800943 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700944 slab_err(s, page, "Attempt to free object(0x%p) "
945 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800946 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700947 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700948 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700949 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700950 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800951 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700952 object_err(s, page, object,
953 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700954 goto fail;
955 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700956
957 /* Special debug activities for freeing objects */
Andy Whitcroft8a380822008-07-23 21:27:18 -0700958 if (!PageSlubFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700959 remove_full(s, page);
960 if (s->flags & SLAB_STORE_USER)
961 set_track(s, object, TRACK_FREE, addr);
962 trace(s, page, object, 0);
963 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700964 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700965
Christoph Lameter81819f02007-05-06 14:49:36 -0700966fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700967 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700968 return 0;
969}
970
Christoph Lameter41ecc552007-05-09 02:32:44 -0700971static int __init setup_slub_debug(char *str)
972{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700973 slub_debug = DEBUG_DEFAULT_FLAGS;
974 if (*str++ != '=' || !*str)
975 /*
976 * No options specified. Switch on full debugging.
977 */
978 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700979
980 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700981 /*
982 * No options but restriction on slabs. This means full
983 * debugging for slabs matching a pattern.
984 */
985 goto check_slabs;
986
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700987 if (tolower(*str) == 'o') {
988 /*
989 * Avoid enabling debugging on caches if its minimum order
990 * would increase as a result.
991 */
992 disable_higher_order_debug = 1;
993 goto out;
994 }
995
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700996 slub_debug = 0;
997 if (*str == '-')
998 /*
999 * Switch off all debugging measures.
1000 */
1001 goto out;
1002
1003 /*
1004 * Determine which debug features should be switched on
1005 */
Pekka Enberg06428782008-01-07 23:20:27 -08001006 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001007 switch (tolower(*str)) {
1008 case 'f':
1009 slub_debug |= SLAB_DEBUG_FREE;
1010 break;
1011 case 'z':
1012 slub_debug |= SLAB_RED_ZONE;
1013 break;
1014 case 'p':
1015 slub_debug |= SLAB_POISON;
1016 break;
1017 case 'u':
1018 slub_debug |= SLAB_STORE_USER;
1019 break;
1020 case 't':
1021 slub_debug |= SLAB_TRACE;
1022 break;
1023 default:
1024 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001025 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001026 }
1027 }
1028
1029check_slabs:
1030 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001031 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001032out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001033 return 1;
1034}
1035
1036__setup("slub_debug", setup_slub_debug);
1037
Christoph Lameterba0268a2007-09-11 15:24:11 -07001038static unsigned long kmem_cache_flags(unsigned long objsize,
1039 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001040 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001041{
1042 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001043 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001044 */
Christoph Lametere1533622008-02-15 23:45:24 -08001045 if (slub_debug && (!slub_debug_slabs ||
David Rientjes3de47212009-07-27 18:30:35 -07001046 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1047 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001048
1049 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001050}
1051#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001052static inline void setup_object_debug(struct kmem_cache *s,
1053 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001054
Christoph Lameter3ec09742007-05-16 22:11:00 -07001055static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001056 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001057
Christoph Lameter3ec09742007-05-16 22:11:00 -07001058static inline int free_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001059 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001060
Christoph Lameter41ecc552007-05-09 02:32:44 -07001061static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1062 { return 1; }
1063static inline int check_object(struct kmem_cache *s, struct page *page,
1064 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001065static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001066static inline unsigned long kmem_cache_flags(unsigned long objsize,
1067 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001068 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001069{
1070 return flags;
1071}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001072#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001073
Ingo Molnarfdaa45e2009-09-15 11:00:26 +02001074#define disable_higher_order_debug 0
1075
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001076static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1077 { return 0; }
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001078static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1079 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001080static inline void inc_slabs_node(struct kmem_cache *s, int node,
1081 int objects) {}
1082static inline void dec_slabs_node(struct kmem_cache *s, int node,
1083 int objects) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001084#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03001085
Christoph Lameter81819f02007-05-06 14:49:36 -07001086/*
1087 * Slab allocation and freeing
1088 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001089static inline struct page *alloc_slab_page(gfp_t flags, int node,
1090 struct kmem_cache_order_objects oo)
1091{
1092 int order = oo_order(oo);
1093
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001094 flags |= __GFP_NOTRACK;
1095
Christoph Lameter65c33762008-04-14 19:11:40 +03001096 if (node == -1)
1097 return alloc_pages(flags, order);
1098 else
1099 return alloc_pages_node(node, flags, order);
1100}
1101
Christoph Lameter81819f02007-05-06 14:49:36 -07001102static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1103{
Pekka Enberg06428782008-01-07 23:20:27 -08001104 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001105 struct kmem_cache_order_objects oo = s->oo;
Pekka Enbergba522702009-06-24 21:59:51 +03001106 gfp_t alloc_gfp;
Christoph Lameter81819f02007-05-06 14:49:36 -07001107
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001108 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001109
Pekka Enbergba522702009-06-24 21:59:51 +03001110 /*
1111 * Let the initial higher-order allocation fail under memory pressure
1112 * so we fall-back to the minimum order allocation.
1113 */
1114 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1115
1116 page = alloc_slab_page(alloc_gfp, node, oo);
Christoph Lameter65c33762008-04-14 19:11:40 +03001117 if (unlikely(!page)) {
1118 oo = s->min;
1119 /*
1120 * Allocation may have failed due to fragmentation.
1121 * Try a lower order alloc if possible
1122 */
1123 page = alloc_slab_page(flags, node, oo);
1124 if (!page)
1125 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001126
Christoph Lameter65c33762008-04-14 19:11:40 +03001127 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1128 }
Vegard Nossum5a896d92008-04-04 00:54:48 +02001129
1130 if (kmemcheck_enabled
Amerigo Wang5086c382009-08-19 21:44:13 +03001131 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001132 int pages = 1 << oo_order(oo);
1133
1134 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1135
1136 /*
1137 * Objects from caches that have a constructor don't get
1138 * cleared when they're allocated, so we need to do it here.
1139 */
1140 if (s->ctor)
1141 kmemcheck_mark_uninitialized_pages(page, pages);
1142 else
1143 kmemcheck_mark_unallocated_pages(page, pages);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001144 }
1145
Christoph Lameter834f3d12008-04-14 19:11:31 +03001146 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001147 mod_zone_page_state(page_zone(page),
1148 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1149 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001150 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001151
1152 return page;
1153}
1154
1155static void setup_object(struct kmem_cache *s, struct page *page,
1156 void *object)
1157{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001158 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001159 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001160 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001161}
1162
1163static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1164{
1165 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001166 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001167 void *last;
1168 void *p;
1169
Christoph Lameter6cb06222007-10-16 01:25:41 -07001170 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001171
Christoph Lameter6cb06222007-10-16 01:25:41 -07001172 page = allocate_slab(s,
1173 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001174 if (!page)
1175 goto out;
1176
Christoph Lameter205ab992008-04-14 19:11:40 +03001177 inc_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001178 page->slab = s;
1179 page->flags |= 1 << PG_slab;
1180 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1181 SLAB_STORE_USER | SLAB_TRACE))
Andy Whitcroft8a380822008-07-23 21:27:18 -07001182 __SetPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001183
1184 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001185
1186 if (unlikely(s->flags & SLAB_POISON))
Christoph Lameter834f3d12008-04-14 19:11:31 +03001187 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001188
1189 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001190 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001191 setup_object(s, page, last);
1192 set_freepointer(s, last, p);
1193 last = p;
1194 }
1195 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001196 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001197
1198 page->freelist = start;
1199 page->inuse = 0;
1200out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001201 return page;
1202}
1203
1204static void __free_slab(struct kmem_cache *s, struct page *page)
1205{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001206 int order = compound_order(page);
1207 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001208
Andy Whitcroft8a380822008-07-23 21:27:18 -07001209 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001210 void *p;
1211
1212 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001213 for_each_object(p, s, page_address(page),
1214 page->objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001215 check_object(s, page, p, 0);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001216 __ClearPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001217 }
1218
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001219 kmemcheck_free_shadow(page, compound_order(page));
Vegard Nossum5a896d92008-04-04 00:54:48 +02001220
Christoph Lameter81819f02007-05-06 14:49:36 -07001221 mod_zone_page_state(page_zone(page),
1222 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1223 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001224 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001225
Christoph Lameter49bd5222008-04-14 18:52:18 +03001226 __ClearPageSlab(page);
1227 reset_page_mapcount(page);
Nick Piggin1eb5ac62009-05-05 19:13:44 +10001228 if (current->reclaim_state)
1229 current->reclaim_state->reclaimed_slab += pages;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001230 __free_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001231}
1232
1233static void rcu_free_slab(struct rcu_head *h)
1234{
1235 struct page *page;
1236
1237 page = container_of((struct list_head *)h, struct page, lru);
1238 __free_slab(page->slab, page);
1239}
1240
1241static void free_slab(struct kmem_cache *s, struct page *page)
1242{
1243 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1244 /*
1245 * RCU free overloads the RCU head over the LRU
1246 */
1247 struct rcu_head *head = (void *)&page->lru;
1248
1249 call_rcu(head, rcu_free_slab);
1250 } else
1251 __free_slab(s, page);
1252}
1253
1254static void discard_slab(struct kmem_cache *s, struct page *page)
1255{
Christoph Lameter205ab992008-04-14 19:11:40 +03001256 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001257 free_slab(s, page);
1258}
1259
1260/*
1261 * Per slab locking using the pagelock
1262 */
1263static __always_inline void slab_lock(struct page *page)
1264{
1265 bit_spin_lock(PG_locked, &page->flags);
1266}
1267
1268static __always_inline void slab_unlock(struct page *page)
1269{
Nick Piggina76d3542008-01-07 23:20:27 -08001270 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001271}
1272
1273static __always_inline int slab_trylock(struct page *page)
1274{
1275 int rc = 1;
1276
1277 rc = bit_spin_trylock(PG_locked, &page->flags);
1278 return rc;
1279}
1280
1281/*
1282 * Management of partially allocated slabs
1283 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001284static void add_partial(struct kmem_cache_node *n,
1285 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001286{
Christoph Lametere95eed52007-05-06 14:49:44 -07001287 spin_lock(&n->list_lock);
1288 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001289 if (tail)
1290 list_add_tail(&page->lru, &n->partial);
1291 else
1292 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001293 spin_unlock(&n->list_lock);
1294}
1295
Christoph Lameter0121c6192008-04-29 16:11:12 -07001296static void remove_partial(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001297{
1298 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1299
1300 spin_lock(&n->list_lock);
1301 list_del(&page->lru);
1302 n->nr_partial--;
1303 spin_unlock(&n->list_lock);
1304}
1305
1306/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001307 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001308 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001309 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001310 */
Christoph Lameter0121c6192008-04-29 16:11:12 -07001311static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1312 struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001313{
1314 if (slab_trylock(page)) {
1315 list_del(&page->lru);
1316 n->nr_partial--;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001317 __SetPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001318 return 1;
1319 }
1320 return 0;
1321}
1322
1323/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001324 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001325 */
1326static struct page *get_partial_node(struct kmem_cache_node *n)
1327{
1328 struct page *page;
1329
1330 /*
1331 * Racy check. If we mistakenly see no partial slabs then we
1332 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001333 * partial slab and there is none available then get_partials()
1334 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001335 */
1336 if (!n || !n->nr_partial)
1337 return NULL;
1338
1339 spin_lock(&n->list_lock);
1340 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001341 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001342 goto out;
1343 page = NULL;
1344out:
1345 spin_unlock(&n->list_lock);
1346 return page;
1347}
1348
1349/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001350 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001351 */
1352static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1353{
1354#ifdef CONFIG_NUMA
1355 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001356 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001357 struct zone *zone;
1358 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001359 struct page *page;
1360
1361 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001362 * The defrag ratio allows a configuration of the tradeoffs between
1363 * inter node defragmentation and node local allocations. A lower
1364 * defrag_ratio increases the tendency to do local allocations
1365 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001366 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001367 * If the defrag_ratio is set to 0 then kmalloc() always
1368 * returns node local objects. If the ratio is higher then kmalloc()
1369 * may return off node objects because partial slabs are obtained
1370 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001371 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001372 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001373 * defrag_ratio = 1000) then every (well almost) allocation will
1374 * first attempt to defrag slab caches on other nodes. This means
1375 * scanning over all nodes to look for partial slabs which may be
1376 * expensive if we do it every time we are trying to find a slab
1377 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001378 */
Christoph Lameter98246012008-01-07 23:20:26 -08001379 if (!s->remote_node_defrag_ratio ||
1380 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001381 return NULL;
1382
Mel Gorman0e884602008-04-28 02:12:14 -07001383 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001384 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001385 struct kmem_cache_node *n;
1386
Mel Gorman54a6eb52008-04-28 02:12:16 -07001387 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001388
Mel Gorman54a6eb52008-04-28 02:12:16 -07001389 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
David Rientjes3b89d7d2009-02-22 17:40:07 -08001390 n->nr_partial > s->min_partial) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001391 page = get_partial_node(n);
1392 if (page)
1393 return page;
1394 }
1395 }
1396#endif
1397 return NULL;
1398}
1399
1400/*
1401 * Get a partial page, lock it and return it.
1402 */
1403static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1404{
1405 struct page *page;
1406 int searchnode = (node == -1) ? numa_node_id() : node;
1407
1408 page = get_partial_node(get_node(s, searchnode));
1409 if (page || (flags & __GFP_THISNODE))
1410 return page;
1411
1412 return get_any_partial(s, flags);
1413}
1414
1415/*
1416 * Move a page back to the lists.
1417 *
1418 * Must be called with the slab lock held.
1419 *
1420 * On exit the slab lock will have been dropped.
1421 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001422static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001423{
Christoph Lametere95eed52007-05-06 14:49:44 -07001424 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001425 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001426
Andy Whitcroft8a380822008-07-23 21:27:18 -07001427 __ClearPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001428 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001429
Christoph Lametera973e9d2008-03-01 13:40:44 -08001430 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001431 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001432 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1433 } else {
1434 stat(c, DEACTIVATE_FULL);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001435 if (SLABDEBUG && PageSlubDebug(page) &&
1436 (s->flags & SLAB_STORE_USER))
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001437 add_full(n, page);
1438 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001439 slab_unlock(page);
1440 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001441 stat(c, DEACTIVATE_EMPTY);
David Rientjes3b89d7d2009-02-22 17:40:07 -08001442 if (n->nr_partial < s->min_partial) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001443 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001444 * Adding an empty slab to the partial slabs in order
1445 * to avoid page allocator overhead. This slab needs
1446 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001447 * so that the others get filled first. That way the
1448 * size of the partial list stays small.
1449 *
Christoph Lameter0121c6192008-04-29 16:11:12 -07001450 * kmem_cache_shrink can reclaim any empty slabs from
1451 * the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001452 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001453 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001454 slab_unlock(page);
1455 } else {
1456 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001457 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001458 discard_slab(s, page);
1459 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001460 }
1461}
1462
1463/*
1464 * Remove the cpu slab
1465 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001466static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001467{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001468 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001469 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001470
Christoph Lameterb773ad72008-03-04 11:10:17 -08001471 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001472 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001473 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001474 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001475 * because both freelists are empty. So this is unlikely
1476 * to occur.
1477 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001478 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001479 void **object;
1480
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001481 tail = 0; /* Hot objects. Put the slab first */
1482
Christoph Lameter894b8782007-05-10 03:15:16 -07001483 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001484 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001485 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001486
1487 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001488 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001489 page->freelist = object;
1490 page->inuse--;
1491 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001492 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001493 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001494}
1495
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001496static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001497{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001498 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001499 slab_lock(c->page);
1500 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001501}
1502
1503/*
1504 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001505 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001506 * Called from IPI handler with interrupts disabled.
1507 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001508static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001509{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001510 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001511
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001512 if (likely(c && c->page))
1513 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001514}
1515
1516static void flush_cpu_slab(void *d)
1517{
1518 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001519
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001520 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001521}
1522
1523static void flush_all(struct kmem_cache *s)
1524{
Jens Axboe15c8b6c2008-05-09 09:39:44 +02001525 on_each_cpu(flush_cpu_slab, s, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07001526}
1527
1528/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001529 * Check if the objects in a per cpu structure fit numa
1530 * locality expectations.
1531 */
1532static inline int node_match(struct kmem_cache_cpu *c, int node)
1533{
1534#ifdef CONFIG_NUMA
1535 if (node != -1 && c->node != node)
1536 return 0;
1537#endif
1538 return 1;
1539}
1540
Pekka Enberg781b2ba2009-06-10 18:50:32 +03001541static int count_free(struct page *page)
1542{
1543 return page->objects - page->inuse;
1544}
1545
1546static unsigned long count_partial(struct kmem_cache_node *n,
1547 int (*get_count)(struct page *))
1548{
1549 unsigned long flags;
1550 unsigned long x = 0;
1551 struct page *page;
1552
1553 spin_lock_irqsave(&n->list_lock, flags);
1554 list_for_each_entry(page, &n->partial, lru)
1555 x += get_count(page);
1556 spin_unlock_irqrestore(&n->list_lock, flags);
1557 return x;
1558}
1559
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001560static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1561{
1562#ifdef CONFIG_SLUB_DEBUG
1563 return atomic_long_read(&n->total_objects);
1564#else
1565 return 0;
1566#endif
1567}
1568
Pekka Enberg781b2ba2009-06-10 18:50:32 +03001569static noinline void
1570slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1571{
1572 int node;
1573
1574 printk(KERN_WARNING
1575 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1576 nid, gfpflags);
1577 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1578 "default order: %d, min order: %d\n", s->name, s->objsize,
1579 s->size, oo_order(s->oo), oo_order(s->min));
1580
David Rientjesfa5ec8a2009-07-07 00:14:14 -07001581 if (oo_order(s->min) > get_order(s->objsize))
1582 printk(KERN_WARNING " %s debugging increased min order, use "
1583 "slub_debug=O to disable.\n", s->name);
1584
Pekka Enberg781b2ba2009-06-10 18:50:32 +03001585 for_each_online_node(node) {
1586 struct kmem_cache_node *n = get_node(s, node);
1587 unsigned long nr_slabs;
1588 unsigned long nr_objs;
1589 unsigned long nr_free;
1590
1591 if (!n)
1592 continue;
1593
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001594 nr_free = count_partial(n, count_free);
1595 nr_slabs = node_nr_slabs(n);
1596 nr_objs = node_nr_objs(n);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03001597
1598 printk(KERN_WARNING
1599 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1600 node, nr_slabs, nr_objs, nr_free);
1601 }
1602}
1603
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001604/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001605 * Slow path. The lockless freelist is empty or we need to perform
1606 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001607 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001608 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001609 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001610 * Processing is still very fast if new objects have been freed to the
1611 * regular freelist. In that case we simply take over the regular freelist
1612 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001613 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001614 * If that is not working then we fall back to the partial lists. We take the
1615 * first element of the freelist as the object to allocate now and move the
1616 * rest of the freelist to the lockless freelist.
1617 *
1618 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001619 * we need to allocate a new slab. This is the slowest path since it involves
1620 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001621 */
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001622static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1623 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001624{
Christoph Lameter81819f02007-05-06 14:49:36 -07001625 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001626 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001627
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001628 /* We handle __GFP_ZERO in the caller */
1629 gfpflags &= ~__GFP_ZERO;
1630
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001631 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001632 goto new_slab;
1633
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001634 slab_lock(c->page);
1635 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001636 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001637
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001638 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001639
Christoph Lameter894b8782007-05-10 03:15:16 -07001640load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001641 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001642 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001643 goto another_slab;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001644 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001645 goto debug;
1646
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001647 c->freelist = object[c->offset];
Christoph Lameter39b26462008-04-14 19:11:30 +03001648 c->page->inuse = c->page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001649 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001650 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001651unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001652 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001653 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001654 return object;
1655
1656another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001657 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001658
1659new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001660 new = get_partial(s, gfpflags, node);
1661 if (new) {
1662 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001663 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001664 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001665 }
1666
Christoph Lameterb811c202007-10-16 23:25:51 -07001667 if (gfpflags & __GFP_WAIT)
1668 local_irq_enable();
1669
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001670 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001671
1672 if (gfpflags & __GFP_WAIT)
1673 local_irq_disable();
1674
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001675 if (new) {
1676 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001677 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001678 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001679 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001680 slab_lock(new);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001681 __SetPageSlubFrozen(new);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001682 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001683 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001684 }
Pekka Enberg95f85982009-06-11 16:18:09 +03001685 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1686 slab_out_of_memory(s, gfpflags, node);
Christoph Lameter71c7a062008-02-14 14:28:01 -08001687 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001688debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001689 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001690 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001691
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001692 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001693 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001694 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001695 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001696}
1697
1698/*
1699 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1700 * have the fastpath folded into their functions. So no function call
1701 * overhead for requests that can be satisfied on the fastpath.
1702 *
1703 * The fastpath works by first checking if the lockless freelist can be used.
1704 * If not then __slab_alloc is called for slow processing.
1705 *
1706 * Otherwise we can simply pick the next object from the lockless free list.
1707 */
Pekka Enberg06428782008-01-07 23:20:27 -08001708static __always_inline void *slab_alloc(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001709 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001710{
Christoph Lameter894b8782007-05-10 03:15:16 -07001711 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001712 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001713 unsigned long flags;
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001714 unsigned int objsize;
Christoph Lameter1f842602008-01-07 23:20:30 -08001715
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10001716 gfpflags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03001717
Nick Piggincf40bd12009-01-21 08:12:39 +01001718 lockdep_trace_alloc(gfpflags);
OGAWA Hirofumi89124d72008-11-19 21:23:59 +09001719 might_sleep_if(gfpflags & __GFP_WAIT);
Pekka Enberg3c506ef2008-12-29 11:47:05 +02001720
Akinobu Mita773ff602008-12-23 19:37:01 +09001721 if (should_failslab(s->objsize, gfpflags))
1722 return NULL;
1723
Christoph Lameter894b8782007-05-10 03:15:16 -07001724 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001725 c = get_cpu_slab(s, smp_processor_id());
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001726 objsize = c->objsize;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001727 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001728
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001729 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001730
1731 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001732 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001733 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001734 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001735 }
1736 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001737
1738 if (unlikely((gfpflags & __GFP_ZERO) && object))
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001739 memset(object, 0, objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001740
Vegard Nossum5a896d92008-04-04 00:54:48 +02001741 kmemcheck_slab_alloc(s, gfpflags, object, c->objsize);
Catalin Marinas06f22f12009-06-11 13:23:18 +01001742 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001743
Christoph Lameter894b8782007-05-10 03:15:16 -07001744 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001745}
1746
1747void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1748{
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03001749 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1750
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02001751 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03001752
1753 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07001754}
1755EXPORT_SYMBOL(kmem_cache_alloc);
1756
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03001757#ifdef CONFIG_KMEMTRACE
1758void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1759{
1760 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1761}
1762EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1763#endif
1764
Christoph Lameter81819f02007-05-06 14:49:36 -07001765#ifdef CONFIG_NUMA
1766void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1767{
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03001768 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1769
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02001770 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1771 s->objsize, s->size, gfpflags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03001772
1773 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07001774}
1775EXPORT_SYMBOL(kmem_cache_alloc_node);
1776#endif
1777
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03001778#ifdef CONFIG_KMEMTRACE
1779void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1780 gfp_t gfpflags,
1781 int node)
1782{
1783 return slab_alloc(s, gfpflags, node, _RET_IP_);
1784}
1785EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1786#endif
1787
Christoph Lameter81819f02007-05-06 14:49:36 -07001788/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001789 * Slow patch handling. This may still be called frequently since objects
1790 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001791 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001792 * So we still attempt to reduce cache line usage. Just take the slab
1793 * lock and free the item. If there is no additional partial page
1794 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001795 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001796static void __slab_free(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001797 void *x, unsigned long addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001798{
1799 void *prior;
1800 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001801 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001802
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001803 c = get_cpu_slab(s, raw_smp_processor_id());
1804 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001805 slab_lock(page);
1806
Andy Whitcroft8a380822008-07-23 21:27:18 -07001807 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001808 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001809
Christoph Lameter81819f02007-05-06 14:49:36 -07001810checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001811 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001812 page->freelist = object;
1813 page->inuse--;
1814
Andy Whitcroft8a380822008-07-23 21:27:18 -07001815 if (unlikely(PageSlubFrozen(page))) {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001816 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001817 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001818 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001819
1820 if (unlikely(!page->inuse))
1821 goto slab_empty;
1822
1823 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001824 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001825 * then add it.
1826 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001827 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001828 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001829 stat(c, FREE_ADD_PARTIAL);
1830 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001831
1832out_unlock:
1833 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001834 return;
1835
1836slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001837 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001838 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001839 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001840 */
1841 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001842 stat(c, FREE_REMOVE_PARTIAL);
1843 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001844 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001845 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001846 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001847 return;
1848
1849debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001850 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001851 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001852 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001853}
1854
Christoph Lameter894b8782007-05-10 03:15:16 -07001855/*
1856 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1857 * can perform fastpath freeing without additional function calls.
1858 *
1859 * The fastpath is only possible if we are freeing to the current cpu slab
1860 * of this processor. This typically the case if we have just allocated
1861 * the item before.
1862 *
1863 * If fastpath is not possible then fall back to __slab_free where we deal
1864 * with all sorts of special processing.
1865 */
Pekka Enberg06428782008-01-07 23:20:27 -08001866static __always_inline void slab_free(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001867 struct page *page, void *x, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001868{
1869 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001870 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001871 unsigned long flags;
1872
Catalin Marinas06f22f12009-06-11 13:23:18 +01001873 kmemleak_free_recursive(x, s->flags);
Christoph Lameter894b8782007-05-10 03:15:16 -07001874 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001875 c = get_cpu_slab(s, smp_processor_id());
Vegard Nossum5a896d92008-04-04 00:54:48 +02001876 kmemcheck_slab_free(s, object, c->objsize);
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001877 debug_check_no_locks_freed(object, c->objsize);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001878 if (!(s->flags & SLAB_DEBUG_OBJECTS))
Pekka Enberg6047a002009-01-14 12:22:25 +02001879 debug_check_no_obj_freed(object, c->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001880 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001881 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001882 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001883 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001884 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001885 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001886
1887 local_irq_restore(flags);
1888}
1889
Christoph Lameter81819f02007-05-06 14:49:36 -07001890void kmem_cache_free(struct kmem_cache *s, void *x)
1891{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001892 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001893
Christoph Lameterb49af682007-05-06 14:49:41 -07001894 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001895
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001896 slab_free(s, page, x, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03001897
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02001898 trace_kmem_cache_free(_RET_IP_, x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001899}
1900EXPORT_SYMBOL(kmem_cache_free);
1901
Cyrill Gorcunove9beef12008-10-28 22:02:26 +03001902/* Figure out on which slab page the object resides */
Christoph Lameter81819f02007-05-06 14:49:36 -07001903static struct page *get_object_page(const void *x)
1904{
Christoph Lameterb49af682007-05-06 14:49:41 -07001905 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001906
1907 if (!PageSlab(page))
1908 return NULL;
1909
1910 return page;
1911}
1912
1913/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001914 * Object placement in a slab is made very easy because we always start at
1915 * offset 0. If we tune the size of the object to the alignment then we can
1916 * get the required alignment by putting one properly sized object after
1917 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001918 *
1919 * Notice that the allocation order determines the sizes of the per cpu
1920 * caches. Each processor has always one slab available for allocations.
1921 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001922 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001923 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001924 */
1925
1926/*
1927 * Mininum / Maximum order of slab pages. This influences locking overhead
1928 * and slab fragmentation. A higher order reduces the number of partial slabs
1929 * and increases the number of allocations possible without having to
1930 * take the list_lock.
1931 */
1932static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03001933static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001934static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001935
1936/*
1937 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001938 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001939 */
1940static int slub_nomerge;
1941
1942/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001943 * Calculate the order of allocation given an slab object size.
1944 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001945 * The order of allocation has significant impact on performance and other
1946 * system components. Generally order 0 allocations should be preferred since
1947 * order 0 does not cause fragmentation in the page allocator. Larger objects
1948 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001949 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07001950 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001951 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001952 * In order to reach satisfactory performance we must ensure that a minimum
1953 * number of objects is in one slab. Otherwise we may generate too much
1954 * activity on the partial lists which requires taking the list_lock. This is
1955 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001956 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001957 * slub_max_order specifies the order where we begin to stop considering the
1958 * number of objects in a slab as critical. If we reach slub_max_order then
1959 * we try to keep the page order as low as possible. So we accept more waste
1960 * of space in favor of a small page order.
1961 *
1962 * Higher order allocations also allow the placement of more objects in a
1963 * slab and thereby reduce object handling overhead. If the user has
1964 * requested a higher mininum order then we start with that one instead of
1965 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001966 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001967static inline int slab_order(int size, int min_objects,
1968 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001969{
1970 int order;
1971 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001972 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001973
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04001974 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1975 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03001976
Christoph Lameter6300ea72007-07-17 04:03:20 -07001977 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001978 fls(min_objects * size - 1) - PAGE_SHIFT);
1979 order <= max_order; order++) {
1980
Christoph Lameter81819f02007-05-06 14:49:36 -07001981 unsigned long slab_size = PAGE_SIZE << order;
1982
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001983 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001984 continue;
1985
Christoph Lameter81819f02007-05-06 14:49:36 -07001986 rem = slab_size % size;
1987
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001988 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001989 break;
1990
1991 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001992
Christoph Lameter81819f02007-05-06 14:49:36 -07001993 return order;
1994}
1995
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001996static inline int calculate_order(int size)
1997{
1998 int order;
1999 int min_objects;
2000 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002001 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002002
2003 /*
2004 * Attempt to find best configuration for a slab. This
2005 * works by first attempting to generate a layout with
2006 * the best configuration and backing off gradually.
2007 *
2008 * First we reduce the acceptable waste in a slab. Then
2009 * we reduce the minimum objects required in a slab.
2010 */
2011 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002012 if (!min_objects)
2013 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002014 max_objects = (PAGE_SIZE << slub_max_order)/size;
2015 min_objects = min(min_objects, max_objects);
2016
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002017 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002018 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002019 while (fraction >= 4) {
2020 order = slab_order(size, min_objects,
2021 slub_max_order, fraction);
2022 if (order <= slub_max_order)
2023 return order;
2024 fraction /= 2;
2025 }
Amerigo Wang5086c382009-08-19 21:44:13 +03002026 min_objects--;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002027 }
2028
2029 /*
2030 * We were unable to place multiple objects in a slab. Now
2031 * lets see if we can place a single object there.
2032 */
2033 order = slab_order(size, 1, slub_max_order, 1);
2034 if (order <= slub_max_order)
2035 return order;
2036
2037 /*
2038 * Doh this slab cannot be placed using slub_max_order.
2039 */
2040 order = slab_order(size, 1, MAX_ORDER, 1);
David Rientjes818cf592009-04-23 09:58:22 +03002041 if (order < MAX_ORDER)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002042 return order;
2043 return -ENOSYS;
2044}
2045
Christoph Lameter81819f02007-05-06 14:49:36 -07002046/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002047 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07002048 */
2049static unsigned long calculate_alignment(unsigned long flags,
2050 unsigned long align, unsigned long size)
2051{
2052 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08002053 * If the user wants hardware cache aligned objects then follow that
2054 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07002055 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08002056 * The hardware cache alignment cannot override the specified
2057 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07002058 */
Nick Pigginb6210382008-03-05 14:05:56 -08002059 if (flags & SLAB_HWCACHE_ALIGN) {
2060 unsigned long ralign = cache_line_size();
2061 while (size <= ralign / 2)
2062 ralign /= 2;
2063 align = max(align, ralign);
2064 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002065
2066 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08002067 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07002068
2069 return ALIGN(align, sizeof(void *));
2070}
2071
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002072static void init_kmem_cache_cpu(struct kmem_cache *s,
2073 struct kmem_cache_cpu *c)
2074{
2075 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08002076 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002077 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07002078 c->offset = s->offset / sizeof(void *);
2079 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03002080#ifdef CONFIG_SLUB_STATS
2081 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
2082#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002083}
2084
Pekka Enberg5595cff2008-08-05 09:28:47 +03002085static void
2086init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002087{
2088 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002089 spin_lock_init(&n->list_lock);
2090 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002091#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002092 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07002093 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07002094 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002095#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002096}
2097
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002098#ifdef CONFIG_SMP
2099/*
2100 * Per cpu array for per cpu structures.
2101 *
2102 * The per cpu array places all kmem_cache_cpu structures from one processor
2103 * close together meaning that it becomes possible that multiple per cpu
2104 * structures are contained in one cacheline. This may be particularly
2105 * beneficial for the kmalloc caches.
2106 *
2107 * A desktop system typically has around 60-80 slabs. With 100 here we are
2108 * likely able to get per cpu structures for all caches from the array defined
2109 * here. We must be able to cover all kmalloc caches during bootstrap.
2110 *
2111 * If the per cpu array is exhausted then fall back to kmalloc
2112 * of individual cachelines. No sharing is possible then.
2113 */
2114#define NR_KMEM_CACHE_CPU 100
2115
Tejun Heo204fba42009-06-24 15:13:45 +09002116static DEFINE_PER_CPU(struct kmem_cache_cpu [NR_KMEM_CACHE_CPU],
2117 kmem_cache_cpu);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002118
2119static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
Rusty Russell174596a2009-01-01 10:12:29 +10302120static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002121
2122static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2123 int cpu, gfp_t flags)
2124{
2125 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2126
2127 if (c)
2128 per_cpu(kmem_cache_cpu_free, cpu) =
2129 (void *)c->freelist;
2130 else {
2131 /* Table overflow: So allocate ourselves */
2132 c = kmalloc_node(
2133 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2134 flags, cpu_to_node(cpu));
2135 if (!c)
2136 return NULL;
2137 }
2138
2139 init_kmem_cache_cpu(s, c);
2140 return c;
2141}
2142
2143static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2144{
2145 if (c < per_cpu(kmem_cache_cpu, cpu) ||
David Rientjes37189092009-01-27 18:59:46 -08002146 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002147 kfree(c);
2148 return;
2149 }
2150 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2151 per_cpu(kmem_cache_cpu_free, cpu) = c;
2152}
2153
2154static void free_kmem_cache_cpus(struct kmem_cache *s)
2155{
2156 int cpu;
2157
2158 for_each_online_cpu(cpu) {
2159 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2160
2161 if (c) {
2162 s->cpu_slab[cpu] = NULL;
2163 free_kmem_cache_cpu(c, cpu);
2164 }
2165 }
2166}
2167
2168static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2169{
2170 int cpu;
2171
2172 for_each_online_cpu(cpu) {
2173 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2174
2175 if (c)
2176 continue;
2177
2178 c = alloc_kmem_cache_cpu(s, cpu, flags);
2179 if (!c) {
2180 free_kmem_cache_cpus(s);
2181 return 0;
2182 }
2183 s->cpu_slab[cpu] = c;
2184 }
2185 return 1;
2186}
2187
2188/*
2189 * Initialize the per cpu array.
2190 */
2191static void init_alloc_cpu_cpu(int cpu)
2192{
2193 int i;
2194
Rusty Russell174596a2009-01-01 10:12:29 +10302195 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002196 return;
2197
2198 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2199 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2200
Rusty Russell174596a2009-01-01 10:12:29 +10302201 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002202}
2203
2204static void __init init_alloc_cpu(void)
2205{
2206 int cpu;
2207
2208 for_each_online_cpu(cpu)
2209 init_alloc_cpu_cpu(cpu);
2210 }
2211
2212#else
2213static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2214static inline void init_alloc_cpu(void) {}
2215
2216static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2217{
2218 init_kmem_cache_cpu(s, &s->cpu_slab);
2219 return 1;
2220}
2221#endif
2222
Christoph Lameter81819f02007-05-06 14:49:36 -07002223#ifdef CONFIG_NUMA
2224/*
2225 * No kmalloc_node yet so do it by hand. We know that this is the first
2226 * slab on the node for this slabcache. There are no concurrent accesses
2227 * possible.
2228 *
2229 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002230 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2231 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002232 */
David Rientjes0094de92008-11-25 19:14:19 -08002233static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002234{
2235 struct page *page;
2236 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002237 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002238
2239 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2240
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002241 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002242
2243 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002244 if (page_to_nid(page) != node) {
2245 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2246 "node %d\n", node);
2247 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2248 "in order to be able to continue\n");
2249 }
2250
Christoph Lameter81819f02007-05-06 14:49:36 -07002251 n = page->freelist;
2252 BUG_ON(!n);
2253 page->freelist = get_freepointer(kmalloc_caches, n);
2254 page->inuse++;
2255 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002256#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002257 init_object(kmalloc_caches, n, 1);
2258 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002259#endif
Pekka Enberg5595cff2008-08-05 09:28:47 +03002260 init_kmem_cache_node(n, kmalloc_caches);
Christoph Lameter205ab992008-04-14 19:11:40 +03002261 inc_slabs_node(kmalloc_caches, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002262
rootba84c732008-01-07 23:20:28 -08002263 /*
2264 * lockdep requires consistent irq usage for each lock
2265 * so even though there cannot be a race this early in
2266 * the boot sequence, we still disable irqs.
2267 */
2268 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002269 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002270 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002271}
2272
2273static void free_kmem_cache_nodes(struct kmem_cache *s)
2274{
2275 int node;
2276
Christoph Lameterf64dc582007-10-16 01:25:33 -07002277 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002278 struct kmem_cache_node *n = s->node[node];
2279 if (n && n != &s->local_node)
2280 kmem_cache_free(kmalloc_caches, n);
2281 s->node[node] = NULL;
2282 }
2283}
2284
2285static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2286{
2287 int node;
2288 int local_node;
2289
2290 if (slab_state >= UP)
2291 local_node = page_to_nid(virt_to_page(s));
2292 else
2293 local_node = 0;
2294
Christoph Lameterf64dc582007-10-16 01:25:33 -07002295 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002296 struct kmem_cache_node *n;
2297
2298 if (local_node == node)
2299 n = &s->local_node;
2300 else {
2301 if (slab_state == DOWN) {
David Rientjes0094de92008-11-25 19:14:19 -08002302 early_kmem_cache_node_alloc(gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002303 continue;
2304 }
2305 n = kmem_cache_alloc_node(kmalloc_caches,
2306 gfpflags, node);
2307
2308 if (!n) {
2309 free_kmem_cache_nodes(s);
2310 return 0;
2311 }
2312
2313 }
2314 s->node[node] = n;
Pekka Enberg5595cff2008-08-05 09:28:47 +03002315 init_kmem_cache_node(n, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002316 }
2317 return 1;
2318}
2319#else
2320static void free_kmem_cache_nodes(struct kmem_cache *s)
2321{
2322}
2323
2324static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2325{
Pekka Enberg5595cff2008-08-05 09:28:47 +03002326 init_kmem_cache_node(&s->local_node, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002327 return 1;
2328}
2329#endif
2330
David Rientjesc0bdb232009-02-25 09:16:35 +02002331static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08002332{
2333 if (min < MIN_PARTIAL)
2334 min = MIN_PARTIAL;
2335 else if (min > MAX_PARTIAL)
2336 min = MAX_PARTIAL;
2337 s->min_partial = min;
2338}
2339
Christoph Lameter81819f02007-05-06 14:49:36 -07002340/*
2341 * calculate_sizes() determines the order and the distribution of data within
2342 * a slab object.
2343 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002344static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002345{
2346 unsigned long flags = s->flags;
2347 unsigned long size = s->objsize;
2348 unsigned long align = s->align;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002349 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002350
2351 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002352 * Round up object size to the next word boundary. We can only
2353 * place the free pointer at word boundaries and this determines
2354 * the possible location of the free pointer.
2355 */
2356 size = ALIGN(size, sizeof(void *));
2357
2358#ifdef CONFIG_SLUB_DEBUG
2359 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002360 * Determine if we can poison the object itself. If the user of
2361 * the slab may touch the object after free or before allocation
2362 * then we should never poison the object itself.
2363 */
2364 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002365 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002366 s->flags |= __OBJECT_POISON;
2367 else
2368 s->flags &= ~__OBJECT_POISON;
2369
Christoph Lameter81819f02007-05-06 14:49:36 -07002370
2371 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002372 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002373 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002374 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002375 */
2376 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2377 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002378#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002379
2380 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002381 * With that we have determined the number of bytes in actual use
2382 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002383 */
2384 s->inuse = size;
2385
2386 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002387 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002388 /*
2389 * Relocate free pointer after the object if it is not
2390 * permitted to overwrite the first word of the object on
2391 * kmem_cache_free.
2392 *
2393 * This is the case if we do RCU, have a constructor or
2394 * destructor or are poisoning the objects.
2395 */
2396 s->offset = size;
2397 size += sizeof(void *);
2398 }
2399
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002400#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002401 if (flags & SLAB_STORE_USER)
2402 /*
2403 * Need to store information about allocs and frees after
2404 * the object.
2405 */
2406 size += 2 * sizeof(struct track);
2407
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002408 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002409 /*
2410 * Add some empty padding so that we can catch
2411 * overwrites from earlier objects rather than let
2412 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01002413 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07002414 * of the object.
2415 */
2416 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002417#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002418
Christoph Lameter81819f02007-05-06 14:49:36 -07002419 /*
2420 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002421 * user specified and the dynamic determination of cache line size
2422 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002423 */
2424 align = calculate_alignment(flags, align, s->objsize);
Zhang, Yanmindcb0ce12009-07-30 11:28:11 +08002425 s->align = align;
Christoph Lameter81819f02007-05-06 14:49:36 -07002426
2427 /*
2428 * SLUB stores one object immediately after another beginning from
2429 * offset 0. In order to align the objects we have to simply size
2430 * each object to conform to the alignment.
2431 */
2432 size = ALIGN(size, align);
2433 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03002434 if (forced_order >= 0)
2435 order = forced_order;
2436 else
2437 order = calculate_order(size);
Christoph Lameter81819f02007-05-06 14:49:36 -07002438
Christoph Lameter834f3d12008-04-14 19:11:31 +03002439 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07002440 return 0;
2441
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002442 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002443 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002444 s->allocflags |= __GFP_COMP;
2445
2446 if (s->flags & SLAB_CACHE_DMA)
2447 s->allocflags |= SLUB_DMA;
2448
2449 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2450 s->allocflags |= __GFP_RECLAIMABLE;
2451
Christoph Lameter81819f02007-05-06 14:49:36 -07002452 /*
2453 * Determine the number of objects per slab
2454 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002455 s->oo = oo_make(order, size);
Christoph Lameter65c33762008-04-14 19:11:40 +03002456 s->min = oo_make(get_order(size), size);
Christoph Lameter205ab992008-04-14 19:11:40 +03002457 if (oo_objects(s->oo) > oo_objects(s->max))
2458 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002459
Christoph Lameter834f3d12008-04-14 19:11:31 +03002460 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07002461
2462}
2463
Christoph Lameter81819f02007-05-06 14:49:36 -07002464static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2465 const char *name, size_t size,
2466 size_t align, unsigned long flags,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002467 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002468{
2469 memset(s, 0, kmem_size);
2470 s->name = name;
2471 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002472 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002473 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002474 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002475
Christoph Lameter06b285d2008-04-14 19:11:41 +03002476 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07002477 goto error;
David Rientjes3de47212009-07-27 18:30:35 -07002478 if (disable_higher_order_debug) {
2479 /*
2480 * Disable debugging flags that store metadata if the min slab
2481 * order increased.
2482 */
2483 if (get_order(s->size) > get_order(s->objsize)) {
2484 s->flags &= ~DEBUG_METADATA_FLAGS;
2485 s->offset = 0;
2486 if (!calculate_sizes(s, -1))
2487 goto error;
2488 }
2489 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002490
David Rientjes3b89d7d2009-02-22 17:40:07 -08002491 /*
2492 * The larger the object size is, the more pages we want on the partial
2493 * list to avoid pounding the page allocator excessively.
2494 */
David Rientjesc0bdb232009-02-25 09:16:35 +02002495 set_min_partial(s, ilog2(s->size));
Christoph Lameter81819f02007-05-06 14:49:36 -07002496 s->refcount = 1;
2497#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05002498 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07002499#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002500 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2501 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002502
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002503 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002504 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002505 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002506error:
2507 if (flags & SLAB_PANIC)
2508 panic("Cannot create slab %s size=%lu realsize=%u "
2509 "order=%u offset=%u flags=%lx\n",
Christoph Lameter834f3d12008-04-14 19:11:31 +03002510 s->name, (unsigned long)size, s->size, oo_order(s->oo),
Christoph Lameter81819f02007-05-06 14:49:36 -07002511 s->offset, flags);
2512 return 0;
2513}
Christoph Lameter81819f02007-05-06 14:49:36 -07002514
2515/*
2516 * Check if a given pointer is valid
2517 */
2518int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2519{
Pekka Enberg06428782008-01-07 23:20:27 -08002520 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002521
2522 page = get_object_page(object);
2523
2524 if (!page || s != page->slab)
2525 /* No slab or wrong slab */
2526 return 0;
2527
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002528 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002529 return 0;
2530
2531 /*
2532 * We could also check if the object is on the slabs freelist.
2533 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002534 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002535 * to a certain slab.
2536 */
2537 return 1;
2538}
2539EXPORT_SYMBOL(kmem_ptr_validate);
2540
2541/*
2542 * Determine the size of a slab object
2543 */
2544unsigned int kmem_cache_size(struct kmem_cache *s)
2545{
2546 return s->objsize;
2547}
2548EXPORT_SYMBOL(kmem_cache_size);
2549
2550const char *kmem_cache_name(struct kmem_cache *s)
2551{
2552 return s->name;
2553}
2554EXPORT_SYMBOL(kmem_cache_name);
2555
Christoph Lameter33b12c32008-04-25 12:22:43 -07002556static void list_slab_objects(struct kmem_cache *s, struct page *page,
2557 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07002558{
Christoph Lameter33b12c32008-04-25 12:22:43 -07002559#ifdef CONFIG_SLUB_DEBUG
2560 void *addr = page_address(page);
2561 void *p;
2562 DECLARE_BITMAP(map, page->objects);
2563
2564 bitmap_zero(map, page->objects);
2565 slab_err(s, page, "%s", text);
2566 slab_lock(page);
2567 for_each_free_object(p, s, page->freelist)
2568 set_bit(slab_index(p, s, addr), map);
2569
2570 for_each_object(p, s, addr, page->objects) {
2571
2572 if (!test_bit(slab_index(p, s, addr), map)) {
2573 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2574 p, p - addr);
2575 print_tracking(s, p);
2576 }
2577 }
2578 slab_unlock(page);
2579#endif
2580}
2581
Christoph Lameter81819f02007-05-06 14:49:36 -07002582/*
Christoph Lameter599870b2008-04-23 12:36:52 -07002583 * Attempt to free all partial slabs on a node.
Christoph Lameter81819f02007-05-06 14:49:36 -07002584 */
Christoph Lameter599870b2008-04-23 12:36:52 -07002585static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002586{
Christoph Lameter81819f02007-05-06 14:49:36 -07002587 unsigned long flags;
2588 struct page *page, *h;
2589
2590 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter33b12c32008-04-25 12:22:43 -07002591 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002592 if (!page->inuse) {
2593 list_del(&page->lru);
2594 discard_slab(s, page);
Christoph Lameter599870b2008-04-23 12:36:52 -07002595 n->nr_partial--;
Christoph Lameter33b12c32008-04-25 12:22:43 -07002596 } else {
2597 list_slab_objects(s, page,
2598 "Objects remaining on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07002599 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07002600 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002601 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002602}
2603
2604/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002605 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002606 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002607static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002608{
2609 int node;
2610
2611 flush_all(s);
2612
2613 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002614 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002615 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002616 struct kmem_cache_node *n = get_node(s, node);
2617
Christoph Lameter599870b2008-04-23 12:36:52 -07002618 free_partial(s, n);
2619 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002620 return 1;
2621 }
2622 free_kmem_cache_nodes(s);
2623 return 0;
2624}
2625
2626/*
2627 * Close a cache and release the kmem_cache structure
2628 * (must be used for caches created using kmem_cache_create)
2629 */
2630void kmem_cache_destroy(struct kmem_cache *s)
2631{
2632 down_write(&slub_lock);
2633 s->refcount--;
2634 if (!s->refcount) {
2635 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002636 up_write(&slub_lock);
Pekka Enbergd629d812008-04-23 22:31:08 +03002637 if (kmem_cache_close(s)) {
2638 printk(KERN_ERR "SLUB %s: %s called for cache that "
2639 "still has objects.\n", s->name, __func__);
2640 dump_stack();
2641 }
Eric Dumazetd76b1592009-09-03 22:38:59 +03002642 if (s->flags & SLAB_DESTROY_BY_RCU)
2643 rcu_barrier();
Christoph Lameter81819f02007-05-06 14:49:36 -07002644 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002645 } else
2646 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002647}
2648EXPORT_SYMBOL(kmem_cache_destroy);
2649
2650/********************************************************************
2651 * Kmalloc subsystem
2652 *******************************************************************/
2653
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002654struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002655EXPORT_SYMBOL(kmalloc_caches);
2656
Christoph Lameter81819f02007-05-06 14:49:36 -07002657static int __init setup_slub_min_order(char *str)
2658{
Pekka Enberg06428782008-01-07 23:20:27 -08002659 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002660
2661 return 1;
2662}
2663
2664__setup("slub_min_order=", setup_slub_min_order);
2665
2666static int __init setup_slub_max_order(char *str)
2667{
Pekka Enberg06428782008-01-07 23:20:27 -08002668 get_option(&str, &slub_max_order);
David Rientjes818cf592009-04-23 09:58:22 +03002669 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002670
2671 return 1;
2672}
2673
2674__setup("slub_max_order=", setup_slub_max_order);
2675
2676static int __init setup_slub_min_objects(char *str)
2677{
Pekka Enberg06428782008-01-07 23:20:27 -08002678 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002679
2680 return 1;
2681}
2682
2683__setup("slub_min_objects=", setup_slub_min_objects);
2684
2685static int __init setup_slub_nomerge(char *str)
2686{
2687 slub_nomerge = 1;
2688 return 1;
2689}
2690
2691__setup("slub_nomerge", setup_slub_nomerge);
2692
Christoph Lameter81819f02007-05-06 14:49:36 -07002693static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2694 const char *name, int size, gfp_t gfp_flags)
2695{
2696 unsigned int flags = 0;
2697
2698 if (gfp_flags & SLUB_DMA)
2699 flags = SLAB_CACHE_DMA;
2700
Pekka Enberg83b519e2009-06-10 19:40:04 +03002701 /*
2702 * This function is called with IRQs disabled during early-boot on
2703 * single CPU so there's no need to take slub_lock here.
2704 */
Christoph Lameter81819f02007-05-06 14:49:36 -07002705 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter319d1e22008-04-14 19:11:41 +03002706 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002707 goto panic;
2708
2709 list_add(&s->list, &slab_caches);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002710
Christoph Lameter81819f02007-05-06 14:49:36 -07002711 if (sysfs_slab_add(s))
2712 goto panic;
2713 return s;
2714
2715panic:
2716 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2717}
2718
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002719#ifdef CONFIG_ZONE_DMA
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002720static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002721
2722static void sysfs_add_func(struct work_struct *w)
2723{
2724 struct kmem_cache *s;
2725
2726 down_write(&slub_lock);
2727 list_for_each_entry(s, &slab_caches, list) {
2728 if (s->flags & __SYSFS_ADD_DEFERRED) {
2729 s->flags &= ~__SYSFS_ADD_DEFERRED;
2730 sysfs_slab_add(s);
2731 }
2732 }
2733 up_write(&slub_lock);
2734}
2735
2736static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2737
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002738static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2739{
2740 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002741 char *text;
2742 size_t realsize;
Nick Piggin964cf352009-06-15 13:35:10 +03002743 unsigned long slabflags;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002744
2745 s = kmalloc_caches_dma[index];
2746 if (s)
2747 return s;
2748
2749 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002750 if (flags & __GFP_WAIT)
2751 down_write(&slub_lock);
2752 else {
2753 if (!down_write_trylock(&slub_lock))
2754 goto out;
2755 }
2756
2757 if (kmalloc_caches_dma[index])
2758 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002759
Christoph Lameter7b55f622007-07-17 04:03:27 -07002760 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002761 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2762 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002763 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2764
Nick Piggin964cf352009-06-15 13:35:10 +03002765 /*
2766 * Must defer sysfs creation to a workqueue because we don't know
2767 * what context we are called from. Before sysfs comes up, we don't
2768 * need to do anything because our sysfs initcall will start by
2769 * adding all existing slabs to sysfs.
2770 */
Pekka Enberg5caf5c72009-06-17 08:30:54 +03002771 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK;
Nick Piggin964cf352009-06-15 13:35:10 +03002772 if (slab_state >= SYSFS)
2773 slabflags |= __SYSFS_ADD_DEFERRED;
2774
Christoph Lameter1ceef402007-08-07 15:11:48 -07002775 if (!s || !text || !kmem_cache_open(s, flags, text,
Nick Piggin964cf352009-06-15 13:35:10 +03002776 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) {
Christoph Lameter1ceef402007-08-07 15:11:48 -07002777 kfree(s);
2778 kfree(text);
2779 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002780 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002781
2782 list_add(&s->list, &slab_caches);
2783 kmalloc_caches_dma[index] = s;
2784
Nick Piggin964cf352009-06-15 13:35:10 +03002785 if (slab_state >= SYSFS)
2786 schedule_work(&sysfs_add_work);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002787
2788unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002789 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002790out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002791 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002792}
2793#endif
2794
Christoph Lameterf1b26332007-07-17 04:03:26 -07002795/*
2796 * Conversion table for small slabs sizes / 8 to the index in the
2797 * kmalloc array. This is necessary for slabs < 192 since we have non power
2798 * of two cache sizes there. The size of larger slabs can be determined using
2799 * fls.
2800 */
2801static s8 size_index[24] = {
2802 3, /* 8 */
2803 4, /* 16 */
2804 5, /* 24 */
2805 5, /* 32 */
2806 6, /* 40 */
2807 6, /* 48 */
2808 6, /* 56 */
2809 6, /* 64 */
2810 1, /* 72 */
2811 1, /* 80 */
2812 1, /* 88 */
2813 1, /* 96 */
2814 7, /* 104 */
2815 7, /* 112 */
2816 7, /* 120 */
2817 7, /* 128 */
2818 2, /* 136 */
2819 2, /* 144 */
2820 2, /* 152 */
2821 2, /* 160 */
2822 2, /* 168 */
2823 2, /* 176 */
2824 2, /* 184 */
2825 2 /* 192 */
2826};
2827
Aaro Koskinenacdfcd02009-08-28 14:28:54 +03002828static inline int size_index_elem(size_t bytes)
2829{
2830 return (bytes - 1) / 8;
2831}
2832
Christoph Lameter81819f02007-05-06 14:49:36 -07002833static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2834{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002835 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002836
Christoph Lameterf1b26332007-07-17 04:03:26 -07002837 if (size <= 192) {
2838 if (!size)
2839 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002840
Aaro Koskinenacdfcd02009-08-28 14:28:54 +03002841 index = size_index[size_index_elem(size)];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002842 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002843 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002844
2845#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002846 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002847 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002848
Christoph Lameter81819f02007-05-06 14:49:36 -07002849#endif
2850 return &kmalloc_caches[index];
2851}
2852
2853void *__kmalloc(size_t size, gfp_t flags)
2854{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002855 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002856 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002857
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002858 if (unlikely(size > SLUB_MAX_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002859 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002860
2861 s = get_slab(size, flags);
2862
2863 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002864 return s;
2865
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002866 ret = slab_alloc(s, flags, -1, _RET_IP_);
2867
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02002868 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002869
2870 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002871}
2872EXPORT_SYMBOL(__kmalloc);
2873
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002874static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2875{
Vegard Nossumb1eeab62008-11-25 16:55:53 +01002876 struct page *page;
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01002877 void *ptr = NULL;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002878
Vegard Nossumb1eeab62008-11-25 16:55:53 +01002879 flags |= __GFP_COMP | __GFP_NOTRACK;
2880 page = alloc_pages_node(node, flags, get_order(size));
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002881 if (page)
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01002882 ptr = page_address(page);
2883
2884 kmemleak_alloc(ptr, size, 1, flags);
2885 return ptr;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002886}
2887
Christoph Lameter81819f02007-05-06 14:49:36 -07002888#ifdef CONFIG_NUMA
2889void *__kmalloc_node(size_t size, gfp_t flags, int node)
2890{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002891 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002892 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002893
Ingo Molnar057685c2009-02-20 12:15:30 +01002894 if (unlikely(size > SLUB_MAX_SIZE)) {
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002895 ret = kmalloc_large_node(size, flags, node);
2896
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02002897 trace_kmalloc_node(_RET_IP_, ret,
2898 size, PAGE_SIZE << get_order(size),
2899 flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002900
2901 return ret;
2902 }
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002903
2904 s = get_slab(size, flags);
2905
2906 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002907 return s;
2908
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002909 ret = slab_alloc(s, flags, node, _RET_IP_);
2910
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02002911 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002912
2913 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002914}
2915EXPORT_SYMBOL(__kmalloc_node);
2916#endif
2917
2918size_t ksize(const void *object)
2919{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002920 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002921 struct kmem_cache *s;
2922
Christoph Lameteref8b4522007-10-16 01:24:46 -07002923 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002924 return 0;
2925
Vegard Nossum294a80a2007-12-04 23:45:30 -08002926 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002927
Pekka Enberg76994412008-05-22 19:22:25 +03002928 if (unlikely(!PageSlab(page))) {
2929 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08002930 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03002931 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002932 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002933
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002934#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002935 /*
2936 * Debugging requires use of the padding between object
2937 * and whatever may come after it.
2938 */
2939 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2940 return s->objsize;
2941
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002942#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002943 /*
2944 * If we have the need to store the freelist pointer
2945 * back there or track user information then we can
2946 * only use the space before that information.
2947 */
2948 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2949 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002950 /*
2951 * Else we can use all the padding etc for the allocation
2952 */
2953 return s->size;
2954}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02002955EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07002956
2957void kfree(const void *x)
2958{
Christoph Lameter81819f02007-05-06 14:49:36 -07002959 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002960 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002961
Pekka Enberg2121db72009-03-25 11:05:57 +02002962 trace_kfree(_RET_IP_, x);
2963
Satyam Sharma2408c552007-10-16 01:24:44 -07002964 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002965 return;
2966
Christoph Lameterb49af682007-05-06 14:49:41 -07002967 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002968 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07002969 BUG_ON(!PageCompound(page));
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01002970 kmemleak_free(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002971 put_page(page);
2972 return;
2973 }
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002974 slab_free(page->slab, page, object, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07002975}
2976EXPORT_SYMBOL(kfree);
2977
Christoph Lameter2086d262007-05-06 14:49:46 -07002978/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002979 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2980 * the remaining slabs by the number of items in use. The slabs with the
2981 * most items in use come first. New allocations will then fill those up
2982 * and thus they can be removed from the partial lists.
2983 *
2984 * The slabs with the least items are placed last. This results in them
2985 * being allocated from last increasing the chance that the last objects
2986 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002987 */
2988int kmem_cache_shrink(struct kmem_cache *s)
2989{
2990 int node;
2991 int i;
2992 struct kmem_cache_node *n;
2993 struct page *page;
2994 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03002995 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07002996 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03002997 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07002998 unsigned long flags;
2999
3000 if (!slabs_by_inuse)
3001 return -ENOMEM;
3002
3003 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003004 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07003005 n = get_node(s, node);
3006
3007 if (!n->nr_partial)
3008 continue;
3009
Christoph Lameter834f3d12008-04-14 19:11:31 +03003010 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07003011 INIT_LIST_HEAD(slabs_by_inuse + i);
3012
3013 spin_lock_irqsave(&n->list_lock, flags);
3014
3015 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003016 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07003017 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003018 * Note that concurrent frees may occur while we hold the
3019 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07003020 */
3021 list_for_each_entry_safe(page, t, &n->partial, lru) {
3022 if (!page->inuse && slab_trylock(page)) {
3023 /*
3024 * Must hold slab lock here because slab_free
3025 * may have freed the last object and be
3026 * waiting to release the slab.
3027 */
3028 list_del(&page->lru);
3029 n->nr_partial--;
3030 slab_unlock(page);
3031 discard_slab(s, page);
3032 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07003033 list_move(&page->lru,
3034 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07003035 }
3036 }
3037
Christoph Lameter2086d262007-05-06 14:49:46 -07003038 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003039 * Rebuild the partial list with the slabs filled up most
3040 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07003041 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03003042 for (i = objects - 1; i >= 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07003043 list_splice(slabs_by_inuse + i, n->partial.prev);
3044
Christoph Lameter2086d262007-05-06 14:49:46 -07003045 spin_unlock_irqrestore(&n->list_lock, flags);
3046 }
3047
3048 kfree(slabs_by_inuse);
3049 return 0;
3050}
3051EXPORT_SYMBOL(kmem_cache_shrink);
3052
Yasunori Gotob9049e22007-10-21 16:41:37 -07003053#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
3054static int slab_mem_going_offline_callback(void *arg)
3055{
3056 struct kmem_cache *s;
3057
3058 down_read(&slub_lock);
3059 list_for_each_entry(s, &slab_caches, list)
3060 kmem_cache_shrink(s);
3061 up_read(&slub_lock);
3062
3063 return 0;
3064}
3065
3066static void slab_mem_offline_callback(void *arg)
3067{
3068 struct kmem_cache_node *n;
3069 struct kmem_cache *s;
3070 struct memory_notify *marg = arg;
3071 int offline_node;
3072
3073 offline_node = marg->status_change_nid;
3074
3075 /*
3076 * If the node still has available memory. we need kmem_cache_node
3077 * for it yet.
3078 */
3079 if (offline_node < 0)
3080 return;
3081
3082 down_read(&slub_lock);
3083 list_for_each_entry(s, &slab_caches, list) {
3084 n = get_node(s, offline_node);
3085 if (n) {
3086 /*
3087 * if n->nr_slabs > 0, slabs still exist on the node
3088 * that is going down. We were unable to free them,
3089 * and offline_pages() function shoudn't call this
3090 * callback. So, we must fail.
3091 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003092 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07003093
3094 s->node[offline_node] = NULL;
3095 kmem_cache_free(kmalloc_caches, n);
3096 }
3097 }
3098 up_read(&slub_lock);
3099}
3100
3101static int slab_mem_going_online_callback(void *arg)
3102{
3103 struct kmem_cache_node *n;
3104 struct kmem_cache *s;
3105 struct memory_notify *marg = arg;
3106 int nid = marg->status_change_nid;
3107 int ret = 0;
3108
3109 /*
3110 * If the node's memory is already available, then kmem_cache_node is
3111 * already created. Nothing to do.
3112 */
3113 if (nid < 0)
3114 return 0;
3115
3116 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07003117 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07003118 * allocate a kmem_cache_node structure in order to bring the node
3119 * online.
3120 */
3121 down_read(&slub_lock);
3122 list_for_each_entry(s, &slab_caches, list) {
3123 /*
3124 * XXX: kmem_cache_alloc_node will fallback to other nodes
3125 * since memory is not yet available from the node that
3126 * is brought up.
3127 */
3128 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3129 if (!n) {
3130 ret = -ENOMEM;
3131 goto out;
3132 }
Pekka Enberg5595cff2008-08-05 09:28:47 +03003133 init_kmem_cache_node(n, s);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003134 s->node[nid] = n;
3135 }
3136out:
3137 up_read(&slub_lock);
3138 return ret;
3139}
3140
3141static int slab_memory_callback(struct notifier_block *self,
3142 unsigned long action, void *arg)
3143{
3144 int ret = 0;
3145
3146 switch (action) {
3147 case MEM_GOING_ONLINE:
3148 ret = slab_mem_going_online_callback(arg);
3149 break;
3150 case MEM_GOING_OFFLINE:
3151 ret = slab_mem_going_offline_callback(arg);
3152 break;
3153 case MEM_OFFLINE:
3154 case MEM_CANCEL_ONLINE:
3155 slab_mem_offline_callback(arg);
3156 break;
3157 case MEM_ONLINE:
3158 case MEM_CANCEL_OFFLINE:
3159 break;
3160 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08003161 if (ret)
3162 ret = notifier_from_errno(ret);
3163 else
3164 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003165 return ret;
3166}
3167
3168#endif /* CONFIG_MEMORY_HOTPLUG */
3169
Christoph Lameter81819f02007-05-06 14:49:36 -07003170/********************************************************************
3171 * Basic setup of slabs
3172 *******************************************************************/
3173
3174void __init kmem_cache_init(void)
3175{
3176 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07003177 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003178
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003179 init_alloc_cpu();
3180
Christoph Lameter81819f02007-05-06 14:49:36 -07003181#ifdef CONFIG_NUMA
3182 /*
3183 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07003184 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07003185 * kmem_cache_open for slab_state == DOWN.
3186 */
3187 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
Pekka Enberg83b519e2009-06-10 19:40:04 +03003188 sizeof(struct kmem_cache_node), GFP_NOWAIT);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003189 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07003190 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003191
Nadia Derbey0c40ba42008-04-29 01:00:41 -07003192 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
Christoph Lameter81819f02007-05-06 14:49:36 -07003193#endif
3194
3195 /* Able to allocate the per node structures */
3196 slab_state = PARTIAL;
3197
3198 /* Caches that are not of the two-to-the-power-of size */
Aaro Koskinenacdfcd02009-08-28 14:28:54 +03003199 if (KMALLOC_MIN_SIZE <= 32) {
Christoph Lameter4b356be2007-06-16 10:16:13 -07003200 create_kmalloc_cache(&kmalloc_caches[1],
Pekka Enberg83b519e2009-06-10 19:40:04 +03003201 "kmalloc-96", 96, GFP_NOWAIT);
Christoph Lameter4b356be2007-06-16 10:16:13 -07003202 caches++;
Aaro Koskinenacdfcd02009-08-28 14:28:54 +03003203 }
3204 if (KMALLOC_MIN_SIZE <= 64) {
Christoph Lameter4b356be2007-06-16 10:16:13 -07003205 create_kmalloc_cache(&kmalloc_caches[2],
Pekka Enberg83b519e2009-06-10 19:40:04 +03003206 "kmalloc-192", 192, GFP_NOWAIT);
Christoph Lameter4b356be2007-06-16 10:16:13 -07003207 caches++;
3208 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003209
Christoph Lameterffadd4d2009-02-17 12:05:07 -05003210 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003211 create_kmalloc_cache(&kmalloc_caches[i],
Pekka Enberg83b519e2009-06-10 19:40:04 +03003212 "kmalloc", 1 << i, GFP_NOWAIT);
Christoph Lameter4b356be2007-06-16 10:16:13 -07003213 caches++;
3214 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003215
Christoph Lameterf1b26332007-07-17 04:03:26 -07003216
3217 /*
3218 * Patch up the size_index table if we have strange large alignment
3219 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08003220 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07003221 *
3222 * Largest permitted alignment is 256 bytes due to the way we
3223 * handle the index determination for the smaller caches.
3224 *
3225 * Make sure that nothing crazy happens if someone starts tinkering
3226 * around with ARCH_KMALLOC_MINALIGN
3227 */
3228 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3229 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3230
Aaro Koskinenacdfcd02009-08-28 14:28:54 +03003231 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3232 int elem = size_index_elem(i);
3233 if (elem >= ARRAY_SIZE(size_index))
3234 break;
3235 size_index[elem] = KMALLOC_SHIFT_LOW;
3236 }
Christoph Lameterf1b26332007-07-17 04:03:26 -07003237
Aaro Koskinenacdfcd02009-08-28 14:28:54 +03003238 if (KMALLOC_MIN_SIZE == 64) {
3239 /*
3240 * The 96 byte size cache is not used if the alignment
3241 * is 64 byte.
3242 */
3243 for (i = 64 + 8; i <= 96; i += 8)
3244 size_index[size_index_elem(i)] = 7;
3245 } else if (KMALLOC_MIN_SIZE == 128) {
Christoph Lameter41d54d32008-07-03 09:14:26 -05003246 /*
3247 * The 192 byte sized cache is not used if the alignment
3248 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3249 * instead.
3250 */
3251 for (i = 128 + 8; i <= 192; i += 8)
Aaro Koskinenacdfcd02009-08-28 14:28:54 +03003252 size_index[size_index_elem(i)] = 8;
Christoph Lameter41d54d32008-07-03 09:14:26 -05003253 }
3254
Christoph Lameter81819f02007-05-06 14:49:36 -07003255 slab_state = UP;
3256
3257 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameterffadd4d2009-02-17 12:05:07 -05003258 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07003259 kmalloc_caches[i]. name =
Pekka Enberg83b519e2009-06-10 19:40:04 +03003260 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
Christoph Lameter81819f02007-05-06 14:49:36 -07003261
3262#ifdef CONFIG_SMP
3263 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003264 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3265 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3266#else
3267 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003268#endif
3269
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003270 printk(KERN_INFO
3271 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003272 " CPUs=%d, Nodes=%d\n",
3273 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003274 slub_min_order, slub_max_order, slub_min_objects,
3275 nr_cpu_ids, nr_node_ids);
3276}
3277
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003278void __init kmem_cache_init_late(void)
3279{
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003280}
3281
Christoph Lameter81819f02007-05-06 14:49:36 -07003282/*
3283 * Find a mergeable slab cache
3284 */
3285static int slab_unmergeable(struct kmem_cache *s)
3286{
3287 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3288 return 1;
3289
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003290 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003291 return 1;
3292
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003293 /*
3294 * We may have set a slab to be unmergeable during bootstrap.
3295 */
3296 if (s->refcount < 0)
3297 return 1;
3298
Christoph Lameter81819f02007-05-06 14:49:36 -07003299 return 0;
3300}
3301
3302static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003303 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003304 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003305{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003306 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003307
3308 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3309 return NULL;
3310
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003311 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003312 return NULL;
3313
3314 size = ALIGN(size, sizeof(void *));
3315 align = calculate_alignment(flags, align, size);
3316 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003317 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003318
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003319 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003320 if (slab_unmergeable(s))
3321 continue;
3322
3323 if (size > s->size)
3324 continue;
3325
Christoph Lameterba0268a2007-09-11 15:24:11 -07003326 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003327 continue;
3328 /*
3329 * Check if alignment is compatible.
3330 * Courtesy of Adrian Drzewiecki
3331 */
Pekka Enberg06428782008-01-07 23:20:27 -08003332 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003333 continue;
3334
3335 if (s->size - size >= sizeof(void *))
3336 continue;
3337
3338 return s;
3339 }
3340 return NULL;
3341}
3342
3343struct kmem_cache *kmem_cache_create(const char *name, size_t size,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003344 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003345{
3346 struct kmem_cache *s;
3347
Benjamin Herrenschmidtfe1ff492009-09-21 17:02:30 -07003348 if (WARN_ON(!name))
3349 return NULL;
3350
Christoph Lameter81819f02007-05-06 14:49:36 -07003351 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003352 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003353 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003354 int cpu;
3355
Christoph Lameter81819f02007-05-06 14:49:36 -07003356 s->refcount++;
3357 /*
3358 * Adjust the object sizes so that we clear
3359 * the complete object on kzalloc.
3360 */
3361 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003362
3363 /*
3364 * And then we need to update the object size in the
3365 * per cpu structures
3366 */
3367 for_each_online_cpu(cpu)
3368 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003369
Christoph Lameter81819f02007-05-06 14:49:36 -07003370 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003371 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003372
David Rientjes7b8f3b62008-12-17 22:09:46 -08003373 if (sysfs_slab_alias(s, name)) {
3374 down_write(&slub_lock);
3375 s->refcount--;
3376 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003377 goto err;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003378 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003379 return s;
3380 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003381
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003382 s = kmalloc(kmem_size, GFP_KERNEL);
3383 if (s) {
3384 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003385 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003386 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003387 up_write(&slub_lock);
David Rientjes7b8f3b62008-12-17 22:09:46 -08003388 if (sysfs_slab_add(s)) {
3389 down_write(&slub_lock);
3390 list_del(&s->list);
3391 up_write(&slub_lock);
3392 kfree(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003393 goto err;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003394 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003395 return s;
3396 }
3397 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003398 }
3399 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003400
3401err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003402 if (flags & SLAB_PANIC)
3403 panic("Cannot create slabcache %s\n", name);
3404 else
3405 s = NULL;
3406 return s;
3407}
3408EXPORT_SYMBOL(kmem_cache_create);
3409
Christoph Lameter81819f02007-05-06 14:49:36 -07003410#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003411/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003412 * Use the cpu notifier to insure that the cpu slabs are flushed when
3413 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003414 */
3415static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3416 unsigned long action, void *hcpu)
3417{
3418 long cpu = (long)hcpu;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003419 struct kmem_cache *s;
3420 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003421
3422 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003423 case CPU_UP_PREPARE:
3424 case CPU_UP_PREPARE_FROZEN:
3425 init_alloc_cpu_cpu(cpu);
3426 down_read(&slub_lock);
3427 list_for_each_entry(s, &slab_caches, list)
3428 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3429 GFP_KERNEL);
3430 up_read(&slub_lock);
3431 break;
3432
Christoph Lameter81819f02007-05-06 14:49:36 -07003433 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003434 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003435 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003436 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003437 down_read(&slub_lock);
3438 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003439 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3440
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003441 local_irq_save(flags);
3442 __flush_cpu_slab(s, cpu);
3443 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003444 free_kmem_cache_cpu(c, cpu);
3445 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003446 }
3447 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003448 break;
3449 default:
3450 break;
3451 }
3452 return NOTIFY_OK;
3453}
3454
Pekka Enberg06428782008-01-07 23:20:27 -08003455static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003456 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003457};
Christoph Lameter81819f02007-05-06 14:49:36 -07003458
3459#endif
3460
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003461void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003462{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003463 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003464 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003465
Christoph Lameterffadd4d2009-02-17 12:05:07 -05003466 if (unlikely(size > SLUB_MAX_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003467 return kmalloc_large(size, gfpflags);
3468
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003469 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003470
Satyam Sharma2408c552007-10-16 01:24:44 -07003471 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003472 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003473
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003474 ret = slab_alloc(s, gfpflags, -1, caller);
3475
3476 /* Honor the call site pointer we recieved. */
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003477 trace_kmalloc(caller, ret, size, s->size, gfpflags);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003478
3479 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003480}
3481
3482void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003483 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003484{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003485 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003486 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003487
Christoph Lameterffadd4d2009-02-17 12:05:07 -05003488 if (unlikely(size > SLUB_MAX_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003489 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003490
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003491 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003492
Satyam Sharma2408c552007-10-16 01:24:44 -07003493 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003494 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003495
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003496 ret = slab_alloc(s, gfpflags, node, caller);
3497
3498 /* Honor the call site pointer we recieved. */
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003499 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003500
3501 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003502}
3503
Christoph Lameterf6acb632008-04-29 16:16:06 -07003504#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03003505static int count_inuse(struct page *page)
3506{
3507 return page->inuse;
3508}
3509
3510static int count_total(struct page *page)
3511{
3512 return page->objects;
3513}
3514
Christoph Lameter434e2452007-07-17 04:03:30 -07003515static int validate_slab(struct kmem_cache *s, struct page *page,
3516 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003517{
3518 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003519 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003520
3521 if (!check_slab(s, page) ||
3522 !on_freelist(s, page, NULL))
3523 return 0;
3524
3525 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003526 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003527
Christoph Lameter7656c722007-05-09 02:32:40 -07003528 for_each_free_object(p, s, page->freelist) {
3529 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003530 if (!check_object(s, page, p, 0))
3531 return 0;
3532 }
3533
Christoph Lameter224a88b2008-04-14 19:11:31 +03003534 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003535 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003536 if (!check_object(s, page, p, 1))
3537 return 0;
3538 return 1;
3539}
3540
Christoph Lameter434e2452007-07-17 04:03:30 -07003541static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3542 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003543{
3544 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003545 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003546 slab_unlock(page);
3547 } else
3548 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3549 s->name, page);
3550
3551 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003552 if (!PageSlubDebug(page))
3553 printk(KERN_ERR "SLUB %s: SlubDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003554 "on slab 0x%p\n", s->name, page);
3555 } else {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003556 if (PageSlubDebug(page))
3557 printk(KERN_ERR "SLUB %s: SlubDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003558 "slab 0x%p\n", s->name, page);
3559 }
3560}
3561
Christoph Lameter434e2452007-07-17 04:03:30 -07003562static int validate_slab_node(struct kmem_cache *s,
3563 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003564{
3565 unsigned long count = 0;
3566 struct page *page;
3567 unsigned long flags;
3568
3569 spin_lock_irqsave(&n->list_lock, flags);
3570
3571 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003572 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003573 count++;
3574 }
3575 if (count != n->nr_partial)
3576 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3577 "counter=%ld\n", s->name, count, n->nr_partial);
3578
3579 if (!(s->flags & SLAB_STORE_USER))
3580 goto out;
3581
3582 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003583 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003584 count++;
3585 }
3586 if (count != atomic_long_read(&n->nr_slabs))
3587 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3588 "counter=%ld\n", s->name, count,
3589 atomic_long_read(&n->nr_slabs));
3590
3591out:
3592 spin_unlock_irqrestore(&n->list_lock, flags);
3593 return count;
3594}
3595
Christoph Lameter434e2452007-07-17 04:03:30 -07003596static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003597{
3598 int node;
3599 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003600 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003601 sizeof(unsigned long), GFP_KERNEL);
3602
3603 if (!map)
3604 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003605
3606 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003607 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003608 struct kmem_cache_node *n = get_node(s, node);
3609
Christoph Lameter434e2452007-07-17 04:03:30 -07003610 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003611 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003612 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003613 return count;
3614}
3615
Christoph Lameterb3459702007-05-09 02:32:41 -07003616#ifdef SLUB_RESILIENCY_TEST
3617static void resiliency_test(void)
3618{
3619 u8 *p;
3620
3621 printk(KERN_ERR "SLUB resiliency testing\n");
3622 printk(KERN_ERR "-----------------------\n");
3623 printk(KERN_ERR "A. Corruption after allocation\n");
3624
3625 p = kzalloc(16, GFP_KERNEL);
3626 p[16] = 0x12;
3627 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3628 " 0x12->0x%p\n\n", p + 16);
3629
3630 validate_slab_cache(kmalloc_caches + 4);
3631
3632 /* Hmmm... The next two are dangerous */
3633 p = kzalloc(32, GFP_KERNEL);
3634 p[32 + sizeof(void *)] = 0x34;
3635 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003636 " 0x34 -> -0x%p\n", p);
3637 printk(KERN_ERR
3638 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003639
3640 validate_slab_cache(kmalloc_caches + 5);
3641 p = kzalloc(64, GFP_KERNEL);
3642 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3643 *p = 0x56;
3644 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3645 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003646 printk(KERN_ERR
3647 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003648 validate_slab_cache(kmalloc_caches + 6);
3649
3650 printk(KERN_ERR "\nB. Corruption after free\n");
3651 p = kzalloc(128, GFP_KERNEL);
3652 kfree(p);
3653 *p = 0x78;
3654 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3655 validate_slab_cache(kmalloc_caches + 7);
3656
3657 p = kzalloc(256, GFP_KERNEL);
3658 kfree(p);
3659 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003660 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3661 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003662 validate_slab_cache(kmalloc_caches + 8);
3663
3664 p = kzalloc(512, GFP_KERNEL);
3665 kfree(p);
3666 p[512] = 0xab;
3667 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3668 validate_slab_cache(kmalloc_caches + 9);
3669}
3670#else
3671static void resiliency_test(void) {};
3672#endif
3673
Christoph Lameter88a420e2007-05-06 14:49:45 -07003674/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003675 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003676 * and freed.
3677 */
3678
3679struct location {
3680 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003681 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003682 long long sum_time;
3683 long min_time;
3684 long max_time;
3685 long min_pid;
3686 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303687 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003688 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003689};
3690
3691struct loc_track {
3692 unsigned long max;
3693 unsigned long count;
3694 struct location *loc;
3695};
3696
3697static void free_loc_track(struct loc_track *t)
3698{
3699 if (t->max)
3700 free_pages((unsigned long)t->loc,
3701 get_order(sizeof(struct location) * t->max));
3702}
3703
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003704static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003705{
3706 struct location *l;
3707 int order;
3708
Christoph Lameter88a420e2007-05-06 14:49:45 -07003709 order = get_order(sizeof(struct location) * max);
3710
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003711 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003712 if (!l)
3713 return 0;
3714
3715 if (t->count) {
3716 memcpy(l, t->loc, sizeof(struct location) * t->count);
3717 free_loc_track(t);
3718 }
3719 t->max = max;
3720 t->loc = l;
3721 return 1;
3722}
3723
3724static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003725 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003726{
3727 long start, end, pos;
3728 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003729 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003730 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003731
3732 start = -1;
3733 end = t->count;
3734
3735 for ( ; ; ) {
3736 pos = start + (end - start + 1) / 2;
3737
3738 /*
3739 * There is nothing at "end". If we end up there
3740 * we need to add something to before end.
3741 */
3742 if (pos == end)
3743 break;
3744
3745 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003746 if (track->addr == caddr) {
3747
3748 l = &t->loc[pos];
3749 l->count++;
3750 if (track->when) {
3751 l->sum_time += age;
3752 if (age < l->min_time)
3753 l->min_time = age;
3754 if (age > l->max_time)
3755 l->max_time = age;
3756
3757 if (track->pid < l->min_pid)
3758 l->min_pid = track->pid;
3759 if (track->pid > l->max_pid)
3760 l->max_pid = track->pid;
3761
Rusty Russell174596a2009-01-01 10:12:29 +10303762 cpumask_set_cpu(track->cpu,
3763 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003764 }
3765 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003766 return 1;
3767 }
3768
Christoph Lameter45edfa52007-05-09 02:32:45 -07003769 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003770 end = pos;
3771 else
3772 start = pos;
3773 }
3774
3775 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003776 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003777 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003778 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003779 return 0;
3780
3781 l = t->loc + pos;
3782 if (pos < t->count)
3783 memmove(l + 1, l,
3784 (t->count - pos) * sizeof(struct location));
3785 t->count++;
3786 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003787 l->addr = track->addr;
3788 l->sum_time = age;
3789 l->min_time = age;
3790 l->max_time = age;
3791 l->min_pid = track->pid;
3792 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303793 cpumask_clear(to_cpumask(l->cpus));
3794 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003795 nodes_clear(l->nodes);
3796 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003797 return 1;
3798}
3799
3800static void process_slab(struct loc_track *t, struct kmem_cache *s,
3801 struct page *page, enum track_item alloc)
3802{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003803 void *addr = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +03003804 DECLARE_BITMAP(map, page->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003805 void *p;
3806
Christoph Lameter39b26462008-04-14 19:11:30 +03003807 bitmap_zero(map, page->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003808 for_each_free_object(p, s, page->freelist)
3809 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003810
Christoph Lameter224a88b2008-04-14 19:11:31 +03003811 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003812 if (!test_bit(slab_index(p, s, addr), map))
3813 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003814}
3815
3816static int list_locations(struct kmem_cache *s, char *buf,
3817 enum track_item alloc)
3818{
Harvey Harrisone374d482008-01-31 15:20:50 -08003819 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003820 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003821 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003822 int node;
3823
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003824 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003825 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003826 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003827
3828 /* Push back cpu slabs */
3829 flush_all(s);
3830
Christoph Lameterf64dc582007-10-16 01:25:33 -07003831 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003832 struct kmem_cache_node *n = get_node(s, node);
3833 unsigned long flags;
3834 struct page *page;
3835
Christoph Lameter9e869432007-08-22 14:01:56 -07003836 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003837 continue;
3838
3839 spin_lock_irqsave(&n->list_lock, flags);
3840 list_for_each_entry(page, &n->partial, lru)
3841 process_slab(&t, s, page, alloc);
3842 list_for_each_entry(page, &n->full, lru)
3843 process_slab(&t, s, page, alloc);
3844 spin_unlock_irqrestore(&n->list_lock, flags);
3845 }
3846
3847 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003848 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003849
Hugh Dickins9c246242008-12-09 13:14:27 -08003850 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003851 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003852 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003853
3854 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003855 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003856 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003857 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003858
3859 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08003860 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07003861 l->min_time,
3862 (long)div_u64(l->sum_time, l->count),
3863 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003864 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003865 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003866 l->min_time);
3867
3868 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003869 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003870 l->min_pid, l->max_pid);
3871 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003872 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003873 l->min_pid);
3874
Rusty Russell174596a2009-01-01 10:12:29 +10303875 if (num_online_cpus() > 1 &&
3876 !cpumask_empty(to_cpumask(l->cpus)) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003877 len < PAGE_SIZE - 60) {
3878 len += sprintf(buf + len, " cpus=");
3879 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Rusty Russell174596a2009-01-01 10:12:29 +10303880 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003881 }
3882
Christoph Lameter62bc62a2009-06-16 15:32:15 -07003883 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003884 len < PAGE_SIZE - 60) {
3885 len += sprintf(buf + len, " nodes=");
3886 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003887 l->nodes);
3888 }
3889
Harvey Harrisone374d482008-01-31 15:20:50 -08003890 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003891 }
3892
3893 free_loc_track(&t);
3894 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003895 len += sprintf(buf, "No data\n");
3896 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003897}
3898
Christoph Lameter81819f02007-05-06 14:49:36 -07003899enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03003900 SL_ALL, /* All slabs */
3901 SL_PARTIAL, /* Only partially allocated slabs */
3902 SL_CPU, /* Only slabs used for cpu caches */
3903 SL_OBJECTS, /* Determine allocated objects not slabs */
3904 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07003905};
3906
Christoph Lameter205ab992008-04-14 19:11:40 +03003907#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003908#define SO_PARTIAL (1 << SL_PARTIAL)
3909#define SO_CPU (1 << SL_CPU)
3910#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03003911#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003912
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003913static ssize_t show_slab_objects(struct kmem_cache *s,
3914 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003915{
3916 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003917 int node;
3918 int x;
3919 unsigned long *nodes;
3920 unsigned long *per_cpu;
3921
3922 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003923 if (!nodes)
3924 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003925 per_cpu = nodes + nr_node_ids;
3926
Christoph Lameter205ab992008-04-14 19:11:40 +03003927 if (flags & SO_CPU) {
3928 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07003929
Christoph Lameter205ab992008-04-14 19:11:40 +03003930 for_each_possible_cpu(cpu) {
3931 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003932
Christoph Lameter205ab992008-04-14 19:11:40 +03003933 if (!c || c->node < 0)
3934 continue;
3935
3936 if (c->page) {
3937 if (flags & SO_TOTAL)
3938 x = c->page->objects;
3939 else if (flags & SO_OBJECTS)
3940 x = c->page->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07003941 else
3942 x = 1;
Christoph Lameter205ab992008-04-14 19:11:40 +03003943
Christoph Lameter81819f02007-05-06 14:49:36 -07003944 total += x;
Christoph Lameter205ab992008-04-14 19:11:40 +03003945 nodes[c->node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003946 }
Christoph Lameter205ab992008-04-14 19:11:40 +03003947 per_cpu[c->node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003948 }
3949 }
3950
Christoph Lameter205ab992008-04-14 19:11:40 +03003951 if (flags & SO_ALL) {
3952 for_each_node_state(node, N_NORMAL_MEMORY) {
3953 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003954
Christoph Lameter205ab992008-04-14 19:11:40 +03003955 if (flags & SO_TOTAL)
3956 x = atomic_long_read(&n->total_objects);
3957 else if (flags & SO_OBJECTS)
3958 x = atomic_long_read(&n->total_objects) -
3959 count_partial(n, count_free);
3960
3961 else
3962 x = atomic_long_read(&n->nr_slabs);
3963 total += x;
3964 nodes[node] += x;
3965 }
3966
3967 } else if (flags & SO_PARTIAL) {
3968 for_each_node_state(node, N_NORMAL_MEMORY) {
3969 struct kmem_cache_node *n = get_node(s, node);
3970
3971 if (flags & SO_TOTAL)
3972 x = count_partial(n, count_total);
3973 else if (flags & SO_OBJECTS)
3974 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07003975 else
3976 x = n->nr_partial;
3977 total += x;
3978 nodes[node] += x;
3979 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003980 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003981 x = sprintf(buf, "%lu", total);
3982#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003983 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003984 if (nodes[node])
3985 x += sprintf(buf + x, " N%d=%lu",
3986 node, nodes[node]);
3987#endif
3988 kfree(nodes);
3989 return x + sprintf(buf + x, "\n");
3990}
3991
3992static int any_slab_objects(struct kmem_cache *s)
3993{
3994 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003995
3996 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003997 struct kmem_cache_node *n = get_node(s, node);
3998
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003999 if (!n)
4000 continue;
4001
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07004002 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07004003 return 1;
4004 }
4005 return 0;
4006}
4007
4008#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4009#define to_slab(n) container_of(n, struct kmem_cache, kobj);
4010
4011struct slab_attribute {
4012 struct attribute attr;
4013 ssize_t (*show)(struct kmem_cache *s, char *buf);
4014 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4015};
4016
4017#define SLAB_ATTR_RO(_name) \
4018 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4019
4020#define SLAB_ATTR(_name) \
4021 static struct slab_attribute _name##_attr = \
4022 __ATTR(_name, 0644, _name##_show, _name##_store)
4023
Christoph Lameter81819f02007-05-06 14:49:36 -07004024static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4025{
4026 return sprintf(buf, "%d\n", s->size);
4027}
4028SLAB_ATTR_RO(slab_size);
4029
4030static ssize_t align_show(struct kmem_cache *s, char *buf)
4031{
4032 return sprintf(buf, "%d\n", s->align);
4033}
4034SLAB_ATTR_RO(align);
4035
4036static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4037{
4038 return sprintf(buf, "%d\n", s->objsize);
4039}
4040SLAB_ATTR_RO(object_size);
4041
4042static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4043{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004044 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004045}
4046SLAB_ATTR_RO(objs_per_slab);
4047
Christoph Lameter06b285d2008-04-14 19:11:41 +03004048static ssize_t order_store(struct kmem_cache *s,
4049 const char *buf, size_t length)
4050{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004051 unsigned long order;
4052 int err;
4053
4054 err = strict_strtoul(buf, 10, &order);
4055 if (err)
4056 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004057
4058 if (order > slub_max_order || order < slub_min_order)
4059 return -EINVAL;
4060
4061 calculate_sizes(s, order);
4062 return length;
4063}
4064
Christoph Lameter81819f02007-05-06 14:49:36 -07004065static ssize_t order_show(struct kmem_cache *s, char *buf)
4066{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004067 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004068}
Christoph Lameter06b285d2008-04-14 19:11:41 +03004069SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07004070
David Rientjes73d342b2009-02-22 17:40:09 -08004071static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4072{
4073 return sprintf(buf, "%lu\n", s->min_partial);
4074}
4075
4076static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4077 size_t length)
4078{
4079 unsigned long min;
4080 int err;
4081
4082 err = strict_strtoul(buf, 10, &min);
4083 if (err)
4084 return err;
4085
David Rientjesc0bdb232009-02-25 09:16:35 +02004086 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08004087 return length;
4088}
4089SLAB_ATTR(min_partial);
4090
Christoph Lameter81819f02007-05-06 14:49:36 -07004091static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4092{
4093 if (s->ctor) {
4094 int n = sprint_symbol(buf, (unsigned long)s->ctor);
4095
4096 return n + sprintf(buf + n, "\n");
4097 }
4098 return 0;
4099}
4100SLAB_ATTR_RO(ctor);
4101
Christoph Lameter81819f02007-05-06 14:49:36 -07004102static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4103{
4104 return sprintf(buf, "%d\n", s->refcount - 1);
4105}
4106SLAB_ATTR_RO(aliases);
4107
4108static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4109{
Christoph Lameter205ab992008-04-14 19:11:40 +03004110 return show_slab_objects(s, buf, SO_ALL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004111}
4112SLAB_ATTR_RO(slabs);
4113
4114static ssize_t partial_show(struct kmem_cache *s, char *buf)
4115{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004116 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004117}
4118SLAB_ATTR_RO(partial);
4119
4120static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4121{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004122 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07004123}
4124SLAB_ATTR_RO(cpu_slabs);
4125
4126static ssize_t objects_show(struct kmem_cache *s, char *buf)
4127{
Christoph Lameter205ab992008-04-14 19:11:40 +03004128 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07004129}
4130SLAB_ATTR_RO(objects);
4131
Christoph Lameter205ab992008-04-14 19:11:40 +03004132static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4133{
4134 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4135}
4136SLAB_ATTR_RO(objects_partial);
4137
4138static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4139{
4140 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4141}
4142SLAB_ATTR_RO(total_objects);
4143
Christoph Lameter81819f02007-05-06 14:49:36 -07004144static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4145{
4146 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4147}
4148
4149static ssize_t sanity_checks_store(struct kmem_cache *s,
4150 const char *buf, size_t length)
4151{
4152 s->flags &= ~SLAB_DEBUG_FREE;
4153 if (buf[0] == '1')
4154 s->flags |= SLAB_DEBUG_FREE;
4155 return length;
4156}
4157SLAB_ATTR(sanity_checks);
4158
4159static ssize_t trace_show(struct kmem_cache *s, char *buf)
4160{
4161 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4162}
4163
4164static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4165 size_t length)
4166{
4167 s->flags &= ~SLAB_TRACE;
4168 if (buf[0] == '1')
4169 s->flags |= SLAB_TRACE;
4170 return length;
4171}
4172SLAB_ATTR(trace);
4173
4174static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4175{
4176 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4177}
4178
4179static ssize_t reclaim_account_store(struct kmem_cache *s,
4180 const char *buf, size_t length)
4181{
4182 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4183 if (buf[0] == '1')
4184 s->flags |= SLAB_RECLAIM_ACCOUNT;
4185 return length;
4186}
4187SLAB_ATTR(reclaim_account);
4188
4189static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4190{
Christoph Lameter5af60832007-05-06 14:49:56 -07004191 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07004192}
4193SLAB_ATTR_RO(hwcache_align);
4194
4195#ifdef CONFIG_ZONE_DMA
4196static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4197{
4198 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4199}
4200SLAB_ATTR_RO(cache_dma);
4201#endif
4202
4203static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4204{
4205 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4206}
4207SLAB_ATTR_RO(destroy_by_rcu);
4208
4209static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4210{
4211 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4212}
4213
4214static ssize_t red_zone_store(struct kmem_cache *s,
4215 const char *buf, size_t length)
4216{
4217 if (any_slab_objects(s))
4218 return -EBUSY;
4219
4220 s->flags &= ~SLAB_RED_ZONE;
4221 if (buf[0] == '1')
4222 s->flags |= SLAB_RED_ZONE;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004223 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004224 return length;
4225}
4226SLAB_ATTR(red_zone);
4227
4228static ssize_t poison_show(struct kmem_cache *s, char *buf)
4229{
4230 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4231}
4232
4233static ssize_t poison_store(struct kmem_cache *s,
4234 const char *buf, size_t length)
4235{
4236 if (any_slab_objects(s))
4237 return -EBUSY;
4238
4239 s->flags &= ~SLAB_POISON;
4240 if (buf[0] == '1')
4241 s->flags |= SLAB_POISON;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004242 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004243 return length;
4244}
4245SLAB_ATTR(poison);
4246
4247static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4248{
4249 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4250}
4251
4252static ssize_t store_user_store(struct kmem_cache *s,
4253 const char *buf, size_t length)
4254{
4255 if (any_slab_objects(s))
4256 return -EBUSY;
4257
4258 s->flags &= ~SLAB_STORE_USER;
4259 if (buf[0] == '1')
4260 s->flags |= SLAB_STORE_USER;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004261 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004262 return length;
4263}
4264SLAB_ATTR(store_user);
4265
Christoph Lameter53e15af2007-05-06 14:49:43 -07004266static ssize_t validate_show(struct kmem_cache *s, char *buf)
4267{
4268 return 0;
4269}
4270
4271static ssize_t validate_store(struct kmem_cache *s,
4272 const char *buf, size_t length)
4273{
Christoph Lameter434e2452007-07-17 04:03:30 -07004274 int ret = -EINVAL;
4275
4276 if (buf[0] == '1') {
4277 ret = validate_slab_cache(s);
4278 if (ret >= 0)
4279 ret = length;
4280 }
4281 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004282}
4283SLAB_ATTR(validate);
4284
Christoph Lameter2086d262007-05-06 14:49:46 -07004285static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4286{
4287 return 0;
4288}
4289
4290static ssize_t shrink_store(struct kmem_cache *s,
4291 const char *buf, size_t length)
4292{
4293 if (buf[0] == '1') {
4294 int rc = kmem_cache_shrink(s);
4295
4296 if (rc)
4297 return rc;
4298 } else
4299 return -EINVAL;
4300 return length;
4301}
4302SLAB_ATTR(shrink);
4303
Christoph Lameter88a420e2007-05-06 14:49:45 -07004304static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4305{
4306 if (!(s->flags & SLAB_STORE_USER))
4307 return -ENOSYS;
4308 return list_locations(s, buf, TRACK_ALLOC);
4309}
4310SLAB_ATTR_RO(alloc_calls);
4311
4312static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4313{
4314 if (!(s->flags & SLAB_STORE_USER))
4315 return -ENOSYS;
4316 return list_locations(s, buf, TRACK_FREE);
4317}
4318SLAB_ATTR_RO(free_calls);
4319
Christoph Lameter81819f02007-05-06 14:49:36 -07004320#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004321static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004322{
Christoph Lameter98246012008-01-07 23:20:26 -08004323 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004324}
4325
Christoph Lameter98246012008-01-07 23:20:26 -08004326static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004327 const char *buf, size_t length)
4328{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004329 unsigned long ratio;
4330 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004331
Christoph Lameter0121c6192008-04-29 16:11:12 -07004332 err = strict_strtoul(buf, 10, &ratio);
4333 if (err)
4334 return err;
4335
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004336 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004337 s->remote_node_defrag_ratio = ratio * 10;
4338
Christoph Lameter81819f02007-05-06 14:49:36 -07004339 return length;
4340}
Christoph Lameter98246012008-01-07 23:20:26 -08004341SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004342#endif
4343
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004344#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004345static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4346{
4347 unsigned long sum = 0;
4348 int cpu;
4349 int len;
4350 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4351
4352 if (!data)
4353 return -ENOMEM;
4354
4355 for_each_online_cpu(cpu) {
4356 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4357
4358 data[cpu] = x;
4359 sum += x;
4360 }
4361
4362 len = sprintf(buf, "%lu", sum);
4363
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004364#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004365 for_each_online_cpu(cpu) {
4366 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004367 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004368 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004369#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004370 kfree(data);
4371 return len + sprintf(buf + len, "\n");
4372}
4373
4374#define STAT_ATTR(si, text) \
4375static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4376{ \
4377 return show_stat(s, buf, si); \
4378} \
4379SLAB_ATTR_RO(text); \
4380
4381STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4382STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4383STAT_ATTR(FREE_FASTPATH, free_fastpath);
4384STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4385STAT_ATTR(FREE_FROZEN, free_frozen);
4386STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4387STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4388STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4389STAT_ATTR(ALLOC_SLAB, alloc_slab);
4390STAT_ATTR(ALLOC_REFILL, alloc_refill);
4391STAT_ATTR(FREE_SLAB, free_slab);
4392STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4393STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4394STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4395STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4396STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4397STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter65c33762008-04-14 19:11:40 +03004398STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004399#endif
4400
Pekka Enberg06428782008-01-07 23:20:27 -08004401static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004402 &slab_size_attr.attr,
4403 &object_size_attr.attr,
4404 &objs_per_slab_attr.attr,
4405 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08004406 &min_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004407 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004408 &objects_partial_attr.attr,
4409 &total_objects_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004410 &slabs_attr.attr,
4411 &partial_attr.attr,
4412 &cpu_slabs_attr.attr,
4413 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004414 &aliases_attr.attr,
4415 &align_attr.attr,
4416 &sanity_checks_attr.attr,
4417 &trace_attr.attr,
4418 &hwcache_align_attr.attr,
4419 &reclaim_account_attr.attr,
4420 &destroy_by_rcu_attr.attr,
4421 &red_zone_attr.attr,
4422 &poison_attr.attr,
4423 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004424 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004425 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004426 &alloc_calls_attr.attr,
4427 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004428#ifdef CONFIG_ZONE_DMA
4429 &cache_dma_attr.attr,
4430#endif
4431#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004432 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004433#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004434#ifdef CONFIG_SLUB_STATS
4435 &alloc_fastpath_attr.attr,
4436 &alloc_slowpath_attr.attr,
4437 &free_fastpath_attr.attr,
4438 &free_slowpath_attr.attr,
4439 &free_frozen_attr.attr,
4440 &free_add_partial_attr.attr,
4441 &free_remove_partial_attr.attr,
4442 &alloc_from_partial_attr.attr,
4443 &alloc_slab_attr.attr,
4444 &alloc_refill_attr.attr,
4445 &free_slab_attr.attr,
4446 &cpuslab_flush_attr.attr,
4447 &deactivate_full_attr.attr,
4448 &deactivate_empty_attr.attr,
4449 &deactivate_to_head_attr.attr,
4450 &deactivate_to_tail_attr.attr,
4451 &deactivate_remote_frees_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004452 &order_fallback_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004453#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004454 NULL
4455};
4456
4457static struct attribute_group slab_attr_group = {
4458 .attrs = slab_attrs,
4459};
4460
4461static ssize_t slab_attr_show(struct kobject *kobj,
4462 struct attribute *attr,
4463 char *buf)
4464{
4465 struct slab_attribute *attribute;
4466 struct kmem_cache *s;
4467 int err;
4468
4469 attribute = to_slab_attr(attr);
4470 s = to_slab(kobj);
4471
4472 if (!attribute->show)
4473 return -EIO;
4474
4475 err = attribute->show(s, buf);
4476
4477 return err;
4478}
4479
4480static ssize_t slab_attr_store(struct kobject *kobj,
4481 struct attribute *attr,
4482 const char *buf, size_t len)
4483{
4484 struct slab_attribute *attribute;
4485 struct kmem_cache *s;
4486 int err;
4487
4488 attribute = to_slab_attr(attr);
4489 s = to_slab(kobj);
4490
4491 if (!attribute->store)
4492 return -EIO;
4493
4494 err = attribute->store(s, buf, len);
4495
4496 return err;
4497}
4498
Christoph Lameter151c6022008-01-07 22:29:05 -08004499static void kmem_cache_release(struct kobject *kobj)
4500{
4501 struct kmem_cache *s = to_slab(kobj);
4502
4503 kfree(s);
4504}
4505
Christoph Lameter81819f02007-05-06 14:49:36 -07004506static struct sysfs_ops slab_sysfs_ops = {
4507 .show = slab_attr_show,
4508 .store = slab_attr_store,
4509};
4510
4511static struct kobj_type slab_ktype = {
4512 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004513 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004514};
4515
4516static int uevent_filter(struct kset *kset, struct kobject *kobj)
4517{
4518 struct kobj_type *ktype = get_ktype(kobj);
4519
4520 if (ktype == &slab_ktype)
4521 return 1;
4522 return 0;
4523}
4524
4525static struct kset_uevent_ops slab_uevent_ops = {
4526 .filter = uevent_filter,
4527};
4528
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004529static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004530
4531#define ID_STR_LENGTH 64
4532
4533/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004534 *
4535 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004536 */
4537static char *create_unique_id(struct kmem_cache *s)
4538{
4539 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4540 char *p = name;
4541
4542 BUG_ON(!name);
4543
4544 *p++ = ':';
4545 /*
4546 * First flags affecting slabcache operations. We will only
4547 * get here for aliasable slabs so we do not need to support
4548 * too many flags. The flags here must cover all flags that
4549 * are matched during merging to guarantee that the id is
4550 * unique.
4551 */
4552 if (s->flags & SLAB_CACHE_DMA)
4553 *p++ = 'd';
4554 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4555 *p++ = 'a';
4556 if (s->flags & SLAB_DEBUG_FREE)
4557 *p++ = 'F';
Vegard Nossum5a896d92008-04-04 00:54:48 +02004558 if (!(s->flags & SLAB_NOTRACK))
4559 *p++ = 't';
Christoph Lameter81819f02007-05-06 14:49:36 -07004560 if (p != name + 1)
4561 *p++ = '-';
4562 p += sprintf(p, "%07d", s->size);
4563 BUG_ON(p > name + ID_STR_LENGTH - 1);
4564 return name;
4565}
4566
4567static int sysfs_slab_add(struct kmem_cache *s)
4568{
4569 int err;
4570 const char *name;
4571 int unmergeable;
4572
4573 if (slab_state < SYSFS)
4574 /* Defer until later */
4575 return 0;
4576
4577 unmergeable = slab_unmergeable(s);
4578 if (unmergeable) {
4579 /*
4580 * Slabcache can never be merged so we can use the name proper.
4581 * This is typically the case for debug situations. In that
4582 * case we can catch duplicate names easily.
4583 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004584 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004585 name = s->name;
4586 } else {
4587 /*
4588 * Create a unique name for the slab as a target
4589 * for the symlinks.
4590 */
4591 name = create_unique_id(s);
4592 }
4593
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004594 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004595 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4596 if (err) {
4597 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004598 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004599 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004600
4601 err = sysfs_create_group(&s->kobj, &slab_attr_group);
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08004602 if (err) {
4603 kobject_del(&s->kobj);
4604 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004605 return err;
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08004606 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004607 kobject_uevent(&s->kobj, KOBJ_ADD);
4608 if (!unmergeable) {
4609 /* Setup first alias */
4610 sysfs_slab_alias(s, s->name);
4611 kfree(name);
4612 }
4613 return 0;
4614}
4615
4616static void sysfs_slab_remove(struct kmem_cache *s)
4617{
4618 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4619 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004620 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004621}
4622
4623/*
4624 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11004625 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07004626 */
4627struct saved_alias {
4628 struct kmem_cache *s;
4629 const char *name;
4630 struct saved_alias *next;
4631};
4632
Adrian Bunk5af328a2007-07-17 04:03:27 -07004633static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004634
4635static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4636{
4637 struct saved_alias *al;
4638
4639 if (slab_state == SYSFS) {
4640 /*
4641 * If we have a leftover link then remove it.
4642 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004643 sysfs_remove_link(&slab_kset->kobj, name);
4644 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004645 }
4646
4647 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4648 if (!al)
4649 return -ENOMEM;
4650
4651 al->s = s;
4652 al->name = name;
4653 al->next = alias_list;
4654 alias_list = al;
4655 return 0;
4656}
4657
4658static int __init slab_sysfs_init(void)
4659{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004660 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004661 int err;
4662
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004663 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004664 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004665 printk(KERN_ERR "Cannot register slab subsystem.\n");
4666 return -ENOSYS;
4667 }
4668
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004669 slab_state = SYSFS;
4670
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004671 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004672 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004673 if (err)
4674 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4675 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004676 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004677
4678 while (alias_list) {
4679 struct saved_alias *al = alias_list;
4680
4681 alias_list = alias_list->next;
4682 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004683 if (err)
4684 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4685 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004686 kfree(al);
4687 }
4688
4689 resiliency_test();
4690 return 0;
4691}
4692
4693__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004694#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004695
4696/*
4697 * The /proc/slabinfo ABI
4698 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004699#ifdef CONFIG_SLABINFO
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004700static void print_slabinfo_header(struct seq_file *m)
4701{
4702 seq_puts(m, "slabinfo - version: 2.1\n");
4703 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4704 "<objperslab> <pagesperslab>");
4705 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4706 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4707 seq_putc(m, '\n');
4708}
4709
4710static void *s_start(struct seq_file *m, loff_t *pos)
4711{
4712 loff_t n = *pos;
4713
4714 down_read(&slub_lock);
4715 if (!n)
4716 print_slabinfo_header(m);
4717
4718 return seq_list_start(&slab_caches, *pos);
4719}
4720
4721static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4722{
4723 return seq_list_next(p, &slab_caches, pos);
4724}
4725
4726static void s_stop(struct seq_file *m, void *p)
4727{
4728 up_read(&slub_lock);
4729}
4730
4731static int s_show(struct seq_file *m, void *p)
4732{
4733 unsigned long nr_partials = 0;
4734 unsigned long nr_slabs = 0;
4735 unsigned long nr_inuse = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004736 unsigned long nr_objs = 0;
4737 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004738 struct kmem_cache *s;
4739 int node;
4740
4741 s = list_entry(p, struct kmem_cache, list);
4742
4743 for_each_online_node(node) {
4744 struct kmem_cache_node *n = get_node(s, node);
4745
4746 if (!n)
4747 continue;
4748
4749 nr_partials += n->nr_partial;
4750 nr_slabs += atomic_long_read(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03004751 nr_objs += atomic_long_read(&n->total_objects);
4752 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004753 }
4754
Christoph Lameter205ab992008-04-14 19:11:40 +03004755 nr_inuse = nr_objs - nr_free;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004756
4757 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
Christoph Lameter834f3d12008-04-14 19:11:31 +03004758 nr_objs, s->size, oo_objects(s->oo),
4759 (1 << oo_order(s->oo)));
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004760 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4761 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4762 0UL);
4763 seq_putc(m, '\n');
4764 return 0;
4765}
4766
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004767static const struct seq_operations slabinfo_op = {
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004768 .start = s_start,
4769 .next = s_next,
4770 .stop = s_stop,
4771 .show = s_show,
4772};
4773
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004774static int slabinfo_open(struct inode *inode, struct file *file)
4775{
4776 return seq_open(file, &slabinfo_op);
4777}
4778
4779static const struct file_operations proc_slabinfo_operations = {
4780 .open = slabinfo_open,
4781 .read = seq_read,
4782 .llseek = seq_lseek,
4783 .release = seq_release,
4784};
4785
4786static int __init slab_proc_init(void)
4787{
WANG Congcf5d1132009-08-18 19:11:40 +03004788 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004789 return 0;
4790}
4791module_init(slab_proc_init);
Linus Torvalds158a9622008-01-02 13:04:48 -08004792#endif /* CONFIG_SLABINFO */