blob: 74c65af0a54f4c112e6f2bc223018deb74771d25 [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 Lameter683d0ba2008-01-07 23:20:29 -0800294/*
295 * The end pointer in a slab is special. It points to the first object in the
296 * slab but has bit 0 set to mark it.
297 *
298 * Note that SLUB relies on page_mapping returning NULL for pages with bit 0
299 * in the mapping set.
300 */
301static inline int is_end(void *addr)
302{
303 return (unsigned long)addr & PAGE_MAPPING_ANON;
304}
305
Adrian Bunkdada1232008-02-13 23:30:32 +0200306static void *slab_address(struct page *page)
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800307{
308 return page->end - PAGE_MAPPING_ANON;
309}
310
Christoph Lameter02cbc872007-05-09 02:32:43 -0700311static inline int check_valid_pointer(struct kmem_cache *s,
312 struct page *page, const void *object)
313{
314 void *base;
315
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800316 if (object == page->end)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700317 return 1;
318
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800319 base = slab_address(page);
Christoph Lameter02cbc872007-05-09 02:32:43 -0700320 if (object < base || object >= base + s->objects * s->size ||
321 (object - base) % s->size) {
322 return 0;
323 }
324
325 return 1;
326}
327
Christoph Lameter81819f02007-05-06 14:49:36 -0700328/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700329 * Slow version of get and set free pointer.
330 *
331 * This version requires touching the cache lines of kmem_cache which
332 * we avoid to do in the fast alloc free paths. There we obtain the offset
333 * from the page struct.
334 */
335static inline void *get_freepointer(struct kmem_cache *s, void *object)
336{
337 return *(void **)(object + s->offset);
338}
339
340static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
341{
342 *(void **)(object + s->offset) = fp;
343}
344
345/* Loop over all objects in a slab */
346#define for_each_object(__p, __s, __addr) \
347 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
348 __p += (__s)->size)
349
350/* Scan freelist */
351#define for_each_free_object(__p, __s, __free) \
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800352 for (__p = (__free); (__p) != page->end; __p = get_freepointer((__s),\
353 __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700354
355/* Determine object index from a given position */
356static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
357{
358 return (p - addr) / s->size;
359}
360
Christoph Lameter41ecc552007-05-09 02:32:44 -0700361#ifdef CONFIG_SLUB_DEBUG
362/*
363 * Debug settings:
364 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700365#ifdef CONFIG_SLUB_DEBUG_ON
366static int slub_debug = DEBUG_DEFAULT_FLAGS;
367#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700368static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700369#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700370
371static char *slub_debug_slabs;
372
Christoph Lameter7656c722007-05-09 02:32:40 -0700373/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700374 * Object debugging
375 */
376static void print_section(char *text, u8 *addr, unsigned int length)
377{
378 int i, offset;
379 int newline = 1;
380 char ascii[17];
381
382 ascii[16] = 0;
383
384 for (i = 0; i < length; i++) {
385 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700386 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700387 newline = 0;
388 }
Pekka Enberg06428782008-01-07 23:20:27 -0800389 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700390 offset = i % 16;
391 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
392 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800393 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700394 newline = 1;
395 }
396 }
397 if (!newline) {
398 i %= 16;
399 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800400 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700401 ascii[i] = ' ';
402 i++;
403 }
Pekka Enberg06428782008-01-07 23:20:27 -0800404 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700405 }
406}
407
Christoph Lameter81819f02007-05-06 14:49:36 -0700408static struct track *get_track(struct kmem_cache *s, void *object,
409 enum track_item alloc)
410{
411 struct track *p;
412
413 if (s->offset)
414 p = object + s->offset + sizeof(void *);
415 else
416 p = object + s->inuse;
417
418 return p + alloc;
419}
420
421static void set_track(struct kmem_cache *s, void *object,
422 enum track_item alloc, void *addr)
423{
424 struct track *p;
425
426 if (s->offset)
427 p = object + s->offset + sizeof(void *);
428 else
429 p = object + s->inuse;
430
431 p += alloc;
432 if (addr) {
433 p->addr = addr;
434 p->cpu = smp_processor_id();
435 p->pid = current ? current->pid : -1;
436 p->when = jiffies;
437 } else
438 memset(p, 0, sizeof(struct track));
439}
440
Christoph Lameter81819f02007-05-06 14:49:36 -0700441static void init_tracking(struct kmem_cache *s, void *object)
442{
Christoph Lameter24922682007-07-17 04:03:18 -0700443 if (!(s->flags & SLAB_STORE_USER))
444 return;
445
446 set_track(s, object, TRACK_FREE, NULL);
447 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700448}
449
450static void print_track(const char *s, struct track *t)
451{
452 if (!t->addr)
453 return;
454
Christoph Lameter24922682007-07-17 04:03:18 -0700455 printk(KERN_ERR "INFO: %s in ", s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700456 __print_symbol("%s", (unsigned long)t->addr);
Christoph Lameter24922682007-07-17 04:03:18 -0700457 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700458}
459
Christoph Lameter24922682007-07-17 04:03:18 -0700460static void print_tracking(struct kmem_cache *s, void *object)
461{
462 if (!(s->flags & SLAB_STORE_USER))
463 return;
464
465 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
466 print_track("Freed", get_track(s, object, TRACK_FREE));
467}
468
469static void print_page_info(struct page *page)
470{
471 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
472 page, page->inuse, page->freelist, page->flags);
473
474}
475
476static void slab_bug(struct kmem_cache *s, char *fmt, ...)
477{
478 va_list args;
479 char buf[100];
480
481 va_start(args, fmt);
482 vsnprintf(buf, sizeof(buf), fmt, args);
483 va_end(args);
484 printk(KERN_ERR "========================================"
485 "=====================================\n");
486 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
487 printk(KERN_ERR "----------------------------------------"
488 "-------------------------------------\n\n");
489}
490
491static void slab_fix(struct kmem_cache *s, char *fmt, ...)
492{
493 va_list args;
494 char buf[100];
495
496 va_start(args, fmt);
497 vsnprintf(buf, sizeof(buf), fmt, args);
498 va_end(args);
499 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
500}
501
502static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700503{
504 unsigned int off; /* Offset of last byte */
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800505 u8 *addr = slab_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700506
507 print_tracking(s, p);
508
509 print_page_info(page);
510
511 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
512 p, p - addr, get_freepointer(s, p));
513
514 if (p > addr + 16)
515 print_section("Bytes b4", p - 16, 16);
516
517 print_section("Object", p, min(s->objsize, 128));
Christoph Lameter81819f02007-05-06 14:49:36 -0700518
519 if (s->flags & SLAB_RED_ZONE)
520 print_section("Redzone", p + s->objsize,
521 s->inuse - s->objsize);
522
Christoph Lameter81819f02007-05-06 14:49:36 -0700523 if (s->offset)
524 off = s->offset + sizeof(void *);
525 else
526 off = s->inuse;
527
Christoph Lameter24922682007-07-17 04:03:18 -0700528 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700529 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700530
531 if (off != s->size)
532 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700533 print_section("Padding", p + off, s->size - off);
534
535 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700536}
537
538static void object_err(struct kmem_cache *s, struct page *page,
539 u8 *object, char *reason)
540{
Christoph Lameter24922682007-07-17 04:03:18 -0700541 slab_bug(s, reason);
542 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700543}
544
Christoph Lameter24922682007-07-17 04:03:18 -0700545static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700546{
547 va_list args;
548 char buf[100];
549
Christoph Lameter24922682007-07-17 04:03:18 -0700550 va_start(args, fmt);
551 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700552 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700553 slab_bug(s, fmt);
554 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700555 dump_stack();
556}
557
558static void init_object(struct kmem_cache *s, void *object, int active)
559{
560 u8 *p = object;
561
562 if (s->flags & __OBJECT_POISON) {
563 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800564 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700565 }
566
567 if (s->flags & SLAB_RED_ZONE)
568 memset(p + s->objsize,
569 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
570 s->inuse - s->objsize);
571}
572
Christoph Lameter24922682007-07-17 04:03:18 -0700573static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700574{
575 while (bytes) {
576 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700577 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700578 start++;
579 bytes--;
580 }
Christoph Lameter24922682007-07-17 04:03:18 -0700581 return NULL;
582}
583
584static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
585 void *from, void *to)
586{
587 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
588 memset(from, data, to - from);
589}
590
591static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
592 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800593 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700594{
595 u8 *fault;
596 u8 *end;
597
598 fault = check_bytes(start, value, bytes);
599 if (!fault)
600 return 1;
601
602 end = start + bytes;
603 while (end > fault && end[-1] == value)
604 end--;
605
606 slab_bug(s, "%s overwritten", what);
607 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
608 fault, end - 1, fault[0], value);
609 print_trailer(s, page, object);
610
611 restore_bytes(s, what, value, fault, end);
612 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700613}
614
Christoph Lameter81819f02007-05-06 14:49:36 -0700615/*
616 * Object layout:
617 *
618 * object address
619 * Bytes of the object to be managed.
620 * If the freepointer may overlay the object then the free
621 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700622 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700623 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
624 * 0xa5 (POISON_END)
625 *
626 * object + s->objsize
627 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700628 * Padding is extended by another word if Redzoning is enabled and
629 * objsize == inuse.
630 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700631 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
632 * 0xcc (RED_ACTIVE) for objects in use.
633 *
634 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700635 * Meta data starts here.
636 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700637 * A. Free pointer (if we cannot overwrite object on free)
638 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700639 * C. Padding to reach required alignment boundary or at mininum
640 * one word if debuggin is on to be able to detect writes
641 * before the word boundary.
642 *
643 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700644 *
645 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700646 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700647 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700648 * If slabcaches are merged then the objsize and inuse boundaries are mostly
649 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700650 * may be used with merged slabcaches.
651 */
652
Christoph Lameter81819f02007-05-06 14:49:36 -0700653static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
654{
655 unsigned long off = s->inuse; /* The end of info */
656
657 if (s->offset)
658 /* Freepointer is placed after the object. */
659 off += sizeof(void *);
660
661 if (s->flags & SLAB_STORE_USER)
662 /* We also have user information there */
663 off += 2 * sizeof(struct track);
664
665 if (s->size == off)
666 return 1;
667
Christoph Lameter24922682007-07-17 04:03:18 -0700668 return check_bytes_and_report(s, page, p, "Object padding",
669 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700670}
671
672static int slab_pad_check(struct kmem_cache *s, struct page *page)
673{
Christoph Lameter24922682007-07-17 04:03:18 -0700674 u8 *start;
675 u8 *fault;
676 u8 *end;
677 int length;
678 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700679
680 if (!(s->flags & SLAB_POISON))
681 return 1;
682
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800683 start = slab_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700684 end = start + (PAGE_SIZE << s->order);
Christoph Lameter81819f02007-05-06 14:49:36 -0700685 length = s->objects * s->size;
Christoph Lameter24922682007-07-17 04:03:18 -0700686 remainder = end - (start + length);
Christoph Lameter81819f02007-05-06 14:49:36 -0700687 if (!remainder)
688 return 1;
689
Christoph Lameter24922682007-07-17 04:03:18 -0700690 fault = check_bytes(start + length, POISON_INUSE, remainder);
691 if (!fault)
692 return 1;
693 while (end > fault && end[-1] == POISON_INUSE)
694 end--;
695
696 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
697 print_section("Padding", start, length);
698
699 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
700 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700701}
702
703static int check_object(struct kmem_cache *s, struct page *page,
704 void *object, int active)
705{
706 u8 *p = object;
707 u8 *endobject = object + s->objsize;
708
709 if (s->flags & SLAB_RED_ZONE) {
710 unsigned int red =
711 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
712
Christoph Lameter24922682007-07-17 04:03:18 -0700713 if (!check_bytes_and_report(s, page, object, "Redzone",
714 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700715 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700716 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800717 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
718 check_bytes_and_report(s, page, p, "Alignment padding",
719 endobject, POISON_INUSE, s->inuse - s->objsize);
720 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700721 }
722
723 if (s->flags & SLAB_POISON) {
724 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700725 (!check_bytes_and_report(s, page, p, "Poison", p,
726 POISON_FREE, s->objsize - 1) ||
727 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800728 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700729 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700730 /*
731 * check_pad_bytes cleans up on its own.
732 */
733 check_pad_bytes(s, page, p);
734 }
735
736 if (!s->offset && active)
737 /*
738 * Object and freepointer overlap. Cannot check
739 * freepointer while object is allocated.
740 */
741 return 1;
742
743 /* Check free pointer validity */
744 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
745 object_err(s, page, p, "Freepointer corrupt");
746 /*
747 * No choice but to zap it and thus loose the remainder
748 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700749 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700750 */
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800751 set_freepointer(s, p, page->end);
Christoph Lameter81819f02007-05-06 14:49:36 -0700752 return 0;
753 }
754 return 1;
755}
756
757static int check_slab(struct kmem_cache *s, struct page *page)
758{
759 VM_BUG_ON(!irqs_disabled());
760
761 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700762 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700763 return 0;
764 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700765 if (page->inuse > s->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700766 slab_err(s, page, "inuse %u > max %u",
767 s->name, page->inuse, s->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700768 return 0;
769 }
770 /* Slab_pad_check fixes things up after itself */
771 slab_pad_check(s, page);
772 return 1;
773}
774
775/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700776 * Determine if a certain object on a page is on the freelist. Must hold the
777 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700778 */
779static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
780{
781 int nr = 0;
782 void *fp = page->freelist;
783 void *object = NULL;
784
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800785 while (fp != page->end && nr <= s->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700786 if (fp == search)
787 return 1;
788 if (!check_valid_pointer(s, page, fp)) {
789 if (object) {
790 object_err(s, page, object,
791 "Freechain corrupt");
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800792 set_freepointer(s, object, page->end);
Christoph Lameter81819f02007-05-06 14:49:36 -0700793 break;
794 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700795 slab_err(s, page, "Freepointer corrupt");
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800796 page->freelist = page->end;
Christoph Lameter81819f02007-05-06 14:49:36 -0700797 page->inuse = s->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700798 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700799 return 0;
800 }
801 break;
802 }
803 object = fp;
804 fp = get_freepointer(s, object);
805 nr++;
806 }
807
808 if (page->inuse != s->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700809 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter24922682007-07-17 04:03:18 -0700810 "counted were %d", page->inuse, s->objects - nr);
Christoph Lameter81819f02007-05-06 14:49:36 -0700811 page->inuse = s->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700812 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700813 }
814 return search == NULL;
815}
816
Christoph Lameter3ec09742007-05-16 22:11:00 -0700817static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
818{
819 if (s->flags & SLAB_TRACE) {
820 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
821 s->name,
822 alloc ? "alloc" : "free",
823 object, page->inuse,
824 page->freelist);
825
826 if (!alloc)
827 print_section("Object", (void *)object, s->objsize);
828
829 dump_stack();
830 }
831}
832
Christoph Lameter643b1132007-05-06 14:49:42 -0700833/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700834 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700835 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700836static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700837{
Christoph Lameter643b1132007-05-06 14:49:42 -0700838 spin_lock(&n->list_lock);
839 list_add(&page->lru, &n->full);
840 spin_unlock(&n->list_lock);
841}
842
843static void remove_full(struct kmem_cache *s, struct page *page)
844{
845 struct kmem_cache_node *n;
846
847 if (!(s->flags & SLAB_STORE_USER))
848 return;
849
850 n = get_node(s, page_to_nid(page));
851
852 spin_lock(&n->list_lock);
853 list_del(&page->lru);
854 spin_unlock(&n->list_lock);
855}
856
Christoph Lameter3ec09742007-05-16 22:11:00 -0700857static void setup_object_debug(struct kmem_cache *s, struct page *page,
858 void *object)
859{
860 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
861 return;
862
863 init_object(s, object, 0);
864 init_tracking(s, object);
865}
866
867static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
868 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700869{
870 if (!check_slab(s, page))
871 goto bad;
872
873 if (object && !on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700874 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700875 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700876 }
877
878 if (!check_valid_pointer(s, page, object)) {
879 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700880 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700881 }
882
Christoph Lameter3ec09742007-05-16 22:11:00 -0700883 if (object && !check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700884 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700885
Christoph Lameter3ec09742007-05-16 22:11:00 -0700886 /* Success perform special debug activities for allocs */
887 if (s->flags & SLAB_STORE_USER)
888 set_track(s, object, TRACK_ALLOC, addr);
889 trace(s, page, object, 1);
890 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700891 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700892
Christoph Lameter81819f02007-05-06 14:49:36 -0700893bad:
894 if (PageSlab(page)) {
895 /*
896 * If this is a slab page then lets do the best we can
897 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700898 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700899 */
Christoph Lameter24922682007-07-17 04:03:18 -0700900 slab_fix(s, "Marking all objects used");
Christoph Lameter81819f02007-05-06 14:49:36 -0700901 page->inuse = s->objects;
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800902 page->freelist = page->end;
Christoph Lameter81819f02007-05-06 14:49:36 -0700903 }
904 return 0;
905}
906
Christoph Lameter3ec09742007-05-16 22:11:00 -0700907static int free_debug_processing(struct kmem_cache *s, struct page *page,
908 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700909{
910 if (!check_slab(s, page))
911 goto fail;
912
913 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700914 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700915 goto fail;
916 }
917
918 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700919 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700920 goto fail;
921 }
922
923 if (!check_object(s, page, object, 1))
924 return 0;
925
926 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800927 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700928 slab_err(s, page, "Attempt to free object(0x%p) "
929 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800930 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700931 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700932 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700933 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700934 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800935 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700936 object_err(s, page, object,
937 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700938 goto fail;
939 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700940
941 /* Special debug activities for freeing objects */
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800942 if (!SlabFrozen(page) && page->freelist == page->end)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700943 remove_full(s, page);
944 if (s->flags & SLAB_STORE_USER)
945 set_track(s, object, TRACK_FREE, addr);
946 trace(s, page, object, 0);
947 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700948 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700949
Christoph Lameter81819f02007-05-06 14:49:36 -0700950fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700951 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700952 return 0;
953}
954
Christoph Lameter41ecc552007-05-09 02:32:44 -0700955static int __init setup_slub_debug(char *str)
956{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700957 slub_debug = DEBUG_DEFAULT_FLAGS;
958 if (*str++ != '=' || !*str)
959 /*
960 * No options specified. Switch on full debugging.
961 */
962 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700963
964 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700965 /*
966 * No options but restriction on slabs. This means full
967 * debugging for slabs matching a pattern.
968 */
969 goto check_slabs;
970
971 slub_debug = 0;
972 if (*str == '-')
973 /*
974 * Switch off all debugging measures.
975 */
976 goto out;
977
978 /*
979 * Determine which debug features should be switched on
980 */
Pekka Enberg06428782008-01-07 23:20:27 -0800981 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700982 switch (tolower(*str)) {
983 case 'f':
984 slub_debug |= SLAB_DEBUG_FREE;
985 break;
986 case 'z':
987 slub_debug |= SLAB_RED_ZONE;
988 break;
989 case 'p':
990 slub_debug |= SLAB_POISON;
991 break;
992 case 'u':
993 slub_debug |= SLAB_STORE_USER;
994 break;
995 case 't':
996 slub_debug |= SLAB_TRACE;
997 break;
998 default:
999 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001000 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001001 }
1002 }
1003
1004check_slabs:
1005 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001006 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001007out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001008 return 1;
1009}
1010
1011__setup("slub_debug", setup_slub_debug);
1012
Christoph Lameterba0268a2007-09-11 15:24:11 -07001013static unsigned long kmem_cache_flags(unsigned long objsize,
1014 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001015 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001016{
1017 /*
1018 * The page->offset field is only 16 bit wide. This is an offset
1019 * in units of words from the beginning of an object. If the slab
1020 * size is bigger then we cannot move the free pointer behind the
1021 * object anymore.
1022 *
1023 * On 32 bit platforms the limit is 256k. On 64bit platforms
1024 * the limit is 512k.
1025 *
Christoph Lameterc59def92007-05-16 22:10:50 -07001026 * Debugging or ctor may create a need to move the free
Christoph Lameter41ecc552007-05-09 02:32:44 -07001027 * pointer. Fail if this happens.
1028 */
Christoph Lameterba0268a2007-09-11 15:24:11 -07001029 if (objsize >= 65535 * sizeof(void *)) {
1030 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
Christoph Lameter41ecc552007-05-09 02:32:44 -07001031 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
Christoph Lameterba0268a2007-09-11 15:24:11 -07001032 BUG_ON(ctor);
1033 } else {
Christoph Lameter41ecc552007-05-09 02:32:44 -07001034 /*
1035 * Enable debugging if selected on the kernel commandline.
1036 */
1037 if (slub_debug && (!slub_debug_slabs ||
Christoph Lameterba0268a2007-09-11 15:24:11 -07001038 strncmp(slub_debug_slabs, name,
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001039 strlen(slub_debug_slabs)) == 0))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001040 flags |= slub_debug;
1041 }
1042
1043 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001044}
1045#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001046static inline void setup_object_debug(struct kmem_cache *s,
1047 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001048
Christoph Lameter3ec09742007-05-16 22:11:00 -07001049static inline int alloc_debug_processing(struct kmem_cache *s,
1050 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001051
Christoph Lameter3ec09742007-05-16 22:11:00 -07001052static inline int free_debug_processing(struct kmem_cache *s,
1053 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001054
Christoph Lameter41ecc552007-05-09 02:32:44 -07001055static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1056 { return 1; }
1057static inline int check_object(struct kmem_cache *s, struct page *page,
1058 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001059static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001060static inline unsigned long kmem_cache_flags(unsigned long objsize,
1061 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001062 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001063{
1064 return flags;
1065}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001066#define slub_debug 0
1067#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001068/*
1069 * Slab allocation and freeing
1070 */
1071static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1072{
Pekka Enberg06428782008-01-07 23:20:27 -08001073 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001074 int pages = 1 << s->order;
1075
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001076 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001077
Christoph Lameter81819f02007-05-06 14:49:36 -07001078 if (node == -1)
1079 page = alloc_pages(flags, s->order);
1080 else
1081 page = alloc_pages_node(node, flags, s->order);
1082
1083 if (!page)
1084 return NULL;
1085
1086 mod_zone_page_state(page_zone(page),
1087 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1088 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1089 pages);
1090
1091 return page;
1092}
1093
1094static void setup_object(struct kmem_cache *s, struct page *page,
1095 void *object)
1096{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001097 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001098 if (unlikely(s->ctor))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001099 s->ctor(s, object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001100}
1101
1102static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1103{
1104 struct page *page;
1105 struct kmem_cache_node *n;
1106 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001107 void *last;
1108 void *p;
1109
Christoph Lameter6cb06222007-10-16 01:25:41 -07001110 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001111
Christoph Lameter6cb06222007-10-16 01:25:41 -07001112 page = allocate_slab(s,
1113 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001114 if (!page)
1115 goto out;
1116
1117 n = get_node(s, page_to_nid(page));
1118 if (n)
1119 atomic_long_inc(&n->nr_slabs);
Christoph Lameter81819f02007-05-06 14:49:36 -07001120 page->slab = s;
1121 page->flags |= 1 << PG_slab;
1122 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1123 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001124 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001125
1126 start = page_address(page);
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001127 page->end = start + 1;
Christoph Lameter81819f02007-05-06 14:49:36 -07001128
1129 if (unlikely(s->flags & SLAB_POISON))
1130 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1131
1132 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001133 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001134 setup_object(s, page, last);
1135 set_freepointer(s, last, p);
1136 last = p;
1137 }
1138 setup_object(s, page, last);
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001139 set_freepointer(s, last, page->end);
Christoph Lameter81819f02007-05-06 14:49:36 -07001140
1141 page->freelist = start;
1142 page->inuse = 0;
1143out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001144 return page;
1145}
1146
1147static void __free_slab(struct kmem_cache *s, struct page *page)
1148{
1149 int pages = 1 << s->order;
1150
Christoph Lameterc59def92007-05-16 22:10:50 -07001151 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001152 void *p;
1153
1154 slab_pad_check(s, page);
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001155 for_each_object(p, s, slab_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001156 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001157 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001158 }
1159
1160 mod_zone_page_state(page_zone(page),
1161 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1162 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001163 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001164
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001165 page->mapping = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001166 __free_pages(page, s->order);
1167}
1168
1169static void rcu_free_slab(struct rcu_head *h)
1170{
1171 struct page *page;
1172
1173 page = container_of((struct list_head *)h, struct page, lru);
1174 __free_slab(page->slab, page);
1175}
1176
1177static void free_slab(struct kmem_cache *s, struct page *page)
1178{
1179 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1180 /*
1181 * RCU free overloads the RCU head over the LRU
1182 */
1183 struct rcu_head *head = (void *)&page->lru;
1184
1185 call_rcu(head, rcu_free_slab);
1186 } else
1187 __free_slab(s, page);
1188}
1189
1190static void discard_slab(struct kmem_cache *s, struct page *page)
1191{
1192 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1193
1194 atomic_long_dec(&n->nr_slabs);
1195 reset_page_mapcount(page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001196 __ClearPageSlab(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001197 free_slab(s, page);
1198}
1199
1200/*
1201 * Per slab locking using the pagelock
1202 */
1203static __always_inline void slab_lock(struct page *page)
1204{
1205 bit_spin_lock(PG_locked, &page->flags);
1206}
1207
1208static __always_inline void slab_unlock(struct page *page)
1209{
Nick Piggina76d3542008-01-07 23:20:27 -08001210 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001211}
1212
1213static __always_inline int slab_trylock(struct page *page)
1214{
1215 int rc = 1;
1216
1217 rc = bit_spin_trylock(PG_locked, &page->flags);
1218 return rc;
1219}
1220
1221/*
1222 * Management of partially allocated slabs
1223 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001224static void add_partial(struct kmem_cache_node *n,
1225 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001226{
Christoph Lametere95eed52007-05-06 14:49:44 -07001227 spin_lock(&n->list_lock);
1228 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001229 if (tail)
1230 list_add_tail(&page->lru, &n->partial);
1231 else
1232 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001233 spin_unlock(&n->list_lock);
1234}
1235
1236static void remove_partial(struct kmem_cache *s,
1237 struct page *page)
1238{
1239 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1240
1241 spin_lock(&n->list_lock);
1242 list_del(&page->lru);
1243 n->nr_partial--;
1244 spin_unlock(&n->list_lock);
1245}
1246
1247/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001248 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001249 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001250 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001251 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001252static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001253{
1254 if (slab_trylock(page)) {
1255 list_del(&page->lru);
1256 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001257 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001258 return 1;
1259 }
1260 return 0;
1261}
1262
1263/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001264 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001265 */
1266static struct page *get_partial_node(struct kmem_cache_node *n)
1267{
1268 struct page *page;
1269
1270 /*
1271 * Racy check. If we mistakenly see no partial slabs then we
1272 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001273 * partial slab and there is none available then get_partials()
1274 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001275 */
1276 if (!n || !n->nr_partial)
1277 return NULL;
1278
1279 spin_lock(&n->list_lock);
1280 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001281 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001282 goto out;
1283 page = NULL;
1284out:
1285 spin_unlock(&n->list_lock);
1286 return page;
1287}
1288
1289/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001290 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001291 */
1292static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1293{
1294#ifdef CONFIG_NUMA
1295 struct zonelist *zonelist;
1296 struct zone **z;
1297 struct page *page;
1298
1299 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001300 * The defrag ratio allows a configuration of the tradeoffs between
1301 * inter node defragmentation and node local allocations. A lower
1302 * defrag_ratio increases the tendency to do local allocations
1303 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001304 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001305 * If the defrag_ratio is set to 0 then kmalloc() always
1306 * returns node local objects. If the ratio is higher then kmalloc()
1307 * may return off node objects because partial slabs are obtained
1308 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001309 *
1310 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001311 * defrag_ratio = 1000) then every (well almost) allocation will
1312 * first attempt to defrag slab caches on other nodes. This means
1313 * scanning over all nodes to look for partial slabs which may be
1314 * expensive if we do it every time we are trying to find a slab
1315 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001316 */
Christoph Lameter98246012008-01-07 23:20:26 -08001317 if (!s->remote_node_defrag_ratio ||
1318 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001319 return NULL;
1320
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001321 zonelist = &NODE_DATA(
1322 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
Christoph Lameter81819f02007-05-06 14:49:36 -07001323 for (z = zonelist->zones; *z; z++) {
1324 struct kmem_cache_node *n;
1325
1326 n = get_node(s, zone_to_nid(*z));
1327
1328 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001329 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001330 page = get_partial_node(n);
1331 if (page)
1332 return page;
1333 }
1334 }
1335#endif
1336 return NULL;
1337}
1338
1339/*
1340 * Get a partial page, lock it and return it.
1341 */
1342static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1343{
1344 struct page *page;
1345 int searchnode = (node == -1) ? numa_node_id() : node;
1346
1347 page = get_partial_node(get_node(s, searchnode));
1348 if (page || (flags & __GFP_THISNODE))
1349 return page;
1350
1351 return get_any_partial(s, flags);
1352}
1353
1354/*
1355 * Move a page back to the lists.
1356 *
1357 * Must be called with the slab lock held.
1358 *
1359 * On exit the slab lock will have been dropped.
1360 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001361static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001362{
Christoph Lametere95eed52007-05-06 14:49:44 -07001363 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001364 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001365
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001366 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001367 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001368
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001369 if (page->freelist != page->end) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001370 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001371 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1372 } else {
1373 stat(c, DEACTIVATE_FULL);
1374 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1375 add_full(n, page);
1376 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001377 slab_unlock(page);
1378 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001379 stat(c, DEACTIVATE_EMPTY);
Christoph Lametere95eed52007-05-06 14:49:44 -07001380 if (n->nr_partial < MIN_PARTIAL) {
1381 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001382 * Adding an empty slab to the partial slabs in order
1383 * to avoid page allocator overhead. This slab needs
1384 * to come after the other slabs with objects in
1385 * order to fill them up. That way the size of the
1386 * partial list stays small. kmem_cache_shrink can
1387 * reclaim empty slabs from the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001388 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001389 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001390 slab_unlock(page);
1391 } else {
1392 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001393 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001394 discard_slab(s, page);
1395 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001396 }
1397}
1398
1399/*
1400 * Remove the cpu slab
1401 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001402static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001403{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001404 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001405 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001406
1407 if (c->freelist)
1408 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001409 /*
1410 * Merge cpu freelist into freelist. Typically we get here
1411 * because both freelists are empty. So this is unlikely
1412 * to occur.
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001413 *
1414 * We need to use _is_end here because deactivate slab may
1415 * be called for a debug slab. Then c->freelist may contain
1416 * a dummy pointer.
Christoph Lameter894b8782007-05-10 03:15:16 -07001417 */
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001418 while (unlikely(!is_end(c->freelist))) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001419 void **object;
1420
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001421 tail = 0; /* Hot objects. Put the slab first */
1422
Christoph Lameter894b8782007-05-10 03:15:16 -07001423 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001424 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001425 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001426
1427 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001428 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001429 page->freelist = object;
1430 page->inuse--;
1431 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001432 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001433 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001434}
1435
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001436static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001437{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001438 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001439 slab_lock(c->page);
1440 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001441}
1442
1443/*
1444 * Flush cpu slab.
1445 * Called from IPI handler with interrupts disabled.
1446 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001447static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001448{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001449 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001450
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001451 if (likely(c && c->page))
1452 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001453}
1454
1455static void flush_cpu_slab(void *d)
1456{
1457 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001458
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001459 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001460}
1461
1462static void flush_all(struct kmem_cache *s)
1463{
1464#ifdef CONFIG_SMP
1465 on_each_cpu(flush_cpu_slab, s, 1, 1);
1466#else
1467 unsigned long flags;
1468
1469 local_irq_save(flags);
1470 flush_cpu_slab(s);
1471 local_irq_restore(flags);
1472#endif
1473}
1474
1475/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001476 * Check if the objects in a per cpu structure fit numa
1477 * locality expectations.
1478 */
1479static inline int node_match(struct kmem_cache_cpu *c, int node)
1480{
1481#ifdef CONFIG_NUMA
1482 if (node != -1 && c->node != node)
1483 return 0;
1484#endif
1485 return 1;
1486}
1487
1488/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001489 * Slow path. The lockless freelist is empty or we need to perform
1490 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001491 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001492 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001493 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001494 * Processing is still very fast if new objects have been freed to the
1495 * regular freelist. In that case we simply take over the regular freelist
1496 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001497 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001498 * If that is not working then we fall back to the partial lists. We take the
1499 * first element of the freelist as the object to allocate now and move the
1500 * rest of the freelist to the lockless freelist.
1501 *
1502 * And if we were unable to get a new slab from the partial slab lists then
1503 * we need to allocate a new slab. This is slowest path since we may sleep.
Christoph Lameter81819f02007-05-06 14:49:36 -07001504 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001505static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001506 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001507{
Christoph Lameter81819f02007-05-06 14:49:36 -07001508 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001509 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001510
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001511 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001512 goto new_slab;
1513
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001514 slab_lock(c->page);
1515 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001516 goto another_slab;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001517 stat(c, ALLOC_REFILL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001518load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001519 object = c->page->freelist;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001520 if (unlikely(object == c->page->end))
Christoph Lameter81819f02007-05-06 14:49:36 -07001521 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001522 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001523 goto debug;
1524
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001525 object = c->page->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001526 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001527 c->page->inuse = s->objects;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001528 c->page->freelist = c->page->end;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001529 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001530unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001531 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001532 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001533 return object;
1534
1535another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001536 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001537
1538new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001539 new = get_partial(s, gfpflags, node);
1540 if (new) {
1541 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001542 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001543 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001544 }
1545
Christoph Lameterb811c202007-10-16 23:25:51 -07001546 if (gfpflags & __GFP_WAIT)
1547 local_irq_enable();
1548
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001549 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001550
1551 if (gfpflags & __GFP_WAIT)
1552 local_irq_disable();
1553
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001554 if (new) {
1555 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001556 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001557 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001558 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001559 slab_lock(new);
1560 SetSlabFrozen(new);
1561 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001562 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001563 }
Linus Torvalds00e962c2008-02-19 09:08:49 -08001564
Christoph Lameter71c7a062008-02-14 14:28:01 -08001565 /*
1566 * No memory available.
1567 *
1568 * If the slab uses higher order allocs but the object is
1569 * smaller than a page size then we can fallback in emergencies
1570 * to the page allocator via kmalloc_large. The page allocator may
1571 * have failed to obtain a higher order page and we can try to
1572 * allocate a single page if the object fits into a single page.
1573 * That is only possible if certain conditions are met that are being
1574 * checked when a slab is created.
1575 */
1576 if (!(gfpflags & __GFP_NORETRY) && (s->flags & __PAGE_ALLOC_FALLBACK))
1577 return kmalloc_large(s->objsize, gfpflags);
1578
1579 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001580debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001581 object = c->page->freelist;
1582 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 Lameter683d0ba2008-01-07 23:20:29 -08001610 if (unlikely(is_end(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;
1662checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001663 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001664 page->freelist = object;
1665 page->inuse--;
1666
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001667 if (unlikely(SlabFrozen(page))) {
1668 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001669 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001670 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001671
1672 if (unlikely(!page->inuse))
1673 goto slab_empty;
1674
1675 /*
1676 * Objects left in the slab. If it
1677 * was not on the partial list before
1678 * then add it.
1679 */
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001680 if (unlikely(prior == page->end)) {
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 Lameter8ff12cf2008-02-07 17:47:41 -08001690 if (prior != page->end) {
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);
Peter Zijlstra02febdf2007-07-26 20:01:38 +02001727 debug_check_no_locks_freed(object, s->objsize);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001728 c = get_cpu_slab(s, smp_processor_id());
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 /*
1891 * If the user wants hardware cache aligned objects then
1892 * follow that suggestion if the object is sufficiently
1893 * large.
1894 *
1895 * The hardware cache alignment cannot override the
1896 * specified alignment though. If that is greater
1897 * then use it.
1898 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001899 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001900 size > cache_line_size() / 2)
1901 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001902
1903 if (align < ARCH_SLAB_MINALIGN)
1904 return ARCH_SLAB_MINALIGN;
1905
1906 return ALIGN(align, sizeof(void *));
1907}
1908
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001909static void init_kmem_cache_cpu(struct kmem_cache *s,
1910 struct kmem_cache_cpu *c)
1911{
1912 c->page = NULL;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001913 c->freelist = (void *)PAGE_MAPPING_ANON;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001914 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001915 c->offset = s->offset / sizeof(void *);
1916 c->objsize = s->objsize;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001917}
1918
Christoph Lameter81819f02007-05-06 14:49:36 -07001919static void init_kmem_cache_node(struct kmem_cache_node *n)
1920{
1921 n->nr_partial = 0;
1922 atomic_long_set(&n->nr_slabs, 0);
1923 spin_lock_init(&n->list_lock);
1924 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001925#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07001926 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001927#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001928}
1929
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001930#ifdef CONFIG_SMP
1931/*
1932 * Per cpu array for per cpu structures.
1933 *
1934 * The per cpu array places all kmem_cache_cpu structures from one processor
1935 * close together meaning that it becomes possible that multiple per cpu
1936 * structures are contained in one cacheline. This may be particularly
1937 * beneficial for the kmalloc caches.
1938 *
1939 * A desktop system typically has around 60-80 slabs. With 100 here we are
1940 * likely able to get per cpu structures for all caches from the array defined
1941 * here. We must be able to cover all kmalloc caches during bootstrap.
1942 *
1943 * If the per cpu array is exhausted then fall back to kmalloc
1944 * of individual cachelines. No sharing is possible then.
1945 */
1946#define NR_KMEM_CACHE_CPU 100
1947
1948static DEFINE_PER_CPU(struct kmem_cache_cpu,
1949 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1950
1951static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1952static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1953
1954static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1955 int cpu, gfp_t flags)
1956{
1957 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1958
1959 if (c)
1960 per_cpu(kmem_cache_cpu_free, cpu) =
1961 (void *)c->freelist;
1962 else {
1963 /* Table overflow: So allocate ourselves */
1964 c = kmalloc_node(
1965 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1966 flags, cpu_to_node(cpu));
1967 if (!c)
1968 return NULL;
1969 }
1970
1971 init_kmem_cache_cpu(s, c);
1972 return c;
1973}
1974
1975static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1976{
1977 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1978 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1979 kfree(c);
1980 return;
1981 }
1982 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1983 per_cpu(kmem_cache_cpu_free, cpu) = c;
1984}
1985
1986static void free_kmem_cache_cpus(struct kmem_cache *s)
1987{
1988 int cpu;
1989
1990 for_each_online_cpu(cpu) {
1991 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1992
1993 if (c) {
1994 s->cpu_slab[cpu] = NULL;
1995 free_kmem_cache_cpu(c, cpu);
1996 }
1997 }
1998}
1999
2000static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2001{
2002 int cpu;
2003
2004 for_each_online_cpu(cpu) {
2005 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2006
2007 if (c)
2008 continue;
2009
2010 c = alloc_kmem_cache_cpu(s, cpu, flags);
2011 if (!c) {
2012 free_kmem_cache_cpus(s);
2013 return 0;
2014 }
2015 s->cpu_slab[cpu] = c;
2016 }
2017 return 1;
2018}
2019
2020/*
2021 * Initialize the per cpu array.
2022 */
2023static void init_alloc_cpu_cpu(int cpu)
2024{
2025 int i;
2026
2027 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2028 return;
2029
2030 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2031 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2032
2033 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2034}
2035
2036static void __init init_alloc_cpu(void)
2037{
2038 int cpu;
2039
2040 for_each_online_cpu(cpu)
2041 init_alloc_cpu_cpu(cpu);
2042 }
2043
2044#else
2045static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2046static inline void init_alloc_cpu(void) {}
2047
2048static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2049{
2050 init_kmem_cache_cpu(s, &s->cpu_slab);
2051 return 1;
2052}
2053#endif
2054
Christoph Lameter81819f02007-05-06 14:49:36 -07002055#ifdef CONFIG_NUMA
2056/*
2057 * No kmalloc_node yet so do it by hand. We know that this is the first
2058 * slab on the node for this slabcache. There are no concurrent accesses
2059 * possible.
2060 *
2061 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002062 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2063 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002064 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002065static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2066 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002067{
2068 struct page *page;
2069 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002070 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002071
2072 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2073
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002074 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002075
2076 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002077 if (page_to_nid(page) != node) {
2078 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2079 "node %d\n", node);
2080 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2081 "in order to be able to continue\n");
2082 }
2083
Christoph Lameter81819f02007-05-06 14:49:36 -07002084 n = page->freelist;
2085 BUG_ON(!n);
2086 page->freelist = get_freepointer(kmalloc_caches, n);
2087 page->inuse++;
2088 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002089#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002090 init_object(kmalloc_caches, n, 1);
2091 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002092#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002093 init_kmem_cache_node(n);
2094 atomic_long_inc(&n->nr_slabs);
rootba84c732008-01-07 23:20:28 -08002095 /*
2096 * lockdep requires consistent irq usage for each lock
2097 * so even though there cannot be a race this early in
2098 * the boot sequence, we still disable irqs.
2099 */
2100 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002101 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002102 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002103 return n;
2104}
2105
2106static void free_kmem_cache_nodes(struct kmem_cache *s)
2107{
2108 int node;
2109
Christoph Lameterf64dc582007-10-16 01:25:33 -07002110 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002111 struct kmem_cache_node *n = s->node[node];
2112 if (n && n != &s->local_node)
2113 kmem_cache_free(kmalloc_caches, n);
2114 s->node[node] = NULL;
2115 }
2116}
2117
2118static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2119{
2120 int node;
2121 int local_node;
2122
2123 if (slab_state >= UP)
2124 local_node = page_to_nid(virt_to_page(s));
2125 else
2126 local_node = 0;
2127
Christoph Lameterf64dc582007-10-16 01:25:33 -07002128 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002129 struct kmem_cache_node *n;
2130
2131 if (local_node == node)
2132 n = &s->local_node;
2133 else {
2134 if (slab_state == DOWN) {
2135 n = early_kmem_cache_node_alloc(gfpflags,
2136 node);
2137 continue;
2138 }
2139 n = kmem_cache_alloc_node(kmalloc_caches,
2140 gfpflags, node);
2141
2142 if (!n) {
2143 free_kmem_cache_nodes(s);
2144 return 0;
2145 }
2146
2147 }
2148 s->node[node] = n;
2149 init_kmem_cache_node(n);
2150 }
2151 return 1;
2152}
2153#else
2154static void free_kmem_cache_nodes(struct kmem_cache *s)
2155{
2156}
2157
2158static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2159{
2160 init_kmem_cache_node(&s->local_node);
2161 return 1;
2162}
2163#endif
2164
2165/*
2166 * calculate_sizes() determines the order and the distribution of data within
2167 * a slab object.
2168 */
2169static int calculate_sizes(struct kmem_cache *s)
2170{
2171 unsigned long flags = s->flags;
2172 unsigned long size = s->objsize;
2173 unsigned long align = s->align;
2174
2175 /*
2176 * Determine if we can poison the object itself. If the user of
2177 * the slab may touch the object after free or before allocation
2178 * then we should never poison the object itself.
2179 */
2180 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002181 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002182 s->flags |= __OBJECT_POISON;
2183 else
2184 s->flags &= ~__OBJECT_POISON;
2185
2186 /*
2187 * Round up object size to the next word boundary. We can only
2188 * place the free pointer at word boundaries and this determines
2189 * the possible location of the free pointer.
2190 */
2191 size = ALIGN(size, sizeof(void *));
2192
Christoph Lameter41ecc552007-05-09 02:32:44 -07002193#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002194 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002195 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002196 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002197 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002198 */
2199 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2200 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002201#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002202
2203 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002204 * With that we have determined the number of bytes in actual use
2205 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002206 */
2207 s->inuse = size;
2208
2209 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002210 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002211 /*
2212 * Relocate free pointer after the object if it is not
2213 * permitted to overwrite the first word of the object on
2214 * kmem_cache_free.
2215 *
2216 * This is the case if we do RCU, have a constructor or
2217 * destructor or are poisoning the objects.
2218 */
2219 s->offset = size;
2220 size += sizeof(void *);
2221 }
2222
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002223#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002224 if (flags & SLAB_STORE_USER)
2225 /*
2226 * Need to store information about allocs and frees after
2227 * the object.
2228 */
2229 size += 2 * sizeof(struct track);
2230
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002231 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002232 /*
2233 * Add some empty padding so that we can catch
2234 * overwrites from earlier objects rather than let
2235 * tracking information or the free pointer be
2236 * corrupted if an user writes before the start
2237 * of the object.
2238 */
2239 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002240#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002241
Christoph Lameter81819f02007-05-06 14:49:36 -07002242 /*
2243 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002244 * user specified and the dynamic determination of cache line size
2245 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002246 */
2247 align = calculate_alignment(flags, align, s->objsize);
2248
2249 /*
2250 * SLUB stores one object immediately after another beginning from
2251 * offset 0. In order to align the objects we have to simply size
2252 * each object to conform to the alignment.
2253 */
2254 size = ALIGN(size, align);
2255 s->size = size;
2256
Christoph Lameter71c7a062008-02-14 14:28:01 -08002257 if ((flags & __KMALLOC_CACHE) &&
2258 PAGE_SIZE / size < slub_min_objects) {
2259 /*
2260 * Kmalloc cache that would not have enough objects in
2261 * an order 0 page. Kmalloc slabs can fallback to
2262 * page allocator order 0 allocs so take a reasonably large
2263 * order that will allows us a good number of objects.
2264 */
2265 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2266 s->flags |= __PAGE_ALLOC_FALLBACK;
2267 s->allocflags |= __GFP_NOWARN;
2268 } else
2269 s->order = calculate_order(size);
2270
Christoph Lameter81819f02007-05-06 14:49:36 -07002271 if (s->order < 0)
2272 return 0;
2273
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002274 s->allocflags = 0;
2275 if (s->order)
2276 s->allocflags |= __GFP_COMP;
2277
2278 if (s->flags & SLAB_CACHE_DMA)
2279 s->allocflags |= SLUB_DMA;
2280
2281 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2282 s->allocflags |= __GFP_RECLAIMABLE;
2283
Christoph Lameter81819f02007-05-06 14:49:36 -07002284 /*
2285 * Determine the number of objects per slab
2286 */
2287 s->objects = (PAGE_SIZE << s->order) / size;
2288
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002289 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002290
2291}
2292
Christoph Lameter81819f02007-05-06 14:49:36 -07002293static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2294 const char *name, size_t size,
2295 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002296 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002297{
2298 memset(s, 0, kmem_size);
2299 s->name = name;
2300 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002301 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002302 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002303 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002304
2305 if (!calculate_sizes(s))
2306 goto error;
2307
2308 s->refcount = 1;
2309#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002310 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002311#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002312 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2313 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002314
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002315 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002316 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002317 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002318error:
2319 if (flags & SLAB_PANIC)
2320 panic("Cannot create slab %s size=%lu realsize=%u "
2321 "order=%u offset=%u flags=%lx\n",
2322 s->name, (unsigned long)size, s->size, s->order,
2323 s->offset, flags);
2324 return 0;
2325}
Christoph Lameter81819f02007-05-06 14:49:36 -07002326
2327/*
2328 * Check if a given pointer is valid
2329 */
2330int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2331{
Pekka Enberg06428782008-01-07 23:20:27 -08002332 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002333
2334 page = get_object_page(object);
2335
2336 if (!page || s != page->slab)
2337 /* No slab or wrong slab */
2338 return 0;
2339
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002340 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002341 return 0;
2342
2343 /*
2344 * We could also check if the object is on the slabs freelist.
2345 * But this would be too expensive and it seems that the main
2346 * purpose of kmem_ptr_valid is to check if the object belongs
2347 * to a certain slab.
2348 */
2349 return 1;
2350}
2351EXPORT_SYMBOL(kmem_ptr_validate);
2352
2353/*
2354 * Determine the size of a slab object
2355 */
2356unsigned int kmem_cache_size(struct kmem_cache *s)
2357{
2358 return s->objsize;
2359}
2360EXPORT_SYMBOL(kmem_cache_size);
2361
2362const char *kmem_cache_name(struct kmem_cache *s)
2363{
2364 return s->name;
2365}
2366EXPORT_SYMBOL(kmem_cache_name);
2367
2368/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002369 * Attempt to free all slabs on a node. Return the number of slabs we
2370 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002371 */
2372static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2373 struct list_head *list)
2374{
2375 int slabs_inuse = 0;
2376 unsigned long flags;
2377 struct page *page, *h;
2378
2379 spin_lock_irqsave(&n->list_lock, flags);
2380 list_for_each_entry_safe(page, h, list, lru)
2381 if (!page->inuse) {
2382 list_del(&page->lru);
2383 discard_slab(s, page);
2384 } else
2385 slabs_inuse++;
2386 spin_unlock_irqrestore(&n->list_lock, flags);
2387 return slabs_inuse;
2388}
2389
2390/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002391 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002392 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002393static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002394{
2395 int node;
2396
2397 flush_all(s);
2398
2399 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002400 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002401 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002402 struct kmem_cache_node *n = get_node(s, node);
2403
Christoph Lameter2086d262007-05-06 14:49:46 -07002404 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002405 if (atomic_long_read(&n->nr_slabs))
2406 return 1;
2407 }
2408 free_kmem_cache_nodes(s);
2409 return 0;
2410}
2411
2412/*
2413 * Close a cache and release the kmem_cache structure
2414 * (must be used for caches created using kmem_cache_create)
2415 */
2416void kmem_cache_destroy(struct kmem_cache *s)
2417{
2418 down_write(&slub_lock);
2419 s->refcount--;
2420 if (!s->refcount) {
2421 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002422 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002423 if (kmem_cache_close(s))
2424 WARN_ON(1);
2425 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002426 } else
2427 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002428}
2429EXPORT_SYMBOL(kmem_cache_destroy);
2430
2431/********************************************************************
2432 * Kmalloc subsystem
2433 *******************************************************************/
2434
Christoph Lameter331dc552008-02-14 14:28:09 -08002435struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002436EXPORT_SYMBOL(kmalloc_caches);
2437
2438#ifdef CONFIG_ZONE_DMA
Christoph Lameter331dc552008-02-14 14:28:09 -08002439static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter81819f02007-05-06 14:49:36 -07002440#endif
2441
2442static int __init setup_slub_min_order(char *str)
2443{
Pekka Enberg06428782008-01-07 23:20:27 -08002444 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002445
2446 return 1;
2447}
2448
2449__setup("slub_min_order=", setup_slub_min_order);
2450
2451static int __init setup_slub_max_order(char *str)
2452{
Pekka Enberg06428782008-01-07 23:20:27 -08002453 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002454
2455 return 1;
2456}
2457
2458__setup("slub_max_order=", setup_slub_max_order);
2459
2460static int __init setup_slub_min_objects(char *str)
2461{
Pekka Enberg06428782008-01-07 23:20:27 -08002462 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002463
2464 return 1;
2465}
2466
2467__setup("slub_min_objects=", setup_slub_min_objects);
2468
2469static int __init setup_slub_nomerge(char *str)
2470{
2471 slub_nomerge = 1;
2472 return 1;
2473}
2474
2475__setup("slub_nomerge", setup_slub_nomerge);
2476
Christoph Lameter81819f02007-05-06 14:49:36 -07002477static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2478 const char *name, int size, gfp_t gfp_flags)
2479{
2480 unsigned int flags = 0;
2481
2482 if (gfp_flags & SLUB_DMA)
2483 flags = SLAB_CACHE_DMA;
2484
2485 down_write(&slub_lock);
2486 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter71c7a062008-02-14 14:28:01 -08002487 flags | __KMALLOC_CACHE, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002488 goto panic;
2489
2490 list_add(&s->list, &slab_caches);
2491 up_write(&slub_lock);
2492 if (sysfs_slab_add(s))
2493 goto panic;
2494 return s;
2495
2496panic:
2497 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2498}
2499
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002500#ifdef CONFIG_ZONE_DMA
Christoph Lameter1ceef402007-08-07 15:11:48 -07002501
2502static void sysfs_add_func(struct work_struct *w)
2503{
2504 struct kmem_cache *s;
2505
2506 down_write(&slub_lock);
2507 list_for_each_entry(s, &slab_caches, list) {
2508 if (s->flags & __SYSFS_ADD_DEFERRED) {
2509 s->flags &= ~__SYSFS_ADD_DEFERRED;
2510 sysfs_slab_add(s);
2511 }
2512 }
2513 up_write(&slub_lock);
2514}
2515
2516static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2517
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002518static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2519{
2520 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002521 char *text;
2522 size_t realsize;
2523
2524 s = kmalloc_caches_dma[index];
2525 if (s)
2526 return s;
2527
2528 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002529 if (flags & __GFP_WAIT)
2530 down_write(&slub_lock);
2531 else {
2532 if (!down_write_trylock(&slub_lock))
2533 goto out;
2534 }
2535
2536 if (kmalloc_caches_dma[index])
2537 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002538
Christoph Lameter7b55f622007-07-17 04:03:27 -07002539 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002540 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2541 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002542 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2543
2544 if (!s || !text || !kmem_cache_open(s, flags, text,
2545 realsize, ARCH_KMALLOC_MINALIGN,
2546 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2547 kfree(s);
2548 kfree(text);
2549 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002550 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002551
2552 list_add(&s->list, &slab_caches);
2553 kmalloc_caches_dma[index] = s;
2554
2555 schedule_work(&sysfs_add_work);
2556
2557unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002558 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002559out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002560 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002561}
2562#endif
2563
Christoph Lameterf1b26332007-07-17 04:03:26 -07002564/*
2565 * Conversion table for small slabs sizes / 8 to the index in the
2566 * kmalloc array. This is necessary for slabs < 192 since we have non power
2567 * of two cache sizes there. The size of larger slabs can be determined using
2568 * fls.
2569 */
2570static s8 size_index[24] = {
2571 3, /* 8 */
2572 4, /* 16 */
2573 5, /* 24 */
2574 5, /* 32 */
2575 6, /* 40 */
2576 6, /* 48 */
2577 6, /* 56 */
2578 6, /* 64 */
2579 1, /* 72 */
2580 1, /* 80 */
2581 1, /* 88 */
2582 1, /* 96 */
2583 7, /* 104 */
2584 7, /* 112 */
2585 7, /* 120 */
2586 7, /* 128 */
2587 2, /* 136 */
2588 2, /* 144 */
2589 2, /* 152 */
2590 2, /* 160 */
2591 2, /* 168 */
2592 2, /* 176 */
2593 2, /* 184 */
2594 2 /* 192 */
2595};
2596
Christoph Lameter81819f02007-05-06 14:49:36 -07002597static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2598{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002599 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002600
Christoph Lameterf1b26332007-07-17 04:03:26 -07002601 if (size <= 192) {
2602 if (!size)
2603 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002604
Christoph Lameterf1b26332007-07-17 04:03:26 -07002605 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002606 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002607 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002608
2609#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002610 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002611 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002612
Christoph Lameter81819f02007-05-06 14:49:36 -07002613#endif
2614 return &kmalloc_caches[index];
2615}
2616
2617void *__kmalloc(size_t size, gfp_t flags)
2618{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002619 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002620
Christoph Lameter331dc552008-02-14 14:28:09 -08002621 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002622 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002623
2624 s = get_slab(size, flags);
2625
2626 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002627 return s;
2628
Christoph Lameterce15fea2007-07-17 04:03:28 -07002629 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002630}
2631EXPORT_SYMBOL(__kmalloc);
2632
2633#ifdef CONFIG_NUMA
2634void *__kmalloc_node(size_t size, gfp_t flags, int node)
2635{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002636 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002637
Christoph Lameter331dc552008-02-14 14:28:09 -08002638 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002639 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002640
2641 s = get_slab(size, flags);
2642
2643 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002644 return s;
2645
Christoph Lameterce15fea2007-07-17 04:03:28 -07002646 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002647}
2648EXPORT_SYMBOL(__kmalloc_node);
2649#endif
2650
2651size_t ksize(const void *object)
2652{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002653 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002654 struct kmem_cache *s;
2655
Christoph Lameteref8b4522007-10-16 01:24:46 -07002656 BUG_ON(!object);
2657 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002658 return 0;
2659
Vegard Nossum294a80a2007-12-04 23:45:30 -08002660 page = virt_to_head_page(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07002661 BUG_ON(!page);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002662
2663 if (unlikely(!PageSlab(page)))
2664 return PAGE_SIZE << compound_order(page);
2665
Christoph Lameter81819f02007-05-06 14:49:36 -07002666 s = page->slab;
2667 BUG_ON(!s);
2668
2669 /*
2670 * Debugging requires use of the padding between object
2671 * and whatever may come after it.
2672 */
2673 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2674 return s->objsize;
2675
2676 /*
2677 * If we have the need to store the freelist pointer
2678 * back there or track user information then we can
2679 * only use the space before that information.
2680 */
2681 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2682 return s->inuse;
2683
2684 /*
2685 * Else we can use all the padding etc for the allocation
2686 */
2687 return s->size;
2688}
2689EXPORT_SYMBOL(ksize);
2690
2691void kfree(const void *x)
2692{
Christoph Lameter81819f02007-05-06 14:49:36 -07002693 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002694 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002695
Satyam Sharma2408c552007-10-16 01:24:44 -07002696 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002697 return;
2698
Christoph Lameterb49af682007-05-06 14:49:41 -07002699 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002700 if (unlikely(!PageSlab(page))) {
2701 put_page(page);
2702 return;
2703 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002704 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002705}
2706EXPORT_SYMBOL(kfree);
2707
Christoph Lameterf61396a2008-01-07 23:20:26 -08002708static unsigned long count_partial(struct kmem_cache_node *n)
2709{
2710 unsigned long flags;
2711 unsigned long x = 0;
2712 struct page *page;
2713
2714 spin_lock_irqsave(&n->list_lock, flags);
2715 list_for_each_entry(page, &n->partial, lru)
2716 x += page->inuse;
2717 spin_unlock_irqrestore(&n->list_lock, flags);
2718 return x;
2719}
2720
Christoph Lameter2086d262007-05-06 14:49:46 -07002721/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002722 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2723 * the remaining slabs by the number of items in use. The slabs with the
2724 * most items in use come first. New allocations will then fill those up
2725 * and thus they can be removed from the partial lists.
2726 *
2727 * The slabs with the least items are placed last. This results in them
2728 * being allocated from last increasing the chance that the last objects
2729 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002730 */
2731int kmem_cache_shrink(struct kmem_cache *s)
2732{
2733 int node;
2734 int i;
2735 struct kmem_cache_node *n;
2736 struct page *page;
2737 struct page *t;
2738 struct list_head *slabs_by_inuse =
2739 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2740 unsigned long flags;
2741
2742 if (!slabs_by_inuse)
2743 return -ENOMEM;
2744
2745 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002746 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002747 n = get_node(s, node);
2748
2749 if (!n->nr_partial)
2750 continue;
2751
2752 for (i = 0; i < s->objects; i++)
2753 INIT_LIST_HEAD(slabs_by_inuse + i);
2754
2755 spin_lock_irqsave(&n->list_lock, flags);
2756
2757 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002758 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002759 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002760 * Note that concurrent frees may occur while we hold the
2761 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002762 */
2763 list_for_each_entry_safe(page, t, &n->partial, lru) {
2764 if (!page->inuse && slab_trylock(page)) {
2765 /*
2766 * Must hold slab lock here because slab_free
2767 * may have freed the last object and be
2768 * waiting to release the slab.
2769 */
2770 list_del(&page->lru);
2771 n->nr_partial--;
2772 slab_unlock(page);
2773 discard_slab(s, page);
2774 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002775 list_move(&page->lru,
2776 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002777 }
2778 }
2779
Christoph Lameter2086d262007-05-06 14:49:46 -07002780 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002781 * Rebuild the partial list with the slabs filled up most
2782 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002783 */
2784 for (i = s->objects - 1; i >= 0; i--)
2785 list_splice(slabs_by_inuse + i, n->partial.prev);
2786
Christoph Lameter2086d262007-05-06 14:49:46 -07002787 spin_unlock_irqrestore(&n->list_lock, flags);
2788 }
2789
2790 kfree(slabs_by_inuse);
2791 return 0;
2792}
2793EXPORT_SYMBOL(kmem_cache_shrink);
2794
Yasunori Gotob9049e22007-10-21 16:41:37 -07002795#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2796static int slab_mem_going_offline_callback(void *arg)
2797{
2798 struct kmem_cache *s;
2799
2800 down_read(&slub_lock);
2801 list_for_each_entry(s, &slab_caches, list)
2802 kmem_cache_shrink(s);
2803 up_read(&slub_lock);
2804
2805 return 0;
2806}
2807
2808static void slab_mem_offline_callback(void *arg)
2809{
2810 struct kmem_cache_node *n;
2811 struct kmem_cache *s;
2812 struct memory_notify *marg = arg;
2813 int offline_node;
2814
2815 offline_node = marg->status_change_nid;
2816
2817 /*
2818 * If the node still has available memory. we need kmem_cache_node
2819 * for it yet.
2820 */
2821 if (offline_node < 0)
2822 return;
2823
2824 down_read(&slub_lock);
2825 list_for_each_entry(s, &slab_caches, list) {
2826 n = get_node(s, offline_node);
2827 if (n) {
2828 /*
2829 * if n->nr_slabs > 0, slabs still exist on the node
2830 * that is going down. We were unable to free them,
2831 * and offline_pages() function shoudn't call this
2832 * callback. So, we must fail.
2833 */
Al Viro27bb6282007-10-29 04:42:55 +00002834 BUG_ON(atomic_long_read(&n->nr_slabs));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002835
2836 s->node[offline_node] = NULL;
2837 kmem_cache_free(kmalloc_caches, n);
2838 }
2839 }
2840 up_read(&slub_lock);
2841}
2842
2843static int slab_mem_going_online_callback(void *arg)
2844{
2845 struct kmem_cache_node *n;
2846 struct kmem_cache *s;
2847 struct memory_notify *marg = arg;
2848 int nid = marg->status_change_nid;
2849 int ret = 0;
2850
2851 /*
2852 * If the node's memory is already available, then kmem_cache_node is
2853 * already created. Nothing to do.
2854 */
2855 if (nid < 0)
2856 return 0;
2857
2858 /*
2859 * We are bringing a node online. No memory is availabe yet. We must
2860 * allocate a kmem_cache_node structure in order to bring the node
2861 * online.
2862 */
2863 down_read(&slub_lock);
2864 list_for_each_entry(s, &slab_caches, list) {
2865 /*
2866 * XXX: kmem_cache_alloc_node will fallback to other nodes
2867 * since memory is not yet available from the node that
2868 * is brought up.
2869 */
2870 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2871 if (!n) {
2872 ret = -ENOMEM;
2873 goto out;
2874 }
2875 init_kmem_cache_node(n);
2876 s->node[nid] = n;
2877 }
2878out:
2879 up_read(&slub_lock);
2880 return ret;
2881}
2882
2883static int slab_memory_callback(struct notifier_block *self,
2884 unsigned long action, void *arg)
2885{
2886 int ret = 0;
2887
2888 switch (action) {
2889 case MEM_GOING_ONLINE:
2890 ret = slab_mem_going_online_callback(arg);
2891 break;
2892 case MEM_GOING_OFFLINE:
2893 ret = slab_mem_going_offline_callback(arg);
2894 break;
2895 case MEM_OFFLINE:
2896 case MEM_CANCEL_ONLINE:
2897 slab_mem_offline_callback(arg);
2898 break;
2899 case MEM_ONLINE:
2900 case MEM_CANCEL_OFFLINE:
2901 break;
2902 }
2903
2904 ret = notifier_from_errno(ret);
2905 return ret;
2906}
2907
2908#endif /* CONFIG_MEMORY_HOTPLUG */
2909
Christoph Lameter81819f02007-05-06 14:49:36 -07002910/********************************************************************
2911 * Basic setup of slabs
2912 *******************************************************************/
2913
2914void __init kmem_cache_init(void)
2915{
2916 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002917 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002918
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002919 init_alloc_cpu();
2920
Christoph Lameter81819f02007-05-06 14:49:36 -07002921#ifdef CONFIG_NUMA
2922 /*
2923 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002924 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002925 * kmem_cache_open for slab_state == DOWN.
2926 */
2927 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2928 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002929 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002930 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002931
2932 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002933#endif
2934
2935 /* Able to allocate the per node structures */
2936 slab_state = PARTIAL;
2937
2938 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002939 if (KMALLOC_MIN_SIZE <= 64) {
2940 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002941 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002942 caches++;
2943 }
2944 if (KMALLOC_MIN_SIZE <= 128) {
2945 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002946 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002947 caches++;
2948 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002949
Christoph Lameter331dc552008-02-14 14:28:09 -08002950 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002951 create_kmalloc_cache(&kmalloc_caches[i],
2952 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002953 caches++;
2954 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002955
Christoph Lameterf1b26332007-07-17 04:03:26 -07002956
2957 /*
2958 * Patch up the size_index table if we have strange large alignment
2959 * requirements for the kmalloc array. This is only the case for
2960 * mips it seems. The standard arches will not generate any code here.
2961 *
2962 * Largest permitted alignment is 256 bytes due to the way we
2963 * handle the index determination for the smaller caches.
2964 *
2965 * Make sure that nothing crazy happens if someone starts tinkering
2966 * around with ARCH_KMALLOC_MINALIGN
2967 */
2968 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2969 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2970
Christoph Lameter12ad6842007-07-17 04:03:28 -07002971 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002972 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2973
Christoph Lameter81819f02007-05-06 14:49:36 -07002974 slab_state = UP;
2975
2976 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08002977 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002978 kmalloc_caches[i]. name =
2979 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2980
2981#ifdef CONFIG_SMP
2982 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002983 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2984 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2985#else
2986 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002987#endif
2988
Christoph Lameter81819f02007-05-06 14:49:36 -07002989
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002990 printk(KERN_INFO
2991 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002992 " CPUs=%d, Nodes=%d\n",
2993 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002994 slub_min_order, slub_max_order, slub_min_objects,
2995 nr_cpu_ids, nr_node_ids);
2996}
2997
2998/*
2999 * Find a mergeable slab cache
3000 */
3001static int slab_unmergeable(struct kmem_cache *s)
3002{
3003 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3004 return 1;
3005
Christoph Lameter331dc552008-02-14 14:28:09 -08003006 if ((s->flags & __PAGE_ALLOC_FALLBACK))
Christoph Lameter71c7a062008-02-14 14:28:01 -08003007 return 1;
3008
Christoph Lameterc59def92007-05-16 22:10:50 -07003009 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003010 return 1;
3011
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003012 /*
3013 * We may have set a slab to be unmergeable during bootstrap.
3014 */
3015 if (s->refcount < 0)
3016 return 1;
3017
Christoph Lameter81819f02007-05-06 14:49:36 -07003018 return 0;
3019}
3020
3021static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003022 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003023 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003024{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003025 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003026
3027 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3028 return NULL;
3029
Christoph Lameterc59def92007-05-16 22:10:50 -07003030 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003031 return NULL;
3032
3033 size = ALIGN(size, sizeof(void *));
3034 align = calculate_alignment(flags, align, size);
3035 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003036 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003037
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003038 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003039 if (slab_unmergeable(s))
3040 continue;
3041
3042 if (size > s->size)
3043 continue;
3044
Christoph Lameterba0268a2007-09-11 15:24:11 -07003045 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003046 continue;
3047 /*
3048 * Check if alignment is compatible.
3049 * Courtesy of Adrian Drzewiecki
3050 */
Pekka Enberg06428782008-01-07 23:20:27 -08003051 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003052 continue;
3053
3054 if (s->size - size >= sizeof(void *))
3055 continue;
3056
3057 return s;
3058 }
3059 return NULL;
3060}
3061
3062struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3063 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003064 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003065{
3066 struct kmem_cache *s;
3067
3068 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003069 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003070 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003071 int cpu;
3072
Christoph Lameter81819f02007-05-06 14:49:36 -07003073 s->refcount++;
3074 /*
3075 * Adjust the object sizes so that we clear
3076 * the complete object on kzalloc.
3077 */
3078 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003079
3080 /*
3081 * And then we need to update the object size in the
3082 * per cpu structures
3083 */
3084 for_each_online_cpu(cpu)
3085 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter81819f02007-05-06 14:49:36 -07003086 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003087 up_write(&slub_lock);
Christoph 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 }
3092 s = kmalloc(kmem_size, GFP_KERNEL);
3093 if (s) {
3094 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003095 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003096 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003097 up_write(&slub_lock);
3098 if (sysfs_slab_add(s))
3099 goto err;
3100 return s;
3101 }
3102 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003103 }
3104 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003105
3106err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003107 if (flags & SLAB_PANIC)
3108 panic("Cannot create slabcache %s\n", name);
3109 else
3110 s = NULL;
3111 return s;
3112}
3113EXPORT_SYMBOL(kmem_cache_create);
3114
Christoph Lameter81819f02007-05-06 14:49:36 -07003115#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003116/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003117 * Use the cpu notifier to insure that the cpu slabs are flushed when
3118 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003119 */
3120static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3121 unsigned long action, void *hcpu)
3122{
3123 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003124 struct kmem_cache *s;
3125 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003126
3127 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003128 case CPU_UP_PREPARE:
3129 case CPU_UP_PREPARE_FROZEN:
3130 init_alloc_cpu_cpu(cpu);
3131 down_read(&slub_lock);
3132 list_for_each_entry(s, &slab_caches, list)
3133 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3134 GFP_KERNEL);
3135 up_read(&slub_lock);
3136 break;
3137
Christoph Lameter81819f02007-05-06 14:49:36 -07003138 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003139 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003140 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003141 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003142 down_read(&slub_lock);
3143 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003144 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3145
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003146 local_irq_save(flags);
3147 __flush_cpu_slab(s, cpu);
3148 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003149 free_kmem_cache_cpu(c, cpu);
3150 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003151 }
3152 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003153 break;
3154 default:
3155 break;
3156 }
3157 return NOTIFY_OK;
3158}
3159
Pekka Enberg06428782008-01-07 23:20:27 -08003160static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003161 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003162};
Christoph Lameter81819f02007-05-06 14:49:36 -07003163
3164#endif
3165
Christoph Lameter81819f02007-05-06 14:49:36 -07003166void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3167{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003168 struct kmem_cache *s;
3169
Christoph Lameter331dc552008-02-14 14:28:09 -08003170 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003171 return kmalloc_large(size, gfpflags);
3172
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003173 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003174
Satyam Sharma2408c552007-10-16 01:24:44 -07003175 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003176 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003177
Christoph Lameterce15fea2007-07-17 04:03:28 -07003178 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003179}
3180
3181void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3182 int node, void *caller)
3183{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003184 struct kmem_cache *s;
3185
Christoph Lameter331dc552008-02-14 14:28:09 -08003186 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003187 return kmalloc_large(size, gfpflags);
3188
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003189 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003190
Satyam Sharma2408c552007-10-16 01:24:44 -07003191 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003192 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003193
Christoph Lameterce15fea2007-07-17 04:03:28 -07003194 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003195}
3196
Christoph Lameter41ecc552007-05-09 02:32:44 -07003197#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003198static int validate_slab(struct kmem_cache *s, struct page *page,
3199 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003200{
3201 void *p;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08003202 void *addr = slab_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003203
3204 if (!check_slab(s, page) ||
3205 !on_freelist(s, page, NULL))
3206 return 0;
3207
3208 /* Now we know that a valid freelist exists */
3209 bitmap_zero(map, s->objects);
3210
Christoph Lameter7656c722007-05-09 02:32:40 -07003211 for_each_free_object(p, s, page->freelist) {
3212 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003213 if (!check_object(s, page, p, 0))
3214 return 0;
3215 }
3216
Christoph Lameter7656c722007-05-09 02:32:40 -07003217 for_each_object(p, s, addr)
3218 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003219 if (!check_object(s, page, p, 1))
3220 return 0;
3221 return 1;
3222}
3223
Christoph Lameter434e2452007-07-17 04:03:30 -07003224static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3225 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003226{
3227 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003228 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003229 slab_unlock(page);
3230 } else
3231 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3232 s->name, page);
3233
3234 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003235 if (!SlabDebug(page))
3236 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003237 "on slab 0x%p\n", s->name, page);
3238 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003239 if (SlabDebug(page))
3240 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003241 "slab 0x%p\n", s->name, page);
3242 }
3243}
3244
Christoph Lameter434e2452007-07-17 04:03:30 -07003245static int validate_slab_node(struct kmem_cache *s,
3246 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003247{
3248 unsigned long count = 0;
3249 struct page *page;
3250 unsigned long flags;
3251
3252 spin_lock_irqsave(&n->list_lock, flags);
3253
3254 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003255 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003256 count++;
3257 }
3258 if (count != n->nr_partial)
3259 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3260 "counter=%ld\n", s->name, count, n->nr_partial);
3261
3262 if (!(s->flags & SLAB_STORE_USER))
3263 goto out;
3264
3265 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003266 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003267 count++;
3268 }
3269 if (count != atomic_long_read(&n->nr_slabs))
3270 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3271 "counter=%ld\n", s->name, count,
3272 atomic_long_read(&n->nr_slabs));
3273
3274out:
3275 spin_unlock_irqrestore(&n->list_lock, flags);
3276 return count;
3277}
3278
Christoph Lameter434e2452007-07-17 04:03:30 -07003279static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003280{
3281 int node;
3282 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003283 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3284 sizeof(unsigned long), GFP_KERNEL);
3285
3286 if (!map)
3287 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003288
3289 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003290 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003291 struct kmem_cache_node *n = get_node(s, node);
3292
Christoph Lameter434e2452007-07-17 04:03:30 -07003293 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003294 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003295 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003296 return count;
3297}
3298
Christoph Lameterb3459702007-05-09 02:32:41 -07003299#ifdef SLUB_RESILIENCY_TEST
3300static void resiliency_test(void)
3301{
3302 u8 *p;
3303
3304 printk(KERN_ERR "SLUB resiliency testing\n");
3305 printk(KERN_ERR "-----------------------\n");
3306 printk(KERN_ERR "A. Corruption after allocation\n");
3307
3308 p = kzalloc(16, GFP_KERNEL);
3309 p[16] = 0x12;
3310 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3311 " 0x12->0x%p\n\n", p + 16);
3312
3313 validate_slab_cache(kmalloc_caches + 4);
3314
3315 /* Hmmm... The next two are dangerous */
3316 p = kzalloc(32, GFP_KERNEL);
3317 p[32 + sizeof(void *)] = 0x34;
3318 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003319 " 0x34 -> -0x%p\n", p);
3320 printk(KERN_ERR
3321 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003322
3323 validate_slab_cache(kmalloc_caches + 5);
3324 p = kzalloc(64, GFP_KERNEL);
3325 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3326 *p = 0x56;
3327 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3328 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003329 printk(KERN_ERR
3330 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003331 validate_slab_cache(kmalloc_caches + 6);
3332
3333 printk(KERN_ERR "\nB. Corruption after free\n");
3334 p = kzalloc(128, GFP_KERNEL);
3335 kfree(p);
3336 *p = 0x78;
3337 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3338 validate_slab_cache(kmalloc_caches + 7);
3339
3340 p = kzalloc(256, GFP_KERNEL);
3341 kfree(p);
3342 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003343 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3344 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003345 validate_slab_cache(kmalloc_caches + 8);
3346
3347 p = kzalloc(512, GFP_KERNEL);
3348 kfree(p);
3349 p[512] = 0xab;
3350 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3351 validate_slab_cache(kmalloc_caches + 9);
3352}
3353#else
3354static void resiliency_test(void) {};
3355#endif
3356
Christoph Lameter88a420e2007-05-06 14:49:45 -07003357/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003358 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003359 * and freed.
3360 */
3361
3362struct location {
3363 unsigned long count;
3364 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003365 long long sum_time;
3366 long min_time;
3367 long max_time;
3368 long min_pid;
3369 long max_pid;
3370 cpumask_t cpus;
3371 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003372};
3373
3374struct loc_track {
3375 unsigned long max;
3376 unsigned long count;
3377 struct location *loc;
3378};
3379
3380static void free_loc_track(struct loc_track *t)
3381{
3382 if (t->max)
3383 free_pages((unsigned long)t->loc,
3384 get_order(sizeof(struct location) * t->max));
3385}
3386
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003387static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003388{
3389 struct location *l;
3390 int order;
3391
Christoph Lameter88a420e2007-05-06 14:49:45 -07003392 order = get_order(sizeof(struct location) * max);
3393
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003394 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003395 if (!l)
3396 return 0;
3397
3398 if (t->count) {
3399 memcpy(l, t->loc, sizeof(struct location) * t->count);
3400 free_loc_track(t);
3401 }
3402 t->max = max;
3403 t->loc = l;
3404 return 1;
3405}
3406
3407static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003408 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003409{
3410 long start, end, pos;
3411 struct location *l;
3412 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003413 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003414
3415 start = -1;
3416 end = t->count;
3417
3418 for ( ; ; ) {
3419 pos = start + (end - start + 1) / 2;
3420
3421 /*
3422 * There is nothing at "end". If we end up there
3423 * we need to add something to before end.
3424 */
3425 if (pos == end)
3426 break;
3427
3428 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003429 if (track->addr == caddr) {
3430
3431 l = &t->loc[pos];
3432 l->count++;
3433 if (track->when) {
3434 l->sum_time += age;
3435 if (age < l->min_time)
3436 l->min_time = age;
3437 if (age > l->max_time)
3438 l->max_time = age;
3439
3440 if (track->pid < l->min_pid)
3441 l->min_pid = track->pid;
3442 if (track->pid > l->max_pid)
3443 l->max_pid = track->pid;
3444
3445 cpu_set(track->cpu, l->cpus);
3446 }
3447 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003448 return 1;
3449 }
3450
Christoph Lameter45edfa52007-05-09 02:32:45 -07003451 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003452 end = pos;
3453 else
3454 start = pos;
3455 }
3456
3457 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003458 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003459 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003460 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003461 return 0;
3462
3463 l = t->loc + pos;
3464 if (pos < t->count)
3465 memmove(l + 1, l,
3466 (t->count - pos) * sizeof(struct location));
3467 t->count++;
3468 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003469 l->addr = track->addr;
3470 l->sum_time = age;
3471 l->min_time = age;
3472 l->max_time = age;
3473 l->min_pid = track->pid;
3474 l->max_pid = track->pid;
3475 cpus_clear(l->cpus);
3476 cpu_set(track->cpu, l->cpus);
3477 nodes_clear(l->nodes);
3478 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003479 return 1;
3480}
3481
3482static void process_slab(struct loc_track *t, struct kmem_cache *s,
3483 struct page *page, enum track_item alloc)
3484{
Christoph Lameter683d0ba2008-01-07 23:20:29 -08003485 void *addr = slab_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003486 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003487 void *p;
3488
3489 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003490 for_each_free_object(p, s, page->freelist)
3491 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003492
Christoph Lameter7656c722007-05-09 02:32:40 -07003493 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003494 if (!test_bit(slab_index(p, s, addr), map))
3495 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003496}
3497
3498static int list_locations(struct kmem_cache *s, char *buf,
3499 enum track_item alloc)
3500{
Harvey Harrisone374d482008-01-31 15:20:50 -08003501 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003502 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003503 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003504 int node;
3505
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003506 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003507 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003508 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003509
3510 /* Push back cpu slabs */
3511 flush_all(s);
3512
Christoph Lameterf64dc582007-10-16 01:25:33 -07003513 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003514 struct kmem_cache_node *n = get_node(s, node);
3515 unsigned long flags;
3516 struct page *page;
3517
Christoph Lameter9e869432007-08-22 14:01:56 -07003518 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003519 continue;
3520
3521 spin_lock_irqsave(&n->list_lock, flags);
3522 list_for_each_entry(page, &n->partial, lru)
3523 process_slab(&t, s, page, alloc);
3524 list_for_each_entry(page, &n->full, lru)
3525 process_slab(&t, s, page, alloc);
3526 spin_unlock_irqrestore(&n->list_lock, flags);
3527 }
3528
3529 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003530 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003531
Harvey Harrisone374d482008-01-31 15:20:50 -08003532 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003533 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003534 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003535
3536 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003537 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003538 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003539 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003540
3541 if (l->sum_time != l->min_time) {
3542 unsigned long remainder;
3543
Harvey Harrisone374d482008-01-31 15:20:50 -08003544 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003545 l->min_time,
3546 div_long_long_rem(l->sum_time, l->count, &remainder),
3547 l->max_time);
3548 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003549 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003550 l->min_time);
3551
3552 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003553 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003554 l->min_pid, l->max_pid);
3555 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003556 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003557 l->min_pid);
3558
Christoph Lameter84966342007-06-23 17:16:32 -07003559 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003560 len < PAGE_SIZE - 60) {
3561 len += sprintf(buf + len, " cpus=");
3562 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003563 l->cpus);
3564 }
3565
Christoph Lameter84966342007-06-23 17:16:32 -07003566 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003567 len < PAGE_SIZE - 60) {
3568 len += sprintf(buf + len, " nodes=");
3569 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003570 l->nodes);
3571 }
3572
Harvey Harrisone374d482008-01-31 15:20:50 -08003573 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003574 }
3575
3576 free_loc_track(&t);
3577 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003578 len += sprintf(buf, "No data\n");
3579 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003580}
3581
Christoph Lameter81819f02007-05-06 14:49:36 -07003582enum slab_stat_type {
3583 SL_FULL,
3584 SL_PARTIAL,
3585 SL_CPU,
3586 SL_OBJECTS
3587};
3588
3589#define SO_FULL (1 << SL_FULL)
3590#define SO_PARTIAL (1 << SL_PARTIAL)
3591#define SO_CPU (1 << SL_CPU)
3592#define SO_OBJECTS (1 << SL_OBJECTS)
3593
3594static unsigned long slab_objects(struct kmem_cache *s,
3595 char *buf, unsigned long flags)
3596{
3597 unsigned long total = 0;
3598 int cpu;
3599 int node;
3600 int x;
3601 unsigned long *nodes;
3602 unsigned long *per_cpu;
3603
3604 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3605 per_cpu = nodes + nr_node_ids;
3606
3607 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003608 struct page *page;
3609 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003610
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003611 if (!c)
3612 continue;
3613
3614 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003615 node = c->node;
3616 if (node < 0)
3617 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003618 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003619 if (flags & SO_CPU) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003620 if (flags & SO_OBJECTS)
3621 x = page->inuse;
3622 else
3623 x = 1;
3624 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003625 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003626 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003627 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003628 }
3629 }
3630
Christoph Lameterf64dc582007-10-16 01:25:33 -07003631 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003632 struct kmem_cache_node *n = get_node(s, node);
3633
3634 if (flags & SO_PARTIAL) {
3635 if (flags & SO_OBJECTS)
3636 x = count_partial(n);
3637 else
3638 x = n->nr_partial;
3639 total += x;
3640 nodes[node] += x;
3641 }
3642
3643 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003644 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003645 - per_cpu[node]
3646 - n->nr_partial;
3647
3648 if (flags & SO_OBJECTS)
3649 x = full_slabs * s->objects;
3650 else
3651 x = full_slabs;
3652 total += x;
3653 nodes[node] += x;
3654 }
3655 }
3656
3657 x = sprintf(buf, "%lu", total);
3658#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003659 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003660 if (nodes[node])
3661 x += sprintf(buf + x, " N%d=%lu",
3662 node, nodes[node]);
3663#endif
3664 kfree(nodes);
3665 return x + sprintf(buf + x, "\n");
3666}
3667
3668static int any_slab_objects(struct kmem_cache *s)
3669{
3670 int node;
3671 int cpu;
3672
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003673 for_each_possible_cpu(cpu) {
3674 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003675
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003676 if (c && c->page)
3677 return 1;
3678 }
3679
3680 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003681 struct kmem_cache_node *n = get_node(s, node);
3682
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003683 if (!n)
3684 continue;
3685
Christoph Lameter9e869432007-08-22 14:01:56 -07003686 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003687 return 1;
3688 }
3689 return 0;
3690}
3691
3692#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3693#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3694
3695struct slab_attribute {
3696 struct attribute attr;
3697 ssize_t (*show)(struct kmem_cache *s, char *buf);
3698 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3699};
3700
3701#define SLAB_ATTR_RO(_name) \
3702 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3703
3704#define SLAB_ATTR(_name) \
3705 static struct slab_attribute _name##_attr = \
3706 __ATTR(_name, 0644, _name##_show, _name##_store)
3707
Christoph Lameter81819f02007-05-06 14:49:36 -07003708static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3709{
3710 return sprintf(buf, "%d\n", s->size);
3711}
3712SLAB_ATTR_RO(slab_size);
3713
3714static ssize_t align_show(struct kmem_cache *s, char *buf)
3715{
3716 return sprintf(buf, "%d\n", s->align);
3717}
3718SLAB_ATTR_RO(align);
3719
3720static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3721{
3722 return sprintf(buf, "%d\n", s->objsize);
3723}
3724SLAB_ATTR_RO(object_size);
3725
3726static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3727{
3728 return sprintf(buf, "%d\n", s->objects);
3729}
3730SLAB_ATTR_RO(objs_per_slab);
3731
3732static ssize_t order_show(struct kmem_cache *s, char *buf)
3733{
3734 return sprintf(buf, "%d\n", s->order);
3735}
3736SLAB_ATTR_RO(order);
3737
3738static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3739{
3740 if (s->ctor) {
3741 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3742
3743 return n + sprintf(buf + n, "\n");
3744 }
3745 return 0;
3746}
3747SLAB_ATTR_RO(ctor);
3748
Christoph Lameter81819f02007-05-06 14:49:36 -07003749static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3750{
3751 return sprintf(buf, "%d\n", s->refcount - 1);
3752}
3753SLAB_ATTR_RO(aliases);
3754
3755static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3756{
3757 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3758}
3759SLAB_ATTR_RO(slabs);
3760
3761static ssize_t partial_show(struct kmem_cache *s, char *buf)
3762{
3763 return slab_objects(s, buf, SO_PARTIAL);
3764}
3765SLAB_ATTR_RO(partial);
3766
3767static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3768{
3769 return slab_objects(s, buf, SO_CPU);
3770}
3771SLAB_ATTR_RO(cpu_slabs);
3772
3773static ssize_t objects_show(struct kmem_cache *s, char *buf)
3774{
3775 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3776}
3777SLAB_ATTR_RO(objects);
3778
3779static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3780{
3781 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3782}
3783
3784static ssize_t sanity_checks_store(struct kmem_cache *s,
3785 const char *buf, size_t length)
3786{
3787 s->flags &= ~SLAB_DEBUG_FREE;
3788 if (buf[0] == '1')
3789 s->flags |= SLAB_DEBUG_FREE;
3790 return length;
3791}
3792SLAB_ATTR(sanity_checks);
3793
3794static ssize_t trace_show(struct kmem_cache *s, char *buf)
3795{
3796 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3797}
3798
3799static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3800 size_t length)
3801{
3802 s->flags &= ~SLAB_TRACE;
3803 if (buf[0] == '1')
3804 s->flags |= SLAB_TRACE;
3805 return length;
3806}
3807SLAB_ATTR(trace);
3808
3809static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3810{
3811 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3812}
3813
3814static ssize_t reclaim_account_store(struct kmem_cache *s,
3815 const char *buf, size_t length)
3816{
3817 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3818 if (buf[0] == '1')
3819 s->flags |= SLAB_RECLAIM_ACCOUNT;
3820 return length;
3821}
3822SLAB_ATTR(reclaim_account);
3823
3824static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3825{
Christoph Lameter5af60832007-05-06 14:49:56 -07003826 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003827}
3828SLAB_ATTR_RO(hwcache_align);
3829
3830#ifdef CONFIG_ZONE_DMA
3831static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3832{
3833 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3834}
3835SLAB_ATTR_RO(cache_dma);
3836#endif
3837
3838static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3839{
3840 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3841}
3842SLAB_ATTR_RO(destroy_by_rcu);
3843
3844static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3845{
3846 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3847}
3848
3849static ssize_t red_zone_store(struct kmem_cache *s,
3850 const char *buf, size_t length)
3851{
3852 if (any_slab_objects(s))
3853 return -EBUSY;
3854
3855 s->flags &= ~SLAB_RED_ZONE;
3856 if (buf[0] == '1')
3857 s->flags |= SLAB_RED_ZONE;
3858 calculate_sizes(s);
3859 return length;
3860}
3861SLAB_ATTR(red_zone);
3862
3863static ssize_t poison_show(struct kmem_cache *s, char *buf)
3864{
3865 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3866}
3867
3868static ssize_t poison_store(struct kmem_cache *s,
3869 const char *buf, size_t length)
3870{
3871 if (any_slab_objects(s))
3872 return -EBUSY;
3873
3874 s->flags &= ~SLAB_POISON;
3875 if (buf[0] == '1')
3876 s->flags |= SLAB_POISON;
3877 calculate_sizes(s);
3878 return length;
3879}
3880SLAB_ATTR(poison);
3881
3882static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3883{
3884 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3885}
3886
3887static ssize_t store_user_store(struct kmem_cache *s,
3888 const char *buf, size_t length)
3889{
3890 if (any_slab_objects(s))
3891 return -EBUSY;
3892
3893 s->flags &= ~SLAB_STORE_USER;
3894 if (buf[0] == '1')
3895 s->flags |= SLAB_STORE_USER;
3896 calculate_sizes(s);
3897 return length;
3898}
3899SLAB_ATTR(store_user);
3900
Christoph Lameter53e15af2007-05-06 14:49:43 -07003901static ssize_t validate_show(struct kmem_cache *s, char *buf)
3902{
3903 return 0;
3904}
3905
3906static ssize_t validate_store(struct kmem_cache *s,
3907 const char *buf, size_t length)
3908{
Christoph Lameter434e2452007-07-17 04:03:30 -07003909 int ret = -EINVAL;
3910
3911 if (buf[0] == '1') {
3912 ret = validate_slab_cache(s);
3913 if (ret >= 0)
3914 ret = length;
3915 }
3916 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003917}
3918SLAB_ATTR(validate);
3919
Christoph Lameter2086d262007-05-06 14:49:46 -07003920static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3921{
3922 return 0;
3923}
3924
3925static ssize_t shrink_store(struct kmem_cache *s,
3926 const char *buf, size_t length)
3927{
3928 if (buf[0] == '1') {
3929 int rc = kmem_cache_shrink(s);
3930
3931 if (rc)
3932 return rc;
3933 } else
3934 return -EINVAL;
3935 return length;
3936}
3937SLAB_ATTR(shrink);
3938
Christoph Lameter88a420e2007-05-06 14:49:45 -07003939static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3940{
3941 if (!(s->flags & SLAB_STORE_USER))
3942 return -ENOSYS;
3943 return list_locations(s, buf, TRACK_ALLOC);
3944}
3945SLAB_ATTR_RO(alloc_calls);
3946
3947static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3948{
3949 if (!(s->flags & SLAB_STORE_USER))
3950 return -ENOSYS;
3951 return list_locations(s, buf, TRACK_FREE);
3952}
3953SLAB_ATTR_RO(free_calls);
3954
Christoph Lameter81819f02007-05-06 14:49:36 -07003955#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08003956static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07003957{
Christoph Lameter98246012008-01-07 23:20:26 -08003958 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07003959}
3960
Christoph Lameter98246012008-01-07 23:20:26 -08003961static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07003962 const char *buf, size_t length)
3963{
3964 int n = simple_strtoul(buf, NULL, 10);
3965
3966 if (n < 100)
Christoph Lameter98246012008-01-07 23:20:26 -08003967 s->remote_node_defrag_ratio = n * 10;
Christoph Lameter81819f02007-05-06 14:49:36 -07003968 return length;
3969}
Christoph Lameter98246012008-01-07 23:20:26 -08003970SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07003971#endif
3972
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003973#ifdef CONFIG_SLUB_STATS
3974
3975static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3976{
3977 unsigned long sum = 0;
3978 int cpu;
3979 int len;
3980 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3981
3982 if (!data)
3983 return -ENOMEM;
3984
3985 for_each_online_cpu(cpu) {
3986 unsigned x = get_cpu_slab(s, cpu)->stat[si];
3987
3988 data[cpu] = x;
3989 sum += x;
3990 }
3991
3992 len = sprintf(buf, "%lu", sum);
3993
3994 for_each_online_cpu(cpu) {
3995 if (data[cpu] && len < PAGE_SIZE - 20)
3996 len += sprintf(buf + len, " c%d=%u", cpu, data[cpu]);
3997 }
3998 kfree(data);
3999 return len + sprintf(buf + len, "\n");
4000}
4001
4002#define STAT_ATTR(si, text) \
4003static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4004{ \
4005 return show_stat(s, buf, si); \
4006} \
4007SLAB_ATTR_RO(text); \
4008
4009STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4010STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4011STAT_ATTR(FREE_FASTPATH, free_fastpath);
4012STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4013STAT_ATTR(FREE_FROZEN, free_frozen);
4014STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4015STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4016STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4017STAT_ATTR(ALLOC_SLAB, alloc_slab);
4018STAT_ATTR(ALLOC_REFILL, alloc_refill);
4019STAT_ATTR(FREE_SLAB, free_slab);
4020STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4021STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4022STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4023STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4024STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4025STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4026
4027#endif
4028
Pekka Enberg06428782008-01-07 23:20:27 -08004029static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004030 &slab_size_attr.attr,
4031 &object_size_attr.attr,
4032 &objs_per_slab_attr.attr,
4033 &order_attr.attr,
4034 &objects_attr.attr,
4035 &slabs_attr.attr,
4036 &partial_attr.attr,
4037 &cpu_slabs_attr.attr,
4038 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004039 &aliases_attr.attr,
4040 &align_attr.attr,
4041 &sanity_checks_attr.attr,
4042 &trace_attr.attr,
4043 &hwcache_align_attr.attr,
4044 &reclaim_account_attr.attr,
4045 &destroy_by_rcu_attr.attr,
4046 &red_zone_attr.attr,
4047 &poison_attr.attr,
4048 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004049 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004050 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004051 &alloc_calls_attr.attr,
4052 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004053#ifdef CONFIG_ZONE_DMA
4054 &cache_dma_attr.attr,
4055#endif
4056#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004057 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004058#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004059#ifdef CONFIG_SLUB_STATS
4060 &alloc_fastpath_attr.attr,
4061 &alloc_slowpath_attr.attr,
4062 &free_fastpath_attr.attr,
4063 &free_slowpath_attr.attr,
4064 &free_frozen_attr.attr,
4065 &free_add_partial_attr.attr,
4066 &free_remove_partial_attr.attr,
4067 &alloc_from_partial_attr.attr,
4068 &alloc_slab_attr.attr,
4069 &alloc_refill_attr.attr,
4070 &free_slab_attr.attr,
4071 &cpuslab_flush_attr.attr,
4072 &deactivate_full_attr.attr,
4073 &deactivate_empty_attr.attr,
4074 &deactivate_to_head_attr.attr,
4075 &deactivate_to_tail_attr.attr,
4076 &deactivate_remote_frees_attr.attr,
4077#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004078 NULL
4079};
4080
4081static struct attribute_group slab_attr_group = {
4082 .attrs = slab_attrs,
4083};
4084
4085static ssize_t slab_attr_show(struct kobject *kobj,
4086 struct attribute *attr,
4087 char *buf)
4088{
4089 struct slab_attribute *attribute;
4090 struct kmem_cache *s;
4091 int err;
4092
4093 attribute = to_slab_attr(attr);
4094 s = to_slab(kobj);
4095
4096 if (!attribute->show)
4097 return -EIO;
4098
4099 err = attribute->show(s, buf);
4100
4101 return err;
4102}
4103
4104static ssize_t slab_attr_store(struct kobject *kobj,
4105 struct attribute *attr,
4106 const char *buf, size_t len)
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->store)
4116 return -EIO;
4117
4118 err = attribute->store(s, buf, len);
4119
4120 return err;
4121}
4122
Christoph Lameter151c6022008-01-07 22:29:05 -08004123static void kmem_cache_release(struct kobject *kobj)
4124{
4125 struct kmem_cache *s = to_slab(kobj);
4126
4127 kfree(s);
4128}
4129
Christoph Lameter81819f02007-05-06 14:49:36 -07004130static struct sysfs_ops slab_sysfs_ops = {
4131 .show = slab_attr_show,
4132 .store = slab_attr_store,
4133};
4134
4135static struct kobj_type slab_ktype = {
4136 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004137 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004138};
4139
4140static int uevent_filter(struct kset *kset, struct kobject *kobj)
4141{
4142 struct kobj_type *ktype = get_ktype(kobj);
4143
4144 if (ktype == &slab_ktype)
4145 return 1;
4146 return 0;
4147}
4148
4149static struct kset_uevent_ops slab_uevent_ops = {
4150 .filter = uevent_filter,
4151};
4152
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004153static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004154
4155#define ID_STR_LENGTH 64
4156
4157/* Create a unique string id for a slab cache:
4158 * format
4159 * :[flags-]size:[memory address of kmemcache]
4160 */
4161static char *create_unique_id(struct kmem_cache *s)
4162{
4163 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4164 char *p = name;
4165
4166 BUG_ON(!name);
4167
4168 *p++ = ':';
4169 /*
4170 * First flags affecting slabcache operations. We will only
4171 * get here for aliasable slabs so we do not need to support
4172 * too many flags. The flags here must cover all flags that
4173 * are matched during merging to guarantee that the id is
4174 * unique.
4175 */
4176 if (s->flags & SLAB_CACHE_DMA)
4177 *p++ = 'd';
4178 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4179 *p++ = 'a';
4180 if (s->flags & SLAB_DEBUG_FREE)
4181 *p++ = 'F';
4182 if (p != name + 1)
4183 *p++ = '-';
4184 p += sprintf(p, "%07d", s->size);
4185 BUG_ON(p > name + ID_STR_LENGTH - 1);
4186 return name;
4187}
4188
4189static int sysfs_slab_add(struct kmem_cache *s)
4190{
4191 int err;
4192 const char *name;
4193 int unmergeable;
4194
4195 if (slab_state < SYSFS)
4196 /* Defer until later */
4197 return 0;
4198
4199 unmergeable = slab_unmergeable(s);
4200 if (unmergeable) {
4201 /*
4202 * Slabcache can never be merged so we can use the name proper.
4203 * This is typically the case for debug situations. In that
4204 * case we can catch duplicate names easily.
4205 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004206 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004207 name = s->name;
4208 } else {
4209 /*
4210 * Create a unique name for the slab as a target
4211 * for the symlinks.
4212 */
4213 name = create_unique_id(s);
4214 }
4215
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004216 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004217 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4218 if (err) {
4219 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004220 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004221 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004222
4223 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4224 if (err)
4225 return err;
4226 kobject_uevent(&s->kobj, KOBJ_ADD);
4227 if (!unmergeable) {
4228 /* Setup first alias */
4229 sysfs_slab_alias(s, s->name);
4230 kfree(name);
4231 }
4232 return 0;
4233}
4234
4235static void sysfs_slab_remove(struct kmem_cache *s)
4236{
4237 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4238 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004239 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004240}
4241
4242/*
4243 * Need to buffer aliases during bootup until sysfs becomes
4244 * available lest we loose that information.
4245 */
4246struct saved_alias {
4247 struct kmem_cache *s;
4248 const char *name;
4249 struct saved_alias *next;
4250};
4251
Adrian Bunk5af328a2007-07-17 04:03:27 -07004252static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004253
4254static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4255{
4256 struct saved_alias *al;
4257
4258 if (slab_state == SYSFS) {
4259 /*
4260 * If we have a leftover link then remove it.
4261 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004262 sysfs_remove_link(&slab_kset->kobj, name);
4263 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004264 }
4265
4266 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4267 if (!al)
4268 return -ENOMEM;
4269
4270 al->s = s;
4271 al->name = name;
4272 al->next = alias_list;
4273 alias_list = al;
4274 return 0;
4275}
4276
4277static int __init slab_sysfs_init(void)
4278{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004279 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004280 int err;
4281
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004282 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004283 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004284 printk(KERN_ERR "Cannot register slab subsystem.\n");
4285 return -ENOSYS;
4286 }
4287
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004288 slab_state = SYSFS;
4289
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004290 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004291 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004292 if (err)
4293 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4294 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004295 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004296
4297 while (alias_list) {
4298 struct saved_alias *al = alias_list;
4299
4300 alias_list = alias_list->next;
4301 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004302 if (err)
4303 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4304 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004305 kfree(al);
4306 }
4307
4308 resiliency_test();
4309 return 0;
4310}
4311
4312__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004313#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004314
4315/*
4316 * The /proc/slabinfo ABI
4317 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004318#ifdef CONFIG_SLABINFO
4319
4320ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4321 size_t count, loff_t *ppos)
4322{
4323 return -EINVAL;
4324}
4325
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004326
4327static void print_slabinfo_header(struct seq_file *m)
4328{
4329 seq_puts(m, "slabinfo - version: 2.1\n");
4330 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4331 "<objperslab> <pagesperslab>");
4332 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4333 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4334 seq_putc(m, '\n');
4335}
4336
4337static void *s_start(struct seq_file *m, loff_t *pos)
4338{
4339 loff_t n = *pos;
4340
4341 down_read(&slub_lock);
4342 if (!n)
4343 print_slabinfo_header(m);
4344
4345 return seq_list_start(&slab_caches, *pos);
4346}
4347
4348static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4349{
4350 return seq_list_next(p, &slab_caches, pos);
4351}
4352
4353static void s_stop(struct seq_file *m, void *p)
4354{
4355 up_read(&slub_lock);
4356}
4357
4358static int s_show(struct seq_file *m, void *p)
4359{
4360 unsigned long nr_partials = 0;
4361 unsigned long nr_slabs = 0;
4362 unsigned long nr_inuse = 0;
4363 unsigned long nr_objs;
4364 struct kmem_cache *s;
4365 int node;
4366
4367 s = list_entry(p, struct kmem_cache, list);
4368
4369 for_each_online_node(node) {
4370 struct kmem_cache_node *n = get_node(s, node);
4371
4372 if (!n)
4373 continue;
4374
4375 nr_partials += n->nr_partial;
4376 nr_slabs += atomic_long_read(&n->nr_slabs);
4377 nr_inuse += count_partial(n);
4378 }
4379
4380 nr_objs = nr_slabs * s->objects;
4381 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4382
4383 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4384 nr_objs, s->size, s->objects, (1 << s->order));
4385 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4386 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4387 0UL);
4388 seq_putc(m, '\n');
4389 return 0;
4390}
4391
4392const struct seq_operations slabinfo_op = {
4393 .start = s_start,
4394 .next = s_next,
4395 .stop = s_stop,
4396 .show = s_show,
4397};
4398
Linus Torvalds158a9622008-01-02 13:04:48 -08004399#endif /* CONFIG_SLABINFO */