blob: bcdb2c8941a5fcb428d4d3d256106f29e921debd [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 Lametere95eed52007-05-06 14:49:44 -0700175#define MIN_PARTIAL 2
176
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;
1083 void *end;
1084 void *last;
1085 void *p;
1086
Christoph Lameter6cb06222007-10-16 01:25:41 -07001087 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001088
Christoph Lameter6cb06222007-10-16 01:25:41 -07001089 page = allocate_slab(s,
1090 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001091 if (!page)
1092 goto out;
1093
1094 n = get_node(s, page_to_nid(page));
1095 if (n)
1096 atomic_long_inc(&n->nr_slabs);
Christoph Lameter81819f02007-05-06 14:49:36 -07001097 page->slab = s;
1098 page->flags |= 1 << PG_slab;
1099 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1100 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001101 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001102
1103 start = page_address(page);
1104 end = start + s->objects * s->size;
1105
1106 if (unlikely(s->flags & SLAB_POISON))
1107 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1108
1109 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001110 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001111 setup_object(s, page, last);
1112 set_freepointer(s, last, p);
1113 last = p;
1114 }
1115 setup_object(s, page, last);
1116 set_freepointer(s, last, NULL);
1117
1118 page->freelist = start;
1119 page->inuse = 0;
1120out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001121 return page;
1122}
1123
1124static void __free_slab(struct kmem_cache *s, struct page *page)
1125{
1126 int pages = 1 << s->order;
1127
Christoph Lameterc59def92007-05-16 22:10:50 -07001128 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001129 void *p;
1130
1131 slab_pad_check(s, page);
Christoph Lameterc59def92007-05-16 22:10:50 -07001132 for_each_object(p, s, page_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001133 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001134 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001135 }
1136
1137 mod_zone_page_state(page_zone(page),
1138 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1139 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1140 - pages);
1141
Christoph Lameter81819f02007-05-06 14:49:36 -07001142 __free_pages(page, s->order);
1143}
1144
1145static void rcu_free_slab(struct rcu_head *h)
1146{
1147 struct page *page;
1148
1149 page = container_of((struct list_head *)h, struct page, lru);
1150 __free_slab(page->slab, page);
1151}
1152
1153static void free_slab(struct kmem_cache *s, struct page *page)
1154{
1155 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1156 /*
1157 * RCU free overloads the RCU head over the LRU
1158 */
1159 struct rcu_head *head = (void *)&page->lru;
1160
1161 call_rcu(head, rcu_free_slab);
1162 } else
1163 __free_slab(s, page);
1164}
1165
1166static void discard_slab(struct kmem_cache *s, struct page *page)
1167{
1168 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1169
1170 atomic_long_dec(&n->nr_slabs);
1171 reset_page_mapcount(page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001172 __ClearPageSlab(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001173 free_slab(s, page);
1174}
1175
1176/*
1177 * Per slab locking using the pagelock
1178 */
1179static __always_inline void slab_lock(struct page *page)
1180{
1181 bit_spin_lock(PG_locked, &page->flags);
1182}
1183
1184static __always_inline void slab_unlock(struct page *page)
1185{
1186 bit_spin_unlock(PG_locked, &page->flags);
1187}
1188
1189static __always_inline int slab_trylock(struct page *page)
1190{
1191 int rc = 1;
1192
1193 rc = bit_spin_trylock(PG_locked, &page->flags);
1194 return rc;
1195}
1196
1197/*
1198 * Management of partially allocated slabs
1199 */
Christoph Lametere95eed52007-05-06 14:49:44 -07001200static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001201{
Christoph Lametere95eed52007-05-06 14:49:44 -07001202 spin_lock(&n->list_lock);
1203 n->nr_partial++;
1204 list_add_tail(&page->lru, &n->partial);
1205 spin_unlock(&n->list_lock);
1206}
Christoph Lameter81819f02007-05-06 14:49:36 -07001207
Christoph Lametere95eed52007-05-06 14:49:44 -07001208static void add_partial(struct kmem_cache_node *n, struct page *page)
1209{
Christoph Lameter81819f02007-05-06 14:49:36 -07001210 spin_lock(&n->list_lock);
1211 n->nr_partial++;
1212 list_add(&page->lru, &n->partial);
1213 spin_unlock(&n->list_lock);
1214}
1215
1216static void remove_partial(struct kmem_cache *s,
1217 struct page *page)
1218{
1219 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1220
1221 spin_lock(&n->list_lock);
1222 list_del(&page->lru);
1223 n->nr_partial--;
1224 spin_unlock(&n->list_lock);
1225}
1226
1227/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001228 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001229 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001230 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001231 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001232static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001233{
1234 if (slab_trylock(page)) {
1235 list_del(&page->lru);
1236 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001237 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001238 return 1;
1239 }
1240 return 0;
1241}
1242
1243/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001244 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001245 */
1246static struct page *get_partial_node(struct kmem_cache_node *n)
1247{
1248 struct page *page;
1249
1250 /*
1251 * Racy check. If we mistakenly see no partial slabs then we
1252 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001253 * partial slab and there is none available then get_partials()
1254 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001255 */
1256 if (!n || !n->nr_partial)
1257 return NULL;
1258
1259 spin_lock(&n->list_lock);
1260 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001261 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001262 goto out;
1263 page = NULL;
1264out:
1265 spin_unlock(&n->list_lock);
1266 return page;
1267}
1268
1269/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001270 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001271 */
1272static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1273{
1274#ifdef CONFIG_NUMA
1275 struct zonelist *zonelist;
1276 struct zone **z;
1277 struct page *page;
1278
1279 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001280 * The defrag ratio allows a configuration of the tradeoffs between
1281 * inter node defragmentation and node local allocations. A lower
1282 * defrag_ratio increases the tendency to do local allocations
1283 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001284 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001285 * If the defrag_ratio is set to 0 then kmalloc() always
1286 * returns node local objects. If the ratio is higher then kmalloc()
1287 * may return off node objects because partial slabs are obtained
1288 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001289 *
1290 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001291 * defrag_ratio = 1000) then every (well almost) allocation will
1292 * first attempt to defrag slab caches on other nodes. This means
1293 * scanning over all nodes to look for partial slabs which may be
1294 * expensive if we do it every time we are trying to find a slab
1295 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001296 */
1297 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1298 return NULL;
1299
1300 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1301 ->node_zonelists[gfp_zone(flags)];
1302 for (z = zonelist->zones; *z; z++) {
1303 struct kmem_cache_node *n;
1304
1305 n = get_node(s, zone_to_nid(*z));
1306
1307 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001308 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001309 page = get_partial_node(n);
1310 if (page)
1311 return page;
1312 }
1313 }
1314#endif
1315 return NULL;
1316}
1317
1318/*
1319 * Get a partial page, lock it and return it.
1320 */
1321static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1322{
1323 struct page *page;
1324 int searchnode = (node == -1) ? numa_node_id() : node;
1325
1326 page = get_partial_node(get_node(s, searchnode));
1327 if (page || (flags & __GFP_THISNODE))
1328 return page;
1329
1330 return get_any_partial(s, flags);
1331}
1332
1333/*
1334 * Move a page back to the lists.
1335 *
1336 * Must be called with the slab lock held.
1337 *
1338 * On exit the slab lock will have been dropped.
1339 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001340static void unfreeze_slab(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001341{
Christoph Lametere95eed52007-05-06 14:49:44 -07001342 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1343
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001344 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001345 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001346
Christoph Lameter81819f02007-05-06 14:49:36 -07001347 if (page->freelist)
Christoph Lametere95eed52007-05-06 14:49:44 -07001348 add_partial(n, page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001349 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
Christoph Lametere95eed52007-05-06 14:49:44 -07001350 add_full(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001351 slab_unlock(page);
Christoph Lametere95eed52007-05-06 14:49:44 -07001352
Christoph Lameter81819f02007-05-06 14:49:36 -07001353 } else {
Christoph Lametere95eed52007-05-06 14:49:44 -07001354 if (n->nr_partial < MIN_PARTIAL) {
1355 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001356 * Adding an empty slab to the partial slabs in order
1357 * to avoid page allocator overhead. This slab needs
1358 * to come after the other slabs with objects in
1359 * order to fill them up. That way the size of the
1360 * partial list stays small. kmem_cache_shrink can
1361 * reclaim empty slabs from the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001362 */
1363 add_partial_tail(n, page);
1364 slab_unlock(page);
1365 } else {
1366 slab_unlock(page);
1367 discard_slab(s, page);
1368 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001369 }
1370}
1371
1372/*
1373 * Remove the cpu slab
1374 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001375static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001376{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001377 struct page *page = c->page;
Christoph Lameter894b8782007-05-10 03:15:16 -07001378 /*
1379 * Merge cpu freelist into freelist. Typically we get here
1380 * because both freelists are empty. So this is unlikely
1381 * to occur.
1382 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001383 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001384 void **object;
1385
1386 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001387 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001388 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001389
1390 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001391 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001392 page->freelist = object;
1393 page->inuse--;
1394 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001395 c->page = NULL;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001396 unfreeze_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001397}
1398
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001399static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001400{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001401 slab_lock(c->page);
1402 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001403}
1404
1405/*
1406 * Flush cpu slab.
1407 * Called from IPI handler with interrupts disabled.
1408 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001409static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001410{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001411 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001412
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001413 if (likely(c && c->page))
1414 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001415}
1416
1417static void flush_cpu_slab(void *d)
1418{
1419 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001420
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001421 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001422}
1423
1424static void flush_all(struct kmem_cache *s)
1425{
1426#ifdef CONFIG_SMP
1427 on_each_cpu(flush_cpu_slab, s, 1, 1);
1428#else
1429 unsigned long flags;
1430
1431 local_irq_save(flags);
1432 flush_cpu_slab(s);
1433 local_irq_restore(flags);
1434#endif
1435}
1436
1437/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001438 * Check if the objects in a per cpu structure fit numa
1439 * locality expectations.
1440 */
1441static inline int node_match(struct kmem_cache_cpu *c, int node)
1442{
1443#ifdef CONFIG_NUMA
1444 if (node != -1 && c->node != node)
1445 return 0;
1446#endif
1447 return 1;
1448}
1449
1450/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001451 * Slow path. The lockless freelist is empty or we need to perform
1452 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001453 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001454 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001455 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001456 * Processing is still very fast if new objects have been freed to the
1457 * regular freelist. In that case we simply take over the regular freelist
1458 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001459 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001460 * If that is not working then we fall back to the partial lists. We take the
1461 * first element of the freelist as the object to allocate now and move the
1462 * rest of the freelist to the lockless freelist.
1463 *
1464 * And if we were unable to get a new slab from the partial slab lists then
1465 * we need to allocate a new slab. This is slowest path since we may sleep.
Christoph Lameter81819f02007-05-06 14:49:36 -07001466 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001467static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001468 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001469{
Christoph Lameter81819f02007-05-06 14:49:36 -07001470 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001471 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001472
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001473 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001474 goto new_slab;
1475
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001476 slab_lock(c->page);
1477 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001478 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001479load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001480 object = c->page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001481 if (unlikely(!object))
1482 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001483 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001484 goto debug;
1485
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001486 object = c->page->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001487 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001488 c->page->inuse = s->objects;
1489 c->page->freelist = NULL;
1490 c->node = page_to_nid(c->page);
1491 slab_unlock(c->page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001492 return object;
1493
1494another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001495 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001496
1497new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001498 new = get_partial(s, gfpflags, node);
1499 if (new) {
1500 c->page = new;
Christoph Lameter894b8782007-05-10 03:15:16 -07001501 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001502 }
1503
Christoph Lameterb811c202007-10-16 23:25:51 -07001504 if (gfpflags & __GFP_WAIT)
1505 local_irq_enable();
1506
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001507 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001508
1509 if (gfpflags & __GFP_WAIT)
1510 local_irq_disable();
1511
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001512 if (new) {
1513 c = get_cpu_slab(s, smp_processor_id());
1514 if (c->page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001515 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001516 * Someone else populated the cpu_slab while we
1517 * enabled interrupts, or we have gotten scheduled
1518 * on another cpu. The page may not be on the
1519 * requested node even if __GFP_THISNODE was
1520 * specified. So we need to recheck.
Christoph Lameter81819f02007-05-06 14:49:36 -07001521 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001522 if (node_match(c, node)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001523 /*
1524 * Current cpuslab is acceptable and we
1525 * want the current one since its cache hot
1526 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001527 discard_slab(s, new);
1528 slab_lock(c->page);
Christoph Lameter894b8782007-05-10 03:15:16 -07001529 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001530 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001531 /* New slab does not fit our expectations */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001532 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001533 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001534 slab_lock(new);
1535 SetSlabFrozen(new);
1536 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001537 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001538 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001539 return NULL;
1540debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001541 object = c->page->freelist;
1542 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001543 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001544
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001545 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001546 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001547 c->node = -1;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001548 slab_unlock(c->page);
Christoph Lameter894b8782007-05-10 03:15:16 -07001549 return object;
1550}
1551
1552/*
1553 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1554 * have the fastpath folded into their functions. So no function call
1555 * overhead for requests that can be satisfied on the fastpath.
1556 *
1557 * The fastpath works by first checking if the lockless freelist can be used.
1558 * If not then __slab_alloc is called for slow processing.
1559 *
1560 * Otherwise we can simply pick the next object from the lockless free list.
1561 */
1562static void __always_inline *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001563 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001564{
Christoph Lameter894b8782007-05-10 03:15:16 -07001565 void **object;
1566 unsigned long flags;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001567 struct kmem_cache_cpu *c;
Christoph Lameter894b8782007-05-10 03:15:16 -07001568
1569 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001570 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001571 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001572
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001573 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001574
1575 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001576 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001577 c->freelist = object[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001578 }
1579 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001580
1581 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001582 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001583
Christoph Lameter894b8782007-05-10 03:15:16 -07001584 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001585}
1586
1587void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1588{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001589 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001590}
1591EXPORT_SYMBOL(kmem_cache_alloc);
1592
1593#ifdef CONFIG_NUMA
1594void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1595{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001596 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001597}
1598EXPORT_SYMBOL(kmem_cache_alloc_node);
1599#endif
1600
1601/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001602 * Slow patch handling. This may still be called frequently since objects
1603 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001604 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001605 * So we still attempt to reduce cache line usage. Just take the slab
1606 * lock and free the item. If there is no additional partial page
1607 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001608 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001609static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001610 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001611{
1612 void *prior;
1613 void **object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07001614
Christoph Lameter81819f02007-05-06 14:49:36 -07001615 slab_lock(page);
1616
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001617 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001618 goto debug;
1619checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001620 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001621 page->freelist = object;
1622 page->inuse--;
1623
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001624 if (unlikely(SlabFrozen(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001625 goto out_unlock;
1626
1627 if (unlikely(!page->inuse))
1628 goto slab_empty;
1629
1630 /*
1631 * Objects left in the slab. If it
1632 * was not on the partial list before
1633 * then add it.
1634 */
1635 if (unlikely(!prior))
Christoph Lametere95eed52007-05-06 14:49:44 -07001636 add_partial(get_node(s, page_to_nid(page)), page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001637
1638out_unlock:
1639 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001640 return;
1641
1642slab_empty:
1643 if (prior)
1644 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001645 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001646 */
1647 remove_partial(s, page);
1648
1649 slab_unlock(page);
1650 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001651 return;
1652
1653debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001654 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001655 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001656 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001657}
1658
Christoph Lameter894b8782007-05-10 03:15:16 -07001659/*
1660 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1661 * can perform fastpath freeing without additional function calls.
1662 *
1663 * The fastpath is only possible if we are freeing to the current cpu slab
1664 * of this processor. This typically the case if we have just allocated
1665 * the item before.
1666 *
1667 * If fastpath is not possible then fall back to __slab_free where we deal
1668 * with all sorts of special processing.
1669 */
1670static void __always_inline slab_free(struct kmem_cache *s,
1671 struct page *page, void *x, void *addr)
1672{
1673 void **object = (void *)x;
1674 unsigned long flags;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001675 struct kmem_cache_cpu *c;
Christoph Lameter894b8782007-05-10 03:15:16 -07001676
1677 local_irq_save(flags);
Peter Zijlstra02febdf2007-07-26 20:01:38 +02001678 debug_check_no_locks_freed(object, s->objsize);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001679 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001680 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001681 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001682 c->freelist = object;
Christoph Lameter894b8782007-05-10 03:15:16 -07001683 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001684 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001685
1686 local_irq_restore(flags);
1687}
1688
Christoph Lameter81819f02007-05-06 14:49:36 -07001689void kmem_cache_free(struct kmem_cache *s, void *x)
1690{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001691 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001692
Christoph Lameterb49af682007-05-06 14:49:41 -07001693 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001694
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001695 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001696}
1697EXPORT_SYMBOL(kmem_cache_free);
1698
1699/* Figure out on which slab object the object resides */
1700static struct page *get_object_page(const void *x)
1701{
Christoph Lameterb49af682007-05-06 14:49:41 -07001702 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001703
1704 if (!PageSlab(page))
1705 return NULL;
1706
1707 return page;
1708}
1709
1710/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001711 * Object placement in a slab is made very easy because we always start at
1712 * offset 0. If we tune the size of the object to the alignment then we can
1713 * get the required alignment by putting one properly sized object after
1714 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001715 *
1716 * Notice that the allocation order determines the sizes of the per cpu
1717 * caches. Each processor has always one slab available for allocations.
1718 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001719 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001720 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001721 */
1722
1723/*
1724 * Mininum / Maximum order of slab pages. This influences locking overhead
1725 * and slab fragmentation. A higher order reduces the number of partial slabs
1726 * and increases the number of allocations possible without having to
1727 * take the list_lock.
1728 */
1729static int slub_min_order;
1730static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001731static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1732
1733/*
1734 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001735 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001736 */
1737static int slub_nomerge;
1738
1739/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001740 * Calculate the order of allocation given an slab object size.
1741 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001742 * The order of allocation has significant impact on performance and other
1743 * system components. Generally order 0 allocations should be preferred since
1744 * order 0 does not cause fragmentation in the page allocator. Larger objects
1745 * be problematic to put into order 0 slabs because there may be too much
1746 * unused space left. We go to a higher order if more than 1/8th of the slab
1747 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001748 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001749 * In order to reach satisfactory performance we must ensure that a minimum
1750 * number of objects is in one slab. Otherwise we may generate too much
1751 * activity on the partial lists which requires taking the list_lock. This is
1752 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001753 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001754 * slub_max_order specifies the order where we begin to stop considering the
1755 * number of objects in a slab as critical. If we reach slub_max_order then
1756 * we try to keep the page order as low as possible. So we accept more waste
1757 * of space in favor of a small page order.
1758 *
1759 * Higher order allocations also allow the placement of more objects in a
1760 * slab and thereby reduce object handling overhead. If the user has
1761 * requested a higher mininum order then we start with that one instead of
1762 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001763 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001764static inline int slab_order(int size, int min_objects,
1765 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001766{
1767 int order;
1768 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001769 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001770
Christoph Lameter6300ea72007-07-17 04:03:20 -07001771 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001772 fls(min_objects * size - 1) - PAGE_SHIFT);
1773 order <= max_order; order++) {
1774
Christoph Lameter81819f02007-05-06 14:49:36 -07001775 unsigned long slab_size = PAGE_SIZE << order;
1776
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001777 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001778 continue;
1779
Christoph Lameter81819f02007-05-06 14:49:36 -07001780 rem = slab_size % size;
1781
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001782 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001783 break;
1784
1785 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001786
Christoph Lameter81819f02007-05-06 14:49:36 -07001787 return order;
1788}
1789
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001790static inline int calculate_order(int size)
1791{
1792 int order;
1793 int min_objects;
1794 int fraction;
1795
1796 /*
1797 * Attempt to find best configuration for a slab. This
1798 * works by first attempting to generate a layout with
1799 * the best configuration and backing off gradually.
1800 *
1801 * First we reduce the acceptable waste in a slab. Then
1802 * we reduce the minimum objects required in a slab.
1803 */
1804 min_objects = slub_min_objects;
1805 while (min_objects > 1) {
1806 fraction = 8;
1807 while (fraction >= 4) {
1808 order = slab_order(size, min_objects,
1809 slub_max_order, fraction);
1810 if (order <= slub_max_order)
1811 return order;
1812 fraction /= 2;
1813 }
1814 min_objects /= 2;
1815 }
1816
1817 /*
1818 * We were unable to place multiple objects in a slab. Now
1819 * lets see if we can place a single object there.
1820 */
1821 order = slab_order(size, 1, slub_max_order, 1);
1822 if (order <= slub_max_order)
1823 return order;
1824
1825 /*
1826 * Doh this slab cannot be placed using slub_max_order.
1827 */
1828 order = slab_order(size, 1, MAX_ORDER, 1);
1829 if (order <= MAX_ORDER)
1830 return order;
1831 return -ENOSYS;
1832}
1833
Christoph Lameter81819f02007-05-06 14:49:36 -07001834/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001835 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001836 */
1837static unsigned long calculate_alignment(unsigned long flags,
1838 unsigned long align, unsigned long size)
1839{
1840 /*
1841 * If the user wants hardware cache aligned objects then
1842 * follow that suggestion if the object is sufficiently
1843 * large.
1844 *
1845 * The hardware cache alignment cannot override the
1846 * specified alignment though. If that is greater
1847 * then use it.
1848 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001849 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001850 size > cache_line_size() / 2)
1851 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001852
1853 if (align < ARCH_SLAB_MINALIGN)
1854 return ARCH_SLAB_MINALIGN;
1855
1856 return ALIGN(align, sizeof(void *));
1857}
1858
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001859static void init_kmem_cache_cpu(struct kmem_cache *s,
1860 struct kmem_cache_cpu *c)
1861{
1862 c->page = NULL;
1863 c->freelist = NULL;
1864 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001865 c->offset = s->offset / sizeof(void *);
1866 c->objsize = s->objsize;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001867}
1868
Christoph Lameter81819f02007-05-06 14:49:36 -07001869static void init_kmem_cache_node(struct kmem_cache_node *n)
1870{
1871 n->nr_partial = 0;
1872 atomic_long_set(&n->nr_slabs, 0);
1873 spin_lock_init(&n->list_lock);
1874 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001875#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07001876 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001877#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001878}
1879
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001880#ifdef CONFIG_SMP
1881/*
1882 * Per cpu array for per cpu structures.
1883 *
1884 * The per cpu array places all kmem_cache_cpu structures from one processor
1885 * close together meaning that it becomes possible that multiple per cpu
1886 * structures are contained in one cacheline. This may be particularly
1887 * beneficial for the kmalloc caches.
1888 *
1889 * A desktop system typically has around 60-80 slabs. With 100 here we are
1890 * likely able to get per cpu structures for all caches from the array defined
1891 * here. We must be able to cover all kmalloc caches during bootstrap.
1892 *
1893 * If the per cpu array is exhausted then fall back to kmalloc
1894 * of individual cachelines. No sharing is possible then.
1895 */
1896#define NR_KMEM_CACHE_CPU 100
1897
1898static DEFINE_PER_CPU(struct kmem_cache_cpu,
1899 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1900
1901static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1902static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1903
1904static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1905 int cpu, gfp_t flags)
1906{
1907 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1908
1909 if (c)
1910 per_cpu(kmem_cache_cpu_free, cpu) =
1911 (void *)c->freelist;
1912 else {
1913 /* Table overflow: So allocate ourselves */
1914 c = kmalloc_node(
1915 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1916 flags, cpu_to_node(cpu));
1917 if (!c)
1918 return NULL;
1919 }
1920
1921 init_kmem_cache_cpu(s, c);
1922 return c;
1923}
1924
1925static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1926{
1927 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1928 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1929 kfree(c);
1930 return;
1931 }
1932 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1933 per_cpu(kmem_cache_cpu_free, cpu) = c;
1934}
1935
1936static void free_kmem_cache_cpus(struct kmem_cache *s)
1937{
1938 int cpu;
1939
1940 for_each_online_cpu(cpu) {
1941 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1942
1943 if (c) {
1944 s->cpu_slab[cpu] = NULL;
1945 free_kmem_cache_cpu(c, cpu);
1946 }
1947 }
1948}
1949
1950static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1951{
1952 int cpu;
1953
1954 for_each_online_cpu(cpu) {
1955 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1956
1957 if (c)
1958 continue;
1959
1960 c = alloc_kmem_cache_cpu(s, cpu, flags);
1961 if (!c) {
1962 free_kmem_cache_cpus(s);
1963 return 0;
1964 }
1965 s->cpu_slab[cpu] = c;
1966 }
1967 return 1;
1968}
1969
1970/*
1971 * Initialize the per cpu array.
1972 */
1973static void init_alloc_cpu_cpu(int cpu)
1974{
1975 int i;
1976
1977 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
1978 return;
1979
1980 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
1981 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
1982
1983 cpu_set(cpu, kmem_cach_cpu_free_init_once);
1984}
1985
1986static void __init init_alloc_cpu(void)
1987{
1988 int cpu;
1989
1990 for_each_online_cpu(cpu)
1991 init_alloc_cpu_cpu(cpu);
1992 }
1993
1994#else
1995static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
1996static inline void init_alloc_cpu(void) {}
1997
1998static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1999{
2000 init_kmem_cache_cpu(s, &s->cpu_slab);
2001 return 1;
2002}
2003#endif
2004
Christoph Lameter81819f02007-05-06 14:49:36 -07002005#ifdef CONFIG_NUMA
2006/*
2007 * No kmalloc_node yet so do it by hand. We know that this is the first
2008 * slab on the node for this slabcache. There are no concurrent accesses
2009 * possible.
2010 *
2011 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002012 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2013 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002014 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002015static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2016 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002017{
2018 struct page *page;
2019 struct kmem_cache_node *n;
2020
2021 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2022
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002023 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002024
2025 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002026 if (page_to_nid(page) != node) {
2027 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2028 "node %d\n", node);
2029 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2030 "in order to be able to continue\n");
2031 }
2032
Christoph Lameter81819f02007-05-06 14:49:36 -07002033 n = page->freelist;
2034 BUG_ON(!n);
2035 page->freelist = get_freepointer(kmalloc_caches, n);
2036 page->inuse++;
2037 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002038#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002039 init_object(kmalloc_caches, n, 1);
2040 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002041#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002042 init_kmem_cache_node(n);
2043 atomic_long_inc(&n->nr_slabs);
Christoph Lametere95eed52007-05-06 14:49:44 -07002044 add_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002045 return n;
2046}
2047
2048static void free_kmem_cache_nodes(struct kmem_cache *s)
2049{
2050 int node;
2051
Christoph Lameterf64dc582007-10-16 01:25:33 -07002052 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002053 struct kmem_cache_node *n = s->node[node];
2054 if (n && n != &s->local_node)
2055 kmem_cache_free(kmalloc_caches, n);
2056 s->node[node] = NULL;
2057 }
2058}
2059
2060static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2061{
2062 int node;
2063 int local_node;
2064
2065 if (slab_state >= UP)
2066 local_node = page_to_nid(virt_to_page(s));
2067 else
2068 local_node = 0;
2069
Christoph Lameterf64dc582007-10-16 01:25:33 -07002070 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002071 struct kmem_cache_node *n;
2072
2073 if (local_node == node)
2074 n = &s->local_node;
2075 else {
2076 if (slab_state == DOWN) {
2077 n = early_kmem_cache_node_alloc(gfpflags,
2078 node);
2079 continue;
2080 }
2081 n = kmem_cache_alloc_node(kmalloc_caches,
2082 gfpflags, node);
2083
2084 if (!n) {
2085 free_kmem_cache_nodes(s);
2086 return 0;
2087 }
2088
2089 }
2090 s->node[node] = n;
2091 init_kmem_cache_node(n);
2092 }
2093 return 1;
2094}
2095#else
2096static void free_kmem_cache_nodes(struct kmem_cache *s)
2097{
2098}
2099
2100static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2101{
2102 init_kmem_cache_node(&s->local_node);
2103 return 1;
2104}
2105#endif
2106
2107/*
2108 * calculate_sizes() determines the order and the distribution of data within
2109 * a slab object.
2110 */
2111static int calculate_sizes(struct kmem_cache *s)
2112{
2113 unsigned long flags = s->flags;
2114 unsigned long size = s->objsize;
2115 unsigned long align = s->align;
2116
2117 /*
2118 * Determine if we can poison the object itself. If the user of
2119 * the slab may touch the object after free or before allocation
2120 * then we should never poison the object itself.
2121 */
2122 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002123 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002124 s->flags |= __OBJECT_POISON;
2125 else
2126 s->flags &= ~__OBJECT_POISON;
2127
2128 /*
2129 * Round up object size to the next word boundary. We can only
2130 * place the free pointer at word boundaries and this determines
2131 * the possible location of the free pointer.
2132 */
2133 size = ALIGN(size, sizeof(void *));
2134
Christoph Lameter41ecc552007-05-09 02:32:44 -07002135#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002136 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002137 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002138 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002139 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002140 */
2141 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2142 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002143#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002144
2145 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002146 * With that we have determined the number of bytes in actual use
2147 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002148 */
2149 s->inuse = size;
2150
2151 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002152 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002153 /*
2154 * Relocate free pointer after the object if it is not
2155 * permitted to overwrite the first word of the object on
2156 * kmem_cache_free.
2157 *
2158 * This is the case if we do RCU, have a constructor or
2159 * destructor or are poisoning the objects.
2160 */
2161 s->offset = size;
2162 size += sizeof(void *);
2163 }
2164
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002165#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002166 if (flags & SLAB_STORE_USER)
2167 /*
2168 * Need to store information about allocs and frees after
2169 * the object.
2170 */
2171 size += 2 * sizeof(struct track);
2172
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002173 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002174 /*
2175 * Add some empty padding so that we can catch
2176 * overwrites from earlier objects rather than let
2177 * tracking information or the free pointer be
2178 * corrupted if an user writes before the start
2179 * of the object.
2180 */
2181 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002182#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002183
Christoph Lameter81819f02007-05-06 14:49:36 -07002184 /*
2185 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002186 * user specified and the dynamic determination of cache line size
2187 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002188 */
2189 align = calculate_alignment(flags, align, s->objsize);
2190
2191 /*
2192 * SLUB stores one object immediately after another beginning from
2193 * offset 0. In order to align the objects we have to simply size
2194 * each object to conform to the alignment.
2195 */
2196 size = ALIGN(size, align);
2197 s->size = size;
2198
2199 s->order = calculate_order(size);
2200 if (s->order < 0)
2201 return 0;
2202
2203 /*
2204 * Determine the number of objects per slab
2205 */
2206 s->objects = (PAGE_SIZE << s->order) / size;
2207
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002208 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002209
2210}
2211
Christoph Lameter81819f02007-05-06 14:49:36 -07002212static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2213 const char *name, size_t size,
2214 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002215 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002216{
2217 memset(s, 0, kmem_size);
2218 s->name = name;
2219 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002220 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002221 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002222 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002223
2224 if (!calculate_sizes(s))
2225 goto error;
2226
2227 s->refcount = 1;
2228#ifdef CONFIG_NUMA
2229 s->defrag_ratio = 100;
2230#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002231 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2232 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002233
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002234 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002235 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002236 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002237error:
2238 if (flags & SLAB_PANIC)
2239 panic("Cannot create slab %s size=%lu realsize=%u "
2240 "order=%u offset=%u flags=%lx\n",
2241 s->name, (unsigned long)size, s->size, s->order,
2242 s->offset, flags);
2243 return 0;
2244}
Christoph Lameter81819f02007-05-06 14:49:36 -07002245
2246/*
2247 * Check if a given pointer is valid
2248 */
2249int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2250{
2251 struct page * page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002252
2253 page = get_object_page(object);
2254
2255 if (!page || s != page->slab)
2256 /* No slab or wrong slab */
2257 return 0;
2258
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002259 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002260 return 0;
2261
2262 /*
2263 * We could also check if the object is on the slabs freelist.
2264 * But this would be too expensive and it seems that the main
2265 * purpose of kmem_ptr_valid is to check if the object belongs
2266 * to a certain slab.
2267 */
2268 return 1;
2269}
2270EXPORT_SYMBOL(kmem_ptr_validate);
2271
2272/*
2273 * Determine the size of a slab object
2274 */
2275unsigned int kmem_cache_size(struct kmem_cache *s)
2276{
2277 return s->objsize;
2278}
2279EXPORT_SYMBOL(kmem_cache_size);
2280
2281const char *kmem_cache_name(struct kmem_cache *s)
2282{
2283 return s->name;
2284}
2285EXPORT_SYMBOL(kmem_cache_name);
2286
2287/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002288 * Attempt to free all slabs on a node. Return the number of slabs we
2289 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002290 */
2291static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2292 struct list_head *list)
2293{
2294 int slabs_inuse = 0;
2295 unsigned long flags;
2296 struct page *page, *h;
2297
2298 spin_lock_irqsave(&n->list_lock, flags);
2299 list_for_each_entry_safe(page, h, list, lru)
2300 if (!page->inuse) {
2301 list_del(&page->lru);
2302 discard_slab(s, page);
2303 } else
2304 slabs_inuse++;
2305 spin_unlock_irqrestore(&n->list_lock, flags);
2306 return slabs_inuse;
2307}
2308
2309/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002310 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002311 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002312static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002313{
2314 int node;
2315
2316 flush_all(s);
2317
2318 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002319 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002320 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002321 struct kmem_cache_node *n = get_node(s, node);
2322
Christoph Lameter2086d262007-05-06 14:49:46 -07002323 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002324 if (atomic_long_read(&n->nr_slabs))
2325 return 1;
2326 }
2327 free_kmem_cache_nodes(s);
2328 return 0;
2329}
2330
2331/*
2332 * Close a cache and release the kmem_cache structure
2333 * (must be used for caches created using kmem_cache_create)
2334 */
2335void kmem_cache_destroy(struct kmem_cache *s)
2336{
2337 down_write(&slub_lock);
2338 s->refcount--;
2339 if (!s->refcount) {
2340 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002341 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002342 if (kmem_cache_close(s))
2343 WARN_ON(1);
2344 sysfs_slab_remove(s);
2345 kfree(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002346 } else
2347 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002348}
2349EXPORT_SYMBOL(kmem_cache_destroy);
2350
2351/********************************************************************
2352 * Kmalloc subsystem
2353 *******************************************************************/
2354
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002355struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002356EXPORT_SYMBOL(kmalloc_caches);
2357
2358#ifdef CONFIG_ZONE_DMA
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002359static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT];
Christoph Lameter81819f02007-05-06 14:49:36 -07002360#endif
2361
2362static int __init setup_slub_min_order(char *str)
2363{
2364 get_option (&str, &slub_min_order);
2365
2366 return 1;
2367}
2368
2369__setup("slub_min_order=", setup_slub_min_order);
2370
2371static int __init setup_slub_max_order(char *str)
2372{
2373 get_option (&str, &slub_max_order);
2374
2375 return 1;
2376}
2377
2378__setup("slub_max_order=", setup_slub_max_order);
2379
2380static int __init setup_slub_min_objects(char *str)
2381{
2382 get_option (&str, &slub_min_objects);
2383
2384 return 1;
2385}
2386
2387__setup("slub_min_objects=", setup_slub_min_objects);
2388
2389static int __init setup_slub_nomerge(char *str)
2390{
2391 slub_nomerge = 1;
2392 return 1;
2393}
2394
2395__setup("slub_nomerge", setup_slub_nomerge);
2396
Christoph Lameter81819f02007-05-06 14:49:36 -07002397static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2398 const char *name, int size, gfp_t gfp_flags)
2399{
2400 unsigned int flags = 0;
2401
2402 if (gfp_flags & SLUB_DMA)
2403 flags = SLAB_CACHE_DMA;
2404
2405 down_write(&slub_lock);
2406 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameterc59def92007-05-16 22:10:50 -07002407 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002408 goto panic;
2409
2410 list_add(&s->list, &slab_caches);
2411 up_write(&slub_lock);
2412 if (sysfs_slab_add(s))
2413 goto panic;
2414 return s;
2415
2416panic:
2417 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2418}
2419
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002420#ifdef CONFIG_ZONE_DMA
Christoph Lameter1ceef402007-08-07 15:11:48 -07002421
2422static void sysfs_add_func(struct work_struct *w)
2423{
2424 struct kmem_cache *s;
2425
2426 down_write(&slub_lock);
2427 list_for_each_entry(s, &slab_caches, list) {
2428 if (s->flags & __SYSFS_ADD_DEFERRED) {
2429 s->flags &= ~__SYSFS_ADD_DEFERRED;
2430 sysfs_slab_add(s);
2431 }
2432 }
2433 up_write(&slub_lock);
2434}
2435
2436static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2437
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002438static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2439{
2440 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002441 char *text;
2442 size_t realsize;
2443
2444 s = kmalloc_caches_dma[index];
2445 if (s)
2446 return s;
2447
2448 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002449 if (flags & __GFP_WAIT)
2450 down_write(&slub_lock);
2451 else {
2452 if (!down_write_trylock(&slub_lock))
2453 goto out;
2454 }
2455
2456 if (kmalloc_caches_dma[index])
2457 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002458
Christoph Lameter7b55f622007-07-17 04:03:27 -07002459 realsize = kmalloc_caches[index].objsize;
Christoph Lameter1ceef402007-08-07 15:11:48 -07002460 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", (unsigned int)realsize),
2461 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2462
2463 if (!s || !text || !kmem_cache_open(s, flags, text,
2464 realsize, ARCH_KMALLOC_MINALIGN,
2465 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2466 kfree(s);
2467 kfree(text);
2468 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002469 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002470
2471 list_add(&s->list, &slab_caches);
2472 kmalloc_caches_dma[index] = s;
2473
2474 schedule_work(&sysfs_add_work);
2475
2476unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002477 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002478out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002479 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002480}
2481#endif
2482
Christoph Lameterf1b26332007-07-17 04:03:26 -07002483/*
2484 * Conversion table for small slabs sizes / 8 to the index in the
2485 * kmalloc array. This is necessary for slabs < 192 since we have non power
2486 * of two cache sizes there. The size of larger slabs can be determined using
2487 * fls.
2488 */
2489static s8 size_index[24] = {
2490 3, /* 8 */
2491 4, /* 16 */
2492 5, /* 24 */
2493 5, /* 32 */
2494 6, /* 40 */
2495 6, /* 48 */
2496 6, /* 56 */
2497 6, /* 64 */
2498 1, /* 72 */
2499 1, /* 80 */
2500 1, /* 88 */
2501 1, /* 96 */
2502 7, /* 104 */
2503 7, /* 112 */
2504 7, /* 120 */
2505 7, /* 128 */
2506 2, /* 136 */
2507 2, /* 144 */
2508 2, /* 152 */
2509 2, /* 160 */
2510 2, /* 168 */
2511 2, /* 176 */
2512 2, /* 184 */
2513 2 /* 192 */
2514};
2515
Christoph Lameter81819f02007-05-06 14:49:36 -07002516static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2517{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002518 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002519
Christoph Lameterf1b26332007-07-17 04:03:26 -07002520 if (size <= 192) {
2521 if (!size)
2522 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002523
Christoph Lameterf1b26332007-07-17 04:03:26 -07002524 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002525 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002526 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002527
2528#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002529 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002530 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002531
Christoph Lameter81819f02007-05-06 14:49:36 -07002532#endif
2533 return &kmalloc_caches[index];
2534}
2535
2536void *__kmalloc(size_t size, gfp_t flags)
2537{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002538 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002539
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002540 if (unlikely(size > PAGE_SIZE / 2))
2541 return (void *)__get_free_pages(flags | __GFP_COMP,
2542 get_order(size));
2543
2544 s = get_slab(size, flags);
2545
2546 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002547 return s;
2548
Christoph Lameterce15fea2007-07-17 04:03:28 -07002549 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002550}
2551EXPORT_SYMBOL(__kmalloc);
2552
2553#ifdef CONFIG_NUMA
2554void *__kmalloc_node(size_t size, gfp_t flags, int node)
2555{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002556 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002557
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002558 if (unlikely(size > PAGE_SIZE / 2))
2559 return (void *)__get_free_pages(flags | __GFP_COMP,
2560 get_order(size));
2561
2562 s = get_slab(size, flags);
2563
2564 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002565 return s;
2566
Christoph Lameterce15fea2007-07-17 04:03:28 -07002567 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002568}
2569EXPORT_SYMBOL(__kmalloc_node);
2570#endif
2571
2572size_t ksize(const void *object)
2573{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002574 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002575 struct kmem_cache *s;
2576
Christoph Lameteref8b4522007-10-16 01:24:46 -07002577 BUG_ON(!object);
2578 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002579 return 0;
2580
2581 page = get_object_page(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07002582 BUG_ON(!page);
2583 s = page->slab;
2584 BUG_ON(!s);
2585
2586 /*
2587 * Debugging requires use of the padding between object
2588 * and whatever may come after it.
2589 */
2590 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2591 return s->objsize;
2592
2593 /*
2594 * If we have the need to store the freelist pointer
2595 * back there or track user information then we can
2596 * only use the space before that information.
2597 */
2598 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2599 return s->inuse;
2600
2601 /*
2602 * Else we can use all the padding etc for the allocation
2603 */
2604 return s->size;
2605}
2606EXPORT_SYMBOL(ksize);
2607
2608void kfree(const void *x)
2609{
Christoph Lameter81819f02007-05-06 14:49:36 -07002610 struct page *page;
2611
Satyam Sharma2408c552007-10-16 01:24:44 -07002612 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002613 return;
2614
Christoph Lameterb49af682007-05-06 14:49:41 -07002615 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002616 if (unlikely(!PageSlab(page))) {
2617 put_page(page);
2618 return;
2619 }
2620 slab_free(page->slab, page, (void *)x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002621}
2622EXPORT_SYMBOL(kfree);
2623
Christoph Lameter2086d262007-05-06 14:49:46 -07002624/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002625 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2626 * the remaining slabs by the number of items in use. The slabs with the
2627 * most items in use come first. New allocations will then fill those up
2628 * and thus they can be removed from the partial lists.
2629 *
2630 * The slabs with the least items are placed last. This results in them
2631 * being allocated from last increasing the chance that the last objects
2632 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002633 */
2634int kmem_cache_shrink(struct kmem_cache *s)
2635{
2636 int node;
2637 int i;
2638 struct kmem_cache_node *n;
2639 struct page *page;
2640 struct page *t;
2641 struct list_head *slabs_by_inuse =
2642 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2643 unsigned long flags;
2644
2645 if (!slabs_by_inuse)
2646 return -ENOMEM;
2647
2648 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002649 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002650 n = get_node(s, node);
2651
2652 if (!n->nr_partial)
2653 continue;
2654
2655 for (i = 0; i < s->objects; i++)
2656 INIT_LIST_HEAD(slabs_by_inuse + i);
2657
2658 spin_lock_irqsave(&n->list_lock, flags);
2659
2660 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002661 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002662 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002663 * Note that concurrent frees may occur while we hold the
2664 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002665 */
2666 list_for_each_entry_safe(page, t, &n->partial, lru) {
2667 if (!page->inuse && slab_trylock(page)) {
2668 /*
2669 * Must hold slab lock here because slab_free
2670 * may have freed the last object and be
2671 * waiting to release the slab.
2672 */
2673 list_del(&page->lru);
2674 n->nr_partial--;
2675 slab_unlock(page);
2676 discard_slab(s, page);
2677 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002678 list_move(&page->lru,
2679 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002680 }
2681 }
2682
Christoph Lameter2086d262007-05-06 14:49:46 -07002683 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002684 * Rebuild the partial list with the slabs filled up most
2685 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002686 */
2687 for (i = s->objects - 1; i >= 0; i--)
2688 list_splice(slabs_by_inuse + i, n->partial.prev);
2689
Christoph Lameter2086d262007-05-06 14:49:46 -07002690 spin_unlock_irqrestore(&n->list_lock, flags);
2691 }
2692
2693 kfree(slabs_by_inuse);
2694 return 0;
2695}
2696EXPORT_SYMBOL(kmem_cache_shrink);
2697
Yasunori Gotob9049e22007-10-21 16:41:37 -07002698#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2699static int slab_mem_going_offline_callback(void *arg)
2700{
2701 struct kmem_cache *s;
2702
2703 down_read(&slub_lock);
2704 list_for_each_entry(s, &slab_caches, list)
2705 kmem_cache_shrink(s);
2706 up_read(&slub_lock);
2707
2708 return 0;
2709}
2710
2711static void slab_mem_offline_callback(void *arg)
2712{
2713 struct kmem_cache_node *n;
2714 struct kmem_cache *s;
2715 struct memory_notify *marg = arg;
2716 int offline_node;
2717
2718 offline_node = marg->status_change_nid;
2719
2720 /*
2721 * If the node still has available memory. we need kmem_cache_node
2722 * for it yet.
2723 */
2724 if (offline_node < 0)
2725 return;
2726
2727 down_read(&slub_lock);
2728 list_for_each_entry(s, &slab_caches, list) {
2729 n = get_node(s, offline_node);
2730 if (n) {
2731 /*
2732 * if n->nr_slabs > 0, slabs still exist on the node
2733 * that is going down. We were unable to free them,
2734 * and offline_pages() function shoudn't call this
2735 * callback. So, we must fail.
2736 */
Al Viro27bb6282007-10-29 04:42:55 +00002737 BUG_ON(atomic_long_read(&n->nr_slabs));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002738
2739 s->node[offline_node] = NULL;
2740 kmem_cache_free(kmalloc_caches, n);
2741 }
2742 }
2743 up_read(&slub_lock);
2744}
2745
2746static int slab_mem_going_online_callback(void *arg)
2747{
2748 struct kmem_cache_node *n;
2749 struct kmem_cache *s;
2750 struct memory_notify *marg = arg;
2751 int nid = marg->status_change_nid;
2752 int ret = 0;
2753
2754 /*
2755 * If the node's memory is already available, then kmem_cache_node is
2756 * already created. Nothing to do.
2757 */
2758 if (nid < 0)
2759 return 0;
2760
2761 /*
2762 * We are bringing a node online. No memory is availabe yet. We must
2763 * allocate a kmem_cache_node structure in order to bring the node
2764 * online.
2765 */
2766 down_read(&slub_lock);
2767 list_for_each_entry(s, &slab_caches, list) {
2768 /*
2769 * XXX: kmem_cache_alloc_node will fallback to other nodes
2770 * since memory is not yet available from the node that
2771 * is brought up.
2772 */
2773 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2774 if (!n) {
2775 ret = -ENOMEM;
2776 goto out;
2777 }
2778 init_kmem_cache_node(n);
2779 s->node[nid] = n;
2780 }
2781out:
2782 up_read(&slub_lock);
2783 return ret;
2784}
2785
2786static int slab_memory_callback(struct notifier_block *self,
2787 unsigned long action, void *arg)
2788{
2789 int ret = 0;
2790
2791 switch (action) {
2792 case MEM_GOING_ONLINE:
2793 ret = slab_mem_going_online_callback(arg);
2794 break;
2795 case MEM_GOING_OFFLINE:
2796 ret = slab_mem_going_offline_callback(arg);
2797 break;
2798 case MEM_OFFLINE:
2799 case MEM_CANCEL_ONLINE:
2800 slab_mem_offline_callback(arg);
2801 break;
2802 case MEM_ONLINE:
2803 case MEM_CANCEL_OFFLINE:
2804 break;
2805 }
2806
2807 ret = notifier_from_errno(ret);
2808 return ret;
2809}
2810
2811#endif /* CONFIG_MEMORY_HOTPLUG */
2812
Christoph Lameter81819f02007-05-06 14:49:36 -07002813/********************************************************************
2814 * Basic setup of slabs
2815 *******************************************************************/
2816
2817void __init kmem_cache_init(void)
2818{
2819 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002820 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002821
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002822 init_alloc_cpu();
2823
Christoph Lameter81819f02007-05-06 14:49:36 -07002824#ifdef CONFIG_NUMA
2825 /*
2826 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002827 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002828 * kmem_cache_open for slab_state == DOWN.
2829 */
2830 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2831 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002832 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002833 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002834
2835 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002836#endif
2837
2838 /* Able to allocate the per node structures */
2839 slab_state = PARTIAL;
2840
2841 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002842 if (KMALLOC_MIN_SIZE <= 64) {
2843 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002844 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002845 caches++;
2846 }
2847 if (KMALLOC_MIN_SIZE <= 128) {
2848 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002849 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002850 caches++;
2851 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002852
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002853 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002854 create_kmalloc_cache(&kmalloc_caches[i],
2855 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002856 caches++;
2857 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002858
Christoph Lameterf1b26332007-07-17 04:03:26 -07002859
2860 /*
2861 * Patch up the size_index table if we have strange large alignment
2862 * requirements for the kmalloc array. This is only the case for
2863 * mips it seems. The standard arches will not generate any code here.
2864 *
2865 * Largest permitted alignment is 256 bytes due to the way we
2866 * handle the index determination for the smaller caches.
2867 *
2868 * Make sure that nothing crazy happens if someone starts tinkering
2869 * around with ARCH_KMALLOC_MINALIGN
2870 */
2871 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2872 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2873
Christoph Lameter12ad6842007-07-17 04:03:28 -07002874 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002875 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2876
Christoph Lameter81819f02007-05-06 14:49:36 -07002877 slab_state = UP;
2878
2879 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002880 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002881 kmalloc_caches[i]. name =
2882 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2883
2884#ifdef CONFIG_SMP
2885 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002886 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2887 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2888#else
2889 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002890#endif
2891
Christoph Lameter81819f02007-05-06 14:49:36 -07002892
2893 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002894 " CPUs=%d, Nodes=%d\n",
2895 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002896 slub_min_order, slub_max_order, slub_min_objects,
2897 nr_cpu_ids, nr_node_ids);
2898}
2899
2900/*
2901 * Find a mergeable slab cache
2902 */
2903static int slab_unmergeable(struct kmem_cache *s)
2904{
2905 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2906 return 1;
2907
Christoph Lameterc59def92007-05-16 22:10:50 -07002908 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002909 return 1;
2910
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002911 /*
2912 * We may have set a slab to be unmergeable during bootstrap.
2913 */
2914 if (s->refcount < 0)
2915 return 1;
2916
Christoph Lameter81819f02007-05-06 14:49:36 -07002917 return 0;
2918}
2919
2920static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07002921 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002922 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002923{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002924 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002925
2926 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2927 return NULL;
2928
Christoph Lameterc59def92007-05-16 22:10:50 -07002929 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002930 return NULL;
2931
2932 size = ALIGN(size, sizeof(void *));
2933 align = calculate_alignment(flags, align, size);
2934 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07002935 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002936
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07002937 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002938 if (slab_unmergeable(s))
2939 continue;
2940
2941 if (size > s->size)
2942 continue;
2943
Christoph Lameterba0268a2007-09-11 15:24:11 -07002944 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07002945 continue;
2946 /*
2947 * Check if alignment is compatible.
2948 * Courtesy of Adrian Drzewiecki
2949 */
2950 if ((s->size & ~(align -1)) != s->size)
2951 continue;
2952
2953 if (s->size - size >= sizeof(void *))
2954 continue;
2955
2956 return s;
2957 }
2958 return NULL;
2959}
2960
2961struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2962 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002963 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002964{
2965 struct kmem_cache *s;
2966
2967 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07002968 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002969 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07002970 int cpu;
2971
Christoph Lameter81819f02007-05-06 14:49:36 -07002972 s->refcount++;
2973 /*
2974 * Adjust the object sizes so that we clear
2975 * the complete object on kzalloc.
2976 */
2977 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07002978
2979 /*
2980 * And then we need to update the object size in the
2981 * per cpu structures
2982 */
2983 for_each_online_cpu(cpu)
2984 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter81819f02007-05-06 14:49:36 -07002985 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002986 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002987 if (sysfs_slab_alias(s, name))
2988 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002989 return s;
2990 }
2991 s = kmalloc(kmem_size, GFP_KERNEL);
2992 if (s) {
2993 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07002994 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002995 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002996 up_write(&slub_lock);
2997 if (sysfs_slab_add(s))
2998 goto err;
2999 return s;
3000 }
3001 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003002 }
3003 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003004
3005err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003006 if (flags & SLAB_PANIC)
3007 panic("Cannot create slabcache %s\n", name);
3008 else
3009 s = NULL;
3010 return s;
3011}
3012EXPORT_SYMBOL(kmem_cache_create);
3013
Christoph Lameter81819f02007-05-06 14:49:36 -07003014#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003015/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003016 * Use the cpu notifier to insure that the cpu slabs are flushed when
3017 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003018 */
3019static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3020 unsigned long action, void *hcpu)
3021{
3022 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003023 struct kmem_cache *s;
3024 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003025
3026 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003027 case CPU_UP_PREPARE:
3028 case CPU_UP_PREPARE_FROZEN:
3029 init_alloc_cpu_cpu(cpu);
3030 down_read(&slub_lock);
3031 list_for_each_entry(s, &slab_caches, list)
3032 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3033 GFP_KERNEL);
3034 up_read(&slub_lock);
3035 break;
3036
Christoph Lameter81819f02007-05-06 14:49:36 -07003037 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003038 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003039 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003040 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003041 down_read(&slub_lock);
3042 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003043 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3044
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003045 local_irq_save(flags);
3046 __flush_cpu_slab(s, cpu);
3047 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003048 free_kmem_cache_cpu(c, cpu);
3049 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003050 }
3051 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003052 break;
3053 default:
3054 break;
3055 }
3056 return NOTIFY_OK;
3057}
3058
3059static struct notifier_block __cpuinitdata slab_notifier =
3060 { &slab_cpuup_callback, NULL, 0 };
3061
3062#endif
3063
Christoph Lameter81819f02007-05-06 14:49:36 -07003064void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, 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, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003077}
3078
3079void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3080 int node, void *caller)
3081{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003082 struct kmem_cache *s;
3083
3084 if (unlikely(size > PAGE_SIZE / 2))
3085 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3086 get_order(size));
3087 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003088
Satyam Sharma2408c552007-10-16 01:24:44 -07003089 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003090 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003091
Christoph Lameterce15fea2007-07-17 04:03:28 -07003092 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003093}
3094
Christoph Lameter41ecc552007-05-09 02:32:44 -07003095#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003096static int validate_slab(struct kmem_cache *s, struct page *page,
3097 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003098{
3099 void *p;
3100 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003101
3102 if (!check_slab(s, page) ||
3103 !on_freelist(s, page, NULL))
3104 return 0;
3105
3106 /* Now we know that a valid freelist exists */
3107 bitmap_zero(map, s->objects);
3108
Christoph Lameter7656c722007-05-09 02:32:40 -07003109 for_each_free_object(p, s, page->freelist) {
3110 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003111 if (!check_object(s, page, p, 0))
3112 return 0;
3113 }
3114
Christoph Lameter7656c722007-05-09 02:32:40 -07003115 for_each_object(p, s, addr)
3116 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003117 if (!check_object(s, page, p, 1))
3118 return 0;
3119 return 1;
3120}
3121
Christoph Lameter434e2452007-07-17 04:03:30 -07003122static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3123 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003124{
3125 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003126 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003127 slab_unlock(page);
3128 } else
3129 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3130 s->name, page);
3131
3132 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003133 if (!SlabDebug(page))
3134 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003135 "on slab 0x%p\n", s->name, page);
3136 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003137 if (SlabDebug(page))
3138 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003139 "slab 0x%p\n", s->name, page);
3140 }
3141}
3142
Christoph Lameter434e2452007-07-17 04:03:30 -07003143static int validate_slab_node(struct kmem_cache *s,
3144 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003145{
3146 unsigned long count = 0;
3147 struct page *page;
3148 unsigned long flags;
3149
3150 spin_lock_irqsave(&n->list_lock, flags);
3151
3152 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003153 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003154 count++;
3155 }
3156 if (count != n->nr_partial)
3157 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3158 "counter=%ld\n", s->name, count, n->nr_partial);
3159
3160 if (!(s->flags & SLAB_STORE_USER))
3161 goto out;
3162
3163 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003164 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003165 count++;
3166 }
3167 if (count != atomic_long_read(&n->nr_slabs))
3168 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3169 "counter=%ld\n", s->name, count,
3170 atomic_long_read(&n->nr_slabs));
3171
3172out:
3173 spin_unlock_irqrestore(&n->list_lock, flags);
3174 return count;
3175}
3176
Christoph Lameter434e2452007-07-17 04:03:30 -07003177static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003178{
3179 int node;
3180 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003181 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3182 sizeof(unsigned long), GFP_KERNEL);
3183
3184 if (!map)
3185 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003186
3187 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003188 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003189 struct kmem_cache_node *n = get_node(s, node);
3190
Christoph Lameter434e2452007-07-17 04:03:30 -07003191 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003192 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003193 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003194 return count;
3195}
3196
Christoph Lameterb3459702007-05-09 02:32:41 -07003197#ifdef SLUB_RESILIENCY_TEST
3198static void resiliency_test(void)
3199{
3200 u8 *p;
3201
3202 printk(KERN_ERR "SLUB resiliency testing\n");
3203 printk(KERN_ERR "-----------------------\n");
3204 printk(KERN_ERR "A. Corruption after allocation\n");
3205
3206 p = kzalloc(16, GFP_KERNEL);
3207 p[16] = 0x12;
3208 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3209 " 0x12->0x%p\n\n", p + 16);
3210
3211 validate_slab_cache(kmalloc_caches + 4);
3212
3213 /* Hmmm... The next two are dangerous */
3214 p = kzalloc(32, GFP_KERNEL);
3215 p[32 + sizeof(void *)] = 0x34;
3216 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3217 " 0x34 -> -0x%p\n", p);
3218 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3219
3220 validate_slab_cache(kmalloc_caches + 5);
3221 p = kzalloc(64, GFP_KERNEL);
3222 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3223 *p = 0x56;
3224 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3225 p);
3226 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3227 validate_slab_cache(kmalloc_caches + 6);
3228
3229 printk(KERN_ERR "\nB. Corruption after free\n");
3230 p = kzalloc(128, GFP_KERNEL);
3231 kfree(p);
3232 *p = 0x78;
3233 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3234 validate_slab_cache(kmalloc_caches + 7);
3235
3236 p = kzalloc(256, GFP_KERNEL);
3237 kfree(p);
3238 p[50] = 0x9a;
3239 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
3240 validate_slab_cache(kmalloc_caches + 8);
3241
3242 p = kzalloc(512, GFP_KERNEL);
3243 kfree(p);
3244 p[512] = 0xab;
3245 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3246 validate_slab_cache(kmalloc_caches + 9);
3247}
3248#else
3249static void resiliency_test(void) {};
3250#endif
3251
Christoph Lameter88a420e2007-05-06 14:49:45 -07003252/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003253 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003254 * and freed.
3255 */
3256
3257struct location {
3258 unsigned long count;
3259 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003260 long long sum_time;
3261 long min_time;
3262 long max_time;
3263 long min_pid;
3264 long max_pid;
3265 cpumask_t cpus;
3266 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003267};
3268
3269struct loc_track {
3270 unsigned long max;
3271 unsigned long count;
3272 struct location *loc;
3273};
3274
3275static void free_loc_track(struct loc_track *t)
3276{
3277 if (t->max)
3278 free_pages((unsigned long)t->loc,
3279 get_order(sizeof(struct location) * t->max));
3280}
3281
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003282static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003283{
3284 struct location *l;
3285 int order;
3286
Christoph Lameter88a420e2007-05-06 14:49:45 -07003287 order = get_order(sizeof(struct location) * max);
3288
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003289 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003290 if (!l)
3291 return 0;
3292
3293 if (t->count) {
3294 memcpy(l, t->loc, sizeof(struct location) * t->count);
3295 free_loc_track(t);
3296 }
3297 t->max = max;
3298 t->loc = l;
3299 return 1;
3300}
3301
3302static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003303 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003304{
3305 long start, end, pos;
3306 struct location *l;
3307 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003308 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003309
3310 start = -1;
3311 end = t->count;
3312
3313 for ( ; ; ) {
3314 pos = start + (end - start + 1) / 2;
3315
3316 /*
3317 * There is nothing at "end". If we end up there
3318 * we need to add something to before end.
3319 */
3320 if (pos == end)
3321 break;
3322
3323 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003324 if (track->addr == caddr) {
3325
3326 l = &t->loc[pos];
3327 l->count++;
3328 if (track->when) {
3329 l->sum_time += age;
3330 if (age < l->min_time)
3331 l->min_time = age;
3332 if (age > l->max_time)
3333 l->max_time = age;
3334
3335 if (track->pid < l->min_pid)
3336 l->min_pid = track->pid;
3337 if (track->pid > l->max_pid)
3338 l->max_pid = track->pid;
3339
3340 cpu_set(track->cpu, l->cpus);
3341 }
3342 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003343 return 1;
3344 }
3345
Christoph Lameter45edfa52007-05-09 02:32:45 -07003346 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003347 end = pos;
3348 else
3349 start = pos;
3350 }
3351
3352 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003353 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003354 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003355 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003356 return 0;
3357
3358 l = t->loc + pos;
3359 if (pos < t->count)
3360 memmove(l + 1, l,
3361 (t->count - pos) * sizeof(struct location));
3362 t->count++;
3363 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003364 l->addr = track->addr;
3365 l->sum_time = age;
3366 l->min_time = age;
3367 l->max_time = age;
3368 l->min_pid = track->pid;
3369 l->max_pid = track->pid;
3370 cpus_clear(l->cpus);
3371 cpu_set(track->cpu, l->cpus);
3372 nodes_clear(l->nodes);
3373 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003374 return 1;
3375}
3376
3377static void process_slab(struct loc_track *t, struct kmem_cache *s,
3378 struct page *page, enum track_item alloc)
3379{
3380 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003381 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003382 void *p;
3383
3384 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003385 for_each_free_object(p, s, page->freelist)
3386 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003387
Christoph Lameter7656c722007-05-09 02:32:40 -07003388 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003389 if (!test_bit(slab_index(p, s, addr), map))
3390 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003391}
3392
3393static int list_locations(struct kmem_cache *s, char *buf,
3394 enum track_item alloc)
3395{
3396 int n = 0;
3397 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003398 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003399 int node;
3400
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003401 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003402 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003403 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003404
3405 /* Push back cpu slabs */
3406 flush_all(s);
3407
Christoph Lameterf64dc582007-10-16 01:25:33 -07003408 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003409 struct kmem_cache_node *n = get_node(s, node);
3410 unsigned long flags;
3411 struct page *page;
3412
Christoph Lameter9e869432007-08-22 14:01:56 -07003413 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003414 continue;
3415
3416 spin_lock_irqsave(&n->list_lock, flags);
3417 list_for_each_entry(page, &n->partial, lru)
3418 process_slab(&t, s, page, alloc);
3419 list_for_each_entry(page, &n->full, lru)
3420 process_slab(&t, s, page, alloc);
3421 spin_unlock_irqrestore(&n->list_lock, flags);
3422 }
3423
3424 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003425 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003426
3427 if (n > PAGE_SIZE - 100)
3428 break;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003429 n += sprintf(buf + n, "%7ld ", l->count);
3430
3431 if (l->addr)
3432 n += sprint_symbol(buf + n, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003433 else
3434 n += sprintf(buf + n, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003435
3436 if (l->sum_time != l->min_time) {
3437 unsigned long remainder;
3438
3439 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3440 l->min_time,
3441 div_long_long_rem(l->sum_time, l->count, &remainder),
3442 l->max_time);
3443 } else
3444 n += sprintf(buf + n, " age=%ld",
3445 l->min_time);
3446
3447 if (l->min_pid != l->max_pid)
3448 n += sprintf(buf + n, " pid=%ld-%ld",
3449 l->min_pid, l->max_pid);
3450 else
3451 n += sprintf(buf + n, " pid=%ld",
3452 l->min_pid);
3453
Christoph Lameter84966342007-06-23 17:16:32 -07003454 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3455 n < PAGE_SIZE - 60) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003456 n += sprintf(buf + n, " cpus=");
3457 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3458 l->cpus);
3459 }
3460
Christoph Lameter84966342007-06-23 17:16:32 -07003461 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3462 n < PAGE_SIZE - 60) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003463 n += sprintf(buf + n, " nodes=");
3464 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3465 l->nodes);
3466 }
3467
Christoph Lameter88a420e2007-05-06 14:49:45 -07003468 n += sprintf(buf + n, "\n");
3469 }
3470
3471 free_loc_track(&t);
3472 if (!t.count)
3473 n += sprintf(buf, "No data\n");
3474 return n;
3475}
3476
Christoph Lameter81819f02007-05-06 14:49:36 -07003477static unsigned long count_partial(struct kmem_cache_node *n)
3478{
3479 unsigned long flags;
3480 unsigned long x = 0;
3481 struct page *page;
3482
3483 spin_lock_irqsave(&n->list_lock, flags);
3484 list_for_each_entry(page, &n->partial, lru)
3485 x += page->inuse;
3486 spin_unlock_irqrestore(&n->list_lock, flags);
3487 return x;
3488}
3489
3490enum slab_stat_type {
3491 SL_FULL,
3492 SL_PARTIAL,
3493 SL_CPU,
3494 SL_OBJECTS
3495};
3496
3497#define SO_FULL (1 << SL_FULL)
3498#define SO_PARTIAL (1 << SL_PARTIAL)
3499#define SO_CPU (1 << SL_CPU)
3500#define SO_OBJECTS (1 << SL_OBJECTS)
3501
3502static unsigned long slab_objects(struct kmem_cache *s,
3503 char *buf, unsigned long flags)
3504{
3505 unsigned long total = 0;
3506 int cpu;
3507 int node;
3508 int x;
3509 unsigned long *nodes;
3510 unsigned long *per_cpu;
3511
3512 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3513 per_cpu = nodes + nr_node_ids;
3514
3515 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003516 struct page *page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003517 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003518 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003519
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003520 if (!c)
3521 continue;
3522
3523 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003524 node = c->node;
3525 if (node < 0)
3526 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003527 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003528 if (flags & SO_CPU) {
3529 int x = 0;
3530
3531 if (flags & SO_OBJECTS)
3532 x = page->inuse;
3533 else
3534 x = 1;
3535 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003536 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003537 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003538 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003539 }
3540 }
3541
Christoph Lameterf64dc582007-10-16 01:25:33 -07003542 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003543 struct kmem_cache_node *n = get_node(s, node);
3544
3545 if (flags & SO_PARTIAL) {
3546 if (flags & SO_OBJECTS)
3547 x = count_partial(n);
3548 else
3549 x = n->nr_partial;
3550 total += x;
3551 nodes[node] += x;
3552 }
3553
3554 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003555 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003556 - per_cpu[node]
3557 - n->nr_partial;
3558
3559 if (flags & SO_OBJECTS)
3560 x = full_slabs * s->objects;
3561 else
3562 x = full_slabs;
3563 total += x;
3564 nodes[node] += x;
3565 }
3566 }
3567
3568 x = sprintf(buf, "%lu", total);
3569#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003570 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003571 if (nodes[node])
3572 x += sprintf(buf + x, " N%d=%lu",
3573 node, nodes[node]);
3574#endif
3575 kfree(nodes);
3576 return x + sprintf(buf + x, "\n");
3577}
3578
3579static int any_slab_objects(struct kmem_cache *s)
3580{
3581 int node;
3582 int cpu;
3583
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003584 for_each_possible_cpu(cpu) {
3585 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003586
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003587 if (c && c->page)
3588 return 1;
3589 }
3590
3591 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003592 struct kmem_cache_node *n = get_node(s, node);
3593
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003594 if (!n)
3595 continue;
3596
Christoph Lameter9e869432007-08-22 14:01:56 -07003597 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003598 return 1;
3599 }
3600 return 0;
3601}
3602
3603#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3604#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3605
3606struct slab_attribute {
3607 struct attribute attr;
3608 ssize_t (*show)(struct kmem_cache *s, char *buf);
3609 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3610};
3611
3612#define SLAB_ATTR_RO(_name) \
3613 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3614
3615#define SLAB_ATTR(_name) \
3616 static struct slab_attribute _name##_attr = \
3617 __ATTR(_name, 0644, _name##_show, _name##_store)
3618
Christoph Lameter81819f02007-05-06 14:49:36 -07003619static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3620{
3621 return sprintf(buf, "%d\n", s->size);
3622}
3623SLAB_ATTR_RO(slab_size);
3624
3625static ssize_t align_show(struct kmem_cache *s, char *buf)
3626{
3627 return sprintf(buf, "%d\n", s->align);
3628}
3629SLAB_ATTR_RO(align);
3630
3631static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3632{
3633 return sprintf(buf, "%d\n", s->objsize);
3634}
3635SLAB_ATTR_RO(object_size);
3636
3637static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3638{
3639 return sprintf(buf, "%d\n", s->objects);
3640}
3641SLAB_ATTR_RO(objs_per_slab);
3642
3643static ssize_t order_show(struct kmem_cache *s, char *buf)
3644{
3645 return sprintf(buf, "%d\n", s->order);
3646}
3647SLAB_ATTR_RO(order);
3648
3649static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3650{
3651 if (s->ctor) {
3652 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3653
3654 return n + sprintf(buf + n, "\n");
3655 }
3656 return 0;
3657}
3658SLAB_ATTR_RO(ctor);
3659
Christoph Lameter81819f02007-05-06 14:49:36 -07003660static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3661{
3662 return sprintf(buf, "%d\n", s->refcount - 1);
3663}
3664SLAB_ATTR_RO(aliases);
3665
3666static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3667{
3668 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3669}
3670SLAB_ATTR_RO(slabs);
3671
3672static ssize_t partial_show(struct kmem_cache *s, char *buf)
3673{
3674 return slab_objects(s, buf, SO_PARTIAL);
3675}
3676SLAB_ATTR_RO(partial);
3677
3678static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3679{
3680 return slab_objects(s, buf, SO_CPU);
3681}
3682SLAB_ATTR_RO(cpu_slabs);
3683
3684static ssize_t objects_show(struct kmem_cache *s, char *buf)
3685{
3686 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3687}
3688SLAB_ATTR_RO(objects);
3689
3690static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3691{
3692 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3693}
3694
3695static ssize_t sanity_checks_store(struct kmem_cache *s,
3696 const char *buf, size_t length)
3697{
3698 s->flags &= ~SLAB_DEBUG_FREE;
3699 if (buf[0] == '1')
3700 s->flags |= SLAB_DEBUG_FREE;
3701 return length;
3702}
3703SLAB_ATTR(sanity_checks);
3704
3705static ssize_t trace_show(struct kmem_cache *s, char *buf)
3706{
3707 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3708}
3709
3710static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3711 size_t length)
3712{
3713 s->flags &= ~SLAB_TRACE;
3714 if (buf[0] == '1')
3715 s->flags |= SLAB_TRACE;
3716 return length;
3717}
3718SLAB_ATTR(trace);
3719
3720static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3721{
3722 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3723}
3724
3725static ssize_t reclaim_account_store(struct kmem_cache *s,
3726 const char *buf, size_t length)
3727{
3728 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3729 if (buf[0] == '1')
3730 s->flags |= SLAB_RECLAIM_ACCOUNT;
3731 return length;
3732}
3733SLAB_ATTR(reclaim_account);
3734
3735static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3736{
Christoph Lameter5af60832007-05-06 14:49:56 -07003737 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003738}
3739SLAB_ATTR_RO(hwcache_align);
3740
3741#ifdef CONFIG_ZONE_DMA
3742static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3743{
3744 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3745}
3746SLAB_ATTR_RO(cache_dma);
3747#endif
3748
3749static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3750{
3751 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3752}
3753SLAB_ATTR_RO(destroy_by_rcu);
3754
3755static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3756{
3757 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3758}
3759
3760static ssize_t red_zone_store(struct kmem_cache *s,
3761 const char *buf, size_t length)
3762{
3763 if (any_slab_objects(s))
3764 return -EBUSY;
3765
3766 s->flags &= ~SLAB_RED_ZONE;
3767 if (buf[0] == '1')
3768 s->flags |= SLAB_RED_ZONE;
3769 calculate_sizes(s);
3770 return length;
3771}
3772SLAB_ATTR(red_zone);
3773
3774static ssize_t poison_show(struct kmem_cache *s, char *buf)
3775{
3776 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3777}
3778
3779static ssize_t poison_store(struct kmem_cache *s,
3780 const char *buf, size_t length)
3781{
3782 if (any_slab_objects(s))
3783 return -EBUSY;
3784
3785 s->flags &= ~SLAB_POISON;
3786 if (buf[0] == '1')
3787 s->flags |= SLAB_POISON;
3788 calculate_sizes(s);
3789 return length;
3790}
3791SLAB_ATTR(poison);
3792
3793static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3794{
3795 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3796}
3797
3798static ssize_t store_user_store(struct kmem_cache *s,
3799 const char *buf, size_t length)
3800{
3801 if (any_slab_objects(s))
3802 return -EBUSY;
3803
3804 s->flags &= ~SLAB_STORE_USER;
3805 if (buf[0] == '1')
3806 s->flags |= SLAB_STORE_USER;
3807 calculate_sizes(s);
3808 return length;
3809}
3810SLAB_ATTR(store_user);
3811
Christoph Lameter53e15af2007-05-06 14:49:43 -07003812static ssize_t validate_show(struct kmem_cache *s, char *buf)
3813{
3814 return 0;
3815}
3816
3817static ssize_t validate_store(struct kmem_cache *s,
3818 const char *buf, size_t length)
3819{
Christoph Lameter434e2452007-07-17 04:03:30 -07003820 int ret = -EINVAL;
3821
3822 if (buf[0] == '1') {
3823 ret = validate_slab_cache(s);
3824 if (ret >= 0)
3825 ret = length;
3826 }
3827 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003828}
3829SLAB_ATTR(validate);
3830
Christoph Lameter2086d262007-05-06 14:49:46 -07003831static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3832{
3833 return 0;
3834}
3835
3836static ssize_t shrink_store(struct kmem_cache *s,
3837 const char *buf, size_t length)
3838{
3839 if (buf[0] == '1') {
3840 int rc = kmem_cache_shrink(s);
3841
3842 if (rc)
3843 return rc;
3844 } else
3845 return -EINVAL;
3846 return length;
3847}
3848SLAB_ATTR(shrink);
3849
Christoph Lameter88a420e2007-05-06 14:49:45 -07003850static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3851{
3852 if (!(s->flags & SLAB_STORE_USER))
3853 return -ENOSYS;
3854 return list_locations(s, buf, TRACK_ALLOC);
3855}
3856SLAB_ATTR_RO(alloc_calls);
3857
3858static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3859{
3860 if (!(s->flags & SLAB_STORE_USER))
3861 return -ENOSYS;
3862 return list_locations(s, buf, TRACK_FREE);
3863}
3864SLAB_ATTR_RO(free_calls);
3865
Christoph Lameter81819f02007-05-06 14:49:36 -07003866#ifdef CONFIG_NUMA
3867static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3868{
3869 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3870}
3871
3872static ssize_t defrag_ratio_store(struct kmem_cache *s,
3873 const char *buf, size_t length)
3874{
3875 int n = simple_strtoul(buf, NULL, 10);
3876
3877 if (n < 100)
3878 s->defrag_ratio = n * 10;
3879 return length;
3880}
3881SLAB_ATTR(defrag_ratio);
3882#endif
3883
3884static struct attribute * slab_attrs[] = {
3885 &slab_size_attr.attr,
3886 &object_size_attr.attr,
3887 &objs_per_slab_attr.attr,
3888 &order_attr.attr,
3889 &objects_attr.attr,
3890 &slabs_attr.attr,
3891 &partial_attr.attr,
3892 &cpu_slabs_attr.attr,
3893 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003894 &aliases_attr.attr,
3895 &align_attr.attr,
3896 &sanity_checks_attr.attr,
3897 &trace_attr.attr,
3898 &hwcache_align_attr.attr,
3899 &reclaim_account_attr.attr,
3900 &destroy_by_rcu_attr.attr,
3901 &red_zone_attr.attr,
3902 &poison_attr.attr,
3903 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07003904 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07003905 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07003906 &alloc_calls_attr.attr,
3907 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003908#ifdef CONFIG_ZONE_DMA
3909 &cache_dma_attr.attr,
3910#endif
3911#ifdef CONFIG_NUMA
3912 &defrag_ratio_attr.attr,
3913#endif
3914 NULL
3915};
3916
3917static struct attribute_group slab_attr_group = {
3918 .attrs = slab_attrs,
3919};
3920
3921static ssize_t slab_attr_show(struct kobject *kobj,
3922 struct attribute *attr,
3923 char *buf)
3924{
3925 struct slab_attribute *attribute;
3926 struct kmem_cache *s;
3927 int err;
3928
3929 attribute = to_slab_attr(attr);
3930 s = to_slab(kobj);
3931
3932 if (!attribute->show)
3933 return -EIO;
3934
3935 err = attribute->show(s, buf);
3936
3937 return err;
3938}
3939
3940static ssize_t slab_attr_store(struct kobject *kobj,
3941 struct attribute *attr,
3942 const char *buf, size_t len)
3943{
3944 struct slab_attribute *attribute;
3945 struct kmem_cache *s;
3946 int err;
3947
3948 attribute = to_slab_attr(attr);
3949 s = to_slab(kobj);
3950
3951 if (!attribute->store)
3952 return -EIO;
3953
3954 err = attribute->store(s, buf, len);
3955
3956 return err;
3957}
3958
3959static struct sysfs_ops slab_sysfs_ops = {
3960 .show = slab_attr_show,
3961 .store = slab_attr_store,
3962};
3963
3964static struct kobj_type slab_ktype = {
3965 .sysfs_ops = &slab_sysfs_ops,
3966};
3967
3968static int uevent_filter(struct kset *kset, struct kobject *kobj)
3969{
3970 struct kobj_type *ktype = get_ktype(kobj);
3971
3972 if (ktype == &slab_ktype)
3973 return 1;
3974 return 0;
3975}
3976
3977static struct kset_uevent_ops slab_uevent_ops = {
3978 .filter = uevent_filter,
3979};
3980
Adrian Bunk5af328a2007-07-17 04:03:27 -07003981static decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
Christoph Lameter81819f02007-05-06 14:49:36 -07003982
3983#define ID_STR_LENGTH 64
3984
3985/* Create a unique string id for a slab cache:
3986 * format
3987 * :[flags-]size:[memory address of kmemcache]
3988 */
3989static char *create_unique_id(struct kmem_cache *s)
3990{
3991 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3992 char *p = name;
3993
3994 BUG_ON(!name);
3995
3996 *p++ = ':';
3997 /*
3998 * First flags affecting slabcache operations. We will only
3999 * get here for aliasable slabs so we do not need to support
4000 * too many flags. The flags here must cover all flags that
4001 * are matched during merging to guarantee that the id is
4002 * unique.
4003 */
4004 if (s->flags & SLAB_CACHE_DMA)
4005 *p++ = 'd';
4006 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4007 *p++ = 'a';
4008 if (s->flags & SLAB_DEBUG_FREE)
4009 *p++ = 'F';
4010 if (p != name + 1)
4011 *p++ = '-';
4012 p += sprintf(p, "%07d", s->size);
4013 BUG_ON(p > name + ID_STR_LENGTH - 1);
4014 return name;
4015}
4016
4017static int sysfs_slab_add(struct kmem_cache *s)
4018{
4019 int err;
4020 const char *name;
4021 int unmergeable;
4022
4023 if (slab_state < SYSFS)
4024 /* Defer until later */
4025 return 0;
4026
4027 unmergeable = slab_unmergeable(s);
4028 if (unmergeable) {
4029 /*
4030 * Slabcache can never be merged so we can use the name proper.
4031 * This is typically the case for debug situations. In that
4032 * case we can catch duplicate names easily.
4033 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07004034 sysfs_remove_link(&slab_subsys.kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004035 name = s->name;
4036 } else {
4037 /*
4038 * Create a unique name for the slab as a target
4039 * for the symlinks.
4040 */
4041 name = create_unique_id(s);
4042 }
4043
4044 kobj_set_kset_s(s, slab_subsys);
4045 kobject_set_name(&s->kobj, name);
4046 kobject_init(&s->kobj);
4047 err = kobject_add(&s->kobj);
4048 if (err)
4049 return err;
4050
4051 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4052 if (err)
4053 return err;
4054 kobject_uevent(&s->kobj, KOBJ_ADD);
4055 if (!unmergeable) {
4056 /* Setup first alias */
4057 sysfs_slab_alias(s, s->name);
4058 kfree(name);
4059 }
4060 return 0;
4061}
4062
4063static void sysfs_slab_remove(struct kmem_cache *s)
4064{
4065 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4066 kobject_del(&s->kobj);
4067}
4068
4069/*
4070 * Need to buffer aliases during bootup until sysfs becomes
4071 * available lest we loose that information.
4072 */
4073struct saved_alias {
4074 struct kmem_cache *s;
4075 const char *name;
4076 struct saved_alias *next;
4077};
4078
Adrian Bunk5af328a2007-07-17 04:03:27 -07004079static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004080
4081static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4082{
4083 struct saved_alias *al;
4084
4085 if (slab_state == SYSFS) {
4086 /*
4087 * If we have a leftover link then remove it.
4088 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07004089 sysfs_remove_link(&slab_subsys.kobj, name);
4090 return sysfs_create_link(&slab_subsys.kobj,
Christoph Lameter81819f02007-05-06 14:49:36 -07004091 &s->kobj, name);
4092 }
4093
4094 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4095 if (!al)
4096 return -ENOMEM;
4097
4098 al->s = s;
4099 al->name = name;
4100 al->next = alias_list;
4101 alias_list = al;
4102 return 0;
4103}
4104
4105static int __init slab_sysfs_init(void)
4106{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004107 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004108 int err;
4109
4110 err = subsystem_register(&slab_subsys);
4111 if (err) {
4112 printk(KERN_ERR "Cannot register slab subsystem.\n");
4113 return -ENOSYS;
4114 }
4115
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004116 slab_state = SYSFS;
4117
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004118 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004119 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004120 if (err)
4121 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4122 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004123 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004124
4125 while (alias_list) {
4126 struct saved_alias *al = alias_list;
4127
4128 alias_list = alias_list->next;
4129 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004130 if (err)
4131 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4132 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004133 kfree(al);
4134 }
4135
4136 resiliency_test();
4137 return 0;
4138}
4139
4140__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004141#endif