blob: e2989ae243b53571bb8ee5695771adbd90185f3d [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
Christoph Lameter1f842602008-01-07 23:20:30 -0800152/*
153 * Currently fastpath is not supported if preemption is enabled.
154 */
155#if defined(CONFIG_FAST_CMPXCHG_LOCAL) && !defined(CONFIG_PREEMPT)
156#define SLUB_FASTPATH
157#endif
158
Christoph Lameter81819f02007-05-06 14:49:36 -0700159#if PAGE_SHIFT <= 12
160
161/*
162 * Small page size. Make sure that we do not fragment memory
163 */
164#define DEFAULT_MAX_ORDER 1
165#define DEFAULT_MIN_OBJECTS 4
166
167#else
168
169/*
170 * Large page machines are customarily able to handle larger
171 * page orders.
172 */
173#define DEFAULT_MAX_ORDER 2
174#define DEFAULT_MIN_OBJECTS 8
175
176#endif
177
178/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700179 * Mininum number of partial slabs. These will be left on the partial
180 * lists even if they are empty. kmem_cache_shrink may reclaim them.
181 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800182#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700183
Christoph Lameter2086d262007-05-06 14:49:46 -0700184/*
185 * Maximum number of desirable partial slabs.
186 * The existence of more partial slabs makes kmem_cache_shrink
187 * sort the partial list by the number of objects in the.
188 */
189#define MAX_PARTIAL 10
190
Christoph Lameter81819f02007-05-06 14:49:36 -0700191#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
192 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700193
Christoph Lameter81819f02007-05-06 14:49:36 -0700194/*
195 * Set of flags that will prevent slab merging
196 */
197#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
198 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
199
200#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
201 SLAB_CACHE_DMA)
202
203#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700204#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700205#endif
206
207#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700208#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700209#endif
210
211/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700212#define __OBJECT_POISON 0x80000000 /* Poison object */
213#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700214
Christoph Lameter65c02d42007-05-09 02:32:35 -0700215/* Not all arches define cache_line_size */
216#ifndef cache_line_size
217#define cache_line_size() L1_CACHE_BYTES
218#endif
219
Christoph Lameter81819f02007-05-06 14:49:36 -0700220static int kmem_size = sizeof(struct kmem_cache);
221
222#ifdef CONFIG_SMP
223static struct notifier_block slab_notifier;
224#endif
225
226static enum {
227 DOWN, /* No slab functionality available */
228 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700229 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700230 SYSFS /* Sysfs up */
231} slab_state = DOWN;
232
233/* A list of all slab caches on the system */
234static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700235static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700236
Christoph Lameter02cbc872007-05-09 02:32:43 -0700237/*
238 * Tracking user of a slab.
239 */
240struct track {
241 void *addr; /* Called from address */
242 int cpu; /* Was running on cpu */
243 int pid; /* Pid context */
244 unsigned long when; /* When did the operation occur */
245};
246
247enum track_item { TRACK_ALLOC, TRACK_FREE };
248
Christoph Lameter41ecc552007-05-09 02:32:44 -0700249#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter81819f02007-05-06 14:49:36 -0700250static int sysfs_slab_add(struct kmem_cache *);
251static int sysfs_slab_alias(struct kmem_cache *, const char *);
252static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800253
Christoph Lameter81819f02007-05-06 14:49:36 -0700254#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700255static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
256static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
257 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800258static inline void sysfs_slab_remove(struct kmem_cache *s)
259{
260 kfree(s);
261}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800262
Christoph Lameter81819f02007-05-06 14:49:36 -0700263#endif
264
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800265static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
266{
267#ifdef CONFIG_SLUB_STATS
268 c->stat[si]++;
269#endif
270}
271
Christoph Lameter81819f02007-05-06 14:49:36 -0700272/********************************************************************
273 * Core slab cache functions
274 *******************************************************************/
275
276int slab_is_available(void)
277{
278 return slab_state >= UP;
279}
280
281static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
282{
283#ifdef CONFIG_NUMA
284 return s->node[node];
285#else
286 return &s->local_node;
287#endif
288}
289
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700290static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
291{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700292#ifdef CONFIG_SMP
293 return s->cpu_slab[cpu];
294#else
295 return &s->cpu_slab;
296#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700297}
298
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800299/*
300 * The end pointer in a slab is special. It points to the first object in the
301 * slab but has bit 0 set to mark it.
302 *
303 * Note that SLUB relies on page_mapping returning NULL for pages with bit 0
304 * in the mapping set.
305 */
306static inline int is_end(void *addr)
307{
308 return (unsigned long)addr & PAGE_MAPPING_ANON;
309}
310
311void *slab_address(struct page *page)
312{
313 return page->end - PAGE_MAPPING_ANON;
314}
315
Christoph Lameter02cbc872007-05-09 02:32:43 -0700316static inline int check_valid_pointer(struct kmem_cache *s,
317 struct page *page, const void *object)
318{
319 void *base;
320
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800321 if (object == page->end)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700322 return 1;
323
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800324 base = slab_address(page);
Christoph Lameter02cbc872007-05-09 02:32:43 -0700325 if (object < base || object >= base + s->objects * s->size ||
326 (object - base) % s->size) {
327 return 0;
328 }
329
330 return 1;
331}
332
Christoph Lameter81819f02007-05-06 14:49:36 -0700333/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700334 * Slow version of get and set free pointer.
335 *
336 * This version requires touching the cache lines of kmem_cache which
337 * we avoid to do in the fast alloc free paths. There we obtain the offset
338 * from the page struct.
339 */
340static inline void *get_freepointer(struct kmem_cache *s, void *object)
341{
342 return *(void **)(object + s->offset);
343}
344
345static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
346{
347 *(void **)(object + s->offset) = fp;
348}
349
350/* Loop over all objects in a slab */
351#define for_each_object(__p, __s, __addr) \
352 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
353 __p += (__s)->size)
354
355/* Scan freelist */
356#define for_each_free_object(__p, __s, __free) \
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800357 for (__p = (__free); (__p) != page->end; __p = get_freepointer((__s),\
358 __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700359
360/* Determine object index from a given position */
361static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
362{
363 return (p - addr) / s->size;
364}
365
Christoph Lameter41ecc552007-05-09 02:32:44 -0700366#ifdef CONFIG_SLUB_DEBUG
367/*
368 * Debug settings:
369 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700370#ifdef CONFIG_SLUB_DEBUG_ON
371static int slub_debug = DEBUG_DEFAULT_FLAGS;
372#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700373static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700374#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700375
376static char *slub_debug_slabs;
377
Christoph Lameter7656c722007-05-09 02:32:40 -0700378/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700379 * Object debugging
380 */
381static void print_section(char *text, u8 *addr, unsigned int length)
382{
383 int i, offset;
384 int newline = 1;
385 char ascii[17];
386
387 ascii[16] = 0;
388
389 for (i = 0; i < length; i++) {
390 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700391 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700392 newline = 0;
393 }
Pekka Enberg06428782008-01-07 23:20:27 -0800394 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700395 offset = i % 16;
396 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
397 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800398 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700399 newline = 1;
400 }
401 }
402 if (!newline) {
403 i %= 16;
404 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800405 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700406 ascii[i] = ' ';
407 i++;
408 }
Pekka Enberg06428782008-01-07 23:20:27 -0800409 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700410 }
411}
412
Christoph Lameter81819f02007-05-06 14:49:36 -0700413static struct track *get_track(struct kmem_cache *s, void *object,
414 enum track_item alloc)
415{
416 struct track *p;
417
418 if (s->offset)
419 p = object + s->offset + sizeof(void *);
420 else
421 p = object + s->inuse;
422
423 return p + alloc;
424}
425
426static void set_track(struct kmem_cache *s, void *object,
427 enum track_item alloc, void *addr)
428{
429 struct track *p;
430
431 if (s->offset)
432 p = object + s->offset + sizeof(void *);
433 else
434 p = object + s->inuse;
435
436 p += alloc;
437 if (addr) {
438 p->addr = addr;
439 p->cpu = smp_processor_id();
440 p->pid = current ? current->pid : -1;
441 p->when = jiffies;
442 } else
443 memset(p, 0, sizeof(struct track));
444}
445
Christoph Lameter81819f02007-05-06 14:49:36 -0700446static void init_tracking(struct kmem_cache *s, void *object)
447{
Christoph Lameter24922682007-07-17 04:03:18 -0700448 if (!(s->flags & SLAB_STORE_USER))
449 return;
450
451 set_track(s, object, TRACK_FREE, NULL);
452 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700453}
454
455static void print_track(const char *s, struct track *t)
456{
457 if (!t->addr)
458 return;
459
Christoph Lameter24922682007-07-17 04:03:18 -0700460 printk(KERN_ERR "INFO: %s in ", s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700461 __print_symbol("%s", (unsigned long)t->addr);
Christoph Lameter24922682007-07-17 04:03:18 -0700462 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700463}
464
Christoph Lameter24922682007-07-17 04:03:18 -0700465static void print_tracking(struct kmem_cache *s, void *object)
466{
467 if (!(s->flags & SLAB_STORE_USER))
468 return;
469
470 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
471 print_track("Freed", get_track(s, object, TRACK_FREE));
472}
473
474static void print_page_info(struct page *page)
475{
476 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
477 page, page->inuse, page->freelist, page->flags);
478
479}
480
481static void slab_bug(struct kmem_cache *s, char *fmt, ...)
482{
483 va_list args;
484 char buf[100];
485
486 va_start(args, fmt);
487 vsnprintf(buf, sizeof(buf), fmt, args);
488 va_end(args);
489 printk(KERN_ERR "========================================"
490 "=====================================\n");
491 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
492 printk(KERN_ERR "----------------------------------------"
493 "-------------------------------------\n\n");
494}
495
496static void slab_fix(struct kmem_cache *s, char *fmt, ...)
497{
498 va_list args;
499 char buf[100];
500
501 va_start(args, fmt);
502 vsnprintf(buf, sizeof(buf), fmt, args);
503 va_end(args);
504 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
505}
506
507static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700508{
509 unsigned int off; /* Offset of last byte */
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800510 u8 *addr = slab_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700511
512 print_tracking(s, p);
513
514 print_page_info(page);
515
516 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
517 p, p - addr, get_freepointer(s, p));
518
519 if (p > addr + 16)
520 print_section("Bytes b4", p - 16, 16);
521
522 print_section("Object", p, min(s->objsize, 128));
Christoph Lameter81819f02007-05-06 14:49:36 -0700523
524 if (s->flags & SLAB_RED_ZONE)
525 print_section("Redzone", p + s->objsize,
526 s->inuse - s->objsize);
527
Christoph Lameter81819f02007-05-06 14:49:36 -0700528 if (s->offset)
529 off = s->offset + sizeof(void *);
530 else
531 off = s->inuse;
532
Christoph Lameter24922682007-07-17 04:03:18 -0700533 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700534 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700535
536 if (off != s->size)
537 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700538 print_section("Padding", p + off, s->size - off);
539
540 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700541}
542
543static void object_err(struct kmem_cache *s, struct page *page,
544 u8 *object, char *reason)
545{
Christoph Lameter24922682007-07-17 04:03:18 -0700546 slab_bug(s, reason);
547 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700548}
549
Christoph Lameter24922682007-07-17 04:03:18 -0700550static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700551{
552 va_list args;
553 char buf[100];
554
Christoph Lameter24922682007-07-17 04:03:18 -0700555 va_start(args, fmt);
556 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700557 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700558 slab_bug(s, fmt);
559 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700560 dump_stack();
561}
562
563static void init_object(struct kmem_cache *s, void *object, int active)
564{
565 u8 *p = object;
566
567 if (s->flags & __OBJECT_POISON) {
568 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800569 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700570 }
571
572 if (s->flags & SLAB_RED_ZONE)
573 memset(p + s->objsize,
574 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
575 s->inuse - s->objsize);
576}
577
Christoph Lameter24922682007-07-17 04:03:18 -0700578static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700579{
580 while (bytes) {
581 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700582 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700583 start++;
584 bytes--;
585 }
Christoph Lameter24922682007-07-17 04:03:18 -0700586 return NULL;
587}
588
589static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
590 void *from, void *to)
591{
592 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
593 memset(from, data, to - from);
594}
595
596static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
597 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800598 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700599{
600 u8 *fault;
601 u8 *end;
602
603 fault = check_bytes(start, value, bytes);
604 if (!fault)
605 return 1;
606
607 end = start + bytes;
608 while (end > fault && end[-1] == value)
609 end--;
610
611 slab_bug(s, "%s overwritten", what);
612 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
613 fault, end - 1, fault[0], value);
614 print_trailer(s, page, object);
615
616 restore_bytes(s, what, value, fault, end);
617 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700618}
619
Christoph Lameter81819f02007-05-06 14:49:36 -0700620/*
621 * Object layout:
622 *
623 * object address
624 * Bytes of the object to be managed.
625 * If the freepointer may overlay the object then the free
626 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700627 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700628 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
629 * 0xa5 (POISON_END)
630 *
631 * object + s->objsize
632 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700633 * Padding is extended by another word if Redzoning is enabled and
634 * objsize == inuse.
635 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700636 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
637 * 0xcc (RED_ACTIVE) for objects in use.
638 *
639 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700640 * Meta data starts here.
641 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700642 * A. Free pointer (if we cannot overwrite object on free)
643 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700644 * C. Padding to reach required alignment boundary or at mininum
645 * one word if debuggin is on to be able to detect writes
646 * before the word boundary.
647 *
648 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700649 *
650 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700651 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700652 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700653 * If slabcaches are merged then the objsize and inuse boundaries are mostly
654 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700655 * may be used with merged slabcaches.
656 */
657
Christoph Lameter81819f02007-05-06 14:49:36 -0700658static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
659{
660 unsigned long off = s->inuse; /* The end of info */
661
662 if (s->offset)
663 /* Freepointer is placed after the object. */
664 off += sizeof(void *);
665
666 if (s->flags & SLAB_STORE_USER)
667 /* We also have user information there */
668 off += 2 * sizeof(struct track);
669
670 if (s->size == off)
671 return 1;
672
Christoph Lameter24922682007-07-17 04:03:18 -0700673 return check_bytes_and_report(s, page, p, "Object padding",
674 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700675}
676
677static int slab_pad_check(struct kmem_cache *s, struct page *page)
678{
Christoph Lameter24922682007-07-17 04:03:18 -0700679 u8 *start;
680 u8 *fault;
681 u8 *end;
682 int length;
683 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700684
685 if (!(s->flags & SLAB_POISON))
686 return 1;
687
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800688 start = slab_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700689 end = start + (PAGE_SIZE << s->order);
Christoph Lameter81819f02007-05-06 14:49:36 -0700690 length = s->objects * s->size;
Christoph Lameter24922682007-07-17 04:03:18 -0700691 remainder = end - (start + length);
Christoph Lameter81819f02007-05-06 14:49:36 -0700692 if (!remainder)
693 return 1;
694
Christoph Lameter24922682007-07-17 04:03:18 -0700695 fault = check_bytes(start + length, POISON_INUSE, remainder);
696 if (!fault)
697 return 1;
698 while (end > fault && end[-1] == POISON_INUSE)
699 end--;
700
701 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
702 print_section("Padding", start, length);
703
704 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
705 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700706}
707
708static int check_object(struct kmem_cache *s, struct page *page,
709 void *object, int active)
710{
711 u8 *p = object;
712 u8 *endobject = object + s->objsize;
713
714 if (s->flags & SLAB_RED_ZONE) {
715 unsigned int red =
716 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
717
Christoph Lameter24922682007-07-17 04:03:18 -0700718 if (!check_bytes_and_report(s, page, object, "Redzone",
719 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700720 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700721 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800722 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
723 check_bytes_and_report(s, page, p, "Alignment padding",
724 endobject, POISON_INUSE, s->inuse - s->objsize);
725 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700726 }
727
728 if (s->flags & SLAB_POISON) {
729 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700730 (!check_bytes_and_report(s, page, p, "Poison", p,
731 POISON_FREE, s->objsize - 1) ||
732 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800733 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700734 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700735 /*
736 * check_pad_bytes cleans up on its own.
737 */
738 check_pad_bytes(s, page, p);
739 }
740
741 if (!s->offset && active)
742 /*
743 * Object and freepointer overlap. Cannot check
744 * freepointer while object is allocated.
745 */
746 return 1;
747
748 /* Check free pointer validity */
749 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
750 object_err(s, page, p, "Freepointer corrupt");
751 /*
752 * No choice but to zap it and thus loose the remainder
753 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700754 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700755 */
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800756 set_freepointer(s, p, page->end);
Christoph Lameter81819f02007-05-06 14:49:36 -0700757 return 0;
758 }
759 return 1;
760}
761
762static int check_slab(struct kmem_cache *s, struct page *page)
763{
764 VM_BUG_ON(!irqs_disabled());
765
766 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700767 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700768 return 0;
769 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700770 if (page->inuse > s->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700771 slab_err(s, page, "inuse %u > max %u",
772 s->name, page->inuse, s->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700773 return 0;
774 }
775 /* Slab_pad_check fixes things up after itself */
776 slab_pad_check(s, page);
777 return 1;
778}
779
780/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700781 * Determine if a certain object on a page is on the freelist. Must hold the
782 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700783 */
784static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
785{
786 int nr = 0;
787 void *fp = page->freelist;
788 void *object = NULL;
789
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800790 while (fp != page->end && nr <= s->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700791 if (fp == search)
792 return 1;
793 if (!check_valid_pointer(s, page, fp)) {
794 if (object) {
795 object_err(s, page, object,
796 "Freechain corrupt");
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800797 set_freepointer(s, object, page->end);
Christoph Lameter81819f02007-05-06 14:49:36 -0700798 break;
799 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700800 slab_err(s, page, "Freepointer corrupt");
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800801 page->freelist = page->end;
Christoph Lameter81819f02007-05-06 14:49:36 -0700802 page->inuse = s->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700803 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700804 return 0;
805 }
806 break;
807 }
808 object = fp;
809 fp = get_freepointer(s, object);
810 nr++;
811 }
812
813 if (page->inuse != s->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700814 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter24922682007-07-17 04:03:18 -0700815 "counted were %d", page->inuse, s->objects - nr);
Christoph Lameter81819f02007-05-06 14:49:36 -0700816 page->inuse = s->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700817 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700818 }
819 return search == NULL;
820}
821
Christoph Lameter3ec09742007-05-16 22:11:00 -0700822static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
823{
824 if (s->flags & SLAB_TRACE) {
825 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
826 s->name,
827 alloc ? "alloc" : "free",
828 object, page->inuse,
829 page->freelist);
830
831 if (!alloc)
832 print_section("Object", (void *)object, s->objsize);
833
834 dump_stack();
835 }
836}
837
Christoph Lameter643b1132007-05-06 14:49:42 -0700838/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700839 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700840 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700841static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700842{
Christoph Lameter643b1132007-05-06 14:49:42 -0700843 spin_lock(&n->list_lock);
844 list_add(&page->lru, &n->full);
845 spin_unlock(&n->list_lock);
846}
847
848static void remove_full(struct kmem_cache *s, struct page *page)
849{
850 struct kmem_cache_node *n;
851
852 if (!(s->flags & SLAB_STORE_USER))
853 return;
854
855 n = get_node(s, page_to_nid(page));
856
857 spin_lock(&n->list_lock);
858 list_del(&page->lru);
859 spin_unlock(&n->list_lock);
860}
861
Christoph Lameter3ec09742007-05-16 22:11:00 -0700862static void setup_object_debug(struct kmem_cache *s, struct page *page,
863 void *object)
864{
865 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
866 return;
867
868 init_object(s, object, 0);
869 init_tracking(s, object);
870}
871
872static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
873 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700874{
875 if (!check_slab(s, page))
876 goto bad;
877
878 if (object && !on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700879 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700880 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700881 }
882
883 if (!check_valid_pointer(s, page, object)) {
884 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700885 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700886 }
887
Christoph Lameter3ec09742007-05-16 22:11:00 -0700888 if (object && !check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700889 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700890
Christoph Lameter3ec09742007-05-16 22:11:00 -0700891 /* Success perform special debug activities for allocs */
892 if (s->flags & SLAB_STORE_USER)
893 set_track(s, object, TRACK_ALLOC, addr);
894 trace(s, page, object, 1);
895 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700896 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700897
Christoph Lameter81819f02007-05-06 14:49:36 -0700898bad:
899 if (PageSlab(page)) {
900 /*
901 * If this is a slab page then lets do the best we can
902 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700903 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700904 */
Christoph Lameter24922682007-07-17 04:03:18 -0700905 slab_fix(s, "Marking all objects used");
Christoph Lameter81819f02007-05-06 14:49:36 -0700906 page->inuse = s->objects;
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800907 page->freelist = page->end;
Christoph Lameter81819f02007-05-06 14:49:36 -0700908 }
909 return 0;
910}
911
Christoph Lameter3ec09742007-05-16 22:11:00 -0700912static int free_debug_processing(struct kmem_cache *s, struct page *page,
913 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700914{
915 if (!check_slab(s, page))
916 goto fail;
917
918 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700919 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700920 goto fail;
921 }
922
923 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700924 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700925 goto fail;
926 }
927
928 if (!check_object(s, page, object, 1))
929 return 0;
930
931 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800932 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700933 slab_err(s, page, "Attempt to free object(0x%p) "
934 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800935 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700936 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700937 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700938 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700939 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800940 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700941 object_err(s, page, object,
942 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700943 goto fail;
944 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700945
946 /* Special debug activities for freeing objects */
Christoph Lameter683d0ba2008-01-07 23:20:29 -0800947 if (!SlabFrozen(page) && page->freelist == page->end)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700948 remove_full(s, page);
949 if (s->flags & SLAB_STORE_USER)
950 set_track(s, object, TRACK_FREE, addr);
951 trace(s, page, object, 0);
952 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700953 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700954
Christoph Lameter81819f02007-05-06 14:49:36 -0700955fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700956 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700957 return 0;
958}
959
Christoph Lameter41ecc552007-05-09 02:32:44 -0700960static int __init setup_slub_debug(char *str)
961{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700962 slub_debug = DEBUG_DEFAULT_FLAGS;
963 if (*str++ != '=' || !*str)
964 /*
965 * No options specified. Switch on full debugging.
966 */
967 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700968
969 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700970 /*
971 * No options but restriction on slabs. This means full
972 * debugging for slabs matching a pattern.
973 */
974 goto check_slabs;
975
976 slub_debug = 0;
977 if (*str == '-')
978 /*
979 * Switch off all debugging measures.
980 */
981 goto out;
982
983 /*
984 * Determine which debug features should be switched on
985 */
Pekka Enberg06428782008-01-07 23:20:27 -0800986 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700987 switch (tolower(*str)) {
988 case 'f':
989 slub_debug |= SLAB_DEBUG_FREE;
990 break;
991 case 'z':
992 slub_debug |= SLAB_RED_ZONE;
993 break;
994 case 'p':
995 slub_debug |= SLAB_POISON;
996 break;
997 case 'u':
998 slub_debug |= SLAB_STORE_USER;
999 break;
1000 case 't':
1001 slub_debug |= SLAB_TRACE;
1002 break;
1003 default:
1004 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001005 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001006 }
1007 }
1008
1009check_slabs:
1010 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001011 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001012out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001013 return 1;
1014}
1015
1016__setup("slub_debug", setup_slub_debug);
1017
Christoph Lameterba0268a2007-09-11 15:24:11 -07001018static unsigned long kmem_cache_flags(unsigned long objsize,
1019 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001020 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001021{
1022 /*
1023 * The page->offset field is only 16 bit wide. This is an offset
1024 * in units of words from the beginning of an object. If the slab
1025 * size is bigger then we cannot move the free pointer behind the
1026 * object anymore.
1027 *
1028 * On 32 bit platforms the limit is 256k. On 64bit platforms
1029 * the limit is 512k.
1030 *
Christoph Lameterc59def92007-05-16 22:10:50 -07001031 * Debugging or ctor may create a need to move the free
Christoph Lameter41ecc552007-05-09 02:32:44 -07001032 * pointer. Fail if this happens.
1033 */
Christoph Lameterba0268a2007-09-11 15:24:11 -07001034 if (objsize >= 65535 * sizeof(void *)) {
1035 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
Christoph Lameter41ecc552007-05-09 02:32:44 -07001036 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
Christoph Lameterba0268a2007-09-11 15:24:11 -07001037 BUG_ON(ctor);
1038 } else {
Christoph Lameter41ecc552007-05-09 02:32:44 -07001039 /*
1040 * Enable debugging if selected on the kernel commandline.
1041 */
1042 if (slub_debug && (!slub_debug_slabs ||
Christoph Lameterba0268a2007-09-11 15:24:11 -07001043 strncmp(slub_debug_slabs, name,
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001044 strlen(slub_debug_slabs)) == 0))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001045 flags |= slub_debug;
1046 }
1047
1048 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001049}
1050#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001051static inline void setup_object_debug(struct kmem_cache *s,
1052 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001053
Christoph Lameter3ec09742007-05-16 22:11:00 -07001054static inline int alloc_debug_processing(struct kmem_cache *s,
1055 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001056
Christoph Lameter3ec09742007-05-16 22:11:00 -07001057static inline int free_debug_processing(struct kmem_cache *s,
1058 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001059
Christoph Lameter41ecc552007-05-09 02:32:44 -07001060static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1061 { return 1; }
1062static inline int check_object(struct kmem_cache *s, struct page *page,
1063 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001064static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001065static inline unsigned long kmem_cache_flags(unsigned long objsize,
1066 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001067 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001068{
1069 return flags;
1070}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001071#define slub_debug 0
1072#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001073/*
1074 * Slab allocation and freeing
1075 */
1076static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1077{
Pekka Enberg06428782008-01-07 23:20:27 -08001078 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001079 int pages = 1 << s->order;
1080
1081 if (s->order)
1082 flags |= __GFP_COMP;
1083
1084 if (s->flags & SLAB_CACHE_DMA)
1085 flags |= SLUB_DMA;
1086
Mel Gormane12ba742007-10-16 01:25:52 -07001087 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1088 flags |= __GFP_RECLAIMABLE;
1089
Christoph Lameter81819f02007-05-06 14:49:36 -07001090 if (node == -1)
1091 page = alloc_pages(flags, s->order);
1092 else
1093 page = alloc_pages_node(node, flags, s->order);
1094
1095 if (!page)
1096 return NULL;
1097
1098 mod_zone_page_state(page_zone(page),
1099 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1100 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1101 pages);
1102
1103 return page;
1104}
1105
1106static void setup_object(struct kmem_cache *s, struct page *page,
1107 void *object)
1108{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001109 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001110 if (unlikely(s->ctor))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001111 s->ctor(s, object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001112}
1113
1114static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1115{
1116 struct page *page;
1117 struct kmem_cache_node *n;
1118 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001119 void *last;
1120 void *p;
1121
Christoph Lameter6cb06222007-10-16 01:25:41 -07001122 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001123
Christoph Lameter6cb06222007-10-16 01:25:41 -07001124 page = allocate_slab(s,
1125 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001126 if (!page)
1127 goto out;
1128
1129 n = get_node(s, page_to_nid(page));
1130 if (n)
1131 atomic_long_inc(&n->nr_slabs);
Christoph Lameter81819f02007-05-06 14:49:36 -07001132 page->slab = s;
1133 page->flags |= 1 << PG_slab;
1134 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1135 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001136 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001137
1138 start = page_address(page);
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001139 page->end = start + 1;
Christoph Lameter81819f02007-05-06 14:49:36 -07001140
1141 if (unlikely(s->flags & SLAB_POISON))
1142 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1143
1144 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001145 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001146 setup_object(s, page, last);
1147 set_freepointer(s, last, p);
1148 last = p;
1149 }
1150 setup_object(s, page, last);
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001151 set_freepointer(s, last, page->end);
Christoph Lameter81819f02007-05-06 14:49:36 -07001152
1153 page->freelist = start;
1154 page->inuse = 0;
1155out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001156 return page;
1157}
1158
1159static void __free_slab(struct kmem_cache *s, struct page *page)
1160{
1161 int pages = 1 << s->order;
1162
Christoph Lameterc59def92007-05-16 22:10:50 -07001163 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001164 void *p;
1165
1166 slab_pad_check(s, page);
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001167 for_each_object(p, s, slab_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001168 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001169 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001170 }
1171
1172 mod_zone_page_state(page_zone(page),
1173 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1174 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001175 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001176
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001177 page->mapping = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001178 __free_pages(page, s->order);
1179}
1180
1181static void rcu_free_slab(struct rcu_head *h)
1182{
1183 struct page *page;
1184
1185 page = container_of((struct list_head *)h, struct page, lru);
1186 __free_slab(page->slab, page);
1187}
1188
1189static void free_slab(struct kmem_cache *s, struct page *page)
1190{
1191 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1192 /*
1193 * RCU free overloads the RCU head over the LRU
1194 */
1195 struct rcu_head *head = (void *)&page->lru;
1196
1197 call_rcu(head, rcu_free_slab);
1198 } else
1199 __free_slab(s, page);
1200}
1201
1202static void discard_slab(struct kmem_cache *s, struct page *page)
1203{
1204 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1205
1206 atomic_long_dec(&n->nr_slabs);
1207 reset_page_mapcount(page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001208 __ClearPageSlab(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001209 free_slab(s, page);
1210}
1211
1212/*
1213 * Per slab locking using the pagelock
1214 */
1215static __always_inline void slab_lock(struct page *page)
1216{
1217 bit_spin_lock(PG_locked, &page->flags);
1218}
1219
1220static __always_inline void slab_unlock(struct page *page)
1221{
Nick Piggina76d3542008-01-07 23:20:27 -08001222 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001223}
1224
1225static __always_inline int slab_trylock(struct page *page)
1226{
1227 int rc = 1;
1228
1229 rc = bit_spin_trylock(PG_locked, &page->flags);
1230 return rc;
1231}
1232
1233/*
1234 * Management of partially allocated slabs
1235 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001236static void add_partial(struct kmem_cache_node *n,
1237 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001238{
Christoph Lametere95eed52007-05-06 14:49:44 -07001239 spin_lock(&n->list_lock);
1240 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001241 if (tail)
1242 list_add_tail(&page->lru, &n->partial);
1243 else
1244 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001245 spin_unlock(&n->list_lock);
1246}
1247
1248static void remove_partial(struct kmem_cache *s,
1249 struct page *page)
1250{
1251 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1252
1253 spin_lock(&n->list_lock);
1254 list_del(&page->lru);
1255 n->nr_partial--;
1256 spin_unlock(&n->list_lock);
1257}
1258
1259/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001260 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001261 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001262 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001263 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001264static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001265{
1266 if (slab_trylock(page)) {
1267 list_del(&page->lru);
1268 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001269 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001270 return 1;
1271 }
1272 return 0;
1273}
1274
1275/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001276 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001277 */
1278static struct page *get_partial_node(struct kmem_cache_node *n)
1279{
1280 struct page *page;
1281
1282 /*
1283 * Racy check. If we mistakenly see no partial slabs then we
1284 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001285 * partial slab and there is none available then get_partials()
1286 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001287 */
1288 if (!n || !n->nr_partial)
1289 return NULL;
1290
1291 spin_lock(&n->list_lock);
1292 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001293 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001294 goto out;
1295 page = NULL;
1296out:
1297 spin_unlock(&n->list_lock);
1298 return page;
1299}
1300
1301/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001302 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001303 */
1304static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1305{
1306#ifdef CONFIG_NUMA
1307 struct zonelist *zonelist;
1308 struct zone **z;
1309 struct page *page;
1310
1311 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001312 * The defrag ratio allows a configuration of the tradeoffs between
1313 * inter node defragmentation and node local allocations. A lower
1314 * defrag_ratio increases the tendency to do local allocations
1315 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001316 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001317 * If the defrag_ratio is set to 0 then kmalloc() always
1318 * returns node local objects. If the ratio is higher then kmalloc()
1319 * may return off node objects because partial slabs are obtained
1320 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001321 *
1322 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001323 * defrag_ratio = 1000) then every (well almost) allocation will
1324 * first attempt to defrag slab caches on other nodes. This means
1325 * scanning over all nodes to look for partial slabs which may be
1326 * expensive if we do it every time we are trying to find a slab
1327 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001328 */
Christoph Lameter98246012008-01-07 23:20:26 -08001329 if (!s->remote_node_defrag_ratio ||
1330 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001331 return NULL;
1332
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001333 zonelist = &NODE_DATA(
1334 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
Christoph Lameter81819f02007-05-06 14:49:36 -07001335 for (z = zonelist->zones; *z; z++) {
1336 struct kmem_cache_node *n;
1337
1338 n = get_node(s, zone_to_nid(*z));
1339
1340 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001341 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001342 page = get_partial_node(n);
1343 if (page)
1344 return page;
1345 }
1346 }
1347#endif
1348 return NULL;
1349}
1350
1351/*
1352 * Get a partial page, lock it and return it.
1353 */
1354static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1355{
1356 struct page *page;
1357 int searchnode = (node == -1) ? numa_node_id() : node;
1358
1359 page = get_partial_node(get_node(s, searchnode));
1360 if (page || (flags & __GFP_THISNODE))
1361 return page;
1362
1363 return get_any_partial(s, flags);
1364}
1365
1366/*
1367 * Move a page back to the lists.
1368 *
1369 * Must be called with the slab lock held.
1370 *
1371 * On exit the slab lock will have been dropped.
1372 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001373static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001374{
Christoph Lametere95eed52007-05-06 14:49:44 -07001375 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001376 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001377
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001378 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001379 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001380
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001381 if (page->freelist != page->end) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001382 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001383 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1384 } else {
1385 stat(c, DEACTIVATE_FULL);
1386 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1387 add_full(n, page);
1388 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001389 slab_unlock(page);
1390 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001391 stat(c, DEACTIVATE_EMPTY);
Christoph Lametere95eed52007-05-06 14:49:44 -07001392 if (n->nr_partial < MIN_PARTIAL) {
1393 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001394 * Adding an empty slab to the partial slabs in order
1395 * to avoid page allocator overhead. This slab needs
1396 * to come after the other slabs with objects in
1397 * order to fill them up. That way the size of the
1398 * partial list stays small. kmem_cache_shrink can
1399 * reclaim empty slabs from the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001400 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001401 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001402 slab_unlock(page);
1403 } else {
1404 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001405 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001406 discard_slab(s, page);
1407 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001408 }
1409}
1410
1411/*
1412 * Remove the cpu slab
1413 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001414static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001415{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001416 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001417 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001418
1419 if (c->freelist)
1420 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001421 /*
1422 * Merge cpu freelist into freelist. Typically we get here
1423 * because both freelists are empty. So this is unlikely
1424 * to occur.
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001425 *
1426 * We need to use _is_end here because deactivate slab may
1427 * be called for a debug slab. Then c->freelist may contain
1428 * a dummy pointer.
Christoph Lameter894b8782007-05-10 03:15:16 -07001429 */
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001430 while (unlikely(!is_end(c->freelist))) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001431 void **object;
1432
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001433 tail = 0; /* Hot objects. Put the slab first */
1434
Christoph Lameter894b8782007-05-10 03:15:16 -07001435 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001436 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001437 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001438
1439 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001440 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001441 page->freelist = object;
1442 page->inuse--;
1443 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001444 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001445 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001446}
1447
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001448static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001449{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001450 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001451 slab_lock(c->page);
1452 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001453}
1454
1455/*
1456 * Flush cpu slab.
1457 * Called from IPI handler with interrupts disabled.
1458 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001459static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001460{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001461 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001462
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001463 if (likely(c && c->page))
1464 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001465}
1466
1467static void flush_cpu_slab(void *d)
1468{
1469 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001470
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001471 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001472}
1473
1474static void flush_all(struct kmem_cache *s)
1475{
1476#ifdef CONFIG_SMP
1477 on_each_cpu(flush_cpu_slab, s, 1, 1);
1478#else
1479 unsigned long flags;
1480
1481 local_irq_save(flags);
1482 flush_cpu_slab(s);
1483 local_irq_restore(flags);
1484#endif
1485}
1486
1487/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001488 * Check if the objects in a per cpu structure fit numa
1489 * locality expectations.
1490 */
1491static inline int node_match(struct kmem_cache_cpu *c, int node)
1492{
1493#ifdef CONFIG_NUMA
1494 if (node != -1 && c->node != node)
1495 return 0;
1496#endif
1497 return 1;
1498}
1499
1500/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001501 * Slow path. The lockless freelist is empty or we need to perform
1502 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001503 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001504 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001505 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001506 * Processing is still very fast if new objects have been freed to the
1507 * regular freelist. In that case we simply take over the regular freelist
1508 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001509 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001510 * If that is not working then we fall back to the partial lists. We take the
1511 * first element of the freelist as the object to allocate now and move the
1512 * rest of the freelist to the lockless freelist.
1513 *
1514 * And if we were unable to get a new slab from the partial slab lists then
1515 * we need to allocate a new slab. This is slowest path since we may sleep.
Christoph Lameter81819f02007-05-06 14:49:36 -07001516 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001517static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001518 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001519{
Christoph Lameter81819f02007-05-06 14:49:36 -07001520 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001521 struct page *new;
Christoph Lameter1f842602008-01-07 23:20:30 -08001522#ifdef SLUB_FASTPATH
1523 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07001524
Christoph Lameter1f842602008-01-07 23:20:30 -08001525 local_irq_save(flags);
1526#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001527 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001528 goto new_slab;
1529
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001530 slab_lock(c->page);
1531 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001532 goto another_slab;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001533 stat(c, ALLOC_REFILL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001534load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001535 object = c->page->freelist;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001536 if (unlikely(object == c->page->end))
Christoph Lameter81819f02007-05-06 14:49:36 -07001537 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001538 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001539 goto debug;
1540
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001541 object = c->page->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001542 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001543 c->page->inuse = s->objects;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001544 c->page->freelist = c->page->end;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001545 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001546unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001547 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001548 stat(c, ALLOC_SLOWPATH);
Christoph Lameter1f842602008-01-07 23:20:30 -08001549out:
1550#ifdef SLUB_FASTPATH
1551 local_irq_restore(flags);
1552#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001553 return object;
1554
1555another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001556 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001557
1558new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001559 new = get_partial(s, gfpflags, node);
1560 if (new) {
1561 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001562 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001563 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001564 }
1565
Christoph Lameterb811c202007-10-16 23:25:51 -07001566 if (gfpflags & __GFP_WAIT)
1567 local_irq_enable();
1568
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001569 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001570
1571 if (gfpflags & __GFP_WAIT)
1572 local_irq_disable();
1573
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001574 if (new) {
1575 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001576 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001577 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001578 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001579 slab_lock(new);
1580 SetSlabFrozen(new);
1581 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001582 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001583 }
Christoph Lameter1f842602008-01-07 23:20:30 -08001584 object = NULL;
1585 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07001586debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001587 object = c->page->freelist;
1588 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001589 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001590
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001591 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001592 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001593 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001594 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001595}
1596
1597/*
1598 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1599 * have the fastpath folded into their functions. So no function call
1600 * overhead for requests that can be satisfied on the fastpath.
1601 *
1602 * The fastpath works by first checking if the lockless freelist can be used.
1603 * If not then __slab_alloc is called for slow processing.
1604 *
1605 * Otherwise we can simply pick the next object from the lockless free list.
1606 */
Pekka Enberg06428782008-01-07 23:20:27 -08001607static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001608 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001609{
Christoph Lameter894b8782007-05-10 03:15:16 -07001610 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001611 struct kmem_cache_cpu *c;
Christoph Lameter894b8782007-05-10 03:15:16 -07001612
Christoph Lameter1f842602008-01-07 23:20:30 -08001613/*
1614 * The SLUB_FASTPATH path is provisional and is currently disabled if the
1615 * kernel is compiled with preemption or if the arch does not support
1616 * fast cmpxchg operations. There are a couple of coming changes that will
1617 * simplify matters and allow preemption. Ultimately we may end up making
1618 * SLUB_FASTPATH the default.
1619 *
1620 * 1. The introduction of the per cpu allocator will avoid array lookups
1621 * through get_cpu_slab(). A special register can be used instead.
1622 *
1623 * 2. The introduction of per cpu atomic operations (cpu_ops) means that
1624 * we can realize the logic here entirely with per cpu atomics. The
1625 * per cpu atomic ops will take care of the preemption issues.
1626 */
1627
1628#ifdef SLUB_FASTPATH
1629 c = get_cpu_slab(s, raw_smp_processor_id());
1630 do {
1631 object = c->freelist;
1632 if (unlikely(is_end(object) || !node_match(c, node))) {
1633 object = __slab_alloc(s, gfpflags, node, addr, c);
1634 break;
1635 }
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001636 stat(c, ALLOC_FASTPATH);
Christoph Lameter1f842602008-01-07 23:20:30 -08001637 } while (cmpxchg_local(&c->freelist, object, object[c->offset])
1638 != object);
1639#else
1640 unsigned long flags;
1641
Christoph Lameter894b8782007-05-10 03:15:16 -07001642 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001643 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001644 if (unlikely(is_end(c->freelist) || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001645
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001646 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001647
1648 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001649 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001650 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001651 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001652 }
1653 local_irq_restore(flags);
Christoph Lameter1f842602008-01-07 23:20:30 -08001654#endif
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001655
1656 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001657 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001658
Christoph Lameter894b8782007-05-10 03:15:16 -07001659 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001660}
1661
1662void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1663{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001664 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001665}
1666EXPORT_SYMBOL(kmem_cache_alloc);
1667
1668#ifdef CONFIG_NUMA
1669void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1670{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001671 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001672}
1673EXPORT_SYMBOL(kmem_cache_alloc_node);
1674#endif
1675
1676/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001677 * Slow patch handling. This may still be called frequently since objects
1678 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001679 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001680 * So we still attempt to reduce cache line usage. Just take the slab
1681 * lock and free the item. If there is no additional partial page
1682 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001683 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001684static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001685 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001686{
1687 void *prior;
1688 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001689 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001690
Christoph Lameter1f842602008-01-07 23:20:30 -08001691#ifdef SLUB_FASTPATH
1692 unsigned long flags;
1693
1694 local_irq_save(flags);
1695#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001696 c = get_cpu_slab(s, raw_smp_processor_id());
1697 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001698 slab_lock(page);
1699
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001700 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001701 goto debug;
1702checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001703 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001704 page->freelist = object;
1705 page->inuse--;
1706
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001707 if (unlikely(SlabFrozen(page))) {
1708 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001709 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001710 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001711
1712 if (unlikely(!page->inuse))
1713 goto slab_empty;
1714
1715 /*
1716 * Objects left in the slab. If it
1717 * was not on the partial list before
1718 * then add it.
1719 */
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001720 if (unlikely(prior == page->end)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001721 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001722 stat(c, FREE_ADD_PARTIAL);
1723 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001724
1725out_unlock:
1726 slab_unlock(page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001727#ifdef SLUB_FASTPATH
1728 local_irq_restore(flags);
1729#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001730 return;
1731
1732slab_empty:
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001733 if (prior != page->end) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001734 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001735 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001736 */
1737 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001738 stat(c, FREE_REMOVE_PARTIAL);
1739 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001740 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001741 stat(c, FREE_SLAB);
Christoph Lameter1f842602008-01-07 23:20:30 -08001742#ifdef SLUB_FASTPATH
1743 local_irq_restore(flags);
1744#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001745 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001746 return;
1747
1748debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001749 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001750 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001751 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001752}
1753
Christoph Lameter894b8782007-05-10 03:15:16 -07001754/*
1755 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1756 * can perform fastpath freeing without additional function calls.
1757 *
1758 * The fastpath is only possible if we are freeing to the current cpu slab
1759 * of this processor. This typically the case if we have just allocated
1760 * the item before.
1761 *
1762 * If fastpath is not possible then fall back to __slab_free where we deal
1763 * with all sorts of special processing.
1764 */
Pekka Enberg06428782008-01-07 23:20:27 -08001765static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001766 struct page *page, void *x, void *addr)
1767{
1768 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001769 struct kmem_cache_cpu *c;
Christoph Lameter894b8782007-05-10 03:15:16 -07001770
Christoph Lameter1f842602008-01-07 23:20:30 -08001771#ifdef SLUB_FASTPATH
1772 void **freelist;
1773
1774 c = get_cpu_slab(s, raw_smp_processor_id());
1775 debug_check_no_locks_freed(object, s->objsize);
1776 do {
1777 freelist = c->freelist;
1778 barrier();
1779 /*
1780 * If the compiler would reorder the retrieval of c->page to
1781 * come before c->freelist then an interrupt could
1782 * change the cpu slab before we retrieve c->freelist. We
1783 * could be matching on a page no longer active and put the
1784 * object onto the freelist of the wrong slab.
1785 *
1786 * On the other hand: If we already have the freelist pointer
1787 * then any change of cpu_slab will cause the cmpxchg to fail
1788 * since the freelist pointers are unique per slab.
1789 */
1790 if (unlikely(page != c->page || c->node < 0)) {
1791 __slab_free(s, page, x, addr, c->offset);
1792 break;
1793 }
1794 object[c->offset] = freelist;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001795 stat(c, FREE_FASTPATH);
Christoph Lameter1f842602008-01-07 23:20:30 -08001796 } while (cmpxchg_local(&c->freelist, freelist, object) != freelist);
1797#else
1798 unsigned long flags;
1799
Christoph Lameter894b8782007-05-10 03:15:16 -07001800 local_irq_save(flags);
Peter Zijlstra02febdf2007-07-26 20:01:38 +02001801 debug_check_no_locks_freed(object, s->objsize);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001802 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001803 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001804 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001805 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001806 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001807 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001808 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001809
1810 local_irq_restore(flags);
Christoph Lameter1f842602008-01-07 23:20:30 -08001811#endif
Christoph Lameter894b8782007-05-10 03:15:16 -07001812}
1813
Christoph Lameter81819f02007-05-06 14:49:36 -07001814void kmem_cache_free(struct kmem_cache *s, void *x)
1815{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001816 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001817
Christoph Lameterb49af682007-05-06 14:49:41 -07001818 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001819
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001820 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001821}
1822EXPORT_SYMBOL(kmem_cache_free);
1823
1824/* Figure out on which slab object the object resides */
1825static struct page *get_object_page(const void *x)
1826{
Christoph Lameterb49af682007-05-06 14:49:41 -07001827 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001828
1829 if (!PageSlab(page))
1830 return NULL;
1831
1832 return page;
1833}
1834
1835/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001836 * Object placement in a slab is made very easy because we always start at
1837 * offset 0. If we tune the size of the object to the alignment then we can
1838 * get the required alignment by putting one properly sized object after
1839 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001840 *
1841 * Notice that the allocation order determines the sizes of the per cpu
1842 * caches. Each processor has always one slab available for allocations.
1843 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001844 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001845 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001846 */
1847
1848/*
1849 * Mininum / Maximum order of slab pages. This influences locking overhead
1850 * and slab fragmentation. A higher order reduces the number of partial slabs
1851 * and increases the number of allocations possible without having to
1852 * take the list_lock.
1853 */
1854static int slub_min_order;
1855static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001856static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1857
1858/*
1859 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001860 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001861 */
1862static int slub_nomerge;
1863
1864/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001865 * Calculate the order of allocation given an slab object size.
1866 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001867 * The order of allocation has significant impact on performance and other
1868 * system components. Generally order 0 allocations should be preferred since
1869 * order 0 does not cause fragmentation in the page allocator. Larger objects
1870 * be problematic to put into order 0 slabs because there may be too much
1871 * unused space left. We go to a higher order if more than 1/8th of the slab
1872 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001873 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001874 * In order to reach satisfactory performance we must ensure that a minimum
1875 * number of objects is in one slab. Otherwise we may generate too much
1876 * activity on the partial lists which requires taking the list_lock. This is
1877 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001878 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001879 * slub_max_order specifies the order where we begin to stop considering the
1880 * number of objects in a slab as critical. If we reach slub_max_order then
1881 * we try to keep the page order as low as possible. So we accept more waste
1882 * of space in favor of a small page order.
1883 *
1884 * Higher order allocations also allow the placement of more objects in a
1885 * slab and thereby reduce object handling overhead. If the user has
1886 * requested a higher mininum order then we start with that one instead of
1887 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001888 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001889static inline int slab_order(int size, int min_objects,
1890 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001891{
1892 int order;
1893 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001894 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001895
Christoph Lameter6300ea72007-07-17 04:03:20 -07001896 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001897 fls(min_objects * size - 1) - PAGE_SHIFT);
1898 order <= max_order; order++) {
1899
Christoph Lameter81819f02007-05-06 14:49:36 -07001900 unsigned long slab_size = PAGE_SIZE << order;
1901
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001902 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001903 continue;
1904
Christoph Lameter81819f02007-05-06 14:49:36 -07001905 rem = slab_size % size;
1906
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001907 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001908 break;
1909
1910 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001911
Christoph Lameter81819f02007-05-06 14:49:36 -07001912 return order;
1913}
1914
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001915static inline int calculate_order(int size)
1916{
1917 int order;
1918 int min_objects;
1919 int fraction;
1920
1921 /*
1922 * Attempt to find best configuration for a slab. This
1923 * works by first attempting to generate a layout with
1924 * the best configuration and backing off gradually.
1925 *
1926 * First we reduce the acceptable waste in a slab. Then
1927 * we reduce the minimum objects required in a slab.
1928 */
1929 min_objects = slub_min_objects;
1930 while (min_objects > 1) {
1931 fraction = 8;
1932 while (fraction >= 4) {
1933 order = slab_order(size, min_objects,
1934 slub_max_order, fraction);
1935 if (order <= slub_max_order)
1936 return order;
1937 fraction /= 2;
1938 }
1939 min_objects /= 2;
1940 }
1941
1942 /*
1943 * We were unable to place multiple objects in a slab. Now
1944 * lets see if we can place a single object there.
1945 */
1946 order = slab_order(size, 1, slub_max_order, 1);
1947 if (order <= slub_max_order)
1948 return order;
1949
1950 /*
1951 * Doh this slab cannot be placed using slub_max_order.
1952 */
1953 order = slab_order(size, 1, MAX_ORDER, 1);
1954 if (order <= MAX_ORDER)
1955 return order;
1956 return -ENOSYS;
1957}
1958
Christoph Lameter81819f02007-05-06 14:49:36 -07001959/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001960 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001961 */
1962static unsigned long calculate_alignment(unsigned long flags,
1963 unsigned long align, unsigned long size)
1964{
1965 /*
1966 * If the user wants hardware cache aligned objects then
1967 * follow that suggestion if the object is sufficiently
1968 * large.
1969 *
1970 * The hardware cache alignment cannot override the
1971 * specified alignment though. If that is greater
1972 * then use it.
1973 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001974 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001975 size > cache_line_size() / 2)
1976 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001977
1978 if (align < ARCH_SLAB_MINALIGN)
1979 return ARCH_SLAB_MINALIGN;
1980
1981 return ALIGN(align, sizeof(void *));
1982}
1983
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001984static void init_kmem_cache_cpu(struct kmem_cache *s,
1985 struct kmem_cache_cpu *c)
1986{
1987 c->page = NULL;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08001988 c->freelist = (void *)PAGE_MAPPING_ANON;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001989 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001990 c->offset = s->offset / sizeof(void *);
1991 c->objsize = s->objsize;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001992}
1993
Christoph Lameter81819f02007-05-06 14:49:36 -07001994static void init_kmem_cache_node(struct kmem_cache_node *n)
1995{
1996 n->nr_partial = 0;
1997 atomic_long_set(&n->nr_slabs, 0);
1998 spin_lock_init(&n->list_lock);
1999 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002000#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07002001 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002002#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002003}
2004
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002005#ifdef CONFIG_SMP
2006/*
2007 * Per cpu array for per cpu structures.
2008 *
2009 * The per cpu array places all kmem_cache_cpu structures from one processor
2010 * close together meaning that it becomes possible that multiple per cpu
2011 * structures are contained in one cacheline. This may be particularly
2012 * beneficial for the kmalloc caches.
2013 *
2014 * A desktop system typically has around 60-80 slabs. With 100 here we are
2015 * likely able to get per cpu structures for all caches from the array defined
2016 * here. We must be able to cover all kmalloc caches during bootstrap.
2017 *
2018 * If the per cpu array is exhausted then fall back to kmalloc
2019 * of individual cachelines. No sharing is possible then.
2020 */
2021#define NR_KMEM_CACHE_CPU 100
2022
2023static DEFINE_PER_CPU(struct kmem_cache_cpu,
2024 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2025
2026static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2027static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
2028
2029static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2030 int cpu, gfp_t flags)
2031{
2032 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2033
2034 if (c)
2035 per_cpu(kmem_cache_cpu_free, cpu) =
2036 (void *)c->freelist;
2037 else {
2038 /* Table overflow: So allocate ourselves */
2039 c = kmalloc_node(
2040 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2041 flags, cpu_to_node(cpu));
2042 if (!c)
2043 return NULL;
2044 }
2045
2046 init_kmem_cache_cpu(s, c);
2047 return c;
2048}
2049
2050static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2051{
2052 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2053 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2054 kfree(c);
2055 return;
2056 }
2057 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2058 per_cpu(kmem_cache_cpu_free, cpu) = c;
2059}
2060
2061static void free_kmem_cache_cpus(struct kmem_cache *s)
2062{
2063 int cpu;
2064
2065 for_each_online_cpu(cpu) {
2066 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2067
2068 if (c) {
2069 s->cpu_slab[cpu] = NULL;
2070 free_kmem_cache_cpu(c, cpu);
2071 }
2072 }
2073}
2074
2075static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2076{
2077 int cpu;
2078
2079 for_each_online_cpu(cpu) {
2080 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2081
2082 if (c)
2083 continue;
2084
2085 c = alloc_kmem_cache_cpu(s, cpu, flags);
2086 if (!c) {
2087 free_kmem_cache_cpus(s);
2088 return 0;
2089 }
2090 s->cpu_slab[cpu] = c;
2091 }
2092 return 1;
2093}
2094
2095/*
2096 * Initialize the per cpu array.
2097 */
2098static void init_alloc_cpu_cpu(int cpu)
2099{
2100 int i;
2101
2102 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2103 return;
2104
2105 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2106 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2107
2108 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2109}
2110
2111static void __init init_alloc_cpu(void)
2112{
2113 int cpu;
2114
2115 for_each_online_cpu(cpu)
2116 init_alloc_cpu_cpu(cpu);
2117 }
2118
2119#else
2120static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2121static inline void init_alloc_cpu(void) {}
2122
2123static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2124{
2125 init_kmem_cache_cpu(s, &s->cpu_slab);
2126 return 1;
2127}
2128#endif
2129
Christoph Lameter81819f02007-05-06 14:49:36 -07002130#ifdef CONFIG_NUMA
2131/*
2132 * No kmalloc_node yet so do it by hand. We know that this is the first
2133 * slab on the node for this slabcache. There are no concurrent accesses
2134 * possible.
2135 *
2136 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002137 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2138 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002139 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002140static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2141 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002142{
2143 struct page *page;
2144 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002145 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002146
2147 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2148
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002149 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002150
2151 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002152 if (page_to_nid(page) != node) {
2153 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2154 "node %d\n", node);
2155 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2156 "in order to be able to continue\n");
2157 }
2158
Christoph Lameter81819f02007-05-06 14:49:36 -07002159 n = page->freelist;
2160 BUG_ON(!n);
2161 page->freelist = get_freepointer(kmalloc_caches, n);
2162 page->inuse++;
2163 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002164#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002165 init_object(kmalloc_caches, n, 1);
2166 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002167#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002168 init_kmem_cache_node(n);
2169 atomic_long_inc(&n->nr_slabs);
rootba84c732008-01-07 23:20:28 -08002170 /*
2171 * lockdep requires consistent irq usage for each lock
2172 * so even though there cannot be a race this early in
2173 * the boot sequence, we still disable irqs.
2174 */
2175 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002176 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002177 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002178 return n;
2179}
2180
2181static void free_kmem_cache_nodes(struct kmem_cache *s)
2182{
2183 int node;
2184
Christoph Lameterf64dc582007-10-16 01:25:33 -07002185 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002186 struct kmem_cache_node *n = s->node[node];
2187 if (n && n != &s->local_node)
2188 kmem_cache_free(kmalloc_caches, n);
2189 s->node[node] = NULL;
2190 }
2191}
2192
2193static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2194{
2195 int node;
2196 int local_node;
2197
2198 if (slab_state >= UP)
2199 local_node = page_to_nid(virt_to_page(s));
2200 else
2201 local_node = 0;
2202
Christoph Lameterf64dc582007-10-16 01:25:33 -07002203 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002204 struct kmem_cache_node *n;
2205
2206 if (local_node == node)
2207 n = &s->local_node;
2208 else {
2209 if (slab_state == DOWN) {
2210 n = early_kmem_cache_node_alloc(gfpflags,
2211 node);
2212 continue;
2213 }
2214 n = kmem_cache_alloc_node(kmalloc_caches,
2215 gfpflags, node);
2216
2217 if (!n) {
2218 free_kmem_cache_nodes(s);
2219 return 0;
2220 }
2221
2222 }
2223 s->node[node] = n;
2224 init_kmem_cache_node(n);
2225 }
2226 return 1;
2227}
2228#else
2229static void free_kmem_cache_nodes(struct kmem_cache *s)
2230{
2231}
2232
2233static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2234{
2235 init_kmem_cache_node(&s->local_node);
2236 return 1;
2237}
2238#endif
2239
2240/*
2241 * calculate_sizes() determines the order and the distribution of data within
2242 * a slab object.
2243 */
2244static int calculate_sizes(struct kmem_cache *s)
2245{
2246 unsigned long flags = s->flags;
2247 unsigned long size = s->objsize;
2248 unsigned long align = s->align;
2249
2250 /*
2251 * Determine if we can poison the object itself. If the user of
2252 * the slab may touch the object after free or before allocation
2253 * then we should never poison the object itself.
2254 */
2255 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002256 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002257 s->flags |= __OBJECT_POISON;
2258 else
2259 s->flags &= ~__OBJECT_POISON;
2260
2261 /*
2262 * Round up object size to the next word boundary. We can only
2263 * place the free pointer at word boundaries and this determines
2264 * the possible location of the free pointer.
2265 */
2266 size = ALIGN(size, sizeof(void *));
2267
Christoph Lameter41ecc552007-05-09 02:32:44 -07002268#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002269 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002270 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002271 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002272 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002273 */
2274 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2275 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002276#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002277
2278 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002279 * With that we have determined the number of bytes in actual use
2280 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002281 */
2282 s->inuse = size;
2283
2284 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002285 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002286 /*
2287 * Relocate free pointer after the object if it is not
2288 * permitted to overwrite the first word of the object on
2289 * kmem_cache_free.
2290 *
2291 * This is the case if we do RCU, have a constructor or
2292 * destructor or are poisoning the objects.
2293 */
2294 s->offset = size;
2295 size += sizeof(void *);
2296 }
2297
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002298#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002299 if (flags & SLAB_STORE_USER)
2300 /*
2301 * Need to store information about allocs and frees after
2302 * the object.
2303 */
2304 size += 2 * sizeof(struct track);
2305
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002306 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002307 /*
2308 * Add some empty padding so that we can catch
2309 * overwrites from earlier objects rather than let
2310 * tracking information or the free pointer be
2311 * corrupted if an user writes before the start
2312 * of the object.
2313 */
2314 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002315#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002316
Christoph Lameter81819f02007-05-06 14:49:36 -07002317 /*
2318 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002319 * user specified and the dynamic determination of cache line size
2320 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002321 */
2322 align = calculate_alignment(flags, align, s->objsize);
2323
2324 /*
2325 * SLUB stores one object immediately after another beginning from
2326 * offset 0. In order to align the objects we have to simply size
2327 * each object to conform to the alignment.
2328 */
2329 size = ALIGN(size, align);
2330 s->size = size;
2331
2332 s->order = calculate_order(size);
2333 if (s->order < 0)
2334 return 0;
2335
2336 /*
2337 * Determine the number of objects per slab
2338 */
2339 s->objects = (PAGE_SIZE << s->order) / size;
2340
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002341 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002342
2343}
2344
Christoph Lameter81819f02007-05-06 14:49:36 -07002345static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2346 const char *name, size_t size,
2347 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002348 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002349{
2350 memset(s, 0, kmem_size);
2351 s->name = name;
2352 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002353 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002354 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002355 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002356
2357 if (!calculate_sizes(s))
2358 goto error;
2359
2360 s->refcount = 1;
2361#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002362 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002363#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002364 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2365 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002366
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002367 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002368 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002369 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002370error:
2371 if (flags & SLAB_PANIC)
2372 panic("Cannot create slab %s size=%lu realsize=%u "
2373 "order=%u offset=%u flags=%lx\n",
2374 s->name, (unsigned long)size, s->size, s->order,
2375 s->offset, flags);
2376 return 0;
2377}
Christoph Lameter81819f02007-05-06 14:49:36 -07002378
2379/*
2380 * Check if a given pointer is valid
2381 */
2382int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2383{
Pekka Enberg06428782008-01-07 23:20:27 -08002384 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002385
2386 page = get_object_page(object);
2387
2388 if (!page || s != page->slab)
2389 /* No slab or wrong slab */
2390 return 0;
2391
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002392 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002393 return 0;
2394
2395 /*
2396 * We could also check if the object is on the slabs freelist.
2397 * But this would be too expensive and it seems that the main
2398 * purpose of kmem_ptr_valid is to check if the object belongs
2399 * to a certain slab.
2400 */
2401 return 1;
2402}
2403EXPORT_SYMBOL(kmem_ptr_validate);
2404
2405/*
2406 * Determine the size of a slab object
2407 */
2408unsigned int kmem_cache_size(struct kmem_cache *s)
2409{
2410 return s->objsize;
2411}
2412EXPORT_SYMBOL(kmem_cache_size);
2413
2414const char *kmem_cache_name(struct kmem_cache *s)
2415{
2416 return s->name;
2417}
2418EXPORT_SYMBOL(kmem_cache_name);
2419
2420/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002421 * Attempt to free all slabs on a node. Return the number of slabs we
2422 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002423 */
2424static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2425 struct list_head *list)
2426{
2427 int slabs_inuse = 0;
2428 unsigned long flags;
2429 struct page *page, *h;
2430
2431 spin_lock_irqsave(&n->list_lock, flags);
2432 list_for_each_entry_safe(page, h, list, lru)
2433 if (!page->inuse) {
2434 list_del(&page->lru);
2435 discard_slab(s, page);
2436 } else
2437 slabs_inuse++;
2438 spin_unlock_irqrestore(&n->list_lock, flags);
2439 return slabs_inuse;
2440}
2441
2442/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002443 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002444 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002445static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002446{
2447 int node;
2448
2449 flush_all(s);
2450
2451 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002452 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002453 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002454 struct kmem_cache_node *n = get_node(s, node);
2455
Christoph Lameter2086d262007-05-06 14:49:46 -07002456 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002457 if (atomic_long_read(&n->nr_slabs))
2458 return 1;
2459 }
2460 free_kmem_cache_nodes(s);
2461 return 0;
2462}
2463
2464/*
2465 * Close a cache and release the kmem_cache structure
2466 * (must be used for caches created using kmem_cache_create)
2467 */
2468void kmem_cache_destroy(struct kmem_cache *s)
2469{
2470 down_write(&slub_lock);
2471 s->refcount--;
2472 if (!s->refcount) {
2473 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002474 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002475 if (kmem_cache_close(s))
2476 WARN_ON(1);
2477 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002478 } else
2479 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002480}
2481EXPORT_SYMBOL(kmem_cache_destroy);
2482
2483/********************************************************************
2484 * Kmalloc subsystem
2485 *******************************************************************/
2486
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002487struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002488EXPORT_SYMBOL(kmalloc_caches);
2489
2490#ifdef CONFIG_ZONE_DMA
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002491static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT];
Christoph Lameter81819f02007-05-06 14:49:36 -07002492#endif
2493
2494static int __init setup_slub_min_order(char *str)
2495{
Pekka Enberg06428782008-01-07 23:20:27 -08002496 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002497
2498 return 1;
2499}
2500
2501__setup("slub_min_order=", setup_slub_min_order);
2502
2503static int __init setup_slub_max_order(char *str)
2504{
Pekka Enberg06428782008-01-07 23:20:27 -08002505 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002506
2507 return 1;
2508}
2509
2510__setup("slub_max_order=", setup_slub_max_order);
2511
2512static int __init setup_slub_min_objects(char *str)
2513{
Pekka Enberg06428782008-01-07 23:20:27 -08002514 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002515
2516 return 1;
2517}
2518
2519__setup("slub_min_objects=", setup_slub_min_objects);
2520
2521static int __init setup_slub_nomerge(char *str)
2522{
2523 slub_nomerge = 1;
2524 return 1;
2525}
2526
2527__setup("slub_nomerge", setup_slub_nomerge);
2528
Christoph Lameter81819f02007-05-06 14:49:36 -07002529static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2530 const char *name, int size, gfp_t gfp_flags)
2531{
2532 unsigned int flags = 0;
2533
2534 if (gfp_flags & SLUB_DMA)
2535 flags = SLAB_CACHE_DMA;
2536
2537 down_write(&slub_lock);
2538 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameterc59def92007-05-16 22:10:50 -07002539 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002540 goto panic;
2541
2542 list_add(&s->list, &slab_caches);
2543 up_write(&slub_lock);
2544 if (sysfs_slab_add(s))
2545 goto panic;
2546 return s;
2547
2548panic:
2549 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2550}
2551
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002552#ifdef CONFIG_ZONE_DMA
Christoph Lameter1ceef402007-08-07 15:11:48 -07002553
2554static void sysfs_add_func(struct work_struct *w)
2555{
2556 struct kmem_cache *s;
2557
2558 down_write(&slub_lock);
2559 list_for_each_entry(s, &slab_caches, list) {
2560 if (s->flags & __SYSFS_ADD_DEFERRED) {
2561 s->flags &= ~__SYSFS_ADD_DEFERRED;
2562 sysfs_slab_add(s);
2563 }
2564 }
2565 up_write(&slub_lock);
2566}
2567
2568static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2569
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002570static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2571{
2572 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002573 char *text;
2574 size_t realsize;
2575
2576 s = kmalloc_caches_dma[index];
2577 if (s)
2578 return s;
2579
2580 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002581 if (flags & __GFP_WAIT)
2582 down_write(&slub_lock);
2583 else {
2584 if (!down_write_trylock(&slub_lock))
2585 goto out;
2586 }
2587
2588 if (kmalloc_caches_dma[index])
2589 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002590
Christoph Lameter7b55f622007-07-17 04:03:27 -07002591 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002592 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2593 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002594 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2595
2596 if (!s || !text || !kmem_cache_open(s, flags, text,
2597 realsize, ARCH_KMALLOC_MINALIGN,
2598 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2599 kfree(s);
2600 kfree(text);
2601 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002602 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002603
2604 list_add(&s->list, &slab_caches);
2605 kmalloc_caches_dma[index] = s;
2606
2607 schedule_work(&sysfs_add_work);
2608
2609unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002610 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002611out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002612 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002613}
2614#endif
2615
Christoph Lameterf1b26332007-07-17 04:03:26 -07002616/*
2617 * Conversion table for small slabs sizes / 8 to the index in the
2618 * kmalloc array. This is necessary for slabs < 192 since we have non power
2619 * of two cache sizes there. The size of larger slabs can be determined using
2620 * fls.
2621 */
2622static s8 size_index[24] = {
2623 3, /* 8 */
2624 4, /* 16 */
2625 5, /* 24 */
2626 5, /* 32 */
2627 6, /* 40 */
2628 6, /* 48 */
2629 6, /* 56 */
2630 6, /* 64 */
2631 1, /* 72 */
2632 1, /* 80 */
2633 1, /* 88 */
2634 1, /* 96 */
2635 7, /* 104 */
2636 7, /* 112 */
2637 7, /* 120 */
2638 7, /* 128 */
2639 2, /* 136 */
2640 2, /* 144 */
2641 2, /* 152 */
2642 2, /* 160 */
2643 2, /* 168 */
2644 2, /* 176 */
2645 2, /* 184 */
2646 2 /* 192 */
2647};
2648
Christoph Lameter81819f02007-05-06 14:49:36 -07002649static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2650{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002651 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002652
Christoph Lameterf1b26332007-07-17 04:03:26 -07002653 if (size <= 192) {
2654 if (!size)
2655 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002656
Christoph Lameterf1b26332007-07-17 04:03:26 -07002657 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002658 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002659 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002660
2661#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002662 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002663 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002664
Christoph Lameter81819f02007-05-06 14:49:36 -07002665#endif
2666 return &kmalloc_caches[index];
2667}
2668
2669void *__kmalloc(size_t size, gfp_t flags)
2670{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002671 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002672
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002673 if (unlikely(size > PAGE_SIZE / 2))
2674 return (void *)__get_free_pages(flags | __GFP_COMP,
2675 get_order(size));
2676
2677 s = get_slab(size, flags);
2678
2679 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002680 return s;
2681
Christoph Lameterce15fea2007-07-17 04:03:28 -07002682 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002683}
2684EXPORT_SYMBOL(__kmalloc);
2685
2686#ifdef CONFIG_NUMA
2687void *__kmalloc_node(size_t size, gfp_t flags, int node)
2688{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002689 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002690
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002691 if (unlikely(size > PAGE_SIZE / 2))
2692 return (void *)__get_free_pages(flags | __GFP_COMP,
2693 get_order(size));
2694
2695 s = get_slab(size, flags);
2696
2697 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002698 return s;
2699
Christoph Lameterce15fea2007-07-17 04:03:28 -07002700 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002701}
2702EXPORT_SYMBOL(__kmalloc_node);
2703#endif
2704
2705size_t ksize(const void *object)
2706{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002707 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002708 struct kmem_cache *s;
2709
Christoph Lameteref8b4522007-10-16 01:24:46 -07002710 BUG_ON(!object);
2711 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002712 return 0;
2713
Vegard Nossum294a80a2007-12-04 23:45:30 -08002714 page = virt_to_head_page(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07002715 BUG_ON(!page);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002716
2717 if (unlikely(!PageSlab(page)))
2718 return PAGE_SIZE << compound_order(page);
2719
Christoph Lameter81819f02007-05-06 14:49:36 -07002720 s = page->slab;
2721 BUG_ON(!s);
2722
2723 /*
2724 * Debugging requires use of the padding between object
2725 * and whatever may come after it.
2726 */
2727 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2728 return s->objsize;
2729
2730 /*
2731 * If we have the need to store the freelist pointer
2732 * back there or track user information then we can
2733 * only use the space before that information.
2734 */
2735 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2736 return s->inuse;
2737
2738 /*
2739 * Else we can use all the padding etc for the allocation
2740 */
2741 return s->size;
2742}
2743EXPORT_SYMBOL(ksize);
2744
2745void kfree(const void *x)
2746{
Christoph Lameter81819f02007-05-06 14:49:36 -07002747 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002748 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002749
Satyam Sharma2408c552007-10-16 01:24:44 -07002750 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002751 return;
2752
Christoph Lameterb49af682007-05-06 14:49:41 -07002753 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002754 if (unlikely(!PageSlab(page))) {
2755 put_page(page);
2756 return;
2757 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002758 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002759}
2760EXPORT_SYMBOL(kfree);
2761
Christoph Lameterf61396a2008-01-07 23:20:26 -08002762static unsigned long count_partial(struct kmem_cache_node *n)
2763{
2764 unsigned long flags;
2765 unsigned long x = 0;
2766 struct page *page;
2767
2768 spin_lock_irqsave(&n->list_lock, flags);
2769 list_for_each_entry(page, &n->partial, lru)
2770 x += page->inuse;
2771 spin_unlock_irqrestore(&n->list_lock, flags);
2772 return x;
2773}
2774
Christoph Lameter2086d262007-05-06 14:49:46 -07002775/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002776 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2777 * the remaining slabs by the number of items in use. The slabs with the
2778 * most items in use come first. New allocations will then fill those up
2779 * and thus they can be removed from the partial lists.
2780 *
2781 * The slabs with the least items are placed last. This results in them
2782 * being allocated from last increasing the chance that the last objects
2783 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002784 */
2785int kmem_cache_shrink(struct kmem_cache *s)
2786{
2787 int node;
2788 int i;
2789 struct kmem_cache_node *n;
2790 struct page *page;
2791 struct page *t;
2792 struct list_head *slabs_by_inuse =
2793 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2794 unsigned long flags;
2795
2796 if (!slabs_by_inuse)
2797 return -ENOMEM;
2798
2799 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002800 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002801 n = get_node(s, node);
2802
2803 if (!n->nr_partial)
2804 continue;
2805
2806 for (i = 0; i < s->objects; i++)
2807 INIT_LIST_HEAD(slabs_by_inuse + i);
2808
2809 spin_lock_irqsave(&n->list_lock, flags);
2810
2811 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002812 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002813 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002814 * Note that concurrent frees may occur while we hold the
2815 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002816 */
2817 list_for_each_entry_safe(page, t, &n->partial, lru) {
2818 if (!page->inuse && slab_trylock(page)) {
2819 /*
2820 * Must hold slab lock here because slab_free
2821 * may have freed the last object and be
2822 * waiting to release the slab.
2823 */
2824 list_del(&page->lru);
2825 n->nr_partial--;
2826 slab_unlock(page);
2827 discard_slab(s, page);
2828 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002829 list_move(&page->lru,
2830 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002831 }
2832 }
2833
Christoph Lameter2086d262007-05-06 14:49:46 -07002834 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002835 * Rebuild the partial list with the slabs filled up most
2836 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002837 */
2838 for (i = s->objects - 1; i >= 0; i--)
2839 list_splice(slabs_by_inuse + i, n->partial.prev);
2840
Christoph Lameter2086d262007-05-06 14:49:46 -07002841 spin_unlock_irqrestore(&n->list_lock, flags);
2842 }
2843
2844 kfree(slabs_by_inuse);
2845 return 0;
2846}
2847EXPORT_SYMBOL(kmem_cache_shrink);
2848
Yasunori Gotob9049e22007-10-21 16:41:37 -07002849#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2850static int slab_mem_going_offline_callback(void *arg)
2851{
2852 struct kmem_cache *s;
2853
2854 down_read(&slub_lock);
2855 list_for_each_entry(s, &slab_caches, list)
2856 kmem_cache_shrink(s);
2857 up_read(&slub_lock);
2858
2859 return 0;
2860}
2861
2862static void slab_mem_offline_callback(void *arg)
2863{
2864 struct kmem_cache_node *n;
2865 struct kmem_cache *s;
2866 struct memory_notify *marg = arg;
2867 int offline_node;
2868
2869 offline_node = marg->status_change_nid;
2870
2871 /*
2872 * If the node still has available memory. we need kmem_cache_node
2873 * for it yet.
2874 */
2875 if (offline_node < 0)
2876 return;
2877
2878 down_read(&slub_lock);
2879 list_for_each_entry(s, &slab_caches, list) {
2880 n = get_node(s, offline_node);
2881 if (n) {
2882 /*
2883 * if n->nr_slabs > 0, slabs still exist on the node
2884 * that is going down. We were unable to free them,
2885 * and offline_pages() function shoudn't call this
2886 * callback. So, we must fail.
2887 */
Al Viro27bb6282007-10-29 04:42:55 +00002888 BUG_ON(atomic_long_read(&n->nr_slabs));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002889
2890 s->node[offline_node] = NULL;
2891 kmem_cache_free(kmalloc_caches, n);
2892 }
2893 }
2894 up_read(&slub_lock);
2895}
2896
2897static int slab_mem_going_online_callback(void *arg)
2898{
2899 struct kmem_cache_node *n;
2900 struct kmem_cache *s;
2901 struct memory_notify *marg = arg;
2902 int nid = marg->status_change_nid;
2903 int ret = 0;
2904
2905 /*
2906 * If the node's memory is already available, then kmem_cache_node is
2907 * already created. Nothing to do.
2908 */
2909 if (nid < 0)
2910 return 0;
2911
2912 /*
2913 * We are bringing a node online. No memory is availabe yet. We must
2914 * allocate a kmem_cache_node structure in order to bring the node
2915 * online.
2916 */
2917 down_read(&slub_lock);
2918 list_for_each_entry(s, &slab_caches, list) {
2919 /*
2920 * XXX: kmem_cache_alloc_node will fallback to other nodes
2921 * since memory is not yet available from the node that
2922 * is brought up.
2923 */
2924 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2925 if (!n) {
2926 ret = -ENOMEM;
2927 goto out;
2928 }
2929 init_kmem_cache_node(n);
2930 s->node[nid] = n;
2931 }
2932out:
2933 up_read(&slub_lock);
2934 return ret;
2935}
2936
2937static int slab_memory_callback(struct notifier_block *self,
2938 unsigned long action, void *arg)
2939{
2940 int ret = 0;
2941
2942 switch (action) {
2943 case MEM_GOING_ONLINE:
2944 ret = slab_mem_going_online_callback(arg);
2945 break;
2946 case MEM_GOING_OFFLINE:
2947 ret = slab_mem_going_offline_callback(arg);
2948 break;
2949 case MEM_OFFLINE:
2950 case MEM_CANCEL_ONLINE:
2951 slab_mem_offline_callback(arg);
2952 break;
2953 case MEM_ONLINE:
2954 case MEM_CANCEL_OFFLINE:
2955 break;
2956 }
2957
2958 ret = notifier_from_errno(ret);
2959 return ret;
2960}
2961
2962#endif /* CONFIG_MEMORY_HOTPLUG */
2963
Christoph Lameter81819f02007-05-06 14:49:36 -07002964/********************************************************************
2965 * Basic setup of slabs
2966 *******************************************************************/
2967
2968void __init kmem_cache_init(void)
2969{
2970 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002971 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002972
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002973 init_alloc_cpu();
2974
Christoph Lameter81819f02007-05-06 14:49:36 -07002975#ifdef CONFIG_NUMA
2976 /*
2977 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002978 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002979 * kmem_cache_open for slab_state == DOWN.
2980 */
2981 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2982 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002983 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002984 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002985
2986 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002987#endif
2988
2989 /* Able to allocate the per node structures */
2990 slab_state = PARTIAL;
2991
2992 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002993 if (KMALLOC_MIN_SIZE <= 64) {
2994 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002995 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002996 caches++;
2997 }
2998 if (KMALLOC_MIN_SIZE <= 128) {
2999 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07003000 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07003001 caches++;
3002 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003003
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003004 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003005 create_kmalloc_cache(&kmalloc_caches[i],
3006 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07003007 caches++;
3008 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003009
Christoph Lameterf1b26332007-07-17 04:03:26 -07003010
3011 /*
3012 * Patch up the size_index table if we have strange large alignment
3013 * requirements for the kmalloc array. This is only the case for
3014 * mips it seems. The standard arches will not generate any code here.
3015 *
3016 * Largest permitted alignment is 256 bytes due to the way we
3017 * handle the index determination for the smaller caches.
3018 *
3019 * Make sure that nothing crazy happens if someone starts tinkering
3020 * around with ARCH_KMALLOC_MINALIGN
3021 */
3022 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3023 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3024
Christoph Lameter12ad6842007-07-17 04:03:28 -07003025 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07003026 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3027
Christoph Lameter81819f02007-05-06 14:49:36 -07003028 slab_state = UP;
3029
3030 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003031 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07003032 kmalloc_caches[i]. name =
3033 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3034
3035#ifdef CONFIG_SMP
3036 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003037 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3038 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3039#else
3040 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003041#endif
3042
Christoph Lameter81819f02007-05-06 14:49:36 -07003043
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003044 printk(KERN_INFO
3045 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003046 " CPUs=%d, Nodes=%d\n",
3047 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003048 slub_min_order, slub_max_order, slub_min_objects,
3049 nr_cpu_ids, nr_node_ids);
3050}
3051
3052/*
3053 * Find a mergeable slab cache
3054 */
3055static int slab_unmergeable(struct kmem_cache *s)
3056{
3057 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3058 return 1;
3059
Christoph Lameterc59def92007-05-16 22:10:50 -07003060 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003061 return 1;
3062
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003063 /*
3064 * We may have set a slab to be unmergeable during bootstrap.
3065 */
3066 if (s->refcount < 0)
3067 return 1;
3068
Christoph Lameter81819f02007-05-06 14:49:36 -07003069 return 0;
3070}
3071
3072static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003073 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003074 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003075{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003076 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003077
3078 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3079 return NULL;
3080
Christoph Lameterc59def92007-05-16 22:10:50 -07003081 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003082 return NULL;
3083
3084 size = ALIGN(size, sizeof(void *));
3085 align = calculate_alignment(flags, align, size);
3086 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003087 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003088
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003089 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003090 if (slab_unmergeable(s))
3091 continue;
3092
3093 if (size > s->size)
3094 continue;
3095
Christoph Lameterba0268a2007-09-11 15:24:11 -07003096 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003097 continue;
3098 /*
3099 * Check if alignment is compatible.
3100 * Courtesy of Adrian Drzewiecki
3101 */
Pekka Enberg06428782008-01-07 23:20:27 -08003102 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003103 continue;
3104
3105 if (s->size - size >= sizeof(void *))
3106 continue;
3107
3108 return s;
3109 }
3110 return NULL;
3111}
3112
3113struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3114 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003115 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003116{
3117 struct kmem_cache *s;
3118
3119 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003120 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003121 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003122 int cpu;
3123
Christoph Lameter81819f02007-05-06 14:49:36 -07003124 s->refcount++;
3125 /*
3126 * Adjust the object sizes so that we clear
3127 * the complete object on kzalloc.
3128 */
3129 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003130
3131 /*
3132 * And then we need to update the object size in the
3133 * per cpu structures
3134 */
3135 for_each_online_cpu(cpu)
3136 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter81819f02007-05-06 14:49:36 -07003137 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003138 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003139 if (sysfs_slab_alias(s, name))
3140 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003141 return s;
3142 }
3143 s = kmalloc(kmem_size, GFP_KERNEL);
3144 if (s) {
3145 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003146 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003147 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003148 up_write(&slub_lock);
3149 if (sysfs_slab_add(s))
3150 goto err;
3151 return s;
3152 }
3153 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003154 }
3155 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003156
3157err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003158 if (flags & SLAB_PANIC)
3159 panic("Cannot create slabcache %s\n", name);
3160 else
3161 s = NULL;
3162 return s;
3163}
3164EXPORT_SYMBOL(kmem_cache_create);
3165
Christoph Lameter81819f02007-05-06 14:49:36 -07003166#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003167/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003168 * Use the cpu notifier to insure that the cpu slabs are flushed when
3169 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003170 */
3171static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3172 unsigned long action, void *hcpu)
3173{
3174 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003175 struct kmem_cache *s;
3176 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003177
3178 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003179 case CPU_UP_PREPARE:
3180 case CPU_UP_PREPARE_FROZEN:
3181 init_alloc_cpu_cpu(cpu);
3182 down_read(&slub_lock);
3183 list_for_each_entry(s, &slab_caches, list)
3184 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3185 GFP_KERNEL);
3186 up_read(&slub_lock);
3187 break;
3188
Christoph Lameter81819f02007-05-06 14:49:36 -07003189 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003190 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003191 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003192 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003193 down_read(&slub_lock);
3194 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003195 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3196
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003197 local_irq_save(flags);
3198 __flush_cpu_slab(s, cpu);
3199 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003200 free_kmem_cache_cpu(c, cpu);
3201 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003202 }
3203 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003204 break;
3205 default:
3206 break;
3207 }
3208 return NOTIFY_OK;
3209}
3210
Pekka Enberg06428782008-01-07 23:20:27 -08003211static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003212 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003213};
Christoph Lameter81819f02007-05-06 14:49:36 -07003214
3215#endif
3216
Christoph Lameter81819f02007-05-06 14:49:36 -07003217void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3218{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003219 struct kmem_cache *s;
3220
3221 if (unlikely(size > PAGE_SIZE / 2))
3222 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3223 get_order(size));
3224 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003225
Satyam Sharma2408c552007-10-16 01:24:44 -07003226 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003227 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003228
Christoph Lameterce15fea2007-07-17 04:03:28 -07003229 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003230}
3231
3232void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3233 int node, void *caller)
3234{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003235 struct kmem_cache *s;
3236
3237 if (unlikely(size > PAGE_SIZE / 2))
3238 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3239 get_order(size));
3240 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003241
Satyam Sharma2408c552007-10-16 01:24:44 -07003242 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003243 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003244
Christoph Lameterce15fea2007-07-17 04:03:28 -07003245 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003246}
3247
Christoph Lameter41ecc552007-05-09 02:32:44 -07003248#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003249static int validate_slab(struct kmem_cache *s, struct page *page,
3250 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003251{
3252 void *p;
Christoph Lameter683d0ba2008-01-07 23:20:29 -08003253 void *addr = slab_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003254
3255 if (!check_slab(s, page) ||
3256 !on_freelist(s, page, NULL))
3257 return 0;
3258
3259 /* Now we know that a valid freelist exists */
3260 bitmap_zero(map, s->objects);
3261
Christoph Lameter7656c722007-05-09 02:32:40 -07003262 for_each_free_object(p, s, page->freelist) {
3263 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003264 if (!check_object(s, page, p, 0))
3265 return 0;
3266 }
3267
Christoph Lameter7656c722007-05-09 02:32:40 -07003268 for_each_object(p, s, addr)
3269 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003270 if (!check_object(s, page, p, 1))
3271 return 0;
3272 return 1;
3273}
3274
Christoph Lameter434e2452007-07-17 04:03:30 -07003275static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3276 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003277{
3278 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003279 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003280 slab_unlock(page);
3281 } else
3282 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3283 s->name, page);
3284
3285 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003286 if (!SlabDebug(page))
3287 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003288 "on slab 0x%p\n", s->name, page);
3289 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003290 if (SlabDebug(page))
3291 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003292 "slab 0x%p\n", s->name, page);
3293 }
3294}
3295
Christoph Lameter434e2452007-07-17 04:03:30 -07003296static int validate_slab_node(struct kmem_cache *s,
3297 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003298{
3299 unsigned long count = 0;
3300 struct page *page;
3301 unsigned long flags;
3302
3303 spin_lock_irqsave(&n->list_lock, flags);
3304
3305 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003306 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003307 count++;
3308 }
3309 if (count != n->nr_partial)
3310 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3311 "counter=%ld\n", s->name, count, n->nr_partial);
3312
3313 if (!(s->flags & SLAB_STORE_USER))
3314 goto out;
3315
3316 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003317 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003318 count++;
3319 }
3320 if (count != atomic_long_read(&n->nr_slabs))
3321 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3322 "counter=%ld\n", s->name, count,
3323 atomic_long_read(&n->nr_slabs));
3324
3325out:
3326 spin_unlock_irqrestore(&n->list_lock, flags);
3327 return count;
3328}
3329
Christoph Lameter434e2452007-07-17 04:03:30 -07003330static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003331{
3332 int node;
3333 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003334 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3335 sizeof(unsigned long), GFP_KERNEL);
3336
3337 if (!map)
3338 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003339
3340 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003341 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003342 struct kmem_cache_node *n = get_node(s, node);
3343
Christoph Lameter434e2452007-07-17 04:03:30 -07003344 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003345 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003346 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003347 return count;
3348}
3349
Christoph Lameterb3459702007-05-09 02:32:41 -07003350#ifdef SLUB_RESILIENCY_TEST
3351static void resiliency_test(void)
3352{
3353 u8 *p;
3354
3355 printk(KERN_ERR "SLUB resiliency testing\n");
3356 printk(KERN_ERR "-----------------------\n");
3357 printk(KERN_ERR "A. Corruption after allocation\n");
3358
3359 p = kzalloc(16, GFP_KERNEL);
3360 p[16] = 0x12;
3361 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3362 " 0x12->0x%p\n\n", p + 16);
3363
3364 validate_slab_cache(kmalloc_caches + 4);
3365
3366 /* Hmmm... The next two are dangerous */
3367 p = kzalloc(32, GFP_KERNEL);
3368 p[32 + sizeof(void *)] = 0x34;
3369 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003370 " 0x34 -> -0x%p\n", p);
3371 printk(KERN_ERR
3372 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003373
3374 validate_slab_cache(kmalloc_caches + 5);
3375 p = kzalloc(64, GFP_KERNEL);
3376 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3377 *p = 0x56;
3378 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3379 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003380 printk(KERN_ERR
3381 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003382 validate_slab_cache(kmalloc_caches + 6);
3383
3384 printk(KERN_ERR "\nB. Corruption after free\n");
3385 p = kzalloc(128, GFP_KERNEL);
3386 kfree(p);
3387 *p = 0x78;
3388 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3389 validate_slab_cache(kmalloc_caches + 7);
3390
3391 p = kzalloc(256, GFP_KERNEL);
3392 kfree(p);
3393 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003394 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3395 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003396 validate_slab_cache(kmalloc_caches + 8);
3397
3398 p = kzalloc(512, GFP_KERNEL);
3399 kfree(p);
3400 p[512] = 0xab;
3401 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3402 validate_slab_cache(kmalloc_caches + 9);
3403}
3404#else
3405static void resiliency_test(void) {};
3406#endif
3407
Christoph Lameter88a420e2007-05-06 14:49:45 -07003408/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003409 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003410 * and freed.
3411 */
3412
3413struct location {
3414 unsigned long count;
3415 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003416 long long sum_time;
3417 long min_time;
3418 long max_time;
3419 long min_pid;
3420 long max_pid;
3421 cpumask_t cpus;
3422 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003423};
3424
3425struct loc_track {
3426 unsigned long max;
3427 unsigned long count;
3428 struct location *loc;
3429};
3430
3431static void free_loc_track(struct loc_track *t)
3432{
3433 if (t->max)
3434 free_pages((unsigned long)t->loc,
3435 get_order(sizeof(struct location) * t->max));
3436}
3437
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003438static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003439{
3440 struct location *l;
3441 int order;
3442
Christoph Lameter88a420e2007-05-06 14:49:45 -07003443 order = get_order(sizeof(struct location) * max);
3444
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003445 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003446 if (!l)
3447 return 0;
3448
3449 if (t->count) {
3450 memcpy(l, t->loc, sizeof(struct location) * t->count);
3451 free_loc_track(t);
3452 }
3453 t->max = max;
3454 t->loc = l;
3455 return 1;
3456}
3457
3458static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003459 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003460{
3461 long start, end, pos;
3462 struct location *l;
3463 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003464 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003465
3466 start = -1;
3467 end = t->count;
3468
3469 for ( ; ; ) {
3470 pos = start + (end - start + 1) / 2;
3471
3472 /*
3473 * There is nothing at "end". If we end up there
3474 * we need to add something to before end.
3475 */
3476 if (pos == end)
3477 break;
3478
3479 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003480 if (track->addr == caddr) {
3481
3482 l = &t->loc[pos];
3483 l->count++;
3484 if (track->when) {
3485 l->sum_time += age;
3486 if (age < l->min_time)
3487 l->min_time = age;
3488 if (age > l->max_time)
3489 l->max_time = age;
3490
3491 if (track->pid < l->min_pid)
3492 l->min_pid = track->pid;
3493 if (track->pid > l->max_pid)
3494 l->max_pid = track->pid;
3495
3496 cpu_set(track->cpu, l->cpus);
3497 }
3498 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003499 return 1;
3500 }
3501
Christoph Lameter45edfa52007-05-09 02:32:45 -07003502 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003503 end = pos;
3504 else
3505 start = pos;
3506 }
3507
3508 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003509 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003510 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003511 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003512 return 0;
3513
3514 l = t->loc + pos;
3515 if (pos < t->count)
3516 memmove(l + 1, l,
3517 (t->count - pos) * sizeof(struct location));
3518 t->count++;
3519 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003520 l->addr = track->addr;
3521 l->sum_time = age;
3522 l->min_time = age;
3523 l->max_time = age;
3524 l->min_pid = track->pid;
3525 l->max_pid = track->pid;
3526 cpus_clear(l->cpus);
3527 cpu_set(track->cpu, l->cpus);
3528 nodes_clear(l->nodes);
3529 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003530 return 1;
3531}
3532
3533static void process_slab(struct loc_track *t, struct kmem_cache *s,
3534 struct page *page, enum track_item alloc)
3535{
Christoph Lameter683d0ba2008-01-07 23:20:29 -08003536 void *addr = slab_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003537 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003538 void *p;
3539
3540 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003541 for_each_free_object(p, s, page->freelist)
3542 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003543
Christoph Lameter7656c722007-05-09 02:32:40 -07003544 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003545 if (!test_bit(slab_index(p, s, addr), map))
3546 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003547}
3548
3549static int list_locations(struct kmem_cache *s, char *buf,
3550 enum track_item alloc)
3551{
Harvey Harrisone374d482008-01-31 15:20:50 -08003552 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003553 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003554 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003555 int node;
3556
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003557 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003558 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003559 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003560
3561 /* Push back cpu slabs */
3562 flush_all(s);
3563
Christoph Lameterf64dc582007-10-16 01:25:33 -07003564 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003565 struct kmem_cache_node *n = get_node(s, node);
3566 unsigned long flags;
3567 struct page *page;
3568
Christoph Lameter9e869432007-08-22 14:01:56 -07003569 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003570 continue;
3571
3572 spin_lock_irqsave(&n->list_lock, flags);
3573 list_for_each_entry(page, &n->partial, lru)
3574 process_slab(&t, s, page, alloc);
3575 list_for_each_entry(page, &n->full, lru)
3576 process_slab(&t, s, page, alloc);
3577 spin_unlock_irqrestore(&n->list_lock, flags);
3578 }
3579
3580 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003581 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003582
Harvey Harrisone374d482008-01-31 15:20:50 -08003583 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003584 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003585 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003586
3587 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003588 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003589 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003590 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003591
3592 if (l->sum_time != l->min_time) {
3593 unsigned long remainder;
3594
Harvey Harrisone374d482008-01-31 15:20:50 -08003595 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003596 l->min_time,
3597 div_long_long_rem(l->sum_time, l->count, &remainder),
3598 l->max_time);
3599 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003600 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003601 l->min_time);
3602
3603 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003604 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003605 l->min_pid, l->max_pid);
3606 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003607 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003608 l->min_pid);
3609
Christoph Lameter84966342007-06-23 17:16:32 -07003610 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003611 len < PAGE_SIZE - 60) {
3612 len += sprintf(buf + len, " cpus=");
3613 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003614 l->cpus);
3615 }
3616
Christoph Lameter84966342007-06-23 17:16:32 -07003617 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003618 len < PAGE_SIZE - 60) {
3619 len += sprintf(buf + len, " nodes=");
3620 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003621 l->nodes);
3622 }
3623
Harvey Harrisone374d482008-01-31 15:20:50 -08003624 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003625 }
3626
3627 free_loc_track(&t);
3628 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003629 len += sprintf(buf, "No data\n");
3630 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003631}
3632
Christoph Lameter81819f02007-05-06 14:49:36 -07003633enum slab_stat_type {
3634 SL_FULL,
3635 SL_PARTIAL,
3636 SL_CPU,
3637 SL_OBJECTS
3638};
3639
3640#define SO_FULL (1 << SL_FULL)
3641#define SO_PARTIAL (1 << SL_PARTIAL)
3642#define SO_CPU (1 << SL_CPU)
3643#define SO_OBJECTS (1 << SL_OBJECTS)
3644
3645static unsigned long slab_objects(struct kmem_cache *s,
3646 char *buf, unsigned long flags)
3647{
3648 unsigned long total = 0;
3649 int cpu;
3650 int node;
3651 int x;
3652 unsigned long *nodes;
3653 unsigned long *per_cpu;
3654
3655 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3656 per_cpu = nodes + nr_node_ids;
3657
3658 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003659 struct page *page;
3660 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003661
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003662 if (!c)
3663 continue;
3664
3665 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003666 node = c->node;
3667 if (node < 0)
3668 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003669 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003670 if (flags & SO_CPU) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003671 if (flags & SO_OBJECTS)
3672 x = page->inuse;
3673 else
3674 x = 1;
3675 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003676 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003677 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003678 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003679 }
3680 }
3681
Christoph Lameterf64dc582007-10-16 01:25:33 -07003682 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003683 struct kmem_cache_node *n = get_node(s, node);
3684
3685 if (flags & SO_PARTIAL) {
3686 if (flags & SO_OBJECTS)
3687 x = count_partial(n);
3688 else
3689 x = n->nr_partial;
3690 total += x;
3691 nodes[node] += x;
3692 }
3693
3694 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003695 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003696 - per_cpu[node]
3697 - n->nr_partial;
3698
3699 if (flags & SO_OBJECTS)
3700 x = full_slabs * s->objects;
3701 else
3702 x = full_slabs;
3703 total += x;
3704 nodes[node] += x;
3705 }
3706 }
3707
3708 x = sprintf(buf, "%lu", total);
3709#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003710 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003711 if (nodes[node])
3712 x += sprintf(buf + x, " N%d=%lu",
3713 node, nodes[node]);
3714#endif
3715 kfree(nodes);
3716 return x + sprintf(buf + x, "\n");
3717}
3718
3719static int any_slab_objects(struct kmem_cache *s)
3720{
3721 int node;
3722 int cpu;
3723
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003724 for_each_possible_cpu(cpu) {
3725 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003726
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003727 if (c && c->page)
3728 return 1;
3729 }
3730
3731 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003732 struct kmem_cache_node *n = get_node(s, node);
3733
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003734 if (!n)
3735 continue;
3736
Christoph Lameter9e869432007-08-22 14:01:56 -07003737 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003738 return 1;
3739 }
3740 return 0;
3741}
3742
3743#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3744#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3745
3746struct slab_attribute {
3747 struct attribute attr;
3748 ssize_t (*show)(struct kmem_cache *s, char *buf);
3749 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3750};
3751
3752#define SLAB_ATTR_RO(_name) \
3753 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3754
3755#define SLAB_ATTR(_name) \
3756 static struct slab_attribute _name##_attr = \
3757 __ATTR(_name, 0644, _name##_show, _name##_store)
3758
Christoph Lameter81819f02007-05-06 14:49:36 -07003759static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3760{
3761 return sprintf(buf, "%d\n", s->size);
3762}
3763SLAB_ATTR_RO(slab_size);
3764
3765static ssize_t align_show(struct kmem_cache *s, char *buf)
3766{
3767 return sprintf(buf, "%d\n", s->align);
3768}
3769SLAB_ATTR_RO(align);
3770
3771static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3772{
3773 return sprintf(buf, "%d\n", s->objsize);
3774}
3775SLAB_ATTR_RO(object_size);
3776
3777static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3778{
3779 return sprintf(buf, "%d\n", s->objects);
3780}
3781SLAB_ATTR_RO(objs_per_slab);
3782
3783static ssize_t order_show(struct kmem_cache *s, char *buf)
3784{
3785 return sprintf(buf, "%d\n", s->order);
3786}
3787SLAB_ATTR_RO(order);
3788
3789static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3790{
3791 if (s->ctor) {
3792 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3793
3794 return n + sprintf(buf + n, "\n");
3795 }
3796 return 0;
3797}
3798SLAB_ATTR_RO(ctor);
3799
Christoph Lameter81819f02007-05-06 14:49:36 -07003800static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3801{
3802 return sprintf(buf, "%d\n", s->refcount - 1);
3803}
3804SLAB_ATTR_RO(aliases);
3805
3806static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3807{
3808 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3809}
3810SLAB_ATTR_RO(slabs);
3811
3812static ssize_t partial_show(struct kmem_cache *s, char *buf)
3813{
3814 return slab_objects(s, buf, SO_PARTIAL);
3815}
3816SLAB_ATTR_RO(partial);
3817
3818static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3819{
3820 return slab_objects(s, buf, SO_CPU);
3821}
3822SLAB_ATTR_RO(cpu_slabs);
3823
3824static ssize_t objects_show(struct kmem_cache *s, char *buf)
3825{
3826 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3827}
3828SLAB_ATTR_RO(objects);
3829
3830static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3831{
3832 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3833}
3834
3835static ssize_t sanity_checks_store(struct kmem_cache *s,
3836 const char *buf, size_t length)
3837{
3838 s->flags &= ~SLAB_DEBUG_FREE;
3839 if (buf[0] == '1')
3840 s->flags |= SLAB_DEBUG_FREE;
3841 return length;
3842}
3843SLAB_ATTR(sanity_checks);
3844
3845static ssize_t trace_show(struct kmem_cache *s, char *buf)
3846{
3847 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3848}
3849
3850static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3851 size_t length)
3852{
3853 s->flags &= ~SLAB_TRACE;
3854 if (buf[0] == '1')
3855 s->flags |= SLAB_TRACE;
3856 return length;
3857}
3858SLAB_ATTR(trace);
3859
3860static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3861{
3862 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3863}
3864
3865static ssize_t reclaim_account_store(struct kmem_cache *s,
3866 const char *buf, size_t length)
3867{
3868 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3869 if (buf[0] == '1')
3870 s->flags |= SLAB_RECLAIM_ACCOUNT;
3871 return length;
3872}
3873SLAB_ATTR(reclaim_account);
3874
3875static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3876{
Christoph Lameter5af60832007-05-06 14:49:56 -07003877 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003878}
3879SLAB_ATTR_RO(hwcache_align);
3880
3881#ifdef CONFIG_ZONE_DMA
3882static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3883{
3884 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3885}
3886SLAB_ATTR_RO(cache_dma);
3887#endif
3888
3889static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3890{
3891 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3892}
3893SLAB_ATTR_RO(destroy_by_rcu);
3894
3895static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3896{
3897 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3898}
3899
3900static ssize_t red_zone_store(struct kmem_cache *s,
3901 const char *buf, size_t length)
3902{
3903 if (any_slab_objects(s))
3904 return -EBUSY;
3905
3906 s->flags &= ~SLAB_RED_ZONE;
3907 if (buf[0] == '1')
3908 s->flags |= SLAB_RED_ZONE;
3909 calculate_sizes(s);
3910 return length;
3911}
3912SLAB_ATTR(red_zone);
3913
3914static ssize_t poison_show(struct kmem_cache *s, char *buf)
3915{
3916 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3917}
3918
3919static ssize_t poison_store(struct kmem_cache *s,
3920 const char *buf, size_t length)
3921{
3922 if (any_slab_objects(s))
3923 return -EBUSY;
3924
3925 s->flags &= ~SLAB_POISON;
3926 if (buf[0] == '1')
3927 s->flags |= SLAB_POISON;
3928 calculate_sizes(s);
3929 return length;
3930}
3931SLAB_ATTR(poison);
3932
3933static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3934{
3935 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3936}
3937
3938static ssize_t store_user_store(struct kmem_cache *s,
3939 const char *buf, size_t length)
3940{
3941 if (any_slab_objects(s))
3942 return -EBUSY;
3943
3944 s->flags &= ~SLAB_STORE_USER;
3945 if (buf[0] == '1')
3946 s->flags |= SLAB_STORE_USER;
3947 calculate_sizes(s);
3948 return length;
3949}
3950SLAB_ATTR(store_user);
3951
Christoph Lameter53e15af2007-05-06 14:49:43 -07003952static ssize_t validate_show(struct kmem_cache *s, char *buf)
3953{
3954 return 0;
3955}
3956
3957static ssize_t validate_store(struct kmem_cache *s,
3958 const char *buf, size_t length)
3959{
Christoph Lameter434e2452007-07-17 04:03:30 -07003960 int ret = -EINVAL;
3961
3962 if (buf[0] == '1') {
3963 ret = validate_slab_cache(s);
3964 if (ret >= 0)
3965 ret = length;
3966 }
3967 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003968}
3969SLAB_ATTR(validate);
3970
Christoph Lameter2086d262007-05-06 14:49:46 -07003971static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3972{
3973 return 0;
3974}
3975
3976static ssize_t shrink_store(struct kmem_cache *s,
3977 const char *buf, size_t length)
3978{
3979 if (buf[0] == '1') {
3980 int rc = kmem_cache_shrink(s);
3981
3982 if (rc)
3983 return rc;
3984 } else
3985 return -EINVAL;
3986 return length;
3987}
3988SLAB_ATTR(shrink);
3989
Christoph Lameter88a420e2007-05-06 14:49:45 -07003990static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3991{
3992 if (!(s->flags & SLAB_STORE_USER))
3993 return -ENOSYS;
3994 return list_locations(s, buf, TRACK_ALLOC);
3995}
3996SLAB_ATTR_RO(alloc_calls);
3997
3998static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3999{
4000 if (!(s->flags & SLAB_STORE_USER))
4001 return -ENOSYS;
4002 return list_locations(s, buf, TRACK_FREE);
4003}
4004SLAB_ATTR_RO(free_calls);
4005
Christoph Lameter81819f02007-05-06 14:49:36 -07004006#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004007static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004008{
Christoph Lameter98246012008-01-07 23:20:26 -08004009 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004010}
4011
Christoph Lameter98246012008-01-07 23:20:26 -08004012static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004013 const char *buf, size_t length)
4014{
4015 int n = simple_strtoul(buf, NULL, 10);
4016
4017 if (n < 100)
Christoph Lameter98246012008-01-07 23:20:26 -08004018 s->remote_node_defrag_ratio = n * 10;
Christoph Lameter81819f02007-05-06 14:49:36 -07004019 return length;
4020}
Christoph Lameter98246012008-01-07 23:20:26 -08004021SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004022#endif
4023
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004024#ifdef CONFIG_SLUB_STATS
4025
4026static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4027{
4028 unsigned long sum = 0;
4029 int cpu;
4030 int len;
4031 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4032
4033 if (!data)
4034 return -ENOMEM;
4035
4036 for_each_online_cpu(cpu) {
4037 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4038
4039 data[cpu] = x;
4040 sum += x;
4041 }
4042
4043 len = sprintf(buf, "%lu", sum);
4044
4045 for_each_online_cpu(cpu) {
4046 if (data[cpu] && len < PAGE_SIZE - 20)
4047 len += sprintf(buf + len, " c%d=%u", cpu, data[cpu]);
4048 }
4049 kfree(data);
4050 return len + sprintf(buf + len, "\n");
4051}
4052
4053#define STAT_ATTR(si, text) \
4054static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4055{ \
4056 return show_stat(s, buf, si); \
4057} \
4058SLAB_ATTR_RO(text); \
4059
4060STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4061STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4062STAT_ATTR(FREE_FASTPATH, free_fastpath);
4063STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4064STAT_ATTR(FREE_FROZEN, free_frozen);
4065STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4066STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4067STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4068STAT_ATTR(ALLOC_SLAB, alloc_slab);
4069STAT_ATTR(ALLOC_REFILL, alloc_refill);
4070STAT_ATTR(FREE_SLAB, free_slab);
4071STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4072STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4073STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4074STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4075STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4076STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4077
4078#endif
4079
Pekka Enberg06428782008-01-07 23:20:27 -08004080static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004081 &slab_size_attr.attr,
4082 &object_size_attr.attr,
4083 &objs_per_slab_attr.attr,
4084 &order_attr.attr,
4085 &objects_attr.attr,
4086 &slabs_attr.attr,
4087 &partial_attr.attr,
4088 &cpu_slabs_attr.attr,
4089 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004090 &aliases_attr.attr,
4091 &align_attr.attr,
4092 &sanity_checks_attr.attr,
4093 &trace_attr.attr,
4094 &hwcache_align_attr.attr,
4095 &reclaim_account_attr.attr,
4096 &destroy_by_rcu_attr.attr,
4097 &red_zone_attr.attr,
4098 &poison_attr.attr,
4099 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004100 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004101 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004102 &alloc_calls_attr.attr,
4103 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004104#ifdef CONFIG_ZONE_DMA
4105 &cache_dma_attr.attr,
4106#endif
4107#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004108 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004109#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004110#ifdef CONFIG_SLUB_STATS
4111 &alloc_fastpath_attr.attr,
4112 &alloc_slowpath_attr.attr,
4113 &free_fastpath_attr.attr,
4114 &free_slowpath_attr.attr,
4115 &free_frozen_attr.attr,
4116 &free_add_partial_attr.attr,
4117 &free_remove_partial_attr.attr,
4118 &alloc_from_partial_attr.attr,
4119 &alloc_slab_attr.attr,
4120 &alloc_refill_attr.attr,
4121 &free_slab_attr.attr,
4122 &cpuslab_flush_attr.attr,
4123 &deactivate_full_attr.attr,
4124 &deactivate_empty_attr.attr,
4125 &deactivate_to_head_attr.attr,
4126 &deactivate_to_tail_attr.attr,
4127 &deactivate_remote_frees_attr.attr,
4128#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004129 NULL
4130};
4131
4132static struct attribute_group slab_attr_group = {
4133 .attrs = slab_attrs,
4134};
4135
4136static ssize_t slab_attr_show(struct kobject *kobj,
4137 struct attribute *attr,
4138 char *buf)
4139{
4140 struct slab_attribute *attribute;
4141 struct kmem_cache *s;
4142 int err;
4143
4144 attribute = to_slab_attr(attr);
4145 s = to_slab(kobj);
4146
4147 if (!attribute->show)
4148 return -EIO;
4149
4150 err = attribute->show(s, buf);
4151
4152 return err;
4153}
4154
4155static ssize_t slab_attr_store(struct kobject *kobj,
4156 struct attribute *attr,
4157 const char *buf, size_t len)
4158{
4159 struct slab_attribute *attribute;
4160 struct kmem_cache *s;
4161 int err;
4162
4163 attribute = to_slab_attr(attr);
4164 s = to_slab(kobj);
4165
4166 if (!attribute->store)
4167 return -EIO;
4168
4169 err = attribute->store(s, buf, len);
4170
4171 return err;
4172}
4173
Christoph Lameter151c6022008-01-07 22:29:05 -08004174static void kmem_cache_release(struct kobject *kobj)
4175{
4176 struct kmem_cache *s = to_slab(kobj);
4177
4178 kfree(s);
4179}
4180
Christoph Lameter81819f02007-05-06 14:49:36 -07004181static struct sysfs_ops slab_sysfs_ops = {
4182 .show = slab_attr_show,
4183 .store = slab_attr_store,
4184};
4185
4186static struct kobj_type slab_ktype = {
4187 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004188 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004189};
4190
4191static int uevent_filter(struct kset *kset, struct kobject *kobj)
4192{
4193 struct kobj_type *ktype = get_ktype(kobj);
4194
4195 if (ktype == &slab_ktype)
4196 return 1;
4197 return 0;
4198}
4199
4200static struct kset_uevent_ops slab_uevent_ops = {
4201 .filter = uevent_filter,
4202};
4203
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004204static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004205
4206#define ID_STR_LENGTH 64
4207
4208/* Create a unique string id for a slab cache:
4209 * format
4210 * :[flags-]size:[memory address of kmemcache]
4211 */
4212static char *create_unique_id(struct kmem_cache *s)
4213{
4214 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4215 char *p = name;
4216
4217 BUG_ON(!name);
4218
4219 *p++ = ':';
4220 /*
4221 * First flags affecting slabcache operations. We will only
4222 * get here for aliasable slabs so we do not need to support
4223 * too many flags. The flags here must cover all flags that
4224 * are matched during merging to guarantee that the id is
4225 * unique.
4226 */
4227 if (s->flags & SLAB_CACHE_DMA)
4228 *p++ = 'd';
4229 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4230 *p++ = 'a';
4231 if (s->flags & SLAB_DEBUG_FREE)
4232 *p++ = 'F';
4233 if (p != name + 1)
4234 *p++ = '-';
4235 p += sprintf(p, "%07d", s->size);
4236 BUG_ON(p > name + ID_STR_LENGTH - 1);
4237 return name;
4238}
4239
4240static int sysfs_slab_add(struct kmem_cache *s)
4241{
4242 int err;
4243 const char *name;
4244 int unmergeable;
4245
4246 if (slab_state < SYSFS)
4247 /* Defer until later */
4248 return 0;
4249
4250 unmergeable = slab_unmergeable(s);
4251 if (unmergeable) {
4252 /*
4253 * Slabcache can never be merged so we can use the name proper.
4254 * This is typically the case for debug situations. In that
4255 * case we can catch duplicate names easily.
4256 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004257 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004258 name = s->name;
4259 } else {
4260 /*
4261 * Create a unique name for the slab as a target
4262 * for the symlinks.
4263 */
4264 name = create_unique_id(s);
4265 }
4266
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004267 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004268 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4269 if (err) {
4270 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004271 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004272 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004273
4274 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4275 if (err)
4276 return err;
4277 kobject_uevent(&s->kobj, KOBJ_ADD);
4278 if (!unmergeable) {
4279 /* Setup first alias */
4280 sysfs_slab_alias(s, s->name);
4281 kfree(name);
4282 }
4283 return 0;
4284}
4285
4286static void sysfs_slab_remove(struct kmem_cache *s)
4287{
4288 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4289 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004290 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004291}
4292
4293/*
4294 * Need to buffer aliases during bootup until sysfs becomes
4295 * available lest we loose that information.
4296 */
4297struct saved_alias {
4298 struct kmem_cache *s;
4299 const char *name;
4300 struct saved_alias *next;
4301};
4302
Adrian Bunk5af328a2007-07-17 04:03:27 -07004303static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004304
4305static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4306{
4307 struct saved_alias *al;
4308
4309 if (slab_state == SYSFS) {
4310 /*
4311 * If we have a leftover link then remove it.
4312 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004313 sysfs_remove_link(&slab_kset->kobj, name);
4314 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004315 }
4316
4317 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4318 if (!al)
4319 return -ENOMEM;
4320
4321 al->s = s;
4322 al->name = name;
4323 al->next = alias_list;
4324 alias_list = al;
4325 return 0;
4326}
4327
4328static int __init slab_sysfs_init(void)
4329{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004330 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004331 int err;
4332
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004333 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004334 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004335 printk(KERN_ERR "Cannot register slab subsystem.\n");
4336 return -ENOSYS;
4337 }
4338
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004339 slab_state = SYSFS;
4340
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004341 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004342 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004343 if (err)
4344 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4345 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004346 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004347
4348 while (alias_list) {
4349 struct saved_alias *al = alias_list;
4350
4351 alias_list = alias_list->next;
4352 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004353 if (err)
4354 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4355 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004356 kfree(al);
4357 }
4358
4359 resiliency_test();
4360 return 0;
4361}
4362
4363__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004364#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004365
4366/*
4367 * The /proc/slabinfo ABI
4368 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004369#ifdef CONFIG_SLABINFO
4370
4371ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4372 size_t count, loff_t *ppos)
4373{
4374 return -EINVAL;
4375}
4376
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004377
4378static void print_slabinfo_header(struct seq_file *m)
4379{
4380 seq_puts(m, "slabinfo - version: 2.1\n");
4381 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4382 "<objperslab> <pagesperslab>");
4383 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4384 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4385 seq_putc(m, '\n');
4386}
4387
4388static void *s_start(struct seq_file *m, loff_t *pos)
4389{
4390 loff_t n = *pos;
4391
4392 down_read(&slub_lock);
4393 if (!n)
4394 print_slabinfo_header(m);
4395
4396 return seq_list_start(&slab_caches, *pos);
4397}
4398
4399static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4400{
4401 return seq_list_next(p, &slab_caches, pos);
4402}
4403
4404static void s_stop(struct seq_file *m, void *p)
4405{
4406 up_read(&slub_lock);
4407}
4408
4409static int s_show(struct seq_file *m, void *p)
4410{
4411 unsigned long nr_partials = 0;
4412 unsigned long nr_slabs = 0;
4413 unsigned long nr_inuse = 0;
4414 unsigned long nr_objs;
4415 struct kmem_cache *s;
4416 int node;
4417
4418 s = list_entry(p, struct kmem_cache, list);
4419
4420 for_each_online_node(node) {
4421 struct kmem_cache_node *n = get_node(s, node);
4422
4423 if (!n)
4424 continue;
4425
4426 nr_partials += n->nr_partial;
4427 nr_slabs += atomic_long_read(&n->nr_slabs);
4428 nr_inuse += count_partial(n);
4429 }
4430
4431 nr_objs = nr_slabs * s->objects;
4432 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4433
4434 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4435 nr_objs, s->size, s->objects, (1 << s->order));
4436 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4437 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4438 0UL);
4439 seq_putc(m, '\n');
4440 return 0;
4441}
4442
4443const struct seq_operations slabinfo_op = {
4444 .start = s_start,
4445 .next = s_next,
4446 .stop = s_stop,
4447 .show = s_show,
4448};
4449
Linus Torvalds158a9622008-01-02 13:04:48 -08004450#endif /* CONFIG_SLABINFO */