blob: 9b2d6178d06ce7259941c6aafa7b20cf81228e40 [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>
23
24/*
25 * Lock order:
26 * 1. slab_lock(page)
27 * 2. slab->list_lock
28 *
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
35 *
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
41 *
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
46 * the list lock.
47 *
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
60 *
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
65 *
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
68 *
Christoph Lameter672bba32007-05-09 02:32:39 -070069 * Slabs with free elements are kept on a partial list and during regular
70 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070071 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070072 * We track full slabs for debugging purposes though because otherwise we
73 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070074 *
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
78 *
79 * Overloading of page flags that are otherwise used for LRU management.
80 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070081 * PageActive The slab is frozen and exempt from list processing.
82 * This means that the slab is dedicated to a purpose
83 * such as satisfying allocations for a specific
84 * processor. Objects may be freed in the slab while
85 * it is frozen but slab_free will then skip the usual
86 * list operations. It is up to the processor holding
87 * the slab to integrate the slab into the slab lists
88 * when the slab is no longer needed.
89 *
90 * One use of this flag is to mark slabs that are
91 * used for allocations. Then such a slab becomes a cpu
92 * slab. The cpu slab may be equipped with an additional
Christoph Lameter894b8782007-05-10 03:15:16 -070093 * lockless_freelist that allows lockless access to
94 * free objects in addition to the regular freelist
95 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070096 *
97 * PageError Slab requires special handling due to debug
98 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -070099 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700100 */
101
Christoph Lameter5577bd82007-05-16 22:10:56 -0700102#define FROZEN (1 << PG_active)
103
104#ifdef CONFIG_SLUB_DEBUG
105#define SLABDEBUG (1 << PG_error)
106#else
107#define SLABDEBUG 0
108#endif
109
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700110static inline int SlabFrozen(struct page *page)
111{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700112 return page->flags & FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700113}
114
115static inline void SetSlabFrozen(struct page *page)
116{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700117 page->flags |= FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700118}
119
120static inline void ClearSlabFrozen(struct page *page)
121{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700122 page->flags &= ~FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700123}
124
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700125static inline int SlabDebug(struct page *page)
126{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700127 return page->flags & SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700128}
129
130static inline void SetSlabDebug(struct page *page)
131{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700132 page->flags |= SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700133}
134
135static inline void ClearSlabDebug(struct page *page)
136{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700137 page->flags &= ~SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700138}
139
Christoph Lameter81819f02007-05-06 14:49:36 -0700140/*
141 * Issues still to be resolved:
142 *
143 * - The per cpu array is updated for each new slab and and is a remote
144 * cacheline for most nodes. This could become a bouncing cacheline given
Christoph Lameter672bba32007-05-09 02:32:39 -0700145 * enough frequent updates. There are 16 pointers in a cacheline, so at
146 * max 16 cpus could compete for the cacheline which may be okay.
Christoph Lameter81819f02007-05-06 14:49:36 -0700147 *
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700150 * - Variable sizing of the per node arrays
151 */
152
153/* Enable to test recovery from slab corruption on boot */
154#undef SLUB_RESILIENCY_TEST
155
156#if PAGE_SHIFT <= 12
157
158/*
159 * Small page size. Make sure that we do not fragment memory
160 */
161#define DEFAULT_MAX_ORDER 1
162#define DEFAULT_MIN_OBJECTS 4
163
164#else
165
166/*
167 * Large page machines are customarily able to handle larger
168 * page orders.
169 */
170#define DEFAULT_MAX_ORDER 2
171#define DEFAULT_MIN_OBJECTS 8
172
173#endif
174
175/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700176 * Mininum number of partial slabs. These will be left on the partial
177 * lists even if they are empty. kmem_cache_shrink may reclaim them.
178 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700179#define MIN_PARTIAL 2
180
Christoph Lameter2086d262007-05-06 14:49:46 -0700181/*
182 * Maximum number of desirable partial slabs.
183 * The existence of more partial slabs makes kmem_cache_shrink
184 * sort the partial list by the number of objects in the.
185 */
186#define MAX_PARTIAL 10
187
Christoph Lameter81819f02007-05-06 14:49:36 -0700188#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
189 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700190
Christoph Lameter81819f02007-05-06 14:49:36 -0700191/*
192 * Set of flags that will prevent slab merging
193 */
194#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
195 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
196
197#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
198 SLAB_CACHE_DMA)
199
200#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700201#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700202#endif
203
204#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700205#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700206#endif
207
Christoph Lameter6300ea72007-07-17 04:03:20 -0700208/*
209 * The page->inuse field is 16 bit thus we have this limitation
210 */
211#define MAX_OBJECTS_PER_SLAB 65535
212
Christoph Lameter81819f02007-05-06 14:49:36 -0700213/* Internal SLUB flags */
214#define __OBJECT_POISON 0x80000000 /* Poison object */
215
Christoph Lameter65c02d42007-05-09 02:32:35 -0700216/* Not all arches define cache_line_size */
217#ifndef cache_line_size
218#define cache_line_size() L1_CACHE_BYTES
219#endif
220
Christoph Lameter81819f02007-05-06 14:49:36 -0700221static int kmem_size = sizeof(struct kmem_cache);
222
223#ifdef CONFIG_SMP
224static struct notifier_block slab_notifier;
225#endif
226
227static enum {
228 DOWN, /* No slab functionality available */
229 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700230 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700231 SYSFS /* Sysfs up */
232} slab_state = DOWN;
233
234/* A list of all slab caches on the system */
235static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700236static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700237
Christoph Lameter02cbc872007-05-09 02:32:43 -0700238/*
239 * Tracking user of a slab.
240 */
241struct track {
242 void *addr; /* Called from address */
243 int cpu; /* Was running on cpu */
244 int pid; /* Pid context */
245 unsigned long when; /* When did the operation occur */
246};
247
248enum track_item { TRACK_ALLOC, TRACK_FREE };
249
Christoph Lameter41ecc552007-05-09 02:32:44 -0700250#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter81819f02007-05-06 14:49:36 -0700251static int sysfs_slab_add(struct kmem_cache *);
252static int sysfs_slab_alias(struct kmem_cache *, const char *);
253static void sysfs_slab_remove(struct kmem_cache *);
254#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; }
258static inline void sysfs_slab_remove(struct kmem_cache *s) {}
Christoph Lameter81819f02007-05-06 14:49:36 -0700259#endif
260
261/********************************************************************
262 * Core slab cache functions
263 *******************************************************************/
264
265int slab_is_available(void)
266{
267 return slab_state >= UP;
268}
269
270static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
271{
272#ifdef CONFIG_NUMA
273 return s->node[node];
274#else
275 return &s->local_node;
276#endif
277}
278
Christoph Lameter02cbc872007-05-09 02:32:43 -0700279static inline int check_valid_pointer(struct kmem_cache *s,
280 struct page *page, const void *object)
281{
282 void *base;
283
284 if (!object)
285 return 1;
286
287 base = page_address(page);
288 if (object < base || object >= base + s->objects * s->size ||
289 (object - base) % s->size) {
290 return 0;
291 }
292
293 return 1;
294}
295
Christoph Lameter81819f02007-05-06 14:49:36 -0700296/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700297 * Slow version of get and set free pointer.
298 *
299 * This version requires touching the cache lines of kmem_cache which
300 * we avoid to do in the fast alloc free paths. There we obtain the offset
301 * from the page struct.
302 */
303static inline void *get_freepointer(struct kmem_cache *s, void *object)
304{
305 return *(void **)(object + s->offset);
306}
307
308static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
309{
310 *(void **)(object + s->offset) = fp;
311}
312
313/* Loop over all objects in a slab */
314#define for_each_object(__p, __s, __addr) \
315 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
316 __p += (__s)->size)
317
318/* Scan freelist */
319#define for_each_free_object(__p, __s, __free) \
320 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
321
322/* Determine object index from a given position */
323static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
324{
325 return (p - addr) / s->size;
326}
327
Christoph Lameter41ecc552007-05-09 02:32:44 -0700328#ifdef CONFIG_SLUB_DEBUG
329/*
330 * Debug settings:
331 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700332#ifdef CONFIG_SLUB_DEBUG_ON
333static int slub_debug = DEBUG_DEFAULT_FLAGS;
334#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700335static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700336#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700337
338static char *slub_debug_slabs;
339
Christoph Lameter7656c722007-05-09 02:32:40 -0700340/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700341 * Object debugging
342 */
343static void print_section(char *text, u8 *addr, unsigned int length)
344{
345 int i, offset;
346 int newline = 1;
347 char ascii[17];
348
349 ascii[16] = 0;
350
351 for (i = 0; i < length; i++) {
352 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700353 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700354 newline = 0;
355 }
356 printk(" %02x", addr[i]);
357 offset = i % 16;
358 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
359 if (offset == 15) {
360 printk(" %s\n",ascii);
361 newline = 1;
362 }
363 }
364 if (!newline) {
365 i %= 16;
366 while (i < 16) {
367 printk(" ");
368 ascii[i] = ' ';
369 i++;
370 }
371 printk(" %s\n", ascii);
372 }
373}
374
Christoph Lameter81819f02007-05-06 14:49:36 -0700375static struct track *get_track(struct kmem_cache *s, void *object,
376 enum track_item alloc)
377{
378 struct track *p;
379
380 if (s->offset)
381 p = object + s->offset + sizeof(void *);
382 else
383 p = object + s->inuse;
384
385 return p + alloc;
386}
387
388static void set_track(struct kmem_cache *s, void *object,
389 enum track_item alloc, void *addr)
390{
391 struct track *p;
392
393 if (s->offset)
394 p = object + s->offset + sizeof(void *);
395 else
396 p = object + s->inuse;
397
398 p += alloc;
399 if (addr) {
400 p->addr = addr;
401 p->cpu = smp_processor_id();
402 p->pid = current ? current->pid : -1;
403 p->when = jiffies;
404 } else
405 memset(p, 0, sizeof(struct track));
406}
407
Christoph Lameter81819f02007-05-06 14:49:36 -0700408static void init_tracking(struct kmem_cache *s, void *object)
409{
Christoph Lameter24922682007-07-17 04:03:18 -0700410 if (!(s->flags & SLAB_STORE_USER))
411 return;
412
413 set_track(s, object, TRACK_FREE, NULL);
414 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700415}
416
417static void print_track(const char *s, struct track *t)
418{
419 if (!t->addr)
420 return;
421
Christoph Lameter24922682007-07-17 04:03:18 -0700422 printk(KERN_ERR "INFO: %s in ", s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700423 __print_symbol("%s", (unsigned long)t->addr);
Christoph Lameter24922682007-07-17 04:03:18 -0700424 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700425}
426
Christoph Lameter24922682007-07-17 04:03:18 -0700427static void print_tracking(struct kmem_cache *s, void *object)
428{
429 if (!(s->flags & SLAB_STORE_USER))
430 return;
431
432 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
433 print_track("Freed", get_track(s, object, TRACK_FREE));
434}
435
436static void print_page_info(struct page *page)
437{
438 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
439 page, page->inuse, page->freelist, page->flags);
440
441}
442
443static void slab_bug(struct kmem_cache *s, char *fmt, ...)
444{
445 va_list args;
446 char buf[100];
447
448 va_start(args, fmt);
449 vsnprintf(buf, sizeof(buf), fmt, args);
450 va_end(args);
451 printk(KERN_ERR "========================================"
452 "=====================================\n");
453 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
454 printk(KERN_ERR "----------------------------------------"
455 "-------------------------------------\n\n");
456}
457
458static void slab_fix(struct kmem_cache *s, char *fmt, ...)
459{
460 va_list args;
461 char buf[100];
462
463 va_start(args, fmt);
464 vsnprintf(buf, sizeof(buf), fmt, args);
465 va_end(args);
466 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
467}
468
469static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700470{
471 unsigned int off; /* Offset of last byte */
Christoph Lameter24922682007-07-17 04:03:18 -0700472 u8 *addr = page_address(page);
473
474 print_tracking(s, p);
475
476 print_page_info(page);
477
478 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
479 p, p - addr, get_freepointer(s, p));
480
481 if (p > addr + 16)
482 print_section("Bytes b4", p - 16, 16);
483
484 print_section("Object", p, min(s->objsize, 128));
Christoph Lameter81819f02007-05-06 14:49:36 -0700485
486 if (s->flags & SLAB_RED_ZONE)
487 print_section("Redzone", p + s->objsize,
488 s->inuse - s->objsize);
489
Christoph Lameter81819f02007-05-06 14:49:36 -0700490 if (s->offset)
491 off = s->offset + sizeof(void *);
492 else
493 off = s->inuse;
494
Christoph Lameter24922682007-07-17 04:03:18 -0700495 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700496 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700497
498 if (off != s->size)
499 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700500 print_section("Padding", p + off, s->size - off);
501
502 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700503}
504
505static void object_err(struct kmem_cache *s, struct page *page,
506 u8 *object, char *reason)
507{
Christoph Lameter24922682007-07-17 04:03:18 -0700508 slab_bug(s, reason);
509 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700510}
511
Christoph Lameter24922682007-07-17 04:03:18 -0700512static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700513{
514 va_list args;
515 char buf[100];
516
Christoph Lameter24922682007-07-17 04:03:18 -0700517 va_start(args, fmt);
518 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700519 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700520 slab_bug(s, fmt);
521 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700522 dump_stack();
523}
524
525static void init_object(struct kmem_cache *s, void *object, int active)
526{
527 u8 *p = object;
528
529 if (s->flags & __OBJECT_POISON) {
530 memset(p, POISON_FREE, s->objsize - 1);
531 p[s->objsize -1] = POISON_END;
532 }
533
534 if (s->flags & SLAB_RED_ZONE)
535 memset(p + s->objsize,
536 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
537 s->inuse - s->objsize);
538}
539
Christoph Lameter24922682007-07-17 04:03:18 -0700540static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700541{
542 while (bytes) {
543 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700544 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700545 start++;
546 bytes--;
547 }
Christoph Lameter24922682007-07-17 04:03:18 -0700548 return NULL;
549}
550
551static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
552 void *from, void *to)
553{
554 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
555 memset(from, data, to - from);
556}
557
558static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
559 u8 *object, char *what,
560 u8* start, unsigned int value, unsigned int bytes)
561{
562 u8 *fault;
563 u8 *end;
564
565 fault = check_bytes(start, value, bytes);
566 if (!fault)
567 return 1;
568
569 end = start + bytes;
570 while (end > fault && end[-1] == value)
571 end--;
572
573 slab_bug(s, "%s overwritten", what);
574 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
575 fault, end - 1, fault[0], value);
576 print_trailer(s, page, object);
577
578 restore_bytes(s, what, value, fault, end);
579 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700580}
581
Christoph Lameter81819f02007-05-06 14:49:36 -0700582/*
583 * Object layout:
584 *
585 * object address
586 * Bytes of the object to be managed.
587 * If the freepointer may overlay the object then the free
588 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700589 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700590 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
591 * 0xa5 (POISON_END)
592 *
593 * object + s->objsize
594 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700595 * Padding is extended by another word if Redzoning is enabled and
596 * objsize == inuse.
597 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700598 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
599 * 0xcc (RED_ACTIVE) for objects in use.
600 *
601 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700602 * Meta data starts here.
603 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700604 * A. Free pointer (if we cannot overwrite object on free)
605 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700606 * C. Padding to reach required alignment boundary or at mininum
607 * one word if debuggin is on to be able to detect writes
608 * before the word boundary.
609 *
610 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700611 *
612 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700613 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700614 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700615 * If slabcaches are merged then the objsize and inuse boundaries are mostly
616 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700617 * may be used with merged slabcaches.
618 */
619
Christoph Lameter81819f02007-05-06 14:49:36 -0700620static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
621{
622 unsigned long off = s->inuse; /* The end of info */
623
624 if (s->offset)
625 /* Freepointer is placed after the object. */
626 off += sizeof(void *);
627
628 if (s->flags & SLAB_STORE_USER)
629 /* We also have user information there */
630 off += 2 * sizeof(struct track);
631
632 if (s->size == off)
633 return 1;
634
Christoph Lameter24922682007-07-17 04:03:18 -0700635 return check_bytes_and_report(s, page, p, "Object padding",
636 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700637}
638
639static int slab_pad_check(struct kmem_cache *s, struct page *page)
640{
Christoph Lameter24922682007-07-17 04:03:18 -0700641 u8 *start;
642 u8 *fault;
643 u8 *end;
644 int length;
645 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700646
647 if (!(s->flags & SLAB_POISON))
648 return 1;
649
Christoph Lameter24922682007-07-17 04:03:18 -0700650 start = page_address(page);
651 end = start + (PAGE_SIZE << s->order);
Christoph Lameter81819f02007-05-06 14:49:36 -0700652 length = s->objects * s->size;
Christoph Lameter24922682007-07-17 04:03:18 -0700653 remainder = end - (start + length);
Christoph Lameter81819f02007-05-06 14:49:36 -0700654 if (!remainder)
655 return 1;
656
Christoph Lameter24922682007-07-17 04:03:18 -0700657 fault = check_bytes(start + length, POISON_INUSE, remainder);
658 if (!fault)
659 return 1;
660 while (end > fault && end[-1] == POISON_INUSE)
661 end--;
662
663 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
664 print_section("Padding", start, length);
665
666 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
667 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700668}
669
670static int check_object(struct kmem_cache *s, struct page *page,
671 void *object, int active)
672{
673 u8 *p = object;
674 u8 *endobject = object + s->objsize;
675
676 if (s->flags & SLAB_RED_ZONE) {
677 unsigned int red =
678 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
679
Christoph Lameter24922682007-07-17 04:03:18 -0700680 if (!check_bytes_and_report(s, page, object, "Redzone",
681 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700682 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700683 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700684 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse)
685 check_bytes_and_report(s, page, p, "Alignment padding", endobject,
686 POISON_INUSE, s->inuse - s->objsize);
Christoph Lameter81819f02007-05-06 14:49:36 -0700687 }
688
689 if (s->flags & SLAB_POISON) {
690 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700691 (!check_bytes_and_report(s, page, p, "Poison", p,
692 POISON_FREE, s->objsize - 1) ||
693 !check_bytes_and_report(s, page, p, "Poison",
694 p + s->objsize -1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700695 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700696 /*
697 * check_pad_bytes cleans up on its own.
698 */
699 check_pad_bytes(s, page, p);
700 }
701
702 if (!s->offset && active)
703 /*
704 * Object and freepointer overlap. Cannot check
705 * freepointer while object is allocated.
706 */
707 return 1;
708
709 /* Check free pointer validity */
710 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
711 object_err(s, page, p, "Freepointer corrupt");
712 /*
713 * No choice but to zap it and thus loose the remainder
714 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700715 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700716 */
717 set_freepointer(s, p, NULL);
718 return 0;
719 }
720 return 1;
721}
722
723static int check_slab(struct kmem_cache *s, struct page *page)
724{
725 VM_BUG_ON(!irqs_disabled());
726
727 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700728 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700729 return 0;
730 }
731 if (page->offset * sizeof(void *) != s->offset) {
Christoph Lameter24922682007-07-17 04:03:18 -0700732 slab_err(s, page, "Corrupted offset %lu",
733 (unsigned long)(page->offset * sizeof(void *)));
Christoph Lameter81819f02007-05-06 14:49:36 -0700734 return 0;
735 }
736 if (page->inuse > s->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700737 slab_err(s, page, "inuse %u > max %u",
738 s->name, page->inuse, s->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700739 return 0;
740 }
741 /* Slab_pad_check fixes things up after itself */
742 slab_pad_check(s, page);
743 return 1;
744}
745
746/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700747 * Determine if a certain object on a page is on the freelist. Must hold the
748 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700749 */
750static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
751{
752 int nr = 0;
753 void *fp = page->freelist;
754 void *object = NULL;
755
756 while (fp && nr <= s->objects) {
757 if (fp == search)
758 return 1;
759 if (!check_valid_pointer(s, page, fp)) {
760 if (object) {
761 object_err(s, page, object,
762 "Freechain corrupt");
763 set_freepointer(s, object, NULL);
764 break;
765 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700766 slab_err(s, page, "Freepointer corrupt");
Christoph Lameter81819f02007-05-06 14:49:36 -0700767 page->freelist = NULL;
768 page->inuse = s->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700769 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700770 return 0;
771 }
772 break;
773 }
774 object = fp;
775 fp = get_freepointer(s, object);
776 nr++;
777 }
778
779 if (page->inuse != s->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700780 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter24922682007-07-17 04:03:18 -0700781 "counted were %d", page->inuse, s->objects - nr);
Christoph Lameter81819f02007-05-06 14:49:36 -0700782 page->inuse = s->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700783 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700784 }
785 return search == NULL;
786}
787
Christoph Lameter3ec09742007-05-16 22:11:00 -0700788static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
789{
790 if (s->flags & SLAB_TRACE) {
791 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
792 s->name,
793 alloc ? "alloc" : "free",
794 object, page->inuse,
795 page->freelist);
796
797 if (!alloc)
798 print_section("Object", (void *)object, s->objsize);
799
800 dump_stack();
801 }
802}
803
Christoph Lameter643b1132007-05-06 14:49:42 -0700804/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700805 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700806 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700807static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700808{
Christoph Lameter643b1132007-05-06 14:49:42 -0700809 spin_lock(&n->list_lock);
810 list_add(&page->lru, &n->full);
811 spin_unlock(&n->list_lock);
812}
813
814static void remove_full(struct kmem_cache *s, struct page *page)
815{
816 struct kmem_cache_node *n;
817
818 if (!(s->flags & SLAB_STORE_USER))
819 return;
820
821 n = get_node(s, page_to_nid(page));
822
823 spin_lock(&n->list_lock);
824 list_del(&page->lru);
825 spin_unlock(&n->list_lock);
826}
827
Christoph Lameter3ec09742007-05-16 22:11:00 -0700828static void setup_object_debug(struct kmem_cache *s, struct page *page,
829 void *object)
830{
831 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
832 return;
833
834 init_object(s, object, 0);
835 init_tracking(s, object);
836}
837
838static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
839 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700840{
841 if (!check_slab(s, page))
842 goto bad;
843
844 if (object && !on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700845 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700846 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700847 }
848
849 if (!check_valid_pointer(s, page, object)) {
850 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700851 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700852 }
853
Christoph Lameter3ec09742007-05-16 22:11:00 -0700854 if (object && !check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700855 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700856
Christoph Lameter3ec09742007-05-16 22:11:00 -0700857 /* Success perform special debug activities for allocs */
858 if (s->flags & SLAB_STORE_USER)
859 set_track(s, object, TRACK_ALLOC, addr);
860 trace(s, page, object, 1);
861 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700862 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700863
Christoph Lameter81819f02007-05-06 14:49:36 -0700864bad:
865 if (PageSlab(page)) {
866 /*
867 * If this is a slab page then lets do the best we can
868 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700869 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700870 */
Christoph Lameter24922682007-07-17 04:03:18 -0700871 slab_fix(s, "Marking all objects used");
Christoph Lameter81819f02007-05-06 14:49:36 -0700872 page->inuse = s->objects;
873 page->freelist = NULL;
874 /* Fix up fields that may be corrupted */
875 page->offset = s->offset / sizeof(void *);
876 }
877 return 0;
878}
879
Christoph Lameter3ec09742007-05-16 22:11:00 -0700880static int free_debug_processing(struct kmem_cache *s, struct page *page,
881 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700882{
883 if (!check_slab(s, page))
884 goto fail;
885
886 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700887 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700888 goto fail;
889 }
890
891 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700892 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700893 goto fail;
894 }
895
896 if (!check_object(s, page, object, 1))
897 return 0;
898
899 if (unlikely(s != page->slab)) {
900 if (!PageSlab(page))
Christoph Lameter70d71222007-05-06 14:49:47 -0700901 slab_err(s, page, "Attempt to free object(0x%p) "
902 "outside of slab", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700903 else
Christoph Lameter70d71222007-05-06 14:49:47 -0700904 if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700905 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700906 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700907 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700908 dump_stack();
909 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700910 else
Christoph Lameter24922682007-07-17 04:03:18 -0700911 object_err(s, page, object,
912 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700913 goto fail;
914 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700915
916 /* Special debug activities for freeing objects */
917 if (!SlabFrozen(page) && !page->freelist)
918 remove_full(s, page);
919 if (s->flags & SLAB_STORE_USER)
920 set_track(s, object, TRACK_FREE, addr);
921 trace(s, page, object, 0);
922 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700923 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700924
Christoph Lameter81819f02007-05-06 14:49:36 -0700925fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700926 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700927 return 0;
928}
929
Christoph Lameter41ecc552007-05-09 02:32:44 -0700930static int __init setup_slub_debug(char *str)
931{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700932 slub_debug = DEBUG_DEFAULT_FLAGS;
933 if (*str++ != '=' || !*str)
934 /*
935 * No options specified. Switch on full debugging.
936 */
937 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700938
939 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700940 /*
941 * No options but restriction on slabs. This means full
942 * debugging for slabs matching a pattern.
943 */
944 goto check_slabs;
945
946 slub_debug = 0;
947 if (*str == '-')
948 /*
949 * Switch off all debugging measures.
950 */
951 goto out;
952
953 /*
954 * Determine which debug features should be switched on
955 */
956 for ( ;*str && *str != ','; str++) {
957 switch (tolower(*str)) {
958 case 'f':
959 slub_debug |= SLAB_DEBUG_FREE;
960 break;
961 case 'z':
962 slub_debug |= SLAB_RED_ZONE;
963 break;
964 case 'p':
965 slub_debug |= SLAB_POISON;
966 break;
967 case 'u':
968 slub_debug |= SLAB_STORE_USER;
969 break;
970 case 't':
971 slub_debug |= SLAB_TRACE;
972 break;
973 default:
974 printk(KERN_ERR "slub_debug option '%c' "
975 "unknown. skipped\n",*str);
976 }
977 }
978
979check_slabs:
980 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -0700981 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700982out:
Christoph Lameter41ecc552007-05-09 02:32:44 -0700983 return 1;
984}
985
986__setup("slub_debug", setup_slub_debug);
987
988static void kmem_cache_open_debug_check(struct kmem_cache *s)
989{
990 /*
991 * The page->offset field is only 16 bit wide. This is an offset
992 * in units of words from the beginning of an object. If the slab
993 * size is bigger then we cannot move the free pointer behind the
994 * object anymore.
995 *
996 * On 32 bit platforms the limit is 256k. On 64bit platforms
997 * the limit is 512k.
998 *
Christoph Lameterc59def92007-05-16 22:10:50 -0700999 * Debugging or ctor may create a need to move the free
Christoph Lameter41ecc552007-05-09 02:32:44 -07001000 * pointer. Fail if this happens.
1001 */
Christoph Lameter33e9e242007-05-23 13:57:56 -07001002 if (s->objsize >= 65535 * sizeof(void *)) {
Christoph Lameter41ecc552007-05-09 02:32:44 -07001003 BUG_ON(s->flags & (SLAB_RED_ZONE | SLAB_POISON |
1004 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
Christoph Lameterc59def92007-05-16 22:10:50 -07001005 BUG_ON(s->ctor);
Christoph Lameter41ecc552007-05-09 02:32:44 -07001006 }
1007 else
1008 /*
1009 * Enable debugging if selected on the kernel commandline.
1010 */
1011 if (slub_debug && (!slub_debug_slabs ||
1012 strncmp(slub_debug_slabs, s->name,
1013 strlen(slub_debug_slabs)) == 0))
1014 s->flags |= slub_debug;
1015}
1016#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001017static inline void setup_object_debug(struct kmem_cache *s,
1018 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001019
Christoph Lameter3ec09742007-05-16 22:11:00 -07001020static inline int alloc_debug_processing(struct kmem_cache *s,
1021 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001022
Christoph Lameter3ec09742007-05-16 22:11:00 -07001023static inline int free_debug_processing(struct kmem_cache *s,
1024 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001025
Christoph Lameter41ecc552007-05-09 02:32:44 -07001026static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1027 { return 1; }
1028static inline int check_object(struct kmem_cache *s, struct page *page,
1029 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001030static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001031static inline void kmem_cache_open_debug_check(struct kmem_cache *s) {}
1032#define slub_debug 0
1033#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001034/*
1035 * Slab allocation and freeing
1036 */
1037static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1038{
1039 struct page * page;
1040 int pages = 1 << s->order;
1041
1042 if (s->order)
1043 flags |= __GFP_COMP;
1044
1045 if (s->flags & SLAB_CACHE_DMA)
1046 flags |= SLUB_DMA;
1047
1048 if (node == -1)
1049 page = alloc_pages(flags, s->order);
1050 else
1051 page = alloc_pages_node(node, flags, s->order);
1052
1053 if (!page)
1054 return NULL;
1055
1056 mod_zone_page_state(page_zone(page),
1057 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1058 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1059 pages);
1060
1061 return page;
1062}
1063
1064static void setup_object(struct kmem_cache *s, struct page *page,
1065 void *object)
1066{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001067 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001068 if (unlikely(s->ctor))
Christoph Lametera35afb82007-05-16 22:10:57 -07001069 s->ctor(object, s, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -07001070}
1071
1072static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1073{
1074 struct page *page;
1075 struct kmem_cache_node *n;
1076 void *start;
1077 void *end;
1078 void *last;
1079 void *p;
1080
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001081 BUG_ON(flags & ~(GFP_DMA | __GFP_ZERO | GFP_LEVEL_MASK));
Christoph Lameter81819f02007-05-06 14:49:36 -07001082
1083 if (flags & __GFP_WAIT)
1084 local_irq_enable();
1085
1086 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
1087 if (!page)
1088 goto out;
1089
1090 n = get_node(s, page_to_nid(page));
1091 if (n)
1092 atomic_long_inc(&n->nr_slabs);
1093 page->offset = s->offset / sizeof(void *);
1094 page->slab = s;
1095 page->flags |= 1 << PG_slab;
1096 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1097 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001098 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001099
1100 start = page_address(page);
1101 end = start + s->objects * s->size;
1102
1103 if (unlikely(s->flags & SLAB_POISON))
1104 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1105
1106 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001107 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001108 setup_object(s, page, last);
1109 set_freepointer(s, last, p);
1110 last = p;
1111 }
1112 setup_object(s, page, last);
1113 set_freepointer(s, last, NULL);
1114
1115 page->freelist = start;
Christoph Lameter894b8782007-05-10 03:15:16 -07001116 page->lockless_freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001117 page->inuse = 0;
1118out:
1119 if (flags & __GFP_WAIT)
1120 local_irq_disable();
1121 return page;
1122}
1123
1124static void __free_slab(struct kmem_cache *s, struct page *page)
1125{
1126 int pages = 1 << s->order;
1127
Christoph Lameterc59def92007-05-16 22:10:50 -07001128 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001129 void *p;
1130
1131 slab_pad_check(s, page);
Christoph Lameterc59def92007-05-16 22:10:50 -07001132 for_each_object(p, s, page_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001133 check_object(s, page, p, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -07001134 }
1135
1136 mod_zone_page_state(page_zone(page),
1137 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1138 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1139 - pages);
1140
1141 page->mapping = NULL;
1142 __free_pages(page, s->order);
1143}
1144
1145static void rcu_free_slab(struct rcu_head *h)
1146{
1147 struct page *page;
1148
1149 page = container_of((struct list_head *)h, struct page, lru);
1150 __free_slab(page->slab, page);
1151}
1152
1153static void free_slab(struct kmem_cache *s, struct page *page)
1154{
1155 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1156 /*
1157 * RCU free overloads the RCU head over the LRU
1158 */
1159 struct rcu_head *head = (void *)&page->lru;
1160
1161 call_rcu(head, rcu_free_slab);
1162 } else
1163 __free_slab(s, page);
1164}
1165
1166static void discard_slab(struct kmem_cache *s, struct page *page)
1167{
1168 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1169
1170 atomic_long_dec(&n->nr_slabs);
1171 reset_page_mapcount(page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001172 ClearSlabDebug(page);
1173 __ClearPageSlab(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001174 free_slab(s, page);
1175}
1176
1177/*
1178 * Per slab locking using the pagelock
1179 */
1180static __always_inline void slab_lock(struct page *page)
1181{
1182 bit_spin_lock(PG_locked, &page->flags);
1183}
1184
1185static __always_inline void slab_unlock(struct page *page)
1186{
1187 bit_spin_unlock(PG_locked, &page->flags);
1188}
1189
1190static __always_inline int slab_trylock(struct page *page)
1191{
1192 int rc = 1;
1193
1194 rc = bit_spin_trylock(PG_locked, &page->flags);
1195 return rc;
1196}
1197
1198/*
1199 * Management of partially allocated slabs
1200 */
Christoph Lametere95eed52007-05-06 14:49:44 -07001201static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001202{
Christoph Lametere95eed52007-05-06 14:49:44 -07001203 spin_lock(&n->list_lock);
1204 n->nr_partial++;
1205 list_add_tail(&page->lru, &n->partial);
1206 spin_unlock(&n->list_lock);
1207}
Christoph Lameter81819f02007-05-06 14:49:36 -07001208
Christoph Lametere95eed52007-05-06 14:49:44 -07001209static void add_partial(struct kmem_cache_node *n, struct page *page)
1210{
Christoph Lameter81819f02007-05-06 14:49:36 -07001211 spin_lock(&n->list_lock);
1212 n->nr_partial++;
1213 list_add(&page->lru, &n->partial);
1214 spin_unlock(&n->list_lock);
1215}
1216
1217static void remove_partial(struct kmem_cache *s,
1218 struct page *page)
1219{
1220 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1221
1222 spin_lock(&n->list_lock);
1223 list_del(&page->lru);
1224 n->nr_partial--;
1225 spin_unlock(&n->list_lock);
1226}
1227
1228/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001229 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001230 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001231 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001232 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001233static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001234{
1235 if (slab_trylock(page)) {
1236 list_del(&page->lru);
1237 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001238 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001239 return 1;
1240 }
1241 return 0;
1242}
1243
1244/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001245 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001246 */
1247static struct page *get_partial_node(struct kmem_cache_node *n)
1248{
1249 struct page *page;
1250
1251 /*
1252 * Racy check. If we mistakenly see no partial slabs then we
1253 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001254 * partial slab and there is none available then get_partials()
1255 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001256 */
1257 if (!n || !n->nr_partial)
1258 return NULL;
1259
1260 spin_lock(&n->list_lock);
1261 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001262 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001263 goto out;
1264 page = NULL;
1265out:
1266 spin_unlock(&n->list_lock);
1267 return page;
1268}
1269
1270/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001271 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001272 */
1273static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1274{
1275#ifdef CONFIG_NUMA
1276 struct zonelist *zonelist;
1277 struct zone **z;
1278 struct page *page;
1279
1280 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001281 * The defrag ratio allows a configuration of the tradeoffs between
1282 * inter node defragmentation and node local allocations. A lower
1283 * defrag_ratio increases the tendency to do local allocations
1284 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001285 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001286 * If the defrag_ratio is set to 0 then kmalloc() always
1287 * returns node local objects. If the ratio is higher then kmalloc()
1288 * may return off node objects because partial slabs are obtained
1289 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001290 *
1291 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001292 * defrag_ratio = 1000) then every (well almost) allocation will
1293 * first attempt to defrag slab caches on other nodes. This means
1294 * scanning over all nodes to look for partial slabs which may be
1295 * expensive if we do it every time we are trying to find a slab
1296 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001297 */
1298 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1299 return NULL;
1300
1301 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1302 ->node_zonelists[gfp_zone(flags)];
1303 for (z = zonelist->zones; *z; z++) {
1304 struct kmem_cache_node *n;
1305
1306 n = get_node(s, zone_to_nid(*z));
1307
1308 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001309 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001310 page = get_partial_node(n);
1311 if (page)
1312 return page;
1313 }
1314 }
1315#endif
1316 return NULL;
1317}
1318
1319/*
1320 * Get a partial page, lock it and return it.
1321 */
1322static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1323{
1324 struct page *page;
1325 int searchnode = (node == -1) ? numa_node_id() : node;
1326
1327 page = get_partial_node(get_node(s, searchnode));
1328 if (page || (flags & __GFP_THISNODE))
1329 return page;
1330
1331 return get_any_partial(s, flags);
1332}
1333
1334/*
1335 * Move a page back to the lists.
1336 *
1337 * Must be called with the slab lock held.
1338 *
1339 * On exit the slab lock will have been dropped.
1340 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001341static void unfreeze_slab(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001342{
Christoph Lametere95eed52007-05-06 14:49:44 -07001343 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1344
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001345 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001346 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001347
Christoph Lameter81819f02007-05-06 14:49:36 -07001348 if (page->freelist)
Christoph Lametere95eed52007-05-06 14:49:44 -07001349 add_partial(n, page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001350 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
Christoph Lametere95eed52007-05-06 14:49:44 -07001351 add_full(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001352 slab_unlock(page);
Christoph Lametere95eed52007-05-06 14:49:44 -07001353
Christoph Lameter81819f02007-05-06 14:49:36 -07001354 } else {
Christoph Lametere95eed52007-05-06 14:49:44 -07001355 if (n->nr_partial < MIN_PARTIAL) {
1356 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001357 * Adding an empty slab to the partial slabs in order
1358 * to avoid page allocator overhead. This slab needs
1359 * to come after the other slabs with objects in
1360 * order to fill them up. That way the size of the
1361 * partial list stays small. kmem_cache_shrink can
1362 * reclaim empty slabs from the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001363 */
1364 add_partial_tail(n, page);
1365 slab_unlock(page);
1366 } else {
1367 slab_unlock(page);
1368 discard_slab(s, page);
1369 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001370 }
1371}
1372
1373/*
1374 * Remove the cpu slab
1375 */
1376static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1377{
Christoph Lameter894b8782007-05-10 03:15:16 -07001378 /*
1379 * Merge cpu freelist into freelist. Typically we get here
1380 * because both freelists are empty. So this is unlikely
1381 * to occur.
1382 */
1383 while (unlikely(page->lockless_freelist)) {
1384 void **object;
1385
1386 /* Retrieve object from cpu_freelist */
1387 object = page->lockless_freelist;
1388 page->lockless_freelist = page->lockless_freelist[page->offset];
1389
1390 /* And put onto the regular freelist */
1391 object[page->offset] = page->freelist;
1392 page->freelist = object;
1393 page->inuse--;
1394 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001395 s->cpu_slab[cpu] = NULL;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001396 unfreeze_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001397}
1398
Christoph Lameter0c710012007-07-17 04:03:24 -07001399static inline void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001400{
1401 slab_lock(page);
1402 deactivate_slab(s, page, cpu);
1403}
1404
1405/*
1406 * Flush cpu slab.
1407 * Called from IPI handler with interrupts disabled.
1408 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001409static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001410{
1411 struct page *page = s->cpu_slab[cpu];
1412
1413 if (likely(page))
1414 flush_slab(s, page, cpu);
1415}
1416
1417static void flush_cpu_slab(void *d)
1418{
1419 struct kmem_cache *s = d;
1420 int cpu = smp_processor_id();
1421
1422 __flush_cpu_slab(s, cpu);
1423}
1424
1425static void flush_all(struct kmem_cache *s)
1426{
1427#ifdef CONFIG_SMP
1428 on_each_cpu(flush_cpu_slab, s, 1, 1);
1429#else
1430 unsigned long flags;
1431
1432 local_irq_save(flags);
1433 flush_cpu_slab(s);
1434 local_irq_restore(flags);
1435#endif
1436}
1437
1438/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001439 * Slow path. The lockless freelist is empty or we need to perform
1440 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001441 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001442 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001443 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001444 * Processing is still very fast if new objects have been freed to the
1445 * regular freelist. In that case we simply take over the regular freelist
1446 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001447 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001448 * If that is not working then we fall back to the partial lists. We take the
1449 * first element of the freelist as the object to allocate now and move the
1450 * rest of the freelist to the lockless freelist.
1451 *
1452 * And if we were unable to get a new slab from the partial slab lists then
1453 * we need to allocate a new slab. This is slowest path since we may sleep.
Christoph Lameter81819f02007-05-06 14:49:36 -07001454 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001455static void *__slab_alloc(struct kmem_cache *s,
1456 gfp_t gfpflags, int node, void *addr, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001457{
Christoph Lameter81819f02007-05-06 14:49:36 -07001458 void **object;
Christoph Lameter894b8782007-05-10 03:15:16 -07001459 int cpu = smp_processor_id();
Christoph Lameter81819f02007-05-06 14:49:36 -07001460
Christoph Lameter81819f02007-05-06 14:49:36 -07001461 if (!page)
1462 goto new_slab;
1463
1464 slab_lock(page);
1465 if (unlikely(node != -1 && page_to_nid(page) != node))
1466 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001467load_freelist:
Christoph Lameter81819f02007-05-06 14:49:36 -07001468 object = page->freelist;
1469 if (unlikely(!object))
1470 goto another_slab;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001471 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001472 goto debug;
1473
Christoph Lameter894b8782007-05-10 03:15:16 -07001474 object = page->freelist;
1475 page->lockless_freelist = object[page->offset];
1476 page->inuse = s->objects;
1477 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001478 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001479 return object;
1480
1481another_slab:
1482 deactivate_slab(s, page, cpu);
1483
1484new_slab:
1485 page = get_partial(s, gfpflags, node);
Christoph Lameter894b8782007-05-10 03:15:16 -07001486 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001487 s->cpu_slab[cpu] = page;
Christoph Lameter894b8782007-05-10 03:15:16 -07001488 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001489 }
1490
1491 page = new_slab(s, gfpflags, node);
1492 if (page) {
1493 cpu = smp_processor_id();
1494 if (s->cpu_slab[cpu]) {
1495 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001496 * Someone else populated the cpu_slab while we
1497 * enabled interrupts, or we have gotten scheduled
1498 * on another cpu. The page may not be on the
1499 * requested node even if __GFP_THISNODE was
1500 * specified. So we need to recheck.
Christoph Lameter81819f02007-05-06 14:49:36 -07001501 */
1502 if (node == -1 ||
1503 page_to_nid(s->cpu_slab[cpu]) == node) {
1504 /*
1505 * Current cpuslab is acceptable and we
1506 * want the current one since its cache hot
1507 */
1508 discard_slab(s, page);
1509 page = s->cpu_slab[cpu];
1510 slab_lock(page);
Christoph Lameter894b8782007-05-10 03:15:16 -07001511 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001512 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001513 /* New slab does not fit our expectations */
Christoph Lameter81819f02007-05-06 14:49:36 -07001514 flush_slab(s, s->cpu_slab[cpu], cpu);
1515 }
1516 slab_lock(page);
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001517 SetSlabFrozen(page);
1518 s->cpu_slab[cpu] = page;
1519 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001520 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001521 return NULL;
1522debug:
Christoph Lameter894b8782007-05-10 03:15:16 -07001523 object = page->freelist;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001524 if (!alloc_debug_processing(s, page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001525 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001526
1527 page->inuse++;
1528 page->freelist = object[page->offset];
1529 slab_unlock(page);
1530 return object;
1531}
1532
1533/*
1534 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1535 * have the fastpath folded into their functions. So no function call
1536 * overhead for requests that can be satisfied on the fastpath.
1537 *
1538 * The fastpath works by first checking if the lockless freelist can be used.
1539 * If not then __slab_alloc is called for slow processing.
1540 *
1541 * Otherwise we can simply pick the next object from the lockless free list.
1542 */
1543static void __always_inline *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001544 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001545{
1546 struct page *page;
1547 void **object;
1548 unsigned long flags;
1549
1550 local_irq_save(flags);
1551 page = s->cpu_slab[smp_processor_id()];
1552 if (unlikely(!page || !page->lockless_freelist ||
1553 (node != -1 && page_to_nid(page) != node)))
1554
1555 object = __slab_alloc(s, gfpflags, node, addr, page);
1556
1557 else {
1558 object = page->lockless_freelist;
1559 page->lockless_freelist = object[page->offset];
1560 }
1561 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001562
1563 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameterce15fea2007-07-17 04:03:28 -07001564 memset(object, 0, s->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001565
Christoph Lameter894b8782007-05-10 03:15:16 -07001566 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001567}
1568
1569void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1570{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001571 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001572}
1573EXPORT_SYMBOL(kmem_cache_alloc);
1574
1575#ifdef CONFIG_NUMA
1576void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1577{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001578 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001579}
1580EXPORT_SYMBOL(kmem_cache_alloc_node);
1581#endif
1582
1583/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001584 * Slow patch handling. This may still be called frequently since objects
1585 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001586 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001587 * So we still attempt to reduce cache line usage. Just take the slab
1588 * lock and free the item. If there is no additional partial page
1589 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001590 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001591static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001592 void *x, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001593{
1594 void *prior;
1595 void **object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07001596
Christoph Lameter81819f02007-05-06 14:49:36 -07001597 slab_lock(page);
1598
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001599 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001600 goto debug;
1601checks_ok:
1602 prior = object[page->offset] = page->freelist;
1603 page->freelist = object;
1604 page->inuse--;
1605
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001606 if (unlikely(SlabFrozen(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001607 goto out_unlock;
1608
1609 if (unlikely(!page->inuse))
1610 goto slab_empty;
1611
1612 /*
1613 * Objects left in the slab. If it
1614 * was not on the partial list before
1615 * then add it.
1616 */
1617 if (unlikely(!prior))
Christoph Lametere95eed52007-05-06 14:49:44 -07001618 add_partial(get_node(s, page_to_nid(page)), page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001619
1620out_unlock:
1621 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001622 return;
1623
1624slab_empty:
1625 if (prior)
1626 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001627 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001628 */
1629 remove_partial(s, page);
1630
1631 slab_unlock(page);
1632 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001633 return;
1634
1635debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001636 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001637 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001638 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001639}
1640
Christoph Lameter894b8782007-05-10 03:15:16 -07001641/*
1642 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1643 * can perform fastpath freeing without additional function calls.
1644 *
1645 * The fastpath is only possible if we are freeing to the current cpu slab
1646 * of this processor. This typically the case if we have just allocated
1647 * the item before.
1648 *
1649 * If fastpath is not possible then fall back to __slab_free where we deal
1650 * with all sorts of special processing.
1651 */
1652static void __always_inline slab_free(struct kmem_cache *s,
1653 struct page *page, void *x, void *addr)
1654{
1655 void **object = (void *)x;
1656 unsigned long flags;
1657
1658 local_irq_save(flags);
1659 if (likely(page == s->cpu_slab[smp_processor_id()] &&
1660 !SlabDebug(page))) {
1661 object[page->offset] = page->lockless_freelist;
1662 page->lockless_freelist = object;
1663 } else
1664 __slab_free(s, page, x, addr);
1665
1666 local_irq_restore(flags);
1667}
1668
Christoph Lameter81819f02007-05-06 14:49:36 -07001669void kmem_cache_free(struct kmem_cache *s, void *x)
1670{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001671 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001672
Christoph Lameterb49af682007-05-06 14:49:41 -07001673 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001674
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001675 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001676}
1677EXPORT_SYMBOL(kmem_cache_free);
1678
1679/* Figure out on which slab object the object resides */
1680static struct page *get_object_page(const void *x)
1681{
Christoph Lameterb49af682007-05-06 14:49:41 -07001682 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001683
1684 if (!PageSlab(page))
1685 return NULL;
1686
1687 return page;
1688}
1689
1690/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001691 * Object placement in a slab is made very easy because we always start at
1692 * offset 0. If we tune the size of the object to the alignment then we can
1693 * get the required alignment by putting one properly sized object after
1694 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001695 *
1696 * Notice that the allocation order determines the sizes of the per cpu
1697 * caches. Each processor has always one slab available for allocations.
1698 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001699 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001700 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001701 */
1702
1703/*
1704 * Mininum / Maximum order of slab pages. This influences locking overhead
1705 * and slab fragmentation. A higher order reduces the number of partial slabs
1706 * and increases the number of allocations possible without having to
1707 * take the list_lock.
1708 */
1709static int slub_min_order;
1710static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001711static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1712
1713/*
1714 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001715 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001716 */
1717static int slub_nomerge;
1718
1719/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001720 * Calculate the order of allocation given an slab object size.
1721 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001722 * The order of allocation has significant impact on performance and other
1723 * system components. Generally order 0 allocations should be preferred since
1724 * order 0 does not cause fragmentation in the page allocator. Larger objects
1725 * be problematic to put into order 0 slabs because there may be too much
1726 * unused space left. We go to a higher order if more than 1/8th of the slab
1727 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001728 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001729 * In order to reach satisfactory performance we must ensure that a minimum
1730 * number of objects is in one slab. Otherwise we may generate too much
1731 * activity on the partial lists which requires taking the list_lock. This is
1732 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001733 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001734 * slub_max_order specifies the order where we begin to stop considering the
1735 * number of objects in a slab as critical. If we reach slub_max_order then
1736 * we try to keep the page order as low as possible. So we accept more waste
1737 * of space in favor of a small page order.
1738 *
1739 * Higher order allocations also allow the placement of more objects in a
1740 * slab and thereby reduce object handling overhead. If the user has
1741 * requested a higher mininum order then we start with that one instead of
1742 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001743 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001744static inline int slab_order(int size, int min_objects,
1745 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001746{
1747 int order;
1748 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001749 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001750
Christoph Lameter6300ea72007-07-17 04:03:20 -07001751 /*
1752 * If we would create too many object per slab then reduce
1753 * the slab order even if it goes below slub_min_order.
1754 */
1755 while (min_order > 0 &&
1756 (PAGE_SIZE << min_order) >= MAX_OBJECTS_PER_SLAB * size)
1757 min_order--;
1758
1759 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001760 fls(min_objects * size - 1) - PAGE_SHIFT);
1761 order <= max_order; order++) {
1762
Christoph Lameter81819f02007-05-06 14:49:36 -07001763 unsigned long slab_size = PAGE_SIZE << order;
1764
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001765 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001766 continue;
1767
Christoph Lameter81819f02007-05-06 14:49:36 -07001768 rem = slab_size % size;
1769
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001770 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001771 break;
1772
Christoph Lameter6300ea72007-07-17 04:03:20 -07001773 /* If the next size is too high then exit now */
1774 if (slab_size * 2 >= MAX_OBJECTS_PER_SLAB * size)
1775 break;
Christoph Lameter81819f02007-05-06 14:49:36 -07001776 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001777
Christoph Lameter81819f02007-05-06 14:49:36 -07001778 return order;
1779}
1780
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001781static inline int calculate_order(int size)
1782{
1783 int order;
1784 int min_objects;
1785 int fraction;
1786
1787 /*
1788 * Attempt to find best configuration for a slab. This
1789 * works by first attempting to generate a layout with
1790 * the best configuration and backing off gradually.
1791 *
1792 * First we reduce the acceptable waste in a slab. Then
1793 * we reduce the minimum objects required in a slab.
1794 */
1795 min_objects = slub_min_objects;
1796 while (min_objects > 1) {
1797 fraction = 8;
1798 while (fraction >= 4) {
1799 order = slab_order(size, min_objects,
1800 slub_max_order, fraction);
1801 if (order <= slub_max_order)
1802 return order;
1803 fraction /= 2;
1804 }
1805 min_objects /= 2;
1806 }
1807
1808 /*
1809 * We were unable to place multiple objects in a slab. Now
1810 * lets see if we can place a single object there.
1811 */
1812 order = slab_order(size, 1, slub_max_order, 1);
1813 if (order <= slub_max_order)
1814 return order;
1815
1816 /*
1817 * Doh this slab cannot be placed using slub_max_order.
1818 */
1819 order = slab_order(size, 1, MAX_ORDER, 1);
1820 if (order <= MAX_ORDER)
1821 return order;
1822 return -ENOSYS;
1823}
1824
Christoph Lameter81819f02007-05-06 14:49:36 -07001825/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001826 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001827 */
1828static unsigned long calculate_alignment(unsigned long flags,
1829 unsigned long align, unsigned long size)
1830{
1831 /*
1832 * If the user wants hardware cache aligned objects then
1833 * follow that suggestion if the object is sufficiently
1834 * large.
1835 *
1836 * The hardware cache alignment cannot override the
1837 * specified alignment though. If that is greater
1838 * then use it.
1839 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001840 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001841 size > cache_line_size() / 2)
1842 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001843
1844 if (align < ARCH_SLAB_MINALIGN)
1845 return ARCH_SLAB_MINALIGN;
1846
1847 return ALIGN(align, sizeof(void *));
1848}
1849
1850static void init_kmem_cache_node(struct kmem_cache_node *n)
1851{
1852 n->nr_partial = 0;
1853 atomic_long_set(&n->nr_slabs, 0);
1854 spin_lock_init(&n->list_lock);
1855 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001856#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07001857 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001858#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001859}
1860
1861#ifdef CONFIG_NUMA
1862/*
1863 * No kmalloc_node yet so do it by hand. We know that this is the first
1864 * slab on the node for this slabcache. There are no concurrent accesses
1865 * possible.
1866 *
1867 * Note that this function only works on the kmalloc_node_cache
1868 * when allocating for the kmalloc_node_cache.
1869 */
1870static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1871 int node)
1872{
1873 struct page *page;
1874 struct kmem_cache_node *n;
1875
1876 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1877
1878 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001879
1880 BUG_ON(!page);
1881 n = page->freelist;
1882 BUG_ON(!n);
1883 page->freelist = get_freepointer(kmalloc_caches, n);
1884 page->inuse++;
1885 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07001886#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07001887 init_object(kmalloc_caches, n, 1);
1888 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001889#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001890 init_kmem_cache_node(n);
1891 atomic_long_inc(&n->nr_slabs);
Christoph Lametere95eed52007-05-06 14:49:44 -07001892 add_partial(n, page);
Christoph Lameterdbc55fa2007-07-03 09:31:04 -07001893
1894 /*
1895 * new_slab() disables interupts. If we do not reenable interrupts here
1896 * then bootup would continue with interrupts disabled.
1897 */
1898 local_irq_enable();
Christoph Lameter81819f02007-05-06 14:49:36 -07001899 return n;
1900}
1901
1902static void free_kmem_cache_nodes(struct kmem_cache *s)
1903{
1904 int node;
1905
1906 for_each_online_node(node) {
1907 struct kmem_cache_node *n = s->node[node];
1908 if (n && n != &s->local_node)
1909 kmem_cache_free(kmalloc_caches, n);
1910 s->node[node] = NULL;
1911 }
1912}
1913
1914static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1915{
1916 int node;
1917 int local_node;
1918
1919 if (slab_state >= UP)
1920 local_node = page_to_nid(virt_to_page(s));
1921 else
1922 local_node = 0;
1923
1924 for_each_online_node(node) {
1925 struct kmem_cache_node *n;
1926
1927 if (local_node == node)
1928 n = &s->local_node;
1929 else {
1930 if (slab_state == DOWN) {
1931 n = early_kmem_cache_node_alloc(gfpflags,
1932 node);
1933 continue;
1934 }
1935 n = kmem_cache_alloc_node(kmalloc_caches,
1936 gfpflags, node);
1937
1938 if (!n) {
1939 free_kmem_cache_nodes(s);
1940 return 0;
1941 }
1942
1943 }
1944 s->node[node] = n;
1945 init_kmem_cache_node(n);
1946 }
1947 return 1;
1948}
1949#else
1950static void free_kmem_cache_nodes(struct kmem_cache *s)
1951{
1952}
1953
1954static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1955{
1956 init_kmem_cache_node(&s->local_node);
1957 return 1;
1958}
1959#endif
1960
1961/*
1962 * calculate_sizes() determines the order and the distribution of data within
1963 * a slab object.
1964 */
1965static int calculate_sizes(struct kmem_cache *s)
1966{
1967 unsigned long flags = s->flags;
1968 unsigned long size = s->objsize;
1969 unsigned long align = s->align;
1970
1971 /*
1972 * Determine if we can poison the object itself. If the user of
1973 * the slab may touch the object after free or before allocation
1974 * then we should never poison the object itself.
1975 */
1976 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07001977 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07001978 s->flags |= __OBJECT_POISON;
1979 else
1980 s->flags &= ~__OBJECT_POISON;
1981
1982 /*
1983 * Round up object size to the next word boundary. We can only
1984 * place the free pointer at word boundaries and this determines
1985 * the possible location of the free pointer.
1986 */
1987 size = ALIGN(size, sizeof(void *));
1988
Christoph Lameter41ecc552007-05-09 02:32:44 -07001989#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07001990 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001991 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07001992 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07001993 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07001994 */
1995 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1996 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07001997#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001998
1999 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002000 * With that we have determined the number of bytes in actual use
2001 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002002 */
2003 s->inuse = size;
2004
2005 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002006 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002007 /*
2008 * Relocate free pointer after the object if it is not
2009 * permitted to overwrite the first word of the object on
2010 * kmem_cache_free.
2011 *
2012 * This is the case if we do RCU, have a constructor or
2013 * destructor or are poisoning the objects.
2014 */
2015 s->offset = size;
2016 size += sizeof(void *);
2017 }
2018
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002019#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002020 if (flags & SLAB_STORE_USER)
2021 /*
2022 * Need to store information about allocs and frees after
2023 * the object.
2024 */
2025 size += 2 * sizeof(struct track);
2026
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002027 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002028 /*
2029 * Add some empty padding so that we can catch
2030 * overwrites from earlier objects rather than let
2031 * tracking information or the free pointer be
2032 * corrupted if an user writes before the start
2033 * of the object.
2034 */
2035 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002036#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002037
Christoph Lameter81819f02007-05-06 14:49:36 -07002038 /*
2039 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002040 * user specified and the dynamic determination of cache line size
2041 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002042 */
2043 align = calculate_alignment(flags, align, s->objsize);
2044
2045 /*
2046 * SLUB stores one object immediately after another beginning from
2047 * offset 0. In order to align the objects we have to simply size
2048 * each object to conform to the alignment.
2049 */
2050 size = ALIGN(size, align);
2051 s->size = size;
2052
2053 s->order = calculate_order(size);
2054 if (s->order < 0)
2055 return 0;
2056
2057 /*
2058 * Determine the number of objects per slab
2059 */
2060 s->objects = (PAGE_SIZE << s->order) / size;
2061
2062 /*
2063 * Verify that the number of objects is within permitted limits.
2064 * The page->inuse field is only 16 bit wide! So we cannot have
2065 * more than 64k objects per slab.
2066 */
Christoph Lameter6300ea72007-07-17 04:03:20 -07002067 if (!s->objects || s->objects > MAX_OBJECTS_PER_SLAB)
Christoph Lameter81819f02007-05-06 14:49:36 -07002068 return 0;
2069 return 1;
2070
2071}
2072
Christoph Lameter81819f02007-05-06 14:49:36 -07002073static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2074 const char *name, size_t size,
2075 size_t align, unsigned long flags,
Christoph Lameterc59def92007-05-16 22:10:50 -07002076 void (*ctor)(void *, struct kmem_cache *, unsigned long))
Christoph Lameter81819f02007-05-06 14:49:36 -07002077{
2078 memset(s, 0, kmem_size);
2079 s->name = name;
2080 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002081 s->objsize = size;
2082 s->flags = flags;
2083 s->align = align;
Christoph Lameter41ecc552007-05-09 02:32:44 -07002084 kmem_cache_open_debug_check(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002085
2086 if (!calculate_sizes(s))
2087 goto error;
2088
2089 s->refcount = 1;
2090#ifdef CONFIG_NUMA
2091 s->defrag_ratio = 100;
2092#endif
2093
2094 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2095 return 1;
2096error:
2097 if (flags & SLAB_PANIC)
2098 panic("Cannot create slab %s size=%lu realsize=%u "
2099 "order=%u offset=%u flags=%lx\n",
2100 s->name, (unsigned long)size, s->size, s->order,
2101 s->offset, flags);
2102 return 0;
2103}
Christoph Lameter81819f02007-05-06 14:49:36 -07002104
2105/*
2106 * Check if a given pointer is valid
2107 */
2108int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2109{
2110 struct page * page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002111
2112 page = get_object_page(object);
2113
2114 if (!page || s != page->slab)
2115 /* No slab or wrong slab */
2116 return 0;
2117
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002118 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002119 return 0;
2120
2121 /*
2122 * We could also check if the object is on the slabs freelist.
2123 * But this would be too expensive and it seems that the main
2124 * purpose of kmem_ptr_valid is to check if the object belongs
2125 * to a certain slab.
2126 */
2127 return 1;
2128}
2129EXPORT_SYMBOL(kmem_ptr_validate);
2130
2131/*
2132 * Determine the size of a slab object
2133 */
2134unsigned int kmem_cache_size(struct kmem_cache *s)
2135{
2136 return s->objsize;
2137}
2138EXPORT_SYMBOL(kmem_cache_size);
2139
2140const char *kmem_cache_name(struct kmem_cache *s)
2141{
2142 return s->name;
2143}
2144EXPORT_SYMBOL(kmem_cache_name);
2145
2146/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002147 * Attempt to free all slabs on a node. Return the number of slabs we
2148 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002149 */
2150static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2151 struct list_head *list)
2152{
2153 int slabs_inuse = 0;
2154 unsigned long flags;
2155 struct page *page, *h;
2156
2157 spin_lock_irqsave(&n->list_lock, flags);
2158 list_for_each_entry_safe(page, h, list, lru)
2159 if (!page->inuse) {
2160 list_del(&page->lru);
2161 discard_slab(s, page);
2162 } else
2163 slabs_inuse++;
2164 spin_unlock_irqrestore(&n->list_lock, flags);
2165 return slabs_inuse;
2166}
2167
2168/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002169 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002170 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002171static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002172{
2173 int node;
2174
2175 flush_all(s);
2176
2177 /* Attempt to free all objects */
2178 for_each_online_node(node) {
2179 struct kmem_cache_node *n = get_node(s, node);
2180
Christoph Lameter2086d262007-05-06 14:49:46 -07002181 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002182 if (atomic_long_read(&n->nr_slabs))
2183 return 1;
2184 }
2185 free_kmem_cache_nodes(s);
2186 return 0;
2187}
2188
2189/*
2190 * Close a cache and release the kmem_cache structure
2191 * (must be used for caches created using kmem_cache_create)
2192 */
2193void kmem_cache_destroy(struct kmem_cache *s)
2194{
2195 down_write(&slub_lock);
2196 s->refcount--;
2197 if (!s->refcount) {
2198 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002199 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002200 if (kmem_cache_close(s))
2201 WARN_ON(1);
2202 sysfs_slab_remove(s);
2203 kfree(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002204 } else
2205 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002206}
2207EXPORT_SYMBOL(kmem_cache_destroy);
2208
2209/********************************************************************
2210 * Kmalloc subsystem
2211 *******************************************************************/
2212
2213struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
2214EXPORT_SYMBOL(kmalloc_caches);
2215
2216#ifdef CONFIG_ZONE_DMA
2217static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
2218#endif
2219
2220static int __init setup_slub_min_order(char *str)
2221{
2222 get_option (&str, &slub_min_order);
2223
2224 return 1;
2225}
2226
2227__setup("slub_min_order=", setup_slub_min_order);
2228
2229static int __init setup_slub_max_order(char *str)
2230{
2231 get_option (&str, &slub_max_order);
2232
2233 return 1;
2234}
2235
2236__setup("slub_max_order=", setup_slub_max_order);
2237
2238static int __init setup_slub_min_objects(char *str)
2239{
2240 get_option (&str, &slub_min_objects);
2241
2242 return 1;
2243}
2244
2245__setup("slub_min_objects=", setup_slub_min_objects);
2246
2247static int __init setup_slub_nomerge(char *str)
2248{
2249 slub_nomerge = 1;
2250 return 1;
2251}
2252
2253__setup("slub_nomerge", setup_slub_nomerge);
2254
Christoph Lameter81819f02007-05-06 14:49:36 -07002255static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2256 const char *name, int size, gfp_t gfp_flags)
2257{
2258 unsigned int flags = 0;
2259
2260 if (gfp_flags & SLUB_DMA)
2261 flags = SLAB_CACHE_DMA;
2262
2263 down_write(&slub_lock);
2264 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameterc59def92007-05-16 22:10:50 -07002265 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002266 goto panic;
2267
2268 list_add(&s->list, &slab_caches);
2269 up_write(&slub_lock);
2270 if (sysfs_slab_add(s))
2271 goto panic;
2272 return s;
2273
2274panic:
2275 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2276}
2277
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002278#ifdef CONFIG_ZONE_DMA
2279static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2280{
2281 struct kmem_cache *s;
2282 struct kmem_cache *x;
2283 char *text;
2284 size_t realsize;
2285
2286 s = kmalloc_caches_dma[index];
2287 if (s)
2288 return s;
2289
2290 /* Dynamically create dma cache */
2291 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2292 if (!x)
2293 panic("Unable to allocate memory for dma cache\n");
2294
Christoph Lameter7b55f622007-07-17 04:03:27 -07002295 realsize = kmalloc_caches[index].objsize;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002296 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2297 (unsigned int)realsize);
2298 s = create_kmalloc_cache(x, text, realsize, flags);
Christoph Lameterdfce8642007-07-17 04:03:25 -07002299 down_write(&slub_lock);
2300 if (!kmalloc_caches_dma[index]) {
2301 kmalloc_caches_dma[index] = s;
2302 up_write(&slub_lock);
2303 return s;
2304 }
2305 up_write(&slub_lock);
2306 kmem_cache_destroy(s);
2307 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002308}
2309#endif
2310
Christoph Lameterf1b26332007-07-17 04:03:26 -07002311/*
2312 * Conversion table for small slabs sizes / 8 to the index in the
2313 * kmalloc array. This is necessary for slabs < 192 since we have non power
2314 * of two cache sizes there. The size of larger slabs can be determined using
2315 * fls.
2316 */
2317static s8 size_index[24] = {
2318 3, /* 8 */
2319 4, /* 16 */
2320 5, /* 24 */
2321 5, /* 32 */
2322 6, /* 40 */
2323 6, /* 48 */
2324 6, /* 56 */
2325 6, /* 64 */
2326 1, /* 72 */
2327 1, /* 80 */
2328 1, /* 88 */
2329 1, /* 96 */
2330 7, /* 104 */
2331 7, /* 112 */
2332 7, /* 120 */
2333 7, /* 128 */
2334 2, /* 136 */
2335 2, /* 144 */
2336 2, /* 152 */
2337 2, /* 160 */
2338 2, /* 168 */
2339 2, /* 176 */
2340 2, /* 184 */
2341 2 /* 192 */
2342};
2343
Christoph Lameter81819f02007-05-06 14:49:36 -07002344static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2345{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002346 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002347
Christoph Lameterf1b26332007-07-17 04:03:26 -07002348 if (size <= 192) {
2349 if (!size)
2350 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002351
Christoph Lameterf1b26332007-07-17 04:03:26 -07002352 index = size_index[(size - 1) / 8];
2353 } else {
2354 if (size > KMALLOC_MAX_SIZE)
2355 return NULL;
2356
2357 index = fls(size - 1);
2358 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002359
2360#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002361 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002362 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002363
Christoph Lameter81819f02007-05-06 14:49:36 -07002364#endif
2365 return &kmalloc_caches[index];
2366}
2367
2368void *__kmalloc(size_t size, gfp_t flags)
2369{
2370 struct kmem_cache *s = get_slab(size, flags);
2371
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002372 if (ZERO_OR_NULL_PTR(s))
2373 return s;
2374
Christoph Lameterce15fea2007-07-17 04:03:28 -07002375 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002376}
2377EXPORT_SYMBOL(__kmalloc);
2378
2379#ifdef CONFIG_NUMA
2380void *__kmalloc_node(size_t size, gfp_t flags, int node)
2381{
2382 struct kmem_cache *s = get_slab(size, flags);
2383
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002384 if (ZERO_OR_NULL_PTR(s))
2385 return s;
2386
Christoph Lameterce15fea2007-07-17 04:03:28 -07002387 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002388}
2389EXPORT_SYMBOL(__kmalloc_node);
2390#endif
2391
2392size_t ksize(const void *object)
2393{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002394 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002395 struct kmem_cache *s;
2396
Linus Torvalds9550b102007-07-19 13:21:34 -07002397 if (ZERO_OR_NULL_PTR(object))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002398 return 0;
2399
2400 page = get_object_page(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07002401 BUG_ON(!page);
2402 s = page->slab;
2403 BUG_ON(!s);
2404
2405 /*
2406 * Debugging requires use of the padding between object
2407 * and whatever may come after it.
2408 */
2409 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2410 return s->objsize;
2411
2412 /*
2413 * If we have the need to store the freelist pointer
2414 * back there or track user information then we can
2415 * only use the space before that information.
2416 */
2417 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2418 return s->inuse;
2419
2420 /*
2421 * Else we can use all the padding etc for the allocation
2422 */
2423 return s->size;
2424}
2425EXPORT_SYMBOL(ksize);
2426
2427void kfree(const void *x)
2428{
2429 struct kmem_cache *s;
2430 struct page *page;
2431
Christoph Lameter272c1d22007-06-08 13:46:49 -07002432 /*
2433 * This has to be an unsigned comparison. According to Linus
2434 * some gcc version treat a pointer as a signed entity. Then
2435 * this comparison would be true for all "negative" pointers
2436 * (which would cover the whole upper half of the address space).
2437 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002438 if (ZERO_OR_NULL_PTR(x))
Christoph Lameter81819f02007-05-06 14:49:36 -07002439 return;
2440
Christoph Lameterb49af682007-05-06 14:49:41 -07002441 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002442 s = page->slab;
2443
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002444 slab_free(s, page, (void *)x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002445}
2446EXPORT_SYMBOL(kfree);
2447
Christoph Lameter2086d262007-05-06 14:49:46 -07002448/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002449 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2450 * the remaining slabs by the number of items in use. The slabs with the
2451 * most items in use come first. New allocations will then fill those up
2452 * and thus they can be removed from the partial lists.
2453 *
2454 * The slabs with the least items are placed last. This results in them
2455 * being allocated from last increasing the chance that the last objects
2456 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002457 */
2458int kmem_cache_shrink(struct kmem_cache *s)
2459{
2460 int node;
2461 int i;
2462 struct kmem_cache_node *n;
2463 struct page *page;
2464 struct page *t;
2465 struct list_head *slabs_by_inuse =
2466 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2467 unsigned long flags;
2468
2469 if (!slabs_by_inuse)
2470 return -ENOMEM;
2471
2472 flush_all(s);
2473 for_each_online_node(node) {
2474 n = get_node(s, node);
2475
2476 if (!n->nr_partial)
2477 continue;
2478
2479 for (i = 0; i < s->objects; i++)
2480 INIT_LIST_HEAD(slabs_by_inuse + i);
2481
2482 spin_lock_irqsave(&n->list_lock, flags);
2483
2484 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002485 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002486 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002487 * Note that concurrent frees may occur while we hold the
2488 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002489 */
2490 list_for_each_entry_safe(page, t, &n->partial, lru) {
2491 if (!page->inuse && slab_trylock(page)) {
2492 /*
2493 * Must hold slab lock here because slab_free
2494 * may have freed the last object and be
2495 * waiting to release the slab.
2496 */
2497 list_del(&page->lru);
2498 n->nr_partial--;
2499 slab_unlock(page);
2500 discard_slab(s, page);
2501 } else {
2502 if (n->nr_partial > MAX_PARTIAL)
2503 list_move(&page->lru,
2504 slabs_by_inuse + page->inuse);
2505 }
2506 }
2507
2508 if (n->nr_partial <= MAX_PARTIAL)
2509 goto out;
2510
2511 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002512 * Rebuild the partial list with the slabs filled up most
2513 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002514 */
2515 for (i = s->objects - 1; i >= 0; i--)
2516 list_splice(slabs_by_inuse + i, n->partial.prev);
2517
2518 out:
2519 spin_unlock_irqrestore(&n->list_lock, flags);
2520 }
2521
2522 kfree(slabs_by_inuse);
2523 return 0;
2524}
2525EXPORT_SYMBOL(kmem_cache_shrink);
2526
Christoph Lameter81819f02007-05-06 14:49:36 -07002527/********************************************************************
2528 * Basic setup of slabs
2529 *******************************************************************/
2530
2531void __init kmem_cache_init(void)
2532{
2533 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002534 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002535
2536#ifdef CONFIG_NUMA
2537 /*
2538 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002539 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002540 * kmem_cache_open for slab_state == DOWN.
2541 */
2542 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2543 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002544 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002545 caches++;
Christoph Lameter81819f02007-05-06 14:49:36 -07002546#endif
2547
2548 /* Able to allocate the per node structures */
2549 slab_state = PARTIAL;
2550
2551 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002552 if (KMALLOC_MIN_SIZE <= 64) {
2553 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002554 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002555 caches++;
2556 }
2557 if (KMALLOC_MIN_SIZE <= 128) {
2558 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002559 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002560 caches++;
2561 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002562
Christoph Lameter4b356be2007-06-16 10:16:13 -07002563 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002564 create_kmalloc_cache(&kmalloc_caches[i],
2565 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002566 caches++;
2567 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002568
Christoph Lameterf1b26332007-07-17 04:03:26 -07002569
2570 /*
2571 * Patch up the size_index table if we have strange large alignment
2572 * requirements for the kmalloc array. This is only the case for
2573 * mips it seems. The standard arches will not generate any code here.
2574 *
2575 * Largest permitted alignment is 256 bytes due to the way we
2576 * handle the index determination for the smaller caches.
2577 *
2578 * Make sure that nothing crazy happens if someone starts tinkering
2579 * around with ARCH_KMALLOC_MINALIGN
2580 */
2581 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2582 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2583
Christoph Lameter12ad6842007-07-17 04:03:28 -07002584 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002585 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2586
Christoph Lameter81819f02007-05-06 14:49:36 -07002587 slab_state = UP;
2588
2589 /* Provide the correct kmalloc names now that the caches are up */
2590 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2591 kmalloc_caches[i]. name =
2592 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2593
2594#ifdef CONFIG_SMP
2595 register_cpu_notifier(&slab_notifier);
2596#endif
2597
Christoph Lameterbcf889f2007-05-10 03:15:44 -07002598 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2599 nr_cpu_ids * sizeof(struct page *);
Christoph Lameter81819f02007-05-06 14:49:36 -07002600
2601 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002602 " CPUs=%d, Nodes=%d\n",
2603 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002604 slub_min_order, slub_max_order, slub_min_objects,
2605 nr_cpu_ids, nr_node_ids);
2606}
2607
2608/*
2609 * Find a mergeable slab cache
2610 */
2611static int slab_unmergeable(struct kmem_cache *s)
2612{
2613 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2614 return 1;
2615
Christoph Lameterc59def92007-05-16 22:10:50 -07002616 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002617 return 1;
2618
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002619 /*
2620 * We may have set a slab to be unmergeable during bootstrap.
2621 */
2622 if (s->refcount < 0)
2623 return 1;
2624
Christoph Lameter81819f02007-05-06 14:49:36 -07002625 return 0;
2626}
2627
2628static struct kmem_cache *find_mergeable(size_t size,
2629 size_t align, unsigned long flags,
Christoph Lameterc59def92007-05-16 22:10:50 -07002630 void (*ctor)(void *, struct kmem_cache *, unsigned long))
Christoph Lameter81819f02007-05-06 14:49:36 -07002631{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002632 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002633
2634 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2635 return NULL;
2636
Christoph Lameterc59def92007-05-16 22:10:50 -07002637 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002638 return NULL;
2639
2640 size = ALIGN(size, sizeof(void *));
2641 align = calculate_alignment(flags, align, size);
2642 size = ALIGN(size, align);
2643
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002644 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002645 if (slab_unmergeable(s))
2646 continue;
2647
2648 if (size > s->size)
2649 continue;
2650
2651 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2652 (s->flags & SLUB_MERGE_SAME))
2653 continue;
2654 /*
2655 * Check if alignment is compatible.
2656 * Courtesy of Adrian Drzewiecki
2657 */
2658 if ((s->size & ~(align -1)) != s->size)
2659 continue;
2660
2661 if (s->size - size >= sizeof(void *))
2662 continue;
2663
2664 return s;
2665 }
2666 return NULL;
2667}
2668
2669struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2670 size_t align, unsigned long flags,
Paul Mundt20c2df82007-07-20 10:11:58 +09002671 void (*ctor)(void *, struct kmem_cache *, unsigned long))
Christoph Lameter81819f02007-05-06 14:49:36 -07002672{
2673 struct kmem_cache *s;
2674
2675 down_write(&slub_lock);
Christoph Lameterc59def92007-05-16 22:10:50 -07002676 s = find_mergeable(size, align, flags, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002677 if (s) {
2678 s->refcount++;
2679 /*
2680 * Adjust the object sizes so that we clear
2681 * the complete object on kzalloc.
2682 */
2683 s->objsize = max(s->objsize, (int)size);
2684 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002685 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002686 if (sysfs_slab_alias(s, name))
2687 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002688 return s;
2689 }
2690 s = kmalloc(kmem_size, GFP_KERNEL);
2691 if (s) {
2692 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07002693 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002694 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002695 up_write(&slub_lock);
2696 if (sysfs_slab_add(s))
2697 goto err;
2698 return s;
2699 }
2700 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002701 }
2702 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002703
2704err:
Christoph Lameter81819f02007-05-06 14:49:36 -07002705 if (flags & SLAB_PANIC)
2706 panic("Cannot create slabcache %s\n", name);
2707 else
2708 s = NULL;
2709 return s;
2710}
2711EXPORT_SYMBOL(kmem_cache_create);
2712
Christoph Lameter81819f02007-05-06 14:49:36 -07002713#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07002714/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002715 * Use the cpu notifier to insure that the cpu slabs are flushed when
2716 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07002717 */
2718static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2719 unsigned long action, void *hcpu)
2720{
2721 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002722 struct kmem_cache *s;
2723 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002724
2725 switch (action) {
2726 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002727 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07002728 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002729 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002730 down_read(&slub_lock);
2731 list_for_each_entry(s, &slab_caches, list) {
2732 local_irq_save(flags);
2733 __flush_cpu_slab(s, cpu);
2734 local_irq_restore(flags);
2735 }
2736 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002737 break;
2738 default:
2739 break;
2740 }
2741 return NOTIFY_OK;
2742}
2743
2744static struct notifier_block __cpuinitdata slab_notifier =
2745 { &slab_cpuup_callback, NULL, 0 };
2746
2747#endif
2748
Christoph Lameter81819f02007-05-06 14:49:36 -07002749void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2750{
2751 struct kmem_cache *s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002752
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002753 if (ZERO_OR_NULL_PTR(s))
2754 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002755
Christoph Lameterce15fea2007-07-17 04:03:28 -07002756 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07002757}
2758
2759void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2760 int node, void *caller)
2761{
2762 struct kmem_cache *s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002763
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002764 if (ZERO_OR_NULL_PTR(s))
2765 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002766
Christoph Lameterce15fea2007-07-17 04:03:28 -07002767 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07002768}
2769
Christoph Lameter41ecc552007-05-09 02:32:44 -07002770#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07002771static int validate_slab(struct kmem_cache *s, struct page *page,
2772 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07002773{
2774 void *p;
2775 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07002776
2777 if (!check_slab(s, page) ||
2778 !on_freelist(s, page, NULL))
2779 return 0;
2780
2781 /* Now we know that a valid freelist exists */
2782 bitmap_zero(map, s->objects);
2783
Christoph Lameter7656c722007-05-09 02:32:40 -07002784 for_each_free_object(p, s, page->freelist) {
2785 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07002786 if (!check_object(s, page, p, 0))
2787 return 0;
2788 }
2789
Christoph Lameter7656c722007-05-09 02:32:40 -07002790 for_each_object(p, s, addr)
2791 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07002792 if (!check_object(s, page, p, 1))
2793 return 0;
2794 return 1;
2795}
2796
Christoph Lameter434e2452007-07-17 04:03:30 -07002797static void validate_slab_slab(struct kmem_cache *s, struct page *page,
2798 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07002799{
2800 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07002801 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07002802 slab_unlock(page);
2803 } else
2804 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2805 s->name, page);
2806
2807 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07002808 if (!SlabDebug(page))
2809 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07002810 "on slab 0x%p\n", s->name, page);
2811 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07002812 if (SlabDebug(page))
2813 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07002814 "slab 0x%p\n", s->name, page);
2815 }
2816}
2817
Christoph Lameter434e2452007-07-17 04:03:30 -07002818static int validate_slab_node(struct kmem_cache *s,
2819 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07002820{
2821 unsigned long count = 0;
2822 struct page *page;
2823 unsigned long flags;
2824
2825 spin_lock_irqsave(&n->list_lock, flags);
2826
2827 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07002828 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07002829 count++;
2830 }
2831 if (count != n->nr_partial)
2832 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2833 "counter=%ld\n", s->name, count, n->nr_partial);
2834
2835 if (!(s->flags & SLAB_STORE_USER))
2836 goto out;
2837
2838 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07002839 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07002840 count++;
2841 }
2842 if (count != atomic_long_read(&n->nr_slabs))
2843 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2844 "counter=%ld\n", s->name, count,
2845 atomic_long_read(&n->nr_slabs));
2846
2847out:
2848 spin_unlock_irqrestore(&n->list_lock, flags);
2849 return count;
2850}
2851
Christoph Lameter434e2452007-07-17 04:03:30 -07002852static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07002853{
2854 int node;
2855 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07002856 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
2857 sizeof(unsigned long), GFP_KERNEL);
2858
2859 if (!map)
2860 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07002861
2862 flush_all(s);
2863 for_each_online_node(node) {
2864 struct kmem_cache_node *n = get_node(s, node);
2865
Christoph Lameter434e2452007-07-17 04:03:30 -07002866 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07002867 }
Christoph Lameter434e2452007-07-17 04:03:30 -07002868 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07002869 return count;
2870}
2871
Christoph Lameterb3459702007-05-09 02:32:41 -07002872#ifdef SLUB_RESILIENCY_TEST
2873static void resiliency_test(void)
2874{
2875 u8 *p;
2876
2877 printk(KERN_ERR "SLUB resiliency testing\n");
2878 printk(KERN_ERR "-----------------------\n");
2879 printk(KERN_ERR "A. Corruption after allocation\n");
2880
2881 p = kzalloc(16, GFP_KERNEL);
2882 p[16] = 0x12;
2883 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2884 " 0x12->0x%p\n\n", p + 16);
2885
2886 validate_slab_cache(kmalloc_caches + 4);
2887
2888 /* Hmmm... The next two are dangerous */
2889 p = kzalloc(32, GFP_KERNEL);
2890 p[32 + sizeof(void *)] = 0x34;
2891 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2892 " 0x34 -> -0x%p\n", p);
2893 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2894
2895 validate_slab_cache(kmalloc_caches + 5);
2896 p = kzalloc(64, GFP_KERNEL);
2897 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2898 *p = 0x56;
2899 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2900 p);
2901 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2902 validate_slab_cache(kmalloc_caches + 6);
2903
2904 printk(KERN_ERR "\nB. Corruption after free\n");
2905 p = kzalloc(128, GFP_KERNEL);
2906 kfree(p);
2907 *p = 0x78;
2908 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2909 validate_slab_cache(kmalloc_caches + 7);
2910
2911 p = kzalloc(256, GFP_KERNEL);
2912 kfree(p);
2913 p[50] = 0x9a;
2914 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2915 validate_slab_cache(kmalloc_caches + 8);
2916
2917 p = kzalloc(512, GFP_KERNEL);
2918 kfree(p);
2919 p[512] = 0xab;
2920 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2921 validate_slab_cache(kmalloc_caches + 9);
2922}
2923#else
2924static void resiliency_test(void) {};
2925#endif
2926
Christoph Lameter88a420e2007-05-06 14:49:45 -07002927/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002928 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07002929 * and freed.
2930 */
2931
2932struct location {
2933 unsigned long count;
2934 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07002935 long long sum_time;
2936 long min_time;
2937 long max_time;
2938 long min_pid;
2939 long max_pid;
2940 cpumask_t cpus;
2941 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07002942};
2943
2944struct loc_track {
2945 unsigned long max;
2946 unsigned long count;
2947 struct location *loc;
2948};
2949
2950static void free_loc_track(struct loc_track *t)
2951{
2952 if (t->max)
2953 free_pages((unsigned long)t->loc,
2954 get_order(sizeof(struct location) * t->max));
2955}
2956
Christoph Lameter68dff6a2007-07-17 04:03:20 -07002957static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07002958{
2959 struct location *l;
2960 int order;
2961
Christoph Lameter88a420e2007-05-06 14:49:45 -07002962 order = get_order(sizeof(struct location) * max);
2963
Christoph Lameter68dff6a2007-07-17 04:03:20 -07002964 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07002965 if (!l)
2966 return 0;
2967
2968 if (t->count) {
2969 memcpy(l, t->loc, sizeof(struct location) * t->count);
2970 free_loc_track(t);
2971 }
2972 t->max = max;
2973 t->loc = l;
2974 return 1;
2975}
2976
2977static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07002978 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07002979{
2980 long start, end, pos;
2981 struct location *l;
2982 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07002983 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07002984
2985 start = -1;
2986 end = t->count;
2987
2988 for ( ; ; ) {
2989 pos = start + (end - start + 1) / 2;
2990
2991 /*
2992 * There is nothing at "end". If we end up there
2993 * we need to add something to before end.
2994 */
2995 if (pos == end)
2996 break;
2997
2998 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07002999 if (track->addr == caddr) {
3000
3001 l = &t->loc[pos];
3002 l->count++;
3003 if (track->when) {
3004 l->sum_time += age;
3005 if (age < l->min_time)
3006 l->min_time = age;
3007 if (age > l->max_time)
3008 l->max_time = age;
3009
3010 if (track->pid < l->min_pid)
3011 l->min_pid = track->pid;
3012 if (track->pid > l->max_pid)
3013 l->max_pid = track->pid;
3014
3015 cpu_set(track->cpu, l->cpus);
3016 }
3017 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003018 return 1;
3019 }
3020
Christoph Lameter45edfa52007-05-09 02:32:45 -07003021 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003022 end = pos;
3023 else
3024 start = pos;
3025 }
3026
3027 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003028 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003029 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003030 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003031 return 0;
3032
3033 l = t->loc + pos;
3034 if (pos < t->count)
3035 memmove(l + 1, l,
3036 (t->count - pos) * sizeof(struct location));
3037 t->count++;
3038 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003039 l->addr = track->addr;
3040 l->sum_time = age;
3041 l->min_time = age;
3042 l->max_time = age;
3043 l->min_pid = track->pid;
3044 l->max_pid = track->pid;
3045 cpus_clear(l->cpus);
3046 cpu_set(track->cpu, l->cpus);
3047 nodes_clear(l->nodes);
3048 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003049 return 1;
3050}
3051
3052static void process_slab(struct loc_track *t, struct kmem_cache *s,
3053 struct page *page, enum track_item alloc)
3054{
3055 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003056 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003057 void *p;
3058
3059 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003060 for_each_free_object(p, s, page->freelist)
3061 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003062
Christoph Lameter7656c722007-05-09 02:32:40 -07003063 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003064 if (!test_bit(slab_index(p, s, addr), map))
3065 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003066}
3067
3068static int list_locations(struct kmem_cache *s, char *buf,
3069 enum track_item alloc)
3070{
3071 int n = 0;
3072 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003073 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003074 int node;
3075
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003076 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3077 GFP_KERNEL))
3078 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003079
3080 /* Push back cpu slabs */
3081 flush_all(s);
3082
3083 for_each_online_node(node) {
3084 struct kmem_cache_node *n = get_node(s, node);
3085 unsigned long flags;
3086 struct page *page;
3087
3088 if (!atomic_read(&n->nr_slabs))
3089 continue;
3090
3091 spin_lock_irqsave(&n->list_lock, flags);
3092 list_for_each_entry(page, &n->partial, lru)
3093 process_slab(&t, s, page, alloc);
3094 list_for_each_entry(page, &n->full, lru)
3095 process_slab(&t, s, page, alloc);
3096 spin_unlock_irqrestore(&n->list_lock, flags);
3097 }
3098
3099 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003100 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003101
3102 if (n > PAGE_SIZE - 100)
3103 break;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003104 n += sprintf(buf + n, "%7ld ", l->count);
3105
3106 if (l->addr)
3107 n += sprint_symbol(buf + n, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003108 else
3109 n += sprintf(buf + n, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003110
3111 if (l->sum_time != l->min_time) {
3112 unsigned long remainder;
3113
3114 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3115 l->min_time,
3116 div_long_long_rem(l->sum_time, l->count, &remainder),
3117 l->max_time);
3118 } else
3119 n += sprintf(buf + n, " age=%ld",
3120 l->min_time);
3121
3122 if (l->min_pid != l->max_pid)
3123 n += sprintf(buf + n, " pid=%ld-%ld",
3124 l->min_pid, l->max_pid);
3125 else
3126 n += sprintf(buf + n, " pid=%ld",
3127 l->min_pid);
3128
Christoph Lameter84966342007-06-23 17:16:32 -07003129 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3130 n < PAGE_SIZE - 60) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003131 n += sprintf(buf + n, " cpus=");
3132 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3133 l->cpus);
3134 }
3135
Christoph Lameter84966342007-06-23 17:16:32 -07003136 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3137 n < PAGE_SIZE - 60) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003138 n += sprintf(buf + n, " nodes=");
3139 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3140 l->nodes);
3141 }
3142
Christoph Lameter88a420e2007-05-06 14:49:45 -07003143 n += sprintf(buf + n, "\n");
3144 }
3145
3146 free_loc_track(&t);
3147 if (!t.count)
3148 n += sprintf(buf, "No data\n");
3149 return n;
3150}
3151
Christoph Lameter81819f02007-05-06 14:49:36 -07003152static unsigned long count_partial(struct kmem_cache_node *n)
3153{
3154 unsigned long flags;
3155 unsigned long x = 0;
3156 struct page *page;
3157
3158 spin_lock_irqsave(&n->list_lock, flags);
3159 list_for_each_entry(page, &n->partial, lru)
3160 x += page->inuse;
3161 spin_unlock_irqrestore(&n->list_lock, flags);
3162 return x;
3163}
3164
3165enum slab_stat_type {
3166 SL_FULL,
3167 SL_PARTIAL,
3168 SL_CPU,
3169 SL_OBJECTS
3170};
3171
3172#define SO_FULL (1 << SL_FULL)
3173#define SO_PARTIAL (1 << SL_PARTIAL)
3174#define SO_CPU (1 << SL_CPU)
3175#define SO_OBJECTS (1 << SL_OBJECTS)
3176
3177static unsigned long slab_objects(struct kmem_cache *s,
3178 char *buf, unsigned long flags)
3179{
3180 unsigned long total = 0;
3181 int cpu;
3182 int node;
3183 int x;
3184 unsigned long *nodes;
3185 unsigned long *per_cpu;
3186
3187 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3188 per_cpu = nodes + nr_node_ids;
3189
3190 for_each_possible_cpu(cpu) {
3191 struct page *page = s->cpu_slab[cpu];
3192 int node;
3193
3194 if (page) {
3195 node = page_to_nid(page);
3196 if (flags & SO_CPU) {
3197 int x = 0;
3198
3199 if (flags & SO_OBJECTS)
3200 x = page->inuse;
3201 else
3202 x = 1;
3203 total += x;
3204 nodes[node] += x;
3205 }
3206 per_cpu[node]++;
3207 }
3208 }
3209
3210 for_each_online_node(node) {
3211 struct kmem_cache_node *n = get_node(s, node);
3212
3213 if (flags & SO_PARTIAL) {
3214 if (flags & SO_OBJECTS)
3215 x = count_partial(n);
3216 else
3217 x = n->nr_partial;
3218 total += x;
3219 nodes[node] += x;
3220 }
3221
3222 if (flags & SO_FULL) {
3223 int full_slabs = atomic_read(&n->nr_slabs)
3224 - per_cpu[node]
3225 - n->nr_partial;
3226
3227 if (flags & SO_OBJECTS)
3228 x = full_slabs * s->objects;
3229 else
3230 x = full_slabs;
3231 total += x;
3232 nodes[node] += x;
3233 }
3234 }
3235
3236 x = sprintf(buf, "%lu", total);
3237#ifdef CONFIG_NUMA
3238 for_each_online_node(node)
3239 if (nodes[node])
3240 x += sprintf(buf + x, " N%d=%lu",
3241 node, nodes[node]);
3242#endif
3243 kfree(nodes);
3244 return x + sprintf(buf + x, "\n");
3245}
3246
3247static int any_slab_objects(struct kmem_cache *s)
3248{
3249 int node;
3250 int cpu;
3251
3252 for_each_possible_cpu(cpu)
3253 if (s->cpu_slab[cpu])
3254 return 1;
3255
3256 for_each_node(node) {
3257 struct kmem_cache_node *n = get_node(s, node);
3258
3259 if (n->nr_partial || atomic_read(&n->nr_slabs))
3260 return 1;
3261 }
3262 return 0;
3263}
3264
3265#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3266#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3267
3268struct slab_attribute {
3269 struct attribute attr;
3270 ssize_t (*show)(struct kmem_cache *s, char *buf);
3271 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3272};
3273
3274#define SLAB_ATTR_RO(_name) \
3275 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3276
3277#define SLAB_ATTR(_name) \
3278 static struct slab_attribute _name##_attr = \
3279 __ATTR(_name, 0644, _name##_show, _name##_store)
3280
Christoph Lameter81819f02007-05-06 14:49:36 -07003281static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3282{
3283 return sprintf(buf, "%d\n", s->size);
3284}
3285SLAB_ATTR_RO(slab_size);
3286
3287static ssize_t align_show(struct kmem_cache *s, char *buf)
3288{
3289 return sprintf(buf, "%d\n", s->align);
3290}
3291SLAB_ATTR_RO(align);
3292
3293static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3294{
3295 return sprintf(buf, "%d\n", s->objsize);
3296}
3297SLAB_ATTR_RO(object_size);
3298
3299static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3300{
3301 return sprintf(buf, "%d\n", s->objects);
3302}
3303SLAB_ATTR_RO(objs_per_slab);
3304
3305static ssize_t order_show(struct kmem_cache *s, char *buf)
3306{
3307 return sprintf(buf, "%d\n", s->order);
3308}
3309SLAB_ATTR_RO(order);
3310
3311static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3312{
3313 if (s->ctor) {
3314 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3315
3316 return n + sprintf(buf + n, "\n");
3317 }
3318 return 0;
3319}
3320SLAB_ATTR_RO(ctor);
3321
Christoph Lameter81819f02007-05-06 14:49:36 -07003322static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3323{
3324 return sprintf(buf, "%d\n", s->refcount - 1);
3325}
3326SLAB_ATTR_RO(aliases);
3327
3328static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3329{
3330 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3331}
3332SLAB_ATTR_RO(slabs);
3333
3334static ssize_t partial_show(struct kmem_cache *s, char *buf)
3335{
3336 return slab_objects(s, buf, SO_PARTIAL);
3337}
3338SLAB_ATTR_RO(partial);
3339
3340static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3341{
3342 return slab_objects(s, buf, SO_CPU);
3343}
3344SLAB_ATTR_RO(cpu_slabs);
3345
3346static ssize_t objects_show(struct kmem_cache *s, char *buf)
3347{
3348 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3349}
3350SLAB_ATTR_RO(objects);
3351
3352static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3353{
3354 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3355}
3356
3357static ssize_t sanity_checks_store(struct kmem_cache *s,
3358 const char *buf, size_t length)
3359{
3360 s->flags &= ~SLAB_DEBUG_FREE;
3361 if (buf[0] == '1')
3362 s->flags |= SLAB_DEBUG_FREE;
3363 return length;
3364}
3365SLAB_ATTR(sanity_checks);
3366
3367static ssize_t trace_show(struct kmem_cache *s, char *buf)
3368{
3369 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3370}
3371
3372static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3373 size_t length)
3374{
3375 s->flags &= ~SLAB_TRACE;
3376 if (buf[0] == '1')
3377 s->flags |= SLAB_TRACE;
3378 return length;
3379}
3380SLAB_ATTR(trace);
3381
3382static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3383{
3384 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3385}
3386
3387static ssize_t reclaim_account_store(struct kmem_cache *s,
3388 const char *buf, size_t length)
3389{
3390 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3391 if (buf[0] == '1')
3392 s->flags |= SLAB_RECLAIM_ACCOUNT;
3393 return length;
3394}
3395SLAB_ATTR(reclaim_account);
3396
3397static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3398{
Christoph Lameter5af60832007-05-06 14:49:56 -07003399 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003400}
3401SLAB_ATTR_RO(hwcache_align);
3402
3403#ifdef CONFIG_ZONE_DMA
3404static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3405{
3406 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3407}
3408SLAB_ATTR_RO(cache_dma);
3409#endif
3410
3411static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3412{
3413 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3414}
3415SLAB_ATTR_RO(destroy_by_rcu);
3416
3417static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3418{
3419 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3420}
3421
3422static ssize_t red_zone_store(struct kmem_cache *s,
3423 const char *buf, size_t length)
3424{
3425 if (any_slab_objects(s))
3426 return -EBUSY;
3427
3428 s->flags &= ~SLAB_RED_ZONE;
3429 if (buf[0] == '1')
3430 s->flags |= SLAB_RED_ZONE;
3431 calculate_sizes(s);
3432 return length;
3433}
3434SLAB_ATTR(red_zone);
3435
3436static ssize_t poison_show(struct kmem_cache *s, char *buf)
3437{
3438 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3439}
3440
3441static ssize_t poison_store(struct kmem_cache *s,
3442 const char *buf, size_t length)
3443{
3444 if (any_slab_objects(s))
3445 return -EBUSY;
3446
3447 s->flags &= ~SLAB_POISON;
3448 if (buf[0] == '1')
3449 s->flags |= SLAB_POISON;
3450 calculate_sizes(s);
3451 return length;
3452}
3453SLAB_ATTR(poison);
3454
3455static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3456{
3457 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3458}
3459
3460static ssize_t store_user_store(struct kmem_cache *s,
3461 const char *buf, size_t length)
3462{
3463 if (any_slab_objects(s))
3464 return -EBUSY;
3465
3466 s->flags &= ~SLAB_STORE_USER;
3467 if (buf[0] == '1')
3468 s->flags |= SLAB_STORE_USER;
3469 calculate_sizes(s);
3470 return length;
3471}
3472SLAB_ATTR(store_user);
3473
Christoph Lameter53e15af2007-05-06 14:49:43 -07003474static ssize_t validate_show(struct kmem_cache *s, char *buf)
3475{
3476 return 0;
3477}
3478
3479static ssize_t validate_store(struct kmem_cache *s,
3480 const char *buf, size_t length)
3481{
Christoph Lameter434e2452007-07-17 04:03:30 -07003482 int ret = -EINVAL;
3483
3484 if (buf[0] == '1') {
3485 ret = validate_slab_cache(s);
3486 if (ret >= 0)
3487 ret = length;
3488 }
3489 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003490}
3491SLAB_ATTR(validate);
3492
Christoph Lameter2086d262007-05-06 14:49:46 -07003493static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3494{
3495 return 0;
3496}
3497
3498static ssize_t shrink_store(struct kmem_cache *s,
3499 const char *buf, size_t length)
3500{
3501 if (buf[0] == '1') {
3502 int rc = kmem_cache_shrink(s);
3503
3504 if (rc)
3505 return rc;
3506 } else
3507 return -EINVAL;
3508 return length;
3509}
3510SLAB_ATTR(shrink);
3511
Christoph Lameter88a420e2007-05-06 14:49:45 -07003512static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3513{
3514 if (!(s->flags & SLAB_STORE_USER))
3515 return -ENOSYS;
3516 return list_locations(s, buf, TRACK_ALLOC);
3517}
3518SLAB_ATTR_RO(alloc_calls);
3519
3520static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3521{
3522 if (!(s->flags & SLAB_STORE_USER))
3523 return -ENOSYS;
3524 return list_locations(s, buf, TRACK_FREE);
3525}
3526SLAB_ATTR_RO(free_calls);
3527
Christoph Lameter81819f02007-05-06 14:49:36 -07003528#ifdef CONFIG_NUMA
3529static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3530{
3531 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3532}
3533
3534static ssize_t defrag_ratio_store(struct kmem_cache *s,
3535 const char *buf, size_t length)
3536{
3537 int n = simple_strtoul(buf, NULL, 10);
3538
3539 if (n < 100)
3540 s->defrag_ratio = n * 10;
3541 return length;
3542}
3543SLAB_ATTR(defrag_ratio);
3544#endif
3545
3546static struct attribute * slab_attrs[] = {
3547 &slab_size_attr.attr,
3548 &object_size_attr.attr,
3549 &objs_per_slab_attr.attr,
3550 &order_attr.attr,
3551 &objects_attr.attr,
3552 &slabs_attr.attr,
3553 &partial_attr.attr,
3554 &cpu_slabs_attr.attr,
3555 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003556 &aliases_attr.attr,
3557 &align_attr.attr,
3558 &sanity_checks_attr.attr,
3559 &trace_attr.attr,
3560 &hwcache_align_attr.attr,
3561 &reclaim_account_attr.attr,
3562 &destroy_by_rcu_attr.attr,
3563 &red_zone_attr.attr,
3564 &poison_attr.attr,
3565 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07003566 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07003567 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07003568 &alloc_calls_attr.attr,
3569 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003570#ifdef CONFIG_ZONE_DMA
3571 &cache_dma_attr.attr,
3572#endif
3573#ifdef CONFIG_NUMA
3574 &defrag_ratio_attr.attr,
3575#endif
3576 NULL
3577};
3578
3579static struct attribute_group slab_attr_group = {
3580 .attrs = slab_attrs,
3581};
3582
3583static ssize_t slab_attr_show(struct kobject *kobj,
3584 struct attribute *attr,
3585 char *buf)
3586{
3587 struct slab_attribute *attribute;
3588 struct kmem_cache *s;
3589 int err;
3590
3591 attribute = to_slab_attr(attr);
3592 s = to_slab(kobj);
3593
3594 if (!attribute->show)
3595 return -EIO;
3596
3597 err = attribute->show(s, buf);
3598
3599 return err;
3600}
3601
3602static ssize_t slab_attr_store(struct kobject *kobj,
3603 struct attribute *attr,
3604 const char *buf, size_t len)
3605{
3606 struct slab_attribute *attribute;
3607 struct kmem_cache *s;
3608 int err;
3609
3610 attribute = to_slab_attr(attr);
3611 s = to_slab(kobj);
3612
3613 if (!attribute->store)
3614 return -EIO;
3615
3616 err = attribute->store(s, buf, len);
3617
3618 return err;
3619}
3620
3621static struct sysfs_ops slab_sysfs_ops = {
3622 .show = slab_attr_show,
3623 .store = slab_attr_store,
3624};
3625
3626static struct kobj_type slab_ktype = {
3627 .sysfs_ops = &slab_sysfs_ops,
3628};
3629
3630static int uevent_filter(struct kset *kset, struct kobject *kobj)
3631{
3632 struct kobj_type *ktype = get_ktype(kobj);
3633
3634 if (ktype == &slab_ktype)
3635 return 1;
3636 return 0;
3637}
3638
3639static struct kset_uevent_ops slab_uevent_ops = {
3640 .filter = uevent_filter,
3641};
3642
Adrian Bunk5af328a2007-07-17 04:03:27 -07003643static decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
Christoph Lameter81819f02007-05-06 14:49:36 -07003644
3645#define ID_STR_LENGTH 64
3646
3647/* Create a unique string id for a slab cache:
3648 * format
3649 * :[flags-]size:[memory address of kmemcache]
3650 */
3651static char *create_unique_id(struct kmem_cache *s)
3652{
3653 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3654 char *p = name;
3655
3656 BUG_ON(!name);
3657
3658 *p++ = ':';
3659 /*
3660 * First flags affecting slabcache operations. We will only
3661 * get here for aliasable slabs so we do not need to support
3662 * too many flags. The flags here must cover all flags that
3663 * are matched during merging to guarantee that the id is
3664 * unique.
3665 */
3666 if (s->flags & SLAB_CACHE_DMA)
3667 *p++ = 'd';
3668 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3669 *p++ = 'a';
3670 if (s->flags & SLAB_DEBUG_FREE)
3671 *p++ = 'F';
3672 if (p != name + 1)
3673 *p++ = '-';
3674 p += sprintf(p, "%07d", s->size);
3675 BUG_ON(p > name + ID_STR_LENGTH - 1);
3676 return name;
3677}
3678
3679static int sysfs_slab_add(struct kmem_cache *s)
3680{
3681 int err;
3682 const char *name;
3683 int unmergeable;
3684
3685 if (slab_state < SYSFS)
3686 /* Defer until later */
3687 return 0;
3688
3689 unmergeable = slab_unmergeable(s);
3690 if (unmergeable) {
3691 /*
3692 * Slabcache can never be merged so we can use the name proper.
3693 * This is typically the case for debug situations. In that
3694 * case we can catch duplicate names easily.
3695 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07003696 sysfs_remove_link(&slab_subsys.kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07003697 name = s->name;
3698 } else {
3699 /*
3700 * Create a unique name for the slab as a target
3701 * for the symlinks.
3702 */
3703 name = create_unique_id(s);
3704 }
3705
3706 kobj_set_kset_s(s, slab_subsys);
3707 kobject_set_name(&s->kobj, name);
3708 kobject_init(&s->kobj);
3709 err = kobject_add(&s->kobj);
3710 if (err)
3711 return err;
3712
3713 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3714 if (err)
3715 return err;
3716 kobject_uevent(&s->kobj, KOBJ_ADD);
3717 if (!unmergeable) {
3718 /* Setup first alias */
3719 sysfs_slab_alias(s, s->name);
3720 kfree(name);
3721 }
3722 return 0;
3723}
3724
3725static void sysfs_slab_remove(struct kmem_cache *s)
3726{
3727 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3728 kobject_del(&s->kobj);
3729}
3730
3731/*
3732 * Need to buffer aliases during bootup until sysfs becomes
3733 * available lest we loose that information.
3734 */
3735struct saved_alias {
3736 struct kmem_cache *s;
3737 const char *name;
3738 struct saved_alias *next;
3739};
3740
Adrian Bunk5af328a2007-07-17 04:03:27 -07003741static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07003742
3743static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3744{
3745 struct saved_alias *al;
3746
3747 if (slab_state == SYSFS) {
3748 /*
3749 * If we have a leftover link then remove it.
3750 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07003751 sysfs_remove_link(&slab_subsys.kobj, name);
3752 return sysfs_create_link(&slab_subsys.kobj,
Christoph Lameter81819f02007-05-06 14:49:36 -07003753 &s->kobj, name);
3754 }
3755
3756 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3757 if (!al)
3758 return -ENOMEM;
3759
3760 al->s = s;
3761 al->name = name;
3762 al->next = alias_list;
3763 alias_list = al;
3764 return 0;
3765}
3766
3767static int __init slab_sysfs_init(void)
3768{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003769 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003770 int err;
3771
3772 err = subsystem_register(&slab_subsys);
3773 if (err) {
3774 printk(KERN_ERR "Cannot register slab subsystem.\n");
3775 return -ENOSYS;
3776 }
3777
Christoph Lameter26a7bd02007-05-09 02:32:39 -07003778 slab_state = SYSFS;
3779
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003780 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07003781 err = sysfs_slab_add(s);
3782 BUG_ON(err);
3783 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003784
3785 while (alias_list) {
3786 struct saved_alias *al = alias_list;
3787
3788 alias_list = alias_list->next;
3789 err = sysfs_slab_alias(al->s, al->name);
3790 BUG_ON(err);
3791 kfree(al);
3792 }
3793
3794 resiliency_test();
3795 return 0;
3796}
3797
3798__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07003799#endif