blob: a96e11c77fd9595deec57b445888fde3b9bf4671 [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 *
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
9 */
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>
22#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070023#include <linux/memory.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024
25/*
26 * Lock order:
27 * 1. slab_lock(page)
28 * 2. slab->list_lock
29 *
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
36 *
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
42 *
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
47 * the list lock.
48 *
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
61 *
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
66 *
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
69 *
Christoph Lameter672bba32007-05-09 02:32:39 -070070 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070072 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070073 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070075 *
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
79 *
80 * Overloading of page flags that are otherwise used for LRU management.
81 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070082 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
90 *
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -070094 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -070095 * free objects in addition to the regular freelist
96 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070097 *
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700100 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700101 */
102
Christoph Lameter5577bd82007-05-16 22:10:56 -0700103#define FROZEN (1 << PG_active)
104
105#ifdef CONFIG_SLUB_DEBUG
106#define SLABDEBUG (1 << PG_error)
107#else
108#define SLABDEBUG 0
109#endif
110
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700111static inline int SlabFrozen(struct page *page)
112{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700113 return page->flags & FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700114}
115
116static inline void SetSlabFrozen(struct page *page)
117{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700118 page->flags |= FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700119}
120
121static inline void ClearSlabFrozen(struct page *page)
122{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700123 page->flags &= ~FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700124}
125
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700126static inline int SlabDebug(struct page *page)
127{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700128 return page->flags & SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700129}
130
131static inline void SetSlabDebug(struct page *page)
132{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700133 page->flags |= SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700134}
135
136static inline void ClearSlabDebug(struct page *page)
137{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700138 page->flags &= ~SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700139}
140
Christoph Lameter81819f02007-05-06 14:49:36 -0700141/*
142 * Issues still to be resolved:
143 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
145 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700146 * - Variable sizing of the per node arrays
147 */
148
149/* Enable to test recovery from slab corruption on boot */
150#undef SLUB_RESILIENCY_TEST
151
152#if PAGE_SHIFT <= 12
153
154/*
155 * Small page size. Make sure that we do not fragment memory
156 */
157#define DEFAULT_MAX_ORDER 1
158#define DEFAULT_MIN_OBJECTS 4
159
160#else
161
162/*
163 * Large page machines are customarily able to handle larger
164 * page orders.
165 */
166#define DEFAULT_MAX_ORDER 2
167#define DEFAULT_MIN_OBJECTS 8
168
169#endif
170
171/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
174 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800175#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700176
Christoph Lameter2086d262007-05-06 14:49:46 -0700177/*
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
181 */
182#define MAX_PARTIAL 10
183
Christoph Lameter81819f02007-05-06 14:49:36 -0700184#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700186
Christoph Lameter81819f02007-05-06 14:49:36 -0700187/*
188 * Set of flags that will prevent slab merging
189 */
190#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
192
193#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
194 SLAB_CACHE_DMA)
195
196#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700197#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700198#endif
199
200#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700201#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700202#endif
203
204/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700205#define __OBJECT_POISON 0x80000000 /* Poison object */
206#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter71c7a062008-02-14 14:28:01 -0800207#define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */
208#define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */
Christoph Lameter81819f02007-05-06 14:49:36 -0700209
Christoph Lameter65c02d42007-05-09 02:32:35 -0700210/* Not all arches define cache_line_size */
211#ifndef cache_line_size
212#define cache_line_size() L1_CACHE_BYTES
213#endif
214
Christoph Lameter81819f02007-05-06 14:49:36 -0700215static int kmem_size = sizeof(struct kmem_cache);
216
217#ifdef CONFIG_SMP
218static struct notifier_block slab_notifier;
219#endif
220
221static enum {
222 DOWN, /* No slab functionality available */
223 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700224 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700225 SYSFS /* Sysfs up */
226} slab_state = DOWN;
227
228/* A list of all slab caches on the system */
229static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700230static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700231
Christoph Lameter02cbc872007-05-09 02:32:43 -0700232/*
233 * Tracking user of a slab.
234 */
235struct track {
236 void *addr; /* Called from address */
237 int cpu; /* Was running on cpu */
238 int pid; /* Pid context */
239 unsigned long when; /* When did the operation occur */
240};
241
242enum track_item { TRACK_ALLOC, TRACK_FREE };
243
Christoph Lameter41ecc552007-05-09 02:32:44 -0700244#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter81819f02007-05-06 14:49:36 -0700245static int sysfs_slab_add(struct kmem_cache *);
246static int sysfs_slab_alias(struct kmem_cache *, const char *);
247static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800248
Christoph Lameter81819f02007-05-06 14:49:36 -0700249#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700250static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
252 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800253static inline void sysfs_slab_remove(struct kmem_cache *s)
254{
255 kfree(s);
256}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800257
Christoph Lameter81819f02007-05-06 14:49:36 -0700258#endif
259
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800260static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
261{
262#ifdef CONFIG_SLUB_STATS
263 c->stat[si]++;
264#endif
265}
266
Christoph Lameter81819f02007-05-06 14:49:36 -0700267/********************************************************************
268 * Core slab cache functions
269 *******************************************************************/
270
271int slab_is_available(void)
272{
273 return slab_state >= UP;
274}
275
276static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
277{
278#ifdef CONFIG_NUMA
279 return s->node[node];
280#else
281 return &s->local_node;
282#endif
283}
284
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700285static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
286{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700287#ifdef CONFIG_SMP
288 return s->cpu_slab[cpu];
289#else
290 return &s->cpu_slab;
291#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700292}
293
Christoph Lameter6446faa2008-02-15 23:45:26 -0800294/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700295static inline int check_valid_pointer(struct kmem_cache *s,
296 struct page *page, const void *object)
297{
298 void *base;
299
Christoph Lametera973e9d2008-03-01 13:40:44 -0800300 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700301 return 1;
302
Christoph Lametera973e9d2008-03-01 13:40:44 -0800303 base = page_address(page);
Christoph Lameter02cbc872007-05-09 02:32:43 -0700304 if (object < base || object >= base + s->objects * s->size ||
305 (object - base) % s->size) {
306 return 0;
307 }
308
309 return 1;
310}
311
Christoph Lameter81819f02007-05-06 14:49:36 -0700312/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700313 * Slow version of get and set free pointer.
314 *
315 * This version requires touching the cache lines of kmem_cache which
316 * we avoid to do in the fast alloc free paths. There we obtain the offset
317 * from the page struct.
318 */
319static inline void *get_freepointer(struct kmem_cache *s, void *object)
320{
321 return *(void **)(object + s->offset);
322}
323
324static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
325{
326 *(void **)(object + s->offset) = fp;
327}
328
329/* Loop over all objects in a slab */
330#define for_each_object(__p, __s, __addr) \
331 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
332 __p += (__s)->size)
333
334/* Scan freelist */
335#define for_each_free_object(__p, __s, __free) \
Christoph Lametera973e9d2008-03-01 13:40:44 -0800336 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700337
338/* Determine object index from a given position */
339static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
340{
341 return (p - addr) / s->size;
342}
343
Christoph Lameter41ecc552007-05-09 02:32:44 -0700344#ifdef CONFIG_SLUB_DEBUG
345/*
346 * Debug settings:
347 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700348#ifdef CONFIG_SLUB_DEBUG_ON
349static int slub_debug = DEBUG_DEFAULT_FLAGS;
350#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700351static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700352#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700353
354static char *slub_debug_slabs;
355
Christoph Lameter7656c722007-05-09 02:32:40 -0700356/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700357 * Object debugging
358 */
359static void print_section(char *text, u8 *addr, unsigned int length)
360{
361 int i, offset;
362 int newline = 1;
363 char ascii[17];
364
365 ascii[16] = 0;
366
367 for (i = 0; i < length; i++) {
368 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700369 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700370 newline = 0;
371 }
Pekka Enberg06428782008-01-07 23:20:27 -0800372 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700373 offset = i % 16;
374 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
375 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800376 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700377 newline = 1;
378 }
379 }
380 if (!newline) {
381 i %= 16;
382 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800383 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700384 ascii[i] = ' ';
385 i++;
386 }
Pekka Enberg06428782008-01-07 23:20:27 -0800387 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700388 }
389}
390
Christoph Lameter81819f02007-05-06 14:49:36 -0700391static struct track *get_track(struct kmem_cache *s, void *object,
392 enum track_item alloc)
393{
394 struct track *p;
395
396 if (s->offset)
397 p = object + s->offset + sizeof(void *);
398 else
399 p = object + s->inuse;
400
401 return p + alloc;
402}
403
404static void set_track(struct kmem_cache *s, void *object,
405 enum track_item alloc, void *addr)
406{
407 struct track *p;
408
409 if (s->offset)
410 p = object + s->offset + sizeof(void *);
411 else
412 p = object + s->inuse;
413
414 p += alloc;
415 if (addr) {
416 p->addr = addr;
417 p->cpu = smp_processor_id();
418 p->pid = current ? current->pid : -1;
419 p->when = jiffies;
420 } else
421 memset(p, 0, sizeof(struct track));
422}
423
Christoph Lameter81819f02007-05-06 14:49:36 -0700424static void init_tracking(struct kmem_cache *s, void *object)
425{
Christoph Lameter24922682007-07-17 04:03:18 -0700426 if (!(s->flags & SLAB_STORE_USER))
427 return;
428
429 set_track(s, object, TRACK_FREE, NULL);
430 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700431}
432
433static void print_track(const char *s, struct track *t)
434{
435 if (!t->addr)
436 return;
437
Christoph Lameter24922682007-07-17 04:03:18 -0700438 printk(KERN_ERR "INFO: %s in ", s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700439 __print_symbol("%s", (unsigned long)t->addr);
Christoph Lameter24922682007-07-17 04:03:18 -0700440 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700441}
442
Christoph Lameter24922682007-07-17 04:03:18 -0700443static void print_tracking(struct kmem_cache *s, void *object)
444{
445 if (!(s->flags & SLAB_STORE_USER))
446 return;
447
448 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
449 print_track("Freed", get_track(s, object, TRACK_FREE));
450}
451
452static void print_page_info(struct page *page)
453{
454 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
455 page, page->inuse, page->freelist, page->flags);
456
457}
458
459static void slab_bug(struct kmem_cache *s, char *fmt, ...)
460{
461 va_list args;
462 char buf[100];
463
464 va_start(args, fmt);
465 vsnprintf(buf, sizeof(buf), fmt, args);
466 va_end(args);
467 printk(KERN_ERR "========================================"
468 "=====================================\n");
469 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
470 printk(KERN_ERR "----------------------------------------"
471 "-------------------------------------\n\n");
472}
473
474static void slab_fix(struct kmem_cache *s, char *fmt, ...)
475{
476 va_list args;
477 char buf[100];
478
479 va_start(args, fmt);
480 vsnprintf(buf, sizeof(buf), fmt, args);
481 va_end(args);
482 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
483}
484
485static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700486{
487 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800488 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700489
490 print_tracking(s, p);
491
492 print_page_info(page);
493
494 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
495 p, p - addr, get_freepointer(s, p));
496
497 if (p > addr + 16)
498 print_section("Bytes b4", p - 16, 16);
499
500 print_section("Object", p, min(s->objsize, 128));
Christoph Lameter81819f02007-05-06 14:49:36 -0700501
502 if (s->flags & SLAB_RED_ZONE)
503 print_section("Redzone", p + s->objsize,
504 s->inuse - s->objsize);
505
Christoph Lameter81819f02007-05-06 14:49:36 -0700506 if (s->offset)
507 off = s->offset + sizeof(void *);
508 else
509 off = s->inuse;
510
Christoph Lameter24922682007-07-17 04:03:18 -0700511 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700512 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700513
514 if (off != s->size)
515 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700516 print_section("Padding", p + off, s->size - off);
517
518 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700519}
520
521static void object_err(struct kmem_cache *s, struct page *page,
522 u8 *object, char *reason)
523{
Christoph Lameter24922682007-07-17 04:03:18 -0700524 slab_bug(s, reason);
525 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700526}
527
Christoph Lameter24922682007-07-17 04:03:18 -0700528static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700529{
530 va_list args;
531 char buf[100];
532
Christoph Lameter24922682007-07-17 04:03:18 -0700533 va_start(args, fmt);
534 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700535 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700536 slab_bug(s, fmt);
537 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700538 dump_stack();
539}
540
541static void init_object(struct kmem_cache *s, void *object, int active)
542{
543 u8 *p = object;
544
545 if (s->flags & __OBJECT_POISON) {
546 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800547 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700548 }
549
550 if (s->flags & SLAB_RED_ZONE)
551 memset(p + s->objsize,
552 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
553 s->inuse - s->objsize);
554}
555
Christoph Lameter24922682007-07-17 04:03:18 -0700556static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700557{
558 while (bytes) {
559 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700560 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700561 start++;
562 bytes--;
563 }
Christoph Lameter24922682007-07-17 04:03:18 -0700564 return NULL;
565}
566
567static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
568 void *from, void *to)
569{
570 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
571 memset(from, data, to - from);
572}
573
574static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
575 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800576 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700577{
578 u8 *fault;
579 u8 *end;
580
581 fault = check_bytes(start, value, bytes);
582 if (!fault)
583 return 1;
584
585 end = start + bytes;
586 while (end > fault && end[-1] == value)
587 end--;
588
589 slab_bug(s, "%s overwritten", what);
590 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
591 fault, end - 1, fault[0], value);
592 print_trailer(s, page, object);
593
594 restore_bytes(s, what, value, fault, end);
595 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700596}
597
Christoph Lameter81819f02007-05-06 14:49:36 -0700598/*
599 * Object layout:
600 *
601 * object address
602 * Bytes of the object to be managed.
603 * If the freepointer may overlay the object then the free
604 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700605 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700606 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
607 * 0xa5 (POISON_END)
608 *
609 * object + s->objsize
610 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700611 * Padding is extended by another word if Redzoning is enabled and
612 * objsize == inuse.
613 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700614 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
615 * 0xcc (RED_ACTIVE) for objects in use.
616 *
617 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700618 * Meta data starts here.
619 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700620 * A. Free pointer (if we cannot overwrite object on free)
621 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700622 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800623 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700624 * before the word boundary.
625 *
626 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700627 *
628 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700629 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700630 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700631 * If slabcaches are merged then the objsize and inuse boundaries are mostly
632 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700633 * may be used with merged slabcaches.
634 */
635
Christoph Lameter81819f02007-05-06 14:49:36 -0700636static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
637{
638 unsigned long off = s->inuse; /* The end of info */
639
640 if (s->offset)
641 /* Freepointer is placed after the object. */
642 off += sizeof(void *);
643
644 if (s->flags & SLAB_STORE_USER)
645 /* We also have user information there */
646 off += 2 * sizeof(struct track);
647
648 if (s->size == off)
649 return 1;
650
Christoph Lameter24922682007-07-17 04:03:18 -0700651 return check_bytes_and_report(s, page, p, "Object padding",
652 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700653}
654
655static int slab_pad_check(struct kmem_cache *s, struct page *page)
656{
Christoph Lameter24922682007-07-17 04:03:18 -0700657 u8 *start;
658 u8 *fault;
659 u8 *end;
660 int length;
661 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700662
663 if (!(s->flags & SLAB_POISON))
664 return 1;
665
Christoph Lametera973e9d2008-03-01 13:40:44 -0800666 start = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700667 end = start + (PAGE_SIZE << s->order);
Christoph Lameter81819f02007-05-06 14:49:36 -0700668 length = s->objects * s->size;
Christoph Lameter24922682007-07-17 04:03:18 -0700669 remainder = end - (start + length);
Christoph Lameter81819f02007-05-06 14:49:36 -0700670 if (!remainder)
671 return 1;
672
Christoph Lameter24922682007-07-17 04:03:18 -0700673 fault = check_bytes(start + length, POISON_INUSE, remainder);
674 if (!fault)
675 return 1;
676 while (end > fault && end[-1] == POISON_INUSE)
677 end--;
678
679 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
680 print_section("Padding", start, length);
681
682 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
683 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700684}
685
686static int check_object(struct kmem_cache *s, struct page *page,
687 void *object, int active)
688{
689 u8 *p = object;
690 u8 *endobject = object + s->objsize;
691
692 if (s->flags & SLAB_RED_ZONE) {
693 unsigned int red =
694 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
695
Christoph Lameter24922682007-07-17 04:03:18 -0700696 if (!check_bytes_and_report(s, page, object, "Redzone",
697 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700698 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700699 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800700 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
701 check_bytes_and_report(s, page, p, "Alignment padding",
702 endobject, POISON_INUSE, s->inuse - s->objsize);
703 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700704 }
705
706 if (s->flags & SLAB_POISON) {
707 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700708 (!check_bytes_and_report(s, page, p, "Poison", p,
709 POISON_FREE, s->objsize - 1) ||
710 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800711 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700712 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700713 /*
714 * check_pad_bytes cleans up on its own.
715 */
716 check_pad_bytes(s, page, p);
717 }
718
719 if (!s->offset && active)
720 /*
721 * Object and freepointer overlap. Cannot check
722 * freepointer while object is allocated.
723 */
724 return 1;
725
726 /* Check free pointer validity */
727 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
728 object_err(s, page, p, "Freepointer corrupt");
729 /*
730 * No choice but to zap it and thus loose the remainder
731 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700732 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700733 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800734 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700735 return 0;
736 }
737 return 1;
738}
739
740static int check_slab(struct kmem_cache *s, struct page *page)
741{
742 VM_BUG_ON(!irqs_disabled());
743
744 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700745 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700746 return 0;
747 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700748 if (page->inuse > s->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700749 slab_err(s, page, "inuse %u > max %u",
750 s->name, page->inuse, s->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700751 return 0;
752 }
753 /* Slab_pad_check fixes things up after itself */
754 slab_pad_check(s, page);
755 return 1;
756}
757
758/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700759 * Determine if a certain object on a page is on the freelist. Must hold the
760 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700761 */
762static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
763{
764 int nr = 0;
765 void *fp = page->freelist;
766 void *object = NULL;
767
Christoph Lametera973e9d2008-03-01 13:40:44 -0800768 while (fp && nr <= s->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700769 if (fp == search)
770 return 1;
771 if (!check_valid_pointer(s, page, fp)) {
772 if (object) {
773 object_err(s, page, object,
774 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800775 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700776 break;
777 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700778 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800779 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700780 page->inuse = s->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700781 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700782 return 0;
783 }
784 break;
785 }
786 object = fp;
787 fp = get_freepointer(s, object);
788 nr++;
789 }
790
791 if (page->inuse != s->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700792 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter24922682007-07-17 04:03:18 -0700793 "counted were %d", page->inuse, s->objects - nr);
Christoph Lameter81819f02007-05-06 14:49:36 -0700794 page->inuse = s->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700795 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700796 }
797 return search == NULL;
798}
799
Christoph Lameter3ec09742007-05-16 22:11:00 -0700800static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
801{
802 if (s->flags & SLAB_TRACE) {
803 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
804 s->name,
805 alloc ? "alloc" : "free",
806 object, page->inuse,
807 page->freelist);
808
809 if (!alloc)
810 print_section("Object", (void *)object, s->objsize);
811
812 dump_stack();
813 }
814}
815
Christoph Lameter643b1132007-05-06 14:49:42 -0700816/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700817 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700818 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700819static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700820{
Christoph Lameter643b1132007-05-06 14:49:42 -0700821 spin_lock(&n->list_lock);
822 list_add(&page->lru, &n->full);
823 spin_unlock(&n->list_lock);
824}
825
826static void remove_full(struct kmem_cache *s, struct page *page)
827{
828 struct kmem_cache_node *n;
829
830 if (!(s->flags & SLAB_STORE_USER))
831 return;
832
833 n = get_node(s, page_to_nid(page));
834
835 spin_lock(&n->list_lock);
836 list_del(&page->lru);
837 spin_unlock(&n->list_lock);
838}
839
Christoph Lameter3ec09742007-05-16 22:11:00 -0700840static void setup_object_debug(struct kmem_cache *s, struct page *page,
841 void *object)
842{
843 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
844 return;
845
846 init_object(s, object, 0);
847 init_tracking(s, object);
848}
849
850static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
851 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700852{
853 if (!check_slab(s, page))
854 goto bad;
855
Christoph Lameterd692ef62008-02-15 23:45:24 -0800856 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700857 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700858 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700859 }
860
861 if (!check_valid_pointer(s, page, object)) {
862 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700863 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700864 }
865
Christoph Lameterd692ef62008-02-15 23:45:24 -0800866 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700867 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700868
Christoph Lameter3ec09742007-05-16 22:11:00 -0700869 /* Success perform special debug activities for allocs */
870 if (s->flags & SLAB_STORE_USER)
871 set_track(s, object, TRACK_ALLOC, addr);
872 trace(s, page, object, 1);
873 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700874 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700875
Christoph Lameter81819f02007-05-06 14:49:36 -0700876bad:
877 if (PageSlab(page)) {
878 /*
879 * If this is a slab page then lets do the best we can
880 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700881 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700882 */
Christoph Lameter24922682007-07-17 04:03:18 -0700883 slab_fix(s, "Marking all objects used");
Christoph Lameter81819f02007-05-06 14:49:36 -0700884 page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800885 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700886 }
887 return 0;
888}
889
Christoph Lameter3ec09742007-05-16 22:11:00 -0700890static int free_debug_processing(struct kmem_cache *s, struct page *page,
891 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700892{
893 if (!check_slab(s, page))
894 goto fail;
895
896 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700897 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700898 goto fail;
899 }
900
901 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700902 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700903 goto fail;
904 }
905
906 if (!check_object(s, page, object, 1))
907 return 0;
908
909 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800910 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700911 slab_err(s, page, "Attempt to free object(0x%p) "
912 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800913 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700914 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700915 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700916 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700917 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800918 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700919 object_err(s, page, object,
920 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700921 goto fail;
922 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700923
924 /* Special debug activities for freeing objects */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800925 if (!SlabFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700926 remove_full(s, page);
927 if (s->flags & SLAB_STORE_USER)
928 set_track(s, object, TRACK_FREE, addr);
929 trace(s, page, object, 0);
930 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700931 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700932
Christoph Lameter81819f02007-05-06 14:49:36 -0700933fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700934 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700935 return 0;
936}
937
Christoph Lameter41ecc552007-05-09 02:32:44 -0700938static int __init setup_slub_debug(char *str)
939{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700940 slub_debug = DEBUG_DEFAULT_FLAGS;
941 if (*str++ != '=' || !*str)
942 /*
943 * No options specified. Switch on full debugging.
944 */
945 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700946
947 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700948 /*
949 * No options but restriction on slabs. This means full
950 * debugging for slabs matching a pattern.
951 */
952 goto check_slabs;
953
954 slub_debug = 0;
955 if (*str == '-')
956 /*
957 * Switch off all debugging measures.
958 */
959 goto out;
960
961 /*
962 * Determine which debug features should be switched on
963 */
Pekka Enberg06428782008-01-07 23:20:27 -0800964 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700965 switch (tolower(*str)) {
966 case 'f':
967 slub_debug |= SLAB_DEBUG_FREE;
968 break;
969 case 'z':
970 slub_debug |= SLAB_RED_ZONE;
971 break;
972 case 'p':
973 slub_debug |= SLAB_POISON;
974 break;
975 case 'u':
976 slub_debug |= SLAB_STORE_USER;
977 break;
978 case 't':
979 slub_debug |= SLAB_TRACE;
980 break;
981 default:
982 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -0800983 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700984 }
985 }
986
987check_slabs:
988 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -0700989 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700990out:
Christoph Lameter41ecc552007-05-09 02:32:44 -0700991 return 1;
992}
993
994__setup("slub_debug", setup_slub_debug);
995
Christoph Lameterba0268a2007-09-11 15:24:11 -0700996static unsigned long kmem_cache_flags(unsigned long objsize,
997 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -0700998 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -0700999{
1000 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001001 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001002 */
Christoph Lametere1533622008-02-15 23:45:24 -08001003 if (slub_debug && (!slub_debug_slabs ||
1004 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1005 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001006
1007 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001008}
1009#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001010static inline void setup_object_debug(struct kmem_cache *s,
1011 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001012
Christoph Lameter3ec09742007-05-16 22:11:00 -07001013static inline int alloc_debug_processing(struct kmem_cache *s,
1014 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001015
Christoph Lameter3ec09742007-05-16 22:11:00 -07001016static inline int free_debug_processing(struct kmem_cache *s,
1017 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001018
Christoph Lameter41ecc552007-05-09 02:32:44 -07001019static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1020 { return 1; }
1021static inline int check_object(struct kmem_cache *s, struct page *page,
1022 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001023static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001024static inline unsigned long kmem_cache_flags(unsigned long objsize,
1025 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001026 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001027{
1028 return flags;
1029}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001030#define slub_debug 0
1031#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001032/*
1033 * Slab allocation and freeing
1034 */
1035static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1036{
Pekka Enberg06428782008-01-07 23:20:27 -08001037 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001038 int pages = 1 << s->order;
1039
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001040 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001041
Christoph Lameter81819f02007-05-06 14:49:36 -07001042 if (node == -1)
1043 page = alloc_pages(flags, s->order);
1044 else
1045 page = alloc_pages_node(node, flags, s->order);
1046
1047 if (!page)
1048 return NULL;
1049
1050 mod_zone_page_state(page_zone(page),
1051 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1052 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1053 pages);
1054
1055 return page;
1056}
1057
1058static void setup_object(struct kmem_cache *s, struct page *page,
1059 void *object)
1060{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001061 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001062 if (unlikely(s->ctor))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001063 s->ctor(s, object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001064}
1065
1066static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1067{
1068 struct page *page;
1069 struct kmem_cache_node *n;
1070 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001071 void *last;
1072 void *p;
1073
Christoph Lameter6cb06222007-10-16 01:25:41 -07001074 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001075
Christoph Lameter6cb06222007-10-16 01:25:41 -07001076 page = allocate_slab(s,
1077 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001078 if (!page)
1079 goto out;
1080
1081 n = get_node(s, page_to_nid(page));
1082 if (n)
1083 atomic_long_inc(&n->nr_slabs);
Christoph Lameter81819f02007-05-06 14:49:36 -07001084 page->slab = s;
1085 page->flags |= 1 << PG_slab;
1086 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1087 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001088 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001089
1090 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001091
1092 if (unlikely(s->flags & SLAB_POISON))
1093 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1094
1095 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001096 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001097 setup_object(s, page, last);
1098 set_freepointer(s, last, p);
1099 last = p;
1100 }
1101 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001102 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001103
1104 page->freelist = start;
1105 page->inuse = 0;
1106out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001107 return page;
1108}
1109
1110static void __free_slab(struct kmem_cache *s, struct page *page)
1111{
1112 int pages = 1 << s->order;
1113
Christoph Lameterc59def92007-05-16 22:10:50 -07001114 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001115 void *p;
1116
1117 slab_pad_check(s, page);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001118 for_each_object(p, s, page_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001119 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001120 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001121 }
1122
1123 mod_zone_page_state(page_zone(page),
1124 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1125 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001126 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001127
Christoph Lameter81819f02007-05-06 14:49:36 -07001128 __free_pages(page, s->order);
1129}
1130
1131static void rcu_free_slab(struct rcu_head *h)
1132{
1133 struct page *page;
1134
1135 page = container_of((struct list_head *)h, struct page, lru);
1136 __free_slab(page->slab, page);
1137}
1138
1139static void free_slab(struct kmem_cache *s, struct page *page)
1140{
1141 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1142 /*
1143 * RCU free overloads the RCU head over the LRU
1144 */
1145 struct rcu_head *head = (void *)&page->lru;
1146
1147 call_rcu(head, rcu_free_slab);
1148 } else
1149 __free_slab(s, page);
1150}
1151
1152static void discard_slab(struct kmem_cache *s, struct page *page)
1153{
1154 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1155
1156 atomic_long_dec(&n->nr_slabs);
1157 reset_page_mapcount(page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001158 __ClearPageSlab(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001159 free_slab(s, page);
1160}
1161
1162/*
1163 * Per slab locking using the pagelock
1164 */
1165static __always_inline void slab_lock(struct page *page)
1166{
1167 bit_spin_lock(PG_locked, &page->flags);
1168}
1169
1170static __always_inline void slab_unlock(struct page *page)
1171{
Nick Piggina76d3542008-01-07 23:20:27 -08001172 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001173}
1174
1175static __always_inline int slab_trylock(struct page *page)
1176{
1177 int rc = 1;
1178
1179 rc = bit_spin_trylock(PG_locked, &page->flags);
1180 return rc;
1181}
1182
1183/*
1184 * Management of partially allocated slabs
1185 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001186static void add_partial(struct kmem_cache_node *n,
1187 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001188{
Christoph Lametere95eed52007-05-06 14:49:44 -07001189 spin_lock(&n->list_lock);
1190 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001191 if (tail)
1192 list_add_tail(&page->lru, &n->partial);
1193 else
1194 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001195 spin_unlock(&n->list_lock);
1196}
1197
1198static void remove_partial(struct kmem_cache *s,
1199 struct page *page)
1200{
1201 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1202
1203 spin_lock(&n->list_lock);
1204 list_del(&page->lru);
1205 n->nr_partial--;
1206 spin_unlock(&n->list_lock);
1207}
1208
1209/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001210 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001211 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001212 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001213 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001214static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001215{
1216 if (slab_trylock(page)) {
1217 list_del(&page->lru);
1218 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001219 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001220 return 1;
1221 }
1222 return 0;
1223}
1224
1225/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001226 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001227 */
1228static struct page *get_partial_node(struct kmem_cache_node *n)
1229{
1230 struct page *page;
1231
1232 /*
1233 * Racy check. If we mistakenly see no partial slabs then we
1234 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001235 * partial slab and there is none available then get_partials()
1236 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001237 */
1238 if (!n || !n->nr_partial)
1239 return NULL;
1240
1241 spin_lock(&n->list_lock);
1242 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001243 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001244 goto out;
1245 page = NULL;
1246out:
1247 spin_unlock(&n->list_lock);
1248 return page;
1249}
1250
1251/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001252 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001253 */
1254static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1255{
1256#ifdef CONFIG_NUMA
1257 struct zonelist *zonelist;
1258 struct zone **z;
1259 struct page *page;
1260
1261 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001262 * The defrag ratio allows a configuration of the tradeoffs between
1263 * inter node defragmentation and node local allocations. A lower
1264 * defrag_ratio increases the tendency to do local allocations
1265 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001266 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001267 * If the defrag_ratio is set to 0 then kmalloc() always
1268 * returns node local objects. If the ratio is higher then kmalloc()
1269 * may return off node objects because partial slabs are obtained
1270 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001271 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001272 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001273 * defrag_ratio = 1000) then every (well almost) allocation will
1274 * first attempt to defrag slab caches on other nodes. This means
1275 * scanning over all nodes to look for partial slabs which may be
1276 * expensive if we do it every time we are trying to find a slab
1277 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001278 */
Christoph Lameter98246012008-01-07 23:20:26 -08001279 if (!s->remote_node_defrag_ratio ||
1280 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001281 return NULL;
1282
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001283 zonelist = &NODE_DATA(
1284 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
Christoph Lameter81819f02007-05-06 14:49:36 -07001285 for (z = zonelist->zones; *z; z++) {
1286 struct kmem_cache_node *n;
1287
1288 n = get_node(s, zone_to_nid(*z));
1289
1290 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001291 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001292 page = get_partial_node(n);
1293 if (page)
1294 return page;
1295 }
1296 }
1297#endif
1298 return NULL;
1299}
1300
1301/*
1302 * Get a partial page, lock it and return it.
1303 */
1304static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1305{
1306 struct page *page;
1307 int searchnode = (node == -1) ? numa_node_id() : node;
1308
1309 page = get_partial_node(get_node(s, searchnode));
1310 if (page || (flags & __GFP_THISNODE))
1311 return page;
1312
1313 return get_any_partial(s, flags);
1314}
1315
1316/*
1317 * Move a page back to the lists.
1318 *
1319 * Must be called with the slab lock held.
1320 *
1321 * On exit the slab lock will have been dropped.
1322 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001323static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001324{
Christoph Lametere95eed52007-05-06 14:49:44 -07001325 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001326 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001327
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001328 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001329 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001330
Christoph Lametera973e9d2008-03-01 13:40:44 -08001331 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001332 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001333 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1334 } else {
1335 stat(c, DEACTIVATE_FULL);
1336 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1337 add_full(n, page);
1338 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001339 slab_unlock(page);
1340 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001341 stat(c, DEACTIVATE_EMPTY);
Christoph Lametere95eed52007-05-06 14:49:44 -07001342 if (n->nr_partial < MIN_PARTIAL) {
1343 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001344 * Adding an empty slab to the partial slabs in order
1345 * to avoid page allocator overhead. This slab needs
1346 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001347 * so that the others get filled first. That way the
1348 * size of the partial list stays small.
1349 *
1350 * kmem_cache_shrink can reclaim any empty slabs from the
1351 * partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001352 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001353 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001354 slab_unlock(page);
1355 } else {
1356 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001357 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001358 discard_slab(s, page);
1359 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001360 }
1361}
1362
1363/*
1364 * Remove the cpu slab
1365 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001366static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001367{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001368 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001369 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001370
Christoph Lameterb773ad72008-03-04 11:10:17 -08001371 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001372 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001373 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001374 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001375 * because both freelists are empty. So this is unlikely
1376 * to occur.
1377 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001378 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001379 void **object;
1380
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001381 tail = 0; /* Hot objects. Put the slab first */
1382
Christoph Lameter894b8782007-05-10 03:15:16 -07001383 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001384 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001385 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001386
1387 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001388 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001389 page->freelist = object;
1390 page->inuse--;
1391 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001392 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001393 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001394}
1395
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001396static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001397{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001398 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001399 slab_lock(c->page);
1400 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001401}
1402
1403/*
1404 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001405 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001406 * Called from IPI handler with interrupts disabled.
1407 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001408static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001409{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001410 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001411
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001412 if (likely(c && c->page))
1413 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001414}
1415
1416static void flush_cpu_slab(void *d)
1417{
1418 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001419
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001420 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001421}
1422
1423static void flush_all(struct kmem_cache *s)
1424{
1425#ifdef CONFIG_SMP
1426 on_each_cpu(flush_cpu_slab, s, 1, 1);
1427#else
1428 unsigned long flags;
1429
1430 local_irq_save(flags);
1431 flush_cpu_slab(s);
1432 local_irq_restore(flags);
1433#endif
1434}
1435
1436/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001437 * Check if the objects in a per cpu structure fit numa
1438 * locality expectations.
1439 */
1440static inline int node_match(struct kmem_cache_cpu *c, int node)
1441{
1442#ifdef CONFIG_NUMA
1443 if (node != -1 && c->node != node)
1444 return 0;
1445#endif
1446 return 1;
1447}
1448
1449/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001450 * Slow path. The lockless freelist is empty or we need to perform
1451 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001452 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001453 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001454 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001455 * Processing is still very fast if new objects have been freed to the
1456 * regular freelist. In that case we simply take over the regular freelist
1457 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001458 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001459 * If that is not working then we fall back to the partial lists. We take the
1460 * first element of the freelist as the object to allocate now and move the
1461 * rest of the freelist to the lockless freelist.
1462 *
1463 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001464 * we need to allocate a new slab. This is the slowest path since it involves
1465 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001466 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001467static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001468 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001469{
Christoph Lameter81819f02007-05-06 14:49:36 -07001470 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001471 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001472
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001473 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001474 goto new_slab;
1475
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001476 slab_lock(c->page);
1477 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001478 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001479
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001480 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001481
Christoph Lameter894b8782007-05-10 03:15:16 -07001482load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001483 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001484 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001485 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001486 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001487 goto debug;
1488
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001489 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001490 c->page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001491 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001492 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001493unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001494 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001495 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001496 return object;
1497
1498another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001499 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001500
1501new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001502 new = get_partial(s, gfpflags, node);
1503 if (new) {
1504 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001505 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001506 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001507 }
1508
Christoph Lameterb811c202007-10-16 23:25:51 -07001509 if (gfpflags & __GFP_WAIT)
1510 local_irq_enable();
1511
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001512 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001513
1514 if (gfpflags & __GFP_WAIT)
1515 local_irq_disable();
1516
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001517 if (new) {
1518 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001519 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001520 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001521 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001522 slab_lock(new);
1523 SetSlabFrozen(new);
1524 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001525 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001526 }
Linus Torvalds00e962c2008-02-19 09:08:49 -08001527
Christoph Lameter71c7a062008-02-14 14:28:01 -08001528 /*
1529 * No memory available.
1530 *
1531 * If the slab uses higher order allocs but the object is
1532 * smaller than a page size then we can fallback in emergencies
1533 * to the page allocator via kmalloc_large. The page allocator may
1534 * have failed to obtain a higher order page and we can try to
1535 * allocate a single page if the object fits into a single page.
1536 * That is only possible if certain conditions are met that are being
1537 * checked when a slab is created.
1538 */
1539 if (!(gfpflags & __GFP_NORETRY) && (s->flags & __PAGE_ALLOC_FALLBACK))
1540 return kmalloc_large(s->objsize, gfpflags);
1541
1542 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001543debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001544 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001545 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001546
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001547 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001548 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001549 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001550 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001551}
1552
1553/*
1554 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1555 * have the fastpath folded into their functions. So no function call
1556 * overhead for requests that can be satisfied on the fastpath.
1557 *
1558 * The fastpath works by first checking if the lockless freelist can be used.
1559 * If not then __slab_alloc is called for slow processing.
1560 *
1561 * Otherwise we can simply pick the next object from the lockless free list.
1562 */
Pekka Enberg06428782008-01-07 23:20:27 -08001563static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001564 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001565{
Christoph Lameter894b8782007-05-10 03:15:16 -07001566 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001567 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001568 unsigned long flags;
1569
Christoph Lameter894b8782007-05-10 03:15:16 -07001570 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001571 c = get_cpu_slab(s, smp_processor_id());
Christoph Lametera973e9d2008-03-01 13:40:44 -08001572 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001573
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001574 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001575
1576 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001577 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001578 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001579 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001580 }
1581 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001582
1583 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001584 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001585
Christoph Lameter894b8782007-05-10 03:15:16 -07001586 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001587}
1588
1589void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1590{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001591 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001592}
1593EXPORT_SYMBOL(kmem_cache_alloc);
1594
1595#ifdef CONFIG_NUMA
1596void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1597{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001598 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001599}
1600EXPORT_SYMBOL(kmem_cache_alloc_node);
1601#endif
1602
1603/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001604 * Slow patch handling. This may still be called frequently since objects
1605 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001606 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001607 * So we still attempt to reduce cache line usage. Just take the slab
1608 * lock and free the item. If there is no additional partial page
1609 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001610 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001611static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001612 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001613{
1614 void *prior;
1615 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001616 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001617
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001618 c = get_cpu_slab(s, raw_smp_processor_id());
1619 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001620 slab_lock(page);
1621
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001622 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001623 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001624
Christoph Lameter81819f02007-05-06 14:49:36 -07001625checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001626 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001627 page->freelist = object;
1628 page->inuse--;
1629
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001630 if (unlikely(SlabFrozen(page))) {
1631 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001632 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001633 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001634
1635 if (unlikely(!page->inuse))
1636 goto slab_empty;
1637
1638 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001639 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001640 * then add it.
1641 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001642 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001643 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001644 stat(c, FREE_ADD_PARTIAL);
1645 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001646
1647out_unlock:
1648 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001649 return;
1650
1651slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001652 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001653 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001654 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001655 */
1656 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001657 stat(c, FREE_REMOVE_PARTIAL);
1658 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001659 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001660 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001661 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001662 return;
1663
1664debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001665 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001666 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001667 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001668}
1669
Christoph Lameter894b8782007-05-10 03:15:16 -07001670/*
1671 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1672 * can perform fastpath freeing without additional function calls.
1673 *
1674 * The fastpath is only possible if we are freeing to the current cpu slab
1675 * of this processor. This typically the case if we have just allocated
1676 * the item before.
1677 *
1678 * If fastpath is not possible then fall back to __slab_free where we deal
1679 * with all sorts of special processing.
1680 */
Pekka Enberg06428782008-01-07 23:20:27 -08001681static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001682 struct page *page, void *x, void *addr)
1683{
1684 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001685 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001686 unsigned long flags;
1687
Christoph Lameter894b8782007-05-10 03:15:16 -07001688 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001689 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001690 debug_check_no_locks_freed(object, c->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001691 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001692 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001693 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001694 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001695 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001696 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001697
1698 local_irq_restore(flags);
1699}
1700
Christoph Lameter81819f02007-05-06 14:49:36 -07001701void kmem_cache_free(struct kmem_cache *s, void *x)
1702{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001703 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001704
Christoph Lameterb49af682007-05-06 14:49:41 -07001705 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001706
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001707 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001708}
1709EXPORT_SYMBOL(kmem_cache_free);
1710
1711/* Figure out on which slab object the object resides */
1712static struct page *get_object_page(const void *x)
1713{
Christoph Lameterb49af682007-05-06 14:49:41 -07001714 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001715
1716 if (!PageSlab(page))
1717 return NULL;
1718
1719 return page;
1720}
1721
1722/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001723 * Object placement in a slab is made very easy because we always start at
1724 * offset 0. If we tune the size of the object to the alignment then we can
1725 * get the required alignment by putting one properly sized object after
1726 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001727 *
1728 * Notice that the allocation order determines the sizes of the per cpu
1729 * caches. Each processor has always one slab available for allocations.
1730 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001731 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001732 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001733 */
1734
1735/*
1736 * Mininum / Maximum order of slab pages. This influences locking overhead
1737 * and slab fragmentation. A higher order reduces the number of partial slabs
1738 * and increases the number of allocations possible without having to
1739 * take the list_lock.
1740 */
1741static int slub_min_order;
1742static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001743static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1744
1745/*
1746 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001747 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001748 */
1749static int slub_nomerge;
1750
1751/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001752 * Calculate the order of allocation given an slab object size.
1753 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001754 * The order of allocation has significant impact on performance and other
1755 * system components. Generally order 0 allocations should be preferred since
1756 * order 0 does not cause fragmentation in the page allocator. Larger objects
1757 * be problematic to put into order 0 slabs because there may be too much
1758 * unused space left. We go to a higher order if more than 1/8th of the slab
1759 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001760 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001761 * In order to reach satisfactory performance we must ensure that a minimum
1762 * number of objects is in one slab. Otherwise we may generate too much
1763 * activity on the partial lists which requires taking the list_lock. This is
1764 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001765 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001766 * slub_max_order specifies the order where we begin to stop considering the
1767 * number of objects in a slab as critical. If we reach slub_max_order then
1768 * we try to keep the page order as low as possible. So we accept more waste
1769 * of space in favor of a small page order.
1770 *
1771 * Higher order allocations also allow the placement of more objects in a
1772 * slab and thereby reduce object handling overhead. If the user has
1773 * requested a higher mininum order then we start with that one instead of
1774 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001775 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001776static inline int slab_order(int size, int min_objects,
1777 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001778{
1779 int order;
1780 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001781 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001782
Christoph Lameter6300ea72007-07-17 04:03:20 -07001783 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001784 fls(min_objects * size - 1) - PAGE_SHIFT);
1785 order <= max_order; order++) {
1786
Christoph Lameter81819f02007-05-06 14:49:36 -07001787 unsigned long slab_size = PAGE_SIZE << order;
1788
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001789 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001790 continue;
1791
Christoph Lameter81819f02007-05-06 14:49:36 -07001792 rem = slab_size % size;
1793
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001794 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001795 break;
1796
1797 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001798
Christoph Lameter81819f02007-05-06 14:49:36 -07001799 return order;
1800}
1801
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001802static inline int calculate_order(int size)
1803{
1804 int order;
1805 int min_objects;
1806 int fraction;
1807
1808 /*
1809 * Attempt to find best configuration for a slab. This
1810 * works by first attempting to generate a layout with
1811 * the best configuration and backing off gradually.
1812 *
1813 * First we reduce the acceptable waste in a slab. Then
1814 * we reduce the minimum objects required in a slab.
1815 */
1816 min_objects = slub_min_objects;
1817 while (min_objects > 1) {
1818 fraction = 8;
1819 while (fraction >= 4) {
1820 order = slab_order(size, min_objects,
1821 slub_max_order, fraction);
1822 if (order <= slub_max_order)
1823 return order;
1824 fraction /= 2;
1825 }
1826 min_objects /= 2;
1827 }
1828
1829 /*
1830 * We were unable to place multiple objects in a slab. Now
1831 * lets see if we can place a single object there.
1832 */
1833 order = slab_order(size, 1, slub_max_order, 1);
1834 if (order <= slub_max_order)
1835 return order;
1836
1837 /*
1838 * Doh this slab cannot be placed using slub_max_order.
1839 */
1840 order = slab_order(size, 1, MAX_ORDER, 1);
1841 if (order <= MAX_ORDER)
1842 return order;
1843 return -ENOSYS;
1844}
1845
Christoph Lameter81819f02007-05-06 14:49:36 -07001846/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001847 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001848 */
1849static unsigned long calculate_alignment(unsigned long flags,
1850 unsigned long align, unsigned long size)
1851{
1852 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001853 * If the user wants hardware cache aligned objects then follow that
1854 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001855 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001856 * The hardware cache alignment cannot override the specified
1857 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001858 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001859 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001860 size > cache_line_size() / 2)
1861 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001862
1863 if (align < ARCH_SLAB_MINALIGN)
1864 return ARCH_SLAB_MINALIGN;
1865
1866 return ALIGN(align, sizeof(void *));
1867}
1868
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001869static void init_kmem_cache_cpu(struct kmem_cache *s,
1870 struct kmem_cache_cpu *c)
1871{
1872 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001873 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001874 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001875 c->offset = s->offset / sizeof(void *);
1876 c->objsize = s->objsize;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001877}
1878
Christoph Lameter81819f02007-05-06 14:49:36 -07001879static void init_kmem_cache_node(struct kmem_cache_node *n)
1880{
1881 n->nr_partial = 0;
1882 atomic_long_set(&n->nr_slabs, 0);
1883 spin_lock_init(&n->list_lock);
1884 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001885#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07001886 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001887#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001888}
1889
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001890#ifdef CONFIG_SMP
1891/*
1892 * Per cpu array for per cpu structures.
1893 *
1894 * The per cpu array places all kmem_cache_cpu structures from one processor
1895 * close together meaning that it becomes possible that multiple per cpu
1896 * structures are contained in one cacheline. This may be particularly
1897 * beneficial for the kmalloc caches.
1898 *
1899 * A desktop system typically has around 60-80 slabs. With 100 here we are
1900 * likely able to get per cpu structures for all caches from the array defined
1901 * here. We must be able to cover all kmalloc caches during bootstrap.
1902 *
1903 * If the per cpu array is exhausted then fall back to kmalloc
1904 * of individual cachelines. No sharing is possible then.
1905 */
1906#define NR_KMEM_CACHE_CPU 100
1907
1908static DEFINE_PER_CPU(struct kmem_cache_cpu,
1909 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1910
1911static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1912static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1913
1914static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1915 int cpu, gfp_t flags)
1916{
1917 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1918
1919 if (c)
1920 per_cpu(kmem_cache_cpu_free, cpu) =
1921 (void *)c->freelist;
1922 else {
1923 /* Table overflow: So allocate ourselves */
1924 c = kmalloc_node(
1925 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1926 flags, cpu_to_node(cpu));
1927 if (!c)
1928 return NULL;
1929 }
1930
1931 init_kmem_cache_cpu(s, c);
1932 return c;
1933}
1934
1935static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1936{
1937 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1938 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1939 kfree(c);
1940 return;
1941 }
1942 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1943 per_cpu(kmem_cache_cpu_free, cpu) = c;
1944}
1945
1946static void free_kmem_cache_cpus(struct kmem_cache *s)
1947{
1948 int cpu;
1949
1950 for_each_online_cpu(cpu) {
1951 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1952
1953 if (c) {
1954 s->cpu_slab[cpu] = NULL;
1955 free_kmem_cache_cpu(c, cpu);
1956 }
1957 }
1958}
1959
1960static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1961{
1962 int cpu;
1963
1964 for_each_online_cpu(cpu) {
1965 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1966
1967 if (c)
1968 continue;
1969
1970 c = alloc_kmem_cache_cpu(s, cpu, flags);
1971 if (!c) {
1972 free_kmem_cache_cpus(s);
1973 return 0;
1974 }
1975 s->cpu_slab[cpu] = c;
1976 }
1977 return 1;
1978}
1979
1980/*
1981 * Initialize the per cpu array.
1982 */
1983static void init_alloc_cpu_cpu(int cpu)
1984{
1985 int i;
1986
1987 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
1988 return;
1989
1990 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
1991 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
1992
1993 cpu_set(cpu, kmem_cach_cpu_free_init_once);
1994}
1995
1996static void __init init_alloc_cpu(void)
1997{
1998 int cpu;
1999
2000 for_each_online_cpu(cpu)
2001 init_alloc_cpu_cpu(cpu);
2002 }
2003
2004#else
2005static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2006static inline void init_alloc_cpu(void) {}
2007
2008static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2009{
2010 init_kmem_cache_cpu(s, &s->cpu_slab);
2011 return 1;
2012}
2013#endif
2014
Christoph Lameter81819f02007-05-06 14:49:36 -07002015#ifdef CONFIG_NUMA
2016/*
2017 * No kmalloc_node yet so do it by hand. We know that this is the first
2018 * slab on the node for this slabcache. There are no concurrent accesses
2019 * possible.
2020 *
2021 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002022 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2023 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002024 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002025static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2026 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002027{
2028 struct page *page;
2029 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002030 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002031
2032 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2033
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002034 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002035
2036 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002037 if (page_to_nid(page) != node) {
2038 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2039 "node %d\n", node);
2040 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2041 "in order to be able to continue\n");
2042 }
2043
Christoph Lameter81819f02007-05-06 14:49:36 -07002044 n = page->freelist;
2045 BUG_ON(!n);
2046 page->freelist = get_freepointer(kmalloc_caches, n);
2047 page->inuse++;
2048 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002049#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002050 init_object(kmalloc_caches, n, 1);
2051 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002052#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002053 init_kmem_cache_node(n);
2054 atomic_long_inc(&n->nr_slabs);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002055
rootba84c732008-01-07 23:20:28 -08002056 /*
2057 * lockdep requires consistent irq usage for each lock
2058 * so even though there cannot be a race this early in
2059 * the boot sequence, we still disable irqs.
2060 */
2061 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002062 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002063 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002064 return n;
2065}
2066
2067static void free_kmem_cache_nodes(struct kmem_cache *s)
2068{
2069 int node;
2070
Christoph Lameterf64dc582007-10-16 01:25:33 -07002071 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002072 struct kmem_cache_node *n = s->node[node];
2073 if (n && n != &s->local_node)
2074 kmem_cache_free(kmalloc_caches, n);
2075 s->node[node] = NULL;
2076 }
2077}
2078
2079static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2080{
2081 int node;
2082 int local_node;
2083
2084 if (slab_state >= UP)
2085 local_node = page_to_nid(virt_to_page(s));
2086 else
2087 local_node = 0;
2088
Christoph Lameterf64dc582007-10-16 01:25:33 -07002089 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002090 struct kmem_cache_node *n;
2091
2092 if (local_node == node)
2093 n = &s->local_node;
2094 else {
2095 if (slab_state == DOWN) {
2096 n = early_kmem_cache_node_alloc(gfpflags,
2097 node);
2098 continue;
2099 }
2100 n = kmem_cache_alloc_node(kmalloc_caches,
2101 gfpflags, node);
2102
2103 if (!n) {
2104 free_kmem_cache_nodes(s);
2105 return 0;
2106 }
2107
2108 }
2109 s->node[node] = n;
2110 init_kmem_cache_node(n);
2111 }
2112 return 1;
2113}
2114#else
2115static void free_kmem_cache_nodes(struct kmem_cache *s)
2116{
2117}
2118
2119static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2120{
2121 init_kmem_cache_node(&s->local_node);
2122 return 1;
2123}
2124#endif
2125
2126/*
2127 * calculate_sizes() determines the order and the distribution of data within
2128 * a slab object.
2129 */
2130static int calculate_sizes(struct kmem_cache *s)
2131{
2132 unsigned long flags = s->flags;
2133 unsigned long size = s->objsize;
2134 unsigned long align = s->align;
2135
2136 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002137 * Round up object size to the next word boundary. We can only
2138 * place the free pointer at word boundaries and this determines
2139 * the possible location of the free pointer.
2140 */
2141 size = ALIGN(size, sizeof(void *));
2142
2143#ifdef CONFIG_SLUB_DEBUG
2144 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002145 * Determine if we can poison the object itself. If the user of
2146 * the slab may touch the object after free or before allocation
2147 * then we should never poison the object itself.
2148 */
2149 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002150 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002151 s->flags |= __OBJECT_POISON;
2152 else
2153 s->flags &= ~__OBJECT_POISON;
2154
Christoph Lameter81819f02007-05-06 14:49:36 -07002155
2156 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002157 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002158 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002159 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002160 */
2161 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2162 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002163#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002164
2165 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002166 * With that we have determined the number of bytes in actual use
2167 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002168 */
2169 s->inuse = size;
2170
2171 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002172 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002173 /*
2174 * Relocate free pointer after the object if it is not
2175 * permitted to overwrite the first word of the object on
2176 * kmem_cache_free.
2177 *
2178 * This is the case if we do RCU, have a constructor or
2179 * destructor or are poisoning the objects.
2180 */
2181 s->offset = size;
2182 size += sizeof(void *);
2183 }
2184
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002185#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002186 if (flags & SLAB_STORE_USER)
2187 /*
2188 * Need to store information about allocs and frees after
2189 * the object.
2190 */
2191 size += 2 * sizeof(struct track);
2192
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002193 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002194 /*
2195 * Add some empty padding so that we can catch
2196 * overwrites from earlier objects rather than let
2197 * tracking information or the free pointer be
2198 * corrupted if an user writes before the start
2199 * of the object.
2200 */
2201 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002202#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002203
Christoph Lameter81819f02007-05-06 14:49:36 -07002204 /*
2205 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002206 * user specified and the dynamic determination of cache line size
2207 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002208 */
2209 align = calculate_alignment(flags, align, s->objsize);
2210
2211 /*
2212 * SLUB stores one object immediately after another beginning from
2213 * offset 0. In order to align the objects we have to simply size
2214 * each object to conform to the alignment.
2215 */
2216 size = ALIGN(size, align);
2217 s->size = size;
2218
Christoph Lameter71c7a062008-02-14 14:28:01 -08002219 if ((flags & __KMALLOC_CACHE) &&
2220 PAGE_SIZE / size < slub_min_objects) {
2221 /*
2222 * Kmalloc cache that would not have enough objects in
2223 * an order 0 page. Kmalloc slabs can fallback to
2224 * page allocator order 0 allocs so take a reasonably large
2225 * order that will allows us a good number of objects.
2226 */
2227 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2228 s->flags |= __PAGE_ALLOC_FALLBACK;
2229 s->allocflags |= __GFP_NOWARN;
2230 } else
2231 s->order = calculate_order(size);
2232
Christoph Lameter81819f02007-05-06 14:49:36 -07002233 if (s->order < 0)
2234 return 0;
2235
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002236 s->allocflags = 0;
2237 if (s->order)
2238 s->allocflags |= __GFP_COMP;
2239
2240 if (s->flags & SLAB_CACHE_DMA)
2241 s->allocflags |= SLUB_DMA;
2242
2243 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2244 s->allocflags |= __GFP_RECLAIMABLE;
2245
Christoph Lameter81819f02007-05-06 14:49:36 -07002246 /*
2247 * Determine the number of objects per slab
2248 */
2249 s->objects = (PAGE_SIZE << s->order) / size;
2250
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002251 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002252
2253}
2254
Christoph Lameter81819f02007-05-06 14:49:36 -07002255static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2256 const char *name, size_t size,
2257 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002258 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002259{
2260 memset(s, 0, kmem_size);
2261 s->name = name;
2262 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002263 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002264 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002265 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002266
2267 if (!calculate_sizes(s))
2268 goto error;
2269
2270 s->refcount = 1;
2271#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002272 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002273#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002274 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2275 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002276
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002277 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002278 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002279 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002280error:
2281 if (flags & SLAB_PANIC)
2282 panic("Cannot create slab %s size=%lu realsize=%u "
2283 "order=%u offset=%u flags=%lx\n",
2284 s->name, (unsigned long)size, s->size, s->order,
2285 s->offset, flags);
2286 return 0;
2287}
Christoph Lameter81819f02007-05-06 14:49:36 -07002288
2289/*
2290 * Check if a given pointer is valid
2291 */
2292int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2293{
Pekka Enberg06428782008-01-07 23:20:27 -08002294 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002295
2296 page = get_object_page(object);
2297
2298 if (!page || s != page->slab)
2299 /* No slab or wrong slab */
2300 return 0;
2301
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002302 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002303 return 0;
2304
2305 /*
2306 * We could also check if the object is on the slabs freelist.
2307 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002308 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002309 * to a certain slab.
2310 */
2311 return 1;
2312}
2313EXPORT_SYMBOL(kmem_ptr_validate);
2314
2315/*
2316 * Determine the size of a slab object
2317 */
2318unsigned int kmem_cache_size(struct kmem_cache *s)
2319{
2320 return s->objsize;
2321}
2322EXPORT_SYMBOL(kmem_cache_size);
2323
2324const char *kmem_cache_name(struct kmem_cache *s)
2325{
2326 return s->name;
2327}
2328EXPORT_SYMBOL(kmem_cache_name);
2329
2330/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002331 * Attempt to free all slabs on a node. Return the number of slabs we
2332 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002333 */
2334static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2335 struct list_head *list)
2336{
2337 int slabs_inuse = 0;
2338 unsigned long flags;
2339 struct page *page, *h;
2340
2341 spin_lock_irqsave(&n->list_lock, flags);
2342 list_for_each_entry_safe(page, h, list, lru)
2343 if (!page->inuse) {
2344 list_del(&page->lru);
2345 discard_slab(s, page);
2346 } else
2347 slabs_inuse++;
2348 spin_unlock_irqrestore(&n->list_lock, flags);
2349 return slabs_inuse;
2350}
2351
2352/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002353 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002354 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002355static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002356{
2357 int node;
2358
2359 flush_all(s);
2360
2361 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002362 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002363 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002364 struct kmem_cache_node *n = get_node(s, node);
2365
Christoph Lameter2086d262007-05-06 14:49:46 -07002366 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002367 if (atomic_long_read(&n->nr_slabs))
2368 return 1;
2369 }
2370 free_kmem_cache_nodes(s);
2371 return 0;
2372}
2373
2374/*
2375 * Close a cache and release the kmem_cache structure
2376 * (must be used for caches created using kmem_cache_create)
2377 */
2378void kmem_cache_destroy(struct kmem_cache *s)
2379{
2380 down_write(&slub_lock);
2381 s->refcount--;
2382 if (!s->refcount) {
2383 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002384 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002385 if (kmem_cache_close(s))
2386 WARN_ON(1);
2387 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002388 } else
2389 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002390}
2391EXPORT_SYMBOL(kmem_cache_destroy);
2392
2393/********************************************************************
2394 * Kmalloc subsystem
2395 *******************************************************************/
2396
Christoph Lameter331dc552008-02-14 14:28:09 -08002397struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002398EXPORT_SYMBOL(kmalloc_caches);
2399
2400#ifdef CONFIG_ZONE_DMA
Christoph Lameter331dc552008-02-14 14:28:09 -08002401static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter81819f02007-05-06 14:49:36 -07002402#endif
2403
2404static int __init setup_slub_min_order(char *str)
2405{
Pekka Enberg06428782008-01-07 23:20:27 -08002406 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002407
2408 return 1;
2409}
2410
2411__setup("slub_min_order=", setup_slub_min_order);
2412
2413static int __init setup_slub_max_order(char *str)
2414{
Pekka Enberg06428782008-01-07 23:20:27 -08002415 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002416
2417 return 1;
2418}
2419
2420__setup("slub_max_order=", setup_slub_max_order);
2421
2422static int __init setup_slub_min_objects(char *str)
2423{
Pekka Enberg06428782008-01-07 23:20:27 -08002424 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002425
2426 return 1;
2427}
2428
2429__setup("slub_min_objects=", setup_slub_min_objects);
2430
2431static int __init setup_slub_nomerge(char *str)
2432{
2433 slub_nomerge = 1;
2434 return 1;
2435}
2436
2437__setup("slub_nomerge", setup_slub_nomerge);
2438
Christoph Lameter81819f02007-05-06 14:49:36 -07002439static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2440 const char *name, int size, gfp_t gfp_flags)
2441{
2442 unsigned int flags = 0;
2443
2444 if (gfp_flags & SLUB_DMA)
2445 flags = SLAB_CACHE_DMA;
2446
2447 down_write(&slub_lock);
2448 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter71c7a062008-02-14 14:28:01 -08002449 flags | __KMALLOC_CACHE, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002450 goto panic;
2451
2452 list_add(&s->list, &slab_caches);
2453 up_write(&slub_lock);
2454 if (sysfs_slab_add(s))
2455 goto panic;
2456 return s;
2457
2458panic:
2459 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2460}
2461
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002462#ifdef CONFIG_ZONE_DMA
Christoph Lameter1ceef402007-08-07 15:11:48 -07002463
2464static void sysfs_add_func(struct work_struct *w)
2465{
2466 struct kmem_cache *s;
2467
2468 down_write(&slub_lock);
2469 list_for_each_entry(s, &slab_caches, list) {
2470 if (s->flags & __SYSFS_ADD_DEFERRED) {
2471 s->flags &= ~__SYSFS_ADD_DEFERRED;
2472 sysfs_slab_add(s);
2473 }
2474 }
2475 up_write(&slub_lock);
2476}
2477
2478static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2479
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002480static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2481{
2482 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002483 char *text;
2484 size_t realsize;
2485
2486 s = kmalloc_caches_dma[index];
2487 if (s)
2488 return s;
2489
2490 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002491 if (flags & __GFP_WAIT)
2492 down_write(&slub_lock);
2493 else {
2494 if (!down_write_trylock(&slub_lock))
2495 goto out;
2496 }
2497
2498 if (kmalloc_caches_dma[index])
2499 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002500
Christoph Lameter7b55f622007-07-17 04:03:27 -07002501 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002502 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2503 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002504 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2505
2506 if (!s || !text || !kmem_cache_open(s, flags, text,
2507 realsize, ARCH_KMALLOC_MINALIGN,
2508 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2509 kfree(s);
2510 kfree(text);
2511 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002512 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002513
2514 list_add(&s->list, &slab_caches);
2515 kmalloc_caches_dma[index] = s;
2516
2517 schedule_work(&sysfs_add_work);
2518
2519unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002520 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002521out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002522 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002523}
2524#endif
2525
Christoph Lameterf1b26332007-07-17 04:03:26 -07002526/*
2527 * Conversion table for small slabs sizes / 8 to the index in the
2528 * kmalloc array. This is necessary for slabs < 192 since we have non power
2529 * of two cache sizes there. The size of larger slabs can be determined using
2530 * fls.
2531 */
2532static s8 size_index[24] = {
2533 3, /* 8 */
2534 4, /* 16 */
2535 5, /* 24 */
2536 5, /* 32 */
2537 6, /* 40 */
2538 6, /* 48 */
2539 6, /* 56 */
2540 6, /* 64 */
2541 1, /* 72 */
2542 1, /* 80 */
2543 1, /* 88 */
2544 1, /* 96 */
2545 7, /* 104 */
2546 7, /* 112 */
2547 7, /* 120 */
2548 7, /* 128 */
2549 2, /* 136 */
2550 2, /* 144 */
2551 2, /* 152 */
2552 2, /* 160 */
2553 2, /* 168 */
2554 2, /* 176 */
2555 2, /* 184 */
2556 2 /* 192 */
2557};
2558
Christoph Lameter81819f02007-05-06 14:49:36 -07002559static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2560{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002561 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002562
Christoph Lameterf1b26332007-07-17 04:03:26 -07002563 if (size <= 192) {
2564 if (!size)
2565 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002566
Christoph Lameterf1b26332007-07-17 04:03:26 -07002567 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002568 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002569 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002570
2571#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002572 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002573 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002574
Christoph Lameter81819f02007-05-06 14:49:36 -07002575#endif
2576 return &kmalloc_caches[index];
2577}
2578
2579void *__kmalloc(size_t size, gfp_t flags)
2580{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002581 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002582
Christoph Lameter331dc552008-02-14 14:28:09 -08002583 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002584 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002585
2586 s = get_slab(size, flags);
2587
2588 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002589 return s;
2590
Christoph Lameterce15fea2007-07-17 04:03:28 -07002591 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002592}
2593EXPORT_SYMBOL(__kmalloc);
2594
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002595static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2596{
2597 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2598 get_order(size));
2599
2600 if (page)
2601 return page_address(page);
2602 else
2603 return NULL;
2604}
2605
Christoph Lameter81819f02007-05-06 14:49:36 -07002606#ifdef CONFIG_NUMA
2607void *__kmalloc_node(size_t size, gfp_t flags, int node)
2608{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002609 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002610
Christoph Lameter331dc552008-02-14 14:28:09 -08002611 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002612 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002613
2614 s = get_slab(size, flags);
2615
2616 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002617 return s;
2618
Christoph Lameterce15fea2007-07-17 04:03:28 -07002619 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002620}
2621EXPORT_SYMBOL(__kmalloc_node);
2622#endif
2623
2624size_t ksize(const void *object)
2625{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002626 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002627 struct kmem_cache *s;
2628
Christoph Lameteref8b4522007-10-16 01:24:46 -07002629 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002630 return 0;
2631
Vegard Nossum294a80a2007-12-04 23:45:30 -08002632 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002633
2634 if (unlikely(!PageSlab(page)))
2635 return PAGE_SIZE << compound_order(page);
2636
Christoph Lameter81819f02007-05-06 14:49:36 -07002637 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002638
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002639#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002640 /*
2641 * Debugging requires use of the padding between object
2642 * and whatever may come after it.
2643 */
2644 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2645 return s->objsize;
2646
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002647#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002648 /*
2649 * If we have the need to store the freelist pointer
2650 * back there or track user information then we can
2651 * only use the space before that information.
2652 */
2653 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2654 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002655 /*
2656 * Else we can use all the padding etc for the allocation
2657 */
2658 return s->size;
2659}
2660EXPORT_SYMBOL(ksize);
2661
2662void kfree(const void *x)
2663{
Christoph Lameter81819f02007-05-06 14:49:36 -07002664 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002665 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002666
Satyam Sharma2408c552007-10-16 01:24:44 -07002667 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002668 return;
2669
Christoph Lameterb49af682007-05-06 14:49:41 -07002670 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002671 if (unlikely(!PageSlab(page))) {
2672 put_page(page);
2673 return;
2674 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002675 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002676}
2677EXPORT_SYMBOL(kfree);
2678
Christoph Lameterf61396a2008-01-07 23:20:26 -08002679static unsigned long count_partial(struct kmem_cache_node *n)
2680{
2681 unsigned long flags;
2682 unsigned long x = 0;
2683 struct page *page;
2684
2685 spin_lock_irqsave(&n->list_lock, flags);
2686 list_for_each_entry(page, &n->partial, lru)
2687 x += page->inuse;
2688 spin_unlock_irqrestore(&n->list_lock, flags);
2689 return x;
2690}
2691
Christoph Lameter2086d262007-05-06 14:49:46 -07002692/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002693 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2694 * the remaining slabs by the number of items in use. The slabs with the
2695 * most items in use come first. New allocations will then fill those up
2696 * and thus they can be removed from the partial lists.
2697 *
2698 * The slabs with the least items are placed last. This results in them
2699 * being allocated from last increasing the chance that the last objects
2700 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002701 */
2702int kmem_cache_shrink(struct kmem_cache *s)
2703{
2704 int node;
2705 int i;
2706 struct kmem_cache_node *n;
2707 struct page *page;
2708 struct page *t;
2709 struct list_head *slabs_by_inuse =
2710 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2711 unsigned long flags;
2712
2713 if (!slabs_by_inuse)
2714 return -ENOMEM;
2715
2716 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002717 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002718 n = get_node(s, node);
2719
2720 if (!n->nr_partial)
2721 continue;
2722
2723 for (i = 0; i < s->objects; i++)
2724 INIT_LIST_HEAD(slabs_by_inuse + i);
2725
2726 spin_lock_irqsave(&n->list_lock, flags);
2727
2728 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002729 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002730 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002731 * Note that concurrent frees may occur while we hold the
2732 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002733 */
2734 list_for_each_entry_safe(page, t, &n->partial, lru) {
2735 if (!page->inuse && slab_trylock(page)) {
2736 /*
2737 * Must hold slab lock here because slab_free
2738 * may have freed the last object and be
2739 * waiting to release the slab.
2740 */
2741 list_del(&page->lru);
2742 n->nr_partial--;
2743 slab_unlock(page);
2744 discard_slab(s, page);
2745 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002746 list_move(&page->lru,
2747 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002748 }
2749 }
2750
Christoph Lameter2086d262007-05-06 14:49:46 -07002751 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002752 * Rebuild the partial list with the slabs filled up most
2753 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002754 */
2755 for (i = s->objects - 1; i >= 0; i--)
2756 list_splice(slabs_by_inuse + i, n->partial.prev);
2757
Christoph Lameter2086d262007-05-06 14:49:46 -07002758 spin_unlock_irqrestore(&n->list_lock, flags);
2759 }
2760
2761 kfree(slabs_by_inuse);
2762 return 0;
2763}
2764EXPORT_SYMBOL(kmem_cache_shrink);
2765
Yasunori Gotob9049e22007-10-21 16:41:37 -07002766#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2767static int slab_mem_going_offline_callback(void *arg)
2768{
2769 struct kmem_cache *s;
2770
2771 down_read(&slub_lock);
2772 list_for_each_entry(s, &slab_caches, list)
2773 kmem_cache_shrink(s);
2774 up_read(&slub_lock);
2775
2776 return 0;
2777}
2778
2779static void slab_mem_offline_callback(void *arg)
2780{
2781 struct kmem_cache_node *n;
2782 struct kmem_cache *s;
2783 struct memory_notify *marg = arg;
2784 int offline_node;
2785
2786 offline_node = marg->status_change_nid;
2787
2788 /*
2789 * If the node still has available memory. we need kmem_cache_node
2790 * for it yet.
2791 */
2792 if (offline_node < 0)
2793 return;
2794
2795 down_read(&slub_lock);
2796 list_for_each_entry(s, &slab_caches, list) {
2797 n = get_node(s, offline_node);
2798 if (n) {
2799 /*
2800 * if n->nr_slabs > 0, slabs still exist on the node
2801 * that is going down. We were unable to free them,
2802 * and offline_pages() function shoudn't call this
2803 * callback. So, we must fail.
2804 */
Al Viro27bb6282007-10-29 04:42:55 +00002805 BUG_ON(atomic_long_read(&n->nr_slabs));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002806
2807 s->node[offline_node] = NULL;
2808 kmem_cache_free(kmalloc_caches, n);
2809 }
2810 }
2811 up_read(&slub_lock);
2812}
2813
2814static int slab_mem_going_online_callback(void *arg)
2815{
2816 struct kmem_cache_node *n;
2817 struct kmem_cache *s;
2818 struct memory_notify *marg = arg;
2819 int nid = marg->status_change_nid;
2820 int ret = 0;
2821
2822 /*
2823 * If the node's memory is already available, then kmem_cache_node is
2824 * already created. Nothing to do.
2825 */
2826 if (nid < 0)
2827 return 0;
2828
2829 /*
2830 * We are bringing a node online. No memory is availabe yet. We must
2831 * allocate a kmem_cache_node structure in order to bring the node
2832 * online.
2833 */
2834 down_read(&slub_lock);
2835 list_for_each_entry(s, &slab_caches, list) {
2836 /*
2837 * XXX: kmem_cache_alloc_node will fallback to other nodes
2838 * since memory is not yet available from the node that
2839 * is brought up.
2840 */
2841 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2842 if (!n) {
2843 ret = -ENOMEM;
2844 goto out;
2845 }
2846 init_kmem_cache_node(n);
2847 s->node[nid] = n;
2848 }
2849out:
2850 up_read(&slub_lock);
2851 return ret;
2852}
2853
2854static int slab_memory_callback(struct notifier_block *self,
2855 unsigned long action, void *arg)
2856{
2857 int ret = 0;
2858
2859 switch (action) {
2860 case MEM_GOING_ONLINE:
2861 ret = slab_mem_going_online_callback(arg);
2862 break;
2863 case MEM_GOING_OFFLINE:
2864 ret = slab_mem_going_offline_callback(arg);
2865 break;
2866 case MEM_OFFLINE:
2867 case MEM_CANCEL_ONLINE:
2868 slab_mem_offline_callback(arg);
2869 break;
2870 case MEM_ONLINE:
2871 case MEM_CANCEL_OFFLINE:
2872 break;
2873 }
2874
2875 ret = notifier_from_errno(ret);
2876 return ret;
2877}
2878
2879#endif /* CONFIG_MEMORY_HOTPLUG */
2880
Christoph Lameter81819f02007-05-06 14:49:36 -07002881/********************************************************************
2882 * Basic setup of slabs
2883 *******************************************************************/
2884
2885void __init kmem_cache_init(void)
2886{
2887 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002888 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002889
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002890 init_alloc_cpu();
2891
Christoph Lameter81819f02007-05-06 14:49:36 -07002892#ifdef CONFIG_NUMA
2893 /*
2894 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002895 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002896 * kmem_cache_open for slab_state == DOWN.
2897 */
2898 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2899 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002900 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002901 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002902
2903 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002904#endif
2905
2906 /* Able to allocate the per node structures */
2907 slab_state = PARTIAL;
2908
2909 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002910 if (KMALLOC_MIN_SIZE <= 64) {
2911 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002912 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002913 caches++;
2914 }
2915 if (KMALLOC_MIN_SIZE <= 128) {
2916 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002917 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002918 caches++;
2919 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002920
Christoph Lameter331dc552008-02-14 14:28:09 -08002921 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002922 create_kmalloc_cache(&kmalloc_caches[i],
2923 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002924 caches++;
2925 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002926
Christoph Lameterf1b26332007-07-17 04:03:26 -07002927
2928 /*
2929 * Patch up the size_index table if we have strange large alignment
2930 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002931 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002932 *
2933 * Largest permitted alignment is 256 bytes due to the way we
2934 * handle the index determination for the smaller caches.
2935 *
2936 * Make sure that nothing crazy happens if someone starts tinkering
2937 * around with ARCH_KMALLOC_MINALIGN
2938 */
2939 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2940 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2941
Christoph Lameter12ad6842007-07-17 04:03:28 -07002942 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002943 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2944
Christoph Lameter81819f02007-05-06 14:49:36 -07002945 slab_state = UP;
2946
2947 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08002948 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002949 kmalloc_caches[i]. name =
2950 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2951
2952#ifdef CONFIG_SMP
2953 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002954 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2955 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2956#else
2957 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002958#endif
2959
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002960 printk(KERN_INFO
2961 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002962 " CPUs=%d, Nodes=%d\n",
2963 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002964 slub_min_order, slub_max_order, slub_min_objects,
2965 nr_cpu_ids, nr_node_ids);
2966}
2967
2968/*
2969 * Find a mergeable slab cache
2970 */
2971static int slab_unmergeable(struct kmem_cache *s)
2972{
2973 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2974 return 1;
2975
Christoph Lameter331dc552008-02-14 14:28:09 -08002976 if ((s->flags & __PAGE_ALLOC_FALLBACK))
Christoph Lameter71c7a062008-02-14 14:28:01 -08002977 return 1;
2978
Christoph Lameterc59def92007-05-16 22:10:50 -07002979 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002980 return 1;
2981
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002982 /*
2983 * We may have set a slab to be unmergeable during bootstrap.
2984 */
2985 if (s->refcount < 0)
2986 return 1;
2987
Christoph Lameter81819f02007-05-06 14:49:36 -07002988 return 0;
2989}
2990
2991static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07002992 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002993 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002994{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002995 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002996
2997 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2998 return NULL;
2999
Christoph Lameterc59def92007-05-16 22:10:50 -07003000 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003001 return NULL;
3002
3003 size = ALIGN(size, sizeof(void *));
3004 align = calculate_alignment(flags, align, size);
3005 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003006 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003007
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003008 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003009 if (slab_unmergeable(s))
3010 continue;
3011
3012 if (size > s->size)
3013 continue;
3014
Christoph Lameterba0268a2007-09-11 15:24:11 -07003015 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003016 continue;
3017 /*
3018 * Check if alignment is compatible.
3019 * Courtesy of Adrian Drzewiecki
3020 */
Pekka Enberg06428782008-01-07 23:20:27 -08003021 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003022 continue;
3023
3024 if (s->size - size >= sizeof(void *))
3025 continue;
3026
3027 return s;
3028 }
3029 return NULL;
3030}
3031
3032struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3033 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003034 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003035{
3036 struct kmem_cache *s;
3037
3038 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003039 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003040 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003041 int cpu;
3042
Christoph Lameter81819f02007-05-06 14:49:36 -07003043 s->refcount++;
3044 /*
3045 * Adjust the object sizes so that we clear
3046 * the complete object on kzalloc.
3047 */
3048 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003049
3050 /*
3051 * And then we need to update the object size in the
3052 * per cpu structures
3053 */
3054 for_each_online_cpu(cpu)
3055 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003056
Christoph Lameter81819f02007-05-06 14:49:36 -07003057 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003058 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003059
Christoph Lameter81819f02007-05-06 14:49:36 -07003060 if (sysfs_slab_alias(s, name))
3061 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003062 return s;
3063 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003064
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003065 s = kmalloc(kmem_size, GFP_KERNEL);
3066 if (s) {
3067 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003068 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003069 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003070 up_write(&slub_lock);
3071 if (sysfs_slab_add(s))
3072 goto err;
3073 return s;
3074 }
3075 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003076 }
3077 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003078
3079err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003080 if (flags & SLAB_PANIC)
3081 panic("Cannot create slabcache %s\n", name);
3082 else
3083 s = NULL;
3084 return s;
3085}
3086EXPORT_SYMBOL(kmem_cache_create);
3087
Christoph Lameter81819f02007-05-06 14:49:36 -07003088#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003089/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003090 * Use the cpu notifier to insure that the cpu slabs are flushed when
3091 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003092 */
3093static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3094 unsigned long action, void *hcpu)
3095{
3096 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003097 struct kmem_cache *s;
3098 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003099
3100 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003101 case CPU_UP_PREPARE:
3102 case CPU_UP_PREPARE_FROZEN:
3103 init_alloc_cpu_cpu(cpu);
3104 down_read(&slub_lock);
3105 list_for_each_entry(s, &slab_caches, list)
3106 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3107 GFP_KERNEL);
3108 up_read(&slub_lock);
3109 break;
3110
Christoph Lameter81819f02007-05-06 14:49:36 -07003111 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003112 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003113 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003114 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003115 down_read(&slub_lock);
3116 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003117 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3118
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003119 local_irq_save(flags);
3120 __flush_cpu_slab(s, cpu);
3121 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003122 free_kmem_cache_cpu(c, cpu);
3123 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003124 }
3125 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003126 break;
3127 default:
3128 break;
3129 }
3130 return NOTIFY_OK;
3131}
3132
Pekka Enberg06428782008-01-07 23:20:27 -08003133static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003134 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003135};
Christoph Lameter81819f02007-05-06 14:49:36 -07003136
3137#endif
3138
Christoph Lameter81819f02007-05-06 14:49:36 -07003139void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3140{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003141 struct kmem_cache *s;
3142
Christoph Lameter331dc552008-02-14 14:28:09 -08003143 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003144 return kmalloc_large(size, gfpflags);
3145
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003146 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003147
Satyam Sharma2408c552007-10-16 01:24:44 -07003148 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003149 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003150
Christoph Lameterce15fea2007-07-17 04:03:28 -07003151 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003152}
3153
3154void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3155 int node, void *caller)
3156{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003157 struct kmem_cache *s;
3158
Christoph Lameter331dc552008-02-14 14:28:09 -08003159 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003160 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003161
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003162 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003163
Satyam Sharma2408c552007-10-16 01:24:44 -07003164 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003165 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003166
Christoph Lameterce15fea2007-07-17 04:03:28 -07003167 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003168}
3169
Christoph Lameter41ecc552007-05-09 02:32:44 -07003170#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003171static int validate_slab(struct kmem_cache *s, struct page *page,
3172 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003173{
3174 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003175 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003176
3177 if (!check_slab(s, page) ||
3178 !on_freelist(s, page, NULL))
3179 return 0;
3180
3181 /* Now we know that a valid freelist exists */
3182 bitmap_zero(map, s->objects);
3183
Christoph Lameter7656c722007-05-09 02:32:40 -07003184 for_each_free_object(p, s, page->freelist) {
3185 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003186 if (!check_object(s, page, p, 0))
3187 return 0;
3188 }
3189
Christoph Lameter7656c722007-05-09 02:32:40 -07003190 for_each_object(p, s, addr)
3191 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003192 if (!check_object(s, page, p, 1))
3193 return 0;
3194 return 1;
3195}
3196
Christoph Lameter434e2452007-07-17 04:03:30 -07003197static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3198 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003199{
3200 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003201 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003202 slab_unlock(page);
3203 } else
3204 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3205 s->name, page);
3206
3207 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003208 if (!SlabDebug(page))
3209 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003210 "on slab 0x%p\n", s->name, page);
3211 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003212 if (SlabDebug(page))
3213 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003214 "slab 0x%p\n", s->name, page);
3215 }
3216}
3217
Christoph Lameter434e2452007-07-17 04:03:30 -07003218static int validate_slab_node(struct kmem_cache *s,
3219 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003220{
3221 unsigned long count = 0;
3222 struct page *page;
3223 unsigned long flags;
3224
3225 spin_lock_irqsave(&n->list_lock, flags);
3226
3227 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003228 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003229 count++;
3230 }
3231 if (count != n->nr_partial)
3232 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3233 "counter=%ld\n", s->name, count, n->nr_partial);
3234
3235 if (!(s->flags & SLAB_STORE_USER))
3236 goto out;
3237
3238 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003239 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003240 count++;
3241 }
3242 if (count != atomic_long_read(&n->nr_slabs))
3243 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3244 "counter=%ld\n", s->name, count,
3245 atomic_long_read(&n->nr_slabs));
3246
3247out:
3248 spin_unlock_irqrestore(&n->list_lock, flags);
3249 return count;
3250}
3251
Christoph Lameter434e2452007-07-17 04:03:30 -07003252static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003253{
3254 int node;
3255 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003256 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3257 sizeof(unsigned long), GFP_KERNEL);
3258
3259 if (!map)
3260 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003261
3262 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003263 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003264 struct kmem_cache_node *n = get_node(s, node);
3265
Christoph Lameter434e2452007-07-17 04:03:30 -07003266 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003267 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003268 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003269 return count;
3270}
3271
Christoph Lameterb3459702007-05-09 02:32:41 -07003272#ifdef SLUB_RESILIENCY_TEST
3273static void resiliency_test(void)
3274{
3275 u8 *p;
3276
3277 printk(KERN_ERR "SLUB resiliency testing\n");
3278 printk(KERN_ERR "-----------------------\n");
3279 printk(KERN_ERR "A. Corruption after allocation\n");
3280
3281 p = kzalloc(16, GFP_KERNEL);
3282 p[16] = 0x12;
3283 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3284 " 0x12->0x%p\n\n", p + 16);
3285
3286 validate_slab_cache(kmalloc_caches + 4);
3287
3288 /* Hmmm... The next two are dangerous */
3289 p = kzalloc(32, GFP_KERNEL);
3290 p[32 + sizeof(void *)] = 0x34;
3291 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003292 " 0x34 -> -0x%p\n", p);
3293 printk(KERN_ERR
3294 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003295
3296 validate_slab_cache(kmalloc_caches + 5);
3297 p = kzalloc(64, GFP_KERNEL);
3298 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3299 *p = 0x56;
3300 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3301 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003302 printk(KERN_ERR
3303 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003304 validate_slab_cache(kmalloc_caches + 6);
3305
3306 printk(KERN_ERR "\nB. Corruption after free\n");
3307 p = kzalloc(128, GFP_KERNEL);
3308 kfree(p);
3309 *p = 0x78;
3310 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3311 validate_slab_cache(kmalloc_caches + 7);
3312
3313 p = kzalloc(256, GFP_KERNEL);
3314 kfree(p);
3315 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003316 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3317 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003318 validate_slab_cache(kmalloc_caches + 8);
3319
3320 p = kzalloc(512, GFP_KERNEL);
3321 kfree(p);
3322 p[512] = 0xab;
3323 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3324 validate_slab_cache(kmalloc_caches + 9);
3325}
3326#else
3327static void resiliency_test(void) {};
3328#endif
3329
Christoph Lameter88a420e2007-05-06 14:49:45 -07003330/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003331 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003332 * and freed.
3333 */
3334
3335struct location {
3336 unsigned long count;
3337 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003338 long long sum_time;
3339 long min_time;
3340 long max_time;
3341 long min_pid;
3342 long max_pid;
3343 cpumask_t cpus;
3344 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003345};
3346
3347struct loc_track {
3348 unsigned long max;
3349 unsigned long count;
3350 struct location *loc;
3351};
3352
3353static void free_loc_track(struct loc_track *t)
3354{
3355 if (t->max)
3356 free_pages((unsigned long)t->loc,
3357 get_order(sizeof(struct location) * t->max));
3358}
3359
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003360static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003361{
3362 struct location *l;
3363 int order;
3364
Christoph Lameter88a420e2007-05-06 14:49:45 -07003365 order = get_order(sizeof(struct location) * max);
3366
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003367 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003368 if (!l)
3369 return 0;
3370
3371 if (t->count) {
3372 memcpy(l, t->loc, sizeof(struct location) * t->count);
3373 free_loc_track(t);
3374 }
3375 t->max = max;
3376 t->loc = l;
3377 return 1;
3378}
3379
3380static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003381 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003382{
3383 long start, end, pos;
3384 struct location *l;
3385 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003386 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003387
3388 start = -1;
3389 end = t->count;
3390
3391 for ( ; ; ) {
3392 pos = start + (end - start + 1) / 2;
3393
3394 /*
3395 * There is nothing at "end". If we end up there
3396 * we need to add something to before end.
3397 */
3398 if (pos == end)
3399 break;
3400
3401 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003402 if (track->addr == caddr) {
3403
3404 l = &t->loc[pos];
3405 l->count++;
3406 if (track->when) {
3407 l->sum_time += age;
3408 if (age < l->min_time)
3409 l->min_time = age;
3410 if (age > l->max_time)
3411 l->max_time = age;
3412
3413 if (track->pid < l->min_pid)
3414 l->min_pid = track->pid;
3415 if (track->pid > l->max_pid)
3416 l->max_pid = track->pid;
3417
3418 cpu_set(track->cpu, l->cpus);
3419 }
3420 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003421 return 1;
3422 }
3423
Christoph Lameter45edfa52007-05-09 02:32:45 -07003424 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003425 end = pos;
3426 else
3427 start = pos;
3428 }
3429
3430 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003431 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003432 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003433 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003434 return 0;
3435
3436 l = t->loc + pos;
3437 if (pos < t->count)
3438 memmove(l + 1, l,
3439 (t->count - pos) * sizeof(struct location));
3440 t->count++;
3441 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003442 l->addr = track->addr;
3443 l->sum_time = age;
3444 l->min_time = age;
3445 l->max_time = age;
3446 l->min_pid = track->pid;
3447 l->max_pid = track->pid;
3448 cpus_clear(l->cpus);
3449 cpu_set(track->cpu, l->cpus);
3450 nodes_clear(l->nodes);
3451 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003452 return 1;
3453}
3454
3455static void process_slab(struct loc_track *t, struct kmem_cache *s,
3456 struct page *page, enum track_item alloc)
3457{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003458 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003459 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003460 void *p;
3461
3462 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003463 for_each_free_object(p, s, page->freelist)
3464 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003465
Christoph Lameter7656c722007-05-09 02:32:40 -07003466 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003467 if (!test_bit(slab_index(p, s, addr), map))
3468 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003469}
3470
3471static int list_locations(struct kmem_cache *s, char *buf,
3472 enum track_item alloc)
3473{
Harvey Harrisone374d482008-01-31 15:20:50 -08003474 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003475 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003476 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003477 int node;
3478
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003479 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003480 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003481 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003482
3483 /* Push back cpu slabs */
3484 flush_all(s);
3485
Christoph Lameterf64dc582007-10-16 01:25:33 -07003486 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003487 struct kmem_cache_node *n = get_node(s, node);
3488 unsigned long flags;
3489 struct page *page;
3490
Christoph Lameter9e869432007-08-22 14:01:56 -07003491 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003492 continue;
3493
3494 spin_lock_irqsave(&n->list_lock, flags);
3495 list_for_each_entry(page, &n->partial, lru)
3496 process_slab(&t, s, page, alloc);
3497 list_for_each_entry(page, &n->full, lru)
3498 process_slab(&t, s, page, alloc);
3499 spin_unlock_irqrestore(&n->list_lock, flags);
3500 }
3501
3502 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003503 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003504
Harvey Harrisone374d482008-01-31 15:20:50 -08003505 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003506 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003507 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003508
3509 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003510 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003511 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003512 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003513
3514 if (l->sum_time != l->min_time) {
3515 unsigned long remainder;
3516
Harvey Harrisone374d482008-01-31 15:20:50 -08003517 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003518 l->min_time,
3519 div_long_long_rem(l->sum_time, l->count, &remainder),
3520 l->max_time);
3521 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003522 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003523 l->min_time);
3524
3525 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003526 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003527 l->min_pid, l->max_pid);
3528 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003529 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003530 l->min_pid);
3531
Christoph Lameter84966342007-06-23 17:16:32 -07003532 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003533 len < PAGE_SIZE - 60) {
3534 len += sprintf(buf + len, " cpus=");
3535 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003536 l->cpus);
3537 }
3538
Christoph Lameter84966342007-06-23 17:16:32 -07003539 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003540 len < PAGE_SIZE - 60) {
3541 len += sprintf(buf + len, " nodes=");
3542 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003543 l->nodes);
3544 }
3545
Harvey Harrisone374d482008-01-31 15:20:50 -08003546 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003547 }
3548
3549 free_loc_track(&t);
3550 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003551 len += sprintf(buf, "No data\n");
3552 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003553}
3554
Christoph Lameter81819f02007-05-06 14:49:36 -07003555enum slab_stat_type {
3556 SL_FULL,
3557 SL_PARTIAL,
3558 SL_CPU,
3559 SL_OBJECTS
3560};
3561
3562#define SO_FULL (1 << SL_FULL)
3563#define SO_PARTIAL (1 << SL_PARTIAL)
3564#define SO_CPU (1 << SL_CPU)
3565#define SO_OBJECTS (1 << SL_OBJECTS)
3566
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003567static ssize_t show_slab_objects(struct kmem_cache *s,
3568 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003569{
3570 unsigned long total = 0;
3571 int cpu;
3572 int node;
3573 int x;
3574 unsigned long *nodes;
3575 unsigned long *per_cpu;
3576
3577 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003578 if (!nodes)
3579 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003580 per_cpu = nodes + nr_node_ids;
3581
3582 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003583 struct page *page;
3584 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003585
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003586 if (!c)
3587 continue;
3588
3589 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003590 node = c->node;
3591 if (node < 0)
3592 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003593 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003594 if (flags & SO_CPU) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003595 if (flags & SO_OBJECTS)
3596 x = page->inuse;
3597 else
3598 x = 1;
3599 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003600 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003601 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003602 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003603 }
3604 }
3605
Christoph Lameterf64dc582007-10-16 01:25:33 -07003606 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003607 struct kmem_cache_node *n = get_node(s, node);
3608
3609 if (flags & SO_PARTIAL) {
3610 if (flags & SO_OBJECTS)
3611 x = count_partial(n);
3612 else
3613 x = n->nr_partial;
3614 total += x;
3615 nodes[node] += x;
3616 }
3617
3618 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003619 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003620 - per_cpu[node]
3621 - n->nr_partial;
3622
3623 if (flags & SO_OBJECTS)
3624 x = full_slabs * s->objects;
3625 else
3626 x = full_slabs;
3627 total += x;
3628 nodes[node] += x;
3629 }
3630 }
3631
3632 x = sprintf(buf, "%lu", total);
3633#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003634 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003635 if (nodes[node])
3636 x += sprintf(buf + x, " N%d=%lu",
3637 node, nodes[node]);
3638#endif
3639 kfree(nodes);
3640 return x + sprintf(buf + x, "\n");
3641}
3642
3643static int any_slab_objects(struct kmem_cache *s)
3644{
3645 int node;
3646 int cpu;
3647
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003648 for_each_possible_cpu(cpu) {
3649 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003650
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003651 if (c && c->page)
3652 return 1;
3653 }
3654
3655 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003656 struct kmem_cache_node *n = get_node(s, node);
3657
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003658 if (!n)
3659 continue;
3660
Christoph Lameter9e869432007-08-22 14:01:56 -07003661 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003662 return 1;
3663 }
3664 return 0;
3665}
3666
3667#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3668#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3669
3670struct slab_attribute {
3671 struct attribute attr;
3672 ssize_t (*show)(struct kmem_cache *s, char *buf);
3673 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3674};
3675
3676#define SLAB_ATTR_RO(_name) \
3677 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3678
3679#define SLAB_ATTR(_name) \
3680 static struct slab_attribute _name##_attr = \
3681 __ATTR(_name, 0644, _name##_show, _name##_store)
3682
Christoph Lameter81819f02007-05-06 14:49:36 -07003683static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3684{
3685 return sprintf(buf, "%d\n", s->size);
3686}
3687SLAB_ATTR_RO(slab_size);
3688
3689static ssize_t align_show(struct kmem_cache *s, char *buf)
3690{
3691 return sprintf(buf, "%d\n", s->align);
3692}
3693SLAB_ATTR_RO(align);
3694
3695static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3696{
3697 return sprintf(buf, "%d\n", s->objsize);
3698}
3699SLAB_ATTR_RO(object_size);
3700
3701static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3702{
3703 return sprintf(buf, "%d\n", s->objects);
3704}
3705SLAB_ATTR_RO(objs_per_slab);
3706
3707static ssize_t order_show(struct kmem_cache *s, char *buf)
3708{
3709 return sprintf(buf, "%d\n", s->order);
3710}
3711SLAB_ATTR_RO(order);
3712
3713static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3714{
3715 if (s->ctor) {
3716 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3717
3718 return n + sprintf(buf + n, "\n");
3719 }
3720 return 0;
3721}
3722SLAB_ATTR_RO(ctor);
3723
Christoph Lameter81819f02007-05-06 14:49:36 -07003724static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3725{
3726 return sprintf(buf, "%d\n", s->refcount - 1);
3727}
3728SLAB_ATTR_RO(aliases);
3729
3730static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3731{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003732 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003733}
3734SLAB_ATTR_RO(slabs);
3735
3736static ssize_t partial_show(struct kmem_cache *s, char *buf)
3737{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003738 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003739}
3740SLAB_ATTR_RO(partial);
3741
3742static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3743{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003744 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003745}
3746SLAB_ATTR_RO(cpu_slabs);
3747
3748static ssize_t objects_show(struct kmem_cache *s, char *buf)
3749{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003750 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003751}
3752SLAB_ATTR_RO(objects);
3753
3754static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3755{
3756 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3757}
3758
3759static ssize_t sanity_checks_store(struct kmem_cache *s,
3760 const char *buf, size_t length)
3761{
3762 s->flags &= ~SLAB_DEBUG_FREE;
3763 if (buf[0] == '1')
3764 s->flags |= SLAB_DEBUG_FREE;
3765 return length;
3766}
3767SLAB_ATTR(sanity_checks);
3768
3769static ssize_t trace_show(struct kmem_cache *s, char *buf)
3770{
3771 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3772}
3773
3774static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3775 size_t length)
3776{
3777 s->flags &= ~SLAB_TRACE;
3778 if (buf[0] == '1')
3779 s->flags |= SLAB_TRACE;
3780 return length;
3781}
3782SLAB_ATTR(trace);
3783
3784static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3785{
3786 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3787}
3788
3789static ssize_t reclaim_account_store(struct kmem_cache *s,
3790 const char *buf, size_t length)
3791{
3792 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3793 if (buf[0] == '1')
3794 s->flags |= SLAB_RECLAIM_ACCOUNT;
3795 return length;
3796}
3797SLAB_ATTR(reclaim_account);
3798
3799static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3800{
Christoph Lameter5af60832007-05-06 14:49:56 -07003801 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003802}
3803SLAB_ATTR_RO(hwcache_align);
3804
3805#ifdef CONFIG_ZONE_DMA
3806static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3807{
3808 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3809}
3810SLAB_ATTR_RO(cache_dma);
3811#endif
3812
3813static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3814{
3815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3816}
3817SLAB_ATTR_RO(destroy_by_rcu);
3818
3819static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3820{
3821 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3822}
3823
3824static ssize_t red_zone_store(struct kmem_cache *s,
3825 const char *buf, size_t length)
3826{
3827 if (any_slab_objects(s))
3828 return -EBUSY;
3829
3830 s->flags &= ~SLAB_RED_ZONE;
3831 if (buf[0] == '1')
3832 s->flags |= SLAB_RED_ZONE;
3833 calculate_sizes(s);
3834 return length;
3835}
3836SLAB_ATTR(red_zone);
3837
3838static ssize_t poison_show(struct kmem_cache *s, char *buf)
3839{
3840 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3841}
3842
3843static ssize_t poison_store(struct kmem_cache *s,
3844 const char *buf, size_t length)
3845{
3846 if (any_slab_objects(s))
3847 return -EBUSY;
3848
3849 s->flags &= ~SLAB_POISON;
3850 if (buf[0] == '1')
3851 s->flags |= SLAB_POISON;
3852 calculate_sizes(s);
3853 return length;
3854}
3855SLAB_ATTR(poison);
3856
3857static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3858{
3859 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3860}
3861
3862static ssize_t store_user_store(struct kmem_cache *s,
3863 const char *buf, size_t length)
3864{
3865 if (any_slab_objects(s))
3866 return -EBUSY;
3867
3868 s->flags &= ~SLAB_STORE_USER;
3869 if (buf[0] == '1')
3870 s->flags |= SLAB_STORE_USER;
3871 calculate_sizes(s);
3872 return length;
3873}
3874SLAB_ATTR(store_user);
3875
Christoph Lameter53e15af2007-05-06 14:49:43 -07003876static ssize_t validate_show(struct kmem_cache *s, char *buf)
3877{
3878 return 0;
3879}
3880
3881static ssize_t validate_store(struct kmem_cache *s,
3882 const char *buf, size_t length)
3883{
Christoph Lameter434e2452007-07-17 04:03:30 -07003884 int ret = -EINVAL;
3885
3886 if (buf[0] == '1') {
3887 ret = validate_slab_cache(s);
3888 if (ret >= 0)
3889 ret = length;
3890 }
3891 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003892}
3893SLAB_ATTR(validate);
3894
Christoph Lameter2086d262007-05-06 14:49:46 -07003895static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3896{
3897 return 0;
3898}
3899
3900static ssize_t shrink_store(struct kmem_cache *s,
3901 const char *buf, size_t length)
3902{
3903 if (buf[0] == '1') {
3904 int rc = kmem_cache_shrink(s);
3905
3906 if (rc)
3907 return rc;
3908 } else
3909 return -EINVAL;
3910 return length;
3911}
3912SLAB_ATTR(shrink);
3913
Christoph Lameter88a420e2007-05-06 14:49:45 -07003914static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3915{
3916 if (!(s->flags & SLAB_STORE_USER))
3917 return -ENOSYS;
3918 return list_locations(s, buf, TRACK_ALLOC);
3919}
3920SLAB_ATTR_RO(alloc_calls);
3921
3922static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3923{
3924 if (!(s->flags & SLAB_STORE_USER))
3925 return -ENOSYS;
3926 return list_locations(s, buf, TRACK_FREE);
3927}
3928SLAB_ATTR_RO(free_calls);
3929
Christoph Lameter81819f02007-05-06 14:49:36 -07003930#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08003931static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07003932{
Christoph Lameter98246012008-01-07 23:20:26 -08003933 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07003934}
3935
Christoph Lameter98246012008-01-07 23:20:26 -08003936static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07003937 const char *buf, size_t length)
3938{
3939 int n = simple_strtoul(buf, NULL, 10);
3940
3941 if (n < 100)
Christoph Lameter98246012008-01-07 23:20:26 -08003942 s->remote_node_defrag_ratio = n * 10;
Christoph Lameter81819f02007-05-06 14:49:36 -07003943 return length;
3944}
Christoph Lameter98246012008-01-07 23:20:26 -08003945SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07003946#endif
3947
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003948#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003949static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3950{
3951 unsigned long sum = 0;
3952 int cpu;
3953 int len;
3954 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3955
3956 if (!data)
3957 return -ENOMEM;
3958
3959 for_each_online_cpu(cpu) {
3960 unsigned x = get_cpu_slab(s, cpu)->stat[si];
3961
3962 data[cpu] = x;
3963 sum += x;
3964 }
3965
3966 len = sprintf(buf, "%lu", sum);
3967
3968 for_each_online_cpu(cpu) {
3969 if (data[cpu] && len < PAGE_SIZE - 20)
3970 len += sprintf(buf + len, " c%d=%u", cpu, data[cpu]);
3971 }
3972 kfree(data);
3973 return len + sprintf(buf + len, "\n");
3974}
3975
3976#define STAT_ATTR(si, text) \
3977static ssize_t text##_show(struct kmem_cache *s, char *buf) \
3978{ \
3979 return show_stat(s, buf, si); \
3980} \
3981SLAB_ATTR_RO(text); \
3982
3983STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
3984STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
3985STAT_ATTR(FREE_FASTPATH, free_fastpath);
3986STAT_ATTR(FREE_SLOWPATH, free_slowpath);
3987STAT_ATTR(FREE_FROZEN, free_frozen);
3988STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
3989STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
3990STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
3991STAT_ATTR(ALLOC_SLAB, alloc_slab);
3992STAT_ATTR(ALLOC_REFILL, alloc_refill);
3993STAT_ATTR(FREE_SLAB, free_slab);
3994STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
3995STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
3996STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
3997STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
3998STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
3999STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4000
4001#endif
4002
Pekka Enberg06428782008-01-07 23:20:27 -08004003static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004004 &slab_size_attr.attr,
4005 &object_size_attr.attr,
4006 &objs_per_slab_attr.attr,
4007 &order_attr.attr,
4008 &objects_attr.attr,
4009 &slabs_attr.attr,
4010 &partial_attr.attr,
4011 &cpu_slabs_attr.attr,
4012 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004013 &aliases_attr.attr,
4014 &align_attr.attr,
4015 &sanity_checks_attr.attr,
4016 &trace_attr.attr,
4017 &hwcache_align_attr.attr,
4018 &reclaim_account_attr.attr,
4019 &destroy_by_rcu_attr.attr,
4020 &red_zone_attr.attr,
4021 &poison_attr.attr,
4022 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004023 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004024 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004025 &alloc_calls_attr.attr,
4026 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004027#ifdef CONFIG_ZONE_DMA
4028 &cache_dma_attr.attr,
4029#endif
4030#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004031 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004032#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004033#ifdef CONFIG_SLUB_STATS
4034 &alloc_fastpath_attr.attr,
4035 &alloc_slowpath_attr.attr,
4036 &free_fastpath_attr.attr,
4037 &free_slowpath_attr.attr,
4038 &free_frozen_attr.attr,
4039 &free_add_partial_attr.attr,
4040 &free_remove_partial_attr.attr,
4041 &alloc_from_partial_attr.attr,
4042 &alloc_slab_attr.attr,
4043 &alloc_refill_attr.attr,
4044 &free_slab_attr.attr,
4045 &cpuslab_flush_attr.attr,
4046 &deactivate_full_attr.attr,
4047 &deactivate_empty_attr.attr,
4048 &deactivate_to_head_attr.attr,
4049 &deactivate_to_tail_attr.attr,
4050 &deactivate_remote_frees_attr.attr,
4051#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004052 NULL
4053};
4054
4055static struct attribute_group slab_attr_group = {
4056 .attrs = slab_attrs,
4057};
4058
4059static ssize_t slab_attr_show(struct kobject *kobj,
4060 struct attribute *attr,
4061 char *buf)
4062{
4063 struct slab_attribute *attribute;
4064 struct kmem_cache *s;
4065 int err;
4066
4067 attribute = to_slab_attr(attr);
4068 s = to_slab(kobj);
4069
4070 if (!attribute->show)
4071 return -EIO;
4072
4073 err = attribute->show(s, buf);
4074
4075 return err;
4076}
4077
4078static ssize_t slab_attr_store(struct kobject *kobj,
4079 struct attribute *attr,
4080 const char *buf, size_t len)
4081{
4082 struct slab_attribute *attribute;
4083 struct kmem_cache *s;
4084 int err;
4085
4086 attribute = to_slab_attr(attr);
4087 s = to_slab(kobj);
4088
4089 if (!attribute->store)
4090 return -EIO;
4091
4092 err = attribute->store(s, buf, len);
4093
4094 return err;
4095}
4096
Christoph Lameter151c6022008-01-07 22:29:05 -08004097static void kmem_cache_release(struct kobject *kobj)
4098{
4099 struct kmem_cache *s = to_slab(kobj);
4100
4101 kfree(s);
4102}
4103
Christoph Lameter81819f02007-05-06 14:49:36 -07004104static struct sysfs_ops slab_sysfs_ops = {
4105 .show = slab_attr_show,
4106 .store = slab_attr_store,
4107};
4108
4109static struct kobj_type slab_ktype = {
4110 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004111 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004112};
4113
4114static int uevent_filter(struct kset *kset, struct kobject *kobj)
4115{
4116 struct kobj_type *ktype = get_ktype(kobj);
4117
4118 if (ktype == &slab_ktype)
4119 return 1;
4120 return 0;
4121}
4122
4123static struct kset_uevent_ops slab_uevent_ops = {
4124 .filter = uevent_filter,
4125};
4126
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004127static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004128
4129#define ID_STR_LENGTH 64
4130
4131/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004132 *
4133 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004134 */
4135static char *create_unique_id(struct kmem_cache *s)
4136{
4137 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4138 char *p = name;
4139
4140 BUG_ON(!name);
4141
4142 *p++ = ':';
4143 /*
4144 * First flags affecting slabcache operations. We will only
4145 * get here for aliasable slabs so we do not need to support
4146 * too many flags. The flags here must cover all flags that
4147 * are matched during merging to guarantee that the id is
4148 * unique.
4149 */
4150 if (s->flags & SLAB_CACHE_DMA)
4151 *p++ = 'd';
4152 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4153 *p++ = 'a';
4154 if (s->flags & SLAB_DEBUG_FREE)
4155 *p++ = 'F';
4156 if (p != name + 1)
4157 *p++ = '-';
4158 p += sprintf(p, "%07d", s->size);
4159 BUG_ON(p > name + ID_STR_LENGTH - 1);
4160 return name;
4161}
4162
4163static int sysfs_slab_add(struct kmem_cache *s)
4164{
4165 int err;
4166 const char *name;
4167 int unmergeable;
4168
4169 if (slab_state < SYSFS)
4170 /* Defer until later */
4171 return 0;
4172
4173 unmergeable = slab_unmergeable(s);
4174 if (unmergeable) {
4175 /*
4176 * Slabcache can never be merged so we can use the name proper.
4177 * This is typically the case for debug situations. In that
4178 * case we can catch duplicate names easily.
4179 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004180 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004181 name = s->name;
4182 } else {
4183 /*
4184 * Create a unique name for the slab as a target
4185 * for the symlinks.
4186 */
4187 name = create_unique_id(s);
4188 }
4189
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004190 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004191 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4192 if (err) {
4193 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004194 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004195 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004196
4197 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4198 if (err)
4199 return err;
4200 kobject_uevent(&s->kobj, KOBJ_ADD);
4201 if (!unmergeable) {
4202 /* Setup first alias */
4203 sysfs_slab_alias(s, s->name);
4204 kfree(name);
4205 }
4206 return 0;
4207}
4208
4209static void sysfs_slab_remove(struct kmem_cache *s)
4210{
4211 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4212 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004213 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004214}
4215
4216/*
4217 * Need to buffer aliases during bootup until sysfs becomes
4218 * available lest we loose that information.
4219 */
4220struct saved_alias {
4221 struct kmem_cache *s;
4222 const char *name;
4223 struct saved_alias *next;
4224};
4225
Adrian Bunk5af328a2007-07-17 04:03:27 -07004226static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004227
4228static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4229{
4230 struct saved_alias *al;
4231
4232 if (slab_state == SYSFS) {
4233 /*
4234 * If we have a leftover link then remove it.
4235 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004236 sysfs_remove_link(&slab_kset->kobj, name);
4237 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004238 }
4239
4240 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4241 if (!al)
4242 return -ENOMEM;
4243
4244 al->s = s;
4245 al->name = name;
4246 al->next = alias_list;
4247 alias_list = al;
4248 return 0;
4249}
4250
4251static int __init slab_sysfs_init(void)
4252{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004253 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004254 int err;
4255
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004256 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004257 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004258 printk(KERN_ERR "Cannot register slab subsystem.\n");
4259 return -ENOSYS;
4260 }
4261
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004262 slab_state = SYSFS;
4263
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004264 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004265 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004266 if (err)
4267 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4268 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004269 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004270
4271 while (alias_list) {
4272 struct saved_alias *al = alias_list;
4273
4274 alias_list = alias_list->next;
4275 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004276 if (err)
4277 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4278 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004279 kfree(al);
4280 }
4281
4282 resiliency_test();
4283 return 0;
4284}
4285
4286__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004287#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004288
4289/*
4290 * The /proc/slabinfo ABI
4291 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004292#ifdef CONFIG_SLABINFO
4293
4294ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4295 size_t count, loff_t *ppos)
4296{
4297 return -EINVAL;
4298}
4299
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004300
4301static void print_slabinfo_header(struct seq_file *m)
4302{
4303 seq_puts(m, "slabinfo - version: 2.1\n");
4304 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4305 "<objperslab> <pagesperslab>");
4306 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4307 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4308 seq_putc(m, '\n');
4309}
4310
4311static void *s_start(struct seq_file *m, loff_t *pos)
4312{
4313 loff_t n = *pos;
4314
4315 down_read(&slub_lock);
4316 if (!n)
4317 print_slabinfo_header(m);
4318
4319 return seq_list_start(&slab_caches, *pos);
4320}
4321
4322static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4323{
4324 return seq_list_next(p, &slab_caches, pos);
4325}
4326
4327static void s_stop(struct seq_file *m, void *p)
4328{
4329 up_read(&slub_lock);
4330}
4331
4332static int s_show(struct seq_file *m, void *p)
4333{
4334 unsigned long nr_partials = 0;
4335 unsigned long nr_slabs = 0;
4336 unsigned long nr_inuse = 0;
4337 unsigned long nr_objs;
4338 struct kmem_cache *s;
4339 int node;
4340
4341 s = list_entry(p, struct kmem_cache, list);
4342
4343 for_each_online_node(node) {
4344 struct kmem_cache_node *n = get_node(s, node);
4345
4346 if (!n)
4347 continue;
4348
4349 nr_partials += n->nr_partial;
4350 nr_slabs += atomic_long_read(&n->nr_slabs);
4351 nr_inuse += count_partial(n);
4352 }
4353
4354 nr_objs = nr_slabs * s->objects;
4355 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4356
4357 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4358 nr_objs, s->size, s->objects, (1 << s->order));
4359 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4360 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4361 0UL);
4362 seq_putc(m, '\n');
4363 return 0;
4364}
4365
4366const struct seq_operations slabinfo_op = {
4367 .start = s_start,
4368 .next = s_next,
4369 .stop = s_stop,
4370 .show = s_show,
4371};
4372
Linus Torvalds158a9622008-01-02 13:04:48 -08004373#endif /* CONFIG_SLABINFO */