blob: 35ab38a94b46279dc018c14841452f533313702b [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>
17#include <linux/seq_file.h>
18#include <linux/cpu.h>
19#include <linux/cpuset.h>
20#include <linux/mempolicy.h>
21#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070022#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070023#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070024#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070025#include <linux/math64.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070026
27/*
28 * Lock order:
29 * 1. slab_lock(page)
30 * 2. slab->list_lock
31 *
32 * The slab_lock protects operations on the object of a particular
33 * slab and its metadata in the page struct. If the slab lock
34 * has been taken then no allocations nor frees can be performed
35 * on the objects in the slab nor can the slab be added or removed
36 * from the partial or full lists since this would mean modifying
37 * the page_struct of the slab.
38 *
39 * The list_lock protects the partial and full list on each node and
40 * the partial slab counter. If taken then no new slabs may be added or
41 * removed from the lists nor make the number of partial slabs be modified.
42 * (Note that the total number of slabs is an atomic value that may be
43 * modified without taking the list lock).
44 *
45 * The list_lock is a centralized lock and thus we avoid taking it as
46 * much as possible. As long as SLUB does not have to handle partial
47 * slabs, operations can continue without any centralized lock. F.e.
48 * allocating a long series of objects that fill up slabs does not require
49 * the list lock.
50 *
51 * The lock order is sometimes inverted when we are trying to get a slab
52 * off a list. We take the list_lock and then look for a page on the list
53 * to use. While we do that objects in the slabs may be freed. We can
54 * only operate on the slab if we have also taken the slab_lock. So we use
55 * a slab_trylock() on the slab. If trylock was successful then no frees
56 * can occur anymore and we can use the slab for allocations etc. If the
57 * slab_trylock() does not succeed then frees are in progress in the slab and
58 * we must stay away from it for a while since we may cause a bouncing
59 * cacheline if we try to acquire the lock. So go onto the next slab.
60 * If all pages are busy then we may allocate a new slab instead of reusing
61 * a partial slab. A new slab has noone operating on it and thus there is
62 * no danger of cacheline contention.
63 *
64 * Interrupts are disabled during allocation and deallocation in order to
65 * make the slab allocator safe to use in the context of an irq. In addition
66 * interrupts are disabled to ensure that the processor does not change
67 * while handling per_cpu slabs, due to kernel preemption.
68 *
69 * SLUB assigns one slab for allocation to each processor.
70 * Allocations only occur from these slabs called cpu slabs.
71 *
Christoph Lameter672bba32007-05-09 02:32:39 -070072 * Slabs with free elements are kept on a partial list and during regular
73 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070074 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070075 * We track full slabs for debugging purposes though because otherwise we
76 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070077 *
78 * Slabs are freed when they become empty. Teardown and setup is
79 * minimal so we rely on the page allocators per cpu caches for
80 * fast frees and allocs.
81 *
82 * Overloading of page flags that are otherwise used for LRU management.
83 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070084 * PageActive The slab is frozen and exempt from list processing.
85 * This means that the slab is dedicated to a purpose
86 * such as satisfying allocations for a specific
87 * processor. Objects may be freed in the slab while
88 * it is frozen but slab_free will then skip the usual
89 * list operations. It is up to the processor holding
90 * the slab to integrate the slab into the slab lists
91 * when the slab is no longer needed.
92 *
93 * One use of this flag is to mark slabs that are
94 * used for allocations. Then such a slab becomes a cpu
95 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -070096 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -070097 * free objects in addition to the regular freelist
98 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070099 *
100 * PageError Slab requires special handling due to debug
101 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700102 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700103 */
104
Christoph Lameter5577bd82007-05-16 22:10:56 -0700105#define FROZEN (1 << PG_active)
106
107#ifdef CONFIG_SLUB_DEBUG
108#define SLABDEBUG (1 << PG_error)
109#else
110#define SLABDEBUG 0
111#endif
112
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700113static inline int SlabFrozen(struct page *page)
114{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700115 return page->flags & FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700116}
117
118static inline void SetSlabFrozen(struct page *page)
119{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700120 page->flags |= FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700121}
122
123static inline void ClearSlabFrozen(struct page *page)
124{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700125 page->flags &= ~FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700126}
127
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700128static inline int SlabDebug(struct page *page)
129{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700130 return page->flags & SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700131}
132
133static inline void SetSlabDebug(struct page *page)
134{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700135 page->flags |= SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700136}
137
138static inline void ClearSlabDebug(struct page *page)
139{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700140 page->flags &= ~SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700141}
142
Christoph Lameter81819f02007-05-06 14:49:36 -0700143/*
144 * Issues still to be resolved:
145 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700146 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
147 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700148 * - Variable sizing of the per node arrays
149 */
150
151/* Enable to test recovery from slab corruption on boot */
152#undef SLUB_RESILIENCY_TEST
153
Christoph Lameter81819f02007-05-06 14:49:36 -0700154/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700155 * Mininum number of partial slabs. These will be left on the partial
156 * lists even if they are empty. kmem_cache_shrink may reclaim them.
157 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800158#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700159
Christoph Lameter2086d262007-05-06 14:49:46 -0700160/*
161 * Maximum number of desirable partial slabs.
162 * The existence of more partial slabs makes kmem_cache_shrink
163 * sort the partial list by the number of objects in the.
164 */
165#define MAX_PARTIAL 10
166
Christoph Lameter81819f02007-05-06 14:49:36 -0700167#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
168 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700169
Christoph Lameter81819f02007-05-06 14:49:36 -0700170/*
171 * Set of flags that will prevent slab merging
172 */
173#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
174 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
175
176#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
177 SLAB_CACHE_DMA)
178
179#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700180#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700181#endif
182
183#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700184#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700185#endif
186
187/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700188#define __OBJECT_POISON 0x80000000 /* Poison object */
189#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700190
191static int kmem_size = sizeof(struct kmem_cache);
192
193#ifdef CONFIG_SMP
194static struct notifier_block slab_notifier;
195#endif
196
197static enum {
198 DOWN, /* No slab functionality available */
199 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700200 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700201 SYSFS /* Sysfs up */
202} slab_state = DOWN;
203
204/* A list of all slab caches on the system */
205static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700206static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700207
Christoph Lameter02cbc872007-05-09 02:32:43 -0700208/*
209 * Tracking user of a slab.
210 */
211struct track {
212 void *addr; /* Called from address */
213 int cpu; /* Was running on cpu */
214 int pid; /* Pid context */
215 unsigned long when; /* When did the operation occur */
216};
217
218enum track_item { TRACK_ALLOC, TRACK_FREE };
219
Christoph Lameterf6acb632008-04-29 16:16:06 -0700220#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -0700221static int sysfs_slab_add(struct kmem_cache *);
222static int sysfs_slab_alias(struct kmem_cache *, const char *);
223static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800224
Christoph Lameter81819f02007-05-06 14:49:36 -0700225#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700226static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
227static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
228 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800229static inline void sysfs_slab_remove(struct kmem_cache *s)
230{
231 kfree(s);
232}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800233
Christoph Lameter81819f02007-05-06 14:49:36 -0700234#endif
235
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800236static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
237{
238#ifdef CONFIG_SLUB_STATS
239 c->stat[si]++;
240#endif
241}
242
Christoph Lameter81819f02007-05-06 14:49:36 -0700243/********************************************************************
244 * Core slab cache functions
245 *******************************************************************/
246
247int slab_is_available(void)
248{
249 return slab_state >= UP;
250}
251
252static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
253{
254#ifdef CONFIG_NUMA
255 return s->node[node];
256#else
257 return &s->local_node;
258#endif
259}
260
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700261static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
262{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700263#ifdef CONFIG_SMP
264 return s->cpu_slab[cpu];
265#else
266 return &s->cpu_slab;
267#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700268}
269
Christoph Lameter6446faa2008-02-15 23:45:26 -0800270/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700271static inline int check_valid_pointer(struct kmem_cache *s,
272 struct page *page, const void *object)
273{
274 void *base;
275
Christoph Lametera973e9d2008-03-01 13:40:44 -0800276 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700277 return 1;
278
Christoph Lametera973e9d2008-03-01 13:40:44 -0800279 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300280 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700281 (object - base) % s->size) {
282 return 0;
283 }
284
285 return 1;
286}
287
Christoph Lameter81819f02007-05-06 14:49:36 -0700288/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700289 * Slow version of get and set free pointer.
290 *
291 * This version requires touching the cache lines of kmem_cache which
292 * we avoid to do in the fast alloc free paths. There we obtain the offset
293 * from the page struct.
294 */
295static inline void *get_freepointer(struct kmem_cache *s, void *object)
296{
297 return *(void **)(object + s->offset);
298}
299
300static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
301{
302 *(void **)(object + s->offset) = fp;
303}
304
305/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300306#define for_each_object(__p, __s, __addr, __objects) \
307 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700308 __p += (__s)->size)
309
310/* Scan freelist */
311#define for_each_free_object(__p, __s, __free) \
Christoph Lametera973e9d2008-03-01 13:40:44 -0800312 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700313
314/* Determine object index from a given position */
315static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
316{
317 return (p - addr) / s->size;
318}
319
Christoph Lameter834f3d12008-04-14 19:11:31 +0300320static inline struct kmem_cache_order_objects oo_make(int order,
321 unsigned long size)
322{
323 struct kmem_cache_order_objects x = {
324 (order << 16) + (PAGE_SIZE << order) / size
325 };
326
327 return x;
328}
329
330static inline int oo_order(struct kmem_cache_order_objects x)
331{
332 return x.x >> 16;
333}
334
335static inline int oo_objects(struct kmem_cache_order_objects x)
336{
337 return x.x & ((1 << 16) - 1);
338}
339
Christoph Lameter41ecc552007-05-09 02:32:44 -0700340#ifdef CONFIG_SLUB_DEBUG
341/*
342 * Debug settings:
343 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700344#ifdef CONFIG_SLUB_DEBUG_ON
345static int slub_debug = DEBUG_DEFAULT_FLAGS;
346#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700347static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700348#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700349
350static char *slub_debug_slabs;
351
Christoph Lameter7656c722007-05-09 02:32:40 -0700352/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700353 * Object debugging
354 */
355static void print_section(char *text, u8 *addr, unsigned int length)
356{
357 int i, offset;
358 int newline = 1;
359 char ascii[17];
360
361 ascii[16] = 0;
362
363 for (i = 0; i < length; i++) {
364 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700365 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700366 newline = 0;
367 }
Pekka Enberg06428782008-01-07 23:20:27 -0800368 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700369 offset = i % 16;
370 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
371 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800372 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700373 newline = 1;
374 }
375 }
376 if (!newline) {
377 i %= 16;
378 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800379 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700380 ascii[i] = ' ';
381 i++;
382 }
Pekka Enberg06428782008-01-07 23:20:27 -0800383 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700384 }
385}
386
Christoph Lameter81819f02007-05-06 14:49:36 -0700387static struct track *get_track(struct kmem_cache *s, void *object,
388 enum track_item alloc)
389{
390 struct track *p;
391
392 if (s->offset)
393 p = object + s->offset + sizeof(void *);
394 else
395 p = object + s->inuse;
396
397 return p + alloc;
398}
399
400static void set_track(struct kmem_cache *s, void *object,
401 enum track_item alloc, void *addr)
402{
403 struct track *p;
404
405 if (s->offset)
406 p = object + s->offset + sizeof(void *);
407 else
408 p = object + s->inuse;
409
410 p += alloc;
411 if (addr) {
412 p->addr = addr;
413 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400414 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700415 p->when = jiffies;
416 } else
417 memset(p, 0, sizeof(struct track));
418}
419
Christoph Lameter81819f02007-05-06 14:49:36 -0700420static void init_tracking(struct kmem_cache *s, void *object)
421{
Christoph Lameter24922682007-07-17 04:03:18 -0700422 if (!(s->flags & SLAB_STORE_USER))
423 return;
424
425 set_track(s, object, TRACK_FREE, NULL);
426 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700427}
428
429static void print_track(const char *s, struct track *t)
430{
431 if (!t->addr)
432 return;
433
Linus Torvalds7daf7052008-07-14 12:12:53 -0700434 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
435 s, t->addr, jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700436}
437
Christoph Lameter24922682007-07-17 04:03:18 -0700438static void print_tracking(struct kmem_cache *s, void *object)
439{
440 if (!(s->flags & SLAB_STORE_USER))
441 return;
442
443 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
444 print_track("Freed", get_track(s, object, TRACK_FREE));
445}
446
447static void print_page_info(struct page *page)
448{
Christoph Lameter39b26462008-04-14 19:11:30 +0300449 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
450 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700451
452}
453
454static void slab_bug(struct kmem_cache *s, char *fmt, ...)
455{
456 va_list args;
457 char buf[100];
458
459 va_start(args, fmt);
460 vsnprintf(buf, sizeof(buf), fmt, args);
461 va_end(args);
462 printk(KERN_ERR "========================================"
463 "=====================================\n");
464 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
465 printk(KERN_ERR "----------------------------------------"
466 "-------------------------------------\n\n");
467}
468
469static void slab_fix(struct kmem_cache *s, char *fmt, ...)
470{
471 va_list args;
472 char buf[100];
473
474 va_start(args, fmt);
475 vsnprintf(buf, sizeof(buf), fmt, args);
476 va_end(args);
477 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
478}
479
480static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700481{
482 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800483 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700484
485 print_tracking(s, p);
486
487 print_page_info(page);
488
489 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
490 p, p - addr, get_freepointer(s, p));
491
492 if (p > addr + 16)
493 print_section("Bytes b4", p - 16, 16);
494
495 print_section("Object", p, min(s->objsize, 128));
Christoph Lameter81819f02007-05-06 14:49:36 -0700496
497 if (s->flags & SLAB_RED_ZONE)
498 print_section("Redzone", p + s->objsize,
499 s->inuse - s->objsize);
500
Christoph Lameter81819f02007-05-06 14:49:36 -0700501 if (s->offset)
502 off = s->offset + sizeof(void *);
503 else
504 off = s->inuse;
505
Christoph Lameter24922682007-07-17 04:03:18 -0700506 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700507 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700508
509 if (off != s->size)
510 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700511 print_section("Padding", p + off, s->size - off);
512
513 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700514}
515
516static void object_err(struct kmem_cache *s, struct page *page,
517 u8 *object, char *reason)
518{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700519 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700520 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700521}
522
Christoph Lameter24922682007-07-17 04:03:18 -0700523static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700524{
525 va_list args;
526 char buf[100];
527
Christoph Lameter24922682007-07-17 04:03:18 -0700528 va_start(args, fmt);
529 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700530 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700531 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700532 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700533 dump_stack();
534}
535
536static void init_object(struct kmem_cache *s, void *object, int active)
537{
538 u8 *p = object;
539
540 if (s->flags & __OBJECT_POISON) {
541 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800542 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700543 }
544
545 if (s->flags & SLAB_RED_ZONE)
546 memset(p + s->objsize,
547 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
548 s->inuse - s->objsize);
549}
550
Christoph Lameter24922682007-07-17 04:03:18 -0700551static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700552{
553 while (bytes) {
554 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700555 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700556 start++;
557 bytes--;
558 }
Christoph Lameter24922682007-07-17 04:03:18 -0700559 return NULL;
560}
561
562static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
563 void *from, void *to)
564{
565 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
566 memset(from, data, to - from);
567}
568
569static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
570 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800571 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700572{
573 u8 *fault;
574 u8 *end;
575
576 fault = check_bytes(start, value, bytes);
577 if (!fault)
578 return 1;
579
580 end = start + bytes;
581 while (end > fault && end[-1] == value)
582 end--;
583
584 slab_bug(s, "%s overwritten", what);
585 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
586 fault, end - 1, fault[0], value);
587 print_trailer(s, page, object);
588
589 restore_bytes(s, what, value, fault, end);
590 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700591}
592
Christoph Lameter81819f02007-05-06 14:49:36 -0700593/*
594 * Object layout:
595 *
596 * object address
597 * Bytes of the object to be managed.
598 * If the freepointer may overlay the object then the free
599 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700600 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700601 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
602 * 0xa5 (POISON_END)
603 *
604 * object + s->objsize
605 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700606 * Padding is extended by another word if Redzoning is enabled and
607 * objsize == inuse.
608 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700609 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
610 * 0xcc (RED_ACTIVE) for objects in use.
611 *
612 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700613 * Meta data starts here.
614 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700615 * A. Free pointer (if we cannot overwrite object on free)
616 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700617 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800618 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700619 * before the word boundary.
620 *
621 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700622 *
623 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700624 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700625 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700626 * If slabcaches are merged then the objsize and inuse boundaries are mostly
627 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700628 * may be used with merged slabcaches.
629 */
630
Christoph Lameter81819f02007-05-06 14:49:36 -0700631static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
632{
633 unsigned long off = s->inuse; /* The end of info */
634
635 if (s->offset)
636 /* Freepointer is placed after the object. */
637 off += sizeof(void *);
638
639 if (s->flags & SLAB_STORE_USER)
640 /* We also have user information there */
641 off += 2 * sizeof(struct track);
642
643 if (s->size == off)
644 return 1;
645
Christoph Lameter24922682007-07-17 04:03:18 -0700646 return check_bytes_and_report(s, page, p, "Object padding",
647 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700648}
649
Christoph Lameter39b26462008-04-14 19:11:30 +0300650/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700651static int slab_pad_check(struct kmem_cache *s, struct page *page)
652{
Christoph Lameter24922682007-07-17 04:03:18 -0700653 u8 *start;
654 u8 *fault;
655 u8 *end;
656 int length;
657 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700658
659 if (!(s->flags & SLAB_POISON))
660 return 1;
661
Christoph Lametera973e9d2008-03-01 13:40:44 -0800662 start = page_address(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +0300663 length = (PAGE_SIZE << compound_order(page));
Christoph Lameter39b26462008-04-14 19:11:30 +0300664 end = start + length;
665 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700666 if (!remainder)
667 return 1;
668
Christoph Lameter39b26462008-04-14 19:11:30 +0300669 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700670 if (!fault)
671 return 1;
672 while (end > fault && end[-1] == POISON_INUSE)
673 end--;
674
675 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Christoph Lameter39b26462008-04-14 19:11:30 +0300676 print_section("Padding", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700677
678 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
679 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700680}
681
682static int check_object(struct kmem_cache *s, struct page *page,
683 void *object, int active)
684{
685 u8 *p = object;
686 u8 *endobject = object + s->objsize;
687
688 if (s->flags & SLAB_RED_ZONE) {
689 unsigned int red =
690 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
691
Christoph Lameter24922682007-07-17 04:03:18 -0700692 if (!check_bytes_and_report(s, page, object, "Redzone",
693 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700694 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700695 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800696 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
697 check_bytes_and_report(s, page, p, "Alignment padding",
698 endobject, POISON_INUSE, s->inuse - s->objsize);
699 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700700 }
701
702 if (s->flags & SLAB_POISON) {
703 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700704 (!check_bytes_and_report(s, page, p, "Poison", p,
705 POISON_FREE, s->objsize - 1) ||
706 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800707 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700708 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700709 /*
710 * check_pad_bytes cleans up on its own.
711 */
712 check_pad_bytes(s, page, p);
713 }
714
715 if (!s->offset && active)
716 /*
717 * Object and freepointer overlap. Cannot check
718 * freepointer while object is allocated.
719 */
720 return 1;
721
722 /* Check free pointer validity */
723 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
724 object_err(s, page, p, "Freepointer corrupt");
725 /*
726 * No choice but to zap it and thus loose the remainder
727 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700728 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700729 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800730 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700731 return 0;
732 }
733 return 1;
734}
735
736static int check_slab(struct kmem_cache *s, struct page *page)
737{
Christoph Lameter39b26462008-04-14 19:11:30 +0300738 int maxobj;
739
Christoph Lameter81819f02007-05-06 14:49:36 -0700740 VM_BUG_ON(!irqs_disabled());
741
742 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700743 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700744 return 0;
745 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300746
747 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
748 if (page->objects > maxobj) {
749 slab_err(s, page, "objects %u > max %u",
750 s->name, page->objects, maxobj);
751 return 0;
752 }
753 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700754 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300755 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700756 return 0;
757 }
758 /* Slab_pad_check fixes things up after itself */
759 slab_pad_check(s, page);
760 return 1;
761}
762
763/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700764 * Determine if a certain object on a page is on the freelist. Must hold the
765 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700766 */
767static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
768{
769 int nr = 0;
770 void *fp = page->freelist;
771 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300772 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700773
Christoph Lameter39b26462008-04-14 19:11:30 +0300774 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700775 if (fp == search)
776 return 1;
777 if (!check_valid_pointer(s, page, fp)) {
778 if (object) {
779 object_err(s, page, object,
780 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800781 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700782 break;
783 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700784 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800785 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300786 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700787 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700788 return 0;
789 }
790 break;
791 }
792 object = fp;
793 fp = get_freepointer(s, object);
794 nr++;
795 }
796
Christoph Lameter224a88b2008-04-14 19:11:31 +0300797 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
798 if (max_objects > 65535)
799 max_objects = 65535;
800
801 if (page->objects != max_objects) {
802 slab_err(s, page, "Wrong number of objects. Found %d but "
803 "should be %d", page->objects, max_objects);
804 page->objects = max_objects;
805 slab_fix(s, "Number of objects adjusted.");
806 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300807 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700808 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300809 "counted were %d", page->inuse, page->objects - nr);
810 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700811 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700812 }
813 return search == NULL;
814}
815
Christoph Lameter0121c6192008-04-29 16:11:12 -0700816static void trace(struct kmem_cache *s, struct page *page, void *object,
817 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700818{
819 if (s->flags & SLAB_TRACE) {
820 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
821 s->name,
822 alloc ? "alloc" : "free",
823 object, page->inuse,
824 page->freelist);
825
826 if (!alloc)
827 print_section("Object", (void *)object, s->objsize);
828
829 dump_stack();
830 }
831}
832
Christoph Lameter643b1132007-05-06 14:49:42 -0700833/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700834 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700835 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700836static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700837{
Christoph Lameter643b1132007-05-06 14:49:42 -0700838 spin_lock(&n->list_lock);
839 list_add(&page->lru, &n->full);
840 spin_unlock(&n->list_lock);
841}
842
843static void remove_full(struct kmem_cache *s, struct page *page)
844{
845 struct kmem_cache_node *n;
846
847 if (!(s->flags & SLAB_STORE_USER))
848 return;
849
850 n = get_node(s, page_to_nid(page));
851
852 spin_lock(&n->list_lock);
853 list_del(&page->lru);
854 spin_unlock(&n->list_lock);
855}
856
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300857/* Tracking of the number of slabs for debugging purposes */
858static inline unsigned long slabs_node(struct kmem_cache *s, int node)
859{
860 struct kmem_cache_node *n = get_node(s, node);
861
862 return atomic_long_read(&n->nr_slabs);
863}
864
Christoph Lameter205ab992008-04-14 19:11:40 +0300865static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300866{
867 struct kmem_cache_node *n = get_node(s, node);
868
869 /*
870 * May be called early in order to allocate a slab for the
871 * kmem_cache_node structure. Solve the chicken-egg
872 * dilemma by deferring the increment of the count during
873 * bootstrap (see early_kmem_cache_node_alloc).
874 */
Christoph Lameter205ab992008-04-14 19:11:40 +0300875 if (!NUMA_BUILD || n) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300876 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300877 atomic_long_add(objects, &n->total_objects);
878 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300879}
Christoph Lameter205ab992008-04-14 19:11:40 +0300880static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300881{
882 struct kmem_cache_node *n = get_node(s, node);
883
884 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300885 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300886}
887
888/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -0700889static void setup_object_debug(struct kmem_cache *s, struct page *page,
890 void *object)
891{
892 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
893 return;
894
895 init_object(s, object, 0);
896 init_tracking(s, object);
897}
898
899static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
900 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700901{
902 if (!check_slab(s, page))
903 goto bad;
904
Christoph Lameterd692ef62008-02-15 23:45:24 -0800905 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700906 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700907 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700908 }
909
910 if (!check_valid_pointer(s, page, object)) {
911 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700912 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700913 }
914
Christoph Lameterd692ef62008-02-15 23:45:24 -0800915 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700916 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700917
Christoph Lameter3ec09742007-05-16 22:11:00 -0700918 /* Success perform special debug activities for allocs */
919 if (s->flags & SLAB_STORE_USER)
920 set_track(s, object, TRACK_ALLOC, addr);
921 trace(s, page, object, 1);
922 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700923 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700924
Christoph Lameter81819f02007-05-06 14:49:36 -0700925bad:
926 if (PageSlab(page)) {
927 /*
928 * If this is a slab page then lets do the best we can
929 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700930 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700931 */
Christoph Lameter24922682007-07-17 04:03:18 -0700932 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +0300933 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800934 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700935 }
936 return 0;
937}
938
Christoph Lameter3ec09742007-05-16 22:11:00 -0700939static int free_debug_processing(struct kmem_cache *s, struct page *page,
940 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700941{
942 if (!check_slab(s, page))
943 goto fail;
944
945 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700946 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700947 goto fail;
948 }
949
950 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700951 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700952 goto fail;
953 }
954
955 if (!check_object(s, page, object, 1))
956 return 0;
957
958 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800959 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700960 slab_err(s, page, "Attempt to free object(0x%p) "
961 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800962 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700963 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700964 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700965 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700966 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800967 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700968 object_err(s, page, object,
969 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700970 goto fail;
971 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700972
973 /* Special debug activities for freeing objects */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800974 if (!SlabFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700975 remove_full(s, page);
976 if (s->flags & SLAB_STORE_USER)
977 set_track(s, object, TRACK_FREE, addr);
978 trace(s, page, object, 0);
979 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700980 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700981
Christoph Lameter81819f02007-05-06 14:49:36 -0700982fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700983 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700984 return 0;
985}
986
Christoph Lameter41ecc552007-05-09 02:32:44 -0700987static int __init setup_slub_debug(char *str)
988{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700989 slub_debug = DEBUG_DEFAULT_FLAGS;
990 if (*str++ != '=' || !*str)
991 /*
992 * No options specified. Switch on full debugging.
993 */
994 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700995
996 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700997 /*
998 * No options but restriction on slabs. This means full
999 * debugging for slabs matching a pattern.
1000 */
1001 goto check_slabs;
1002
1003 slub_debug = 0;
1004 if (*str == '-')
1005 /*
1006 * Switch off all debugging measures.
1007 */
1008 goto out;
1009
1010 /*
1011 * Determine which debug features should be switched on
1012 */
Pekka Enberg06428782008-01-07 23:20:27 -08001013 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001014 switch (tolower(*str)) {
1015 case 'f':
1016 slub_debug |= SLAB_DEBUG_FREE;
1017 break;
1018 case 'z':
1019 slub_debug |= SLAB_RED_ZONE;
1020 break;
1021 case 'p':
1022 slub_debug |= SLAB_POISON;
1023 break;
1024 case 'u':
1025 slub_debug |= SLAB_STORE_USER;
1026 break;
1027 case 't':
1028 slub_debug |= SLAB_TRACE;
1029 break;
1030 default:
1031 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001032 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001033 }
1034 }
1035
1036check_slabs:
1037 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001038 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001039out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001040 return 1;
1041}
1042
1043__setup("slub_debug", setup_slub_debug);
1044
Christoph Lameterba0268a2007-09-11 15:24:11 -07001045static unsigned long kmem_cache_flags(unsigned long objsize,
1046 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001047 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001048{
1049 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001050 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001051 */
Christoph Lametere1533622008-02-15 23:45:24 -08001052 if (slub_debug && (!slub_debug_slabs ||
1053 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1054 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001055
1056 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001057}
1058#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001059static inline void setup_object_debug(struct kmem_cache *s,
1060 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001061
Christoph Lameter3ec09742007-05-16 22:11:00 -07001062static inline int alloc_debug_processing(struct kmem_cache *s,
1063 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001064
Christoph Lameter3ec09742007-05-16 22:11:00 -07001065static inline int free_debug_processing(struct kmem_cache *s,
1066 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001067
Christoph Lameter41ecc552007-05-09 02:32:44 -07001068static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1069 { return 1; }
1070static inline int check_object(struct kmem_cache *s, struct page *page,
1071 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001072static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001073static inline unsigned long kmem_cache_flags(unsigned long objsize,
1074 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001075 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001076{
1077 return flags;
1078}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001079#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001080
1081static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1082 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001083static inline void inc_slabs_node(struct kmem_cache *s, int node,
1084 int objects) {}
1085static inline void dec_slabs_node(struct kmem_cache *s, int node,
1086 int objects) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001087#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03001088
Christoph Lameter81819f02007-05-06 14:49:36 -07001089/*
1090 * Slab allocation and freeing
1091 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001092static inline struct page *alloc_slab_page(gfp_t flags, int node,
1093 struct kmem_cache_order_objects oo)
1094{
1095 int order = oo_order(oo);
1096
1097 if (node == -1)
1098 return alloc_pages(flags, order);
1099 else
1100 return alloc_pages_node(node, flags, order);
1101}
1102
Christoph Lameter81819f02007-05-06 14:49:36 -07001103static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1104{
Pekka Enberg06428782008-01-07 23:20:27 -08001105 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001106 struct kmem_cache_order_objects oo = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07001107
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001108 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001109
Christoph Lameter65c33762008-04-14 19:11:40 +03001110 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1111 oo);
1112 if (unlikely(!page)) {
1113 oo = s->min;
1114 /*
1115 * Allocation may have failed due to fragmentation.
1116 * Try a lower order alloc if possible
1117 */
1118 page = alloc_slab_page(flags, node, oo);
1119 if (!page)
1120 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001121
Christoph Lameter65c33762008-04-14 19:11:40 +03001122 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1123 }
Christoph Lameter834f3d12008-04-14 19:11:31 +03001124 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001125 mod_zone_page_state(page_zone(page),
1126 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1127 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001128 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001129
1130 return page;
1131}
1132
1133static void setup_object(struct kmem_cache *s, struct page *page,
1134 void *object)
1135{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001136 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001137 if (unlikely(s->ctor))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001138 s->ctor(s, object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001139}
1140
1141static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1142{
1143 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001144 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001145 void *last;
1146 void *p;
1147
Christoph Lameter6cb06222007-10-16 01:25:41 -07001148 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001149
Christoph Lameter6cb06222007-10-16 01:25:41 -07001150 page = allocate_slab(s,
1151 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001152 if (!page)
1153 goto out;
1154
Christoph Lameter205ab992008-04-14 19:11:40 +03001155 inc_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001156 page->slab = s;
1157 page->flags |= 1 << PG_slab;
1158 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1159 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001160 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001161
1162 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001163
1164 if (unlikely(s->flags & SLAB_POISON))
Christoph Lameter834f3d12008-04-14 19:11:31 +03001165 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001166
1167 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001168 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001169 setup_object(s, page, last);
1170 set_freepointer(s, last, p);
1171 last = p;
1172 }
1173 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001174 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001175
1176 page->freelist = start;
1177 page->inuse = 0;
1178out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001179 return page;
1180}
1181
1182static void __free_slab(struct kmem_cache *s, struct page *page)
1183{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001184 int order = compound_order(page);
1185 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001186
Christoph Lameterc59def92007-05-16 22:10:50 -07001187 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001188 void *p;
1189
1190 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001191 for_each_object(p, s, page_address(page),
1192 page->objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001193 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001194 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001195 }
1196
1197 mod_zone_page_state(page_zone(page),
1198 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1199 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001200 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001201
Christoph Lameter49bd5222008-04-14 18:52:18 +03001202 __ClearPageSlab(page);
1203 reset_page_mapcount(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +03001204 __free_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001205}
1206
1207static void rcu_free_slab(struct rcu_head *h)
1208{
1209 struct page *page;
1210
1211 page = container_of((struct list_head *)h, struct page, lru);
1212 __free_slab(page->slab, page);
1213}
1214
1215static void free_slab(struct kmem_cache *s, struct page *page)
1216{
1217 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1218 /*
1219 * RCU free overloads the RCU head over the LRU
1220 */
1221 struct rcu_head *head = (void *)&page->lru;
1222
1223 call_rcu(head, rcu_free_slab);
1224 } else
1225 __free_slab(s, page);
1226}
1227
1228static void discard_slab(struct kmem_cache *s, struct page *page)
1229{
Christoph Lameter205ab992008-04-14 19:11:40 +03001230 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001231 free_slab(s, page);
1232}
1233
1234/*
1235 * Per slab locking using the pagelock
1236 */
1237static __always_inline void slab_lock(struct page *page)
1238{
1239 bit_spin_lock(PG_locked, &page->flags);
1240}
1241
1242static __always_inline void slab_unlock(struct page *page)
1243{
Nick Piggina76d3542008-01-07 23:20:27 -08001244 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001245}
1246
1247static __always_inline int slab_trylock(struct page *page)
1248{
1249 int rc = 1;
1250
1251 rc = bit_spin_trylock(PG_locked, &page->flags);
1252 return rc;
1253}
1254
1255/*
1256 * Management of partially allocated slabs
1257 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001258static void add_partial(struct kmem_cache_node *n,
1259 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001260{
Christoph Lametere95eed52007-05-06 14:49:44 -07001261 spin_lock(&n->list_lock);
1262 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001263 if (tail)
1264 list_add_tail(&page->lru, &n->partial);
1265 else
1266 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001267 spin_unlock(&n->list_lock);
1268}
1269
Christoph Lameter0121c6192008-04-29 16:11:12 -07001270static void remove_partial(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001271{
1272 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1273
1274 spin_lock(&n->list_lock);
1275 list_del(&page->lru);
1276 n->nr_partial--;
1277 spin_unlock(&n->list_lock);
1278}
1279
1280/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001281 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001282 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001283 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001284 */
Christoph Lameter0121c6192008-04-29 16:11:12 -07001285static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1286 struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001287{
1288 if (slab_trylock(page)) {
1289 list_del(&page->lru);
1290 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001291 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001292 return 1;
1293 }
1294 return 0;
1295}
1296
1297/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001298 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001299 */
1300static struct page *get_partial_node(struct kmem_cache_node *n)
1301{
1302 struct page *page;
1303
1304 /*
1305 * Racy check. If we mistakenly see no partial slabs then we
1306 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001307 * partial slab and there is none available then get_partials()
1308 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001309 */
1310 if (!n || !n->nr_partial)
1311 return NULL;
1312
1313 spin_lock(&n->list_lock);
1314 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001315 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001316 goto out;
1317 page = NULL;
1318out:
1319 spin_unlock(&n->list_lock);
1320 return page;
1321}
1322
1323/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001324 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001325 */
1326static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1327{
1328#ifdef CONFIG_NUMA
1329 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001330 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001331 struct zone *zone;
1332 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001333 struct page *page;
1334
1335 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001336 * The defrag ratio allows a configuration of the tradeoffs between
1337 * inter node defragmentation and node local allocations. A lower
1338 * defrag_ratio increases the tendency to do local allocations
1339 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001340 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001341 * If the defrag_ratio is set to 0 then kmalloc() always
1342 * returns node local objects. If the ratio is higher then kmalloc()
1343 * may return off node objects because partial slabs are obtained
1344 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001345 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001346 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001347 * defrag_ratio = 1000) then every (well almost) allocation will
1348 * first attempt to defrag slab caches on other nodes. This means
1349 * scanning over all nodes to look for partial slabs which may be
1350 * expensive if we do it every time we are trying to find a slab
1351 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001352 */
Christoph Lameter98246012008-01-07 23:20:26 -08001353 if (!s->remote_node_defrag_ratio ||
1354 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001355 return NULL;
1356
Mel Gorman0e884602008-04-28 02:12:14 -07001357 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001358 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001359 struct kmem_cache_node *n;
1360
Mel Gorman54a6eb52008-04-28 02:12:16 -07001361 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001362
Mel Gorman54a6eb52008-04-28 02:12:16 -07001363 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001364 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001365 page = get_partial_node(n);
1366 if (page)
1367 return page;
1368 }
1369 }
1370#endif
1371 return NULL;
1372}
1373
1374/*
1375 * Get a partial page, lock it and return it.
1376 */
1377static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1378{
1379 struct page *page;
1380 int searchnode = (node == -1) ? numa_node_id() : node;
1381
1382 page = get_partial_node(get_node(s, searchnode));
1383 if (page || (flags & __GFP_THISNODE))
1384 return page;
1385
1386 return get_any_partial(s, flags);
1387}
1388
1389/*
1390 * Move a page back to the lists.
1391 *
1392 * Must be called with the slab lock held.
1393 *
1394 * On exit the slab lock will have been dropped.
1395 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001396static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001397{
Christoph Lametere95eed52007-05-06 14:49:44 -07001398 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001399 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001400
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001401 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001402 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001403
Christoph Lametera973e9d2008-03-01 13:40:44 -08001404 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001405 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001406 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1407 } else {
1408 stat(c, DEACTIVATE_FULL);
1409 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1410 add_full(n, page);
1411 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001412 slab_unlock(page);
1413 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001414 stat(c, DEACTIVATE_EMPTY);
Christoph Lametere95eed52007-05-06 14:49:44 -07001415 if (n->nr_partial < MIN_PARTIAL) {
1416 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001417 * Adding an empty slab to the partial slabs in order
1418 * to avoid page allocator overhead. This slab needs
1419 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001420 * so that the others get filled first. That way the
1421 * size of the partial list stays small.
1422 *
Christoph Lameter0121c6192008-04-29 16:11:12 -07001423 * kmem_cache_shrink can reclaim any empty slabs from
1424 * the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001425 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001426 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001427 slab_unlock(page);
1428 } else {
1429 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001430 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001431 discard_slab(s, page);
1432 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001433 }
1434}
1435
1436/*
1437 * Remove the cpu slab
1438 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001439static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001440{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001441 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001442 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001443
Christoph Lameterb773ad72008-03-04 11:10:17 -08001444 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001445 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001446 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001447 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001448 * because both freelists are empty. So this is unlikely
1449 * to occur.
1450 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001451 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001452 void **object;
1453
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001454 tail = 0; /* Hot objects. Put the slab first */
1455
Christoph Lameter894b8782007-05-10 03:15:16 -07001456 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001457 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001458 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001459
1460 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001461 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001462 page->freelist = object;
1463 page->inuse--;
1464 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001465 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001466 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001467}
1468
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001469static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001470{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001471 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001472 slab_lock(c->page);
1473 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001474}
1475
1476/*
1477 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001478 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001479 * Called from IPI handler with interrupts disabled.
1480 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001481static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001482{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001483 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001484
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001485 if (likely(c && c->page))
1486 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001487}
1488
1489static void flush_cpu_slab(void *d)
1490{
1491 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001492
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001493 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001494}
1495
1496static void flush_all(struct kmem_cache *s)
1497{
1498#ifdef CONFIG_SMP
Jens Axboe15c8b6c2008-05-09 09:39:44 +02001499 on_each_cpu(flush_cpu_slab, s, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07001500#else
1501 unsigned long flags;
1502
1503 local_irq_save(flags);
1504 flush_cpu_slab(s);
1505 local_irq_restore(flags);
1506#endif
1507}
1508
1509/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001510 * Check if the objects in a per cpu structure fit numa
1511 * locality expectations.
1512 */
1513static inline int node_match(struct kmem_cache_cpu *c, int node)
1514{
1515#ifdef CONFIG_NUMA
1516 if (node != -1 && c->node != node)
1517 return 0;
1518#endif
1519 return 1;
1520}
1521
1522/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001523 * Slow path. The lockless freelist is empty or we need to perform
1524 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001525 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001526 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001527 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001528 * Processing is still very fast if new objects have been freed to the
1529 * regular freelist. In that case we simply take over the regular freelist
1530 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001531 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001532 * If that is not working then we fall back to the partial lists. We take the
1533 * first element of the freelist as the object to allocate now and move the
1534 * rest of the freelist to the lockless freelist.
1535 *
1536 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001537 * we need to allocate a new slab. This is the slowest path since it involves
1538 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001539 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001540static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001541 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001542{
Christoph Lameter81819f02007-05-06 14:49:36 -07001543 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001544 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001545
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001546 /* We handle __GFP_ZERO in the caller */
1547 gfpflags &= ~__GFP_ZERO;
1548
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001549 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001550 goto new_slab;
1551
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001552 slab_lock(c->page);
1553 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001554 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001555
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001556 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001557
Christoph Lameter894b8782007-05-10 03:15:16 -07001558load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001559 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001560 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001561 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001562 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001563 goto debug;
1564
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001565 c->freelist = object[c->offset];
Christoph Lameter39b26462008-04-14 19:11:30 +03001566 c->page->inuse = c->page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001567 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001568 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001569unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001570 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001571 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001572 return object;
1573
1574another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001575 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001576
1577new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001578 new = get_partial(s, gfpflags, node);
1579 if (new) {
1580 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001581 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001582 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001583 }
1584
Christoph Lameterb811c202007-10-16 23:25:51 -07001585 if (gfpflags & __GFP_WAIT)
1586 local_irq_enable();
1587
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001588 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001589
1590 if (gfpflags & __GFP_WAIT)
1591 local_irq_disable();
1592
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001593 if (new) {
1594 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001595 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001596 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001597 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001598 slab_lock(new);
1599 SetSlabFrozen(new);
1600 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001601 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001602 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001603 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001604debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001605 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001606 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001607
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001608 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001609 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001610 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001611 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001612}
1613
1614/*
1615 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1616 * have the fastpath folded into their functions. So no function call
1617 * overhead for requests that can be satisfied on the fastpath.
1618 *
1619 * The fastpath works by first checking if the lockless freelist can be used.
1620 * If not then __slab_alloc is called for slow processing.
1621 *
1622 * Otherwise we can simply pick the next object from the lockless free list.
1623 */
Pekka Enberg06428782008-01-07 23:20:27 -08001624static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001625 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001626{
Christoph Lameter894b8782007-05-10 03:15:16 -07001627 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001628 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001629 unsigned long flags;
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001630 unsigned int objsize;
Christoph Lameter1f842602008-01-07 23:20:30 -08001631
Christoph Lameter894b8782007-05-10 03:15:16 -07001632 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001633 c = get_cpu_slab(s, smp_processor_id());
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001634 objsize = c->objsize;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001635 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001636
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001637 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001638
1639 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001640 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001641 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001642 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001643 }
1644 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001645
1646 if (unlikely((gfpflags & __GFP_ZERO) && object))
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001647 memset(object, 0, objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001648
Christoph Lameter894b8782007-05-10 03:15:16 -07001649 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001650}
1651
1652void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1653{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001654 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001655}
1656EXPORT_SYMBOL(kmem_cache_alloc);
1657
1658#ifdef CONFIG_NUMA
1659void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1660{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001661 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001662}
1663EXPORT_SYMBOL(kmem_cache_alloc_node);
1664#endif
1665
1666/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001667 * Slow patch handling. This may still be called frequently since objects
1668 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001669 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001670 * So we still attempt to reduce cache line usage. Just take the slab
1671 * lock and free the item. If there is no additional partial page
1672 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001673 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001674static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001675 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001676{
1677 void *prior;
1678 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001679 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001680
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001681 c = get_cpu_slab(s, raw_smp_processor_id());
1682 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001683 slab_lock(page);
1684
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001685 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001686 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001687
Christoph Lameter81819f02007-05-06 14:49:36 -07001688checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001689 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001690 page->freelist = object;
1691 page->inuse--;
1692
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001693 if (unlikely(SlabFrozen(page))) {
1694 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001695 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001696 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001697
1698 if (unlikely(!page->inuse))
1699 goto slab_empty;
1700
1701 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001702 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001703 * then add it.
1704 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001705 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001706 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001707 stat(c, FREE_ADD_PARTIAL);
1708 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001709
1710out_unlock:
1711 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001712 return;
1713
1714slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001715 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001716 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001717 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001718 */
1719 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001720 stat(c, FREE_REMOVE_PARTIAL);
1721 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001722 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001723 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001724 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001725 return;
1726
1727debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001728 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001729 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001730 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001731}
1732
Christoph Lameter894b8782007-05-10 03:15:16 -07001733/*
1734 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1735 * can perform fastpath freeing without additional function calls.
1736 *
1737 * The fastpath is only possible if we are freeing to the current cpu slab
1738 * of this processor. This typically the case if we have just allocated
1739 * the item before.
1740 *
1741 * If fastpath is not possible then fall back to __slab_free where we deal
1742 * with all sorts of special processing.
1743 */
Pekka Enberg06428782008-01-07 23:20:27 -08001744static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001745 struct page *page, void *x, void *addr)
1746{
1747 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001748 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001749 unsigned long flags;
1750
Christoph Lameter894b8782007-05-10 03:15:16 -07001751 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001752 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001753 debug_check_no_locks_freed(object, c->objsize);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001754 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1755 debug_check_no_obj_freed(object, s->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001756 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001757 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001758 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001759 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001760 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001761 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001762
1763 local_irq_restore(flags);
1764}
1765
Christoph Lameter81819f02007-05-06 14:49:36 -07001766void kmem_cache_free(struct kmem_cache *s, void *x)
1767{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001768 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001769
Christoph Lameterb49af682007-05-06 14:49:41 -07001770 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001771
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001772 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001773}
1774EXPORT_SYMBOL(kmem_cache_free);
1775
1776/* Figure out on which slab object the object resides */
1777static struct page *get_object_page(const void *x)
1778{
Christoph Lameterb49af682007-05-06 14:49:41 -07001779 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001780
1781 if (!PageSlab(page))
1782 return NULL;
1783
1784 return page;
1785}
1786
1787/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001788 * Object placement in a slab is made very easy because we always start at
1789 * offset 0. If we tune the size of the object to the alignment then we can
1790 * get the required alignment by putting one properly sized object after
1791 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001792 *
1793 * Notice that the allocation order determines the sizes of the per cpu
1794 * caches. Each processor has always one slab available for allocations.
1795 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001796 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001797 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001798 */
1799
1800/*
1801 * Mininum / Maximum order of slab pages. This influences locking overhead
1802 * and slab fragmentation. A higher order reduces the number of partial slabs
1803 * and increases the number of allocations possible without having to
1804 * take the list_lock.
1805 */
1806static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03001807static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001808static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001809
1810/*
1811 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001812 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001813 */
1814static int slub_nomerge;
1815
1816/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001817 * Calculate the order of allocation given an slab object size.
1818 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001819 * The order of allocation has significant impact on performance and other
1820 * system components. Generally order 0 allocations should be preferred since
1821 * order 0 does not cause fragmentation in the page allocator. Larger objects
1822 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001823 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07001824 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001825 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001826 * In order to reach satisfactory performance we must ensure that a minimum
1827 * number of objects is in one slab. Otherwise we may generate too much
1828 * activity on the partial lists which requires taking the list_lock. This is
1829 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001830 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001831 * slub_max_order specifies the order where we begin to stop considering the
1832 * number of objects in a slab as critical. If we reach slub_max_order then
1833 * we try to keep the page order as low as possible. So we accept more waste
1834 * of space in favor of a small page order.
1835 *
1836 * Higher order allocations also allow the placement of more objects in a
1837 * slab and thereby reduce object handling overhead. If the user has
1838 * requested a higher mininum order then we start with that one instead of
1839 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001840 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001841static inline int slab_order(int size, int min_objects,
1842 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001843{
1844 int order;
1845 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001846 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001847
Christoph Lameter39b26462008-04-14 19:11:30 +03001848 if ((PAGE_SIZE << min_order) / size > 65535)
1849 return get_order(size * 65535) - 1;
1850
Christoph Lameter6300ea72007-07-17 04:03:20 -07001851 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001852 fls(min_objects * size - 1) - PAGE_SHIFT);
1853 order <= max_order; order++) {
1854
Christoph Lameter81819f02007-05-06 14:49:36 -07001855 unsigned long slab_size = PAGE_SIZE << order;
1856
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001857 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001858 continue;
1859
Christoph Lameter81819f02007-05-06 14:49:36 -07001860 rem = slab_size % size;
1861
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001862 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001863 break;
1864
1865 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001866
Christoph Lameter81819f02007-05-06 14:49:36 -07001867 return order;
1868}
1869
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001870static inline int calculate_order(int size)
1871{
1872 int order;
1873 int min_objects;
1874 int fraction;
1875
1876 /*
1877 * Attempt to find best configuration for a slab. This
1878 * works by first attempting to generate a layout with
1879 * the best configuration and backing off gradually.
1880 *
1881 * First we reduce the acceptable waste in a slab. Then
1882 * we reduce the minimum objects required in a slab.
1883 */
1884 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001885 if (!min_objects)
1886 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001887 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001888 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001889 while (fraction >= 4) {
1890 order = slab_order(size, min_objects,
1891 slub_max_order, fraction);
1892 if (order <= slub_max_order)
1893 return order;
1894 fraction /= 2;
1895 }
1896 min_objects /= 2;
1897 }
1898
1899 /*
1900 * We were unable to place multiple objects in a slab. Now
1901 * lets see if we can place a single object there.
1902 */
1903 order = slab_order(size, 1, slub_max_order, 1);
1904 if (order <= slub_max_order)
1905 return order;
1906
1907 /*
1908 * Doh this slab cannot be placed using slub_max_order.
1909 */
1910 order = slab_order(size, 1, MAX_ORDER, 1);
1911 if (order <= MAX_ORDER)
1912 return order;
1913 return -ENOSYS;
1914}
1915
Christoph Lameter81819f02007-05-06 14:49:36 -07001916/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001917 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001918 */
1919static unsigned long calculate_alignment(unsigned long flags,
1920 unsigned long align, unsigned long size)
1921{
1922 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001923 * If the user wants hardware cache aligned objects then follow that
1924 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001925 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001926 * The hardware cache alignment cannot override the specified
1927 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001928 */
Nick Pigginb6210382008-03-05 14:05:56 -08001929 if (flags & SLAB_HWCACHE_ALIGN) {
1930 unsigned long ralign = cache_line_size();
1931 while (size <= ralign / 2)
1932 ralign /= 2;
1933 align = max(align, ralign);
1934 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001935
1936 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001937 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001938
1939 return ALIGN(align, sizeof(void *));
1940}
1941
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001942static void init_kmem_cache_cpu(struct kmem_cache *s,
1943 struct kmem_cache_cpu *c)
1944{
1945 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001946 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001947 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001948 c->offset = s->offset / sizeof(void *);
1949 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03001950#ifdef CONFIG_SLUB_STATS
1951 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1952#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001953}
1954
Christoph Lameter81819f02007-05-06 14:49:36 -07001955static void init_kmem_cache_node(struct kmem_cache_node *n)
1956{
1957 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001958 spin_lock_init(&n->list_lock);
1959 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001960#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001961 atomic_long_set(&n->nr_slabs, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07001962 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001963#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001964}
1965
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001966#ifdef CONFIG_SMP
1967/*
1968 * Per cpu array for per cpu structures.
1969 *
1970 * The per cpu array places all kmem_cache_cpu structures from one processor
1971 * close together meaning that it becomes possible that multiple per cpu
1972 * structures are contained in one cacheline. This may be particularly
1973 * beneficial for the kmalloc caches.
1974 *
1975 * A desktop system typically has around 60-80 slabs. With 100 here we are
1976 * likely able to get per cpu structures for all caches from the array defined
1977 * here. We must be able to cover all kmalloc caches during bootstrap.
1978 *
1979 * If the per cpu array is exhausted then fall back to kmalloc
1980 * of individual cachelines. No sharing is possible then.
1981 */
1982#define NR_KMEM_CACHE_CPU 100
1983
1984static DEFINE_PER_CPU(struct kmem_cache_cpu,
1985 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1986
1987static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1988static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1989
1990static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1991 int cpu, gfp_t flags)
1992{
1993 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1994
1995 if (c)
1996 per_cpu(kmem_cache_cpu_free, cpu) =
1997 (void *)c->freelist;
1998 else {
1999 /* Table overflow: So allocate ourselves */
2000 c = kmalloc_node(
2001 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2002 flags, cpu_to_node(cpu));
2003 if (!c)
2004 return NULL;
2005 }
2006
2007 init_kmem_cache_cpu(s, c);
2008 return c;
2009}
2010
2011static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2012{
2013 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2014 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2015 kfree(c);
2016 return;
2017 }
2018 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2019 per_cpu(kmem_cache_cpu_free, cpu) = c;
2020}
2021
2022static void free_kmem_cache_cpus(struct kmem_cache *s)
2023{
2024 int cpu;
2025
2026 for_each_online_cpu(cpu) {
2027 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2028
2029 if (c) {
2030 s->cpu_slab[cpu] = NULL;
2031 free_kmem_cache_cpu(c, cpu);
2032 }
2033 }
2034}
2035
2036static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2037{
2038 int cpu;
2039
2040 for_each_online_cpu(cpu) {
2041 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2042
2043 if (c)
2044 continue;
2045
2046 c = alloc_kmem_cache_cpu(s, cpu, flags);
2047 if (!c) {
2048 free_kmem_cache_cpus(s);
2049 return 0;
2050 }
2051 s->cpu_slab[cpu] = c;
2052 }
2053 return 1;
2054}
2055
2056/*
2057 * Initialize the per cpu array.
2058 */
2059static void init_alloc_cpu_cpu(int cpu)
2060{
2061 int i;
2062
2063 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2064 return;
2065
2066 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2067 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2068
2069 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2070}
2071
2072static void __init init_alloc_cpu(void)
2073{
2074 int cpu;
2075
2076 for_each_online_cpu(cpu)
2077 init_alloc_cpu_cpu(cpu);
2078 }
2079
2080#else
2081static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2082static inline void init_alloc_cpu(void) {}
2083
2084static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2085{
2086 init_kmem_cache_cpu(s, &s->cpu_slab);
2087 return 1;
2088}
2089#endif
2090
Christoph Lameter81819f02007-05-06 14:49:36 -07002091#ifdef CONFIG_NUMA
2092/*
2093 * No kmalloc_node yet so do it by hand. We know that this is the first
2094 * slab on the node for this slabcache. There are no concurrent accesses
2095 * possible.
2096 *
2097 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002098 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2099 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002100 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002101static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2102 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002103{
2104 struct page *page;
2105 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002106 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002107
2108 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2109
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002110 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002111
2112 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002113 if (page_to_nid(page) != node) {
2114 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2115 "node %d\n", node);
2116 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2117 "in order to be able to continue\n");
2118 }
2119
Christoph Lameter81819f02007-05-06 14:49:36 -07002120 n = page->freelist;
2121 BUG_ON(!n);
2122 page->freelist = get_freepointer(kmalloc_caches, n);
2123 page->inuse++;
2124 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002125#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002126 init_object(kmalloc_caches, n, 1);
2127 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002128#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002129 init_kmem_cache_node(n);
Christoph Lameter205ab992008-04-14 19:11:40 +03002130 inc_slabs_node(kmalloc_caches, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002131
rootba84c732008-01-07 23:20:28 -08002132 /*
2133 * lockdep requires consistent irq usage for each lock
2134 * so even though there cannot be a race this early in
2135 * the boot sequence, we still disable irqs.
2136 */
2137 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002138 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002139 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002140 return n;
2141}
2142
2143static void free_kmem_cache_nodes(struct kmem_cache *s)
2144{
2145 int node;
2146
Christoph Lameterf64dc582007-10-16 01:25:33 -07002147 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002148 struct kmem_cache_node *n = s->node[node];
2149 if (n && n != &s->local_node)
2150 kmem_cache_free(kmalloc_caches, n);
2151 s->node[node] = NULL;
2152 }
2153}
2154
2155static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2156{
2157 int node;
2158 int local_node;
2159
2160 if (slab_state >= UP)
2161 local_node = page_to_nid(virt_to_page(s));
2162 else
2163 local_node = 0;
2164
Christoph Lameterf64dc582007-10-16 01:25:33 -07002165 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002166 struct kmem_cache_node *n;
2167
2168 if (local_node == node)
2169 n = &s->local_node;
2170 else {
2171 if (slab_state == DOWN) {
2172 n = early_kmem_cache_node_alloc(gfpflags,
2173 node);
2174 continue;
2175 }
2176 n = kmem_cache_alloc_node(kmalloc_caches,
2177 gfpflags, node);
2178
2179 if (!n) {
2180 free_kmem_cache_nodes(s);
2181 return 0;
2182 }
2183
2184 }
2185 s->node[node] = n;
2186 init_kmem_cache_node(n);
2187 }
2188 return 1;
2189}
2190#else
2191static void free_kmem_cache_nodes(struct kmem_cache *s)
2192{
2193}
2194
2195static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2196{
2197 init_kmem_cache_node(&s->local_node);
2198 return 1;
2199}
2200#endif
2201
2202/*
2203 * calculate_sizes() determines the order and the distribution of data within
2204 * a slab object.
2205 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002206static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002207{
2208 unsigned long flags = s->flags;
2209 unsigned long size = s->objsize;
2210 unsigned long align = s->align;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002211 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002212
2213 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002214 * Round up object size to the next word boundary. We can only
2215 * place the free pointer at word boundaries and this determines
2216 * the possible location of the free pointer.
2217 */
2218 size = ALIGN(size, sizeof(void *));
2219
2220#ifdef CONFIG_SLUB_DEBUG
2221 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002222 * Determine if we can poison the object itself. If the user of
2223 * the slab may touch the object after free or before allocation
2224 * then we should never poison the object itself.
2225 */
2226 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002227 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002228 s->flags |= __OBJECT_POISON;
2229 else
2230 s->flags &= ~__OBJECT_POISON;
2231
Christoph Lameter81819f02007-05-06 14:49:36 -07002232
2233 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002234 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002235 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002236 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002237 */
2238 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2239 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002240#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002241
2242 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002243 * With that we have determined the number of bytes in actual use
2244 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002245 */
2246 s->inuse = size;
2247
2248 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002249 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002250 /*
2251 * Relocate free pointer after the object if it is not
2252 * permitted to overwrite the first word of the object on
2253 * kmem_cache_free.
2254 *
2255 * This is the case if we do RCU, have a constructor or
2256 * destructor or are poisoning the objects.
2257 */
2258 s->offset = size;
2259 size += sizeof(void *);
2260 }
2261
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002262#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002263 if (flags & SLAB_STORE_USER)
2264 /*
2265 * Need to store information about allocs and frees after
2266 * the object.
2267 */
2268 size += 2 * sizeof(struct track);
2269
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002270 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002271 /*
2272 * Add some empty padding so that we can catch
2273 * overwrites from earlier objects rather than let
2274 * tracking information or the free pointer be
2275 * corrupted if an user writes before the start
2276 * of the object.
2277 */
2278 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002279#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002280
Christoph Lameter81819f02007-05-06 14:49:36 -07002281 /*
2282 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002283 * user specified and the dynamic determination of cache line size
2284 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002285 */
2286 align = calculate_alignment(flags, align, s->objsize);
2287
2288 /*
2289 * SLUB stores one object immediately after another beginning from
2290 * offset 0. In order to align the objects we have to simply size
2291 * each object to conform to the alignment.
2292 */
2293 size = ALIGN(size, align);
2294 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03002295 if (forced_order >= 0)
2296 order = forced_order;
2297 else
2298 order = calculate_order(size);
Christoph Lameter81819f02007-05-06 14:49:36 -07002299
Christoph Lameter834f3d12008-04-14 19:11:31 +03002300 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07002301 return 0;
2302
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002303 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002304 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002305 s->allocflags |= __GFP_COMP;
2306
2307 if (s->flags & SLAB_CACHE_DMA)
2308 s->allocflags |= SLUB_DMA;
2309
2310 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2311 s->allocflags |= __GFP_RECLAIMABLE;
2312
Christoph Lameter81819f02007-05-06 14:49:36 -07002313 /*
2314 * Determine the number of objects per slab
2315 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002316 s->oo = oo_make(order, size);
Christoph Lameter65c33762008-04-14 19:11:40 +03002317 s->min = oo_make(get_order(size), size);
Christoph Lameter205ab992008-04-14 19:11:40 +03002318 if (oo_objects(s->oo) > oo_objects(s->max))
2319 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002320
Christoph Lameter834f3d12008-04-14 19:11:31 +03002321 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07002322
2323}
2324
Christoph Lameter81819f02007-05-06 14:49:36 -07002325static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2326 const char *name, size_t size,
2327 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002328 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002329{
2330 memset(s, 0, kmem_size);
2331 s->name = name;
2332 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002333 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002334 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002335 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002336
Christoph Lameter06b285d2008-04-14 19:11:41 +03002337 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07002338 goto error;
2339
2340 s->refcount = 1;
2341#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002342 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002343#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002344 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2345 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002346
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002347 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002348 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002349 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002350error:
2351 if (flags & SLAB_PANIC)
2352 panic("Cannot create slab %s size=%lu realsize=%u "
2353 "order=%u offset=%u flags=%lx\n",
Christoph Lameter834f3d12008-04-14 19:11:31 +03002354 s->name, (unsigned long)size, s->size, oo_order(s->oo),
Christoph Lameter81819f02007-05-06 14:49:36 -07002355 s->offset, flags);
2356 return 0;
2357}
Christoph Lameter81819f02007-05-06 14:49:36 -07002358
2359/*
2360 * Check if a given pointer is valid
2361 */
2362int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2363{
Pekka Enberg06428782008-01-07 23:20:27 -08002364 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002365
2366 page = get_object_page(object);
2367
2368 if (!page || s != page->slab)
2369 /* No slab or wrong slab */
2370 return 0;
2371
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002372 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002373 return 0;
2374
2375 /*
2376 * We could also check if the object is on the slabs freelist.
2377 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002378 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002379 * to a certain slab.
2380 */
2381 return 1;
2382}
2383EXPORT_SYMBOL(kmem_ptr_validate);
2384
2385/*
2386 * Determine the size of a slab object
2387 */
2388unsigned int kmem_cache_size(struct kmem_cache *s)
2389{
2390 return s->objsize;
2391}
2392EXPORT_SYMBOL(kmem_cache_size);
2393
2394const char *kmem_cache_name(struct kmem_cache *s)
2395{
2396 return s->name;
2397}
2398EXPORT_SYMBOL(kmem_cache_name);
2399
Christoph Lameter33b12c32008-04-25 12:22:43 -07002400static void list_slab_objects(struct kmem_cache *s, struct page *page,
2401 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07002402{
Christoph Lameter33b12c32008-04-25 12:22:43 -07002403#ifdef CONFIG_SLUB_DEBUG
2404 void *addr = page_address(page);
2405 void *p;
2406 DECLARE_BITMAP(map, page->objects);
2407
2408 bitmap_zero(map, page->objects);
2409 slab_err(s, page, "%s", text);
2410 slab_lock(page);
2411 for_each_free_object(p, s, page->freelist)
2412 set_bit(slab_index(p, s, addr), map);
2413
2414 for_each_object(p, s, addr, page->objects) {
2415
2416 if (!test_bit(slab_index(p, s, addr), map)) {
2417 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2418 p, p - addr);
2419 print_tracking(s, p);
2420 }
2421 }
2422 slab_unlock(page);
2423#endif
2424}
2425
Christoph Lameter81819f02007-05-06 14:49:36 -07002426/*
Christoph Lameter599870b2008-04-23 12:36:52 -07002427 * Attempt to free all partial slabs on a node.
Christoph Lameter81819f02007-05-06 14:49:36 -07002428 */
Christoph Lameter599870b2008-04-23 12:36:52 -07002429static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002430{
Christoph Lameter81819f02007-05-06 14:49:36 -07002431 unsigned long flags;
2432 struct page *page, *h;
2433
2434 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter33b12c32008-04-25 12:22:43 -07002435 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002436 if (!page->inuse) {
2437 list_del(&page->lru);
2438 discard_slab(s, page);
Christoph Lameter599870b2008-04-23 12:36:52 -07002439 n->nr_partial--;
Christoph Lameter33b12c32008-04-25 12:22:43 -07002440 } else {
2441 list_slab_objects(s, page,
2442 "Objects remaining on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07002443 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07002444 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002445 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002446}
2447
2448/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002449 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002450 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002451static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002452{
2453 int node;
2454
2455 flush_all(s);
2456
2457 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002458 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002459 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002460 struct kmem_cache_node *n = get_node(s, node);
2461
Christoph Lameter599870b2008-04-23 12:36:52 -07002462 free_partial(s, n);
2463 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002464 return 1;
2465 }
2466 free_kmem_cache_nodes(s);
2467 return 0;
2468}
2469
2470/*
2471 * Close a cache and release the kmem_cache structure
2472 * (must be used for caches created using kmem_cache_create)
2473 */
2474void kmem_cache_destroy(struct kmem_cache *s)
2475{
2476 down_write(&slub_lock);
2477 s->refcount--;
2478 if (!s->refcount) {
2479 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002480 up_write(&slub_lock);
Pekka Enbergd629d812008-04-23 22:31:08 +03002481 if (kmem_cache_close(s)) {
2482 printk(KERN_ERR "SLUB %s: %s called for cache that "
2483 "still has objects.\n", s->name, __func__);
2484 dump_stack();
2485 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002486 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002487 } else
2488 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002489}
2490EXPORT_SYMBOL(kmem_cache_destroy);
2491
2492/********************************************************************
2493 * Kmalloc subsystem
2494 *******************************************************************/
2495
Christoph Lameter331dc552008-02-14 14:28:09 -08002496struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002497EXPORT_SYMBOL(kmalloc_caches);
2498
Christoph Lameter81819f02007-05-06 14:49:36 -07002499static int __init setup_slub_min_order(char *str)
2500{
Pekka Enberg06428782008-01-07 23:20:27 -08002501 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002502
2503 return 1;
2504}
2505
2506__setup("slub_min_order=", setup_slub_min_order);
2507
2508static int __init setup_slub_max_order(char *str)
2509{
Pekka Enberg06428782008-01-07 23:20:27 -08002510 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002511
2512 return 1;
2513}
2514
2515__setup("slub_max_order=", setup_slub_max_order);
2516
2517static int __init setup_slub_min_objects(char *str)
2518{
Pekka Enberg06428782008-01-07 23:20:27 -08002519 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002520
2521 return 1;
2522}
2523
2524__setup("slub_min_objects=", setup_slub_min_objects);
2525
2526static int __init setup_slub_nomerge(char *str)
2527{
2528 slub_nomerge = 1;
2529 return 1;
2530}
2531
2532__setup("slub_nomerge", setup_slub_nomerge);
2533
Christoph Lameter81819f02007-05-06 14:49:36 -07002534static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2535 const char *name, int size, gfp_t gfp_flags)
2536{
2537 unsigned int flags = 0;
2538
2539 if (gfp_flags & SLUB_DMA)
2540 flags = SLAB_CACHE_DMA;
2541
2542 down_write(&slub_lock);
2543 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter319d1e22008-04-14 19:11:41 +03002544 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002545 goto panic;
2546
2547 list_add(&s->list, &slab_caches);
2548 up_write(&slub_lock);
2549 if (sysfs_slab_add(s))
2550 goto panic;
2551 return s;
2552
2553panic:
2554 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2555}
2556
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002557#ifdef CONFIG_ZONE_DMA
Christoph Lameter4097d602008-04-14 18:51:18 +03002558static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002559
2560static void sysfs_add_func(struct work_struct *w)
2561{
2562 struct kmem_cache *s;
2563
2564 down_write(&slub_lock);
2565 list_for_each_entry(s, &slab_caches, list) {
2566 if (s->flags & __SYSFS_ADD_DEFERRED) {
2567 s->flags &= ~__SYSFS_ADD_DEFERRED;
2568 sysfs_slab_add(s);
2569 }
2570 }
2571 up_write(&slub_lock);
2572}
2573
2574static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2575
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002576static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2577{
2578 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002579 char *text;
2580 size_t realsize;
2581
2582 s = kmalloc_caches_dma[index];
2583 if (s)
2584 return s;
2585
2586 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002587 if (flags & __GFP_WAIT)
2588 down_write(&slub_lock);
2589 else {
2590 if (!down_write_trylock(&slub_lock))
2591 goto out;
2592 }
2593
2594 if (kmalloc_caches_dma[index])
2595 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002596
Christoph Lameter7b55f622007-07-17 04:03:27 -07002597 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002598 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2599 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002600 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2601
2602 if (!s || !text || !kmem_cache_open(s, flags, text,
2603 realsize, ARCH_KMALLOC_MINALIGN,
2604 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2605 kfree(s);
2606 kfree(text);
2607 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002608 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002609
2610 list_add(&s->list, &slab_caches);
2611 kmalloc_caches_dma[index] = s;
2612
2613 schedule_work(&sysfs_add_work);
2614
2615unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002616 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002617out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002618 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002619}
2620#endif
2621
Christoph Lameterf1b26332007-07-17 04:03:26 -07002622/*
2623 * Conversion table for small slabs sizes / 8 to the index in the
2624 * kmalloc array. This is necessary for slabs < 192 since we have non power
2625 * of two cache sizes there. The size of larger slabs can be determined using
2626 * fls.
2627 */
2628static s8 size_index[24] = {
2629 3, /* 8 */
2630 4, /* 16 */
2631 5, /* 24 */
2632 5, /* 32 */
2633 6, /* 40 */
2634 6, /* 48 */
2635 6, /* 56 */
2636 6, /* 64 */
2637 1, /* 72 */
2638 1, /* 80 */
2639 1, /* 88 */
2640 1, /* 96 */
2641 7, /* 104 */
2642 7, /* 112 */
2643 7, /* 120 */
2644 7, /* 128 */
2645 2, /* 136 */
2646 2, /* 144 */
2647 2, /* 152 */
2648 2, /* 160 */
2649 2, /* 168 */
2650 2, /* 176 */
2651 2, /* 184 */
2652 2 /* 192 */
2653};
2654
Christoph Lameter81819f02007-05-06 14:49:36 -07002655static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2656{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002657 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002658
Christoph Lameterf1b26332007-07-17 04:03:26 -07002659 if (size <= 192) {
2660 if (!size)
2661 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002662
Christoph Lameterf1b26332007-07-17 04:03:26 -07002663 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002664 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002665 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002666
2667#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002668 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002669 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002670
Christoph Lameter81819f02007-05-06 14:49:36 -07002671#endif
2672 return &kmalloc_caches[index];
2673}
2674
2675void *__kmalloc(size_t size, gfp_t flags)
2676{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002677 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002678
Christoph Lameter331dc552008-02-14 14:28:09 -08002679 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002680 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002681
2682 s = get_slab(size, flags);
2683
2684 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002685 return s;
2686
Christoph Lameterce15fea2007-07-17 04:03:28 -07002687 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002688}
2689EXPORT_SYMBOL(__kmalloc);
2690
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002691static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2692{
2693 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2694 get_order(size));
2695
2696 if (page)
2697 return page_address(page);
2698 else
2699 return NULL;
2700}
2701
Christoph Lameter81819f02007-05-06 14:49:36 -07002702#ifdef CONFIG_NUMA
2703void *__kmalloc_node(size_t size, gfp_t flags, int node)
2704{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002705 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002706
Christoph Lameter331dc552008-02-14 14:28:09 -08002707 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002708 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002709
2710 s = get_slab(size, flags);
2711
2712 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002713 return s;
2714
Christoph Lameterce15fea2007-07-17 04:03:28 -07002715 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002716}
2717EXPORT_SYMBOL(__kmalloc_node);
2718#endif
2719
2720size_t ksize(const void *object)
2721{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002722 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002723 struct kmem_cache *s;
2724
Christoph Lameteref8b4522007-10-16 01:24:46 -07002725 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002726 return 0;
2727
Vegard Nossum294a80a2007-12-04 23:45:30 -08002728 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002729
Pekka Enberg76994412008-05-22 19:22:25 +03002730 if (unlikely(!PageSlab(page))) {
2731 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08002732 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03002733 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002734 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002735
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002736#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002737 /*
2738 * Debugging requires use of the padding between object
2739 * and whatever may come after it.
2740 */
2741 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2742 return s->objsize;
2743
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002744#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002745 /*
2746 * If we have the need to store the freelist pointer
2747 * back there or track user information then we can
2748 * only use the space before that information.
2749 */
2750 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2751 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002752 /*
2753 * Else we can use all the padding etc for the allocation
2754 */
2755 return s->size;
2756}
2757EXPORT_SYMBOL(ksize);
2758
2759void kfree(const void *x)
2760{
Christoph Lameter81819f02007-05-06 14:49:36 -07002761 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002762 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002763
Satyam Sharma2408c552007-10-16 01:24:44 -07002764 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002765 return;
2766
Christoph Lameterb49af682007-05-06 14:49:41 -07002767 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002768 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07002769 BUG_ON(!PageCompound(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002770 put_page(page);
2771 return;
2772 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002773 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002774}
2775EXPORT_SYMBOL(kfree);
2776
Christoph Lameter2086d262007-05-06 14:49:46 -07002777/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002778 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2779 * the remaining slabs by the number of items in use. The slabs with the
2780 * most items in use come first. New allocations will then fill those up
2781 * and thus they can be removed from the partial lists.
2782 *
2783 * The slabs with the least items are placed last. This results in them
2784 * being allocated from last increasing the chance that the last objects
2785 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002786 */
2787int kmem_cache_shrink(struct kmem_cache *s)
2788{
2789 int node;
2790 int i;
2791 struct kmem_cache_node *n;
2792 struct page *page;
2793 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03002794 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07002795 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03002796 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07002797 unsigned long flags;
2798
2799 if (!slabs_by_inuse)
2800 return -ENOMEM;
2801
2802 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002803 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002804 n = get_node(s, node);
2805
2806 if (!n->nr_partial)
2807 continue;
2808
Christoph Lameter834f3d12008-04-14 19:11:31 +03002809 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07002810 INIT_LIST_HEAD(slabs_by_inuse + i);
2811
2812 spin_lock_irqsave(&n->list_lock, flags);
2813
2814 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002815 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002816 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002817 * Note that concurrent frees may occur while we hold the
2818 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002819 */
2820 list_for_each_entry_safe(page, t, &n->partial, lru) {
2821 if (!page->inuse && slab_trylock(page)) {
2822 /*
2823 * Must hold slab lock here because slab_free
2824 * may have freed the last object and be
2825 * waiting to release the slab.
2826 */
2827 list_del(&page->lru);
2828 n->nr_partial--;
2829 slab_unlock(page);
2830 discard_slab(s, page);
2831 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002832 list_move(&page->lru,
2833 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002834 }
2835 }
2836
Christoph Lameter2086d262007-05-06 14:49:46 -07002837 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002838 * Rebuild the partial list with the slabs filled up most
2839 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002840 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002841 for (i = objects - 1; i >= 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07002842 list_splice(slabs_by_inuse + i, n->partial.prev);
2843
Christoph Lameter2086d262007-05-06 14:49:46 -07002844 spin_unlock_irqrestore(&n->list_lock, flags);
2845 }
2846
2847 kfree(slabs_by_inuse);
2848 return 0;
2849}
2850EXPORT_SYMBOL(kmem_cache_shrink);
2851
Yasunori Gotob9049e22007-10-21 16:41:37 -07002852#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2853static int slab_mem_going_offline_callback(void *arg)
2854{
2855 struct kmem_cache *s;
2856
2857 down_read(&slub_lock);
2858 list_for_each_entry(s, &slab_caches, list)
2859 kmem_cache_shrink(s);
2860 up_read(&slub_lock);
2861
2862 return 0;
2863}
2864
2865static void slab_mem_offline_callback(void *arg)
2866{
2867 struct kmem_cache_node *n;
2868 struct kmem_cache *s;
2869 struct memory_notify *marg = arg;
2870 int offline_node;
2871
2872 offline_node = marg->status_change_nid;
2873
2874 /*
2875 * If the node still has available memory. we need kmem_cache_node
2876 * for it yet.
2877 */
2878 if (offline_node < 0)
2879 return;
2880
2881 down_read(&slub_lock);
2882 list_for_each_entry(s, &slab_caches, list) {
2883 n = get_node(s, offline_node);
2884 if (n) {
2885 /*
2886 * if n->nr_slabs > 0, slabs still exist on the node
2887 * that is going down. We were unable to free them,
2888 * and offline_pages() function shoudn't call this
2889 * callback. So, we must fail.
2890 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002891 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002892
2893 s->node[offline_node] = NULL;
2894 kmem_cache_free(kmalloc_caches, n);
2895 }
2896 }
2897 up_read(&slub_lock);
2898}
2899
2900static int slab_mem_going_online_callback(void *arg)
2901{
2902 struct kmem_cache_node *n;
2903 struct kmem_cache *s;
2904 struct memory_notify *marg = arg;
2905 int nid = marg->status_change_nid;
2906 int ret = 0;
2907
2908 /*
2909 * If the node's memory is already available, then kmem_cache_node is
2910 * already created. Nothing to do.
2911 */
2912 if (nid < 0)
2913 return 0;
2914
2915 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07002916 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07002917 * allocate a kmem_cache_node structure in order to bring the node
2918 * online.
2919 */
2920 down_read(&slub_lock);
2921 list_for_each_entry(s, &slab_caches, list) {
2922 /*
2923 * XXX: kmem_cache_alloc_node will fallback to other nodes
2924 * since memory is not yet available from the node that
2925 * is brought up.
2926 */
2927 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2928 if (!n) {
2929 ret = -ENOMEM;
2930 goto out;
2931 }
2932 init_kmem_cache_node(n);
2933 s->node[nid] = n;
2934 }
2935out:
2936 up_read(&slub_lock);
2937 return ret;
2938}
2939
2940static int slab_memory_callback(struct notifier_block *self,
2941 unsigned long action, void *arg)
2942{
2943 int ret = 0;
2944
2945 switch (action) {
2946 case MEM_GOING_ONLINE:
2947 ret = slab_mem_going_online_callback(arg);
2948 break;
2949 case MEM_GOING_OFFLINE:
2950 ret = slab_mem_going_offline_callback(arg);
2951 break;
2952 case MEM_OFFLINE:
2953 case MEM_CANCEL_ONLINE:
2954 slab_mem_offline_callback(arg);
2955 break;
2956 case MEM_ONLINE:
2957 case MEM_CANCEL_OFFLINE:
2958 break;
2959 }
2960
2961 ret = notifier_from_errno(ret);
2962 return ret;
2963}
2964
2965#endif /* CONFIG_MEMORY_HOTPLUG */
2966
Christoph Lameter81819f02007-05-06 14:49:36 -07002967/********************************************************************
2968 * Basic setup of slabs
2969 *******************************************************************/
2970
2971void __init kmem_cache_init(void)
2972{
2973 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002974 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002975
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002976 init_alloc_cpu();
2977
Christoph Lameter81819f02007-05-06 14:49:36 -07002978#ifdef CONFIG_NUMA
2979 /*
2980 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002981 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002982 * kmem_cache_open for slab_state == DOWN.
2983 */
2984 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2985 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002986 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002987 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002988
Nadia Derbey0c40ba42008-04-29 01:00:41 -07002989 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
Christoph Lameter81819f02007-05-06 14:49:36 -07002990#endif
2991
2992 /* Able to allocate the per node structures */
2993 slab_state = PARTIAL;
2994
2995 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002996 if (KMALLOC_MIN_SIZE <= 64) {
2997 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002998 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002999 caches++;
Christoph Lameter4b356be2007-06-16 10:16:13 -07003000 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07003001 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07003002 caches++;
3003 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003004
Christoph Lameter331dc552008-02-14 14:28:09 -08003005 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003006 create_kmalloc_cache(&kmalloc_caches[i],
3007 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07003008 caches++;
3009 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003010
Christoph Lameterf1b26332007-07-17 04:03:26 -07003011
3012 /*
3013 * Patch up the size_index table if we have strange large alignment
3014 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08003015 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07003016 *
3017 * Largest permitted alignment is 256 bytes due to the way we
3018 * handle the index determination for the smaller caches.
3019 *
3020 * Make sure that nothing crazy happens if someone starts tinkering
3021 * around with ARCH_KMALLOC_MINALIGN
3022 */
3023 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3024 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3025
Christoph Lameter12ad6842007-07-17 04:03:28 -07003026 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07003027 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3028
Christoph Lameter41d54d32008-07-03 09:14:26 -05003029 if (KMALLOC_MIN_SIZE == 128) {
3030 /*
3031 * The 192 byte sized cache is not used if the alignment
3032 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3033 * instead.
3034 */
3035 for (i = 128 + 8; i <= 192; i += 8)
3036 size_index[(i - 1) / 8] = 8;
3037 }
3038
Christoph Lameter81819f02007-05-06 14:49:36 -07003039 slab_state = UP;
3040
3041 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08003042 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07003043 kmalloc_caches[i]. name =
3044 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3045
3046#ifdef CONFIG_SMP
3047 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003048 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3049 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3050#else
3051 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003052#endif
3053
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003054 printk(KERN_INFO
3055 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003056 " CPUs=%d, Nodes=%d\n",
3057 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003058 slub_min_order, slub_max_order, slub_min_objects,
3059 nr_cpu_ids, nr_node_ids);
3060}
3061
3062/*
3063 * Find a mergeable slab cache
3064 */
3065static int slab_unmergeable(struct kmem_cache *s)
3066{
3067 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3068 return 1;
3069
Christoph Lameterc59def92007-05-16 22:10:50 -07003070 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003071 return 1;
3072
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003073 /*
3074 * We may have set a slab to be unmergeable during bootstrap.
3075 */
3076 if (s->refcount < 0)
3077 return 1;
3078
Christoph Lameter81819f02007-05-06 14:49:36 -07003079 return 0;
3080}
3081
3082static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003083 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003084 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003085{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003086 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003087
3088 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3089 return NULL;
3090
Christoph Lameterc59def92007-05-16 22:10:50 -07003091 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003092 return NULL;
3093
3094 size = ALIGN(size, sizeof(void *));
3095 align = calculate_alignment(flags, align, size);
3096 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003097 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003098
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003099 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003100 if (slab_unmergeable(s))
3101 continue;
3102
3103 if (size > s->size)
3104 continue;
3105
Christoph Lameterba0268a2007-09-11 15:24:11 -07003106 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003107 continue;
3108 /*
3109 * Check if alignment is compatible.
3110 * Courtesy of Adrian Drzewiecki
3111 */
Pekka Enberg06428782008-01-07 23:20:27 -08003112 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003113 continue;
3114
3115 if (s->size - size >= sizeof(void *))
3116 continue;
3117
3118 return s;
3119 }
3120 return NULL;
3121}
3122
3123struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3124 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003125 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003126{
3127 struct kmem_cache *s;
3128
3129 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003130 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003131 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003132 int cpu;
3133
Christoph Lameter81819f02007-05-06 14:49:36 -07003134 s->refcount++;
3135 /*
3136 * Adjust the object sizes so that we clear
3137 * the complete object on kzalloc.
3138 */
3139 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003140
3141 /*
3142 * And then we need to update the object size in the
3143 * per cpu structures
3144 */
3145 for_each_online_cpu(cpu)
3146 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003147
Christoph Lameter81819f02007-05-06 14:49:36 -07003148 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003149 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003150
Christoph Lameter81819f02007-05-06 14:49:36 -07003151 if (sysfs_slab_alias(s, name))
3152 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003153 return s;
3154 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003155
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003156 s = kmalloc(kmem_size, GFP_KERNEL);
3157 if (s) {
3158 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003159 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003160 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003161 up_write(&slub_lock);
3162 if (sysfs_slab_add(s))
3163 goto err;
3164 return s;
3165 }
3166 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003167 }
3168 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003169
3170err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003171 if (flags & SLAB_PANIC)
3172 panic("Cannot create slabcache %s\n", name);
3173 else
3174 s = NULL;
3175 return s;
3176}
3177EXPORT_SYMBOL(kmem_cache_create);
3178
Christoph Lameter81819f02007-05-06 14:49:36 -07003179#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003180/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003181 * Use the cpu notifier to insure that the cpu slabs are flushed when
3182 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003183 */
3184static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3185 unsigned long action, void *hcpu)
3186{
3187 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003188 struct kmem_cache *s;
3189 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003190
3191 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003192 case CPU_UP_PREPARE:
3193 case CPU_UP_PREPARE_FROZEN:
3194 init_alloc_cpu_cpu(cpu);
3195 down_read(&slub_lock);
3196 list_for_each_entry(s, &slab_caches, list)
3197 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3198 GFP_KERNEL);
3199 up_read(&slub_lock);
3200 break;
3201
Christoph Lameter81819f02007-05-06 14:49:36 -07003202 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003203 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003204 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003205 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003206 down_read(&slub_lock);
3207 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003208 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3209
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003210 local_irq_save(flags);
3211 __flush_cpu_slab(s, cpu);
3212 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003213 free_kmem_cache_cpu(c, cpu);
3214 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003215 }
3216 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003217 break;
3218 default:
3219 break;
3220 }
3221 return NOTIFY_OK;
3222}
3223
Pekka Enberg06428782008-01-07 23:20:27 -08003224static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003225 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003226};
Christoph Lameter81819f02007-05-06 14:49:36 -07003227
3228#endif
3229
Christoph Lameter81819f02007-05-06 14:49:36 -07003230void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3231{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003232 struct kmem_cache *s;
3233
Christoph Lameter331dc552008-02-14 14:28:09 -08003234 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003235 return kmalloc_large(size, gfpflags);
3236
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003237 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003238
Satyam Sharma2408c552007-10-16 01:24:44 -07003239 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003240 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003241
Christoph Lameterce15fea2007-07-17 04:03:28 -07003242 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003243}
3244
3245void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3246 int node, void *caller)
3247{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003248 struct kmem_cache *s;
3249
Christoph Lameter331dc552008-02-14 14:28:09 -08003250 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003251 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003252
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003253 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003254
Satyam Sharma2408c552007-10-16 01:24:44 -07003255 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003256 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003257
Christoph Lameterce15fea2007-07-17 04:03:28 -07003258 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003259}
3260
Christoph Lameterf6acb632008-04-29 16:16:06 -07003261#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03003262static unsigned long count_partial(struct kmem_cache_node *n,
3263 int (*get_count)(struct page *))
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003264{
3265 unsigned long flags;
3266 unsigned long x = 0;
3267 struct page *page;
3268
3269 spin_lock_irqsave(&n->list_lock, flags);
3270 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter205ab992008-04-14 19:11:40 +03003271 x += get_count(page);
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003272 spin_unlock_irqrestore(&n->list_lock, flags);
3273 return x;
3274}
Christoph Lameter205ab992008-04-14 19:11:40 +03003275
3276static int count_inuse(struct page *page)
3277{
3278 return page->inuse;
3279}
3280
3281static int count_total(struct page *page)
3282{
3283 return page->objects;
3284}
3285
3286static int count_free(struct page *page)
3287{
3288 return page->objects - page->inuse;
3289}
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003290
Christoph Lameter434e2452007-07-17 04:03:30 -07003291static int validate_slab(struct kmem_cache *s, struct page *page,
3292 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003293{
3294 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003295 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003296
3297 if (!check_slab(s, page) ||
3298 !on_freelist(s, page, NULL))
3299 return 0;
3300
3301 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003302 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003303
Christoph Lameter7656c722007-05-09 02:32:40 -07003304 for_each_free_object(p, s, page->freelist) {
3305 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003306 if (!check_object(s, page, p, 0))
3307 return 0;
3308 }
3309
Christoph Lameter224a88b2008-04-14 19:11:31 +03003310 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003311 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003312 if (!check_object(s, page, p, 1))
3313 return 0;
3314 return 1;
3315}
3316
Christoph Lameter434e2452007-07-17 04:03:30 -07003317static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3318 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003319{
3320 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003321 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003322 slab_unlock(page);
3323 } else
3324 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3325 s->name, page);
3326
3327 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003328 if (!SlabDebug(page))
3329 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003330 "on slab 0x%p\n", s->name, page);
3331 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003332 if (SlabDebug(page))
3333 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003334 "slab 0x%p\n", s->name, page);
3335 }
3336}
3337
Christoph Lameter434e2452007-07-17 04:03:30 -07003338static int validate_slab_node(struct kmem_cache *s,
3339 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003340{
3341 unsigned long count = 0;
3342 struct page *page;
3343 unsigned long flags;
3344
3345 spin_lock_irqsave(&n->list_lock, flags);
3346
3347 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003348 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003349 count++;
3350 }
3351 if (count != n->nr_partial)
3352 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3353 "counter=%ld\n", s->name, count, n->nr_partial);
3354
3355 if (!(s->flags & SLAB_STORE_USER))
3356 goto out;
3357
3358 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003359 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003360 count++;
3361 }
3362 if (count != atomic_long_read(&n->nr_slabs))
3363 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3364 "counter=%ld\n", s->name, count,
3365 atomic_long_read(&n->nr_slabs));
3366
3367out:
3368 spin_unlock_irqrestore(&n->list_lock, flags);
3369 return count;
3370}
3371
Christoph Lameter434e2452007-07-17 04:03:30 -07003372static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003373{
3374 int node;
3375 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003376 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003377 sizeof(unsigned long), GFP_KERNEL);
3378
3379 if (!map)
3380 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003381
3382 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003383 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003384 struct kmem_cache_node *n = get_node(s, node);
3385
Christoph Lameter434e2452007-07-17 04:03:30 -07003386 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003387 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003388 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003389 return count;
3390}
3391
Christoph Lameterb3459702007-05-09 02:32:41 -07003392#ifdef SLUB_RESILIENCY_TEST
3393static void resiliency_test(void)
3394{
3395 u8 *p;
3396
3397 printk(KERN_ERR "SLUB resiliency testing\n");
3398 printk(KERN_ERR "-----------------------\n");
3399 printk(KERN_ERR "A. Corruption after allocation\n");
3400
3401 p = kzalloc(16, GFP_KERNEL);
3402 p[16] = 0x12;
3403 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3404 " 0x12->0x%p\n\n", p + 16);
3405
3406 validate_slab_cache(kmalloc_caches + 4);
3407
3408 /* Hmmm... The next two are dangerous */
3409 p = kzalloc(32, GFP_KERNEL);
3410 p[32 + sizeof(void *)] = 0x34;
3411 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003412 " 0x34 -> -0x%p\n", p);
3413 printk(KERN_ERR
3414 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003415
3416 validate_slab_cache(kmalloc_caches + 5);
3417 p = kzalloc(64, GFP_KERNEL);
3418 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3419 *p = 0x56;
3420 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3421 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003422 printk(KERN_ERR
3423 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003424 validate_slab_cache(kmalloc_caches + 6);
3425
3426 printk(KERN_ERR "\nB. Corruption after free\n");
3427 p = kzalloc(128, GFP_KERNEL);
3428 kfree(p);
3429 *p = 0x78;
3430 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3431 validate_slab_cache(kmalloc_caches + 7);
3432
3433 p = kzalloc(256, GFP_KERNEL);
3434 kfree(p);
3435 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003436 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3437 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003438 validate_slab_cache(kmalloc_caches + 8);
3439
3440 p = kzalloc(512, GFP_KERNEL);
3441 kfree(p);
3442 p[512] = 0xab;
3443 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3444 validate_slab_cache(kmalloc_caches + 9);
3445}
3446#else
3447static void resiliency_test(void) {};
3448#endif
3449
Christoph Lameter88a420e2007-05-06 14:49:45 -07003450/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003451 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003452 * and freed.
3453 */
3454
3455struct location {
3456 unsigned long count;
3457 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003458 long long sum_time;
3459 long min_time;
3460 long max_time;
3461 long min_pid;
3462 long max_pid;
3463 cpumask_t cpus;
3464 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003465};
3466
3467struct loc_track {
3468 unsigned long max;
3469 unsigned long count;
3470 struct location *loc;
3471};
3472
3473static void free_loc_track(struct loc_track *t)
3474{
3475 if (t->max)
3476 free_pages((unsigned long)t->loc,
3477 get_order(sizeof(struct location) * t->max));
3478}
3479
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003480static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003481{
3482 struct location *l;
3483 int order;
3484
Christoph Lameter88a420e2007-05-06 14:49:45 -07003485 order = get_order(sizeof(struct location) * max);
3486
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003487 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003488 if (!l)
3489 return 0;
3490
3491 if (t->count) {
3492 memcpy(l, t->loc, sizeof(struct location) * t->count);
3493 free_loc_track(t);
3494 }
3495 t->max = max;
3496 t->loc = l;
3497 return 1;
3498}
3499
3500static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003501 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003502{
3503 long start, end, pos;
3504 struct location *l;
3505 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003506 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003507
3508 start = -1;
3509 end = t->count;
3510
3511 for ( ; ; ) {
3512 pos = start + (end - start + 1) / 2;
3513
3514 /*
3515 * There is nothing at "end". If we end up there
3516 * we need to add something to before end.
3517 */
3518 if (pos == end)
3519 break;
3520
3521 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003522 if (track->addr == caddr) {
3523
3524 l = &t->loc[pos];
3525 l->count++;
3526 if (track->when) {
3527 l->sum_time += age;
3528 if (age < l->min_time)
3529 l->min_time = age;
3530 if (age > l->max_time)
3531 l->max_time = age;
3532
3533 if (track->pid < l->min_pid)
3534 l->min_pid = track->pid;
3535 if (track->pid > l->max_pid)
3536 l->max_pid = track->pid;
3537
3538 cpu_set(track->cpu, l->cpus);
3539 }
3540 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003541 return 1;
3542 }
3543
Christoph Lameter45edfa52007-05-09 02:32:45 -07003544 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003545 end = pos;
3546 else
3547 start = pos;
3548 }
3549
3550 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003551 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003552 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003553 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003554 return 0;
3555
3556 l = t->loc + pos;
3557 if (pos < t->count)
3558 memmove(l + 1, l,
3559 (t->count - pos) * sizeof(struct location));
3560 t->count++;
3561 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003562 l->addr = track->addr;
3563 l->sum_time = age;
3564 l->min_time = age;
3565 l->max_time = age;
3566 l->min_pid = track->pid;
3567 l->max_pid = track->pid;
3568 cpus_clear(l->cpus);
3569 cpu_set(track->cpu, l->cpus);
3570 nodes_clear(l->nodes);
3571 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003572 return 1;
3573}
3574
3575static void process_slab(struct loc_track *t, struct kmem_cache *s,
3576 struct page *page, enum track_item alloc)
3577{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003578 void *addr = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +03003579 DECLARE_BITMAP(map, page->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003580 void *p;
3581
Christoph Lameter39b26462008-04-14 19:11:30 +03003582 bitmap_zero(map, page->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003583 for_each_free_object(p, s, page->freelist)
3584 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003585
Christoph Lameter224a88b2008-04-14 19:11:31 +03003586 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003587 if (!test_bit(slab_index(p, s, addr), map))
3588 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003589}
3590
3591static int list_locations(struct kmem_cache *s, char *buf,
3592 enum track_item alloc)
3593{
Harvey Harrisone374d482008-01-31 15:20:50 -08003594 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003595 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003596 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003597 int node;
3598
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003599 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003600 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003601 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003602
3603 /* Push back cpu slabs */
3604 flush_all(s);
3605
Christoph Lameterf64dc582007-10-16 01:25:33 -07003606 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003607 struct kmem_cache_node *n = get_node(s, node);
3608 unsigned long flags;
3609 struct page *page;
3610
Christoph Lameter9e869432007-08-22 14:01:56 -07003611 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003612 continue;
3613
3614 spin_lock_irqsave(&n->list_lock, flags);
3615 list_for_each_entry(page, &n->partial, lru)
3616 process_slab(&t, s, page, alloc);
3617 list_for_each_entry(page, &n->full, lru)
3618 process_slab(&t, s, page, alloc);
3619 spin_unlock_irqrestore(&n->list_lock, flags);
3620 }
3621
3622 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003623 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003624
Harvey Harrisone374d482008-01-31 15:20:50 -08003625 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003626 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003627 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003628
3629 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003630 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003631 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003632 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003633
3634 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08003635 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07003636 l->min_time,
3637 (long)div_u64(l->sum_time, l->count),
3638 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003639 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003640 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003641 l->min_time);
3642
3643 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003644 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003645 l->min_pid, l->max_pid);
3646 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003647 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003648 l->min_pid);
3649
Christoph Lameter84966342007-06-23 17:16:32 -07003650 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003651 len < PAGE_SIZE - 60) {
3652 len += sprintf(buf + len, " cpus=");
3653 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003654 l->cpus);
3655 }
3656
Christoph Lameter84966342007-06-23 17:16:32 -07003657 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003658 len < PAGE_SIZE - 60) {
3659 len += sprintf(buf + len, " nodes=");
3660 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003661 l->nodes);
3662 }
3663
Harvey Harrisone374d482008-01-31 15:20:50 -08003664 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003665 }
3666
3667 free_loc_track(&t);
3668 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003669 len += sprintf(buf, "No data\n");
3670 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003671}
3672
Christoph Lameter81819f02007-05-06 14:49:36 -07003673enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03003674 SL_ALL, /* All slabs */
3675 SL_PARTIAL, /* Only partially allocated slabs */
3676 SL_CPU, /* Only slabs used for cpu caches */
3677 SL_OBJECTS, /* Determine allocated objects not slabs */
3678 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07003679};
3680
Christoph Lameter205ab992008-04-14 19:11:40 +03003681#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003682#define SO_PARTIAL (1 << SL_PARTIAL)
3683#define SO_CPU (1 << SL_CPU)
3684#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03003685#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003686
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003687static ssize_t show_slab_objects(struct kmem_cache *s,
3688 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003689{
3690 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003691 int node;
3692 int x;
3693 unsigned long *nodes;
3694 unsigned long *per_cpu;
3695
3696 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003697 if (!nodes)
3698 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003699 per_cpu = nodes + nr_node_ids;
3700
Christoph Lameter205ab992008-04-14 19:11:40 +03003701 if (flags & SO_CPU) {
3702 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07003703
Christoph Lameter205ab992008-04-14 19:11:40 +03003704 for_each_possible_cpu(cpu) {
3705 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003706
Christoph Lameter205ab992008-04-14 19:11:40 +03003707 if (!c || c->node < 0)
3708 continue;
3709
3710 if (c->page) {
3711 if (flags & SO_TOTAL)
3712 x = c->page->objects;
3713 else if (flags & SO_OBJECTS)
3714 x = c->page->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07003715 else
3716 x = 1;
Christoph Lameter205ab992008-04-14 19:11:40 +03003717
Christoph Lameter81819f02007-05-06 14:49:36 -07003718 total += x;
Christoph Lameter205ab992008-04-14 19:11:40 +03003719 nodes[c->node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003720 }
Christoph Lameter205ab992008-04-14 19:11:40 +03003721 per_cpu[c->node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003722 }
3723 }
3724
Christoph Lameter205ab992008-04-14 19:11:40 +03003725 if (flags & SO_ALL) {
3726 for_each_node_state(node, N_NORMAL_MEMORY) {
3727 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003728
Christoph Lameter205ab992008-04-14 19:11:40 +03003729 if (flags & SO_TOTAL)
3730 x = atomic_long_read(&n->total_objects);
3731 else if (flags & SO_OBJECTS)
3732 x = atomic_long_read(&n->total_objects) -
3733 count_partial(n, count_free);
3734
3735 else
3736 x = atomic_long_read(&n->nr_slabs);
3737 total += x;
3738 nodes[node] += x;
3739 }
3740
3741 } else if (flags & SO_PARTIAL) {
3742 for_each_node_state(node, N_NORMAL_MEMORY) {
3743 struct kmem_cache_node *n = get_node(s, node);
3744
3745 if (flags & SO_TOTAL)
3746 x = count_partial(n, count_total);
3747 else if (flags & SO_OBJECTS)
3748 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07003749 else
3750 x = n->nr_partial;
3751 total += x;
3752 nodes[node] += x;
3753 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003754 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003755 x = sprintf(buf, "%lu", total);
3756#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003757 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003758 if (nodes[node])
3759 x += sprintf(buf + x, " N%d=%lu",
3760 node, nodes[node]);
3761#endif
3762 kfree(nodes);
3763 return x + sprintf(buf + x, "\n");
3764}
3765
3766static int any_slab_objects(struct kmem_cache *s)
3767{
3768 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003769
3770 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003771 struct kmem_cache_node *n = get_node(s, node);
3772
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003773 if (!n)
3774 continue;
3775
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07003776 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07003777 return 1;
3778 }
3779 return 0;
3780}
3781
3782#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3783#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3784
3785struct slab_attribute {
3786 struct attribute attr;
3787 ssize_t (*show)(struct kmem_cache *s, char *buf);
3788 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3789};
3790
3791#define SLAB_ATTR_RO(_name) \
3792 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3793
3794#define SLAB_ATTR(_name) \
3795 static struct slab_attribute _name##_attr = \
3796 __ATTR(_name, 0644, _name##_show, _name##_store)
3797
Christoph Lameter81819f02007-05-06 14:49:36 -07003798static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3799{
3800 return sprintf(buf, "%d\n", s->size);
3801}
3802SLAB_ATTR_RO(slab_size);
3803
3804static ssize_t align_show(struct kmem_cache *s, char *buf)
3805{
3806 return sprintf(buf, "%d\n", s->align);
3807}
3808SLAB_ATTR_RO(align);
3809
3810static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3811{
3812 return sprintf(buf, "%d\n", s->objsize);
3813}
3814SLAB_ATTR_RO(object_size);
3815
3816static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3817{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003818 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003819}
3820SLAB_ATTR_RO(objs_per_slab);
3821
Christoph Lameter06b285d2008-04-14 19:11:41 +03003822static ssize_t order_store(struct kmem_cache *s,
3823 const char *buf, size_t length)
3824{
Christoph Lameter0121c6192008-04-29 16:11:12 -07003825 unsigned long order;
3826 int err;
3827
3828 err = strict_strtoul(buf, 10, &order);
3829 if (err)
3830 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003831
3832 if (order > slub_max_order || order < slub_min_order)
3833 return -EINVAL;
3834
3835 calculate_sizes(s, order);
3836 return length;
3837}
3838
Christoph Lameter81819f02007-05-06 14:49:36 -07003839static ssize_t order_show(struct kmem_cache *s, char *buf)
3840{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003841 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003842}
Christoph Lameter06b285d2008-04-14 19:11:41 +03003843SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003844
3845static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3846{
3847 if (s->ctor) {
3848 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3849
3850 return n + sprintf(buf + n, "\n");
3851 }
3852 return 0;
3853}
3854SLAB_ATTR_RO(ctor);
3855
Christoph Lameter81819f02007-05-06 14:49:36 -07003856static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3857{
3858 return sprintf(buf, "%d\n", s->refcount - 1);
3859}
3860SLAB_ATTR_RO(aliases);
3861
3862static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3863{
Christoph Lameter205ab992008-04-14 19:11:40 +03003864 return show_slab_objects(s, buf, SO_ALL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003865}
3866SLAB_ATTR_RO(slabs);
3867
3868static ssize_t partial_show(struct kmem_cache *s, char *buf)
3869{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003870 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003871}
3872SLAB_ATTR_RO(partial);
3873
3874static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3875{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003876 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003877}
3878SLAB_ATTR_RO(cpu_slabs);
3879
3880static ssize_t objects_show(struct kmem_cache *s, char *buf)
3881{
Christoph Lameter205ab992008-04-14 19:11:40 +03003882 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003883}
3884SLAB_ATTR_RO(objects);
3885
Christoph Lameter205ab992008-04-14 19:11:40 +03003886static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3887{
3888 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3889}
3890SLAB_ATTR_RO(objects_partial);
3891
3892static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3893{
3894 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3895}
3896SLAB_ATTR_RO(total_objects);
3897
Christoph Lameter81819f02007-05-06 14:49:36 -07003898static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3899{
3900 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3901}
3902
3903static ssize_t sanity_checks_store(struct kmem_cache *s,
3904 const char *buf, size_t length)
3905{
3906 s->flags &= ~SLAB_DEBUG_FREE;
3907 if (buf[0] == '1')
3908 s->flags |= SLAB_DEBUG_FREE;
3909 return length;
3910}
3911SLAB_ATTR(sanity_checks);
3912
3913static ssize_t trace_show(struct kmem_cache *s, char *buf)
3914{
3915 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3916}
3917
3918static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3919 size_t length)
3920{
3921 s->flags &= ~SLAB_TRACE;
3922 if (buf[0] == '1')
3923 s->flags |= SLAB_TRACE;
3924 return length;
3925}
3926SLAB_ATTR(trace);
3927
3928static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3929{
3930 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3931}
3932
3933static ssize_t reclaim_account_store(struct kmem_cache *s,
3934 const char *buf, size_t length)
3935{
3936 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3937 if (buf[0] == '1')
3938 s->flags |= SLAB_RECLAIM_ACCOUNT;
3939 return length;
3940}
3941SLAB_ATTR(reclaim_account);
3942
3943static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3944{
Christoph Lameter5af60832007-05-06 14:49:56 -07003945 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003946}
3947SLAB_ATTR_RO(hwcache_align);
3948
3949#ifdef CONFIG_ZONE_DMA
3950static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3951{
3952 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3953}
3954SLAB_ATTR_RO(cache_dma);
3955#endif
3956
3957static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3958{
3959 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3960}
3961SLAB_ATTR_RO(destroy_by_rcu);
3962
3963static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3964{
3965 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3966}
3967
3968static ssize_t red_zone_store(struct kmem_cache *s,
3969 const char *buf, size_t length)
3970{
3971 if (any_slab_objects(s))
3972 return -EBUSY;
3973
3974 s->flags &= ~SLAB_RED_ZONE;
3975 if (buf[0] == '1')
3976 s->flags |= SLAB_RED_ZONE;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003977 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003978 return length;
3979}
3980SLAB_ATTR(red_zone);
3981
3982static ssize_t poison_show(struct kmem_cache *s, char *buf)
3983{
3984 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3985}
3986
3987static ssize_t poison_store(struct kmem_cache *s,
3988 const char *buf, size_t length)
3989{
3990 if (any_slab_objects(s))
3991 return -EBUSY;
3992
3993 s->flags &= ~SLAB_POISON;
3994 if (buf[0] == '1')
3995 s->flags |= SLAB_POISON;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003996 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003997 return length;
3998}
3999SLAB_ATTR(poison);
4000
4001static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4002{
4003 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4004}
4005
4006static ssize_t store_user_store(struct kmem_cache *s,
4007 const char *buf, size_t length)
4008{
4009 if (any_slab_objects(s))
4010 return -EBUSY;
4011
4012 s->flags &= ~SLAB_STORE_USER;
4013 if (buf[0] == '1')
4014 s->flags |= SLAB_STORE_USER;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004015 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004016 return length;
4017}
4018SLAB_ATTR(store_user);
4019
Christoph Lameter53e15af2007-05-06 14:49:43 -07004020static ssize_t validate_show(struct kmem_cache *s, char *buf)
4021{
4022 return 0;
4023}
4024
4025static ssize_t validate_store(struct kmem_cache *s,
4026 const char *buf, size_t length)
4027{
Christoph Lameter434e2452007-07-17 04:03:30 -07004028 int ret = -EINVAL;
4029
4030 if (buf[0] == '1') {
4031 ret = validate_slab_cache(s);
4032 if (ret >= 0)
4033 ret = length;
4034 }
4035 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004036}
4037SLAB_ATTR(validate);
4038
Christoph Lameter2086d262007-05-06 14:49:46 -07004039static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4040{
4041 return 0;
4042}
4043
4044static ssize_t shrink_store(struct kmem_cache *s,
4045 const char *buf, size_t length)
4046{
4047 if (buf[0] == '1') {
4048 int rc = kmem_cache_shrink(s);
4049
4050 if (rc)
4051 return rc;
4052 } else
4053 return -EINVAL;
4054 return length;
4055}
4056SLAB_ATTR(shrink);
4057
Christoph Lameter88a420e2007-05-06 14:49:45 -07004058static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4059{
4060 if (!(s->flags & SLAB_STORE_USER))
4061 return -ENOSYS;
4062 return list_locations(s, buf, TRACK_ALLOC);
4063}
4064SLAB_ATTR_RO(alloc_calls);
4065
4066static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4067{
4068 if (!(s->flags & SLAB_STORE_USER))
4069 return -ENOSYS;
4070 return list_locations(s, buf, TRACK_FREE);
4071}
4072SLAB_ATTR_RO(free_calls);
4073
Christoph Lameter81819f02007-05-06 14:49:36 -07004074#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004075static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004076{
Christoph Lameter98246012008-01-07 23:20:26 -08004077 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004078}
4079
Christoph Lameter98246012008-01-07 23:20:26 -08004080static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004081 const char *buf, size_t length)
4082{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004083 unsigned long ratio;
4084 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004085
Christoph Lameter0121c6192008-04-29 16:11:12 -07004086 err = strict_strtoul(buf, 10, &ratio);
4087 if (err)
4088 return err;
4089
4090 if (ratio < 100)
4091 s->remote_node_defrag_ratio = ratio * 10;
4092
Christoph Lameter81819f02007-05-06 14:49:36 -07004093 return length;
4094}
Christoph Lameter98246012008-01-07 23:20:26 -08004095SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004096#endif
4097
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004098#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004099static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4100{
4101 unsigned long sum = 0;
4102 int cpu;
4103 int len;
4104 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4105
4106 if (!data)
4107 return -ENOMEM;
4108
4109 for_each_online_cpu(cpu) {
4110 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4111
4112 data[cpu] = x;
4113 sum += x;
4114 }
4115
4116 len = sprintf(buf, "%lu", sum);
4117
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004118#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004119 for_each_online_cpu(cpu) {
4120 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004121 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004122 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004123#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004124 kfree(data);
4125 return len + sprintf(buf + len, "\n");
4126}
4127
4128#define STAT_ATTR(si, text) \
4129static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4130{ \
4131 return show_stat(s, buf, si); \
4132} \
4133SLAB_ATTR_RO(text); \
4134
4135STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4136STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4137STAT_ATTR(FREE_FASTPATH, free_fastpath);
4138STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4139STAT_ATTR(FREE_FROZEN, free_frozen);
4140STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4141STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4142STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4143STAT_ATTR(ALLOC_SLAB, alloc_slab);
4144STAT_ATTR(ALLOC_REFILL, alloc_refill);
4145STAT_ATTR(FREE_SLAB, free_slab);
4146STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4147STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4148STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4149STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4150STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4151STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter65c33762008-04-14 19:11:40 +03004152STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004153#endif
4154
Pekka Enberg06428782008-01-07 23:20:27 -08004155static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004156 &slab_size_attr.attr,
4157 &object_size_attr.attr,
4158 &objs_per_slab_attr.attr,
4159 &order_attr.attr,
4160 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004161 &objects_partial_attr.attr,
4162 &total_objects_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004163 &slabs_attr.attr,
4164 &partial_attr.attr,
4165 &cpu_slabs_attr.attr,
4166 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004167 &aliases_attr.attr,
4168 &align_attr.attr,
4169 &sanity_checks_attr.attr,
4170 &trace_attr.attr,
4171 &hwcache_align_attr.attr,
4172 &reclaim_account_attr.attr,
4173 &destroy_by_rcu_attr.attr,
4174 &red_zone_attr.attr,
4175 &poison_attr.attr,
4176 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004177 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004178 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004179 &alloc_calls_attr.attr,
4180 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004181#ifdef CONFIG_ZONE_DMA
4182 &cache_dma_attr.attr,
4183#endif
4184#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004185 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004186#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004187#ifdef CONFIG_SLUB_STATS
4188 &alloc_fastpath_attr.attr,
4189 &alloc_slowpath_attr.attr,
4190 &free_fastpath_attr.attr,
4191 &free_slowpath_attr.attr,
4192 &free_frozen_attr.attr,
4193 &free_add_partial_attr.attr,
4194 &free_remove_partial_attr.attr,
4195 &alloc_from_partial_attr.attr,
4196 &alloc_slab_attr.attr,
4197 &alloc_refill_attr.attr,
4198 &free_slab_attr.attr,
4199 &cpuslab_flush_attr.attr,
4200 &deactivate_full_attr.attr,
4201 &deactivate_empty_attr.attr,
4202 &deactivate_to_head_attr.attr,
4203 &deactivate_to_tail_attr.attr,
4204 &deactivate_remote_frees_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004205 &order_fallback_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004206#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004207 NULL
4208};
4209
4210static struct attribute_group slab_attr_group = {
4211 .attrs = slab_attrs,
4212};
4213
4214static ssize_t slab_attr_show(struct kobject *kobj,
4215 struct attribute *attr,
4216 char *buf)
4217{
4218 struct slab_attribute *attribute;
4219 struct kmem_cache *s;
4220 int err;
4221
4222 attribute = to_slab_attr(attr);
4223 s = to_slab(kobj);
4224
4225 if (!attribute->show)
4226 return -EIO;
4227
4228 err = attribute->show(s, buf);
4229
4230 return err;
4231}
4232
4233static ssize_t slab_attr_store(struct kobject *kobj,
4234 struct attribute *attr,
4235 const char *buf, size_t len)
4236{
4237 struct slab_attribute *attribute;
4238 struct kmem_cache *s;
4239 int err;
4240
4241 attribute = to_slab_attr(attr);
4242 s = to_slab(kobj);
4243
4244 if (!attribute->store)
4245 return -EIO;
4246
4247 err = attribute->store(s, buf, len);
4248
4249 return err;
4250}
4251
Christoph Lameter151c6022008-01-07 22:29:05 -08004252static void kmem_cache_release(struct kobject *kobj)
4253{
4254 struct kmem_cache *s = to_slab(kobj);
4255
4256 kfree(s);
4257}
4258
Christoph Lameter81819f02007-05-06 14:49:36 -07004259static struct sysfs_ops slab_sysfs_ops = {
4260 .show = slab_attr_show,
4261 .store = slab_attr_store,
4262};
4263
4264static struct kobj_type slab_ktype = {
4265 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004266 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004267};
4268
4269static int uevent_filter(struct kset *kset, struct kobject *kobj)
4270{
4271 struct kobj_type *ktype = get_ktype(kobj);
4272
4273 if (ktype == &slab_ktype)
4274 return 1;
4275 return 0;
4276}
4277
4278static struct kset_uevent_ops slab_uevent_ops = {
4279 .filter = uevent_filter,
4280};
4281
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004282static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004283
4284#define ID_STR_LENGTH 64
4285
4286/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004287 *
4288 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004289 */
4290static char *create_unique_id(struct kmem_cache *s)
4291{
4292 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4293 char *p = name;
4294
4295 BUG_ON(!name);
4296
4297 *p++ = ':';
4298 /*
4299 * First flags affecting slabcache operations. We will only
4300 * get here for aliasable slabs so we do not need to support
4301 * too many flags. The flags here must cover all flags that
4302 * are matched during merging to guarantee that the id is
4303 * unique.
4304 */
4305 if (s->flags & SLAB_CACHE_DMA)
4306 *p++ = 'd';
4307 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4308 *p++ = 'a';
4309 if (s->flags & SLAB_DEBUG_FREE)
4310 *p++ = 'F';
4311 if (p != name + 1)
4312 *p++ = '-';
4313 p += sprintf(p, "%07d", s->size);
4314 BUG_ON(p > name + ID_STR_LENGTH - 1);
4315 return name;
4316}
4317
4318static int sysfs_slab_add(struct kmem_cache *s)
4319{
4320 int err;
4321 const char *name;
4322 int unmergeable;
4323
4324 if (slab_state < SYSFS)
4325 /* Defer until later */
4326 return 0;
4327
4328 unmergeable = slab_unmergeable(s);
4329 if (unmergeable) {
4330 /*
4331 * Slabcache can never be merged so we can use the name proper.
4332 * This is typically the case for debug situations. In that
4333 * case we can catch duplicate names easily.
4334 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004335 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004336 name = s->name;
4337 } else {
4338 /*
4339 * Create a unique name for the slab as a target
4340 * for the symlinks.
4341 */
4342 name = create_unique_id(s);
4343 }
4344
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004345 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004346 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4347 if (err) {
4348 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004349 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004350 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004351
4352 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4353 if (err)
4354 return err;
4355 kobject_uevent(&s->kobj, KOBJ_ADD);
4356 if (!unmergeable) {
4357 /* Setup first alias */
4358 sysfs_slab_alias(s, s->name);
4359 kfree(name);
4360 }
4361 return 0;
4362}
4363
4364static void sysfs_slab_remove(struct kmem_cache *s)
4365{
4366 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4367 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004368 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004369}
4370
4371/*
4372 * Need to buffer aliases during bootup until sysfs becomes
4373 * available lest we loose that information.
4374 */
4375struct saved_alias {
4376 struct kmem_cache *s;
4377 const char *name;
4378 struct saved_alias *next;
4379};
4380
Adrian Bunk5af328a2007-07-17 04:03:27 -07004381static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004382
4383static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4384{
4385 struct saved_alias *al;
4386
4387 if (slab_state == SYSFS) {
4388 /*
4389 * If we have a leftover link then remove it.
4390 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004391 sysfs_remove_link(&slab_kset->kobj, name);
4392 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004393 }
4394
4395 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4396 if (!al)
4397 return -ENOMEM;
4398
4399 al->s = s;
4400 al->name = name;
4401 al->next = alias_list;
4402 alias_list = al;
4403 return 0;
4404}
4405
4406static int __init slab_sysfs_init(void)
4407{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004408 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004409 int err;
4410
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004411 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004412 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004413 printk(KERN_ERR "Cannot register slab subsystem.\n");
4414 return -ENOSYS;
4415 }
4416
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004417 slab_state = SYSFS;
4418
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004419 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004420 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004421 if (err)
4422 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4423 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004424 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004425
4426 while (alias_list) {
4427 struct saved_alias *al = alias_list;
4428
4429 alias_list = alias_list->next;
4430 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004431 if (err)
4432 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4433 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004434 kfree(al);
4435 }
4436
4437 resiliency_test();
4438 return 0;
4439}
4440
4441__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004442#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004443
4444/*
4445 * The /proc/slabinfo ABI
4446 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004447#ifdef CONFIG_SLABINFO
4448
Christoph Lameter0121c6192008-04-29 16:11:12 -07004449ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4450 size_t count, loff_t *ppos)
Linus Torvalds158a9622008-01-02 13:04:48 -08004451{
4452 return -EINVAL;
4453}
4454
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004455
4456static void print_slabinfo_header(struct seq_file *m)
4457{
4458 seq_puts(m, "slabinfo - version: 2.1\n");
4459 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4460 "<objperslab> <pagesperslab>");
4461 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4462 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4463 seq_putc(m, '\n');
4464}
4465
4466static void *s_start(struct seq_file *m, loff_t *pos)
4467{
4468 loff_t n = *pos;
4469
4470 down_read(&slub_lock);
4471 if (!n)
4472 print_slabinfo_header(m);
4473
4474 return seq_list_start(&slab_caches, *pos);
4475}
4476
4477static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4478{
4479 return seq_list_next(p, &slab_caches, pos);
4480}
4481
4482static void s_stop(struct seq_file *m, void *p)
4483{
4484 up_read(&slub_lock);
4485}
4486
4487static int s_show(struct seq_file *m, void *p)
4488{
4489 unsigned long nr_partials = 0;
4490 unsigned long nr_slabs = 0;
4491 unsigned long nr_inuse = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004492 unsigned long nr_objs = 0;
4493 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004494 struct kmem_cache *s;
4495 int node;
4496
4497 s = list_entry(p, struct kmem_cache, list);
4498
4499 for_each_online_node(node) {
4500 struct kmem_cache_node *n = get_node(s, node);
4501
4502 if (!n)
4503 continue;
4504
4505 nr_partials += n->nr_partial;
4506 nr_slabs += atomic_long_read(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03004507 nr_objs += atomic_long_read(&n->total_objects);
4508 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004509 }
4510
Christoph Lameter205ab992008-04-14 19:11:40 +03004511 nr_inuse = nr_objs - nr_free;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004512
4513 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
Christoph Lameter834f3d12008-04-14 19:11:31 +03004514 nr_objs, s->size, oo_objects(s->oo),
4515 (1 << oo_order(s->oo)));
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004516 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4517 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4518 0UL);
4519 seq_putc(m, '\n');
4520 return 0;
4521}
4522
4523const struct seq_operations slabinfo_op = {
4524 .start = s_start,
4525 .next = s_next,
4526 .stop = s_stop,
4527 .show = s_show,
4528};
4529
Linus Torvalds158a9622008-01-02 13:04:48 -08004530#endif /* CONFIG_SLABINFO */