blob: c4ea9158c9fbd0e4630062aa0098406000be45c3 [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{
Akinobu Mita1a00df42009-03-07 00:36:21 +0900377 struct track *p = get_track(s, object, alloc);
Christoph Lameter81819f02007-05-06 14:49:36 -0700378
Christoph Lameter81819f02007-05-06 14:49:36 -0700379 if (addr) {
380 p->addr = addr;
381 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400382 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700383 p->when = jiffies;
384 } else
385 memset(p, 0, sizeof(struct track));
386}
387
Christoph Lameter81819f02007-05-06 14:49:36 -0700388static void init_tracking(struct kmem_cache *s, void *object)
389{
Christoph Lameter24922682007-07-17 04:03:18 -0700390 if (!(s->flags & SLAB_STORE_USER))
391 return;
392
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300393 set_track(s, object, TRACK_FREE, 0UL);
394 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700395}
396
397static void print_track(const char *s, struct track *t)
398{
399 if (!t->addr)
400 return;
401
Linus Torvalds7daf7052008-07-14 12:12:53 -0700402 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300403 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700404}
405
Christoph Lameter24922682007-07-17 04:03:18 -0700406static void print_tracking(struct kmem_cache *s, void *object)
407{
408 if (!(s->flags & SLAB_STORE_USER))
409 return;
410
411 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
412 print_track("Freed", get_track(s, object, TRACK_FREE));
413}
414
415static void print_page_info(struct page *page)
416{
Christoph Lameter39b26462008-04-14 19:11:30 +0300417 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
418 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700419
420}
421
422static void slab_bug(struct kmem_cache *s, char *fmt, ...)
423{
424 va_list args;
425 char buf[100];
426
427 va_start(args, fmt);
428 vsnprintf(buf, sizeof(buf), fmt, args);
429 va_end(args);
430 printk(KERN_ERR "========================================"
431 "=====================================\n");
432 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
433 printk(KERN_ERR "----------------------------------------"
434 "-------------------------------------\n\n");
435}
436
437static void slab_fix(struct kmem_cache *s, char *fmt, ...)
438{
439 va_list args;
440 char buf[100];
441
442 va_start(args, fmt);
443 vsnprintf(buf, sizeof(buf), fmt, args);
444 va_end(args);
445 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
446}
447
448static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700449{
450 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800451 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700452
453 print_tracking(s, p);
454
455 print_page_info(page);
456
457 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
458 p, p - addr, get_freepointer(s, p));
459
460 if (p > addr + 16)
461 print_section("Bytes b4", p - 16, 16);
462
Pekka Enberg0ebd6522008-07-19 14:17:22 +0300463 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700464
465 if (s->flags & SLAB_RED_ZONE)
466 print_section("Redzone", p + s->objsize,
467 s->inuse - s->objsize);
468
Christoph Lameter81819f02007-05-06 14:49:36 -0700469 if (s->offset)
470 off = s->offset + sizeof(void *);
471 else
472 off = s->inuse;
473
Christoph Lameter24922682007-07-17 04:03:18 -0700474 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700475 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700476
477 if (off != s->size)
478 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700479 print_section("Padding", p + off, s->size - off);
480
481 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700482}
483
484static void object_err(struct kmem_cache *s, struct page *page,
485 u8 *object, char *reason)
486{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700487 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700488 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700489}
490
Christoph Lameter24922682007-07-17 04:03:18 -0700491static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700492{
493 va_list args;
494 char buf[100];
495
Christoph Lameter24922682007-07-17 04:03:18 -0700496 va_start(args, fmt);
497 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700498 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700499 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700500 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700501 dump_stack();
502}
503
504static void init_object(struct kmem_cache *s, void *object, int active)
505{
506 u8 *p = object;
507
508 if (s->flags & __OBJECT_POISON) {
509 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800510 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700511 }
512
513 if (s->flags & SLAB_RED_ZONE)
514 memset(p + s->objsize,
515 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
516 s->inuse - s->objsize);
517}
518
Christoph Lameter24922682007-07-17 04:03:18 -0700519static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700520{
521 while (bytes) {
522 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700523 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700524 start++;
525 bytes--;
526 }
Christoph Lameter24922682007-07-17 04:03:18 -0700527 return NULL;
528}
529
530static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
531 void *from, void *to)
532{
533 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
534 memset(from, data, to - from);
535}
536
537static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
538 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800539 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700540{
541 u8 *fault;
542 u8 *end;
543
544 fault = check_bytes(start, value, bytes);
545 if (!fault)
546 return 1;
547
548 end = start + bytes;
549 while (end > fault && end[-1] == value)
550 end--;
551
552 slab_bug(s, "%s overwritten", what);
553 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
554 fault, end - 1, fault[0], value);
555 print_trailer(s, page, object);
556
557 restore_bytes(s, what, value, fault, end);
558 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700559}
560
Christoph Lameter81819f02007-05-06 14:49:36 -0700561/*
562 * Object layout:
563 *
564 * object address
565 * Bytes of the object to be managed.
566 * If the freepointer may overlay the object then the free
567 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700568 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700569 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
570 * 0xa5 (POISON_END)
571 *
572 * object + s->objsize
573 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700574 * Padding is extended by another word if Redzoning is enabled and
575 * objsize == inuse.
576 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700577 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
578 * 0xcc (RED_ACTIVE) for objects in use.
579 *
580 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700581 * Meta data starts here.
582 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700583 * A. Free pointer (if we cannot overwrite object on free)
584 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700585 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800586 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700587 * before the word boundary.
588 *
589 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700590 *
591 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700592 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700593 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700594 * If slabcaches are merged then the objsize and inuse boundaries are mostly
595 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700596 * may be used with merged slabcaches.
597 */
598
Christoph Lameter81819f02007-05-06 14:49:36 -0700599static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
600{
601 unsigned long off = s->inuse; /* The end of info */
602
603 if (s->offset)
604 /* Freepointer is placed after the object. */
605 off += sizeof(void *);
606
607 if (s->flags & SLAB_STORE_USER)
608 /* We also have user information there */
609 off += 2 * sizeof(struct track);
610
611 if (s->size == off)
612 return 1;
613
Christoph Lameter24922682007-07-17 04:03:18 -0700614 return check_bytes_and_report(s, page, p, "Object padding",
615 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700616}
617
Christoph Lameter39b26462008-04-14 19:11:30 +0300618/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700619static int slab_pad_check(struct kmem_cache *s, struct page *page)
620{
Christoph Lameter24922682007-07-17 04:03:18 -0700621 u8 *start;
622 u8 *fault;
623 u8 *end;
624 int length;
625 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700626
627 if (!(s->flags & SLAB_POISON))
628 return 1;
629
Christoph Lametera973e9d2008-03-01 13:40:44 -0800630 start = page_address(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +0300631 length = (PAGE_SIZE << compound_order(page));
Christoph Lameter39b26462008-04-14 19:11:30 +0300632 end = start + length;
633 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700634 if (!remainder)
635 return 1;
636
Christoph Lameter39b26462008-04-14 19:11:30 +0300637 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700638 if (!fault)
639 return 1;
640 while (end > fault && end[-1] == POISON_INUSE)
641 end--;
642
643 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Christoph Lameter39b26462008-04-14 19:11:30 +0300644 print_section("Padding", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700645
646 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
647 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700648}
649
650static int check_object(struct kmem_cache *s, struct page *page,
651 void *object, int active)
652{
653 u8 *p = object;
654 u8 *endobject = object + s->objsize;
655
656 if (s->flags & SLAB_RED_ZONE) {
657 unsigned int red =
658 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
659
Christoph Lameter24922682007-07-17 04:03:18 -0700660 if (!check_bytes_and_report(s, page, object, "Redzone",
661 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700662 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700663 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800664 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
665 check_bytes_and_report(s, page, p, "Alignment padding",
666 endobject, POISON_INUSE, s->inuse - s->objsize);
667 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700668 }
669
670 if (s->flags & SLAB_POISON) {
671 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700672 (!check_bytes_and_report(s, page, p, "Poison", p,
673 POISON_FREE, s->objsize - 1) ||
674 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800675 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700676 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700677 /*
678 * check_pad_bytes cleans up on its own.
679 */
680 check_pad_bytes(s, page, p);
681 }
682
683 if (!s->offset && active)
684 /*
685 * Object and freepointer overlap. Cannot check
686 * freepointer while object is allocated.
687 */
688 return 1;
689
690 /* Check free pointer validity */
691 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
692 object_err(s, page, p, "Freepointer corrupt");
693 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100694 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700695 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700696 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700697 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800698 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700699 return 0;
700 }
701 return 1;
702}
703
704static int check_slab(struct kmem_cache *s, struct page *page)
705{
Christoph Lameter39b26462008-04-14 19:11:30 +0300706 int maxobj;
707
Christoph Lameter81819f02007-05-06 14:49:36 -0700708 VM_BUG_ON(!irqs_disabled());
709
710 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700711 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700712 return 0;
713 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300714
715 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
716 if (page->objects > maxobj) {
717 slab_err(s, page, "objects %u > max %u",
718 s->name, page->objects, maxobj);
719 return 0;
720 }
721 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700722 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300723 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700724 return 0;
725 }
726 /* Slab_pad_check fixes things up after itself */
727 slab_pad_check(s, page);
728 return 1;
729}
730
731/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700732 * Determine if a certain object on a page is on the freelist. Must hold the
733 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700734 */
735static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
736{
737 int nr = 0;
738 void *fp = page->freelist;
739 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300740 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700741
Christoph Lameter39b26462008-04-14 19:11:30 +0300742 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700743 if (fp == search)
744 return 1;
745 if (!check_valid_pointer(s, page, fp)) {
746 if (object) {
747 object_err(s, page, object,
748 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800749 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700750 break;
751 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700752 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800753 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300754 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700755 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700756 return 0;
757 }
758 break;
759 }
760 object = fp;
761 fp = get_freepointer(s, object);
762 nr++;
763 }
764
Christoph Lameter224a88b2008-04-14 19:11:31 +0300765 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400766 if (max_objects > MAX_OBJS_PER_PAGE)
767 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300768
769 if (page->objects != max_objects) {
770 slab_err(s, page, "Wrong number of objects. Found %d but "
771 "should be %d", page->objects, max_objects);
772 page->objects = max_objects;
773 slab_fix(s, "Number of objects adjusted.");
774 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300775 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700776 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300777 "counted were %d", page->inuse, page->objects - nr);
778 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700779 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700780 }
781 return search == NULL;
782}
783
Christoph Lameter0121c6192008-04-29 16:11:12 -0700784static void trace(struct kmem_cache *s, struct page *page, void *object,
785 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700786{
787 if (s->flags & SLAB_TRACE) {
788 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
789 s->name,
790 alloc ? "alloc" : "free",
791 object, page->inuse,
792 page->freelist);
793
794 if (!alloc)
795 print_section("Object", (void *)object, s->objsize);
796
797 dump_stack();
798 }
799}
800
Christoph Lameter643b1132007-05-06 14:49:42 -0700801/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700802 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700803 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700804static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700805{
Christoph Lameter643b1132007-05-06 14:49:42 -0700806 spin_lock(&n->list_lock);
807 list_add(&page->lru, &n->full);
808 spin_unlock(&n->list_lock);
809}
810
811static void remove_full(struct kmem_cache *s, struct page *page)
812{
813 struct kmem_cache_node *n;
814
815 if (!(s->flags & SLAB_STORE_USER))
816 return;
817
818 n = get_node(s, page_to_nid(page));
819
820 spin_lock(&n->list_lock);
821 list_del(&page->lru);
822 spin_unlock(&n->list_lock);
823}
824
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300825/* Tracking of the number of slabs for debugging purposes */
826static inline unsigned long slabs_node(struct kmem_cache *s, int node)
827{
828 struct kmem_cache_node *n = get_node(s, node);
829
830 return atomic_long_read(&n->nr_slabs);
831}
832
Christoph Lameter205ab992008-04-14 19:11:40 +0300833static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300834{
835 struct kmem_cache_node *n = get_node(s, node);
836
837 /*
838 * May be called early in order to allocate a slab for the
839 * kmem_cache_node structure. Solve the chicken-egg
840 * dilemma by deferring the increment of the count during
841 * bootstrap (see early_kmem_cache_node_alloc).
842 */
Christoph Lameter205ab992008-04-14 19:11:40 +0300843 if (!NUMA_BUILD || n) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300844 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300845 atomic_long_add(objects, &n->total_objects);
846 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300847}
Christoph Lameter205ab992008-04-14 19:11:40 +0300848static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300849{
850 struct kmem_cache_node *n = get_node(s, node);
851
852 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300853 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300854}
855
856/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -0700857static void setup_object_debug(struct kmem_cache *s, struct page *page,
858 void *object)
859{
860 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
861 return;
862
863 init_object(s, object, 0);
864 init_tracking(s, object);
865}
866
867static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300868 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700869{
870 if (!check_slab(s, page))
871 goto bad;
872
Christoph Lameterd692ef62008-02-15 23:45:24 -0800873 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700874 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700875 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700876 }
877
878 if (!check_valid_pointer(s, page, object)) {
879 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700880 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700881 }
882
Christoph Lameterd692ef62008-02-15 23:45:24 -0800883 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700884 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700885
Christoph Lameter3ec09742007-05-16 22:11:00 -0700886 /* Success perform special debug activities for allocs */
887 if (s->flags & SLAB_STORE_USER)
888 set_track(s, object, TRACK_ALLOC, addr);
889 trace(s, page, object, 1);
890 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700891 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700892
Christoph Lameter81819f02007-05-06 14:49:36 -0700893bad:
894 if (PageSlab(page)) {
895 /*
896 * If this is a slab page then lets do the best we can
897 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700898 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700899 */
Christoph Lameter24922682007-07-17 04:03:18 -0700900 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +0300901 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800902 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700903 }
904 return 0;
905}
906
Christoph Lameter3ec09742007-05-16 22:11:00 -0700907static int free_debug_processing(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300908 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700909{
910 if (!check_slab(s, page))
911 goto fail;
912
913 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700914 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700915 goto fail;
916 }
917
918 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700919 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700920 goto fail;
921 }
922
923 if (!check_object(s, page, object, 1))
924 return 0;
925
926 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800927 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700928 slab_err(s, page, "Attempt to free object(0x%p) "
929 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800930 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700931 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700932 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700933 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700934 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800935 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700936 object_err(s, page, object,
937 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700938 goto fail;
939 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700940
941 /* Special debug activities for freeing objects */
Andy Whitcroft8a380822008-07-23 21:27:18 -0700942 if (!PageSlubFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700943 remove_full(s, page);
944 if (s->flags & SLAB_STORE_USER)
945 set_track(s, object, TRACK_FREE, addr);
946 trace(s, page, object, 0);
947 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700948 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700949
Christoph Lameter81819f02007-05-06 14:49:36 -0700950fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700951 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700952 return 0;
953}
954
Christoph Lameter41ecc552007-05-09 02:32:44 -0700955static int __init setup_slub_debug(char *str)
956{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700957 slub_debug = DEBUG_DEFAULT_FLAGS;
958 if (*str++ != '=' || !*str)
959 /*
960 * No options specified. Switch on full debugging.
961 */
962 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700963
964 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700965 /*
966 * No options but restriction on slabs. This means full
967 * debugging for slabs matching a pattern.
968 */
969 goto check_slabs;
970
971 slub_debug = 0;
972 if (*str == '-')
973 /*
974 * Switch off all debugging measures.
975 */
976 goto out;
977
978 /*
979 * Determine which debug features should be switched on
980 */
Pekka Enberg06428782008-01-07 23:20:27 -0800981 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700982 switch (tolower(*str)) {
983 case 'f':
984 slub_debug |= SLAB_DEBUG_FREE;
985 break;
986 case 'z':
987 slub_debug |= SLAB_RED_ZONE;
988 break;
989 case 'p':
990 slub_debug |= SLAB_POISON;
991 break;
992 case 'u':
993 slub_debug |= SLAB_STORE_USER;
994 break;
995 case 't':
996 slub_debug |= SLAB_TRACE;
997 break;
998 default:
999 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001000 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001001 }
1002 }
1003
1004check_slabs:
1005 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001006 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001007out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001008 return 1;
1009}
1010
1011__setup("slub_debug", setup_slub_debug);
1012
Christoph Lameterba0268a2007-09-11 15:24:11 -07001013static unsigned long kmem_cache_flags(unsigned long objsize,
1014 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001015 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001016{
1017 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001018 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001019 */
Christoph Lametere1533622008-02-15 23:45:24 -08001020 if (slub_debug && (!slub_debug_slabs ||
1021 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1022 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001023
1024 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001025}
1026#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001027static inline void setup_object_debug(struct kmem_cache *s,
1028 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001029
Christoph Lameter3ec09742007-05-16 22:11:00 -07001030static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001031 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001032
Christoph Lameter3ec09742007-05-16 22:11:00 -07001033static inline int free_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001034 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001035
Christoph Lameter41ecc552007-05-09 02:32:44 -07001036static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1037 { return 1; }
1038static inline int check_object(struct kmem_cache *s, struct page *page,
1039 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001040static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001041static inline unsigned long kmem_cache_flags(unsigned long objsize,
1042 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001043 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001044{
1045 return flags;
1046}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001047#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001048
1049static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1050 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001051static inline void inc_slabs_node(struct kmem_cache *s, int node,
1052 int objects) {}
1053static inline void dec_slabs_node(struct kmem_cache *s, int node,
1054 int objects) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001055#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03001056
Christoph Lameter81819f02007-05-06 14:49:36 -07001057/*
1058 * Slab allocation and freeing
1059 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001060static inline struct page *alloc_slab_page(gfp_t flags, int node,
1061 struct kmem_cache_order_objects oo)
1062{
1063 int order = oo_order(oo);
1064
1065 if (node == -1)
1066 return alloc_pages(flags, order);
1067 else
1068 return alloc_pages_node(node, flags, order);
1069}
1070
Christoph Lameter81819f02007-05-06 14:49:36 -07001071static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1072{
Pekka Enberg06428782008-01-07 23:20:27 -08001073 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001074 struct kmem_cache_order_objects oo = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07001075
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001076 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001077
Christoph Lameter65c33762008-04-14 19:11:40 +03001078 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1079 oo);
1080 if (unlikely(!page)) {
1081 oo = s->min;
1082 /*
1083 * Allocation may have failed due to fragmentation.
1084 * Try a lower order alloc if possible
1085 */
1086 page = alloc_slab_page(flags, node, oo);
1087 if (!page)
1088 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001089
Christoph Lameter65c33762008-04-14 19:11:40 +03001090 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1091 }
Christoph Lameter834f3d12008-04-14 19:11:31 +03001092 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001093 mod_zone_page_state(page_zone(page),
1094 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1095 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001096 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001097
1098 return page;
1099}
1100
1101static void setup_object(struct kmem_cache *s, struct page *page,
1102 void *object)
1103{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001104 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001105 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001106 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001107}
1108
1109static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1110{
1111 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001112 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001113 void *last;
1114 void *p;
1115
Christoph Lameter6cb06222007-10-16 01:25:41 -07001116 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001117
Christoph Lameter6cb06222007-10-16 01:25:41 -07001118 page = allocate_slab(s,
1119 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001120 if (!page)
1121 goto out;
1122
Christoph Lameter205ab992008-04-14 19:11:40 +03001123 inc_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001124 page->slab = s;
1125 page->flags |= 1 << PG_slab;
1126 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1127 SLAB_STORE_USER | SLAB_TRACE))
Andy Whitcroft8a380822008-07-23 21:27:18 -07001128 __SetPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001129
1130 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001131
1132 if (unlikely(s->flags & SLAB_POISON))
Christoph Lameter834f3d12008-04-14 19:11:31 +03001133 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001134
1135 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001136 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001137 setup_object(s, page, last);
1138 set_freepointer(s, last, p);
1139 last = p;
1140 }
1141 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001142 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001143
1144 page->freelist = start;
1145 page->inuse = 0;
1146out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001147 return page;
1148}
1149
1150static void __free_slab(struct kmem_cache *s, struct page *page)
1151{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001152 int order = compound_order(page);
1153 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001154
Andy Whitcroft8a380822008-07-23 21:27:18 -07001155 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001156 void *p;
1157
1158 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001159 for_each_object(p, s, page_address(page),
1160 page->objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001161 check_object(s, page, p, 0);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001162 __ClearPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001163 }
1164
1165 mod_zone_page_state(page_zone(page),
1166 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1167 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001168 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001169
Christoph Lameter49bd5222008-04-14 18:52:18 +03001170 __ClearPageSlab(page);
1171 reset_page_mapcount(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +03001172 __free_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001173}
1174
1175static void rcu_free_slab(struct rcu_head *h)
1176{
1177 struct page *page;
1178
1179 page = container_of((struct list_head *)h, struct page, lru);
1180 __free_slab(page->slab, page);
1181}
1182
1183static void free_slab(struct kmem_cache *s, struct page *page)
1184{
1185 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1186 /*
1187 * RCU free overloads the RCU head over the LRU
1188 */
1189 struct rcu_head *head = (void *)&page->lru;
1190
1191 call_rcu(head, rcu_free_slab);
1192 } else
1193 __free_slab(s, page);
1194}
1195
1196static void discard_slab(struct kmem_cache *s, struct page *page)
1197{
Christoph Lameter205ab992008-04-14 19:11:40 +03001198 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001199 free_slab(s, page);
1200}
1201
1202/*
1203 * Per slab locking using the pagelock
1204 */
1205static __always_inline void slab_lock(struct page *page)
1206{
1207 bit_spin_lock(PG_locked, &page->flags);
1208}
1209
1210static __always_inline void slab_unlock(struct page *page)
1211{
Nick Piggina76d3542008-01-07 23:20:27 -08001212 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001213}
1214
1215static __always_inline int slab_trylock(struct page *page)
1216{
1217 int rc = 1;
1218
1219 rc = bit_spin_trylock(PG_locked, &page->flags);
1220 return rc;
1221}
1222
1223/*
1224 * Management of partially allocated slabs
1225 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001226static void add_partial(struct kmem_cache_node *n,
1227 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001228{
Christoph Lametere95eed52007-05-06 14:49:44 -07001229 spin_lock(&n->list_lock);
1230 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001231 if (tail)
1232 list_add_tail(&page->lru, &n->partial);
1233 else
1234 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001235 spin_unlock(&n->list_lock);
1236}
1237
Christoph Lameter0121c6192008-04-29 16:11:12 -07001238static void remove_partial(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001239{
1240 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1241
1242 spin_lock(&n->list_lock);
1243 list_del(&page->lru);
1244 n->nr_partial--;
1245 spin_unlock(&n->list_lock);
1246}
1247
1248/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001249 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001250 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001251 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001252 */
Christoph Lameter0121c6192008-04-29 16:11:12 -07001253static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1254 struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001255{
1256 if (slab_trylock(page)) {
1257 list_del(&page->lru);
1258 n->nr_partial--;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001259 __SetPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001260 return 1;
1261 }
1262 return 0;
1263}
1264
1265/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001266 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001267 */
1268static struct page *get_partial_node(struct kmem_cache_node *n)
1269{
1270 struct page *page;
1271
1272 /*
1273 * Racy check. If we mistakenly see no partial slabs then we
1274 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001275 * partial slab and there is none available then get_partials()
1276 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001277 */
1278 if (!n || !n->nr_partial)
1279 return NULL;
1280
1281 spin_lock(&n->list_lock);
1282 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001283 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001284 goto out;
1285 page = NULL;
1286out:
1287 spin_unlock(&n->list_lock);
1288 return page;
1289}
1290
1291/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001292 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001293 */
1294static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1295{
1296#ifdef CONFIG_NUMA
1297 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001298 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001299 struct zone *zone;
1300 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001301 struct page *page;
1302
1303 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001304 * The defrag ratio allows a configuration of the tradeoffs between
1305 * inter node defragmentation and node local allocations. A lower
1306 * defrag_ratio increases the tendency to do local allocations
1307 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001308 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001309 * If the defrag_ratio is set to 0 then kmalloc() always
1310 * returns node local objects. If the ratio is higher then kmalloc()
1311 * may return off node objects because partial slabs are obtained
1312 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001313 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001314 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001315 * defrag_ratio = 1000) then every (well almost) allocation will
1316 * first attempt to defrag slab caches on other nodes. This means
1317 * scanning over all nodes to look for partial slabs which may be
1318 * expensive if we do it every time we are trying to find a slab
1319 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001320 */
Christoph Lameter98246012008-01-07 23:20:26 -08001321 if (!s->remote_node_defrag_ratio ||
1322 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001323 return NULL;
1324
Mel Gorman0e884602008-04-28 02:12:14 -07001325 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001326 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001327 struct kmem_cache_node *n;
1328
Mel Gorman54a6eb52008-04-28 02:12:16 -07001329 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001330
Mel Gorman54a6eb52008-04-28 02:12:16 -07001331 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
David Rientjes3b89d7d2009-02-22 17:40:07 -08001332 n->nr_partial > s->min_partial) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001333 page = get_partial_node(n);
1334 if (page)
1335 return page;
1336 }
1337 }
1338#endif
1339 return NULL;
1340}
1341
1342/*
1343 * Get a partial page, lock it and return it.
1344 */
1345static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1346{
1347 struct page *page;
1348 int searchnode = (node == -1) ? numa_node_id() : node;
1349
1350 page = get_partial_node(get_node(s, searchnode));
1351 if (page || (flags & __GFP_THISNODE))
1352 return page;
1353
1354 return get_any_partial(s, flags);
1355}
1356
1357/*
1358 * Move a page back to the lists.
1359 *
1360 * Must be called with the slab lock held.
1361 *
1362 * On exit the slab lock will have been dropped.
1363 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001364static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001365{
Christoph Lametere95eed52007-05-06 14:49:44 -07001366 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001367 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001368
Andy Whitcroft8a380822008-07-23 21:27:18 -07001369 __ClearPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001370 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001371
Christoph Lametera973e9d2008-03-01 13:40:44 -08001372 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001373 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001374 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1375 } else {
1376 stat(c, DEACTIVATE_FULL);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001377 if (SLABDEBUG && PageSlubDebug(page) &&
1378 (s->flags & SLAB_STORE_USER))
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001379 add_full(n, page);
1380 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001381 slab_unlock(page);
1382 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001383 stat(c, DEACTIVATE_EMPTY);
David Rientjes3b89d7d2009-02-22 17:40:07 -08001384 if (n->nr_partial < s->min_partial) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001385 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001386 * Adding an empty slab to the partial slabs in order
1387 * to avoid page allocator overhead. This slab needs
1388 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001389 * so that the others get filled first. That way the
1390 * size of the partial list stays small.
1391 *
Christoph Lameter0121c6192008-04-29 16:11:12 -07001392 * kmem_cache_shrink can reclaim any empty slabs from
1393 * the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001394 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001395 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001396 slab_unlock(page);
1397 } else {
1398 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001399 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001400 discard_slab(s, page);
1401 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001402 }
1403}
1404
1405/*
1406 * Remove the cpu slab
1407 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001408static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001409{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001410 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001411 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001412
Christoph Lameterb773ad72008-03-04 11:10:17 -08001413 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001414 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001415 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001416 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001417 * because both freelists are empty. So this is unlikely
1418 * to occur.
1419 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001420 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001421 void **object;
1422
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001423 tail = 0; /* Hot objects. Put the slab first */
1424
Christoph Lameter894b8782007-05-10 03:15:16 -07001425 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001426 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001427 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001428
1429 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001430 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001431 page->freelist = object;
1432 page->inuse--;
1433 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001434 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001435 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001436}
1437
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001438static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001439{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001440 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001441 slab_lock(c->page);
1442 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001443}
1444
1445/*
1446 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001447 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001448 * Called from IPI handler with interrupts disabled.
1449 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001450static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001451{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001452 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001453
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001454 if (likely(c && c->page))
1455 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001456}
1457
1458static void flush_cpu_slab(void *d)
1459{
1460 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001461
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001462 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001463}
1464
1465static void flush_all(struct kmem_cache *s)
1466{
Jens Axboe15c8b6c2008-05-09 09:39:44 +02001467 on_each_cpu(flush_cpu_slab, s, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07001468}
1469
1470/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001471 * Check if the objects in a per cpu structure fit numa
1472 * locality expectations.
1473 */
1474static inline int node_match(struct kmem_cache_cpu *c, int node)
1475{
1476#ifdef CONFIG_NUMA
1477 if (node != -1 && c->node != node)
1478 return 0;
1479#endif
1480 return 1;
1481}
1482
1483/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001484 * Slow path. The lockless freelist is empty or we need to perform
1485 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001486 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001487 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001488 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001489 * Processing is still very fast if new objects have been freed to the
1490 * regular freelist. In that case we simply take over the regular freelist
1491 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001492 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001493 * If that is not working then we fall back to the partial lists. We take the
1494 * first element of the freelist as the object to allocate now and move the
1495 * rest of the freelist to the lockless freelist.
1496 *
1497 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001498 * we need to allocate a new slab. This is the slowest path since it involves
1499 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001500 */
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001501static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1502 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001503{
Christoph Lameter81819f02007-05-06 14:49:36 -07001504 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001505 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001506
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001507 /* We handle __GFP_ZERO in the caller */
1508 gfpflags &= ~__GFP_ZERO;
1509
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001510 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001511 goto new_slab;
1512
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001513 slab_lock(c->page);
1514 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001515 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001516
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001517 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001518
Christoph Lameter894b8782007-05-10 03:15:16 -07001519load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001520 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001521 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001522 goto another_slab;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001523 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001524 goto debug;
1525
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001526 c->freelist = object[c->offset];
Christoph Lameter39b26462008-04-14 19:11:30 +03001527 c->page->inuse = c->page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001528 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001529 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001530unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001531 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001532 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001533 return object;
1534
1535another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001536 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001537
1538new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001539 new = get_partial(s, gfpflags, node);
1540 if (new) {
1541 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001542 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001543 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001544 }
1545
Christoph Lameterb811c202007-10-16 23:25:51 -07001546 if (gfpflags & __GFP_WAIT)
1547 local_irq_enable();
1548
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001549 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001550
1551 if (gfpflags & __GFP_WAIT)
1552 local_irq_disable();
1553
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001554 if (new) {
1555 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001556 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001557 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001558 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001559 slab_lock(new);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001560 __SetPageSlubFrozen(new);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001561 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001562 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001563 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001564 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001565debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001566 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001567 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001568
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001569 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001570 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001571 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001572 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001573}
1574
1575/*
1576 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1577 * have the fastpath folded into their functions. So no function call
1578 * overhead for requests that can be satisfied on the fastpath.
1579 *
1580 * The fastpath works by first checking if the lockless freelist can be used.
1581 * If not then __slab_alloc is called for slow processing.
1582 *
1583 * Otherwise we can simply pick the next object from the lockless free list.
1584 */
Pekka Enberg06428782008-01-07 23:20:27 -08001585static __always_inline void *slab_alloc(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001586 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001587{
Christoph Lameter894b8782007-05-10 03:15:16 -07001588 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001589 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001590 unsigned long flags;
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001591 unsigned int objsize;
Christoph Lameter1f842602008-01-07 23:20:30 -08001592
Nick Piggincf40bd12009-01-21 08:12:39 +01001593 lockdep_trace_alloc(gfpflags);
OGAWA Hirofumi89124d72008-11-19 21:23:59 +09001594 might_sleep_if(gfpflags & __GFP_WAIT);
Pekka Enberg3c506ef2008-12-29 11:47:05 +02001595
Akinobu Mita773ff602008-12-23 19:37:01 +09001596 if (should_failslab(s->objsize, gfpflags))
1597 return NULL;
1598
Christoph Lameter894b8782007-05-10 03:15:16 -07001599 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001600 c = get_cpu_slab(s, smp_processor_id());
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001601 objsize = c->objsize;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001602 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001603
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001604 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001605
1606 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001607 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001608 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001609 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001610 }
1611 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001612
1613 if (unlikely((gfpflags & __GFP_ZERO) && object))
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001614 memset(object, 0, objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001615
Christoph Lameter894b8782007-05-10 03:15:16 -07001616 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001617}
1618
1619void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1620{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001621 return slab_alloc(s, gfpflags, -1, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07001622}
1623EXPORT_SYMBOL(kmem_cache_alloc);
1624
1625#ifdef CONFIG_NUMA
1626void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1627{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001628 return slab_alloc(s, gfpflags, node, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07001629}
1630EXPORT_SYMBOL(kmem_cache_alloc_node);
1631#endif
1632
1633/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001634 * Slow patch handling. This may still be called frequently since objects
1635 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001636 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001637 * So we still attempt to reduce cache line usage. Just take the slab
1638 * lock and free the item. If there is no additional partial page
1639 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001640 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001641static void __slab_free(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001642 void *x, unsigned long addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001643{
1644 void *prior;
1645 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001646 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001647
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001648 c = get_cpu_slab(s, raw_smp_processor_id());
1649 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001650 slab_lock(page);
1651
Andy Whitcroft8a380822008-07-23 21:27:18 -07001652 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001653 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001654
Christoph Lameter81819f02007-05-06 14:49:36 -07001655checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001656 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001657 page->freelist = object;
1658 page->inuse--;
1659
Andy Whitcroft8a380822008-07-23 21:27:18 -07001660 if (unlikely(PageSlubFrozen(page))) {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001661 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001662 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001663 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001664
1665 if (unlikely(!page->inuse))
1666 goto slab_empty;
1667
1668 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001669 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001670 * then add it.
1671 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001672 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001673 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001674 stat(c, FREE_ADD_PARTIAL);
1675 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001676
1677out_unlock:
1678 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001679 return;
1680
1681slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001682 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001683 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001684 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001685 */
1686 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001687 stat(c, FREE_REMOVE_PARTIAL);
1688 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001689 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001690 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001691 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001692 return;
1693
1694debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001695 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001696 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001697 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001698}
1699
Christoph Lameter894b8782007-05-10 03:15:16 -07001700/*
1701 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1702 * can perform fastpath freeing without additional function calls.
1703 *
1704 * The fastpath is only possible if we are freeing to the current cpu slab
1705 * of this processor. This typically the case if we have just allocated
1706 * the item before.
1707 *
1708 * If fastpath is not possible then fall back to __slab_free where we deal
1709 * with all sorts of special processing.
1710 */
Pekka Enberg06428782008-01-07 23:20:27 -08001711static __always_inline void slab_free(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001712 struct page *page, void *x, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001713{
1714 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001715 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001716 unsigned long flags;
1717
Christoph Lameter894b8782007-05-10 03:15:16 -07001718 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001719 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001720 debug_check_no_locks_freed(object, c->objsize);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001721 if (!(s->flags & SLAB_DEBUG_OBJECTS))
Pekka Enberg6047a002009-01-14 12:22:25 +02001722 debug_check_no_obj_freed(object, c->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001723 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001724 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001725 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001726 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001727 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001728 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001729
1730 local_irq_restore(flags);
1731}
1732
Christoph Lameter81819f02007-05-06 14:49:36 -07001733void kmem_cache_free(struct kmem_cache *s, void *x)
1734{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001735 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001736
Christoph Lameterb49af682007-05-06 14:49:41 -07001737 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001738
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001739 slab_free(s, page, x, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07001740}
1741EXPORT_SYMBOL(kmem_cache_free);
1742
Cyrill Gorcunove9beef12008-10-28 22:02:26 +03001743/* Figure out on which slab page the object resides */
Christoph Lameter81819f02007-05-06 14:49:36 -07001744static struct page *get_object_page(const void *x)
1745{
Christoph Lameterb49af682007-05-06 14:49:41 -07001746 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001747
1748 if (!PageSlab(page))
1749 return NULL;
1750
1751 return page;
1752}
1753
1754/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001755 * Object placement in a slab is made very easy because we always start at
1756 * offset 0. If we tune the size of the object to the alignment then we can
1757 * get the required alignment by putting one properly sized object after
1758 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001759 *
1760 * Notice that the allocation order determines the sizes of the per cpu
1761 * caches. Each processor has always one slab available for allocations.
1762 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001763 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001764 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001765 */
1766
1767/*
1768 * Mininum / Maximum order of slab pages. This influences locking overhead
1769 * and slab fragmentation. A higher order reduces the number of partial slabs
1770 * and increases the number of allocations possible without having to
1771 * take the list_lock.
1772 */
1773static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03001774static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001775static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001776
1777/*
1778 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001779 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001780 */
1781static int slub_nomerge;
1782
1783/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001784 * Calculate the order of allocation given an slab object size.
1785 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001786 * The order of allocation has significant impact on performance and other
1787 * system components. Generally order 0 allocations should be preferred since
1788 * order 0 does not cause fragmentation in the page allocator. Larger objects
1789 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001790 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07001791 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001792 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001793 * In order to reach satisfactory performance we must ensure that a minimum
1794 * number of objects is in one slab. Otherwise we may generate too much
1795 * activity on the partial lists which requires taking the list_lock. This is
1796 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001797 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001798 * slub_max_order specifies the order where we begin to stop considering the
1799 * number of objects in a slab as critical. If we reach slub_max_order then
1800 * we try to keep the page order as low as possible. So we accept more waste
1801 * of space in favor of a small page order.
1802 *
1803 * Higher order allocations also allow the placement of more objects in a
1804 * slab and thereby reduce object handling overhead. If the user has
1805 * requested a higher mininum order then we start with that one instead of
1806 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001807 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001808static inline int slab_order(int size, int min_objects,
1809 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001810{
1811 int order;
1812 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001813 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001814
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04001815 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1816 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03001817
Christoph Lameter6300ea72007-07-17 04:03:20 -07001818 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001819 fls(min_objects * size - 1) - PAGE_SHIFT);
1820 order <= max_order; order++) {
1821
Christoph Lameter81819f02007-05-06 14:49:36 -07001822 unsigned long slab_size = PAGE_SIZE << order;
1823
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001824 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001825 continue;
1826
Christoph Lameter81819f02007-05-06 14:49:36 -07001827 rem = slab_size % size;
1828
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001829 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001830 break;
1831
1832 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001833
Christoph Lameter81819f02007-05-06 14:49:36 -07001834 return order;
1835}
1836
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001837static inline int calculate_order(int size)
1838{
1839 int order;
1840 int min_objects;
1841 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02001842 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001843
1844 /*
1845 * Attempt to find best configuration for a slab. This
1846 * works by first attempting to generate a layout with
1847 * the best configuration and backing off gradually.
1848 *
1849 * First we reduce the acceptable waste in a slab. Then
1850 * we reduce the minimum objects required in a slab.
1851 */
1852 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001853 if (!min_objects)
1854 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02001855 max_objects = (PAGE_SIZE << slub_max_order)/size;
1856 min_objects = min(min_objects, max_objects);
1857
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001858 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001859 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001860 while (fraction >= 4) {
1861 order = slab_order(size, min_objects,
1862 slub_max_order, fraction);
1863 if (order <= slub_max_order)
1864 return order;
1865 fraction /= 2;
1866 }
Zhang Yanmine8120ff2009-02-12 18:00:17 +02001867 min_objects --;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001868 }
1869
1870 /*
1871 * We were unable to place multiple objects in a slab. Now
1872 * lets see if we can place a single object there.
1873 */
1874 order = slab_order(size, 1, slub_max_order, 1);
1875 if (order <= slub_max_order)
1876 return order;
1877
1878 /*
1879 * Doh this slab cannot be placed using slub_max_order.
1880 */
1881 order = slab_order(size, 1, MAX_ORDER, 1);
1882 if (order <= MAX_ORDER)
1883 return order;
1884 return -ENOSYS;
1885}
1886
Christoph Lameter81819f02007-05-06 14:49:36 -07001887/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001888 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001889 */
1890static unsigned long calculate_alignment(unsigned long flags,
1891 unsigned long align, unsigned long size)
1892{
1893 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001894 * If the user wants hardware cache aligned objects then follow that
1895 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001896 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001897 * The hardware cache alignment cannot override the specified
1898 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001899 */
Nick Pigginb6210382008-03-05 14:05:56 -08001900 if (flags & SLAB_HWCACHE_ALIGN) {
1901 unsigned long ralign = cache_line_size();
1902 while (size <= ralign / 2)
1903 ralign /= 2;
1904 align = max(align, ralign);
1905 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001906
1907 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001908 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001909
1910 return ALIGN(align, sizeof(void *));
1911}
1912
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001913static void init_kmem_cache_cpu(struct kmem_cache *s,
1914 struct kmem_cache_cpu *c)
1915{
1916 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001917 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001918 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001919 c->offset = s->offset / sizeof(void *);
1920 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03001921#ifdef CONFIG_SLUB_STATS
1922 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1923#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001924}
1925
Pekka Enberg5595cff2008-08-05 09:28:47 +03001926static void
1927init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07001928{
1929 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001930 spin_lock_init(&n->list_lock);
1931 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001932#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001933 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07001934 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07001935 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001936#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001937}
1938
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001939#ifdef CONFIG_SMP
1940/*
1941 * Per cpu array for per cpu structures.
1942 *
1943 * The per cpu array places all kmem_cache_cpu structures from one processor
1944 * close together meaning that it becomes possible that multiple per cpu
1945 * structures are contained in one cacheline. This may be particularly
1946 * beneficial for the kmalloc caches.
1947 *
1948 * A desktop system typically has around 60-80 slabs. With 100 here we are
1949 * likely able to get per cpu structures for all caches from the array defined
1950 * here. We must be able to cover all kmalloc caches during bootstrap.
1951 *
1952 * If the per cpu array is exhausted then fall back to kmalloc
1953 * of individual cachelines. No sharing is possible then.
1954 */
1955#define NR_KMEM_CACHE_CPU 100
1956
1957static DEFINE_PER_CPU(struct kmem_cache_cpu,
1958 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1959
1960static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
Rusty Russell174596a2009-01-01 10:12:29 +10301961static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001962
1963static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1964 int cpu, gfp_t flags)
1965{
1966 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1967
1968 if (c)
1969 per_cpu(kmem_cache_cpu_free, cpu) =
1970 (void *)c->freelist;
1971 else {
1972 /* Table overflow: So allocate ourselves */
1973 c = kmalloc_node(
1974 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1975 flags, cpu_to_node(cpu));
1976 if (!c)
1977 return NULL;
1978 }
1979
1980 init_kmem_cache_cpu(s, c);
1981 return c;
1982}
1983
1984static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1985{
1986 if (c < per_cpu(kmem_cache_cpu, cpu) ||
David Rientjes37189092009-01-27 18:59:46 -08001987 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001988 kfree(c);
1989 return;
1990 }
1991 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1992 per_cpu(kmem_cache_cpu_free, cpu) = c;
1993}
1994
1995static void free_kmem_cache_cpus(struct kmem_cache *s)
1996{
1997 int cpu;
1998
1999 for_each_online_cpu(cpu) {
2000 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2001
2002 if (c) {
2003 s->cpu_slab[cpu] = NULL;
2004 free_kmem_cache_cpu(c, cpu);
2005 }
2006 }
2007}
2008
2009static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2010{
2011 int cpu;
2012
2013 for_each_online_cpu(cpu) {
2014 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2015
2016 if (c)
2017 continue;
2018
2019 c = alloc_kmem_cache_cpu(s, cpu, flags);
2020 if (!c) {
2021 free_kmem_cache_cpus(s);
2022 return 0;
2023 }
2024 s->cpu_slab[cpu] = c;
2025 }
2026 return 1;
2027}
2028
2029/*
2030 * Initialize the per cpu array.
2031 */
2032static void init_alloc_cpu_cpu(int cpu)
2033{
2034 int i;
2035
Rusty Russell174596a2009-01-01 10:12:29 +10302036 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002037 return;
2038
2039 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2040 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2041
Rusty Russell174596a2009-01-01 10:12:29 +10302042 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002043}
2044
2045static void __init init_alloc_cpu(void)
2046{
2047 int cpu;
2048
2049 for_each_online_cpu(cpu)
2050 init_alloc_cpu_cpu(cpu);
2051 }
2052
2053#else
2054static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2055static inline void init_alloc_cpu(void) {}
2056
2057static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2058{
2059 init_kmem_cache_cpu(s, &s->cpu_slab);
2060 return 1;
2061}
2062#endif
2063
Christoph Lameter81819f02007-05-06 14:49:36 -07002064#ifdef CONFIG_NUMA
2065/*
2066 * No kmalloc_node yet so do it by hand. We know that this is the first
2067 * slab on the node for this slabcache. There are no concurrent accesses
2068 * possible.
2069 *
2070 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002071 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2072 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002073 */
David Rientjes0094de92008-11-25 19:14:19 -08002074static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002075{
2076 struct page *page;
2077 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002078 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002079
2080 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2081
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002082 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002083
2084 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002085 if (page_to_nid(page) != node) {
2086 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2087 "node %d\n", node);
2088 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2089 "in order to be able to continue\n");
2090 }
2091
Christoph Lameter81819f02007-05-06 14:49:36 -07002092 n = page->freelist;
2093 BUG_ON(!n);
2094 page->freelist = get_freepointer(kmalloc_caches, n);
2095 page->inuse++;
2096 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002097#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002098 init_object(kmalloc_caches, n, 1);
2099 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002100#endif
Pekka Enberg5595cff2008-08-05 09:28:47 +03002101 init_kmem_cache_node(n, kmalloc_caches);
Christoph Lameter205ab992008-04-14 19:11:40 +03002102 inc_slabs_node(kmalloc_caches, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002103
rootba84c732008-01-07 23:20:28 -08002104 /*
2105 * lockdep requires consistent irq usage for each lock
2106 * so even though there cannot be a race this early in
2107 * the boot sequence, we still disable irqs.
2108 */
2109 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002110 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002111 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002112}
2113
2114static void free_kmem_cache_nodes(struct kmem_cache *s)
2115{
2116 int node;
2117
Christoph Lameterf64dc582007-10-16 01:25:33 -07002118 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002119 struct kmem_cache_node *n = s->node[node];
2120 if (n && n != &s->local_node)
2121 kmem_cache_free(kmalloc_caches, n);
2122 s->node[node] = NULL;
2123 }
2124}
2125
2126static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2127{
2128 int node;
2129 int local_node;
2130
2131 if (slab_state >= UP)
2132 local_node = page_to_nid(virt_to_page(s));
2133 else
2134 local_node = 0;
2135
Christoph Lameterf64dc582007-10-16 01:25:33 -07002136 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002137 struct kmem_cache_node *n;
2138
2139 if (local_node == node)
2140 n = &s->local_node;
2141 else {
2142 if (slab_state == DOWN) {
David Rientjes0094de92008-11-25 19:14:19 -08002143 early_kmem_cache_node_alloc(gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002144 continue;
2145 }
2146 n = kmem_cache_alloc_node(kmalloc_caches,
2147 gfpflags, node);
2148
2149 if (!n) {
2150 free_kmem_cache_nodes(s);
2151 return 0;
2152 }
2153
2154 }
2155 s->node[node] = n;
Pekka Enberg5595cff2008-08-05 09:28:47 +03002156 init_kmem_cache_node(n, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002157 }
2158 return 1;
2159}
2160#else
2161static void free_kmem_cache_nodes(struct kmem_cache *s)
2162{
2163}
2164
2165static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2166{
Pekka Enberg5595cff2008-08-05 09:28:47 +03002167 init_kmem_cache_node(&s->local_node, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002168 return 1;
2169}
2170#endif
2171
David Rientjesc0bdb232009-02-25 09:16:35 +02002172static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08002173{
2174 if (min < MIN_PARTIAL)
2175 min = MIN_PARTIAL;
2176 else if (min > MAX_PARTIAL)
2177 min = MAX_PARTIAL;
2178 s->min_partial = min;
2179}
2180
Christoph Lameter81819f02007-05-06 14:49:36 -07002181/*
2182 * calculate_sizes() determines the order and the distribution of data within
2183 * a slab object.
2184 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002185static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002186{
2187 unsigned long flags = s->flags;
2188 unsigned long size = s->objsize;
2189 unsigned long align = s->align;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002190 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002191
2192 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002193 * Round up object size to the next word boundary. We can only
2194 * place the free pointer at word boundaries and this determines
2195 * the possible location of the free pointer.
2196 */
2197 size = ALIGN(size, sizeof(void *));
2198
2199#ifdef CONFIG_SLUB_DEBUG
2200 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002201 * Determine if we can poison the object itself. If the user of
2202 * the slab may touch the object after free or before allocation
2203 * then we should never poison the object itself.
2204 */
2205 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002206 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002207 s->flags |= __OBJECT_POISON;
2208 else
2209 s->flags &= ~__OBJECT_POISON;
2210
Christoph Lameter81819f02007-05-06 14:49:36 -07002211
2212 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002213 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002214 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002215 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002216 */
2217 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2218 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002219#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002220
2221 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002222 * With that we have determined the number of bytes in actual use
2223 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002224 */
2225 s->inuse = size;
2226
2227 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002228 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002229 /*
2230 * Relocate free pointer after the object if it is not
2231 * permitted to overwrite the first word of the object on
2232 * kmem_cache_free.
2233 *
2234 * This is the case if we do RCU, have a constructor or
2235 * destructor or are poisoning the objects.
2236 */
2237 s->offset = size;
2238 size += sizeof(void *);
2239 }
2240
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002241#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002242 if (flags & SLAB_STORE_USER)
2243 /*
2244 * Need to store information about allocs and frees after
2245 * the object.
2246 */
2247 size += 2 * sizeof(struct track);
2248
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002249 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002250 /*
2251 * Add some empty padding so that we can catch
2252 * overwrites from earlier objects rather than let
2253 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01002254 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07002255 * of the object.
2256 */
2257 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002258#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002259
Christoph Lameter81819f02007-05-06 14:49:36 -07002260 /*
2261 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002262 * user specified and the dynamic determination of cache line size
2263 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002264 */
2265 align = calculate_alignment(flags, align, s->objsize);
2266
2267 /*
2268 * SLUB stores one object immediately after another beginning from
2269 * offset 0. In order to align the objects we have to simply size
2270 * each object to conform to the alignment.
2271 */
2272 size = ALIGN(size, align);
2273 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03002274 if (forced_order >= 0)
2275 order = forced_order;
2276 else
2277 order = calculate_order(size);
Christoph Lameter81819f02007-05-06 14:49:36 -07002278
Christoph Lameter834f3d12008-04-14 19:11:31 +03002279 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07002280 return 0;
2281
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002282 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002283 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002284 s->allocflags |= __GFP_COMP;
2285
2286 if (s->flags & SLAB_CACHE_DMA)
2287 s->allocflags |= SLUB_DMA;
2288
2289 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2290 s->allocflags |= __GFP_RECLAIMABLE;
2291
Christoph Lameter81819f02007-05-06 14:49:36 -07002292 /*
2293 * Determine the number of objects per slab
2294 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002295 s->oo = oo_make(order, size);
Christoph Lameter65c33762008-04-14 19:11:40 +03002296 s->min = oo_make(get_order(size), size);
Christoph Lameter205ab992008-04-14 19:11:40 +03002297 if (oo_objects(s->oo) > oo_objects(s->max))
2298 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002299
Christoph Lameter834f3d12008-04-14 19:11:31 +03002300 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07002301
2302}
2303
Christoph Lameter81819f02007-05-06 14:49:36 -07002304static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2305 const char *name, size_t size,
2306 size_t align, unsigned long flags,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002307 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002308{
2309 memset(s, 0, kmem_size);
2310 s->name = name;
2311 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002312 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002313 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002314 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002315
Christoph Lameter06b285d2008-04-14 19:11:41 +03002316 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07002317 goto error;
2318
David Rientjes3b89d7d2009-02-22 17:40:07 -08002319 /*
2320 * The larger the object size is, the more pages we want on the partial
2321 * list to avoid pounding the page allocator excessively.
2322 */
David Rientjesc0bdb232009-02-25 09:16:35 +02002323 set_min_partial(s, ilog2(s->size));
Christoph Lameter81819f02007-05-06 14:49:36 -07002324 s->refcount = 1;
2325#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05002326 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07002327#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002328 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2329 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002330
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002331 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002332 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002333 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002334error:
2335 if (flags & SLAB_PANIC)
2336 panic("Cannot create slab %s size=%lu realsize=%u "
2337 "order=%u offset=%u flags=%lx\n",
Christoph Lameter834f3d12008-04-14 19:11:31 +03002338 s->name, (unsigned long)size, s->size, oo_order(s->oo),
Christoph Lameter81819f02007-05-06 14:49:36 -07002339 s->offset, flags);
2340 return 0;
2341}
Christoph Lameter81819f02007-05-06 14:49:36 -07002342
2343/*
2344 * Check if a given pointer is valid
2345 */
2346int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2347{
Pekka Enberg06428782008-01-07 23:20:27 -08002348 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002349
2350 page = get_object_page(object);
2351
2352 if (!page || s != page->slab)
2353 /* No slab or wrong slab */
2354 return 0;
2355
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002356 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002357 return 0;
2358
2359 /*
2360 * We could also check if the object is on the slabs freelist.
2361 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002362 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002363 * to a certain slab.
2364 */
2365 return 1;
2366}
2367EXPORT_SYMBOL(kmem_ptr_validate);
2368
2369/*
2370 * Determine the size of a slab object
2371 */
2372unsigned int kmem_cache_size(struct kmem_cache *s)
2373{
2374 return s->objsize;
2375}
2376EXPORT_SYMBOL(kmem_cache_size);
2377
2378const char *kmem_cache_name(struct kmem_cache *s)
2379{
2380 return s->name;
2381}
2382EXPORT_SYMBOL(kmem_cache_name);
2383
Christoph Lameter33b12c32008-04-25 12:22:43 -07002384static void list_slab_objects(struct kmem_cache *s, struct page *page,
2385 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07002386{
Christoph Lameter33b12c32008-04-25 12:22:43 -07002387#ifdef CONFIG_SLUB_DEBUG
2388 void *addr = page_address(page);
2389 void *p;
2390 DECLARE_BITMAP(map, page->objects);
2391
2392 bitmap_zero(map, page->objects);
2393 slab_err(s, page, "%s", text);
2394 slab_lock(page);
2395 for_each_free_object(p, s, page->freelist)
2396 set_bit(slab_index(p, s, addr), map);
2397
2398 for_each_object(p, s, addr, page->objects) {
2399
2400 if (!test_bit(slab_index(p, s, addr), map)) {
2401 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2402 p, p - addr);
2403 print_tracking(s, p);
2404 }
2405 }
2406 slab_unlock(page);
2407#endif
2408}
2409
Christoph Lameter81819f02007-05-06 14:49:36 -07002410/*
Christoph Lameter599870b2008-04-23 12:36:52 -07002411 * Attempt to free all partial slabs on a node.
Christoph Lameter81819f02007-05-06 14:49:36 -07002412 */
Christoph Lameter599870b2008-04-23 12:36:52 -07002413static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002414{
Christoph Lameter81819f02007-05-06 14:49:36 -07002415 unsigned long flags;
2416 struct page *page, *h;
2417
2418 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter33b12c32008-04-25 12:22:43 -07002419 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002420 if (!page->inuse) {
2421 list_del(&page->lru);
2422 discard_slab(s, page);
Christoph Lameter599870b2008-04-23 12:36:52 -07002423 n->nr_partial--;
Christoph Lameter33b12c32008-04-25 12:22:43 -07002424 } else {
2425 list_slab_objects(s, page,
2426 "Objects remaining on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07002427 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07002428 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002429 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002430}
2431
2432/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002433 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002434 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002435static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002436{
2437 int node;
2438
2439 flush_all(s);
2440
2441 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002442 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002443 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002444 struct kmem_cache_node *n = get_node(s, node);
2445
Christoph Lameter599870b2008-04-23 12:36:52 -07002446 free_partial(s, n);
2447 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002448 return 1;
2449 }
2450 free_kmem_cache_nodes(s);
2451 return 0;
2452}
2453
2454/*
2455 * Close a cache and release the kmem_cache structure
2456 * (must be used for caches created using kmem_cache_create)
2457 */
2458void kmem_cache_destroy(struct kmem_cache *s)
2459{
2460 down_write(&slub_lock);
2461 s->refcount--;
2462 if (!s->refcount) {
2463 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002464 up_write(&slub_lock);
Pekka Enbergd629d812008-04-23 22:31:08 +03002465 if (kmem_cache_close(s)) {
2466 printk(KERN_ERR "SLUB %s: %s called for cache that "
2467 "still has objects.\n", s->name, __func__);
2468 dump_stack();
2469 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002470 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002471 } else
2472 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002473}
2474EXPORT_SYMBOL(kmem_cache_destroy);
2475
2476/********************************************************************
2477 * Kmalloc subsystem
2478 *******************************************************************/
2479
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002480struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002481EXPORT_SYMBOL(kmalloc_caches);
2482
Christoph Lameter81819f02007-05-06 14:49:36 -07002483static int __init setup_slub_min_order(char *str)
2484{
Pekka Enberg06428782008-01-07 23:20:27 -08002485 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002486
2487 return 1;
2488}
2489
2490__setup("slub_min_order=", setup_slub_min_order);
2491
2492static int __init setup_slub_max_order(char *str)
2493{
Pekka Enberg06428782008-01-07 23:20:27 -08002494 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002495
2496 return 1;
2497}
2498
2499__setup("slub_max_order=", setup_slub_max_order);
2500
2501static int __init setup_slub_min_objects(char *str)
2502{
Pekka Enberg06428782008-01-07 23:20:27 -08002503 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002504
2505 return 1;
2506}
2507
2508__setup("slub_min_objects=", setup_slub_min_objects);
2509
2510static int __init setup_slub_nomerge(char *str)
2511{
2512 slub_nomerge = 1;
2513 return 1;
2514}
2515
2516__setup("slub_nomerge", setup_slub_nomerge);
2517
Christoph Lameter81819f02007-05-06 14:49:36 -07002518static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2519 const char *name, int size, gfp_t gfp_flags)
2520{
2521 unsigned int flags = 0;
2522
2523 if (gfp_flags & SLUB_DMA)
2524 flags = SLAB_CACHE_DMA;
2525
2526 down_write(&slub_lock);
2527 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter319d1e22008-04-14 19:11:41 +03002528 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002529 goto panic;
2530
2531 list_add(&s->list, &slab_caches);
2532 up_write(&slub_lock);
2533 if (sysfs_slab_add(s))
2534 goto panic;
2535 return s;
2536
2537panic:
2538 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2539}
2540
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002541#ifdef CONFIG_ZONE_DMA
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002542static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002543
2544static void sysfs_add_func(struct work_struct *w)
2545{
2546 struct kmem_cache *s;
2547
2548 down_write(&slub_lock);
2549 list_for_each_entry(s, &slab_caches, list) {
2550 if (s->flags & __SYSFS_ADD_DEFERRED) {
2551 s->flags &= ~__SYSFS_ADD_DEFERRED;
2552 sysfs_slab_add(s);
2553 }
2554 }
2555 up_write(&slub_lock);
2556}
2557
2558static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2559
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002560static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2561{
2562 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002563 char *text;
2564 size_t realsize;
2565
2566 s = kmalloc_caches_dma[index];
2567 if (s)
2568 return s;
2569
2570 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002571 if (flags & __GFP_WAIT)
2572 down_write(&slub_lock);
2573 else {
2574 if (!down_write_trylock(&slub_lock))
2575 goto out;
2576 }
2577
2578 if (kmalloc_caches_dma[index])
2579 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002580
Christoph Lameter7b55f622007-07-17 04:03:27 -07002581 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002582 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2583 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002584 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2585
2586 if (!s || !text || !kmem_cache_open(s, flags, text,
2587 realsize, ARCH_KMALLOC_MINALIGN,
2588 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2589 kfree(s);
2590 kfree(text);
2591 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002592 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002593
2594 list_add(&s->list, &slab_caches);
2595 kmalloc_caches_dma[index] = s;
2596
2597 schedule_work(&sysfs_add_work);
2598
2599unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002600 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002601out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002602 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002603}
2604#endif
2605
Christoph Lameterf1b26332007-07-17 04:03:26 -07002606/*
2607 * Conversion table for small slabs sizes / 8 to the index in the
2608 * kmalloc array. This is necessary for slabs < 192 since we have non power
2609 * of two cache sizes there. The size of larger slabs can be determined using
2610 * fls.
2611 */
2612static s8 size_index[24] = {
2613 3, /* 8 */
2614 4, /* 16 */
2615 5, /* 24 */
2616 5, /* 32 */
2617 6, /* 40 */
2618 6, /* 48 */
2619 6, /* 56 */
2620 6, /* 64 */
2621 1, /* 72 */
2622 1, /* 80 */
2623 1, /* 88 */
2624 1, /* 96 */
2625 7, /* 104 */
2626 7, /* 112 */
2627 7, /* 120 */
2628 7, /* 128 */
2629 2, /* 136 */
2630 2, /* 144 */
2631 2, /* 152 */
2632 2, /* 160 */
2633 2, /* 168 */
2634 2, /* 176 */
2635 2, /* 184 */
2636 2 /* 192 */
2637};
2638
Christoph Lameter81819f02007-05-06 14:49:36 -07002639static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2640{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002641 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002642
Christoph Lameterf1b26332007-07-17 04:03:26 -07002643 if (size <= 192) {
2644 if (!size)
2645 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002646
Christoph Lameterf1b26332007-07-17 04:03:26 -07002647 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002648 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002649 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002650
2651#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002652 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002653 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002654
Christoph Lameter81819f02007-05-06 14:49:36 -07002655#endif
2656 return &kmalloc_caches[index];
2657}
2658
2659void *__kmalloc(size_t size, gfp_t flags)
2660{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002661 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002662
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002663 if (unlikely(size > SLUB_MAX_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002664 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002665
2666 s = get_slab(size, flags);
2667
2668 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002669 return s;
2670
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002671 return slab_alloc(s, flags, -1, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07002672}
2673EXPORT_SYMBOL(__kmalloc);
2674
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002675static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2676{
2677 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2678 get_order(size));
2679
2680 if (page)
2681 return page_address(page);
2682 else
2683 return NULL;
2684}
2685
Christoph Lameter81819f02007-05-06 14:49:36 -07002686#ifdef CONFIG_NUMA
2687void *__kmalloc_node(size_t size, gfp_t flags, int node)
2688{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002689 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002690
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002691 if (unlikely(size > SLUB_MAX_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002692 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002693
2694 s = get_slab(size, flags);
2695
2696 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002697 return s;
2698
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002699 return slab_alloc(s, flags, node, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07002700}
2701EXPORT_SYMBOL(__kmalloc_node);
2702#endif
2703
2704size_t ksize(const void *object)
2705{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002706 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002707 struct kmem_cache *s;
2708
Christoph Lameteref8b4522007-10-16 01:24:46 -07002709 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002710 return 0;
2711
Vegard Nossum294a80a2007-12-04 23:45:30 -08002712 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002713
Pekka Enberg76994412008-05-22 19:22:25 +03002714 if (unlikely(!PageSlab(page))) {
2715 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08002716 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03002717 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002718 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002719
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002720#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002721 /*
2722 * Debugging requires use of the padding between object
2723 * and whatever may come after it.
2724 */
2725 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2726 return s->objsize;
2727
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002728#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002729 /*
2730 * If we have the need to store the freelist pointer
2731 * back there or track user information then we can
2732 * only use the space before that information.
2733 */
2734 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2735 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002736 /*
2737 * Else we can use all the padding etc for the allocation
2738 */
2739 return s->size;
2740}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02002741EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07002742
2743void kfree(const void *x)
2744{
Christoph Lameter81819f02007-05-06 14:49:36 -07002745 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002746 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002747
Satyam Sharma2408c552007-10-16 01:24:44 -07002748 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002749 return;
2750
Christoph Lameterb49af682007-05-06 14:49:41 -07002751 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002752 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07002753 BUG_ON(!PageCompound(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002754 put_page(page);
2755 return;
2756 }
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002757 slab_free(page->slab, page, object, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07002758}
2759EXPORT_SYMBOL(kfree);
2760
Christoph Lameter2086d262007-05-06 14:49:46 -07002761/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002762 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2763 * the remaining slabs by the number of items in use. The slabs with the
2764 * most items in use come first. New allocations will then fill those up
2765 * and thus they can be removed from the partial lists.
2766 *
2767 * The slabs with the least items are placed last. This results in them
2768 * being allocated from last increasing the chance that the last objects
2769 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002770 */
2771int kmem_cache_shrink(struct kmem_cache *s)
2772{
2773 int node;
2774 int i;
2775 struct kmem_cache_node *n;
2776 struct page *page;
2777 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03002778 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07002779 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03002780 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07002781 unsigned long flags;
2782
2783 if (!slabs_by_inuse)
2784 return -ENOMEM;
2785
2786 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002787 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002788 n = get_node(s, node);
2789
2790 if (!n->nr_partial)
2791 continue;
2792
Christoph Lameter834f3d12008-04-14 19:11:31 +03002793 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07002794 INIT_LIST_HEAD(slabs_by_inuse + i);
2795
2796 spin_lock_irqsave(&n->list_lock, flags);
2797
2798 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002799 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002800 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002801 * Note that concurrent frees may occur while we hold the
2802 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002803 */
2804 list_for_each_entry_safe(page, t, &n->partial, lru) {
2805 if (!page->inuse && slab_trylock(page)) {
2806 /*
2807 * Must hold slab lock here because slab_free
2808 * may have freed the last object and be
2809 * waiting to release the slab.
2810 */
2811 list_del(&page->lru);
2812 n->nr_partial--;
2813 slab_unlock(page);
2814 discard_slab(s, page);
2815 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002816 list_move(&page->lru,
2817 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002818 }
2819 }
2820
Christoph Lameter2086d262007-05-06 14:49:46 -07002821 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002822 * Rebuild the partial list with the slabs filled up most
2823 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002824 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002825 for (i = objects - 1; i >= 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07002826 list_splice(slabs_by_inuse + i, n->partial.prev);
2827
Christoph Lameter2086d262007-05-06 14:49:46 -07002828 spin_unlock_irqrestore(&n->list_lock, flags);
2829 }
2830
2831 kfree(slabs_by_inuse);
2832 return 0;
2833}
2834EXPORT_SYMBOL(kmem_cache_shrink);
2835
Yasunori Gotob9049e22007-10-21 16:41:37 -07002836#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2837static int slab_mem_going_offline_callback(void *arg)
2838{
2839 struct kmem_cache *s;
2840
2841 down_read(&slub_lock);
2842 list_for_each_entry(s, &slab_caches, list)
2843 kmem_cache_shrink(s);
2844 up_read(&slub_lock);
2845
2846 return 0;
2847}
2848
2849static void slab_mem_offline_callback(void *arg)
2850{
2851 struct kmem_cache_node *n;
2852 struct kmem_cache *s;
2853 struct memory_notify *marg = arg;
2854 int offline_node;
2855
2856 offline_node = marg->status_change_nid;
2857
2858 /*
2859 * If the node still has available memory. we need kmem_cache_node
2860 * for it yet.
2861 */
2862 if (offline_node < 0)
2863 return;
2864
2865 down_read(&slub_lock);
2866 list_for_each_entry(s, &slab_caches, list) {
2867 n = get_node(s, offline_node);
2868 if (n) {
2869 /*
2870 * if n->nr_slabs > 0, slabs still exist on the node
2871 * that is going down. We were unable to free them,
2872 * and offline_pages() function shoudn't call this
2873 * callback. So, we must fail.
2874 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002875 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002876
2877 s->node[offline_node] = NULL;
2878 kmem_cache_free(kmalloc_caches, n);
2879 }
2880 }
2881 up_read(&slub_lock);
2882}
2883
2884static int slab_mem_going_online_callback(void *arg)
2885{
2886 struct kmem_cache_node *n;
2887 struct kmem_cache *s;
2888 struct memory_notify *marg = arg;
2889 int nid = marg->status_change_nid;
2890 int ret = 0;
2891
2892 /*
2893 * If the node's memory is already available, then kmem_cache_node is
2894 * already created. Nothing to do.
2895 */
2896 if (nid < 0)
2897 return 0;
2898
2899 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07002900 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07002901 * allocate a kmem_cache_node structure in order to bring the node
2902 * online.
2903 */
2904 down_read(&slub_lock);
2905 list_for_each_entry(s, &slab_caches, list) {
2906 /*
2907 * XXX: kmem_cache_alloc_node will fallback to other nodes
2908 * since memory is not yet available from the node that
2909 * is brought up.
2910 */
2911 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2912 if (!n) {
2913 ret = -ENOMEM;
2914 goto out;
2915 }
Pekka Enberg5595cff2008-08-05 09:28:47 +03002916 init_kmem_cache_node(n, s);
Yasunori Gotob9049e22007-10-21 16:41:37 -07002917 s->node[nid] = n;
2918 }
2919out:
2920 up_read(&slub_lock);
2921 return ret;
2922}
2923
2924static int slab_memory_callback(struct notifier_block *self,
2925 unsigned long action, void *arg)
2926{
2927 int ret = 0;
2928
2929 switch (action) {
2930 case MEM_GOING_ONLINE:
2931 ret = slab_mem_going_online_callback(arg);
2932 break;
2933 case MEM_GOING_OFFLINE:
2934 ret = slab_mem_going_offline_callback(arg);
2935 break;
2936 case MEM_OFFLINE:
2937 case MEM_CANCEL_ONLINE:
2938 slab_mem_offline_callback(arg);
2939 break;
2940 case MEM_ONLINE:
2941 case MEM_CANCEL_OFFLINE:
2942 break;
2943 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08002944 if (ret)
2945 ret = notifier_from_errno(ret);
2946 else
2947 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002948 return ret;
2949}
2950
2951#endif /* CONFIG_MEMORY_HOTPLUG */
2952
Christoph Lameter81819f02007-05-06 14:49:36 -07002953/********************************************************************
2954 * Basic setup of slabs
2955 *******************************************************************/
2956
2957void __init kmem_cache_init(void)
2958{
2959 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002960 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002961
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002962 init_alloc_cpu();
2963
Christoph Lameter81819f02007-05-06 14:49:36 -07002964#ifdef CONFIG_NUMA
2965 /*
2966 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002967 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002968 * kmem_cache_open for slab_state == DOWN.
2969 */
2970 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2971 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002972 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002973 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002974
Nadia Derbey0c40ba42008-04-29 01:00:41 -07002975 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
Christoph Lameter81819f02007-05-06 14:49:36 -07002976#endif
2977
2978 /* Able to allocate the per node structures */
2979 slab_state = PARTIAL;
2980
2981 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002982 if (KMALLOC_MIN_SIZE <= 64) {
2983 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002984 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002985 caches++;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002986 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002987 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002988 caches++;
2989 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002990
Christoph Lameterffadd4d2009-02-17 12:05:07 -05002991 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002992 create_kmalloc_cache(&kmalloc_caches[i],
2993 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002994 caches++;
2995 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002996
Christoph Lameterf1b26332007-07-17 04:03:26 -07002997
2998 /*
2999 * Patch up the size_index table if we have strange large alignment
3000 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08003001 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07003002 *
3003 * Largest permitted alignment is 256 bytes due to the way we
3004 * handle the index determination for the smaller caches.
3005 *
3006 * Make sure that nothing crazy happens if someone starts tinkering
3007 * around with ARCH_KMALLOC_MINALIGN
3008 */
3009 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3010 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3011
Christoph Lameter12ad6842007-07-17 04:03:28 -07003012 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07003013 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3014
Christoph Lameter41d54d32008-07-03 09:14:26 -05003015 if (KMALLOC_MIN_SIZE == 128) {
3016 /*
3017 * The 192 byte sized cache is not used if the alignment
3018 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3019 * instead.
3020 */
3021 for (i = 128 + 8; i <= 192; i += 8)
3022 size_index[(i - 1) / 8] = 8;
3023 }
3024
Christoph Lameter81819f02007-05-06 14:49:36 -07003025 slab_state = UP;
3026
3027 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameterffadd4d2009-02-17 12:05:07 -05003028 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07003029 kmalloc_caches[i]. name =
3030 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3031
3032#ifdef CONFIG_SMP
3033 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003034 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3035 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3036#else
3037 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003038#endif
3039
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003040 printk(KERN_INFO
3041 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003042 " CPUs=%d, Nodes=%d\n",
3043 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003044 slub_min_order, slub_max_order, slub_min_objects,
3045 nr_cpu_ids, nr_node_ids);
3046}
3047
3048/*
3049 * Find a mergeable slab cache
3050 */
3051static int slab_unmergeable(struct kmem_cache *s)
3052{
3053 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3054 return 1;
3055
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003056 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003057 return 1;
3058
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003059 /*
3060 * We may have set a slab to be unmergeable during bootstrap.
3061 */
3062 if (s->refcount < 0)
3063 return 1;
3064
Christoph Lameter81819f02007-05-06 14:49:36 -07003065 return 0;
3066}
3067
3068static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003069 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003070 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003071{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003072 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003073
3074 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3075 return NULL;
3076
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003077 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003078 return NULL;
3079
3080 size = ALIGN(size, sizeof(void *));
3081 align = calculate_alignment(flags, align, size);
3082 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003083 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003084
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003085 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003086 if (slab_unmergeable(s))
3087 continue;
3088
3089 if (size > s->size)
3090 continue;
3091
Christoph Lameterba0268a2007-09-11 15:24:11 -07003092 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003093 continue;
3094 /*
3095 * Check if alignment is compatible.
3096 * Courtesy of Adrian Drzewiecki
3097 */
Pekka Enberg06428782008-01-07 23:20:27 -08003098 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003099 continue;
3100
3101 if (s->size - size >= sizeof(void *))
3102 continue;
3103
3104 return s;
3105 }
3106 return NULL;
3107}
3108
3109struct kmem_cache *kmem_cache_create(const char *name, size_t size,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003110 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003111{
3112 struct kmem_cache *s;
3113
3114 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003115 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003116 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003117 int cpu;
3118
Christoph Lameter81819f02007-05-06 14:49:36 -07003119 s->refcount++;
3120 /*
3121 * Adjust the object sizes so that we clear
3122 * the complete object on kzalloc.
3123 */
3124 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003125
3126 /*
3127 * And then we need to update the object size in the
3128 * per cpu structures
3129 */
3130 for_each_online_cpu(cpu)
3131 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003132
Christoph Lameter81819f02007-05-06 14:49:36 -07003133 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003134 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003135
David Rientjes7b8f3b62008-12-17 22:09:46 -08003136 if (sysfs_slab_alias(s, name)) {
3137 down_write(&slub_lock);
3138 s->refcount--;
3139 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003140 goto err;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003141 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003142 return s;
3143 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003144
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003145 s = kmalloc(kmem_size, GFP_KERNEL);
3146 if (s) {
3147 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003148 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003149 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003150 up_write(&slub_lock);
David Rientjes7b8f3b62008-12-17 22:09:46 -08003151 if (sysfs_slab_add(s)) {
3152 down_write(&slub_lock);
3153 list_del(&s->list);
3154 up_write(&slub_lock);
3155 kfree(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003156 goto err;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003157 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003158 return s;
3159 }
3160 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003161 }
3162 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003163
3164err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003165 if (flags & SLAB_PANIC)
3166 panic("Cannot create slabcache %s\n", name);
3167 else
3168 s = NULL;
3169 return s;
3170}
3171EXPORT_SYMBOL(kmem_cache_create);
3172
Christoph Lameter81819f02007-05-06 14:49:36 -07003173#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003174/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003175 * Use the cpu notifier to insure that the cpu slabs are flushed when
3176 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003177 */
3178static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3179 unsigned long action, void *hcpu)
3180{
3181 long cpu = (long)hcpu;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003182 struct kmem_cache *s;
3183 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003184
3185 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003186 case CPU_UP_PREPARE:
3187 case CPU_UP_PREPARE_FROZEN:
3188 init_alloc_cpu_cpu(cpu);
3189 down_read(&slub_lock);
3190 list_for_each_entry(s, &slab_caches, list)
3191 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3192 GFP_KERNEL);
3193 up_read(&slub_lock);
3194 break;
3195
Christoph Lameter81819f02007-05-06 14:49:36 -07003196 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003197 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003198 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003199 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003200 down_read(&slub_lock);
3201 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003202 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3203
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003204 local_irq_save(flags);
3205 __flush_cpu_slab(s, cpu);
3206 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003207 free_kmem_cache_cpu(c, cpu);
3208 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003209 }
3210 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003211 break;
3212 default:
3213 break;
3214 }
3215 return NOTIFY_OK;
3216}
3217
Pekka Enberg06428782008-01-07 23:20:27 -08003218static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003219 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003220};
Christoph Lameter81819f02007-05-06 14:49:36 -07003221
3222#endif
3223
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003224void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003225{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003226 struct kmem_cache *s;
3227
Christoph Lameterffadd4d2009-02-17 12:05:07 -05003228 if (unlikely(size > SLUB_MAX_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003229 return kmalloc_large(size, gfpflags);
3230
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003231 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003232
Satyam Sharma2408c552007-10-16 01:24:44 -07003233 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003234 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003235
Christoph Lameterce15fea2007-07-17 04:03:28 -07003236 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003237}
3238
3239void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003240 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003241{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003242 struct kmem_cache *s;
3243
Christoph Lameterffadd4d2009-02-17 12:05:07 -05003244 if (unlikely(size > SLUB_MAX_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003245 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003246
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003247 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003248
Satyam Sharma2408c552007-10-16 01:24:44 -07003249 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003250 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003251
Christoph Lameterce15fea2007-07-17 04:03:28 -07003252 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003253}
3254
Christoph Lameterf6acb632008-04-29 16:16:06 -07003255#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03003256static unsigned long count_partial(struct kmem_cache_node *n,
3257 int (*get_count)(struct page *))
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003258{
3259 unsigned long flags;
3260 unsigned long x = 0;
3261 struct page *page;
3262
3263 spin_lock_irqsave(&n->list_lock, flags);
3264 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter205ab992008-04-14 19:11:40 +03003265 x += get_count(page);
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003266 spin_unlock_irqrestore(&n->list_lock, flags);
3267 return x;
3268}
Christoph Lameter205ab992008-04-14 19:11:40 +03003269
3270static int count_inuse(struct page *page)
3271{
3272 return page->inuse;
3273}
3274
3275static int count_total(struct page *page)
3276{
3277 return page->objects;
3278}
3279
3280static int count_free(struct page *page)
3281{
3282 return page->objects - page->inuse;
3283}
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003284
Christoph Lameter434e2452007-07-17 04:03:30 -07003285static int validate_slab(struct kmem_cache *s, struct page *page,
3286 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003287{
3288 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003289 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003290
3291 if (!check_slab(s, page) ||
3292 !on_freelist(s, page, NULL))
3293 return 0;
3294
3295 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003296 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003297
Christoph Lameter7656c722007-05-09 02:32:40 -07003298 for_each_free_object(p, s, page->freelist) {
3299 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003300 if (!check_object(s, page, p, 0))
3301 return 0;
3302 }
3303
Christoph Lameter224a88b2008-04-14 19:11:31 +03003304 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003305 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003306 if (!check_object(s, page, p, 1))
3307 return 0;
3308 return 1;
3309}
3310
Christoph Lameter434e2452007-07-17 04:03:30 -07003311static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3312 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003313{
3314 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003315 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003316 slab_unlock(page);
3317 } else
3318 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3319 s->name, page);
3320
3321 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003322 if (!PageSlubDebug(page))
3323 printk(KERN_ERR "SLUB %s: SlubDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003324 "on slab 0x%p\n", s->name, page);
3325 } else {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003326 if (PageSlubDebug(page))
3327 printk(KERN_ERR "SLUB %s: SlubDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003328 "slab 0x%p\n", s->name, page);
3329 }
3330}
3331
Christoph Lameter434e2452007-07-17 04:03:30 -07003332static int validate_slab_node(struct kmem_cache *s,
3333 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003334{
3335 unsigned long count = 0;
3336 struct page *page;
3337 unsigned long flags;
3338
3339 spin_lock_irqsave(&n->list_lock, flags);
3340
3341 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003342 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003343 count++;
3344 }
3345 if (count != n->nr_partial)
3346 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3347 "counter=%ld\n", s->name, count, n->nr_partial);
3348
3349 if (!(s->flags & SLAB_STORE_USER))
3350 goto out;
3351
3352 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003353 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003354 count++;
3355 }
3356 if (count != atomic_long_read(&n->nr_slabs))
3357 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3358 "counter=%ld\n", s->name, count,
3359 atomic_long_read(&n->nr_slabs));
3360
3361out:
3362 spin_unlock_irqrestore(&n->list_lock, flags);
3363 return count;
3364}
3365
Christoph Lameter434e2452007-07-17 04:03:30 -07003366static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003367{
3368 int node;
3369 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003370 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003371 sizeof(unsigned long), GFP_KERNEL);
3372
3373 if (!map)
3374 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003375
3376 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003377 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003378 struct kmem_cache_node *n = get_node(s, node);
3379
Christoph Lameter434e2452007-07-17 04:03:30 -07003380 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003381 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003382 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003383 return count;
3384}
3385
Christoph Lameterb3459702007-05-09 02:32:41 -07003386#ifdef SLUB_RESILIENCY_TEST
3387static void resiliency_test(void)
3388{
3389 u8 *p;
3390
3391 printk(KERN_ERR "SLUB resiliency testing\n");
3392 printk(KERN_ERR "-----------------------\n");
3393 printk(KERN_ERR "A. Corruption after allocation\n");
3394
3395 p = kzalloc(16, GFP_KERNEL);
3396 p[16] = 0x12;
3397 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3398 " 0x12->0x%p\n\n", p + 16);
3399
3400 validate_slab_cache(kmalloc_caches + 4);
3401
3402 /* Hmmm... The next two are dangerous */
3403 p = kzalloc(32, GFP_KERNEL);
3404 p[32 + sizeof(void *)] = 0x34;
3405 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003406 " 0x34 -> -0x%p\n", p);
3407 printk(KERN_ERR
3408 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003409
3410 validate_slab_cache(kmalloc_caches + 5);
3411 p = kzalloc(64, GFP_KERNEL);
3412 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3413 *p = 0x56;
3414 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3415 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003416 printk(KERN_ERR
3417 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003418 validate_slab_cache(kmalloc_caches + 6);
3419
3420 printk(KERN_ERR "\nB. Corruption after free\n");
3421 p = kzalloc(128, GFP_KERNEL);
3422 kfree(p);
3423 *p = 0x78;
3424 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3425 validate_slab_cache(kmalloc_caches + 7);
3426
3427 p = kzalloc(256, GFP_KERNEL);
3428 kfree(p);
3429 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003430 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3431 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003432 validate_slab_cache(kmalloc_caches + 8);
3433
3434 p = kzalloc(512, GFP_KERNEL);
3435 kfree(p);
3436 p[512] = 0xab;
3437 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3438 validate_slab_cache(kmalloc_caches + 9);
3439}
3440#else
3441static void resiliency_test(void) {};
3442#endif
3443
Christoph Lameter88a420e2007-05-06 14:49:45 -07003444/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003445 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003446 * and freed.
3447 */
3448
3449struct location {
3450 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003451 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003452 long long sum_time;
3453 long min_time;
3454 long max_time;
3455 long min_pid;
3456 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303457 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003458 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003459};
3460
3461struct loc_track {
3462 unsigned long max;
3463 unsigned long count;
3464 struct location *loc;
3465};
3466
3467static void free_loc_track(struct loc_track *t)
3468{
3469 if (t->max)
3470 free_pages((unsigned long)t->loc,
3471 get_order(sizeof(struct location) * t->max));
3472}
3473
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003474static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003475{
3476 struct location *l;
3477 int order;
3478
Christoph Lameter88a420e2007-05-06 14:49:45 -07003479 order = get_order(sizeof(struct location) * max);
3480
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003481 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003482 if (!l)
3483 return 0;
3484
3485 if (t->count) {
3486 memcpy(l, t->loc, sizeof(struct location) * t->count);
3487 free_loc_track(t);
3488 }
3489 t->max = max;
3490 t->loc = l;
3491 return 1;
3492}
3493
3494static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003495 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003496{
3497 long start, end, pos;
3498 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003499 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003500 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003501
3502 start = -1;
3503 end = t->count;
3504
3505 for ( ; ; ) {
3506 pos = start + (end - start + 1) / 2;
3507
3508 /*
3509 * There is nothing at "end". If we end up there
3510 * we need to add something to before end.
3511 */
3512 if (pos == end)
3513 break;
3514
3515 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003516 if (track->addr == caddr) {
3517
3518 l = &t->loc[pos];
3519 l->count++;
3520 if (track->when) {
3521 l->sum_time += age;
3522 if (age < l->min_time)
3523 l->min_time = age;
3524 if (age > l->max_time)
3525 l->max_time = age;
3526
3527 if (track->pid < l->min_pid)
3528 l->min_pid = track->pid;
3529 if (track->pid > l->max_pid)
3530 l->max_pid = track->pid;
3531
Rusty Russell174596a2009-01-01 10:12:29 +10303532 cpumask_set_cpu(track->cpu,
3533 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003534 }
3535 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003536 return 1;
3537 }
3538
Christoph Lameter45edfa52007-05-09 02:32:45 -07003539 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003540 end = pos;
3541 else
3542 start = pos;
3543 }
3544
3545 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003546 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003547 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003548 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003549 return 0;
3550
3551 l = t->loc + pos;
3552 if (pos < t->count)
3553 memmove(l + 1, l,
3554 (t->count - pos) * sizeof(struct location));
3555 t->count++;
3556 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003557 l->addr = track->addr;
3558 l->sum_time = age;
3559 l->min_time = age;
3560 l->max_time = age;
3561 l->min_pid = track->pid;
3562 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303563 cpumask_clear(to_cpumask(l->cpus));
3564 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003565 nodes_clear(l->nodes);
3566 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003567 return 1;
3568}
3569
3570static void process_slab(struct loc_track *t, struct kmem_cache *s,
3571 struct page *page, enum track_item alloc)
3572{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003573 void *addr = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +03003574 DECLARE_BITMAP(map, page->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003575 void *p;
3576
Christoph Lameter39b26462008-04-14 19:11:30 +03003577 bitmap_zero(map, page->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003578 for_each_free_object(p, s, page->freelist)
3579 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003580
Christoph Lameter224a88b2008-04-14 19:11:31 +03003581 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003582 if (!test_bit(slab_index(p, s, addr), map))
3583 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003584}
3585
3586static int list_locations(struct kmem_cache *s, char *buf,
3587 enum track_item alloc)
3588{
Harvey Harrisone374d482008-01-31 15:20:50 -08003589 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003590 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003591 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003592 int node;
3593
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003594 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003595 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003596 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003597
3598 /* Push back cpu slabs */
3599 flush_all(s);
3600
Christoph Lameterf64dc582007-10-16 01:25:33 -07003601 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003602 struct kmem_cache_node *n = get_node(s, node);
3603 unsigned long flags;
3604 struct page *page;
3605
Christoph Lameter9e869432007-08-22 14:01:56 -07003606 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003607 continue;
3608
3609 spin_lock_irqsave(&n->list_lock, flags);
3610 list_for_each_entry(page, &n->partial, lru)
3611 process_slab(&t, s, page, alloc);
3612 list_for_each_entry(page, &n->full, lru)
3613 process_slab(&t, s, page, alloc);
3614 spin_unlock_irqrestore(&n->list_lock, flags);
3615 }
3616
3617 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003618 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003619
Hugh Dickins9c246242008-12-09 13:14:27 -08003620 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003621 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003622 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003623
3624 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003625 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003626 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003627 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003628
3629 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08003630 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07003631 l->min_time,
3632 (long)div_u64(l->sum_time, l->count),
3633 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003634 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003635 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003636 l->min_time);
3637
3638 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003639 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003640 l->min_pid, l->max_pid);
3641 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003642 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003643 l->min_pid);
3644
Rusty Russell174596a2009-01-01 10:12:29 +10303645 if (num_online_cpus() > 1 &&
3646 !cpumask_empty(to_cpumask(l->cpus)) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003647 len < PAGE_SIZE - 60) {
3648 len += sprintf(buf + len, " cpus=");
3649 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Rusty Russell174596a2009-01-01 10:12:29 +10303650 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07003651 }
3652
Christoph Lameter84966342007-06-23 17:16:32 -07003653 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003654 len < PAGE_SIZE - 60) {
3655 len += sprintf(buf + len, " nodes=");
3656 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003657 l->nodes);
3658 }
3659
Harvey Harrisone374d482008-01-31 15:20:50 -08003660 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003661 }
3662
3663 free_loc_track(&t);
3664 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003665 len += sprintf(buf, "No data\n");
3666 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003667}
3668
Christoph Lameter81819f02007-05-06 14:49:36 -07003669enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03003670 SL_ALL, /* All slabs */
3671 SL_PARTIAL, /* Only partially allocated slabs */
3672 SL_CPU, /* Only slabs used for cpu caches */
3673 SL_OBJECTS, /* Determine allocated objects not slabs */
3674 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07003675};
3676
Christoph Lameter205ab992008-04-14 19:11:40 +03003677#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003678#define SO_PARTIAL (1 << SL_PARTIAL)
3679#define SO_CPU (1 << SL_CPU)
3680#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03003681#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003682
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003683static ssize_t show_slab_objects(struct kmem_cache *s,
3684 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003685{
3686 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003687 int node;
3688 int x;
3689 unsigned long *nodes;
3690 unsigned long *per_cpu;
3691
3692 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003693 if (!nodes)
3694 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003695 per_cpu = nodes + nr_node_ids;
3696
Christoph Lameter205ab992008-04-14 19:11:40 +03003697 if (flags & SO_CPU) {
3698 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07003699
Christoph Lameter205ab992008-04-14 19:11:40 +03003700 for_each_possible_cpu(cpu) {
3701 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003702
Christoph Lameter205ab992008-04-14 19:11:40 +03003703 if (!c || c->node < 0)
3704 continue;
3705
3706 if (c->page) {
3707 if (flags & SO_TOTAL)
3708 x = c->page->objects;
3709 else if (flags & SO_OBJECTS)
3710 x = c->page->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07003711 else
3712 x = 1;
Christoph Lameter205ab992008-04-14 19:11:40 +03003713
Christoph Lameter81819f02007-05-06 14:49:36 -07003714 total += x;
Christoph Lameter205ab992008-04-14 19:11:40 +03003715 nodes[c->node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003716 }
Christoph Lameter205ab992008-04-14 19:11:40 +03003717 per_cpu[c->node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003718 }
3719 }
3720
Christoph Lameter205ab992008-04-14 19:11:40 +03003721 if (flags & SO_ALL) {
3722 for_each_node_state(node, N_NORMAL_MEMORY) {
3723 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003724
Christoph Lameter205ab992008-04-14 19:11:40 +03003725 if (flags & SO_TOTAL)
3726 x = atomic_long_read(&n->total_objects);
3727 else if (flags & SO_OBJECTS)
3728 x = atomic_long_read(&n->total_objects) -
3729 count_partial(n, count_free);
3730
3731 else
3732 x = atomic_long_read(&n->nr_slabs);
3733 total += x;
3734 nodes[node] += x;
3735 }
3736
3737 } else if (flags & SO_PARTIAL) {
3738 for_each_node_state(node, N_NORMAL_MEMORY) {
3739 struct kmem_cache_node *n = get_node(s, node);
3740
3741 if (flags & SO_TOTAL)
3742 x = count_partial(n, count_total);
3743 else if (flags & SO_OBJECTS)
3744 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07003745 else
3746 x = n->nr_partial;
3747 total += x;
3748 nodes[node] += x;
3749 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003750 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003751 x = sprintf(buf, "%lu", total);
3752#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003753 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003754 if (nodes[node])
3755 x += sprintf(buf + x, " N%d=%lu",
3756 node, nodes[node]);
3757#endif
3758 kfree(nodes);
3759 return x + sprintf(buf + x, "\n");
3760}
3761
3762static int any_slab_objects(struct kmem_cache *s)
3763{
3764 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003765
3766 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003767 struct kmem_cache_node *n = get_node(s, node);
3768
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003769 if (!n)
3770 continue;
3771
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07003772 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07003773 return 1;
3774 }
3775 return 0;
3776}
3777
3778#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3779#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3780
3781struct slab_attribute {
3782 struct attribute attr;
3783 ssize_t (*show)(struct kmem_cache *s, char *buf);
3784 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3785};
3786
3787#define SLAB_ATTR_RO(_name) \
3788 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3789
3790#define SLAB_ATTR(_name) \
3791 static struct slab_attribute _name##_attr = \
3792 __ATTR(_name, 0644, _name##_show, _name##_store)
3793
Christoph Lameter81819f02007-05-06 14:49:36 -07003794static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3795{
3796 return sprintf(buf, "%d\n", s->size);
3797}
3798SLAB_ATTR_RO(slab_size);
3799
3800static ssize_t align_show(struct kmem_cache *s, char *buf)
3801{
3802 return sprintf(buf, "%d\n", s->align);
3803}
3804SLAB_ATTR_RO(align);
3805
3806static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3807{
3808 return sprintf(buf, "%d\n", s->objsize);
3809}
3810SLAB_ATTR_RO(object_size);
3811
3812static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3813{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003814 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003815}
3816SLAB_ATTR_RO(objs_per_slab);
3817
Christoph Lameter06b285d2008-04-14 19:11:41 +03003818static ssize_t order_store(struct kmem_cache *s,
3819 const char *buf, size_t length)
3820{
Christoph Lameter0121c6192008-04-29 16:11:12 -07003821 unsigned long order;
3822 int err;
3823
3824 err = strict_strtoul(buf, 10, &order);
3825 if (err)
3826 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003827
3828 if (order > slub_max_order || order < slub_min_order)
3829 return -EINVAL;
3830
3831 calculate_sizes(s, order);
3832 return length;
3833}
3834
Christoph Lameter81819f02007-05-06 14:49:36 -07003835static ssize_t order_show(struct kmem_cache *s, char *buf)
3836{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003837 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003838}
Christoph Lameter06b285d2008-04-14 19:11:41 +03003839SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003840
David Rientjes73d342b2009-02-22 17:40:09 -08003841static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3842{
3843 return sprintf(buf, "%lu\n", s->min_partial);
3844}
3845
3846static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3847 size_t length)
3848{
3849 unsigned long min;
3850 int err;
3851
3852 err = strict_strtoul(buf, 10, &min);
3853 if (err)
3854 return err;
3855
David Rientjesc0bdb232009-02-25 09:16:35 +02003856 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08003857 return length;
3858}
3859SLAB_ATTR(min_partial);
3860
Christoph Lameter81819f02007-05-06 14:49:36 -07003861static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3862{
3863 if (s->ctor) {
3864 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3865
3866 return n + sprintf(buf + n, "\n");
3867 }
3868 return 0;
3869}
3870SLAB_ATTR_RO(ctor);
3871
Christoph Lameter81819f02007-05-06 14:49:36 -07003872static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3873{
3874 return sprintf(buf, "%d\n", s->refcount - 1);
3875}
3876SLAB_ATTR_RO(aliases);
3877
3878static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3879{
Christoph Lameter205ab992008-04-14 19:11:40 +03003880 return show_slab_objects(s, buf, SO_ALL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003881}
3882SLAB_ATTR_RO(slabs);
3883
3884static ssize_t partial_show(struct kmem_cache *s, char *buf)
3885{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003886 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003887}
3888SLAB_ATTR_RO(partial);
3889
3890static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3891{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003892 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003893}
3894SLAB_ATTR_RO(cpu_slabs);
3895
3896static ssize_t objects_show(struct kmem_cache *s, char *buf)
3897{
Christoph Lameter205ab992008-04-14 19:11:40 +03003898 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003899}
3900SLAB_ATTR_RO(objects);
3901
Christoph Lameter205ab992008-04-14 19:11:40 +03003902static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3903{
3904 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3905}
3906SLAB_ATTR_RO(objects_partial);
3907
3908static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3909{
3910 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3911}
3912SLAB_ATTR_RO(total_objects);
3913
Christoph Lameter81819f02007-05-06 14:49:36 -07003914static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3915{
3916 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3917}
3918
3919static ssize_t sanity_checks_store(struct kmem_cache *s,
3920 const char *buf, size_t length)
3921{
3922 s->flags &= ~SLAB_DEBUG_FREE;
3923 if (buf[0] == '1')
3924 s->flags |= SLAB_DEBUG_FREE;
3925 return length;
3926}
3927SLAB_ATTR(sanity_checks);
3928
3929static ssize_t trace_show(struct kmem_cache *s, char *buf)
3930{
3931 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3932}
3933
3934static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3935 size_t length)
3936{
3937 s->flags &= ~SLAB_TRACE;
3938 if (buf[0] == '1')
3939 s->flags |= SLAB_TRACE;
3940 return length;
3941}
3942SLAB_ATTR(trace);
3943
3944static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3945{
3946 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3947}
3948
3949static ssize_t reclaim_account_store(struct kmem_cache *s,
3950 const char *buf, size_t length)
3951{
3952 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3953 if (buf[0] == '1')
3954 s->flags |= SLAB_RECLAIM_ACCOUNT;
3955 return length;
3956}
3957SLAB_ATTR(reclaim_account);
3958
3959static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3960{
Christoph Lameter5af60832007-05-06 14:49:56 -07003961 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003962}
3963SLAB_ATTR_RO(hwcache_align);
3964
3965#ifdef CONFIG_ZONE_DMA
3966static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3967{
3968 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3969}
3970SLAB_ATTR_RO(cache_dma);
3971#endif
3972
3973static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3974{
3975 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3976}
3977SLAB_ATTR_RO(destroy_by_rcu);
3978
3979static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3980{
3981 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3982}
3983
3984static ssize_t red_zone_store(struct kmem_cache *s,
3985 const char *buf, size_t length)
3986{
3987 if (any_slab_objects(s))
3988 return -EBUSY;
3989
3990 s->flags &= ~SLAB_RED_ZONE;
3991 if (buf[0] == '1')
3992 s->flags |= SLAB_RED_ZONE;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003993 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003994 return length;
3995}
3996SLAB_ATTR(red_zone);
3997
3998static ssize_t poison_show(struct kmem_cache *s, char *buf)
3999{
4000 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4001}
4002
4003static ssize_t poison_store(struct kmem_cache *s,
4004 const char *buf, size_t length)
4005{
4006 if (any_slab_objects(s))
4007 return -EBUSY;
4008
4009 s->flags &= ~SLAB_POISON;
4010 if (buf[0] == '1')
4011 s->flags |= SLAB_POISON;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004012 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004013 return length;
4014}
4015SLAB_ATTR(poison);
4016
4017static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4018{
4019 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4020}
4021
4022static ssize_t store_user_store(struct kmem_cache *s,
4023 const char *buf, size_t length)
4024{
4025 if (any_slab_objects(s))
4026 return -EBUSY;
4027
4028 s->flags &= ~SLAB_STORE_USER;
4029 if (buf[0] == '1')
4030 s->flags |= SLAB_STORE_USER;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004031 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004032 return length;
4033}
4034SLAB_ATTR(store_user);
4035
Christoph Lameter53e15af2007-05-06 14:49:43 -07004036static ssize_t validate_show(struct kmem_cache *s, char *buf)
4037{
4038 return 0;
4039}
4040
4041static ssize_t validate_store(struct kmem_cache *s,
4042 const char *buf, size_t length)
4043{
Christoph Lameter434e2452007-07-17 04:03:30 -07004044 int ret = -EINVAL;
4045
4046 if (buf[0] == '1') {
4047 ret = validate_slab_cache(s);
4048 if (ret >= 0)
4049 ret = length;
4050 }
4051 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004052}
4053SLAB_ATTR(validate);
4054
Christoph Lameter2086d262007-05-06 14:49:46 -07004055static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4056{
4057 return 0;
4058}
4059
4060static ssize_t shrink_store(struct kmem_cache *s,
4061 const char *buf, size_t length)
4062{
4063 if (buf[0] == '1') {
4064 int rc = kmem_cache_shrink(s);
4065
4066 if (rc)
4067 return rc;
4068 } else
4069 return -EINVAL;
4070 return length;
4071}
4072SLAB_ATTR(shrink);
4073
Christoph Lameter88a420e2007-05-06 14:49:45 -07004074static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4075{
4076 if (!(s->flags & SLAB_STORE_USER))
4077 return -ENOSYS;
4078 return list_locations(s, buf, TRACK_ALLOC);
4079}
4080SLAB_ATTR_RO(alloc_calls);
4081
4082static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4083{
4084 if (!(s->flags & SLAB_STORE_USER))
4085 return -ENOSYS;
4086 return list_locations(s, buf, TRACK_FREE);
4087}
4088SLAB_ATTR_RO(free_calls);
4089
Christoph Lameter81819f02007-05-06 14:49:36 -07004090#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004091static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004092{
Christoph Lameter98246012008-01-07 23:20:26 -08004093 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004094}
4095
Christoph Lameter98246012008-01-07 23:20:26 -08004096static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004097 const char *buf, size_t length)
4098{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004099 unsigned long ratio;
4100 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004101
Christoph Lameter0121c6192008-04-29 16:11:12 -07004102 err = strict_strtoul(buf, 10, &ratio);
4103 if (err)
4104 return err;
4105
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004106 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004107 s->remote_node_defrag_ratio = ratio * 10;
4108
Christoph Lameter81819f02007-05-06 14:49:36 -07004109 return length;
4110}
Christoph Lameter98246012008-01-07 23:20:26 -08004111SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004112#endif
4113
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004114#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004115static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4116{
4117 unsigned long sum = 0;
4118 int cpu;
4119 int len;
4120 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4121
4122 if (!data)
4123 return -ENOMEM;
4124
4125 for_each_online_cpu(cpu) {
4126 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4127
4128 data[cpu] = x;
4129 sum += x;
4130 }
4131
4132 len = sprintf(buf, "%lu", sum);
4133
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004134#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004135 for_each_online_cpu(cpu) {
4136 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004137 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004138 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004139#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004140 kfree(data);
4141 return len + sprintf(buf + len, "\n");
4142}
4143
4144#define STAT_ATTR(si, text) \
4145static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4146{ \
4147 return show_stat(s, buf, si); \
4148} \
4149SLAB_ATTR_RO(text); \
4150
4151STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4152STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4153STAT_ATTR(FREE_FASTPATH, free_fastpath);
4154STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4155STAT_ATTR(FREE_FROZEN, free_frozen);
4156STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4157STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4158STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4159STAT_ATTR(ALLOC_SLAB, alloc_slab);
4160STAT_ATTR(ALLOC_REFILL, alloc_refill);
4161STAT_ATTR(FREE_SLAB, free_slab);
4162STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4163STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4164STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4165STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4166STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4167STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter65c33762008-04-14 19:11:40 +03004168STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004169#endif
4170
Pekka Enberg06428782008-01-07 23:20:27 -08004171static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004172 &slab_size_attr.attr,
4173 &object_size_attr.attr,
4174 &objs_per_slab_attr.attr,
4175 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08004176 &min_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004177 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004178 &objects_partial_attr.attr,
4179 &total_objects_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004180 &slabs_attr.attr,
4181 &partial_attr.attr,
4182 &cpu_slabs_attr.attr,
4183 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004184 &aliases_attr.attr,
4185 &align_attr.attr,
4186 &sanity_checks_attr.attr,
4187 &trace_attr.attr,
4188 &hwcache_align_attr.attr,
4189 &reclaim_account_attr.attr,
4190 &destroy_by_rcu_attr.attr,
4191 &red_zone_attr.attr,
4192 &poison_attr.attr,
4193 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004194 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004195 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004196 &alloc_calls_attr.attr,
4197 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004198#ifdef CONFIG_ZONE_DMA
4199 &cache_dma_attr.attr,
4200#endif
4201#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004202 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004203#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004204#ifdef CONFIG_SLUB_STATS
4205 &alloc_fastpath_attr.attr,
4206 &alloc_slowpath_attr.attr,
4207 &free_fastpath_attr.attr,
4208 &free_slowpath_attr.attr,
4209 &free_frozen_attr.attr,
4210 &free_add_partial_attr.attr,
4211 &free_remove_partial_attr.attr,
4212 &alloc_from_partial_attr.attr,
4213 &alloc_slab_attr.attr,
4214 &alloc_refill_attr.attr,
4215 &free_slab_attr.attr,
4216 &cpuslab_flush_attr.attr,
4217 &deactivate_full_attr.attr,
4218 &deactivate_empty_attr.attr,
4219 &deactivate_to_head_attr.attr,
4220 &deactivate_to_tail_attr.attr,
4221 &deactivate_remote_frees_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004222 &order_fallback_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004223#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004224 NULL
4225};
4226
4227static struct attribute_group slab_attr_group = {
4228 .attrs = slab_attrs,
4229};
4230
4231static ssize_t slab_attr_show(struct kobject *kobj,
4232 struct attribute *attr,
4233 char *buf)
4234{
4235 struct slab_attribute *attribute;
4236 struct kmem_cache *s;
4237 int err;
4238
4239 attribute = to_slab_attr(attr);
4240 s = to_slab(kobj);
4241
4242 if (!attribute->show)
4243 return -EIO;
4244
4245 err = attribute->show(s, buf);
4246
4247 return err;
4248}
4249
4250static ssize_t slab_attr_store(struct kobject *kobj,
4251 struct attribute *attr,
4252 const char *buf, size_t len)
4253{
4254 struct slab_attribute *attribute;
4255 struct kmem_cache *s;
4256 int err;
4257
4258 attribute = to_slab_attr(attr);
4259 s = to_slab(kobj);
4260
4261 if (!attribute->store)
4262 return -EIO;
4263
4264 err = attribute->store(s, buf, len);
4265
4266 return err;
4267}
4268
Christoph Lameter151c6022008-01-07 22:29:05 -08004269static void kmem_cache_release(struct kobject *kobj)
4270{
4271 struct kmem_cache *s = to_slab(kobj);
4272
4273 kfree(s);
4274}
4275
Christoph Lameter81819f02007-05-06 14:49:36 -07004276static struct sysfs_ops slab_sysfs_ops = {
4277 .show = slab_attr_show,
4278 .store = slab_attr_store,
4279};
4280
4281static struct kobj_type slab_ktype = {
4282 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004283 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004284};
4285
4286static int uevent_filter(struct kset *kset, struct kobject *kobj)
4287{
4288 struct kobj_type *ktype = get_ktype(kobj);
4289
4290 if (ktype == &slab_ktype)
4291 return 1;
4292 return 0;
4293}
4294
4295static struct kset_uevent_ops slab_uevent_ops = {
4296 .filter = uevent_filter,
4297};
4298
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004299static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004300
4301#define ID_STR_LENGTH 64
4302
4303/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004304 *
4305 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004306 */
4307static char *create_unique_id(struct kmem_cache *s)
4308{
4309 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4310 char *p = name;
4311
4312 BUG_ON(!name);
4313
4314 *p++ = ':';
4315 /*
4316 * First flags affecting slabcache operations. We will only
4317 * get here for aliasable slabs so we do not need to support
4318 * too many flags. The flags here must cover all flags that
4319 * are matched during merging to guarantee that the id is
4320 * unique.
4321 */
4322 if (s->flags & SLAB_CACHE_DMA)
4323 *p++ = 'd';
4324 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4325 *p++ = 'a';
4326 if (s->flags & SLAB_DEBUG_FREE)
4327 *p++ = 'F';
4328 if (p != name + 1)
4329 *p++ = '-';
4330 p += sprintf(p, "%07d", s->size);
4331 BUG_ON(p > name + ID_STR_LENGTH - 1);
4332 return name;
4333}
4334
4335static int sysfs_slab_add(struct kmem_cache *s)
4336{
4337 int err;
4338 const char *name;
4339 int unmergeable;
4340
4341 if (slab_state < SYSFS)
4342 /* Defer until later */
4343 return 0;
4344
4345 unmergeable = slab_unmergeable(s);
4346 if (unmergeable) {
4347 /*
4348 * Slabcache can never be merged so we can use the name proper.
4349 * This is typically the case for debug situations. In that
4350 * case we can catch duplicate names easily.
4351 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004352 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004353 name = s->name;
4354 } else {
4355 /*
4356 * Create a unique name for the slab as a target
4357 * for the symlinks.
4358 */
4359 name = create_unique_id(s);
4360 }
4361
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004362 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004363 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4364 if (err) {
4365 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004366 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004367 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004368
4369 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4370 if (err)
4371 return err;
4372 kobject_uevent(&s->kobj, KOBJ_ADD);
4373 if (!unmergeable) {
4374 /* Setup first alias */
4375 sysfs_slab_alias(s, s->name);
4376 kfree(name);
4377 }
4378 return 0;
4379}
4380
4381static void sysfs_slab_remove(struct kmem_cache *s)
4382{
4383 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4384 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004385 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004386}
4387
4388/*
4389 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11004390 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07004391 */
4392struct saved_alias {
4393 struct kmem_cache *s;
4394 const char *name;
4395 struct saved_alias *next;
4396};
4397
Adrian Bunk5af328a2007-07-17 04:03:27 -07004398static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004399
4400static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4401{
4402 struct saved_alias *al;
4403
4404 if (slab_state == SYSFS) {
4405 /*
4406 * If we have a leftover link then remove it.
4407 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004408 sysfs_remove_link(&slab_kset->kobj, name);
4409 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004410 }
4411
4412 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4413 if (!al)
4414 return -ENOMEM;
4415
4416 al->s = s;
4417 al->name = name;
4418 al->next = alias_list;
4419 alias_list = al;
4420 return 0;
4421}
4422
4423static int __init slab_sysfs_init(void)
4424{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004425 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004426 int err;
4427
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004428 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004429 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004430 printk(KERN_ERR "Cannot register slab subsystem.\n");
4431 return -ENOSYS;
4432 }
4433
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004434 slab_state = SYSFS;
4435
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004436 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004437 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004438 if (err)
4439 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4440 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004441 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004442
4443 while (alias_list) {
4444 struct saved_alias *al = alias_list;
4445
4446 alias_list = alias_list->next;
4447 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004448 if (err)
4449 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4450 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004451 kfree(al);
4452 }
4453
4454 resiliency_test();
4455 return 0;
4456}
4457
4458__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004459#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004460
4461/*
4462 * The /proc/slabinfo ABI
4463 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004464#ifdef CONFIG_SLABINFO
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004465static void print_slabinfo_header(struct seq_file *m)
4466{
4467 seq_puts(m, "slabinfo - version: 2.1\n");
4468 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4469 "<objperslab> <pagesperslab>");
4470 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4471 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4472 seq_putc(m, '\n');
4473}
4474
4475static void *s_start(struct seq_file *m, loff_t *pos)
4476{
4477 loff_t n = *pos;
4478
4479 down_read(&slub_lock);
4480 if (!n)
4481 print_slabinfo_header(m);
4482
4483 return seq_list_start(&slab_caches, *pos);
4484}
4485
4486static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4487{
4488 return seq_list_next(p, &slab_caches, pos);
4489}
4490
4491static void s_stop(struct seq_file *m, void *p)
4492{
4493 up_read(&slub_lock);
4494}
4495
4496static int s_show(struct seq_file *m, void *p)
4497{
4498 unsigned long nr_partials = 0;
4499 unsigned long nr_slabs = 0;
4500 unsigned long nr_inuse = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004501 unsigned long nr_objs = 0;
4502 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004503 struct kmem_cache *s;
4504 int node;
4505
4506 s = list_entry(p, struct kmem_cache, list);
4507
4508 for_each_online_node(node) {
4509 struct kmem_cache_node *n = get_node(s, node);
4510
4511 if (!n)
4512 continue;
4513
4514 nr_partials += n->nr_partial;
4515 nr_slabs += atomic_long_read(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03004516 nr_objs += atomic_long_read(&n->total_objects);
4517 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004518 }
4519
Christoph Lameter205ab992008-04-14 19:11:40 +03004520 nr_inuse = nr_objs - nr_free;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004521
4522 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
Christoph Lameter834f3d12008-04-14 19:11:31 +03004523 nr_objs, s->size, oo_objects(s->oo),
4524 (1 << oo_order(s->oo)));
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004525 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4526 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4527 0UL);
4528 seq_putc(m, '\n');
4529 return 0;
4530}
4531
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004532static const struct seq_operations slabinfo_op = {
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004533 .start = s_start,
4534 .next = s_next,
4535 .stop = s_stop,
4536 .show = s_show,
4537};
4538
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004539static int slabinfo_open(struct inode *inode, struct file *file)
4540{
4541 return seq_open(file, &slabinfo_op);
4542}
4543
4544static const struct file_operations proc_slabinfo_operations = {
4545 .open = slabinfo_open,
4546 .read = seq_read,
4547 .llseek = seq_lseek,
4548 .release = seq_release,
4549};
4550
4551static int __init slab_proc_init(void)
4552{
4553 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4554 return 0;
4555}
4556module_init(slab_proc_init);
Linus Torvalds158a9622008-01-02 13:04:48 -08004557#endif /* CONFIG_SLABINFO */