blob: bdc9abb08a230f2724ad842a0bc4928da0e48d23 [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>
12#include <linux/module.h>
13#include <linux/bit_spinlock.h>
14#include <linux/interrupt.h>
15#include <linux/bitops.h>
16#include <linux/slab.h>
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +040017#include <linux/proc_fs.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070018#include <linux/seq_file.h>
19#include <linux/cpu.h>
20#include <linux/cpuset.h>
21#include <linux/mempolicy.h>
22#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070023#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070025#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070026#include <linux/math64.h>
Akinobu Mita773ff602008-12-23 19:37:01 +090027#include <linux/fault-inject.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070028
29/*
30 * Lock order:
31 * 1. slab_lock(page)
32 * 2. slab->list_lock
33 *
34 * The slab_lock protects operations on the object of a particular
35 * slab and its metadata in the page struct. If the slab lock
36 * has been taken then no allocations nor frees can be performed
37 * on the objects in the slab nor can the slab be added or removed
38 * from the partial or full lists since this would mean modifying
39 * the page_struct of the slab.
40 *
41 * The list_lock protects the partial and full list on each node and
42 * the partial slab counter. If taken then no new slabs may be added or
43 * removed from the lists nor make the number of partial slabs be modified.
44 * (Note that the total number of slabs is an atomic value that may be
45 * modified without taking the list lock).
46 *
47 * The list_lock is a centralized lock and thus we avoid taking it as
48 * much as possible. As long as SLUB does not have to handle partial
49 * slabs, operations can continue without any centralized lock. F.e.
50 * allocating a long series of objects that fill up slabs does not require
51 * the list lock.
52 *
53 * The lock order is sometimes inverted when we are trying to get a slab
54 * off a list. We take the list_lock and then look for a page on the list
55 * to use. While we do that objects in the slabs may be freed. We can
56 * only operate on the slab if we have also taken the slab_lock. So we use
57 * a slab_trylock() on the slab. If trylock was successful then no frees
58 * can occur anymore and we can use the slab for allocations etc. If the
59 * slab_trylock() does not succeed then frees are in progress in the slab and
60 * we must stay away from it for a while since we may cause a bouncing
61 * cacheline if we try to acquire the lock. So go onto the next slab.
62 * If all pages are busy then we may allocate a new slab instead of reusing
63 * a partial slab. A new slab has noone operating on it and thus there is
64 * no danger of cacheline contention.
65 *
66 * Interrupts are disabled during allocation and deallocation in order to
67 * make the slab allocator safe to use in the context of an irq. In addition
68 * interrupts are disabled to ensure that the processor does not change
69 * while handling per_cpu slabs, due to kernel preemption.
70 *
71 * SLUB assigns one slab for allocation to each processor.
72 * Allocations only occur from these slabs called cpu slabs.
73 *
Christoph Lameter672bba32007-05-09 02:32:39 -070074 * Slabs with free elements are kept on a partial list and during regular
75 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070076 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070077 * We track full slabs for debugging purposes though because otherwise we
78 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070079 *
80 * Slabs are freed when they become empty. Teardown and setup is
81 * minimal so we rely on the page allocators per cpu caches for
82 * fast frees and allocs.
83 *
84 * Overloading of page flags that are otherwise used for LRU management.
85 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070086 * PageActive The slab is frozen and exempt from list processing.
87 * This means that the slab is dedicated to a purpose
88 * such as satisfying allocations for a specific
89 * processor. Objects may be freed in the slab while
90 * it is frozen but slab_free will then skip the usual
91 * list operations. It is up to the processor holding
92 * the slab to integrate the slab into the slab lists
93 * when the slab is no longer needed.
94 *
95 * One use of this flag is to mark slabs that are
96 * used for allocations. Then such a slab becomes a cpu
97 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -070098 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -070099 * free objects in addition to the regular freelist
100 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -0700101 *
102 * PageError Slab requires special handling due to debug
103 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700104 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700105 */
106
Christoph Lameter5577bd82007-05-16 22:10:56 -0700107#ifdef CONFIG_SLUB_DEBUG
Andy Whitcroft8a380822008-07-23 21:27:18 -0700108#define SLABDEBUG 1
Christoph Lameter5577bd82007-05-16 22:10:56 -0700109#else
110#define SLABDEBUG 0
111#endif
112
Christoph Lameter81819f02007-05-06 14:49:36 -0700113/*
114 * Issues still to be resolved:
115 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700116 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
117 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700118 * - Variable sizing of the per node arrays
119 */
120
121/* Enable to test recovery from slab corruption on boot */
122#undef SLUB_RESILIENCY_TEST
123
Christoph Lameter81819f02007-05-06 14:49:36 -0700124/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700125 * Mininum number of partial slabs. These will be left on the partial
126 * lists even if they are empty. kmem_cache_shrink may reclaim them.
127 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800128#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700129
Christoph Lameter2086d262007-05-06 14:49:46 -0700130/*
131 * Maximum number of desirable partial slabs.
132 * The existence of more partial slabs makes kmem_cache_shrink
133 * sort the partial list by the number of objects in the.
134 */
135#define MAX_PARTIAL 10
136
Christoph Lameter81819f02007-05-06 14:49:36 -0700137#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
138 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700139
Christoph Lameter81819f02007-05-06 14:49:36 -0700140/*
141 * Set of flags that will prevent slab merging
142 */
143#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
144 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
145
146#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
147 SLAB_CACHE_DMA)
148
149#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700150#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700151#endif
152
153#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700154#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700155#endif
156
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400157#define OO_SHIFT 16
158#define OO_MASK ((1 << OO_SHIFT) - 1)
159#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
160
Christoph Lameter81819f02007-05-06 14:49:36 -0700161/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700162#define __OBJECT_POISON 0x80000000 /* Poison object */
163#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700164
165static int kmem_size = sizeof(struct kmem_cache);
166
167#ifdef CONFIG_SMP
168static struct notifier_block slab_notifier;
169#endif
170
171static enum {
172 DOWN, /* No slab functionality available */
173 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700174 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700175 SYSFS /* Sysfs up */
176} slab_state = DOWN;
177
178/* A list of all slab caches on the system */
179static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700180static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700181
Christoph Lameter02cbc872007-05-09 02:32:43 -0700182/*
183 * Tracking user of a slab.
184 */
185struct track {
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300186 unsigned long addr; /* Called from address */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700187 int cpu; /* Was running on cpu */
188 int pid; /* Pid context */
189 unsigned long when; /* When did the operation occur */
190};
191
192enum track_item { TRACK_ALLOC, TRACK_FREE };
193
Christoph Lameterf6acb632008-04-29 16:16:06 -0700194#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -0700195static int sysfs_slab_add(struct kmem_cache *);
196static int sysfs_slab_alias(struct kmem_cache *, const char *);
197static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800198
Christoph Lameter81819f02007-05-06 14:49:36 -0700199#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700200static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
201static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
202 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800203static inline void sysfs_slab_remove(struct kmem_cache *s)
204{
205 kfree(s);
206}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800207
Christoph Lameter81819f02007-05-06 14:49:36 -0700208#endif
209
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800210static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
211{
212#ifdef CONFIG_SLUB_STATS
213 c->stat[si]++;
214#endif
215}
216
Christoph Lameter81819f02007-05-06 14:49:36 -0700217/********************************************************************
218 * Core slab cache functions
219 *******************************************************************/
220
221int slab_is_available(void)
222{
223 return slab_state >= UP;
224}
225
226static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
227{
228#ifdef CONFIG_NUMA
229 return s->node[node];
230#else
231 return &s->local_node;
232#endif
233}
234
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700235static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
236{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700237#ifdef CONFIG_SMP
238 return s->cpu_slab[cpu];
239#else
240 return &s->cpu_slab;
241#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700242}
243
Christoph Lameter6446faa2008-02-15 23:45:26 -0800244/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700245static inline int check_valid_pointer(struct kmem_cache *s,
246 struct page *page, const void *object)
247{
248 void *base;
249
Christoph Lametera973e9d2008-03-01 13:40:44 -0800250 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700251 return 1;
252
Christoph Lametera973e9d2008-03-01 13:40:44 -0800253 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300254 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700255 (object - base) % s->size) {
256 return 0;
257 }
258
259 return 1;
260}
261
Christoph Lameter81819f02007-05-06 14:49:36 -0700262/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700263 * Slow version of get and set free pointer.
264 *
265 * This version requires touching the cache lines of kmem_cache which
266 * we avoid to do in the fast alloc free paths. There we obtain the offset
267 * from the page struct.
268 */
269static inline void *get_freepointer(struct kmem_cache *s, void *object)
270{
271 return *(void **)(object + s->offset);
272}
273
274static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
275{
276 *(void **)(object + s->offset) = fp;
277}
278
279/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300280#define for_each_object(__p, __s, __addr, __objects) \
281 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700282 __p += (__s)->size)
283
284/* Scan freelist */
285#define for_each_free_object(__p, __s, __free) \
Christoph Lametera973e9d2008-03-01 13:40:44 -0800286 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700287
288/* Determine object index from a given position */
289static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
290{
291 return (p - addr) / s->size;
292}
293
Christoph Lameter834f3d12008-04-14 19:11:31 +0300294static inline struct kmem_cache_order_objects oo_make(int order,
295 unsigned long size)
296{
297 struct kmem_cache_order_objects x = {
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400298 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
Christoph Lameter834f3d12008-04-14 19:11:31 +0300299 };
300
301 return x;
302}
303
304static inline int oo_order(struct kmem_cache_order_objects x)
305{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400306 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300307}
308
309static inline int oo_objects(struct kmem_cache_order_objects x)
310{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400311 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300312}
313
Christoph Lameter41ecc552007-05-09 02:32:44 -0700314#ifdef CONFIG_SLUB_DEBUG
315/*
316 * Debug settings:
317 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700318#ifdef CONFIG_SLUB_DEBUG_ON
319static int slub_debug = DEBUG_DEFAULT_FLAGS;
320#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700321static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700322#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700323
324static char *slub_debug_slabs;
325
Christoph Lameter7656c722007-05-09 02:32:40 -0700326/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700327 * Object debugging
328 */
329static void print_section(char *text, u8 *addr, unsigned int length)
330{
331 int i, offset;
332 int newline = 1;
333 char ascii[17];
334
335 ascii[16] = 0;
336
337 for (i = 0; i < length; i++) {
338 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700339 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700340 newline = 0;
341 }
Pekka Enberg06428782008-01-07 23:20:27 -0800342 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700343 offset = i % 16;
344 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
345 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800346 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700347 newline = 1;
348 }
349 }
350 if (!newline) {
351 i %= 16;
352 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800353 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700354 ascii[i] = ' ';
355 i++;
356 }
Pekka Enberg06428782008-01-07 23:20:27 -0800357 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700358 }
359}
360
Christoph Lameter81819f02007-05-06 14:49:36 -0700361static struct track *get_track(struct kmem_cache *s, void *object,
362 enum track_item alloc)
363{
364 struct track *p;
365
366 if (s->offset)
367 p = object + s->offset + sizeof(void *);
368 else
369 p = object + s->inuse;
370
371 return p + alloc;
372}
373
374static void set_track(struct kmem_cache *s, void *object,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300375 enum track_item alloc, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700376{
377 struct track *p;
378
379 if (s->offset)
380 p = object + s->offset + sizeof(void *);
381 else
382 p = object + s->inuse;
383
384 p += alloc;
385 if (addr) {
386 p->addr = addr;
387 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400388 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700389 p->when = jiffies;
390 } else
391 memset(p, 0, sizeof(struct track));
392}
393
Christoph Lameter81819f02007-05-06 14:49:36 -0700394static void init_tracking(struct kmem_cache *s, void *object)
395{
Christoph Lameter24922682007-07-17 04:03:18 -0700396 if (!(s->flags & SLAB_STORE_USER))
397 return;
398
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300399 set_track(s, object, TRACK_FREE, 0UL);
400 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700401}
402
403static void print_track(const char *s, struct track *t)
404{
405 if (!t->addr)
406 return;
407
Linus Torvalds7daf7052008-07-14 12:12:53 -0700408 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300409 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700410}
411
Christoph Lameter24922682007-07-17 04:03:18 -0700412static void print_tracking(struct kmem_cache *s, void *object)
413{
414 if (!(s->flags & SLAB_STORE_USER))
415 return;
416
417 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
418 print_track("Freed", get_track(s, object, TRACK_FREE));
419}
420
421static void print_page_info(struct page *page)
422{
Christoph Lameter39b26462008-04-14 19:11:30 +0300423 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
424 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700425
426}
427
428static void slab_bug(struct kmem_cache *s, char *fmt, ...)
429{
430 va_list args;
431 char buf[100];
432
433 va_start(args, fmt);
434 vsnprintf(buf, sizeof(buf), fmt, args);
435 va_end(args);
436 printk(KERN_ERR "========================================"
437 "=====================================\n");
438 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
439 printk(KERN_ERR "----------------------------------------"
440 "-------------------------------------\n\n");
441}
442
443static void slab_fix(struct kmem_cache *s, char *fmt, ...)
444{
445 va_list args;
446 char buf[100];
447
448 va_start(args, fmt);
449 vsnprintf(buf, sizeof(buf), fmt, args);
450 va_end(args);
451 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
452}
453
454static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700455{
456 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800457 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700458
459 print_tracking(s, p);
460
461 print_page_info(page);
462
463 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
464 p, p - addr, get_freepointer(s, p));
465
466 if (p > addr + 16)
467 print_section("Bytes b4", p - 16, 16);
468
Pekka Enberg0ebd6522008-07-19 14:17:22 +0300469 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700470
471 if (s->flags & SLAB_RED_ZONE)
472 print_section("Redzone", p + s->objsize,
473 s->inuse - s->objsize);
474
Christoph Lameter81819f02007-05-06 14:49:36 -0700475 if (s->offset)
476 off = s->offset + sizeof(void *);
477 else
478 off = s->inuse;
479
Christoph Lameter24922682007-07-17 04:03:18 -0700480 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700481 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700482
483 if (off != s->size)
484 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700485 print_section("Padding", p + off, s->size - off);
486
487 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700488}
489
490static void object_err(struct kmem_cache *s, struct page *page,
491 u8 *object, char *reason)
492{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700493 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700494 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700495}
496
Christoph Lameter24922682007-07-17 04:03:18 -0700497static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700498{
499 va_list args;
500 char buf[100];
501
Christoph Lameter24922682007-07-17 04:03:18 -0700502 va_start(args, fmt);
503 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700504 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700505 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700506 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700507 dump_stack();
508}
509
510static void init_object(struct kmem_cache *s, void *object, int active)
511{
512 u8 *p = object;
513
514 if (s->flags & __OBJECT_POISON) {
515 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800516 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700517 }
518
519 if (s->flags & SLAB_RED_ZONE)
520 memset(p + s->objsize,
521 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
522 s->inuse - s->objsize);
523}
524
Christoph Lameter24922682007-07-17 04:03:18 -0700525static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700526{
527 while (bytes) {
528 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700529 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700530 start++;
531 bytes--;
532 }
Christoph Lameter24922682007-07-17 04:03:18 -0700533 return NULL;
534}
535
536static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
537 void *from, void *to)
538{
539 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
540 memset(from, data, to - from);
541}
542
543static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
544 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800545 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700546{
547 u8 *fault;
548 u8 *end;
549
550 fault = check_bytes(start, value, bytes);
551 if (!fault)
552 return 1;
553
554 end = start + bytes;
555 while (end > fault && end[-1] == value)
556 end--;
557
558 slab_bug(s, "%s overwritten", what);
559 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
560 fault, end - 1, fault[0], value);
561 print_trailer(s, page, object);
562
563 restore_bytes(s, what, value, fault, end);
564 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700565}
566
Christoph Lameter81819f02007-05-06 14:49:36 -0700567/*
568 * Object layout:
569 *
570 * object address
571 * Bytes of the object to be managed.
572 * If the freepointer may overlay the object then the free
573 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700574 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700575 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
576 * 0xa5 (POISON_END)
577 *
578 * object + s->objsize
579 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700580 * Padding is extended by another word if Redzoning is enabled and
581 * objsize == inuse.
582 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700583 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
584 * 0xcc (RED_ACTIVE) for objects in use.
585 *
586 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700587 * Meta data starts here.
588 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700589 * A. Free pointer (if we cannot overwrite object on free)
590 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700591 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800592 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700593 * before the word boundary.
594 *
595 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700596 *
597 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700598 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700599 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700600 * If slabcaches are merged then the objsize and inuse boundaries are mostly
601 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700602 * may be used with merged slabcaches.
603 */
604
Christoph Lameter81819f02007-05-06 14:49:36 -0700605static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
606{
607 unsigned long off = s->inuse; /* The end of info */
608
609 if (s->offset)
610 /* Freepointer is placed after the object. */
611 off += sizeof(void *);
612
613 if (s->flags & SLAB_STORE_USER)
614 /* We also have user information there */
615 off += 2 * sizeof(struct track);
616
617 if (s->size == off)
618 return 1;
619
Christoph Lameter24922682007-07-17 04:03:18 -0700620 return check_bytes_and_report(s, page, p, "Object padding",
621 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700622}
623
Christoph Lameter39b26462008-04-14 19:11:30 +0300624/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700625static int slab_pad_check(struct kmem_cache *s, struct page *page)
626{
Christoph Lameter24922682007-07-17 04:03:18 -0700627 u8 *start;
628 u8 *fault;
629 u8 *end;
630 int length;
631 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700632
633 if (!(s->flags & SLAB_POISON))
634 return 1;
635
Christoph Lametera973e9d2008-03-01 13:40:44 -0800636 start = page_address(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +0300637 length = (PAGE_SIZE << compound_order(page));
Christoph Lameter39b26462008-04-14 19:11:30 +0300638 end = start + length;
639 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700640 if (!remainder)
641 return 1;
642
Christoph Lameter39b26462008-04-14 19:11:30 +0300643 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700644 if (!fault)
645 return 1;
646 while (end > fault && end[-1] == POISON_INUSE)
647 end--;
648
649 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Christoph Lameter39b26462008-04-14 19:11:30 +0300650 print_section("Padding", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700651
652 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
653 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700654}
655
656static int check_object(struct kmem_cache *s, struct page *page,
657 void *object, int active)
658{
659 u8 *p = object;
660 u8 *endobject = object + s->objsize;
661
662 if (s->flags & SLAB_RED_ZONE) {
663 unsigned int red =
664 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
665
Christoph Lameter24922682007-07-17 04:03:18 -0700666 if (!check_bytes_and_report(s, page, object, "Redzone",
667 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700668 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700669 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800670 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
671 check_bytes_and_report(s, page, p, "Alignment padding",
672 endobject, POISON_INUSE, s->inuse - s->objsize);
673 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700674 }
675
676 if (s->flags & SLAB_POISON) {
677 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700678 (!check_bytes_and_report(s, page, p, "Poison", p,
679 POISON_FREE, s->objsize - 1) ||
680 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800681 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700682 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700683 /*
684 * check_pad_bytes cleans up on its own.
685 */
686 check_pad_bytes(s, page, p);
687 }
688
689 if (!s->offset && active)
690 /*
691 * Object and freepointer overlap. Cannot check
692 * freepointer while object is allocated.
693 */
694 return 1;
695
696 /* Check free pointer validity */
697 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
698 object_err(s, page, p, "Freepointer corrupt");
699 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100700 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700701 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700702 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700703 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800704 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700705 return 0;
706 }
707 return 1;
708}
709
710static int check_slab(struct kmem_cache *s, struct page *page)
711{
Christoph Lameter39b26462008-04-14 19:11:30 +0300712 int maxobj;
713
Christoph Lameter81819f02007-05-06 14:49:36 -0700714 VM_BUG_ON(!irqs_disabled());
715
716 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700717 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700718 return 0;
719 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300720
721 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
722 if (page->objects > maxobj) {
723 slab_err(s, page, "objects %u > max %u",
724 s->name, page->objects, maxobj);
725 return 0;
726 }
727 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700728 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300729 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700730 return 0;
731 }
732 /* Slab_pad_check fixes things up after itself */
733 slab_pad_check(s, page);
734 return 1;
735}
736
737/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700738 * Determine if a certain object on a page is on the freelist. Must hold the
739 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700740 */
741static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
742{
743 int nr = 0;
744 void *fp = page->freelist;
745 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300746 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700747
Christoph Lameter39b26462008-04-14 19:11:30 +0300748 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700749 if (fp == search)
750 return 1;
751 if (!check_valid_pointer(s, page, fp)) {
752 if (object) {
753 object_err(s, page, object,
754 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800755 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700756 break;
757 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700758 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800759 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300760 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700761 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700762 return 0;
763 }
764 break;
765 }
766 object = fp;
767 fp = get_freepointer(s, object);
768 nr++;
769 }
770
Christoph Lameter224a88b2008-04-14 19:11:31 +0300771 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400772 if (max_objects > MAX_OBJS_PER_PAGE)
773 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300774
775 if (page->objects != max_objects) {
776 slab_err(s, page, "Wrong number of objects. Found %d but "
777 "should be %d", page->objects, max_objects);
778 page->objects = max_objects;
779 slab_fix(s, "Number of objects adjusted.");
780 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300781 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700782 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300783 "counted were %d", page->inuse, page->objects - nr);
784 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700785 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700786 }
787 return search == NULL;
788}
789
Christoph Lameter0121c6192008-04-29 16:11:12 -0700790static void trace(struct kmem_cache *s, struct page *page, void *object,
791 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700792{
793 if (s->flags & SLAB_TRACE) {
794 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
795 s->name,
796 alloc ? "alloc" : "free",
797 object, page->inuse,
798 page->freelist);
799
800 if (!alloc)
801 print_section("Object", (void *)object, s->objsize);
802
803 dump_stack();
804 }
805}
806
Christoph Lameter643b1132007-05-06 14:49:42 -0700807/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700808 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700809 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700810static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700811{
Christoph Lameter643b1132007-05-06 14:49:42 -0700812 spin_lock(&n->list_lock);
813 list_add(&page->lru, &n->full);
814 spin_unlock(&n->list_lock);
815}
816
817static void remove_full(struct kmem_cache *s, struct page *page)
818{
819 struct kmem_cache_node *n;
820
821 if (!(s->flags & SLAB_STORE_USER))
822 return;
823
824 n = get_node(s, page_to_nid(page));
825
826 spin_lock(&n->list_lock);
827 list_del(&page->lru);
828 spin_unlock(&n->list_lock);
829}
830
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300831/* Tracking of the number of slabs for debugging purposes */
832static inline unsigned long slabs_node(struct kmem_cache *s, int node)
833{
834 struct kmem_cache_node *n = get_node(s, node);
835
836 return atomic_long_read(&n->nr_slabs);
837}
838
Christoph Lameter205ab992008-04-14 19:11:40 +0300839static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300840{
841 struct kmem_cache_node *n = get_node(s, node);
842
843 /*
844 * May be called early in order to allocate a slab for the
845 * kmem_cache_node structure. Solve the chicken-egg
846 * dilemma by deferring the increment of the count during
847 * bootstrap (see early_kmem_cache_node_alloc).
848 */
Christoph Lameter205ab992008-04-14 19:11:40 +0300849 if (!NUMA_BUILD || n) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300850 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300851 atomic_long_add(objects, &n->total_objects);
852 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300853}
Christoph Lameter205ab992008-04-14 19:11:40 +0300854static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300855{
856 struct kmem_cache_node *n = get_node(s, node);
857
858 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300859 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300860}
861
862/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -0700863static void setup_object_debug(struct kmem_cache *s, struct page *page,
864 void *object)
865{
866 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
867 return;
868
869 init_object(s, object, 0);
870 init_tracking(s, object);
871}
872
873static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300874 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700875{
876 if (!check_slab(s, page))
877 goto bad;
878
Christoph Lameterd692ef62008-02-15 23:45:24 -0800879 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700880 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700881 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700882 }
883
884 if (!check_valid_pointer(s, page, object)) {
885 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700886 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700887 }
888
Christoph Lameterd692ef62008-02-15 23:45:24 -0800889 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700890 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700891
Christoph Lameter3ec09742007-05-16 22:11:00 -0700892 /* Success perform special debug activities for allocs */
893 if (s->flags & SLAB_STORE_USER)
894 set_track(s, object, TRACK_ALLOC, addr);
895 trace(s, page, object, 1);
896 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700897 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700898
Christoph Lameter81819f02007-05-06 14:49:36 -0700899bad:
900 if (PageSlab(page)) {
901 /*
902 * If this is a slab page then lets do the best we can
903 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700904 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700905 */
Christoph Lameter24922682007-07-17 04:03:18 -0700906 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +0300907 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800908 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700909 }
910 return 0;
911}
912
Christoph Lameter3ec09742007-05-16 22:11:00 -0700913static int free_debug_processing(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300914 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700915{
916 if (!check_slab(s, page))
917 goto fail;
918
919 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700920 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700921 goto fail;
922 }
923
924 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700925 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700926 goto fail;
927 }
928
929 if (!check_object(s, page, object, 1))
930 return 0;
931
932 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800933 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700934 slab_err(s, page, "Attempt to free object(0x%p) "
935 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800936 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700937 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700938 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700939 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700940 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800941 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700942 object_err(s, page, object,
943 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700944 goto fail;
945 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700946
947 /* Special debug activities for freeing objects */
Andy Whitcroft8a380822008-07-23 21:27:18 -0700948 if (!PageSlubFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700949 remove_full(s, page);
950 if (s->flags & SLAB_STORE_USER)
951 set_track(s, object, TRACK_FREE, addr);
952 trace(s, page, object, 0);
953 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700954 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700955
Christoph Lameter81819f02007-05-06 14:49:36 -0700956fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700957 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700958 return 0;
959}
960
Christoph Lameter41ecc552007-05-09 02:32:44 -0700961static int __init setup_slub_debug(char *str)
962{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700963 slub_debug = DEBUG_DEFAULT_FLAGS;
964 if (*str++ != '=' || !*str)
965 /*
966 * No options specified. Switch on full debugging.
967 */
968 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700969
970 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700971 /*
972 * No options but restriction on slabs. This means full
973 * debugging for slabs matching a pattern.
974 */
975 goto check_slabs;
976
977 slub_debug = 0;
978 if (*str == '-')
979 /*
980 * Switch off all debugging measures.
981 */
982 goto out;
983
984 /*
985 * Determine which debug features should be switched on
986 */
Pekka Enberg06428782008-01-07 23:20:27 -0800987 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700988 switch (tolower(*str)) {
989 case 'f':
990 slub_debug |= SLAB_DEBUG_FREE;
991 break;
992 case 'z':
993 slub_debug |= SLAB_RED_ZONE;
994 break;
995 case 'p':
996 slub_debug |= SLAB_POISON;
997 break;
998 case 'u':
999 slub_debug |= SLAB_STORE_USER;
1000 break;
1001 case 't':
1002 slub_debug |= SLAB_TRACE;
1003 break;
1004 default:
1005 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001006 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001007 }
1008 }
1009
1010check_slabs:
1011 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001012 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001013out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001014 return 1;
1015}
1016
1017__setup("slub_debug", setup_slub_debug);
1018
Christoph Lameterba0268a2007-09-11 15:24:11 -07001019static unsigned long kmem_cache_flags(unsigned long objsize,
1020 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001021 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001022{
1023 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001024 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001025 */
Christoph Lametere1533622008-02-15 23:45:24 -08001026 if (slub_debug && (!slub_debug_slabs ||
1027 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1028 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001029
1030 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001031}
1032#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001033static inline void setup_object_debug(struct kmem_cache *s,
1034 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001035
Christoph Lameter3ec09742007-05-16 22:11:00 -07001036static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001037 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001038
Christoph Lameter3ec09742007-05-16 22:11:00 -07001039static inline int free_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001040 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001041
Christoph Lameter41ecc552007-05-09 02:32:44 -07001042static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1043 { return 1; }
1044static inline int check_object(struct kmem_cache *s, struct page *page,
1045 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001046static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001047static inline unsigned long kmem_cache_flags(unsigned long objsize,
1048 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001049 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001050{
1051 return flags;
1052}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001053#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001054
1055static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1056 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001057static inline void inc_slabs_node(struct kmem_cache *s, int node,
1058 int objects) {}
1059static inline void dec_slabs_node(struct kmem_cache *s, int node,
1060 int objects) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001061#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03001062
Christoph Lameter81819f02007-05-06 14:49:36 -07001063/*
1064 * Slab allocation and freeing
1065 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001066static inline struct page *alloc_slab_page(gfp_t flags, int node,
1067 struct kmem_cache_order_objects oo)
1068{
1069 int order = oo_order(oo);
1070
1071 if (node == -1)
1072 return alloc_pages(flags, order);
1073 else
1074 return alloc_pages_node(node, flags, order);
1075}
1076
Christoph Lameter81819f02007-05-06 14:49:36 -07001077static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1078{
Pekka Enberg06428782008-01-07 23:20:27 -08001079 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001080 struct kmem_cache_order_objects oo = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07001081
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001082 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001083
Christoph Lameter65c33762008-04-14 19:11:40 +03001084 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1085 oo);
1086 if (unlikely(!page)) {
1087 oo = s->min;
1088 /*
1089 * Allocation may have failed due to fragmentation.
1090 * Try a lower order alloc if possible
1091 */
1092 page = alloc_slab_page(flags, node, oo);
1093 if (!page)
1094 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001095
Christoph Lameter65c33762008-04-14 19:11:40 +03001096 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1097 }
Christoph Lameter834f3d12008-04-14 19:11:31 +03001098 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001099 mod_zone_page_state(page_zone(page),
1100 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1101 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001102 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001103
1104 return page;
1105}
1106
1107static void setup_object(struct kmem_cache *s, struct page *page,
1108 void *object)
1109{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001110 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001111 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001112 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001113}
1114
1115static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1116{
1117 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001118 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001119 void *last;
1120 void *p;
1121
Christoph Lameter6cb06222007-10-16 01:25:41 -07001122 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001123
Christoph Lameter6cb06222007-10-16 01:25:41 -07001124 page = allocate_slab(s,
1125 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001126 if (!page)
1127 goto out;
1128
Christoph Lameter205ab992008-04-14 19:11:40 +03001129 inc_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001130 page->slab = s;
1131 page->flags |= 1 << PG_slab;
1132 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1133 SLAB_STORE_USER | SLAB_TRACE))
Andy Whitcroft8a380822008-07-23 21:27:18 -07001134 __SetPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001135
1136 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001137
1138 if (unlikely(s->flags & SLAB_POISON))
Christoph Lameter834f3d12008-04-14 19:11:31 +03001139 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001140
1141 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001142 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001143 setup_object(s, page, last);
1144 set_freepointer(s, last, p);
1145 last = p;
1146 }
1147 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001148 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001149
1150 page->freelist = start;
1151 page->inuse = 0;
1152out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001153 return page;
1154}
1155
1156static void __free_slab(struct kmem_cache *s, struct page *page)
1157{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001158 int order = compound_order(page);
1159 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001160
Andy Whitcroft8a380822008-07-23 21:27:18 -07001161 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001162 void *p;
1163
1164 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001165 for_each_object(p, s, page_address(page),
1166 page->objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001167 check_object(s, page, p, 0);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001168 __ClearPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001169 }
1170
1171 mod_zone_page_state(page_zone(page),
1172 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1173 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001174 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001175
Christoph Lameter49bd5222008-04-14 18:52:18 +03001176 __ClearPageSlab(page);
1177 reset_page_mapcount(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +03001178 __free_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001179}
1180
1181static void rcu_free_slab(struct rcu_head *h)
1182{
1183 struct page *page;
1184
1185 page = container_of((struct list_head *)h, struct page, lru);
1186 __free_slab(page->slab, page);
1187}
1188
1189static void free_slab(struct kmem_cache *s, struct page *page)
1190{
1191 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1192 /*
1193 * RCU free overloads the RCU head over the LRU
1194 */
1195 struct rcu_head *head = (void *)&page->lru;
1196
1197 call_rcu(head, rcu_free_slab);
1198 } else
1199 __free_slab(s, page);
1200}
1201
1202static void discard_slab(struct kmem_cache *s, struct page *page)
1203{
Christoph Lameter205ab992008-04-14 19:11:40 +03001204 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001205 free_slab(s, page);
1206}
1207
1208/*
1209 * Per slab locking using the pagelock
1210 */
1211static __always_inline void slab_lock(struct page *page)
1212{
1213 bit_spin_lock(PG_locked, &page->flags);
1214}
1215
1216static __always_inline void slab_unlock(struct page *page)
1217{
Nick Piggina76d3542008-01-07 23:20:27 -08001218 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001219}
1220
1221static __always_inline int slab_trylock(struct page *page)
1222{
1223 int rc = 1;
1224
1225 rc = bit_spin_trylock(PG_locked, &page->flags);
1226 return rc;
1227}
1228
1229/*
1230 * Management of partially allocated slabs
1231 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001232static void add_partial(struct kmem_cache_node *n,
1233 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001234{
Christoph Lametere95eed52007-05-06 14:49:44 -07001235 spin_lock(&n->list_lock);
1236 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001237 if (tail)
1238 list_add_tail(&page->lru, &n->partial);
1239 else
1240 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001241 spin_unlock(&n->list_lock);
1242}
1243
Christoph Lameter0121c6192008-04-29 16:11:12 -07001244static void remove_partial(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001245{
1246 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1247
1248 spin_lock(&n->list_lock);
1249 list_del(&page->lru);
1250 n->nr_partial--;
1251 spin_unlock(&n->list_lock);
1252}
1253
1254/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001255 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001256 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001257 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001258 */
Christoph Lameter0121c6192008-04-29 16:11:12 -07001259static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1260 struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001261{
1262 if (slab_trylock(page)) {
1263 list_del(&page->lru);
1264 n->nr_partial--;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001265 __SetPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001266 return 1;
1267 }
1268 return 0;
1269}
1270
1271/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001272 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001273 */
1274static struct page *get_partial_node(struct kmem_cache_node *n)
1275{
1276 struct page *page;
1277
1278 /*
1279 * Racy check. If we mistakenly see no partial slabs then we
1280 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001281 * partial slab and there is none available then get_partials()
1282 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001283 */
1284 if (!n || !n->nr_partial)
1285 return NULL;
1286
1287 spin_lock(&n->list_lock);
1288 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001289 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001290 goto out;
1291 page = NULL;
1292out:
1293 spin_unlock(&n->list_lock);
1294 return page;
1295}
1296
1297/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001298 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001299 */
1300static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1301{
1302#ifdef CONFIG_NUMA
1303 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001304 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001305 struct zone *zone;
1306 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001307 struct page *page;
1308
1309 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001310 * The defrag ratio allows a configuration of the tradeoffs between
1311 * inter node defragmentation and node local allocations. A lower
1312 * defrag_ratio increases the tendency to do local allocations
1313 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001314 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001315 * If the defrag_ratio is set to 0 then kmalloc() always
1316 * returns node local objects. If the ratio is higher then kmalloc()
1317 * may return off node objects because partial slabs are obtained
1318 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001319 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001320 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001321 * defrag_ratio = 1000) then every (well almost) allocation will
1322 * first attempt to defrag slab caches on other nodes. This means
1323 * scanning over all nodes to look for partial slabs which may be
1324 * expensive if we do it every time we are trying to find a slab
1325 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001326 */
Christoph Lameter98246012008-01-07 23:20:26 -08001327 if (!s->remote_node_defrag_ratio ||
1328 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001329 return NULL;
1330
Mel Gorman0e884602008-04-28 02:12:14 -07001331 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001332 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001333 struct kmem_cache_node *n;
1334
Mel Gorman54a6eb52008-04-28 02:12:16 -07001335 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001336
Mel Gorman54a6eb52008-04-28 02:12:16 -07001337 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
Pekka Enberg5595cff2008-08-05 09:28:47 +03001338 n->nr_partial > n->min_partial) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001339 page = get_partial_node(n);
1340 if (page)
1341 return page;
1342 }
1343 }
1344#endif
1345 return NULL;
1346}
1347
1348/*
1349 * Get a partial page, lock it and return it.
1350 */
1351static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1352{
1353 struct page *page;
1354 int searchnode = (node == -1) ? numa_node_id() : node;
1355
1356 page = get_partial_node(get_node(s, searchnode));
1357 if (page || (flags & __GFP_THISNODE))
1358 return page;
1359
1360 return get_any_partial(s, flags);
1361}
1362
1363/*
1364 * Move a page back to the lists.
1365 *
1366 * Must be called with the slab lock held.
1367 *
1368 * On exit the slab lock will have been dropped.
1369 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001370static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001371{
Christoph Lametere95eed52007-05-06 14:49:44 -07001372 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001373 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001374
Andy Whitcroft8a380822008-07-23 21:27:18 -07001375 __ClearPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001376 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001377
Christoph Lametera973e9d2008-03-01 13:40:44 -08001378 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001379 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001380 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1381 } else {
1382 stat(c, DEACTIVATE_FULL);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001383 if (SLABDEBUG && PageSlubDebug(page) &&
1384 (s->flags & SLAB_STORE_USER))
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001385 add_full(n, page);
1386 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001387 slab_unlock(page);
1388 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001389 stat(c, DEACTIVATE_EMPTY);
Pekka Enberg5595cff2008-08-05 09:28:47 +03001390 if (n->nr_partial < n->min_partial) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001391 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001392 * Adding an empty slab to the partial slabs in order
1393 * to avoid page allocator overhead. This slab needs
1394 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001395 * so that the others get filled first. That way the
1396 * size of the partial list stays small.
1397 *
Christoph Lameter0121c6192008-04-29 16:11:12 -07001398 * kmem_cache_shrink can reclaim any empty slabs from
1399 * the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001400 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001401 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001402 slab_unlock(page);
1403 } else {
1404 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001405 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001406 discard_slab(s, page);
1407 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001408 }
1409}
1410
1411/*
1412 * Remove the cpu slab
1413 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001414static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001415{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001416 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001417 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001418
Christoph Lameterb773ad72008-03-04 11:10:17 -08001419 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001420 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001421 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001422 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001423 * because both freelists are empty. So this is unlikely
1424 * to occur.
1425 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001426 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001427 void **object;
1428
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001429 tail = 0; /* Hot objects. Put the slab first */
1430
Christoph Lameter894b8782007-05-10 03:15:16 -07001431 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001432 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001433 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001434
1435 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001436 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001437 page->freelist = object;
1438 page->inuse--;
1439 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001440 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001441 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001442}
1443
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001444static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001445{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001446 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001447 slab_lock(c->page);
1448 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001449}
1450
1451/*
1452 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001453 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001454 * Called from IPI handler with interrupts disabled.
1455 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001456static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001457{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001458 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001459
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001460 if (likely(c && c->page))
1461 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001462}
1463
1464static void flush_cpu_slab(void *d)
1465{
1466 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001467
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001468 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001469}
1470
1471static void flush_all(struct kmem_cache *s)
1472{
Jens Axboe15c8b6c2008-05-09 09:39:44 +02001473 on_each_cpu(flush_cpu_slab, s, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07001474}
1475
1476/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001477 * Check if the objects in a per cpu structure fit numa
1478 * locality expectations.
1479 */
1480static inline int node_match(struct kmem_cache_cpu *c, int node)
1481{
1482#ifdef CONFIG_NUMA
1483 if (node != -1 && c->node != node)
1484 return 0;
1485#endif
1486 return 1;
1487}
1488
1489/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001490 * Slow path. The lockless freelist is empty or we need to perform
1491 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001492 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001493 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001494 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001495 * Processing is still very fast if new objects have been freed to the
1496 * regular freelist. In that case we simply take over the regular freelist
1497 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001498 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001499 * If that is not working then we fall back to the partial lists. We take the
1500 * first element of the freelist as the object to allocate now and move the
1501 * rest of the freelist to the lockless freelist.
1502 *
1503 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001504 * we need to allocate a new slab. This is the slowest path since it involves
1505 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001506 */
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001507static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1508 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001509{
Christoph Lameter81819f02007-05-06 14:49:36 -07001510 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001511 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001512
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001513 /* We handle __GFP_ZERO in the caller */
1514 gfpflags &= ~__GFP_ZERO;
1515
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001516 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001517 goto new_slab;
1518
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001519 slab_lock(c->page);
1520 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001521 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001522
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001523 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001524
Christoph Lameter894b8782007-05-10 03:15:16 -07001525load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001526 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001527 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001528 goto another_slab;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001529 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001530 goto debug;
1531
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001532 c->freelist = object[c->offset];
Christoph Lameter39b26462008-04-14 19:11:30 +03001533 c->page->inuse = c->page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001534 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001535 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001536unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001537 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001538 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001539 return object;
1540
1541another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001542 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001543
1544new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001545 new = get_partial(s, gfpflags, node);
1546 if (new) {
1547 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001548 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001549 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001550 }
1551
Christoph Lameterb811c202007-10-16 23:25:51 -07001552 if (gfpflags & __GFP_WAIT)
1553 local_irq_enable();
1554
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001555 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001556
1557 if (gfpflags & __GFP_WAIT)
1558 local_irq_disable();
1559
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001560 if (new) {
1561 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001562 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001563 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001564 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001565 slab_lock(new);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001566 __SetPageSlubFrozen(new);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001567 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001568 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001569 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001570 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001571debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001572 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001573 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001574
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001575 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001576 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001577 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001578 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001579}
1580
1581/*
1582 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1583 * have the fastpath folded into their functions. So no function call
1584 * overhead for requests that can be satisfied on the fastpath.
1585 *
1586 * The fastpath works by first checking if the lockless freelist can be used.
1587 * If not then __slab_alloc is called for slow processing.
1588 *
1589 * Otherwise we can simply pick the next object from the lockless free list.
1590 */
Pekka Enberg06428782008-01-07 23:20:27 -08001591static __always_inline void *slab_alloc(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001592 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001593{
Christoph Lameter894b8782007-05-10 03:15:16 -07001594 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001595 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001596 unsigned long flags;
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001597 unsigned int objsize;
Christoph Lameter1f842602008-01-07 23:20:30 -08001598
OGAWA Hirofumi89124d72008-11-19 21:23:59 +09001599 might_sleep_if(gfpflags & __GFP_WAIT);
Pekka Enberg3c506ef2008-12-29 11:47:05 +02001600
Akinobu Mita773ff602008-12-23 19:37:01 +09001601 if (should_failslab(s->objsize, gfpflags))
1602 return NULL;
1603
Christoph Lameter894b8782007-05-10 03:15:16 -07001604 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001605 c = get_cpu_slab(s, smp_processor_id());
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001606 objsize = c->objsize;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001607 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001608
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001609 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001610
1611 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001612 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001613 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001614 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001615 }
1616 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001617
1618 if (unlikely((gfpflags & __GFP_ZERO) && object))
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001619 memset(object, 0, objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001620
Christoph Lameter894b8782007-05-10 03:15:16 -07001621 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001622}
1623
1624void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1625{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001626 return slab_alloc(s, gfpflags, -1, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07001627}
1628EXPORT_SYMBOL(kmem_cache_alloc);
1629
1630#ifdef CONFIG_NUMA
1631void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1632{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001633 return slab_alloc(s, gfpflags, node, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07001634}
1635EXPORT_SYMBOL(kmem_cache_alloc_node);
1636#endif
1637
1638/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001639 * Slow patch handling. This may still be called frequently since objects
1640 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001641 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001642 * So we still attempt to reduce cache line usage. Just take the slab
1643 * lock and free the item. If there is no additional partial page
1644 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001645 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001646static void __slab_free(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001647 void *x, unsigned long addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001648{
1649 void *prior;
1650 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001651 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001652
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001653 c = get_cpu_slab(s, raw_smp_processor_id());
1654 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001655 slab_lock(page);
1656
Andy Whitcroft8a380822008-07-23 21:27:18 -07001657 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001658 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001659
Christoph Lameter81819f02007-05-06 14:49:36 -07001660checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001661 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001662 page->freelist = object;
1663 page->inuse--;
1664
Andy Whitcroft8a380822008-07-23 21:27:18 -07001665 if (unlikely(PageSlubFrozen(page))) {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001666 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001667 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001668 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001669
1670 if (unlikely(!page->inuse))
1671 goto slab_empty;
1672
1673 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001674 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001675 * then add it.
1676 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001677 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001678 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001679 stat(c, FREE_ADD_PARTIAL);
1680 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001681
1682out_unlock:
1683 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001684 return;
1685
1686slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001687 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001688 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001689 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001690 */
1691 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001692 stat(c, FREE_REMOVE_PARTIAL);
1693 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001694 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001695 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001696 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001697 return;
1698
1699debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001700 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001701 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001702 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001703}
1704
Christoph Lameter894b8782007-05-10 03:15:16 -07001705/*
1706 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1707 * can perform fastpath freeing without additional function calls.
1708 *
1709 * The fastpath is only possible if we are freeing to the current cpu slab
1710 * of this processor. This typically the case if we have just allocated
1711 * the item before.
1712 *
1713 * If fastpath is not possible then fall back to __slab_free where we deal
1714 * with all sorts of special processing.
1715 */
Pekka Enberg06428782008-01-07 23:20:27 -08001716static __always_inline void slab_free(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001717 struct page *page, void *x, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001718{
1719 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001720 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001721 unsigned long flags;
1722
Christoph Lameter894b8782007-05-10 03:15:16 -07001723 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001724 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001725 debug_check_no_locks_freed(object, c->objsize);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001726 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1727 debug_check_no_obj_freed(object, s->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001728 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001729 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001730 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001731 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001732 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001733 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001734
1735 local_irq_restore(flags);
1736}
1737
Christoph Lameter81819f02007-05-06 14:49:36 -07001738void kmem_cache_free(struct kmem_cache *s, void *x)
1739{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001740 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001741
Christoph Lameterb49af682007-05-06 14:49:41 -07001742 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001743
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001744 slab_free(s, page, x, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07001745}
1746EXPORT_SYMBOL(kmem_cache_free);
1747
Cyrill Gorcunove9beef12008-10-28 22:02:26 +03001748/* Figure out on which slab page the object resides */
Christoph Lameter81819f02007-05-06 14:49:36 -07001749static struct page *get_object_page(const void *x)
1750{
Christoph Lameterb49af682007-05-06 14:49:41 -07001751 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001752
1753 if (!PageSlab(page))
1754 return NULL;
1755
1756 return page;
1757}
1758
1759/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001760 * Object placement in a slab is made very easy because we always start at
1761 * offset 0. If we tune the size of the object to the alignment then we can
1762 * get the required alignment by putting one properly sized object after
1763 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001764 *
1765 * Notice that the allocation order determines the sizes of the per cpu
1766 * caches. Each processor has always one slab available for allocations.
1767 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001768 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001769 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001770 */
1771
1772/*
1773 * Mininum / Maximum order of slab pages. This influences locking overhead
1774 * and slab fragmentation. A higher order reduces the number of partial slabs
1775 * and increases the number of allocations possible without having to
1776 * take the list_lock.
1777 */
1778static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03001779static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001780static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001781
1782/*
1783 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001784 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001785 */
1786static int slub_nomerge;
1787
1788/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001789 * Calculate the order of allocation given an slab object size.
1790 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001791 * The order of allocation has significant impact on performance and other
1792 * system components. Generally order 0 allocations should be preferred since
1793 * order 0 does not cause fragmentation in the page allocator. Larger objects
1794 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001795 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07001796 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001797 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001798 * In order to reach satisfactory performance we must ensure that a minimum
1799 * number of objects is in one slab. Otherwise we may generate too much
1800 * activity on the partial lists which requires taking the list_lock. This is
1801 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001802 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001803 * slub_max_order specifies the order where we begin to stop considering the
1804 * number of objects in a slab as critical. If we reach slub_max_order then
1805 * we try to keep the page order as low as possible. So we accept more waste
1806 * of space in favor of a small page order.
1807 *
1808 * Higher order allocations also allow the placement of more objects in a
1809 * slab and thereby reduce object handling overhead. If the user has
1810 * requested a higher mininum order then we start with that one instead of
1811 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001812 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001813static inline int slab_order(int size, int min_objects,
1814 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001815{
1816 int order;
1817 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001818 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001819
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04001820 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1821 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03001822
Christoph Lameter6300ea72007-07-17 04:03:20 -07001823 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001824 fls(min_objects * size - 1) - PAGE_SHIFT);
1825 order <= max_order; order++) {
1826
Christoph Lameter81819f02007-05-06 14:49:36 -07001827 unsigned long slab_size = PAGE_SIZE << order;
1828
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001829 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001830 continue;
1831
Christoph Lameter81819f02007-05-06 14:49:36 -07001832 rem = slab_size % size;
1833
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001834 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001835 break;
1836
1837 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001838
Christoph Lameter81819f02007-05-06 14:49:36 -07001839 return order;
1840}
1841
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001842static inline int calculate_order(int size)
1843{
1844 int order;
1845 int min_objects;
1846 int fraction;
1847
1848 /*
1849 * Attempt to find best configuration for a slab. This
1850 * works by first attempting to generate a layout with
1851 * the best configuration and backing off gradually.
1852 *
1853 * First we reduce the acceptable waste in a slab. Then
1854 * we reduce the minimum objects required in a slab.
1855 */
1856 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001857 if (!min_objects)
1858 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001859 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001860 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001861 while (fraction >= 4) {
1862 order = slab_order(size, min_objects,
1863 slub_max_order, fraction);
1864 if (order <= slub_max_order)
1865 return order;
1866 fraction /= 2;
1867 }
1868 min_objects /= 2;
1869 }
1870
1871 /*
1872 * We were unable to place multiple objects in a slab. Now
1873 * lets see if we can place a single object there.
1874 */
1875 order = slab_order(size, 1, slub_max_order, 1);
1876 if (order <= slub_max_order)
1877 return order;
1878
1879 /*
1880 * Doh this slab cannot be placed using slub_max_order.
1881 */
1882 order = slab_order(size, 1, MAX_ORDER, 1);
1883 if (order <= MAX_ORDER)
1884 return order;
1885 return -ENOSYS;
1886}
1887
Christoph Lameter81819f02007-05-06 14:49:36 -07001888/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001889 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001890 */
1891static unsigned long calculate_alignment(unsigned long flags,
1892 unsigned long align, unsigned long size)
1893{
1894 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001895 * If the user wants hardware cache aligned objects then follow that
1896 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001897 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001898 * The hardware cache alignment cannot override the specified
1899 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001900 */
Nick Pigginb6210382008-03-05 14:05:56 -08001901 if (flags & SLAB_HWCACHE_ALIGN) {
1902 unsigned long ralign = cache_line_size();
1903 while (size <= ralign / 2)
1904 ralign /= 2;
1905 align = max(align, ralign);
1906 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001907
1908 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001909 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001910
1911 return ALIGN(align, sizeof(void *));
1912}
1913
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001914static void init_kmem_cache_cpu(struct kmem_cache *s,
1915 struct kmem_cache_cpu *c)
1916{
1917 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001918 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001919 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001920 c->offset = s->offset / sizeof(void *);
1921 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03001922#ifdef CONFIG_SLUB_STATS
1923 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1924#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001925}
1926
Pekka Enberg5595cff2008-08-05 09:28:47 +03001927static void
1928init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07001929{
1930 n->nr_partial = 0;
Pekka Enberg5595cff2008-08-05 09:28:47 +03001931
1932 /*
1933 * The larger the object size is, the more pages we want on the partial
1934 * list to avoid pounding the page allocator excessively.
1935 */
1936 n->min_partial = ilog2(s->size);
1937 if (n->min_partial < MIN_PARTIAL)
1938 n->min_partial = MIN_PARTIAL;
1939 else if (n->min_partial > MAX_PARTIAL)
1940 n->min_partial = MAX_PARTIAL;
1941
Christoph Lameter81819f02007-05-06 14:49:36 -07001942 spin_lock_init(&n->list_lock);
1943 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001944#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001945 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07001946 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07001947 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001948#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001949}
1950
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001951#ifdef CONFIG_SMP
1952/*
1953 * Per cpu array for per cpu structures.
1954 *
1955 * The per cpu array places all kmem_cache_cpu structures from one processor
1956 * close together meaning that it becomes possible that multiple per cpu
1957 * structures are contained in one cacheline. This may be particularly
1958 * beneficial for the kmalloc caches.
1959 *
1960 * A desktop system typically has around 60-80 slabs. With 100 here we are
1961 * likely able to get per cpu structures for all caches from the array defined
1962 * here. We must be able to cover all kmalloc caches during bootstrap.
1963 *
1964 * If the per cpu array is exhausted then fall back to kmalloc
1965 * of individual cachelines. No sharing is possible then.
1966 */
1967#define NR_KMEM_CACHE_CPU 100
1968
1969static DEFINE_PER_CPU(struct kmem_cache_cpu,
1970 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1971
1972static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
Rusty Russell174596a2009-01-01 10:12:29 +10301973static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001974
1975static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1976 int cpu, gfp_t flags)
1977{
1978 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1979
1980 if (c)
1981 per_cpu(kmem_cache_cpu_free, cpu) =
1982 (void *)c->freelist;
1983 else {
1984 /* Table overflow: So allocate ourselves */
1985 c = kmalloc_node(
1986 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1987 flags, cpu_to_node(cpu));
1988 if (!c)
1989 return NULL;
1990 }
1991
1992 init_kmem_cache_cpu(s, c);
1993 return c;
1994}
1995
1996static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1997{
1998 if (c < per_cpu(kmem_cache_cpu, cpu) ||
David Rientjes37189092009-01-27 18:59:46 -08001999 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002000 kfree(c);
2001 return;
2002 }
2003 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2004 per_cpu(kmem_cache_cpu_free, cpu) = c;
2005}
2006
2007static void free_kmem_cache_cpus(struct kmem_cache *s)
2008{
2009 int cpu;
2010
2011 for_each_online_cpu(cpu) {
2012 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2013
2014 if (c) {
2015 s->cpu_slab[cpu] = NULL;
2016 free_kmem_cache_cpu(c, cpu);
2017 }
2018 }
2019}
2020
2021static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2022{
2023 int cpu;
2024
2025 for_each_online_cpu(cpu) {
2026 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2027
2028 if (c)
2029 continue;
2030
2031 c = alloc_kmem_cache_cpu(s, cpu, flags);
2032 if (!c) {
2033 free_kmem_cache_cpus(s);
2034 return 0;
2035 }
2036 s->cpu_slab[cpu] = c;
2037 }
2038 return 1;
2039}
2040
2041/*
2042 * Initialize the per cpu array.
2043 */
2044static void init_alloc_cpu_cpu(int cpu)
2045{
2046 int i;
2047
Rusty Russell174596a2009-01-01 10:12:29 +10302048 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002049 return;
2050
2051 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2052 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2053
Rusty Russell174596a2009-01-01 10:12:29 +10302054 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002055}
2056
2057static void __init init_alloc_cpu(void)
2058{
2059 int cpu;
2060
2061 for_each_online_cpu(cpu)
2062 init_alloc_cpu_cpu(cpu);
2063 }
2064
2065#else
2066static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2067static inline void init_alloc_cpu(void) {}
2068
2069static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2070{
2071 init_kmem_cache_cpu(s, &s->cpu_slab);
2072 return 1;
2073}
2074#endif
2075
Christoph Lameter81819f02007-05-06 14:49:36 -07002076#ifdef CONFIG_NUMA
2077/*
2078 * No kmalloc_node yet so do it by hand. We know that this is the first
2079 * slab on the node for this slabcache. There are no concurrent accesses
2080 * possible.
2081 *
2082 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002083 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2084 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002085 */
David Rientjes0094de92008-11-25 19:14:19 -08002086static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002087{
2088 struct page *page;
2089 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002090 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002091
2092 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2093
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002094 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002095
2096 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002097 if (page_to_nid(page) != node) {
2098 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2099 "node %d\n", node);
2100 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2101 "in order to be able to continue\n");
2102 }
2103
Christoph Lameter81819f02007-05-06 14:49:36 -07002104 n = page->freelist;
2105 BUG_ON(!n);
2106 page->freelist = get_freepointer(kmalloc_caches, n);
2107 page->inuse++;
2108 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002109#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002110 init_object(kmalloc_caches, n, 1);
2111 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002112#endif
Pekka Enberg5595cff2008-08-05 09:28:47 +03002113 init_kmem_cache_node(n, kmalloc_caches);
Christoph Lameter205ab992008-04-14 19:11:40 +03002114 inc_slabs_node(kmalloc_caches, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002115
rootba84c732008-01-07 23:20:28 -08002116 /*
2117 * lockdep requires consistent irq usage for each lock
2118 * so even though there cannot be a race this early in
2119 * the boot sequence, we still disable irqs.
2120 */
2121 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002122 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002123 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002124}
2125
2126static void free_kmem_cache_nodes(struct kmem_cache *s)
2127{
2128 int node;
2129
Christoph Lameterf64dc582007-10-16 01:25:33 -07002130 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002131 struct kmem_cache_node *n = s->node[node];
2132 if (n && n != &s->local_node)
2133 kmem_cache_free(kmalloc_caches, n);
2134 s->node[node] = NULL;
2135 }
2136}
2137
2138static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2139{
2140 int node;
2141 int local_node;
2142
2143 if (slab_state >= UP)
2144 local_node = page_to_nid(virt_to_page(s));
2145 else
2146 local_node = 0;
2147
Christoph Lameterf64dc582007-10-16 01:25:33 -07002148 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002149 struct kmem_cache_node *n;
2150
2151 if (local_node == node)
2152 n = &s->local_node;
2153 else {
2154 if (slab_state == DOWN) {
David Rientjes0094de92008-11-25 19:14:19 -08002155 early_kmem_cache_node_alloc(gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002156 continue;
2157 }
2158 n = kmem_cache_alloc_node(kmalloc_caches,
2159 gfpflags, node);
2160
2161 if (!n) {
2162 free_kmem_cache_nodes(s);
2163 return 0;
2164 }
2165
2166 }
2167 s->node[node] = n;
Pekka Enberg5595cff2008-08-05 09:28:47 +03002168 init_kmem_cache_node(n, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002169 }
2170 return 1;
2171}
2172#else
2173static void free_kmem_cache_nodes(struct kmem_cache *s)
2174{
2175}
2176
2177static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2178{
Pekka Enberg5595cff2008-08-05 09:28:47 +03002179 init_kmem_cache_node(&s->local_node, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002180 return 1;
2181}
2182#endif
2183
2184/*
2185 * calculate_sizes() determines the order and the distribution of data within
2186 * a slab object.
2187 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002188static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002189{
2190 unsigned long flags = s->flags;
2191 unsigned long size = s->objsize;
2192 unsigned long align = s->align;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002193 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002194
2195 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002196 * Round up object size to the next word boundary. We can only
2197 * place the free pointer at word boundaries and this determines
2198 * the possible location of the free pointer.
2199 */
2200 size = ALIGN(size, sizeof(void *));
2201
2202#ifdef CONFIG_SLUB_DEBUG
2203 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002204 * Determine if we can poison the object itself. If the user of
2205 * the slab may touch the object after free or before allocation
2206 * then we should never poison the object itself.
2207 */
2208 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002209 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002210 s->flags |= __OBJECT_POISON;
2211 else
2212 s->flags &= ~__OBJECT_POISON;
2213
Christoph Lameter81819f02007-05-06 14:49:36 -07002214
2215 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002216 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002217 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002218 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002219 */
2220 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2221 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002222#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002223
2224 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002225 * With that we have determined the number of bytes in actual use
2226 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002227 */
2228 s->inuse = size;
2229
2230 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002231 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002232 /*
2233 * Relocate free pointer after the object if it is not
2234 * permitted to overwrite the first word of the object on
2235 * kmem_cache_free.
2236 *
2237 * This is the case if we do RCU, have a constructor or
2238 * destructor or are poisoning the objects.
2239 */
2240 s->offset = size;
2241 size += sizeof(void *);
2242 }
2243
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002244#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002245 if (flags & SLAB_STORE_USER)
2246 /*
2247 * Need to store information about allocs and frees after
2248 * the object.
2249 */
2250 size += 2 * sizeof(struct track);
2251
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002252 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002253 /*
2254 * Add some empty padding so that we can catch
2255 * overwrites from earlier objects rather than let
2256 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01002257 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07002258 * of the object.
2259 */
2260 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002261#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002262
Christoph Lameter81819f02007-05-06 14:49:36 -07002263 /*
2264 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002265 * user specified and the dynamic determination of cache line size
2266 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002267 */
2268 align = calculate_alignment(flags, align, s->objsize);
2269
2270 /*
2271 * SLUB stores one object immediately after another beginning from
2272 * offset 0. In order to align the objects we have to simply size
2273 * each object to conform to the alignment.
2274 */
2275 size = ALIGN(size, align);
2276 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03002277 if (forced_order >= 0)
2278 order = forced_order;
2279 else
2280 order = calculate_order(size);
Christoph Lameter81819f02007-05-06 14:49:36 -07002281
Christoph Lameter834f3d12008-04-14 19:11:31 +03002282 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07002283 return 0;
2284
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002285 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002286 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002287 s->allocflags |= __GFP_COMP;
2288
2289 if (s->flags & SLAB_CACHE_DMA)
2290 s->allocflags |= SLUB_DMA;
2291
2292 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2293 s->allocflags |= __GFP_RECLAIMABLE;
2294
Christoph Lameter81819f02007-05-06 14:49:36 -07002295 /*
2296 * Determine the number of objects per slab
2297 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002298 s->oo = oo_make(order, size);
Christoph Lameter65c33762008-04-14 19:11:40 +03002299 s->min = oo_make(get_order(size), size);
Christoph Lameter205ab992008-04-14 19:11:40 +03002300 if (oo_objects(s->oo) > oo_objects(s->max))
2301 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002302
Christoph Lameter834f3d12008-04-14 19:11:31 +03002303 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07002304
2305}
2306
Christoph Lameter81819f02007-05-06 14:49:36 -07002307static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2308 const char *name, size_t size,
2309 size_t align, unsigned long flags,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002310 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002311{
2312 memset(s, 0, kmem_size);
2313 s->name = name;
2314 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002315 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002316 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002317 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002318
Christoph Lameter06b285d2008-04-14 19:11:41 +03002319 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07002320 goto error;
2321
2322 s->refcount = 1;
2323#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05002324 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07002325#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002326 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2327 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002328
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002329 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002330 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002331 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002332error:
2333 if (flags & SLAB_PANIC)
2334 panic("Cannot create slab %s size=%lu realsize=%u "
2335 "order=%u offset=%u flags=%lx\n",
Christoph Lameter834f3d12008-04-14 19:11:31 +03002336 s->name, (unsigned long)size, s->size, oo_order(s->oo),
Christoph Lameter81819f02007-05-06 14:49:36 -07002337 s->offset, flags);
2338 return 0;
2339}
Christoph Lameter81819f02007-05-06 14:49:36 -07002340
2341/*
2342 * Check if a given pointer is valid
2343 */
2344int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2345{
Pekka Enberg06428782008-01-07 23:20:27 -08002346 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002347
2348 page = get_object_page(object);
2349
2350 if (!page || s != page->slab)
2351 /* No slab or wrong slab */
2352 return 0;
2353
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002354 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002355 return 0;
2356
2357 /*
2358 * We could also check if the object is on the slabs freelist.
2359 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002360 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002361 * to a certain slab.
2362 */
2363 return 1;
2364}
2365EXPORT_SYMBOL(kmem_ptr_validate);
2366
2367/*
2368 * Determine the size of a slab object
2369 */
2370unsigned int kmem_cache_size(struct kmem_cache *s)
2371{
2372 return s->objsize;
2373}
2374EXPORT_SYMBOL(kmem_cache_size);
2375
2376const char *kmem_cache_name(struct kmem_cache *s)
2377{
2378 return s->name;
2379}
2380EXPORT_SYMBOL(kmem_cache_name);
2381
Christoph Lameter33b12c32008-04-25 12:22:43 -07002382static void list_slab_objects(struct kmem_cache *s, struct page *page,
2383 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07002384{
Christoph Lameter33b12c32008-04-25 12:22:43 -07002385#ifdef CONFIG_SLUB_DEBUG
2386 void *addr = page_address(page);
2387 void *p;
2388 DECLARE_BITMAP(map, page->objects);
2389
2390 bitmap_zero(map, page->objects);
2391 slab_err(s, page, "%s", text);
2392 slab_lock(page);
2393 for_each_free_object(p, s, page->freelist)
2394 set_bit(slab_index(p, s, addr), map);
2395
2396 for_each_object(p, s, addr, page->objects) {
2397
2398 if (!test_bit(slab_index(p, s, addr), map)) {
2399 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2400 p, p - addr);
2401 print_tracking(s, p);
2402 }
2403 }
2404 slab_unlock(page);
2405#endif
2406}
2407
Christoph Lameter81819f02007-05-06 14:49:36 -07002408/*
Christoph Lameter599870b2008-04-23 12:36:52 -07002409 * Attempt to free all partial slabs on a node.
Christoph Lameter81819f02007-05-06 14:49:36 -07002410 */
Christoph Lameter599870b2008-04-23 12:36:52 -07002411static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002412{
Christoph Lameter81819f02007-05-06 14:49:36 -07002413 unsigned long flags;
2414 struct page *page, *h;
2415
2416 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter33b12c32008-04-25 12:22:43 -07002417 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002418 if (!page->inuse) {
2419 list_del(&page->lru);
2420 discard_slab(s, page);
Christoph Lameter599870b2008-04-23 12:36:52 -07002421 n->nr_partial--;
Christoph Lameter33b12c32008-04-25 12:22:43 -07002422 } else {
2423 list_slab_objects(s, page,
2424 "Objects remaining on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07002425 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07002426 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002427 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002428}
2429
2430/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002431 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002432 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002433static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002434{
2435 int node;
2436
2437 flush_all(s);
2438
2439 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002440 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002441 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002442 struct kmem_cache_node *n = get_node(s, node);
2443
Christoph Lameter599870b2008-04-23 12:36:52 -07002444 free_partial(s, n);
2445 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002446 return 1;
2447 }
2448 free_kmem_cache_nodes(s);
2449 return 0;
2450}
2451
2452/*
2453 * Close a cache and release the kmem_cache structure
2454 * (must be used for caches created using kmem_cache_create)
2455 */
2456void kmem_cache_destroy(struct kmem_cache *s)
2457{
2458 down_write(&slub_lock);
2459 s->refcount--;
2460 if (!s->refcount) {
2461 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002462 up_write(&slub_lock);
Pekka Enbergd629d812008-04-23 22:31:08 +03002463 if (kmem_cache_close(s)) {
2464 printk(KERN_ERR "SLUB %s: %s called for cache that "
2465 "still has objects.\n", s->name, __func__);
2466 dump_stack();
2467 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002468 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002469 } else
2470 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002471}
2472EXPORT_SYMBOL(kmem_cache_destroy);
2473
2474/********************************************************************
2475 * Kmalloc subsystem
2476 *******************************************************************/
2477
Christoph Lameter331dc552008-02-14 14:28:09 -08002478struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002479EXPORT_SYMBOL(kmalloc_caches);
2480
Christoph Lameter81819f02007-05-06 14:49:36 -07002481static int __init setup_slub_min_order(char *str)
2482{
Pekka Enberg06428782008-01-07 23:20:27 -08002483 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002484
2485 return 1;
2486}
2487
2488__setup("slub_min_order=", setup_slub_min_order);
2489
2490static int __init setup_slub_max_order(char *str)
2491{
Pekka Enberg06428782008-01-07 23:20:27 -08002492 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002493
2494 return 1;
2495}
2496
2497__setup("slub_max_order=", setup_slub_max_order);
2498
2499static int __init setup_slub_min_objects(char *str)
2500{
Pekka Enberg06428782008-01-07 23:20:27 -08002501 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002502
2503 return 1;
2504}
2505
2506__setup("slub_min_objects=", setup_slub_min_objects);
2507
2508static int __init setup_slub_nomerge(char *str)
2509{
2510 slub_nomerge = 1;
2511 return 1;
2512}
2513
2514__setup("slub_nomerge", setup_slub_nomerge);
2515
Christoph Lameter81819f02007-05-06 14:49:36 -07002516static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2517 const char *name, int size, gfp_t gfp_flags)
2518{
2519 unsigned int flags = 0;
2520
2521 if (gfp_flags & SLUB_DMA)
2522 flags = SLAB_CACHE_DMA;
2523
2524 down_write(&slub_lock);
2525 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter319d1e22008-04-14 19:11:41 +03002526 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002527 goto panic;
2528
2529 list_add(&s->list, &slab_caches);
2530 up_write(&slub_lock);
2531 if (sysfs_slab_add(s))
2532 goto panic;
2533 return s;
2534
2535panic:
2536 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2537}
2538
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002539#ifdef CONFIG_ZONE_DMA
Christoph Lameter4097d602008-04-14 18:51:18 +03002540static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002541
2542static void sysfs_add_func(struct work_struct *w)
2543{
2544 struct kmem_cache *s;
2545
2546 down_write(&slub_lock);
2547 list_for_each_entry(s, &slab_caches, list) {
2548 if (s->flags & __SYSFS_ADD_DEFERRED) {
2549 s->flags &= ~__SYSFS_ADD_DEFERRED;
2550 sysfs_slab_add(s);
2551 }
2552 }
2553 up_write(&slub_lock);
2554}
2555
2556static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2557
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002558static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2559{
2560 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002561 char *text;
2562 size_t realsize;
2563
2564 s = kmalloc_caches_dma[index];
2565 if (s)
2566 return s;
2567
2568 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002569 if (flags & __GFP_WAIT)
2570 down_write(&slub_lock);
2571 else {
2572 if (!down_write_trylock(&slub_lock))
2573 goto out;
2574 }
2575
2576 if (kmalloc_caches_dma[index])
2577 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002578
Christoph Lameter7b55f622007-07-17 04:03:27 -07002579 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002580 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2581 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002582 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2583
2584 if (!s || !text || !kmem_cache_open(s, flags, text,
2585 realsize, ARCH_KMALLOC_MINALIGN,
2586 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2587 kfree(s);
2588 kfree(text);
2589 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002590 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002591
2592 list_add(&s->list, &slab_caches);
2593 kmalloc_caches_dma[index] = s;
2594
2595 schedule_work(&sysfs_add_work);
2596
2597unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002598 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002599out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002600 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002601}
2602#endif
2603
Christoph Lameterf1b26332007-07-17 04:03:26 -07002604/*
2605 * Conversion table for small slabs sizes / 8 to the index in the
2606 * kmalloc array. This is necessary for slabs < 192 since we have non power
2607 * of two cache sizes there. The size of larger slabs can be determined using
2608 * fls.
2609 */
2610static s8 size_index[24] = {
2611 3, /* 8 */
2612 4, /* 16 */
2613 5, /* 24 */
2614 5, /* 32 */
2615 6, /* 40 */
2616 6, /* 48 */
2617 6, /* 56 */
2618 6, /* 64 */
2619 1, /* 72 */
2620 1, /* 80 */
2621 1, /* 88 */
2622 1, /* 96 */
2623 7, /* 104 */
2624 7, /* 112 */
2625 7, /* 120 */
2626 7, /* 128 */
2627 2, /* 136 */
2628 2, /* 144 */
2629 2, /* 152 */
2630 2, /* 160 */
2631 2, /* 168 */
2632 2, /* 176 */
2633 2, /* 184 */
2634 2 /* 192 */
2635};
2636
Christoph Lameter81819f02007-05-06 14:49:36 -07002637static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2638{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002639 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002640
Christoph Lameterf1b26332007-07-17 04:03:26 -07002641 if (size <= 192) {
2642 if (!size)
2643 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002644
Christoph Lameterf1b26332007-07-17 04:03:26 -07002645 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002646 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002647 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002648
2649#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002650 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002651 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002652
Christoph Lameter81819f02007-05-06 14:49:36 -07002653#endif
2654 return &kmalloc_caches[index];
2655}
2656
2657void *__kmalloc(size_t size, gfp_t flags)
2658{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002659 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002660
Christoph Lameter331dc552008-02-14 14:28:09 -08002661 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002662 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002663
2664 s = get_slab(size, flags);
2665
2666 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002667 return s;
2668
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002669 return slab_alloc(s, flags, -1, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07002670}
2671EXPORT_SYMBOL(__kmalloc);
2672
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002673static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2674{
2675 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2676 get_order(size));
2677
2678 if (page)
2679 return page_address(page);
2680 else
2681 return NULL;
2682}
2683
Christoph Lameter81819f02007-05-06 14:49:36 -07002684#ifdef CONFIG_NUMA
2685void *__kmalloc_node(size_t size, gfp_t flags, int node)
2686{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002687 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002688
Christoph Lameter331dc552008-02-14 14:28:09 -08002689 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002690 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002691
2692 s = get_slab(size, flags);
2693
2694 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002695 return s;
2696
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002697 return slab_alloc(s, flags, node, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07002698}
2699EXPORT_SYMBOL(__kmalloc_node);
2700#endif
2701
2702size_t ksize(const void *object)
2703{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002704 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002705 struct kmem_cache *s;
2706
Christoph Lameteref8b4522007-10-16 01:24:46 -07002707 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002708 return 0;
2709
Vegard Nossum294a80a2007-12-04 23:45:30 -08002710 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002711
Pekka Enberg76994412008-05-22 19:22:25 +03002712 if (unlikely(!PageSlab(page))) {
2713 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08002714 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03002715 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002716 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002717
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002718#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002719 /*
2720 * Debugging requires use of the padding between object
2721 * and whatever may come after it.
2722 */
2723 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2724 return s->objsize;
2725
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002726#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002727 /*
2728 * If we have the need to store the freelist pointer
2729 * back there or track user information then we can
2730 * only use the space before that information.
2731 */
2732 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2733 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002734 /*
2735 * Else we can use all the padding etc for the allocation
2736 */
2737 return s->size;
2738}
Christoph Lameter81819f02007-05-06 14:49:36 -07002739
2740void kfree(const void *x)
2741{
Christoph Lameter81819f02007-05-06 14:49:36 -07002742 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002743 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002744
Satyam Sharma2408c552007-10-16 01:24:44 -07002745 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002746 return;
2747
Christoph Lameterb49af682007-05-06 14:49:41 -07002748 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002749 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07002750 BUG_ON(!PageCompound(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002751 put_page(page);
2752 return;
2753 }
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002754 slab_free(page->slab, page, object, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07002755}
2756EXPORT_SYMBOL(kfree);
2757
Christoph Lameter2086d262007-05-06 14:49:46 -07002758/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002759 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2760 * the remaining slabs by the number of items in use. The slabs with the
2761 * most items in use come first. New allocations will then fill those up
2762 * and thus they can be removed from the partial lists.
2763 *
2764 * The slabs with the least items are placed last. This results in them
2765 * being allocated from last increasing the chance that the last objects
2766 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002767 */
2768int kmem_cache_shrink(struct kmem_cache *s)
2769{
2770 int node;
2771 int i;
2772 struct kmem_cache_node *n;
2773 struct page *page;
2774 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03002775 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07002776 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03002777 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07002778 unsigned long flags;
2779
2780 if (!slabs_by_inuse)
2781 return -ENOMEM;
2782
2783 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002784 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002785 n = get_node(s, node);
2786
2787 if (!n->nr_partial)
2788 continue;
2789
Christoph Lameter834f3d12008-04-14 19:11:31 +03002790 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07002791 INIT_LIST_HEAD(slabs_by_inuse + i);
2792
2793 spin_lock_irqsave(&n->list_lock, flags);
2794
2795 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002796 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002797 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002798 * Note that concurrent frees may occur while we hold the
2799 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002800 */
2801 list_for_each_entry_safe(page, t, &n->partial, lru) {
2802 if (!page->inuse && slab_trylock(page)) {
2803 /*
2804 * Must hold slab lock here because slab_free
2805 * may have freed the last object and be
2806 * waiting to release the slab.
2807 */
2808 list_del(&page->lru);
2809 n->nr_partial--;
2810 slab_unlock(page);
2811 discard_slab(s, page);
2812 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002813 list_move(&page->lru,
2814 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002815 }
2816 }
2817
Christoph Lameter2086d262007-05-06 14:49:46 -07002818 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002819 * Rebuild the partial list with the slabs filled up most
2820 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002821 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002822 for (i = objects - 1; i >= 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07002823 list_splice(slabs_by_inuse + i, n->partial.prev);
2824
Christoph Lameter2086d262007-05-06 14:49:46 -07002825 spin_unlock_irqrestore(&n->list_lock, flags);
2826 }
2827
2828 kfree(slabs_by_inuse);
2829 return 0;
2830}
2831EXPORT_SYMBOL(kmem_cache_shrink);
2832
Yasunori Gotob9049e22007-10-21 16:41:37 -07002833#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2834static int slab_mem_going_offline_callback(void *arg)
2835{
2836 struct kmem_cache *s;
2837
2838 down_read(&slub_lock);
2839 list_for_each_entry(s, &slab_caches, list)
2840 kmem_cache_shrink(s);
2841 up_read(&slub_lock);
2842
2843 return 0;
2844}
2845
2846static void slab_mem_offline_callback(void *arg)
2847{
2848 struct kmem_cache_node *n;
2849 struct kmem_cache *s;
2850 struct memory_notify *marg = arg;
2851 int offline_node;
2852
2853 offline_node = marg->status_change_nid;
2854
2855 /*
2856 * If the node still has available memory. we need kmem_cache_node
2857 * for it yet.
2858 */
2859 if (offline_node < 0)
2860 return;
2861
2862 down_read(&slub_lock);
2863 list_for_each_entry(s, &slab_caches, list) {
2864 n = get_node(s, offline_node);
2865 if (n) {
2866 /*
2867 * if n->nr_slabs > 0, slabs still exist on the node
2868 * that is going down. We were unable to free them,
2869 * and offline_pages() function shoudn't call this
2870 * callback. So, we must fail.
2871 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002872 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002873
2874 s->node[offline_node] = NULL;
2875 kmem_cache_free(kmalloc_caches, n);
2876 }
2877 }
2878 up_read(&slub_lock);
2879}
2880
2881static int slab_mem_going_online_callback(void *arg)
2882{
2883 struct kmem_cache_node *n;
2884 struct kmem_cache *s;
2885 struct memory_notify *marg = arg;
2886 int nid = marg->status_change_nid;
2887 int ret = 0;
2888
2889 /*
2890 * If the node's memory is already available, then kmem_cache_node is
2891 * already created. Nothing to do.
2892 */
2893 if (nid < 0)
2894 return 0;
2895
2896 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07002897 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07002898 * allocate a kmem_cache_node structure in order to bring the node
2899 * online.
2900 */
2901 down_read(&slub_lock);
2902 list_for_each_entry(s, &slab_caches, list) {
2903 /*
2904 * XXX: kmem_cache_alloc_node will fallback to other nodes
2905 * since memory is not yet available from the node that
2906 * is brought up.
2907 */
2908 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2909 if (!n) {
2910 ret = -ENOMEM;
2911 goto out;
2912 }
Pekka Enberg5595cff2008-08-05 09:28:47 +03002913 init_kmem_cache_node(n, s);
Yasunori Gotob9049e22007-10-21 16:41:37 -07002914 s->node[nid] = n;
2915 }
2916out:
2917 up_read(&slub_lock);
2918 return ret;
2919}
2920
2921static int slab_memory_callback(struct notifier_block *self,
2922 unsigned long action, void *arg)
2923{
2924 int ret = 0;
2925
2926 switch (action) {
2927 case MEM_GOING_ONLINE:
2928 ret = slab_mem_going_online_callback(arg);
2929 break;
2930 case MEM_GOING_OFFLINE:
2931 ret = slab_mem_going_offline_callback(arg);
2932 break;
2933 case MEM_OFFLINE:
2934 case MEM_CANCEL_ONLINE:
2935 slab_mem_offline_callback(arg);
2936 break;
2937 case MEM_ONLINE:
2938 case MEM_CANCEL_OFFLINE:
2939 break;
2940 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08002941 if (ret)
2942 ret = notifier_from_errno(ret);
2943 else
2944 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002945 return ret;
2946}
2947
2948#endif /* CONFIG_MEMORY_HOTPLUG */
2949
Christoph Lameter81819f02007-05-06 14:49:36 -07002950/********************************************************************
2951 * Basic setup of slabs
2952 *******************************************************************/
2953
2954void __init kmem_cache_init(void)
2955{
2956 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002957 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002958
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002959 init_alloc_cpu();
2960
Christoph Lameter81819f02007-05-06 14:49:36 -07002961#ifdef CONFIG_NUMA
2962 /*
2963 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002964 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002965 * kmem_cache_open for slab_state == DOWN.
2966 */
2967 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2968 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002969 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002970 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002971
Nadia Derbey0c40ba42008-04-29 01:00:41 -07002972 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
Christoph Lameter81819f02007-05-06 14:49:36 -07002973#endif
2974
2975 /* Able to allocate the per node structures */
2976 slab_state = PARTIAL;
2977
2978 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002979 if (KMALLOC_MIN_SIZE <= 64) {
2980 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002981 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002982 caches++;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002983 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002984 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002985 caches++;
2986 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002987
Christoph Lameter331dc552008-02-14 14:28:09 -08002988 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002989 create_kmalloc_cache(&kmalloc_caches[i],
2990 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002991 caches++;
2992 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002993
Christoph Lameterf1b26332007-07-17 04:03:26 -07002994
2995 /*
2996 * Patch up the size_index table if we have strange large alignment
2997 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002998 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002999 *
3000 * Largest permitted alignment is 256 bytes due to the way we
3001 * handle the index determination for the smaller caches.
3002 *
3003 * Make sure that nothing crazy happens if someone starts tinkering
3004 * around with ARCH_KMALLOC_MINALIGN
3005 */
3006 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3007 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3008
Christoph Lameter12ad6842007-07-17 04:03:28 -07003009 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07003010 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3011
Christoph Lameter41d54d32008-07-03 09:14:26 -05003012 if (KMALLOC_MIN_SIZE == 128) {
3013 /*
3014 * The 192 byte sized cache is not used if the alignment
3015 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3016 * instead.
3017 */
3018 for (i = 128 + 8; i <= 192; i += 8)
3019 size_index[(i - 1) / 8] = 8;
3020 }
3021
Christoph Lameter81819f02007-05-06 14:49:36 -07003022 slab_state = UP;
3023
3024 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08003025 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07003026 kmalloc_caches[i]. name =
3027 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3028
3029#ifdef CONFIG_SMP
3030 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003031 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3032 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3033#else
3034 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003035#endif
3036
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003037 printk(KERN_INFO
3038 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003039 " CPUs=%d, Nodes=%d\n",
3040 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003041 slub_min_order, slub_max_order, slub_min_objects,
3042 nr_cpu_ids, nr_node_ids);
3043}
3044
3045/*
3046 * Find a mergeable slab cache
3047 */
3048static int slab_unmergeable(struct kmem_cache *s)
3049{
3050 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3051 return 1;
3052
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003053 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003054 return 1;
3055
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003056 /*
3057 * We may have set a slab to be unmergeable during bootstrap.
3058 */
3059 if (s->refcount < 0)
3060 return 1;
3061
Christoph Lameter81819f02007-05-06 14:49:36 -07003062 return 0;
3063}
3064
3065static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003066 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003067 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003068{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003069 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003070
3071 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3072 return NULL;
3073
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003074 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003075 return NULL;
3076
3077 size = ALIGN(size, sizeof(void *));
3078 align = calculate_alignment(flags, align, size);
3079 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003080 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003081
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003082 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003083 if (slab_unmergeable(s))
3084 continue;
3085
3086 if (size > s->size)
3087 continue;
3088
Christoph Lameterba0268a2007-09-11 15:24:11 -07003089 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003090 continue;
3091 /*
3092 * Check if alignment is compatible.
3093 * Courtesy of Adrian Drzewiecki
3094 */
Pekka Enberg06428782008-01-07 23:20:27 -08003095 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003096 continue;
3097
3098 if (s->size - size >= sizeof(void *))
3099 continue;
3100
3101 return s;
3102 }
3103 return NULL;
3104}
3105
3106struct kmem_cache *kmem_cache_create(const char *name, size_t size,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003107 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003108{
3109 struct kmem_cache *s;
3110
3111 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003112 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003113 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003114 int cpu;
3115
Christoph Lameter81819f02007-05-06 14:49:36 -07003116 s->refcount++;
3117 /*
3118 * Adjust the object sizes so that we clear
3119 * the complete object on kzalloc.
3120 */
3121 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003122
3123 /*
3124 * And then we need to update the object size in the
3125 * per cpu structures
3126 */
3127 for_each_online_cpu(cpu)
3128 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003129
Christoph Lameter81819f02007-05-06 14:49:36 -07003130 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003131 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003132
David Rientjes7b8f3b62008-12-17 22:09:46 -08003133 if (sysfs_slab_alias(s, name)) {
3134 down_write(&slub_lock);
3135 s->refcount--;
3136 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003137 goto err;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003138 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003139 return s;
3140 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003141
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003142 s = kmalloc(kmem_size, GFP_KERNEL);
3143 if (s) {
3144 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003145 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003146 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003147 up_write(&slub_lock);
David Rientjes7b8f3b62008-12-17 22:09:46 -08003148 if (sysfs_slab_add(s)) {
3149 down_write(&slub_lock);
3150 list_del(&s->list);
3151 up_write(&slub_lock);
3152 kfree(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003153 goto err;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003154 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003155 return s;
3156 }
3157 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003158 }
3159 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003160
3161err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003162 if (flags & SLAB_PANIC)
3163 panic("Cannot create slabcache %s\n", name);
3164 else
3165 s = NULL;
3166 return s;
3167}
3168EXPORT_SYMBOL(kmem_cache_create);
3169
Christoph Lameter81819f02007-05-06 14:49:36 -07003170#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003171/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003172 * Use the cpu notifier to insure that the cpu slabs are flushed when
3173 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003174 */
3175static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3176 unsigned long action, void *hcpu)
3177{
3178 long cpu = (long)hcpu;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003179 struct kmem_cache *s;
3180 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003181
3182 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003183 case CPU_UP_PREPARE:
3184 case CPU_UP_PREPARE_FROZEN:
3185 init_alloc_cpu_cpu(cpu);
3186 down_read(&slub_lock);
3187 list_for_each_entry(s, &slab_caches, list)
3188 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3189 GFP_KERNEL);
3190 up_read(&slub_lock);
3191 break;
3192
Christoph Lameter81819f02007-05-06 14:49:36 -07003193 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003194 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003195 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003196 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003197 down_read(&slub_lock);
3198 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003199 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3200
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003201 local_irq_save(flags);
3202 __flush_cpu_slab(s, cpu);
3203 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003204 free_kmem_cache_cpu(c, cpu);
3205 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003206 }
3207 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003208 break;
3209 default:
3210 break;
3211 }
3212 return NOTIFY_OK;
3213}
3214
Pekka Enberg06428782008-01-07 23:20:27 -08003215static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003216 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003217};
Christoph Lameter81819f02007-05-06 14:49:36 -07003218
3219#endif
3220
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003221void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003222{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003223 struct kmem_cache *s;
3224
Christoph Lameter331dc552008-02-14 14:28:09 -08003225 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003226 return kmalloc_large(size, gfpflags);
3227
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003228 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003229
Satyam Sharma2408c552007-10-16 01:24:44 -07003230 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003231 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003232
Christoph Lameterce15fea2007-07-17 04:03:28 -07003233 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003234}
3235
3236void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003237 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003238{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003239 struct kmem_cache *s;
3240
Christoph Lameter331dc552008-02-14 14:28:09 -08003241 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003242 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003243
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003244 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003245
Satyam Sharma2408c552007-10-16 01:24:44 -07003246 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003247 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003248
Christoph Lameterce15fea2007-07-17 04:03:28 -07003249 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003250}
3251
Christoph Lameterf6acb632008-04-29 16:16:06 -07003252#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03003253static unsigned long count_partial(struct kmem_cache_node *n,
3254 int (*get_count)(struct page *))
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003255{
3256 unsigned long flags;
3257 unsigned long x = 0;
3258 struct page *page;
3259
3260 spin_lock_irqsave(&n->list_lock, flags);
3261 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter205ab992008-04-14 19:11:40 +03003262 x += get_count(page);
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003263 spin_unlock_irqrestore(&n->list_lock, flags);
3264 return x;
3265}
Christoph Lameter205ab992008-04-14 19:11:40 +03003266
3267static int count_inuse(struct page *page)
3268{
3269 return page->inuse;
3270}
3271
3272static int count_total(struct page *page)
3273{
3274 return page->objects;
3275}
3276
3277static int count_free(struct page *page)
3278{
3279 return page->objects - page->inuse;
3280}
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003281
Christoph Lameter434e2452007-07-17 04:03:30 -07003282static int validate_slab(struct kmem_cache *s, struct page *page,
3283 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003284{
3285 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003286 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003287
3288 if (!check_slab(s, page) ||
3289 !on_freelist(s, page, NULL))
3290 return 0;
3291
3292 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003293 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003294
Christoph Lameter7656c722007-05-09 02:32:40 -07003295 for_each_free_object(p, s, page->freelist) {
3296 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003297 if (!check_object(s, page, p, 0))
3298 return 0;
3299 }
3300
Christoph Lameter224a88b2008-04-14 19:11:31 +03003301 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003302 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003303 if (!check_object(s, page, p, 1))
3304 return 0;
3305 return 1;
3306}
3307
Christoph Lameter434e2452007-07-17 04:03:30 -07003308static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3309 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003310{
3311 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003312 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003313 slab_unlock(page);
3314 } else
3315 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3316 s->name, page);
3317
3318 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003319 if (!PageSlubDebug(page))
3320 printk(KERN_ERR "SLUB %s: SlubDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003321 "on slab 0x%p\n", s->name, page);
3322 } else {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003323 if (PageSlubDebug(page))
3324 printk(KERN_ERR "SLUB %s: SlubDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003325 "slab 0x%p\n", s->name, page);
3326 }
3327}
3328
Christoph Lameter434e2452007-07-17 04:03:30 -07003329static int validate_slab_node(struct kmem_cache *s,
3330 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003331{
3332 unsigned long count = 0;
3333 struct page *page;
3334 unsigned long flags;
3335
3336 spin_lock_irqsave(&n->list_lock, flags);
3337
3338 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003339 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003340 count++;
3341 }
3342 if (count != n->nr_partial)
3343 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3344 "counter=%ld\n", s->name, count, n->nr_partial);
3345
3346 if (!(s->flags & SLAB_STORE_USER))
3347 goto out;
3348
3349 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003350 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003351 count++;
3352 }
3353 if (count != atomic_long_read(&n->nr_slabs))
3354 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3355 "counter=%ld\n", s->name, count,
3356 atomic_long_read(&n->nr_slabs));
3357
3358out:
3359 spin_unlock_irqrestore(&n->list_lock, flags);
3360 return count;
3361}
3362
Christoph Lameter434e2452007-07-17 04:03:30 -07003363static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003364{
3365 int node;
3366 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003367 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003368 sizeof(unsigned long), GFP_KERNEL);
3369
3370 if (!map)
3371 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003372
3373 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003374 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003375 struct kmem_cache_node *n = get_node(s, node);
3376
Christoph Lameter434e2452007-07-17 04:03:30 -07003377 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003378 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003379 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003380 return count;
3381}
3382
Christoph Lameterb3459702007-05-09 02:32:41 -07003383#ifdef SLUB_RESILIENCY_TEST
3384static void resiliency_test(void)
3385{
3386 u8 *p;
3387
3388 printk(KERN_ERR "SLUB resiliency testing\n");
3389 printk(KERN_ERR "-----------------------\n");
3390 printk(KERN_ERR "A. Corruption after allocation\n");
3391
3392 p = kzalloc(16, GFP_KERNEL);
3393 p[16] = 0x12;
3394 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3395 " 0x12->0x%p\n\n", p + 16);
3396
3397 validate_slab_cache(kmalloc_caches + 4);
3398
3399 /* Hmmm... The next two are dangerous */
3400 p = kzalloc(32, GFP_KERNEL);
3401 p[32 + sizeof(void *)] = 0x34;
3402 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003403 " 0x34 -> -0x%p\n", p);
3404 printk(KERN_ERR
3405 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003406
3407 validate_slab_cache(kmalloc_caches + 5);
3408 p = kzalloc(64, GFP_KERNEL);
3409 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3410 *p = 0x56;
3411 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3412 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003413 printk(KERN_ERR
3414 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003415 validate_slab_cache(kmalloc_caches + 6);
3416
3417 printk(KERN_ERR "\nB. Corruption after free\n");
3418 p = kzalloc(128, GFP_KERNEL);
3419 kfree(p);
3420 *p = 0x78;
3421 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3422 validate_slab_cache(kmalloc_caches + 7);
3423
3424 p = kzalloc(256, GFP_KERNEL);
3425 kfree(p);
3426 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003427 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3428 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003429 validate_slab_cache(kmalloc_caches + 8);
3430
3431 p = kzalloc(512, GFP_KERNEL);
3432 kfree(p);
3433 p[512] = 0xab;
3434 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3435 validate_slab_cache(kmalloc_caches + 9);
3436}
3437#else
3438static void resiliency_test(void) {};
3439#endif
3440
Christoph Lameter88a420e2007-05-06 14:49:45 -07003441/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003442 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003443 * and freed.
3444 */
3445
3446struct location {
3447 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003448 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003449 long long sum_time;
3450 long min_time;
3451 long max_time;
3452 long min_pid;
3453 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303454 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003455 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003456};
3457
3458struct loc_track {
3459 unsigned long max;
3460 unsigned long count;
3461 struct location *loc;
3462};
3463
3464static void free_loc_track(struct loc_track *t)
3465{
3466 if (t->max)
3467 free_pages((unsigned long)t->loc,
3468 get_order(sizeof(struct location) * t->max));
3469}
3470
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003471static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003472{
3473 struct location *l;
3474 int order;
3475
Christoph Lameter88a420e2007-05-06 14:49:45 -07003476 order = get_order(sizeof(struct location) * max);
3477
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003478 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003479 if (!l)
3480 return 0;
3481
3482 if (t->count) {
3483 memcpy(l, t->loc, sizeof(struct location) * t->count);
3484 free_loc_track(t);
3485 }
3486 t->max = max;
3487 t->loc = l;
3488 return 1;
3489}
3490
3491static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003492 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003493{
3494 long start, end, pos;
3495 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003496 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003497 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003498
3499 start = -1;
3500 end = t->count;
3501
3502 for ( ; ; ) {
3503 pos = start + (end - start + 1) / 2;
3504
3505 /*
3506 * There is nothing at "end". If we end up there
3507 * we need to add something to before end.
3508 */
3509 if (pos == end)
3510 break;
3511
3512 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003513 if (track->addr == caddr) {
3514
3515 l = &t->loc[pos];
3516 l->count++;
3517 if (track->when) {
3518 l->sum_time += age;
3519 if (age < l->min_time)
3520 l->min_time = age;
3521 if (age > l->max_time)
3522 l->max_time = age;
3523
3524 if (track->pid < l->min_pid)
3525 l->min_pid = track->pid;
3526 if (track->pid > l->max_pid)
3527 l->max_pid = track->pid;
3528
Rusty Russell174596a2009-01-01 10:12:29 +10303529 cpumask_set_cpu(track->cpu,
3530 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003531 }
3532 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003533 return 1;
3534 }
3535
Christoph Lameter45edfa52007-05-09 02:32:45 -07003536 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003537 end = pos;
3538 else
3539 start = pos;
3540 }
3541
3542 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003543 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003544 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003545 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003546 return 0;
3547
3548 l = t->loc + pos;
3549 if (pos < t->count)
3550 memmove(l + 1, l,
3551 (t->count - pos) * sizeof(struct location));
3552 t->count++;
3553 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003554 l->addr = track->addr;
3555 l->sum_time = age;
3556 l->min_time = age;
3557 l->max_time = age;
3558 l->min_pid = track->pid;
3559 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303560 cpumask_clear(to_cpumask(l->cpus));
3561 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003562 nodes_clear(l->nodes);
3563 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003564 return 1;
3565}
3566
3567static void process_slab(struct loc_track *t, struct kmem_cache *s,
3568 struct page *page, enum track_item alloc)
3569{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003570 void *addr = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +03003571 DECLARE_BITMAP(map, page->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003572 void *p;
3573
Christoph Lameter39b26462008-04-14 19:11:30 +03003574 bitmap_zero(map, page->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003575 for_each_free_object(p, s, page->freelist)
3576 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003577
Christoph Lameter224a88b2008-04-14 19:11:31 +03003578 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003579 if (!test_bit(slab_index(p, s, addr), map))
3580 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003581}
3582
3583static int list_locations(struct kmem_cache *s, char *buf,
3584 enum track_item alloc)
3585{
Harvey Harrisone374d482008-01-31 15:20:50 -08003586 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003587 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003588 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003589 int node;
3590
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003591 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003592 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003593 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003594
3595 /* Push back cpu slabs */
3596 flush_all(s);
3597
Christoph Lameterf64dc582007-10-16 01:25:33 -07003598 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003599 struct kmem_cache_node *n = get_node(s, node);
3600 unsigned long flags;
3601 struct page *page;
3602
Christoph Lameter9e869432007-08-22 14:01:56 -07003603 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003604 continue;
3605
3606 spin_lock_irqsave(&n->list_lock, flags);
3607 list_for_each_entry(page, &n->partial, lru)
3608 process_slab(&t, s, page, alloc);
3609 list_for_each_entry(page, &n->full, lru)
3610 process_slab(&t, s, page, alloc);
3611 spin_unlock_irqrestore(&n->list_lock, flags);
3612 }
3613
3614 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003615 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003616
Hugh Dickins9c246242008-12-09 13:14:27 -08003617 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003618 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003619 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003620
3621 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003622 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003623 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003624 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003625
3626 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08003627 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07003628 l->min_time,
3629 (long)div_u64(l->sum_time, l->count),
3630 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003631 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003632 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003633 l->min_time);
3634
3635 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003636 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003637 l->min_pid, l->max_pid);
3638 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003639 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003640 l->min_pid);
3641
Rusty Russell174596a2009-01-01 10:12:29 +10303642 if (num_online_cpus() > 1 &&
3643 !cpumask_empty(to_cpumask(l->cpus)) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003644 len < PAGE_SIZE - 60) {
3645 len += sprintf(buf + len, " cpus=");
3646 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Rusty Russell174596a2009-01-01 10:12:29 +10303647 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003648 }
3649
Christoph Lameter84966342007-06-23 17:16:32 -07003650 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003651 len < PAGE_SIZE - 60) {
3652 len += sprintf(buf + len, " nodes=");
3653 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003654 l->nodes);
3655 }
3656
Harvey Harrisone374d482008-01-31 15:20:50 -08003657 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003658 }
3659
3660 free_loc_track(&t);
3661 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003662 len += sprintf(buf, "No data\n");
3663 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003664}
3665
Christoph Lameter81819f02007-05-06 14:49:36 -07003666enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03003667 SL_ALL, /* All slabs */
3668 SL_PARTIAL, /* Only partially allocated slabs */
3669 SL_CPU, /* Only slabs used for cpu caches */
3670 SL_OBJECTS, /* Determine allocated objects not slabs */
3671 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07003672};
3673
Christoph Lameter205ab992008-04-14 19:11:40 +03003674#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003675#define SO_PARTIAL (1 << SL_PARTIAL)
3676#define SO_CPU (1 << SL_CPU)
3677#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03003678#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003679
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003680static ssize_t show_slab_objects(struct kmem_cache *s,
3681 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003682{
3683 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003684 int node;
3685 int x;
3686 unsigned long *nodes;
3687 unsigned long *per_cpu;
3688
3689 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003690 if (!nodes)
3691 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003692 per_cpu = nodes + nr_node_ids;
3693
Christoph Lameter205ab992008-04-14 19:11:40 +03003694 if (flags & SO_CPU) {
3695 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07003696
Christoph Lameter205ab992008-04-14 19:11:40 +03003697 for_each_possible_cpu(cpu) {
3698 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003699
Christoph Lameter205ab992008-04-14 19:11:40 +03003700 if (!c || c->node < 0)
3701 continue;
3702
3703 if (c->page) {
3704 if (flags & SO_TOTAL)
3705 x = c->page->objects;
3706 else if (flags & SO_OBJECTS)
3707 x = c->page->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07003708 else
3709 x = 1;
Christoph Lameter205ab992008-04-14 19:11:40 +03003710
Christoph Lameter81819f02007-05-06 14:49:36 -07003711 total += x;
Christoph Lameter205ab992008-04-14 19:11:40 +03003712 nodes[c->node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003713 }
Christoph Lameter205ab992008-04-14 19:11:40 +03003714 per_cpu[c->node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003715 }
3716 }
3717
Christoph Lameter205ab992008-04-14 19:11:40 +03003718 if (flags & SO_ALL) {
3719 for_each_node_state(node, N_NORMAL_MEMORY) {
3720 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003721
Christoph Lameter205ab992008-04-14 19:11:40 +03003722 if (flags & SO_TOTAL)
3723 x = atomic_long_read(&n->total_objects);
3724 else if (flags & SO_OBJECTS)
3725 x = atomic_long_read(&n->total_objects) -
3726 count_partial(n, count_free);
3727
3728 else
3729 x = atomic_long_read(&n->nr_slabs);
3730 total += x;
3731 nodes[node] += x;
3732 }
3733
3734 } else if (flags & SO_PARTIAL) {
3735 for_each_node_state(node, N_NORMAL_MEMORY) {
3736 struct kmem_cache_node *n = get_node(s, node);
3737
3738 if (flags & SO_TOTAL)
3739 x = count_partial(n, count_total);
3740 else if (flags & SO_OBJECTS)
3741 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07003742 else
3743 x = n->nr_partial;
3744 total += x;
3745 nodes[node] += x;
3746 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003747 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003748 x = sprintf(buf, "%lu", total);
3749#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003750 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003751 if (nodes[node])
3752 x += sprintf(buf + x, " N%d=%lu",
3753 node, nodes[node]);
3754#endif
3755 kfree(nodes);
3756 return x + sprintf(buf + x, "\n");
3757}
3758
3759static int any_slab_objects(struct kmem_cache *s)
3760{
3761 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003762
3763 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003764 struct kmem_cache_node *n = get_node(s, node);
3765
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003766 if (!n)
3767 continue;
3768
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07003769 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07003770 return 1;
3771 }
3772 return 0;
3773}
3774
3775#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3776#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3777
3778struct slab_attribute {
3779 struct attribute attr;
3780 ssize_t (*show)(struct kmem_cache *s, char *buf);
3781 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3782};
3783
3784#define SLAB_ATTR_RO(_name) \
3785 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3786
3787#define SLAB_ATTR(_name) \
3788 static struct slab_attribute _name##_attr = \
3789 __ATTR(_name, 0644, _name##_show, _name##_store)
3790
Christoph Lameter81819f02007-05-06 14:49:36 -07003791static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3792{
3793 return sprintf(buf, "%d\n", s->size);
3794}
3795SLAB_ATTR_RO(slab_size);
3796
3797static ssize_t align_show(struct kmem_cache *s, char *buf)
3798{
3799 return sprintf(buf, "%d\n", s->align);
3800}
3801SLAB_ATTR_RO(align);
3802
3803static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3804{
3805 return sprintf(buf, "%d\n", s->objsize);
3806}
3807SLAB_ATTR_RO(object_size);
3808
3809static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3810{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003811 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003812}
3813SLAB_ATTR_RO(objs_per_slab);
3814
Christoph Lameter06b285d2008-04-14 19:11:41 +03003815static ssize_t order_store(struct kmem_cache *s,
3816 const char *buf, size_t length)
3817{
Christoph Lameter0121c6192008-04-29 16:11:12 -07003818 unsigned long order;
3819 int err;
3820
3821 err = strict_strtoul(buf, 10, &order);
3822 if (err)
3823 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003824
3825 if (order > slub_max_order || order < slub_min_order)
3826 return -EINVAL;
3827
3828 calculate_sizes(s, order);
3829 return length;
3830}
3831
Christoph Lameter81819f02007-05-06 14:49:36 -07003832static ssize_t order_show(struct kmem_cache *s, char *buf)
3833{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003834 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003835}
Christoph Lameter06b285d2008-04-14 19:11:41 +03003836SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003837
3838static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3839{
3840 if (s->ctor) {
3841 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3842
3843 return n + sprintf(buf + n, "\n");
3844 }
3845 return 0;
3846}
3847SLAB_ATTR_RO(ctor);
3848
Christoph Lameter81819f02007-05-06 14:49:36 -07003849static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3850{
3851 return sprintf(buf, "%d\n", s->refcount - 1);
3852}
3853SLAB_ATTR_RO(aliases);
3854
3855static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3856{
Christoph Lameter205ab992008-04-14 19:11:40 +03003857 return show_slab_objects(s, buf, SO_ALL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003858}
3859SLAB_ATTR_RO(slabs);
3860
3861static ssize_t partial_show(struct kmem_cache *s, char *buf)
3862{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003863 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003864}
3865SLAB_ATTR_RO(partial);
3866
3867static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3868{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003869 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003870}
3871SLAB_ATTR_RO(cpu_slabs);
3872
3873static ssize_t objects_show(struct kmem_cache *s, char *buf)
3874{
Christoph Lameter205ab992008-04-14 19:11:40 +03003875 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003876}
3877SLAB_ATTR_RO(objects);
3878
Christoph Lameter205ab992008-04-14 19:11:40 +03003879static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3880{
3881 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3882}
3883SLAB_ATTR_RO(objects_partial);
3884
3885static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3886{
3887 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3888}
3889SLAB_ATTR_RO(total_objects);
3890
Christoph Lameter81819f02007-05-06 14:49:36 -07003891static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3892{
3893 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3894}
3895
3896static ssize_t sanity_checks_store(struct kmem_cache *s,
3897 const char *buf, size_t length)
3898{
3899 s->flags &= ~SLAB_DEBUG_FREE;
3900 if (buf[0] == '1')
3901 s->flags |= SLAB_DEBUG_FREE;
3902 return length;
3903}
3904SLAB_ATTR(sanity_checks);
3905
3906static ssize_t trace_show(struct kmem_cache *s, char *buf)
3907{
3908 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3909}
3910
3911static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3912 size_t length)
3913{
3914 s->flags &= ~SLAB_TRACE;
3915 if (buf[0] == '1')
3916 s->flags |= SLAB_TRACE;
3917 return length;
3918}
3919SLAB_ATTR(trace);
3920
3921static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3922{
3923 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3924}
3925
3926static ssize_t reclaim_account_store(struct kmem_cache *s,
3927 const char *buf, size_t length)
3928{
3929 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3930 if (buf[0] == '1')
3931 s->flags |= SLAB_RECLAIM_ACCOUNT;
3932 return length;
3933}
3934SLAB_ATTR(reclaim_account);
3935
3936static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3937{
Christoph Lameter5af60832007-05-06 14:49:56 -07003938 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003939}
3940SLAB_ATTR_RO(hwcache_align);
3941
3942#ifdef CONFIG_ZONE_DMA
3943static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3944{
3945 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3946}
3947SLAB_ATTR_RO(cache_dma);
3948#endif
3949
3950static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3951{
3952 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3953}
3954SLAB_ATTR_RO(destroy_by_rcu);
3955
3956static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3957{
3958 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3959}
3960
3961static ssize_t red_zone_store(struct kmem_cache *s,
3962 const char *buf, size_t length)
3963{
3964 if (any_slab_objects(s))
3965 return -EBUSY;
3966
3967 s->flags &= ~SLAB_RED_ZONE;
3968 if (buf[0] == '1')
3969 s->flags |= SLAB_RED_ZONE;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003970 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003971 return length;
3972}
3973SLAB_ATTR(red_zone);
3974
3975static ssize_t poison_show(struct kmem_cache *s, char *buf)
3976{
3977 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3978}
3979
3980static ssize_t poison_store(struct kmem_cache *s,
3981 const char *buf, size_t length)
3982{
3983 if (any_slab_objects(s))
3984 return -EBUSY;
3985
3986 s->flags &= ~SLAB_POISON;
3987 if (buf[0] == '1')
3988 s->flags |= SLAB_POISON;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003989 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003990 return length;
3991}
3992SLAB_ATTR(poison);
3993
3994static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3995{
3996 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3997}
3998
3999static ssize_t store_user_store(struct kmem_cache *s,
4000 const char *buf, size_t length)
4001{
4002 if (any_slab_objects(s))
4003 return -EBUSY;
4004
4005 s->flags &= ~SLAB_STORE_USER;
4006 if (buf[0] == '1')
4007 s->flags |= SLAB_STORE_USER;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004008 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004009 return length;
4010}
4011SLAB_ATTR(store_user);
4012
Christoph Lameter53e15af2007-05-06 14:49:43 -07004013static ssize_t validate_show(struct kmem_cache *s, char *buf)
4014{
4015 return 0;
4016}
4017
4018static ssize_t validate_store(struct kmem_cache *s,
4019 const char *buf, size_t length)
4020{
Christoph Lameter434e2452007-07-17 04:03:30 -07004021 int ret = -EINVAL;
4022
4023 if (buf[0] == '1') {
4024 ret = validate_slab_cache(s);
4025 if (ret >= 0)
4026 ret = length;
4027 }
4028 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004029}
4030SLAB_ATTR(validate);
4031
Christoph Lameter2086d262007-05-06 14:49:46 -07004032static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4033{
4034 return 0;
4035}
4036
4037static ssize_t shrink_store(struct kmem_cache *s,
4038 const char *buf, size_t length)
4039{
4040 if (buf[0] == '1') {
4041 int rc = kmem_cache_shrink(s);
4042
4043 if (rc)
4044 return rc;
4045 } else
4046 return -EINVAL;
4047 return length;
4048}
4049SLAB_ATTR(shrink);
4050
Christoph Lameter88a420e2007-05-06 14:49:45 -07004051static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4052{
4053 if (!(s->flags & SLAB_STORE_USER))
4054 return -ENOSYS;
4055 return list_locations(s, buf, TRACK_ALLOC);
4056}
4057SLAB_ATTR_RO(alloc_calls);
4058
4059static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4060{
4061 if (!(s->flags & SLAB_STORE_USER))
4062 return -ENOSYS;
4063 return list_locations(s, buf, TRACK_FREE);
4064}
4065SLAB_ATTR_RO(free_calls);
4066
Christoph Lameter81819f02007-05-06 14:49:36 -07004067#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004068static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004069{
Christoph Lameter98246012008-01-07 23:20:26 -08004070 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004071}
4072
Christoph Lameter98246012008-01-07 23:20:26 -08004073static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004074 const char *buf, size_t length)
4075{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004076 unsigned long ratio;
4077 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004078
Christoph Lameter0121c6192008-04-29 16:11:12 -07004079 err = strict_strtoul(buf, 10, &ratio);
4080 if (err)
4081 return err;
4082
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004083 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004084 s->remote_node_defrag_ratio = ratio * 10;
4085
Christoph Lameter81819f02007-05-06 14:49:36 -07004086 return length;
4087}
Christoph Lameter98246012008-01-07 23:20:26 -08004088SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004089#endif
4090
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004091#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004092static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4093{
4094 unsigned long sum = 0;
4095 int cpu;
4096 int len;
4097 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4098
4099 if (!data)
4100 return -ENOMEM;
4101
4102 for_each_online_cpu(cpu) {
4103 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4104
4105 data[cpu] = x;
4106 sum += x;
4107 }
4108
4109 len = sprintf(buf, "%lu", sum);
4110
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004111#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004112 for_each_online_cpu(cpu) {
4113 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004114 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004115 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004116#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004117 kfree(data);
4118 return len + sprintf(buf + len, "\n");
4119}
4120
4121#define STAT_ATTR(si, text) \
4122static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4123{ \
4124 return show_stat(s, buf, si); \
4125} \
4126SLAB_ATTR_RO(text); \
4127
4128STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4129STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4130STAT_ATTR(FREE_FASTPATH, free_fastpath);
4131STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4132STAT_ATTR(FREE_FROZEN, free_frozen);
4133STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4134STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4135STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4136STAT_ATTR(ALLOC_SLAB, alloc_slab);
4137STAT_ATTR(ALLOC_REFILL, alloc_refill);
4138STAT_ATTR(FREE_SLAB, free_slab);
4139STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4140STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4141STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4142STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4143STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4144STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter65c33762008-04-14 19:11:40 +03004145STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004146#endif
4147
Pekka Enberg06428782008-01-07 23:20:27 -08004148static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004149 &slab_size_attr.attr,
4150 &object_size_attr.attr,
4151 &objs_per_slab_attr.attr,
4152 &order_attr.attr,
4153 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004154 &objects_partial_attr.attr,
4155 &total_objects_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004156 &slabs_attr.attr,
4157 &partial_attr.attr,
4158 &cpu_slabs_attr.attr,
4159 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004160 &aliases_attr.attr,
4161 &align_attr.attr,
4162 &sanity_checks_attr.attr,
4163 &trace_attr.attr,
4164 &hwcache_align_attr.attr,
4165 &reclaim_account_attr.attr,
4166 &destroy_by_rcu_attr.attr,
4167 &red_zone_attr.attr,
4168 &poison_attr.attr,
4169 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004170 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004171 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004172 &alloc_calls_attr.attr,
4173 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004174#ifdef CONFIG_ZONE_DMA
4175 &cache_dma_attr.attr,
4176#endif
4177#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004178 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004179#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004180#ifdef CONFIG_SLUB_STATS
4181 &alloc_fastpath_attr.attr,
4182 &alloc_slowpath_attr.attr,
4183 &free_fastpath_attr.attr,
4184 &free_slowpath_attr.attr,
4185 &free_frozen_attr.attr,
4186 &free_add_partial_attr.attr,
4187 &free_remove_partial_attr.attr,
4188 &alloc_from_partial_attr.attr,
4189 &alloc_slab_attr.attr,
4190 &alloc_refill_attr.attr,
4191 &free_slab_attr.attr,
4192 &cpuslab_flush_attr.attr,
4193 &deactivate_full_attr.attr,
4194 &deactivate_empty_attr.attr,
4195 &deactivate_to_head_attr.attr,
4196 &deactivate_to_tail_attr.attr,
4197 &deactivate_remote_frees_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004198 &order_fallback_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004199#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004200 NULL
4201};
4202
4203static struct attribute_group slab_attr_group = {
4204 .attrs = slab_attrs,
4205};
4206
4207static ssize_t slab_attr_show(struct kobject *kobj,
4208 struct attribute *attr,
4209 char *buf)
4210{
4211 struct slab_attribute *attribute;
4212 struct kmem_cache *s;
4213 int err;
4214
4215 attribute = to_slab_attr(attr);
4216 s = to_slab(kobj);
4217
4218 if (!attribute->show)
4219 return -EIO;
4220
4221 err = attribute->show(s, buf);
4222
4223 return err;
4224}
4225
4226static ssize_t slab_attr_store(struct kobject *kobj,
4227 struct attribute *attr,
4228 const char *buf, size_t len)
4229{
4230 struct slab_attribute *attribute;
4231 struct kmem_cache *s;
4232 int err;
4233
4234 attribute = to_slab_attr(attr);
4235 s = to_slab(kobj);
4236
4237 if (!attribute->store)
4238 return -EIO;
4239
4240 err = attribute->store(s, buf, len);
4241
4242 return err;
4243}
4244
Christoph Lameter151c6022008-01-07 22:29:05 -08004245static void kmem_cache_release(struct kobject *kobj)
4246{
4247 struct kmem_cache *s = to_slab(kobj);
4248
4249 kfree(s);
4250}
4251
Christoph Lameter81819f02007-05-06 14:49:36 -07004252static struct sysfs_ops slab_sysfs_ops = {
4253 .show = slab_attr_show,
4254 .store = slab_attr_store,
4255};
4256
4257static struct kobj_type slab_ktype = {
4258 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004259 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004260};
4261
4262static int uevent_filter(struct kset *kset, struct kobject *kobj)
4263{
4264 struct kobj_type *ktype = get_ktype(kobj);
4265
4266 if (ktype == &slab_ktype)
4267 return 1;
4268 return 0;
4269}
4270
4271static struct kset_uevent_ops slab_uevent_ops = {
4272 .filter = uevent_filter,
4273};
4274
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004275static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004276
4277#define ID_STR_LENGTH 64
4278
4279/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004280 *
4281 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004282 */
4283static char *create_unique_id(struct kmem_cache *s)
4284{
4285 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4286 char *p = name;
4287
4288 BUG_ON(!name);
4289
4290 *p++ = ':';
4291 /*
4292 * First flags affecting slabcache operations. We will only
4293 * get here for aliasable slabs so we do not need to support
4294 * too many flags. The flags here must cover all flags that
4295 * are matched during merging to guarantee that the id is
4296 * unique.
4297 */
4298 if (s->flags & SLAB_CACHE_DMA)
4299 *p++ = 'd';
4300 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4301 *p++ = 'a';
4302 if (s->flags & SLAB_DEBUG_FREE)
4303 *p++ = 'F';
4304 if (p != name + 1)
4305 *p++ = '-';
4306 p += sprintf(p, "%07d", s->size);
4307 BUG_ON(p > name + ID_STR_LENGTH - 1);
4308 return name;
4309}
4310
4311static int sysfs_slab_add(struct kmem_cache *s)
4312{
4313 int err;
4314 const char *name;
4315 int unmergeable;
4316
4317 if (slab_state < SYSFS)
4318 /* Defer until later */
4319 return 0;
4320
4321 unmergeable = slab_unmergeable(s);
4322 if (unmergeable) {
4323 /*
4324 * Slabcache can never be merged so we can use the name proper.
4325 * This is typically the case for debug situations. In that
4326 * case we can catch duplicate names easily.
4327 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004328 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004329 name = s->name;
4330 } else {
4331 /*
4332 * Create a unique name for the slab as a target
4333 * for the symlinks.
4334 */
4335 name = create_unique_id(s);
4336 }
4337
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004338 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004339 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4340 if (err) {
4341 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004342 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004343 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004344
4345 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4346 if (err)
4347 return err;
4348 kobject_uevent(&s->kobj, KOBJ_ADD);
4349 if (!unmergeable) {
4350 /* Setup first alias */
4351 sysfs_slab_alias(s, s->name);
4352 kfree(name);
4353 }
4354 return 0;
4355}
4356
4357static void sysfs_slab_remove(struct kmem_cache *s)
4358{
4359 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4360 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004361 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004362}
4363
4364/*
4365 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11004366 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07004367 */
4368struct saved_alias {
4369 struct kmem_cache *s;
4370 const char *name;
4371 struct saved_alias *next;
4372};
4373
Adrian Bunk5af328a2007-07-17 04:03:27 -07004374static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004375
4376static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4377{
4378 struct saved_alias *al;
4379
4380 if (slab_state == SYSFS) {
4381 /*
4382 * If we have a leftover link then remove it.
4383 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004384 sysfs_remove_link(&slab_kset->kobj, name);
4385 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004386 }
4387
4388 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4389 if (!al)
4390 return -ENOMEM;
4391
4392 al->s = s;
4393 al->name = name;
4394 al->next = alias_list;
4395 alias_list = al;
4396 return 0;
4397}
4398
4399static int __init slab_sysfs_init(void)
4400{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004401 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004402 int err;
4403
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004404 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004405 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004406 printk(KERN_ERR "Cannot register slab subsystem.\n");
4407 return -ENOSYS;
4408 }
4409
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004410 slab_state = SYSFS;
4411
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004412 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004413 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004414 if (err)
4415 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4416 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004417 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004418
4419 while (alias_list) {
4420 struct saved_alias *al = alias_list;
4421
4422 alias_list = alias_list->next;
4423 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004424 if (err)
4425 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4426 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004427 kfree(al);
4428 }
4429
4430 resiliency_test();
4431 return 0;
4432}
4433
4434__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004435#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004436
4437/*
4438 * The /proc/slabinfo ABI
4439 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004440#ifdef CONFIG_SLABINFO
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004441static void print_slabinfo_header(struct seq_file *m)
4442{
4443 seq_puts(m, "slabinfo - version: 2.1\n");
4444 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4445 "<objperslab> <pagesperslab>");
4446 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4447 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4448 seq_putc(m, '\n');
4449}
4450
4451static void *s_start(struct seq_file *m, loff_t *pos)
4452{
4453 loff_t n = *pos;
4454
4455 down_read(&slub_lock);
4456 if (!n)
4457 print_slabinfo_header(m);
4458
4459 return seq_list_start(&slab_caches, *pos);
4460}
4461
4462static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4463{
4464 return seq_list_next(p, &slab_caches, pos);
4465}
4466
4467static void s_stop(struct seq_file *m, void *p)
4468{
4469 up_read(&slub_lock);
4470}
4471
4472static int s_show(struct seq_file *m, void *p)
4473{
4474 unsigned long nr_partials = 0;
4475 unsigned long nr_slabs = 0;
4476 unsigned long nr_inuse = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004477 unsigned long nr_objs = 0;
4478 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004479 struct kmem_cache *s;
4480 int node;
4481
4482 s = list_entry(p, struct kmem_cache, list);
4483
4484 for_each_online_node(node) {
4485 struct kmem_cache_node *n = get_node(s, node);
4486
4487 if (!n)
4488 continue;
4489
4490 nr_partials += n->nr_partial;
4491 nr_slabs += atomic_long_read(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03004492 nr_objs += atomic_long_read(&n->total_objects);
4493 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004494 }
4495
Christoph Lameter205ab992008-04-14 19:11:40 +03004496 nr_inuse = nr_objs - nr_free;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004497
4498 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
Christoph Lameter834f3d12008-04-14 19:11:31 +03004499 nr_objs, s->size, oo_objects(s->oo),
4500 (1 << oo_order(s->oo)));
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004501 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4502 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4503 0UL);
4504 seq_putc(m, '\n');
4505 return 0;
4506}
4507
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004508static const struct seq_operations slabinfo_op = {
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004509 .start = s_start,
4510 .next = s_next,
4511 .stop = s_stop,
4512 .show = s_show,
4513};
4514
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004515static int slabinfo_open(struct inode *inode, struct file *file)
4516{
4517 return seq_open(file, &slabinfo_op);
4518}
4519
4520static const struct file_operations proc_slabinfo_operations = {
4521 .open = slabinfo_open,
4522 .read = seq_read,
4523 .llseek = seq_lseek,
4524 .release = seq_release,
4525};
4526
4527static int __init slab_proc_init(void)
4528{
4529 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4530 return 0;
4531}
4532module_init(slab_proc_init);
Linus Torvalds158a9622008-01-02 13:04:48 -08004533#endif /* CONFIG_SLABINFO */