blob: 7ec2888bffc1efc6e7a02b771056699391e506af [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 *
Christoph Lametercde53532008-07-04 09:59:22 -07008 * (C) 2007 SGI, Christoph Lameter
Christoph Lameter81819f02007-05-06 14:49:36 -07009 */
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>
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +040017#include <linux/proc_fs.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070018#include <linux/seq_file.h>
19#include <linux/cpu.h>
20#include <linux/cpuset.h>
21#include <linux/mempolicy.h>
22#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070023#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070025#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070026#include <linux/math64.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070027
28/*
29 * Lock order:
30 * 1. slab_lock(page)
31 * 2. slab->list_lock
32 *
33 * The slab_lock protects operations on the object of a particular
34 * slab and its metadata in the page struct. If the slab lock
35 * has been taken then no allocations nor frees can be performed
36 * on the objects in the slab nor can the slab be added or removed
37 * from the partial or full lists since this would mean modifying
38 * the page_struct of the slab.
39 *
40 * The list_lock protects the partial and full list on each node and
41 * the partial slab counter. If taken then no new slabs may be added or
42 * removed from the lists nor make the number of partial slabs be modified.
43 * (Note that the total number of slabs is an atomic value that may be
44 * modified without taking the list lock).
45 *
46 * The list_lock is a centralized lock and thus we avoid taking it as
47 * much as possible. As long as SLUB does not have to handle partial
48 * slabs, operations can continue without any centralized lock. F.e.
49 * allocating a long series of objects that fill up slabs does not require
50 * the list lock.
51 *
52 * The lock order is sometimes inverted when we are trying to get a slab
53 * off a list. We take the list_lock and then look for a page on the list
54 * to use. While we do that objects in the slabs may be freed. We can
55 * only operate on the slab if we have also taken the slab_lock. So we use
56 * a slab_trylock() on the slab. If trylock was successful then no frees
57 * can occur anymore and we can use the slab for allocations etc. If the
58 * slab_trylock() does not succeed then frees are in progress in the slab and
59 * we must stay away from it for a while since we may cause a bouncing
60 * cacheline if we try to acquire the lock. So go onto the next slab.
61 * If all pages are busy then we may allocate a new slab instead of reusing
62 * a partial slab. A new slab has noone operating on it and thus there is
63 * no danger of cacheline contention.
64 *
65 * Interrupts are disabled during allocation and deallocation in order to
66 * make the slab allocator safe to use in the context of an irq. In addition
67 * interrupts are disabled to ensure that the processor does not change
68 * while handling per_cpu slabs, due to kernel preemption.
69 *
70 * SLUB assigns one slab for allocation to each processor.
71 * Allocations only occur from these slabs called cpu slabs.
72 *
Christoph Lameter672bba32007-05-09 02:32:39 -070073 * Slabs with free elements are kept on a partial list and during regular
74 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070075 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070076 * We track full slabs for debugging purposes though because otherwise we
77 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070078 *
79 * Slabs are freed when they become empty. Teardown and setup is
80 * minimal so we rely on the page allocators per cpu caches for
81 * fast frees and allocs.
82 *
83 * Overloading of page flags that are otherwise used for LRU management.
84 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070085 * PageActive The slab is frozen and exempt from list processing.
86 * This means that the slab is dedicated to a purpose
87 * such as satisfying allocations for a specific
88 * processor. Objects may be freed in the slab while
89 * it is frozen but slab_free will then skip the usual
90 * list operations. It is up to the processor holding
91 * the slab to integrate the slab into the slab lists
92 * when the slab is no longer needed.
93 *
94 * One use of this flag is to mark slabs that are
95 * used for allocations. Then such a slab becomes a cpu
96 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -070097 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -070098 * free objects in addition to the regular freelist
99 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -0700100 *
101 * PageError Slab requires special handling due to debug
102 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700103 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700104 */
105
Christoph Lameter5577bd82007-05-16 22:10:56 -0700106#ifdef CONFIG_SLUB_DEBUG
Andy Whitcroft8a380822008-07-23 21:27:18 -0700107#define SLABDEBUG 1
Christoph Lameter5577bd82007-05-16 22:10:56 -0700108#else
109#define SLABDEBUG 0
110#endif
111
Christoph Lameter81819f02007-05-06 14:49:36 -0700112/*
113 * Issues still to be resolved:
114 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700115 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
116 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700117 * - Variable sizing of the per node arrays
118 */
119
120/* Enable to test recovery from slab corruption on boot */
121#undef SLUB_RESILIENCY_TEST
122
Christoph Lameter81819f02007-05-06 14:49:36 -0700123/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700124 * Mininum number of partial slabs. These will be left on the partial
125 * lists even if they are empty. kmem_cache_shrink may reclaim them.
126 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800127#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700128
Christoph Lameter2086d262007-05-06 14:49:46 -0700129/*
130 * Maximum number of desirable partial slabs.
131 * The existence of more partial slabs makes kmem_cache_shrink
132 * sort the partial list by the number of objects in the.
133 */
134#define MAX_PARTIAL 10
135
Christoph Lameter81819f02007-05-06 14:49:36 -0700136#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
137 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700138
Christoph Lameter81819f02007-05-06 14:49:36 -0700139/*
140 * Set of flags that will prevent slab merging
141 */
142#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
143 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
144
145#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
146 SLAB_CACHE_DMA)
147
148#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700149#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700150#endif
151
152#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700153#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700154#endif
155
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400156#define OO_SHIFT 16
157#define OO_MASK ((1 << OO_SHIFT) - 1)
158#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
159
Christoph Lameter81819f02007-05-06 14:49:36 -0700160/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700161#define __OBJECT_POISON 0x80000000 /* Poison object */
162#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700163
164static int kmem_size = sizeof(struct kmem_cache);
165
166#ifdef CONFIG_SMP
167static struct notifier_block slab_notifier;
168#endif
169
170static enum {
171 DOWN, /* No slab functionality available */
172 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700173 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700174 SYSFS /* Sysfs up */
175} slab_state = DOWN;
176
177/* A list of all slab caches on the system */
178static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700179static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700180
Christoph Lameter02cbc872007-05-09 02:32:43 -0700181/*
182 * Tracking user of a slab.
183 */
184struct track {
185 void *addr; /* Called from address */
186 int cpu; /* Was running on cpu */
187 int pid; /* Pid context */
188 unsigned long when; /* When did the operation occur */
189};
190
191enum track_item { TRACK_ALLOC, TRACK_FREE };
192
Christoph Lameterf6acb632008-04-29 16:16:06 -0700193#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -0700194static int sysfs_slab_add(struct kmem_cache *);
195static int sysfs_slab_alias(struct kmem_cache *, const char *);
196static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800197
Christoph Lameter81819f02007-05-06 14:49:36 -0700198#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700199static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
200static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
201 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800202static inline void sysfs_slab_remove(struct kmem_cache *s)
203{
204 kfree(s);
205}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800206
Christoph Lameter81819f02007-05-06 14:49:36 -0700207#endif
208
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800209static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
210{
211#ifdef CONFIG_SLUB_STATS
212 c->stat[si]++;
213#endif
214}
215
Christoph Lameter81819f02007-05-06 14:49:36 -0700216/********************************************************************
217 * Core slab cache functions
218 *******************************************************************/
219
220int slab_is_available(void)
221{
222 return slab_state >= UP;
223}
224
225static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
226{
227#ifdef CONFIG_NUMA
228 return s->node[node];
229#else
230 return &s->local_node;
231#endif
232}
233
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700234static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
235{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700236#ifdef CONFIG_SMP
237 return s->cpu_slab[cpu];
238#else
239 return &s->cpu_slab;
240#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700241}
242
Christoph Lameter6446faa2008-02-15 23:45:26 -0800243/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700244static inline int check_valid_pointer(struct kmem_cache *s,
245 struct page *page, const void *object)
246{
247 void *base;
248
Christoph Lametera973e9d2008-03-01 13:40:44 -0800249 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700250 return 1;
251
Christoph Lametera973e9d2008-03-01 13:40:44 -0800252 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300253 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700254 (object - base) % s->size) {
255 return 0;
256 }
257
258 return 1;
259}
260
Christoph Lameter81819f02007-05-06 14:49:36 -0700261/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700262 * Slow version of get and set free pointer.
263 *
264 * This version requires touching the cache lines of kmem_cache which
265 * we avoid to do in the fast alloc free paths. There we obtain the offset
266 * from the page struct.
267 */
268static inline void *get_freepointer(struct kmem_cache *s, void *object)
269{
270 return *(void **)(object + s->offset);
271}
272
273static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
274{
275 *(void **)(object + s->offset) = fp;
276}
277
278/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300279#define for_each_object(__p, __s, __addr, __objects) \
280 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700281 __p += (__s)->size)
282
283/* Scan freelist */
284#define for_each_free_object(__p, __s, __free) \
Christoph Lametera973e9d2008-03-01 13:40:44 -0800285 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700286
287/* Determine object index from a given position */
288static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
289{
290 return (p - addr) / s->size;
291}
292
Christoph Lameter834f3d12008-04-14 19:11:31 +0300293static inline struct kmem_cache_order_objects oo_make(int order,
294 unsigned long size)
295{
296 struct kmem_cache_order_objects x = {
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400297 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
Christoph Lameter834f3d12008-04-14 19:11:31 +0300298 };
299
300 return x;
301}
302
303static inline int oo_order(struct kmem_cache_order_objects x)
304{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400305 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300306}
307
308static inline int oo_objects(struct kmem_cache_order_objects x)
309{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400310 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300311}
312
Christoph Lameter41ecc552007-05-09 02:32:44 -0700313#ifdef CONFIG_SLUB_DEBUG
314/*
315 * Debug settings:
316 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700317#ifdef CONFIG_SLUB_DEBUG_ON
318static int slub_debug = DEBUG_DEFAULT_FLAGS;
319#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700320static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700321#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700322
323static char *slub_debug_slabs;
324
Christoph Lameter7656c722007-05-09 02:32:40 -0700325/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700326 * Object debugging
327 */
328static void print_section(char *text, u8 *addr, unsigned int length)
329{
330 int i, offset;
331 int newline = 1;
332 char ascii[17];
333
334 ascii[16] = 0;
335
336 for (i = 0; i < length; i++) {
337 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700338 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700339 newline = 0;
340 }
Pekka Enberg06428782008-01-07 23:20:27 -0800341 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700342 offset = i % 16;
343 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
344 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800345 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700346 newline = 1;
347 }
348 }
349 if (!newline) {
350 i %= 16;
351 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800352 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700353 ascii[i] = ' ';
354 i++;
355 }
Pekka Enberg06428782008-01-07 23:20:27 -0800356 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700357 }
358}
359
Christoph Lameter81819f02007-05-06 14:49:36 -0700360static struct track *get_track(struct kmem_cache *s, void *object,
361 enum track_item alloc)
362{
363 struct track *p;
364
365 if (s->offset)
366 p = object + s->offset + sizeof(void *);
367 else
368 p = object + s->inuse;
369
370 return p + alloc;
371}
372
373static void set_track(struct kmem_cache *s, void *object,
374 enum track_item alloc, void *addr)
375{
376 struct track *p;
377
378 if (s->offset)
379 p = object + s->offset + sizeof(void *);
380 else
381 p = object + s->inuse;
382
383 p += alloc;
384 if (addr) {
385 p->addr = addr;
386 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400387 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700388 p->when = jiffies;
389 } else
390 memset(p, 0, sizeof(struct track));
391}
392
Christoph Lameter81819f02007-05-06 14:49:36 -0700393static void init_tracking(struct kmem_cache *s, void *object)
394{
Christoph Lameter24922682007-07-17 04:03:18 -0700395 if (!(s->flags & SLAB_STORE_USER))
396 return;
397
398 set_track(s, object, TRACK_FREE, NULL);
399 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700400}
401
402static void print_track(const char *s, struct track *t)
403{
404 if (!t->addr)
405 return;
406
Linus Torvalds7daf7052008-07-14 12:12:53 -0700407 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
408 s, t->addr, jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700409}
410
Christoph Lameter24922682007-07-17 04:03:18 -0700411static void print_tracking(struct kmem_cache *s, void *object)
412{
413 if (!(s->flags & SLAB_STORE_USER))
414 return;
415
416 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
417 print_track("Freed", get_track(s, object, TRACK_FREE));
418}
419
420static void print_page_info(struct page *page)
421{
Christoph Lameter39b26462008-04-14 19:11:30 +0300422 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
423 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700424
425}
426
427static void slab_bug(struct kmem_cache *s, char *fmt, ...)
428{
429 va_list args;
430 char buf[100];
431
432 va_start(args, fmt);
433 vsnprintf(buf, sizeof(buf), fmt, args);
434 va_end(args);
435 printk(KERN_ERR "========================================"
436 "=====================================\n");
437 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
438 printk(KERN_ERR "----------------------------------------"
439 "-------------------------------------\n\n");
440}
441
442static void slab_fix(struct kmem_cache *s, char *fmt, ...)
443{
444 va_list args;
445 char buf[100];
446
447 va_start(args, fmt);
448 vsnprintf(buf, sizeof(buf), fmt, args);
449 va_end(args);
450 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
451}
452
453static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700454{
455 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800456 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700457
458 print_tracking(s, p);
459
460 print_page_info(page);
461
462 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
463 p, p - addr, get_freepointer(s, p));
464
465 if (p > addr + 16)
466 print_section("Bytes b4", p - 16, 16);
467
Pekka Enberg0ebd6522008-07-19 14:17:22 +0300468 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700469
470 if (s->flags & SLAB_RED_ZONE)
471 print_section("Redzone", p + s->objsize,
472 s->inuse - s->objsize);
473
Christoph Lameter81819f02007-05-06 14:49:36 -0700474 if (s->offset)
475 off = s->offset + sizeof(void *);
476 else
477 off = s->inuse;
478
Christoph Lameter24922682007-07-17 04:03:18 -0700479 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700480 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700481
482 if (off != s->size)
483 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700484 print_section("Padding", p + off, s->size - off);
485
486 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700487}
488
489static void object_err(struct kmem_cache *s, struct page *page,
490 u8 *object, char *reason)
491{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700492 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700493 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700494}
495
Christoph Lameter24922682007-07-17 04:03:18 -0700496static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700497{
498 va_list args;
499 char buf[100];
500
Christoph Lameter24922682007-07-17 04:03:18 -0700501 va_start(args, fmt);
502 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700503 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700504 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700505 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700506 dump_stack();
507}
508
509static void init_object(struct kmem_cache *s, void *object, int active)
510{
511 u8 *p = object;
512
513 if (s->flags & __OBJECT_POISON) {
514 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800515 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700516 }
517
518 if (s->flags & SLAB_RED_ZONE)
519 memset(p + s->objsize,
520 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
521 s->inuse - s->objsize);
522}
523
Christoph Lameter24922682007-07-17 04:03:18 -0700524static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700525{
526 while (bytes) {
527 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700528 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700529 start++;
530 bytes--;
531 }
Christoph Lameter24922682007-07-17 04:03:18 -0700532 return NULL;
533}
534
535static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
536 void *from, void *to)
537{
538 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
539 memset(from, data, to - from);
540}
541
542static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
543 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800544 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700545{
546 u8 *fault;
547 u8 *end;
548
549 fault = check_bytes(start, value, bytes);
550 if (!fault)
551 return 1;
552
553 end = start + bytes;
554 while (end > fault && end[-1] == value)
555 end--;
556
557 slab_bug(s, "%s overwritten", what);
558 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
559 fault, end - 1, fault[0], value);
560 print_trailer(s, page, object);
561
562 restore_bytes(s, what, value, fault, end);
563 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700564}
565
Christoph Lameter81819f02007-05-06 14:49:36 -0700566/*
567 * Object layout:
568 *
569 * object address
570 * Bytes of the object to be managed.
571 * If the freepointer may overlay the object then the free
572 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700573 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700574 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
575 * 0xa5 (POISON_END)
576 *
577 * object + s->objsize
578 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700579 * Padding is extended by another word if Redzoning is enabled and
580 * objsize == inuse.
581 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700582 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
583 * 0xcc (RED_ACTIVE) for objects in use.
584 *
585 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700586 * Meta data starts here.
587 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700588 * A. Free pointer (if we cannot overwrite object on free)
589 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700590 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800591 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700592 * before the word boundary.
593 *
594 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700595 *
596 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700597 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700598 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700599 * If slabcaches are merged then the objsize and inuse boundaries are mostly
600 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700601 * may be used with merged slabcaches.
602 */
603
Christoph Lameter81819f02007-05-06 14:49:36 -0700604static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
605{
606 unsigned long off = s->inuse; /* The end of info */
607
608 if (s->offset)
609 /* Freepointer is placed after the object. */
610 off += sizeof(void *);
611
612 if (s->flags & SLAB_STORE_USER)
613 /* We also have user information there */
614 off += 2 * sizeof(struct track);
615
616 if (s->size == off)
617 return 1;
618
Christoph Lameter24922682007-07-17 04:03:18 -0700619 return check_bytes_and_report(s, page, p, "Object padding",
620 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700621}
622
Christoph Lameter39b26462008-04-14 19:11:30 +0300623/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700624static int slab_pad_check(struct kmem_cache *s, struct page *page)
625{
Christoph Lameter24922682007-07-17 04:03:18 -0700626 u8 *start;
627 u8 *fault;
628 u8 *end;
629 int length;
630 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700631
632 if (!(s->flags & SLAB_POISON))
633 return 1;
634
Christoph Lametera973e9d2008-03-01 13:40:44 -0800635 start = page_address(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +0300636 length = (PAGE_SIZE << compound_order(page));
Christoph Lameter39b26462008-04-14 19:11:30 +0300637 end = start + length;
638 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700639 if (!remainder)
640 return 1;
641
Christoph Lameter39b26462008-04-14 19:11:30 +0300642 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700643 if (!fault)
644 return 1;
645 while (end > fault && end[-1] == POISON_INUSE)
646 end--;
647
648 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Christoph Lameter39b26462008-04-14 19:11:30 +0300649 print_section("Padding", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700650
651 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
652 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700653}
654
655static int check_object(struct kmem_cache *s, struct page *page,
656 void *object, int active)
657{
658 u8 *p = object;
659 u8 *endobject = object + s->objsize;
660
661 if (s->flags & SLAB_RED_ZONE) {
662 unsigned int red =
663 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
664
Christoph Lameter24922682007-07-17 04:03:18 -0700665 if (!check_bytes_and_report(s, page, object, "Redzone",
666 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700667 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700668 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800669 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
670 check_bytes_and_report(s, page, p, "Alignment padding",
671 endobject, POISON_INUSE, s->inuse - s->objsize);
672 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700673 }
674
675 if (s->flags & SLAB_POISON) {
676 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700677 (!check_bytes_and_report(s, page, p, "Poison", p,
678 POISON_FREE, s->objsize - 1) ||
679 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800680 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700681 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700682 /*
683 * check_pad_bytes cleans up on its own.
684 */
685 check_pad_bytes(s, page, p);
686 }
687
688 if (!s->offset && active)
689 /*
690 * Object and freepointer overlap. Cannot check
691 * freepointer while object is allocated.
692 */
693 return 1;
694
695 /* Check free pointer validity */
696 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
697 object_err(s, page, p, "Freepointer corrupt");
698 /*
699 * No choice but to zap it and thus loose the remainder
700 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700701 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700702 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800703 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700704 return 0;
705 }
706 return 1;
707}
708
709static int check_slab(struct kmem_cache *s, struct page *page)
710{
Christoph Lameter39b26462008-04-14 19:11:30 +0300711 int maxobj;
712
Christoph Lameter81819f02007-05-06 14:49:36 -0700713 VM_BUG_ON(!irqs_disabled());
714
715 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700716 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700717 return 0;
718 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300719
720 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
721 if (page->objects > maxobj) {
722 slab_err(s, page, "objects %u > max %u",
723 s->name, page->objects, maxobj);
724 return 0;
725 }
726 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700727 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300728 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700729 return 0;
730 }
731 /* Slab_pad_check fixes things up after itself */
732 slab_pad_check(s, page);
733 return 1;
734}
735
736/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700737 * Determine if a certain object on a page is on the freelist. Must hold the
738 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700739 */
740static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
741{
742 int nr = 0;
743 void *fp = page->freelist;
744 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300745 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700746
Christoph Lameter39b26462008-04-14 19:11:30 +0300747 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700748 if (fp == search)
749 return 1;
750 if (!check_valid_pointer(s, page, fp)) {
751 if (object) {
752 object_err(s, page, object,
753 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800754 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700755 break;
756 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700757 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800758 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300759 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700760 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700761 return 0;
762 }
763 break;
764 }
765 object = fp;
766 fp = get_freepointer(s, object);
767 nr++;
768 }
769
Christoph Lameter224a88b2008-04-14 19:11:31 +0300770 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400771 if (max_objects > MAX_OBJS_PER_PAGE)
772 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300773
774 if (page->objects != max_objects) {
775 slab_err(s, page, "Wrong number of objects. Found %d but "
776 "should be %d", page->objects, max_objects);
777 page->objects = max_objects;
778 slab_fix(s, "Number of objects adjusted.");
779 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300780 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700781 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300782 "counted were %d", page->inuse, page->objects - nr);
783 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700784 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700785 }
786 return search == NULL;
787}
788
Christoph Lameter0121c6192008-04-29 16:11:12 -0700789static void trace(struct kmem_cache *s, struct page *page, void *object,
790 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700791{
792 if (s->flags & SLAB_TRACE) {
793 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
794 s->name,
795 alloc ? "alloc" : "free",
796 object, page->inuse,
797 page->freelist);
798
799 if (!alloc)
800 print_section("Object", (void *)object, s->objsize);
801
802 dump_stack();
803 }
804}
805
Christoph Lameter643b1132007-05-06 14:49:42 -0700806/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700807 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700808 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700809static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700810{
Christoph Lameter643b1132007-05-06 14:49:42 -0700811 spin_lock(&n->list_lock);
812 list_add(&page->lru, &n->full);
813 spin_unlock(&n->list_lock);
814}
815
816static void remove_full(struct kmem_cache *s, struct page *page)
817{
818 struct kmem_cache_node *n;
819
820 if (!(s->flags & SLAB_STORE_USER))
821 return;
822
823 n = get_node(s, page_to_nid(page));
824
825 spin_lock(&n->list_lock);
826 list_del(&page->lru);
827 spin_unlock(&n->list_lock);
828}
829
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300830/* Tracking of the number of slabs for debugging purposes */
831static inline unsigned long slabs_node(struct kmem_cache *s, int node)
832{
833 struct kmem_cache_node *n = get_node(s, node);
834
835 return atomic_long_read(&n->nr_slabs);
836}
837
Christoph Lameter205ab992008-04-14 19:11:40 +0300838static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300839{
840 struct kmem_cache_node *n = get_node(s, node);
841
842 /*
843 * May be called early in order to allocate a slab for the
844 * kmem_cache_node structure. Solve the chicken-egg
845 * dilemma by deferring the increment of the count during
846 * bootstrap (see early_kmem_cache_node_alloc).
847 */
Christoph Lameter205ab992008-04-14 19:11:40 +0300848 if (!NUMA_BUILD || n) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300849 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300850 atomic_long_add(objects, &n->total_objects);
851 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300852}
Christoph Lameter205ab992008-04-14 19:11:40 +0300853static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300854{
855 struct kmem_cache_node *n = get_node(s, node);
856
857 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300858 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300859}
860
861/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -0700862static void setup_object_debug(struct kmem_cache *s, struct page *page,
863 void *object)
864{
865 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
866 return;
867
868 init_object(s, object, 0);
869 init_tracking(s, object);
870}
871
872static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
873 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700874{
875 if (!check_slab(s, page))
876 goto bad;
877
Christoph Lameterd692ef62008-02-15 23:45:24 -0800878 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700879 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700880 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700881 }
882
883 if (!check_valid_pointer(s, page, object)) {
884 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700885 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700886 }
887
Christoph Lameterd692ef62008-02-15 23:45:24 -0800888 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700889 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700890
Christoph Lameter3ec09742007-05-16 22:11:00 -0700891 /* Success perform special debug activities for allocs */
892 if (s->flags & SLAB_STORE_USER)
893 set_track(s, object, TRACK_ALLOC, addr);
894 trace(s, page, object, 1);
895 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700896 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700897
Christoph Lameter81819f02007-05-06 14:49:36 -0700898bad:
899 if (PageSlab(page)) {
900 /*
901 * If this is a slab page then lets do the best we can
902 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700903 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700904 */
Christoph Lameter24922682007-07-17 04:03:18 -0700905 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +0300906 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800907 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700908 }
909 return 0;
910}
911
Christoph Lameter3ec09742007-05-16 22:11:00 -0700912static int free_debug_processing(struct kmem_cache *s, struct page *page,
913 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700914{
915 if (!check_slab(s, page))
916 goto fail;
917
918 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700919 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700920 goto fail;
921 }
922
923 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700924 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700925 goto fail;
926 }
927
928 if (!check_object(s, page, object, 1))
929 return 0;
930
931 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800932 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700933 slab_err(s, page, "Attempt to free object(0x%p) "
934 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800935 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700936 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700937 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700938 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700939 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800940 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700941 object_err(s, page, object,
942 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700943 goto fail;
944 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700945
946 /* Special debug activities for freeing objects */
Andy Whitcroft8a380822008-07-23 21:27:18 -0700947 if (!PageSlubFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700948 remove_full(s, page);
949 if (s->flags & SLAB_STORE_USER)
950 set_track(s, object, TRACK_FREE, addr);
951 trace(s, page, object, 0);
952 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700953 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700954
Christoph Lameter81819f02007-05-06 14:49:36 -0700955fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700956 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700957 return 0;
958}
959
Christoph Lameter41ecc552007-05-09 02:32:44 -0700960static int __init setup_slub_debug(char *str)
961{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700962 slub_debug = DEBUG_DEFAULT_FLAGS;
963 if (*str++ != '=' || !*str)
964 /*
965 * No options specified. Switch on full debugging.
966 */
967 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700968
969 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700970 /*
971 * No options but restriction on slabs. This means full
972 * debugging for slabs matching a pattern.
973 */
974 goto check_slabs;
975
976 slub_debug = 0;
977 if (*str == '-')
978 /*
979 * Switch off all debugging measures.
980 */
981 goto out;
982
983 /*
984 * Determine which debug features should be switched on
985 */
Pekka Enberg06428782008-01-07 23:20:27 -0800986 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700987 switch (tolower(*str)) {
988 case 'f':
989 slub_debug |= SLAB_DEBUG_FREE;
990 break;
991 case 'z':
992 slub_debug |= SLAB_RED_ZONE;
993 break;
994 case 'p':
995 slub_debug |= SLAB_POISON;
996 break;
997 case 'u':
998 slub_debug |= SLAB_STORE_USER;
999 break;
1000 case 't':
1001 slub_debug |= SLAB_TRACE;
1002 break;
1003 default:
1004 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001005 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001006 }
1007 }
1008
1009check_slabs:
1010 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001011 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001012out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001013 return 1;
1014}
1015
1016__setup("slub_debug", setup_slub_debug);
1017
Christoph Lameterba0268a2007-09-11 15:24:11 -07001018static unsigned long kmem_cache_flags(unsigned long objsize,
1019 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001020 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001021{
1022 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001023 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001024 */
Christoph Lametere1533622008-02-15 23:45:24 -08001025 if (slub_debug && (!slub_debug_slabs ||
1026 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1027 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001028
1029 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001030}
1031#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001032static inline void setup_object_debug(struct kmem_cache *s,
1033 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001034
Christoph Lameter3ec09742007-05-16 22:11:00 -07001035static inline int alloc_debug_processing(struct kmem_cache *s,
1036 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001037
Christoph Lameter3ec09742007-05-16 22:11:00 -07001038static inline int free_debug_processing(struct kmem_cache *s,
1039 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001040
Christoph Lameter41ecc552007-05-09 02:32:44 -07001041static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1042 { return 1; }
1043static inline int check_object(struct kmem_cache *s, struct page *page,
1044 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001045static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001046static inline unsigned long kmem_cache_flags(unsigned long objsize,
1047 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001048 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001049{
1050 return flags;
1051}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001052#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001053
1054static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1055 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001056static inline void inc_slabs_node(struct kmem_cache *s, int node,
1057 int objects) {}
1058static inline void dec_slabs_node(struct kmem_cache *s, int node,
1059 int objects) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001060#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03001061
Christoph Lameter81819f02007-05-06 14:49:36 -07001062/*
1063 * Slab allocation and freeing
1064 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001065static inline struct page *alloc_slab_page(gfp_t flags, int node,
1066 struct kmem_cache_order_objects oo)
1067{
1068 int order = oo_order(oo);
1069
1070 if (node == -1)
1071 return alloc_pages(flags, order);
1072 else
1073 return alloc_pages_node(node, flags, order);
1074}
1075
Christoph Lameter81819f02007-05-06 14:49:36 -07001076static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1077{
Pekka Enberg06428782008-01-07 23:20:27 -08001078 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001079 struct kmem_cache_order_objects oo = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07001080
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001081 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001082
Christoph Lameter65c33762008-04-14 19:11:40 +03001083 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1084 oo);
1085 if (unlikely(!page)) {
1086 oo = s->min;
1087 /*
1088 * Allocation may have failed due to fragmentation.
1089 * Try a lower order alloc if possible
1090 */
1091 page = alloc_slab_page(flags, node, oo);
1092 if (!page)
1093 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001094
Christoph Lameter65c33762008-04-14 19:11:40 +03001095 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1096 }
Christoph Lameter834f3d12008-04-14 19:11:31 +03001097 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001098 mod_zone_page_state(page_zone(page),
1099 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1100 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001101 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001102
1103 return page;
1104}
1105
1106static void setup_object(struct kmem_cache *s, struct page *page,
1107 void *object)
1108{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001109 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001110 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001111 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001112}
1113
1114static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1115{
1116 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001117 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001118 void *last;
1119 void *p;
1120
Christoph Lameter6cb06222007-10-16 01:25:41 -07001121 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001122
Christoph Lameter6cb06222007-10-16 01:25:41 -07001123 page = allocate_slab(s,
1124 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001125 if (!page)
1126 goto out;
1127
Christoph Lameter205ab992008-04-14 19:11:40 +03001128 inc_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001129 page->slab = s;
1130 page->flags |= 1 << PG_slab;
1131 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1132 SLAB_STORE_USER | SLAB_TRACE))
Andy Whitcroft8a380822008-07-23 21:27:18 -07001133 __SetPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001134
1135 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001136
1137 if (unlikely(s->flags & SLAB_POISON))
Christoph Lameter834f3d12008-04-14 19:11:31 +03001138 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001139
1140 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001141 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001142 setup_object(s, page, last);
1143 set_freepointer(s, last, p);
1144 last = p;
1145 }
1146 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001147 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001148
1149 page->freelist = start;
1150 page->inuse = 0;
1151out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001152 return page;
1153}
1154
1155static void __free_slab(struct kmem_cache *s, struct page *page)
1156{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001157 int order = compound_order(page);
1158 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001159
Andy Whitcroft8a380822008-07-23 21:27:18 -07001160 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001161 void *p;
1162
1163 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001164 for_each_object(p, s, page_address(page),
1165 page->objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001166 check_object(s, page, p, 0);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001167 __ClearPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001168 }
1169
1170 mod_zone_page_state(page_zone(page),
1171 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1172 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001173 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001174
Christoph Lameter49bd5222008-04-14 18:52:18 +03001175 __ClearPageSlab(page);
1176 reset_page_mapcount(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +03001177 __free_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001178}
1179
1180static void rcu_free_slab(struct rcu_head *h)
1181{
1182 struct page *page;
1183
1184 page = container_of((struct list_head *)h, struct page, lru);
1185 __free_slab(page->slab, page);
1186}
1187
1188static void free_slab(struct kmem_cache *s, struct page *page)
1189{
1190 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1191 /*
1192 * RCU free overloads the RCU head over the LRU
1193 */
1194 struct rcu_head *head = (void *)&page->lru;
1195
1196 call_rcu(head, rcu_free_slab);
1197 } else
1198 __free_slab(s, page);
1199}
1200
1201static void discard_slab(struct kmem_cache *s, struct page *page)
1202{
Christoph Lameter205ab992008-04-14 19:11:40 +03001203 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001204 free_slab(s, page);
1205}
1206
1207/*
1208 * Per slab locking using the pagelock
1209 */
1210static __always_inline void slab_lock(struct page *page)
1211{
1212 bit_spin_lock(PG_locked, &page->flags);
1213}
1214
1215static __always_inline void slab_unlock(struct page *page)
1216{
Nick Piggina76d3542008-01-07 23:20:27 -08001217 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001218}
1219
1220static __always_inline int slab_trylock(struct page *page)
1221{
1222 int rc = 1;
1223
1224 rc = bit_spin_trylock(PG_locked, &page->flags);
1225 return rc;
1226}
1227
1228/*
1229 * Management of partially allocated slabs
1230 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001231static void add_partial(struct kmem_cache_node *n,
1232 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001233{
Christoph Lametere95eed52007-05-06 14:49:44 -07001234 spin_lock(&n->list_lock);
1235 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001236 if (tail)
1237 list_add_tail(&page->lru, &n->partial);
1238 else
1239 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001240 spin_unlock(&n->list_lock);
1241}
1242
Christoph Lameter0121c6192008-04-29 16:11:12 -07001243static void remove_partial(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001244{
1245 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1246
1247 spin_lock(&n->list_lock);
1248 list_del(&page->lru);
1249 n->nr_partial--;
1250 spin_unlock(&n->list_lock);
1251}
1252
1253/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001254 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001255 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001256 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001257 */
Christoph Lameter0121c6192008-04-29 16:11:12 -07001258static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1259 struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001260{
1261 if (slab_trylock(page)) {
1262 list_del(&page->lru);
1263 n->nr_partial--;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001264 __SetPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001265 return 1;
1266 }
1267 return 0;
1268}
1269
1270/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001271 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001272 */
1273static struct page *get_partial_node(struct kmem_cache_node *n)
1274{
1275 struct page *page;
1276
1277 /*
1278 * Racy check. If we mistakenly see no partial slabs then we
1279 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001280 * partial slab and there is none available then get_partials()
1281 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001282 */
1283 if (!n || !n->nr_partial)
1284 return NULL;
1285
1286 spin_lock(&n->list_lock);
1287 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001288 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001289 goto out;
1290 page = NULL;
1291out:
1292 spin_unlock(&n->list_lock);
1293 return page;
1294}
1295
1296/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001297 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001298 */
1299static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1300{
1301#ifdef CONFIG_NUMA
1302 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001303 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001304 struct zone *zone;
1305 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001306 struct page *page;
1307
1308 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001309 * The defrag ratio allows a configuration of the tradeoffs between
1310 * inter node defragmentation and node local allocations. A lower
1311 * defrag_ratio increases the tendency to do local allocations
1312 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001313 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001314 * If the defrag_ratio is set to 0 then kmalloc() always
1315 * returns node local objects. If the ratio is higher then kmalloc()
1316 * may return off node objects because partial slabs are obtained
1317 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001318 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001319 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001320 * defrag_ratio = 1000) then every (well almost) allocation will
1321 * first attempt to defrag slab caches on other nodes. This means
1322 * scanning over all nodes to look for partial slabs which may be
1323 * expensive if we do it every time we are trying to find a slab
1324 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001325 */
Christoph Lameter98246012008-01-07 23:20:26 -08001326 if (!s->remote_node_defrag_ratio ||
1327 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001328 return NULL;
1329
Mel Gorman0e884602008-04-28 02:12:14 -07001330 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001331 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001332 struct kmem_cache_node *n;
1333
Mel Gorman54a6eb52008-04-28 02:12:16 -07001334 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001335
Mel Gorman54a6eb52008-04-28 02:12:16 -07001336 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
Pekka Enberg5595cff2008-08-05 09:28:47 +03001337 n->nr_partial > n->min_partial) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001338 page = get_partial_node(n);
1339 if (page)
1340 return page;
1341 }
1342 }
1343#endif
1344 return NULL;
1345}
1346
1347/*
1348 * Get a partial page, lock it and return it.
1349 */
1350static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1351{
1352 struct page *page;
1353 int searchnode = (node == -1) ? numa_node_id() : node;
1354
1355 page = get_partial_node(get_node(s, searchnode));
1356 if (page || (flags & __GFP_THISNODE))
1357 return page;
1358
1359 return get_any_partial(s, flags);
1360}
1361
1362/*
1363 * Move a page back to the lists.
1364 *
1365 * Must be called with the slab lock held.
1366 *
1367 * On exit the slab lock will have been dropped.
1368 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001369static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001370{
Christoph Lametere95eed52007-05-06 14:49:44 -07001371 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001372 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001373
Andy Whitcroft8a380822008-07-23 21:27:18 -07001374 __ClearPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001375 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001376
Christoph Lametera973e9d2008-03-01 13:40:44 -08001377 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001378 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001379 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1380 } else {
1381 stat(c, DEACTIVATE_FULL);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001382 if (SLABDEBUG && PageSlubDebug(page) &&
1383 (s->flags & SLAB_STORE_USER))
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001384 add_full(n, page);
1385 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001386 slab_unlock(page);
1387 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001388 stat(c, DEACTIVATE_EMPTY);
Pekka Enberg5595cff2008-08-05 09:28:47 +03001389 if (n->nr_partial < n->min_partial) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001390 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001391 * Adding an empty slab to the partial slabs in order
1392 * to avoid page allocator overhead. This slab needs
1393 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001394 * so that the others get filled first. That way the
1395 * size of the partial list stays small.
1396 *
Christoph Lameter0121c6192008-04-29 16:11:12 -07001397 * kmem_cache_shrink can reclaim any empty slabs from
1398 * the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001399 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001400 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001401 slab_unlock(page);
1402 } else {
1403 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001404 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001405 discard_slab(s, page);
1406 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001407 }
1408}
1409
1410/*
1411 * Remove the cpu slab
1412 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001413static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001414{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001415 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001416 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001417
Christoph Lameterb773ad72008-03-04 11:10:17 -08001418 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001419 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001420 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001421 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001422 * because both freelists are empty. So this is unlikely
1423 * to occur.
1424 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001425 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001426 void **object;
1427
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001428 tail = 0; /* Hot objects. Put the slab first */
1429
Christoph Lameter894b8782007-05-10 03:15:16 -07001430 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001431 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001432 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001433
1434 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001435 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001436 page->freelist = object;
1437 page->inuse--;
1438 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001439 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001440 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001441}
1442
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001443static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001444{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001445 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001446 slab_lock(c->page);
1447 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001448}
1449
1450/*
1451 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001452 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001453 * Called from IPI handler with interrupts disabled.
1454 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001455static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001456{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001457 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001458
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001459 if (likely(c && c->page))
1460 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001461}
1462
1463static void flush_cpu_slab(void *d)
1464{
1465 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001466
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001467 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001468}
1469
1470static void flush_all(struct kmem_cache *s)
1471{
Jens Axboe15c8b6c2008-05-09 09:39:44 +02001472 on_each_cpu(flush_cpu_slab, s, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07001473}
1474
1475/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001476 * Check if the objects in a per cpu structure fit numa
1477 * locality expectations.
1478 */
1479static inline int node_match(struct kmem_cache_cpu *c, int node)
1480{
1481#ifdef CONFIG_NUMA
1482 if (node != -1 && c->node != node)
1483 return 0;
1484#endif
1485 return 1;
1486}
1487
1488/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001489 * Slow path. The lockless freelist is empty or we need to perform
1490 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001491 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001492 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001493 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001494 * Processing is still very fast if new objects have been freed to the
1495 * regular freelist. In that case we simply take over the regular freelist
1496 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001497 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001498 * If that is not working then we fall back to the partial lists. We take the
1499 * first element of the freelist as the object to allocate now and move the
1500 * rest of the freelist to the lockless freelist.
1501 *
1502 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001503 * we need to allocate a new slab. This is the slowest path since it involves
1504 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001505 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001506static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001507 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001508{
Christoph Lameter81819f02007-05-06 14:49:36 -07001509 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001510 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001511
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001512 /* We handle __GFP_ZERO in the caller */
1513 gfpflags &= ~__GFP_ZERO;
1514
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001515 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001516 goto new_slab;
1517
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001518 slab_lock(c->page);
1519 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001520 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001521
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001522 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001523
Christoph Lameter894b8782007-05-10 03:15:16 -07001524load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001525 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001526 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001527 goto another_slab;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001528 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001529 goto debug;
1530
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001531 c->freelist = object[c->offset];
Christoph Lameter39b26462008-04-14 19:11:30 +03001532 c->page->inuse = c->page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001533 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001534 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001535unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001536 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001537 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001538 return object;
1539
1540another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001541 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001542
1543new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001544 new = get_partial(s, gfpflags, node);
1545 if (new) {
1546 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001547 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001548 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001549 }
1550
Christoph Lameterb811c202007-10-16 23:25:51 -07001551 if (gfpflags & __GFP_WAIT)
1552 local_irq_enable();
1553
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001554 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001555
1556 if (gfpflags & __GFP_WAIT)
1557 local_irq_disable();
1558
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001559 if (new) {
1560 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001561 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001562 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001563 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001564 slab_lock(new);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001565 __SetPageSlubFrozen(new);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001566 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001567 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001568 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001569 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001570debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001571 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001572 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001573
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001574 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001575 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001576 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001577 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001578}
1579
1580/*
1581 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1582 * have the fastpath folded into their functions. So no function call
1583 * overhead for requests that can be satisfied on the fastpath.
1584 *
1585 * The fastpath works by first checking if the lockless freelist can be used.
1586 * If not then __slab_alloc is called for slow processing.
1587 *
1588 * Otherwise we can simply pick the next object from the lockless free list.
1589 */
Pekka Enberg06428782008-01-07 23:20:27 -08001590static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001591 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001592{
Christoph Lameter894b8782007-05-10 03:15:16 -07001593 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001594 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001595 unsigned long flags;
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001596 unsigned int objsize;
Christoph Lameter1f842602008-01-07 23:20:30 -08001597
Christoph Lameter894b8782007-05-10 03:15:16 -07001598 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001599 c = get_cpu_slab(s, smp_processor_id());
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001600 objsize = c->objsize;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001601 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001602
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001603 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001604
1605 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001606 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001607 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001608 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001609 }
1610 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001611
1612 if (unlikely((gfpflags & __GFP_ZERO) && object))
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001613 memset(object, 0, objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001614
Christoph Lameter894b8782007-05-10 03:15:16 -07001615 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001616}
1617
1618void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1619{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001620 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001621}
1622EXPORT_SYMBOL(kmem_cache_alloc);
1623
1624#ifdef CONFIG_NUMA
1625void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1626{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001627 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001628}
1629EXPORT_SYMBOL(kmem_cache_alloc_node);
1630#endif
1631
1632/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001633 * Slow patch handling. This may still be called frequently since objects
1634 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001635 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001636 * So we still attempt to reduce cache line usage. Just take the slab
1637 * lock and free the item. If there is no additional partial page
1638 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001639 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001640static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001641 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001642{
1643 void *prior;
1644 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001645 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001646
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001647 c = get_cpu_slab(s, raw_smp_processor_id());
1648 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001649 slab_lock(page);
1650
Andy Whitcroft8a380822008-07-23 21:27:18 -07001651 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001652 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001653
Christoph Lameter81819f02007-05-06 14:49:36 -07001654checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001655 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001656 page->freelist = object;
1657 page->inuse--;
1658
Andy Whitcroft8a380822008-07-23 21:27:18 -07001659 if (unlikely(PageSlubFrozen(page))) {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001660 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001661 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001662 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001663
1664 if (unlikely(!page->inuse))
1665 goto slab_empty;
1666
1667 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001668 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001669 * then add it.
1670 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001671 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001672 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001673 stat(c, FREE_ADD_PARTIAL);
1674 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001675
1676out_unlock:
1677 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001678 return;
1679
1680slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001681 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001682 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001683 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001684 */
1685 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001686 stat(c, FREE_REMOVE_PARTIAL);
1687 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001688 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001689 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001690 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001691 return;
1692
1693debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001694 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001695 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001696 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001697}
1698
Christoph Lameter894b8782007-05-10 03:15:16 -07001699/*
1700 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1701 * can perform fastpath freeing without additional function calls.
1702 *
1703 * The fastpath is only possible if we are freeing to the current cpu slab
1704 * of this processor. This typically the case if we have just allocated
1705 * the item before.
1706 *
1707 * If fastpath is not possible then fall back to __slab_free where we deal
1708 * with all sorts of special processing.
1709 */
Pekka Enberg06428782008-01-07 23:20:27 -08001710static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001711 struct page *page, void *x, void *addr)
1712{
1713 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001714 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001715 unsigned long flags;
1716
Christoph Lameter894b8782007-05-10 03:15:16 -07001717 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001718 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001719 debug_check_no_locks_freed(object, c->objsize);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001720 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1721 debug_check_no_obj_freed(object, s->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001722 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001723 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001724 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001725 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001726 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001727 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001728
1729 local_irq_restore(flags);
1730}
1731
Christoph Lameter81819f02007-05-06 14:49:36 -07001732void kmem_cache_free(struct kmem_cache *s, void *x)
1733{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001734 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001735
Christoph Lameterb49af682007-05-06 14:49:41 -07001736 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001737
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001738 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001739}
1740EXPORT_SYMBOL(kmem_cache_free);
1741
1742/* Figure out on which slab object the object resides */
1743static struct page *get_object_page(const void *x)
1744{
Christoph Lameterb49af682007-05-06 14:49:41 -07001745 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001746
1747 if (!PageSlab(page))
1748 return NULL;
1749
1750 return page;
1751}
1752
1753/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001754 * Object placement in a slab is made very easy because we always start at
1755 * offset 0. If we tune the size of the object to the alignment then we can
1756 * get the required alignment by putting one properly sized object after
1757 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001758 *
1759 * Notice that the allocation order determines the sizes of the per cpu
1760 * caches. Each processor has always one slab available for allocations.
1761 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001762 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001763 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001764 */
1765
1766/*
1767 * Mininum / Maximum order of slab pages. This influences locking overhead
1768 * and slab fragmentation. A higher order reduces the number of partial slabs
1769 * and increases the number of allocations possible without having to
1770 * take the list_lock.
1771 */
1772static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03001773static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001774static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001775
1776/*
1777 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001778 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001779 */
1780static int slub_nomerge;
1781
1782/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001783 * Calculate the order of allocation given an slab object size.
1784 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001785 * The order of allocation has significant impact on performance and other
1786 * system components. Generally order 0 allocations should be preferred since
1787 * order 0 does not cause fragmentation in the page allocator. Larger objects
1788 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001789 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07001790 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001791 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001792 * In order to reach satisfactory performance we must ensure that a minimum
1793 * number of objects is in one slab. Otherwise we may generate too much
1794 * activity on the partial lists which requires taking the list_lock. This is
1795 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001796 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001797 * slub_max_order specifies the order where we begin to stop considering the
1798 * number of objects in a slab as critical. If we reach slub_max_order then
1799 * we try to keep the page order as low as possible. So we accept more waste
1800 * of space in favor of a small page order.
1801 *
1802 * Higher order allocations also allow the placement of more objects in a
1803 * slab and thereby reduce object handling overhead. If the user has
1804 * requested a higher mininum order then we start with that one instead of
1805 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001806 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001807static inline int slab_order(int size, int min_objects,
1808 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001809{
1810 int order;
1811 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001812 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001813
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04001814 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1815 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03001816
Christoph Lameter6300ea72007-07-17 04:03:20 -07001817 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001818 fls(min_objects * size - 1) - PAGE_SHIFT);
1819 order <= max_order; order++) {
1820
Christoph Lameter81819f02007-05-06 14:49:36 -07001821 unsigned long slab_size = PAGE_SIZE << order;
1822
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001823 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001824 continue;
1825
Christoph Lameter81819f02007-05-06 14:49:36 -07001826 rem = slab_size % size;
1827
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001828 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001829 break;
1830
1831 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001832
Christoph Lameter81819f02007-05-06 14:49:36 -07001833 return order;
1834}
1835
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001836static inline int calculate_order(int size)
1837{
1838 int order;
1839 int min_objects;
1840 int fraction;
1841
1842 /*
1843 * Attempt to find best configuration for a slab. This
1844 * works by first attempting to generate a layout with
1845 * the best configuration and backing off gradually.
1846 *
1847 * First we reduce the acceptable waste in a slab. Then
1848 * we reduce the minimum objects required in a slab.
1849 */
1850 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001851 if (!min_objects)
1852 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001853 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001854 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001855 while (fraction >= 4) {
1856 order = slab_order(size, min_objects,
1857 slub_max_order, fraction);
1858 if (order <= slub_max_order)
1859 return order;
1860 fraction /= 2;
1861 }
1862 min_objects /= 2;
1863 }
1864
1865 /*
1866 * We were unable to place multiple objects in a slab. Now
1867 * lets see if we can place a single object there.
1868 */
1869 order = slab_order(size, 1, slub_max_order, 1);
1870 if (order <= slub_max_order)
1871 return order;
1872
1873 /*
1874 * Doh this slab cannot be placed using slub_max_order.
1875 */
1876 order = slab_order(size, 1, MAX_ORDER, 1);
1877 if (order <= MAX_ORDER)
1878 return order;
1879 return -ENOSYS;
1880}
1881
Christoph Lameter81819f02007-05-06 14:49:36 -07001882/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001883 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001884 */
1885static unsigned long calculate_alignment(unsigned long flags,
1886 unsigned long align, unsigned long size)
1887{
1888 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001889 * If the user wants hardware cache aligned objects then follow that
1890 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001891 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001892 * The hardware cache alignment cannot override the specified
1893 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001894 */
Nick Pigginb6210382008-03-05 14:05:56 -08001895 if (flags & SLAB_HWCACHE_ALIGN) {
1896 unsigned long ralign = cache_line_size();
1897 while (size <= ralign / 2)
1898 ralign /= 2;
1899 align = max(align, ralign);
1900 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001901
1902 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001903 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001904
1905 return ALIGN(align, sizeof(void *));
1906}
1907
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001908static void init_kmem_cache_cpu(struct kmem_cache *s,
1909 struct kmem_cache_cpu *c)
1910{
1911 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001912 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001913 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001914 c->offset = s->offset / sizeof(void *);
1915 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03001916#ifdef CONFIG_SLUB_STATS
1917 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1918#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001919}
1920
Pekka Enberg5595cff2008-08-05 09:28:47 +03001921static void
1922init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07001923{
1924 n->nr_partial = 0;
Pekka Enberg5595cff2008-08-05 09:28:47 +03001925
1926 /*
1927 * The larger the object size is, the more pages we want on the partial
1928 * list to avoid pounding the page allocator excessively.
1929 */
1930 n->min_partial = ilog2(s->size);
1931 if (n->min_partial < MIN_PARTIAL)
1932 n->min_partial = MIN_PARTIAL;
1933 else if (n->min_partial > MAX_PARTIAL)
1934 n->min_partial = MAX_PARTIAL;
1935
Christoph Lameter81819f02007-05-06 14:49:36 -07001936 spin_lock_init(&n->list_lock);
1937 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001938#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001939 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07001940 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07001941 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001942#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001943}
1944
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001945#ifdef CONFIG_SMP
1946/*
1947 * Per cpu array for per cpu structures.
1948 *
1949 * The per cpu array places all kmem_cache_cpu structures from one processor
1950 * close together meaning that it becomes possible that multiple per cpu
1951 * structures are contained in one cacheline. This may be particularly
1952 * beneficial for the kmalloc caches.
1953 *
1954 * A desktop system typically has around 60-80 slabs. With 100 here we are
1955 * likely able to get per cpu structures for all caches from the array defined
1956 * here. We must be able to cover all kmalloc caches during bootstrap.
1957 *
1958 * If the per cpu array is exhausted then fall back to kmalloc
1959 * of individual cachelines. No sharing is possible then.
1960 */
1961#define NR_KMEM_CACHE_CPU 100
1962
1963static DEFINE_PER_CPU(struct kmem_cache_cpu,
1964 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1965
1966static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1967static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1968
1969static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1970 int cpu, gfp_t flags)
1971{
1972 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1973
1974 if (c)
1975 per_cpu(kmem_cache_cpu_free, cpu) =
1976 (void *)c->freelist;
1977 else {
1978 /* Table overflow: So allocate ourselves */
1979 c = kmalloc_node(
1980 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1981 flags, cpu_to_node(cpu));
1982 if (!c)
1983 return NULL;
1984 }
1985
1986 init_kmem_cache_cpu(s, c);
1987 return c;
1988}
1989
1990static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1991{
1992 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1993 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1994 kfree(c);
1995 return;
1996 }
1997 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1998 per_cpu(kmem_cache_cpu_free, cpu) = c;
1999}
2000
2001static void free_kmem_cache_cpus(struct kmem_cache *s)
2002{
2003 int cpu;
2004
2005 for_each_online_cpu(cpu) {
2006 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2007
2008 if (c) {
2009 s->cpu_slab[cpu] = NULL;
2010 free_kmem_cache_cpu(c, cpu);
2011 }
2012 }
2013}
2014
2015static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2016{
2017 int cpu;
2018
2019 for_each_online_cpu(cpu) {
2020 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2021
2022 if (c)
2023 continue;
2024
2025 c = alloc_kmem_cache_cpu(s, cpu, flags);
2026 if (!c) {
2027 free_kmem_cache_cpus(s);
2028 return 0;
2029 }
2030 s->cpu_slab[cpu] = c;
2031 }
2032 return 1;
2033}
2034
2035/*
2036 * Initialize the per cpu array.
2037 */
2038static void init_alloc_cpu_cpu(int cpu)
2039{
2040 int i;
2041
2042 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2043 return;
2044
2045 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2046 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2047
2048 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2049}
2050
2051static void __init init_alloc_cpu(void)
2052{
2053 int cpu;
2054
2055 for_each_online_cpu(cpu)
2056 init_alloc_cpu_cpu(cpu);
2057 }
2058
2059#else
2060static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2061static inline void init_alloc_cpu(void) {}
2062
2063static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2064{
2065 init_kmem_cache_cpu(s, &s->cpu_slab);
2066 return 1;
2067}
2068#endif
2069
Christoph Lameter81819f02007-05-06 14:49:36 -07002070#ifdef CONFIG_NUMA
2071/*
2072 * No kmalloc_node yet so do it by hand. We know that this is the first
2073 * slab on the node for this slabcache. There are no concurrent accesses
2074 * possible.
2075 *
2076 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002077 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2078 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002079 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002080static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2081 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002082{
2083 struct page *page;
2084 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002085 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002086
2087 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2088
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002089 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002090
2091 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002092 if (page_to_nid(page) != node) {
2093 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2094 "node %d\n", node);
2095 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2096 "in order to be able to continue\n");
2097 }
2098
Christoph Lameter81819f02007-05-06 14:49:36 -07002099 n = page->freelist;
2100 BUG_ON(!n);
2101 page->freelist = get_freepointer(kmalloc_caches, n);
2102 page->inuse++;
2103 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002104#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002105 init_object(kmalloc_caches, n, 1);
2106 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002107#endif
Pekka Enberg5595cff2008-08-05 09:28:47 +03002108 init_kmem_cache_node(n, kmalloc_caches);
Christoph Lameter205ab992008-04-14 19:11:40 +03002109 inc_slabs_node(kmalloc_caches, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002110
rootba84c732008-01-07 23:20:28 -08002111 /*
2112 * lockdep requires consistent irq usage for each lock
2113 * so even though there cannot be a race this early in
2114 * the boot sequence, we still disable irqs.
2115 */
2116 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002117 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002118 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002119 return n;
2120}
2121
2122static void free_kmem_cache_nodes(struct kmem_cache *s)
2123{
2124 int node;
2125
Christoph Lameterf64dc582007-10-16 01:25:33 -07002126 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002127 struct kmem_cache_node *n = s->node[node];
2128 if (n && n != &s->local_node)
2129 kmem_cache_free(kmalloc_caches, n);
2130 s->node[node] = NULL;
2131 }
2132}
2133
2134static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2135{
2136 int node;
2137 int local_node;
2138
2139 if (slab_state >= UP)
2140 local_node = page_to_nid(virt_to_page(s));
2141 else
2142 local_node = 0;
2143
Christoph Lameterf64dc582007-10-16 01:25:33 -07002144 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002145 struct kmem_cache_node *n;
2146
2147 if (local_node == node)
2148 n = &s->local_node;
2149 else {
2150 if (slab_state == DOWN) {
2151 n = early_kmem_cache_node_alloc(gfpflags,
2152 node);
2153 continue;
2154 }
2155 n = kmem_cache_alloc_node(kmalloc_caches,
2156 gfpflags, node);
2157
2158 if (!n) {
2159 free_kmem_cache_nodes(s);
2160 return 0;
2161 }
2162
2163 }
2164 s->node[node] = n;
Pekka Enberg5595cff2008-08-05 09:28:47 +03002165 init_kmem_cache_node(n, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002166 }
2167 return 1;
2168}
2169#else
2170static void free_kmem_cache_nodes(struct kmem_cache *s)
2171{
2172}
2173
2174static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2175{
Pekka Enberg5595cff2008-08-05 09:28:47 +03002176 init_kmem_cache_node(&s->local_node, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002177 return 1;
2178}
2179#endif
2180
2181/*
2182 * calculate_sizes() determines the order and the distribution of data within
2183 * a slab object.
2184 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002185static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002186{
2187 unsigned long flags = s->flags;
2188 unsigned long size = s->objsize;
2189 unsigned long align = s->align;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002190 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002191
2192 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002193 * Round up object size to the next word boundary. We can only
2194 * place the free pointer at word boundaries and this determines
2195 * the possible location of the free pointer.
2196 */
2197 size = ALIGN(size, sizeof(void *));
2198
2199#ifdef CONFIG_SLUB_DEBUG
2200 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002201 * Determine if we can poison the object itself. If the user of
2202 * the slab may touch the object after free or before allocation
2203 * then we should never poison the object itself.
2204 */
2205 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002206 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002207 s->flags |= __OBJECT_POISON;
2208 else
2209 s->flags &= ~__OBJECT_POISON;
2210
Christoph Lameter81819f02007-05-06 14:49:36 -07002211
2212 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002213 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002214 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002215 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002216 */
2217 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2218 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002219#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002220
2221 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002222 * With that we have determined the number of bytes in actual use
2223 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002224 */
2225 s->inuse = size;
2226
2227 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002228 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002229 /*
2230 * Relocate free pointer after the object if it is not
2231 * permitted to overwrite the first word of the object on
2232 * kmem_cache_free.
2233 *
2234 * This is the case if we do RCU, have a constructor or
2235 * destructor or are poisoning the objects.
2236 */
2237 s->offset = size;
2238 size += sizeof(void *);
2239 }
2240
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002241#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002242 if (flags & SLAB_STORE_USER)
2243 /*
2244 * Need to store information about allocs and frees after
2245 * the object.
2246 */
2247 size += 2 * sizeof(struct track);
2248
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002249 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002250 /*
2251 * Add some empty padding so that we can catch
2252 * overwrites from earlier objects rather than let
2253 * tracking information or the free pointer be
2254 * corrupted if an user writes before the start
2255 * of the object.
2256 */
2257 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002258#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002259
Christoph Lameter81819f02007-05-06 14:49:36 -07002260 /*
2261 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002262 * user specified and the dynamic determination of cache line size
2263 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002264 */
2265 align = calculate_alignment(flags, align, s->objsize);
2266
2267 /*
2268 * SLUB stores one object immediately after another beginning from
2269 * offset 0. In order to align the objects we have to simply size
2270 * each object to conform to the alignment.
2271 */
2272 size = ALIGN(size, align);
2273 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03002274 if (forced_order >= 0)
2275 order = forced_order;
2276 else
2277 order = calculate_order(size);
Christoph Lameter81819f02007-05-06 14:49:36 -07002278
Christoph Lameter834f3d12008-04-14 19:11:31 +03002279 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07002280 return 0;
2281
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002282 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002283 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002284 s->allocflags |= __GFP_COMP;
2285
2286 if (s->flags & SLAB_CACHE_DMA)
2287 s->allocflags |= SLUB_DMA;
2288
2289 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2290 s->allocflags |= __GFP_RECLAIMABLE;
2291
Christoph Lameter81819f02007-05-06 14:49:36 -07002292 /*
2293 * Determine the number of objects per slab
2294 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002295 s->oo = oo_make(order, size);
Christoph Lameter65c33762008-04-14 19:11:40 +03002296 s->min = oo_make(get_order(size), size);
Christoph Lameter205ab992008-04-14 19:11:40 +03002297 if (oo_objects(s->oo) > oo_objects(s->max))
2298 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002299
Christoph Lameter834f3d12008-04-14 19:11:31 +03002300 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07002301
2302}
2303
Christoph Lameter81819f02007-05-06 14:49:36 -07002304static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2305 const char *name, size_t size,
2306 size_t align, unsigned long flags,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002307 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002308{
2309 memset(s, 0, kmem_size);
2310 s->name = name;
2311 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002312 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002313 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002314 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002315
Christoph Lameter06b285d2008-04-14 19:11:41 +03002316 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07002317 goto error;
2318
2319 s->refcount = 1;
2320#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05002321 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07002322#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002323 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2324 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002325
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002326 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002327 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002328 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002329error:
2330 if (flags & SLAB_PANIC)
2331 panic("Cannot create slab %s size=%lu realsize=%u "
2332 "order=%u offset=%u flags=%lx\n",
Christoph Lameter834f3d12008-04-14 19:11:31 +03002333 s->name, (unsigned long)size, s->size, oo_order(s->oo),
Christoph Lameter81819f02007-05-06 14:49:36 -07002334 s->offset, flags);
2335 return 0;
2336}
Christoph Lameter81819f02007-05-06 14:49:36 -07002337
2338/*
2339 * Check if a given pointer is valid
2340 */
2341int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2342{
Pekka Enberg06428782008-01-07 23:20:27 -08002343 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002344
2345 page = get_object_page(object);
2346
2347 if (!page || s != page->slab)
2348 /* No slab or wrong slab */
2349 return 0;
2350
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002351 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002352 return 0;
2353
2354 /*
2355 * We could also check if the object is on the slabs freelist.
2356 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002357 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002358 * to a certain slab.
2359 */
2360 return 1;
2361}
2362EXPORT_SYMBOL(kmem_ptr_validate);
2363
2364/*
2365 * Determine the size of a slab object
2366 */
2367unsigned int kmem_cache_size(struct kmem_cache *s)
2368{
2369 return s->objsize;
2370}
2371EXPORT_SYMBOL(kmem_cache_size);
2372
2373const char *kmem_cache_name(struct kmem_cache *s)
2374{
2375 return s->name;
2376}
2377EXPORT_SYMBOL(kmem_cache_name);
2378
Christoph Lameter33b12c32008-04-25 12:22:43 -07002379static void list_slab_objects(struct kmem_cache *s, struct page *page,
2380 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07002381{
Christoph Lameter33b12c32008-04-25 12:22:43 -07002382#ifdef CONFIG_SLUB_DEBUG
2383 void *addr = page_address(page);
2384 void *p;
2385 DECLARE_BITMAP(map, page->objects);
2386
2387 bitmap_zero(map, page->objects);
2388 slab_err(s, page, "%s", text);
2389 slab_lock(page);
2390 for_each_free_object(p, s, page->freelist)
2391 set_bit(slab_index(p, s, addr), map);
2392
2393 for_each_object(p, s, addr, page->objects) {
2394
2395 if (!test_bit(slab_index(p, s, addr), map)) {
2396 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2397 p, p - addr);
2398 print_tracking(s, p);
2399 }
2400 }
2401 slab_unlock(page);
2402#endif
2403}
2404
Christoph Lameter81819f02007-05-06 14:49:36 -07002405/*
Christoph Lameter599870b2008-04-23 12:36:52 -07002406 * Attempt to free all partial slabs on a node.
Christoph Lameter81819f02007-05-06 14:49:36 -07002407 */
Christoph Lameter599870b2008-04-23 12:36:52 -07002408static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002409{
Christoph Lameter81819f02007-05-06 14:49:36 -07002410 unsigned long flags;
2411 struct page *page, *h;
2412
2413 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter33b12c32008-04-25 12:22:43 -07002414 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002415 if (!page->inuse) {
2416 list_del(&page->lru);
2417 discard_slab(s, page);
Christoph Lameter599870b2008-04-23 12:36:52 -07002418 n->nr_partial--;
Christoph Lameter33b12c32008-04-25 12:22:43 -07002419 } else {
2420 list_slab_objects(s, page,
2421 "Objects remaining on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07002422 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07002423 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002424 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002425}
2426
2427/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002428 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002429 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002430static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002431{
2432 int node;
2433
2434 flush_all(s);
2435
2436 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002437 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002438 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002439 struct kmem_cache_node *n = get_node(s, node);
2440
Christoph Lameter599870b2008-04-23 12:36:52 -07002441 free_partial(s, n);
2442 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002443 return 1;
2444 }
2445 free_kmem_cache_nodes(s);
2446 return 0;
2447}
2448
2449/*
2450 * Close a cache and release the kmem_cache structure
2451 * (must be used for caches created using kmem_cache_create)
2452 */
2453void kmem_cache_destroy(struct kmem_cache *s)
2454{
2455 down_write(&slub_lock);
2456 s->refcount--;
2457 if (!s->refcount) {
2458 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002459 up_write(&slub_lock);
Pekka Enbergd629d812008-04-23 22:31:08 +03002460 if (kmem_cache_close(s)) {
2461 printk(KERN_ERR "SLUB %s: %s called for cache that "
2462 "still has objects.\n", s->name, __func__);
2463 dump_stack();
2464 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002465 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002466 } else
2467 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002468}
2469EXPORT_SYMBOL(kmem_cache_destroy);
2470
2471/********************************************************************
2472 * Kmalloc subsystem
2473 *******************************************************************/
2474
Christoph Lameter331dc552008-02-14 14:28:09 -08002475struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002476EXPORT_SYMBOL(kmalloc_caches);
2477
Christoph Lameter81819f02007-05-06 14:49:36 -07002478static int __init setup_slub_min_order(char *str)
2479{
Pekka Enberg06428782008-01-07 23:20:27 -08002480 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002481
2482 return 1;
2483}
2484
2485__setup("slub_min_order=", setup_slub_min_order);
2486
2487static int __init setup_slub_max_order(char *str)
2488{
Pekka Enberg06428782008-01-07 23:20:27 -08002489 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002490
2491 return 1;
2492}
2493
2494__setup("slub_max_order=", setup_slub_max_order);
2495
2496static int __init setup_slub_min_objects(char *str)
2497{
Pekka Enberg06428782008-01-07 23:20:27 -08002498 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002499
2500 return 1;
2501}
2502
2503__setup("slub_min_objects=", setup_slub_min_objects);
2504
2505static int __init setup_slub_nomerge(char *str)
2506{
2507 slub_nomerge = 1;
2508 return 1;
2509}
2510
2511__setup("slub_nomerge", setup_slub_nomerge);
2512
Christoph Lameter81819f02007-05-06 14:49:36 -07002513static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2514 const char *name, int size, gfp_t gfp_flags)
2515{
2516 unsigned int flags = 0;
2517
2518 if (gfp_flags & SLUB_DMA)
2519 flags = SLAB_CACHE_DMA;
2520
2521 down_write(&slub_lock);
2522 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter319d1e22008-04-14 19:11:41 +03002523 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002524 goto panic;
2525
2526 list_add(&s->list, &slab_caches);
2527 up_write(&slub_lock);
2528 if (sysfs_slab_add(s))
2529 goto panic;
2530 return s;
2531
2532panic:
2533 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2534}
2535
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002536#ifdef CONFIG_ZONE_DMA
Christoph Lameter4097d602008-04-14 18:51:18 +03002537static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002538
2539static void sysfs_add_func(struct work_struct *w)
2540{
2541 struct kmem_cache *s;
2542
2543 down_write(&slub_lock);
2544 list_for_each_entry(s, &slab_caches, list) {
2545 if (s->flags & __SYSFS_ADD_DEFERRED) {
2546 s->flags &= ~__SYSFS_ADD_DEFERRED;
2547 sysfs_slab_add(s);
2548 }
2549 }
2550 up_write(&slub_lock);
2551}
2552
2553static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2554
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002555static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2556{
2557 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002558 char *text;
2559 size_t realsize;
2560
2561 s = kmalloc_caches_dma[index];
2562 if (s)
2563 return s;
2564
2565 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002566 if (flags & __GFP_WAIT)
2567 down_write(&slub_lock);
2568 else {
2569 if (!down_write_trylock(&slub_lock))
2570 goto out;
2571 }
2572
2573 if (kmalloc_caches_dma[index])
2574 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002575
Christoph Lameter7b55f622007-07-17 04:03:27 -07002576 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002577 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2578 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002579 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2580
2581 if (!s || !text || !kmem_cache_open(s, flags, text,
2582 realsize, ARCH_KMALLOC_MINALIGN,
2583 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2584 kfree(s);
2585 kfree(text);
2586 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002587 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002588
2589 list_add(&s->list, &slab_caches);
2590 kmalloc_caches_dma[index] = s;
2591
2592 schedule_work(&sysfs_add_work);
2593
2594unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002595 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002596out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002597 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002598}
2599#endif
2600
Christoph Lameterf1b26332007-07-17 04:03:26 -07002601/*
2602 * Conversion table for small slabs sizes / 8 to the index in the
2603 * kmalloc array. This is necessary for slabs < 192 since we have non power
2604 * of two cache sizes there. The size of larger slabs can be determined using
2605 * fls.
2606 */
2607static s8 size_index[24] = {
2608 3, /* 8 */
2609 4, /* 16 */
2610 5, /* 24 */
2611 5, /* 32 */
2612 6, /* 40 */
2613 6, /* 48 */
2614 6, /* 56 */
2615 6, /* 64 */
2616 1, /* 72 */
2617 1, /* 80 */
2618 1, /* 88 */
2619 1, /* 96 */
2620 7, /* 104 */
2621 7, /* 112 */
2622 7, /* 120 */
2623 7, /* 128 */
2624 2, /* 136 */
2625 2, /* 144 */
2626 2, /* 152 */
2627 2, /* 160 */
2628 2, /* 168 */
2629 2, /* 176 */
2630 2, /* 184 */
2631 2 /* 192 */
2632};
2633
Christoph Lameter81819f02007-05-06 14:49:36 -07002634static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2635{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002636 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002637
Christoph Lameterf1b26332007-07-17 04:03:26 -07002638 if (size <= 192) {
2639 if (!size)
2640 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002641
Christoph Lameterf1b26332007-07-17 04:03:26 -07002642 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002643 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002644 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002645
2646#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002647 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002648 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002649
Christoph Lameter81819f02007-05-06 14:49:36 -07002650#endif
2651 return &kmalloc_caches[index];
2652}
2653
2654void *__kmalloc(size_t size, gfp_t flags)
2655{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002656 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002657
Christoph Lameter331dc552008-02-14 14:28:09 -08002658 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002659 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002660
2661 s = get_slab(size, flags);
2662
2663 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002664 return s;
2665
Christoph Lameterce15fea2007-07-17 04:03:28 -07002666 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002667}
2668EXPORT_SYMBOL(__kmalloc);
2669
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002670static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2671{
2672 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2673 get_order(size));
2674
2675 if (page)
2676 return page_address(page);
2677 else
2678 return NULL;
2679}
2680
Christoph Lameter81819f02007-05-06 14:49:36 -07002681#ifdef CONFIG_NUMA
2682void *__kmalloc_node(size_t size, gfp_t flags, int node)
2683{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002684 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002685
Christoph Lameter331dc552008-02-14 14:28:09 -08002686 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002687 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002688
2689 s = get_slab(size, flags);
2690
2691 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002692 return s;
2693
Christoph Lameterce15fea2007-07-17 04:03:28 -07002694 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002695}
2696EXPORT_SYMBOL(__kmalloc_node);
2697#endif
2698
2699size_t ksize(const void *object)
2700{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002701 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002702 struct kmem_cache *s;
2703
Christoph Lameteref8b4522007-10-16 01:24:46 -07002704 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002705 return 0;
2706
Vegard Nossum294a80a2007-12-04 23:45:30 -08002707 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002708
Pekka Enberg76994412008-05-22 19:22:25 +03002709 if (unlikely(!PageSlab(page))) {
2710 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08002711 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03002712 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002713 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002714
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002715#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002716 /*
2717 * Debugging requires use of the padding between object
2718 * and whatever may come after it.
2719 */
2720 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2721 return s->objsize;
2722
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002723#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002724 /*
2725 * If we have the need to store the freelist pointer
2726 * back there or track user information then we can
2727 * only use the space before that information.
2728 */
2729 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2730 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002731 /*
2732 * Else we can use all the padding etc for the allocation
2733 */
2734 return s->size;
2735}
Christoph Lameter81819f02007-05-06 14:49:36 -07002736
2737void kfree(const void *x)
2738{
Christoph Lameter81819f02007-05-06 14:49:36 -07002739 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002740 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002741
Satyam Sharma2408c552007-10-16 01:24:44 -07002742 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002743 return;
2744
Christoph Lameterb49af682007-05-06 14:49:41 -07002745 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002746 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07002747 BUG_ON(!PageCompound(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002748 put_page(page);
2749 return;
2750 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002751 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002752}
2753EXPORT_SYMBOL(kfree);
2754
Christoph Lameter2086d262007-05-06 14:49:46 -07002755/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002756 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2757 * the remaining slabs by the number of items in use. The slabs with the
2758 * most items in use come first. New allocations will then fill those up
2759 * and thus they can be removed from the partial lists.
2760 *
2761 * The slabs with the least items are placed last. This results in them
2762 * being allocated from last increasing the chance that the last objects
2763 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002764 */
2765int kmem_cache_shrink(struct kmem_cache *s)
2766{
2767 int node;
2768 int i;
2769 struct kmem_cache_node *n;
2770 struct page *page;
2771 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03002772 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07002773 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03002774 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07002775 unsigned long flags;
2776
2777 if (!slabs_by_inuse)
2778 return -ENOMEM;
2779
2780 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002781 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002782 n = get_node(s, node);
2783
2784 if (!n->nr_partial)
2785 continue;
2786
Christoph Lameter834f3d12008-04-14 19:11:31 +03002787 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07002788 INIT_LIST_HEAD(slabs_by_inuse + i);
2789
2790 spin_lock_irqsave(&n->list_lock, flags);
2791
2792 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002793 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002794 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002795 * Note that concurrent frees may occur while we hold the
2796 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002797 */
2798 list_for_each_entry_safe(page, t, &n->partial, lru) {
2799 if (!page->inuse && slab_trylock(page)) {
2800 /*
2801 * Must hold slab lock here because slab_free
2802 * may have freed the last object and be
2803 * waiting to release the slab.
2804 */
2805 list_del(&page->lru);
2806 n->nr_partial--;
2807 slab_unlock(page);
2808 discard_slab(s, page);
2809 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002810 list_move(&page->lru,
2811 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002812 }
2813 }
2814
Christoph Lameter2086d262007-05-06 14:49:46 -07002815 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002816 * Rebuild the partial list with the slabs filled up most
2817 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002818 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002819 for (i = objects - 1; i >= 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07002820 list_splice(slabs_by_inuse + i, n->partial.prev);
2821
Christoph Lameter2086d262007-05-06 14:49:46 -07002822 spin_unlock_irqrestore(&n->list_lock, flags);
2823 }
2824
2825 kfree(slabs_by_inuse);
2826 return 0;
2827}
2828EXPORT_SYMBOL(kmem_cache_shrink);
2829
Yasunori Gotob9049e22007-10-21 16:41:37 -07002830#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2831static int slab_mem_going_offline_callback(void *arg)
2832{
2833 struct kmem_cache *s;
2834
2835 down_read(&slub_lock);
2836 list_for_each_entry(s, &slab_caches, list)
2837 kmem_cache_shrink(s);
2838 up_read(&slub_lock);
2839
2840 return 0;
2841}
2842
2843static void slab_mem_offline_callback(void *arg)
2844{
2845 struct kmem_cache_node *n;
2846 struct kmem_cache *s;
2847 struct memory_notify *marg = arg;
2848 int offline_node;
2849
2850 offline_node = marg->status_change_nid;
2851
2852 /*
2853 * If the node still has available memory. we need kmem_cache_node
2854 * for it yet.
2855 */
2856 if (offline_node < 0)
2857 return;
2858
2859 down_read(&slub_lock);
2860 list_for_each_entry(s, &slab_caches, list) {
2861 n = get_node(s, offline_node);
2862 if (n) {
2863 /*
2864 * if n->nr_slabs > 0, slabs still exist on the node
2865 * that is going down. We were unable to free them,
2866 * and offline_pages() function shoudn't call this
2867 * callback. So, we must fail.
2868 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002869 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002870
2871 s->node[offline_node] = NULL;
2872 kmem_cache_free(kmalloc_caches, n);
2873 }
2874 }
2875 up_read(&slub_lock);
2876}
2877
2878static int slab_mem_going_online_callback(void *arg)
2879{
2880 struct kmem_cache_node *n;
2881 struct kmem_cache *s;
2882 struct memory_notify *marg = arg;
2883 int nid = marg->status_change_nid;
2884 int ret = 0;
2885
2886 /*
2887 * If the node's memory is already available, then kmem_cache_node is
2888 * already created. Nothing to do.
2889 */
2890 if (nid < 0)
2891 return 0;
2892
2893 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07002894 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07002895 * allocate a kmem_cache_node structure in order to bring the node
2896 * online.
2897 */
2898 down_read(&slub_lock);
2899 list_for_each_entry(s, &slab_caches, list) {
2900 /*
2901 * XXX: kmem_cache_alloc_node will fallback to other nodes
2902 * since memory is not yet available from the node that
2903 * is brought up.
2904 */
2905 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2906 if (!n) {
2907 ret = -ENOMEM;
2908 goto out;
2909 }
Pekka Enberg5595cff2008-08-05 09:28:47 +03002910 init_kmem_cache_node(n, s);
Yasunori Gotob9049e22007-10-21 16:41:37 -07002911 s->node[nid] = n;
2912 }
2913out:
2914 up_read(&slub_lock);
2915 return ret;
2916}
2917
2918static int slab_memory_callback(struct notifier_block *self,
2919 unsigned long action, void *arg)
2920{
2921 int ret = 0;
2922
2923 switch (action) {
2924 case MEM_GOING_ONLINE:
2925 ret = slab_mem_going_online_callback(arg);
2926 break;
2927 case MEM_GOING_OFFLINE:
2928 ret = slab_mem_going_offline_callback(arg);
2929 break;
2930 case MEM_OFFLINE:
2931 case MEM_CANCEL_ONLINE:
2932 slab_mem_offline_callback(arg);
2933 break;
2934 case MEM_ONLINE:
2935 case MEM_CANCEL_OFFLINE:
2936 break;
2937 }
2938
2939 ret = notifier_from_errno(ret);
2940 return ret;
2941}
2942
2943#endif /* CONFIG_MEMORY_HOTPLUG */
2944
Christoph Lameter81819f02007-05-06 14:49:36 -07002945/********************************************************************
2946 * Basic setup of slabs
2947 *******************************************************************/
2948
2949void __init kmem_cache_init(void)
2950{
2951 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002952 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002953
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002954 init_alloc_cpu();
2955
Christoph Lameter81819f02007-05-06 14:49:36 -07002956#ifdef CONFIG_NUMA
2957 /*
2958 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002959 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002960 * kmem_cache_open for slab_state == DOWN.
2961 */
2962 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2963 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002964 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002965 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002966
Nadia Derbey0c40ba42008-04-29 01:00:41 -07002967 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
Christoph Lameter81819f02007-05-06 14:49:36 -07002968#endif
2969
2970 /* Able to allocate the per node structures */
2971 slab_state = PARTIAL;
2972
2973 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002974 if (KMALLOC_MIN_SIZE <= 64) {
2975 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002976 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002977 caches++;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002978 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002979 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002980 caches++;
2981 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002982
Christoph Lameter331dc552008-02-14 14:28:09 -08002983 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002984 create_kmalloc_cache(&kmalloc_caches[i],
2985 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002986 caches++;
2987 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002988
Christoph Lameterf1b26332007-07-17 04:03:26 -07002989
2990 /*
2991 * Patch up the size_index table if we have strange large alignment
2992 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002993 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002994 *
2995 * Largest permitted alignment is 256 bytes due to the way we
2996 * handle the index determination for the smaller caches.
2997 *
2998 * Make sure that nothing crazy happens if someone starts tinkering
2999 * around with ARCH_KMALLOC_MINALIGN
3000 */
3001 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3002 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3003
Christoph Lameter12ad6842007-07-17 04:03:28 -07003004 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07003005 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3006
Christoph Lameter41d54d32008-07-03 09:14:26 -05003007 if (KMALLOC_MIN_SIZE == 128) {
3008 /*
3009 * The 192 byte sized cache is not used if the alignment
3010 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3011 * instead.
3012 */
3013 for (i = 128 + 8; i <= 192; i += 8)
3014 size_index[(i - 1) / 8] = 8;
3015 }
3016
Christoph Lameter81819f02007-05-06 14:49:36 -07003017 slab_state = UP;
3018
3019 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08003020 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07003021 kmalloc_caches[i]. name =
3022 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3023
3024#ifdef CONFIG_SMP
3025 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003026 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3027 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3028#else
3029 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003030#endif
3031
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003032 printk(KERN_INFO
3033 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003034 " CPUs=%d, Nodes=%d\n",
3035 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003036 slub_min_order, slub_max_order, slub_min_objects,
3037 nr_cpu_ids, nr_node_ids);
3038}
3039
3040/*
3041 * Find a mergeable slab cache
3042 */
3043static int slab_unmergeable(struct kmem_cache *s)
3044{
3045 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3046 return 1;
3047
Christoph Lameterc59def92007-05-16 22:10:50 -07003048 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003049 return 1;
3050
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003051 /*
3052 * We may have set a slab to be unmergeable during bootstrap.
3053 */
3054 if (s->refcount < 0)
3055 return 1;
3056
Christoph Lameter81819f02007-05-06 14:49:36 -07003057 return 0;
3058}
3059
3060static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003061 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003062 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003063{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003064 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003065
3066 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3067 return NULL;
3068
Christoph Lameterc59def92007-05-16 22:10:50 -07003069 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003070 return NULL;
3071
3072 size = ALIGN(size, sizeof(void *));
3073 align = calculate_alignment(flags, align, size);
3074 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003075 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003076
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003077 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003078 if (slab_unmergeable(s))
3079 continue;
3080
3081 if (size > s->size)
3082 continue;
3083
Christoph Lameterba0268a2007-09-11 15:24:11 -07003084 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003085 continue;
3086 /*
3087 * Check if alignment is compatible.
3088 * Courtesy of Adrian Drzewiecki
3089 */
Pekka Enberg06428782008-01-07 23:20:27 -08003090 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003091 continue;
3092
3093 if (s->size - size >= sizeof(void *))
3094 continue;
3095
3096 return s;
3097 }
3098 return NULL;
3099}
3100
3101struct kmem_cache *kmem_cache_create(const char *name, size_t size,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003102 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003103{
3104 struct kmem_cache *s;
3105
3106 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003107 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003108 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003109 int cpu;
3110
Christoph Lameter81819f02007-05-06 14:49:36 -07003111 s->refcount++;
3112 /*
3113 * Adjust the object sizes so that we clear
3114 * the complete object on kzalloc.
3115 */
3116 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003117
3118 /*
3119 * And then we need to update the object size in the
3120 * per cpu structures
3121 */
3122 for_each_online_cpu(cpu)
3123 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003124
Christoph Lameter81819f02007-05-06 14:49:36 -07003125 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003126 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003127
Christoph Lameter81819f02007-05-06 14:49:36 -07003128 if (sysfs_slab_alias(s, name))
3129 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003130 return s;
3131 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003132
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003133 s = kmalloc(kmem_size, GFP_KERNEL);
3134 if (s) {
3135 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003136 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003137 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003138 up_write(&slub_lock);
3139 if (sysfs_slab_add(s))
3140 goto err;
3141 return s;
3142 }
3143 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003144 }
3145 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003146
3147err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003148 if (flags & SLAB_PANIC)
3149 panic("Cannot create slabcache %s\n", name);
3150 else
3151 s = NULL;
3152 return s;
3153}
3154EXPORT_SYMBOL(kmem_cache_create);
3155
Christoph Lameter81819f02007-05-06 14:49:36 -07003156#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003157/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003158 * Use the cpu notifier to insure that the cpu slabs are flushed when
3159 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003160 */
3161static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3162 unsigned long action, void *hcpu)
3163{
3164 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003165 struct kmem_cache *s;
3166 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003167
3168 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003169 case CPU_UP_PREPARE:
3170 case CPU_UP_PREPARE_FROZEN:
3171 init_alloc_cpu_cpu(cpu);
3172 down_read(&slub_lock);
3173 list_for_each_entry(s, &slab_caches, list)
3174 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3175 GFP_KERNEL);
3176 up_read(&slub_lock);
3177 break;
3178
Christoph Lameter81819f02007-05-06 14:49:36 -07003179 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003180 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003181 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003182 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003183 down_read(&slub_lock);
3184 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003185 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3186
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003187 local_irq_save(flags);
3188 __flush_cpu_slab(s, cpu);
3189 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003190 free_kmem_cache_cpu(c, cpu);
3191 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003192 }
3193 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003194 break;
3195 default:
3196 break;
3197 }
3198 return NOTIFY_OK;
3199}
3200
Pekka Enberg06428782008-01-07 23:20:27 -08003201static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003202 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003203};
Christoph Lameter81819f02007-05-06 14:49:36 -07003204
3205#endif
3206
Christoph Lameter81819f02007-05-06 14:49:36 -07003207void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3208{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003209 struct kmem_cache *s;
3210
Christoph Lameter331dc552008-02-14 14:28:09 -08003211 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003212 return kmalloc_large(size, gfpflags);
3213
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003214 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003215
Satyam Sharma2408c552007-10-16 01:24:44 -07003216 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003217 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003218
Christoph Lameterce15fea2007-07-17 04:03:28 -07003219 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003220}
3221
3222void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3223 int node, void *caller)
3224{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003225 struct kmem_cache *s;
3226
Christoph Lameter331dc552008-02-14 14:28:09 -08003227 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003228 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003229
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003230 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003231
Satyam Sharma2408c552007-10-16 01:24:44 -07003232 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003233 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003234
Christoph Lameterce15fea2007-07-17 04:03:28 -07003235 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003236}
3237
Christoph Lameterf6acb632008-04-29 16:16:06 -07003238#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03003239static unsigned long count_partial(struct kmem_cache_node *n,
3240 int (*get_count)(struct page *))
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003241{
3242 unsigned long flags;
3243 unsigned long x = 0;
3244 struct page *page;
3245
3246 spin_lock_irqsave(&n->list_lock, flags);
3247 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter205ab992008-04-14 19:11:40 +03003248 x += get_count(page);
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003249 spin_unlock_irqrestore(&n->list_lock, flags);
3250 return x;
3251}
Christoph Lameter205ab992008-04-14 19:11:40 +03003252
3253static int count_inuse(struct page *page)
3254{
3255 return page->inuse;
3256}
3257
3258static int count_total(struct page *page)
3259{
3260 return page->objects;
3261}
3262
3263static int count_free(struct page *page)
3264{
3265 return page->objects - page->inuse;
3266}
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003267
Christoph Lameter434e2452007-07-17 04:03:30 -07003268static int validate_slab(struct kmem_cache *s, struct page *page,
3269 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003270{
3271 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003272 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003273
3274 if (!check_slab(s, page) ||
3275 !on_freelist(s, page, NULL))
3276 return 0;
3277
3278 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003279 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003280
Christoph Lameter7656c722007-05-09 02:32:40 -07003281 for_each_free_object(p, s, page->freelist) {
3282 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003283 if (!check_object(s, page, p, 0))
3284 return 0;
3285 }
3286
Christoph Lameter224a88b2008-04-14 19:11:31 +03003287 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003288 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003289 if (!check_object(s, page, p, 1))
3290 return 0;
3291 return 1;
3292}
3293
Christoph Lameter434e2452007-07-17 04:03:30 -07003294static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3295 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003296{
3297 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003298 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003299 slab_unlock(page);
3300 } else
3301 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3302 s->name, page);
3303
3304 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003305 if (!PageSlubDebug(page))
3306 printk(KERN_ERR "SLUB %s: SlubDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003307 "on slab 0x%p\n", s->name, page);
3308 } else {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003309 if (PageSlubDebug(page))
3310 printk(KERN_ERR "SLUB %s: SlubDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003311 "slab 0x%p\n", s->name, page);
3312 }
3313}
3314
Christoph Lameter434e2452007-07-17 04:03:30 -07003315static int validate_slab_node(struct kmem_cache *s,
3316 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003317{
3318 unsigned long count = 0;
3319 struct page *page;
3320 unsigned long flags;
3321
3322 spin_lock_irqsave(&n->list_lock, flags);
3323
3324 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003325 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003326 count++;
3327 }
3328 if (count != n->nr_partial)
3329 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3330 "counter=%ld\n", s->name, count, n->nr_partial);
3331
3332 if (!(s->flags & SLAB_STORE_USER))
3333 goto out;
3334
3335 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003336 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003337 count++;
3338 }
3339 if (count != atomic_long_read(&n->nr_slabs))
3340 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3341 "counter=%ld\n", s->name, count,
3342 atomic_long_read(&n->nr_slabs));
3343
3344out:
3345 spin_unlock_irqrestore(&n->list_lock, flags);
3346 return count;
3347}
3348
Christoph Lameter434e2452007-07-17 04:03:30 -07003349static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003350{
3351 int node;
3352 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003353 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003354 sizeof(unsigned long), GFP_KERNEL);
3355
3356 if (!map)
3357 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003358
3359 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003360 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003361 struct kmem_cache_node *n = get_node(s, node);
3362
Christoph Lameter434e2452007-07-17 04:03:30 -07003363 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003364 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003365 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003366 return count;
3367}
3368
Christoph Lameterb3459702007-05-09 02:32:41 -07003369#ifdef SLUB_RESILIENCY_TEST
3370static void resiliency_test(void)
3371{
3372 u8 *p;
3373
3374 printk(KERN_ERR "SLUB resiliency testing\n");
3375 printk(KERN_ERR "-----------------------\n");
3376 printk(KERN_ERR "A. Corruption after allocation\n");
3377
3378 p = kzalloc(16, GFP_KERNEL);
3379 p[16] = 0x12;
3380 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3381 " 0x12->0x%p\n\n", p + 16);
3382
3383 validate_slab_cache(kmalloc_caches + 4);
3384
3385 /* Hmmm... The next two are dangerous */
3386 p = kzalloc(32, GFP_KERNEL);
3387 p[32 + sizeof(void *)] = 0x34;
3388 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003389 " 0x34 -> -0x%p\n", p);
3390 printk(KERN_ERR
3391 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003392
3393 validate_slab_cache(kmalloc_caches + 5);
3394 p = kzalloc(64, GFP_KERNEL);
3395 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3396 *p = 0x56;
3397 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3398 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003399 printk(KERN_ERR
3400 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003401 validate_slab_cache(kmalloc_caches + 6);
3402
3403 printk(KERN_ERR "\nB. Corruption after free\n");
3404 p = kzalloc(128, GFP_KERNEL);
3405 kfree(p);
3406 *p = 0x78;
3407 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3408 validate_slab_cache(kmalloc_caches + 7);
3409
3410 p = kzalloc(256, GFP_KERNEL);
3411 kfree(p);
3412 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003413 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3414 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003415 validate_slab_cache(kmalloc_caches + 8);
3416
3417 p = kzalloc(512, GFP_KERNEL);
3418 kfree(p);
3419 p[512] = 0xab;
3420 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3421 validate_slab_cache(kmalloc_caches + 9);
3422}
3423#else
3424static void resiliency_test(void) {};
3425#endif
3426
Christoph Lameter88a420e2007-05-06 14:49:45 -07003427/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003428 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003429 * and freed.
3430 */
3431
3432struct location {
3433 unsigned long count;
3434 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003435 long long sum_time;
3436 long min_time;
3437 long max_time;
3438 long min_pid;
3439 long max_pid;
3440 cpumask_t cpus;
3441 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003442};
3443
3444struct loc_track {
3445 unsigned long max;
3446 unsigned long count;
3447 struct location *loc;
3448};
3449
3450static void free_loc_track(struct loc_track *t)
3451{
3452 if (t->max)
3453 free_pages((unsigned long)t->loc,
3454 get_order(sizeof(struct location) * t->max));
3455}
3456
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003457static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003458{
3459 struct location *l;
3460 int order;
3461
Christoph Lameter88a420e2007-05-06 14:49:45 -07003462 order = get_order(sizeof(struct location) * max);
3463
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003464 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003465 if (!l)
3466 return 0;
3467
3468 if (t->count) {
3469 memcpy(l, t->loc, sizeof(struct location) * t->count);
3470 free_loc_track(t);
3471 }
3472 t->max = max;
3473 t->loc = l;
3474 return 1;
3475}
3476
3477static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003478 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003479{
3480 long start, end, pos;
3481 struct location *l;
3482 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003483 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003484
3485 start = -1;
3486 end = t->count;
3487
3488 for ( ; ; ) {
3489 pos = start + (end - start + 1) / 2;
3490
3491 /*
3492 * There is nothing at "end". If we end up there
3493 * we need to add something to before end.
3494 */
3495 if (pos == end)
3496 break;
3497
3498 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003499 if (track->addr == caddr) {
3500
3501 l = &t->loc[pos];
3502 l->count++;
3503 if (track->when) {
3504 l->sum_time += age;
3505 if (age < l->min_time)
3506 l->min_time = age;
3507 if (age > l->max_time)
3508 l->max_time = age;
3509
3510 if (track->pid < l->min_pid)
3511 l->min_pid = track->pid;
3512 if (track->pid > l->max_pid)
3513 l->max_pid = track->pid;
3514
3515 cpu_set(track->cpu, l->cpus);
3516 }
3517 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003518 return 1;
3519 }
3520
Christoph Lameter45edfa52007-05-09 02:32:45 -07003521 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003522 end = pos;
3523 else
3524 start = pos;
3525 }
3526
3527 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003528 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003529 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003530 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003531 return 0;
3532
3533 l = t->loc + pos;
3534 if (pos < t->count)
3535 memmove(l + 1, l,
3536 (t->count - pos) * sizeof(struct location));
3537 t->count++;
3538 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003539 l->addr = track->addr;
3540 l->sum_time = age;
3541 l->min_time = age;
3542 l->max_time = age;
3543 l->min_pid = track->pid;
3544 l->max_pid = track->pid;
3545 cpus_clear(l->cpus);
3546 cpu_set(track->cpu, l->cpus);
3547 nodes_clear(l->nodes);
3548 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003549 return 1;
3550}
3551
3552static void process_slab(struct loc_track *t, struct kmem_cache *s,
3553 struct page *page, enum track_item alloc)
3554{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003555 void *addr = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +03003556 DECLARE_BITMAP(map, page->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003557 void *p;
3558
Christoph Lameter39b26462008-04-14 19:11:30 +03003559 bitmap_zero(map, page->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003560 for_each_free_object(p, s, page->freelist)
3561 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003562
Christoph Lameter224a88b2008-04-14 19:11:31 +03003563 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003564 if (!test_bit(slab_index(p, s, addr), map))
3565 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003566}
3567
3568static int list_locations(struct kmem_cache *s, char *buf,
3569 enum track_item alloc)
3570{
Harvey Harrisone374d482008-01-31 15:20:50 -08003571 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003572 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003573 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003574 int node;
3575
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003576 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003577 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003578 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003579
3580 /* Push back cpu slabs */
3581 flush_all(s);
3582
Christoph Lameterf64dc582007-10-16 01:25:33 -07003583 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003584 struct kmem_cache_node *n = get_node(s, node);
3585 unsigned long flags;
3586 struct page *page;
3587
Christoph Lameter9e869432007-08-22 14:01:56 -07003588 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003589 continue;
3590
3591 spin_lock_irqsave(&n->list_lock, flags);
3592 list_for_each_entry(page, &n->partial, lru)
3593 process_slab(&t, s, page, alloc);
3594 list_for_each_entry(page, &n->full, lru)
3595 process_slab(&t, s, page, alloc);
3596 spin_unlock_irqrestore(&n->list_lock, flags);
3597 }
3598
3599 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003600 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003601
Harvey Harrisone374d482008-01-31 15:20:50 -08003602 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003603 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003604 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003605
3606 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003607 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003608 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003609 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003610
3611 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08003612 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07003613 l->min_time,
3614 (long)div_u64(l->sum_time, l->count),
3615 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003616 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003617 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003618 l->min_time);
3619
3620 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003621 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003622 l->min_pid, l->max_pid);
3623 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003624 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003625 l->min_pid);
3626
Christoph Lameter84966342007-06-23 17:16:32 -07003627 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003628 len < PAGE_SIZE - 60) {
3629 len += sprintf(buf + len, " cpus=");
3630 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003631 l->cpus);
3632 }
3633
Christoph Lameter84966342007-06-23 17:16:32 -07003634 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003635 len < PAGE_SIZE - 60) {
3636 len += sprintf(buf + len, " nodes=");
3637 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003638 l->nodes);
3639 }
3640
Harvey Harrisone374d482008-01-31 15:20:50 -08003641 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003642 }
3643
3644 free_loc_track(&t);
3645 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003646 len += sprintf(buf, "No data\n");
3647 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003648}
3649
Christoph Lameter81819f02007-05-06 14:49:36 -07003650enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03003651 SL_ALL, /* All slabs */
3652 SL_PARTIAL, /* Only partially allocated slabs */
3653 SL_CPU, /* Only slabs used for cpu caches */
3654 SL_OBJECTS, /* Determine allocated objects not slabs */
3655 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07003656};
3657
Christoph Lameter205ab992008-04-14 19:11:40 +03003658#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003659#define SO_PARTIAL (1 << SL_PARTIAL)
3660#define SO_CPU (1 << SL_CPU)
3661#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03003662#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003663
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003664static ssize_t show_slab_objects(struct kmem_cache *s,
3665 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003666{
3667 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003668 int node;
3669 int x;
3670 unsigned long *nodes;
3671 unsigned long *per_cpu;
3672
3673 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003674 if (!nodes)
3675 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003676 per_cpu = nodes + nr_node_ids;
3677
Christoph Lameter205ab992008-04-14 19:11:40 +03003678 if (flags & SO_CPU) {
3679 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07003680
Christoph Lameter205ab992008-04-14 19:11:40 +03003681 for_each_possible_cpu(cpu) {
3682 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003683
Christoph Lameter205ab992008-04-14 19:11:40 +03003684 if (!c || c->node < 0)
3685 continue;
3686
3687 if (c->page) {
3688 if (flags & SO_TOTAL)
3689 x = c->page->objects;
3690 else if (flags & SO_OBJECTS)
3691 x = c->page->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07003692 else
3693 x = 1;
Christoph Lameter205ab992008-04-14 19:11:40 +03003694
Christoph Lameter81819f02007-05-06 14:49:36 -07003695 total += x;
Christoph Lameter205ab992008-04-14 19:11:40 +03003696 nodes[c->node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003697 }
Christoph Lameter205ab992008-04-14 19:11:40 +03003698 per_cpu[c->node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003699 }
3700 }
3701
Christoph Lameter205ab992008-04-14 19:11:40 +03003702 if (flags & SO_ALL) {
3703 for_each_node_state(node, N_NORMAL_MEMORY) {
3704 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003705
Christoph Lameter205ab992008-04-14 19:11:40 +03003706 if (flags & SO_TOTAL)
3707 x = atomic_long_read(&n->total_objects);
3708 else if (flags & SO_OBJECTS)
3709 x = atomic_long_read(&n->total_objects) -
3710 count_partial(n, count_free);
3711
3712 else
3713 x = atomic_long_read(&n->nr_slabs);
3714 total += x;
3715 nodes[node] += x;
3716 }
3717
3718 } else if (flags & SO_PARTIAL) {
3719 for_each_node_state(node, N_NORMAL_MEMORY) {
3720 struct kmem_cache_node *n = get_node(s, node);
3721
3722 if (flags & SO_TOTAL)
3723 x = count_partial(n, count_total);
3724 else if (flags & SO_OBJECTS)
3725 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07003726 else
3727 x = n->nr_partial;
3728 total += x;
3729 nodes[node] += x;
3730 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003731 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003732 x = sprintf(buf, "%lu", total);
3733#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003734 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003735 if (nodes[node])
3736 x += sprintf(buf + x, " N%d=%lu",
3737 node, nodes[node]);
3738#endif
3739 kfree(nodes);
3740 return x + sprintf(buf + x, "\n");
3741}
3742
3743static int any_slab_objects(struct kmem_cache *s)
3744{
3745 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003746
3747 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003748 struct kmem_cache_node *n = get_node(s, node);
3749
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003750 if (!n)
3751 continue;
3752
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07003753 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07003754 return 1;
3755 }
3756 return 0;
3757}
3758
3759#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3760#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3761
3762struct slab_attribute {
3763 struct attribute attr;
3764 ssize_t (*show)(struct kmem_cache *s, char *buf);
3765 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3766};
3767
3768#define SLAB_ATTR_RO(_name) \
3769 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3770
3771#define SLAB_ATTR(_name) \
3772 static struct slab_attribute _name##_attr = \
3773 __ATTR(_name, 0644, _name##_show, _name##_store)
3774
Christoph Lameter81819f02007-05-06 14:49:36 -07003775static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3776{
3777 return sprintf(buf, "%d\n", s->size);
3778}
3779SLAB_ATTR_RO(slab_size);
3780
3781static ssize_t align_show(struct kmem_cache *s, char *buf)
3782{
3783 return sprintf(buf, "%d\n", s->align);
3784}
3785SLAB_ATTR_RO(align);
3786
3787static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3788{
3789 return sprintf(buf, "%d\n", s->objsize);
3790}
3791SLAB_ATTR_RO(object_size);
3792
3793static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3794{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003795 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003796}
3797SLAB_ATTR_RO(objs_per_slab);
3798
Christoph Lameter06b285d2008-04-14 19:11:41 +03003799static ssize_t order_store(struct kmem_cache *s,
3800 const char *buf, size_t length)
3801{
Christoph Lameter0121c6192008-04-29 16:11:12 -07003802 unsigned long order;
3803 int err;
3804
3805 err = strict_strtoul(buf, 10, &order);
3806 if (err)
3807 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003808
3809 if (order > slub_max_order || order < slub_min_order)
3810 return -EINVAL;
3811
3812 calculate_sizes(s, order);
3813 return length;
3814}
3815
Christoph Lameter81819f02007-05-06 14:49:36 -07003816static ssize_t order_show(struct kmem_cache *s, char *buf)
3817{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003818 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003819}
Christoph Lameter06b285d2008-04-14 19:11:41 +03003820SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003821
3822static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3823{
3824 if (s->ctor) {
3825 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3826
3827 return n + sprintf(buf + n, "\n");
3828 }
3829 return 0;
3830}
3831SLAB_ATTR_RO(ctor);
3832
Christoph Lameter81819f02007-05-06 14:49:36 -07003833static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3834{
3835 return sprintf(buf, "%d\n", s->refcount - 1);
3836}
3837SLAB_ATTR_RO(aliases);
3838
3839static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3840{
Christoph Lameter205ab992008-04-14 19:11:40 +03003841 return show_slab_objects(s, buf, SO_ALL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003842}
3843SLAB_ATTR_RO(slabs);
3844
3845static ssize_t partial_show(struct kmem_cache *s, char *buf)
3846{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003847 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003848}
3849SLAB_ATTR_RO(partial);
3850
3851static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3852{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003853 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003854}
3855SLAB_ATTR_RO(cpu_slabs);
3856
3857static ssize_t objects_show(struct kmem_cache *s, char *buf)
3858{
Christoph Lameter205ab992008-04-14 19:11:40 +03003859 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003860}
3861SLAB_ATTR_RO(objects);
3862
Christoph Lameter205ab992008-04-14 19:11:40 +03003863static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3864{
3865 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3866}
3867SLAB_ATTR_RO(objects_partial);
3868
3869static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3870{
3871 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3872}
3873SLAB_ATTR_RO(total_objects);
3874
Christoph Lameter81819f02007-05-06 14:49:36 -07003875static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3876{
3877 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3878}
3879
3880static ssize_t sanity_checks_store(struct kmem_cache *s,
3881 const char *buf, size_t length)
3882{
3883 s->flags &= ~SLAB_DEBUG_FREE;
3884 if (buf[0] == '1')
3885 s->flags |= SLAB_DEBUG_FREE;
3886 return length;
3887}
3888SLAB_ATTR(sanity_checks);
3889
3890static ssize_t trace_show(struct kmem_cache *s, char *buf)
3891{
3892 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3893}
3894
3895static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3896 size_t length)
3897{
3898 s->flags &= ~SLAB_TRACE;
3899 if (buf[0] == '1')
3900 s->flags |= SLAB_TRACE;
3901 return length;
3902}
3903SLAB_ATTR(trace);
3904
3905static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3906{
3907 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3908}
3909
3910static ssize_t reclaim_account_store(struct kmem_cache *s,
3911 const char *buf, size_t length)
3912{
3913 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3914 if (buf[0] == '1')
3915 s->flags |= SLAB_RECLAIM_ACCOUNT;
3916 return length;
3917}
3918SLAB_ATTR(reclaim_account);
3919
3920static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3921{
Christoph Lameter5af60832007-05-06 14:49:56 -07003922 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003923}
3924SLAB_ATTR_RO(hwcache_align);
3925
3926#ifdef CONFIG_ZONE_DMA
3927static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3928{
3929 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3930}
3931SLAB_ATTR_RO(cache_dma);
3932#endif
3933
3934static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3935{
3936 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3937}
3938SLAB_ATTR_RO(destroy_by_rcu);
3939
3940static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3941{
3942 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3943}
3944
3945static ssize_t red_zone_store(struct kmem_cache *s,
3946 const char *buf, size_t length)
3947{
3948 if (any_slab_objects(s))
3949 return -EBUSY;
3950
3951 s->flags &= ~SLAB_RED_ZONE;
3952 if (buf[0] == '1')
3953 s->flags |= SLAB_RED_ZONE;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003954 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003955 return length;
3956}
3957SLAB_ATTR(red_zone);
3958
3959static ssize_t poison_show(struct kmem_cache *s, char *buf)
3960{
3961 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3962}
3963
3964static ssize_t poison_store(struct kmem_cache *s,
3965 const char *buf, size_t length)
3966{
3967 if (any_slab_objects(s))
3968 return -EBUSY;
3969
3970 s->flags &= ~SLAB_POISON;
3971 if (buf[0] == '1')
3972 s->flags |= SLAB_POISON;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003973 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003974 return length;
3975}
3976SLAB_ATTR(poison);
3977
3978static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3979{
3980 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3981}
3982
3983static ssize_t store_user_store(struct kmem_cache *s,
3984 const char *buf, size_t length)
3985{
3986 if (any_slab_objects(s))
3987 return -EBUSY;
3988
3989 s->flags &= ~SLAB_STORE_USER;
3990 if (buf[0] == '1')
3991 s->flags |= SLAB_STORE_USER;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003992 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003993 return length;
3994}
3995SLAB_ATTR(store_user);
3996
Christoph Lameter53e15af2007-05-06 14:49:43 -07003997static ssize_t validate_show(struct kmem_cache *s, char *buf)
3998{
3999 return 0;
4000}
4001
4002static ssize_t validate_store(struct kmem_cache *s,
4003 const char *buf, size_t length)
4004{
Christoph Lameter434e2452007-07-17 04:03:30 -07004005 int ret = -EINVAL;
4006
4007 if (buf[0] == '1') {
4008 ret = validate_slab_cache(s);
4009 if (ret >= 0)
4010 ret = length;
4011 }
4012 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004013}
4014SLAB_ATTR(validate);
4015
Christoph Lameter2086d262007-05-06 14:49:46 -07004016static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4017{
4018 return 0;
4019}
4020
4021static ssize_t shrink_store(struct kmem_cache *s,
4022 const char *buf, size_t length)
4023{
4024 if (buf[0] == '1') {
4025 int rc = kmem_cache_shrink(s);
4026
4027 if (rc)
4028 return rc;
4029 } else
4030 return -EINVAL;
4031 return length;
4032}
4033SLAB_ATTR(shrink);
4034
Christoph Lameter88a420e2007-05-06 14:49:45 -07004035static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4036{
4037 if (!(s->flags & SLAB_STORE_USER))
4038 return -ENOSYS;
4039 return list_locations(s, buf, TRACK_ALLOC);
4040}
4041SLAB_ATTR_RO(alloc_calls);
4042
4043static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4044{
4045 if (!(s->flags & SLAB_STORE_USER))
4046 return -ENOSYS;
4047 return list_locations(s, buf, TRACK_FREE);
4048}
4049SLAB_ATTR_RO(free_calls);
4050
Christoph Lameter81819f02007-05-06 14:49:36 -07004051#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004052static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004053{
Christoph Lameter98246012008-01-07 23:20:26 -08004054 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004055}
4056
Christoph Lameter98246012008-01-07 23:20:26 -08004057static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004058 const char *buf, size_t length)
4059{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004060 unsigned long ratio;
4061 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004062
Christoph Lameter0121c6192008-04-29 16:11:12 -07004063 err = strict_strtoul(buf, 10, &ratio);
4064 if (err)
4065 return err;
4066
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004067 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004068 s->remote_node_defrag_ratio = ratio * 10;
4069
Christoph Lameter81819f02007-05-06 14:49:36 -07004070 return length;
4071}
Christoph Lameter98246012008-01-07 23:20:26 -08004072SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004073#endif
4074
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004075#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004076static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4077{
4078 unsigned long sum = 0;
4079 int cpu;
4080 int len;
4081 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4082
4083 if (!data)
4084 return -ENOMEM;
4085
4086 for_each_online_cpu(cpu) {
4087 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4088
4089 data[cpu] = x;
4090 sum += x;
4091 }
4092
4093 len = sprintf(buf, "%lu", sum);
4094
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004095#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004096 for_each_online_cpu(cpu) {
4097 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004098 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004099 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004100#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004101 kfree(data);
4102 return len + sprintf(buf + len, "\n");
4103}
4104
4105#define STAT_ATTR(si, text) \
4106static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4107{ \
4108 return show_stat(s, buf, si); \
4109} \
4110SLAB_ATTR_RO(text); \
4111
4112STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4113STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4114STAT_ATTR(FREE_FASTPATH, free_fastpath);
4115STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4116STAT_ATTR(FREE_FROZEN, free_frozen);
4117STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4118STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4119STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4120STAT_ATTR(ALLOC_SLAB, alloc_slab);
4121STAT_ATTR(ALLOC_REFILL, alloc_refill);
4122STAT_ATTR(FREE_SLAB, free_slab);
4123STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4124STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4125STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4126STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4127STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4128STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter65c33762008-04-14 19:11:40 +03004129STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004130#endif
4131
Pekka Enberg06428782008-01-07 23:20:27 -08004132static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004133 &slab_size_attr.attr,
4134 &object_size_attr.attr,
4135 &objs_per_slab_attr.attr,
4136 &order_attr.attr,
4137 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004138 &objects_partial_attr.attr,
4139 &total_objects_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004140 &slabs_attr.attr,
4141 &partial_attr.attr,
4142 &cpu_slabs_attr.attr,
4143 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004144 &aliases_attr.attr,
4145 &align_attr.attr,
4146 &sanity_checks_attr.attr,
4147 &trace_attr.attr,
4148 &hwcache_align_attr.attr,
4149 &reclaim_account_attr.attr,
4150 &destroy_by_rcu_attr.attr,
4151 &red_zone_attr.attr,
4152 &poison_attr.attr,
4153 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004154 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004155 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004156 &alloc_calls_attr.attr,
4157 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004158#ifdef CONFIG_ZONE_DMA
4159 &cache_dma_attr.attr,
4160#endif
4161#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004162 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004163#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004164#ifdef CONFIG_SLUB_STATS
4165 &alloc_fastpath_attr.attr,
4166 &alloc_slowpath_attr.attr,
4167 &free_fastpath_attr.attr,
4168 &free_slowpath_attr.attr,
4169 &free_frozen_attr.attr,
4170 &free_add_partial_attr.attr,
4171 &free_remove_partial_attr.attr,
4172 &alloc_from_partial_attr.attr,
4173 &alloc_slab_attr.attr,
4174 &alloc_refill_attr.attr,
4175 &free_slab_attr.attr,
4176 &cpuslab_flush_attr.attr,
4177 &deactivate_full_attr.attr,
4178 &deactivate_empty_attr.attr,
4179 &deactivate_to_head_attr.attr,
4180 &deactivate_to_tail_attr.attr,
4181 &deactivate_remote_frees_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004182 &order_fallback_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004183#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004184 NULL
4185};
4186
4187static struct attribute_group slab_attr_group = {
4188 .attrs = slab_attrs,
4189};
4190
4191static ssize_t slab_attr_show(struct kobject *kobj,
4192 struct attribute *attr,
4193 char *buf)
4194{
4195 struct slab_attribute *attribute;
4196 struct kmem_cache *s;
4197 int err;
4198
4199 attribute = to_slab_attr(attr);
4200 s = to_slab(kobj);
4201
4202 if (!attribute->show)
4203 return -EIO;
4204
4205 err = attribute->show(s, buf);
4206
4207 return err;
4208}
4209
4210static ssize_t slab_attr_store(struct kobject *kobj,
4211 struct attribute *attr,
4212 const char *buf, size_t len)
4213{
4214 struct slab_attribute *attribute;
4215 struct kmem_cache *s;
4216 int err;
4217
4218 attribute = to_slab_attr(attr);
4219 s = to_slab(kobj);
4220
4221 if (!attribute->store)
4222 return -EIO;
4223
4224 err = attribute->store(s, buf, len);
4225
4226 return err;
4227}
4228
Christoph Lameter151c6022008-01-07 22:29:05 -08004229static void kmem_cache_release(struct kobject *kobj)
4230{
4231 struct kmem_cache *s = to_slab(kobj);
4232
4233 kfree(s);
4234}
4235
Christoph Lameter81819f02007-05-06 14:49:36 -07004236static struct sysfs_ops slab_sysfs_ops = {
4237 .show = slab_attr_show,
4238 .store = slab_attr_store,
4239};
4240
4241static struct kobj_type slab_ktype = {
4242 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004243 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004244};
4245
4246static int uevent_filter(struct kset *kset, struct kobject *kobj)
4247{
4248 struct kobj_type *ktype = get_ktype(kobj);
4249
4250 if (ktype == &slab_ktype)
4251 return 1;
4252 return 0;
4253}
4254
4255static struct kset_uevent_ops slab_uevent_ops = {
4256 .filter = uevent_filter,
4257};
4258
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004259static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004260
4261#define ID_STR_LENGTH 64
4262
4263/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004264 *
4265 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004266 */
4267static char *create_unique_id(struct kmem_cache *s)
4268{
4269 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4270 char *p = name;
4271
4272 BUG_ON(!name);
4273
4274 *p++ = ':';
4275 /*
4276 * First flags affecting slabcache operations. We will only
4277 * get here for aliasable slabs so we do not need to support
4278 * too many flags. The flags here must cover all flags that
4279 * are matched during merging to guarantee that the id is
4280 * unique.
4281 */
4282 if (s->flags & SLAB_CACHE_DMA)
4283 *p++ = 'd';
4284 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4285 *p++ = 'a';
4286 if (s->flags & SLAB_DEBUG_FREE)
4287 *p++ = 'F';
4288 if (p != name + 1)
4289 *p++ = '-';
4290 p += sprintf(p, "%07d", s->size);
4291 BUG_ON(p > name + ID_STR_LENGTH - 1);
4292 return name;
4293}
4294
4295static int sysfs_slab_add(struct kmem_cache *s)
4296{
4297 int err;
4298 const char *name;
4299 int unmergeable;
4300
4301 if (slab_state < SYSFS)
4302 /* Defer until later */
4303 return 0;
4304
4305 unmergeable = slab_unmergeable(s);
4306 if (unmergeable) {
4307 /*
4308 * Slabcache can never be merged so we can use the name proper.
4309 * This is typically the case for debug situations. In that
4310 * case we can catch duplicate names easily.
4311 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004312 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004313 name = s->name;
4314 } else {
4315 /*
4316 * Create a unique name for the slab as a target
4317 * for the symlinks.
4318 */
4319 name = create_unique_id(s);
4320 }
4321
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004322 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004323 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4324 if (err) {
4325 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004326 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004327 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004328
4329 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4330 if (err)
4331 return err;
4332 kobject_uevent(&s->kobj, KOBJ_ADD);
4333 if (!unmergeable) {
4334 /* Setup first alias */
4335 sysfs_slab_alias(s, s->name);
4336 kfree(name);
4337 }
4338 return 0;
4339}
4340
4341static void sysfs_slab_remove(struct kmem_cache *s)
4342{
4343 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4344 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004345 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004346}
4347
4348/*
4349 * Need to buffer aliases during bootup until sysfs becomes
4350 * available lest we loose that information.
4351 */
4352struct saved_alias {
4353 struct kmem_cache *s;
4354 const char *name;
4355 struct saved_alias *next;
4356};
4357
Adrian Bunk5af328a2007-07-17 04:03:27 -07004358static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004359
4360static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4361{
4362 struct saved_alias *al;
4363
4364 if (slab_state == SYSFS) {
4365 /*
4366 * If we have a leftover link then remove it.
4367 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004368 sysfs_remove_link(&slab_kset->kobj, name);
4369 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004370 }
4371
4372 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4373 if (!al)
4374 return -ENOMEM;
4375
4376 al->s = s;
4377 al->name = name;
4378 al->next = alias_list;
4379 alias_list = al;
4380 return 0;
4381}
4382
4383static int __init slab_sysfs_init(void)
4384{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004385 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004386 int err;
4387
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004388 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004389 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004390 printk(KERN_ERR "Cannot register slab subsystem.\n");
4391 return -ENOSYS;
4392 }
4393
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004394 slab_state = SYSFS;
4395
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004396 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004397 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004398 if (err)
4399 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4400 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004401 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004402
4403 while (alias_list) {
4404 struct saved_alias *al = alias_list;
4405
4406 alias_list = alias_list->next;
4407 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004408 if (err)
4409 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4410 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004411 kfree(al);
4412 }
4413
4414 resiliency_test();
4415 return 0;
4416}
4417
4418__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004419#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004420
4421/*
4422 * The /proc/slabinfo ABI
4423 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004424#ifdef CONFIG_SLABINFO
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004425static void print_slabinfo_header(struct seq_file *m)
4426{
4427 seq_puts(m, "slabinfo - version: 2.1\n");
4428 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4429 "<objperslab> <pagesperslab>");
4430 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4431 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4432 seq_putc(m, '\n');
4433}
4434
4435static void *s_start(struct seq_file *m, loff_t *pos)
4436{
4437 loff_t n = *pos;
4438
4439 down_read(&slub_lock);
4440 if (!n)
4441 print_slabinfo_header(m);
4442
4443 return seq_list_start(&slab_caches, *pos);
4444}
4445
4446static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4447{
4448 return seq_list_next(p, &slab_caches, pos);
4449}
4450
4451static void s_stop(struct seq_file *m, void *p)
4452{
4453 up_read(&slub_lock);
4454}
4455
4456static int s_show(struct seq_file *m, void *p)
4457{
4458 unsigned long nr_partials = 0;
4459 unsigned long nr_slabs = 0;
4460 unsigned long nr_inuse = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004461 unsigned long nr_objs = 0;
4462 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004463 struct kmem_cache *s;
4464 int node;
4465
4466 s = list_entry(p, struct kmem_cache, list);
4467
4468 for_each_online_node(node) {
4469 struct kmem_cache_node *n = get_node(s, node);
4470
4471 if (!n)
4472 continue;
4473
4474 nr_partials += n->nr_partial;
4475 nr_slabs += atomic_long_read(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03004476 nr_objs += atomic_long_read(&n->total_objects);
4477 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004478 }
4479
Christoph Lameter205ab992008-04-14 19:11:40 +03004480 nr_inuse = nr_objs - nr_free;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004481
4482 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
Christoph Lameter834f3d12008-04-14 19:11:31 +03004483 nr_objs, s->size, oo_objects(s->oo),
4484 (1 << oo_order(s->oo)));
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004485 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4486 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4487 0UL);
4488 seq_putc(m, '\n');
4489 return 0;
4490}
4491
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004492static const struct seq_operations slabinfo_op = {
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004493 .start = s_start,
4494 .next = s_next,
4495 .stop = s_stop,
4496 .show = s_show,
4497};
4498
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004499static int slabinfo_open(struct inode *inode, struct file *file)
4500{
4501 return seq_open(file, &slabinfo_op);
4502}
4503
4504static const struct file_operations proc_slabinfo_operations = {
4505 .open = slabinfo_open,
4506 .read = seq_read,
4507 .llseek = seq_lseek,
4508 .release = seq_release,
4509};
4510
4511static int __init slab_proc_init(void)
4512{
4513 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4514 return 0;
4515}
4516module_init(slab_proc_init);
Linus Torvalds158a9622008-01-02 13:04:48 -08004517#endif /* CONFIG_SLABINFO */