blob: 39592b5ce68adc07493ee8ae077f1dcd98839080 [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;
1288 struct page *page;
1289
1290 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001291 * The defrag ratio allows a configuration of the tradeoffs between
1292 * inter node defragmentation and node local allocations. A lower
1293 * defrag_ratio increases the tendency to do local allocations
1294 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001295 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001296 * If the defrag_ratio is set to 0 then kmalloc() always
1297 * returns node local objects. If the ratio is higher then kmalloc()
1298 * may return off node objects because partial slabs are obtained
1299 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001300 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001301 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001302 * defrag_ratio = 1000) then every (well almost) allocation will
1303 * first attempt to defrag slab caches on other nodes. This means
1304 * scanning over all nodes to look for partial slabs which may be
1305 * expensive if we do it every time we are trying to find a slab
1306 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001307 */
Christoph Lameter98246012008-01-07 23:20:26 -08001308 if (!s->remote_node_defrag_ratio ||
1309 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001310 return NULL;
1311
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001312 zonelist = &NODE_DATA(
1313 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
Christoph Lameter81819f02007-05-06 14:49:36 -07001314 for (z = zonelist->zones; *z; z++) {
1315 struct kmem_cache_node *n;
1316
1317 n = get_node(s, zone_to_nid(*z));
1318
1319 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001320 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001321 page = get_partial_node(n);
1322 if (page)
1323 return page;
1324 }
1325 }
1326#endif
1327 return NULL;
1328}
1329
1330/*
1331 * Get a partial page, lock it and return it.
1332 */
1333static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1334{
1335 struct page *page;
1336 int searchnode = (node == -1) ? numa_node_id() : node;
1337
1338 page = get_partial_node(get_node(s, searchnode));
1339 if (page || (flags & __GFP_THISNODE))
1340 return page;
1341
1342 return get_any_partial(s, flags);
1343}
1344
1345/*
1346 * Move a page back to the lists.
1347 *
1348 * Must be called with the slab lock held.
1349 *
1350 * On exit the slab lock will have been dropped.
1351 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001352static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001353{
Christoph Lametere95eed52007-05-06 14:49:44 -07001354 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001355 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001356
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001357 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001358 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001359
Christoph Lametera973e9d2008-03-01 13:40:44 -08001360 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001361 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001362 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1363 } else {
1364 stat(c, DEACTIVATE_FULL);
1365 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1366 add_full(n, page);
1367 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001368 slab_unlock(page);
1369 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001370 stat(c, DEACTIVATE_EMPTY);
Christoph Lametere95eed52007-05-06 14:49:44 -07001371 if (n->nr_partial < MIN_PARTIAL) {
1372 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001373 * Adding an empty slab to the partial slabs in order
1374 * to avoid page allocator overhead. This slab needs
1375 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001376 * so that the others get filled first. That way the
1377 * size of the partial list stays small.
1378 *
1379 * kmem_cache_shrink can reclaim any empty slabs from the
1380 * partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001381 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001382 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001383 slab_unlock(page);
1384 } else {
1385 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001386 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001387 discard_slab(s, page);
1388 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001389 }
1390}
1391
1392/*
1393 * Remove the cpu slab
1394 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001395static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001396{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001397 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001398 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001399
Christoph Lameterb773ad72008-03-04 11:10:17 -08001400 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001401 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001402 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001403 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001404 * because both freelists are empty. So this is unlikely
1405 * to occur.
1406 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001407 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001408 void **object;
1409
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001410 tail = 0; /* Hot objects. Put the slab first */
1411
Christoph Lameter894b8782007-05-10 03:15:16 -07001412 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001413 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001414 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001415
1416 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001417 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001418 page->freelist = object;
1419 page->inuse--;
1420 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001421 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001422 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001423}
1424
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001425static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001426{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001427 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001428 slab_lock(c->page);
1429 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001430}
1431
1432/*
1433 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001434 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001435 * Called from IPI handler with interrupts disabled.
1436 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001437static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001438{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001439 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001440
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001441 if (likely(c && c->page))
1442 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001443}
1444
1445static void flush_cpu_slab(void *d)
1446{
1447 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001448
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001449 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001450}
1451
1452static void flush_all(struct kmem_cache *s)
1453{
1454#ifdef CONFIG_SMP
1455 on_each_cpu(flush_cpu_slab, s, 1, 1);
1456#else
1457 unsigned long flags;
1458
1459 local_irq_save(flags);
1460 flush_cpu_slab(s);
1461 local_irq_restore(flags);
1462#endif
1463}
1464
1465/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001466 * Check if the objects in a per cpu structure fit numa
1467 * locality expectations.
1468 */
1469static inline int node_match(struct kmem_cache_cpu *c, int node)
1470{
1471#ifdef CONFIG_NUMA
1472 if (node != -1 && c->node != node)
1473 return 0;
1474#endif
1475 return 1;
1476}
1477
1478/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001479 * Slow path. The lockless freelist is empty or we need to perform
1480 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001481 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001482 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001483 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001484 * Processing is still very fast if new objects have been freed to the
1485 * regular freelist. In that case we simply take over the regular freelist
1486 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001487 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001488 * If that is not working then we fall back to the partial lists. We take the
1489 * first element of the freelist as the object to allocate now and move the
1490 * rest of the freelist to the lockless freelist.
1491 *
1492 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001493 * we need to allocate a new slab. This is the slowest path since it involves
1494 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001495 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001496static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001497 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001498{
Christoph Lameter81819f02007-05-06 14:49:36 -07001499 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001500 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001501
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001502 /* We handle __GFP_ZERO in the caller */
1503 gfpflags &= ~__GFP_ZERO;
1504
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001505 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001506 goto new_slab;
1507
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001508 slab_lock(c->page);
1509 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001510 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001511
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001512 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001513
Christoph Lameter894b8782007-05-10 03:15:16 -07001514load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001515 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001516 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001517 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001518 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001519 goto debug;
1520
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001521 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001522 c->page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001523 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001524 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001525unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001526 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001527 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001528 return object;
1529
1530another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001531 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001532
1533new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001534 new = get_partial(s, gfpflags, node);
1535 if (new) {
1536 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001537 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001538 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001539 }
1540
Christoph Lameterb811c202007-10-16 23:25:51 -07001541 if (gfpflags & __GFP_WAIT)
1542 local_irq_enable();
1543
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001544 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001545
1546 if (gfpflags & __GFP_WAIT)
1547 local_irq_disable();
1548
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001549 if (new) {
1550 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001551 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001552 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001553 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001554 slab_lock(new);
1555 SetSlabFrozen(new);
1556 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001557 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001558 }
Linus Torvalds00e962c2008-02-19 09:08:49 -08001559
Christoph Lameter71c7a062008-02-14 14:28:01 -08001560 /*
1561 * No memory available.
1562 *
1563 * If the slab uses higher order allocs but the object is
1564 * smaller than a page size then we can fallback in emergencies
1565 * to the page allocator via kmalloc_large. The page allocator may
1566 * have failed to obtain a higher order page and we can try to
1567 * allocate a single page if the object fits into a single page.
1568 * That is only possible if certain conditions are met that are being
1569 * checked when a slab is created.
1570 */
Christoph Lametercaeab082008-03-12 23:57:49 -07001571 if (!(gfpflags & __GFP_NORETRY) &&
1572 (s->flags & __PAGE_ALLOC_FALLBACK)) {
1573 if (gfpflags & __GFP_WAIT)
1574 local_irq_enable();
1575 object = kmalloc_large(s->objsize, gfpflags);
1576 if (gfpflags & __GFP_WAIT)
1577 local_irq_disable();
1578 return object;
1579 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001580 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001581debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001582 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001583 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001584
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001585 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001586 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001587 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001588 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001589}
1590
1591/*
1592 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1593 * have the fastpath folded into their functions. So no function call
1594 * overhead for requests that can be satisfied on the fastpath.
1595 *
1596 * The fastpath works by first checking if the lockless freelist can be used.
1597 * If not then __slab_alloc is called for slow processing.
1598 *
1599 * Otherwise we can simply pick the next object from the lockless free list.
1600 */
Pekka Enberg06428782008-01-07 23:20:27 -08001601static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001602 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001603{
Christoph Lameter894b8782007-05-10 03:15:16 -07001604 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001605 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001606 unsigned long flags;
1607
Christoph Lameter894b8782007-05-10 03:15:16 -07001608 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001609 c = get_cpu_slab(s, smp_processor_id());
Christoph Lametera973e9d2008-03-01 13:40:44 -08001610 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001611
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001612 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001613
1614 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001615 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001616 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001617 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001618 }
1619 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001620
1621 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001622 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001623
Christoph Lameter894b8782007-05-10 03:15:16 -07001624 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001625}
1626
1627void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1628{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001629 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001630}
1631EXPORT_SYMBOL(kmem_cache_alloc);
1632
1633#ifdef CONFIG_NUMA
1634void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1635{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001636 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001637}
1638EXPORT_SYMBOL(kmem_cache_alloc_node);
1639#endif
1640
1641/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001642 * Slow patch handling. This may still be called frequently since objects
1643 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001644 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001645 * So we still attempt to reduce cache line usage. Just take the slab
1646 * lock and free the item. If there is no additional partial page
1647 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001648 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001649static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001650 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001651{
1652 void *prior;
1653 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001654 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001655
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001656 c = get_cpu_slab(s, raw_smp_processor_id());
1657 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001658 slab_lock(page);
1659
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001660 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001661 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001662
Christoph Lameter81819f02007-05-06 14:49:36 -07001663checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001664 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001665 page->freelist = object;
1666 page->inuse--;
1667
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001668 if (unlikely(SlabFrozen(page))) {
1669 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001670 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001671 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001672
1673 if (unlikely(!page->inuse))
1674 goto slab_empty;
1675
1676 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001677 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001678 * then add it.
1679 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001680 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001681 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001682 stat(c, FREE_ADD_PARTIAL);
1683 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001684
1685out_unlock:
1686 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001687 return;
1688
1689slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001690 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001691 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001692 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001693 */
1694 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001695 stat(c, FREE_REMOVE_PARTIAL);
1696 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001697 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001698 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001699 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001700 return;
1701
1702debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001703 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001704 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001705 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001706}
1707
Christoph Lameter894b8782007-05-10 03:15:16 -07001708/*
1709 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1710 * can perform fastpath freeing without additional function calls.
1711 *
1712 * The fastpath is only possible if we are freeing to the current cpu slab
1713 * of this processor. This typically the case if we have just allocated
1714 * the item before.
1715 *
1716 * If fastpath is not possible then fall back to __slab_free where we deal
1717 * with all sorts of special processing.
1718 */
Pekka Enberg06428782008-01-07 23:20:27 -08001719static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001720 struct page *page, void *x, void *addr)
1721{
1722 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001723 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001724 unsigned long flags;
1725
Christoph Lameter894b8782007-05-10 03:15:16 -07001726 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001727 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001728 debug_check_no_locks_freed(object, c->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001729 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001730 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001731 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001732 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001733 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001734 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001735
1736 local_irq_restore(flags);
1737}
1738
Christoph Lameter81819f02007-05-06 14:49:36 -07001739void kmem_cache_free(struct kmem_cache *s, void *x)
1740{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001741 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001742
Christoph Lameterb49af682007-05-06 14:49:41 -07001743 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001744
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001745 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001746}
1747EXPORT_SYMBOL(kmem_cache_free);
1748
1749/* Figure out on which slab object the object resides */
1750static struct page *get_object_page(const void *x)
1751{
Christoph Lameterb49af682007-05-06 14:49:41 -07001752 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001753
1754 if (!PageSlab(page))
1755 return NULL;
1756
1757 return page;
1758}
1759
1760/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001761 * Object placement in a slab is made very easy because we always start at
1762 * offset 0. If we tune the size of the object to the alignment then we can
1763 * get the required alignment by putting one properly sized object after
1764 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001765 *
1766 * Notice that the allocation order determines the sizes of the per cpu
1767 * caches. Each processor has always one slab available for allocations.
1768 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001769 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001770 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001771 */
1772
1773/*
1774 * Mininum / Maximum order of slab pages. This influences locking overhead
1775 * and slab fragmentation. A higher order reduces the number of partial slabs
1776 * and increases the number of allocations possible without having to
1777 * take the list_lock.
1778 */
1779static int slub_min_order;
1780static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001781static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1782
1783/*
1784 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001785 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001786 */
1787static int slub_nomerge;
1788
1789/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001790 * Calculate the order of allocation given an slab object size.
1791 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001792 * The order of allocation has significant impact on performance and other
1793 * system components. Generally order 0 allocations should be preferred since
1794 * order 0 does not cause fragmentation in the page allocator. Larger objects
1795 * be problematic to put into order 0 slabs because there may be too much
1796 * unused space left. We go to a higher order if more than 1/8th of the slab
1797 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001798 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001799 * In order to reach satisfactory performance we must ensure that a minimum
1800 * number of objects is in one slab. Otherwise we may generate too much
1801 * activity on the partial lists which requires taking the list_lock. This is
1802 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001803 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001804 * slub_max_order specifies the order where we begin to stop considering the
1805 * number of objects in a slab as critical. If we reach slub_max_order then
1806 * we try to keep the page order as low as possible. So we accept more waste
1807 * of space in favor of a small page order.
1808 *
1809 * Higher order allocations also allow the placement of more objects in a
1810 * slab and thereby reduce object handling overhead. If the user has
1811 * requested a higher mininum order then we start with that one instead of
1812 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001813 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001814static inline int slab_order(int size, int min_objects,
1815 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001816{
1817 int order;
1818 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001819 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001820
Christoph Lameter6300ea72007-07-17 04:03:20 -07001821 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001822 fls(min_objects * size - 1) - PAGE_SHIFT);
1823 order <= max_order; order++) {
1824
Christoph Lameter81819f02007-05-06 14:49:36 -07001825 unsigned long slab_size = PAGE_SIZE << order;
1826
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001827 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001828 continue;
1829
Christoph Lameter81819f02007-05-06 14:49:36 -07001830 rem = slab_size % size;
1831
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001832 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001833 break;
1834
1835 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001836
Christoph Lameter81819f02007-05-06 14:49:36 -07001837 return order;
1838}
1839
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001840static inline int calculate_order(int size)
1841{
1842 int order;
1843 int min_objects;
1844 int fraction;
1845
1846 /*
1847 * Attempt to find best configuration for a slab. This
1848 * works by first attempting to generate a layout with
1849 * the best configuration and backing off gradually.
1850 *
1851 * First we reduce the acceptable waste in a slab. Then
1852 * we reduce the minimum objects required in a slab.
1853 */
1854 min_objects = slub_min_objects;
1855 while (min_objects > 1) {
1856 fraction = 8;
1857 while (fraction >= 4) {
1858 order = slab_order(size, min_objects,
1859 slub_max_order, fraction);
1860 if (order <= slub_max_order)
1861 return order;
1862 fraction /= 2;
1863 }
1864 min_objects /= 2;
1865 }
1866
1867 /*
1868 * We were unable to place multiple objects in a slab. Now
1869 * lets see if we can place a single object there.
1870 */
1871 order = slab_order(size, 1, slub_max_order, 1);
1872 if (order <= slub_max_order)
1873 return order;
1874
1875 /*
1876 * Doh this slab cannot be placed using slub_max_order.
1877 */
1878 order = slab_order(size, 1, MAX_ORDER, 1);
1879 if (order <= MAX_ORDER)
1880 return order;
1881 return -ENOSYS;
1882}
1883
Christoph Lameter81819f02007-05-06 14:49:36 -07001884/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001885 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001886 */
1887static unsigned long calculate_alignment(unsigned long flags,
1888 unsigned long align, unsigned long size)
1889{
1890 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001891 * If the user wants hardware cache aligned objects then follow that
1892 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001893 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001894 * The hardware cache alignment cannot override the specified
1895 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001896 */
Nick Pigginb6210382008-03-05 14:05:56 -08001897 if (flags & SLAB_HWCACHE_ALIGN) {
1898 unsigned long ralign = cache_line_size();
1899 while (size <= ralign / 2)
1900 ralign /= 2;
1901 align = max(align, ralign);
1902 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001903
1904 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001905 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001906
1907 return ALIGN(align, sizeof(void *));
1908}
1909
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001910static void init_kmem_cache_cpu(struct kmem_cache *s,
1911 struct kmem_cache_cpu *c)
1912{
1913 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001914 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001915 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001916 c->offset = s->offset / sizeof(void *);
1917 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03001918#ifdef CONFIG_SLUB_STATS
1919 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1920#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001921}
1922
Christoph Lameter81819f02007-05-06 14:49:36 -07001923static void init_kmem_cache_node(struct kmem_cache_node *n)
1924{
1925 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001926 spin_lock_init(&n->list_lock);
1927 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001928#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001929 atomic_long_set(&n->nr_slabs, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07001930 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001931#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001932}
1933
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001934#ifdef CONFIG_SMP
1935/*
1936 * Per cpu array for per cpu structures.
1937 *
1938 * The per cpu array places all kmem_cache_cpu structures from one processor
1939 * close together meaning that it becomes possible that multiple per cpu
1940 * structures are contained in one cacheline. This may be particularly
1941 * beneficial for the kmalloc caches.
1942 *
1943 * A desktop system typically has around 60-80 slabs. With 100 here we are
1944 * likely able to get per cpu structures for all caches from the array defined
1945 * here. We must be able to cover all kmalloc caches during bootstrap.
1946 *
1947 * If the per cpu array is exhausted then fall back to kmalloc
1948 * of individual cachelines. No sharing is possible then.
1949 */
1950#define NR_KMEM_CACHE_CPU 100
1951
1952static DEFINE_PER_CPU(struct kmem_cache_cpu,
1953 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1954
1955static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1956static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1957
1958static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1959 int cpu, gfp_t flags)
1960{
1961 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1962
1963 if (c)
1964 per_cpu(kmem_cache_cpu_free, cpu) =
1965 (void *)c->freelist;
1966 else {
1967 /* Table overflow: So allocate ourselves */
1968 c = kmalloc_node(
1969 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1970 flags, cpu_to_node(cpu));
1971 if (!c)
1972 return NULL;
1973 }
1974
1975 init_kmem_cache_cpu(s, c);
1976 return c;
1977}
1978
1979static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1980{
1981 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1982 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1983 kfree(c);
1984 return;
1985 }
1986 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1987 per_cpu(kmem_cache_cpu_free, cpu) = c;
1988}
1989
1990static void free_kmem_cache_cpus(struct kmem_cache *s)
1991{
1992 int cpu;
1993
1994 for_each_online_cpu(cpu) {
1995 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1996
1997 if (c) {
1998 s->cpu_slab[cpu] = NULL;
1999 free_kmem_cache_cpu(c, cpu);
2000 }
2001 }
2002}
2003
2004static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2005{
2006 int cpu;
2007
2008 for_each_online_cpu(cpu) {
2009 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2010
2011 if (c)
2012 continue;
2013
2014 c = alloc_kmem_cache_cpu(s, cpu, flags);
2015 if (!c) {
2016 free_kmem_cache_cpus(s);
2017 return 0;
2018 }
2019 s->cpu_slab[cpu] = c;
2020 }
2021 return 1;
2022}
2023
2024/*
2025 * Initialize the per cpu array.
2026 */
2027static void init_alloc_cpu_cpu(int cpu)
2028{
2029 int i;
2030
2031 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2032 return;
2033
2034 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2035 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2036
2037 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2038}
2039
2040static void __init init_alloc_cpu(void)
2041{
2042 int cpu;
2043
2044 for_each_online_cpu(cpu)
2045 init_alloc_cpu_cpu(cpu);
2046 }
2047
2048#else
2049static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2050static inline void init_alloc_cpu(void) {}
2051
2052static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2053{
2054 init_kmem_cache_cpu(s, &s->cpu_slab);
2055 return 1;
2056}
2057#endif
2058
Christoph Lameter81819f02007-05-06 14:49:36 -07002059#ifdef CONFIG_NUMA
2060/*
2061 * No kmalloc_node yet so do it by hand. We know that this is the first
2062 * slab on the node for this slabcache. There are no concurrent accesses
2063 * possible.
2064 *
2065 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002066 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2067 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002068 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002069static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2070 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002071{
2072 struct page *page;
2073 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002074 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002075
2076 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2077
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002078 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002079
2080 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002081 if (page_to_nid(page) != node) {
2082 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2083 "node %d\n", node);
2084 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2085 "in order to be able to continue\n");
2086 }
2087
Christoph Lameter81819f02007-05-06 14:49:36 -07002088 n = page->freelist;
2089 BUG_ON(!n);
2090 page->freelist = get_freepointer(kmalloc_caches, n);
2091 page->inuse++;
2092 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002093#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002094 init_object(kmalloc_caches, n, 1);
2095 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002096#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002097 init_kmem_cache_node(n);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002098 inc_slabs_node(kmalloc_caches, node);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002099
rootba84c732008-01-07 23:20:28 -08002100 /*
2101 * lockdep requires consistent irq usage for each lock
2102 * so even though there cannot be a race this early in
2103 * the boot sequence, we still disable irqs.
2104 */
2105 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002106 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002107 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002108 return n;
2109}
2110
2111static void free_kmem_cache_nodes(struct kmem_cache *s)
2112{
2113 int node;
2114
Christoph Lameterf64dc582007-10-16 01:25:33 -07002115 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002116 struct kmem_cache_node *n = s->node[node];
2117 if (n && n != &s->local_node)
2118 kmem_cache_free(kmalloc_caches, n);
2119 s->node[node] = NULL;
2120 }
2121}
2122
2123static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2124{
2125 int node;
2126 int local_node;
2127
2128 if (slab_state >= UP)
2129 local_node = page_to_nid(virt_to_page(s));
2130 else
2131 local_node = 0;
2132
Christoph Lameterf64dc582007-10-16 01:25:33 -07002133 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002134 struct kmem_cache_node *n;
2135
2136 if (local_node == node)
2137 n = &s->local_node;
2138 else {
2139 if (slab_state == DOWN) {
2140 n = early_kmem_cache_node_alloc(gfpflags,
2141 node);
2142 continue;
2143 }
2144 n = kmem_cache_alloc_node(kmalloc_caches,
2145 gfpflags, node);
2146
2147 if (!n) {
2148 free_kmem_cache_nodes(s);
2149 return 0;
2150 }
2151
2152 }
2153 s->node[node] = n;
2154 init_kmem_cache_node(n);
2155 }
2156 return 1;
2157}
2158#else
2159static void free_kmem_cache_nodes(struct kmem_cache *s)
2160{
2161}
2162
2163static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2164{
2165 init_kmem_cache_node(&s->local_node);
2166 return 1;
2167}
2168#endif
2169
2170/*
2171 * calculate_sizes() determines the order and the distribution of data within
2172 * a slab object.
2173 */
2174static int calculate_sizes(struct kmem_cache *s)
2175{
2176 unsigned long flags = s->flags;
2177 unsigned long size = s->objsize;
2178 unsigned long align = s->align;
2179
2180 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002181 * Round up object size to the next word boundary. We can only
2182 * place the free pointer at word boundaries and this determines
2183 * the possible location of the free pointer.
2184 */
2185 size = ALIGN(size, sizeof(void *));
2186
2187#ifdef CONFIG_SLUB_DEBUG
2188 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002189 * Determine if we can poison the object itself. If the user of
2190 * the slab may touch the object after free or before allocation
2191 * then we should never poison the object itself.
2192 */
2193 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002194 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002195 s->flags |= __OBJECT_POISON;
2196 else
2197 s->flags &= ~__OBJECT_POISON;
2198
Christoph Lameter81819f02007-05-06 14:49:36 -07002199
2200 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002201 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002202 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002203 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002204 */
2205 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2206 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002207#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002208
2209 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002210 * With that we have determined the number of bytes in actual use
2211 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002212 */
2213 s->inuse = size;
2214
2215 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002216 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002217 /*
2218 * Relocate free pointer after the object if it is not
2219 * permitted to overwrite the first word of the object on
2220 * kmem_cache_free.
2221 *
2222 * This is the case if we do RCU, have a constructor or
2223 * destructor or are poisoning the objects.
2224 */
2225 s->offset = size;
2226 size += sizeof(void *);
2227 }
2228
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002229#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002230 if (flags & SLAB_STORE_USER)
2231 /*
2232 * Need to store information about allocs and frees after
2233 * the object.
2234 */
2235 size += 2 * sizeof(struct track);
2236
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002237 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002238 /*
2239 * Add some empty padding so that we can catch
2240 * overwrites from earlier objects rather than let
2241 * tracking information or the free pointer be
2242 * corrupted if an user writes before the start
2243 * of the object.
2244 */
2245 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002246#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002247
Christoph Lameter81819f02007-05-06 14:49:36 -07002248 /*
2249 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002250 * user specified and the dynamic determination of cache line size
2251 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002252 */
2253 align = calculate_alignment(flags, align, s->objsize);
2254
2255 /*
2256 * SLUB stores one object immediately after another beginning from
2257 * offset 0. In order to align the objects we have to simply size
2258 * each object to conform to the alignment.
2259 */
2260 size = ALIGN(size, align);
2261 s->size = size;
2262
Christoph Lameter71c7a062008-02-14 14:28:01 -08002263 if ((flags & __KMALLOC_CACHE) &&
2264 PAGE_SIZE / size < slub_min_objects) {
2265 /*
2266 * Kmalloc cache that would not have enough objects in
2267 * an order 0 page. Kmalloc slabs can fallback to
2268 * page allocator order 0 allocs so take a reasonably large
2269 * order that will allows us a good number of objects.
2270 */
2271 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2272 s->flags |= __PAGE_ALLOC_FALLBACK;
2273 s->allocflags |= __GFP_NOWARN;
2274 } else
2275 s->order = calculate_order(size);
2276
Christoph Lameter81819f02007-05-06 14:49:36 -07002277 if (s->order < 0)
2278 return 0;
2279
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002280 s->allocflags = 0;
2281 if (s->order)
2282 s->allocflags |= __GFP_COMP;
2283
2284 if (s->flags & SLAB_CACHE_DMA)
2285 s->allocflags |= SLUB_DMA;
2286
2287 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2288 s->allocflags |= __GFP_RECLAIMABLE;
2289
Christoph Lameter81819f02007-05-06 14:49:36 -07002290 /*
2291 * Determine the number of objects per slab
2292 */
2293 s->objects = (PAGE_SIZE << s->order) / size;
2294
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002295 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002296
2297}
2298
Christoph Lameter81819f02007-05-06 14:49:36 -07002299static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2300 const char *name, size_t size,
2301 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002302 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002303{
2304 memset(s, 0, kmem_size);
2305 s->name = name;
2306 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002307 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002308 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002309 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002310
2311 if (!calculate_sizes(s))
2312 goto error;
2313
2314 s->refcount = 1;
2315#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002316 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002317#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002318 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2319 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002320
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002321 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002322 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002323 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002324error:
2325 if (flags & SLAB_PANIC)
2326 panic("Cannot create slab %s size=%lu realsize=%u "
2327 "order=%u offset=%u flags=%lx\n",
2328 s->name, (unsigned long)size, s->size, s->order,
2329 s->offset, flags);
2330 return 0;
2331}
Christoph Lameter81819f02007-05-06 14:49:36 -07002332
2333/*
2334 * Check if a given pointer is valid
2335 */
2336int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2337{
Pekka Enberg06428782008-01-07 23:20:27 -08002338 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002339
2340 page = get_object_page(object);
2341
2342 if (!page || s != page->slab)
2343 /* No slab or wrong slab */
2344 return 0;
2345
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002346 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002347 return 0;
2348
2349 /*
2350 * We could also check if the object is on the slabs freelist.
2351 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002352 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002353 * to a certain slab.
2354 */
2355 return 1;
2356}
2357EXPORT_SYMBOL(kmem_ptr_validate);
2358
2359/*
2360 * Determine the size of a slab object
2361 */
2362unsigned int kmem_cache_size(struct kmem_cache *s)
2363{
2364 return s->objsize;
2365}
2366EXPORT_SYMBOL(kmem_cache_size);
2367
2368const char *kmem_cache_name(struct kmem_cache *s)
2369{
2370 return s->name;
2371}
2372EXPORT_SYMBOL(kmem_cache_name);
2373
2374/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002375 * Attempt to free all slabs on a node. Return the number of slabs we
2376 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002377 */
2378static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2379 struct list_head *list)
2380{
2381 int slabs_inuse = 0;
2382 unsigned long flags;
2383 struct page *page, *h;
2384
2385 spin_lock_irqsave(&n->list_lock, flags);
2386 list_for_each_entry_safe(page, h, list, lru)
2387 if (!page->inuse) {
2388 list_del(&page->lru);
2389 discard_slab(s, page);
2390 } else
2391 slabs_inuse++;
2392 spin_unlock_irqrestore(&n->list_lock, flags);
2393 return slabs_inuse;
2394}
2395
2396/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002397 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002398 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002399static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002400{
2401 int node;
2402
2403 flush_all(s);
2404
2405 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002406 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002407 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002408 struct kmem_cache_node *n = get_node(s, node);
2409
Christoph Lameter2086d262007-05-06 14:49:46 -07002410 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002411 if (slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002412 return 1;
2413 }
2414 free_kmem_cache_nodes(s);
2415 return 0;
2416}
2417
2418/*
2419 * Close a cache and release the kmem_cache structure
2420 * (must be used for caches created using kmem_cache_create)
2421 */
2422void kmem_cache_destroy(struct kmem_cache *s)
2423{
2424 down_write(&slub_lock);
2425 s->refcount--;
2426 if (!s->refcount) {
2427 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002428 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002429 if (kmem_cache_close(s))
2430 WARN_ON(1);
2431 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002432 } else
2433 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002434}
2435EXPORT_SYMBOL(kmem_cache_destroy);
2436
2437/********************************************************************
2438 * Kmalloc subsystem
2439 *******************************************************************/
2440
Christoph Lameter331dc552008-02-14 14:28:09 -08002441struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002442EXPORT_SYMBOL(kmalloc_caches);
2443
Christoph Lameter81819f02007-05-06 14:49:36 -07002444static int __init setup_slub_min_order(char *str)
2445{
Pekka Enberg06428782008-01-07 23:20:27 -08002446 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002447
2448 return 1;
2449}
2450
2451__setup("slub_min_order=", setup_slub_min_order);
2452
2453static int __init setup_slub_max_order(char *str)
2454{
Pekka Enberg06428782008-01-07 23:20:27 -08002455 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002456
2457 return 1;
2458}
2459
2460__setup("slub_max_order=", setup_slub_max_order);
2461
2462static int __init setup_slub_min_objects(char *str)
2463{
Pekka Enberg06428782008-01-07 23:20:27 -08002464 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002465
2466 return 1;
2467}
2468
2469__setup("slub_min_objects=", setup_slub_min_objects);
2470
2471static int __init setup_slub_nomerge(char *str)
2472{
2473 slub_nomerge = 1;
2474 return 1;
2475}
2476
2477__setup("slub_nomerge", setup_slub_nomerge);
2478
Christoph Lameter81819f02007-05-06 14:49:36 -07002479static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2480 const char *name, int size, gfp_t gfp_flags)
2481{
2482 unsigned int flags = 0;
2483
2484 if (gfp_flags & SLUB_DMA)
2485 flags = SLAB_CACHE_DMA;
2486
2487 down_write(&slub_lock);
2488 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter71c7a062008-02-14 14:28:01 -08002489 flags | __KMALLOC_CACHE, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002490 goto panic;
2491
2492 list_add(&s->list, &slab_caches);
2493 up_write(&slub_lock);
2494 if (sysfs_slab_add(s))
2495 goto panic;
2496 return s;
2497
2498panic:
2499 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2500}
2501
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002502#ifdef CONFIG_ZONE_DMA
Christoph Lameter4097d602008-04-14 18:51:18 +03002503static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002504
2505static void sysfs_add_func(struct work_struct *w)
2506{
2507 struct kmem_cache *s;
2508
2509 down_write(&slub_lock);
2510 list_for_each_entry(s, &slab_caches, list) {
2511 if (s->flags & __SYSFS_ADD_DEFERRED) {
2512 s->flags &= ~__SYSFS_ADD_DEFERRED;
2513 sysfs_slab_add(s);
2514 }
2515 }
2516 up_write(&slub_lock);
2517}
2518
2519static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2520
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002521static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2522{
2523 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002524 char *text;
2525 size_t realsize;
2526
2527 s = kmalloc_caches_dma[index];
2528 if (s)
2529 return s;
2530
2531 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002532 if (flags & __GFP_WAIT)
2533 down_write(&slub_lock);
2534 else {
2535 if (!down_write_trylock(&slub_lock))
2536 goto out;
2537 }
2538
2539 if (kmalloc_caches_dma[index])
2540 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002541
Christoph Lameter7b55f622007-07-17 04:03:27 -07002542 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002543 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2544 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002545 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2546
2547 if (!s || !text || !kmem_cache_open(s, flags, text,
2548 realsize, ARCH_KMALLOC_MINALIGN,
2549 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2550 kfree(s);
2551 kfree(text);
2552 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002553 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002554
2555 list_add(&s->list, &slab_caches);
2556 kmalloc_caches_dma[index] = s;
2557
2558 schedule_work(&sysfs_add_work);
2559
2560unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002561 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002562out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002563 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002564}
2565#endif
2566
Christoph Lameterf1b26332007-07-17 04:03:26 -07002567/*
2568 * Conversion table for small slabs sizes / 8 to the index in the
2569 * kmalloc array. This is necessary for slabs < 192 since we have non power
2570 * of two cache sizes there. The size of larger slabs can be determined using
2571 * fls.
2572 */
2573static s8 size_index[24] = {
2574 3, /* 8 */
2575 4, /* 16 */
2576 5, /* 24 */
2577 5, /* 32 */
2578 6, /* 40 */
2579 6, /* 48 */
2580 6, /* 56 */
2581 6, /* 64 */
2582 1, /* 72 */
2583 1, /* 80 */
2584 1, /* 88 */
2585 1, /* 96 */
2586 7, /* 104 */
2587 7, /* 112 */
2588 7, /* 120 */
2589 7, /* 128 */
2590 2, /* 136 */
2591 2, /* 144 */
2592 2, /* 152 */
2593 2, /* 160 */
2594 2, /* 168 */
2595 2, /* 176 */
2596 2, /* 184 */
2597 2 /* 192 */
2598};
2599
Christoph Lameter81819f02007-05-06 14:49:36 -07002600static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2601{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002602 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002603
Christoph Lameterf1b26332007-07-17 04:03:26 -07002604 if (size <= 192) {
2605 if (!size)
2606 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002607
Christoph Lameterf1b26332007-07-17 04:03:26 -07002608 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002609 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002610 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002611
2612#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002613 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002614 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002615
Christoph Lameter81819f02007-05-06 14:49:36 -07002616#endif
2617 return &kmalloc_caches[index];
2618}
2619
2620void *__kmalloc(size_t size, gfp_t flags)
2621{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002622 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002623
Christoph Lameter331dc552008-02-14 14:28:09 -08002624 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002625 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002626
2627 s = get_slab(size, flags);
2628
2629 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002630 return s;
2631
Christoph Lameterce15fea2007-07-17 04:03:28 -07002632 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002633}
2634EXPORT_SYMBOL(__kmalloc);
2635
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002636static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2637{
2638 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2639 get_order(size));
2640
2641 if (page)
2642 return page_address(page);
2643 else
2644 return NULL;
2645}
2646
Christoph Lameter81819f02007-05-06 14:49:36 -07002647#ifdef CONFIG_NUMA
2648void *__kmalloc_node(size_t size, gfp_t flags, int node)
2649{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002650 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002651
Christoph Lameter331dc552008-02-14 14:28:09 -08002652 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002653 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002654
2655 s = get_slab(size, flags);
2656
2657 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002658 return s;
2659
Christoph Lameterce15fea2007-07-17 04:03:28 -07002660 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002661}
2662EXPORT_SYMBOL(__kmalloc_node);
2663#endif
2664
2665size_t ksize(const void *object)
2666{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002667 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002668 struct kmem_cache *s;
2669
Christoph Lameteref8b4522007-10-16 01:24:46 -07002670 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002671 return 0;
2672
Vegard Nossum294a80a2007-12-04 23:45:30 -08002673 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002674
2675 if (unlikely(!PageSlab(page)))
2676 return PAGE_SIZE << compound_order(page);
2677
Christoph Lameter81819f02007-05-06 14:49:36 -07002678 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002679
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002680#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002681 /*
2682 * Debugging requires use of the padding between object
2683 * and whatever may come after it.
2684 */
2685 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2686 return s->objsize;
2687
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002688#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002689 /*
2690 * If we have the need to store the freelist pointer
2691 * back there or track user information then we can
2692 * only use the space before that information.
2693 */
2694 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2695 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002696 /*
2697 * Else we can use all the padding etc for the allocation
2698 */
2699 return s->size;
2700}
2701EXPORT_SYMBOL(ksize);
2702
2703void kfree(const void *x)
2704{
Christoph Lameter81819f02007-05-06 14:49:36 -07002705 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002706 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002707
Satyam Sharma2408c552007-10-16 01:24:44 -07002708 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002709 return;
2710
Christoph Lameterb49af682007-05-06 14:49:41 -07002711 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002712 if (unlikely(!PageSlab(page))) {
2713 put_page(page);
2714 return;
2715 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002716 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002717}
2718EXPORT_SYMBOL(kfree);
2719
Christoph Lameter2086d262007-05-06 14:49:46 -07002720/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002721 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2722 * the remaining slabs by the number of items in use. The slabs with the
2723 * most items in use come first. New allocations will then fill those up
2724 * and thus they can be removed from the partial lists.
2725 *
2726 * The slabs with the least items are placed last. This results in them
2727 * being allocated from last increasing the chance that the last objects
2728 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002729 */
2730int kmem_cache_shrink(struct kmem_cache *s)
2731{
2732 int node;
2733 int i;
2734 struct kmem_cache_node *n;
2735 struct page *page;
2736 struct page *t;
2737 struct list_head *slabs_by_inuse =
2738 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2739 unsigned long flags;
2740
2741 if (!slabs_by_inuse)
2742 return -ENOMEM;
2743
2744 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002745 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002746 n = get_node(s, node);
2747
2748 if (!n->nr_partial)
2749 continue;
2750
2751 for (i = 0; i < s->objects; i++)
2752 INIT_LIST_HEAD(slabs_by_inuse + i);
2753
2754 spin_lock_irqsave(&n->list_lock, flags);
2755
2756 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002757 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002758 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002759 * Note that concurrent frees may occur while we hold the
2760 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002761 */
2762 list_for_each_entry_safe(page, t, &n->partial, lru) {
2763 if (!page->inuse && slab_trylock(page)) {
2764 /*
2765 * Must hold slab lock here because slab_free
2766 * may have freed the last object and be
2767 * waiting to release the slab.
2768 */
2769 list_del(&page->lru);
2770 n->nr_partial--;
2771 slab_unlock(page);
2772 discard_slab(s, page);
2773 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002774 list_move(&page->lru,
2775 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002776 }
2777 }
2778
Christoph Lameter2086d262007-05-06 14:49:46 -07002779 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002780 * Rebuild the partial list with the slabs filled up most
2781 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002782 */
2783 for (i = s->objects - 1; i >= 0; i--)
2784 list_splice(slabs_by_inuse + i, n->partial.prev);
2785
Christoph Lameter2086d262007-05-06 14:49:46 -07002786 spin_unlock_irqrestore(&n->list_lock, flags);
2787 }
2788
2789 kfree(slabs_by_inuse);
2790 return 0;
2791}
2792EXPORT_SYMBOL(kmem_cache_shrink);
2793
Yasunori Gotob9049e22007-10-21 16:41:37 -07002794#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2795static int slab_mem_going_offline_callback(void *arg)
2796{
2797 struct kmem_cache *s;
2798
2799 down_read(&slub_lock);
2800 list_for_each_entry(s, &slab_caches, list)
2801 kmem_cache_shrink(s);
2802 up_read(&slub_lock);
2803
2804 return 0;
2805}
2806
2807static void slab_mem_offline_callback(void *arg)
2808{
2809 struct kmem_cache_node *n;
2810 struct kmem_cache *s;
2811 struct memory_notify *marg = arg;
2812 int offline_node;
2813
2814 offline_node = marg->status_change_nid;
2815
2816 /*
2817 * If the node still has available memory. we need kmem_cache_node
2818 * for it yet.
2819 */
2820 if (offline_node < 0)
2821 return;
2822
2823 down_read(&slub_lock);
2824 list_for_each_entry(s, &slab_caches, list) {
2825 n = get_node(s, offline_node);
2826 if (n) {
2827 /*
2828 * if n->nr_slabs > 0, slabs still exist on the node
2829 * that is going down. We were unable to free them,
2830 * and offline_pages() function shoudn't call this
2831 * callback. So, we must fail.
2832 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002833 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002834
2835 s->node[offline_node] = NULL;
2836 kmem_cache_free(kmalloc_caches, n);
2837 }
2838 }
2839 up_read(&slub_lock);
2840}
2841
2842static int slab_mem_going_online_callback(void *arg)
2843{
2844 struct kmem_cache_node *n;
2845 struct kmem_cache *s;
2846 struct memory_notify *marg = arg;
2847 int nid = marg->status_change_nid;
2848 int ret = 0;
2849
2850 /*
2851 * If the node's memory is already available, then kmem_cache_node is
2852 * already created. Nothing to do.
2853 */
2854 if (nid < 0)
2855 return 0;
2856
2857 /*
2858 * We are bringing a node online. No memory is availabe yet. We must
2859 * allocate a kmem_cache_node structure in order to bring the node
2860 * online.
2861 */
2862 down_read(&slub_lock);
2863 list_for_each_entry(s, &slab_caches, list) {
2864 /*
2865 * XXX: kmem_cache_alloc_node will fallback to other nodes
2866 * since memory is not yet available from the node that
2867 * is brought up.
2868 */
2869 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2870 if (!n) {
2871 ret = -ENOMEM;
2872 goto out;
2873 }
2874 init_kmem_cache_node(n);
2875 s->node[nid] = n;
2876 }
2877out:
2878 up_read(&slub_lock);
2879 return ret;
2880}
2881
2882static int slab_memory_callback(struct notifier_block *self,
2883 unsigned long action, void *arg)
2884{
2885 int ret = 0;
2886
2887 switch (action) {
2888 case MEM_GOING_ONLINE:
2889 ret = slab_mem_going_online_callback(arg);
2890 break;
2891 case MEM_GOING_OFFLINE:
2892 ret = slab_mem_going_offline_callback(arg);
2893 break;
2894 case MEM_OFFLINE:
2895 case MEM_CANCEL_ONLINE:
2896 slab_mem_offline_callback(arg);
2897 break;
2898 case MEM_ONLINE:
2899 case MEM_CANCEL_OFFLINE:
2900 break;
2901 }
2902
2903 ret = notifier_from_errno(ret);
2904 return ret;
2905}
2906
2907#endif /* CONFIG_MEMORY_HOTPLUG */
2908
Christoph Lameter81819f02007-05-06 14:49:36 -07002909/********************************************************************
2910 * Basic setup of slabs
2911 *******************************************************************/
2912
2913void __init kmem_cache_init(void)
2914{
2915 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002916 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002917
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002918 init_alloc_cpu();
2919
Christoph Lameter81819f02007-05-06 14:49:36 -07002920#ifdef CONFIG_NUMA
2921 /*
2922 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002923 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002924 * kmem_cache_open for slab_state == DOWN.
2925 */
2926 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2927 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002928 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002929 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002930
2931 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002932#endif
2933
2934 /* Able to allocate the per node structures */
2935 slab_state = PARTIAL;
2936
2937 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002938 if (KMALLOC_MIN_SIZE <= 64) {
2939 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002940 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002941 caches++;
2942 }
2943 if (KMALLOC_MIN_SIZE <= 128) {
2944 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002945 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002946 caches++;
2947 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002948
Christoph Lameter331dc552008-02-14 14:28:09 -08002949 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002950 create_kmalloc_cache(&kmalloc_caches[i],
2951 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002952 caches++;
2953 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002954
Christoph Lameterf1b26332007-07-17 04:03:26 -07002955
2956 /*
2957 * Patch up the size_index table if we have strange large alignment
2958 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002959 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002960 *
2961 * Largest permitted alignment is 256 bytes due to the way we
2962 * handle the index determination for the smaller caches.
2963 *
2964 * Make sure that nothing crazy happens if someone starts tinkering
2965 * around with ARCH_KMALLOC_MINALIGN
2966 */
2967 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2968 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2969
Christoph Lameter12ad6842007-07-17 04:03:28 -07002970 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002971 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2972
Christoph Lameter81819f02007-05-06 14:49:36 -07002973 slab_state = UP;
2974
2975 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08002976 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002977 kmalloc_caches[i]. name =
2978 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2979
2980#ifdef CONFIG_SMP
2981 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002982 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2983 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2984#else
2985 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002986#endif
2987
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002988 printk(KERN_INFO
2989 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002990 " CPUs=%d, Nodes=%d\n",
2991 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002992 slub_min_order, slub_max_order, slub_min_objects,
2993 nr_cpu_ids, nr_node_ids);
2994}
2995
2996/*
2997 * Find a mergeable slab cache
2998 */
2999static int slab_unmergeable(struct kmem_cache *s)
3000{
3001 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3002 return 1;
3003
Christoph Lameter331dc552008-02-14 14:28:09 -08003004 if ((s->flags & __PAGE_ALLOC_FALLBACK))
Christoph Lameter71c7a062008-02-14 14:28:01 -08003005 return 1;
3006
Christoph Lameterc59def92007-05-16 22:10:50 -07003007 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003008 return 1;
3009
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003010 /*
3011 * We may have set a slab to be unmergeable during bootstrap.
3012 */
3013 if (s->refcount < 0)
3014 return 1;
3015
Christoph Lameter81819f02007-05-06 14:49:36 -07003016 return 0;
3017}
3018
3019static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003020 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003021 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003022{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003023 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003024
3025 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3026 return NULL;
3027
Christoph Lameterc59def92007-05-16 22:10:50 -07003028 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003029 return NULL;
3030
3031 size = ALIGN(size, sizeof(void *));
3032 align = calculate_alignment(flags, align, size);
3033 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003034 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003035
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003036 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003037 if (slab_unmergeable(s))
3038 continue;
3039
3040 if (size > s->size)
3041 continue;
3042
Christoph Lameterba0268a2007-09-11 15:24:11 -07003043 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003044 continue;
3045 /*
3046 * Check if alignment is compatible.
3047 * Courtesy of Adrian Drzewiecki
3048 */
Pekka Enberg06428782008-01-07 23:20:27 -08003049 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003050 continue;
3051
3052 if (s->size - size >= sizeof(void *))
3053 continue;
3054
3055 return s;
3056 }
3057 return NULL;
3058}
3059
3060struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3061 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003062 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003063{
3064 struct kmem_cache *s;
3065
3066 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003067 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003068 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003069 int cpu;
3070
Christoph Lameter81819f02007-05-06 14:49:36 -07003071 s->refcount++;
3072 /*
3073 * Adjust the object sizes so that we clear
3074 * the complete object on kzalloc.
3075 */
3076 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003077
3078 /*
3079 * And then we need to update the object size in the
3080 * per cpu structures
3081 */
3082 for_each_online_cpu(cpu)
3083 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003084
Christoph Lameter81819f02007-05-06 14:49:36 -07003085 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003086 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003087
Christoph Lameter81819f02007-05-06 14:49:36 -07003088 if (sysfs_slab_alias(s, name))
3089 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003090 return s;
3091 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003092
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003093 s = kmalloc(kmem_size, GFP_KERNEL);
3094 if (s) {
3095 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003096 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003097 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003098 up_write(&slub_lock);
3099 if (sysfs_slab_add(s))
3100 goto err;
3101 return s;
3102 }
3103 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003104 }
3105 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003106
3107err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003108 if (flags & SLAB_PANIC)
3109 panic("Cannot create slabcache %s\n", name);
3110 else
3111 s = NULL;
3112 return s;
3113}
3114EXPORT_SYMBOL(kmem_cache_create);
3115
Christoph Lameter81819f02007-05-06 14:49:36 -07003116#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003117/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003118 * Use the cpu notifier to insure that the cpu slabs are flushed when
3119 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003120 */
3121static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3122 unsigned long action, void *hcpu)
3123{
3124 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003125 struct kmem_cache *s;
3126 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003127
3128 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003129 case CPU_UP_PREPARE:
3130 case CPU_UP_PREPARE_FROZEN:
3131 init_alloc_cpu_cpu(cpu);
3132 down_read(&slub_lock);
3133 list_for_each_entry(s, &slab_caches, list)
3134 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3135 GFP_KERNEL);
3136 up_read(&slub_lock);
3137 break;
3138
Christoph Lameter81819f02007-05-06 14:49:36 -07003139 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003140 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003141 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003142 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003143 down_read(&slub_lock);
3144 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003145 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3146
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003147 local_irq_save(flags);
3148 __flush_cpu_slab(s, cpu);
3149 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003150 free_kmem_cache_cpu(c, cpu);
3151 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003152 }
3153 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003154 break;
3155 default:
3156 break;
3157 }
3158 return NOTIFY_OK;
3159}
3160
Pekka Enberg06428782008-01-07 23:20:27 -08003161static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003162 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003163};
Christoph Lameter81819f02007-05-06 14:49:36 -07003164
3165#endif
3166
Christoph Lameter81819f02007-05-06 14:49:36 -07003167void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3168{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003169 struct kmem_cache *s;
3170
Christoph Lameter331dc552008-02-14 14:28:09 -08003171 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003172 return kmalloc_large(size, gfpflags);
3173
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003174 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003175
Satyam Sharma2408c552007-10-16 01:24:44 -07003176 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003177 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003178
Christoph Lameterce15fea2007-07-17 04:03:28 -07003179 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003180}
3181
3182void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3183 int node, void *caller)
3184{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003185 struct kmem_cache *s;
3186
Christoph Lameter331dc552008-02-14 14:28:09 -08003187 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003188 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003189
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003190 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003191
Satyam Sharma2408c552007-10-16 01:24:44 -07003192 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003193 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003194
Christoph Lameterce15fea2007-07-17 04:03:28 -07003195 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003196}
3197
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003198#if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3199static unsigned long count_partial(struct kmem_cache_node *n)
3200{
3201 unsigned long flags;
3202 unsigned long x = 0;
3203 struct page *page;
3204
3205 spin_lock_irqsave(&n->list_lock, flags);
3206 list_for_each_entry(page, &n->partial, lru)
3207 x += page->inuse;
3208 spin_unlock_irqrestore(&n->list_lock, flags);
3209 return x;
3210}
3211#endif
3212
Christoph Lameter41ecc552007-05-09 02:32:44 -07003213#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003214static int validate_slab(struct kmem_cache *s, struct page *page,
3215 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003216{
3217 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003218 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003219
3220 if (!check_slab(s, page) ||
3221 !on_freelist(s, page, NULL))
3222 return 0;
3223
3224 /* Now we know that a valid freelist exists */
3225 bitmap_zero(map, s->objects);
3226
Christoph Lameter7656c722007-05-09 02:32:40 -07003227 for_each_free_object(p, s, page->freelist) {
3228 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003229 if (!check_object(s, page, p, 0))
3230 return 0;
3231 }
3232
Christoph Lameter7656c722007-05-09 02:32:40 -07003233 for_each_object(p, s, addr)
3234 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003235 if (!check_object(s, page, p, 1))
3236 return 0;
3237 return 1;
3238}
3239
Christoph Lameter434e2452007-07-17 04:03:30 -07003240static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3241 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003242{
3243 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003244 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003245 slab_unlock(page);
3246 } else
3247 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3248 s->name, page);
3249
3250 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003251 if (!SlabDebug(page))
3252 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003253 "on slab 0x%p\n", s->name, page);
3254 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003255 if (SlabDebug(page))
3256 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003257 "slab 0x%p\n", s->name, page);
3258 }
3259}
3260
Christoph Lameter434e2452007-07-17 04:03:30 -07003261static int validate_slab_node(struct kmem_cache *s,
3262 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003263{
3264 unsigned long count = 0;
3265 struct page *page;
3266 unsigned long flags;
3267
3268 spin_lock_irqsave(&n->list_lock, flags);
3269
3270 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003271 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003272 count++;
3273 }
3274 if (count != n->nr_partial)
3275 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3276 "counter=%ld\n", s->name, count, n->nr_partial);
3277
3278 if (!(s->flags & SLAB_STORE_USER))
3279 goto out;
3280
3281 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003282 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003283 count++;
3284 }
3285 if (count != atomic_long_read(&n->nr_slabs))
3286 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3287 "counter=%ld\n", s->name, count,
3288 atomic_long_read(&n->nr_slabs));
3289
3290out:
3291 spin_unlock_irqrestore(&n->list_lock, flags);
3292 return count;
3293}
3294
Christoph Lameter434e2452007-07-17 04:03:30 -07003295static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003296{
3297 int node;
3298 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003299 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3300 sizeof(unsigned long), GFP_KERNEL);
3301
3302 if (!map)
3303 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003304
3305 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003306 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003307 struct kmem_cache_node *n = get_node(s, node);
3308
Christoph Lameter434e2452007-07-17 04:03:30 -07003309 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003310 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003311 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003312 return count;
3313}
3314
Christoph Lameterb3459702007-05-09 02:32:41 -07003315#ifdef SLUB_RESILIENCY_TEST
3316static void resiliency_test(void)
3317{
3318 u8 *p;
3319
3320 printk(KERN_ERR "SLUB resiliency testing\n");
3321 printk(KERN_ERR "-----------------------\n");
3322 printk(KERN_ERR "A. Corruption after allocation\n");
3323
3324 p = kzalloc(16, GFP_KERNEL);
3325 p[16] = 0x12;
3326 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3327 " 0x12->0x%p\n\n", p + 16);
3328
3329 validate_slab_cache(kmalloc_caches + 4);
3330
3331 /* Hmmm... The next two are dangerous */
3332 p = kzalloc(32, GFP_KERNEL);
3333 p[32 + sizeof(void *)] = 0x34;
3334 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003335 " 0x34 -> -0x%p\n", p);
3336 printk(KERN_ERR
3337 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003338
3339 validate_slab_cache(kmalloc_caches + 5);
3340 p = kzalloc(64, GFP_KERNEL);
3341 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3342 *p = 0x56;
3343 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3344 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003345 printk(KERN_ERR
3346 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003347 validate_slab_cache(kmalloc_caches + 6);
3348
3349 printk(KERN_ERR "\nB. Corruption after free\n");
3350 p = kzalloc(128, GFP_KERNEL);
3351 kfree(p);
3352 *p = 0x78;
3353 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3354 validate_slab_cache(kmalloc_caches + 7);
3355
3356 p = kzalloc(256, GFP_KERNEL);
3357 kfree(p);
3358 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003359 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3360 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003361 validate_slab_cache(kmalloc_caches + 8);
3362
3363 p = kzalloc(512, GFP_KERNEL);
3364 kfree(p);
3365 p[512] = 0xab;
3366 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3367 validate_slab_cache(kmalloc_caches + 9);
3368}
3369#else
3370static void resiliency_test(void) {};
3371#endif
3372
Christoph Lameter88a420e2007-05-06 14:49:45 -07003373/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003374 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003375 * and freed.
3376 */
3377
3378struct location {
3379 unsigned long count;
3380 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003381 long long sum_time;
3382 long min_time;
3383 long max_time;
3384 long min_pid;
3385 long max_pid;
3386 cpumask_t cpus;
3387 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003388};
3389
3390struct loc_track {
3391 unsigned long max;
3392 unsigned long count;
3393 struct location *loc;
3394};
3395
3396static void free_loc_track(struct loc_track *t)
3397{
3398 if (t->max)
3399 free_pages((unsigned long)t->loc,
3400 get_order(sizeof(struct location) * t->max));
3401}
3402
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003403static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003404{
3405 struct location *l;
3406 int order;
3407
Christoph Lameter88a420e2007-05-06 14:49:45 -07003408 order = get_order(sizeof(struct location) * max);
3409
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003410 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003411 if (!l)
3412 return 0;
3413
3414 if (t->count) {
3415 memcpy(l, t->loc, sizeof(struct location) * t->count);
3416 free_loc_track(t);
3417 }
3418 t->max = max;
3419 t->loc = l;
3420 return 1;
3421}
3422
3423static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003424 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003425{
3426 long start, end, pos;
3427 struct location *l;
3428 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003429 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003430
3431 start = -1;
3432 end = t->count;
3433
3434 for ( ; ; ) {
3435 pos = start + (end - start + 1) / 2;
3436
3437 /*
3438 * There is nothing at "end". If we end up there
3439 * we need to add something to before end.
3440 */
3441 if (pos == end)
3442 break;
3443
3444 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003445 if (track->addr == caddr) {
3446
3447 l = &t->loc[pos];
3448 l->count++;
3449 if (track->when) {
3450 l->sum_time += age;
3451 if (age < l->min_time)
3452 l->min_time = age;
3453 if (age > l->max_time)
3454 l->max_time = age;
3455
3456 if (track->pid < l->min_pid)
3457 l->min_pid = track->pid;
3458 if (track->pid > l->max_pid)
3459 l->max_pid = track->pid;
3460
3461 cpu_set(track->cpu, l->cpus);
3462 }
3463 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003464 return 1;
3465 }
3466
Christoph Lameter45edfa52007-05-09 02:32:45 -07003467 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003468 end = pos;
3469 else
3470 start = pos;
3471 }
3472
3473 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003474 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003475 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003476 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003477 return 0;
3478
3479 l = t->loc + pos;
3480 if (pos < t->count)
3481 memmove(l + 1, l,
3482 (t->count - pos) * sizeof(struct location));
3483 t->count++;
3484 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003485 l->addr = track->addr;
3486 l->sum_time = age;
3487 l->min_time = age;
3488 l->max_time = age;
3489 l->min_pid = track->pid;
3490 l->max_pid = track->pid;
3491 cpus_clear(l->cpus);
3492 cpu_set(track->cpu, l->cpus);
3493 nodes_clear(l->nodes);
3494 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003495 return 1;
3496}
3497
3498static void process_slab(struct loc_track *t, struct kmem_cache *s,
3499 struct page *page, enum track_item alloc)
3500{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003501 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003502 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003503 void *p;
3504
3505 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003506 for_each_free_object(p, s, page->freelist)
3507 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003508
Christoph Lameter7656c722007-05-09 02:32:40 -07003509 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003510 if (!test_bit(slab_index(p, s, addr), map))
3511 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003512}
3513
3514static int list_locations(struct kmem_cache *s, char *buf,
3515 enum track_item alloc)
3516{
Harvey Harrisone374d482008-01-31 15:20:50 -08003517 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003518 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003519 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003520 int node;
3521
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003522 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003523 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003524 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003525
3526 /* Push back cpu slabs */
3527 flush_all(s);
3528
Christoph Lameterf64dc582007-10-16 01:25:33 -07003529 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003530 struct kmem_cache_node *n = get_node(s, node);
3531 unsigned long flags;
3532 struct page *page;
3533
Christoph Lameter9e869432007-08-22 14:01:56 -07003534 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003535 continue;
3536
3537 spin_lock_irqsave(&n->list_lock, flags);
3538 list_for_each_entry(page, &n->partial, lru)
3539 process_slab(&t, s, page, alloc);
3540 list_for_each_entry(page, &n->full, lru)
3541 process_slab(&t, s, page, alloc);
3542 spin_unlock_irqrestore(&n->list_lock, flags);
3543 }
3544
3545 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003546 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003547
Harvey Harrisone374d482008-01-31 15:20:50 -08003548 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003549 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003550 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003551
3552 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003553 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003554 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003555 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003556
3557 if (l->sum_time != l->min_time) {
3558 unsigned long remainder;
3559
Harvey Harrisone374d482008-01-31 15:20:50 -08003560 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003561 l->min_time,
3562 div_long_long_rem(l->sum_time, l->count, &remainder),
3563 l->max_time);
3564 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003565 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003566 l->min_time);
3567
3568 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003569 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003570 l->min_pid, l->max_pid);
3571 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003572 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003573 l->min_pid);
3574
Christoph Lameter84966342007-06-23 17:16:32 -07003575 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003576 len < PAGE_SIZE - 60) {
3577 len += sprintf(buf + len, " cpus=");
3578 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003579 l->cpus);
3580 }
3581
Christoph Lameter84966342007-06-23 17:16:32 -07003582 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003583 len < PAGE_SIZE - 60) {
3584 len += sprintf(buf + len, " nodes=");
3585 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003586 l->nodes);
3587 }
3588
Harvey Harrisone374d482008-01-31 15:20:50 -08003589 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003590 }
3591
3592 free_loc_track(&t);
3593 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003594 len += sprintf(buf, "No data\n");
3595 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003596}
3597
Christoph Lameter81819f02007-05-06 14:49:36 -07003598enum slab_stat_type {
3599 SL_FULL,
3600 SL_PARTIAL,
3601 SL_CPU,
3602 SL_OBJECTS
3603};
3604
3605#define SO_FULL (1 << SL_FULL)
3606#define SO_PARTIAL (1 << SL_PARTIAL)
3607#define SO_CPU (1 << SL_CPU)
3608#define SO_OBJECTS (1 << SL_OBJECTS)
3609
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003610static ssize_t show_slab_objects(struct kmem_cache *s,
3611 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003612{
3613 unsigned long total = 0;
3614 int cpu;
3615 int node;
3616 int x;
3617 unsigned long *nodes;
3618 unsigned long *per_cpu;
3619
3620 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003621 if (!nodes)
3622 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003623 per_cpu = nodes + nr_node_ids;
3624
3625 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003626 struct page *page;
3627 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003628
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003629 if (!c)
3630 continue;
3631
3632 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003633 node = c->node;
3634 if (node < 0)
3635 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003636 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003637 if (flags & SO_CPU) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003638 if (flags & SO_OBJECTS)
3639 x = page->inuse;
3640 else
3641 x = 1;
3642 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003643 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003644 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003645 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003646 }
3647 }
3648
Christoph Lameterf64dc582007-10-16 01:25:33 -07003649 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003650 struct kmem_cache_node *n = get_node(s, node);
3651
3652 if (flags & SO_PARTIAL) {
3653 if (flags & SO_OBJECTS)
3654 x = count_partial(n);
3655 else
3656 x = n->nr_partial;
3657 total += x;
3658 nodes[node] += x;
3659 }
3660
3661 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003662 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003663 - per_cpu[node]
3664 - n->nr_partial;
3665
3666 if (flags & SO_OBJECTS)
3667 x = full_slabs * s->objects;
3668 else
3669 x = full_slabs;
3670 total += x;
3671 nodes[node] += x;
3672 }
3673 }
3674
3675 x = sprintf(buf, "%lu", total);
3676#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003677 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003678 if (nodes[node])
3679 x += sprintf(buf + x, " N%d=%lu",
3680 node, nodes[node]);
3681#endif
3682 kfree(nodes);
3683 return x + sprintf(buf + x, "\n");
3684}
3685
3686static int any_slab_objects(struct kmem_cache *s)
3687{
3688 int node;
3689 int cpu;
3690
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003691 for_each_possible_cpu(cpu) {
3692 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003693
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003694 if (c && c->page)
3695 return 1;
3696 }
3697
3698 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003699 struct kmem_cache_node *n = get_node(s, node);
3700
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003701 if (!n)
3702 continue;
3703
Christoph Lameter9e869432007-08-22 14:01:56 -07003704 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003705 return 1;
3706 }
3707 return 0;
3708}
3709
3710#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3711#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3712
3713struct slab_attribute {
3714 struct attribute attr;
3715 ssize_t (*show)(struct kmem_cache *s, char *buf);
3716 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3717};
3718
3719#define SLAB_ATTR_RO(_name) \
3720 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3721
3722#define SLAB_ATTR(_name) \
3723 static struct slab_attribute _name##_attr = \
3724 __ATTR(_name, 0644, _name##_show, _name##_store)
3725
Christoph Lameter81819f02007-05-06 14:49:36 -07003726static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3727{
3728 return sprintf(buf, "%d\n", s->size);
3729}
3730SLAB_ATTR_RO(slab_size);
3731
3732static ssize_t align_show(struct kmem_cache *s, char *buf)
3733{
3734 return sprintf(buf, "%d\n", s->align);
3735}
3736SLAB_ATTR_RO(align);
3737
3738static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3739{
3740 return sprintf(buf, "%d\n", s->objsize);
3741}
3742SLAB_ATTR_RO(object_size);
3743
3744static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3745{
3746 return sprintf(buf, "%d\n", s->objects);
3747}
3748SLAB_ATTR_RO(objs_per_slab);
3749
3750static ssize_t order_show(struct kmem_cache *s, char *buf)
3751{
3752 return sprintf(buf, "%d\n", s->order);
3753}
3754SLAB_ATTR_RO(order);
3755
3756static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3757{
3758 if (s->ctor) {
3759 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3760
3761 return n + sprintf(buf + n, "\n");
3762 }
3763 return 0;
3764}
3765SLAB_ATTR_RO(ctor);
3766
Christoph Lameter81819f02007-05-06 14:49:36 -07003767static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3768{
3769 return sprintf(buf, "%d\n", s->refcount - 1);
3770}
3771SLAB_ATTR_RO(aliases);
3772
3773static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3774{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003775 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003776}
3777SLAB_ATTR_RO(slabs);
3778
3779static ssize_t partial_show(struct kmem_cache *s, char *buf)
3780{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003781 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003782}
3783SLAB_ATTR_RO(partial);
3784
3785static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3786{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003787 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003788}
3789SLAB_ATTR_RO(cpu_slabs);
3790
3791static ssize_t objects_show(struct kmem_cache *s, char *buf)
3792{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003793 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003794}
3795SLAB_ATTR_RO(objects);
3796
3797static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3798{
3799 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3800}
3801
3802static ssize_t sanity_checks_store(struct kmem_cache *s,
3803 const char *buf, size_t length)
3804{
3805 s->flags &= ~SLAB_DEBUG_FREE;
3806 if (buf[0] == '1')
3807 s->flags |= SLAB_DEBUG_FREE;
3808 return length;
3809}
3810SLAB_ATTR(sanity_checks);
3811
3812static ssize_t trace_show(struct kmem_cache *s, char *buf)
3813{
3814 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3815}
3816
3817static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3818 size_t length)
3819{
3820 s->flags &= ~SLAB_TRACE;
3821 if (buf[0] == '1')
3822 s->flags |= SLAB_TRACE;
3823 return length;
3824}
3825SLAB_ATTR(trace);
3826
3827static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3828{
3829 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3830}
3831
3832static ssize_t reclaim_account_store(struct kmem_cache *s,
3833 const char *buf, size_t length)
3834{
3835 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3836 if (buf[0] == '1')
3837 s->flags |= SLAB_RECLAIM_ACCOUNT;
3838 return length;
3839}
3840SLAB_ATTR(reclaim_account);
3841
3842static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3843{
Christoph Lameter5af60832007-05-06 14:49:56 -07003844 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003845}
3846SLAB_ATTR_RO(hwcache_align);
3847
3848#ifdef CONFIG_ZONE_DMA
3849static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3850{
3851 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3852}
3853SLAB_ATTR_RO(cache_dma);
3854#endif
3855
3856static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3857{
3858 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3859}
3860SLAB_ATTR_RO(destroy_by_rcu);
3861
3862static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3863{
3864 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3865}
3866
3867static ssize_t red_zone_store(struct kmem_cache *s,
3868 const char *buf, size_t length)
3869{
3870 if (any_slab_objects(s))
3871 return -EBUSY;
3872
3873 s->flags &= ~SLAB_RED_ZONE;
3874 if (buf[0] == '1')
3875 s->flags |= SLAB_RED_ZONE;
3876 calculate_sizes(s);
3877 return length;
3878}
3879SLAB_ATTR(red_zone);
3880
3881static ssize_t poison_show(struct kmem_cache *s, char *buf)
3882{
3883 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3884}
3885
3886static ssize_t poison_store(struct kmem_cache *s,
3887 const char *buf, size_t length)
3888{
3889 if (any_slab_objects(s))
3890 return -EBUSY;
3891
3892 s->flags &= ~SLAB_POISON;
3893 if (buf[0] == '1')
3894 s->flags |= SLAB_POISON;
3895 calculate_sizes(s);
3896 return length;
3897}
3898SLAB_ATTR(poison);
3899
3900static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3901{
3902 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3903}
3904
3905static ssize_t store_user_store(struct kmem_cache *s,
3906 const char *buf, size_t length)
3907{
3908 if (any_slab_objects(s))
3909 return -EBUSY;
3910
3911 s->flags &= ~SLAB_STORE_USER;
3912 if (buf[0] == '1')
3913 s->flags |= SLAB_STORE_USER;
3914 calculate_sizes(s);
3915 return length;
3916}
3917SLAB_ATTR(store_user);
3918
Christoph Lameter53e15af2007-05-06 14:49:43 -07003919static ssize_t validate_show(struct kmem_cache *s, char *buf)
3920{
3921 return 0;
3922}
3923
3924static ssize_t validate_store(struct kmem_cache *s,
3925 const char *buf, size_t length)
3926{
Christoph Lameter434e2452007-07-17 04:03:30 -07003927 int ret = -EINVAL;
3928
3929 if (buf[0] == '1') {
3930 ret = validate_slab_cache(s);
3931 if (ret >= 0)
3932 ret = length;
3933 }
3934 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003935}
3936SLAB_ATTR(validate);
3937
Christoph Lameter2086d262007-05-06 14:49:46 -07003938static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3939{
3940 return 0;
3941}
3942
3943static ssize_t shrink_store(struct kmem_cache *s,
3944 const char *buf, size_t length)
3945{
3946 if (buf[0] == '1') {
3947 int rc = kmem_cache_shrink(s);
3948
3949 if (rc)
3950 return rc;
3951 } else
3952 return -EINVAL;
3953 return length;
3954}
3955SLAB_ATTR(shrink);
3956
Christoph Lameter88a420e2007-05-06 14:49:45 -07003957static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3958{
3959 if (!(s->flags & SLAB_STORE_USER))
3960 return -ENOSYS;
3961 return list_locations(s, buf, TRACK_ALLOC);
3962}
3963SLAB_ATTR_RO(alloc_calls);
3964
3965static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3966{
3967 if (!(s->flags & SLAB_STORE_USER))
3968 return -ENOSYS;
3969 return list_locations(s, buf, TRACK_FREE);
3970}
3971SLAB_ATTR_RO(free_calls);
3972
Christoph Lameter81819f02007-05-06 14:49:36 -07003973#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08003974static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07003975{
Christoph Lameter98246012008-01-07 23:20:26 -08003976 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07003977}
3978
Christoph Lameter98246012008-01-07 23:20:26 -08003979static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07003980 const char *buf, size_t length)
3981{
3982 int n = simple_strtoul(buf, NULL, 10);
3983
3984 if (n < 100)
Christoph Lameter98246012008-01-07 23:20:26 -08003985 s->remote_node_defrag_ratio = n * 10;
Christoph Lameter81819f02007-05-06 14:49:36 -07003986 return length;
3987}
Christoph Lameter98246012008-01-07 23:20:26 -08003988SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07003989#endif
3990
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003991#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003992static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3993{
3994 unsigned long sum = 0;
3995 int cpu;
3996 int len;
3997 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3998
3999 if (!data)
4000 return -ENOMEM;
4001
4002 for_each_online_cpu(cpu) {
4003 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4004
4005 data[cpu] = x;
4006 sum += x;
4007 }
4008
4009 len = sprintf(buf, "%lu", sum);
4010
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004011#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004012 for_each_online_cpu(cpu) {
4013 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004014 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004015 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004016#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004017 kfree(data);
4018 return len + sprintf(buf + len, "\n");
4019}
4020
4021#define STAT_ATTR(si, text) \
4022static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4023{ \
4024 return show_stat(s, buf, si); \
4025} \
4026SLAB_ATTR_RO(text); \
4027
4028STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4029STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4030STAT_ATTR(FREE_FASTPATH, free_fastpath);
4031STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4032STAT_ATTR(FREE_FROZEN, free_frozen);
4033STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4034STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4035STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4036STAT_ATTR(ALLOC_SLAB, alloc_slab);
4037STAT_ATTR(ALLOC_REFILL, alloc_refill);
4038STAT_ATTR(FREE_SLAB, free_slab);
4039STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4040STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4041STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4042STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4043STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4044STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4045
4046#endif
4047
Pekka Enberg06428782008-01-07 23:20:27 -08004048static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004049 &slab_size_attr.attr,
4050 &object_size_attr.attr,
4051 &objs_per_slab_attr.attr,
4052 &order_attr.attr,
4053 &objects_attr.attr,
4054 &slabs_attr.attr,
4055 &partial_attr.attr,
4056 &cpu_slabs_attr.attr,
4057 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004058 &aliases_attr.attr,
4059 &align_attr.attr,
4060 &sanity_checks_attr.attr,
4061 &trace_attr.attr,
4062 &hwcache_align_attr.attr,
4063 &reclaim_account_attr.attr,
4064 &destroy_by_rcu_attr.attr,
4065 &red_zone_attr.attr,
4066 &poison_attr.attr,
4067 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004068 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004069 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004070 &alloc_calls_attr.attr,
4071 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004072#ifdef CONFIG_ZONE_DMA
4073 &cache_dma_attr.attr,
4074#endif
4075#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004076 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004077#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004078#ifdef CONFIG_SLUB_STATS
4079 &alloc_fastpath_attr.attr,
4080 &alloc_slowpath_attr.attr,
4081 &free_fastpath_attr.attr,
4082 &free_slowpath_attr.attr,
4083 &free_frozen_attr.attr,
4084 &free_add_partial_attr.attr,
4085 &free_remove_partial_attr.attr,
4086 &alloc_from_partial_attr.attr,
4087 &alloc_slab_attr.attr,
4088 &alloc_refill_attr.attr,
4089 &free_slab_attr.attr,
4090 &cpuslab_flush_attr.attr,
4091 &deactivate_full_attr.attr,
4092 &deactivate_empty_attr.attr,
4093 &deactivate_to_head_attr.attr,
4094 &deactivate_to_tail_attr.attr,
4095 &deactivate_remote_frees_attr.attr,
4096#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004097 NULL
4098};
4099
4100static struct attribute_group slab_attr_group = {
4101 .attrs = slab_attrs,
4102};
4103
4104static ssize_t slab_attr_show(struct kobject *kobj,
4105 struct attribute *attr,
4106 char *buf)
4107{
4108 struct slab_attribute *attribute;
4109 struct kmem_cache *s;
4110 int err;
4111
4112 attribute = to_slab_attr(attr);
4113 s = to_slab(kobj);
4114
4115 if (!attribute->show)
4116 return -EIO;
4117
4118 err = attribute->show(s, buf);
4119
4120 return err;
4121}
4122
4123static ssize_t slab_attr_store(struct kobject *kobj,
4124 struct attribute *attr,
4125 const char *buf, size_t len)
4126{
4127 struct slab_attribute *attribute;
4128 struct kmem_cache *s;
4129 int err;
4130
4131 attribute = to_slab_attr(attr);
4132 s = to_slab(kobj);
4133
4134 if (!attribute->store)
4135 return -EIO;
4136
4137 err = attribute->store(s, buf, len);
4138
4139 return err;
4140}
4141
Christoph Lameter151c6022008-01-07 22:29:05 -08004142static void kmem_cache_release(struct kobject *kobj)
4143{
4144 struct kmem_cache *s = to_slab(kobj);
4145
4146 kfree(s);
4147}
4148
Christoph Lameter81819f02007-05-06 14:49:36 -07004149static struct sysfs_ops slab_sysfs_ops = {
4150 .show = slab_attr_show,
4151 .store = slab_attr_store,
4152};
4153
4154static struct kobj_type slab_ktype = {
4155 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004156 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004157};
4158
4159static int uevent_filter(struct kset *kset, struct kobject *kobj)
4160{
4161 struct kobj_type *ktype = get_ktype(kobj);
4162
4163 if (ktype == &slab_ktype)
4164 return 1;
4165 return 0;
4166}
4167
4168static struct kset_uevent_ops slab_uevent_ops = {
4169 .filter = uevent_filter,
4170};
4171
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004172static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004173
4174#define ID_STR_LENGTH 64
4175
4176/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004177 *
4178 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004179 */
4180static char *create_unique_id(struct kmem_cache *s)
4181{
4182 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4183 char *p = name;
4184
4185 BUG_ON(!name);
4186
4187 *p++ = ':';
4188 /*
4189 * First flags affecting slabcache operations. We will only
4190 * get here for aliasable slabs so we do not need to support
4191 * too many flags. The flags here must cover all flags that
4192 * are matched during merging to guarantee that the id is
4193 * unique.
4194 */
4195 if (s->flags & SLAB_CACHE_DMA)
4196 *p++ = 'd';
4197 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4198 *p++ = 'a';
4199 if (s->flags & SLAB_DEBUG_FREE)
4200 *p++ = 'F';
4201 if (p != name + 1)
4202 *p++ = '-';
4203 p += sprintf(p, "%07d", s->size);
4204 BUG_ON(p > name + ID_STR_LENGTH - 1);
4205 return name;
4206}
4207
4208static int sysfs_slab_add(struct kmem_cache *s)
4209{
4210 int err;
4211 const char *name;
4212 int unmergeable;
4213
4214 if (slab_state < SYSFS)
4215 /* Defer until later */
4216 return 0;
4217
4218 unmergeable = slab_unmergeable(s);
4219 if (unmergeable) {
4220 /*
4221 * Slabcache can never be merged so we can use the name proper.
4222 * This is typically the case for debug situations. In that
4223 * case we can catch duplicate names easily.
4224 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004225 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004226 name = s->name;
4227 } else {
4228 /*
4229 * Create a unique name for the slab as a target
4230 * for the symlinks.
4231 */
4232 name = create_unique_id(s);
4233 }
4234
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004235 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004236 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4237 if (err) {
4238 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004239 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004240 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004241
4242 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4243 if (err)
4244 return err;
4245 kobject_uevent(&s->kobj, KOBJ_ADD);
4246 if (!unmergeable) {
4247 /* Setup first alias */
4248 sysfs_slab_alias(s, s->name);
4249 kfree(name);
4250 }
4251 return 0;
4252}
4253
4254static void sysfs_slab_remove(struct kmem_cache *s)
4255{
4256 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4257 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004258 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004259}
4260
4261/*
4262 * Need to buffer aliases during bootup until sysfs becomes
4263 * available lest we loose that information.
4264 */
4265struct saved_alias {
4266 struct kmem_cache *s;
4267 const char *name;
4268 struct saved_alias *next;
4269};
4270
Adrian Bunk5af328a2007-07-17 04:03:27 -07004271static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004272
4273static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4274{
4275 struct saved_alias *al;
4276
4277 if (slab_state == SYSFS) {
4278 /*
4279 * If we have a leftover link then remove it.
4280 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004281 sysfs_remove_link(&slab_kset->kobj, name);
4282 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004283 }
4284
4285 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4286 if (!al)
4287 return -ENOMEM;
4288
4289 al->s = s;
4290 al->name = name;
4291 al->next = alias_list;
4292 alias_list = al;
4293 return 0;
4294}
4295
4296static int __init slab_sysfs_init(void)
4297{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004298 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004299 int err;
4300
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004301 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004302 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004303 printk(KERN_ERR "Cannot register slab subsystem.\n");
4304 return -ENOSYS;
4305 }
4306
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004307 slab_state = SYSFS;
4308
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004309 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004310 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004311 if (err)
4312 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4313 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004314 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004315
4316 while (alias_list) {
4317 struct saved_alias *al = alias_list;
4318
4319 alias_list = alias_list->next;
4320 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004321 if (err)
4322 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4323 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004324 kfree(al);
4325 }
4326
4327 resiliency_test();
4328 return 0;
4329}
4330
4331__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004332#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004333
4334/*
4335 * The /proc/slabinfo ABI
4336 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004337#ifdef CONFIG_SLABINFO
4338
4339ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4340 size_t count, loff_t *ppos)
4341{
4342 return -EINVAL;
4343}
4344
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004345
4346static void print_slabinfo_header(struct seq_file *m)
4347{
4348 seq_puts(m, "slabinfo - version: 2.1\n");
4349 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4350 "<objperslab> <pagesperslab>");
4351 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4352 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4353 seq_putc(m, '\n');
4354}
4355
4356static void *s_start(struct seq_file *m, loff_t *pos)
4357{
4358 loff_t n = *pos;
4359
4360 down_read(&slub_lock);
4361 if (!n)
4362 print_slabinfo_header(m);
4363
4364 return seq_list_start(&slab_caches, *pos);
4365}
4366
4367static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4368{
4369 return seq_list_next(p, &slab_caches, pos);
4370}
4371
4372static void s_stop(struct seq_file *m, void *p)
4373{
4374 up_read(&slub_lock);
4375}
4376
4377static int s_show(struct seq_file *m, void *p)
4378{
4379 unsigned long nr_partials = 0;
4380 unsigned long nr_slabs = 0;
4381 unsigned long nr_inuse = 0;
4382 unsigned long nr_objs;
4383 struct kmem_cache *s;
4384 int node;
4385
4386 s = list_entry(p, struct kmem_cache, list);
4387
4388 for_each_online_node(node) {
4389 struct kmem_cache_node *n = get_node(s, node);
4390
4391 if (!n)
4392 continue;
4393
4394 nr_partials += n->nr_partial;
4395 nr_slabs += atomic_long_read(&n->nr_slabs);
4396 nr_inuse += count_partial(n);
4397 }
4398
4399 nr_objs = nr_slabs * s->objects;
4400 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4401
4402 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4403 nr_objs, s->size, s->objects, (1 << s->order));
4404 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4405 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4406 0UL);
4407 seq_putc(m, '\n');
4408 return 0;
4409}
4410
4411const struct seq_operations slabinfo_op = {
4412 .start = s_start,
4413 .next = s_next,
4414 .stop = s_stop,
4415 .show = s_show,
4416};
4417
Linus Torvalds158a9622008-01-02 13:04:48 -08004418#endif /* CONFIG_SLABINFO */