blob: acc975fcc8cc9f96a9d3ce7d37a1e1e9b7bbf2ca [file] [log] [blame]
Christoph Lameter81819f02007-05-06 14:49:36 -07001/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
9 */
10
11#include <linux/mm.h>
12#include <linux/module.h>
13#include <linux/bit_spinlock.h>
14#include <linux/interrupt.h>
15#include <linux/bitops.h>
16#include <linux/slab.h>
17#include <linux/seq_file.h>
18#include <linux/cpu.h>
19#include <linux/cpuset.h>
20#include <linux/mempolicy.h>
21#include <linux/ctype.h>
22#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070023#include <linux/memory.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024
25/*
26 * Lock order:
27 * 1. slab_lock(page)
28 * 2. slab->list_lock
29 *
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
36 *
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
42 *
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
47 * the list lock.
48 *
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
61 *
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
66 *
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
69 *
Christoph Lameter672bba32007-05-09 02:32:39 -070070 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070072 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070073 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070075 *
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
79 *
80 * Overloading of page flags that are otherwise used for LRU management.
81 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070082 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
90 *
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -070094 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -070095 * free objects in addition to the regular freelist
96 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070097 *
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700100 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700101 */
102
Christoph Lameter5577bd82007-05-16 22:10:56 -0700103#define FROZEN (1 << PG_active)
104
105#ifdef CONFIG_SLUB_DEBUG
106#define SLABDEBUG (1 << PG_error)
107#else
108#define SLABDEBUG 0
109#endif
110
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700111static inline int SlabFrozen(struct page *page)
112{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700113 return page->flags & FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700114}
115
116static inline void SetSlabFrozen(struct page *page)
117{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700118 page->flags |= FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700119}
120
121static inline void ClearSlabFrozen(struct page *page)
122{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700123 page->flags &= ~FROZEN;
Christoph Lameter4b6f0752007-05-16 22:10:53 -0700124}
125
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700126static inline int SlabDebug(struct page *page)
127{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700128 return page->flags & SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700129}
130
131static inline void SetSlabDebug(struct page *page)
132{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700133 page->flags |= SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700134}
135
136static inline void ClearSlabDebug(struct page *page)
137{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700138 page->flags &= ~SLABDEBUG;
Christoph Lameter35e5d7e2007-05-09 02:32:42 -0700139}
140
Christoph Lameter81819f02007-05-06 14:49:36 -0700141/*
142 * Issues still to be resolved:
143 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
145 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700146 * - Variable sizing of the per node arrays
147 */
148
149/* Enable to test recovery from slab corruption on boot */
150#undef SLUB_RESILIENCY_TEST
151
152#if PAGE_SHIFT <= 12
153
154/*
155 * Small page size. Make sure that we do not fragment memory
156 */
157#define DEFAULT_MAX_ORDER 1
158#define DEFAULT_MIN_OBJECTS 4
159
160#else
161
162/*
163 * Large page machines are customarily able to handle larger
164 * page orders.
165 */
166#define DEFAULT_MAX_ORDER 2
167#define DEFAULT_MIN_OBJECTS 8
168
169#endif
170
171/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
174 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800175#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700176
Christoph Lameter2086d262007-05-06 14:49:46 -0700177/*
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
181 */
182#define MAX_PARTIAL 10
183
Christoph Lameter81819f02007-05-06 14:49:36 -0700184#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700186
Christoph Lameter81819f02007-05-06 14:49:36 -0700187/*
188 * Set of flags that will prevent slab merging
189 */
190#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
192
193#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
194 SLAB_CACHE_DMA)
195
196#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700197#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700198#endif
199
200#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700201#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700202#endif
203
204/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700205#define __OBJECT_POISON 0x80000000 /* Poison object */
206#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter71c7a062008-02-14 14:28:01 -0800207#define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */
208#define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */
Christoph Lameter81819f02007-05-06 14:49:36 -0700209
Christoph Lameter65c02d42007-05-09 02:32:35 -0700210/* Not all arches define cache_line_size */
211#ifndef cache_line_size
212#define cache_line_size() L1_CACHE_BYTES
213#endif
214
Christoph Lameter81819f02007-05-06 14:49:36 -0700215static int kmem_size = sizeof(struct kmem_cache);
216
217#ifdef CONFIG_SMP
218static struct notifier_block slab_notifier;
219#endif
220
221static enum {
222 DOWN, /* No slab functionality available */
223 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700224 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700225 SYSFS /* Sysfs up */
226} slab_state = DOWN;
227
228/* A list of all slab caches on the system */
229static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700230static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700231
Christoph Lameter02cbc872007-05-09 02:32:43 -0700232/*
233 * Tracking user of a slab.
234 */
235struct track {
236 void *addr; /* Called from address */
237 int cpu; /* Was running on cpu */
238 int pid; /* Pid context */
239 unsigned long when; /* When did the operation occur */
240};
241
242enum track_item { TRACK_ALLOC, TRACK_FREE };
243
Christoph Lameter41ecc552007-05-09 02:32:44 -0700244#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter81819f02007-05-06 14:49:36 -0700245static int sysfs_slab_add(struct kmem_cache *);
246static int sysfs_slab_alias(struct kmem_cache *, const char *);
247static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800248
Christoph Lameter81819f02007-05-06 14:49:36 -0700249#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700250static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
252 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800253static inline void sysfs_slab_remove(struct kmem_cache *s)
254{
255 kfree(s);
256}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800257
Christoph Lameter81819f02007-05-06 14:49:36 -0700258#endif
259
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800260static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
261{
262#ifdef CONFIG_SLUB_STATS
263 c->stat[si]++;
264#endif
265}
266
Christoph Lameter81819f02007-05-06 14:49:36 -0700267/********************************************************************
268 * Core slab cache functions
269 *******************************************************************/
270
271int slab_is_available(void)
272{
273 return slab_state >= UP;
274}
275
276static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
277{
278#ifdef CONFIG_NUMA
279 return s->node[node];
280#else
281 return &s->local_node;
282#endif
283}
284
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700285static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
286{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700287#ifdef CONFIG_SMP
288 return s->cpu_slab[cpu];
289#else
290 return &s->cpu_slab;
291#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700292}
293
Christoph Lameter6446faa2008-02-15 23:45:26 -0800294/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700295static inline int check_valid_pointer(struct kmem_cache *s,
296 struct page *page, const void *object)
297{
298 void *base;
299
Christoph Lametera973e9d2008-03-01 13:40:44 -0800300 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700301 return 1;
302
Christoph Lametera973e9d2008-03-01 13:40:44 -0800303 base = page_address(page);
Christoph Lameter02cbc872007-05-09 02:32:43 -0700304 if (object < base || object >= base + s->objects * s->size ||
305 (object - base) % s->size) {
306 return 0;
307 }
308
309 return 1;
310}
311
Christoph Lameter81819f02007-05-06 14:49:36 -0700312/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700313 * Slow version of get and set free pointer.
314 *
315 * This version requires touching the cache lines of kmem_cache which
316 * we avoid to do in the fast alloc free paths. There we obtain the offset
317 * from the page struct.
318 */
319static inline void *get_freepointer(struct kmem_cache *s, void *object)
320{
321 return *(void **)(object + s->offset);
322}
323
324static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
325{
326 *(void **)(object + s->offset) = fp;
327}
328
329/* Loop over all objects in a slab */
330#define for_each_object(__p, __s, __addr) \
331 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
332 __p += (__s)->size)
333
334/* Scan freelist */
335#define for_each_free_object(__p, __s, __free) \
Christoph Lametera973e9d2008-03-01 13:40:44 -0800336 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700337
338/* Determine object index from a given position */
339static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
340{
341 return (p - addr) / s->size;
342}
343
Christoph Lameter41ecc552007-05-09 02:32:44 -0700344#ifdef CONFIG_SLUB_DEBUG
345/*
346 * Debug settings:
347 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700348#ifdef CONFIG_SLUB_DEBUG_ON
349static int slub_debug = DEBUG_DEFAULT_FLAGS;
350#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700351static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700352#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700353
354static char *slub_debug_slabs;
355
Christoph Lameter7656c722007-05-09 02:32:40 -0700356/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700357 * Object debugging
358 */
359static void print_section(char *text, u8 *addr, unsigned int length)
360{
361 int i, offset;
362 int newline = 1;
363 char ascii[17];
364
365 ascii[16] = 0;
366
367 for (i = 0; i < length; i++) {
368 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700369 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700370 newline = 0;
371 }
Pekka Enberg06428782008-01-07 23:20:27 -0800372 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700373 offset = i % 16;
374 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
375 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800376 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700377 newline = 1;
378 }
379 }
380 if (!newline) {
381 i %= 16;
382 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800383 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700384 ascii[i] = ' ';
385 i++;
386 }
Pekka Enberg06428782008-01-07 23:20:27 -0800387 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700388 }
389}
390
Christoph Lameter81819f02007-05-06 14:49:36 -0700391static struct track *get_track(struct kmem_cache *s, void *object,
392 enum track_item alloc)
393{
394 struct track *p;
395
396 if (s->offset)
397 p = object + s->offset + sizeof(void *);
398 else
399 p = object + s->inuse;
400
401 return p + alloc;
402}
403
404static void set_track(struct kmem_cache *s, void *object,
405 enum track_item alloc, void *addr)
406{
407 struct track *p;
408
409 if (s->offset)
410 p = object + s->offset + sizeof(void *);
411 else
412 p = object + s->inuse;
413
414 p += alloc;
415 if (addr) {
416 p->addr = addr;
417 p->cpu = smp_processor_id();
418 p->pid = current ? current->pid : -1;
419 p->when = jiffies;
420 } else
421 memset(p, 0, sizeof(struct track));
422}
423
Christoph Lameter81819f02007-05-06 14:49:36 -0700424static void init_tracking(struct kmem_cache *s, void *object)
425{
Christoph Lameter24922682007-07-17 04:03:18 -0700426 if (!(s->flags & SLAB_STORE_USER))
427 return;
428
429 set_track(s, object, TRACK_FREE, NULL);
430 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700431}
432
433static void print_track(const char *s, struct track *t)
434{
435 if (!t->addr)
436 return;
437
Christoph Lameter24922682007-07-17 04:03:18 -0700438 printk(KERN_ERR "INFO: %s in ", s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700439 __print_symbol("%s", (unsigned long)t->addr);
Christoph Lameter24922682007-07-17 04:03:18 -0700440 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700441}
442
Christoph Lameter24922682007-07-17 04:03:18 -0700443static void print_tracking(struct kmem_cache *s, void *object)
444{
445 if (!(s->flags & SLAB_STORE_USER))
446 return;
447
448 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
449 print_track("Freed", get_track(s, object, TRACK_FREE));
450}
451
452static void print_page_info(struct page *page)
453{
454 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
455 page, page->inuse, page->freelist, page->flags);
456
457}
458
459static void slab_bug(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 "========================================"
468 "=====================================\n");
469 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
470 printk(KERN_ERR "----------------------------------------"
471 "-------------------------------------\n\n");
472}
473
474static void slab_fix(struct kmem_cache *s, char *fmt, ...)
475{
476 va_list args;
477 char buf[100];
478
479 va_start(args, fmt);
480 vsnprintf(buf, sizeof(buf), fmt, args);
481 va_end(args);
482 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
483}
484
485static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700486{
487 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800488 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700489
490 print_tracking(s, p);
491
492 print_page_info(page);
493
494 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
495 p, p - addr, get_freepointer(s, p));
496
497 if (p > addr + 16)
498 print_section("Bytes b4", p - 16, 16);
499
500 print_section("Object", p, min(s->objsize, 128));
Christoph Lameter81819f02007-05-06 14:49:36 -0700501
502 if (s->flags & SLAB_RED_ZONE)
503 print_section("Redzone", p + s->objsize,
504 s->inuse - s->objsize);
505
Christoph Lameter81819f02007-05-06 14:49:36 -0700506 if (s->offset)
507 off = s->offset + sizeof(void *);
508 else
509 off = s->inuse;
510
Christoph Lameter24922682007-07-17 04:03:18 -0700511 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700512 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700513
514 if (off != s->size)
515 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700516 print_section("Padding", p + off, s->size - off);
517
518 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700519}
520
521static void object_err(struct kmem_cache *s, struct page *page,
522 u8 *object, char *reason)
523{
Christoph Lameter24922682007-07-17 04:03:18 -0700524 slab_bug(s, reason);
525 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700526}
527
Christoph Lameter24922682007-07-17 04:03:18 -0700528static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700529{
530 va_list args;
531 char buf[100];
532
Christoph Lameter24922682007-07-17 04:03:18 -0700533 va_start(args, fmt);
534 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700535 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700536 slab_bug(s, fmt);
537 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700538 dump_stack();
539}
540
541static void init_object(struct kmem_cache *s, void *object, int active)
542{
543 u8 *p = object;
544
545 if (s->flags & __OBJECT_POISON) {
546 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800547 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700548 }
549
550 if (s->flags & SLAB_RED_ZONE)
551 memset(p + s->objsize,
552 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
553 s->inuse - s->objsize);
554}
555
Christoph Lameter24922682007-07-17 04:03:18 -0700556static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700557{
558 while (bytes) {
559 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700560 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700561 start++;
562 bytes--;
563 }
Christoph Lameter24922682007-07-17 04:03:18 -0700564 return NULL;
565}
566
567static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
568 void *from, void *to)
569{
570 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
571 memset(from, data, to - from);
572}
573
574static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
575 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800576 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700577{
578 u8 *fault;
579 u8 *end;
580
581 fault = check_bytes(start, value, bytes);
582 if (!fault)
583 return 1;
584
585 end = start + bytes;
586 while (end > fault && end[-1] == value)
587 end--;
588
589 slab_bug(s, "%s overwritten", what);
590 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
591 fault, end - 1, fault[0], value);
592 print_trailer(s, page, object);
593
594 restore_bytes(s, what, value, fault, end);
595 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700596}
597
Christoph Lameter81819f02007-05-06 14:49:36 -0700598/*
599 * Object layout:
600 *
601 * object address
602 * Bytes of the object to be managed.
603 * If the freepointer may overlay the object then the free
604 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700605 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700606 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
607 * 0xa5 (POISON_END)
608 *
609 * object + s->objsize
610 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700611 * Padding is extended by another word if Redzoning is enabled and
612 * objsize == inuse.
613 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700614 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
615 * 0xcc (RED_ACTIVE) for objects in use.
616 *
617 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700618 * Meta data starts here.
619 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700620 * A. Free pointer (if we cannot overwrite object on free)
621 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700622 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800623 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700624 * before the word boundary.
625 *
626 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700627 *
628 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700629 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700630 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700631 * If slabcaches are merged then the objsize and inuse boundaries are mostly
632 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700633 * may be used with merged slabcaches.
634 */
635
Christoph Lameter81819f02007-05-06 14:49:36 -0700636static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
637{
638 unsigned long off = s->inuse; /* The end of info */
639
640 if (s->offset)
641 /* Freepointer is placed after the object. */
642 off += sizeof(void *);
643
644 if (s->flags & SLAB_STORE_USER)
645 /* We also have user information there */
646 off += 2 * sizeof(struct track);
647
648 if (s->size == off)
649 return 1;
650
Christoph Lameter24922682007-07-17 04:03:18 -0700651 return check_bytes_and_report(s, page, p, "Object padding",
652 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700653}
654
655static int slab_pad_check(struct kmem_cache *s, struct page *page)
656{
Christoph Lameter24922682007-07-17 04:03:18 -0700657 u8 *start;
658 u8 *fault;
659 u8 *end;
660 int length;
661 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700662
663 if (!(s->flags & SLAB_POISON))
664 return 1;
665
Christoph Lametera973e9d2008-03-01 13:40:44 -0800666 start = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700667 end = start + (PAGE_SIZE << s->order);
Christoph Lameter81819f02007-05-06 14:49:36 -0700668 length = s->objects * s->size;
Christoph Lameter24922682007-07-17 04:03:18 -0700669 remainder = end - (start + length);
Christoph Lameter81819f02007-05-06 14:49:36 -0700670 if (!remainder)
671 return 1;
672
Christoph Lameter24922682007-07-17 04:03:18 -0700673 fault = check_bytes(start + length, POISON_INUSE, remainder);
674 if (!fault)
675 return 1;
676 while (end > fault && end[-1] == POISON_INUSE)
677 end--;
678
679 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
680 print_section("Padding", start, length);
681
682 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
683 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700684}
685
686static int check_object(struct kmem_cache *s, struct page *page,
687 void *object, int active)
688{
689 u8 *p = object;
690 u8 *endobject = object + s->objsize;
691
692 if (s->flags & SLAB_RED_ZONE) {
693 unsigned int red =
694 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
695
Christoph Lameter24922682007-07-17 04:03:18 -0700696 if (!check_bytes_and_report(s, page, object, "Redzone",
697 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700698 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700699 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800700 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
701 check_bytes_and_report(s, page, p, "Alignment padding",
702 endobject, POISON_INUSE, s->inuse - s->objsize);
703 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700704 }
705
706 if (s->flags & SLAB_POISON) {
707 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700708 (!check_bytes_and_report(s, page, p, "Poison", p,
709 POISON_FREE, s->objsize - 1) ||
710 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800711 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700712 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700713 /*
714 * check_pad_bytes cleans up on its own.
715 */
716 check_pad_bytes(s, page, p);
717 }
718
719 if (!s->offset && active)
720 /*
721 * Object and freepointer overlap. Cannot check
722 * freepointer while object is allocated.
723 */
724 return 1;
725
726 /* Check free pointer validity */
727 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
728 object_err(s, page, p, "Freepointer corrupt");
729 /*
730 * No choice but to zap it and thus loose the remainder
731 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700732 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700733 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800734 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700735 return 0;
736 }
737 return 1;
738}
739
740static int check_slab(struct kmem_cache *s, struct page *page)
741{
742 VM_BUG_ON(!irqs_disabled());
743
744 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700745 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700746 return 0;
747 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700748 if (page->inuse > s->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700749 slab_err(s, page, "inuse %u > max %u",
750 s->name, page->inuse, s->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700751 return 0;
752 }
753 /* Slab_pad_check fixes things up after itself */
754 slab_pad_check(s, page);
755 return 1;
756}
757
758/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700759 * Determine if a certain object on a page is on the freelist. Must hold the
760 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700761 */
762static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
763{
764 int nr = 0;
765 void *fp = page->freelist;
766 void *object = NULL;
767
Christoph Lametera973e9d2008-03-01 13:40:44 -0800768 while (fp && nr <= s->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700769 if (fp == search)
770 return 1;
771 if (!check_valid_pointer(s, page, fp)) {
772 if (object) {
773 object_err(s, page, object,
774 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800775 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700776 break;
777 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700778 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800779 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700780 page->inuse = s->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700781 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700782 return 0;
783 }
784 break;
785 }
786 object = fp;
787 fp = get_freepointer(s, object);
788 nr++;
789 }
790
791 if (page->inuse != s->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700792 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter24922682007-07-17 04:03:18 -0700793 "counted were %d", page->inuse, s->objects - nr);
Christoph Lameter81819f02007-05-06 14:49:36 -0700794 page->inuse = s->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700795 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700796 }
797 return search == NULL;
798}
799
Christoph Lameter3ec09742007-05-16 22:11:00 -0700800static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
801{
802 if (s->flags & SLAB_TRACE) {
803 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
804 s->name,
805 alloc ? "alloc" : "free",
806 object, page->inuse,
807 page->freelist);
808
809 if (!alloc)
810 print_section("Object", (void *)object, s->objsize);
811
812 dump_stack();
813 }
814}
815
Christoph Lameter643b1132007-05-06 14:49:42 -0700816/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700817 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700818 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700819static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700820{
Christoph Lameter643b1132007-05-06 14:49:42 -0700821 spin_lock(&n->list_lock);
822 list_add(&page->lru, &n->full);
823 spin_unlock(&n->list_lock);
824}
825
826static void remove_full(struct kmem_cache *s, struct page *page)
827{
828 struct kmem_cache_node *n;
829
830 if (!(s->flags & SLAB_STORE_USER))
831 return;
832
833 n = get_node(s, page_to_nid(page));
834
835 spin_lock(&n->list_lock);
836 list_del(&page->lru);
837 spin_unlock(&n->list_lock);
838}
839
Christoph Lameter3ec09742007-05-16 22:11:00 -0700840static void setup_object_debug(struct kmem_cache *s, struct page *page,
841 void *object)
842{
843 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
844 return;
845
846 init_object(s, object, 0);
847 init_tracking(s, object);
848}
849
850static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
851 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700852{
853 if (!check_slab(s, page))
854 goto bad;
855
Christoph Lameterd692ef62008-02-15 23:45:24 -0800856 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700857 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700858 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700859 }
860
861 if (!check_valid_pointer(s, page, object)) {
862 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700863 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700864 }
865
Christoph Lameterd692ef62008-02-15 23:45:24 -0800866 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700867 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700868
Christoph Lameter3ec09742007-05-16 22:11:00 -0700869 /* Success perform special debug activities for allocs */
870 if (s->flags & SLAB_STORE_USER)
871 set_track(s, object, TRACK_ALLOC, addr);
872 trace(s, page, object, 1);
873 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700874 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700875
Christoph Lameter81819f02007-05-06 14:49:36 -0700876bad:
877 if (PageSlab(page)) {
878 /*
879 * If this is a slab page then lets do the best we can
880 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700881 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700882 */
Christoph Lameter24922682007-07-17 04:03:18 -0700883 slab_fix(s, "Marking all objects used");
Christoph Lameter81819f02007-05-06 14:49:36 -0700884 page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800885 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700886 }
887 return 0;
888}
889
Christoph Lameter3ec09742007-05-16 22:11:00 -0700890static int free_debug_processing(struct kmem_cache *s, struct page *page,
891 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700892{
893 if (!check_slab(s, page))
894 goto fail;
895
896 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700897 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700898 goto fail;
899 }
900
901 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700902 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700903 goto fail;
904 }
905
906 if (!check_object(s, page, object, 1))
907 return 0;
908
909 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800910 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700911 slab_err(s, page, "Attempt to free object(0x%p) "
912 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800913 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700914 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700915 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700916 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700917 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800918 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700919 object_err(s, page, object,
920 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700921 goto fail;
922 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700923
924 /* Special debug activities for freeing objects */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800925 if (!SlabFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700926 remove_full(s, page);
927 if (s->flags & SLAB_STORE_USER)
928 set_track(s, object, TRACK_FREE, addr);
929 trace(s, page, object, 0);
930 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700931 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700932
Christoph Lameter81819f02007-05-06 14:49:36 -0700933fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700934 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700935 return 0;
936}
937
Christoph Lameter41ecc552007-05-09 02:32:44 -0700938static int __init setup_slub_debug(char *str)
939{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700940 slub_debug = DEBUG_DEFAULT_FLAGS;
941 if (*str++ != '=' || !*str)
942 /*
943 * No options specified. Switch on full debugging.
944 */
945 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700946
947 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700948 /*
949 * No options but restriction on slabs. This means full
950 * debugging for slabs matching a pattern.
951 */
952 goto check_slabs;
953
954 slub_debug = 0;
955 if (*str == '-')
956 /*
957 * Switch off all debugging measures.
958 */
959 goto out;
960
961 /*
962 * Determine which debug features should be switched on
963 */
Pekka Enberg06428782008-01-07 23:20:27 -0800964 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700965 switch (tolower(*str)) {
966 case 'f':
967 slub_debug |= SLAB_DEBUG_FREE;
968 break;
969 case 'z':
970 slub_debug |= SLAB_RED_ZONE;
971 break;
972 case 'p':
973 slub_debug |= SLAB_POISON;
974 break;
975 case 'u':
976 slub_debug |= SLAB_STORE_USER;
977 break;
978 case 't':
979 slub_debug |= SLAB_TRACE;
980 break;
981 default:
982 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -0800983 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700984 }
985 }
986
987check_slabs:
988 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -0700989 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700990out:
Christoph Lameter41ecc552007-05-09 02:32:44 -0700991 return 1;
992}
993
994__setup("slub_debug", setup_slub_debug);
995
Christoph Lameterba0268a2007-09-11 15:24:11 -0700996static unsigned long kmem_cache_flags(unsigned long objsize,
997 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -0700998 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -0700999{
1000 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001001 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001002 */
Christoph Lametere1533622008-02-15 23:45:24 -08001003 if (slub_debug && (!slub_debug_slabs ||
1004 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1005 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001006
1007 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001008}
1009#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001010static inline void setup_object_debug(struct kmem_cache *s,
1011 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001012
Christoph Lameter3ec09742007-05-16 22:11:00 -07001013static inline int alloc_debug_processing(struct kmem_cache *s,
1014 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001015
Christoph Lameter3ec09742007-05-16 22:11:00 -07001016static inline int free_debug_processing(struct kmem_cache *s,
1017 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001018
Christoph Lameter41ecc552007-05-09 02:32:44 -07001019static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1020 { return 1; }
1021static inline int check_object(struct kmem_cache *s, struct page *page,
1022 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001023static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001024static inline unsigned long kmem_cache_flags(unsigned long objsize,
1025 unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001026 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001027{
1028 return flags;
1029}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001030#define slub_debug 0
1031#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001032/*
1033 * Slab allocation and freeing
1034 */
1035static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1036{
Pekka Enberg06428782008-01-07 23:20:27 -08001037 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001038 int pages = 1 << s->order;
1039
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001040 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001041
Christoph Lameter81819f02007-05-06 14:49:36 -07001042 if (node == -1)
1043 page = alloc_pages(flags, s->order);
1044 else
1045 page = alloc_pages_node(node, flags, s->order);
1046
1047 if (!page)
1048 return NULL;
1049
1050 mod_zone_page_state(page_zone(page),
1051 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1052 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1053 pages);
1054
1055 return page;
1056}
1057
1058static void setup_object(struct kmem_cache *s, struct page *page,
1059 void *object)
1060{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001061 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001062 if (unlikely(s->ctor))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07001063 s->ctor(s, object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001064}
1065
1066static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1067{
1068 struct page *page;
1069 struct kmem_cache_node *n;
1070 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001071 void *last;
1072 void *p;
1073
Christoph Lameter6cb06222007-10-16 01:25:41 -07001074 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001075
Christoph Lameter6cb06222007-10-16 01:25:41 -07001076 page = allocate_slab(s,
1077 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001078 if (!page)
1079 goto out;
1080
1081 n = get_node(s, page_to_nid(page));
1082 if (n)
1083 atomic_long_inc(&n->nr_slabs);
Christoph Lameter81819f02007-05-06 14:49:36 -07001084 page->slab = s;
1085 page->flags |= 1 << PG_slab;
1086 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1087 SLAB_STORE_USER | SLAB_TRACE))
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001088 SetSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001089
1090 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001091
1092 if (unlikely(s->flags & SLAB_POISON))
1093 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1094
1095 last = start;
Christoph Lameter7656c722007-05-09 02:32:40 -07001096 for_each_object(p, s, start) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001097 setup_object(s, page, last);
1098 set_freepointer(s, last, p);
1099 last = p;
1100 }
1101 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001102 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001103
1104 page->freelist = start;
1105 page->inuse = 0;
1106out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001107 return page;
1108}
1109
1110static void __free_slab(struct kmem_cache *s, struct page *page)
1111{
1112 int pages = 1 << s->order;
1113
Christoph Lameterc59def9f2007-05-16 22:10:50 -07001114 if (unlikely(SlabDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001115 void *p;
1116
1117 slab_pad_check(s, page);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001118 for_each_object(p, s, page_address(page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001119 check_object(s, page, p, 0);
Peter Zijlstra2208b762007-07-26 20:54:34 +02001120 ClearSlabDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001121 }
1122
1123 mod_zone_page_state(page_zone(page),
1124 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1125 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001126 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001127
Christoph Lameter81819f02007-05-06 14:49:36 -07001128 __free_pages(page, s->order);
1129}
1130
1131static void rcu_free_slab(struct rcu_head *h)
1132{
1133 struct page *page;
1134
1135 page = container_of((struct list_head *)h, struct page, lru);
1136 __free_slab(page->slab, page);
1137}
1138
1139static void free_slab(struct kmem_cache *s, struct page *page)
1140{
1141 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1142 /*
1143 * RCU free overloads the RCU head over the LRU
1144 */
1145 struct rcu_head *head = (void *)&page->lru;
1146
1147 call_rcu(head, rcu_free_slab);
1148 } else
1149 __free_slab(s, page);
1150}
1151
1152static void discard_slab(struct kmem_cache *s, struct page *page)
1153{
1154 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1155
1156 atomic_long_dec(&n->nr_slabs);
1157 reset_page_mapcount(page);
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001158 __ClearPageSlab(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001159 free_slab(s, page);
1160}
1161
1162/*
1163 * Per slab locking using the pagelock
1164 */
1165static __always_inline void slab_lock(struct page *page)
1166{
1167 bit_spin_lock(PG_locked, &page->flags);
1168}
1169
1170static __always_inline void slab_unlock(struct page *page)
1171{
Nick Piggina76d3542008-01-07 23:20:27 -08001172 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001173}
1174
1175static __always_inline int slab_trylock(struct page *page)
1176{
1177 int rc = 1;
1178
1179 rc = bit_spin_trylock(PG_locked, &page->flags);
1180 return rc;
1181}
1182
1183/*
1184 * Management of partially allocated slabs
1185 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001186static void add_partial(struct kmem_cache_node *n,
1187 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001188{
Christoph Lametere95eed52007-05-06 14:49:44 -07001189 spin_lock(&n->list_lock);
1190 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001191 if (tail)
1192 list_add_tail(&page->lru, &n->partial);
1193 else
1194 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001195 spin_unlock(&n->list_lock);
1196}
1197
1198static void remove_partial(struct kmem_cache *s,
1199 struct page *page)
1200{
1201 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1202
1203 spin_lock(&n->list_lock);
1204 list_del(&page->lru);
1205 n->nr_partial--;
1206 spin_unlock(&n->list_lock);
1207}
1208
1209/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001210 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001211 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001212 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001213 */
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001214static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001215{
1216 if (slab_trylock(page)) {
1217 list_del(&page->lru);
1218 n->nr_partial--;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001219 SetSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001220 return 1;
1221 }
1222 return 0;
1223}
1224
1225/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001226 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001227 */
1228static struct page *get_partial_node(struct kmem_cache_node *n)
1229{
1230 struct page *page;
1231
1232 /*
1233 * Racy check. If we mistakenly see no partial slabs then we
1234 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001235 * partial slab and there is none available then get_partials()
1236 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001237 */
1238 if (!n || !n->nr_partial)
1239 return NULL;
1240
1241 spin_lock(&n->list_lock);
1242 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001243 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001244 goto out;
1245 page = NULL;
1246out:
1247 spin_unlock(&n->list_lock);
1248 return page;
1249}
1250
1251/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001252 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001253 */
1254static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1255{
1256#ifdef CONFIG_NUMA
1257 struct zonelist *zonelist;
1258 struct zone **z;
1259 struct page *page;
1260
1261 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001262 * The defrag ratio allows a configuration of the tradeoffs between
1263 * inter node defragmentation and node local allocations. A lower
1264 * defrag_ratio increases the tendency to do local allocations
1265 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001266 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001267 * If the defrag_ratio is set to 0 then kmalloc() always
1268 * returns node local objects. If the ratio is higher then kmalloc()
1269 * may return off node objects because partial slabs are obtained
1270 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001271 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001272 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001273 * defrag_ratio = 1000) then every (well almost) allocation will
1274 * first attempt to defrag slab caches on other nodes. This means
1275 * scanning over all nodes to look for partial slabs which may be
1276 * expensive if we do it every time we are trying to find a slab
1277 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001278 */
Christoph Lameter98246012008-01-07 23:20:26 -08001279 if (!s->remote_node_defrag_ratio ||
1280 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001281 return NULL;
1282
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001283 zonelist = &NODE_DATA(
1284 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
Christoph Lameter81819f02007-05-06 14:49:36 -07001285 for (z = zonelist->zones; *z; z++) {
1286 struct kmem_cache_node *n;
1287
1288 n = get_node(s, zone_to_nid(*z));
1289
1290 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001291 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001292 page = get_partial_node(n);
1293 if (page)
1294 return page;
1295 }
1296 }
1297#endif
1298 return NULL;
1299}
1300
1301/*
1302 * Get a partial page, lock it and return it.
1303 */
1304static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1305{
1306 struct page *page;
1307 int searchnode = (node == -1) ? numa_node_id() : node;
1308
1309 page = get_partial_node(get_node(s, searchnode));
1310 if (page || (flags & __GFP_THISNODE))
1311 return page;
1312
1313 return get_any_partial(s, flags);
1314}
1315
1316/*
1317 * Move a page back to the lists.
1318 *
1319 * Must be called with the slab lock held.
1320 *
1321 * On exit the slab lock will have been dropped.
1322 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001323static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001324{
Christoph Lametere95eed52007-05-06 14:49:44 -07001325 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001326 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001327
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001328 ClearSlabFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001329 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001330
Christoph Lametera973e9d2008-03-01 13:40:44 -08001331 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001332 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001333 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1334 } else {
1335 stat(c, DEACTIVATE_FULL);
1336 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1337 add_full(n, page);
1338 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001339 slab_unlock(page);
1340 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001341 stat(c, DEACTIVATE_EMPTY);
Christoph Lametere95eed52007-05-06 14:49:44 -07001342 if (n->nr_partial < MIN_PARTIAL) {
1343 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001344 * Adding an empty slab to the partial slabs in order
1345 * to avoid page allocator overhead. This slab needs
1346 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001347 * so that the others get filled first. That way the
1348 * size of the partial list stays small.
1349 *
1350 * kmem_cache_shrink can reclaim any empty slabs from the
1351 * partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001352 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001353 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001354 slab_unlock(page);
1355 } else {
1356 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001357 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001358 discard_slab(s, page);
1359 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001360 }
1361}
1362
1363/*
1364 * Remove the cpu slab
1365 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001366static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001367{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001368 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001369 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001370
Christoph Lameterb773ad72008-03-04 11:10:17 -08001371 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001372 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001373 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001374 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001375 * because both freelists are empty. So this is unlikely
1376 * to occur.
1377 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001378 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001379 void **object;
1380
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001381 tail = 0; /* Hot objects. Put the slab first */
1382
Christoph Lameter894b8782007-05-10 03:15:16 -07001383 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001384 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001385 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001386
1387 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001388 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001389 page->freelist = object;
1390 page->inuse--;
1391 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001392 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001393 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001394}
1395
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001396static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001397{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001398 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001399 slab_lock(c->page);
1400 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001401}
1402
1403/*
1404 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001405 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001406 * Called from IPI handler with interrupts disabled.
1407 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001408static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001409{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001410 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001411
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001412 if (likely(c && c->page))
1413 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001414}
1415
1416static void flush_cpu_slab(void *d)
1417{
1418 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001419
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001420 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001421}
1422
1423static void flush_all(struct kmem_cache *s)
1424{
1425#ifdef CONFIG_SMP
1426 on_each_cpu(flush_cpu_slab, s, 1, 1);
1427#else
1428 unsigned long flags;
1429
1430 local_irq_save(flags);
1431 flush_cpu_slab(s);
1432 local_irq_restore(flags);
1433#endif
1434}
1435
1436/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001437 * Check if the objects in a per cpu structure fit numa
1438 * locality expectations.
1439 */
1440static inline int node_match(struct kmem_cache_cpu *c, int node)
1441{
1442#ifdef CONFIG_NUMA
1443 if (node != -1 && c->node != node)
1444 return 0;
1445#endif
1446 return 1;
1447}
1448
1449/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001450 * Slow path. The lockless freelist is empty or we need to perform
1451 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001452 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001453 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001454 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001455 * Processing is still very fast if new objects have been freed to the
1456 * regular freelist. In that case we simply take over the regular freelist
1457 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001458 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001459 * If that is not working then we fall back to the partial lists. We take the
1460 * first element of the freelist as the object to allocate now and move the
1461 * rest of the freelist to the lockless freelist.
1462 *
1463 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001464 * we need to allocate a new slab. This is the slowest path since it involves
1465 * a call to the page allocator and the setup of a new slab.
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
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001473 /* We handle __GFP_ZERO in the caller */
1474 gfpflags &= ~__GFP_ZERO;
1475
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001476 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001477 goto new_slab;
1478
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001479 slab_lock(c->page);
1480 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001481 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001482
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001483 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001484
Christoph Lameter894b8782007-05-10 03:15:16 -07001485load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001486 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001487 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001488 goto another_slab;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001489 if (unlikely(SlabDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001490 goto debug;
1491
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001492 c->freelist = object[c->offset];
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001493 c->page->inuse = s->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001494 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001495 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001496unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001497 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001498 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001499 return object;
1500
1501another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001502 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001503
1504new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001505 new = get_partial(s, gfpflags, node);
1506 if (new) {
1507 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001508 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001509 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001510 }
1511
Christoph Lameterb811c202007-10-16 23:25:51 -07001512 if (gfpflags & __GFP_WAIT)
1513 local_irq_enable();
1514
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001515 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001516
1517 if (gfpflags & __GFP_WAIT)
1518 local_irq_disable();
1519
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001520 if (new) {
1521 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001522 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001523 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001524 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001525 slab_lock(new);
1526 SetSlabFrozen(new);
1527 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001528 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001529 }
Linus Torvalds00e962c2008-02-19 09:08:49 -08001530
Christoph Lameter71c7a062008-02-14 14:28:01 -08001531 /*
1532 * No memory available.
1533 *
1534 * If the slab uses higher order allocs but the object is
1535 * smaller than a page size then we can fallback in emergencies
1536 * to the page allocator via kmalloc_large. The page allocator may
1537 * have failed to obtain a higher order page and we can try to
1538 * allocate a single page if the object fits into a single page.
1539 * That is only possible if certain conditions are met that are being
1540 * checked when a slab is created.
1541 */
Christoph Lametercaeab082008-03-12 23:57:49 -07001542 if (!(gfpflags & __GFP_NORETRY) &&
1543 (s->flags & __PAGE_ALLOC_FALLBACK)) {
1544 if (gfpflags & __GFP_WAIT)
1545 local_irq_enable();
1546 object = kmalloc_large(s->objsize, gfpflags);
1547 if (gfpflags & __GFP_WAIT)
1548 local_irq_disable();
1549 return object;
1550 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001551 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001552debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001553 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001554 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001555
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001556 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001557 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001558 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001559 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001560}
1561
1562/*
1563 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1564 * have the fastpath folded into their functions. So no function call
1565 * overhead for requests that can be satisfied on the fastpath.
1566 *
1567 * The fastpath works by first checking if the lockless freelist can be used.
1568 * If not then __slab_alloc is called for slow processing.
1569 *
1570 * Otherwise we can simply pick the next object from the lockless free list.
1571 */
Pekka Enberg06428782008-01-07 23:20:27 -08001572static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001573 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001574{
Christoph Lameter894b8782007-05-10 03:15:16 -07001575 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001576 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001577 unsigned long flags;
1578
Christoph Lameter894b8782007-05-10 03:15:16 -07001579 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001580 c = get_cpu_slab(s, smp_processor_id());
Christoph Lametera973e9d2008-03-01 13:40:44 -08001581 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001582
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001583 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001584
1585 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001586 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001587 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001588 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001589 }
1590 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001591
1592 if (unlikely((gfpflags & __GFP_ZERO) && object))
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001593 memset(object, 0, c->objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001594
Christoph Lameter894b8782007-05-10 03:15:16 -07001595 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001596}
1597
1598void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1599{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001600 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001601}
1602EXPORT_SYMBOL(kmem_cache_alloc);
1603
1604#ifdef CONFIG_NUMA
1605void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1606{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001607 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001608}
1609EXPORT_SYMBOL(kmem_cache_alloc_node);
1610#endif
1611
1612/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001613 * Slow patch handling. This may still be called frequently since objects
1614 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001615 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001616 * So we still attempt to reduce cache line usage. Just take the slab
1617 * lock and free the item. If there is no additional partial page
1618 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001619 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001620static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001621 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001622{
1623 void *prior;
1624 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001625 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001626
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001627 c = get_cpu_slab(s, raw_smp_processor_id());
1628 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001629 slab_lock(page);
1630
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07001631 if (unlikely(SlabDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001632 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001633
Christoph Lameter81819f02007-05-06 14:49:36 -07001634checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001635 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001636 page->freelist = object;
1637 page->inuse--;
1638
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001639 if (unlikely(SlabFrozen(page))) {
1640 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001641 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001642 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001643
1644 if (unlikely(!page->inuse))
1645 goto slab_empty;
1646
1647 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001648 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001649 * then add it.
1650 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001651 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001652 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001653 stat(c, FREE_ADD_PARTIAL);
1654 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001655
1656out_unlock:
1657 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001658 return;
1659
1660slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001661 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001662 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001663 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001664 */
1665 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001666 stat(c, FREE_REMOVE_PARTIAL);
1667 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001668 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001669 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001670 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001671 return;
1672
1673debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001674 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001675 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001676 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001677}
1678
Christoph Lameter894b8782007-05-10 03:15:16 -07001679/*
1680 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1681 * can perform fastpath freeing without additional function calls.
1682 *
1683 * The fastpath is only possible if we are freeing to the current cpu slab
1684 * of this processor. This typically the case if we have just allocated
1685 * the item before.
1686 *
1687 * If fastpath is not possible then fall back to __slab_free where we deal
1688 * with all sorts of special processing.
1689 */
Pekka Enberg06428782008-01-07 23:20:27 -08001690static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001691 struct page *page, void *x, void *addr)
1692{
1693 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001694 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001695 unsigned long flags;
1696
Christoph Lameter894b8782007-05-10 03:15:16 -07001697 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001698 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001699 debug_check_no_locks_freed(object, c->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001700 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001701 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001702 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001703 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001704 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001705 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001706
1707 local_irq_restore(flags);
1708}
1709
Christoph Lameter81819f02007-05-06 14:49:36 -07001710void kmem_cache_free(struct kmem_cache *s, void *x)
1711{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001712 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001713
Christoph Lameterb49af682007-05-06 14:49:41 -07001714 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001715
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001716 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001717}
1718EXPORT_SYMBOL(kmem_cache_free);
1719
1720/* Figure out on which slab object the object resides */
1721static struct page *get_object_page(const void *x)
1722{
Christoph Lameterb49af682007-05-06 14:49:41 -07001723 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001724
1725 if (!PageSlab(page))
1726 return NULL;
1727
1728 return page;
1729}
1730
1731/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001732 * Object placement in a slab is made very easy because we always start at
1733 * offset 0. If we tune the size of the object to the alignment then we can
1734 * get the required alignment by putting one properly sized object after
1735 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001736 *
1737 * Notice that the allocation order determines the sizes of the per cpu
1738 * caches. Each processor has always one slab available for allocations.
1739 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001740 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001741 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001742 */
1743
1744/*
1745 * Mininum / Maximum order of slab pages. This influences locking overhead
1746 * and slab fragmentation. A higher order reduces the number of partial slabs
1747 * and increases the number of allocations possible without having to
1748 * take the list_lock.
1749 */
1750static int slub_min_order;
1751static int slub_max_order = DEFAULT_MAX_ORDER;
Christoph Lameter81819f02007-05-06 14:49:36 -07001752static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1753
1754/*
1755 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001756 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001757 */
1758static int slub_nomerge;
1759
1760/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001761 * Calculate the order of allocation given an slab object size.
1762 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001763 * The order of allocation has significant impact on performance and other
1764 * system components. Generally order 0 allocations should be preferred since
1765 * order 0 does not cause fragmentation in the page allocator. Larger objects
1766 * be problematic to put into order 0 slabs because there may be too much
1767 * unused space left. We go to a higher order if more than 1/8th of the slab
1768 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001769 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001770 * In order to reach satisfactory performance we must ensure that a minimum
1771 * number of objects is in one slab. Otherwise we may generate too much
1772 * activity on the partial lists which requires taking the list_lock. This is
1773 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001774 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001775 * slub_max_order specifies the order where we begin to stop considering the
1776 * number of objects in a slab as critical. If we reach slub_max_order then
1777 * we try to keep the page order as low as possible. So we accept more waste
1778 * of space in favor of a small page order.
1779 *
1780 * Higher order allocations also allow the placement of more objects in a
1781 * slab and thereby reduce object handling overhead. If the user has
1782 * requested a higher mininum order then we start with that one instead of
1783 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001784 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001785static inline int slab_order(int size, int min_objects,
1786 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001787{
1788 int order;
1789 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001790 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001791
Christoph Lameter6300ea72007-07-17 04:03:20 -07001792 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001793 fls(min_objects * size - 1) - PAGE_SHIFT);
1794 order <= max_order; order++) {
1795
Christoph Lameter81819f02007-05-06 14:49:36 -07001796 unsigned long slab_size = PAGE_SIZE << order;
1797
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001798 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001799 continue;
1800
Christoph Lameter81819f02007-05-06 14:49:36 -07001801 rem = slab_size % size;
1802
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001803 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001804 break;
1805
1806 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001807
Christoph Lameter81819f02007-05-06 14:49:36 -07001808 return order;
1809}
1810
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001811static inline int calculate_order(int size)
1812{
1813 int order;
1814 int min_objects;
1815 int fraction;
1816
1817 /*
1818 * Attempt to find best configuration for a slab. This
1819 * works by first attempting to generate a layout with
1820 * the best configuration and backing off gradually.
1821 *
1822 * First we reduce the acceptable waste in a slab. Then
1823 * we reduce the minimum objects required in a slab.
1824 */
1825 min_objects = slub_min_objects;
1826 while (min_objects > 1) {
1827 fraction = 8;
1828 while (fraction >= 4) {
1829 order = slab_order(size, min_objects,
1830 slub_max_order, fraction);
1831 if (order <= slub_max_order)
1832 return order;
1833 fraction /= 2;
1834 }
1835 min_objects /= 2;
1836 }
1837
1838 /*
1839 * We were unable to place multiple objects in a slab. Now
1840 * lets see if we can place a single object there.
1841 */
1842 order = slab_order(size, 1, slub_max_order, 1);
1843 if (order <= slub_max_order)
1844 return order;
1845
1846 /*
1847 * Doh this slab cannot be placed using slub_max_order.
1848 */
1849 order = slab_order(size, 1, MAX_ORDER, 1);
1850 if (order <= MAX_ORDER)
1851 return order;
1852 return -ENOSYS;
1853}
1854
Christoph Lameter81819f02007-05-06 14:49:36 -07001855/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001856 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001857 */
1858static unsigned long calculate_alignment(unsigned long flags,
1859 unsigned long align, unsigned long size)
1860{
1861 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001862 * If the user wants hardware cache aligned objects then follow that
1863 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001864 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001865 * The hardware cache alignment cannot override the specified
1866 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001867 */
Nick Pigginb6210382008-03-05 14:05:56 -08001868 if (flags & SLAB_HWCACHE_ALIGN) {
1869 unsigned long ralign = cache_line_size();
1870 while (size <= ralign / 2)
1871 ralign /= 2;
1872 align = max(align, ralign);
1873 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001874
1875 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001876 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001877
1878 return ALIGN(align, sizeof(void *));
1879}
1880
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001881static void init_kmem_cache_cpu(struct kmem_cache *s,
1882 struct kmem_cache_cpu *c)
1883{
1884 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001885 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001886 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001887 c->offset = s->offset / sizeof(void *);
1888 c->objsize = s->objsize;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001889}
1890
Christoph Lameter81819f02007-05-06 14:49:36 -07001891static void init_kmem_cache_node(struct kmem_cache_node *n)
1892{
1893 n->nr_partial = 0;
1894 atomic_long_set(&n->nr_slabs, 0);
1895 spin_lock_init(&n->list_lock);
1896 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001897#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter643b1132007-05-06 14:49:42 -07001898 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001899#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001900}
1901
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001902#ifdef CONFIG_SMP
1903/*
1904 * Per cpu array for per cpu structures.
1905 *
1906 * The per cpu array places all kmem_cache_cpu structures from one processor
1907 * close together meaning that it becomes possible that multiple per cpu
1908 * structures are contained in one cacheline. This may be particularly
1909 * beneficial for the kmalloc caches.
1910 *
1911 * A desktop system typically has around 60-80 slabs. With 100 here we are
1912 * likely able to get per cpu structures for all caches from the array defined
1913 * here. We must be able to cover all kmalloc caches during bootstrap.
1914 *
1915 * If the per cpu array is exhausted then fall back to kmalloc
1916 * of individual cachelines. No sharing is possible then.
1917 */
1918#define NR_KMEM_CACHE_CPU 100
1919
1920static DEFINE_PER_CPU(struct kmem_cache_cpu,
1921 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1922
1923static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1924static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1925
1926static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1927 int cpu, gfp_t flags)
1928{
1929 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1930
1931 if (c)
1932 per_cpu(kmem_cache_cpu_free, cpu) =
1933 (void *)c->freelist;
1934 else {
1935 /* Table overflow: So allocate ourselves */
1936 c = kmalloc_node(
1937 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1938 flags, cpu_to_node(cpu));
1939 if (!c)
1940 return NULL;
1941 }
1942
1943 init_kmem_cache_cpu(s, c);
1944 return c;
1945}
1946
1947static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1948{
1949 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1950 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1951 kfree(c);
1952 return;
1953 }
1954 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1955 per_cpu(kmem_cache_cpu_free, cpu) = c;
1956}
1957
1958static void free_kmem_cache_cpus(struct kmem_cache *s)
1959{
1960 int cpu;
1961
1962 for_each_online_cpu(cpu) {
1963 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1964
1965 if (c) {
1966 s->cpu_slab[cpu] = NULL;
1967 free_kmem_cache_cpu(c, cpu);
1968 }
1969 }
1970}
1971
1972static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1973{
1974 int cpu;
1975
1976 for_each_online_cpu(cpu) {
1977 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1978
1979 if (c)
1980 continue;
1981
1982 c = alloc_kmem_cache_cpu(s, cpu, flags);
1983 if (!c) {
1984 free_kmem_cache_cpus(s);
1985 return 0;
1986 }
1987 s->cpu_slab[cpu] = c;
1988 }
1989 return 1;
1990}
1991
1992/*
1993 * Initialize the per cpu array.
1994 */
1995static void init_alloc_cpu_cpu(int cpu)
1996{
1997 int i;
1998
1999 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2000 return;
2001
2002 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2003 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2004
2005 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2006}
2007
2008static void __init init_alloc_cpu(void)
2009{
2010 int cpu;
2011
2012 for_each_online_cpu(cpu)
2013 init_alloc_cpu_cpu(cpu);
2014 }
2015
2016#else
2017static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2018static inline void init_alloc_cpu(void) {}
2019
2020static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2021{
2022 init_kmem_cache_cpu(s, &s->cpu_slab);
2023 return 1;
2024}
2025#endif
2026
Christoph Lameter81819f02007-05-06 14:49:36 -07002027#ifdef CONFIG_NUMA
2028/*
2029 * No kmalloc_node yet so do it by hand. We know that this is the first
2030 * slab on the node for this slabcache. There are no concurrent accesses
2031 * possible.
2032 *
2033 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002034 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2035 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002036 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002037static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2038 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002039{
2040 struct page *page;
2041 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002042 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002043
2044 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2045
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002046 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002047
2048 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002049 if (page_to_nid(page) != node) {
2050 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2051 "node %d\n", node);
2052 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2053 "in order to be able to continue\n");
2054 }
2055
Christoph Lameter81819f02007-05-06 14:49:36 -07002056 n = page->freelist;
2057 BUG_ON(!n);
2058 page->freelist = get_freepointer(kmalloc_caches, n);
2059 page->inuse++;
2060 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002061#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002062 init_object(kmalloc_caches, n, 1);
2063 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002064#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002065 init_kmem_cache_node(n);
2066 atomic_long_inc(&n->nr_slabs);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002067
rootba84c732008-01-07 23:20:28 -08002068 /*
2069 * lockdep requires consistent irq usage for each lock
2070 * so even though there cannot be a race this early in
2071 * the boot sequence, we still disable irqs.
2072 */
2073 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002074 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002075 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002076 return n;
2077}
2078
2079static void free_kmem_cache_nodes(struct kmem_cache *s)
2080{
2081 int node;
2082
Christoph Lameterf64dc582007-10-16 01:25:33 -07002083 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002084 struct kmem_cache_node *n = s->node[node];
2085 if (n && n != &s->local_node)
2086 kmem_cache_free(kmalloc_caches, n);
2087 s->node[node] = NULL;
2088 }
2089}
2090
2091static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2092{
2093 int node;
2094 int local_node;
2095
2096 if (slab_state >= UP)
2097 local_node = page_to_nid(virt_to_page(s));
2098 else
2099 local_node = 0;
2100
Christoph Lameterf64dc582007-10-16 01:25:33 -07002101 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002102 struct kmem_cache_node *n;
2103
2104 if (local_node == node)
2105 n = &s->local_node;
2106 else {
2107 if (slab_state == DOWN) {
2108 n = early_kmem_cache_node_alloc(gfpflags,
2109 node);
2110 continue;
2111 }
2112 n = kmem_cache_alloc_node(kmalloc_caches,
2113 gfpflags, node);
2114
2115 if (!n) {
2116 free_kmem_cache_nodes(s);
2117 return 0;
2118 }
2119
2120 }
2121 s->node[node] = n;
2122 init_kmem_cache_node(n);
2123 }
2124 return 1;
2125}
2126#else
2127static void free_kmem_cache_nodes(struct kmem_cache *s)
2128{
2129}
2130
2131static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2132{
2133 init_kmem_cache_node(&s->local_node);
2134 return 1;
2135}
2136#endif
2137
2138/*
2139 * calculate_sizes() determines the order and the distribution of data within
2140 * a slab object.
2141 */
2142static int calculate_sizes(struct kmem_cache *s)
2143{
2144 unsigned long flags = s->flags;
2145 unsigned long size = s->objsize;
2146 unsigned long align = s->align;
2147
2148 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002149 * Round up object size to the next word boundary. We can only
2150 * place the free pointer at word boundaries and this determines
2151 * the possible location of the free pointer.
2152 */
2153 size = ALIGN(size, sizeof(void *));
2154
2155#ifdef CONFIG_SLUB_DEBUG
2156 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002157 * Determine if we can poison the object itself. If the user of
2158 * the slab may touch the object after free or before allocation
2159 * then we should never poison the object itself.
2160 */
2161 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002162 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002163 s->flags |= __OBJECT_POISON;
2164 else
2165 s->flags &= ~__OBJECT_POISON;
2166
Christoph Lameter81819f02007-05-06 14:49:36 -07002167
2168 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002169 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002170 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002171 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002172 */
2173 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2174 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002175#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002176
2177 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002178 * With that we have determined the number of bytes in actual use
2179 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002180 */
2181 s->inuse = size;
2182
2183 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002184 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002185 /*
2186 * Relocate free pointer after the object if it is not
2187 * permitted to overwrite the first word of the object on
2188 * kmem_cache_free.
2189 *
2190 * This is the case if we do RCU, have a constructor or
2191 * destructor or are poisoning the objects.
2192 */
2193 s->offset = size;
2194 size += sizeof(void *);
2195 }
2196
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002197#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002198 if (flags & SLAB_STORE_USER)
2199 /*
2200 * Need to store information about allocs and frees after
2201 * the object.
2202 */
2203 size += 2 * sizeof(struct track);
2204
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002205 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002206 /*
2207 * Add some empty padding so that we can catch
2208 * overwrites from earlier objects rather than let
2209 * tracking information or the free pointer be
2210 * corrupted if an user writes before the start
2211 * of the object.
2212 */
2213 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002214#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002215
Christoph Lameter81819f02007-05-06 14:49:36 -07002216 /*
2217 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002218 * user specified and the dynamic determination of cache line size
2219 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002220 */
2221 align = calculate_alignment(flags, align, s->objsize);
2222
2223 /*
2224 * SLUB stores one object immediately after another beginning from
2225 * offset 0. In order to align the objects we have to simply size
2226 * each object to conform to the alignment.
2227 */
2228 size = ALIGN(size, align);
2229 s->size = size;
2230
Christoph Lameter71c7a062008-02-14 14:28:01 -08002231 if ((flags & __KMALLOC_CACHE) &&
2232 PAGE_SIZE / size < slub_min_objects) {
2233 /*
2234 * Kmalloc cache that would not have enough objects in
2235 * an order 0 page. Kmalloc slabs can fallback to
2236 * page allocator order 0 allocs so take a reasonably large
2237 * order that will allows us a good number of objects.
2238 */
2239 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2240 s->flags |= __PAGE_ALLOC_FALLBACK;
2241 s->allocflags |= __GFP_NOWARN;
2242 } else
2243 s->order = calculate_order(size);
2244
Christoph Lameter81819f02007-05-06 14:49:36 -07002245 if (s->order < 0)
2246 return 0;
2247
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002248 s->allocflags = 0;
2249 if (s->order)
2250 s->allocflags |= __GFP_COMP;
2251
2252 if (s->flags & SLAB_CACHE_DMA)
2253 s->allocflags |= SLUB_DMA;
2254
2255 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2256 s->allocflags |= __GFP_RECLAIMABLE;
2257
Christoph Lameter81819f02007-05-06 14:49:36 -07002258 /*
2259 * Determine the number of objects per slab
2260 */
2261 s->objects = (PAGE_SIZE << s->order) / size;
2262
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07002263 return !!s->objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002264
2265}
2266
Christoph Lameter81819f02007-05-06 14:49:36 -07002267static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2268 const char *name, size_t size,
2269 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002270 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002271{
2272 memset(s, 0, kmem_size);
2273 s->name = name;
2274 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002275 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002276 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002277 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002278
2279 if (!calculate_sizes(s))
2280 goto error;
2281
2282 s->refcount = 1;
2283#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08002284 s->remote_node_defrag_ratio = 100;
Christoph Lameter81819f02007-05-06 14:49:36 -07002285#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002286 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2287 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002288
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002289 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002290 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002291 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002292error:
2293 if (flags & SLAB_PANIC)
2294 panic("Cannot create slab %s size=%lu realsize=%u "
2295 "order=%u offset=%u flags=%lx\n",
2296 s->name, (unsigned long)size, s->size, s->order,
2297 s->offset, flags);
2298 return 0;
2299}
Christoph Lameter81819f02007-05-06 14:49:36 -07002300
2301/*
2302 * Check if a given pointer is valid
2303 */
2304int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2305{
Pekka Enberg06428782008-01-07 23:20:27 -08002306 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002307
2308 page = get_object_page(object);
2309
2310 if (!page || s != page->slab)
2311 /* No slab or wrong slab */
2312 return 0;
2313
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002314 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002315 return 0;
2316
2317 /*
2318 * We could also check if the object is on the slabs freelist.
2319 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002320 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002321 * to a certain slab.
2322 */
2323 return 1;
2324}
2325EXPORT_SYMBOL(kmem_ptr_validate);
2326
2327/*
2328 * Determine the size of a slab object
2329 */
2330unsigned int kmem_cache_size(struct kmem_cache *s)
2331{
2332 return s->objsize;
2333}
2334EXPORT_SYMBOL(kmem_cache_size);
2335
2336const char *kmem_cache_name(struct kmem_cache *s)
2337{
2338 return s->name;
2339}
2340EXPORT_SYMBOL(kmem_cache_name);
2341
2342/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002343 * Attempt to free all slabs on a node. Return the number of slabs we
2344 * were unable to free.
Christoph Lameter81819f02007-05-06 14:49:36 -07002345 */
2346static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2347 struct list_head *list)
2348{
2349 int slabs_inuse = 0;
2350 unsigned long flags;
2351 struct page *page, *h;
2352
2353 spin_lock_irqsave(&n->list_lock, flags);
2354 list_for_each_entry_safe(page, h, list, lru)
2355 if (!page->inuse) {
2356 list_del(&page->lru);
2357 discard_slab(s, page);
2358 } else
2359 slabs_inuse++;
2360 spin_unlock_irqrestore(&n->list_lock, flags);
2361 return slabs_inuse;
2362}
2363
2364/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002365 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002366 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002367static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002368{
2369 int node;
2370
2371 flush_all(s);
2372
2373 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002374 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002375 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002376 struct kmem_cache_node *n = get_node(s, node);
2377
Christoph Lameter2086d262007-05-06 14:49:46 -07002378 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07002379 if (atomic_long_read(&n->nr_slabs))
2380 return 1;
2381 }
2382 free_kmem_cache_nodes(s);
2383 return 0;
2384}
2385
2386/*
2387 * Close a cache and release the kmem_cache structure
2388 * (must be used for caches created using kmem_cache_create)
2389 */
2390void kmem_cache_destroy(struct kmem_cache *s)
2391{
2392 down_write(&slub_lock);
2393 s->refcount--;
2394 if (!s->refcount) {
2395 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002396 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002397 if (kmem_cache_close(s))
2398 WARN_ON(1);
2399 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002400 } else
2401 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002402}
2403EXPORT_SYMBOL(kmem_cache_destroy);
2404
2405/********************************************************************
2406 * Kmalloc subsystem
2407 *******************************************************************/
2408
Christoph Lameter331dc552008-02-14 14:28:09 -08002409struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002410EXPORT_SYMBOL(kmalloc_caches);
2411
2412#ifdef CONFIG_ZONE_DMA
Christoph Lameter331dc552008-02-14 14:28:09 -08002413static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter81819f02007-05-06 14:49:36 -07002414#endif
2415
2416static int __init setup_slub_min_order(char *str)
2417{
Pekka Enberg06428782008-01-07 23:20:27 -08002418 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002419
2420 return 1;
2421}
2422
2423__setup("slub_min_order=", setup_slub_min_order);
2424
2425static int __init setup_slub_max_order(char *str)
2426{
Pekka Enberg06428782008-01-07 23:20:27 -08002427 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002428
2429 return 1;
2430}
2431
2432__setup("slub_max_order=", setup_slub_max_order);
2433
2434static int __init setup_slub_min_objects(char *str)
2435{
Pekka Enberg06428782008-01-07 23:20:27 -08002436 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002437
2438 return 1;
2439}
2440
2441__setup("slub_min_objects=", setup_slub_min_objects);
2442
2443static int __init setup_slub_nomerge(char *str)
2444{
2445 slub_nomerge = 1;
2446 return 1;
2447}
2448
2449__setup("slub_nomerge", setup_slub_nomerge);
2450
Christoph Lameter81819f02007-05-06 14:49:36 -07002451static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2452 const char *name, int size, gfp_t gfp_flags)
2453{
2454 unsigned int flags = 0;
2455
2456 if (gfp_flags & SLUB_DMA)
2457 flags = SLAB_CACHE_DMA;
2458
2459 down_write(&slub_lock);
2460 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter71c7a062008-02-14 14:28:01 -08002461 flags | __KMALLOC_CACHE, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002462 goto panic;
2463
2464 list_add(&s->list, &slab_caches);
2465 up_write(&slub_lock);
2466 if (sysfs_slab_add(s))
2467 goto panic;
2468 return s;
2469
2470panic:
2471 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2472}
2473
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002474#ifdef CONFIG_ZONE_DMA
Christoph Lameter1ceef402007-08-07 15:11:48 -07002475
2476static void sysfs_add_func(struct work_struct *w)
2477{
2478 struct kmem_cache *s;
2479
2480 down_write(&slub_lock);
2481 list_for_each_entry(s, &slab_caches, list) {
2482 if (s->flags & __SYSFS_ADD_DEFERRED) {
2483 s->flags &= ~__SYSFS_ADD_DEFERRED;
2484 sysfs_slab_add(s);
2485 }
2486 }
2487 up_write(&slub_lock);
2488}
2489
2490static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2491
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002492static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2493{
2494 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002495 char *text;
2496 size_t realsize;
2497
2498 s = kmalloc_caches_dma[index];
2499 if (s)
2500 return s;
2501
2502 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002503 if (flags & __GFP_WAIT)
2504 down_write(&slub_lock);
2505 else {
2506 if (!down_write_trylock(&slub_lock))
2507 goto out;
2508 }
2509
2510 if (kmalloc_caches_dma[index])
2511 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002512
Christoph Lameter7b55f622007-07-17 04:03:27 -07002513 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002514 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2515 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002516 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2517
2518 if (!s || !text || !kmem_cache_open(s, flags, text,
2519 realsize, ARCH_KMALLOC_MINALIGN,
2520 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2521 kfree(s);
2522 kfree(text);
2523 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002524 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002525
2526 list_add(&s->list, &slab_caches);
2527 kmalloc_caches_dma[index] = s;
2528
2529 schedule_work(&sysfs_add_work);
2530
2531unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002532 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002533out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002534 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002535}
2536#endif
2537
Christoph Lameterf1b26332007-07-17 04:03:26 -07002538/*
2539 * Conversion table for small slabs sizes / 8 to the index in the
2540 * kmalloc array. This is necessary for slabs < 192 since we have non power
2541 * of two cache sizes there. The size of larger slabs can be determined using
2542 * fls.
2543 */
2544static s8 size_index[24] = {
2545 3, /* 8 */
2546 4, /* 16 */
2547 5, /* 24 */
2548 5, /* 32 */
2549 6, /* 40 */
2550 6, /* 48 */
2551 6, /* 56 */
2552 6, /* 64 */
2553 1, /* 72 */
2554 1, /* 80 */
2555 1, /* 88 */
2556 1, /* 96 */
2557 7, /* 104 */
2558 7, /* 112 */
2559 7, /* 120 */
2560 7, /* 128 */
2561 2, /* 136 */
2562 2, /* 144 */
2563 2, /* 152 */
2564 2, /* 160 */
2565 2, /* 168 */
2566 2, /* 176 */
2567 2, /* 184 */
2568 2 /* 192 */
2569};
2570
Christoph Lameter81819f02007-05-06 14:49:36 -07002571static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2572{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002573 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002574
Christoph Lameterf1b26332007-07-17 04:03:26 -07002575 if (size <= 192) {
2576 if (!size)
2577 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002578
Christoph Lameterf1b26332007-07-17 04:03:26 -07002579 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002580 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002581 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002582
2583#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002584 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002585 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002586
Christoph Lameter81819f02007-05-06 14:49:36 -07002587#endif
2588 return &kmalloc_caches[index];
2589}
2590
2591void *__kmalloc(size_t size, gfp_t flags)
2592{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002593 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002594
Christoph Lameter331dc552008-02-14 14:28:09 -08002595 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002596 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002597
2598 s = get_slab(size, flags);
2599
2600 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002601 return s;
2602
Christoph Lameterce15fea2007-07-17 04:03:28 -07002603 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002604}
2605EXPORT_SYMBOL(__kmalloc);
2606
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002607static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2608{
2609 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2610 get_order(size));
2611
2612 if (page)
2613 return page_address(page);
2614 else
2615 return NULL;
2616}
2617
Christoph Lameter81819f02007-05-06 14:49:36 -07002618#ifdef CONFIG_NUMA
2619void *__kmalloc_node(size_t size, gfp_t flags, int node)
2620{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002621 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002622
Christoph Lameter331dc552008-02-14 14:28:09 -08002623 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002624 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002625
2626 s = get_slab(size, flags);
2627
2628 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002629 return s;
2630
Christoph Lameterce15fea2007-07-17 04:03:28 -07002631 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002632}
2633EXPORT_SYMBOL(__kmalloc_node);
2634#endif
2635
2636size_t ksize(const void *object)
2637{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002638 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002639 struct kmem_cache *s;
2640
Christoph Lameteref8b4522007-10-16 01:24:46 -07002641 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002642 return 0;
2643
Vegard Nossum294a80a2007-12-04 23:45:30 -08002644 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002645
2646 if (unlikely(!PageSlab(page)))
2647 return PAGE_SIZE << compound_order(page);
2648
Christoph Lameter81819f02007-05-06 14:49:36 -07002649 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002650
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002651#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002652 /*
2653 * Debugging requires use of the padding between object
2654 * and whatever may come after it.
2655 */
2656 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2657 return s->objsize;
2658
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002659#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002660 /*
2661 * If we have the need to store the freelist pointer
2662 * back there or track user information then we can
2663 * only use the space before that information.
2664 */
2665 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2666 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002667 /*
2668 * Else we can use all the padding etc for the allocation
2669 */
2670 return s->size;
2671}
2672EXPORT_SYMBOL(ksize);
2673
2674void kfree(const void *x)
2675{
Christoph Lameter81819f02007-05-06 14:49:36 -07002676 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002677 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002678
Satyam Sharma2408c552007-10-16 01:24:44 -07002679 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002680 return;
2681
Christoph Lameterb49af682007-05-06 14:49:41 -07002682 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002683 if (unlikely(!PageSlab(page))) {
2684 put_page(page);
2685 return;
2686 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002687 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002688}
2689EXPORT_SYMBOL(kfree);
2690
Christoph Lameter00460dd2008-04-01 12:07:41 -07002691#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SLABINFO)
Christoph Lameterf61396a2008-01-07 23:20:26 -08002692static unsigned long count_partial(struct kmem_cache_node *n)
2693{
2694 unsigned long flags;
2695 unsigned long x = 0;
2696 struct page *page;
2697
2698 spin_lock_irqsave(&n->list_lock, flags);
2699 list_for_each_entry(page, &n->partial, lru)
2700 x += page->inuse;
2701 spin_unlock_irqrestore(&n->list_lock, flags);
2702 return x;
2703}
Christoph Lameter53625b42008-03-19 13:42:07 -07002704#endif
Christoph Lameterf61396a2008-01-07 23:20:26 -08002705
Christoph Lameter2086d262007-05-06 14:49:46 -07002706/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002707 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2708 * the remaining slabs by the number of items in use. The slabs with the
2709 * most items in use come first. New allocations will then fill those up
2710 * and thus they can be removed from the partial lists.
2711 *
2712 * The slabs with the least items are placed last. This results in them
2713 * being allocated from last increasing the chance that the last objects
2714 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002715 */
2716int kmem_cache_shrink(struct kmem_cache *s)
2717{
2718 int node;
2719 int i;
2720 struct kmem_cache_node *n;
2721 struct page *page;
2722 struct page *t;
2723 struct list_head *slabs_by_inuse =
2724 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2725 unsigned long flags;
2726
2727 if (!slabs_by_inuse)
2728 return -ENOMEM;
2729
2730 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002731 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002732 n = get_node(s, node);
2733
2734 if (!n->nr_partial)
2735 continue;
2736
2737 for (i = 0; i < s->objects; i++)
2738 INIT_LIST_HEAD(slabs_by_inuse + i);
2739
2740 spin_lock_irqsave(&n->list_lock, flags);
2741
2742 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002743 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002744 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002745 * Note that concurrent frees may occur while we hold the
2746 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002747 */
2748 list_for_each_entry_safe(page, t, &n->partial, lru) {
2749 if (!page->inuse && slab_trylock(page)) {
2750 /*
2751 * Must hold slab lock here because slab_free
2752 * may have freed the last object and be
2753 * waiting to release the slab.
2754 */
2755 list_del(&page->lru);
2756 n->nr_partial--;
2757 slab_unlock(page);
2758 discard_slab(s, page);
2759 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002760 list_move(&page->lru,
2761 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002762 }
2763 }
2764
Christoph Lameter2086d262007-05-06 14:49:46 -07002765 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002766 * Rebuild the partial list with the slabs filled up most
2767 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002768 */
2769 for (i = s->objects - 1; i >= 0; i--)
2770 list_splice(slabs_by_inuse + i, n->partial.prev);
2771
Christoph Lameter2086d262007-05-06 14:49:46 -07002772 spin_unlock_irqrestore(&n->list_lock, flags);
2773 }
2774
2775 kfree(slabs_by_inuse);
2776 return 0;
2777}
2778EXPORT_SYMBOL(kmem_cache_shrink);
2779
Yasunori Gotob9049e22007-10-21 16:41:37 -07002780#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2781static int slab_mem_going_offline_callback(void *arg)
2782{
2783 struct kmem_cache *s;
2784
2785 down_read(&slub_lock);
2786 list_for_each_entry(s, &slab_caches, list)
2787 kmem_cache_shrink(s);
2788 up_read(&slub_lock);
2789
2790 return 0;
2791}
2792
2793static void slab_mem_offline_callback(void *arg)
2794{
2795 struct kmem_cache_node *n;
2796 struct kmem_cache *s;
2797 struct memory_notify *marg = arg;
2798 int offline_node;
2799
2800 offline_node = marg->status_change_nid;
2801
2802 /*
2803 * If the node still has available memory. we need kmem_cache_node
2804 * for it yet.
2805 */
2806 if (offline_node < 0)
2807 return;
2808
2809 down_read(&slub_lock);
2810 list_for_each_entry(s, &slab_caches, list) {
2811 n = get_node(s, offline_node);
2812 if (n) {
2813 /*
2814 * if n->nr_slabs > 0, slabs still exist on the node
2815 * that is going down. We were unable to free them,
2816 * and offline_pages() function shoudn't call this
2817 * callback. So, we must fail.
2818 */
Al Viro27bb6282007-10-29 04:42:55 +00002819 BUG_ON(atomic_long_read(&n->nr_slabs));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002820
2821 s->node[offline_node] = NULL;
2822 kmem_cache_free(kmalloc_caches, n);
2823 }
2824 }
2825 up_read(&slub_lock);
2826}
2827
2828static int slab_mem_going_online_callback(void *arg)
2829{
2830 struct kmem_cache_node *n;
2831 struct kmem_cache *s;
2832 struct memory_notify *marg = arg;
2833 int nid = marg->status_change_nid;
2834 int ret = 0;
2835
2836 /*
2837 * If the node's memory is already available, then kmem_cache_node is
2838 * already created. Nothing to do.
2839 */
2840 if (nid < 0)
2841 return 0;
2842
2843 /*
2844 * We are bringing a node online. No memory is availabe yet. We must
2845 * allocate a kmem_cache_node structure in order to bring the node
2846 * online.
2847 */
2848 down_read(&slub_lock);
2849 list_for_each_entry(s, &slab_caches, list) {
2850 /*
2851 * XXX: kmem_cache_alloc_node will fallback to other nodes
2852 * since memory is not yet available from the node that
2853 * is brought up.
2854 */
2855 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2856 if (!n) {
2857 ret = -ENOMEM;
2858 goto out;
2859 }
2860 init_kmem_cache_node(n);
2861 s->node[nid] = n;
2862 }
2863out:
2864 up_read(&slub_lock);
2865 return ret;
2866}
2867
2868static int slab_memory_callback(struct notifier_block *self,
2869 unsigned long action, void *arg)
2870{
2871 int ret = 0;
2872
2873 switch (action) {
2874 case MEM_GOING_ONLINE:
2875 ret = slab_mem_going_online_callback(arg);
2876 break;
2877 case MEM_GOING_OFFLINE:
2878 ret = slab_mem_going_offline_callback(arg);
2879 break;
2880 case MEM_OFFLINE:
2881 case MEM_CANCEL_ONLINE:
2882 slab_mem_offline_callback(arg);
2883 break;
2884 case MEM_ONLINE:
2885 case MEM_CANCEL_OFFLINE:
2886 break;
2887 }
2888
2889 ret = notifier_from_errno(ret);
2890 return ret;
2891}
2892
2893#endif /* CONFIG_MEMORY_HOTPLUG */
2894
Christoph Lameter81819f02007-05-06 14:49:36 -07002895/********************************************************************
2896 * Basic setup of slabs
2897 *******************************************************************/
2898
2899void __init kmem_cache_init(void)
2900{
2901 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002902 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002903
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002904 init_alloc_cpu();
2905
Christoph Lameter81819f02007-05-06 14:49:36 -07002906#ifdef CONFIG_NUMA
2907 /*
2908 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002909 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002910 * kmem_cache_open for slab_state == DOWN.
2911 */
2912 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2913 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002914 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002915 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002916
2917 hotplug_memory_notifier(slab_memory_callback, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002918#endif
2919
2920 /* Able to allocate the per node structures */
2921 slab_state = PARTIAL;
2922
2923 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002924 if (KMALLOC_MIN_SIZE <= 64) {
2925 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002926 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002927 caches++;
2928 }
2929 if (KMALLOC_MIN_SIZE <= 128) {
2930 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002931 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002932 caches++;
2933 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002934
Christoph Lameter331dc552008-02-14 14:28:09 -08002935 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002936 create_kmalloc_cache(&kmalloc_caches[i],
2937 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002938 caches++;
2939 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002940
Christoph Lameterf1b26332007-07-17 04:03:26 -07002941
2942 /*
2943 * Patch up the size_index table if we have strange large alignment
2944 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002945 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002946 *
2947 * Largest permitted alignment is 256 bytes due to the way we
2948 * handle the index determination for the smaller caches.
2949 *
2950 * Make sure that nothing crazy happens if someone starts tinkering
2951 * around with ARCH_KMALLOC_MINALIGN
2952 */
2953 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2954 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2955
Christoph Lameter12ad6842007-07-17 04:03:28 -07002956 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07002957 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2958
Christoph Lameter81819f02007-05-06 14:49:36 -07002959 slab_state = UP;
2960
2961 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08002962 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07002963 kmalloc_caches[i]. name =
2964 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2965
2966#ifdef CONFIG_SMP
2967 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002968 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2969 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2970#else
2971 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07002972#endif
2973
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002974 printk(KERN_INFO
2975 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07002976 " CPUs=%d, Nodes=%d\n",
2977 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002978 slub_min_order, slub_max_order, slub_min_objects,
2979 nr_cpu_ids, nr_node_ids);
2980}
2981
2982/*
2983 * Find a mergeable slab cache
2984 */
2985static int slab_unmergeable(struct kmem_cache *s)
2986{
2987 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2988 return 1;
2989
Christoph Lameter331dc552008-02-14 14:28:09 -08002990 if ((s->flags & __PAGE_ALLOC_FALLBACK))
Christoph Lameter71c7a062008-02-14 14:28:01 -08002991 return 1;
2992
Christoph Lameterc59def9f2007-05-16 22:10:50 -07002993 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002994 return 1;
2995
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002996 /*
2997 * We may have set a slab to be unmergeable during bootstrap.
2998 */
2999 if (s->refcount < 0)
3000 return 1;
3001
Christoph Lameter81819f02007-05-06 14:49:36 -07003002 return 0;
3003}
3004
3005static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003006 size_t align, unsigned long flags, const char *name,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003007 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003008{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003009 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003010
3011 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3012 return NULL;
3013
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003014 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003015 return NULL;
3016
3017 size = ALIGN(size, sizeof(void *));
3018 align = calculate_alignment(flags, align, size);
3019 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003020 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003021
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003022 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003023 if (slab_unmergeable(s))
3024 continue;
3025
3026 if (size > s->size)
3027 continue;
3028
Christoph Lameterba0268a2007-09-11 15:24:11 -07003029 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003030 continue;
3031 /*
3032 * Check if alignment is compatible.
3033 * Courtesy of Adrian Drzewiecki
3034 */
Pekka Enberg06428782008-01-07 23:20:27 -08003035 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003036 continue;
3037
3038 if (s->size - size >= sizeof(void *))
3039 continue;
3040
3041 return s;
3042 }
3043 return NULL;
3044}
3045
3046struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3047 size_t align, unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003048 void (*ctor)(struct kmem_cache *, void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003049{
3050 struct kmem_cache *s;
3051
3052 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003053 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003054 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003055 int cpu;
3056
Christoph Lameter81819f02007-05-06 14:49:36 -07003057 s->refcount++;
3058 /*
3059 * Adjust the object sizes so that we clear
3060 * the complete object on kzalloc.
3061 */
3062 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003063
3064 /*
3065 * And then we need to update the object size in the
3066 * per cpu structures
3067 */
3068 for_each_online_cpu(cpu)
3069 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003070
Christoph Lameter81819f02007-05-06 14:49:36 -07003071 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003072 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003073
Christoph Lameter81819f02007-05-06 14:49:36 -07003074 if (sysfs_slab_alias(s, name))
3075 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003076 return s;
3077 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003078
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003079 s = kmalloc(kmem_size, GFP_KERNEL);
3080 if (s) {
3081 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def9f2007-05-16 22:10:50 -07003082 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003083 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003084 up_write(&slub_lock);
3085 if (sysfs_slab_add(s))
3086 goto err;
3087 return s;
3088 }
3089 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003090 }
3091 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003092
3093err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003094 if (flags & SLAB_PANIC)
3095 panic("Cannot create slabcache %s\n", name);
3096 else
3097 s = NULL;
3098 return s;
3099}
3100EXPORT_SYMBOL(kmem_cache_create);
3101
Christoph Lameter81819f02007-05-06 14:49:36 -07003102#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003103/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003104 * Use the cpu notifier to insure that the cpu slabs are flushed when
3105 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003106 */
3107static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3108 unsigned long action, void *hcpu)
3109{
3110 long cpu = (long)hcpu;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003111 struct kmem_cache *s;
3112 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003113
3114 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003115 case CPU_UP_PREPARE:
3116 case CPU_UP_PREPARE_FROZEN:
3117 init_alloc_cpu_cpu(cpu);
3118 down_read(&slub_lock);
3119 list_for_each_entry(s, &slab_caches, list)
3120 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3121 GFP_KERNEL);
3122 up_read(&slub_lock);
3123 break;
3124
Christoph Lameter81819f02007-05-06 14:49:36 -07003125 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003126 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003127 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003128 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003129 down_read(&slub_lock);
3130 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003131 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3132
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003133 local_irq_save(flags);
3134 __flush_cpu_slab(s, cpu);
3135 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003136 free_kmem_cache_cpu(c, cpu);
3137 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07003138 }
3139 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003140 break;
3141 default:
3142 break;
3143 }
3144 return NOTIFY_OK;
3145}
3146
Pekka Enberg06428782008-01-07 23:20:27 -08003147static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003148 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003149};
Christoph Lameter81819f02007-05-06 14:49:36 -07003150
3151#endif
3152
Christoph Lameter81819f02007-05-06 14:49:36 -07003153void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3154{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003155 struct kmem_cache *s;
3156
Christoph Lameter331dc552008-02-14 14:28:09 -08003157 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003158 return kmalloc_large(size, gfpflags);
3159
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003160 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003161
Satyam Sharma2408c552007-10-16 01:24:44 -07003162 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003163 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003164
Christoph Lameterce15fea2007-07-17 04:03:28 -07003165 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003166}
3167
3168void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3169 int node, void *caller)
3170{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003171 struct kmem_cache *s;
3172
Christoph Lameter331dc552008-02-14 14:28:09 -08003173 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003174 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003175
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003176 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003177
Satyam Sharma2408c552007-10-16 01:24:44 -07003178 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003179 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003180
Christoph Lameterce15fea2007-07-17 04:03:28 -07003181 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003182}
3183
Christoph Lameter41ecc552007-05-09 02:32:44 -07003184#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
Christoph Lameter434e2452007-07-17 04:03:30 -07003185static int validate_slab(struct kmem_cache *s, struct page *page,
3186 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003187{
3188 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003189 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003190
3191 if (!check_slab(s, page) ||
3192 !on_freelist(s, page, NULL))
3193 return 0;
3194
3195 /* Now we know that a valid freelist exists */
3196 bitmap_zero(map, s->objects);
3197
Christoph Lameter7656c722007-05-09 02:32:40 -07003198 for_each_free_object(p, s, page->freelist) {
3199 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003200 if (!check_object(s, page, p, 0))
3201 return 0;
3202 }
3203
Christoph Lameter7656c722007-05-09 02:32:40 -07003204 for_each_object(p, s, addr)
3205 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003206 if (!check_object(s, page, p, 1))
3207 return 0;
3208 return 1;
3209}
3210
Christoph Lameter434e2452007-07-17 04:03:30 -07003211static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3212 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003213{
3214 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003215 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003216 slab_unlock(page);
3217 } else
3218 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3219 s->name, page);
3220
3221 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003222 if (!SlabDebug(page))
3223 printk(KERN_ERR "SLUB %s: SlabDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003224 "on slab 0x%p\n", s->name, page);
3225 } else {
Christoph Lameter35e5d7e2007-05-09 02:32:42 -07003226 if (SlabDebug(page))
3227 printk(KERN_ERR "SLUB %s: SlabDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003228 "slab 0x%p\n", s->name, page);
3229 }
3230}
3231
Christoph Lameter434e2452007-07-17 04:03:30 -07003232static int validate_slab_node(struct kmem_cache *s,
3233 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003234{
3235 unsigned long count = 0;
3236 struct page *page;
3237 unsigned long flags;
3238
3239 spin_lock_irqsave(&n->list_lock, flags);
3240
3241 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003242 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003243 count++;
3244 }
3245 if (count != n->nr_partial)
3246 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3247 "counter=%ld\n", s->name, count, n->nr_partial);
3248
3249 if (!(s->flags & SLAB_STORE_USER))
3250 goto out;
3251
3252 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003253 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003254 count++;
3255 }
3256 if (count != atomic_long_read(&n->nr_slabs))
3257 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3258 "counter=%ld\n", s->name, count,
3259 atomic_long_read(&n->nr_slabs));
3260
3261out:
3262 spin_unlock_irqrestore(&n->list_lock, flags);
3263 return count;
3264}
3265
Christoph Lameter434e2452007-07-17 04:03:30 -07003266static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003267{
3268 int node;
3269 unsigned long count = 0;
Christoph Lameter434e2452007-07-17 04:03:30 -07003270 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3271 sizeof(unsigned long), GFP_KERNEL);
3272
3273 if (!map)
3274 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003275
3276 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003277 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003278 struct kmem_cache_node *n = get_node(s, node);
3279
Christoph Lameter434e2452007-07-17 04:03:30 -07003280 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003281 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003282 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003283 return count;
3284}
3285
Christoph Lameterb3459702007-05-09 02:32:41 -07003286#ifdef SLUB_RESILIENCY_TEST
3287static void resiliency_test(void)
3288{
3289 u8 *p;
3290
3291 printk(KERN_ERR "SLUB resiliency testing\n");
3292 printk(KERN_ERR "-----------------------\n");
3293 printk(KERN_ERR "A. Corruption after allocation\n");
3294
3295 p = kzalloc(16, GFP_KERNEL);
3296 p[16] = 0x12;
3297 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3298 " 0x12->0x%p\n\n", p + 16);
3299
3300 validate_slab_cache(kmalloc_caches + 4);
3301
3302 /* Hmmm... The next two are dangerous */
3303 p = kzalloc(32, GFP_KERNEL);
3304 p[32 + sizeof(void *)] = 0x34;
3305 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003306 " 0x34 -> -0x%p\n", p);
3307 printk(KERN_ERR
3308 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003309
3310 validate_slab_cache(kmalloc_caches + 5);
3311 p = kzalloc(64, GFP_KERNEL);
3312 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3313 *p = 0x56;
3314 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3315 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003316 printk(KERN_ERR
3317 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003318 validate_slab_cache(kmalloc_caches + 6);
3319
3320 printk(KERN_ERR "\nB. Corruption after free\n");
3321 p = kzalloc(128, GFP_KERNEL);
3322 kfree(p);
3323 *p = 0x78;
3324 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3325 validate_slab_cache(kmalloc_caches + 7);
3326
3327 p = kzalloc(256, GFP_KERNEL);
3328 kfree(p);
3329 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003330 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3331 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003332 validate_slab_cache(kmalloc_caches + 8);
3333
3334 p = kzalloc(512, GFP_KERNEL);
3335 kfree(p);
3336 p[512] = 0xab;
3337 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3338 validate_slab_cache(kmalloc_caches + 9);
3339}
3340#else
3341static void resiliency_test(void) {};
3342#endif
3343
Christoph Lameter88a420e2007-05-06 14:49:45 -07003344/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003345 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003346 * and freed.
3347 */
3348
3349struct location {
3350 unsigned long count;
3351 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003352 long long sum_time;
3353 long min_time;
3354 long max_time;
3355 long min_pid;
3356 long max_pid;
3357 cpumask_t cpus;
3358 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003359};
3360
3361struct loc_track {
3362 unsigned long max;
3363 unsigned long count;
3364 struct location *loc;
3365};
3366
3367static void free_loc_track(struct loc_track *t)
3368{
3369 if (t->max)
3370 free_pages((unsigned long)t->loc,
3371 get_order(sizeof(struct location) * t->max));
3372}
3373
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003374static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003375{
3376 struct location *l;
3377 int order;
3378
Christoph Lameter88a420e2007-05-06 14:49:45 -07003379 order = get_order(sizeof(struct location) * max);
3380
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003381 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003382 if (!l)
3383 return 0;
3384
3385 if (t->count) {
3386 memcpy(l, t->loc, sizeof(struct location) * t->count);
3387 free_loc_track(t);
3388 }
3389 t->max = max;
3390 t->loc = l;
3391 return 1;
3392}
3393
3394static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003395 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003396{
3397 long start, end, pos;
3398 struct location *l;
3399 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003400 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003401
3402 start = -1;
3403 end = t->count;
3404
3405 for ( ; ; ) {
3406 pos = start + (end - start + 1) / 2;
3407
3408 /*
3409 * There is nothing at "end". If we end up there
3410 * we need to add something to before end.
3411 */
3412 if (pos == end)
3413 break;
3414
3415 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003416 if (track->addr == caddr) {
3417
3418 l = &t->loc[pos];
3419 l->count++;
3420 if (track->when) {
3421 l->sum_time += age;
3422 if (age < l->min_time)
3423 l->min_time = age;
3424 if (age > l->max_time)
3425 l->max_time = age;
3426
3427 if (track->pid < l->min_pid)
3428 l->min_pid = track->pid;
3429 if (track->pid > l->max_pid)
3430 l->max_pid = track->pid;
3431
3432 cpu_set(track->cpu, l->cpus);
3433 }
3434 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003435 return 1;
3436 }
3437
Christoph Lameter45edfa52007-05-09 02:32:45 -07003438 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003439 end = pos;
3440 else
3441 start = pos;
3442 }
3443
3444 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003445 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003446 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003447 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003448 return 0;
3449
3450 l = t->loc + pos;
3451 if (pos < t->count)
3452 memmove(l + 1, l,
3453 (t->count - pos) * sizeof(struct location));
3454 t->count++;
3455 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003456 l->addr = track->addr;
3457 l->sum_time = age;
3458 l->min_time = age;
3459 l->max_time = age;
3460 l->min_pid = track->pid;
3461 l->max_pid = track->pid;
3462 cpus_clear(l->cpus);
3463 cpu_set(track->cpu, l->cpus);
3464 nodes_clear(l->nodes);
3465 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003466 return 1;
3467}
3468
3469static void process_slab(struct loc_track *t, struct kmem_cache *s,
3470 struct page *page, enum track_item alloc)
3471{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003472 void *addr = page_address(page);
Christoph Lameter7656c722007-05-09 02:32:40 -07003473 DECLARE_BITMAP(map, s->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003474 void *p;
3475
3476 bitmap_zero(map, s->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003477 for_each_free_object(p, s, page->freelist)
3478 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003479
Christoph Lameter7656c722007-05-09 02:32:40 -07003480 for_each_object(p, s, addr)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003481 if (!test_bit(slab_index(p, s, addr), map))
3482 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003483}
3484
3485static int list_locations(struct kmem_cache *s, char *buf,
3486 enum track_item alloc)
3487{
Harvey Harrisone374d482008-01-31 15:20:50 -08003488 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003489 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003490 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003491 int node;
3492
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003493 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003494 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003495 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003496
3497 /* Push back cpu slabs */
3498 flush_all(s);
3499
Christoph Lameterf64dc582007-10-16 01:25:33 -07003500 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003501 struct kmem_cache_node *n = get_node(s, node);
3502 unsigned long flags;
3503 struct page *page;
3504
Christoph Lameter9e869432007-08-22 14:01:56 -07003505 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003506 continue;
3507
3508 spin_lock_irqsave(&n->list_lock, flags);
3509 list_for_each_entry(page, &n->partial, lru)
3510 process_slab(&t, s, page, alloc);
3511 list_for_each_entry(page, &n->full, lru)
3512 process_slab(&t, s, page, alloc);
3513 spin_unlock_irqrestore(&n->list_lock, flags);
3514 }
3515
3516 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003517 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003518
Harvey Harrisone374d482008-01-31 15:20:50 -08003519 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003520 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003521 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003522
3523 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003524 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003525 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003526 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003527
3528 if (l->sum_time != l->min_time) {
3529 unsigned long remainder;
3530
Harvey Harrisone374d482008-01-31 15:20:50 -08003531 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003532 l->min_time,
3533 div_long_long_rem(l->sum_time, l->count, &remainder),
3534 l->max_time);
3535 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003536 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003537 l->min_time);
3538
3539 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003540 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003541 l->min_pid, l->max_pid);
3542 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003543 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003544 l->min_pid);
3545
Christoph Lameter84966342007-06-23 17:16:32 -07003546 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003547 len < PAGE_SIZE - 60) {
3548 len += sprintf(buf + len, " cpus=");
3549 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003550 l->cpus);
3551 }
3552
Christoph Lameter84966342007-06-23 17:16:32 -07003553 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003554 len < PAGE_SIZE - 60) {
3555 len += sprintf(buf + len, " nodes=");
3556 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003557 l->nodes);
3558 }
3559
Harvey Harrisone374d482008-01-31 15:20:50 -08003560 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003561 }
3562
3563 free_loc_track(&t);
3564 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003565 len += sprintf(buf, "No data\n");
3566 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003567}
3568
Christoph Lameter81819f02007-05-06 14:49:36 -07003569enum slab_stat_type {
3570 SL_FULL,
3571 SL_PARTIAL,
3572 SL_CPU,
3573 SL_OBJECTS
3574};
3575
3576#define SO_FULL (1 << SL_FULL)
3577#define SO_PARTIAL (1 << SL_PARTIAL)
3578#define SO_CPU (1 << SL_CPU)
3579#define SO_OBJECTS (1 << SL_OBJECTS)
3580
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003581static ssize_t show_slab_objects(struct kmem_cache *s,
3582 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003583{
3584 unsigned long total = 0;
3585 int cpu;
3586 int node;
3587 int x;
3588 unsigned long *nodes;
3589 unsigned long *per_cpu;
3590
3591 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003592 if (!nodes)
3593 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003594 per_cpu = nodes + nr_node_ids;
3595
3596 for_each_possible_cpu(cpu) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003597 struct page *page;
3598 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003599
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003600 if (!c)
3601 continue;
3602
3603 page = c->page;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003604 node = c->node;
3605 if (node < 0)
3606 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003607 if (page) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003608 if (flags & SO_CPU) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003609 if (flags & SO_OBJECTS)
3610 x = page->inuse;
3611 else
3612 x = 1;
3613 total += x;
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003614 nodes[node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003615 }
Christoph Lameteree3c72a2007-10-16 01:26:07 -07003616 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003617 }
3618 }
3619
Christoph Lameterf64dc582007-10-16 01:25:33 -07003620 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003621 struct kmem_cache_node *n = get_node(s, node);
3622
3623 if (flags & SO_PARTIAL) {
3624 if (flags & SO_OBJECTS)
3625 x = count_partial(n);
3626 else
3627 x = n->nr_partial;
3628 total += x;
3629 nodes[node] += x;
3630 }
3631
3632 if (flags & SO_FULL) {
Christoph Lameter9e869432007-08-22 14:01:56 -07003633 int full_slabs = atomic_long_read(&n->nr_slabs)
Christoph Lameter81819f02007-05-06 14:49:36 -07003634 - per_cpu[node]
3635 - n->nr_partial;
3636
3637 if (flags & SO_OBJECTS)
3638 x = full_slabs * s->objects;
3639 else
3640 x = full_slabs;
3641 total += x;
3642 nodes[node] += x;
3643 }
3644 }
3645
3646 x = sprintf(buf, "%lu", total);
3647#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003648 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003649 if (nodes[node])
3650 x += sprintf(buf + x, " N%d=%lu",
3651 node, nodes[node]);
3652#endif
3653 kfree(nodes);
3654 return x + sprintf(buf + x, "\n");
3655}
3656
3657static int any_slab_objects(struct kmem_cache *s)
3658{
3659 int node;
3660 int cpu;
3661
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003662 for_each_possible_cpu(cpu) {
3663 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07003664
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003665 if (c && c->page)
3666 return 1;
3667 }
3668
3669 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003670 struct kmem_cache_node *n = get_node(s, node);
3671
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003672 if (!n)
3673 continue;
3674
Christoph Lameter9e869432007-08-22 14:01:56 -07003675 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
Christoph Lameter81819f02007-05-06 14:49:36 -07003676 return 1;
3677 }
3678 return 0;
3679}
3680
3681#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3682#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3683
3684struct slab_attribute {
3685 struct attribute attr;
3686 ssize_t (*show)(struct kmem_cache *s, char *buf);
3687 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3688};
3689
3690#define SLAB_ATTR_RO(_name) \
3691 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3692
3693#define SLAB_ATTR(_name) \
3694 static struct slab_attribute _name##_attr = \
3695 __ATTR(_name, 0644, _name##_show, _name##_store)
3696
Christoph Lameter81819f02007-05-06 14:49:36 -07003697static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3698{
3699 return sprintf(buf, "%d\n", s->size);
3700}
3701SLAB_ATTR_RO(slab_size);
3702
3703static ssize_t align_show(struct kmem_cache *s, char *buf)
3704{
3705 return sprintf(buf, "%d\n", s->align);
3706}
3707SLAB_ATTR_RO(align);
3708
3709static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3710{
3711 return sprintf(buf, "%d\n", s->objsize);
3712}
3713SLAB_ATTR_RO(object_size);
3714
3715static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3716{
3717 return sprintf(buf, "%d\n", s->objects);
3718}
3719SLAB_ATTR_RO(objs_per_slab);
3720
3721static ssize_t order_show(struct kmem_cache *s, char *buf)
3722{
3723 return sprintf(buf, "%d\n", s->order);
3724}
3725SLAB_ATTR_RO(order);
3726
3727static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3728{
3729 if (s->ctor) {
3730 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3731
3732 return n + sprintf(buf + n, "\n");
3733 }
3734 return 0;
3735}
3736SLAB_ATTR_RO(ctor);
3737
Christoph Lameter81819f02007-05-06 14:49:36 -07003738static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3739{
3740 return sprintf(buf, "%d\n", s->refcount - 1);
3741}
3742SLAB_ATTR_RO(aliases);
3743
3744static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3745{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003746 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003747}
3748SLAB_ATTR_RO(slabs);
3749
3750static ssize_t partial_show(struct kmem_cache *s, char *buf)
3751{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003752 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003753}
3754SLAB_ATTR_RO(partial);
3755
3756static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3757{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003758 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003759}
3760SLAB_ATTR_RO(cpu_slabs);
3761
3762static ssize_t objects_show(struct kmem_cache *s, char *buf)
3763{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003764 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003765}
3766SLAB_ATTR_RO(objects);
3767
3768static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3769{
3770 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3771}
3772
3773static ssize_t sanity_checks_store(struct kmem_cache *s,
3774 const char *buf, size_t length)
3775{
3776 s->flags &= ~SLAB_DEBUG_FREE;
3777 if (buf[0] == '1')
3778 s->flags |= SLAB_DEBUG_FREE;
3779 return length;
3780}
3781SLAB_ATTR(sanity_checks);
3782
3783static ssize_t trace_show(struct kmem_cache *s, char *buf)
3784{
3785 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3786}
3787
3788static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3789 size_t length)
3790{
3791 s->flags &= ~SLAB_TRACE;
3792 if (buf[0] == '1')
3793 s->flags |= SLAB_TRACE;
3794 return length;
3795}
3796SLAB_ATTR(trace);
3797
3798static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3799{
3800 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3801}
3802
3803static ssize_t reclaim_account_store(struct kmem_cache *s,
3804 const char *buf, size_t length)
3805{
3806 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3807 if (buf[0] == '1')
3808 s->flags |= SLAB_RECLAIM_ACCOUNT;
3809 return length;
3810}
3811SLAB_ATTR(reclaim_account);
3812
3813static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3814{
Christoph Lameter5af60832007-05-06 14:49:56 -07003815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003816}
3817SLAB_ATTR_RO(hwcache_align);
3818
3819#ifdef CONFIG_ZONE_DMA
3820static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3821{
3822 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3823}
3824SLAB_ATTR_RO(cache_dma);
3825#endif
3826
3827static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3828{
3829 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3830}
3831SLAB_ATTR_RO(destroy_by_rcu);
3832
3833static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3834{
3835 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3836}
3837
3838static ssize_t red_zone_store(struct kmem_cache *s,
3839 const char *buf, size_t length)
3840{
3841 if (any_slab_objects(s))
3842 return -EBUSY;
3843
3844 s->flags &= ~SLAB_RED_ZONE;
3845 if (buf[0] == '1')
3846 s->flags |= SLAB_RED_ZONE;
3847 calculate_sizes(s);
3848 return length;
3849}
3850SLAB_ATTR(red_zone);
3851
3852static ssize_t poison_show(struct kmem_cache *s, char *buf)
3853{
3854 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3855}
3856
3857static ssize_t poison_store(struct kmem_cache *s,
3858 const char *buf, size_t length)
3859{
3860 if (any_slab_objects(s))
3861 return -EBUSY;
3862
3863 s->flags &= ~SLAB_POISON;
3864 if (buf[0] == '1')
3865 s->flags |= SLAB_POISON;
3866 calculate_sizes(s);
3867 return length;
3868}
3869SLAB_ATTR(poison);
3870
3871static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3872{
3873 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3874}
3875
3876static ssize_t store_user_store(struct kmem_cache *s,
3877 const char *buf, size_t length)
3878{
3879 if (any_slab_objects(s))
3880 return -EBUSY;
3881
3882 s->flags &= ~SLAB_STORE_USER;
3883 if (buf[0] == '1')
3884 s->flags |= SLAB_STORE_USER;
3885 calculate_sizes(s);
3886 return length;
3887}
3888SLAB_ATTR(store_user);
3889
Christoph Lameter53e15af2007-05-06 14:49:43 -07003890static ssize_t validate_show(struct kmem_cache *s, char *buf)
3891{
3892 return 0;
3893}
3894
3895static ssize_t validate_store(struct kmem_cache *s,
3896 const char *buf, size_t length)
3897{
Christoph Lameter434e2452007-07-17 04:03:30 -07003898 int ret = -EINVAL;
3899
3900 if (buf[0] == '1') {
3901 ret = validate_slab_cache(s);
3902 if (ret >= 0)
3903 ret = length;
3904 }
3905 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003906}
3907SLAB_ATTR(validate);
3908
Christoph Lameter2086d262007-05-06 14:49:46 -07003909static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3910{
3911 return 0;
3912}
3913
3914static ssize_t shrink_store(struct kmem_cache *s,
3915 const char *buf, size_t length)
3916{
3917 if (buf[0] == '1') {
3918 int rc = kmem_cache_shrink(s);
3919
3920 if (rc)
3921 return rc;
3922 } else
3923 return -EINVAL;
3924 return length;
3925}
3926SLAB_ATTR(shrink);
3927
Christoph Lameter88a420e2007-05-06 14:49:45 -07003928static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3929{
3930 if (!(s->flags & SLAB_STORE_USER))
3931 return -ENOSYS;
3932 return list_locations(s, buf, TRACK_ALLOC);
3933}
3934SLAB_ATTR_RO(alloc_calls);
3935
3936static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3937{
3938 if (!(s->flags & SLAB_STORE_USER))
3939 return -ENOSYS;
3940 return list_locations(s, buf, TRACK_FREE);
3941}
3942SLAB_ATTR_RO(free_calls);
3943
Christoph Lameter81819f02007-05-06 14:49:36 -07003944#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08003945static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07003946{
Christoph Lameter98246012008-01-07 23:20:26 -08003947 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07003948}
3949
Christoph Lameter98246012008-01-07 23:20:26 -08003950static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07003951 const char *buf, size_t length)
3952{
3953 int n = simple_strtoul(buf, NULL, 10);
3954
3955 if (n < 100)
Christoph Lameter98246012008-01-07 23:20:26 -08003956 s->remote_node_defrag_ratio = n * 10;
Christoph Lameter81819f02007-05-06 14:49:36 -07003957 return length;
3958}
Christoph Lameter98246012008-01-07 23:20:26 -08003959SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07003960#endif
3961
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003962#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08003963static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3964{
3965 unsigned long sum = 0;
3966 int cpu;
3967 int len;
3968 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3969
3970 if (!data)
3971 return -ENOMEM;
3972
3973 for_each_online_cpu(cpu) {
3974 unsigned x = get_cpu_slab(s, cpu)->stat[si];
3975
3976 data[cpu] = x;
3977 sum += x;
3978 }
3979
3980 len = sprintf(buf, "%lu", sum);
3981
3982 for_each_online_cpu(cpu) {
3983 if (data[cpu] && len < PAGE_SIZE - 20)
3984 len += sprintf(buf + len, " c%d=%u", cpu, data[cpu]);
3985 }
3986 kfree(data);
3987 return len + sprintf(buf + len, "\n");
3988}
3989
3990#define STAT_ATTR(si, text) \
3991static ssize_t text##_show(struct kmem_cache *s, char *buf) \
3992{ \
3993 return show_stat(s, buf, si); \
3994} \
3995SLAB_ATTR_RO(text); \
3996
3997STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
3998STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
3999STAT_ATTR(FREE_FASTPATH, free_fastpath);
4000STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4001STAT_ATTR(FREE_FROZEN, free_frozen);
4002STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4003STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4004STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4005STAT_ATTR(ALLOC_SLAB, alloc_slab);
4006STAT_ATTR(ALLOC_REFILL, alloc_refill);
4007STAT_ATTR(FREE_SLAB, free_slab);
4008STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4009STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4010STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4011STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4012STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4013STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4014
4015#endif
4016
Pekka Enberg06428782008-01-07 23:20:27 -08004017static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004018 &slab_size_attr.attr,
4019 &object_size_attr.attr,
4020 &objs_per_slab_attr.attr,
4021 &order_attr.attr,
4022 &objects_attr.attr,
4023 &slabs_attr.attr,
4024 &partial_attr.attr,
4025 &cpu_slabs_attr.attr,
4026 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004027 &aliases_attr.attr,
4028 &align_attr.attr,
4029 &sanity_checks_attr.attr,
4030 &trace_attr.attr,
4031 &hwcache_align_attr.attr,
4032 &reclaim_account_attr.attr,
4033 &destroy_by_rcu_attr.attr,
4034 &red_zone_attr.attr,
4035 &poison_attr.attr,
4036 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004037 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004038 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004039 &alloc_calls_attr.attr,
4040 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004041#ifdef CONFIG_ZONE_DMA
4042 &cache_dma_attr.attr,
4043#endif
4044#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004045 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004046#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004047#ifdef CONFIG_SLUB_STATS
4048 &alloc_fastpath_attr.attr,
4049 &alloc_slowpath_attr.attr,
4050 &free_fastpath_attr.attr,
4051 &free_slowpath_attr.attr,
4052 &free_frozen_attr.attr,
4053 &free_add_partial_attr.attr,
4054 &free_remove_partial_attr.attr,
4055 &alloc_from_partial_attr.attr,
4056 &alloc_slab_attr.attr,
4057 &alloc_refill_attr.attr,
4058 &free_slab_attr.attr,
4059 &cpuslab_flush_attr.attr,
4060 &deactivate_full_attr.attr,
4061 &deactivate_empty_attr.attr,
4062 &deactivate_to_head_attr.attr,
4063 &deactivate_to_tail_attr.attr,
4064 &deactivate_remote_frees_attr.attr,
4065#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004066 NULL
4067};
4068
4069static struct attribute_group slab_attr_group = {
4070 .attrs = slab_attrs,
4071};
4072
4073static ssize_t slab_attr_show(struct kobject *kobj,
4074 struct attribute *attr,
4075 char *buf)
4076{
4077 struct slab_attribute *attribute;
4078 struct kmem_cache *s;
4079 int err;
4080
4081 attribute = to_slab_attr(attr);
4082 s = to_slab(kobj);
4083
4084 if (!attribute->show)
4085 return -EIO;
4086
4087 err = attribute->show(s, buf);
4088
4089 return err;
4090}
4091
4092static ssize_t slab_attr_store(struct kobject *kobj,
4093 struct attribute *attr,
4094 const char *buf, size_t len)
4095{
4096 struct slab_attribute *attribute;
4097 struct kmem_cache *s;
4098 int err;
4099
4100 attribute = to_slab_attr(attr);
4101 s = to_slab(kobj);
4102
4103 if (!attribute->store)
4104 return -EIO;
4105
4106 err = attribute->store(s, buf, len);
4107
4108 return err;
4109}
4110
Christoph Lameter151c6022008-01-07 22:29:05 -08004111static void kmem_cache_release(struct kobject *kobj)
4112{
4113 struct kmem_cache *s = to_slab(kobj);
4114
4115 kfree(s);
4116}
4117
Christoph Lameter81819f02007-05-06 14:49:36 -07004118static struct sysfs_ops slab_sysfs_ops = {
4119 .show = slab_attr_show,
4120 .store = slab_attr_store,
4121};
4122
4123static struct kobj_type slab_ktype = {
4124 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004125 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004126};
4127
4128static int uevent_filter(struct kset *kset, struct kobject *kobj)
4129{
4130 struct kobj_type *ktype = get_ktype(kobj);
4131
4132 if (ktype == &slab_ktype)
4133 return 1;
4134 return 0;
4135}
4136
4137static struct kset_uevent_ops slab_uevent_ops = {
4138 .filter = uevent_filter,
4139};
4140
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004141static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004142
4143#define ID_STR_LENGTH 64
4144
4145/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004146 *
4147 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004148 */
4149static char *create_unique_id(struct kmem_cache *s)
4150{
4151 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4152 char *p = name;
4153
4154 BUG_ON(!name);
4155
4156 *p++ = ':';
4157 /*
4158 * First flags affecting slabcache operations. We will only
4159 * get here for aliasable slabs so we do not need to support
4160 * too many flags. The flags here must cover all flags that
4161 * are matched during merging to guarantee that the id is
4162 * unique.
4163 */
4164 if (s->flags & SLAB_CACHE_DMA)
4165 *p++ = 'd';
4166 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4167 *p++ = 'a';
4168 if (s->flags & SLAB_DEBUG_FREE)
4169 *p++ = 'F';
4170 if (p != name + 1)
4171 *p++ = '-';
4172 p += sprintf(p, "%07d", s->size);
4173 BUG_ON(p > name + ID_STR_LENGTH - 1);
4174 return name;
4175}
4176
4177static int sysfs_slab_add(struct kmem_cache *s)
4178{
4179 int err;
4180 const char *name;
4181 int unmergeable;
4182
4183 if (slab_state < SYSFS)
4184 /* Defer until later */
4185 return 0;
4186
4187 unmergeable = slab_unmergeable(s);
4188 if (unmergeable) {
4189 /*
4190 * Slabcache can never be merged so we can use the name proper.
4191 * This is typically the case for debug situations. In that
4192 * case we can catch duplicate names easily.
4193 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004194 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004195 name = s->name;
4196 } else {
4197 /*
4198 * Create a unique name for the slab as a target
4199 * for the symlinks.
4200 */
4201 name = create_unique_id(s);
4202 }
4203
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004204 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004205 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4206 if (err) {
4207 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004208 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004209 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004210
4211 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4212 if (err)
4213 return err;
4214 kobject_uevent(&s->kobj, KOBJ_ADD);
4215 if (!unmergeable) {
4216 /* Setup first alias */
4217 sysfs_slab_alias(s, s->name);
4218 kfree(name);
4219 }
4220 return 0;
4221}
4222
4223static void sysfs_slab_remove(struct kmem_cache *s)
4224{
4225 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4226 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004227 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004228}
4229
4230/*
4231 * Need to buffer aliases during bootup until sysfs becomes
4232 * available lest we loose that information.
4233 */
4234struct saved_alias {
4235 struct kmem_cache *s;
4236 const char *name;
4237 struct saved_alias *next;
4238};
4239
Adrian Bunk5af328a2007-07-17 04:03:27 -07004240static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004241
4242static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4243{
4244 struct saved_alias *al;
4245
4246 if (slab_state == SYSFS) {
4247 /*
4248 * If we have a leftover link then remove it.
4249 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004250 sysfs_remove_link(&slab_kset->kobj, name);
4251 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004252 }
4253
4254 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4255 if (!al)
4256 return -ENOMEM;
4257
4258 al->s = s;
4259 al->name = name;
4260 al->next = alias_list;
4261 alias_list = al;
4262 return 0;
4263}
4264
4265static int __init slab_sysfs_init(void)
4266{
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004267 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004268 int err;
4269
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004270 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004271 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004272 printk(KERN_ERR "Cannot register slab subsystem.\n");
4273 return -ENOSYS;
4274 }
4275
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004276 slab_state = SYSFS;
4277
Christoph Lameter5b95a4a2007-07-17 04:03:19 -07004278 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004279 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004280 if (err)
4281 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4282 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004283 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004284
4285 while (alias_list) {
4286 struct saved_alias *al = alias_list;
4287
4288 alias_list = alias_list->next;
4289 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004290 if (err)
4291 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4292 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004293 kfree(al);
4294 }
4295
4296 resiliency_test();
4297 return 0;
4298}
4299
4300__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004301#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004302
4303/*
4304 * The /proc/slabinfo ABI
4305 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004306#ifdef CONFIG_SLABINFO
4307
4308ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4309 size_t count, loff_t *ppos)
4310{
4311 return -EINVAL;
4312}
4313
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004314
4315static void print_slabinfo_header(struct seq_file *m)
4316{
4317 seq_puts(m, "slabinfo - version: 2.1\n");
4318 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4319 "<objperslab> <pagesperslab>");
4320 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4321 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4322 seq_putc(m, '\n');
4323}
4324
4325static void *s_start(struct seq_file *m, loff_t *pos)
4326{
4327 loff_t n = *pos;
4328
4329 down_read(&slub_lock);
4330 if (!n)
4331 print_slabinfo_header(m);
4332
4333 return seq_list_start(&slab_caches, *pos);
4334}
4335
4336static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4337{
4338 return seq_list_next(p, &slab_caches, pos);
4339}
4340
4341static void s_stop(struct seq_file *m, void *p)
4342{
4343 up_read(&slub_lock);
4344}
4345
4346static int s_show(struct seq_file *m, void *p)
4347{
4348 unsigned long nr_partials = 0;
4349 unsigned long nr_slabs = 0;
4350 unsigned long nr_inuse = 0;
4351 unsigned long nr_objs;
4352 struct kmem_cache *s;
4353 int node;
4354
4355 s = list_entry(p, struct kmem_cache, list);
4356
4357 for_each_online_node(node) {
4358 struct kmem_cache_node *n = get_node(s, node);
4359
4360 if (!n)
4361 continue;
4362
4363 nr_partials += n->nr_partial;
4364 nr_slabs += atomic_long_read(&n->nr_slabs);
4365 nr_inuse += count_partial(n);
4366 }
4367
4368 nr_objs = nr_slabs * s->objects;
4369 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4370
4371 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4372 nr_objs, s->size, s->objects, (1 << s->order));
4373 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4374 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4375 0UL);
4376 seq_putc(m, '\n');
4377 return 0;
4378}
4379
4380const struct seq_operations slabinfo_op = {
4381 .start = s_start,
4382 .next = s_next,
4383 .stop = s_stop,
4384 .show = s_show,
4385};
4386
Linus Torvalds158a9622008-01-02 13:04:48 -08004387#endif /* CONFIG_SLABINFO */