blob: 80d20cc1c0f8db0a2affcf2a16216616c7497082 [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 Lameter3dc50632008-04-23 12:28:01 -0700524 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700525 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 Lameter3dc50632008-04-23 12:28:01 -0700536 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700537 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 Lameter0f389ec2008-04-14 18:53:02 +0300840/* Tracking of the number of slabs for debugging purposes */
841static inline unsigned long slabs_node(struct kmem_cache *s, int node)
842{
843 struct kmem_cache_node *n = get_node(s, node);
844
845 return atomic_long_read(&n->nr_slabs);
846}
847
848static inline void inc_slabs_node(struct kmem_cache *s, int node)
849{
850 struct kmem_cache_node *n = get_node(s, node);
851
852 /*
853 * May be called early in order to allocate a slab for the
854 * kmem_cache_node structure. Solve the chicken-egg
855 * dilemma by deferring the increment of the count during
856 * bootstrap (see early_kmem_cache_node_alloc).
857 */
858 if (!NUMA_BUILD || n)
859 atomic_long_inc(&n->nr_slabs);
860}
861static inline void dec_slabs_node(struct kmem_cache *s, int node)
862{
863 struct kmem_cache_node *n = get_node(s, node);
864
865 atomic_long_dec(&n->nr_slabs);
866}
867
868/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -0700869static void setup_object_debug(struct kmem_cache *s, struct page *page,
870 void *object)
871{
872 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
873 return;
874
875 init_object(s, object, 0);
876 init_tracking(s, object);
877}
878
879static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
880 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700881{
882 if (!check_slab(s, page))
883 goto bad;
884
Christoph Lameterd692ef62008-02-15 23:45:24 -0800885 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700886 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700887 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700888 }
889
890 if (!check_valid_pointer(s, page, object)) {
891 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700892 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700893 }
894
Christoph Lameterd692ef62008-02-15 23:45:24 -0800895 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700896 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700897
Christoph Lameter3ec09742007-05-16 22:11:00 -0700898 /* Success perform special debug activities for allocs */
899 if (s->flags & SLAB_STORE_USER)
900 set_track(s, object, TRACK_ALLOC, addr);
901 trace(s, page, object, 1);
902 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700903 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700904
Christoph Lameter81819f02007-05-06 14:49:36 -0700905bad:
906 if (PageSlab(page)) {
907 /*
908 * If this is a slab page then lets do the best we can
909 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700910 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700911 */
Christoph Lameter24922682007-07-17 04:03:18 -0700912 slab_fix(s, "Marking all objects used");
Christoph Lameter81819f02007-05-06 14:49:36 -0700913 page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800914 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700915 }
916 return 0;
917}
918
Christoph Lameter3ec09742007-05-16 22:11:00 -0700919static int free_debug_processing(struct kmem_cache *s, struct page *page,
920 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700921{
922 if (!check_slab(s, page))
923 goto fail;
924
925 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700926 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700927 goto fail;
928 }
929
930 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700931 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700932 goto fail;
933 }
934
935 if (!check_object(s, page, object, 1))
936 return 0;
937
938 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800939 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700940 slab_err(s, page, "Attempt to free object(0x%p) "
941 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800942 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700943 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700944 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700945 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700946 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800947 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700948 object_err(s, page, object,
949 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700950 goto fail;
951 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700952
953 /* Special debug activities for freeing objects */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800954 if (!SlabFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700955 remove_full(s, page);
956 if (s->flags & SLAB_STORE_USER)
957 set_track(s, object, TRACK_FREE, addr);
958 trace(s, page, object, 0);
959 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700960 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700961
Christoph Lameter81819f02007-05-06 14:49:36 -0700962fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700963 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700964 return 0;
965}
966
Christoph Lameter41ecc552007-05-09 02:32:44 -0700967static int __init setup_slub_debug(char *str)
968{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700969 slub_debug = DEBUG_DEFAULT_FLAGS;
970 if (*str++ != '=' || !*str)
971 /*
972 * No options specified. Switch on full debugging.
973 */
974 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700975
976 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700977 /*
978 * No options but restriction on slabs. This means full
979 * debugging for slabs matching a pattern.
980 */
981 goto check_slabs;
982
983 slub_debug = 0;
984 if (*str == '-')
985 /*
986 * Switch off all debugging measures.
987 */
988 goto out;
989
990 /*
991 * Determine which debug features should be switched on
992 */
Pekka Enberg06428782008-01-07 23:20:27 -0800993 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700994 switch (tolower(*str)) {
995 case 'f':
996 slub_debug |= SLAB_DEBUG_FREE;
997 break;
998 case 'z':
999 slub_debug |= SLAB_RED_ZONE;
1000 break;
1001 case 'p':
1002 slub_debug |= SLAB_POISON;
1003 break;
1004 case 'u':
1005 slub_debug |= SLAB_STORE_USER;
1006 break;
1007 case 't':
1008 slub_debug |= SLAB_TRACE;
1009 break;
1010 default:
1011 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001012 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001013 }
1014 }
1015
1016check_slabs:
1017 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001018 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001019out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001020 return 1;
1021}
1022
1023__setup("slub_debug", setup_slub_debug);
1024
Christoph Lameterba0268a2007-09-11 15:24:11 -07001025static unsigned long kmem_cache_flags(unsigned long objsize,
1026 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001027 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001028{
1029 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001030 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001031 */
Christoph Lametere1533622008-02-15 23:45:24 -08001032 if (slub_debug && (!slub_debug_slabs ||
1033 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1034 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001035
1036 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001037}
1038#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001039static inline void setup_object_debug(struct kmem_cache *s,
1040 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001041
Christoph Lameter3ec09742007-05-16 22:11:00 -07001042static inline int alloc_debug_processing(struct kmem_cache *s,
1043 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001044
Christoph Lameter3ec09742007-05-16 22:11:00 -07001045static inline int free_debug_processing(struct kmem_cache *s,
1046 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001047
Christoph Lameter41ecc552007-05-09 02:32:44 -07001048static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1049 { return 1; }
1050static inline int check_object(struct kmem_cache *s, struct page *page,
1051 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001052static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001053static inline unsigned long kmem_cache_flags(unsigned long objsize,
1054 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001055 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001056{
1057 return flags;
1058}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001059#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001060
1061static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1062 { return 0; }
1063static inline void inc_slabs_node(struct kmem_cache *s, int node) {}
1064static inline void dec_slabs_node(struct kmem_cache *s, int node) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001065#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001066/*
1067 * Slab allocation and freeing
1068 */
1069static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1070{
Pekka Enberg06428782008-01-07 23:20:27 -08001071 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001072 int pages = 1 << s->order;
1073
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001074 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001075
Christoph Lameter81819f02007-05-06 14:49:36 -07001076 if (node == -1)
1077 page = alloc_pages(flags, s->order);
1078 else
1079 page = alloc_pages_node(node, flags, s->order);
1080
1081 if (!page)
1082 return NULL;
1083
1084 mod_zone_page_state(page_zone(page),
1085 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1086 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1087 pages);
1088
1089 return page;
1090}
1091
1092static void setup_object(struct kmem_cache *s, struct page *page,
1093 void *object)
1094{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001095 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001096 if (unlikely(s->ctor))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001097 s->ctor(s, object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001098}
1099
1100static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1101{
1102 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001103 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001104 void *last;
1105 void *p;
1106
Christoph Lameter6cb06222007-10-16 01:25:41 -07001107 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001108
Christoph Lameter6cb06222007-10-16 01:25:41 -07001109 page = allocate_slab(s,
1110 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001111 if (!page)
1112 goto out;
1113
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001114 inc_slabs_node(s, page_to_nid(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001115 page->slab = s;
1116 page->flags |= 1 << PG_slab;
1117 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1118 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001119 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001120
1121 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001122
1123 if (unlikely(s->flags & SLAB_POISON))
1124 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1125
1126 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001127 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001128 setup_object(s, page, last);
1129 set_freepointer(s, last, p);
1130 last = p;
1131 }
1132 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001133 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001134
1135 page->freelist = start;
1136 page->inuse = 0;
1137out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001138 return page;
1139}
1140
1141static void __free_slab(struct kmem_cache *s, struct page *page)
1142{
1143 int pages = 1 << s->order;
1144
Christoph Lameterc59def92007-05-16 22:10:50 -07001145 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001146 void *p;
1147
1148 slab_pad_check(s, page);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001149 for_each_object(p, s, page_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001150 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001151 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001152 }
1153
1154 mod_zone_page_state(page_zone(page),
1155 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1156 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001157 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001158
Christoph Lameter49bd5222008-04-14 18:52:18 +03001159 __ClearPageSlab(page);
1160 reset_page_mapcount(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001161 __free_pages(page, s->order);
1162}
1163
1164static void rcu_free_slab(struct rcu_head *h)
1165{
1166 struct page *page;
1167
1168 page = container_of((struct list_head *)h, struct page, lru);
1169 __free_slab(page->slab, page);
1170}
1171
1172static void free_slab(struct kmem_cache *s, struct page *page)
1173{
1174 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1175 /*
1176 * RCU free overloads the RCU head over the LRU
1177 */
1178 struct rcu_head *head = (void *)&page->lru;
1179
1180 call_rcu(head, rcu_free_slab);
1181 } else
1182 __free_slab(s, page);
1183}
1184
1185static void discard_slab(struct kmem_cache *s, struct page *page)
1186{
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001187 dec_slabs_node(s, page_to_nid(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001188 free_slab(s, page);
1189}
1190
1191/*
1192 * Per slab locking using the pagelock
1193 */
1194static __always_inline void slab_lock(struct page *page)
1195{
1196 bit_spin_lock(PG_locked, &page->flags);
1197}
1198
1199static __always_inline void slab_unlock(struct page *page)
1200{
Nick Piggina76d3542008-01-07 23:20:27 -08001201 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001202}
1203
1204static __always_inline int slab_trylock(struct page *page)
1205{
1206 int rc = 1;
1207
1208 rc = bit_spin_trylock(PG_locked, &page->flags);
1209 return rc;
1210}
1211
1212/*
1213 * Management of partially allocated slabs
1214 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001215static void add_partial(struct kmem_cache_node *n,
1216 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001217{
Christoph Lametere95eed52007-05-06 14:49:44 -07001218 spin_lock(&n->list_lock);
1219 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001220 if (tail)
1221 list_add_tail(&page->lru, &n->partial);
1222 else
1223 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001224 spin_unlock(&n->list_lock);
1225}
1226
1227static void remove_partial(struct kmem_cache *s,
1228 struct page *page)
1229{
1230 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1231
1232 spin_lock(&n->list_lock);
1233 list_del(&page->lru);
1234 n->nr_partial--;
1235 spin_unlock(&n->list_lock);
1236}
1237
1238/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001239 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001240 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001241 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001242 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001243static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001244{
1245 if (slab_trylock(page)) {
1246 list_del(&page->lru);
1247 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001248 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001249 return 1;
1250 }
1251 return 0;
1252}
1253
1254/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001255 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001256 */
1257static struct page *get_partial_node(struct kmem_cache_node *n)
1258{
1259 struct page *page;
1260
1261 /*
1262 * Racy check. If we mistakenly see no partial slabs then we
1263 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001264 * partial slab and there is none available then get_partials()
1265 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001266 */
1267 if (!n || !n->nr_partial)
1268 return NULL;
1269
1270 spin_lock(&n->list_lock);
1271 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001272 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001273 goto out;
1274 page = NULL;
1275out:
1276 spin_unlock(&n->list_lock);
1277 return page;
1278}
1279
1280/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001281 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001282 */
1283static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1284{
1285#ifdef CONFIG_NUMA
1286 struct zonelist *zonelist;
1287 struct zone **z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001288 struct zone *zone;
1289 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001290 struct page *page;
1291
1292 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001293 * The defrag ratio allows a configuration of the tradeoffs between
1294 * inter node defragmentation and node local allocations. A lower
1295 * defrag_ratio increases the tendency to do local allocations
1296 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001297 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001298 * If the defrag_ratio is set to 0 then kmalloc() always
1299 * returns node local objects. If the ratio is higher then kmalloc()
1300 * may return off node objects because partial slabs are obtained
1301 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001302 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001303 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001304 * defrag_ratio = 1000) then every (well almost) allocation will
1305 * first attempt to defrag slab caches on other nodes. This means
1306 * scanning over all nodes to look for partial slabs which may be
1307 * expensive if we do it every time we are trying to find a slab
1308 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001309 */
Christoph Lameter98246012008-01-07 23:20:26 -08001310 if (!s->remote_node_defrag_ratio ||
1311 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001312 return NULL;
1313
Mel Gorman0e884602008-04-28 02:12:14 -07001314 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001315 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001316 struct kmem_cache_node *n;
1317
Mel Gorman54a6eb52008-04-28 02:12:16 -07001318 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001319
Mel Gorman54a6eb52008-04-28 02:12:16 -07001320 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001321 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001322 page = get_partial_node(n);
1323 if (page)
1324 return page;
1325 }
1326 }
1327#endif
1328 return NULL;
1329}
1330
1331/*
1332 * Get a partial page, lock it and return it.
1333 */
1334static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1335{
1336 struct page *page;
1337 int searchnode = (node == -1) ? numa_node_id() : node;
1338
1339 page = get_partial_node(get_node(s, searchnode));
1340 if (page || (flags & __GFP_THISNODE))
1341 return page;
1342
1343 return get_any_partial(s, flags);
1344}
1345
1346/*
1347 * Move a page back to the lists.
1348 *
1349 * Must be called with the slab lock held.
1350 *
1351 * On exit the slab lock will have been dropped.
1352 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001353static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001354{
Christoph Lametere95eed52007-05-06 14:49:44 -07001355 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001356 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001357
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001358 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001359 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001360
Christoph Lametera973e9d2008-03-01 13:40:44 -08001361 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001362 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001363 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1364 } else {
1365 stat(c, DEACTIVATE_FULL);
1366 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1367 add_full(n, page);
1368 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001369 slab_unlock(page);
1370 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001371 stat(c, DEACTIVATE_EMPTY);
Christoph Lametere95eed52007-05-06 14:49:44 -07001372 if (n->nr_partial < MIN_PARTIAL) {
1373 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001374 * Adding an empty slab to the partial slabs in order
1375 * to avoid page allocator overhead. This slab needs
1376 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001377 * so that the others get filled first. That way the
1378 * size of the partial list stays small.
1379 *
1380 * kmem_cache_shrink can reclaim any empty slabs from the
1381 * partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001382 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001383 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001384 slab_unlock(page);
1385 } else {
1386 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001387 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001388 discard_slab(s, page);
1389 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001390 }
1391}
1392
1393/*
1394 * Remove the cpu slab
1395 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001396static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001397{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001398 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001399 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001400
Christoph Lameterb773ad72008-03-04 11:10:17 -08001401 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001402 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001403 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001404 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001405 * because both freelists are empty. So this is unlikely
1406 * to occur.
1407 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001408 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001409 void **object;
1410
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001411 tail = 0; /* Hot objects. Put the slab first */
1412
Christoph Lameter894b8782007-05-10 03:15:16 -07001413 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001414 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001415 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001416
1417 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001418 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001419 page->freelist = object;
1420 page->inuse--;
1421 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001422 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001423 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001424}
1425
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001426static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001427{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001428 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001429 slab_lock(c->page);
1430 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001431}
1432
1433/*
1434 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001435 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001436 * Called from IPI handler with interrupts disabled.
1437 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001438static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001439{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001440 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001441
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001442 if (likely(c && c->page))
1443 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001444}
1445
1446static void flush_cpu_slab(void *d)
1447{
1448 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001449
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001450 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001451}
1452
1453static void flush_all(struct kmem_cache *s)
1454{
1455#ifdef CONFIG_SMP
1456 on_each_cpu(flush_cpu_slab, s, 1, 1);
1457#else
1458 unsigned long flags;
1459
1460 local_irq_save(flags);
1461 flush_cpu_slab(s);
1462 local_irq_restore(flags);
1463#endif
1464}
1465
1466/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001467 * Check if the objects in a per cpu structure fit numa
1468 * locality expectations.
1469 */
1470static inline int node_match(struct kmem_cache_cpu *c, int node)
1471{
1472#ifdef CONFIG_NUMA
1473 if (node != -1 && c->node != node)
1474 return 0;
1475#endif
1476 return 1;
1477}
1478
1479/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001480 * Slow path. The lockless freelist is empty or we need to perform
1481 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001482 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001483 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001484 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001485 * Processing is still very fast if new objects have been freed to the
1486 * regular freelist. In that case we simply take over the regular freelist
1487 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001488 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001489 * If that is not working then we fall back to the partial lists. We take the
1490 * first element of the freelist as the object to allocate now and move the
1491 * rest of the freelist to the lockless freelist.
1492 *
1493 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001494 * we need to allocate a new slab. This is the slowest path since it involves
1495 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001496 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001497static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001498 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001499{
Christoph Lameter81819f02007-05-06 14:49:36 -07001500 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001501 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001502
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001503 /* We handle __GFP_ZERO in the caller */
1504 gfpflags &= ~__GFP_ZERO;
1505
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001506 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001507 goto new_slab;
1508
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001509 slab_lock(c->page);
1510 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001511 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001512
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001513 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001514
Christoph Lameter894b8782007-05-10 03:15:16 -07001515load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001516 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001517 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001518 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001519 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001520 goto debug;
1521
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001522 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001523 c->page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001524 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001525 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001526unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001527 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001528 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001529 return object;
1530
1531another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001532 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001533
1534new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001535 new = get_partial(s, gfpflags, node);
1536 if (new) {
1537 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001538 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001539 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001540 }
1541
Christoph Lameterb811c202007-10-16 23:25:51 -07001542 if (gfpflags & __GFP_WAIT)
1543 local_irq_enable();
1544
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001545 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001546
1547 if (gfpflags & __GFP_WAIT)
1548 local_irq_disable();
1549
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001550 if (new) {
1551 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001552 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001553 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001554 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001555 slab_lock(new);
1556 SetSlabFrozen(new);
1557 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001558 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001559 }
Linus Torvalds00e962c2008-02-19 09:08:49 -08001560
Christoph Lameter71c7a062008-02-14 14:28:01 -08001561 /*
1562 * No memory available.
1563 *
1564 * If the slab uses higher order allocs but the object is
1565 * smaller than a page size then we can fallback in emergencies
1566 * to the page allocator via kmalloc_large. The page allocator may
1567 * have failed to obtain a higher order page and we can try to
1568 * allocate a single page if the object fits into a single page.
1569 * That is only possible if certain conditions are met that are being
1570 * checked when a slab is created.
1571 */
Christoph Lametercaeab082008-03-12 23:57:49 -07001572 if (!(gfpflags & __GFP_NORETRY) &&
1573 (s->flags & __PAGE_ALLOC_FALLBACK)) {
1574 if (gfpflags & __GFP_WAIT)
1575 local_irq_enable();
1576 object = kmalloc_large(s->objsize, gfpflags);
1577 if (gfpflags & __GFP_WAIT)
1578 local_irq_disable();
1579 return object;
1580 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001581 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001582debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001583 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001584 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001585
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001586 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001587 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001588 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001589 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001590}
1591
1592/*
1593 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1594 * have the fastpath folded into their functions. So no function call
1595 * overhead for requests that can be satisfied on the fastpath.
1596 *
1597 * The fastpath works by first checking if the lockless freelist can be used.
1598 * If not then __slab_alloc is called for slow processing.
1599 *
1600 * Otherwise we can simply pick the next object from the lockless free list.
1601 */
Pekka Enberg06428782008-01-07 23:20:27 -08001602static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001603 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001604{
Christoph Lameter894b8782007-05-10 03:15:16 -07001605 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001606 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001607 unsigned long flags;
1608
Christoph Lameter894b8782007-05-10 03:15:16 -07001609 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001610 c = get_cpu_slab(s, smp_processor_id());
Christoph Lametera973e9d2008-03-01 13:40:44 -08001611 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001612
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001613 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001614
1615 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001616 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001617 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001618 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001619 }
1620 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001621
1622 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001623 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001624
Christoph Lameter894b8782007-05-10 03:15:16 -07001625 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001626}
1627
1628void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1629{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001630 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001631}
1632EXPORT_SYMBOL(kmem_cache_alloc);
1633
1634#ifdef CONFIG_NUMA
1635void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1636{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001637 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001638}
1639EXPORT_SYMBOL(kmem_cache_alloc_node);
1640#endif
1641
1642/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001643 * Slow patch handling. This may still be called frequently since objects
1644 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001645 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001646 * So we still attempt to reduce cache line usage. Just take the slab
1647 * lock and free the item. If there is no additional partial page
1648 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001649 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001650static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001651 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001652{
1653 void *prior;
1654 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001655 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001656
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001657 c = get_cpu_slab(s, raw_smp_processor_id());
1658 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001659 slab_lock(page);
1660
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001661 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001662 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001663
Christoph Lameter81819f02007-05-06 14:49:36 -07001664checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001665 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001666 page->freelist = object;
1667 page->inuse--;
1668
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001669 if (unlikely(SlabFrozen(page))) {
1670 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001671 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001672 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001673
1674 if (unlikely(!page->inuse))
1675 goto slab_empty;
1676
1677 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001678 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001679 * then add it.
1680 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001681 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001682 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001683 stat(c, FREE_ADD_PARTIAL);
1684 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001685
1686out_unlock:
1687 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001688 return;
1689
1690slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001691 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001692 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001693 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001694 */
1695 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001696 stat(c, FREE_REMOVE_PARTIAL);
1697 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001698 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001699 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001700 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001701 return;
1702
1703debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001704 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001705 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001706 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001707}
1708
Christoph Lameter894b8782007-05-10 03:15:16 -07001709/*
1710 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1711 * can perform fastpath freeing without additional function calls.
1712 *
1713 * The fastpath is only possible if we are freeing to the current cpu slab
1714 * of this processor. This typically the case if we have just allocated
1715 * the item before.
1716 *
1717 * If fastpath is not possible then fall back to __slab_free where we deal
1718 * with all sorts of special processing.
1719 */
Pekka Enberg06428782008-01-07 23:20:27 -08001720static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001721 struct page *page, void *x, void *addr)
1722{
1723 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001724 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001725 unsigned long flags;
1726
Christoph Lameter894b8782007-05-10 03:15:16 -07001727 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001728 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001729 debug_check_no_locks_freed(object, c->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001730 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001731 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001732 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001733 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001734 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001735 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001736
1737 local_irq_restore(flags);
1738}
1739
Christoph Lameter81819f02007-05-06 14:49:36 -07001740void kmem_cache_free(struct kmem_cache *s, void *x)
1741{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001742 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001743
Christoph Lameterb49af682007-05-06 14:49:41 -07001744 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001745
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001746 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001747}
1748EXPORT_SYMBOL(kmem_cache_free);
1749
1750/* Figure out on which slab object the object resides */
1751static struct page *get_object_page(const void *x)
1752{
Christoph Lameterb49af682007-05-06 14:49:41 -07001753 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001754
1755 if (!PageSlab(page))
1756 return NULL;
1757
1758 return page;
1759}
1760
1761/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001762 * Object placement in a slab is made very easy because we always start at
1763 * offset 0. If we tune the size of the object to the alignment then we can
1764 * get the required alignment by putting one properly sized object after
1765 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001766 *
1767 * Notice that the allocation order determines the sizes of the per cpu
1768 * caches. Each processor has always one slab available for allocations.
1769 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001770 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001771 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001772 */
1773
1774/*
1775 * Mininum / Maximum order of slab pages. This influences locking overhead
1776 * and slab fragmentation. A higher order reduces the number of partial slabs
1777 * and increases the number of allocations possible without having to
1778 * take the list_lock.
1779 */
1780static int slub_min_order;
1781static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001782static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1783
1784/*
1785 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001786 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001787 */
1788static int slub_nomerge;
1789
1790/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001791 * Calculate the order of allocation given an slab object size.
1792 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001793 * The order of allocation has significant impact on performance and other
1794 * system components. Generally order 0 allocations should be preferred since
1795 * order 0 does not cause fragmentation in the page allocator. Larger objects
1796 * be problematic to put into order 0 slabs because there may be too much
1797 * unused space left. We go to a higher order if more than 1/8th of the slab
1798 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001799 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001800 * In order to reach satisfactory performance we must ensure that a minimum
1801 * number of objects is in one slab. Otherwise we may generate too much
1802 * activity on the partial lists which requires taking the list_lock. This is
1803 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001804 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001805 * slub_max_order specifies the order where we begin to stop considering the
1806 * number of objects in a slab as critical. If we reach slub_max_order then
1807 * we try to keep the page order as low as possible. So we accept more waste
1808 * of space in favor of a small page order.
1809 *
1810 * Higher order allocations also allow the placement of more objects in a
1811 * slab and thereby reduce object handling overhead. If the user has
1812 * requested a higher mininum order then we start with that one instead of
1813 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001814 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001815static inline int slab_order(int size, int min_objects,
1816 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001817{
1818 int order;
1819 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001820 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001821
Christoph Lameter6300ea72007-07-17 04:03:20 -07001822 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001823 fls(min_objects * size - 1) - PAGE_SHIFT);
1824 order <= max_order; order++) {
1825
Christoph Lameter81819f02007-05-06 14:49:36 -07001826 unsigned long slab_size = PAGE_SIZE << order;
1827
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001828 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001829 continue;
1830
Christoph Lameter81819f02007-05-06 14:49:36 -07001831 rem = slab_size % size;
1832
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001833 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001834 break;
1835
1836 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001837
Christoph Lameter81819f02007-05-06 14:49:36 -07001838 return order;
1839}
1840
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001841static inline int calculate_order(int size)
1842{
1843 int order;
1844 int min_objects;
1845 int fraction;
1846
1847 /*
1848 * Attempt to find best configuration for a slab. This
1849 * works by first attempting to generate a layout with
1850 * the best configuration and backing off gradually.
1851 *
1852 * First we reduce the acceptable waste in a slab. Then
1853 * we reduce the minimum objects required in a slab.
1854 */
1855 min_objects = slub_min_objects;
1856 while (min_objects > 1) {
1857 fraction = 8;
1858 while (fraction >= 4) {
1859 order = slab_order(size, min_objects,
1860 slub_max_order, fraction);
1861 if (order <= slub_max_order)
1862 return order;
1863 fraction /= 2;
1864 }
1865 min_objects /= 2;
1866 }
1867
1868 /*
1869 * We were unable to place multiple objects in a slab. Now
1870 * lets see if we can place a single object there.
1871 */
1872 order = slab_order(size, 1, slub_max_order, 1);
1873 if (order <= slub_max_order)
1874 return order;
1875
1876 /*
1877 * Doh this slab cannot be placed using slub_max_order.
1878 */
1879 order = slab_order(size, 1, MAX_ORDER, 1);
1880 if (order <= MAX_ORDER)
1881 return order;
1882 return -ENOSYS;
1883}
1884
Christoph Lameter81819f02007-05-06 14:49:36 -07001885/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001886 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001887 */
1888static unsigned long calculate_alignment(unsigned long flags,
1889 unsigned long align, unsigned long size)
1890{
1891 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001892 * If the user wants hardware cache aligned objects then follow that
1893 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001894 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001895 * The hardware cache alignment cannot override the specified
1896 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001897 */
Nick Pigginb6210382008-03-05 14:05:56 -08001898 if (flags & SLAB_HWCACHE_ALIGN) {
1899 unsigned long ralign = cache_line_size();
1900 while (size <= ralign / 2)
1901 ralign /= 2;
1902 align = max(align, ralign);
1903 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001904
1905 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001906 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001907
1908 return ALIGN(align, sizeof(void *));
1909}
1910
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001911static void init_kmem_cache_cpu(struct kmem_cache *s,
1912 struct kmem_cache_cpu *c)
1913{
1914 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001915 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001916 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001917 c->offset = s->offset / sizeof(void *);
1918 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03001919#ifdef CONFIG_SLUB_STATS
1920 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1921#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001922}
1923
Christoph Lameter81819f02007-05-06 14:49:36 -07001924static void init_kmem_cache_node(struct kmem_cache_node *n)
1925{
1926 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001927 spin_lock_init(&n->list_lock);
1928 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001929#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001930 atomic_long_set(&n->nr_slabs, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07001931 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001932#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001933}
1934
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001935#ifdef CONFIG_SMP
1936/*
1937 * Per cpu array for per cpu structures.
1938 *
1939 * The per cpu array places all kmem_cache_cpu structures from one processor
1940 * close together meaning that it becomes possible that multiple per cpu
1941 * structures are contained in one cacheline. This may be particularly
1942 * beneficial for the kmalloc caches.
1943 *
1944 * A desktop system typically has around 60-80 slabs. With 100 here we are
1945 * likely able to get per cpu structures for all caches from the array defined
1946 * here. We must be able to cover all kmalloc caches during bootstrap.
1947 *
1948 * If the per cpu array is exhausted then fall back to kmalloc
1949 * of individual cachelines. No sharing is possible then.
1950 */
1951#define NR_KMEM_CACHE_CPU 100
1952
1953static DEFINE_PER_CPU(struct kmem_cache_cpu,
1954 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1955
1956static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1957static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1958
1959static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1960 int cpu, gfp_t flags)
1961{
1962 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1963
1964 if (c)
1965 per_cpu(kmem_cache_cpu_free, cpu) =
1966 (void *)c->freelist;
1967 else {
1968 /* Table overflow: So allocate ourselves */
1969 c = kmalloc_node(
1970 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1971 flags, cpu_to_node(cpu));
1972 if (!c)
1973 return NULL;
1974 }
1975
1976 init_kmem_cache_cpu(s, c);
1977 return c;
1978}
1979
1980static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1981{
1982 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1983 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1984 kfree(c);
1985 return;
1986 }
1987 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1988 per_cpu(kmem_cache_cpu_free, cpu) = c;
1989}
1990
1991static void free_kmem_cache_cpus(struct kmem_cache *s)
1992{
1993 int cpu;
1994
1995 for_each_online_cpu(cpu) {
1996 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1997
1998 if (c) {
1999 s->cpu_slab[cpu] = NULL;
2000 free_kmem_cache_cpu(c, cpu);
2001 }
2002 }
2003}
2004
2005static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2006{
2007 int cpu;
2008
2009 for_each_online_cpu(cpu) {
2010 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2011
2012 if (c)
2013 continue;
2014
2015 c = alloc_kmem_cache_cpu(s, cpu, flags);
2016 if (!c) {
2017 free_kmem_cache_cpus(s);
2018 return 0;
2019 }
2020 s->cpu_slab[cpu] = c;
2021 }
2022 return 1;
2023}
2024
2025/*
2026 * Initialize the per cpu array.
2027 */
2028static void init_alloc_cpu_cpu(int cpu)
2029{
2030 int i;
2031
2032 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2033 return;
2034
2035 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2036 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2037
2038 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2039}
2040
2041static void __init init_alloc_cpu(void)
2042{
2043 int cpu;
2044
2045 for_each_online_cpu(cpu)
2046 init_alloc_cpu_cpu(cpu);
2047 }
2048
2049#else
2050static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2051static inline void init_alloc_cpu(void) {}
2052
2053static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2054{
2055 init_kmem_cache_cpu(s, &s->cpu_slab);
2056 return 1;
2057}
2058#endif
2059
Christoph Lameter81819f02007-05-06 14:49:36 -07002060#ifdef CONFIG_NUMA
2061/*
2062 * No kmalloc_node yet so do it by hand. We know that this is the first
2063 * slab on the node for this slabcache. There are no concurrent accesses
2064 * possible.
2065 *
2066 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002067 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2068 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002069 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002070static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2071 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002072{
2073 struct page *page;
2074 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002075 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002076
2077 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2078
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002079 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002080
2081 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002082 if (page_to_nid(page) != node) {
2083 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2084 "node %d\n", node);
2085 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2086 "in order to be able to continue\n");
2087 }
2088
Christoph Lameter81819f02007-05-06 14:49:36 -07002089 n = page->freelist;
2090 BUG_ON(!n);
2091 page->freelist = get_freepointer(kmalloc_caches, n);
2092 page->inuse++;
2093 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002094#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002095 init_object(kmalloc_caches, n, 1);
2096 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002097#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002098 init_kmem_cache_node(n);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002099 inc_slabs_node(kmalloc_caches, node);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002100
rootba84c732008-01-07 23:20:28 -08002101 /*
2102 * lockdep requires consistent irq usage for each lock
2103 * so even though there cannot be a race this early in
2104 * the boot sequence, we still disable irqs.
2105 */
2106 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002107 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002108 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002109 return n;
2110}
2111
2112static void free_kmem_cache_nodes(struct kmem_cache *s)
2113{
2114 int node;
2115
Christoph Lameterf64dc582007-10-16 01:25:33 -07002116 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002117 struct kmem_cache_node *n = s->node[node];
2118 if (n && n != &s->local_node)
2119 kmem_cache_free(kmalloc_caches, n);
2120 s->node[node] = NULL;
2121 }
2122}
2123
2124static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2125{
2126 int node;
2127 int local_node;
2128
2129 if (slab_state >= UP)
2130 local_node = page_to_nid(virt_to_page(s));
2131 else
2132 local_node = 0;
2133
Christoph Lameterf64dc582007-10-16 01:25:33 -07002134 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002135 struct kmem_cache_node *n;
2136
2137 if (local_node == node)
2138 n = &s->local_node;
2139 else {
2140 if (slab_state == DOWN) {
2141 n = early_kmem_cache_node_alloc(gfpflags,
2142 node);
2143 continue;
2144 }
2145 n = kmem_cache_alloc_node(kmalloc_caches,
2146 gfpflags, node);
2147
2148 if (!n) {
2149 free_kmem_cache_nodes(s);
2150 return 0;
2151 }
2152
2153 }
2154 s->node[node] = n;
2155 init_kmem_cache_node(n);
2156 }
2157 return 1;
2158}
2159#else
2160static void free_kmem_cache_nodes(struct kmem_cache *s)
2161{
2162}
2163
2164static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2165{
2166 init_kmem_cache_node(&s->local_node);
2167 return 1;
2168}
2169#endif
2170
2171/*
2172 * calculate_sizes() determines the order and the distribution of data within
2173 * a slab object.
2174 */
2175static int calculate_sizes(struct kmem_cache *s)
2176{
2177 unsigned long flags = s->flags;
2178 unsigned long size = s->objsize;
2179 unsigned long align = s->align;
2180
2181 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002182 * Round up object size to the next word boundary. We can only
2183 * place the free pointer at word boundaries and this determines
2184 * the possible location of the free pointer.
2185 */
2186 size = ALIGN(size, sizeof(void *));
2187
2188#ifdef CONFIG_SLUB_DEBUG
2189 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002190 * Determine if we can poison the object itself. If the user of
2191 * the slab may touch the object after free or before allocation
2192 * then we should never poison the object itself.
2193 */
2194 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002195 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002196 s->flags |= __OBJECT_POISON;
2197 else
2198 s->flags &= ~__OBJECT_POISON;
2199
Christoph Lameter81819f02007-05-06 14:49:36 -07002200
2201 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002202 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002203 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002204 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002205 */
2206 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2207 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002208#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002209
2210 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002211 * With that we have determined the number of bytes in actual use
2212 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002213 */
2214 s->inuse = size;
2215
2216 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002217 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002218 /*
2219 * Relocate free pointer after the object if it is not
2220 * permitted to overwrite the first word of the object on
2221 * kmem_cache_free.
2222 *
2223 * This is the case if we do RCU, have a constructor or
2224 * destructor or are poisoning the objects.
2225 */
2226 s->offset = size;
2227 size += sizeof(void *);
2228 }
2229
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002230#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002231 if (flags & SLAB_STORE_USER)
2232 /*
2233 * Need to store information about allocs and frees after
2234 * the object.
2235 */
2236 size += 2 * sizeof(struct track);
2237
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002238 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002239 /*
2240 * Add some empty padding so that we can catch
2241 * overwrites from earlier objects rather than let
2242 * tracking information or the free pointer be
2243 * corrupted if an user writes before the start
2244 * of the object.
2245 */
2246 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002247#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002248
Christoph Lameter81819f02007-05-06 14:49:36 -07002249 /*
2250 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002251 * user specified and the dynamic determination of cache line size
2252 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002253 */
2254 align = calculate_alignment(flags, align, s->objsize);
2255
2256 /*
2257 * SLUB stores one object immediately after another beginning from
2258 * offset 0. In order to align the objects we have to simply size
2259 * each object to conform to the alignment.
2260 */
2261 size = ALIGN(size, align);
2262 s->size = size;
2263
Christoph Lameter71c7a062008-02-14 14:28:01 -08002264 if ((flags & __KMALLOC_CACHE) &&
2265 PAGE_SIZE / size < slub_min_objects) {
2266 /*
2267 * Kmalloc cache that would not have enough objects in
2268 * an order 0 page. Kmalloc slabs can fallback to
2269 * page allocator order 0 allocs so take a reasonably large
2270 * order that will allows us a good number of objects.
2271 */
2272 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2273 s->flags |= __PAGE_ALLOC_FALLBACK;
2274 s->allocflags |= __GFP_NOWARN;
2275 } else
2276 s->order = calculate_order(size);
2277
Christoph Lameter81819f02007-05-06 14:49:36 -07002278 if (s->order < 0)
2279 return 0;
2280
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002281 s->allocflags = 0;
2282 if (s->order)
2283 s->allocflags |= __GFP_COMP;
2284
2285 if (s->flags & SLAB_CACHE_DMA)
2286 s->allocflags |= SLUB_DMA;
2287
2288 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2289 s->allocflags |= __GFP_RECLAIMABLE;
2290
Christoph Lameter81819f02007-05-06 14:49:36 -07002291 /*
2292 * Determine the number of objects per slab
2293 */
2294 s->objects = (PAGE_SIZE << s->order) / size;
2295
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002296 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002297
2298}
2299
Christoph Lameter81819f02007-05-06 14:49:36 -07002300static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2301 const char *name, size_t size,
2302 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002303 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002304{
2305 memset(s, 0, kmem_size);
2306 s->name = name;
2307 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002308 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002309 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002310 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002311
2312 if (!calculate_sizes(s))
2313 goto error;
2314
2315 s->refcount = 1;
2316#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002317 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002318#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002319 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2320 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002321
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002322 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002323 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002324 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002325error:
2326 if (flags & SLAB_PANIC)
2327 panic("Cannot create slab %s size=%lu realsize=%u "
2328 "order=%u offset=%u flags=%lx\n",
2329 s->name, (unsigned long)size, s->size, s->order,
2330 s->offset, flags);
2331 return 0;
2332}
Christoph Lameter81819f02007-05-06 14:49:36 -07002333
2334/*
2335 * Check if a given pointer is valid
2336 */
2337int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2338{
Pekka Enberg06428782008-01-07 23:20:27 -08002339 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002340
2341 page = get_object_page(object);
2342
2343 if (!page || s != page->slab)
2344 /* No slab or wrong slab */
2345 return 0;
2346
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002347 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002348 return 0;
2349
2350 /*
2351 * We could also check if the object is on the slabs freelist.
2352 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002353 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002354 * to a certain slab.
2355 */
2356 return 1;
2357}
2358EXPORT_SYMBOL(kmem_ptr_validate);
2359
2360/*
2361 * Determine the size of a slab object
2362 */
2363unsigned int kmem_cache_size(struct kmem_cache *s)
2364{
2365 return s->objsize;
2366}
2367EXPORT_SYMBOL(kmem_cache_size);
2368
2369const char *kmem_cache_name(struct kmem_cache *s)
2370{
2371 return s->name;
2372}
2373EXPORT_SYMBOL(kmem_cache_name);
2374
2375/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002376 * Attempt to free all slabs on a node. Return the number of slabs we
2377 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002378 */
2379static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2380 struct list_head *list)
2381{
2382 int slabs_inuse = 0;
2383 unsigned long flags;
2384 struct page *page, *h;
2385
2386 spin_lock_irqsave(&n->list_lock, flags);
2387 list_for_each_entry_safe(page, h, list, lru)
2388 if (!page->inuse) {
2389 list_del(&page->lru);
2390 discard_slab(s, page);
2391 } else
2392 slabs_inuse++;
2393 spin_unlock_irqrestore(&n->list_lock, flags);
2394 return slabs_inuse;
2395}
2396
2397/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002398 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002399 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002400static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002401{
2402 int node;
2403
2404 flush_all(s);
2405
2406 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002407 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002408 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002409 struct kmem_cache_node *n = get_node(s, node);
2410
Christoph Lameter2086d262007-05-06 14:49:46 -07002411 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002412 if (slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002413 return 1;
2414 }
2415 free_kmem_cache_nodes(s);
2416 return 0;
2417}
2418
2419/*
2420 * Close a cache and release the kmem_cache structure
2421 * (must be used for caches created using kmem_cache_create)
2422 */
2423void kmem_cache_destroy(struct kmem_cache *s)
2424{
2425 down_write(&slub_lock);
2426 s->refcount--;
2427 if (!s->refcount) {
2428 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002429 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002430 if (kmem_cache_close(s))
2431 WARN_ON(1);
2432 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002433 } else
2434 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002435}
2436EXPORT_SYMBOL(kmem_cache_destroy);
2437
2438/********************************************************************
2439 * Kmalloc subsystem
2440 *******************************************************************/
2441
Christoph Lameter331dc552008-02-14 14:28:09 -08002442struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002443EXPORT_SYMBOL(kmalloc_caches);
2444
Christoph Lameter81819f02007-05-06 14:49:36 -07002445static int __init setup_slub_min_order(char *str)
2446{
Pekka Enberg06428782008-01-07 23:20:27 -08002447 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002448
2449 return 1;
2450}
2451
2452__setup("slub_min_order=", setup_slub_min_order);
2453
2454static int __init setup_slub_max_order(char *str)
2455{
Pekka Enberg06428782008-01-07 23:20:27 -08002456 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002457
2458 return 1;
2459}
2460
2461__setup("slub_max_order=", setup_slub_max_order);
2462
2463static int __init setup_slub_min_objects(char *str)
2464{
Pekka Enberg06428782008-01-07 23:20:27 -08002465 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002466
2467 return 1;
2468}
2469
2470__setup("slub_min_objects=", setup_slub_min_objects);
2471
2472static int __init setup_slub_nomerge(char *str)
2473{
2474 slub_nomerge = 1;
2475 return 1;
2476}
2477
2478__setup("slub_nomerge", setup_slub_nomerge);
2479
Christoph Lameter81819f02007-05-06 14:49:36 -07002480static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2481 const char *name, int size, gfp_t gfp_flags)
2482{
2483 unsigned int flags = 0;
2484
2485 if (gfp_flags & SLUB_DMA)
2486 flags = SLAB_CACHE_DMA;
2487
2488 down_write(&slub_lock);
2489 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter71c7a062008-02-14 14:28:01 -08002490 flags | __KMALLOC_CACHE, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002491 goto panic;
2492
2493 list_add(&s->list, &slab_caches);
2494 up_write(&slub_lock);
2495 if (sysfs_slab_add(s))
2496 goto panic;
2497 return s;
2498
2499panic:
2500 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2501}
2502
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002503#ifdef CONFIG_ZONE_DMA
Christoph Lameter4097d602008-04-14 18:51:18 +03002504static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002505
2506static void sysfs_add_func(struct work_struct *w)
2507{
2508 struct kmem_cache *s;
2509
2510 down_write(&slub_lock);
2511 list_for_each_entry(s, &slab_caches, list) {
2512 if (s->flags & __SYSFS_ADD_DEFERRED) {
2513 s->flags &= ~__SYSFS_ADD_DEFERRED;
2514 sysfs_slab_add(s);
2515 }
2516 }
2517 up_write(&slub_lock);
2518}
2519
2520static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2521
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002522static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2523{
2524 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002525 char *text;
2526 size_t realsize;
2527
2528 s = kmalloc_caches_dma[index];
2529 if (s)
2530 return s;
2531
2532 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002533 if (flags & __GFP_WAIT)
2534 down_write(&slub_lock);
2535 else {
2536 if (!down_write_trylock(&slub_lock))
2537 goto out;
2538 }
2539
2540 if (kmalloc_caches_dma[index])
2541 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002542
Christoph Lameter7b55f622007-07-17 04:03:27 -07002543 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002544 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2545 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002546 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2547
2548 if (!s || !text || !kmem_cache_open(s, flags, text,
2549 realsize, ARCH_KMALLOC_MINALIGN,
2550 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2551 kfree(s);
2552 kfree(text);
2553 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002554 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002555
2556 list_add(&s->list, &slab_caches);
2557 kmalloc_caches_dma[index] = s;
2558
2559 schedule_work(&sysfs_add_work);
2560
2561unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002562 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002563out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002564 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002565}
2566#endif
2567
Christoph Lameterf1b26332007-07-17 04:03:26 -07002568/*
2569 * Conversion table for small slabs sizes / 8 to the index in the
2570 * kmalloc array. This is necessary for slabs < 192 since we have non power
2571 * of two cache sizes there. The size of larger slabs can be determined using
2572 * fls.
2573 */
2574static s8 size_index[24] = {
2575 3, /* 8 */
2576 4, /* 16 */
2577 5, /* 24 */
2578 5, /* 32 */
2579 6, /* 40 */
2580 6, /* 48 */
2581 6, /* 56 */
2582 6, /* 64 */
2583 1, /* 72 */
2584 1, /* 80 */
2585 1, /* 88 */
2586 1, /* 96 */
2587 7, /* 104 */
2588 7, /* 112 */
2589 7, /* 120 */
2590 7, /* 128 */
2591 2, /* 136 */
2592 2, /* 144 */
2593 2, /* 152 */
2594 2, /* 160 */
2595 2, /* 168 */
2596 2, /* 176 */
2597 2, /* 184 */
2598 2 /* 192 */
2599};
2600
Christoph Lameter81819f02007-05-06 14:49:36 -07002601static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2602{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002603 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002604
Christoph Lameterf1b26332007-07-17 04:03:26 -07002605 if (size <= 192) {
2606 if (!size)
2607 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002608
Christoph Lameterf1b26332007-07-17 04:03:26 -07002609 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002610 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002611 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002612
2613#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002614 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002615 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002616
Christoph Lameter81819f02007-05-06 14:49:36 -07002617#endif
2618 return &kmalloc_caches[index];
2619}
2620
2621void *__kmalloc(size_t size, gfp_t flags)
2622{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002623 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002624
Christoph Lameter331dc552008-02-14 14:28:09 -08002625 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002626 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002627
2628 s = get_slab(size, flags);
2629
2630 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002631 return s;
2632
Christoph Lameterce15fea2007-07-17 04:03:28 -07002633 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002634}
2635EXPORT_SYMBOL(__kmalloc);
2636
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002637static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2638{
2639 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2640 get_order(size));
2641
2642 if (page)
2643 return page_address(page);
2644 else
2645 return NULL;
2646}
2647
Christoph Lameter81819f02007-05-06 14:49:36 -07002648#ifdef CONFIG_NUMA
2649void *__kmalloc_node(size_t size, gfp_t flags, int node)
2650{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002651 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002652
Christoph Lameter331dc552008-02-14 14:28:09 -08002653 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002654 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002655
2656 s = get_slab(size, flags);
2657
2658 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002659 return s;
2660
Christoph Lameterce15fea2007-07-17 04:03:28 -07002661 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002662}
2663EXPORT_SYMBOL(__kmalloc_node);
2664#endif
2665
2666size_t ksize(const void *object)
2667{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002668 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002669 struct kmem_cache *s;
2670
Christoph Lameteref8b4522007-10-16 01:24:46 -07002671 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002672 return 0;
2673
Vegard Nossum294a80a2007-12-04 23:45:30 -08002674 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002675
2676 if (unlikely(!PageSlab(page)))
2677 return PAGE_SIZE << compound_order(page);
2678
Christoph Lameter81819f02007-05-06 14:49:36 -07002679 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002680
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002681#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002682 /*
2683 * Debugging requires use of the padding between object
2684 * and whatever may come after it.
2685 */
2686 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2687 return s->objsize;
2688
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002689#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002690 /*
2691 * If we have the need to store the freelist pointer
2692 * back there or track user information then we can
2693 * only use the space before that information.
2694 */
2695 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2696 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002697 /*
2698 * Else we can use all the padding etc for the allocation
2699 */
2700 return s->size;
2701}
2702EXPORT_SYMBOL(ksize);
2703
2704void kfree(const void *x)
2705{
Christoph Lameter81819f02007-05-06 14:49:36 -07002706 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002707 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002708
Satyam Sharma2408c552007-10-16 01:24:44 -07002709 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002710 return;
2711
Christoph Lameterb49af682007-05-06 14:49:41 -07002712 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002713 if (unlikely(!PageSlab(page))) {
2714 put_page(page);
2715 return;
2716 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002717 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002718}
2719EXPORT_SYMBOL(kfree);
2720
Christoph Lameter2086d262007-05-06 14:49:46 -07002721/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002722 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2723 * the remaining slabs by the number of items in use. The slabs with the
2724 * most items in use come first. New allocations will then fill those up
2725 * and thus they can be removed from the partial lists.
2726 *
2727 * The slabs with the least items are placed last. This results in them
2728 * being allocated from last increasing the chance that the last objects
2729 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002730 */
2731int kmem_cache_shrink(struct kmem_cache *s)
2732{
2733 int node;
2734 int i;
2735 struct kmem_cache_node *n;
2736 struct page *page;
2737 struct page *t;
2738 struct list_head *slabs_by_inuse =
2739 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2740 unsigned long flags;
2741
2742 if (!slabs_by_inuse)
2743 return -ENOMEM;
2744
2745 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002746 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002747 n = get_node(s, node);
2748
2749 if (!n->nr_partial)
2750 continue;
2751
2752 for (i = 0; i < s->objects; i++)
2753 INIT_LIST_HEAD(slabs_by_inuse + i);
2754
2755 spin_lock_irqsave(&n->list_lock, flags);
2756
2757 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002758 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002759 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002760 * Note that concurrent frees may occur while we hold the
2761 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002762 */
2763 list_for_each_entry_safe(page, t, &n->partial, lru) {
2764 if (!page->inuse && slab_trylock(page)) {
2765 /*
2766 * Must hold slab lock here because slab_free
2767 * may have freed the last object and be
2768 * waiting to release the slab.
2769 */
2770 list_del(&page->lru);
2771 n->nr_partial--;
2772 slab_unlock(page);
2773 discard_slab(s, page);
2774 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002775 list_move(&page->lru,
2776 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002777 }
2778 }
2779
Christoph Lameter2086d262007-05-06 14:49:46 -07002780 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002781 * Rebuild the partial list with the slabs filled up most
2782 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002783 */
2784 for (i = s->objects - 1; i >= 0; i--)
2785 list_splice(slabs_by_inuse + i, n->partial.prev);
2786
Christoph Lameter2086d262007-05-06 14:49:46 -07002787 spin_unlock_irqrestore(&n->list_lock, flags);
2788 }
2789
2790 kfree(slabs_by_inuse);
2791 return 0;
2792}
2793EXPORT_SYMBOL(kmem_cache_shrink);
2794
Yasunori Gotob9049e22007-10-21 16:41:37 -07002795#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2796static int slab_mem_going_offline_callback(void *arg)
2797{
2798 struct kmem_cache *s;
2799
2800 down_read(&slub_lock);
2801 list_for_each_entry(s, &slab_caches, list)
2802 kmem_cache_shrink(s);
2803 up_read(&slub_lock);
2804
2805 return 0;
2806}
2807
2808static void slab_mem_offline_callback(void *arg)
2809{
2810 struct kmem_cache_node *n;
2811 struct kmem_cache *s;
2812 struct memory_notify *marg = arg;
2813 int offline_node;
2814
2815 offline_node = marg->status_change_nid;
2816
2817 /*
2818 * If the node still has available memory. we need kmem_cache_node
2819 * for it yet.
2820 */
2821 if (offline_node < 0)
2822 return;
2823
2824 down_read(&slub_lock);
2825 list_for_each_entry(s, &slab_caches, list) {
2826 n = get_node(s, offline_node);
2827 if (n) {
2828 /*
2829 * if n->nr_slabs > 0, slabs still exist on the node
2830 * that is going down. We were unable to free them,
2831 * and offline_pages() function shoudn't call this
2832 * callback. So, we must fail.
2833 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002834 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002835
2836 s->node[offline_node] = NULL;
2837 kmem_cache_free(kmalloc_caches, n);
2838 }
2839 }
2840 up_read(&slub_lock);
2841}
2842
2843static int slab_mem_going_online_callback(void *arg)
2844{
2845 struct kmem_cache_node *n;
2846 struct kmem_cache *s;
2847 struct memory_notify *marg = arg;
2848 int nid = marg->status_change_nid;
2849 int ret = 0;
2850
2851 /*
2852 * If the node's memory is already available, then kmem_cache_node is
2853 * already created. Nothing to do.
2854 */
2855 if (nid < 0)
2856 return 0;
2857
2858 /*
2859 * We are bringing a node online. No memory is availabe yet. We must
2860 * allocate a kmem_cache_node structure in order to bring the node
2861 * online.
2862 */
2863 down_read(&slub_lock);
2864 list_for_each_entry(s, &slab_caches, list) {
2865 /*
2866 * XXX: kmem_cache_alloc_node will fallback to other nodes
2867 * since memory is not yet available from the node that
2868 * is brought up.
2869 */
2870 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2871 if (!n) {
2872 ret = -ENOMEM;
2873 goto out;
2874 }
2875 init_kmem_cache_node(n);
2876 s->node[nid] = n;
2877 }
2878out:
2879 up_read(&slub_lock);
2880 return ret;
2881}
2882
2883static int slab_memory_callback(struct notifier_block *self,
2884 unsigned long action, void *arg)
2885{
2886 int ret = 0;
2887
2888 switch (action) {
2889 case MEM_GOING_ONLINE:
2890 ret = slab_mem_going_online_callback(arg);
2891 break;
2892 case MEM_GOING_OFFLINE:
2893 ret = slab_mem_going_offline_callback(arg);
2894 break;
2895 case MEM_OFFLINE:
2896 case MEM_CANCEL_ONLINE:
2897 slab_mem_offline_callback(arg);
2898 break;
2899 case MEM_ONLINE:
2900 case MEM_CANCEL_OFFLINE:
2901 break;
2902 }
2903
2904 ret = notifier_from_errno(ret);
2905 return ret;
2906}
2907
2908#endif /* CONFIG_MEMORY_HOTPLUG */
2909
Christoph Lameter81819f02007-05-06 14:49:36 -07002910/********************************************************************
2911 * Basic setup of slabs
2912 *******************************************************************/
2913
2914void __init kmem_cache_init(void)
2915{
2916 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002917 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002918
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002919 init_alloc_cpu();
2920
Christoph Lameter81819f02007-05-06 14:49:36 -07002921#ifdef CONFIG_NUMA
2922 /*
2923 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002924 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002925 * kmem_cache_open for slab_state == DOWN.
2926 */
2927 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2928 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002929 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002930 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002931
2932 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002933#endif
2934
2935 /* Able to allocate the per node structures */
2936 slab_state = PARTIAL;
2937
2938 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002939 if (KMALLOC_MIN_SIZE <= 64) {
2940 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002941 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002942 caches++;
2943 }
2944 if (KMALLOC_MIN_SIZE <= 128) {
2945 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002946 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002947 caches++;
2948 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002949
Christoph Lameter331dc552008-02-14 14:28:09 -08002950 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002951 create_kmalloc_cache(&kmalloc_caches[i],
2952 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002953 caches++;
2954 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002955
Christoph Lameterf1b26332007-07-17 04:03:26 -07002956
2957 /*
2958 * Patch up the size_index table if we have strange large alignment
2959 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002960 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002961 *
2962 * Largest permitted alignment is 256 bytes due to the way we
2963 * handle the index determination for the smaller caches.
2964 *
2965 * Make sure that nothing crazy happens if someone starts tinkering
2966 * around with ARCH_KMALLOC_MINALIGN
2967 */
2968 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2969 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2970
Christoph Lameter12ad6842007-07-17 04:03:28 -07002971 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002972 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2973
Christoph Lameter81819f02007-05-06 14:49:36 -07002974 slab_state = UP;
2975
2976 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08002977 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002978 kmalloc_caches[i]. name =
2979 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2980
2981#ifdef CONFIG_SMP
2982 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002983 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2984 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2985#else
2986 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002987#endif
2988
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002989 printk(KERN_INFO
2990 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002991 " CPUs=%d, Nodes=%d\n",
2992 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002993 slub_min_order, slub_max_order, slub_min_objects,
2994 nr_cpu_ids, nr_node_ids);
2995}
2996
2997/*
2998 * Find a mergeable slab cache
2999 */
3000static int slab_unmergeable(struct kmem_cache *s)
3001{
3002 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3003 return 1;
3004
Christoph Lameter331dc552008-02-14 14:28:09 -08003005 if ((s->flags & __PAGE_ALLOC_FALLBACK))
Christoph Lameter71c7a062008-02-14 14:28:01 -08003006 return 1;
3007
Christoph Lameterc59def92007-05-16 22:10:50 -07003008 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003009 return 1;
3010
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003011 /*
3012 * We may have set a slab to be unmergeable during bootstrap.
3013 */
3014 if (s->refcount < 0)
3015 return 1;
3016
Christoph Lameter81819f02007-05-06 14:49:36 -07003017 return 0;
3018}
3019
3020static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003021 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003022 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003023{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003024 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003025
3026 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3027 return NULL;
3028
Christoph Lameterc59def92007-05-16 22:10:50 -07003029 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003030 return NULL;
3031
3032 size = ALIGN(size, sizeof(void *));
3033 align = calculate_alignment(flags, align, size);
3034 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003035 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003036
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003037 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003038 if (slab_unmergeable(s))
3039 continue;
3040
3041 if (size > s->size)
3042 continue;
3043
Christoph Lameterba0268a2007-09-11 15:24:11 -07003044 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003045 continue;
3046 /*
3047 * Check if alignment is compatible.
3048 * Courtesy of Adrian Drzewiecki
3049 */
Pekka Enberg06428782008-01-07 23:20:27 -08003050 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003051 continue;
3052
3053 if (s->size - size >= sizeof(void *))
3054 continue;
3055
3056 return s;
3057 }
3058 return NULL;
3059}
3060
3061struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3062 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003063 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003064{
3065 struct kmem_cache *s;
3066
3067 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003068 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003069 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003070 int cpu;
3071
Christoph Lameter81819f02007-05-06 14:49:36 -07003072 s->refcount++;
3073 /*
3074 * Adjust the object sizes so that we clear
3075 * the complete object on kzalloc.
3076 */
3077 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003078
3079 /*
3080 * And then we need to update the object size in the
3081 * per cpu structures
3082 */
3083 for_each_online_cpu(cpu)
3084 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003085
Christoph Lameter81819f02007-05-06 14:49:36 -07003086 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003087 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003088
Christoph Lameter81819f02007-05-06 14:49:36 -07003089 if (sysfs_slab_alias(s, name))
3090 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003091 return s;
3092 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003093
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003094 s = kmalloc(kmem_size, GFP_KERNEL);
3095 if (s) {
3096 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003097 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003098 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003099 up_write(&slub_lock);
3100 if (sysfs_slab_add(s))
3101 goto err;
3102 return s;
3103 }
3104 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003105 }
3106 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003107
3108err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003109 if (flags & SLAB_PANIC)
3110 panic("Cannot create slabcache %s\n", name);
3111 else
3112 s = NULL;
3113 return s;
3114}
3115EXPORT_SYMBOL(kmem_cache_create);
3116
Christoph Lameter81819f02007-05-06 14:49:36 -07003117#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003118/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003119 * Use the cpu notifier to insure that the cpu slabs are flushed when
3120 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003121 */
3122static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3123 unsigned long action, void *hcpu)
3124{
3125 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003126 struct kmem_cache *s;
3127 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003128
3129 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003130 case CPU_UP_PREPARE:
3131 case CPU_UP_PREPARE_FROZEN:
3132 init_alloc_cpu_cpu(cpu);
3133 down_read(&slub_lock);
3134 list_for_each_entry(s, &slab_caches, list)
3135 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3136 GFP_KERNEL);
3137 up_read(&slub_lock);
3138 break;
3139
Christoph Lameter81819f02007-05-06 14:49:36 -07003140 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003141 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003142 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003143 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003144 down_read(&slub_lock);
3145 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003146 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3147
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003148 local_irq_save(flags);
3149 __flush_cpu_slab(s, cpu);
3150 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003151 free_kmem_cache_cpu(c, cpu);
3152 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003153 }
3154 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003155 break;
3156 default:
3157 break;
3158 }
3159 return NOTIFY_OK;
3160}
3161
Pekka Enberg06428782008-01-07 23:20:27 -08003162static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003163 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003164};
Christoph Lameter81819f02007-05-06 14:49:36 -07003165
3166#endif
3167
Christoph Lameter81819f02007-05-06 14:49:36 -07003168void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3169{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003170 struct kmem_cache *s;
3171
Christoph Lameter331dc552008-02-14 14:28:09 -08003172 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003173 return kmalloc_large(size, gfpflags);
3174
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003175 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003176
Satyam Sharma2408c552007-10-16 01:24:44 -07003177 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003178 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003179
Christoph Lameterce15fea2007-07-17 04:03:28 -07003180 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003181}
3182
3183void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3184 int node, void *caller)
3185{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003186 struct kmem_cache *s;
3187
Christoph Lameter331dc552008-02-14 14:28:09 -08003188 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003189 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003190
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003191 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003192
Satyam Sharma2408c552007-10-16 01:24:44 -07003193 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003194 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003195
Christoph Lameterce15fea2007-07-17 04:03:28 -07003196 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003197}
3198
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003199#if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3200static unsigned long count_partial(struct kmem_cache_node *n)
3201{
3202 unsigned long flags;
3203 unsigned long x = 0;
3204 struct page *page;
3205
3206 spin_lock_irqsave(&n->list_lock, flags);
3207 list_for_each_entry(page, &n->partial, lru)
3208 x += page->inuse;
3209 spin_unlock_irqrestore(&n->list_lock, flags);
3210 return x;
3211}
3212#endif
3213
Christoph Lameter41ecc552007-05-09 02:32:44 -07003214#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003215static int validate_slab(struct kmem_cache *s, struct page *page,
3216 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003217{
3218 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003219 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003220
3221 if (!check_slab(s, page) ||
3222 !on_freelist(s, page, NULL))
3223 return 0;
3224
3225 /* Now we know that a valid freelist exists */
3226 bitmap_zero(map, s->objects);
3227
Christoph Lameter7656c722007-05-09 02:32:40 -07003228 for_each_free_object(p, s, page->freelist) {
3229 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003230 if (!check_object(s, page, p, 0))
3231 return 0;
3232 }
3233
Christoph Lameter7656c722007-05-09 02:32:40 -07003234 for_each_object(p, s, addr)
3235 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003236 if (!check_object(s, page, p, 1))
3237 return 0;
3238 return 1;
3239}
3240
Christoph Lameter434e2452007-07-17 04:03:30 -07003241static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3242 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003243{
3244 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003245 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003246 slab_unlock(page);
3247 } else
3248 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3249 s->name, page);
3250
3251 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003252 if (!SlabDebug(page))
3253 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003254 "on slab 0x%p\n", s->name, page);
3255 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003256 if (SlabDebug(page))
3257 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003258 "slab 0x%p\n", s->name, page);
3259 }
3260}
3261
Christoph Lameter434e2452007-07-17 04:03:30 -07003262static int validate_slab_node(struct kmem_cache *s,
3263 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003264{
3265 unsigned long count = 0;
3266 struct page *page;
3267 unsigned long flags;
3268
3269 spin_lock_irqsave(&n->list_lock, flags);
3270
3271 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003272 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003273 count++;
3274 }
3275 if (count != n->nr_partial)
3276 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3277 "counter=%ld\n", s->name, count, n->nr_partial);
3278
3279 if (!(s->flags & SLAB_STORE_USER))
3280 goto out;
3281
3282 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003283 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003284 count++;
3285 }
3286 if (count != atomic_long_read(&n->nr_slabs))
3287 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3288 "counter=%ld\n", s->name, count,
3289 atomic_long_read(&n->nr_slabs));
3290
3291out:
3292 spin_unlock_irqrestore(&n->list_lock, flags);
3293 return count;
3294}
3295
Christoph Lameter434e2452007-07-17 04:03:30 -07003296static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003297{
3298 int node;
3299 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003300 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3301 sizeof(unsigned long), GFP_KERNEL);
3302
3303 if (!map)
3304 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003305
3306 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003307 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003308 struct kmem_cache_node *n = get_node(s, node);
3309
Christoph Lameter434e2452007-07-17 04:03:30 -07003310 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003311 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003312 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003313 return count;
3314}
3315
Christoph Lameterb3459702007-05-09 02:32:41 -07003316#ifdef SLUB_RESILIENCY_TEST
3317static void resiliency_test(void)
3318{
3319 u8 *p;
3320
3321 printk(KERN_ERR "SLUB resiliency testing\n");
3322 printk(KERN_ERR "-----------------------\n");
3323 printk(KERN_ERR "A. Corruption after allocation\n");
3324
3325 p = kzalloc(16, GFP_KERNEL);
3326 p[16] = 0x12;
3327 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3328 " 0x12->0x%p\n\n", p + 16);
3329
3330 validate_slab_cache(kmalloc_caches + 4);
3331
3332 /* Hmmm... The next two are dangerous */
3333 p = kzalloc(32, GFP_KERNEL);
3334 p[32 + sizeof(void *)] = 0x34;
3335 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003336 " 0x34 -> -0x%p\n", p);
3337 printk(KERN_ERR
3338 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003339
3340 validate_slab_cache(kmalloc_caches + 5);
3341 p = kzalloc(64, GFP_KERNEL);
3342 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3343 *p = 0x56;
3344 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3345 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003346 printk(KERN_ERR
3347 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003348 validate_slab_cache(kmalloc_caches + 6);
3349
3350 printk(KERN_ERR "\nB. Corruption after free\n");
3351 p = kzalloc(128, GFP_KERNEL);
3352 kfree(p);
3353 *p = 0x78;
3354 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3355 validate_slab_cache(kmalloc_caches + 7);
3356
3357 p = kzalloc(256, GFP_KERNEL);
3358 kfree(p);
3359 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003360 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3361 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003362 validate_slab_cache(kmalloc_caches + 8);
3363
3364 p = kzalloc(512, GFP_KERNEL);
3365 kfree(p);
3366 p[512] = 0xab;
3367 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3368 validate_slab_cache(kmalloc_caches + 9);
3369}
3370#else
3371static void resiliency_test(void) {};
3372#endif
3373
Christoph Lameter88a420e2007-05-06 14:49:45 -07003374/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003375 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003376 * and freed.
3377 */
3378
3379struct location {
3380 unsigned long count;
3381 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003382 long long sum_time;
3383 long min_time;
3384 long max_time;
3385 long min_pid;
3386 long max_pid;
3387 cpumask_t cpus;
3388 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003389};
3390
3391struct loc_track {
3392 unsigned long max;
3393 unsigned long count;
3394 struct location *loc;
3395};
3396
3397static void free_loc_track(struct loc_track *t)
3398{
3399 if (t->max)
3400 free_pages((unsigned long)t->loc,
3401 get_order(sizeof(struct location) * t->max));
3402}
3403
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003404static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003405{
3406 struct location *l;
3407 int order;
3408
Christoph Lameter88a420e2007-05-06 14:49:45 -07003409 order = get_order(sizeof(struct location) * max);
3410
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003411 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003412 if (!l)
3413 return 0;
3414
3415 if (t->count) {
3416 memcpy(l, t->loc, sizeof(struct location) * t->count);
3417 free_loc_track(t);
3418 }
3419 t->max = max;
3420 t->loc = l;
3421 return 1;
3422}
3423
3424static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003425 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003426{
3427 long start, end, pos;
3428 struct location *l;
3429 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003430 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003431
3432 start = -1;
3433 end = t->count;
3434
3435 for ( ; ; ) {
3436 pos = start + (end - start + 1) / 2;
3437
3438 /*
3439 * There is nothing at "end". If we end up there
3440 * we need to add something to before end.
3441 */
3442 if (pos == end)
3443 break;
3444
3445 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003446 if (track->addr == caddr) {
3447
3448 l = &t->loc[pos];
3449 l->count++;
3450 if (track->when) {
3451 l->sum_time += age;
3452 if (age < l->min_time)
3453 l->min_time = age;
3454 if (age > l->max_time)
3455 l->max_time = age;
3456
3457 if (track->pid < l->min_pid)
3458 l->min_pid = track->pid;
3459 if (track->pid > l->max_pid)
3460 l->max_pid = track->pid;
3461
3462 cpu_set(track->cpu, l->cpus);
3463 }
3464 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003465 return 1;
3466 }
3467
Christoph Lameter45edfa52007-05-09 02:32:45 -07003468 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003469 end = pos;
3470 else
3471 start = pos;
3472 }
3473
3474 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003475 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003476 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003477 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003478 return 0;
3479
3480 l = t->loc + pos;
3481 if (pos < t->count)
3482 memmove(l + 1, l,
3483 (t->count - pos) * sizeof(struct location));
3484 t->count++;
3485 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003486 l->addr = track->addr;
3487 l->sum_time = age;
3488 l->min_time = age;
3489 l->max_time = age;
3490 l->min_pid = track->pid;
3491 l->max_pid = track->pid;
3492 cpus_clear(l->cpus);
3493 cpu_set(track->cpu, l->cpus);
3494 nodes_clear(l->nodes);
3495 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003496 return 1;
3497}
3498
3499static void process_slab(struct loc_track *t, struct kmem_cache *s,
3500 struct page *page, enum track_item alloc)
3501{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003502 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003503 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003504 void *p;
3505
3506 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003507 for_each_free_object(p, s, page->freelist)
3508 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003509
Christoph Lameter7656c722007-05-09 02:32:40 -07003510 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003511 if (!test_bit(slab_index(p, s, addr), map))
3512 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003513}
3514
3515static int list_locations(struct kmem_cache *s, char *buf,
3516 enum track_item alloc)
3517{
Harvey Harrisone374d482008-01-31 15:20:50 -08003518 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003519 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003520 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003521 int node;
3522
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003523 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003524 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003525 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003526
3527 /* Push back cpu slabs */
3528 flush_all(s);
3529
Christoph Lameterf64dc582007-10-16 01:25:33 -07003530 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003531 struct kmem_cache_node *n = get_node(s, node);
3532 unsigned long flags;
3533 struct page *page;
3534
Christoph Lameter9e869432007-08-22 14:01:56 -07003535 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003536 continue;
3537
3538 spin_lock_irqsave(&n->list_lock, flags);
3539 list_for_each_entry(page, &n->partial, lru)
3540 process_slab(&t, s, page, alloc);
3541 list_for_each_entry(page, &n->full, lru)
3542 process_slab(&t, s, page, alloc);
3543 spin_unlock_irqrestore(&n->list_lock, flags);
3544 }
3545
3546 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003547 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003548
Harvey Harrisone374d482008-01-31 15:20:50 -08003549 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003550 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003551 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003552
3553 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003554 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003555 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003556 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003557
3558 if (l->sum_time != l->min_time) {
3559 unsigned long remainder;
3560
Harvey Harrisone374d482008-01-31 15:20:50 -08003561 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003562 l->min_time,
3563 div_long_long_rem(l->sum_time, l->count, &remainder),
3564 l->max_time);
3565 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003566 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003567 l->min_time);
3568
3569 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003570 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003571 l->min_pid, l->max_pid);
3572 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003573 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003574 l->min_pid);
3575
Christoph Lameter84966342007-06-23 17:16:32 -07003576 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003577 len < PAGE_SIZE - 60) {
3578 len += sprintf(buf + len, " cpus=");
3579 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003580 l->cpus);
3581 }
3582
Christoph Lameter84966342007-06-23 17:16:32 -07003583 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003584 len < PAGE_SIZE - 60) {
3585 len += sprintf(buf + len, " nodes=");
3586 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003587 l->nodes);
3588 }
3589
Harvey Harrisone374d482008-01-31 15:20:50 -08003590 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003591 }
3592
3593 free_loc_track(&t);
3594 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003595 len += sprintf(buf, "No data\n");
3596 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003597}
3598
Christoph Lameter81819f02007-05-06 14:49:36 -07003599enum slab_stat_type {
3600 SL_FULL,
3601 SL_PARTIAL,
3602 SL_CPU,
3603 SL_OBJECTS
3604};
3605
3606#define SO_FULL (1 << SL_FULL)
3607#define SO_PARTIAL (1 << SL_PARTIAL)
3608#define SO_CPU (1 << SL_CPU)
3609#define SO_OBJECTS (1 << SL_OBJECTS)
3610
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003611static ssize_t show_slab_objects(struct kmem_cache *s,
3612 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003613{
3614 unsigned long total = 0;
3615 int cpu;
3616 int node;
3617 int x;
3618 unsigned long *nodes;
3619 unsigned long *per_cpu;
3620
3621 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003622 if (!nodes)
3623 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003624 per_cpu = nodes + nr_node_ids;
3625
3626 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003627 struct page *page;
3628 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003629
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003630 if (!c)
3631 continue;
3632
3633 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003634 node = c->node;
3635 if (node < 0)
3636 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003637 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003638 if (flags & SO_CPU) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003639 if (flags & SO_OBJECTS)
3640 x = page->inuse;
3641 else
3642 x = 1;
3643 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003644 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003645 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003646 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003647 }
3648 }
3649
Christoph Lameterf64dc582007-10-16 01:25:33 -07003650 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003651 struct kmem_cache_node *n = get_node(s, node);
3652
3653 if (flags & SO_PARTIAL) {
3654 if (flags & SO_OBJECTS)
3655 x = count_partial(n);
3656 else
3657 x = n->nr_partial;
3658 total += x;
3659 nodes[node] += x;
3660 }
3661
3662 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003663 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003664 - per_cpu[node]
3665 - n->nr_partial;
3666
3667 if (flags & SO_OBJECTS)
3668 x = full_slabs * s->objects;
3669 else
3670 x = full_slabs;
3671 total += x;
3672 nodes[node] += x;
3673 }
3674 }
3675
3676 x = sprintf(buf, "%lu", total);
3677#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003678 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003679 if (nodes[node])
3680 x += sprintf(buf + x, " N%d=%lu",
3681 node, nodes[node]);
3682#endif
3683 kfree(nodes);
3684 return x + sprintf(buf + x, "\n");
3685}
3686
3687static int any_slab_objects(struct kmem_cache *s)
3688{
3689 int node;
3690 int cpu;
3691
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003692 for_each_possible_cpu(cpu) {
3693 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003694
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003695 if (c && c->page)
3696 return 1;
3697 }
3698
3699 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003700 struct kmem_cache_node *n = get_node(s, node);
3701
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003702 if (!n)
3703 continue;
3704
Christoph Lameter9e869432007-08-22 14:01:56 -07003705 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003706 return 1;
3707 }
3708 return 0;
3709}
3710
3711#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3712#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3713
3714struct slab_attribute {
3715 struct attribute attr;
3716 ssize_t (*show)(struct kmem_cache *s, char *buf);
3717 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3718};
3719
3720#define SLAB_ATTR_RO(_name) \
3721 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3722
3723#define SLAB_ATTR(_name) \
3724 static struct slab_attribute _name##_attr = \
3725 __ATTR(_name, 0644, _name##_show, _name##_store)
3726
Christoph Lameter81819f02007-05-06 14:49:36 -07003727static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3728{
3729 return sprintf(buf, "%d\n", s->size);
3730}
3731SLAB_ATTR_RO(slab_size);
3732
3733static ssize_t align_show(struct kmem_cache *s, char *buf)
3734{
3735 return sprintf(buf, "%d\n", s->align);
3736}
3737SLAB_ATTR_RO(align);
3738
3739static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3740{
3741 return sprintf(buf, "%d\n", s->objsize);
3742}
3743SLAB_ATTR_RO(object_size);
3744
3745static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3746{
3747 return sprintf(buf, "%d\n", s->objects);
3748}
3749SLAB_ATTR_RO(objs_per_slab);
3750
3751static ssize_t order_show(struct kmem_cache *s, char *buf)
3752{
3753 return sprintf(buf, "%d\n", s->order);
3754}
3755SLAB_ATTR_RO(order);
3756
3757static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3758{
3759 if (s->ctor) {
3760 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3761
3762 return n + sprintf(buf + n, "\n");
3763 }
3764 return 0;
3765}
3766SLAB_ATTR_RO(ctor);
3767
Christoph Lameter81819f02007-05-06 14:49:36 -07003768static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3769{
3770 return sprintf(buf, "%d\n", s->refcount - 1);
3771}
3772SLAB_ATTR_RO(aliases);
3773
3774static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3775{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003776 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003777}
3778SLAB_ATTR_RO(slabs);
3779
3780static ssize_t partial_show(struct kmem_cache *s, char *buf)
3781{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003782 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003783}
3784SLAB_ATTR_RO(partial);
3785
3786static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3787{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003788 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003789}
3790SLAB_ATTR_RO(cpu_slabs);
3791
3792static ssize_t objects_show(struct kmem_cache *s, char *buf)
3793{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003794 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003795}
3796SLAB_ATTR_RO(objects);
3797
3798static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3799{
3800 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3801}
3802
3803static ssize_t sanity_checks_store(struct kmem_cache *s,
3804 const char *buf, size_t length)
3805{
3806 s->flags &= ~SLAB_DEBUG_FREE;
3807 if (buf[0] == '1')
3808 s->flags |= SLAB_DEBUG_FREE;
3809 return length;
3810}
3811SLAB_ATTR(sanity_checks);
3812
3813static ssize_t trace_show(struct kmem_cache *s, char *buf)
3814{
3815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3816}
3817
3818static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3819 size_t length)
3820{
3821 s->flags &= ~SLAB_TRACE;
3822 if (buf[0] == '1')
3823 s->flags |= SLAB_TRACE;
3824 return length;
3825}
3826SLAB_ATTR(trace);
3827
3828static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3829{
3830 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3831}
3832
3833static ssize_t reclaim_account_store(struct kmem_cache *s,
3834 const char *buf, size_t length)
3835{
3836 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3837 if (buf[0] == '1')
3838 s->flags |= SLAB_RECLAIM_ACCOUNT;
3839 return length;
3840}
3841SLAB_ATTR(reclaim_account);
3842
3843static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3844{
Christoph Lameter5af60832007-05-06 14:49:56 -07003845 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003846}
3847SLAB_ATTR_RO(hwcache_align);
3848
3849#ifdef CONFIG_ZONE_DMA
3850static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3851{
3852 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3853}
3854SLAB_ATTR_RO(cache_dma);
3855#endif
3856
3857static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3858{
3859 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3860}
3861SLAB_ATTR_RO(destroy_by_rcu);
3862
3863static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3864{
3865 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3866}
3867
3868static ssize_t red_zone_store(struct kmem_cache *s,
3869 const char *buf, size_t length)
3870{
3871 if (any_slab_objects(s))
3872 return -EBUSY;
3873
3874 s->flags &= ~SLAB_RED_ZONE;
3875 if (buf[0] == '1')
3876 s->flags |= SLAB_RED_ZONE;
3877 calculate_sizes(s);
3878 return length;
3879}
3880SLAB_ATTR(red_zone);
3881
3882static ssize_t poison_show(struct kmem_cache *s, char *buf)
3883{
3884 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3885}
3886
3887static ssize_t poison_store(struct kmem_cache *s,
3888 const char *buf, size_t length)
3889{
3890 if (any_slab_objects(s))
3891 return -EBUSY;
3892
3893 s->flags &= ~SLAB_POISON;
3894 if (buf[0] == '1')
3895 s->flags |= SLAB_POISON;
3896 calculate_sizes(s);
3897 return length;
3898}
3899SLAB_ATTR(poison);
3900
3901static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3902{
3903 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3904}
3905
3906static ssize_t store_user_store(struct kmem_cache *s,
3907 const char *buf, size_t length)
3908{
3909 if (any_slab_objects(s))
3910 return -EBUSY;
3911
3912 s->flags &= ~SLAB_STORE_USER;
3913 if (buf[0] == '1')
3914 s->flags |= SLAB_STORE_USER;
3915 calculate_sizes(s);
3916 return length;
3917}
3918SLAB_ATTR(store_user);
3919
Christoph Lameter53e15af2007-05-06 14:49:43 -07003920static ssize_t validate_show(struct kmem_cache *s, char *buf)
3921{
3922 return 0;
3923}
3924
3925static ssize_t validate_store(struct kmem_cache *s,
3926 const char *buf, size_t length)
3927{
Christoph Lameter434e2452007-07-17 04:03:30 -07003928 int ret = -EINVAL;
3929
3930 if (buf[0] == '1') {
3931 ret = validate_slab_cache(s);
3932 if (ret >= 0)
3933 ret = length;
3934 }
3935 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003936}
3937SLAB_ATTR(validate);
3938
Christoph Lameter2086d262007-05-06 14:49:46 -07003939static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3940{
3941 return 0;
3942}
3943
3944static ssize_t shrink_store(struct kmem_cache *s,
3945 const char *buf, size_t length)
3946{
3947 if (buf[0] == '1') {
3948 int rc = kmem_cache_shrink(s);
3949
3950 if (rc)
3951 return rc;
3952 } else
3953 return -EINVAL;
3954 return length;
3955}
3956SLAB_ATTR(shrink);
3957
Christoph Lameter88a420e2007-05-06 14:49:45 -07003958static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3959{
3960 if (!(s->flags & SLAB_STORE_USER))
3961 return -ENOSYS;
3962 return list_locations(s, buf, TRACK_ALLOC);
3963}
3964SLAB_ATTR_RO(alloc_calls);
3965
3966static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3967{
3968 if (!(s->flags & SLAB_STORE_USER))
3969 return -ENOSYS;
3970 return list_locations(s, buf, TRACK_FREE);
3971}
3972SLAB_ATTR_RO(free_calls);
3973
Christoph Lameter81819f02007-05-06 14:49:36 -07003974#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08003975static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07003976{
Christoph Lameter98246012008-01-07 23:20:26 -08003977 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07003978}
3979
Christoph Lameter98246012008-01-07 23:20:26 -08003980static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07003981 const char *buf, size_t length)
3982{
3983 int n = simple_strtoul(buf, NULL, 10);
3984
3985 if (n < 100)
Christoph Lameter98246012008-01-07 23:20:26 -08003986 s->remote_node_defrag_ratio = n * 10;
Christoph Lameter81819f02007-05-06 14:49:36 -07003987 return length;
3988}
Christoph Lameter98246012008-01-07 23:20:26 -08003989SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07003990#endif
3991
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003992#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003993static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3994{
3995 unsigned long sum = 0;
3996 int cpu;
3997 int len;
3998 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3999
4000 if (!data)
4001 return -ENOMEM;
4002
4003 for_each_online_cpu(cpu) {
4004 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4005
4006 data[cpu] = x;
4007 sum += x;
4008 }
4009
4010 len = sprintf(buf, "%lu", sum);
4011
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004012#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004013 for_each_online_cpu(cpu) {
4014 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004015 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004016 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004017#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004018 kfree(data);
4019 return len + sprintf(buf + len, "\n");
4020}
4021
4022#define STAT_ATTR(si, text) \
4023static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4024{ \
4025 return show_stat(s, buf, si); \
4026} \
4027SLAB_ATTR_RO(text); \
4028
4029STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4030STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4031STAT_ATTR(FREE_FASTPATH, free_fastpath);
4032STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4033STAT_ATTR(FREE_FROZEN, free_frozen);
4034STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4035STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4036STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4037STAT_ATTR(ALLOC_SLAB, alloc_slab);
4038STAT_ATTR(ALLOC_REFILL, alloc_refill);
4039STAT_ATTR(FREE_SLAB, free_slab);
4040STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4041STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4042STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4043STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4044STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4045STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4046
4047#endif
4048
Pekka Enberg06428782008-01-07 23:20:27 -08004049static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004050 &slab_size_attr.attr,
4051 &object_size_attr.attr,
4052 &objs_per_slab_attr.attr,
4053 &order_attr.attr,
4054 &objects_attr.attr,
4055 &slabs_attr.attr,
4056 &partial_attr.attr,
4057 &cpu_slabs_attr.attr,
4058 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004059 &aliases_attr.attr,
4060 &align_attr.attr,
4061 &sanity_checks_attr.attr,
4062 &trace_attr.attr,
4063 &hwcache_align_attr.attr,
4064 &reclaim_account_attr.attr,
4065 &destroy_by_rcu_attr.attr,
4066 &red_zone_attr.attr,
4067 &poison_attr.attr,
4068 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004069 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004070 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004071 &alloc_calls_attr.attr,
4072 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004073#ifdef CONFIG_ZONE_DMA
4074 &cache_dma_attr.attr,
4075#endif
4076#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004077 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004078#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004079#ifdef CONFIG_SLUB_STATS
4080 &alloc_fastpath_attr.attr,
4081 &alloc_slowpath_attr.attr,
4082 &free_fastpath_attr.attr,
4083 &free_slowpath_attr.attr,
4084 &free_frozen_attr.attr,
4085 &free_add_partial_attr.attr,
4086 &free_remove_partial_attr.attr,
4087 &alloc_from_partial_attr.attr,
4088 &alloc_slab_attr.attr,
4089 &alloc_refill_attr.attr,
4090 &free_slab_attr.attr,
4091 &cpuslab_flush_attr.attr,
4092 &deactivate_full_attr.attr,
4093 &deactivate_empty_attr.attr,
4094 &deactivate_to_head_attr.attr,
4095 &deactivate_to_tail_attr.attr,
4096 &deactivate_remote_frees_attr.attr,
4097#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004098 NULL
4099};
4100
4101static struct attribute_group slab_attr_group = {
4102 .attrs = slab_attrs,
4103};
4104
4105static ssize_t slab_attr_show(struct kobject *kobj,
4106 struct attribute *attr,
4107 char *buf)
4108{
4109 struct slab_attribute *attribute;
4110 struct kmem_cache *s;
4111 int err;
4112
4113 attribute = to_slab_attr(attr);
4114 s = to_slab(kobj);
4115
4116 if (!attribute->show)
4117 return -EIO;
4118
4119 err = attribute->show(s, buf);
4120
4121 return err;
4122}
4123
4124static ssize_t slab_attr_store(struct kobject *kobj,
4125 struct attribute *attr,
4126 const char *buf, size_t len)
4127{
4128 struct slab_attribute *attribute;
4129 struct kmem_cache *s;
4130 int err;
4131
4132 attribute = to_slab_attr(attr);
4133 s = to_slab(kobj);
4134
4135 if (!attribute->store)
4136 return -EIO;
4137
4138 err = attribute->store(s, buf, len);
4139
4140 return err;
4141}
4142
Christoph Lameter151c6022008-01-07 22:29:05 -08004143static void kmem_cache_release(struct kobject *kobj)
4144{
4145 struct kmem_cache *s = to_slab(kobj);
4146
4147 kfree(s);
4148}
4149
Christoph Lameter81819f02007-05-06 14:49:36 -07004150static struct sysfs_ops slab_sysfs_ops = {
4151 .show = slab_attr_show,
4152 .store = slab_attr_store,
4153};
4154
4155static struct kobj_type slab_ktype = {
4156 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004157 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004158};
4159
4160static int uevent_filter(struct kset *kset, struct kobject *kobj)
4161{
4162 struct kobj_type *ktype = get_ktype(kobj);
4163
4164 if (ktype == &slab_ktype)
4165 return 1;
4166 return 0;
4167}
4168
4169static struct kset_uevent_ops slab_uevent_ops = {
4170 .filter = uevent_filter,
4171};
4172
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004173static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004174
4175#define ID_STR_LENGTH 64
4176
4177/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004178 *
4179 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004180 */
4181static char *create_unique_id(struct kmem_cache *s)
4182{
4183 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4184 char *p = name;
4185
4186 BUG_ON(!name);
4187
4188 *p++ = ':';
4189 /*
4190 * First flags affecting slabcache operations. We will only
4191 * get here for aliasable slabs so we do not need to support
4192 * too many flags. The flags here must cover all flags that
4193 * are matched during merging to guarantee that the id is
4194 * unique.
4195 */
4196 if (s->flags & SLAB_CACHE_DMA)
4197 *p++ = 'd';
4198 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4199 *p++ = 'a';
4200 if (s->flags & SLAB_DEBUG_FREE)
4201 *p++ = 'F';
4202 if (p != name + 1)
4203 *p++ = '-';
4204 p += sprintf(p, "%07d", s->size);
4205 BUG_ON(p > name + ID_STR_LENGTH - 1);
4206 return name;
4207}
4208
4209static int sysfs_slab_add(struct kmem_cache *s)
4210{
4211 int err;
4212 const char *name;
4213 int unmergeable;
4214
4215 if (slab_state < SYSFS)
4216 /* Defer until later */
4217 return 0;
4218
4219 unmergeable = slab_unmergeable(s);
4220 if (unmergeable) {
4221 /*
4222 * Slabcache can never be merged so we can use the name proper.
4223 * This is typically the case for debug situations. In that
4224 * case we can catch duplicate names easily.
4225 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004226 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004227 name = s->name;
4228 } else {
4229 /*
4230 * Create a unique name for the slab as a target
4231 * for the symlinks.
4232 */
4233 name = create_unique_id(s);
4234 }
4235
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004236 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004237 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4238 if (err) {
4239 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004240 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004241 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004242
4243 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4244 if (err)
4245 return err;
4246 kobject_uevent(&s->kobj, KOBJ_ADD);
4247 if (!unmergeable) {
4248 /* Setup first alias */
4249 sysfs_slab_alias(s, s->name);
4250 kfree(name);
4251 }
4252 return 0;
4253}
4254
4255static void sysfs_slab_remove(struct kmem_cache *s)
4256{
4257 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4258 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004259 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004260}
4261
4262/*
4263 * Need to buffer aliases during bootup until sysfs becomes
4264 * available lest we loose that information.
4265 */
4266struct saved_alias {
4267 struct kmem_cache *s;
4268 const char *name;
4269 struct saved_alias *next;
4270};
4271
Adrian Bunk5af328a2007-07-17 04:03:27 -07004272static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004273
4274static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4275{
4276 struct saved_alias *al;
4277
4278 if (slab_state == SYSFS) {
4279 /*
4280 * If we have a leftover link then remove it.
4281 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004282 sysfs_remove_link(&slab_kset->kobj, name);
4283 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004284 }
4285
4286 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4287 if (!al)
4288 return -ENOMEM;
4289
4290 al->s = s;
4291 al->name = name;
4292 al->next = alias_list;
4293 alias_list = al;
4294 return 0;
4295}
4296
4297static int __init slab_sysfs_init(void)
4298{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004299 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004300 int err;
4301
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004302 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004303 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004304 printk(KERN_ERR "Cannot register slab subsystem.\n");
4305 return -ENOSYS;
4306 }
4307
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004308 slab_state = SYSFS;
4309
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004310 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004311 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004312 if (err)
4313 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4314 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004315 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004316
4317 while (alias_list) {
4318 struct saved_alias *al = alias_list;
4319
4320 alias_list = alias_list->next;
4321 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004322 if (err)
4323 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4324 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004325 kfree(al);
4326 }
4327
4328 resiliency_test();
4329 return 0;
4330}
4331
4332__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004333#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004334
4335/*
4336 * The /proc/slabinfo ABI
4337 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004338#ifdef CONFIG_SLABINFO
4339
4340ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4341 size_t count, loff_t *ppos)
4342{
4343 return -EINVAL;
4344}
4345
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004346
4347static void print_slabinfo_header(struct seq_file *m)
4348{
4349 seq_puts(m, "slabinfo - version: 2.1\n");
4350 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4351 "<objperslab> <pagesperslab>");
4352 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4353 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4354 seq_putc(m, '\n');
4355}
4356
4357static void *s_start(struct seq_file *m, loff_t *pos)
4358{
4359 loff_t n = *pos;
4360
4361 down_read(&slub_lock);
4362 if (!n)
4363 print_slabinfo_header(m);
4364
4365 return seq_list_start(&slab_caches, *pos);
4366}
4367
4368static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4369{
4370 return seq_list_next(p, &slab_caches, pos);
4371}
4372
4373static void s_stop(struct seq_file *m, void *p)
4374{
4375 up_read(&slub_lock);
4376}
4377
4378static int s_show(struct seq_file *m, void *p)
4379{
4380 unsigned long nr_partials = 0;
4381 unsigned long nr_slabs = 0;
4382 unsigned long nr_inuse = 0;
4383 unsigned long nr_objs;
4384 struct kmem_cache *s;
4385 int node;
4386
4387 s = list_entry(p, struct kmem_cache, list);
4388
4389 for_each_online_node(node) {
4390 struct kmem_cache_node *n = get_node(s, node);
4391
4392 if (!n)
4393 continue;
4394
4395 nr_partials += n->nr_partial;
4396 nr_slabs += atomic_long_read(&n->nr_slabs);
4397 nr_inuse += count_partial(n);
4398 }
4399
4400 nr_objs = nr_slabs * s->objects;
4401 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4402
4403 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4404 nr_objs, s->size, s->objects, (1 << s->order));
4405 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4406 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4407 0UL);
4408 seq_putc(m, '\n');
4409 return 0;
4410}
4411
4412const struct seq_operations slabinfo_op = {
4413 .start = s_start,
4414 .next = s_next,
4415 .stop = s_stop,
4416 .show = s_show,
4417};
4418
Linus Torvalds158a9622008-01-02 13:04:48 -08004419#endif /* CONFIG_SLABINFO */