blob: 474945ecd89d828478fd8048558c2a8d980604b3 [file] [log] [blame]
Christoph Lameter81819f02007-05-06 14:49:36 -07001/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
9 */
10
11#include <linux/mm.h>
12#include <linux/module.h>
13#include <linux/bit_spinlock.h>
14#include <linux/interrupt.h>
15#include <linux/bitops.h>
16#include <linux/slab.h>
17#include <linux/seq_file.h>
18#include <linux/cpu.h>
19#include <linux/cpuset.h>
20#include <linux/mempolicy.h>
21#include <linux/ctype.h>
22#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070023#include <linux/memory.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024
25/*
26 * Lock order:
27 * 1. slab_lock(page)
28 * 2. slab->list_lock
29 *
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
36 *
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
42 *
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
47 * the list lock.
48 *
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
61 *
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
66 *
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
69 *
Christoph Lameter672bba32007-05-09 02:32:39 -070070 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070072 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070073 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070075 *
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
79 *
80 * Overloading of page flags that are otherwise used for LRU management.
81 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070082 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
90 *
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -070094 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -070095 * free objects in addition to the regular freelist
96 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070097 *
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700100 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700101 */
102
Christoph Lameter5577bd82007-05-16 22:10:56 -0700103#define FROZEN (1 << PG_active)
104
105#ifdef CONFIG_SLUB_DEBUG
106#define SLABDEBUG (1 << PG_error)
107#else
108#define SLABDEBUG 0
109#endif
110
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700111static inline int SlabFrozen(struct page *page)
112{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700113 return page->flags & FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700114}
115
116static inline void SetSlabFrozen(struct page *page)
117{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700118 page->flags |= FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700119}
120
121static inline void ClearSlabFrozen(struct page *page)
122{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700123 page->flags &= ~FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700124}
125
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700126static inline int SlabDebug(struct page *page)
127{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700128 return page->flags & SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700129}
130
131static inline void SetSlabDebug(struct page *page)
132{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700133 page->flags |= SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700134}
135
136static inline void ClearSlabDebug(struct page *page)
137{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700138 page->flags &= ~SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700139}
140
Christoph Lameter81819f02007-05-06 14:49:36 -0700141/*
142 * Issues still to be resolved:
143 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
145 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700146 * - Variable sizing of the per node arrays
147 */
148
149/* Enable to test recovery from slab corruption on boot */
150#undef SLUB_RESILIENCY_TEST
151
152#if PAGE_SHIFT <= 12
153
154/*
155 * Small page size. Make sure that we do not fragment memory
156 */
157#define DEFAULT_MAX_ORDER 1
158#define DEFAULT_MIN_OBJECTS 4
159
160#else
161
162/*
163 * Large page machines are customarily able to handle larger
164 * page orders.
165 */
166#define DEFAULT_MAX_ORDER 2
167#define DEFAULT_MIN_OBJECTS 8
168
169#endif
170
171/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
174 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800175#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700176
Christoph Lameter2086d262007-05-06 14:49:46 -0700177/*
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
181 */
182#define MAX_PARTIAL 10
183
Christoph Lameter81819f02007-05-06 14:49:36 -0700184#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700186
Christoph Lameter81819f02007-05-06 14:49:36 -0700187/*
188 * Set of flags that will prevent slab merging
189 */
190#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
192
193#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
194 SLAB_CACHE_DMA)
195
196#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700197#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700198#endif
199
200#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700201#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700202#endif
203
204/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700205#define __OBJECT_POISON 0x80000000 /* Poison object */
206#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700207
Christoph Lameter65c02d42007-05-09 02:32:35 -0700208/* Not all arches define cache_line_size */
209#ifndef cache_line_size
210#define cache_line_size() L1_CACHE_BYTES
211#endif
212
Christoph Lameter81819f02007-05-06 14:49:36 -0700213static int kmem_size = sizeof(struct kmem_cache);
214
215#ifdef CONFIG_SMP
216static struct notifier_block slab_notifier;
217#endif
218
219static enum {
220 DOWN, /* No slab functionality available */
221 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700222 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700223 SYSFS /* Sysfs up */
224} slab_state = DOWN;
225
226/* A list of all slab caches on the system */
227static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700228static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700229
Christoph Lameter02cbc872007-05-09 02:32:43 -0700230/*
231 * Tracking user of a slab.
232 */
233struct track {
234 void *addr; /* Called from address */
235 int cpu; /* Was running on cpu */
236 int pid; /* Pid context */
237 unsigned long when; /* When did the operation occur */
238};
239
240enum track_item { TRACK_ALLOC, TRACK_FREE };
241
Christoph Lameter41ecc552007-05-09 02:32:44 -0700242#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter81819f02007-05-06 14:49:36 -0700243static int sysfs_slab_add(struct kmem_cache *);
244static int sysfs_slab_alias(struct kmem_cache *, const char *);
245static void sysfs_slab_remove(struct kmem_cache *);
246#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700247static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
248static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
249 { return 0; }
250static inline void sysfs_slab_remove(struct kmem_cache *s) {}
Christoph Lameter81819f02007-05-06 14:49:36 -0700251#endif
252
253/********************************************************************
254 * Core slab cache functions
255 *******************************************************************/
256
257int slab_is_available(void)
258{
259 return slab_state >= UP;
260}
261
262static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
263{
264#ifdef CONFIG_NUMA
265 return s->node[node];
266#else
267 return &s->local_node;
268#endif
269}
270
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700271static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
272{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700273#ifdef CONFIG_SMP
274 return s->cpu_slab[cpu];
275#else
276 return &s->cpu_slab;
277#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700278}
279
Christoph Lameter02cbc872007-05-09 02:32:43 -0700280static inline int check_valid_pointer(struct kmem_cache *s,
281 struct page *page, const void *object)
282{
283 void *base;
284
285 if (!object)
286 return 1;
287
288 base = page_address(page);
289 if (object < base || object >= base + s->objects * s->size ||
290 (object - base) % s->size) {
291 return 0;
292 }
293
294 return 1;
295}
296
Christoph Lameter81819f02007-05-06 14:49:36 -0700297/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700298 * Slow version of get and set free pointer.
299 *
300 * This version requires touching the cache lines of kmem_cache which
301 * we avoid to do in the fast alloc free paths. There we obtain the offset
302 * from the page struct.
303 */
304static inline void *get_freepointer(struct kmem_cache *s, void *object)
305{
306 return *(void **)(object + s->offset);
307}
308
309static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
310{
311 *(void **)(object + s->offset) = fp;
312}
313
314/* Loop over all objects in a slab */
315#define for_each_object(__p, __s, __addr) \
316 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
317 __p += (__s)->size)
318
319/* Scan freelist */
320#define for_each_free_object(__p, __s, __free) \
321 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
322
323/* Determine object index from a given position */
324static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
325{
326 return (p - addr) / s->size;
327}
328
Christoph Lameter41ecc552007-05-09 02:32:44 -0700329#ifdef CONFIG_SLUB_DEBUG
330/*
331 * Debug settings:
332 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700333#ifdef CONFIG_SLUB_DEBUG_ON
334static int slub_debug = DEBUG_DEFAULT_FLAGS;
335#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700336static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700337#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700338
339static char *slub_debug_slabs;
340
Christoph Lameter7656c722007-05-09 02:32:40 -0700341/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700342 * Object debugging
343 */
344static void print_section(char *text, u8 *addr, unsigned int length)
345{
346 int i, offset;
347 int newline = 1;
348 char ascii[17];
349
350 ascii[16] = 0;
351
352 for (i = 0; i < length; i++) {
353 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700354 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700355 newline = 0;
356 }
357 printk(" %02x", addr[i]);
358 offset = i % 16;
359 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
360 if (offset == 15) {
361 printk(" %s\n",ascii);
362 newline = 1;
363 }
364 }
365 if (!newline) {
366 i %= 16;
367 while (i < 16) {
368 printk(" ");
369 ascii[i] = ' ';
370 i++;
371 }
372 printk(" %s\n", ascii);
373 }
374}
375
Christoph Lameter81819f02007-05-06 14:49:36 -0700376static struct track *get_track(struct kmem_cache *s, void *object,
377 enum track_item alloc)
378{
379 struct track *p;
380
381 if (s->offset)
382 p = object + s->offset + sizeof(void *);
383 else
384 p = object + s->inuse;
385
386 return p + alloc;
387}
388
389static void set_track(struct kmem_cache *s, void *object,
390 enum track_item alloc, void *addr)
391{
392 struct track *p;
393
394 if (s->offset)
395 p = object + s->offset + sizeof(void *);
396 else
397 p = object + s->inuse;
398
399 p += alloc;
400 if (addr) {
401 p->addr = addr;
402 p->cpu = smp_processor_id();
403 p->pid = current ? current->pid : -1;
404 p->when = jiffies;
405 } else
406 memset(p, 0, sizeof(struct track));
407}
408
Christoph Lameter81819f02007-05-06 14:49:36 -0700409static void init_tracking(struct kmem_cache *s, void *object)
410{
Christoph Lameter24922682007-07-17 04:03:18 -0700411 if (!(s->flags & SLAB_STORE_USER))
412 return;
413
414 set_track(s, object, TRACK_FREE, NULL);
415 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700416}
417
418static void print_track(const char *s, struct track *t)
419{
420 if (!t->addr)
421 return;
422
Christoph Lameter24922682007-07-17 04:03:18 -0700423 printk(KERN_ERR "INFO: %s in ", s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700424 __print_symbol("%s", (unsigned long)t->addr);
Christoph Lameter24922682007-07-17 04:03:18 -0700425 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700426}
427
Christoph Lameter24922682007-07-17 04:03:18 -0700428static void print_tracking(struct kmem_cache *s, void *object)
429{
430 if (!(s->flags & SLAB_STORE_USER))
431 return;
432
433 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
434 print_track("Freed", get_track(s, object, TRACK_FREE));
435}
436
437static void print_page_info(struct page *page)
438{
439 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
440 page, page->inuse, page->freelist, page->flags);
441
442}
443
444static void slab_bug(struct kmem_cache *s, char *fmt, ...)
445{
446 va_list args;
447 char buf[100];
448
449 va_start(args, fmt);
450 vsnprintf(buf, sizeof(buf), fmt, args);
451 va_end(args);
452 printk(KERN_ERR "========================================"
453 "=====================================\n");
454 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
455 printk(KERN_ERR "----------------------------------------"
456 "-------------------------------------\n\n");
457}
458
459static void slab_fix(struct kmem_cache *s, char *fmt, ...)
460{
461 va_list args;
462 char buf[100];
463
464 va_start(args, fmt);
465 vsnprintf(buf, sizeof(buf), fmt, args);
466 va_end(args);
467 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
468}
469
470static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700471{
472 unsigned int off; /* Offset of last byte */
Christoph Lameter24922682007-07-17 04:03:18 -0700473 u8 *addr = page_address(page);
474
475 print_tracking(s, p);
476
477 print_page_info(page);
478
479 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
480 p, p - addr, get_freepointer(s, p));
481
482 if (p > addr + 16)
483 print_section("Bytes b4", p - 16, 16);
484
485 print_section("Object", p, min(s->objsize, 128));
Christoph Lameter81819f02007-05-06 14:49:36 -0700486
487 if (s->flags & SLAB_RED_ZONE)
488 print_section("Redzone", p + s->objsize,
489 s->inuse - s->objsize);
490
Christoph Lameter81819f02007-05-06 14:49:36 -0700491 if (s->offset)
492 off = s->offset + sizeof(void *);
493 else
494 off = s->inuse;
495
Christoph Lameter24922682007-07-17 04:03:18 -0700496 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700497 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700498
499 if (off != s->size)
500 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700501 print_section("Padding", p + off, s->size - off);
502
503 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700504}
505
506static void object_err(struct kmem_cache *s, struct page *page,
507 u8 *object, char *reason)
508{
Christoph Lameter24922682007-07-17 04:03:18 -0700509 slab_bug(s, reason);
510 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700511}
512
Christoph Lameter24922682007-07-17 04:03:18 -0700513static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700514{
515 va_list args;
516 char buf[100];
517
Christoph Lameter24922682007-07-17 04:03:18 -0700518 va_start(args, fmt);
519 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700520 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700521 slab_bug(s, fmt);
522 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700523 dump_stack();
524}
525
526static void init_object(struct kmem_cache *s, void *object, int active)
527{
528 u8 *p = object;
529
530 if (s->flags & __OBJECT_POISON) {
531 memset(p, POISON_FREE, s->objsize - 1);
532 p[s->objsize -1] = POISON_END;
533 }
534
535 if (s->flags & SLAB_RED_ZONE)
536 memset(p + s->objsize,
537 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
538 s->inuse - s->objsize);
539}
540
Christoph Lameter24922682007-07-17 04:03:18 -0700541static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700542{
543 while (bytes) {
544 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700545 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700546 start++;
547 bytes--;
548 }
Christoph Lameter24922682007-07-17 04:03:18 -0700549 return NULL;
550}
551
552static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
553 void *from, void *to)
554{
555 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
556 memset(from, data, to - from);
557}
558
559static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
560 u8 *object, char *what,
561 u8* start, unsigned int value, unsigned int bytes)
562{
563 u8 *fault;
564 u8 *end;
565
566 fault = check_bytes(start, value, bytes);
567 if (!fault)
568 return 1;
569
570 end = start + bytes;
571 while (end > fault && end[-1] == value)
572 end--;
573
574 slab_bug(s, "%s overwritten", what);
575 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
576 fault, end - 1, fault[0], value);
577 print_trailer(s, page, object);
578
579 restore_bytes(s, what, value, fault, end);
580 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700581}
582
Christoph Lameter81819f02007-05-06 14:49:36 -0700583/*
584 * Object layout:
585 *
586 * object address
587 * Bytes of the object to be managed.
588 * If the freepointer may overlay the object then the free
589 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700590 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700591 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
592 * 0xa5 (POISON_END)
593 *
594 * object + s->objsize
595 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700596 * Padding is extended by another word if Redzoning is enabled and
597 * objsize == inuse.
598 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700599 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
600 * 0xcc (RED_ACTIVE) for objects in use.
601 *
602 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700603 * Meta data starts here.
604 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700605 * A. Free pointer (if we cannot overwrite object on free)
606 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700607 * C. Padding to reach required alignment boundary or at mininum
608 * one word if debuggin is on to be able to detect writes
609 * before the word boundary.
610 *
611 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700612 *
613 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700614 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700615 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700616 * If slabcaches are merged then the objsize and inuse boundaries are mostly
617 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700618 * may be used with merged slabcaches.
619 */
620
Christoph Lameter81819f02007-05-06 14:49:36 -0700621static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
622{
623 unsigned long off = s->inuse; /* The end of info */
624
625 if (s->offset)
626 /* Freepointer is placed after the object. */
627 off += sizeof(void *);
628
629 if (s->flags & SLAB_STORE_USER)
630 /* We also have user information there */
631 off += 2 * sizeof(struct track);
632
633 if (s->size == off)
634 return 1;
635
Christoph Lameter24922682007-07-17 04:03:18 -0700636 return check_bytes_and_report(s, page, p, "Object padding",
637 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700638}
639
640static int slab_pad_check(struct kmem_cache *s, struct page *page)
641{
Christoph Lameter24922682007-07-17 04:03:18 -0700642 u8 *start;
643 u8 *fault;
644 u8 *end;
645 int length;
646 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700647
648 if (!(s->flags & SLAB_POISON))
649 return 1;
650
Christoph Lameter24922682007-07-17 04:03:18 -0700651 start = page_address(page);
652 end = start + (PAGE_SIZE << s->order);
Christoph Lameter81819f02007-05-06 14:49:36 -0700653 length = s->objects * s->size;
Christoph Lameter24922682007-07-17 04:03:18 -0700654 remainder = end - (start + length);
Christoph Lameter81819f02007-05-06 14:49:36 -0700655 if (!remainder)
656 return 1;
657
Christoph Lameter24922682007-07-17 04:03:18 -0700658 fault = check_bytes(start + length, POISON_INUSE, remainder);
659 if (!fault)
660 return 1;
661 while (end > fault && end[-1] == POISON_INUSE)
662 end--;
663
664 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
665 print_section("Padding", start, length);
666
667 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
668 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700669}
670
671static int check_object(struct kmem_cache *s, struct page *page,
672 void *object, int active)
673{
674 u8 *p = object;
675 u8 *endobject = object + s->objsize;
676
677 if (s->flags & SLAB_RED_ZONE) {
678 unsigned int red =
679 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
680
Christoph Lameter24922682007-07-17 04:03:18 -0700681 if (!check_bytes_and_report(s, page, object, "Redzone",
682 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700683 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700684 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700685 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse)
686 check_bytes_and_report(s, page, p, "Alignment padding", endobject,
687 POISON_INUSE, s->inuse - s->objsize);
Christoph Lameter81819f02007-05-06 14:49:36 -0700688 }
689
690 if (s->flags & SLAB_POISON) {
691 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700692 (!check_bytes_and_report(s, page, p, "Poison", p,
693 POISON_FREE, s->objsize - 1) ||
694 !check_bytes_and_report(s, page, p, "Poison",
695 p + s->objsize -1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700696 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700697 /*
698 * check_pad_bytes cleans up on its own.
699 */
700 check_pad_bytes(s, page, p);
701 }
702
703 if (!s->offset && active)
704 /*
705 * Object and freepointer overlap. Cannot check
706 * freepointer while object is allocated.
707 */
708 return 1;
709
710 /* Check free pointer validity */
711 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
712 object_err(s, page, p, "Freepointer corrupt");
713 /*
714 * No choice but to zap it and thus loose the remainder
715 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700716 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700717 */
718 set_freepointer(s, p, NULL);
719 return 0;
720 }
721 return 1;
722}
723
724static int check_slab(struct kmem_cache *s, struct page *page)
725{
726 VM_BUG_ON(!irqs_disabled());
727
728 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700729 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700730 return 0;
731 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700732 if (page->inuse > s->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700733 slab_err(s, page, "inuse %u > max %u",
734 s->name, page->inuse, s->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700735 return 0;
736 }
737 /* Slab_pad_check fixes things up after itself */
738 slab_pad_check(s, page);
739 return 1;
740}
741
742/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700743 * Determine if a certain object on a page is on the freelist. Must hold the
744 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700745 */
746static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
747{
748 int nr = 0;
749 void *fp = page->freelist;
750 void *object = NULL;
751
752 while (fp && nr <= s->objects) {
753 if (fp == search)
754 return 1;
755 if (!check_valid_pointer(s, page, fp)) {
756 if (object) {
757 object_err(s, page, object,
758 "Freechain corrupt");
759 set_freepointer(s, object, NULL);
760 break;
761 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700762 slab_err(s, page, "Freepointer corrupt");
Christoph Lameter81819f02007-05-06 14:49:36 -0700763 page->freelist = NULL;
764 page->inuse = s->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700765 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700766 return 0;
767 }
768 break;
769 }
770 object = fp;
771 fp = get_freepointer(s, object);
772 nr++;
773 }
774
775 if (page->inuse != s->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700776 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter24922682007-07-17 04:03:18 -0700777 "counted were %d", page->inuse, s->objects - nr);
Christoph Lameter81819f02007-05-06 14:49:36 -0700778 page->inuse = s->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700779 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700780 }
781 return search == NULL;
782}
783
Christoph Lameter3ec09742007-05-16 22:11:00 -0700784static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
785{
786 if (s->flags & SLAB_TRACE) {
787 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
788 s->name,
789 alloc ? "alloc" : "free",
790 object, page->inuse,
791 page->freelist);
792
793 if (!alloc)
794 print_section("Object", (void *)object, s->objsize);
795
796 dump_stack();
797 }
798}
799
Christoph Lameter643b1132007-05-06 14:49:42 -0700800/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700801 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700802 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700803static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700804{
Christoph Lameter643b1132007-05-06 14:49:42 -0700805 spin_lock(&n->list_lock);
806 list_add(&page->lru, &n->full);
807 spin_unlock(&n->list_lock);
808}
809
810static void remove_full(struct kmem_cache *s, struct page *page)
811{
812 struct kmem_cache_node *n;
813
814 if (!(s->flags & SLAB_STORE_USER))
815 return;
816
817 n = get_node(s, page_to_nid(page));
818
819 spin_lock(&n->list_lock);
820 list_del(&page->lru);
821 spin_unlock(&n->list_lock);
822}
823
Christoph Lameter3ec09742007-05-16 22:11:00 -0700824static void setup_object_debug(struct kmem_cache *s, struct page *page,
825 void *object)
826{
827 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
828 return;
829
830 init_object(s, object, 0);
831 init_tracking(s, object);
832}
833
834static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
835 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700836{
837 if (!check_slab(s, page))
838 goto bad;
839
840 if (object && !on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700841 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700842 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700843 }
844
845 if (!check_valid_pointer(s, page, object)) {
846 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700847 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700848 }
849
Christoph Lameter3ec09742007-05-16 22:11:00 -0700850 if (object && !check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700851 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700852
Christoph Lameter3ec09742007-05-16 22:11:00 -0700853 /* Success perform special debug activities for allocs */
854 if (s->flags & SLAB_STORE_USER)
855 set_track(s, object, TRACK_ALLOC, addr);
856 trace(s, page, object, 1);
857 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700858 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700859
Christoph Lameter81819f02007-05-06 14:49:36 -0700860bad:
861 if (PageSlab(page)) {
862 /*
863 * If this is a slab page then lets do the best we can
864 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700865 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700866 */
Christoph Lameter24922682007-07-17 04:03:18 -0700867 slab_fix(s, "Marking all objects used");
Christoph Lameter81819f02007-05-06 14:49:36 -0700868 page->inuse = s->objects;
869 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700870 }
871 return 0;
872}
873
Christoph Lameter3ec09742007-05-16 22:11:00 -0700874static int free_debug_processing(struct kmem_cache *s, struct page *page,
875 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700876{
877 if (!check_slab(s, page))
878 goto fail;
879
880 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700881 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700882 goto fail;
883 }
884
885 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700886 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700887 goto fail;
888 }
889
890 if (!check_object(s, page, object, 1))
891 return 0;
892
893 if (unlikely(s != page->slab)) {
894 if (!PageSlab(page))
Christoph Lameter70d71222007-05-06 14:49:47 -0700895 slab_err(s, page, "Attempt to free object(0x%p) "
896 "outside of slab", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700897 else
Christoph Lameter70d71222007-05-06 14:49:47 -0700898 if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700899 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700900 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700901 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700902 dump_stack();
903 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700904 else
Christoph Lameter24922682007-07-17 04:03:18 -0700905 object_err(s, page, object,
906 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700907 goto fail;
908 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700909
910 /* Special debug activities for freeing objects */
911 if (!SlabFrozen(page) && !page->freelist)
912 remove_full(s, page);
913 if (s->flags & SLAB_STORE_USER)
914 set_track(s, object, TRACK_FREE, addr);
915 trace(s, page, object, 0);
916 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700917 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700918
Christoph Lameter81819f02007-05-06 14:49:36 -0700919fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700920 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700921 return 0;
922}
923
Christoph Lameter41ecc552007-05-09 02:32:44 -0700924static int __init setup_slub_debug(char *str)
925{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700926 slub_debug = DEBUG_DEFAULT_FLAGS;
927 if (*str++ != '=' || !*str)
928 /*
929 * No options specified. Switch on full debugging.
930 */
931 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700932
933 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700934 /*
935 * No options but restriction on slabs. This means full
936 * debugging for slabs matching a pattern.
937 */
938 goto check_slabs;
939
940 slub_debug = 0;
941 if (*str == '-')
942 /*
943 * Switch off all debugging measures.
944 */
945 goto out;
946
947 /*
948 * Determine which debug features should be switched on
949 */
950 for ( ;*str && *str != ','; str++) {
951 switch (tolower(*str)) {
952 case 'f':
953 slub_debug |= SLAB_DEBUG_FREE;
954 break;
955 case 'z':
956 slub_debug |= SLAB_RED_ZONE;
957 break;
958 case 'p':
959 slub_debug |= SLAB_POISON;
960 break;
961 case 'u':
962 slub_debug |= SLAB_STORE_USER;
963 break;
964 case 't':
965 slub_debug |= SLAB_TRACE;
966 break;
967 default:
968 printk(KERN_ERR "slub_debug option '%c' "
969 "unknown. skipped\n",*str);
970 }
971 }
972
973check_slabs:
974 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -0700975 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700976out:
Christoph Lameter41ecc552007-05-09 02:32:44 -0700977 return 1;
978}
979
980__setup("slub_debug", setup_slub_debug);
981
Christoph Lameterba0268a2007-09-11 15:24:11 -0700982static unsigned long kmem_cache_flags(unsigned long objsize,
983 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -0700984 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -0700985{
986 /*
987 * The page->offset field is only 16 bit wide. This is an offset
988 * in units of words from the beginning of an object. If the slab
989 * size is bigger then we cannot move the free pointer behind the
990 * object anymore.
991 *
992 * On 32 bit platforms the limit is 256k. On 64bit platforms
993 * the limit is 512k.
994 *
Christoph Lameterc59def92007-05-16 22:10:50 -0700995 * Debugging or ctor may create a need to move the free
Christoph Lameter41ecc552007-05-09 02:32:44 -0700996 * pointer. Fail if this happens.
997 */
Christoph Lameterba0268a2007-09-11 15:24:11 -0700998 if (objsize >= 65535 * sizeof(void *)) {
999 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
Christoph Lameter41ecc552007-05-09 02:32:44 -07001000 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
Christoph Lameterba0268a2007-09-11 15:24:11 -07001001 BUG_ON(ctor);
1002 } else {
Christoph Lameter41ecc552007-05-09 02:32:44 -07001003 /*
1004 * Enable debugging if selected on the kernel commandline.
1005 */
1006 if (slub_debug && (!slub_debug_slabs ||
Christoph Lameterba0268a2007-09-11 15:24:11 -07001007 strncmp(slub_debug_slabs, name,
Christoph Lameter41ecc552007-05-09 02:32:44 -07001008 strlen(slub_debug_slabs)) == 0))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001009 flags |= slub_debug;
1010 }
1011
1012 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001013}
1014#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001015static inline void setup_object_debug(struct kmem_cache *s,
1016 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001017
Christoph Lameter3ec09742007-05-16 22:11:00 -07001018static inline int alloc_debug_processing(struct kmem_cache *s,
1019 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001020
Christoph Lameter3ec09742007-05-16 22:11:00 -07001021static inline int free_debug_processing(struct kmem_cache *s,
1022 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001023
Christoph Lameter41ecc552007-05-09 02:32:44 -07001024static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1025 { return 1; }
1026static inline int check_object(struct kmem_cache *s, struct page *page,
1027 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001028static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001029static inline unsigned long kmem_cache_flags(unsigned long objsize,
1030 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001031 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001032{
1033 return flags;
1034}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001035#define slub_debug 0
1036#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001037/*
1038 * Slab allocation and freeing
1039 */
1040static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1041{
1042 struct page * page;
1043 int pages = 1 << s->order;
1044
1045 if (s->order)
1046 flags |= __GFP_COMP;
1047
1048 if (s->flags & SLAB_CACHE_DMA)
1049 flags |= SLUB_DMA;
1050
Mel Gormane12ba742007-10-16 01:25:52 -07001051 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1052 flags |= __GFP_RECLAIMABLE;
1053
Christoph Lameter81819f02007-05-06 14:49:36 -07001054 if (node == -1)
1055 page = alloc_pages(flags, s->order);
1056 else
1057 page = alloc_pages_node(node, flags, s->order);
1058
1059 if (!page)
1060 return NULL;
1061
1062 mod_zone_page_state(page_zone(page),
1063 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1064 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1065 pages);
1066
1067 return page;
1068}
1069
1070static void setup_object(struct kmem_cache *s, struct page *page,
1071 void *object)
1072{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001073 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001074 if (unlikely(s->ctor))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001075 s->ctor(s, object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001076}
1077
1078static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1079{
1080 struct page *page;
1081 struct kmem_cache_node *n;
1082 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001083 void *last;
1084 void *p;
1085
Christoph Lameter6cb06222007-10-16 01:25:41 -07001086 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001087
Christoph Lameter6cb06222007-10-16 01:25:41 -07001088 page = allocate_slab(s,
1089 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001090 if (!page)
1091 goto out;
1092
1093 n = get_node(s, page_to_nid(page));
1094 if (n)
1095 atomic_long_inc(&n->nr_slabs);
Christoph Lameter81819f02007-05-06 14:49:36 -07001096 page->slab = s;
1097 page->flags |= 1 << PG_slab;
1098 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1099 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001100 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001101
1102 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001103
1104 if (unlikely(s->flags & SLAB_POISON))
1105 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1106
1107 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001108 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001109 setup_object(s, page, last);
1110 set_freepointer(s, last, p);
1111 last = p;
1112 }
1113 setup_object(s, page, last);
1114 set_freepointer(s, last, NULL);
1115
1116 page->freelist = start;
1117 page->inuse = 0;
1118out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001119 return page;
1120}
1121
1122static void __free_slab(struct kmem_cache *s, struct page *page)
1123{
1124 int pages = 1 << s->order;
1125
Christoph Lameterc59def92007-05-16 22:10:50 -07001126 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001127 void *p;
1128
1129 slab_pad_check(s, page);
Christoph Lameterc59def92007-05-16 22:10:50 -07001130 for_each_object(p, s, page_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001131 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001132 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001133 }
1134
1135 mod_zone_page_state(page_zone(page),
1136 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1137 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1138 - pages);
1139
Christoph Lameter81819f02007-05-06 14:49:36 -07001140 __free_pages(page, s->order);
1141}
1142
1143static void rcu_free_slab(struct rcu_head *h)
1144{
1145 struct page *page;
1146
1147 page = container_of((struct list_head *)h, struct page, lru);
1148 __free_slab(page->slab, page);
1149}
1150
1151static void free_slab(struct kmem_cache *s, struct page *page)
1152{
1153 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1154 /*
1155 * RCU free overloads the RCU head over the LRU
1156 */
1157 struct rcu_head *head = (void *)&page->lru;
1158
1159 call_rcu(head, rcu_free_slab);
1160 } else
1161 __free_slab(s, page);
1162}
1163
1164static void discard_slab(struct kmem_cache *s, struct page *page)
1165{
1166 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1167
1168 atomic_long_dec(&n->nr_slabs);
1169 reset_page_mapcount(page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001170 __ClearPageSlab(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001171 free_slab(s, page);
1172}
1173
1174/*
1175 * Per slab locking using the pagelock
1176 */
1177static __always_inline void slab_lock(struct page *page)
1178{
1179 bit_spin_lock(PG_locked, &page->flags);
1180}
1181
1182static __always_inline void slab_unlock(struct page *page)
1183{
1184 bit_spin_unlock(PG_locked, &page->flags);
1185}
1186
1187static __always_inline int slab_trylock(struct page *page)
1188{
1189 int rc = 1;
1190
1191 rc = bit_spin_trylock(PG_locked, &page->flags);
1192 return rc;
1193}
1194
1195/*
1196 * Management of partially allocated slabs
1197 */
Christoph Lametere95eed52007-05-06 14:49:44 -07001198static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001199{
Christoph Lametere95eed52007-05-06 14:49:44 -07001200 spin_lock(&n->list_lock);
1201 n->nr_partial++;
1202 list_add_tail(&page->lru, &n->partial);
1203 spin_unlock(&n->list_lock);
1204}
Christoph Lameter81819f02007-05-06 14:49:36 -07001205
Christoph Lametere95eed52007-05-06 14:49:44 -07001206static void add_partial(struct kmem_cache_node *n, struct page *page)
1207{
Christoph Lameter81819f02007-05-06 14:49:36 -07001208 spin_lock(&n->list_lock);
1209 n->nr_partial++;
1210 list_add(&page->lru, &n->partial);
1211 spin_unlock(&n->list_lock);
1212}
1213
1214static void remove_partial(struct kmem_cache *s,
1215 struct page *page)
1216{
1217 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1218
1219 spin_lock(&n->list_lock);
1220 list_del(&page->lru);
1221 n->nr_partial--;
1222 spin_unlock(&n->list_lock);
1223}
1224
1225/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001226 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001227 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001228 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001229 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001230static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001231{
1232 if (slab_trylock(page)) {
1233 list_del(&page->lru);
1234 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001235 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001236 return 1;
1237 }
1238 return 0;
1239}
1240
1241/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001242 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001243 */
1244static struct page *get_partial_node(struct kmem_cache_node *n)
1245{
1246 struct page *page;
1247
1248 /*
1249 * Racy check. If we mistakenly see no partial slabs then we
1250 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001251 * partial slab and there is none available then get_partials()
1252 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001253 */
1254 if (!n || !n->nr_partial)
1255 return NULL;
1256
1257 spin_lock(&n->list_lock);
1258 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001259 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001260 goto out;
1261 page = NULL;
1262out:
1263 spin_unlock(&n->list_lock);
1264 return page;
1265}
1266
1267/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001268 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001269 */
1270static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1271{
1272#ifdef CONFIG_NUMA
1273 struct zonelist *zonelist;
1274 struct zone **z;
1275 struct page *page;
1276
1277 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001278 * The defrag ratio allows a configuration of the tradeoffs between
1279 * inter node defragmentation and node local allocations. A lower
1280 * defrag_ratio increases the tendency to do local allocations
1281 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001282 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001283 * If the defrag_ratio is set to 0 then kmalloc() always
1284 * returns node local objects. If the ratio is higher then kmalloc()
1285 * may return off node objects because partial slabs are obtained
1286 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001287 *
1288 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001289 * defrag_ratio = 1000) then every (well almost) allocation will
1290 * first attempt to defrag slab caches on other nodes. This means
1291 * scanning over all nodes to look for partial slabs which may be
1292 * expensive if we do it every time we are trying to find a slab
1293 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001294 */
1295 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1296 return NULL;
1297
1298 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1299 ->node_zonelists[gfp_zone(flags)];
1300 for (z = zonelist->zones; *z; z++) {
1301 struct kmem_cache_node *n;
1302
1303 n = get_node(s, zone_to_nid(*z));
1304
1305 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001306 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001307 page = get_partial_node(n);
1308 if (page)
1309 return page;
1310 }
1311 }
1312#endif
1313 return NULL;
1314}
1315
1316/*
1317 * Get a partial page, lock it and return it.
1318 */
1319static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1320{
1321 struct page *page;
1322 int searchnode = (node == -1) ? numa_node_id() : node;
1323
1324 page = get_partial_node(get_node(s, searchnode));
1325 if (page || (flags & __GFP_THISNODE))
1326 return page;
1327
1328 return get_any_partial(s, flags);
1329}
1330
1331/*
1332 * Move a page back to the lists.
1333 *
1334 * Must be called with the slab lock held.
1335 *
1336 * On exit the slab lock will have been dropped.
1337 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001338static void unfreeze_slab(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001339{
Christoph Lametere95eed52007-05-06 14:49:44 -07001340 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1341
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001342 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001343 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001344
Christoph Lameter81819f02007-05-06 14:49:36 -07001345 if (page->freelist)
Christoph Lametere95eed52007-05-06 14:49:44 -07001346 add_partial(n, page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001347 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
Christoph Lametere95eed52007-05-06 14:49:44 -07001348 add_full(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001349 slab_unlock(page);
Christoph Lametere95eed52007-05-06 14:49:44 -07001350
Christoph Lameter81819f02007-05-06 14:49:36 -07001351 } else {
Christoph Lametere95eed52007-05-06 14:49:44 -07001352 if (n->nr_partial < MIN_PARTIAL) {
1353 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001354 * Adding an empty slab to the partial slabs in order
1355 * to avoid page allocator overhead. This slab needs
1356 * to come after the other slabs with objects in
1357 * order to fill them up. That way the size of the
1358 * partial list stays small. kmem_cache_shrink can
1359 * reclaim empty slabs from the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001360 */
1361 add_partial_tail(n, page);
1362 slab_unlock(page);
1363 } else {
1364 slab_unlock(page);
1365 discard_slab(s, page);
1366 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001367 }
1368}
1369
1370/*
1371 * Remove the cpu slab
1372 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001373static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001374{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001375 struct page *page = c->page;
Christoph Lameter894b8782007-05-10 03:15:16 -07001376 /*
1377 * Merge cpu freelist into freelist. Typically we get here
1378 * because both freelists are empty. So this is unlikely
1379 * to occur.
1380 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001381 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001382 void **object;
1383
1384 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001385 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001386 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001387
1388 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001389 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001390 page->freelist = object;
1391 page->inuse--;
1392 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001393 c->page = NULL;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001394 unfreeze_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001395}
1396
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001397static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001398{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001399 slab_lock(c->page);
1400 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001401}
1402
1403/*
1404 * Flush cpu slab.
1405 * Called from IPI handler with interrupts disabled.
1406 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001407static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001408{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001409 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001410
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001411 if (likely(c && c->page))
1412 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001413}
1414
1415static void flush_cpu_slab(void *d)
1416{
1417 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001418
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001419 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001420}
1421
1422static void flush_all(struct kmem_cache *s)
1423{
1424#ifdef CONFIG_SMP
1425 on_each_cpu(flush_cpu_slab, s, 1, 1);
1426#else
1427 unsigned long flags;
1428
1429 local_irq_save(flags);
1430 flush_cpu_slab(s);
1431 local_irq_restore(flags);
1432#endif
1433}
1434
1435/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001436 * Check if the objects in a per cpu structure fit numa
1437 * locality expectations.
1438 */
1439static inline int node_match(struct kmem_cache_cpu *c, int node)
1440{
1441#ifdef CONFIG_NUMA
1442 if (node != -1 && c->node != node)
1443 return 0;
1444#endif
1445 return 1;
1446}
1447
1448/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001449 * Slow path. The lockless freelist is empty or we need to perform
1450 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001451 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001452 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001453 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001454 * Processing is still very fast if new objects have been freed to the
1455 * regular freelist. In that case we simply take over the regular freelist
1456 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001457 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001458 * If that is not working then we fall back to the partial lists. We take the
1459 * first element of the freelist as the object to allocate now and move the
1460 * rest of the freelist to the lockless freelist.
1461 *
1462 * And if we were unable to get a new slab from the partial slab lists then
1463 * we need to allocate a new slab. This is slowest path since we may sleep.
Christoph Lameter81819f02007-05-06 14:49:36 -07001464 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001465static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001466 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001467{
Christoph Lameter81819f02007-05-06 14:49:36 -07001468 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001469 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001470
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001471 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001472 goto new_slab;
1473
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001474 slab_lock(c->page);
1475 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001476 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001477load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001478 object = c->page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001479 if (unlikely(!object))
1480 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001481 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001482 goto debug;
1483
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001484 object = c->page->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001485 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001486 c->page->inuse = s->objects;
1487 c->page->freelist = NULL;
1488 c->node = page_to_nid(c->page);
1489 slab_unlock(c->page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001490 return object;
1491
1492another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001493 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001494
1495new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001496 new = get_partial(s, gfpflags, node);
1497 if (new) {
1498 c->page = new;
Christoph Lameter894b8782007-05-10 03:15:16 -07001499 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001500 }
1501
Christoph Lameterb811c202007-10-16 23:25:51 -07001502 if (gfpflags & __GFP_WAIT)
1503 local_irq_enable();
1504
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001505 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001506
1507 if (gfpflags & __GFP_WAIT)
1508 local_irq_disable();
1509
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001510 if (new) {
1511 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter05aa3452007-11-05 11:31:58 -08001512 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001513 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001514 slab_lock(new);
1515 SetSlabFrozen(new);
1516 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001517 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001518 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001519 return NULL;
1520debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001521 object = c->page->freelist;
1522 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001523 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001524
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001525 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001526 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001527 c->node = -1;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001528 slab_unlock(c->page);
Christoph Lameter894b8782007-05-10 03:15:16 -07001529 return object;
1530}
1531
1532/*
1533 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1534 * have the fastpath folded into their functions. So no function call
1535 * overhead for requests that can be satisfied on the fastpath.
1536 *
1537 * The fastpath works by first checking if the lockless freelist can be used.
1538 * If not then __slab_alloc is called for slow processing.
1539 *
1540 * Otherwise we can simply pick the next object from the lockless free list.
1541 */
1542static void __always_inline *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001543 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001544{
Christoph Lameter894b8782007-05-10 03:15:16 -07001545 void **object;
1546 unsigned long flags;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001547 struct kmem_cache_cpu *c;
Christoph Lameter894b8782007-05-10 03:15:16 -07001548
1549 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001550 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001551 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001552
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001553 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001554
1555 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001556 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001557 c->freelist = object[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001558 }
1559 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001560
1561 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001562 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001563
Christoph Lameter894b8782007-05-10 03:15:16 -07001564 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001565}
1566
1567void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1568{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001569 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001570}
1571EXPORT_SYMBOL(kmem_cache_alloc);
1572
1573#ifdef CONFIG_NUMA
1574void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1575{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001576 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001577}
1578EXPORT_SYMBOL(kmem_cache_alloc_node);
1579#endif
1580
1581/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001582 * Slow patch handling. This may still be called frequently since objects
1583 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001584 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001585 * So we still attempt to reduce cache line usage. Just take the slab
1586 * lock and free the item. If there is no additional partial page
1587 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001588 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001589static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001590 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001591{
1592 void *prior;
1593 void **object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07001594
Christoph Lameter81819f02007-05-06 14:49:36 -07001595 slab_lock(page);
1596
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001597 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001598 goto debug;
1599checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001600 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001601 page->freelist = object;
1602 page->inuse--;
1603
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001604 if (unlikely(SlabFrozen(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001605 goto out_unlock;
1606
1607 if (unlikely(!page->inuse))
1608 goto slab_empty;
1609
1610 /*
1611 * Objects left in the slab. If it
1612 * was not on the partial list before
1613 * then add it.
1614 */
1615 if (unlikely(!prior))
Christoph Lameter76be8952007-12-21 14:37:37 -08001616 add_partial_tail(get_node(s, page_to_nid(page)), page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001617
1618out_unlock:
1619 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001620 return;
1621
1622slab_empty:
1623 if (prior)
1624 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001625 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001626 */
1627 remove_partial(s, page);
1628
1629 slab_unlock(page);
1630 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001631 return;
1632
1633debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001634 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001635 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001636 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001637}
1638
Christoph Lameter894b8782007-05-10 03:15:16 -07001639/*
1640 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1641 * can perform fastpath freeing without additional function calls.
1642 *
1643 * The fastpath is only possible if we are freeing to the current cpu slab
1644 * of this processor. This typically the case if we have just allocated
1645 * the item before.
1646 *
1647 * If fastpath is not possible then fall back to __slab_free where we deal
1648 * with all sorts of special processing.
1649 */
1650static void __always_inline slab_free(struct kmem_cache *s,
1651 struct page *page, void *x, void *addr)
1652{
1653 void **object = (void *)x;
1654 unsigned long flags;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001655 struct kmem_cache_cpu *c;
Christoph Lameter894b8782007-05-10 03:15:16 -07001656
1657 local_irq_save(flags);
Peter Zijlstra02febdf2007-07-26 20:01:38 +02001658 debug_check_no_locks_freed(object, s->objsize);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001659 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001660 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001661 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001662 c->freelist = object;
Christoph Lameter894b8782007-05-10 03:15:16 -07001663 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001664 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001665
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 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001752 fls(min_objects * size - 1) - PAGE_SHIFT);
1753 order <= max_order; order++) {
1754
Christoph Lameter81819f02007-05-06 14:49:36 -07001755 unsigned long slab_size = PAGE_SIZE << order;
1756
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001757 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001758 continue;
1759
Christoph Lameter81819f02007-05-06 14:49:36 -07001760 rem = slab_size % size;
1761
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001762 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001763 break;
1764
1765 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001766
Christoph Lameter81819f02007-05-06 14:49:36 -07001767 return order;
1768}
1769
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001770static inline int calculate_order(int size)
1771{
1772 int order;
1773 int min_objects;
1774 int fraction;
1775
1776 /*
1777 * Attempt to find best configuration for a slab. This
1778 * works by first attempting to generate a layout with
1779 * the best configuration and backing off gradually.
1780 *
1781 * First we reduce the acceptable waste in a slab. Then
1782 * we reduce the minimum objects required in a slab.
1783 */
1784 min_objects = slub_min_objects;
1785 while (min_objects > 1) {
1786 fraction = 8;
1787 while (fraction >= 4) {
1788 order = slab_order(size, min_objects,
1789 slub_max_order, fraction);
1790 if (order <= slub_max_order)
1791 return order;
1792 fraction /= 2;
1793 }
1794 min_objects /= 2;
1795 }
1796
1797 /*
1798 * We were unable to place multiple objects in a slab. Now
1799 * lets see if we can place a single object there.
1800 */
1801 order = slab_order(size, 1, slub_max_order, 1);
1802 if (order <= slub_max_order)
1803 return order;
1804
1805 /*
1806 * Doh this slab cannot be placed using slub_max_order.
1807 */
1808 order = slab_order(size, 1, MAX_ORDER, 1);
1809 if (order <= MAX_ORDER)
1810 return order;
1811 return -ENOSYS;
1812}
1813
Christoph Lameter81819f02007-05-06 14:49:36 -07001814/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001815 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001816 */
1817static unsigned long calculate_alignment(unsigned long flags,
1818 unsigned long align, unsigned long size)
1819{
1820 /*
1821 * If the user wants hardware cache aligned objects then
1822 * follow that suggestion if the object is sufficiently
1823 * large.
1824 *
1825 * The hardware cache alignment cannot override the
1826 * specified alignment though. If that is greater
1827 * then use it.
1828 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001829 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001830 size > cache_line_size() / 2)
1831 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001832
1833 if (align < ARCH_SLAB_MINALIGN)
1834 return ARCH_SLAB_MINALIGN;
1835
1836 return ALIGN(align, sizeof(void *));
1837}
1838
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001839static void init_kmem_cache_cpu(struct kmem_cache *s,
1840 struct kmem_cache_cpu *c)
1841{
1842 c->page = NULL;
1843 c->freelist = NULL;
1844 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001845 c->offset = s->offset / sizeof(void *);
1846 c->objsize = s->objsize;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001847}
1848
Christoph Lameter81819f02007-05-06 14:49:36 -07001849static void init_kmem_cache_node(struct kmem_cache_node *n)
1850{
1851 n->nr_partial = 0;
1852 atomic_long_set(&n->nr_slabs, 0);
1853 spin_lock_init(&n->list_lock);
1854 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001855#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07001856 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001857#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001858}
1859
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001860#ifdef CONFIG_SMP
1861/*
1862 * Per cpu array for per cpu structures.
1863 *
1864 * The per cpu array places all kmem_cache_cpu structures from one processor
1865 * close together meaning that it becomes possible that multiple per cpu
1866 * structures are contained in one cacheline. This may be particularly
1867 * beneficial for the kmalloc caches.
1868 *
1869 * A desktop system typically has around 60-80 slabs. With 100 here we are
1870 * likely able to get per cpu structures for all caches from the array defined
1871 * here. We must be able to cover all kmalloc caches during bootstrap.
1872 *
1873 * If the per cpu array is exhausted then fall back to kmalloc
1874 * of individual cachelines. No sharing is possible then.
1875 */
1876#define NR_KMEM_CACHE_CPU 100
1877
1878static DEFINE_PER_CPU(struct kmem_cache_cpu,
1879 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1880
1881static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1882static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1883
1884static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1885 int cpu, gfp_t flags)
1886{
1887 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1888
1889 if (c)
1890 per_cpu(kmem_cache_cpu_free, cpu) =
1891 (void *)c->freelist;
1892 else {
1893 /* Table overflow: So allocate ourselves */
1894 c = kmalloc_node(
1895 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1896 flags, cpu_to_node(cpu));
1897 if (!c)
1898 return NULL;
1899 }
1900
1901 init_kmem_cache_cpu(s, c);
1902 return c;
1903}
1904
1905static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1906{
1907 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1908 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1909 kfree(c);
1910 return;
1911 }
1912 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1913 per_cpu(kmem_cache_cpu_free, cpu) = c;
1914}
1915
1916static void free_kmem_cache_cpus(struct kmem_cache *s)
1917{
1918 int cpu;
1919
1920 for_each_online_cpu(cpu) {
1921 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1922
1923 if (c) {
1924 s->cpu_slab[cpu] = NULL;
1925 free_kmem_cache_cpu(c, cpu);
1926 }
1927 }
1928}
1929
1930static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1931{
1932 int cpu;
1933
1934 for_each_online_cpu(cpu) {
1935 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1936
1937 if (c)
1938 continue;
1939
1940 c = alloc_kmem_cache_cpu(s, cpu, flags);
1941 if (!c) {
1942 free_kmem_cache_cpus(s);
1943 return 0;
1944 }
1945 s->cpu_slab[cpu] = c;
1946 }
1947 return 1;
1948}
1949
1950/*
1951 * Initialize the per cpu array.
1952 */
1953static void init_alloc_cpu_cpu(int cpu)
1954{
1955 int i;
1956
1957 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
1958 return;
1959
1960 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
1961 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
1962
1963 cpu_set(cpu, kmem_cach_cpu_free_init_once);
1964}
1965
1966static void __init init_alloc_cpu(void)
1967{
1968 int cpu;
1969
1970 for_each_online_cpu(cpu)
1971 init_alloc_cpu_cpu(cpu);
1972 }
1973
1974#else
1975static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
1976static inline void init_alloc_cpu(void) {}
1977
1978static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1979{
1980 init_kmem_cache_cpu(s, &s->cpu_slab);
1981 return 1;
1982}
1983#endif
1984
Christoph Lameter81819f02007-05-06 14:49:36 -07001985#ifdef CONFIG_NUMA
1986/*
1987 * No kmalloc_node yet so do it by hand. We know that this is the first
1988 * slab on the node for this slabcache. There are no concurrent accesses
1989 * possible.
1990 *
1991 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001992 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
1993 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07001994 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07001995static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
1996 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07001997{
1998 struct page *page;
1999 struct kmem_cache_node *n;
2000
2001 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2002
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002003 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002004
2005 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002006 if (page_to_nid(page) != node) {
2007 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2008 "node %d\n", node);
2009 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2010 "in order to be able to continue\n");
2011 }
2012
Christoph Lameter81819f02007-05-06 14:49:36 -07002013 n = page->freelist;
2014 BUG_ON(!n);
2015 page->freelist = get_freepointer(kmalloc_caches, n);
2016 page->inuse++;
2017 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002018#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002019 init_object(kmalloc_caches, n, 1);
2020 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002021#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002022 init_kmem_cache_node(n);
2023 atomic_long_inc(&n->nr_slabs);
Christoph Lametere95eed52007-05-06 14:49:44 -07002024 add_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002025 return n;
2026}
2027
2028static void free_kmem_cache_nodes(struct kmem_cache *s)
2029{
2030 int node;
2031
Christoph Lameterf64dc582007-10-16 01:25:33 -07002032 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002033 struct kmem_cache_node *n = s->node[node];
2034 if (n && n != &s->local_node)
2035 kmem_cache_free(kmalloc_caches, n);
2036 s->node[node] = NULL;
2037 }
2038}
2039
2040static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2041{
2042 int node;
2043 int local_node;
2044
2045 if (slab_state >= UP)
2046 local_node = page_to_nid(virt_to_page(s));
2047 else
2048 local_node = 0;
2049
Christoph Lameterf64dc582007-10-16 01:25:33 -07002050 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002051 struct kmem_cache_node *n;
2052
2053 if (local_node == node)
2054 n = &s->local_node;
2055 else {
2056 if (slab_state == DOWN) {
2057 n = early_kmem_cache_node_alloc(gfpflags,
2058 node);
2059 continue;
2060 }
2061 n = kmem_cache_alloc_node(kmalloc_caches,
2062 gfpflags, node);
2063
2064 if (!n) {
2065 free_kmem_cache_nodes(s);
2066 return 0;
2067 }
2068
2069 }
2070 s->node[node] = n;
2071 init_kmem_cache_node(n);
2072 }
2073 return 1;
2074}
2075#else
2076static void free_kmem_cache_nodes(struct kmem_cache *s)
2077{
2078}
2079
2080static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2081{
2082 init_kmem_cache_node(&s->local_node);
2083 return 1;
2084}
2085#endif
2086
2087/*
2088 * calculate_sizes() determines the order and the distribution of data within
2089 * a slab object.
2090 */
2091static int calculate_sizes(struct kmem_cache *s)
2092{
2093 unsigned long flags = s->flags;
2094 unsigned long size = s->objsize;
2095 unsigned long align = s->align;
2096
2097 /*
2098 * Determine if we can poison the object itself. If the user of
2099 * the slab may touch the object after free or before allocation
2100 * then we should never poison the object itself.
2101 */
2102 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002103 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002104 s->flags |= __OBJECT_POISON;
2105 else
2106 s->flags &= ~__OBJECT_POISON;
2107
2108 /*
2109 * Round up object size to the next word boundary. We can only
2110 * place the free pointer at word boundaries and this determines
2111 * the possible location of the free pointer.
2112 */
2113 size = ALIGN(size, sizeof(void *));
2114
Christoph Lameter41ecc552007-05-09 02:32:44 -07002115#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002116 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002117 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002118 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002119 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002120 */
2121 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2122 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002123#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002124
2125 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002126 * With that we have determined the number of bytes in actual use
2127 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002128 */
2129 s->inuse = size;
2130
2131 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002132 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002133 /*
2134 * Relocate free pointer after the object if it is not
2135 * permitted to overwrite the first word of the object on
2136 * kmem_cache_free.
2137 *
2138 * This is the case if we do RCU, have a constructor or
2139 * destructor or are poisoning the objects.
2140 */
2141 s->offset = size;
2142 size += sizeof(void *);
2143 }
2144
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002145#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002146 if (flags & SLAB_STORE_USER)
2147 /*
2148 * Need to store information about allocs and frees after
2149 * the object.
2150 */
2151 size += 2 * sizeof(struct track);
2152
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002153 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002154 /*
2155 * Add some empty padding so that we can catch
2156 * overwrites from earlier objects rather than let
2157 * tracking information or the free pointer be
2158 * corrupted if an user writes before the start
2159 * of the object.
2160 */
2161 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002162#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002163
Christoph Lameter81819f02007-05-06 14:49:36 -07002164 /*
2165 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002166 * user specified and the dynamic determination of cache line size
2167 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002168 */
2169 align = calculate_alignment(flags, align, s->objsize);
2170
2171 /*
2172 * SLUB stores one object immediately after another beginning from
2173 * offset 0. In order to align the objects we have to simply size
2174 * each object to conform to the alignment.
2175 */
2176 size = ALIGN(size, align);
2177 s->size = size;
2178
2179 s->order = calculate_order(size);
2180 if (s->order < 0)
2181 return 0;
2182
2183 /*
2184 * Determine the number of objects per slab
2185 */
2186 s->objects = (PAGE_SIZE << s->order) / size;
2187
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002188 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002189
2190}
2191
Christoph Lameter81819f02007-05-06 14:49:36 -07002192static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2193 const char *name, size_t size,
2194 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002195 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002196{
2197 memset(s, 0, kmem_size);
2198 s->name = name;
2199 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002200 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002201 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002202 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002203
2204 if (!calculate_sizes(s))
2205 goto error;
2206
2207 s->refcount = 1;
2208#ifdef CONFIG_NUMA
2209 s->defrag_ratio = 100;
2210#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002211 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2212 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002213
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002214 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002215 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002216 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002217error:
2218 if (flags & SLAB_PANIC)
2219 panic("Cannot create slab %s size=%lu realsize=%u "
2220 "order=%u offset=%u flags=%lx\n",
2221 s->name, (unsigned long)size, s->size, s->order,
2222 s->offset, flags);
2223 return 0;
2224}
Christoph Lameter81819f02007-05-06 14:49:36 -07002225
2226/*
2227 * Check if a given pointer is valid
2228 */
2229int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2230{
2231 struct page * page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002232
2233 page = get_object_page(object);
2234
2235 if (!page || s != page->slab)
2236 /* No slab or wrong slab */
2237 return 0;
2238
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002239 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002240 return 0;
2241
2242 /*
2243 * We could also check if the object is on the slabs freelist.
2244 * But this would be too expensive and it seems that the main
2245 * purpose of kmem_ptr_valid is to check if the object belongs
2246 * to a certain slab.
2247 */
2248 return 1;
2249}
2250EXPORT_SYMBOL(kmem_ptr_validate);
2251
2252/*
2253 * Determine the size of a slab object
2254 */
2255unsigned int kmem_cache_size(struct kmem_cache *s)
2256{
2257 return s->objsize;
2258}
2259EXPORT_SYMBOL(kmem_cache_size);
2260
2261const char *kmem_cache_name(struct kmem_cache *s)
2262{
2263 return s->name;
2264}
2265EXPORT_SYMBOL(kmem_cache_name);
2266
2267/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002268 * Attempt to free all slabs on a node. Return the number of slabs we
2269 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002270 */
2271static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2272 struct list_head *list)
2273{
2274 int slabs_inuse = 0;
2275 unsigned long flags;
2276 struct page *page, *h;
2277
2278 spin_lock_irqsave(&n->list_lock, flags);
2279 list_for_each_entry_safe(page, h, list, lru)
2280 if (!page->inuse) {
2281 list_del(&page->lru);
2282 discard_slab(s, page);
2283 } else
2284 slabs_inuse++;
2285 spin_unlock_irqrestore(&n->list_lock, flags);
2286 return slabs_inuse;
2287}
2288
2289/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002290 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002291 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002292static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002293{
2294 int node;
2295
2296 flush_all(s);
2297
2298 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002299 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002300 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002301 struct kmem_cache_node *n = get_node(s, node);
2302
Christoph Lameter2086d262007-05-06 14:49:46 -07002303 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002304 if (atomic_long_read(&n->nr_slabs))
2305 return 1;
2306 }
2307 free_kmem_cache_nodes(s);
2308 return 0;
2309}
2310
2311/*
2312 * Close a cache and release the kmem_cache structure
2313 * (must be used for caches created using kmem_cache_create)
2314 */
2315void kmem_cache_destroy(struct kmem_cache *s)
2316{
2317 down_write(&slub_lock);
2318 s->refcount--;
2319 if (!s->refcount) {
2320 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002321 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002322 if (kmem_cache_close(s))
2323 WARN_ON(1);
2324 sysfs_slab_remove(s);
2325 kfree(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002326 } else
2327 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002328}
2329EXPORT_SYMBOL(kmem_cache_destroy);
2330
2331/********************************************************************
2332 * Kmalloc subsystem
2333 *******************************************************************/
2334
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002335struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002336EXPORT_SYMBOL(kmalloc_caches);
2337
2338#ifdef CONFIG_ZONE_DMA
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002339static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT];
Christoph Lameter81819f02007-05-06 14:49:36 -07002340#endif
2341
2342static int __init setup_slub_min_order(char *str)
2343{
2344 get_option (&str, &slub_min_order);
2345
2346 return 1;
2347}
2348
2349__setup("slub_min_order=", setup_slub_min_order);
2350
2351static int __init setup_slub_max_order(char *str)
2352{
2353 get_option (&str, &slub_max_order);
2354
2355 return 1;
2356}
2357
2358__setup("slub_max_order=", setup_slub_max_order);
2359
2360static int __init setup_slub_min_objects(char *str)
2361{
2362 get_option (&str, &slub_min_objects);
2363
2364 return 1;
2365}
2366
2367__setup("slub_min_objects=", setup_slub_min_objects);
2368
2369static int __init setup_slub_nomerge(char *str)
2370{
2371 slub_nomerge = 1;
2372 return 1;
2373}
2374
2375__setup("slub_nomerge", setup_slub_nomerge);
2376
Christoph Lameter81819f02007-05-06 14:49:36 -07002377static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2378 const char *name, int size, gfp_t gfp_flags)
2379{
2380 unsigned int flags = 0;
2381
2382 if (gfp_flags & SLUB_DMA)
2383 flags = SLAB_CACHE_DMA;
2384
2385 down_write(&slub_lock);
2386 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameterc59def92007-05-16 22:10:50 -07002387 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002388 goto panic;
2389
2390 list_add(&s->list, &slab_caches);
2391 up_write(&slub_lock);
2392 if (sysfs_slab_add(s))
2393 goto panic;
2394 return s;
2395
2396panic:
2397 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2398}
2399
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002400#ifdef CONFIG_ZONE_DMA
Christoph Lameter1ceef402007-08-07 15:11:48 -07002401
2402static void sysfs_add_func(struct work_struct *w)
2403{
2404 struct kmem_cache *s;
2405
2406 down_write(&slub_lock);
2407 list_for_each_entry(s, &slab_caches, list) {
2408 if (s->flags & __SYSFS_ADD_DEFERRED) {
2409 s->flags &= ~__SYSFS_ADD_DEFERRED;
2410 sysfs_slab_add(s);
2411 }
2412 }
2413 up_write(&slub_lock);
2414}
2415
2416static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2417
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002418static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2419{
2420 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002421 char *text;
2422 size_t realsize;
2423
2424 s = kmalloc_caches_dma[index];
2425 if (s)
2426 return s;
2427
2428 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002429 if (flags & __GFP_WAIT)
2430 down_write(&slub_lock);
2431 else {
2432 if (!down_write_trylock(&slub_lock))
2433 goto out;
2434 }
2435
2436 if (kmalloc_caches_dma[index])
2437 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002438
Christoph Lameter7b55f622007-07-17 04:03:27 -07002439 realsize = kmalloc_caches[index].objsize;
Christoph Lameter1ceef402007-08-07 15:11:48 -07002440 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", (unsigned int)realsize),
2441 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2442
2443 if (!s || !text || !kmem_cache_open(s, flags, text,
2444 realsize, ARCH_KMALLOC_MINALIGN,
2445 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2446 kfree(s);
2447 kfree(text);
2448 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002449 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002450
2451 list_add(&s->list, &slab_caches);
2452 kmalloc_caches_dma[index] = s;
2453
2454 schedule_work(&sysfs_add_work);
2455
2456unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002457 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002458out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002459 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002460}
2461#endif
2462
Christoph Lameterf1b26332007-07-17 04:03:26 -07002463/*
2464 * Conversion table for small slabs sizes / 8 to the index in the
2465 * kmalloc array. This is necessary for slabs < 192 since we have non power
2466 * of two cache sizes there. The size of larger slabs can be determined using
2467 * fls.
2468 */
2469static s8 size_index[24] = {
2470 3, /* 8 */
2471 4, /* 16 */
2472 5, /* 24 */
2473 5, /* 32 */
2474 6, /* 40 */
2475 6, /* 48 */
2476 6, /* 56 */
2477 6, /* 64 */
2478 1, /* 72 */
2479 1, /* 80 */
2480 1, /* 88 */
2481 1, /* 96 */
2482 7, /* 104 */
2483 7, /* 112 */
2484 7, /* 120 */
2485 7, /* 128 */
2486 2, /* 136 */
2487 2, /* 144 */
2488 2, /* 152 */
2489 2, /* 160 */
2490 2, /* 168 */
2491 2, /* 176 */
2492 2, /* 184 */
2493 2 /* 192 */
2494};
2495
Christoph Lameter81819f02007-05-06 14:49:36 -07002496static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2497{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002498 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002499
Christoph Lameterf1b26332007-07-17 04:03:26 -07002500 if (size <= 192) {
2501 if (!size)
2502 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002503
Christoph Lameterf1b26332007-07-17 04:03:26 -07002504 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002505 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002506 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002507
2508#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002509 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002510 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002511
Christoph Lameter81819f02007-05-06 14:49:36 -07002512#endif
2513 return &kmalloc_caches[index];
2514}
2515
2516void *__kmalloc(size_t size, gfp_t flags)
2517{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002518 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002519
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002520 if (unlikely(size > PAGE_SIZE / 2))
2521 return (void *)__get_free_pages(flags | __GFP_COMP,
2522 get_order(size));
2523
2524 s = get_slab(size, flags);
2525
2526 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002527 return s;
2528
Christoph Lameterce15fea2007-07-17 04:03:28 -07002529 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002530}
2531EXPORT_SYMBOL(__kmalloc);
2532
2533#ifdef CONFIG_NUMA
2534void *__kmalloc_node(size_t size, gfp_t flags, int node)
2535{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002536 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002537
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002538 if (unlikely(size > PAGE_SIZE / 2))
2539 return (void *)__get_free_pages(flags | __GFP_COMP,
2540 get_order(size));
2541
2542 s = get_slab(size, flags);
2543
2544 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002545 return s;
2546
Christoph Lameterce15fea2007-07-17 04:03:28 -07002547 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002548}
2549EXPORT_SYMBOL(__kmalloc_node);
2550#endif
2551
2552size_t ksize(const void *object)
2553{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002554 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002555 struct kmem_cache *s;
2556
Christoph Lameteref8b4522007-10-16 01:24:46 -07002557 BUG_ON(!object);
2558 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002559 return 0;
2560
Vegard Nossum294a80a2007-12-04 23:45:30 -08002561 page = virt_to_head_page(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07002562 BUG_ON(!page);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002563
2564 if (unlikely(!PageSlab(page)))
2565 return PAGE_SIZE << compound_order(page);
2566
Christoph Lameter81819f02007-05-06 14:49:36 -07002567 s = page->slab;
2568 BUG_ON(!s);
2569
2570 /*
2571 * Debugging requires use of the padding between object
2572 * and whatever may come after it.
2573 */
2574 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2575 return s->objsize;
2576
2577 /*
2578 * If we have the need to store the freelist pointer
2579 * back there or track user information then we can
2580 * only use the space before that information.
2581 */
2582 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2583 return s->inuse;
2584
2585 /*
2586 * Else we can use all the padding etc for the allocation
2587 */
2588 return s->size;
2589}
2590EXPORT_SYMBOL(ksize);
2591
2592void kfree(const void *x)
2593{
Christoph Lameter81819f02007-05-06 14:49:36 -07002594 struct page *page;
2595
Satyam Sharma2408c552007-10-16 01:24:44 -07002596 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002597 return;
2598
Christoph Lameterb49af682007-05-06 14:49:41 -07002599 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002600 if (unlikely(!PageSlab(page))) {
2601 put_page(page);
2602 return;
2603 }
2604 slab_free(page->slab, page, (void *)x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002605}
2606EXPORT_SYMBOL(kfree);
2607
Christoph Lameter2086d262007-05-06 14:49:46 -07002608/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002609 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2610 * the remaining slabs by the number of items in use. The slabs with the
2611 * most items in use come first. New allocations will then fill those up
2612 * and thus they can be removed from the partial lists.
2613 *
2614 * The slabs with the least items are placed last. This results in them
2615 * being allocated from last increasing the chance that the last objects
2616 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002617 */
2618int kmem_cache_shrink(struct kmem_cache *s)
2619{
2620 int node;
2621 int i;
2622 struct kmem_cache_node *n;
2623 struct page *page;
2624 struct page *t;
2625 struct list_head *slabs_by_inuse =
2626 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2627 unsigned long flags;
2628
2629 if (!slabs_by_inuse)
2630 return -ENOMEM;
2631
2632 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002633 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002634 n = get_node(s, node);
2635
2636 if (!n->nr_partial)
2637 continue;
2638
2639 for (i = 0; i < s->objects; i++)
2640 INIT_LIST_HEAD(slabs_by_inuse + i);
2641
2642 spin_lock_irqsave(&n->list_lock, flags);
2643
2644 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002645 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002646 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002647 * Note that concurrent frees may occur while we hold the
2648 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002649 */
2650 list_for_each_entry_safe(page, t, &n->partial, lru) {
2651 if (!page->inuse && slab_trylock(page)) {
2652 /*
2653 * Must hold slab lock here because slab_free
2654 * may have freed the last object and be
2655 * waiting to release the slab.
2656 */
2657 list_del(&page->lru);
2658 n->nr_partial--;
2659 slab_unlock(page);
2660 discard_slab(s, page);
2661 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002662 list_move(&page->lru,
2663 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002664 }
2665 }
2666
Christoph Lameter2086d262007-05-06 14:49:46 -07002667 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002668 * Rebuild the partial list with the slabs filled up most
2669 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002670 */
2671 for (i = s->objects - 1; i >= 0; i--)
2672 list_splice(slabs_by_inuse + i, n->partial.prev);
2673
Christoph Lameter2086d262007-05-06 14:49:46 -07002674 spin_unlock_irqrestore(&n->list_lock, flags);
2675 }
2676
2677 kfree(slabs_by_inuse);
2678 return 0;
2679}
2680EXPORT_SYMBOL(kmem_cache_shrink);
2681
Yasunori Gotob9049e22007-10-21 16:41:37 -07002682#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2683static int slab_mem_going_offline_callback(void *arg)
2684{
2685 struct kmem_cache *s;
2686
2687 down_read(&slub_lock);
2688 list_for_each_entry(s, &slab_caches, list)
2689 kmem_cache_shrink(s);
2690 up_read(&slub_lock);
2691
2692 return 0;
2693}
2694
2695static void slab_mem_offline_callback(void *arg)
2696{
2697 struct kmem_cache_node *n;
2698 struct kmem_cache *s;
2699 struct memory_notify *marg = arg;
2700 int offline_node;
2701
2702 offline_node = marg->status_change_nid;
2703
2704 /*
2705 * If the node still has available memory. we need kmem_cache_node
2706 * for it yet.
2707 */
2708 if (offline_node < 0)
2709 return;
2710
2711 down_read(&slub_lock);
2712 list_for_each_entry(s, &slab_caches, list) {
2713 n = get_node(s, offline_node);
2714 if (n) {
2715 /*
2716 * if n->nr_slabs > 0, slabs still exist on the node
2717 * that is going down. We were unable to free them,
2718 * and offline_pages() function shoudn't call this
2719 * callback. So, we must fail.
2720 */
Al Viro27bb6282007-10-29 04:42:55 +00002721 BUG_ON(atomic_long_read(&n->nr_slabs));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002722
2723 s->node[offline_node] = NULL;
2724 kmem_cache_free(kmalloc_caches, n);
2725 }
2726 }
2727 up_read(&slub_lock);
2728}
2729
2730static int slab_mem_going_online_callback(void *arg)
2731{
2732 struct kmem_cache_node *n;
2733 struct kmem_cache *s;
2734 struct memory_notify *marg = arg;
2735 int nid = marg->status_change_nid;
2736 int ret = 0;
2737
2738 /*
2739 * If the node's memory is already available, then kmem_cache_node is
2740 * already created. Nothing to do.
2741 */
2742 if (nid < 0)
2743 return 0;
2744
2745 /*
2746 * We are bringing a node online. No memory is availabe yet. We must
2747 * allocate a kmem_cache_node structure in order to bring the node
2748 * online.
2749 */
2750 down_read(&slub_lock);
2751 list_for_each_entry(s, &slab_caches, list) {
2752 /*
2753 * XXX: kmem_cache_alloc_node will fallback to other nodes
2754 * since memory is not yet available from the node that
2755 * is brought up.
2756 */
2757 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2758 if (!n) {
2759 ret = -ENOMEM;
2760 goto out;
2761 }
2762 init_kmem_cache_node(n);
2763 s->node[nid] = n;
2764 }
2765out:
2766 up_read(&slub_lock);
2767 return ret;
2768}
2769
2770static int slab_memory_callback(struct notifier_block *self,
2771 unsigned long action, void *arg)
2772{
2773 int ret = 0;
2774
2775 switch (action) {
2776 case MEM_GOING_ONLINE:
2777 ret = slab_mem_going_online_callback(arg);
2778 break;
2779 case MEM_GOING_OFFLINE:
2780 ret = slab_mem_going_offline_callback(arg);
2781 break;
2782 case MEM_OFFLINE:
2783 case MEM_CANCEL_ONLINE:
2784 slab_mem_offline_callback(arg);
2785 break;
2786 case MEM_ONLINE:
2787 case MEM_CANCEL_OFFLINE:
2788 break;
2789 }
2790
2791 ret = notifier_from_errno(ret);
2792 return ret;
2793}
2794
2795#endif /* CONFIG_MEMORY_HOTPLUG */
2796
Christoph Lameter81819f02007-05-06 14:49:36 -07002797/********************************************************************
2798 * Basic setup of slabs
2799 *******************************************************************/
2800
2801void __init kmem_cache_init(void)
2802{
2803 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002804 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002805
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002806 init_alloc_cpu();
2807
Christoph Lameter81819f02007-05-06 14:49:36 -07002808#ifdef CONFIG_NUMA
2809 /*
2810 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002811 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002812 * kmem_cache_open for slab_state == DOWN.
2813 */
2814 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2815 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002816 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002817 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002818
2819 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002820#endif
2821
2822 /* Able to allocate the per node structures */
2823 slab_state = PARTIAL;
2824
2825 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002826 if (KMALLOC_MIN_SIZE <= 64) {
2827 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002828 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002829 caches++;
2830 }
2831 if (KMALLOC_MIN_SIZE <= 128) {
2832 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002833 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002834 caches++;
2835 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002836
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002837 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002838 create_kmalloc_cache(&kmalloc_caches[i],
2839 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002840 caches++;
2841 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002842
Christoph Lameterf1b26332007-07-17 04:03:26 -07002843
2844 /*
2845 * Patch up the size_index table if we have strange large alignment
2846 * requirements for the kmalloc array. This is only the case for
2847 * mips it seems. The standard arches will not generate any code here.
2848 *
2849 * Largest permitted alignment is 256 bytes due to the way we
2850 * handle the index determination for the smaller caches.
2851 *
2852 * Make sure that nothing crazy happens if someone starts tinkering
2853 * around with ARCH_KMALLOC_MINALIGN
2854 */
2855 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2856 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2857
Christoph Lameter12ad6842007-07-17 04:03:28 -07002858 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002859 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2860
Christoph Lameter81819f02007-05-06 14:49:36 -07002861 slab_state = UP;
2862
2863 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002864 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002865 kmalloc_caches[i]. name =
2866 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2867
2868#ifdef CONFIG_SMP
2869 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002870 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2871 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2872#else
2873 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002874#endif
2875
Christoph Lameter81819f02007-05-06 14:49:36 -07002876
2877 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002878 " CPUs=%d, Nodes=%d\n",
2879 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002880 slub_min_order, slub_max_order, slub_min_objects,
2881 nr_cpu_ids, nr_node_ids);
2882}
2883
2884/*
2885 * Find a mergeable slab cache
2886 */
2887static int slab_unmergeable(struct kmem_cache *s)
2888{
2889 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2890 return 1;
2891
Christoph Lameterc59def92007-05-16 22:10:50 -07002892 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002893 return 1;
2894
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002895 /*
2896 * We may have set a slab to be unmergeable during bootstrap.
2897 */
2898 if (s->refcount < 0)
2899 return 1;
2900
Christoph Lameter81819f02007-05-06 14:49:36 -07002901 return 0;
2902}
2903
2904static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07002905 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002906 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002907{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002908 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002909
2910 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2911 return NULL;
2912
Christoph Lameterc59def92007-05-16 22:10:50 -07002913 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002914 return NULL;
2915
2916 size = ALIGN(size, sizeof(void *));
2917 align = calculate_alignment(flags, align, size);
2918 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07002919 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002920
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002921 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002922 if (slab_unmergeable(s))
2923 continue;
2924
2925 if (size > s->size)
2926 continue;
2927
Christoph Lameterba0268a2007-09-11 15:24:11 -07002928 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07002929 continue;
2930 /*
2931 * Check if alignment is compatible.
2932 * Courtesy of Adrian Drzewiecki
2933 */
2934 if ((s->size & ~(align -1)) != s->size)
2935 continue;
2936
2937 if (s->size - size >= sizeof(void *))
2938 continue;
2939
2940 return s;
2941 }
2942 return NULL;
2943}
2944
2945struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2946 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002947 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002948{
2949 struct kmem_cache *s;
2950
2951 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07002952 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002953 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07002954 int cpu;
2955
Christoph Lameter81819f02007-05-06 14:49:36 -07002956 s->refcount++;
2957 /*
2958 * Adjust the object sizes so that we clear
2959 * the complete object on kzalloc.
2960 */
2961 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07002962
2963 /*
2964 * And then we need to update the object size in the
2965 * per cpu structures
2966 */
2967 for_each_online_cpu(cpu)
2968 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter81819f02007-05-06 14:49:36 -07002969 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002970 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002971 if (sysfs_slab_alias(s, name))
2972 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002973 return s;
2974 }
2975 s = kmalloc(kmem_size, GFP_KERNEL);
2976 if (s) {
2977 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07002978 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002979 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002980 up_write(&slub_lock);
2981 if (sysfs_slab_add(s))
2982 goto err;
2983 return s;
2984 }
2985 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002986 }
2987 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002988
2989err:
Christoph Lameter81819f02007-05-06 14:49:36 -07002990 if (flags & SLAB_PANIC)
2991 panic("Cannot create slabcache %s\n", name);
2992 else
2993 s = NULL;
2994 return s;
2995}
2996EXPORT_SYMBOL(kmem_cache_create);
2997
Christoph Lameter81819f02007-05-06 14:49:36 -07002998#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07002999/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003000 * Use the cpu notifier to insure that the cpu slabs are flushed when
3001 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003002 */
3003static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3004 unsigned long action, void *hcpu)
3005{
3006 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003007 struct kmem_cache *s;
3008 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003009
3010 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003011 case CPU_UP_PREPARE:
3012 case CPU_UP_PREPARE_FROZEN:
3013 init_alloc_cpu_cpu(cpu);
3014 down_read(&slub_lock);
3015 list_for_each_entry(s, &slab_caches, list)
3016 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3017 GFP_KERNEL);
3018 up_read(&slub_lock);
3019 break;
3020
Christoph Lameter81819f02007-05-06 14:49:36 -07003021 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003022 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003023 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003024 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003025 down_read(&slub_lock);
3026 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003027 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3028
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003029 local_irq_save(flags);
3030 __flush_cpu_slab(s, cpu);
3031 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003032 free_kmem_cache_cpu(c, cpu);
3033 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003034 }
3035 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003036 break;
3037 default:
3038 break;
3039 }
3040 return NOTIFY_OK;
3041}
3042
3043static struct notifier_block __cpuinitdata slab_notifier =
3044 { &slab_cpuup_callback, NULL, 0 };
3045
3046#endif
3047
Christoph Lameter81819f02007-05-06 14:49:36 -07003048void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3049{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003050 struct kmem_cache *s;
3051
3052 if (unlikely(size > PAGE_SIZE / 2))
3053 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3054 get_order(size));
3055 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003056
Satyam Sharma2408c552007-10-16 01:24:44 -07003057 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003058 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003059
Christoph Lameterce15fea2007-07-17 04:03:28 -07003060 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003061}
3062
3063void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3064 int node, void *caller)
3065{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003066 struct kmem_cache *s;
3067
3068 if (unlikely(size > PAGE_SIZE / 2))
3069 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3070 get_order(size));
3071 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003072
Satyam Sharma2408c552007-10-16 01:24:44 -07003073 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003074 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003075
Christoph Lameterce15fea2007-07-17 04:03:28 -07003076 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003077}
3078
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01003079static unsigned long count_partial(struct kmem_cache_node *n)
3080{
3081 unsigned long flags;
3082 unsigned long x = 0;
3083 struct page *page;
3084
3085 spin_lock_irqsave(&n->list_lock, flags);
3086 list_for_each_entry(page, &n->partial, lru)
3087 x += page->inuse;
3088 spin_unlock_irqrestore(&n->list_lock, flags);
3089 return x;
3090}
3091
Christoph Lameter41ecc552007-05-09 02:32:44 -07003092#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003093static int validate_slab(struct kmem_cache *s, struct page *page,
3094 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003095{
3096 void *p;
3097 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003098
3099 if (!check_slab(s, page) ||
3100 !on_freelist(s, page, NULL))
3101 return 0;
3102
3103 /* Now we know that a valid freelist exists */
3104 bitmap_zero(map, s->objects);
3105
Christoph Lameter7656c722007-05-09 02:32:40 -07003106 for_each_free_object(p, s, page->freelist) {
3107 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003108 if (!check_object(s, page, p, 0))
3109 return 0;
3110 }
3111
Christoph Lameter7656c722007-05-09 02:32:40 -07003112 for_each_object(p, s, addr)
3113 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003114 if (!check_object(s, page, p, 1))
3115 return 0;
3116 return 1;
3117}
3118
Christoph Lameter434e2452007-07-17 04:03:30 -07003119static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3120 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003121{
3122 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003123 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003124 slab_unlock(page);
3125 } else
3126 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3127 s->name, page);
3128
3129 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003130 if (!SlabDebug(page))
3131 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003132 "on slab 0x%p\n", s->name, page);
3133 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003134 if (SlabDebug(page))
3135 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003136 "slab 0x%p\n", s->name, page);
3137 }
3138}
3139
Christoph Lameter434e2452007-07-17 04:03:30 -07003140static int validate_slab_node(struct kmem_cache *s,
3141 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003142{
3143 unsigned long count = 0;
3144 struct page *page;
3145 unsigned long flags;
3146
3147 spin_lock_irqsave(&n->list_lock, flags);
3148
3149 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003150 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003151 count++;
3152 }
3153 if (count != n->nr_partial)
3154 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3155 "counter=%ld\n", s->name, count, n->nr_partial);
3156
3157 if (!(s->flags & SLAB_STORE_USER))
3158 goto out;
3159
3160 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003161 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003162 count++;
3163 }
3164 if (count != atomic_long_read(&n->nr_slabs))
3165 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3166 "counter=%ld\n", s->name, count,
3167 atomic_long_read(&n->nr_slabs));
3168
3169out:
3170 spin_unlock_irqrestore(&n->list_lock, flags);
3171 return count;
3172}
3173
Christoph Lameter434e2452007-07-17 04:03:30 -07003174static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003175{
3176 int node;
3177 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003178 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3179 sizeof(unsigned long), GFP_KERNEL);
3180
3181 if (!map)
3182 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003183
3184 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003185 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003186 struct kmem_cache_node *n = get_node(s, node);
3187
Christoph Lameter434e2452007-07-17 04:03:30 -07003188 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003189 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003190 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003191 return count;
3192}
3193
Christoph Lameterb3459702007-05-09 02:32:41 -07003194#ifdef SLUB_RESILIENCY_TEST
3195static void resiliency_test(void)
3196{
3197 u8 *p;
3198
3199 printk(KERN_ERR "SLUB resiliency testing\n");
3200 printk(KERN_ERR "-----------------------\n");
3201 printk(KERN_ERR "A. Corruption after allocation\n");
3202
3203 p = kzalloc(16, GFP_KERNEL);
3204 p[16] = 0x12;
3205 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3206 " 0x12->0x%p\n\n", p + 16);
3207
3208 validate_slab_cache(kmalloc_caches + 4);
3209
3210 /* Hmmm... The next two are dangerous */
3211 p = kzalloc(32, GFP_KERNEL);
3212 p[32 + sizeof(void *)] = 0x34;
3213 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3214 " 0x34 -> -0x%p\n", p);
3215 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3216
3217 validate_slab_cache(kmalloc_caches + 5);
3218 p = kzalloc(64, GFP_KERNEL);
3219 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3220 *p = 0x56;
3221 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3222 p);
3223 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3224 validate_slab_cache(kmalloc_caches + 6);
3225
3226 printk(KERN_ERR "\nB. Corruption after free\n");
3227 p = kzalloc(128, GFP_KERNEL);
3228 kfree(p);
3229 *p = 0x78;
3230 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3231 validate_slab_cache(kmalloc_caches + 7);
3232
3233 p = kzalloc(256, GFP_KERNEL);
3234 kfree(p);
3235 p[50] = 0x9a;
3236 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
3237 validate_slab_cache(kmalloc_caches + 8);
3238
3239 p = kzalloc(512, GFP_KERNEL);
3240 kfree(p);
3241 p[512] = 0xab;
3242 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3243 validate_slab_cache(kmalloc_caches + 9);
3244}
3245#else
3246static void resiliency_test(void) {};
3247#endif
3248
Christoph Lameter88a420e2007-05-06 14:49:45 -07003249/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003250 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003251 * and freed.
3252 */
3253
3254struct location {
3255 unsigned long count;
3256 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003257 long long sum_time;
3258 long min_time;
3259 long max_time;
3260 long min_pid;
3261 long max_pid;
3262 cpumask_t cpus;
3263 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003264};
3265
3266struct loc_track {
3267 unsigned long max;
3268 unsigned long count;
3269 struct location *loc;
3270};
3271
3272static void free_loc_track(struct loc_track *t)
3273{
3274 if (t->max)
3275 free_pages((unsigned long)t->loc,
3276 get_order(sizeof(struct location) * t->max));
3277}
3278
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003279static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003280{
3281 struct location *l;
3282 int order;
3283
Christoph Lameter88a420e2007-05-06 14:49:45 -07003284 order = get_order(sizeof(struct location) * max);
3285
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003286 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003287 if (!l)
3288 return 0;
3289
3290 if (t->count) {
3291 memcpy(l, t->loc, sizeof(struct location) * t->count);
3292 free_loc_track(t);
3293 }
3294 t->max = max;
3295 t->loc = l;
3296 return 1;
3297}
3298
3299static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003300 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003301{
3302 long start, end, pos;
3303 struct location *l;
3304 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003305 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003306
3307 start = -1;
3308 end = t->count;
3309
3310 for ( ; ; ) {
3311 pos = start + (end - start + 1) / 2;
3312
3313 /*
3314 * There is nothing at "end". If we end up there
3315 * we need to add something to before end.
3316 */
3317 if (pos == end)
3318 break;
3319
3320 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003321 if (track->addr == caddr) {
3322
3323 l = &t->loc[pos];
3324 l->count++;
3325 if (track->when) {
3326 l->sum_time += age;
3327 if (age < l->min_time)
3328 l->min_time = age;
3329 if (age > l->max_time)
3330 l->max_time = age;
3331
3332 if (track->pid < l->min_pid)
3333 l->min_pid = track->pid;
3334 if (track->pid > l->max_pid)
3335 l->max_pid = track->pid;
3336
3337 cpu_set(track->cpu, l->cpus);
3338 }
3339 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003340 return 1;
3341 }
3342
Christoph Lameter45edfa52007-05-09 02:32:45 -07003343 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003344 end = pos;
3345 else
3346 start = pos;
3347 }
3348
3349 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003350 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003351 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003352 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003353 return 0;
3354
3355 l = t->loc + pos;
3356 if (pos < t->count)
3357 memmove(l + 1, l,
3358 (t->count - pos) * sizeof(struct location));
3359 t->count++;
3360 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003361 l->addr = track->addr;
3362 l->sum_time = age;
3363 l->min_time = age;
3364 l->max_time = age;
3365 l->min_pid = track->pid;
3366 l->max_pid = track->pid;
3367 cpus_clear(l->cpus);
3368 cpu_set(track->cpu, l->cpus);
3369 nodes_clear(l->nodes);
3370 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003371 return 1;
3372}
3373
3374static void process_slab(struct loc_track *t, struct kmem_cache *s,
3375 struct page *page, enum track_item alloc)
3376{
3377 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003378 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003379 void *p;
3380
3381 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003382 for_each_free_object(p, s, page->freelist)
3383 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003384
Christoph Lameter7656c722007-05-09 02:32:40 -07003385 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003386 if (!test_bit(slab_index(p, s, addr), map))
3387 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003388}
3389
3390static int list_locations(struct kmem_cache *s, char *buf,
3391 enum track_item alloc)
3392{
3393 int n = 0;
3394 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003395 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003396 int node;
3397
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003398 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003399 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003400 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003401
3402 /* Push back cpu slabs */
3403 flush_all(s);
3404
Christoph Lameterf64dc582007-10-16 01:25:33 -07003405 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003406 struct kmem_cache_node *n = get_node(s, node);
3407 unsigned long flags;
3408 struct page *page;
3409
Christoph Lameter9e869432007-08-22 14:01:56 -07003410 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003411 continue;
3412
3413 spin_lock_irqsave(&n->list_lock, flags);
3414 list_for_each_entry(page, &n->partial, lru)
3415 process_slab(&t, s, page, alloc);
3416 list_for_each_entry(page, &n->full, lru)
3417 process_slab(&t, s, page, alloc);
3418 spin_unlock_irqrestore(&n->list_lock, flags);
3419 }
3420
3421 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003422 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003423
3424 if (n > PAGE_SIZE - 100)
3425 break;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003426 n += sprintf(buf + n, "%7ld ", l->count);
3427
3428 if (l->addr)
3429 n += sprint_symbol(buf + n, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003430 else
3431 n += sprintf(buf + n, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003432
3433 if (l->sum_time != l->min_time) {
3434 unsigned long remainder;
3435
3436 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3437 l->min_time,
3438 div_long_long_rem(l->sum_time, l->count, &remainder),
3439 l->max_time);
3440 } else
3441 n += sprintf(buf + n, " age=%ld",
3442 l->min_time);
3443
3444 if (l->min_pid != l->max_pid)
3445 n += sprintf(buf + n, " pid=%ld-%ld",
3446 l->min_pid, l->max_pid);
3447 else
3448 n += sprintf(buf + n, " pid=%ld",
3449 l->min_pid);
3450
Christoph Lameter84966342007-06-23 17:16:32 -07003451 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3452 n < PAGE_SIZE - 60) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003453 n += sprintf(buf + n, " cpus=");
3454 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3455 l->cpus);
3456 }
3457
Christoph Lameter84966342007-06-23 17:16:32 -07003458 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3459 n < PAGE_SIZE - 60) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003460 n += sprintf(buf + n, " nodes=");
3461 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3462 l->nodes);
3463 }
3464
Christoph Lameter88a420e2007-05-06 14:49:45 -07003465 n += sprintf(buf + n, "\n");
3466 }
3467
3468 free_loc_track(&t);
3469 if (!t.count)
3470 n += sprintf(buf, "No data\n");
3471 return n;
3472}
3473
Christoph Lameter81819f02007-05-06 14:49:36 -07003474enum slab_stat_type {
3475 SL_FULL,
3476 SL_PARTIAL,
3477 SL_CPU,
3478 SL_OBJECTS
3479};
3480
3481#define SO_FULL (1 << SL_FULL)
3482#define SO_PARTIAL (1 << SL_PARTIAL)
3483#define SO_CPU (1 << SL_CPU)
3484#define SO_OBJECTS (1 << SL_OBJECTS)
3485
3486static unsigned long slab_objects(struct kmem_cache *s,
3487 char *buf, unsigned long flags)
3488{
3489 unsigned long total = 0;
3490 int cpu;
3491 int node;
3492 int x;
3493 unsigned long *nodes;
3494 unsigned long *per_cpu;
3495
3496 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3497 per_cpu = nodes + nr_node_ids;
3498
3499 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003500 struct page *page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003501 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003502 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003503
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003504 if (!c)
3505 continue;
3506
3507 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003508 node = c->node;
3509 if (node < 0)
3510 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003511 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003512 if (flags & SO_CPU) {
3513 int x = 0;
3514
3515 if (flags & SO_OBJECTS)
3516 x = page->inuse;
3517 else
3518 x = 1;
3519 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003520 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003521 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003522 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003523 }
3524 }
3525
Christoph Lameterf64dc582007-10-16 01:25:33 -07003526 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003527 struct kmem_cache_node *n = get_node(s, node);
3528
3529 if (flags & SO_PARTIAL) {
3530 if (flags & SO_OBJECTS)
3531 x = count_partial(n);
3532 else
3533 x = n->nr_partial;
3534 total += x;
3535 nodes[node] += x;
3536 }
3537
3538 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003539 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003540 - per_cpu[node]
3541 - n->nr_partial;
3542
3543 if (flags & SO_OBJECTS)
3544 x = full_slabs * s->objects;
3545 else
3546 x = full_slabs;
3547 total += x;
3548 nodes[node] += x;
3549 }
3550 }
3551
3552 x = sprintf(buf, "%lu", total);
3553#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003554 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003555 if (nodes[node])
3556 x += sprintf(buf + x, " N%d=%lu",
3557 node, nodes[node]);
3558#endif
3559 kfree(nodes);
3560 return x + sprintf(buf + x, "\n");
3561}
3562
3563static int any_slab_objects(struct kmem_cache *s)
3564{
3565 int node;
3566 int cpu;
3567
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003568 for_each_possible_cpu(cpu) {
3569 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003570
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003571 if (c && c->page)
3572 return 1;
3573 }
3574
3575 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003576 struct kmem_cache_node *n = get_node(s, node);
3577
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003578 if (!n)
3579 continue;
3580
Christoph Lameter9e869432007-08-22 14:01:56 -07003581 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003582 return 1;
3583 }
3584 return 0;
3585}
3586
3587#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3588#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3589
3590struct slab_attribute {
3591 struct attribute attr;
3592 ssize_t (*show)(struct kmem_cache *s, char *buf);
3593 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3594};
3595
3596#define SLAB_ATTR_RO(_name) \
3597 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3598
3599#define SLAB_ATTR(_name) \
3600 static struct slab_attribute _name##_attr = \
3601 __ATTR(_name, 0644, _name##_show, _name##_store)
3602
Christoph Lameter81819f02007-05-06 14:49:36 -07003603static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3604{
3605 return sprintf(buf, "%d\n", s->size);
3606}
3607SLAB_ATTR_RO(slab_size);
3608
3609static ssize_t align_show(struct kmem_cache *s, char *buf)
3610{
3611 return sprintf(buf, "%d\n", s->align);
3612}
3613SLAB_ATTR_RO(align);
3614
3615static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3616{
3617 return sprintf(buf, "%d\n", s->objsize);
3618}
3619SLAB_ATTR_RO(object_size);
3620
3621static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3622{
3623 return sprintf(buf, "%d\n", s->objects);
3624}
3625SLAB_ATTR_RO(objs_per_slab);
3626
3627static ssize_t order_show(struct kmem_cache *s, char *buf)
3628{
3629 return sprintf(buf, "%d\n", s->order);
3630}
3631SLAB_ATTR_RO(order);
3632
3633static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3634{
3635 if (s->ctor) {
3636 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3637
3638 return n + sprintf(buf + n, "\n");
3639 }
3640 return 0;
3641}
3642SLAB_ATTR_RO(ctor);
3643
Christoph Lameter81819f02007-05-06 14:49:36 -07003644static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3645{
3646 return sprintf(buf, "%d\n", s->refcount - 1);
3647}
3648SLAB_ATTR_RO(aliases);
3649
3650static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3651{
3652 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3653}
3654SLAB_ATTR_RO(slabs);
3655
3656static ssize_t partial_show(struct kmem_cache *s, char *buf)
3657{
3658 return slab_objects(s, buf, SO_PARTIAL);
3659}
3660SLAB_ATTR_RO(partial);
3661
3662static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3663{
3664 return slab_objects(s, buf, SO_CPU);
3665}
3666SLAB_ATTR_RO(cpu_slabs);
3667
3668static ssize_t objects_show(struct kmem_cache *s, char *buf)
3669{
3670 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3671}
3672SLAB_ATTR_RO(objects);
3673
3674static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3675{
3676 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3677}
3678
3679static ssize_t sanity_checks_store(struct kmem_cache *s,
3680 const char *buf, size_t length)
3681{
3682 s->flags &= ~SLAB_DEBUG_FREE;
3683 if (buf[0] == '1')
3684 s->flags |= SLAB_DEBUG_FREE;
3685 return length;
3686}
3687SLAB_ATTR(sanity_checks);
3688
3689static ssize_t trace_show(struct kmem_cache *s, char *buf)
3690{
3691 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3692}
3693
3694static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3695 size_t length)
3696{
3697 s->flags &= ~SLAB_TRACE;
3698 if (buf[0] == '1')
3699 s->flags |= SLAB_TRACE;
3700 return length;
3701}
3702SLAB_ATTR(trace);
3703
3704static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3705{
3706 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3707}
3708
3709static ssize_t reclaim_account_store(struct kmem_cache *s,
3710 const char *buf, size_t length)
3711{
3712 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3713 if (buf[0] == '1')
3714 s->flags |= SLAB_RECLAIM_ACCOUNT;
3715 return length;
3716}
3717SLAB_ATTR(reclaim_account);
3718
3719static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3720{
Christoph Lameter5af60832007-05-06 14:49:56 -07003721 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003722}
3723SLAB_ATTR_RO(hwcache_align);
3724
3725#ifdef CONFIG_ZONE_DMA
3726static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3727{
3728 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3729}
3730SLAB_ATTR_RO(cache_dma);
3731#endif
3732
3733static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3734{
3735 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3736}
3737SLAB_ATTR_RO(destroy_by_rcu);
3738
3739static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3740{
3741 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3742}
3743
3744static ssize_t red_zone_store(struct kmem_cache *s,
3745 const char *buf, size_t length)
3746{
3747 if (any_slab_objects(s))
3748 return -EBUSY;
3749
3750 s->flags &= ~SLAB_RED_ZONE;
3751 if (buf[0] == '1')
3752 s->flags |= SLAB_RED_ZONE;
3753 calculate_sizes(s);
3754 return length;
3755}
3756SLAB_ATTR(red_zone);
3757
3758static ssize_t poison_show(struct kmem_cache *s, char *buf)
3759{
3760 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3761}
3762
3763static ssize_t poison_store(struct kmem_cache *s,
3764 const char *buf, size_t length)
3765{
3766 if (any_slab_objects(s))
3767 return -EBUSY;
3768
3769 s->flags &= ~SLAB_POISON;
3770 if (buf[0] == '1')
3771 s->flags |= SLAB_POISON;
3772 calculate_sizes(s);
3773 return length;
3774}
3775SLAB_ATTR(poison);
3776
3777static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3778{
3779 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3780}
3781
3782static ssize_t store_user_store(struct kmem_cache *s,
3783 const char *buf, size_t length)
3784{
3785 if (any_slab_objects(s))
3786 return -EBUSY;
3787
3788 s->flags &= ~SLAB_STORE_USER;
3789 if (buf[0] == '1')
3790 s->flags |= SLAB_STORE_USER;
3791 calculate_sizes(s);
3792 return length;
3793}
3794SLAB_ATTR(store_user);
3795
Christoph Lameter53e15af2007-05-06 14:49:43 -07003796static ssize_t validate_show(struct kmem_cache *s, char *buf)
3797{
3798 return 0;
3799}
3800
3801static ssize_t validate_store(struct kmem_cache *s,
3802 const char *buf, size_t length)
3803{
Christoph Lameter434e2452007-07-17 04:03:30 -07003804 int ret = -EINVAL;
3805
3806 if (buf[0] == '1') {
3807 ret = validate_slab_cache(s);
3808 if (ret >= 0)
3809 ret = length;
3810 }
3811 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003812}
3813SLAB_ATTR(validate);
3814
Christoph Lameter2086d262007-05-06 14:49:46 -07003815static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3816{
3817 return 0;
3818}
3819
3820static ssize_t shrink_store(struct kmem_cache *s,
3821 const char *buf, size_t length)
3822{
3823 if (buf[0] == '1') {
3824 int rc = kmem_cache_shrink(s);
3825
3826 if (rc)
3827 return rc;
3828 } else
3829 return -EINVAL;
3830 return length;
3831}
3832SLAB_ATTR(shrink);
3833
Christoph Lameter88a420e2007-05-06 14:49:45 -07003834static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3835{
3836 if (!(s->flags & SLAB_STORE_USER))
3837 return -ENOSYS;
3838 return list_locations(s, buf, TRACK_ALLOC);
3839}
3840SLAB_ATTR_RO(alloc_calls);
3841
3842static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3843{
3844 if (!(s->flags & SLAB_STORE_USER))
3845 return -ENOSYS;
3846 return list_locations(s, buf, TRACK_FREE);
3847}
3848SLAB_ATTR_RO(free_calls);
3849
Christoph Lameter81819f02007-05-06 14:49:36 -07003850#ifdef CONFIG_NUMA
3851static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3852{
3853 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3854}
3855
3856static ssize_t defrag_ratio_store(struct kmem_cache *s,
3857 const char *buf, size_t length)
3858{
3859 int n = simple_strtoul(buf, NULL, 10);
3860
3861 if (n < 100)
3862 s->defrag_ratio = n * 10;
3863 return length;
3864}
3865SLAB_ATTR(defrag_ratio);
3866#endif
3867
3868static struct attribute * slab_attrs[] = {
3869 &slab_size_attr.attr,
3870 &object_size_attr.attr,
3871 &objs_per_slab_attr.attr,
3872 &order_attr.attr,
3873 &objects_attr.attr,
3874 &slabs_attr.attr,
3875 &partial_attr.attr,
3876 &cpu_slabs_attr.attr,
3877 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003878 &aliases_attr.attr,
3879 &align_attr.attr,
3880 &sanity_checks_attr.attr,
3881 &trace_attr.attr,
3882 &hwcache_align_attr.attr,
3883 &reclaim_account_attr.attr,
3884 &destroy_by_rcu_attr.attr,
3885 &red_zone_attr.attr,
3886 &poison_attr.attr,
3887 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07003888 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07003889 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07003890 &alloc_calls_attr.attr,
3891 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003892#ifdef CONFIG_ZONE_DMA
3893 &cache_dma_attr.attr,
3894#endif
3895#ifdef CONFIG_NUMA
3896 &defrag_ratio_attr.attr,
3897#endif
3898 NULL
3899};
3900
3901static struct attribute_group slab_attr_group = {
3902 .attrs = slab_attrs,
3903};
3904
3905static ssize_t slab_attr_show(struct kobject *kobj,
3906 struct attribute *attr,
3907 char *buf)
3908{
3909 struct slab_attribute *attribute;
3910 struct kmem_cache *s;
3911 int err;
3912
3913 attribute = to_slab_attr(attr);
3914 s = to_slab(kobj);
3915
3916 if (!attribute->show)
3917 return -EIO;
3918
3919 err = attribute->show(s, buf);
3920
3921 return err;
3922}
3923
3924static ssize_t slab_attr_store(struct kobject *kobj,
3925 struct attribute *attr,
3926 const char *buf, size_t len)
3927{
3928 struct slab_attribute *attribute;
3929 struct kmem_cache *s;
3930 int err;
3931
3932 attribute = to_slab_attr(attr);
3933 s = to_slab(kobj);
3934
3935 if (!attribute->store)
3936 return -EIO;
3937
3938 err = attribute->store(s, buf, len);
3939
3940 return err;
3941}
3942
3943static struct sysfs_ops slab_sysfs_ops = {
3944 .show = slab_attr_show,
3945 .store = slab_attr_store,
3946};
3947
3948static struct kobj_type slab_ktype = {
3949 .sysfs_ops = &slab_sysfs_ops,
3950};
3951
3952static int uevent_filter(struct kset *kset, struct kobject *kobj)
3953{
3954 struct kobj_type *ktype = get_ktype(kobj);
3955
3956 if (ktype == &slab_ktype)
3957 return 1;
3958 return 0;
3959}
3960
3961static struct kset_uevent_ops slab_uevent_ops = {
3962 .filter = uevent_filter,
3963};
3964
Adrian Bunk5af328a2007-07-17 04:03:27 -07003965static decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
Christoph Lameter81819f02007-05-06 14:49:36 -07003966
3967#define ID_STR_LENGTH 64
3968
3969/* Create a unique string id for a slab cache:
3970 * format
3971 * :[flags-]size:[memory address of kmemcache]
3972 */
3973static char *create_unique_id(struct kmem_cache *s)
3974{
3975 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3976 char *p = name;
3977
3978 BUG_ON(!name);
3979
3980 *p++ = ':';
3981 /*
3982 * First flags affecting slabcache operations. We will only
3983 * get here for aliasable slabs so we do not need to support
3984 * too many flags. The flags here must cover all flags that
3985 * are matched during merging to guarantee that the id is
3986 * unique.
3987 */
3988 if (s->flags & SLAB_CACHE_DMA)
3989 *p++ = 'd';
3990 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3991 *p++ = 'a';
3992 if (s->flags & SLAB_DEBUG_FREE)
3993 *p++ = 'F';
3994 if (p != name + 1)
3995 *p++ = '-';
3996 p += sprintf(p, "%07d", s->size);
3997 BUG_ON(p > name + ID_STR_LENGTH - 1);
3998 return name;
3999}
4000
4001static int sysfs_slab_add(struct kmem_cache *s)
4002{
4003 int err;
4004 const char *name;
4005 int unmergeable;
4006
4007 if (slab_state < SYSFS)
4008 /* Defer until later */
4009 return 0;
4010
4011 unmergeable = slab_unmergeable(s);
4012 if (unmergeable) {
4013 /*
4014 * Slabcache can never be merged so we can use the name proper.
4015 * This is typically the case for debug situations. In that
4016 * case we can catch duplicate names easily.
4017 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07004018 sysfs_remove_link(&slab_subsys.kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004019 name = s->name;
4020 } else {
4021 /*
4022 * Create a unique name for the slab as a target
4023 * for the symlinks.
4024 */
4025 name = create_unique_id(s);
4026 }
4027
4028 kobj_set_kset_s(s, slab_subsys);
4029 kobject_set_name(&s->kobj, name);
4030 kobject_init(&s->kobj);
4031 err = kobject_add(&s->kobj);
4032 if (err)
4033 return err;
4034
4035 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4036 if (err)
4037 return err;
4038 kobject_uevent(&s->kobj, KOBJ_ADD);
4039 if (!unmergeable) {
4040 /* Setup first alias */
4041 sysfs_slab_alias(s, s->name);
4042 kfree(name);
4043 }
4044 return 0;
4045}
4046
4047static void sysfs_slab_remove(struct kmem_cache *s)
4048{
4049 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4050 kobject_del(&s->kobj);
4051}
4052
4053/*
4054 * Need to buffer aliases during bootup until sysfs becomes
4055 * available lest we loose that information.
4056 */
4057struct saved_alias {
4058 struct kmem_cache *s;
4059 const char *name;
4060 struct saved_alias *next;
4061};
4062
Adrian Bunk5af328a2007-07-17 04:03:27 -07004063static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004064
4065static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4066{
4067 struct saved_alias *al;
4068
4069 if (slab_state == SYSFS) {
4070 /*
4071 * If we have a leftover link then remove it.
4072 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07004073 sysfs_remove_link(&slab_subsys.kobj, name);
4074 return sysfs_create_link(&slab_subsys.kobj,
Christoph Lameter81819f02007-05-06 14:49:36 -07004075 &s->kobj, name);
4076 }
4077
4078 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4079 if (!al)
4080 return -ENOMEM;
4081
4082 al->s = s;
4083 al->name = name;
4084 al->next = alias_list;
4085 alias_list = al;
4086 return 0;
4087}
4088
4089static int __init slab_sysfs_init(void)
4090{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004091 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004092 int err;
4093
4094 err = subsystem_register(&slab_subsys);
4095 if (err) {
4096 printk(KERN_ERR "Cannot register slab subsystem.\n");
4097 return -ENOSYS;
4098 }
4099
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004100 slab_state = SYSFS;
4101
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004102 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004103 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004104 if (err)
4105 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4106 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004107 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004108
4109 while (alias_list) {
4110 struct saved_alias *al = alias_list;
4111
4112 alias_list = alias_list->next;
4113 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004114 if (err)
4115 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4116 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004117 kfree(al);
4118 }
4119
4120 resiliency_test();
4121 return 0;
4122}
4123
4124__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004125#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004126
4127/*
4128 * The /proc/slabinfo ABI
4129 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004130#ifdef CONFIG_SLABINFO
4131
4132ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4133 size_t count, loff_t *ppos)
4134{
4135 return -EINVAL;
4136}
4137
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004138
4139static void print_slabinfo_header(struct seq_file *m)
4140{
4141 seq_puts(m, "slabinfo - version: 2.1\n");
4142 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4143 "<objperslab> <pagesperslab>");
4144 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4145 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4146 seq_putc(m, '\n');
4147}
4148
4149static void *s_start(struct seq_file *m, loff_t *pos)
4150{
4151 loff_t n = *pos;
4152
4153 down_read(&slub_lock);
4154 if (!n)
4155 print_slabinfo_header(m);
4156
4157 return seq_list_start(&slab_caches, *pos);
4158}
4159
4160static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4161{
4162 return seq_list_next(p, &slab_caches, pos);
4163}
4164
4165static void s_stop(struct seq_file *m, void *p)
4166{
4167 up_read(&slub_lock);
4168}
4169
4170static int s_show(struct seq_file *m, void *p)
4171{
4172 unsigned long nr_partials = 0;
4173 unsigned long nr_slabs = 0;
4174 unsigned long nr_inuse = 0;
4175 unsigned long nr_objs;
4176 struct kmem_cache *s;
4177 int node;
4178
4179 s = list_entry(p, struct kmem_cache, list);
4180
4181 for_each_online_node(node) {
4182 struct kmem_cache_node *n = get_node(s, node);
4183
4184 if (!n)
4185 continue;
4186
4187 nr_partials += n->nr_partial;
4188 nr_slabs += atomic_long_read(&n->nr_slabs);
4189 nr_inuse += count_partial(n);
4190 }
4191
4192 nr_objs = nr_slabs * s->objects;
4193 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4194
4195 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4196 nr_objs, s->size, s->objects, (1 << s->order));
4197 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4198 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4199 0UL);
4200 seq_putc(m, '\n');
4201 return 0;
4202}
4203
4204const struct seq_operations slabinfo_op = {
4205 .start = s_start,
4206 .next = s_next,
4207 .stop = s_stop,
4208 .show = s_show,
4209};
4210
Linus Torvalds158a9622008-01-02 13:04:48 -08004211#endif /* CONFIG_SLABINFO */