blob: 10d546954efa81e6cf1c46b8376842ed7754a06b [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
1371 if (c->freelist)
1372 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 Lameterdfb4f092007-10-16 01:26:05 -07001489 object = c->page->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001490 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001491 c->page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001492 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001493 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001494unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001495 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001496 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001497 return object;
1498
1499another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001500 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001501
1502new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001503 new = get_partial(s, gfpflags, node);
1504 if (new) {
1505 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001506 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001507 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001508 }
1509
Christoph Lameterb811c202007-10-16 23:25:51 -07001510 if (gfpflags & __GFP_WAIT)
1511 local_irq_enable();
1512
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001513 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001514
1515 if (gfpflags & __GFP_WAIT)
1516 local_irq_disable();
1517
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001518 if (new) {
1519 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001520 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001521 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001522 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001523 slab_lock(new);
1524 SetSlabFrozen(new);
1525 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001526 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001527 }
Linus Torvalds00e962c2008-02-19 09:08:49 -08001528
Christoph Lameter71c7a062008-02-14 14:28:01 -08001529 /*
1530 * No memory available.
1531 *
1532 * If the slab uses higher order allocs but the object is
1533 * smaller than a page size then we can fallback in emergencies
1534 * to the page allocator via kmalloc_large. The page allocator may
1535 * have failed to obtain a higher order page and we can try to
1536 * allocate a single page if the object fits into a single page.
1537 * That is only possible if certain conditions are met that are being
1538 * checked when a slab is created.
1539 */
1540 if (!(gfpflags & __GFP_NORETRY) && (s->flags & __PAGE_ALLOC_FALLBACK))
1541 return kmalloc_large(s->objsize, gfpflags);
1542
1543 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001544debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001545 object = c->page->freelist;
1546 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001547 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001548
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001549 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001550 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001551 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001552 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001553}
1554
1555/*
1556 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1557 * have the fastpath folded into their functions. So no function call
1558 * overhead for requests that can be satisfied on the fastpath.
1559 *
1560 * The fastpath works by first checking if the lockless freelist can be used.
1561 * If not then __slab_alloc is called for slow processing.
1562 *
1563 * Otherwise we can simply pick the next object from the lockless free list.
1564 */
Pekka Enberg06428782008-01-07 23:20:27 -08001565static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001566 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001567{
Christoph Lameter894b8782007-05-10 03:15:16 -07001568 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001569 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001570 unsigned long flags;
1571
Christoph Lameter894b8782007-05-10 03:15:16 -07001572 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001573 c = get_cpu_slab(s, smp_processor_id());
Christoph Lametera973e9d2008-03-01 13:40:44 -08001574 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001575
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001576 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001577
1578 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001579 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001580 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001581 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001582 }
1583 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001584
1585 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001586 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001587
Christoph Lameter894b8782007-05-10 03:15:16 -07001588 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001589}
1590
1591void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1592{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001593 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001594}
1595EXPORT_SYMBOL(kmem_cache_alloc);
1596
1597#ifdef CONFIG_NUMA
1598void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1599{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001600 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001601}
1602EXPORT_SYMBOL(kmem_cache_alloc_node);
1603#endif
1604
1605/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001606 * Slow patch handling. This may still be called frequently since objects
1607 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001608 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001609 * So we still attempt to reduce cache line usage. Just take the slab
1610 * lock and free the item. If there is no additional partial page
1611 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001612 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001613static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001614 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001615{
1616 void *prior;
1617 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001618 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001619
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001620 c = get_cpu_slab(s, raw_smp_processor_id());
1621 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001622 slab_lock(page);
1623
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001624 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001625 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001626
Christoph Lameter81819f02007-05-06 14:49:36 -07001627checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001628 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001629 page->freelist = object;
1630 page->inuse--;
1631
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001632 if (unlikely(SlabFrozen(page))) {
1633 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001634 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001635 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001636
1637 if (unlikely(!page->inuse))
1638 goto slab_empty;
1639
1640 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001641 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001642 * then add it.
1643 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001644 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001645 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001646 stat(c, FREE_ADD_PARTIAL);
1647 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001648
1649out_unlock:
1650 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001651 return;
1652
1653slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001654 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001655 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001656 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001657 */
1658 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001659 stat(c, FREE_REMOVE_PARTIAL);
1660 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001661 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001662 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001663 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001664 return;
1665
1666debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001667 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001668 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001669 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001670}
1671
Christoph Lameter894b8782007-05-10 03:15:16 -07001672/*
1673 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1674 * can perform fastpath freeing without additional function calls.
1675 *
1676 * The fastpath is only possible if we are freeing to the current cpu slab
1677 * of this processor. This typically the case if we have just allocated
1678 * the item before.
1679 *
1680 * If fastpath is not possible then fall back to __slab_free where we deal
1681 * with all sorts of special processing.
1682 */
Pekka Enberg06428782008-01-07 23:20:27 -08001683static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001684 struct page *page, void *x, void *addr)
1685{
1686 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001687 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001688 unsigned long flags;
1689
Christoph Lameter894b8782007-05-10 03:15:16 -07001690 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001691 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001692 debug_check_no_locks_freed(object, c->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001693 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001694 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001695 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001696 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001697 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001698 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001699
1700 local_irq_restore(flags);
1701}
1702
Christoph Lameter81819f02007-05-06 14:49:36 -07001703void kmem_cache_free(struct kmem_cache *s, void *x)
1704{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001705 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001706
Christoph Lameterb49af682007-05-06 14:49:41 -07001707 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001708
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001709 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001710}
1711EXPORT_SYMBOL(kmem_cache_free);
1712
1713/* Figure out on which slab object the object resides */
1714static struct page *get_object_page(const void *x)
1715{
Christoph Lameterb49af682007-05-06 14:49:41 -07001716 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001717
1718 if (!PageSlab(page))
1719 return NULL;
1720
1721 return page;
1722}
1723
1724/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001725 * Object placement in a slab is made very easy because we always start at
1726 * offset 0. If we tune the size of the object to the alignment then we can
1727 * get the required alignment by putting one properly sized object after
1728 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001729 *
1730 * Notice that the allocation order determines the sizes of the per cpu
1731 * caches. Each processor has always one slab available for allocations.
1732 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001733 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001734 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001735 */
1736
1737/*
1738 * Mininum / Maximum order of slab pages. This influences locking overhead
1739 * and slab fragmentation. A higher order reduces the number of partial slabs
1740 * and increases the number of allocations possible without having to
1741 * take the list_lock.
1742 */
1743static int slub_min_order;
1744static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001745static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1746
1747/*
1748 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001749 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001750 */
1751static int slub_nomerge;
1752
1753/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001754 * Calculate the order of allocation given an slab object size.
1755 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001756 * The order of allocation has significant impact on performance and other
1757 * system components. Generally order 0 allocations should be preferred since
1758 * order 0 does not cause fragmentation in the page allocator. Larger objects
1759 * be problematic to put into order 0 slabs because there may be too much
1760 * unused space left. We go to a higher order if more than 1/8th of the slab
1761 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001762 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001763 * In order to reach satisfactory performance we must ensure that a minimum
1764 * number of objects is in one slab. Otherwise we may generate too much
1765 * activity on the partial lists which requires taking the list_lock. This is
1766 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001767 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001768 * slub_max_order specifies the order where we begin to stop considering the
1769 * number of objects in a slab as critical. If we reach slub_max_order then
1770 * we try to keep the page order as low as possible. So we accept more waste
1771 * of space in favor of a small page order.
1772 *
1773 * Higher order allocations also allow the placement of more objects in a
1774 * slab and thereby reduce object handling overhead. If the user has
1775 * requested a higher mininum order then we start with that one instead of
1776 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001777 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001778static inline int slab_order(int size, int min_objects,
1779 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001780{
1781 int order;
1782 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001783 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001784
Christoph Lameter6300ea72007-07-17 04:03:20 -07001785 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001786 fls(min_objects * size - 1) - PAGE_SHIFT);
1787 order <= max_order; order++) {
1788
Christoph Lameter81819f02007-05-06 14:49:36 -07001789 unsigned long slab_size = PAGE_SIZE << order;
1790
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001791 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001792 continue;
1793
Christoph Lameter81819f02007-05-06 14:49:36 -07001794 rem = slab_size % size;
1795
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001796 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001797 break;
1798
1799 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001800
Christoph Lameter81819f02007-05-06 14:49:36 -07001801 return order;
1802}
1803
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001804static inline int calculate_order(int size)
1805{
1806 int order;
1807 int min_objects;
1808 int fraction;
1809
1810 /*
1811 * Attempt to find best configuration for a slab. This
1812 * works by first attempting to generate a layout with
1813 * the best configuration and backing off gradually.
1814 *
1815 * First we reduce the acceptable waste in a slab. Then
1816 * we reduce the minimum objects required in a slab.
1817 */
1818 min_objects = slub_min_objects;
1819 while (min_objects > 1) {
1820 fraction = 8;
1821 while (fraction >= 4) {
1822 order = slab_order(size, min_objects,
1823 slub_max_order, fraction);
1824 if (order <= slub_max_order)
1825 return order;
1826 fraction /= 2;
1827 }
1828 min_objects /= 2;
1829 }
1830
1831 /*
1832 * We were unable to place multiple objects in a slab. Now
1833 * lets see if we can place a single object there.
1834 */
1835 order = slab_order(size, 1, slub_max_order, 1);
1836 if (order <= slub_max_order)
1837 return order;
1838
1839 /*
1840 * Doh this slab cannot be placed using slub_max_order.
1841 */
1842 order = slab_order(size, 1, MAX_ORDER, 1);
1843 if (order <= MAX_ORDER)
1844 return order;
1845 return -ENOSYS;
1846}
1847
Christoph Lameter81819f02007-05-06 14:49:36 -07001848/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001849 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001850 */
1851static unsigned long calculate_alignment(unsigned long flags,
1852 unsigned long align, unsigned long size)
1853{
1854 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001855 * If the user wants hardware cache aligned objects then follow that
1856 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001857 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001858 * The hardware cache alignment cannot override the specified
1859 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001860 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001861 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001862 size > cache_line_size() / 2)
1863 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001864
1865 if (align < ARCH_SLAB_MINALIGN)
1866 return ARCH_SLAB_MINALIGN;
1867
1868 return ALIGN(align, sizeof(void *));
1869}
1870
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001871static void init_kmem_cache_cpu(struct kmem_cache *s,
1872 struct kmem_cache_cpu *c)
1873{
1874 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001875 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001876 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001877 c->offset = s->offset / sizeof(void *);
1878 c->objsize = s->objsize;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001879}
1880
Christoph Lameter81819f02007-05-06 14:49:36 -07001881static void init_kmem_cache_node(struct kmem_cache_node *n)
1882{
1883 n->nr_partial = 0;
1884 atomic_long_set(&n->nr_slabs, 0);
1885 spin_lock_init(&n->list_lock);
1886 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001887#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07001888 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001889#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001890}
1891
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001892#ifdef CONFIG_SMP
1893/*
1894 * Per cpu array for per cpu structures.
1895 *
1896 * The per cpu array places all kmem_cache_cpu structures from one processor
1897 * close together meaning that it becomes possible that multiple per cpu
1898 * structures are contained in one cacheline. This may be particularly
1899 * beneficial for the kmalloc caches.
1900 *
1901 * A desktop system typically has around 60-80 slabs. With 100 here we are
1902 * likely able to get per cpu structures for all caches from the array defined
1903 * here. We must be able to cover all kmalloc caches during bootstrap.
1904 *
1905 * If the per cpu array is exhausted then fall back to kmalloc
1906 * of individual cachelines. No sharing is possible then.
1907 */
1908#define NR_KMEM_CACHE_CPU 100
1909
1910static DEFINE_PER_CPU(struct kmem_cache_cpu,
1911 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1912
1913static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1914static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1915
1916static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1917 int cpu, gfp_t flags)
1918{
1919 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1920
1921 if (c)
1922 per_cpu(kmem_cache_cpu_free, cpu) =
1923 (void *)c->freelist;
1924 else {
1925 /* Table overflow: So allocate ourselves */
1926 c = kmalloc_node(
1927 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1928 flags, cpu_to_node(cpu));
1929 if (!c)
1930 return NULL;
1931 }
1932
1933 init_kmem_cache_cpu(s, c);
1934 return c;
1935}
1936
1937static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1938{
1939 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1940 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1941 kfree(c);
1942 return;
1943 }
1944 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1945 per_cpu(kmem_cache_cpu_free, cpu) = c;
1946}
1947
1948static void free_kmem_cache_cpus(struct kmem_cache *s)
1949{
1950 int cpu;
1951
1952 for_each_online_cpu(cpu) {
1953 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1954
1955 if (c) {
1956 s->cpu_slab[cpu] = NULL;
1957 free_kmem_cache_cpu(c, cpu);
1958 }
1959 }
1960}
1961
1962static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1963{
1964 int cpu;
1965
1966 for_each_online_cpu(cpu) {
1967 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1968
1969 if (c)
1970 continue;
1971
1972 c = alloc_kmem_cache_cpu(s, cpu, flags);
1973 if (!c) {
1974 free_kmem_cache_cpus(s);
1975 return 0;
1976 }
1977 s->cpu_slab[cpu] = c;
1978 }
1979 return 1;
1980}
1981
1982/*
1983 * Initialize the per cpu array.
1984 */
1985static void init_alloc_cpu_cpu(int cpu)
1986{
1987 int i;
1988
1989 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
1990 return;
1991
1992 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
1993 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
1994
1995 cpu_set(cpu, kmem_cach_cpu_free_init_once);
1996}
1997
1998static void __init init_alloc_cpu(void)
1999{
2000 int cpu;
2001
2002 for_each_online_cpu(cpu)
2003 init_alloc_cpu_cpu(cpu);
2004 }
2005
2006#else
2007static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2008static inline void init_alloc_cpu(void) {}
2009
2010static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2011{
2012 init_kmem_cache_cpu(s, &s->cpu_slab);
2013 return 1;
2014}
2015#endif
2016
Christoph Lameter81819f02007-05-06 14:49:36 -07002017#ifdef CONFIG_NUMA
2018/*
2019 * No kmalloc_node yet so do it by hand. We know that this is the first
2020 * slab on the node for this slabcache. There are no concurrent accesses
2021 * possible.
2022 *
2023 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002024 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2025 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002026 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002027static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2028 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002029{
2030 struct page *page;
2031 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002032 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002033
2034 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2035
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002036 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002037
2038 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002039 if (page_to_nid(page) != node) {
2040 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2041 "node %d\n", node);
2042 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2043 "in order to be able to continue\n");
2044 }
2045
Christoph Lameter81819f02007-05-06 14:49:36 -07002046 n = page->freelist;
2047 BUG_ON(!n);
2048 page->freelist = get_freepointer(kmalloc_caches, n);
2049 page->inuse++;
2050 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002051#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002052 init_object(kmalloc_caches, n, 1);
2053 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002054#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002055 init_kmem_cache_node(n);
2056 atomic_long_inc(&n->nr_slabs);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002057
rootba84c732008-01-07 23:20:28 -08002058 /*
2059 * lockdep requires consistent irq usage for each lock
2060 * so even though there cannot be a race this early in
2061 * the boot sequence, we still disable irqs.
2062 */
2063 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002064 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002065 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002066 return n;
2067}
2068
2069static void free_kmem_cache_nodes(struct kmem_cache *s)
2070{
2071 int node;
2072
Christoph Lameterf64dc582007-10-16 01:25:33 -07002073 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002074 struct kmem_cache_node *n = s->node[node];
2075 if (n && n != &s->local_node)
2076 kmem_cache_free(kmalloc_caches, n);
2077 s->node[node] = NULL;
2078 }
2079}
2080
2081static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2082{
2083 int node;
2084 int local_node;
2085
2086 if (slab_state >= UP)
2087 local_node = page_to_nid(virt_to_page(s));
2088 else
2089 local_node = 0;
2090
Christoph Lameterf64dc582007-10-16 01:25:33 -07002091 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002092 struct kmem_cache_node *n;
2093
2094 if (local_node == node)
2095 n = &s->local_node;
2096 else {
2097 if (slab_state == DOWN) {
2098 n = early_kmem_cache_node_alloc(gfpflags,
2099 node);
2100 continue;
2101 }
2102 n = kmem_cache_alloc_node(kmalloc_caches,
2103 gfpflags, node);
2104
2105 if (!n) {
2106 free_kmem_cache_nodes(s);
2107 return 0;
2108 }
2109
2110 }
2111 s->node[node] = n;
2112 init_kmem_cache_node(n);
2113 }
2114 return 1;
2115}
2116#else
2117static void free_kmem_cache_nodes(struct kmem_cache *s)
2118{
2119}
2120
2121static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2122{
2123 init_kmem_cache_node(&s->local_node);
2124 return 1;
2125}
2126#endif
2127
2128/*
2129 * calculate_sizes() determines the order and the distribution of data within
2130 * a slab object.
2131 */
2132static int calculate_sizes(struct kmem_cache *s)
2133{
2134 unsigned long flags = s->flags;
2135 unsigned long size = s->objsize;
2136 unsigned long align = s->align;
2137
2138 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002139 * Round up object size to the next word boundary. We can only
2140 * place the free pointer at word boundaries and this determines
2141 * the possible location of the free pointer.
2142 */
2143 size = ALIGN(size, sizeof(void *));
2144
2145#ifdef CONFIG_SLUB_DEBUG
2146 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002147 * Determine if we can poison the object itself. If the user of
2148 * the slab may touch the object after free or before allocation
2149 * then we should never poison the object itself.
2150 */
2151 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002152 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002153 s->flags |= __OBJECT_POISON;
2154 else
2155 s->flags &= ~__OBJECT_POISON;
2156
Christoph Lameter81819f02007-05-06 14:49:36 -07002157
2158 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002159 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002160 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002161 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002162 */
2163 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2164 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002165#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002166
2167 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002168 * With that we have determined the number of bytes in actual use
2169 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002170 */
2171 s->inuse = size;
2172
2173 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002174 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002175 /*
2176 * Relocate free pointer after the object if it is not
2177 * permitted to overwrite the first word of the object on
2178 * kmem_cache_free.
2179 *
2180 * This is the case if we do RCU, have a constructor or
2181 * destructor or are poisoning the objects.
2182 */
2183 s->offset = size;
2184 size += sizeof(void *);
2185 }
2186
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002187#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002188 if (flags & SLAB_STORE_USER)
2189 /*
2190 * Need to store information about allocs and frees after
2191 * the object.
2192 */
2193 size += 2 * sizeof(struct track);
2194
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002195 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002196 /*
2197 * Add some empty padding so that we can catch
2198 * overwrites from earlier objects rather than let
2199 * tracking information or the free pointer be
2200 * corrupted if an user writes before the start
2201 * of the object.
2202 */
2203 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002204#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002205
Christoph Lameter81819f02007-05-06 14:49:36 -07002206 /*
2207 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002208 * user specified and the dynamic determination of cache line size
2209 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002210 */
2211 align = calculate_alignment(flags, align, s->objsize);
2212
2213 /*
2214 * SLUB stores one object immediately after another beginning from
2215 * offset 0. In order to align the objects we have to simply size
2216 * each object to conform to the alignment.
2217 */
2218 size = ALIGN(size, align);
2219 s->size = size;
2220
Christoph Lameter71c7a062008-02-14 14:28:01 -08002221 if ((flags & __KMALLOC_CACHE) &&
2222 PAGE_SIZE / size < slub_min_objects) {
2223 /*
2224 * Kmalloc cache that would not have enough objects in
2225 * an order 0 page. Kmalloc slabs can fallback to
2226 * page allocator order 0 allocs so take a reasonably large
2227 * order that will allows us a good number of objects.
2228 */
2229 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2230 s->flags |= __PAGE_ALLOC_FALLBACK;
2231 s->allocflags |= __GFP_NOWARN;
2232 } else
2233 s->order = calculate_order(size);
2234
Christoph Lameter81819f02007-05-06 14:49:36 -07002235 if (s->order < 0)
2236 return 0;
2237
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002238 s->allocflags = 0;
2239 if (s->order)
2240 s->allocflags |= __GFP_COMP;
2241
2242 if (s->flags & SLAB_CACHE_DMA)
2243 s->allocflags |= SLUB_DMA;
2244
2245 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2246 s->allocflags |= __GFP_RECLAIMABLE;
2247
Christoph Lameter81819f02007-05-06 14:49:36 -07002248 /*
2249 * Determine the number of objects per slab
2250 */
2251 s->objects = (PAGE_SIZE << s->order) / size;
2252
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002253 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002254
2255}
2256
Christoph Lameter81819f02007-05-06 14:49:36 -07002257static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2258 const char *name, size_t size,
2259 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002260 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002261{
2262 memset(s, 0, kmem_size);
2263 s->name = name;
2264 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002265 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002266 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002267 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002268
2269 if (!calculate_sizes(s))
2270 goto error;
2271
2272 s->refcount = 1;
2273#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002274 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002275#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002276 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2277 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002278
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002279 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002280 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002281 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002282error:
2283 if (flags & SLAB_PANIC)
2284 panic("Cannot create slab %s size=%lu realsize=%u "
2285 "order=%u offset=%u flags=%lx\n",
2286 s->name, (unsigned long)size, s->size, s->order,
2287 s->offset, flags);
2288 return 0;
2289}
Christoph Lameter81819f02007-05-06 14:49:36 -07002290
2291/*
2292 * Check if a given pointer is valid
2293 */
2294int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2295{
Pekka Enberg06428782008-01-07 23:20:27 -08002296 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002297
2298 page = get_object_page(object);
2299
2300 if (!page || s != page->slab)
2301 /* No slab or wrong slab */
2302 return 0;
2303
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002304 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002305 return 0;
2306
2307 /*
2308 * We could also check if the object is on the slabs freelist.
2309 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002310 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002311 * to a certain slab.
2312 */
2313 return 1;
2314}
2315EXPORT_SYMBOL(kmem_ptr_validate);
2316
2317/*
2318 * Determine the size of a slab object
2319 */
2320unsigned int kmem_cache_size(struct kmem_cache *s)
2321{
2322 return s->objsize;
2323}
2324EXPORT_SYMBOL(kmem_cache_size);
2325
2326const char *kmem_cache_name(struct kmem_cache *s)
2327{
2328 return s->name;
2329}
2330EXPORT_SYMBOL(kmem_cache_name);
2331
2332/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002333 * Attempt to free all slabs on a node. Return the number of slabs we
2334 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002335 */
2336static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2337 struct list_head *list)
2338{
2339 int slabs_inuse = 0;
2340 unsigned long flags;
2341 struct page *page, *h;
2342
2343 spin_lock_irqsave(&n->list_lock, flags);
2344 list_for_each_entry_safe(page, h, list, lru)
2345 if (!page->inuse) {
2346 list_del(&page->lru);
2347 discard_slab(s, page);
2348 } else
2349 slabs_inuse++;
2350 spin_unlock_irqrestore(&n->list_lock, flags);
2351 return slabs_inuse;
2352}
2353
2354/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002355 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002356 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002357static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002358{
2359 int node;
2360
2361 flush_all(s);
2362
2363 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002364 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002365 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002366 struct kmem_cache_node *n = get_node(s, node);
2367
Christoph Lameter2086d262007-05-06 14:49:46 -07002368 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002369 if (atomic_long_read(&n->nr_slabs))
2370 return 1;
2371 }
2372 free_kmem_cache_nodes(s);
2373 return 0;
2374}
2375
2376/*
2377 * Close a cache and release the kmem_cache structure
2378 * (must be used for caches created using kmem_cache_create)
2379 */
2380void kmem_cache_destroy(struct kmem_cache *s)
2381{
2382 down_write(&slub_lock);
2383 s->refcount--;
2384 if (!s->refcount) {
2385 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002386 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002387 if (kmem_cache_close(s))
2388 WARN_ON(1);
2389 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002390 } else
2391 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002392}
2393EXPORT_SYMBOL(kmem_cache_destroy);
2394
2395/********************************************************************
2396 * Kmalloc subsystem
2397 *******************************************************************/
2398
Christoph Lameter331dc552008-02-14 14:28:09 -08002399struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002400EXPORT_SYMBOL(kmalloc_caches);
2401
2402#ifdef CONFIG_ZONE_DMA
Christoph Lameter331dc552008-02-14 14:28:09 -08002403static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter81819f02007-05-06 14:49:36 -07002404#endif
2405
2406static int __init setup_slub_min_order(char *str)
2407{
Pekka Enberg06428782008-01-07 23:20:27 -08002408 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002409
2410 return 1;
2411}
2412
2413__setup("slub_min_order=", setup_slub_min_order);
2414
2415static int __init setup_slub_max_order(char *str)
2416{
Pekka Enberg06428782008-01-07 23:20:27 -08002417 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002418
2419 return 1;
2420}
2421
2422__setup("slub_max_order=", setup_slub_max_order);
2423
2424static int __init setup_slub_min_objects(char *str)
2425{
Pekka Enberg06428782008-01-07 23:20:27 -08002426 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002427
2428 return 1;
2429}
2430
2431__setup("slub_min_objects=", setup_slub_min_objects);
2432
2433static int __init setup_slub_nomerge(char *str)
2434{
2435 slub_nomerge = 1;
2436 return 1;
2437}
2438
2439__setup("slub_nomerge", setup_slub_nomerge);
2440
Christoph Lameter81819f02007-05-06 14:49:36 -07002441static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2442 const char *name, int size, gfp_t gfp_flags)
2443{
2444 unsigned int flags = 0;
2445
2446 if (gfp_flags & SLUB_DMA)
2447 flags = SLAB_CACHE_DMA;
2448
2449 down_write(&slub_lock);
2450 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter71c7a062008-02-14 14:28:01 -08002451 flags | __KMALLOC_CACHE, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002452 goto panic;
2453
2454 list_add(&s->list, &slab_caches);
2455 up_write(&slub_lock);
2456 if (sysfs_slab_add(s))
2457 goto panic;
2458 return s;
2459
2460panic:
2461 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2462}
2463
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002464#ifdef CONFIG_ZONE_DMA
Christoph Lameter1ceef402007-08-07 15:11:48 -07002465
2466static void sysfs_add_func(struct work_struct *w)
2467{
2468 struct kmem_cache *s;
2469
2470 down_write(&slub_lock);
2471 list_for_each_entry(s, &slab_caches, list) {
2472 if (s->flags & __SYSFS_ADD_DEFERRED) {
2473 s->flags &= ~__SYSFS_ADD_DEFERRED;
2474 sysfs_slab_add(s);
2475 }
2476 }
2477 up_write(&slub_lock);
2478}
2479
2480static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2481
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002482static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2483{
2484 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002485 char *text;
2486 size_t realsize;
2487
2488 s = kmalloc_caches_dma[index];
2489 if (s)
2490 return s;
2491
2492 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002493 if (flags & __GFP_WAIT)
2494 down_write(&slub_lock);
2495 else {
2496 if (!down_write_trylock(&slub_lock))
2497 goto out;
2498 }
2499
2500 if (kmalloc_caches_dma[index])
2501 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002502
Christoph Lameter7b55f622007-07-17 04:03:27 -07002503 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002504 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2505 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002506 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2507
2508 if (!s || !text || !kmem_cache_open(s, flags, text,
2509 realsize, ARCH_KMALLOC_MINALIGN,
2510 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2511 kfree(s);
2512 kfree(text);
2513 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002514 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002515
2516 list_add(&s->list, &slab_caches);
2517 kmalloc_caches_dma[index] = s;
2518
2519 schedule_work(&sysfs_add_work);
2520
2521unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002522 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002523out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002524 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002525}
2526#endif
2527
Christoph Lameterf1b26332007-07-17 04:03:26 -07002528/*
2529 * Conversion table for small slabs sizes / 8 to the index in the
2530 * kmalloc array. This is necessary for slabs < 192 since we have non power
2531 * of two cache sizes there. The size of larger slabs can be determined using
2532 * fls.
2533 */
2534static s8 size_index[24] = {
2535 3, /* 8 */
2536 4, /* 16 */
2537 5, /* 24 */
2538 5, /* 32 */
2539 6, /* 40 */
2540 6, /* 48 */
2541 6, /* 56 */
2542 6, /* 64 */
2543 1, /* 72 */
2544 1, /* 80 */
2545 1, /* 88 */
2546 1, /* 96 */
2547 7, /* 104 */
2548 7, /* 112 */
2549 7, /* 120 */
2550 7, /* 128 */
2551 2, /* 136 */
2552 2, /* 144 */
2553 2, /* 152 */
2554 2, /* 160 */
2555 2, /* 168 */
2556 2, /* 176 */
2557 2, /* 184 */
2558 2 /* 192 */
2559};
2560
Christoph Lameter81819f02007-05-06 14:49:36 -07002561static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2562{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002563 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002564
Christoph Lameterf1b26332007-07-17 04:03:26 -07002565 if (size <= 192) {
2566 if (!size)
2567 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002568
Christoph Lameterf1b26332007-07-17 04:03:26 -07002569 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002570 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002571 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002572
2573#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002574 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002575 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002576
Christoph Lameter81819f02007-05-06 14:49:36 -07002577#endif
2578 return &kmalloc_caches[index];
2579}
2580
2581void *__kmalloc(size_t size, gfp_t flags)
2582{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002583 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002584
Christoph Lameter331dc552008-02-14 14:28:09 -08002585 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002586 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002587
2588 s = get_slab(size, flags);
2589
2590 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002591 return s;
2592
Christoph Lameterce15fea2007-07-17 04:03:28 -07002593 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002594}
2595EXPORT_SYMBOL(__kmalloc);
2596
2597#ifdef CONFIG_NUMA
2598void *__kmalloc_node(size_t size, gfp_t flags, int node)
2599{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002600 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002601
Christoph Lameter331dc552008-02-14 14:28:09 -08002602 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002603 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002604
2605 s = get_slab(size, flags);
2606
2607 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002608 return s;
2609
Christoph Lameterce15fea2007-07-17 04:03:28 -07002610 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002611}
2612EXPORT_SYMBOL(__kmalloc_node);
2613#endif
2614
2615size_t ksize(const void *object)
2616{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002617 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002618 struct kmem_cache *s;
2619
Christoph Lameteref8b4522007-10-16 01:24:46 -07002620 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002621 return 0;
2622
Vegard Nossum294a80a2007-12-04 23:45:30 -08002623 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002624
2625 if (unlikely(!PageSlab(page)))
2626 return PAGE_SIZE << compound_order(page);
2627
Christoph Lameter81819f02007-05-06 14:49:36 -07002628 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002629
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002630#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002631 /*
2632 * Debugging requires use of the padding between object
2633 * and whatever may come after it.
2634 */
2635 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2636 return s->objsize;
2637
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002638#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002639 /*
2640 * If we have the need to store the freelist pointer
2641 * back there or track user information then we can
2642 * only use the space before that information.
2643 */
2644 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2645 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002646 /*
2647 * Else we can use all the padding etc for the allocation
2648 */
2649 return s->size;
2650}
2651EXPORT_SYMBOL(ksize);
2652
2653void kfree(const void *x)
2654{
Christoph Lameter81819f02007-05-06 14:49:36 -07002655 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002656 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002657
Satyam Sharma2408c552007-10-16 01:24:44 -07002658 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002659 return;
2660
Christoph Lameterb49af682007-05-06 14:49:41 -07002661 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002662 if (unlikely(!PageSlab(page))) {
2663 put_page(page);
2664 return;
2665 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002666 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002667}
2668EXPORT_SYMBOL(kfree);
2669
Christoph Lameterf61396a2008-01-07 23:20:26 -08002670static unsigned long count_partial(struct kmem_cache_node *n)
2671{
2672 unsigned long flags;
2673 unsigned long x = 0;
2674 struct page *page;
2675
2676 spin_lock_irqsave(&n->list_lock, flags);
2677 list_for_each_entry(page, &n->partial, lru)
2678 x += page->inuse;
2679 spin_unlock_irqrestore(&n->list_lock, flags);
2680 return x;
2681}
2682
Christoph Lameter2086d262007-05-06 14:49:46 -07002683/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002684 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2685 * the remaining slabs by the number of items in use. The slabs with the
2686 * most items in use come first. New allocations will then fill those up
2687 * and thus they can be removed from the partial lists.
2688 *
2689 * The slabs with the least items are placed last. This results in them
2690 * being allocated from last increasing the chance that the last objects
2691 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002692 */
2693int kmem_cache_shrink(struct kmem_cache *s)
2694{
2695 int node;
2696 int i;
2697 struct kmem_cache_node *n;
2698 struct page *page;
2699 struct page *t;
2700 struct list_head *slabs_by_inuse =
2701 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2702 unsigned long flags;
2703
2704 if (!slabs_by_inuse)
2705 return -ENOMEM;
2706
2707 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002708 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002709 n = get_node(s, node);
2710
2711 if (!n->nr_partial)
2712 continue;
2713
2714 for (i = 0; i < s->objects; i++)
2715 INIT_LIST_HEAD(slabs_by_inuse + i);
2716
2717 spin_lock_irqsave(&n->list_lock, flags);
2718
2719 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002720 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002721 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002722 * Note that concurrent frees may occur while we hold the
2723 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002724 */
2725 list_for_each_entry_safe(page, t, &n->partial, lru) {
2726 if (!page->inuse && slab_trylock(page)) {
2727 /*
2728 * Must hold slab lock here because slab_free
2729 * may have freed the last object and be
2730 * waiting to release the slab.
2731 */
2732 list_del(&page->lru);
2733 n->nr_partial--;
2734 slab_unlock(page);
2735 discard_slab(s, page);
2736 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002737 list_move(&page->lru,
2738 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002739 }
2740 }
2741
Christoph Lameter2086d262007-05-06 14:49:46 -07002742 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002743 * Rebuild the partial list with the slabs filled up most
2744 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002745 */
2746 for (i = s->objects - 1; i >= 0; i--)
2747 list_splice(slabs_by_inuse + i, n->partial.prev);
2748
Christoph Lameter2086d262007-05-06 14:49:46 -07002749 spin_unlock_irqrestore(&n->list_lock, flags);
2750 }
2751
2752 kfree(slabs_by_inuse);
2753 return 0;
2754}
2755EXPORT_SYMBOL(kmem_cache_shrink);
2756
Yasunori Gotob9049e22007-10-21 16:41:37 -07002757#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2758static int slab_mem_going_offline_callback(void *arg)
2759{
2760 struct kmem_cache *s;
2761
2762 down_read(&slub_lock);
2763 list_for_each_entry(s, &slab_caches, list)
2764 kmem_cache_shrink(s);
2765 up_read(&slub_lock);
2766
2767 return 0;
2768}
2769
2770static void slab_mem_offline_callback(void *arg)
2771{
2772 struct kmem_cache_node *n;
2773 struct kmem_cache *s;
2774 struct memory_notify *marg = arg;
2775 int offline_node;
2776
2777 offline_node = marg->status_change_nid;
2778
2779 /*
2780 * If the node still has available memory. we need kmem_cache_node
2781 * for it yet.
2782 */
2783 if (offline_node < 0)
2784 return;
2785
2786 down_read(&slub_lock);
2787 list_for_each_entry(s, &slab_caches, list) {
2788 n = get_node(s, offline_node);
2789 if (n) {
2790 /*
2791 * if n->nr_slabs > 0, slabs still exist on the node
2792 * that is going down. We were unable to free them,
2793 * and offline_pages() function shoudn't call this
2794 * callback. So, we must fail.
2795 */
Al Viro27bb6282007-10-29 04:42:55 +00002796 BUG_ON(atomic_long_read(&n->nr_slabs));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002797
2798 s->node[offline_node] = NULL;
2799 kmem_cache_free(kmalloc_caches, n);
2800 }
2801 }
2802 up_read(&slub_lock);
2803}
2804
2805static int slab_mem_going_online_callback(void *arg)
2806{
2807 struct kmem_cache_node *n;
2808 struct kmem_cache *s;
2809 struct memory_notify *marg = arg;
2810 int nid = marg->status_change_nid;
2811 int ret = 0;
2812
2813 /*
2814 * If the node's memory is already available, then kmem_cache_node is
2815 * already created. Nothing to do.
2816 */
2817 if (nid < 0)
2818 return 0;
2819
2820 /*
2821 * We are bringing a node online. No memory is availabe yet. We must
2822 * allocate a kmem_cache_node structure in order to bring the node
2823 * online.
2824 */
2825 down_read(&slub_lock);
2826 list_for_each_entry(s, &slab_caches, list) {
2827 /*
2828 * XXX: kmem_cache_alloc_node will fallback to other nodes
2829 * since memory is not yet available from the node that
2830 * is brought up.
2831 */
2832 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2833 if (!n) {
2834 ret = -ENOMEM;
2835 goto out;
2836 }
2837 init_kmem_cache_node(n);
2838 s->node[nid] = n;
2839 }
2840out:
2841 up_read(&slub_lock);
2842 return ret;
2843}
2844
2845static int slab_memory_callback(struct notifier_block *self,
2846 unsigned long action, void *arg)
2847{
2848 int ret = 0;
2849
2850 switch (action) {
2851 case MEM_GOING_ONLINE:
2852 ret = slab_mem_going_online_callback(arg);
2853 break;
2854 case MEM_GOING_OFFLINE:
2855 ret = slab_mem_going_offline_callback(arg);
2856 break;
2857 case MEM_OFFLINE:
2858 case MEM_CANCEL_ONLINE:
2859 slab_mem_offline_callback(arg);
2860 break;
2861 case MEM_ONLINE:
2862 case MEM_CANCEL_OFFLINE:
2863 break;
2864 }
2865
2866 ret = notifier_from_errno(ret);
2867 return ret;
2868}
2869
2870#endif /* CONFIG_MEMORY_HOTPLUG */
2871
Christoph Lameter81819f02007-05-06 14:49:36 -07002872/********************************************************************
2873 * Basic setup of slabs
2874 *******************************************************************/
2875
2876void __init kmem_cache_init(void)
2877{
2878 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002879 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002880
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002881 init_alloc_cpu();
2882
Christoph Lameter81819f02007-05-06 14:49:36 -07002883#ifdef CONFIG_NUMA
2884 /*
2885 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002886 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002887 * kmem_cache_open for slab_state == DOWN.
2888 */
2889 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2890 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002891 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002892 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002893
2894 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002895#endif
2896
2897 /* Able to allocate the per node structures */
2898 slab_state = PARTIAL;
2899
2900 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002901 if (KMALLOC_MIN_SIZE <= 64) {
2902 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002903 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002904 caches++;
2905 }
2906 if (KMALLOC_MIN_SIZE <= 128) {
2907 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002908 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002909 caches++;
2910 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002911
Christoph Lameter331dc552008-02-14 14:28:09 -08002912 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002913 create_kmalloc_cache(&kmalloc_caches[i],
2914 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002915 caches++;
2916 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002917
Christoph Lameterf1b26332007-07-17 04:03:26 -07002918
2919 /*
2920 * Patch up the size_index table if we have strange large alignment
2921 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002922 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002923 *
2924 * Largest permitted alignment is 256 bytes due to the way we
2925 * handle the index determination for the smaller caches.
2926 *
2927 * Make sure that nothing crazy happens if someone starts tinkering
2928 * around with ARCH_KMALLOC_MINALIGN
2929 */
2930 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2931 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2932
Christoph Lameter12ad6842007-07-17 04:03:28 -07002933 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002934 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2935
Christoph Lameter81819f02007-05-06 14:49:36 -07002936 slab_state = UP;
2937
2938 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08002939 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002940 kmalloc_caches[i]. name =
2941 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2942
2943#ifdef CONFIG_SMP
2944 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002945 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2946 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2947#else
2948 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002949#endif
2950
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002951 printk(KERN_INFO
2952 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002953 " CPUs=%d, Nodes=%d\n",
2954 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002955 slub_min_order, slub_max_order, slub_min_objects,
2956 nr_cpu_ids, nr_node_ids);
2957}
2958
2959/*
2960 * Find a mergeable slab cache
2961 */
2962static int slab_unmergeable(struct kmem_cache *s)
2963{
2964 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2965 return 1;
2966
Christoph Lameter331dc552008-02-14 14:28:09 -08002967 if ((s->flags & __PAGE_ALLOC_FALLBACK))
Christoph Lameter71c7a062008-02-14 14:28:01 -08002968 return 1;
2969
Christoph Lameterc59def92007-05-16 22:10:50 -07002970 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002971 return 1;
2972
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002973 /*
2974 * We may have set a slab to be unmergeable during bootstrap.
2975 */
2976 if (s->refcount < 0)
2977 return 1;
2978
Christoph Lameter81819f02007-05-06 14:49:36 -07002979 return 0;
2980}
2981
2982static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07002983 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002984 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002985{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002986 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002987
2988 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2989 return NULL;
2990
Christoph Lameterc59def92007-05-16 22:10:50 -07002991 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002992 return NULL;
2993
2994 size = ALIGN(size, sizeof(void *));
2995 align = calculate_alignment(flags, align, size);
2996 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07002997 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002998
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002999 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003000 if (slab_unmergeable(s))
3001 continue;
3002
3003 if (size > s->size)
3004 continue;
3005
Christoph Lameterba0268a2007-09-11 15:24:11 -07003006 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003007 continue;
3008 /*
3009 * Check if alignment is compatible.
3010 * Courtesy of Adrian Drzewiecki
3011 */
Pekka Enberg06428782008-01-07 23:20:27 -08003012 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003013 continue;
3014
3015 if (s->size - size >= sizeof(void *))
3016 continue;
3017
3018 return s;
3019 }
3020 return NULL;
3021}
3022
3023struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3024 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003025 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003026{
3027 struct kmem_cache *s;
3028
3029 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003030 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003031 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003032 int cpu;
3033
Christoph Lameter81819f02007-05-06 14:49:36 -07003034 s->refcount++;
3035 /*
3036 * Adjust the object sizes so that we clear
3037 * the complete object on kzalloc.
3038 */
3039 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003040
3041 /*
3042 * And then we need to update the object size in the
3043 * per cpu structures
3044 */
3045 for_each_online_cpu(cpu)
3046 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003047
Christoph Lameter81819f02007-05-06 14:49:36 -07003048 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003049 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003050
Christoph Lameter81819f02007-05-06 14:49:36 -07003051 if (sysfs_slab_alias(s, name))
3052 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003053 return s;
3054 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003055
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003056 s = kmalloc(kmem_size, GFP_KERNEL);
3057 if (s) {
3058 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003059 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003060 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003061 up_write(&slub_lock);
3062 if (sysfs_slab_add(s))
3063 goto err;
3064 return s;
3065 }
3066 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003067 }
3068 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003069
3070err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003071 if (flags & SLAB_PANIC)
3072 panic("Cannot create slabcache %s\n", name);
3073 else
3074 s = NULL;
3075 return s;
3076}
3077EXPORT_SYMBOL(kmem_cache_create);
3078
Christoph Lameter81819f02007-05-06 14:49:36 -07003079#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003080/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003081 * Use the cpu notifier to insure that the cpu slabs are flushed when
3082 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003083 */
3084static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3085 unsigned long action, void *hcpu)
3086{
3087 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003088 struct kmem_cache *s;
3089 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003090
3091 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003092 case CPU_UP_PREPARE:
3093 case CPU_UP_PREPARE_FROZEN:
3094 init_alloc_cpu_cpu(cpu);
3095 down_read(&slub_lock);
3096 list_for_each_entry(s, &slab_caches, list)
3097 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3098 GFP_KERNEL);
3099 up_read(&slub_lock);
3100 break;
3101
Christoph Lameter81819f02007-05-06 14:49:36 -07003102 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003103 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003104 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003105 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003106 down_read(&slub_lock);
3107 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003108 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3109
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003110 local_irq_save(flags);
3111 __flush_cpu_slab(s, cpu);
3112 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003113 free_kmem_cache_cpu(c, cpu);
3114 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003115 }
3116 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003117 break;
3118 default:
3119 break;
3120 }
3121 return NOTIFY_OK;
3122}
3123
Pekka Enberg06428782008-01-07 23:20:27 -08003124static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003125 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003126};
Christoph Lameter81819f02007-05-06 14:49:36 -07003127
3128#endif
3129
Christoph Lameter81819f02007-05-06 14:49:36 -07003130void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3131{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003132 struct kmem_cache *s;
3133
Christoph Lameter331dc552008-02-14 14:28:09 -08003134 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003135 return kmalloc_large(size, gfpflags);
3136
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003137 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003138
Satyam Sharma2408c552007-10-16 01:24:44 -07003139 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003140 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003141
Christoph Lameterce15fea2007-07-17 04:03:28 -07003142 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003143}
3144
3145void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3146 int node, void *caller)
3147{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003148 struct kmem_cache *s;
3149
Christoph Lameter331dc552008-02-14 14:28:09 -08003150 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003151 return kmalloc_large(size, gfpflags);
3152
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003153 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003154
Satyam Sharma2408c552007-10-16 01:24:44 -07003155 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003156 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003157
Christoph Lameterce15fea2007-07-17 04:03:28 -07003158 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003159}
3160
Christoph Lameter41ecc552007-05-09 02:32:44 -07003161#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003162static int validate_slab(struct kmem_cache *s, struct page *page,
3163 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003164{
3165 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003166 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003167
3168 if (!check_slab(s, page) ||
3169 !on_freelist(s, page, NULL))
3170 return 0;
3171
3172 /* Now we know that a valid freelist exists */
3173 bitmap_zero(map, s->objects);
3174
Christoph Lameter7656c722007-05-09 02:32:40 -07003175 for_each_free_object(p, s, page->freelist) {
3176 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003177 if (!check_object(s, page, p, 0))
3178 return 0;
3179 }
3180
Christoph Lameter7656c722007-05-09 02:32:40 -07003181 for_each_object(p, s, addr)
3182 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003183 if (!check_object(s, page, p, 1))
3184 return 0;
3185 return 1;
3186}
3187
Christoph Lameter434e2452007-07-17 04:03:30 -07003188static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3189 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003190{
3191 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003192 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003193 slab_unlock(page);
3194 } else
3195 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3196 s->name, page);
3197
3198 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003199 if (!SlabDebug(page))
3200 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003201 "on slab 0x%p\n", s->name, page);
3202 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003203 if (SlabDebug(page))
3204 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003205 "slab 0x%p\n", s->name, page);
3206 }
3207}
3208
Christoph Lameter434e2452007-07-17 04:03:30 -07003209static int validate_slab_node(struct kmem_cache *s,
3210 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003211{
3212 unsigned long count = 0;
3213 struct page *page;
3214 unsigned long flags;
3215
3216 spin_lock_irqsave(&n->list_lock, flags);
3217
3218 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003219 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003220 count++;
3221 }
3222 if (count != n->nr_partial)
3223 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3224 "counter=%ld\n", s->name, count, n->nr_partial);
3225
3226 if (!(s->flags & SLAB_STORE_USER))
3227 goto out;
3228
3229 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003230 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003231 count++;
3232 }
3233 if (count != atomic_long_read(&n->nr_slabs))
3234 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3235 "counter=%ld\n", s->name, count,
3236 atomic_long_read(&n->nr_slabs));
3237
3238out:
3239 spin_unlock_irqrestore(&n->list_lock, flags);
3240 return count;
3241}
3242
Christoph Lameter434e2452007-07-17 04:03:30 -07003243static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003244{
3245 int node;
3246 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003247 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3248 sizeof(unsigned long), GFP_KERNEL);
3249
3250 if (!map)
3251 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003252
3253 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003254 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003255 struct kmem_cache_node *n = get_node(s, node);
3256
Christoph Lameter434e2452007-07-17 04:03:30 -07003257 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003258 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003259 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003260 return count;
3261}
3262
Christoph Lameterb3459702007-05-09 02:32:41 -07003263#ifdef SLUB_RESILIENCY_TEST
3264static void resiliency_test(void)
3265{
3266 u8 *p;
3267
3268 printk(KERN_ERR "SLUB resiliency testing\n");
3269 printk(KERN_ERR "-----------------------\n");
3270 printk(KERN_ERR "A. Corruption after allocation\n");
3271
3272 p = kzalloc(16, GFP_KERNEL);
3273 p[16] = 0x12;
3274 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3275 " 0x12->0x%p\n\n", p + 16);
3276
3277 validate_slab_cache(kmalloc_caches + 4);
3278
3279 /* Hmmm... The next two are dangerous */
3280 p = kzalloc(32, GFP_KERNEL);
3281 p[32 + sizeof(void *)] = 0x34;
3282 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003283 " 0x34 -> -0x%p\n", p);
3284 printk(KERN_ERR
3285 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003286
3287 validate_slab_cache(kmalloc_caches + 5);
3288 p = kzalloc(64, GFP_KERNEL);
3289 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3290 *p = 0x56;
3291 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3292 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003293 printk(KERN_ERR
3294 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003295 validate_slab_cache(kmalloc_caches + 6);
3296
3297 printk(KERN_ERR "\nB. Corruption after free\n");
3298 p = kzalloc(128, GFP_KERNEL);
3299 kfree(p);
3300 *p = 0x78;
3301 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3302 validate_slab_cache(kmalloc_caches + 7);
3303
3304 p = kzalloc(256, GFP_KERNEL);
3305 kfree(p);
3306 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003307 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3308 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003309 validate_slab_cache(kmalloc_caches + 8);
3310
3311 p = kzalloc(512, GFP_KERNEL);
3312 kfree(p);
3313 p[512] = 0xab;
3314 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3315 validate_slab_cache(kmalloc_caches + 9);
3316}
3317#else
3318static void resiliency_test(void) {};
3319#endif
3320
Christoph Lameter88a420e2007-05-06 14:49:45 -07003321/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003322 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003323 * and freed.
3324 */
3325
3326struct location {
3327 unsigned long count;
3328 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003329 long long sum_time;
3330 long min_time;
3331 long max_time;
3332 long min_pid;
3333 long max_pid;
3334 cpumask_t cpus;
3335 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003336};
3337
3338struct loc_track {
3339 unsigned long max;
3340 unsigned long count;
3341 struct location *loc;
3342};
3343
3344static void free_loc_track(struct loc_track *t)
3345{
3346 if (t->max)
3347 free_pages((unsigned long)t->loc,
3348 get_order(sizeof(struct location) * t->max));
3349}
3350
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003351static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003352{
3353 struct location *l;
3354 int order;
3355
Christoph Lameter88a420e2007-05-06 14:49:45 -07003356 order = get_order(sizeof(struct location) * max);
3357
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003358 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003359 if (!l)
3360 return 0;
3361
3362 if (t->count) {
3363 memcpy(l, t->loc, sizeof(struct location) * t->count);
3364 free_loc_track(t);
3365 }
3366 t->max = max;
3367 t->loc = l;
3368 return 1;
3369}
3370
3371static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003372 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003373{
3374 long start, end, pos;
3375 struct location *l;
3376 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003377 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003378
3379 start = -1;
3380 end = t->count;
3381
3382 for ( ; ; ) {
3383 pos = start + (end - start + 1) / 2;
3384
3385 /*
3386 * There is nothing at "end". If we end up there
3387 * we need to add something to before end.
3388 */
3389 if (pos == end)
3390 break;
3391
3392 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003393 if (track->addr == caddr) {
3394
3395 l = &t->loc[pos];
3396 l->count++;
3397 if (track->when) {
3398 l->sum_time += age;
3399 if (age < l->min_time)
3400 l->min_time = age;
3401 if (age > l->max_time)
3402 l->max_time = age;
3403
3404 if (track->pid < l->min_pid)
3405 l->min_pid = track->pid;
3406 if (track->pid > l->max_pid)
3407 l->max_pid = track->pid;
3408
3409 cpu_set(track->cpu, l->cpus);
3410 }
3411 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003412 return 1;
3413 }
3414
Christoph Lameter45edfa52007-05-09 02:32:45 -07003415 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003416 end = pos;
3417 else
3418 start = pos;
3419 }
3420
3421 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003422 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003423 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003424 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003425 return 0;
3426
3427 l = t->loc + pos;
3428 if (pos < t->count)
3429 memmove(l + 1, l,
3430 (t->count - pos) * sizeof(struct location));
3431 t->count++;
3432 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003433 l->addr = track->addr;
3434 l->sum_time = age;
3435 l->min_time = age;
3436 l->max_time = age;
3437 l->min_pid = track->pid;
3438 l->max_pid = track->pid;
3439 cpus_clear(l->cpus);
3440 cpu_set(track->cpu, l->cpus);
3441 nodes_clear(l->nodes);
3442 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003443 return 1;
3444}
3445
3446static void process_slab(struct loc_track *t, struct kmem_cache *s,
3447 struct page *page, enum track_item alloc)
3448{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003449 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003450 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003451 void *p;
3452
3453 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003454 for_each_free_object(p, s, page->freelist)
3455 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003456
Christoph Lameter7656c722007-05-09 02:32:40 -07003457 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003458 if (!test_bit(slab_index(p, s, addr), map))
3459 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003460}
3461
3462static int list_locations(struct kmem_cache *s, char *buf,
3463 enum track_item alloc)
3464{
Harvey Harrisone374d482008-01-31 15:20:50 -08003465 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003466 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003467 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003468 int node;
3469
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003470 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003471 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003472 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003473
3474 /* Push back cpu slabs */
3475 flush_all(s);
3476
Christoph Lameterf64dc582007-10-16 01:25:33 -07003477 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003478 struct kmem_cache_node *n = get_node(s, node);
3479 unsigned long flags;
3480 struct page *page;
3481
Christoph Lameter9e869432007-08-22 14:01:56 -07003482 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003483 continue;
3484
3485 spin_lock_irqsave(&n->list_lock, flags);
3486 list_for_each_entry(page, &n->partial, lru)
3487 process_slab(&t, s, page, alloc);
3488 list_for_each_entry(page, &n->full, lru)
3489 process_slab(&t, s, page, alloc);
3490 spin_unlock_irqrestore(&n->list_lock, flags);
3491 }
3492
3493 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003494 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003495
Harvey Harrisone374d482008-01-31 15:20:50 -08003496 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003497 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003498 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003499
3500 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003501 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003502 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003503 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003504
3505 if (l->sum_time != l->min_time) {
3506 unsigned long remainder;
3507
Harvey Harrisone374d482008-01-31 15:20:50 -08003508 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003509 l->min_time,
3510 div_long_long_rem(l->sum_time, l->count, &remainder),
3511 l->max_time);
3512 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003513 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003514 l->min_time);
3515
3516 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003517 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003518 l->min_pid, l->max_pid);
3519 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003520 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003521 l->min_pid);
3522
Christoph Lameter84966342007-06-23 17:16:32 -07003523 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003524 len < PAGE_SIZE - 60) {
3525 len += sprintf(buf + len, " cpus=");
3526 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003527 l->cpus);
3528 }
3529
Christoph Lameter84966342007-06-23 17:16:32 -07003530 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003531 len < PAGE_SIZE - 60) {
3532 len += sprintf(buf + len, " nodes=");
3533 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003534 l->nodes);
3535 }
3536
Harvey Harrisone374d482008-01-31 15:20:50 -08003537 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003538 }
3539
3540 free_loc_track(&t);
3541 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003542 len += sprintf(buf, "No data\n");
3543 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003544}
3545
Christoph Lameter81819f02007-05-06 14:49:36 -07003546enum slab_stat_type {
3547 SL_FULL,
3548 SL_PARTIAL,
3549 SL_CPU,
3550 SL_OBJECTS
3551};
3552
3553#define SO_FULL (1 << SL_FULL)
3554#define SO_PARTIAL (1 << SL_PARTIAL)
3555#define SO_CPU (1 << SL_CPU)
3556#define SO_OBJECTS (1 << SL_OBJECTS)
3557
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003558static unsigned long show_slab_objects(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07003559 char *buf, unsigned long flags)
3560{
3561 unsigned long total = 0;
3562 int cpu;
3563 int node;
3564 int x;
3565 unsigned long *nodes;
3566 unsigned long *per_cpu;
3567
3568 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3569 per_cpu = nodes + nr_node_ids;
3570
3571 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003572 struct page *page;
3573 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003574
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003575 if (!c)
3576 continue;
3577
3578 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003579 node = c->node;
3580 if (node < 0)
3581 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003582 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003583 if (flags & SO_CPU) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003584 if (flags & SO_OBJECTS)
3585 x = page->inuse;
3586 else
3587 x = 1;
3588 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003589 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003590 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003591 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003592 }
3593 }
3594
Christoph Lameterf64dc582007-10-16 01:25:33 -07003595 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003596 struct kmem_cache_node *n = get_node(s, node);
3597
3598 if (flags & SO_PARTIAL) {
3599 if (flags & SO_OBJECTS)
3600 x = count_partial(n);
3601 else
3602 x = n->nr_partial;
3603 total += x;
3604 nodes[node] += x;
3605 }
3606
3607 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003608 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003609 - per_cpu[node]
3610 - n->nr_partial;
3611
3612 if (flags & SO_OBJECTS)
3613 x = full_slabs * s->objects;
3614 else
3615 x = full_slabs;
3616 total += x;
3617 nodes[node] += x;
3618 }
3619 }
3620
3621 x = sprintf(buf, "%lu", total);
3622#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003623 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003624 if (nodes[node])
3625 x += sprintf(buf + x, " N%d=%lu",
3626 node, nodes[node]);
3627#endif
3628 kfree(nodes);
3629 return x + sprintf(buf + x, "\n");
3630}
3631
3632static int any_slab_objects(struct kmem_cache *s)
3633{
3634 int node;
3635 int cpu;
3636
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003637 for_each_possible_cpu(cpu) {
3638 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003639
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003640 if (c && c->page)
3641 return 1;
3642 }
3643
3644 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003645 struct kmem_cache_node *n = get_node(s, node);
3646
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003647 if (!n)
3648 continue;
3649
Christoph Lameter9e869432007-08-22 14:01:56 -07003650 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003651 return 1;
3652 }
3653 return 0;
3654}
3655
3656#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3657#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3658
3659struct slab_attribute {
3660 struct attribute attr;
3661 ssize_t (*show)(struct kmem_cache *s, char *buf);
3662 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3663};
3664
3665#define SLAB_ATTR_RO(_name) \
3666 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3667
3668#define SLAB_ATTR(_name) \
3669 static struct slab_attribute _name##_attr = \
3670 __ATTR(_name, 0644, _name##_show, _name##_store)
3671
Christoph Lameter81819f02007-05-06 14:49:36 -07003672static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3673{
3674 return sprintf(buf, "%d\n", s->size);
3675}
3676SLAB_ATTR_RO(slab_size);
3677
3678static ssize_t align_show(struct kmem_cache *s, char *buf)
3679{
3680 return sprintf(buf, "%d\n", s->align);
3681}
3682SLAB_ATTR_RO(align);
3683
3684static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3685{
3686 return sprintf(buf, "%d\n", s->objsize);
3687}
3688SLAB_ATTR_RO(object_size);
3689
3690static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3691{
3692 return sprintf(buf, "%d\n", s->objects);
3693}
3694SLAB_ATTR_RO(objs_per_slab);
3695
3696static ssize_t order_show(struct kmem_cache *s, char *buf)
3697{
3698 return sprintf(buf, "%d\n", s->order);
3699}
3700SLAB_ATTR_RO(order);
3701
3702static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3703{
3704 if (s->ctor) {
3705 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3706
3707 return n + sprintf(buf + n, "\n");
3708 }
3709 return 0;
3710}
3711SLAB_ATTR_RO(ctor);
3712
Christoph Lameter81819f02007-05-06 14:49:36 -07003713static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3714{
3715 return sprintf(buf, "%d\n", s->refcount - 1);
3716}
3717SLAB_ATTR_RO(aliases);
3718
3719static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3720{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003721 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003722}
3723SLAB_ATTR_RO(slabs);
3724
3725static ssize_t partial_show(struct kmem_cache *s, char *buf)
3726{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003727 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003728}
3729SLAB_ATTR_RO(partial);
3730
3731static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3732{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003733 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003734}
3735SLAB_ATTR_RO(cpu_slabs);
3736
3737static ssize_t objects_show(struct kmem_cache *s, char *buf)
3738{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003739 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003740}
3741SLAB_ATTR_RO(objects);
3742
3743static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3744{
3745 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3746}
3747
3748static ssize_t sanity_checks_store(struct kmem_cache *s,
3749 const char *buf, size_t length)
3750{
3751 s->flags &= ~SLAB_DEBUG_FREE;
3752 if (buf[0] == '1')
3753 s->flags |= SLAB_DEBUG_FREE;
3754 return length;
3755}
3756SLAB_ATTR(sanity_checks);
3757
3758static ssize_t trace_show(struct kmem_cache *s, char *buf)
3759{
3760 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3761}
3762
3763static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3764 size_t length)
3765{
3766 s->flags &= ~SLAB_TRACE;
3767 if (buf[0] == '1')
3768 s->flags |= SLAB_TRACE;
3769 return length;
3770}
3771SLAB_ATTR(trace);
3772
3773static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3774{
3775 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3776}
3777
3778static ssize_t reclaim_account_store(struct kmem_cache *s,
3779 const char *buf, size_t length)
3780{
3781 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3782 if (buf[0] == '1')
3783 s->flags |= SLAB_RECLAIM_ACCOUNT;
3784 return length;
3785}
3786SLAB_ATTR(reclaim_account);
3787
3788static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3789{
Christoph Lameter5af60832007-05-06 14:49:56 -07003790 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003791}
3792SLAB_ATTR_RO(hwcache_align);
3793
3794#ifdef CONFIG_ZONE_DMA
3795static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3796{
3797 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3798}
3799SLAB_ATTR_RO(cache_dma);
3800#endif
3801
3802static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3803{
3804 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3805}
3806SLAB_ATTR_RO(destroy_by_rcu);
3807
3808static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3809{
3810 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3811}
3812
3813static ssize_t red_zone_store(struct kmem_cache *s,
3814 const char *buf, size_t length)
3815{
3816 if (any_slab_objects(s))
3817 return -EBUSY;
3818
3819 s->flags &= ~SLAB_RED_ZONE;
3820 if (buf[0] == '1')
3821 s->flags |= SLAB_RED_ZONE;
3822 calculate_sizes(s);
3823 return length;
3824}
3825SLAB_ATTR(red_zone);
3826
3827static ssize_t poison_show(struct kmem_cache *s, char *buf)
3828{
3829 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3830}
3831
3832static ssize_t poison_store(struct kmem_cache *s,
3833 const char *buf, size_t length)
3834{
3835 if (any_slab_objects(s))
3836 return -EBUSY;
3837
3838 s->flags &= ~SLAB_POISON;
3839 if (buf[0] == '1')
3840 s->flags |= SLAB_POISON;
3841 calculate_sizes(s);
3842 return length;
3843}
3844SLAB_ATTR(poison);
3845
3846static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3847{
3848 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3849}
3850
3851static ssize_t store_user_store(struct kmem_cache *s,
3852 const char *buf, size_t length)
3853{
3854 if (any_slab_objects(s))
3855 return -EBUSY;
3856
3857 s->flags &= ~SLAB_STORE_USER;
3858 if (buf[0] == '1')
3859 s->flags |= SLAB_STORE_USER;
3860 calculate_sizes(s);
3861 return length;
3862}
3863SLAB_ATTR(store_user);
3864
Christoph Lameter53e15af2007-05-06 14:49:43 -07003865static ssize_t validate_show(struct kmem_cache *s, char *buf)
3866{
3867 return 0;
3868}
3869
3870static ssize_t validate_store(struct kmem_cache *s,
3871 const char *buf, size_t length)
3872{
Christoph Lameter434e2452007-07-17 04:03:30 -07003873 int ret = -EINVAL;
3874
3875 if (buf[0] == '1') {
3876 ret = validate_slab_cache(s);
3877 if (ret >= 0)
3878 ret = length;
3879 }
3880 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003881}
3882SLAB_ATTR(validate);
3883
Christoph Lameter2086d262007-05-06 14:49:46 -07003884static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3885{
3886 return 0;
3887}
3888
3889static ssize_t shrink_store(struct kmem_cache *s,
3890 const char *buf, size_t length)
3891{
3892 if (buf[0] == '1') {
3893 int rc = kmem_cache_shrink(s);
3894
3895 if (rc)
3896 return rc;
3897 } else
3898 return -EINVAL;
3899 return length;
3900}
3901SLAB_ATTR(shrink);
3902
Christoph Lameter88a420e2007-05-06 14:49:45 -07003903static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3904{
3905 if (!(s->flags & SLAB_STORE_USER))
3906 return -ENOSYS;
3907 return list_locations(s, buf, TRACK_ALLOC);
3908}
3909SLAB_ATTR_RO(alloc_calls);
3910
3911static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3912{
3913 if (!(s->flags & SLAB_STORE_USER))
3914 return -ENOSYS;
3915 return list_locations(s, buf, TRACK_FREE);
3916}
3917SLAB_ATTR_RO(free_calls);
3918
Christoph Lameter81819f02007-05-06 14:49:36 -07003919#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08003920static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07003921{
Christoph Lameter98246012008-01-07 23:20:26 -08003922 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07003923}
3924
Christoph Lameter98246012008-01-07 23:20:26 -08003925static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07003926 const char *buf, size_t length)
3927{
3928 int n = simple_strtoul(buf, NULL, 10);
3929
3930 if (n < 100)
Christoph Lameter98246012008-01-07 23:20:26 -08003931 s->remote_node_defrag_ratio = n * 10;
Christoph Lameter81819f02007-05-06 14:49:36 -07003932 return length;
3933}
Christoph Lameter98246012008-01-07 23:20:26 -08003934SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07003935#endif
3936
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003937#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003938static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3939{
3940 unsigned long sum = 0;
3941 int cpu;
3942 int len;
3943 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3944
3945 if (!data)
3946 return -ENOMEM;
3947
3948 for_each_online_cpu(cpu) {
3949 unsigned x = get_cpu_slab(s, cpu)->stat[si];
3950
3951 data[cpu] = x;
3952 sum += x;
3953 }
3954
3955 len = sprintf(buf, "%lu", sum);
3956
3957 for_each_online_cpu(cpu) {
3958 if (data[cpu] && len < PAGE_SIZE - 20)
3959 len += sprintf(buf + len, " c%d=%u", cpu, data[cpu]);
3960 }
3961 kfree(data);
3962 return len + sprintf(buf + len, "\n");
3963}
3964
3965#define STAT_ATTR(si, text) \
3966static ssize_t text##_show(struct kmem_cache *s, char *buf) \
3967{ \
3968 return show_stat(s, buf, si); \
3969} \
3970SLAB_ATTR_RO(text); \
3971
3972STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
3973STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
3974STAT_ATTR(FREE_FASTPATH, free_fastpath);
3975STAT_ATTR(FREE_SLOWPATH, free_slowpath);
3976STAT_ATTR(FREE_FROZEN, free_frozen);
3977STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
3978STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
3979STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
3980STAT_ATTR(ALLOC_SLAB, alloc_slab);
3981STAT_ATTR(ALLOC_REFILL, alloc_refill);
3982STAT_ATTR(FREE_SLAB, free_slab);
3983STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
3984STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
3985STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
3986STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
3987STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
3988STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
3989
3990#endif
3991
Pekka Enberg06428782008-01-07 23:20:27 -08003992static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07003993 &slab_size_attr.attr,
3994 &object_size_attr.attr,
3995 &objs_per_slab_attr.attr,
3996 &order_attr.attr,
3997 &objects_attr.attr,
3998 &slabs_attr.attr,
3999 &partial_attr.attr,
4000 &cpu_slabs_attr.attr,
4001 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004002 &aliases_attr.attr,
4003 &align_attr.attr,
4004 &sanity_checks_attr.attr,
4005 &trace_attr.attr,
4006 &hwcache_align_attr.attr,
4007 &reclaim_account_attr.attr,
4008 &destroy_by_rcu_attr.attr,
4009 &red_zone_attr.attr,
4010 &poison_attr.attr,
4011 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004012 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004013 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004014 &alloc_calls_attr.attr,
4015 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004016#ifdef CONFIG_ZONE_DMA
4017 &cache_dma_attr.attr,
4018#endif
4019#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004020 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004021#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004022#ifdef CONFIG_SLUB_STATS
4023 &alloc_fastpath_attr.attr,
4024 &alloc_slowpath_attr.attr,
4025 &free_fastpath_attr.attr,
4026 &free_slowpath_attr.attr,
4027 &free_frozen_attr.attr,
4028 &free_add_partial_attr.attr,
4029 &free_remove_partial_attr.attr,
4030 &alloc_from_partial_attr.attr,
4031 &alloc_slab_attr.attr,
4032 &alloc_refill_attr.attr,
4033 &free_slab_attr.attr,
4034 &cpuslab_flush_attr.attr,
4035 &deactivate_full_attr.attr,
4036 &deactivate_empty_attr.attr,
4037 &deactivate_to_head_attr.attr,
4038 &deactivate_to_tail_attr.attr,
4039 &deactivate_remote_frees_attr.attr,
4040#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004041 NULL
4042};
4043
4044static struct attribute_group slab_attr_group = {
4045 .attrs = slab_attrs,
4046};
4047
4048static ssize_t slab_attr_show(struct kobject *kobj,
4049 struct attribute *attr,
4050 char *buf)
4051{
4052 struct slab_attribute *attribute;
4053 struct kmem_cache *s;
4054 int err;
4055
4056 attribute = to_slab_attr(attr);
4057 s = to_slab(kobj);
4058
4059 if (!attribute->show)
4060 return -EIO;
4061
4062 err = attribute->show(s, buf);
4063
4064 return err;
4065}
4066
4067static ssize_t slab_attr_store(struct kobject *kobj,
4068 struct attribute *attr,
4069 const char *buf, size_t len)
4070{
4071 struct slab_attribute *attribute;
4072 struct kmem_cache *s;
4073 int err;
4074
4075 attribute = to_slab_attr(attr);
4076 s = to_slab(kobj);
4077
4078 if (!attribute->store)
4079 return -EIO;
4080
4081 err = attribute->store(s, buf, len);
4082
4083 return err;
4084}
4085
Christoph Lameter151c6022008-01-07 22:29:05 -08004086static void kmem_cache_release(struct kobject *kobj)
4087{
4088 struct kmem_cache *s = to_slab(kobj);
4089
4090 kfree(s);
4091}
4092
Christoph Lameter81819f02007-05-06 14:49:36 -07004093static struct sysfs_ops slab_sysfs_ops = {
4094 .show = slab_attr_show,
4095 .store = slab_attr_store,
4096};
4097
4098static struct kobj_type slab_ktype = {
4099 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004100 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004101};
4102
4103static int uevent_filter(struct kset *kset, struct kobject *kobj)
4104{
4105 struct kobj_type *ktype = get_ktype(kobj);
4106
4107 if (ktype == &slab_ktype)
4108 return 1;
4109 return 0;
4110}
4111
4112static struct kset_uevent_ops slab_uevent_ops = {
4113 .filter = uevent_filter,
4114};
4115
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004116static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004117
4118#define ID_STR_LENGTH 64
4119
4120/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004121 *
4122 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004123 */
4124static char *create_unique_id(struct kmem_cache *s)
4125{
4126 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4127 char *p = name;
4128
4129 BUG_ON(!name);
4130
4131 *p++ = ':';
4132 /*
4133 * First flags affecting slabcache operations. We will only
4134 * get here for aliasable slabs so we do not need to support
4135 * too many flags. The flags here must cover all flags that
4136 * are matched during merging to guarantee that the id is
4137 * unique.
4138 */
4139 if (s->flags & SLAB_CACHE_DMA)
4140 *p++ = 'd';
4141 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4142 *p++ = 'a';
4143 if (s->flags & SLAB_DEBUG_FREE)
4144 *p++ = 'F';
4145 if (p != name + 1)
4146 *p++ = '-';
4147 p += sprintf(p, "%07d", s->size);
4148 BUG_ON(p > name + ID_STR_LENGTH - 1);
4149 return name;
4150}
4151
4152static int sysfs_slab_add(struct kmem_cache *s)
4153{
4154 int err;
4155 const char *name;
4156 int unmergeable;
4157
4158 if (slab_state < SYSFS)
4159 /* Defer until later */
4160 return 0;
4161
4162 unmergeable = slab_unmergeable(s);
4163 if (unmergeable) {
4164 /*
4165 * Slabcache can never be merged so we can use the name proper.
4166 * This is typically the case for debug situations. In that
4167 * case we can catch duplicate names easily.
4168 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004169 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004170 name = s->name;
4171 } else {
4172 /*
4173 * Create a unique name for the slab as a target
4174 * for the symlinks.
4175 */
4176 name = create_unique_id(s);
4177 }
4178
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004179 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004180 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4181 if (err) {
4182 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004183 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004184 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004185
4186 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4187 if (err)
4188 return err;
4189 kobject_uevent(&s->kobj, KOBJ_ADD);
4190 if (!unmergeable) {
4191 /* Setup first alias */
4192 sysfs_slab_alias(s, s->name);
4193 kfree(name);
4194 }
4195 return 0;
4196}
4197
4198static void sysfs_slab_remove(struct kmem_cache *s)
4199{
4200 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4201 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004202 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004203}
4204
4205/*
4206 * Need to buffer aliases during bootup until sysfs becomes
4207 * available lest we loose that information.
4208 */
4209struct saved_alias {
4210 struct kmem_cache *s;
4211 const char *name;
4212 struct saved_alias *next;
4213};
4214
Adrian Bunk5af328a2007-07-17 04:03:27 -07004215static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004216
4217static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4218{
4219 struct saved_alias *al;
4220
4221 if (slab_state == SYSFS) {
4222 /*
4223 * If we have a leftover link then remove it.
4224 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004225 sysfs_remove_link(&slab_kset->kobj, name);
4226 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004227 }
4228
4229 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4230 if (!al)
4231 return -ENOMEM;
4232
4233 al->s = s;
4234 al->name = name;
4235 al->next = alias_list;
4236 alias_list = al;
4237 return 0;
4238}
4239
4240static int __init slab_sysfs_init(void)
4241{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004242 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004243 int err;
4244
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004245 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004246 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004247 printk(KERN_ERR "Cannot register slab subsystem.\n");
4248 return -ENOSYS;
4249 }
4250
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004251 slab_state = SYSFS;
4252
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004253 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004254 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004255 if (err)
4256 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4257 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004258 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004259
4260 while (alias_list) {
4261 struct saved_alias *al = alias_list;
4262
4263 alias_list = alias_list->next;
4264 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004265 if (err)
4266 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4267 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004268 kfree(al);
4269 }
4270
4271 resiliency_test();
4272 return 0;
4273}
4274
4275__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004276#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004277
4278/*
4279 * The /proc/slabinfo ABI
4280 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004281#ifdef CONFIG_SLABINFO
4282
4283ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4284 size_t count, loff_t *ppos)
4285{
4286 return -EINVAL;
4287}
4288
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004289
4290static void print_slabinfo_header(struct seq_file *m)
4291{
4292 seq_puts(m, "slabinfo - version: 2.1\n");
4293 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4294 "<objperslab> <pagesperslab>");
4295 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4296 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4297 seq_putc(m, '\n');
4298}
4299
4300static void *s_start(struct seq_file *m, loff_t *pos)
4301{
4302 loff_t n = *pos;
4303
4304 down_read(&slub_lock);
4305 if (!n)
4306 print_slabinfo_header(m);
4307
4308 return seq_list_start(&slab_caches, *pos);
4309}
4310
4311static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4312{
4313 return seq_list_next(p, &slab_caches, pos);
4314}
4315
4316static void s_stop(struct seq_file *m, void *p)
4317{
4318 up_read(&slub_lock);
4319}
4320
4321static int s_show(struct seq_file *m, void *p)
4322{
4323 unsigned long nr_partials = 0;
4324 unsigned long nr_slabs = 0;
4325 unsigned long nr_inuse = 0;
4326 unsigned long nr_objs;
4327 struct kmem_cache *s;
4328 int node;
4329
4330 s = list_entry(p, struct kmem_cache, list);
4331
4332 for_each_online_node(node) {
4333 struct kmem_cache_node *n = get_node(s, node);
4334
4335 if (!n)
4336 continue;
4337
4338 nr_partials += n->nr_partial;
4339 nr_slabs += atomic_long_read(&n->nr_slabs);
4340 nr_inuse += count_partial(n);
4341 }
4342
4343 nr_objs = nr_slabs * s->objects;
4344 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4345
4346 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4347 nr_objs, s->size, s->objects, (1 << s->order));
4348 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4349 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4350 0UL);
4351 seq_putc(m, '\n');
4352 return 0;
4353}
4354
4355const struct seq_operations slabinfo_op = {
4356 .start = s_start,
4357 .next = s_next,
4358 .stop = s_stop,
4359 .show = s_show,
4360};
4361
Linus Torvalds158a9622008-01-02 13:04:48 -08004362#endif /* CONFIG_SLABINFO */