blob: 7ad489af95614602a2f5fec8756460108a1f6be1 [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
156/* Internal SLUB flags */
Christoph Lameter1ceef402007-08-07 15:11:48 -0700157#define __OBJECT_POISON 0x80000000 /* Poison object */
158#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700159
160static int kmem_size = sizeof(struct kmem_cache);
161
162#ifdef CONFIG_SMP
163static struct notifier_block slab_notifier;
164#endif
165
166static enum {
167 DOWN, /* No slab functionality available */
168 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
Christoph Lameter672bba32007-05-09 02:32:39 -0700169 UP, /* Everything works but does not show up in sysfs */
Christoph Lameter81819f02007-05-06 14:49:36 -0700170 SYSFS /* Sysfs up */
171} slab_state = DOWN;
172
173/* A list of all slab caches on the system */
174static DECLARE_RWSEM(slub_lock);
Adrian Bunk5af328a2007-07-17 04:03:27 -0700175static LIST_HEAD(slab_caches);
Christoph Lameter81819f02007-05-06 14:49:36 -0700176
Christoph Lameter02cbc872007-05-09 02:32:43 -0700177/*
178 * Tracking user of a slab.
179 */
180struct track {
181 void *addr; /* Called from address */
182 int cpu; /* Was running on cpu */
183 int pid; /* Pid context */
184 unsigned long when; /* When did the operation occur */
185};
186
187enum track_item { TRACK_ALLOC, TRACK_FREE };
188
Christoph Lameterf6acb632008-04-29 16:16:06 -0700189#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -0700190static int sysfs_slab_add(struct kmem_cache *);
191static int sysfs_slab_alias(struct kmem_cache *, const char *);
192static void sysfs_slab_remove(struct kmem_cache *);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800193
Christoph Lameter81819f02007-05-06 14:49:36 -0700194#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700195static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
196static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
197 { return 0; }
Christoph Lameter151c6022008-01-07 22:29:05 -0800198static inline void sysfs_slab_remove(struct kmem_cache *s)
199{
200 kfree(s);
201}
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800202
Christoph Lameter81819f02007-05-06 14:49:36 -0700203#endif
204
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800205static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
206{
207#ifdef CONFIG_SLUB_STATS
208 c->stat[si]++;
209#endif
210}
211
Christoph Lameter81819f02007-05-06 14:49:36 -0700212/********************************************************************
213 * Core slab cache functions
214 *******************************************************************/
215
216int slab_is_available(void)
217{
218 return slab_state >= UP;
219}
220
221static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
222{
223#ifdef CONFIG_NUMA
224 return s->node[node];
225#else
226 return &s->local_node;
227#endif
228}
229
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700230static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
231{
Christoph Lameter4c93c3552007-10-16 01:26:08 -0700232#ifdef CONFIG_SMP
233 return s->cpu_slab[cpu];
234#else
235 return &s->cpu_slab;
236#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700237}
238
Christoph Lameter6446faa2008-02-15 23:45:26 -0800239/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700240static inline int check_valid_pointer(struct kmem_cache *s,
241 struct page *page, const void *object)
242{
243 void *base;
244
Christoph Lametera973e9d2008-03-01 13:40:44 -0800245 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700246 return 1;
247
Christoph Lametera973e9d2008-03-01 13:40:44 -0800248 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300249 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700250 (object - base) % s->size) {
251 return 0;
252 }
253
254 return 1;
255}
256
Christoph Lameter81819f02007-05-06 14:49:36 -0700257/*
Christoph Lameter7656c722007-05-09 02:32:40 -0700258 * Slow version of get and set free pointer.
259 *
260 * This version requires touching the cache lines of kmem_cache which
261 * we avoid to do in the fast alloc free paths. There we obtain the offset
262 * from the page struct.
263 */
264static inline void *get_freepointer(struct kmem_cache *s, void *object)
265{
266 return *(void **)(object + s->offset);
267}
268
269static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
270{
271 *(void **)(object + s->offset) = fp;
272}
273
274/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300275#define for_each_object(__p, __s, __addr, __objects) \
276 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700277 __p += (__s)->size)
278
279/* Scan freelist */
280#define for_each_free_object(__p, __s, __free) \
Christoph Lametera973e9d2008-03-01 13:40:44 -0800281 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
Christoph Lameter7656c722007-05-09 02:32:40 -0700282
283/* Determine object index from a given position */
284static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
285{
286 return (p - addr) / s->size;
287}
288
Christoph Lameter834f3d12008-04-14 19:11:31 +0300289static inline struct kmem_cache_order_objects oo_make(int order,
290 unsigned long size)
291{
292 struct kmem_cache_order_objects x = {
293 (order << 16) + (PAGE_SIZE << order) / size
294 };
295
296 return x;
297}
298
299static inline int oo_order(struct kmem_cache_order_objects x)
300{
301 return x.x >> 16;
302}
303
304static inline int oo_objects(struct kmem_cache_order_objects x)
305{
306 return x.x & ((1 << 16) - 1);
307}
308
Christoph Lameter41ecc552007-05-09 02:32:44 -0700309#ifdef CONFIG_SLUB_DEBUG
310/*
311 * Debug settings:
312 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700313#ifdef CONFIG_SLUB_DEBUG_ON
314static int slub_debug = DEBUG_DEFAULT_FLAGS;
315#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700316static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700317#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700318
319static char *slub_debug_slabs;
320
Christoph Lameter7656c722007-05-09 02:32:40 -0700321/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700322 * Object debugging
323 */
324static void print_section(char *text, u8 *addr, unsigned int length)
325{
326 int i, offset;
327 int newline = 1;
328 char ascii[17];
329
330 ascii[16] = 0;
331
332 for (i = 0; i < length; i++) {
333 if (newline) {
Christoph Lameter24922682007-07-17 04:03:18 -0700334 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
Christoph Lameter81819f02007-05-06 14:49:36 -0700335 newline = 0;
336 }
Pekka Enberg06428782008-01-07 23:20:27 -0800337 printk(KERN_CONT " %02x", addr[i]);
Christoph Lameter81819f02007-05-06 14:49:36 -0700338 offset = i % 16;
339 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
340 if (offset == 15) {
Pekka Enberg06428782008-01-07 23:20:27 -0800341 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700342 newline = 1;
343 }
344 }
345 if (!newline) {
346 i %= 16;
347 while (i < 16) {
Pekka Enberg06428782008-01-07 23:20:27 -0800348 printk(KERN_CONT " ");
Christoph Lameter81819f02007-05-06 14:49:36 -0700349 ascii[i] = ' ';
350 i++;
351 }
Pekka Enberg06428782008-01-07 23:20:27 -0800352 printk(KERN_CONT " %s\n", ascii);
Christoph Lameter81819f02007-05-06 14:49:36 -0700353 }
354}
355
Christoph Lameter81819f02007-05-06 14:49:36 -0700356static struct track *get_track(struct kmem_cache *s, void *object,
357 enum track_item alloc)
358{
359 struct track *p;
360
361 if (s->offset)
362 p = object + s->offset + sizeof(void *);
363 else
364 p = object + s->inuse;
365
366 return p + alloc;
367}
368
369static void set_track(struct kmem_cache *s, void *object,
370 enum track_item alloc, void *addr)
371{
372 struct track *p;
373
374 if (s->offset)
375 p = object + s->offset + sizeof(void *);
376 else
377 p = object + s->inuse;
378
379 p += alloc;
380 if (addr) {
381 p->addr = addr;
382 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400383 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700384 p->when = jiffies;
385 } else
386 memset(p, 0, sizeof(struct track));
387}
388
Christoph Lameter81819f02007-05-06 14:49:36 -0700389static void init_tracking(struct kmem_cache *s, void *object)
390{
Christoph Lameter24922682007-07-17 04:03:18 -0700391 if (!(s->flags & SLAB_STORE_USER))
392 return;
393
394 set_track(s, object, TRACK_FREE, NULL);
395 set_track(s, object, TRACK_ALLOC, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700396}
397
398static void print_track(const char *s, struct track *t)
399{
400 if (!t->addr)
401 return;
402
Linus Torvalds7daf7052008-07-14 12:12:53 -0700403 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
404 s, t->addr, jiffies - t->when, t->cpu, t->pid);
Christoph Lameter81819f02007-05-06 14:49:36 -0700405}
406
Christoph Lameter24922682007-07-17 04:03:18 -0700407static void print_tracking(struct kmem_cache *s, void *object)
408{
409 if (!(s->flags & SLAB_STORE_USER))
410 return;
411
412 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
413 print_track("Freed", get_track(s, object, TRACK_FREE));
414}
415
416static void print_page_info(struct page *page)
417{
Christoph Lameter39b26462008-04-14 19:11:30 +0300418 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
419 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700420
421}
422
423static void slab_bug(struct kmem_cache *s, char *fmt, ...)
424{
425 va_list args;
426 char buf[100];
427
428 va_start(args, fmt);
429 vsnprintf(buf, sizeof(buf), fmt, args);
430 va_end(args);
431 printk(KERN_ERR "========================================"
432 "=====================================\n");
433 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
434 printk(KERN_ERR "----------------------------------------"
435 "-------------------------------------\n\n");
436}
437
438static void slab_fix(struct kmem_cache *s, char *fmt, ...)
439{
440 va_list args;
441 char buf[100];
442
443 va_start(args, fmt);
444 vsnprintf(buf, sizeof(buf), fmt, args);
445 va_end(args);
446 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
447}
448
449static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700450{
451 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800452 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700453
454 print_tracking(s, p);
455
456 print_page_info(page);
457
458 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
459 p, p - addr, get_freepointer(s, p));
460
461 if (p > addr + 16)
462 print_section("Bytes b4", p - 16, 16);
463
Pekka Enberg0ebd6522008-07-19 14:17:22 +0300464 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700465
466 if (s->flags & SLAB_RED_ZONE)
467 print_section("Redzone", p + s->objsize,
468 s->inuse - s->objsize);
469
Christoph Lameter81819f02007-05-06 14:49:36 -0700470 if (s->offset)
471 off = s->offset + sizeof(void *);
472 else
473 off = s->inuse;
474
Christoph Lameter24922682007-07-17 04:03:18 -0700475 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700476 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700477
478 if (off != s->size)
479 /* Beginning of the filler is the free pointer */
Christoph Lameter24922682007-07-17 04:03:18 -0700480 print_section("Padding", p + off, s->size - off);
481
482 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700483}
484
485static void object_err(struct kmem_cache *s, struct page *page,
486 u8 *object, char *reason)
487{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700488 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700489 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700490}
491
Christoph Lameter24922682007-07-17 04:03:18 -0700492static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700493{
494 va_list args;
495 char buf[100];
496
Christoph Lameter24922682007-07-17 04:03:18 -0700497 va_start(args, fmt);
498 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700499 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700500 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700501 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700502 dump_stack();
503}
504
505static void init_object(struct kmem_cache *s, void *object, int active)
506{
507 u8 *p = object;
508
509 if (s->flags & __OBJECT_POISON) {
510 memset(p, POISON_FREE, s->objsize - 1);
Pekka Enberg06428782008-01-07 23:20:27 -0800511 p[s->objsize - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700512 }
513
514 if (s->flags & SLAB_RED_ZONE)
515 memset(p + s->objsize,
516 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
517 s->inuse - s->objsize);
518}
519
Christoph Lameter24922682007-07-17 04:03:18 -0700520static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter81819f02007-05-06 14:49:36 -0700521{
522 while (bytes) {
523 if (*start != (u8)value)
Christoph Lameter24922682007-07-17 04:03:18 -0700524 return start;
Christoph Lameter81819f02007-05-06 14:49:36 -0700525 start++;
526 bytes--;
527 }
Christoph Lameter24922682007-07-17 04:03:18 -0700528 return NULL;
529}
530
531static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
532 void *from, void *to)
533{
534 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
535 memset(from, data, to - from);
536}
537
538static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
539 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800540 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700541{
542 u8 *fault;
543 u8 *end;
544
545 fault = check_bytes(start, value, bytes);
546 if (!fault)
547 return 1;
548
549 end = start + bytes;
550 while (end > fault && end[-1] == value)
551 end--;
552
553 slab_bug(s, "%s overwritten", what);
554 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
555 fault, end - 1, fault[0], value);
556 print_trailer(s, page, object);
557
558 restore_bytes(s, what, value, fault, end);
559 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700560}
561
Christoph Lameter81819f02007-05-06 14:49:36 -0700562/*
563 * Object layout:
564 *
565 * object address
566 * Bytes of the object to be managed.
567 * If the freepointer may overlay the object then the free
568 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700569 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700570 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
571 * 0xa5 (POISON_END)
572 *
573 * object + s->objsize
574 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700575 * Padding is extended by another word if Redzoning is enabled and
576 * objsize == inuse.
577 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700578 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
579 * 0xcc (RED_ACTIVE) for objects in use.
580 *
581 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700582 * Meta data starts here.
583 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700584 * A. Free pointer (if we cannot overwrite object on free)
585 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700586 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800587 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700588 * before the word boundary.
589 *
590 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700591 *
592 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700593 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700594 *
Christoph Lameter672bba32007-05-09 02:32:39 -0700595 * If slabcaches are merged then the objsize and inuse boundaries are mostly
596 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700597 * may be used with merged slabcaches.
598 */
599
Christoph Lameter81819f02007-05-06 14:49:36 -0700600static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
601{
602 unsigned long off = s->inuse; /* The end of info */
603
604 if (s->offset)
605 /* Freepointer is placed after the object. */
606 off += sizeof(void *);
607
608 if (s->flags & SLAB_STORE_USER)
609 /* We also have user information there */
610 off += 2 * sizeof(struct track);
611
612 if (s->size == off)
613 return 1;
614
Christoph Lameter24922682007-07-17 04:03:18 -0700615 return check_bytes_and_report(s, page, p, "Object padding",
616 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700617}
618
Christoph Lameter39b26462008-04-14 19:11:30 +0300619/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700620static int slab_pad_check(struct kmem_cache *s, struct page *page)
621{
Christoph Lameter24922682007-07-17 04:03:18 -0700622 u8 *start;
623 u8 *fault;
624 u8 *end;
625 int length;
626 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700627
628 if (!(s->flags & SLAB_POISON))
629 return 1;
630
Christoph Lametera973e9d2008-03-01 13:40:44 -0800631 start = page_address(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +0300632 length = (PAGE_SIZE << compound_order(page));
Christoph Lameter39b26462008-04-14 19:11:30 +0300633 end = start + length;
634 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700635 if (!remainder)
636 return 1;
637
Christoph Lameter39b26462008-04-14 19:11:30 +0300638 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700639 if (!fault)
640 return 1;
641 while (end > fault && end[-1] == POISON_INUSE)
642 end--;
643
644 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Christoph Lameter39b26462008-04-14 19:11:30 +0300645 print_section("Padding", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700646
647 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
648 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700649}
650
651static int check_object(struct kmem_cache *s, struct page *page,
652 void *object, int active)
653{
654 u8 *p = object;
655 u8 *endobject = object + s->objsize;
656
657 if (s->flags & SLAB_RED_ZONE) {
658 unsigned int red =
659 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
660
Christoph Lameter24922682007-07-17 04:03:18 -0700661 if (!check_bytes_and_report(s, page, object, "Redzone",
662 endobject, red, s->inuse - s->objsize))
Christoph Lameter81819f02007-05-06 14:49:36 -0700663 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700664 } else {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800665 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
666 check_bytes_and_report(s, page, p, "Alignment padding",
667 endobject, POISON_INUSE, s->inuse - s->objsize);
668 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700669 }
670
671 if (s->flags & SLAB_POISON) {
672 if (!active && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700673 (!check_bytes_and_report(s, page, p, "Poison", p,
674 POISON_FREE, s->objsize - 1) ||
675 !check_bytes_and_report(s, page, p, "Poison",
Pekka Enberg06428782008-01-07 23:20:27 -0800676 p + s->objsize - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700677 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700678 /*
679 * check_pad_bytes cleans up on its own.
680 */
681 check_pad_bytes(s, page, p);
682 }
683
684 if (!s->offset && active)
685 /*
686 * Object and freepointer overlap. Cannot check
687 * freepointer while object is allocated.
688 */
689 return 1;
690
691 /* Check free pointer validity */
692 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
693 object_err(s, page, p, "Freepointer corrupt");
694 /*
695 * No choice but to zap it and thus loose the remainder
696 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700697 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700698 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800699 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700700 return 0;
701 }
702 return 1;
703}
704
705static int check_slab(struct kmem_cache *s, struct page *page)
706{
Christoph Lameter39b26462008-04-14 19:11:30 +0300707 int maxobj;
708
Christoph Lameter81819f02007-05-06 14:49:36 -0700709 VM_BUG_ON(!irqs_disabled());
710
711 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700712 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700713 return 0;
714 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300715
716 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
717 if (page->objects > maxobj) {
718 slab_err(s, page, "objects %u > max %u",
719 s->name, page->objects, maxobj);
720 return 0;
721 }
722 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700723 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300724 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700725 return 0;
726 }
727 /* Slab_pad_check fixes things up after itself */
728 slab_pad_check(s, page);
729 return 1;
730}
731
732/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700733 * Determine if a certain object on a page is on the freelist. Must hold the
734 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700735 */
736static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
737{
738 int nr = 0;
739 void *fp = page->freelist;
740 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300741 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700742
Christoph Lameter39b26462008-04-14 19:11:30 +0300743 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700744 if (fp == search)
745 return 1;
746 if (!check_valid_pointer(s, page, fp)) {
747 if (object) {
748 object_err(s, page, object,
749 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800750 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700751 break;
752 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700753 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800754 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300755 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700756 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700757 return 0;
758 }
759 break;
760 }
761 object = fp;
762 fp = get_freepointer(s, object);
763 nr++;
764 }
765
Christoph Lameter224a88b2008-04-14 19:11:31 +0300766 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
767 if (max_objects > 65535)
768 max_objects = 65535;
769
770 if (page->objects != max_objects) {
771 slab_err(s, page, "Wrong number of objects. Found %d but "
772 "should be %d", page->objects, max_objects);
773 page->objects = max_objects;
774 slab_fix(s, "Number of objects adjusted.");
775 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300776 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700777 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300778 "counted were %d", page->inuse, page->objects - nr);
779 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700780 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700781 }
782 return search == NULL;
783}
784
Christoph Lameter0121c6192008-04-29 16:11:12 -0700785static void trace(struct kmem_cache *s, struct page *page, void *object,
786 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700787{
788 if (s->flags & SLAB_TRACE) {
789 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
790 s->name,
791 alloc ? "alloc" : "free",
792 object, page->inuse,
793 page->freelist);
794
795 if (!alloc)
796 print_section("Object", (void *)object, s->objsize);
797
798 dump_stack();
799 }
800}
801
Christoph Lameter643b1132007-05-06 14:49:42 -0700802/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700803 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700804 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700805static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700806{
Christoph Lameter643b1132007-05-06 14:49:42 -0700807 spin_lock(&n->list_lock);
808 list_add(&page->lru, &n->full);
809 spin_unlock(&n->list_lock);
810}
811
812static void remove_full(struct kmem_cache *s, struct page *page)
813{
814 struct kmem_cache_node *n;
815
816 if (!(s->flags & SLAB_STORE_USER))
817 return;
818
819 n = get_node(s, page_to_nid(page));
820
821 spin_lock(&n->list_lock);
822 list_del(&page->lru);
823 spin_unlock(&n->list_lock);
824}
825
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300826/* Tracking of the number of slabs for debugging purposes */
827static inline unsigned long slabs_node(struct kmem_cache *s, int node)
828{
829 struct kmem_cache_node *n = get_node(s, node);
830
831 return atomic_long_read(&n->nr_slabs);
832}
833
Christoph Lameter205ab992008-04-14 19:11:40 +0300834static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300835{
836 struct kmem_cache_node *n = get_node(s, node);
837
838 /*
839 * May be called early in order to allocate a slab for the
840 * kmem_cache_node structure. Solve the chicken-egg
841 * dilemma by deferring the increment of the count during
842 * bootstrap (see early_kmem_cache_node_alloc).
843 */
Christoph Lameter205ab992008-04-14 19:11:40 +0300844 if (!NUMA_BUILD || n) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300845 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300846 atomic_long_add(objects, &n->total_objects);
847 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300848}
Christoph Lameter205ab992008-04-14 19:11:40 +0300849static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300850{
851 struct kmem_cache_node *n = get_node(s, node);
852
853 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300854 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300855}
856
857/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -0700858static void setup_object_debug(struct kmem_cache *s, struct page *page,
859 void *object)
860{
861 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
862 return;
863
864 init_object(s, object, 0);
865 init_tracking(s, object);
866}
867
868static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
869 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700870{
871 if (!check_slab(s, page))
872 goto bad;
873
Christoph Lameterd692ef62008-02-15 23:45:24 -0800874 if (!on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700875 object_err(s, page, object, "Object already allocated");
Christoph Lameter70d71222007-05-06 14:49:47 -0700876 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700877 }
878
879 if (!check_valid_pointer(s, page, object)) {
880 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700881 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700882 }
883
Christoph Lameterd692ef62008-02-15 23:45:24 -0800884 if (!check_object(s, page, object, 0))
Christoph Lameter81819f02007-05-06 14:49:36 -0700885 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700886
Christoph Lameter3ec09742007-05-16 22:11:00 -0700887 /* Success perform special debug activities for allocs */
888 if (s->flags & SLAB_STORE_USER)
889 set_track(s, object, TRACK_ALLOC, addr);
890 trace(s, page, object, 1);
891 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700892 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700893
Christoph Lameter81819f02007-05-06 14:49:36 -0700894bad:
895 if (PageSlab(page)) {
896 /*
897 * If this is a slab page then lets do the best we can
898 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -0700899 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -0700900 */
Christoph Lameter24922682007-07-17 04:03:18 -0700901 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +0300902 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -0800903 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -0700904 }
905 return 0;
906}
907
Christoph Lameter3ec09742007-05-16 22:11:00 -0700908static int free_debug_processing(struct kmem_cache *s, struct page *page,
909 void *object, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700910{
911 if (!check_slab(s, page))
912 goto fail;
913
914 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700915 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700916 goto fail;
917 }
918
919 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700920 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -0700921 goto fail;
922 }
923
924 if (!check_object(s, page, object, 1))
925 return 0;
926
927 if (unlikely(s != page->slab)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800928 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700929 slab_err(s, page, "Attempt to free object(0x%p) "
930 "outside of slab", object);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800931 } else if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700932 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700933 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700934 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700935 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -0800936 } else
Christoph Lameter24922682007-07-17 04:03:18 -0700937 object_err(s, page, object,
938 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700939 goto fail;
940 }
Christoph Lameter3ec09742007-05-16 22:11:00 -0700941
942 /* Special debug activities for freeing objects */
Andy Whitcroft8a380822008-07-23 21:27:18 -0700943 if (!PageSlubFrozen(page) && !page->freelist)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700944 remove_full(s, page);
945 if (s->flags & SLAB_STORE_USER)
946 set_track(s, object, TRACK_FREE, addr);
947 trace(s, page, object, 0);
948 init_object(s, object, 0);
Christoph Lameter81819f02007-05-06 14:49:36 -0700949 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -0700950
Christoph Lameter81819f02007-05-06 14:49:36 -0700951fail:
Christoph Lameter24922682007-07-17 04:03:18 -0700952 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700953 return 0;
954}
955
Christoph Lameter41ecc552007-05-09 02:32:44 -0700956static int __init setup_slub_debug(char *str)
957{
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700958 slub_debug = DEBUG_DEFAULT_FLAGS;
959 if (*str++ != '=' || !*str)
960 /*
961 * No options specified. Switch on full debugging.
962 */
963 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700964
965 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700966 /*
967 * No options but restriction on slabs. This means full
968 * debugging for slabs matching a pattern.
969 */
970 goto check_slabs;
971
972 slub_debug = 0;
973 if (*str == '-')
974 /*
975 * Switch off all debugging measures.
976 */
977 goto out;
978
979 /*
980 * Determine which debug features should be switched on
981 */
Pekka Enberg06428782008-01-07 23:20:27 -0800982 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700983 switch (tolower(*str)) {
984 case 'f':
985 slub_debug |= SLAB_DEBUG_FREE;
986 break;
987 case 'z':
988 slub_debug |= SLAB_RED_ZONE;
989 break;
990 case 'p':
991 slub_debug |= SLAB_POISON;
992 break;
993 case 'u':
994 slub_debug |= SLAB_STORE_USER;
995 break;
996 case 't':
997 slub_debug |= SLAB_TRACE;
998 break;
999 default:
1000 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001001 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001002 }
1003 }
1004
1005check_slabs:
1006 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001007 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001008out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001009 return 1;
1010}
1011
1012__setup("slub_debug", setup_slub_debug);
1013
Christoph Lameterba0268a2007-09-11 15:24:11 -07001014static unsigned long kmem_cache_flags(unsigned long objsize,
1015 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001016 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001017{
1018 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001019 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001020 */
Christoph Lametere1533622008-02-15 23:45:24 -08001021 if (slub_debug && (!slub_debug_slabs ||
1022 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1023 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001024
1025 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001026}
1027#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001028static inline void setup_object_debug(struct kmem_cache *s,
1029 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001030
Christoph Lameter3ec09742007-05-16 22:11:00 -07001031static inline int alloc_debug_processing(struct kmem_cache *s,
1032 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001033
Christoph Lameter3ec09742007-05-16 22:11:00 -07001034static inline int free_debug_processing(struct kmem_cache *s,
1035 struct page *page, void *object, void *addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001036
Christoph Lameter41ecc552007-05-09 02:32:44 -07001037static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1038 { return 1; }
1039static inline int check_object(struct kmem_cache *s, struct page *page,
1040 void *object, int active) { return 1; }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001041static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
Christoph Lameterba0268a2007-09-11 15:24:11 -07001042static inline unsigned long kmem_cache_flags(unsigned long objsize,
1043 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001044 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001045{
1046 return flags;
1047}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001048#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001049
1050static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1051 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001052static inline void inc_slabs_node(struct kmem_cache *s, int node,
1053 int objects) {}
1054static inline void dec_slabs_node(struct kmem_cache *s, int node,
1055 int objects) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001056#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03001057
Christoph Lameter81819f02007-05-06 14:49:36 -07001058/*
1059 * Slab allocation and freeing
1060 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001061static inline struct page *alloc_slab_page(gfp_t flags, int node,
1062 struct kmem_cache_order_objects oo)
1063{
1064 int order = oo_order(oo);
1065
1066 if (node == -1)
1067 return alloc_pages(flags, order);
1068 else
1069 return alloc_pages_node(node, flags, order);
1070}
1071
Christoph Lameter81819f02007-05-06 14:49:36 -07001072static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1073{
Pekka Enberg06428782008-01-07 23:20:27 -08001074 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001075 struct kmem_cache_order_objects oo = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07001076
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001077 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001078
Christoph Lameter65c33762008-04-14 19:11:40 +03001079 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1080 oo);
1081 if (unlikely(!page)) {
1082 oo = s->min;
1083 /*
1084 * Allocation may have failed due to fragmentation.
1085 * Try a lower order alloc if possible
1086 */
1087 page = alloc_slab_page(flags, node, oo);
1088 if (!page)
1089 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001090
Christoph Lameter65c33762008-04-14 19:11:40 +03001091 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1092 }
Christoph Lameter834f3d12008-04-14 19:11:31 +03001093 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001094 mod_zone_page_state(page_zone(page),
1095 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1096 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001097 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001098
1099 return page;
1100}
1101
1102static void setup_object(struct kmem_cache *s, struct page *page,
1103 void *object)
1104{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001105 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001106 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001107 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001108}
1109
1110static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1111{
1112 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001113 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001114 void *last;
1115 void *p;
1116
Christoph Lameter6cb06222007-10-16 01:25:41 -07001117 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001118
Christoph Lameter6cb06222007-10-16 01:25:41 -07001119 page = allocate_slab(s,
1120 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001121 if (!page)
1122 goto out;
1123
Christoph Lameter205ab992008-04-14 19:11:40 +03001124 inc_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001125 page->slab = s;
1126 page->flags |= 1 << PG_slab;
1127 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1128 SLAB_STORE_USER | SLAB_TRACE))
Andy Whitcroft8a380822008-07-23 21:27:18 -07001129 __SetPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001130
1131 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001132
1133 if (unlikely(s->flags & SLAB_POISON))
Christoph Lameter834f3d12008-04-14 19:11:31 +03001134 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
Christoph Lameter81819f02007-05-06 14:49:36 -07001135
1136 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001137 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001138 setup_object(s, page, last);
1139 set_freepointer(s, last, p);
1140 last = p;
1141 }
1142 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001143 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001144
1145 page->freelist = start;
1146 page->inuse = 0;
1147out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001148 return page;
1149}
1150
1151static void __free_slab(struct kmem_cache *s, struct page *page)
1152{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001153 int order = compound_order(page);
1154 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001155
Andy Whitcroft8a380822008-07-23 21:27:18 -07001156 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001157 void *p;
1158
1159 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001160 for_each_object(p, s, page_address(page),
1161 page->objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001162 check_object(s, page, p, 0);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001163 __ClearPageSlubDebug(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001164 }
1165
1166 mod_zone_page_state(page_zone(page),
1167 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1168 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001169 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001170
Christoph Lameter49bd5222008-04-14 18:52:18 +03001171 __ClearPageSlab(page);
1172 reset_page_mapcount(page);
Christoph Lameter834f3d12008-04-14 19:11:31 +03001173 __free_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001174}
1175
1176static void rcu_free_slab(struct rcu_head *h)
1177{
1178 struct page *page;
1179
1180 page = container_of((struct list_head *)h, struct page, lru);
1181 __free_slab(page->slab, page);
1182}
1183
1184static void free_slab(struct kmem_cache *s, struct page *page)
1185{
1186 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1187 /*
1188 * RCU free overloads the RCU head over the LRU
1189 */
1190 struct rcu_head *head = (void *)&page->lru;
1191
1192 call_rcu(head, rcu_free_slab);
1193 } else
1194 __free_slab(s, page);
1195}
1196
1197static void discard_slab(struct kmem_cache *s, struct page *page)
1198{
Christoph Lameter205ab992008-04-14 19:11:40 +03001199 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001200 free_slab(s, page);
1201}
1202
1203/*
1204 * Per slab locking using the pagelock
1205 */
1206static __always_inline void slab_lock(struct page *page)
1207{
1208 bit_spin_lock(PG_locked, &page->flags);
1209}
1210
1211static __always_inline void slab_unlock(struct page *page)
1212{
Nick Piggina76d3542008-01-07 23:20:27 -08001213 __bit_spin_unlock(PG_locked, &page->flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001214}
1215
1216static __always_inline int slab_trylock(struct page *page)
1217{
1218 int rc = 1;
1219
1220 rc = bit_spin_trylock(PG_locked, &page->flags);
1221 return rc;
1222}
1223
1224/*
1225 * Management of partially allocated slabs
1226 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001227static void add_partial(struct kmem_cache_node *n,
1228 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001229{
Christoph Lametere95eed52007-05-06 14:49:44 -07001230 spin_lock(&n->list_lock);
1231 n->nr_partial++;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001232 if (tail)
1233 list_add_tail(&page->lru, &n->partial);
1234 else
1235 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001236 spin_unlock(&n->list_lock);
1237}
1238
Christoph Lameter0121c6192008-04-29 16:11:12 -07001239static void remove_partial(struct kmem_cache *s, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001240{
1241 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1242
1243 spin_lock(&n->list_lock);
1244 list_del(&page->lru);
1245 n->nr_partial--;
1246 spin_unlock(&n->list_lock);
1247}
1248
1249/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001250 * Lock slab and remove from the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001251 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001252 * Must hold list_lock.
Christoph Lameter81819f02007-05-06 14:49:36 -07001253 */
Christoph Lameter0121c6192008-04-29 16:11:12 -07001254static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1255 struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001256{
1257 if (slab_trylock(page)) {
1258 list_del(&page->lru);
1259 n->nr_partial--;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001260 __SetPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001261 return 1;
1262 }
1263 return 0;
1264}
1265
1266/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001267 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001268 */
1269static struct page *get_partial_node(struct kmem_cache_node *n)
1270{
1271 struct page *page;
1272
1273 /*
1274 * Racy check. If we mistakenly see no partial slabs then we
1275 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001276 * partial slab and there is none available then get_partials()
1277 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001278 */
1279 if (!n || !n->nr_partial)
1280 return NULL;
1281
1282 spin_lock(&n->list_lock);
1283 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001284 if (lock_and_freeze_slab(n, page))
Christoph Lameter81819f02007-05-06 14:49:36 -07001285 goto out;
1286 page = NULL;
1287out:
1288 spin_unlock(&n->list_lock);
1289 return page;
1290}
1291
1292/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001293 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001294 */
1295static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1296{
1297#ifdef CONFIG_NUMA
1298 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001299 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001300 struct zone *zone;
1301 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07001302 struct page *page;
1303
1304 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001305 * The defrag ratio allows a configuration of the tradeoffs between
1306 * inter node defragmentation and node local allocations. A lower
1307 * defrag_ratio increases the tendency to do local allocations
1308 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001309 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001310 * If the defrag_ratio is set to 0 then kmalloc() always
1311 * returns node local objects. If the ratio is higher then kmalloc()
1312 * may return off node objects because partial slabs are obtained
1313 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001314 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001315 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001316 * defrag_ratio = 1000) then every (well almost) allocation will
1317 * first attempt to defrag slab caches on other nodes. This means
1318 * scanning over all nodes to look for partial slabs which may be
1319 * expensive if we do it every time we are trying to find a slab
1320 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001321 */
Christoph Lameter98246012008-01-07 23:20:26 -08001322 if (!s->remote_node_defrag_ratio ||
1323 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001324 return NULL;
1325
Mel Gorman0e884602008-04-28 02:12:14 -07001326 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001327 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001328 struct kmem_cache_node *n;
1329
Mel Gorman54a6eb52008-04-28 02:12:16 -07001330 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001331
Mel Gorman54a6eb52008-04-28 02:12:16 -07001332 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
Pekka Enberg5595cff2008-08-05 09:28:47 +03001333 n->nr_partial > n->min_partial) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001334 page = get_partial_node(n);
1335 if (page)
1336 return page;
1337 }
1338 }
1339#endif
1340 return NULL;
1341}
1342
1343/*
1344 * Get a partial page, lock it and return it.
1345 */
1346static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1347{
1348 struct page *page;
1349 int searchnode = (node == -1) ? numa_node_id() : node;
1350
1351 page = get_partial_node(get_node(s, searchnode));
1352 if (page || (flags & __GFP_THISNODE))
1353 return page;
1354
1355 return get_any_partial(s, flags);
1356}
1357
1358/*
1359 * Move a page back to the lists.
1360 *
1361 * Must be called with the slab lock held.
1362 *
1363 * On exit the slab lock will have been dropped.
1364 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001365static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001366{
Christoph Lametere95eed52007-05-06 14:49:44 -07001367 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001368 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
Christoph Lametere95eed52007-05-06 14:49:44 -07001369
Andy Whitcroft8a380822008-07-23 21:27:18 -07001370 __ClearPageSlubFrozen(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001371 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001372
Christoph Lametera973e9d2008-03-01 13:40:44 -08001373 if (page->freelist) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001374 add_partial(n, page, tail);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001375 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1376 } else {
1377 stat(c, DEACTIVATE_FULL);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001378 if (SLABDEBUG && PageSlubDebug(page) &&
1379 (s->flags & SLAB_STORE_USER))
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001380 add_full(n, page);
1381 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001382 slab_unlock(page);
1383 } else {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001384 stat(c, DEACTIVATE_EMPTY);
Pekka Enberg5595cff2008-08-05 09:28:47 +03001385 if (n->nr_partial < n->min_partial) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001386 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001387 * Adding an empty slab to the partial slabs in order
1388 * to avoid page allocator overhead. This slab needs
1389 * to come after the other slabs with objects in
Christoph Lameter6446faa2008-02-15 23:45:26 -08001390 * so that the others get filled first. That way the
1391 * size of the partial list stays small.
1392 *
Christoph Lameter0121c6192008-04-29 16:11:12 -07001393 * kmem_cache_shrink can reclaim any empty slabs from
1394 * the partial list.
Christoph Lametere95eed52007-05-06 14:49:44 -07001395 */
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001396 add_partial(n, page, 1);
Christoph Lametere95eed52007-05-06 14:49:44 -07001397 slab_unlock(page);
1398 } else {
1399 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001400 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
Christoph Lametere95eed52007-05-06 14:49:44 -07001401 discard_slab(s, page);
1402 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001403 }
1404}
1405
1406/*
1407 * Remove the cpu slab
1408 */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001409static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001410{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001411 struct page *page = c->page;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001412 int tail = 1;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001413
Christoph Lameterb773ad72008-03-04 11:10:17 -08001414 if (page->freelist)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001415 stat(c, DEACTIVATE_REMOTE_FREES);
Christoph Lameter894b8782007-05-10 03:15:16 -07001416 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001417 * Merge cpu freelist into slab freelist. Typically we get here
Christoph Lameter894b8782007-05-10 03:15:16 -07001418 * because both freelists are empty. So this is unlikely
1419 * to occur.
1420 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001421 while (unlikely(c->freelist)) {
Christoph Lameter894b8782007-05-10 03:15:16 -07001422 void **object;
1423
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001424 tail = 0; /* Hot objects. Put the slab first */
1425
Christoph Lameter894b8782007-05-10 03:15:16 -07001426 /* Retrieve object from cpu_freelist */
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001427 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001428 c->freelist = c->freelist[c->offset];
Christoph Lameter894b8782007-05-10 03:15:16 -07001429
1430 /* And put onto the regular freelist */
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001431 object[c->offset] = page->freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07001432 page->freelist = object;
1433 page->inuse--;
1434 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001435 c->page = NULL;
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001436 unfreeze_slab(s, page, tail);
Christoph Lameter81819f02007-05-06 14:49:36 -07001437}
1438
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001439static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001440{
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001441 stat(c, CPUSLAB_FLUSH);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001442 slab_lock(c->page);
1443 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001444}
1445
1446/*
1447 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08001448 *
Christoph Lameter81819f02007-05-06 14:49:36 -07001449 * Called from IPI handler with interrupts disabled.
1450 */
Christoph Lameter0c710012007-07-17 04:03:24 -07001451static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07001452{
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001453 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07001454
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001455 if (likely(c && c->page))
1456 flush_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001457}
1458
1459static void flush_cpu_slab(void *d)
1460{
1461 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07001462
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001463 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07001464}
1465
1466static void flush_all(struct kmem_cache *s)
1467{
Jens Axboe15c8b6c2008-05-09 09:39:44 +02001468 on_each_cpu(flush_cpu_slab, s, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07001469}
1470
1471/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001472 * Check if the objects in a per cpu structure fit numa
1473 * locality expectations.
1474 */
1475static inline int node_match(struct kmem_cache_cpu *c, int node)
1476{
1477#ifdef CONFIG_NUMA
1478 if (node != -1 && c->node != node)
1479 return 0;
1480#endif
1481 return 1;
1482}
1483
1484/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001485 * Slow path. The lockless freelist is empty or we need to perform
1486 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07001487 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001488 * Interrupts are disabled.
Christoph Lameter81819f02007-05-06 14:49:36 -07001489 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001490 * Processing is still very fast if new objects have been freed to the
1491 * regular freelist. In that case we simply take over the regular freelist
1492 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001493 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001494 * If that is not working then we fall back to the partial lists. We take the
1495 * first element of the freelist as the object to allocate now and move the
1496 * rest of the freelist to the lockless freelist.
1497 *
1498 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08001499 * we need to allocate a new slab. This is the slowest path since it involves
1500 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07001501 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001502static void *__slab_alloc(struct kmem_cache *s,
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001503 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001504{
Christoph Lameter81819f02007-05-06 14:49:36 -07001505 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001506 struct page *new;
Christoph Lameter81819f02007-05-06 14:49:36 -07001507
Linus Torvaldse72e9c22008-03-27 20:56:33 -07001508 /* We handle __GFP_ZERO in the caller */
1509 gfpflags &= ~__GFP_ZERO;
1510
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001511 if (!c->page)
Christoph Lameter81819f02007-05-06 14:49:36 -07001512 goto new_slab;
1513
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001514 slab_lock(c->page);
1515 if (unlikely(!node_match(c, node)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001516 goto another_slab;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001517
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001518 stat(c, ALLOC_REFILL);
Christoph Lameter6446faa2008-02-15 23:45:26 -08001519
Christoph Lameter894b8782007-05-10 03:15:16 -07001520load_freelist:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001521 object = c->page->freelist;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001522 if (unlikely(!object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001523 goto another_slab;
Andy Whitcroft8a380822008-07-23 21:27:18 -07001524 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001525 goto debug;
1526
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001527 c->freelist = object[c->offset];
Christoph Lameter39b26462008-04-14 19:11:30 +03001528 c->page->inuse = c->page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001529 c->page->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001530 c->node = page_to_nid(c->page);
Christoph Lameter1f842602008-01-07 23:20:30 -08001531unlock_out:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001532 slab_unlock(c->page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001533 stat(c, ALLOC_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001534 return object;
1535
1536another_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001537 deactivate_slab(s, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001538
1539new_slab:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001540 new = get_partial(s, gfpflags, node);
1541 if (new) {
1542 c->page = new;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001543 stat(c, ALLOC_FROM_PARTIAL);
Christoph Lameter894b8782007-05-10 03:15:16 -07001544 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001545 }
1546
Christoph Lameterb811c202007-10-16 23:25:51 -07001547 if (gfpflags & __GFP_WAIT)
1548 local_irq_enable();
1549
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001550 new = new_slab(s, gfpflags, node);
Christoph Lameterb811c202007-10-16 23:25:51 -07001551
1552 if (gfpflags & __GFP_WAIT)
1553 local_irq_disable();
1554
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001555 if (new) {
1556 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001557 stat(c, ALLOC_SLAB);
Christoph Lameter05aa3452007-11-05 11:31:58 -08001558 if (c->page)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001559 flush_slab(s, c);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001560 slab_lock(new);
Andy Whitcroft8a380822008-07-23 21:27:18 -07001561 __SetPageSlubFrozen(new);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001562 c->page = new;
Christoph Lameter4b6f0752007-05-16 22:10:53 -07001563 goto load_freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001564 }
Christoph Lameter71c7a062008-02-14 14:28:01 -08001565 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001566debug:
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001567 if (!alloc_debug_processing(s, c->page, object, addr))
Christoph Lameter81819f02007-05-06 14:49:36 -07001568 goto another_slab;
Christoph Lameter894b8782007-05-10 03:15:16 -07001569
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001570 c->page->inuse++;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001571 c->page->freelist = object[c->offset];
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001572 c->node = -1;
Christoph Lameter1f842602008-01-07 23:20:30 -08001573 goto unlock_out;
Christoph Lameter894b8782007-05-10 03:15:16 -07001574}
1575
1576/*
1577 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1578 * have the fastpath folded into their functions. So no function call
1579 * overhead for requests that can be satisfied on the fastpath.
1580 *
1581 * The fastpath works by first checking if the lockless freelist can be used.
1582 * If not then __slab_alloc is called for slow processing.
1583 *
1584 * Otherwise we can simply pick the next object from the lockless free list.
1585 */
Pekka Enberg06428782008-01-07 23:20:27 -08001586static __always_inline void *slab_alloc(struct kmem_cache *s,
Christoph Lameterce15fea2007-07-17 04:03:28 -07001587 gfp_t gfpflags, int node, void *addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07001588{
Christoph Lameter894b8782007-05-10 03:15:16 -07001589 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001590 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001591 unsigned long flags;
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001592 unsigned int objsize;
Christoph Lameter1f842602008-01-07 23:20:30 -08001593
Christoph Lameter894b8782007-05-10 03:15:16 -07001594 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001595 c = get_cpu_slab(s, smp_processor_id());
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001596 objsize = c->objsize;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001597 if (unlikely(!c->freelist || !node_match(c, node)))
Christoph Lameter894b8782007-05-10 03:15:16 -07001598
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001599 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07001600
1601 else {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001602 object = c->freelist;
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001603 c->freelist = object[c->offset];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001604 stat(c, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001605 }
1606 local_irq_restore(flags);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001607
1608 if (unlikely((gfpflags & __GFP_ZERO) && object))
Dmitry Adamushkobdb21922008-07-10 22:21:58 +02001609 memset(object, 0, objsize);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07001610
Christoph Lameter894b8782007-05-10 03:15:16 -07001611 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001612}
1613
1614void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1615{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001616 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001617}
1618EXPORT_SYMBOL(kmem_cache_alloc);
1619
1620#ifdef CONFIG_NUMA
1621void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1622{
Christoph Lameterce15fea2007-07-17 04:03:28 -07001623 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001624}
1625EXPORT_SYMBOL(kmem_cache_alloc_node);
1626#endif
1627
1628/*
Christoph Lameter894b8782007-05-10 03:15:16 -07001629 * Slow patch handling. This may still be called frequently since objects
1630 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07001631 *
Christoph Lameter894b8782007-05-10 03:15:16 -07001632 * So we still attempt to reduce cache line usage. Just take the slab
1633 * lock and free the item. If there is no additional partial page
1634 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07001635 */
Christoph Lameter894b8782007-05-10 03:15:16 -07001636static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001637 void *x, void *addr, unsigned int offset)
Christoph Lameter81819f02007-05-06 14:49:36 -07001638{
1639 void *prior;
1640 void **object = (void *)x;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001641 struct kmem_cache_cpu *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07001642
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001643 c = get_cpu_slab(s, raw_smp_processor_id());
1644 stat(c, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07001645 slab_lock(page);
1646
Andy Whitcroft8a380822008-07-23 21:27:18 -07001647 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
Christoph Lameter81819f02007-05-06 14:49:36 -07001648 goto debug;
Christoph Lameter6446faa2008-02-15 23:45:26 -08001649
Christoph Lameter81819f02007-05-06 14:49:36 -07001650checks_ok:
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001651 prior = object[offset] = page->freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001652 page->freelist = object;
1653 page->inuse--;
1654
Andy Whitcroft8a380822008-07-23 21:27:18 -07001655 if (unlikely(PageSlubFrozen(page))) {
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001656 stat(c, FREE_FROZEN);
Christoph Lameter81819f02007-05-06 14:49:36 -07001657 goto out_unlock;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001658 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001659
1660 if (unlikely(!page->inuse))
1661 goto slab_empty;
1662
1663 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001664 * Objects left in the slab. If it was not on the partial list before
Christoph Lameter81819f02007-05-06 14:49:36 -07001665 * then add it.
1666 */
Christoph Lametera973e9d2008-03-01 13:40:44 -08001667 if (unlikely(!prior)) {
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001668 add_partial(get_node(s, page_to_nid(page)), page, 1);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001669 stat(c, FREE_ADD_PARTIAL);
1670 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001671
1672out_unlock:
1673 slab_unlock(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001674 return;
1675
1676slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08001677 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001678 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001679 * Slab still on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001680 */
1681 remove_partial(s, page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001682 stat(c, FREE_REMOVE_PARTIAL);
1683 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001684 slab_unlock(page);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001685 stat(c, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07001686 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001687 return;
1688
1689debug:
Christoph Lameter3ec09742007-05-16 22:11:00 -07001690 if (!free_debug_processing(s, page, x, addr))
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001691 goto out_unlock;
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001692 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001693}
1694
Christoph Lameter894b8782007-05-10 03:15:16 -07001695/*
1696 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1697 * can perform fastpath freeing without additional function calls.
1698 *
1699 * The fastpath is only possible if we are freeing to the current cpu slab
1700 * of this processor. This typically the case if we have just allocated
1701 * the item before.
1702 *
1703 * If fastpath is not possible then fall back to __slab_free where we deal
1704 * with all sorts of special processing.
1705 */
Pekka Enberg06428782008-01-07 23:20:27 -08001706static __always_inline void slab_free(struct kmem_cache *s,
Christoph Lameter894b8782007-05-10 03:15:16 -07001707 struct page *page, void *x, void *addr)
1708{
1709 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001710 struct kmem_cache_cpu *c;
Christoph Lameter1f842602008-01-07 23:20:30 -08001711 unsigned long flags;
1712
Christoph Lameter894b8782007-05-10 03:15:16 -07001713 local_irq_save(flags);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001714 c = get_cpu_slab(s, smp_processor_id());
Christoph Lameter27d9e4e2008-02-15 23:45:25 -08001715 debug_check_no_locks_freed(object, c->objsize);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001716 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1717 debug_check_no_obj_freed(object, s->objsize);
Christoph Lameteree3c72a2007-10-16 01:26:07 -07001718 if (likely(page == c->page && c->node >= 0)) {
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001719 object[c->offset] = c->freelist;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001720 c->freelist = object;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001721 stat(c, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07001722 } else
Christoph Lameterb3fba8d2007-10-16 01:26:06 -07001723 __slab_free(s, page, x, addr, c->offset);
Christoph Lameter894b8782007-05-10 03:15:16 -07001724
1725 local_irq_restore(flags);
1726}
1727
Christoph Lameter81819f02007-05-06 14:49:36 -07001728void kmem_cache_free(struct kmem_cache *s, void *x)
1729{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001730 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001731
Christoph Lameterb49af682007-05-06 14:49:41 -07001732 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001733
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001734 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001735}
1736EXPORT_SYMBOL(kmem_cache_free);
1737
1738/* Figure out on which slab object the object resides */
1739static struct page *get_object_page(const void *x)
1740{
Christoph Lameterb49af682007-05-06 14:49:41 -07001741 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001742
1743 if (!PageSlab(page))
1744 return NULL;
1745
1746 return page;
1747}
1748
1749/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001750 * Object placement in a slab is made very easy because we always start at
1751 * offset 0. If we tune the size of the object to the alignment then we can
1752 * get the required alignment by putting one properly sized object after
1753 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07001754 *
1755 * Notice that the allocation order determines the sizes of the per cpu
1756 * caches. Each processor has always one slab available for allocations.
1757 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07001758 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07001759 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07001760 */
1761
1762/*
1763 * Mininum / Maximum order of slab pages. This influences locking overhead
1764 * and slab fragmentation. A higher order reduces the number of partial slabs
1765 * and increases the number of allocations possible without having to
1766 * take the list_lock.
1767 */
1768static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03001769static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001770static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001771
1772/*
1773 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07001774 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07001775 */
1776static int slub_nomerge;
1777
1778/*
Christoph Lameter81819f02007-05-06 14:49:36 -07001779 * Calculate the order of allocation given an slab object size.
1780 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001781 * The order of allocation has significant impact on performance and other
1782 * system components. Generally order 0 allocations should be preferred since
1783 * order 0 does not cause fragmentation in the page allocator. Larger objects
1784 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001785 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07001786 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07001787 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001788 * In order to reach satisfactory performance we must ensure that a minimum
1789 * number of objects is in one slab. Otherwise we may generate too much
1790 * activity on the partial lists which requires taking the list_lock. This is
1791 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07001792 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001793 * slub_max_order specifies the order where we begin to stop considering the
1794 * number of objects in a slab as critical. If we reach slub_max_order then
1795 * we try to keep the page order as low as possible. So we accept more waste
1796 * of space in favor of a small page order.
1797 *
1798 * Higher order allocations also allow the placement of more objects in a
1799 * slab and thereby reduce object handling overhead. If the user has
1800 * requested a higher mininum order then we start with that one instead of
1801 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07001802 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001803static inline int slab_order(int size, int min_objects,
1804 int max_order, int fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001805{
1806 int order;
1807 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07001808 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001809
Christoph Lameter39b26462008-04-14 19:11:30 +03001810 if ((PAGE_SIZE << min_order) / size > 65535)
1811 return get_order(size * 65535) - 1;
1812
Christoph Lameter6300ea72007-07-17 04:03:20 -07001813 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001814 fls(min_objects * size - 1) - PAGE_SHIFT);
1815 order <= max_order; order++) {
1816
Christoph Lameter81819f02007-05-06 14:49:36 -07001817 unsigned long slab_size = PAGE_SIZE << order;
1818
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001819 if (slab_size < min_objects * size)
Christoph Lameter81819f02007-05-06 14:49:36 -07001820 continue;
1821
Christoph Lameter81819f02007-05-06 14:49:36 -07001822 rem = slab_size % size;
1823
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001824 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07001825 break;
1826
1827 }
Christoph Lameter672bba32007-05-09 02:32:39 -07001828
Christoph Lameter81819f02007-05-06 14:49:36 -07001829 return order;
1830}
1831
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001832static inline int calculate_order(int size)
1833{
1834 int order;
1835 int min_objects;
1836 int fraction;
1837
1838 /*
1839 * Attempt to find best configuration for a slab. This
1840 * works by first attempting to generate a layout with
1841 * the best configuration and backing off gradually.
1842 *
1843 * First we reduce the acceptable waste in a slab. Then
1844 * we reduce the minimum objects required in a slab.
1845 */
1846 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03001847 if (!min_objects)
1848 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001849 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03001850 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07001851 while (fraction >= 4) {
1852 order = slab_order(size, min_objects,
1853 slub_max_order, fraction);
1854 if (order <= slub_max_order)
1855 return order;
1856 fraction /= 2;
1857 }
1858 min_objects /= 2;
1859 }
1860
1861 /*
1862 * We were unable to place multiple objects in a slab. Now
1863 * lets see if we can place a single object there.
1864 */
1865 order = slab_order(size, 1, slub_max_order, 1);
1866 if (order <= slub_max_order)
1867 return order;
1868
1869 /*
1870 * Doh this slab cannot be placed using slub_max_order.
1871 */
1872 order = slab_order(size, 1, MAX_ORDER, 1);
1873 if (order <= MAX_ORDER)
1874 return order;
1875 return -ENOSYS;
1876}
1877
Christoph Lameter81819f02007-05-06 14:49:36 -07001878/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001879 * Figure out what the alignment of the objects will be.
Christoph Lameter81819f02007-05-06 14:49:36 -07001880 */
1881static unsigned long calculate_alignment(unsigned long flags,
1882 unsigned long align, unsigned long size)
1883{
1884 /*
Christoph Lameter6446faa2008-02-15 23:45:26 -08001885 * If the user wants hardware cache aligned objects then follow that
1886 * suggestion if the object is sufficiently large.
Christoph Lameter81819f02007-05-06 14:49:36 -07001887 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001888 * The hardware cache alignment cannot override the specified
1889 * alignment though. If that is greater then use it.
Christoph Lameter81819f02007-05-06 14:49:36 -07001890 */
Nick Pigginb6210382008-03-05 14:05:56 -08001891 if (flags & SLAB_HWCACHE_ALIGN) {
1892 unsigned long ralign = cache_line_size();
1893 while (size <= ralign / 2)
1894 ralign /= 2;
1895 align = max(align, ralign);
1896 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001897
1898 if (align < ARCH_SLAB_MINALIGN)
Nick Pigginb6210382008-03-05 14:05:56 -08001899 align = ARCH_SLAB_MINALIGN;
Christoph Lameter81819f02007-05-06 14:49:36 -07001900
1901 return ALIGN(align, sizeof(void *));
1902}
1903
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001904static void init_kmem_cache_cpu(struct kmem_cache *s,
1905 struct kmem_cache_cpu *c)
1906{
1907 c->page = NULL;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001908 c->freelist = NULL;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001909 c->node = 0;
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07001910 c->offset = s->offset / sizeof(void *);
1911 c->objsize = s->objsize;
Pekka Enberg62f75532008-04-14 18:50:44 +03001912#ifdef CONFIG_SLUB_STATS
1913 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1914#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07001915}
1916
Pekka Enberg5595cff2008-08-05 09:28:47 +03001917static void
1918init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07001919{
1920 n->nr_partial = 0;
Pekka Enberg5595cff2008-08-05 09:28:47 +03001921
1922 /*
1923 * The larger the object size is, the more pages we want on the partial
1924 * list to avoid pounding the page allocator excessively.
1925 */
1926 n->min_partial = ilog2(s->size);
1927 if (n->min_partial < MIN_PARTIAL)
1928 n->min_partial = MIN_PARTIAL;
1929 else if (n->min_partial > MAX_PARTIAL)
1930 n->min_partial = MAX_PARTIAL;
1931
Christoph Lameter81819f02007-05-06 14:49:36 -07001932 spin_lock_init(&n->list_lock);
1933 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001934#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001935 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07001936 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07001937 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07001938#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07001939}
1940
Christoph Lameter4c93c3552007-10-16 01:26:08 -07001941#ifdef CONFIG_SMP
1942/*
1943 * Per cpu array for per cpu structures.
1944 *
1945 * The per cpu array places all kmem_cache_cpu structures from one processor
1946 * close together meaning that it becomes possible that multiple per cpu
1947 * structures are contained in one cacheline. This may be particularly
1948 * beneficial for the kmalloc caches.
1949 *
1950 * A desktop system typically has around 60-80 slabs. With 100 here we are
1951 * likely able to get per cpu structures for all caches from the array defined
1952 * here. We must be able to cover all kmalloc caches during bootstrap.
1953 *
1954 * If the per cpu array is exhausted then fall back to kmalloc
1955 * of individual cachelines. No sharing is possible then.
1956 */
1957#define NR_KMEM_CACHE_CPU 100
1958
1959static DEFINE_PER_CPU(struct kmem_cache_cpu,
1960 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1961
1962static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1963static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1964
1965static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1966 int cpu, gfp_t flags)
1967{
1968 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1969
1970 if (c)
1971 per_cpu(kmem_cache_cpu_free, cpu) =
1972 (void *)c->freelist;
1973 else {
1974 /* Table overflow: So allocate ourselves */
1975 c = kmalloc_node(
1976 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1977 flags, cpu_to_node(cpu));
1978 if (!c)
1979 return NULL;
1980 }
1981
1982 init_kmem_cache_cpu(s, c);
1983 return c;
1984}
1985
1986static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1987{
1988 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1989 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1990 kfree(c);
1991 return;
1992 }
1993 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1994 per_cpu(kmem_cache_cpu_free, cpu) = c;
1995}
1996
1997static void free_kmem_cache_cpus(struct kmem_cache *s)
1998{
1999 int cpu;
2000
2001 for_each_online_cpu(cpu) {
2002 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2003
2004 if (c) {
2005 s->cpu_slab[cpu] = NULL;
2006 free_kmem_cache_cpu(c, cpu);
2007 }
2008 }
2009}
2010
2011static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2012{
2013 int cpu;
2014
2015 for_each_online_cpu(cpu) {
2016 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2017
2018 if (c)
2019 continue;
2020
2021 c = alloc_kmem_cache_cpu(s, cpu, flags);
2022 if (!c) {
2023 free_kmem_cache_cpus(s);
2024 return 0;
2025 }
2026 s->cpu_slab[cpu] = c;
2027 }
2028 return 1;
2029}
2030
2031/*
2032 * Initialize the per cpu array.
2033 */
2034static void init_alloc_cpu_cpu(int cpu)
2035{
2036 int i;
2037
2038 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2039 return;
2040
2041 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2042 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2043
2044 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2045}
2046
2047static void __init init_alloc_cpu(void)
2048{
2049 int cpu;
2050
2051 for_each_online_cpu(cpu)
2052 init_alloc_cpu_cpu(cpu);
2053 }
2054
2055#else
2056static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2057static inline void init_alloc_cpu(void) {}
2058
2059static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2060{
2061 init_kmem_cache_cpu(s, &s->cpu_slab);
2062 return 1;
2063}
2064#endif
2065
Christoph Lameter81819f02007-05-06 14:49:36 -07002066#ifdef CONFIG_NUMA
2067/*
2068 * No kmalloc_node yet so do it by hand. We know that this is the first
2069 * slab on the node for this slabcache. There are no concurrent accesses
2070 * possible.
2071 *
2072 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002073 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2074 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002075 */
Adrian Bunk1cd7daa2007-10-16 01:24:18 -07002076static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2077 int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002078{
2079 struct page *page;
2080 struct kmem_cache_node *n;
rootba84c732008-01-07 23:20:28 -08002081 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07002082
2083 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2084
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002085 page = new_slab(kmalloc_caches, gfpflags, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002086
2087 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002088 if (page_to_nid(page) != node) {
2089 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2090 "node %d\n", node);
2091 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2092 "in order to be able to continue\n");
2093 }
2094
Christoph Lameter81819f02007-05-06 14:49:36 -07002095 n = page->freelist;
2096 BUG_ON(!n);
2097 page->freelist = get_freepointer(kmalloc_caches, n);
2098 page->inuse++;
2099 kmalloc_caches->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002100#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterd45f39c2007-07-17 04:03:21 -07002101 init_object(kmalloc_caches, n, 1);
2102 init_tracking(kmalloc_caches, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002103#endif
Pekka Enberg5595cff2008-08-05 09:28:47 +03002104 init_kmem_cache_node(n, kmalloc_caches);
Christoph Lameter205ab992008-04-14 19:11:40 +03002105 inc_slabs_node(kmalloc_caches, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002106
rootba84c732008-01-07 23:20:28 -08002107 /*
2108 * lockdep requires consistent irq usage for each lock
2109 * so even though there cannot be a race this early in
2110 * the boot sequence, we still disable irqs.
2111 */
2112 local_irq_save(flags);
Christoph Lameter7c2e1322008-01-07 23:20:27 -08002113 add_partial(n, page, 0);
rootba84c732008-01-07 23:20:28 -08002114 local_irq_restore(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002115 return n;
2116}
2117
2118static void free_kmem_cache_nodes(struct kmem_cache *s)
2119{
2120 int node;
2121
Christoph Lameterf64dc582007-10-16 01:25:33 -07002122 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002123 struct kmem_cache_node *n = s->node[node];
2124 if (n && n != &s->local_node)
2125 kmem_cache_free(kmalloc_caches, n);
2126 s->node[node] = NULL;
2127 }
2128}
2129
2130static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2131{
2132 int node;
2133 int local_node;
2134
2135 if (slab_state >= UP)
2136 local_node = page_to_nid(virt_to_page(s));
2137 else
2138 local_node = 0;
2139
Christoph Lameterf64dc582007-10-16 01:25:33 -07002140 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002141 struct kmem_cache_node *n;
2142
2143 if (local_node == node)
2144 n = &s->local_node;
2145 else {
2146 if (slab_state == DOWN) {
2147 n = early_kmem_cache_node_alloc(gfpflags,
2148 node);
2149 continue;
2150 }
2151 n = kmem_cache_alloc_node(kmalloc_caches,
2152 gfpflags, node);
2153
2154 if (!n) {
2155 free_kmem_cache_nodes(s);
2156 return 0;
2157 }
2158
2159 }
2160 s->node[node] = n;
Pekka Enberg5595cff2008-08-05 09:28:47 +03002161 init_kmem_cache_node(n, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002162 }
2163 return 1;
2164}
2165#else
2166static void free_kmem_cache_nodes(struct kmem_cache *s)
2167{
2168}
2169
2170static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2171{
Pekka Enberg5595cff2008-08-05 09:28:47 +03002172 init_kmem_cache_node(&s->local_node, s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002173 return 1;
2174}
2175#endif
2176
2177/*
2178 * calculate_sizes() determines the order and the distribution of data within
2179 * a slab object.
2180 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002181static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002182{
2183 unsigned long flags = s->flags;
2184 unsigned long size = s->objsize;
2185 unsigned long align = s->align;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002186 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002187
2188 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002189 * Round up object size to the next word boundary. We can only
2190 * place the free pointer at word boundaries and this determines
2191 * the possible location of the free pointer.
2192 */
2193 size = ALIGN(size, sizeof(void *));
2194
2195#ifdef CONFIG_SLUB_DEBUG
2196 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002197 * Determine if we can poison the object itself. If the user of
2198 * the slab may touch the object after free or before allocation
2199 * then we should never poison the object itself.
2200 */
2201 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002202 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002203 s->flags |= __OBJECT_POISON;
2204 else
2205 s->flags &= ~__OBJECT_POISON;
2206
Christoph Lameter81819f02007-05-06 14:49:36 -07002207
2208 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002209 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002210 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002211 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002212 */
2213 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2214 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002215#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002216
2217 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002218 * With that we have determined the number of bytes in actual use
2219 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002220 */
2221 s->inuse = size;
2222
2223 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002224 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002225 /*
2226 * Relocate free pointer after the object if it is not
2227 * permitted to overwrite the first word of the object on
2228 * kmem_cache_free.
2229 *
2230 * This is the case if we do RCU, have a constructor or
2231 * destructor or are poisoning the objects.
2232 */
2233 s->offset = size;
2234 size += sizeof(void *);
2235 }
2236
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002237#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002238 if (flags & SLAB_STORE_USER)
2239 /*
2240 * Need to store information about allocs and frees after
2241 * the object.
2242 */
2243 size += 2 * sizeof(struct track);
2244
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002245 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002246 /*
2247 * Add some empty padding so that we can catch
2248 * overwrites from earlier objects rather than let
2249 * tracking information or the free pointer be
2250 * corrupted if an user writes before the start
2251 * of the object.
2252 */
2253 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002254#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002255
Christoph Lameter81819f02007-05-06 14:49:36 -07002256 /*
2257 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07002258 * user specified and the dynamic determination of cache line size
2259 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07002260 */
2261 align = calculate_alignment(flags, align, s->objsize);
2262
2263 /*
2264 * SLUB stores one object immediately after another beginning from
2265 * offset 0. In order to align the objects we have to simply size
2266 * each object to conform to the alignment.
2267 */
2268 size = ALIGN(size, align);
2269 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03002270 if (forced_order >= 0)
2271 order = forced_order;
2272 else
2273 order = calculate_order(size);
Christoph Lameter81819f02007-05-06 14:49:36 -07002274
Christoph Lameter834f3d12008-04-14 19:11:31 +03002275 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07002276 return 0;
2277
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002278 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002279 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08002280 s->allocflags |= __GFP_COMP;
2281
2282 if (s->flags & SLAB_CACHE_DMA)
2283 s->allocflags |= SLUB_DMA;
2284
2285 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2286 s->allocflags |= __GFP_RECLAIMABLE;
2287
Christoph Lameter81819f02007-05-06 14:49:36 -07002288 /*
2289 * Determine the number of objects per slab
2290 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002291 s->oo = oo_make(order, size);
Christoph Lameter65c33762008-04-14 19:11:40 +03002292 s->min = oo_make(get_order(size), size);
Christoph Lameter205ab992008-04-14 19:11:40 +03002293 if (oo_objects(s->oo) > oo_objects(s->max))
2294 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002295
Christoph Lameter834f3d12008-04-14 19:11:31 +03002296 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07002297
2298}
2299
Christoph Lameter81819f02007-05-06 14:49:36 -07002300static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2301 const char *name, size_t size,
2302 size_t align, unsigned long flags,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002303 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07002304{
2305 memset(s, 0, kmem_size);
2306 s->name = name;
2307 s->ctor = ctor;
Christoph Lameter81819f02007-05-06 14:49:36 -07002308 s->objsize = size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002309 s->align = align;
Christoph Lameterba0268a2007-09-11 15:24:11 -07002310 s->flags = kmem_cache_flags(size, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07002311
Christoph Lameter06b285d2008-04-14 19:11:41 +03002312 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07002313 goto error;
2314
2315 s->refcount = 1;
2316#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05002317 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07002318#endif
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002319 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2320 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07002321
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002322 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
Christoph Lameter81819f02007-05-06 14:49:36 -07002323 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002324 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07002325error:
2326 if (flags & SLAB_PANIC)
2327 panic("Cannot create slab %s size=%lu realsize=%u "
2328 "order=%u offset=%u flags=%lx\n",
Christoph Lameter834f3d12008-04-14 19:11:31 +03002329 s->name, (unsigned long)size, s->size, oo_order(s->oo),
Christoph Lameter81819f02007-05-06 14:49:36 -07002330 s->offset, flags);
2331 return 0;
2332}
Christoph Lameter81819f02007-05-06 14:49:36 -07002333
2334/*
2335 * Check if a given pointer is valid
2336 */
2337int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2338{
Pekka Enberg06428782008-01-07 23:20:27 -08002339 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002340
2341 page = get_object_page(object);
2342
2343 if (!page || s != page->slab)
2344 /* No slab or wrong slab */
2345 return 0;
2346
Christoph Lameterabcd08a2007-05-09 02:32:37 -07002347 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07002348 return 0;
2349
2350 /*
2351 * We could also check if the object is on the slabs freelist.
2352 * But this would be too expensive and it seems that the main
Christoph Lameter6446faa2008-02-15 23:45:26 -08002353 * purpose of kmem_ptr_valid() is to check if the object belongs
Christoph Lameter81819f02007-05-06 14:49:36 -07002354 * to a certain slab.
2355 */
2356 return 1;
2357}
2358EXPORT_SYMBOL(kmem_ptr_validate);
2359
2360/*
2361 * Determine the size of a slab object
2362 */
2363unsigned int kmem_cache_size(struct kmem_cache *s)
2364{
2365 return s->objsize;
2366}
2367EXPORT_SYMBOL(kmem_cache_size);
2368
2369const char *kmem_cache_name(struct kmem_cache *s)
2370{
2371 return s->name;
2372}
2373EXPORT_SYMBOL(kmem_cache_name);
2374
Christoph Lameter33b12c32008-04-25 12:22:43 -07002375static void list_slab_objects(struct kmem_cache *s, struct page *page,
2376 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07002377{
Christoph Lameter33b12c32008-04-25 12:22:43 -07002378#ifdef CONFIG_SLUB_DEBUG
2379 void *addr = page_address(page);
2380 void *p;
2381 DECLARE_BITMAP(map, page->objects);
2382
2383 bitmap_zero(map, page->objects);
2384 slab_err(s, page, "%s", text);
2385 slab_lock(page);
2386 for_each_free_object(p, s, page->freelist)
2387 set_bit(slab_index(p, s, addr), map);
2388
2389 for_each_object(p, s, addr, page->objects) {
2390
2391 if (!test_bit(slab_index(p, s, addr), map)) {
2392 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2393 p, p - addr);
2394 print_tracking(s, p);
2395 }
2396 }
2397 slab_unlock(page);
2398#endif
2399}
2400
Christoph Lameter81819f02007-05-06 14:49:36 -07002401/*
Christoph Lameter599870b2008-04-23 12:36:52 -07002402 * Attempt to free all partial slabs on a node.
Christoph Lameter81819f02007-05-06 14:49:36 -07002403 */
Christoph Lameter599870b2008-04-23 12:36:52 -07002404static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002405{
Christoph Lameter81819f02007-05-06 14:49:36 -07002406 unsigned long flags;
2407 struct page *page, *h;
2408
2409 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter33b12c32008-04-25 12:22:43 -07002410 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002411 if (!page->inuse) {
2412 list_del(&page->lru);
2413 discard_slab(s, page);
Christoph Lameter599870b2008-04-23 12:36:52 -07002414 n->nr_partial--;
Christoph Lameter33b12c32008-04-25 12:22:43 -07002415 } else {
2416 list_slab_objects(s, page,
2417 "Objects remaining on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07002418 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07002419 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002420 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002421}
2422
2423/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002424 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07002425 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002426static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002427{
2428 int node;
2429
2430 flush_all(s);
2431
2432 /* Attempt to free all objects */
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002433 free_kmem_cache_cpus(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002434 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002435 struct kmem_cache_node *n = get_node(s, node);
2436
Christoph Lameter599870b2008-04-23 12:36:52 -07002437 free_partial(s, n);
2438 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07002439 return 1;
2440 }
2441 free_kmem_cache_nodes(s);
2442 return 0;
2443}
2444
2445/*
2446 * Close a cache and release the kmem_cache structure
2447 * (must be used for caches created using kmem_cache_create)
2448 */
2449void kmem_cache_destroy(struct kmem_cache *s)
2450{
2451 down_write(&slub_lock);
2452 s->refcount--;
2453 if (!s->refcount) {
2454 list_del(&s->list);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002455 up_write(&slub_lock);
Pekka Enbergd629d812008-04-23 22:31:08 +03002456 if (kmem_cache_close(s)) {
2457 printk(KERN_ERR "SLUB %s: %s called for cache that "
2458 "still has objects.\n", s->name, __func__);
2459 dump_stack();
2460 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002461 sysfs_slab_remove(s);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07002462 } else
2463 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07002464}
2465EXPORT_SYMBOL(kmem_cache_destroy);
2466
2467/********************************************************************
2468 * Kmalloc subsystem
2469 *******************************************************************/
2470
Christoph Lameter331dc552008-02-14 14:28:09 -08002471struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
Christoph Lameter81819f02007-05-06 14:49:36 -07002472EXPORT_SYMBOL(kmalloc_caches);
2473
Christoph Lameter81819f02007-05-06 14:49:36 -07002474static int __init setup_slub_min_order(char *str)
2475{
Pekka Enberg06428782008-01-07 23:20:27 -08002476 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002477
2478 return 1;
2479}
2480
2481__setup("slub_min_order=", setup_slub_min_order);
2482
2483static int __init setup_slub_max_order(char *str)
2484{
Pekka Enberg06428782008-01-07 23:20:27 -08002485 get_option(&str, &slub_max_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07002486
2487 return 1;
2488}
2489
2490__setup("slub_max_order=", setup_slub_max_order);
2491
2492static int __init setup_slub_min_objects(char *str)
2493{
Pekka Enberg06428782008-01-07 23:20:27 -08002494 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07002495
2496 return 1;
2497}
2498
2499__setup("slub_min_objects=", setup_slub_min_objects);
2500
2501static int __init setup_slub_nomerge(char *str)
2502{
2503 slub_nomerge = 1;
2504 return 1;
2505}
2506
2507__setup("slub_nomerge", setup_slub_nomerge);
2508
Christoph Lameter81819f02007-05-06 14:49:36 -07002509static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2510 const char *name, int size, gfp_t gfp_flags)
2511{
2512 unsigned int flags = 0;
2513
2514 if (gfp_flags & SLUB_DMA)
2515 flags = SLAB_CACHE_DMA;
2516
2517 down_write(&slub_lock);
2518 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
Christoph Lameter319d1e22008-04-14 19:11:41 +03002519 flags, NULL))
Christoph Lameter81819f02007-05-06 14:49:36 -07002520 goto panic;
2521
2522 list_add(&s->list, &slab_caches);
2523 up_write(&slub_lock);
2524 if (sysfs_slab_add(s))
2525 goto panic;
2526 return s;
2527
2528panic:
2529 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2530}
2531
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002532#ifdef CONFIG_ZONE_DMA
Christoph Lameter4097d602008-04-14 18:51:18 +03002533static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
Christoph Lameter1ceef402007-08-07 15:11:48 -07002534
2535static void sysfs_add_func(struct work_struct *w)
2536{
2537 struct kmem_cache *s;
2538
2539 down_write(&slub_lock);
2540 list_for_each_entry(s, &slab_caches, list) {
2541 if (s->flags & __SYSFS_ADD_DEFERRED) {
2542 s->flags &= ~__SYSFS_ADD_DEFERRED;
2543 sysfs_slab_add(s);
2544 }
2545 }
2546 up_write(&slub_lock);
2547}
2548
2549static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2550
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002551static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2552{
2553 struct kmem_cache *s;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002554 char *text;
2555 size_t realsize;
2556
2557 s = kmalloc_caches_dma[index];
2558 if (s)
2559 return s;
2560
2561 /* Dynamically create dma cache */
Christoph Lameter1ceef402007-08-07 15:11:48 -07002562 if (flags & __GFP_WAIT)
2563 down_write(&slub_lock);
2564 else {
2565 if (!down_write_trylock(&slub_lock))
2566 goto out;
2567 }
2568
2569 if (kmalloc_caches_dma[index])
2570 goto unlock_out;
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002571
Christoph Lameter7b55f622007-07-17 04:03:27 -07002572 realsize = kmalloc_caches[index].objsize;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08002573 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2574 (unsigned int)realsize);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002575 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2576
2577 if (!s || !text || !kmem_cache_open(s, flags, text,
2578 realsize, ARCH_KMALLOC_MINALIGN,
2579 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2580 kfree(s);
2581 kfree(text);
2582 goto unlock_out;
Christoph Lameterdfce8642007-07-17 04:03:25 -07002583 }
Christoph Lameter1ceef402007-08-07 15:11:48 -07002584
2585 list_add(&s->list, &slab_caches);
2586 kmalloc_caches_dma[index] = s;
2587
2588 schedule_work(&sysfs_add_work);
2589
2590unlock_out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002591 up_write(&slub_lock);
Christoph Lameter1ceef402007-08-07 15:11:48 -07002592out:
Christoph Lameterdfce8642007-07-17 04:03:25 -07002593 return kmalloc_caches_dma[index];
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002594}
2595#endif
2596
Christoph Lameterf1b26332007-07-17 04:03:26 -07002597/*
2598 * Conversion table for small slabs sizes / 8 to the index in the
2599 * kmalloc array. This is necessary for slabs < 192 since we have non power
2600 * of two cache sizes there. The size of larger slabs can be determined using
2601 * fls.
2602 */
2603static s8 size_index[24] = {
2604 3, /* 8 */
2605 4, /* 16 */
2606 5, /* 24 */
2607 5, /* 32 */
2608 6, /* 40 */
2609 6, /* 48 */
2610 6, /* 56 */
2611 6, /* 64 */
2612 1, /* 72 */
2613 1, /* 80 */
2614 1, /* 88 */
2615 1, /* 96 */
2616 7, /* 104 */
2617 7, /* 112 */
2618 7, /* 120 */
2619 7, /* 128 */
2620 2, /* 136 */
2621 2, /* 144 */
2622 2, /* 152 */
2623 2, /* 160 */
2624 2, /* 168 */
2625 2, /* 176 */
2626 2, /* 184 */
2627 2 /* 192 */
2628};
2629
Christoph Lameter81819f02007-05-06 14:49:36 -07002630static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2631{
Christoph Lameterf1b26332007-07-17 04:03:26 -07002632 int index;
Christoph Lameter81819f02007-05-06 14:49:36 -07002633
Christoph Lameterf1b26332007-07-17 04:03:26 -07002634 if (size <= 192) {
2635 if (!size)
2636 return ZERO_SIZE_PTR;
Christoph Lameter81819f02007-05-06 14:49:36 -07002637
Christoph Lameterf1b26332007-07-17 04:03:26 -07002638 index = size_index[(size - 1) / 8];
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002639 } else
Christoph Lameterf1b26332007-07-17 04:03:26 -07002640 index = fls(size - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07002641
2642#ifdef CONFIG_ZONE_DMA
Christoph Lameterf1b26332007-07-17 04:03:26 -07002643 if (unlikely((flags & SLUB_DMA)))
Christoph Lameter2e443fd2007-07-17 04:03:24 -07002644 return dma_kmalloc_cache(index, flags);
Christoph Lameterf1b26332007-07-17 04:03:26 -07002645
Christoph Lameter81819f02007-05-06 14:49:36 -07002646#endif
2647 return &kmalloc_caches[index];
2648}
2649
2650void *__kmalloc(size_t size, gfp_t flags)
2651{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002652 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002653
Christoph Lameter331dc552008-02-14 14:28:09 -08002654 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02002655 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002656
2657 s = get_slab(size, flags);
2658
2659 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002660 return s;
2661
Christoph Lameterce15fea2007-07-17 04:03:28 -07002662 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002663}
2664EXPORT_SYMBOL(__kmalloc);
2665
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002666static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2667{
2668 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2669 get_order(size));
2670
2671 if (page)
2672 return page_address(page);
2673 else
2674 return NULL;
2675}
2676
Christoph Lameter81819f02007-05-06 14:49:36 -07002677#ifdef CONFIG_NUMA
2678void *__kmalloc_node(size_t size, gfp_t flags, int node)
2679{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002680 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07002681
Christoph Lameter331dc552008-02-14 14:28:09 -08002682 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08002683 return kmalloc_large_node(size, flags, node);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002684
2685 s = get_slab(size, flags);
2686
2687 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002688 return s;
2689
Christoph Lameterce15fea2007-07-17 04:03:28 -07002690 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002691}
2692EXPORT_SYMBOL(__kmalloc_node);
2693#endif
2694
2695size_t ksize(const void *object)
2696{
Christoph Lameter272c1d22007-06-08 13:46:49 -07002697 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002698 struct kmem_cache *s;
2699
Christoph Lameteref8b4522007-10-16 01:24:46 -07002700 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07002701 return 0;
2702
Vegard Nossum294a80a2007-12-04 23:45:30 -08002703 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08002704
Pekka Enberg76994412008-05-22 19:22:25 +03002705 if (unlikely(!PageSlab(page))) {
2706 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08002707 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03002708 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002709 s = page->slab;
Christoph Lameter81819f02007-05-06 14:49:36 -07002710
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002711#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002712 /*
2713 * Debugging requires use of the padding between object
2714 * and whatever may come after it.
2715 */
2716 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2717 return s->objsize;
2718
Christoph Lameterae20bfd2008-02-15 23:45:25 -08002719#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002720 /*
2721 * If we have the need to store the freelist pointer
2722 * back there or track user information then we can
2723 * only use the space before that information.
2724 */
2725 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2726 return s->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07002727 /*
2728 * Else we can use all the padding etc for the allocation
2729 */
2730 return s->size;
2731}
Christoph Lameter81819f02007-05-06 14:49:36 -07002732
2733void kfree(const void *x)
2734{
Christoph Lameter81819f02007-05-06 14:49:36 -07002735 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002736 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07002737
Satyam Sharma2408c552007-10-16 01:24:44 -07002738 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002739 return;
2740
Christoph Lameterb49af682007-05-06 14:49:41 -07002741 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002742 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07002743 BUG_ON(!PageCompound(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07002744 put_page(page);
2745 return;
2746 }
Christoph Lameter5bb983b2008-02-07 17:47:41 -08002747 slab_free(page->slab, page, object, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002748}
2749EXPORT_SYMBOL(kfree);
2750
Christoph Lameter2086d262007-05-06 14:49:46 -07002751/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002752 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2753 * the remaining slabs by the number of items in use. The slabs with the
2754 * most items in use come first. New allocations will then fill those up
2755 * and thus they can be removed from the partial lists.
2756 *
2757 * The slabs with the least items are placed last. This results in them
2758 * being allocated from last increasing the chance that the last objects
2759 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07002760 */
2761int kmem_cache_shrink(struct kmem_cache *s)
2762{
2763 int node;
2764 int i;
2765 struct kmem_cache_node *n;
2766 struct page *page;
2767 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03002768 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07002769 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03002770 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07002771 unsigned long flags;
2772
2773 if (!slabs_by_inuse)
2774 return -ENOMEM;
2775
2776 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07002777 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07002778 n = get_node(s, node);
2779
2780 if (!n->nr_partial)
2781 continue;
2782
Christoph Lameter834f3d12008-04-14 19:11:31 +03002783 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07002784 INIT_LIST_HEAD(slabs_by_inuse + i);
2785
2786 spin_lock_irqsave(&n->list_lock, flags);
2787
2788 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002789 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07002790 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002791 * Note that concurrent frees may occur while we hold the
2792 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07002793 */
2794 list_for_each_entry_safe(page, t, &n->partial, lru) {
2795 if (!page->inuse && slab_trylock(page)) {
2796 /*
2797 * Must hold slab lock here because slab_free
2798 * may have freed the last object and be
2799 * waiting to release the slab.
2800 */
2801 list_del(&page->lru);
2802 n->nr_partial--;
2803 slab_unlock(page);
2804 discard_slab(s, page);
2805 } else {
Christoph Lameterfcda3d82007-07-30 13:06:46 -07002806 list_move(&page->lru,
2807 slabs_by_inuse + page->inuse);
Christoph Lameter2086d262007-05-06 14:49:46 -07002808 }
2809 }
2810
Christoph Lameter2086d262007-05-06 14:49:46 -07002811 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002812 * Rebuild the partial list with the slabs filled up most
2813 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07002814 */
Christoph Lameter834f3d12008-04-14 19:11:31 +03002815 for (i = objects - 1; i >= 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07002816 list_splice(slabs_by_inuse + i, n->partial.prev);
2817
Christoph Lameter2086d262007-05-06 14:49:46 -07002818 spin_unlock_irqrestore(&n->list_lock, flags);
2819 }
2820
2821 kfree(slabs_by_inuse);
2822 return 0;
2823}
2824EXPORT_SYMBOL(kmem_cache_shrink);
2825
Yasunori Gotob9049e22007-10-21 16:41:37 -07002826#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2827static int slab_mem_going_offline_callback(void *arg)
2828{
2829 struct kmem_cache *s;
2830
2831 down_read(&slub_lock);
2832 list_for_each_entry(s, &slab_caches, list)
2833 kmem_cache_shrink(s);
2834 up_read(&slub_lock);
2835
2836 return 0;
2837}
2838
2839static void slab_mem_offline_callback(void *arg)
2840{
2841 struct kmem_cache_node *n;
2842 struct kmem_cache *s;
2843 struct memory_notify *marg = arg;
2844 int offline_node;
2845
2846 offline_node = marg->status_change_nid;
2847
2848 /*
2849 * If the node still has available memory. we need kmem_cache_node
2850 * for it yet.
2851 */
2852 if (offline_node < 0)
2853 return;
2854
2855 down_read(&slub_lock);
2856 list_for_each_entry(s, &slab_caches, list) {
2857 n = get_node(s, offline_node);
2858 if (n) {
2859 /*
2860 * if n->nr_slabs > 0, slabs still exist on the node
2861 * that is going down. We were unable to free them,
2862 * and offline_pages() function shoudn't call this
2863 * callback. So, we must fail.
2864 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002865 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07002866
2867 s->node[offline_node] = NULL;
2868 kmem_cache_free(kmalloc_caches, n);
2869 }
2870 }
2871 up_read(&slub_lock);
2872}
2873
2874static int slab_mem_going_online_callback(void *arg)
2875{
2876 struct kmem_cache_node *n;
2877 struct kmem_cache *s;
2878 struct memory_notify *marg = arg;
2879 int nid = marg->status_change_nid;
2880 int ret = 0;
2881
2882 /*
2883 * If the node's memory is already available, then kmem_cache_node is
2884 * already created. Nothing to do.
2885 */
2886 if (nid < 0)
2887 return 0;
2888
2889 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07002890 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07002891 * allocate a kmem_cache_node structure in order to bring the node
2892 * online.
2893 */
2894 down_read(&slub_lock);
2895 list_for_each_entry(s, &slab_caches, list) {
2896 /*
2897 * XXX: kmem_cache_alloc_node will fallback to other nodes
2898 * since memory is not yet available from the node that
2899 * is brought up.
2900 */
2901 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2902 if (!n) {
2903 ret = -ENOMEM;
2904 goto out;
2905 }
Pekka Enberg5595cff2008-08-05 09:28:47 +03002906 init_kmem_cache_node(n, s);
Yasunori Gotob9049e22007-10-21 16:41:37 -07002907 s->node[nid] = n;
2908 }
2909out:
2910 up_read(&slub_lock);
2911 return ret;
2912}
2913
2914static int slab_memory_callback(struct notifier_block *self,
2915 unsigned long action, void *arg)
2916{
2917 int ret = 0;
2918
2919 switch (action) {
2920 case MEM_GOING_ONLINE:
2921 ret = slab_mem_going_online_callback(arg);
2922 break;
2923 case MEM_GOING_OFFLINE:
2924 ret = slab_mem_going_offline_callback(arg);
2925 break;
2926 case MEM_OFFLINE:
2927 case MEM_CANCEL_ONLINE:
2928 slab_mem_offline_callback(arg);
2929 break;
2930 case MEM_ONLINE:
2931 case MEM_CANCEL_OFFLINE:
2932 break;
2933 }
2934
2935 ret = notifier_from_errno(ret);
2936 return ret;
2937}
2938
2939#endif /* CONFIG_MEMORY_HOTPLUG */
2940
Christoph Lameter81819f02007-05-06 14:49:36 -07002941/********************************************************************
2942 * Basic setup of slabs
2943 *******************************************************************/
2944
2945void __init kmem_cache_init(void)
2946{
2947 int i;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002948 int caches = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002949
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002950 init_alloc_cpu();
2951
Christoph Lameter81819f02007-05-06 14:49:36 -07002952#ifdef CONFIG_NUMA
2953 /*
2954 * Must first have the slab cache available for the allocations of the
Christoph Lameter672bba32007-05-09 02:32:39 -07002955 * struct kmem_cache_node's. There is special bootstrap code in
Christoph Lameter81819f02007-05-06 14:49:36 -07002956 * kmem_cache_open for slab_state == DOWN.
2957 */
2958 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2959 sizeof(struct kmem_cache_node), GFP_KERNEL);
Christoph Lameter8ffa6872007-05-31 00:40:51 -07002960 kmalloc_caches[0].refcount = -1;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002961 caches++;
Yasunori Gotob9049e22007-10-21 16:41:37 -07002962
Nadia Derbey0c40ba42008-04-29 01:00:41 -07002963 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
Christoph Lameter81819f02007-05-06 14:49:36 -07002964#endif
2965
2966 /* Able to allocate the per node structures */
2967 slab_state = PARTIAL;
2968
2969 /* Caches that are not of the two-to-the-power-of size */
Christoph Lameter4b356be2007-06-16 10:16:13 -07002970 if (KMALLOC_MIN_SIZE <= 64) {
2971 create_kmalloc_cache(&kmalloc_caches[1],
Christoph Lameter81819f02007-05-06 14:49:36 -07002972 "kmalloc-96", 96, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002973 caches++;
Christoph Lameter4b356be2007-06-16 10:16:13 -07002974 create_kmalloc_cache(&kmalloc_caches[2],
Christoph Lameter81819f02007-05-06 14:49:36 -07002975 "kmalloc-192", 192, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002976 caches++;
2977 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002978
Christoph Lameter331dc552008-02-14 14:28:09 -08002979 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002980 create_kmalloc_cache(&kmalloc_caches[i],
2981 "kmalloc", 1 << i, GFP_KERNEL);
Christoph Lameter4b356be2007-06-16 10:16:13 -07002982 caches++;
2983 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002984
Christoph Lameterf1b26332007-07-17 04:03:26 -07002985
2986 /*
2987 * Patch up the size_index table if we have strange large alignment
2988 * requirements for the kmalloc array. This is only the case for
Christoph Lameter6446faa2008-02-15 23:45:26 -08002989 * MIPS it seems. The standard arches will not generate any code here.
Christoph Lameterf1b26332007-07-17 04:03:26 -07002990 *
2991 * Largest permitted alignment is 256 bytes due to the way we
2992 * handle the index determination for the smaller caches.
2993 *
2994 * Make sure that nothing crazy happens if someone starts tinkering
2995 * around with ARCH_KMALLOC_MINALIGN
2996 */
2997 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2998 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2999
Christoph Lameter12ad6842007-07-17 04:03:28 -07003000 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
Christoph Lameterf1b26332007-07-17 04:03:26 -07003001 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3002
Christoph Lameter41d54d32008-07-03 09:14:26 -05003003 if (KMALLOC_MIN_SIZE == 128) {
3004 /*
3005 * The 192 byte sized cache is not used if the alignment
3006 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3007 * instead.
3008 */
3009 for (i = 128 + 8; i <= 192; i += 8)
3010 size_index[(i - 1) / 8] = 8;
3011 }
3012
Christoph Lameter81819f02007-05-06 14:49:36 -07003013 slab_state = UP;
3014
3015 /* Provide the correct kmalloc names now that the caches are up */
Christoph Lameter331dc552008-02-14 14:28:09 -08003016 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
Christoph Lameter81819f02007-05-06 14:49:36 -07003017 kmalloc_caches[i]. name =
3018 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3019
3020#ifdef CONFIG_SMP
3021 register_cpu_notifier(&slab_notifier);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003022 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3023 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3024#else
3025 kmem_size = sizeof(struct kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003026#endif
3027
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003028 printk(KERN_INFO
3029 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003030 " CPUs=%d, Nodes=%d\n",
3031 caches, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003032 slub_min_order, slub_max_order, slub_min_objects,
3033 nr_cpu_ids, nr_node_ids);
3034}
3035
3036/*
3037 * Find a mergeable slab cache
3038 */
3039static int slab_unmergeable(struct kmem_cache *s)
3040{
3041 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3042 return 1;
3043
Christoph Lameterc59def92007-05-16 22:10:50 -07003044 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003045 return 1;
3046
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003047 /*
3048 * We may have set a slab to be unmergeable during bootstrap.
3049 */
3050 if (s->refcount < 0)
3051 return 1;
3052
Christoph Lameter81819f02007-05-06 14:49:36 -07003053 return 0;
3054}
3055
3056static struct kmem_cache *find_mergeable(size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003057 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003058 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003059{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003060 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003061
3062 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3063 return NULL;
3064
Christoph Lameterc59def92007-05-16 22:10:50 -07003065 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003066 return NULL;
3067
3068 size = ALIGN(size, sizeof(void *));
3069 align = calculate_alignment(flags, align, size);
3070 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003071 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003072
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003073 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003074 if (slab_unmergeable(s))
3075 continue;
3076
3077 if (size > s->size)
3078 continue;
3079
Christoph Lameterba0268a2007-09-11 15:24:11 -07003080 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003081 continue;
3082 /*
3083 * Check if alignment is compatible.
3084 * Courtesy of Adrian Drzewiecki
3085 */
Pekka Enberg06428782008-01-07 23:20:27 -08003086 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003087 continue;
3088
3089 if (s->size - size >= sizeof(void *))
3090 continue;
3091
3092 return s;
3093 }
3094 return NULL;
3095}
3096
3097struct kmem_cache *kmem_cache_create(const char *name, size_t size,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003098 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003099{
3100 struct kmem_cache *s;
3101
3102 down_write(&slub_lock);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003103 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003104 if (s) {
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003105 int cpu;
3106
Christoph Lameter81819f02007-05-06 14:49:36 -07003107 s->refcount++;
3108 /*
3109 * Adjust the object sizes so that we clear
3110 * the complete object on kzalloc.
3111 */
3112 s->objsize = max(s->objsize, (int)size);
Christoph Lameter42a9fdb2007-10-16 01:26:09 -07003113
3114 /*
3115 * And then we need to update the object size in the
3116 * per cpu structures
3117 */
3118 for_each_online_cpu(cpu)
3119 get_cpu_slab(s, cpu)->objsize = s->objsize;
Christoph Lameter6446faa2008-02-15 23:45:26 -08003120
Christoph Lameter81819f02007-05-06 14:49:36 -07003121 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003122 up_write(&slub_lock);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003123
Christoph Lameter81819f02007-05-06 14:49:36 -07003124 if (sysfs_slab_alias(s, name))
3125 goto err;
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003126 return s;
3127 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003128
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003129 s = kmalloc(kmem_size, GFP_KERNEL);
3130 if (s) {
3131 if (kmem_cache_open(s, GFP_KERNEL, name,
Christoph Lameterc59def92007-05-16 22:10:50 -07003132 size, align, flags, ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003133 list_add(&s->list, &slab_caches);
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003134 up_write(&slub_lock);
3135 if (sysfs_slab_add(s))
3136 goto err;
3137 return s;
3138 }
3139 kfree(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003140 }
3141 up_write(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003142
3143err:
Christoph Lameter81819f02007-05-06 14:49:36 -07003144 if (flags & SLAB_PANIC)
3145 panic("Cannot create slabcache %s\n", name);
3146 else
3147 s = NULL;
3148 return s;
3149}
3150EXPORT_SYMBOL(kmem_cache_create);
3151
Christoph Lameter81819f02007-05-06 14:49:36 -07003152#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003153/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003154 * Use the cpu notifier to insure that the cpu slabs are flushed when
3155 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003156 */
3157static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3158 unsigned long action, void *hcpu)
3159{
3160 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003161 struct kmem_cache *s;
3162 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003163
3164 switch (action) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003165 case CPU_UP_PREPARE:
3166 case CPU_UP_PREPARE_FROZEN:
3167 init_alloc_cpu_cpu(cpu);
3168 down_read(&slub_lock);
3169 list_for_each_entry(s, &slab_caches, list)
3170 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3171 GFP_KERNEL);
3172 up_read(&slub_lock);
3173 break;
3174
Christoph Lameter81819f02007-05-06 14:49:36 -07003175 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003176 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003177 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003178 case CPU_DEAD_FROZEN:
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003179 down_read(&slub_lock);
3180 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003181 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3182
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003183 local_irq_save(flags);
3184 __flush_cpu_slab(s, cpu);
3185 local_irq_restore(flags);
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003186 free_kmem_cache_cpu(c, cpu);
3187 s->cpu_slab[cpu] = NULL;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003188 }
3189 up_read(&slub_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003190 break;
3191 default:
3192 break;
3193 }
3194 return NOTIFY_OK;
3195}
3196
Pekka Enberg06428782008-01-07 23:20:27 -08003197static struct notifier_block __cpuinitdata slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003198 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003199};
Christoph Lameter81819f02007-05-06 14:49:36 -07003200
3201#endif
3202
Christoph Lameter81819f02007-05-06 14:49:36 -07003203void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3204{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003205 struct kmem_cache *s;
3206
Christoph Lameter331dc552008-02-14 14:28:09 -08003207 if (unlikely(size > PAGE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003208 return kmalloc_large(size, gfpflags);
3209
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003210 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003211
Satyam Sharma2408c552007-10-16 01:24:44 -07003212 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003213 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003214
Christoph Lameterce15fea2007-07-17 04:03:28 -07003215 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003216}
3217
3218void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3219 int node, void *caller)
3220{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003221 struct kmem_cache *s;
3222
Christoph Lameter331dc552008-02-14 14:28:09 -08003223 if (unlikely(size > PAGE_SIZE))
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003224 return kmalloc_large_node(size, gfpflags, node);
Pekka Enbergeada35e2008-02-11 22:47:46 +02003225
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003226 s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003227
Satyam Sharma2408c552007-10-16 01:24:44 -07003228 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003229 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003230
Christoph Lameterce15fea2007-07-17 04:03:28 -07003231 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07003232}
3233
Christoph Lameterf6acb632008-04-29 16:16:06 -07003234#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03003235static unsigned long count_partial(struct kmem_cache_node *n,
3236 int (*get_count)(struct page *))
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003237{
3238 unsigned long flags;
3239 unsigned long x = 0;
3240 struct page *page;
3241
3242 spin_lock_irqsave(&n->list_lock, flags);
3243 list_for_each_entry(page, &n->partial, lru)
Christoph Lameter205ab992008-04-14 19:11:40 +03003244 x += get_count(page);
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003245 spin_unlock_irqrestore(&n->list_lock, flags);
3246 return x;
3247}
Christoph Lameter205ab992008-04-14 19:11:40 +03003248
3249static int count_inuse(struct page *page)
3250{
3251 return page->inuse;
3252}
3253
3254static int count_total(struct page *page)
3255{
3256 return page->objects;
3257}
3258
3259static int count_free(struct page *page)
3260{
3261 return page->objects - page->inuse;
3262}
Christoph Lameter5b06c8532008-04-14 18:51:34 +03003263
Christoph Lameter434e2452007-07-17 04:03:30 -07003264static int validate_slab(struct kmem_cache *s, struct page *page,
3265 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003266{
3267 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003268 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003269
3270 if (!check_slab(s, page) ||
3271 !on_freelist(s, page, NULL))
3272 return 0;
3273
3274 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003275 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003276
Christoph Lameter7656c722007-05-09 02:32:40 -07003277 for_each_free_object(p, s, page->freelist) {
3278 set_bit(slab_index(p, s, addr), map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003279 if (!check_object(s, page, p, 0))
3280 return 0;
3281 }
3282
Christoph Lameter224a88b2008-04-14 19:11:31 +03003283 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003284 if (!test_bit(slab_index(p, s, addr), map))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003285 if (!check_object(s, page, p, 1))
3286 return 0;
3287 return 1;
3288}
3289
Christoph Lameter434e2452007-07-17 04:03:30 -07003290static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3291 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003292{
3293 if (slab_trylock(page)) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003294 validate_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003295 slab_unlock(page);
3296 } else
3297 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3298 s->name, page);
3299
3300 if (s->flags & DEBUG_DEFAULT_FLAGS) {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003301 if (!PageSlubDebug(page))
3302 printk(KERN_ERR "SLUB %s: SlubDebug not set "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003303 "on slab 0x%p\n", s->name, page);
3304 } else {
Andy Whitcroft8a380822008-07-23 21:27:18 -07003305 if (PageSlubDebug(page))
3306 printk(KERN_ERR "SLUB %s: SlubDebug set on "
Christoph Lameter53e15af2007-05-06 14:49:43 -07003307 "slab 0x%p\n", s->name, page);
3308 }
3309}
3310
Christoph Lameter434e2452007-07-17 04:03:30 -07003311static int validate_slab_node(struct kmem_cache *s,
3312 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003313{
3314 unsigned long count = 0;
3315 struct page *page;
3316 unsigned long flags;
3317
3318 spin_lock_irqsave(&n->list_lock, flags);
3319
3320 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003321 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003322 count++;
3323 }
3324 if (count != n->nr_partial)
3325 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3326 "counter=%ld\n", s->name, count, n->nr_partial);
3327
3328 if (!(s->flags & SLAB_STORE_USER))
3329 goto out;
3330
3331 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003332 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003333 count++;
3334 }
3335 if (count != atomic_long_read(&n->nr_slabs))
3336 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3337 "counter=%ld\n", s->name, count,
3338 atomic_long_read(&n->nr_slabs));
3339
3340out:
3341 spin_unlock_irqrestore(&n->list_lock, flags);
3342 return count;
3343}
3344
Christoph Lameter434e2452007-07-17 04:03:30 -07003345static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003346{
3347 int node;
3348 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003349 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003350 sizeof(unsigned long), GFP_KERNEL);
3351
3352 if (!map)
3353 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003354
3355 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003356 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003357 struct kmem_cache_node *n = get_node(s, node);
3358
Christoph Lameter434e2452007-07-17 04:03:30 -07003359 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003360 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003361 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003362 return count;
3363}
3364
Christoph Lameterb3459702007-05-09 02:32:41 -07003365#ifdef SLUB_RESILIENCY_TEST
3366static void resiliency_test(void)
3367{
3368 u8 *p;
3369
3370 printk(KERN_ERR "SLUB resiliency testing\n");
3371 printk(KERN_ERR "-----------------------\n");
3372 printk(KERN_ERR "A. Corruption after allocation\n");
3373
3374 p = kzalloc(16, GFP_KERNEL);
3375 p[16] = 0x12;
3376 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3377 " 0x12->0x%p\n\n", p + 16);
3378
3379 validate_slab_cache(kmalloc_caches + 4);
3380
3381 /* Hmmm... The next two are dangerous */
3382 p = kzalloc(32, GFP_KERNEL);
3383 p[32 + sizeof(void *)] = 0x34;
3384 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003385 " 0x34 -> -0x%p\n", p);
3386 printk(KERN_ERR
3387 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003388
3389 validate_slab_cache(kmalloc_caches + 5);
3390 p = kzalloc(64, GFP_KERNEL);
3391 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3392 *p = 0x56;
3393 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3394 p);
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003395 printk(KERN_ERR
3396 "If allocated object is overwritten then not detectable\n\n");
Christoph Lameterb3459702007-05-09 02:32:41 -07003397 validate_slab_cache(kmalloc_caches + 6);
3398
3399 printk(KERN_ERR "\nB. Corruption after free\n");
3400 p = kzalloc(128, GFP_KERNEL);
3401 kfree(p);
3402 *p = 0x78;
3403 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3404 validate_slab_cache(kmalloc_caches + 7);
3405
3406 p = kzalloc(256, GFP_KERNEL);
3407 kfree(p);
3408 p[50] = 0x9a;
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003409 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3410 p);
Christoph Lameterb3459702007-05-09 02:32:41 -07003411 validate_slab_cache(kmalloc_caches + 8);
3412
3413 p = kzalloc(512, GFP_KERNEL);
3414 kfree(p);
3415 p[512] = 0xab;
3416 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3417 validate_slab_cache(kmalloc_caches + 9);
3418}
3419#else
3420static void resiliency_test(void) {};
3421#endif
3422
Christoph Lameter88a420e2007-05-06 14:49:45 -07003423/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003424 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003425 * and freed.
3426 */
3427
3428struct location {
3429 unsigned long count;
3430 void *addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003431 long long sum_time;
3432 long min_time;
3433 long max_time;
3434 long min_pid;
3435 long max_pid;
3436 cpumask_t cpus;
3437 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003438};
3439
3440struct loc_track {
3441 unsigned long max;
3442 unsigned long count;
3443 struct location *loc;
3444};
3445
3446static void free_loc_track(struct loc_track *t)
3447{
3448 if (t->max)
3449 free_pages((unsigned long)t->loc,
3450 get_order(sizeof(struct location) * t->max));
3451}
3452
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003453static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003454{
3455 struct location *l;
3456 int order;
3457
Christoph Lameter88a420e2007-05-06 14:49:45 -07003458 order = get_order(sizeof(struct location) * max);
3459
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003460 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003461 if (!l)
3462 return 0;
3463
3464 if (t->count) {
3465 memcpy(l, t->loc, sizeof(struct location) * t->count);
3466 free_loc_track(t);
3467 }
3468 t->max = max;
3469 t->loc = l;
3470 return 1;
3471}
3472
3473static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003474 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003475{
3476 long start, end, pos;
3477 struct location *l;
3478 void *caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003479 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003480
3481 start = -1;
3482 end = t->count;
3483
3484 for ( ; ; ) {
3485 pos = start + (end - start + 1) / 2;
3486
3487 /*
3488 * There is nothing at "end". If we end up there
3489 * we need to add something to before end.
3490 */
3491 if (pos == end)
3492 break;
3493
3494 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003495 if (track->addr == caddr) {
3496
3497 l = &t->loc[pos];
3498 l->count++;
3499 if (track->when) {
3500 l->sum_time += age;
3501 if (age < l->min_time)
3502 l->min_time = age;
3503 if (age > l->max_time)
3504 l->max_time = age;
3505
3506 if (track->pid < l->min_pid)
3507 l->min_pid = track->pid;
3508 if (track->pid > l->max_pid)
3509 l->max_pid = track->pid;
3510
3511 cpu_set(track->cpu, l->cpus);
3512 }
3513 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003514 return 1;
3515 }
3516
Christoph Lameter45edfa52007-05-09 02:32:45 -07003517 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003518 end = pos;
3519 else
3520 start = pos;
3521 }
3522
3523 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003524 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07003525 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003526 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003527 return 0;
3528
3529 l = t->loc + pos;
3530 if (pos < t->count)
3531 memmove(l + 1, l,
3532 (t->count - pos) * sizeof(struct location));
3533 t->count++;
3534 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003535 l->addr = track->addr;
3536 l->sum_time = age;
3537 l->min_time = age;
3538 l->max_time = age;
3539 l->min_pid = track->pid;
3540 l->max_pid = track->pid;
3541 cpus_clear(l->cpus);
3542 cpu_set(track->cpu, l->cpus);
3543 nodes_clear(l->nodes);
3544 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003545 return 1;
3546}
3547
3548static void process_slab(struct loc_track *t, struct kmem_cache *s,
3549 struct page *page, enum track_item alloc)
3550{
Christoph Lametera973e9d2008-03-01 13:40:44 -08003551 void *addr = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +03003552 DECLARE_BITMAP(map, page->objects);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003553 void *p;
3554
Christoph Lameter39b26462008-04-14 19:11:30 +03003555 bitmap_zero(map, page->objects);
Christoph Lameter7656c722007-05-09 02:32:40 -07003556 for_each_free_object(p, s, page->freelist)
3557 set_bit(slab_index(p, s, addr), map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003558
Christoph Lameter224a88b2008-04-14 19:11:31 +03003559 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07003560 if (!test_bit(slab_index(p, s, addr), map))
3561 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07003562}
3563
3564static int list_locations(struct kmem_cache *s, char *buf,
3565 enum track_item alloc)
3566{
Harvey Harrisone374d482008-01-31 15:20:50 -08003567 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003568 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003569 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07003570 int node;
3571
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003572 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
Andrew Mortonea3061d2007-10-16 01:26:09 -07003573 GFP_TEMPORARY))
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003574 return sprintf(buf, "Out of memory\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003575
3576 /* Push back cpu slabs */
3577 flush_all(s);
3578
Christoph Lameterf64dc582007-10-16 01:25:33 -07003579 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07003580 struct kmem_cache_node *n = get_node(s, node);
3581 unsigned long flags;
3582 struct page *page;
3583
Christoph Lameter9e869432007-08-22 14:01:56 -07003584 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07003585 continue;
3586
3587 spin_lock_irqsave(&n->list_lock, flags);
3588 list_for_each_entry(page, &n->partial, lru)
3589 process_slab(&t, s, page, alloc);
3590 list_for_each_entry(page, &n->full, lru)
3591 process_slab(&t, s, page, alloc);
3592 spin_unlock_irqrestore(&n->list_lock, flags);
3593 }
3594
3595 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07003596 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07003597
Harvey Harrisone374d482008-01-31 15:20:50 -08003598 if (len > PAGE_SIZE - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003599 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08003600 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003601
3602 if (l->addr)
Harvey Harrisone374d482008-01-31 15:20:50 -08003603 len += sprint_symbol(buf + len, (unsigned long)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003604 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003605 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07003606
3607 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08003608 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07003609 l->min_time,
3610 (long)div_u64(l->sum_time, l->count),
3611 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003612 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08003613 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003614 l->min_time);
3615
3616 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08003617 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003618 l->min_pid, l->max_pid);
3619 else
Harvey Harrisone374d482008-01-31 15:20:50 -08003620 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07003621 l->min_pid);
3622
Christoph Lameter84966342007-06-23 17:16:32 -07003623 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003624 len < PAGE_SIZE - 60) {
3625 len += sprintf(buf + len, " cpus=");
3626 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003627 l->cpus);
3628 }
3629
Christoph Lameter84966342007-06-23 17:16:32 -07003630 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08003631 len < PAGE_SIZE - 60) {
3632 len += sprintf(buf + len, " nodes=");
3633 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003634 l->nodes);
3635 }
3636
Harvey Harrisone374d482008-01-31 15:20:50 -08003637 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07003638 }
3639
3640 free_loc_track(&t);
3641 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08003642 len += sprintf(buf, "No data\n");
3643 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003644}
3645
Christoph Lameter81819f02007-05-06 14:49:36 -07003646enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03003647 SL_ALL, /* All slabs */
3648 SL_PARTIAL, /* Only partially allocated slabs */
3649 SL_CPU, /* Only slabs used for cpu caches */
3650 SL_OBJECTS, /* Determine allocated objects not slabs */
3651 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07003652};
3653
Christoph Lameter205ab992008-04-14 19:11:40 +03003654#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003655#define SO_PARTIAL (1 << SL_PARTIAL)
3656#define SO_CPU (1 << SL_CPU)
3657#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03003658#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07003659
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003660static ssize_t show_slab_objects(struct kmem_cache *s,
3661 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003662{
3663 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003664 int node;
3665 int x;
3666 unsigned long *nodes;
3667 unsigned long *per_cpu;
3668
3669 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03003670 if (!nodes)
3671 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07003672 per_cpu = nodes + nr_node_ids;
3673
Christoph Lameter205ab992008-04-14 19:11:40 +03003674 if (flags & SO_CPU) {
3675 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07003676
Christoph Lameter205ab992008-04-14 19:11:40 +03003677 for_each_possible_cpu(cpu) {
3678 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003679
Christoph Lameter205ab992008-04-14 19:11:40 +03003680 if (!c || c->node < 0)
3681 continue;
3682
3683 if (c->page) {
3684 if (flags & SO_TOTAL)
3685 x = c->page->objects;
3686 else if (flags & SO_OBJECTS)
3687 x = c->page->inuse;
Christoph Lameter81819f02007-05-06 14:49:36 -07003688 else
3689 x = 1;
Christoph Lameter205ab992008-04-14 19:11:40 +03003690
Christoph Lameter81819f02007-05-06 14:49:36 -07003691 total += x;
Christoph Lameter205ab992008-04-14 19:11:40 +03003692 nodes[c->node] += x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003693 }
Christoph Lameter205ab992008-04-14 19:11:40 +03003694 per_cpu[c->node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07003695 }
3696 }
3697
Christoph Lameter205ab992008-04-14 19:11:40 +03003698 if (flags & SO_ALL) {
3699 for_each_node_state(node, N_NORMAL_MEMORY) {
3700 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003701
Christoph Lameter205ab992008-04-14 19:11:40 +03003702 if (flags & SO_TOTAL)
3703 x = atomic_long_read(&n->total_objects);
3704 else if (flags & SO_OBJECTS)
3705 x = atomic_long_read(&n->total_objects) -
3706 count_partial(n, count_free);
3707
3708 else
3709 x = atomic_long_read(&n->nr_slabs);
3710 total += x;
3711 nodes[node] += x;
3712 }
3713
3714 } else if (flags & SO_PARTIAL) {
3715 for_each_node_state(node, N_NORMAL_MEMORY) {
3716 struct kmem_cache_node *n = get_node(s, node);
3717
3718 if (flags & SO_TOTAL)
3719 x = count_partial(n, count_total);
3720 else if (flags & SO_OBJECTS)
3721 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07003722 else
3723 x = n->nr_partial;
3724 total += x;
3725 nodes[node] += x;
3726 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003727 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003728 x = sprintf(buf, "%lu", total);
3729#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07003730 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07003731 if (nodes[node])
3732 x += sprintf(buf + x, " N%d=%lu",
3733 node, nodes[node]);
3734#endif
3735 kfree(nodes);
3736 return x + sprintf(buf + x, "\n");
3737}
3738
3739static int any_slab_objects(struct kmem_cache *s)
3740{
3741 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003742
3743 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003744 struct kmem_cache_node *n = get_node(s, node);
3745
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003746 if (!n)
3747 continue;
3748
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07003749 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07003750 return 1;
3751 }
3752 return 0;
3753}
3754
3755#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3756#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3757
3758struct slab_attribute {
3759 struct attribute attr;
3760 ssize_t (*show)(struct kmem_cache *s, char *buf);
3761 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3762};
3763
3764#define SLAB_ATTR_RO(_name) \
3765 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3766
3767#define SLAB_ATTR(_name) \
3768 static struct slab_attribute _name##_attr = \
3769 __ATTR(_name, 0644, _name##_show, _name##_store)
3770
Christoph Lameter81819f02007-05-06 14:49:36 -07003771static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3772{
3773 return sprintf(buf, "%d\n", s->size);
3774}
3775SLAB_ATTR_RO(slab_size);
3776
3777static ssize_t align_show(struct kmem_cache *s, char *buf)
3778{
3779 return sprintf(buf, "%d\n", s->align);
3780}
3781SLAB_ATTR_RO(align);
3782
3783static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3784{
3785 return sprintf(buf, "%d\n", s->objsize);
3786}
3787SLAB_ATTR_RO(object_size);
3788
3789static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3790{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003791 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003792}
3793SLAB_ATTR_RO(objs_per_slab);
3794
Christoph Lameter06b285d2008-04-14 19:11:41 +03003795static ssize_t order_store(struct kmem_cache *s,
3796 const char *buf, size_t length)
3797{
Christoph Lameter0121c6192008-04-29 16:11:12 -07003798 unsigned long order;
3799 int err;
3800
3801 err = strict_strtoul(buf, 10, &order);
3802 if (err)
3803 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003804
3805 if (order > slub_max_order || order < slub_min_order)
3806 return -EINVAL;
3807
3808 calculate_sizes(s, order);
3809 return length;
3810}
3811
Christoph Lameter81819f02007-05-06 14:49:36 -07003812static ssize_t order_show(struct kmem_cache *s, char *buf)
3813{
Christoph Lameter834f3d12008-04-14 19:11:31 +03003814 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07003815}
Christoph Lameter06b285d2008-04-14 19:11:41 +03003816SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003817
3818static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3819{
3820 if (s->ctor) {
3821 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3822
3823 return n + sprintf(buf + n, "\n");
3824 }
3825 return 0;
3826}
3827SLAB_ATTR_RO(ctor);
3828
Christoph Lameter81819f02007-05-06 14:49:36 -07003829static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3830{
3831 return sprintf(buf, "%d\n", s->refcount - 1);
3832}
3833SLAB_ATTR_RO(aliases);
3834
3835static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3836{
Christoph Lameter205ab992008-04-14 19:11:40 +03003837 return show_slab_objects(s, buf, SO_ALL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003838}
3839SLAB_ATTR_RO(slabs);
3840
3841static ssize_t partial_show(struct kmem_cache *s, char *buf)
3842{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003843 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003844}
3845SLAB_ATTR_RO(partial);
3846
3847static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3848{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08003849 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07003850}
3851SLAB_ATTR_RO(cpu_slabs);
3852
3853static ssize_t objects_show(struct kmem_cache *s, char *buf)
3854{
Christoph Lameter205ab992008-04-14 19:11:40 +03003855 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07003856}
3857SLAB_ATTR_RO(objects);
3858
Christoph Lameter205ab992008-04-14 19:11:40 +03003859static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3860{
3861 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3862}
3863SLAB_ATTR_RO(objects_partial);
3864
3865static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3866{
3867 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3868}
3869SLAB_ATTR_RO(total_objects);
3870
Christoph Lameter81819f02007-05-06 14:49:36 -07003871static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3872{
3873 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3874}
3875
3876static ssize_t sanity_checks_store(struct kmem_cache *s,
3877 const char *buf, size_t length)
3878{
3879 s->flags &= ~SLAB_DEBUG_FREE;
3880 if (buf[0] == '1')
3881 s->flags |= SLAB_DEBUG_FREE;
3882 return length;
3883}
3884SLAB_ATTR(sanity_checks);
3885
3886static ssize_t trace_show(struct kmem_cache *s, char *buf)
3887{
3888 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3889}
3890
3891static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3892 size_t length)
3893{
3894 s->flags &= ~SLAB_TRACE;
3895 if (buf[0] == '1')
3896 s->flags |= SLAB_TRACE;
3897 return length;
3898}
3899SLAB_ATTR(trace);
3900
3901static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3902{
3903 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3904}
3905
3906static ssize_t reclaim_account_store(struct kmem_cache *s,
3907 const char *buf, size_t length)
3908{
3909 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3910 if (buf[0] == '1')
3911 s->flags |= SLAB_RECLAIM_ACCOUNT;
3912 return length;
3913}
3914SLAB_ATTR(reclaim_account);
3915
3916static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3917{
Christoph Lameter5af60832007-05-06 14:49:56 -07003918 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003919}
3920SLAB_ATTR_RO(hwcache_align);
3921
3922#ifdef CONFIG_ZONE_DMA
3923static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3924{
3925 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3926}
3927SLAB_ATTR_RO(cache_dma);
3928#endif
3929
3930static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3931{
3932 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3933}
3934SLAB_ATTR_RO(destroy_by_rcu);
3935
3936static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3937{
3938 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3939}
3940
3941static ssize_t red_zone_store(struct kmem_cache *s,
3942 const char *buf, size_t length)
3943{
3944 if (any_slab_objects(s))
3945 return -EBUSY;
3946
3947 s->flags &= ~SLAB_RED_ZONE;
3948 if (buf[0] == '1')
3949 s->flags |= SLAB_RED_ZONE;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003950 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003951 return length;
3952}
3953SLAB_ATTR(red_zone);
3954
3955static ssize_t poison_show(struct kmem_cache *s, char *buf)
3956{
3957 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3958}
3959
3960static ssize_t poison_store(struct kmem_cache *s,
3961 const char *buf, size_t length)
3962{
3963 if (any_slab_objects(s))
3964 return -EBUSY;
3965
3966 s->flags &= ~SLAB_POISON;
3967 if (buf[0] == '1')
3968 s->flags |= SLAB_POISON;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003969 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003970 return length;
3971}
3972SLAB_ATTR(poison);
3973
3974static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3975{
3976 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3977}
3978
3979static ssize_t store_user_store(struct kmem_cache *s,
3980 const char *buf, size_t length)
3981{
3982 if (any_slab_objects(s))
3983 return -EBUSY;
3984
3985 s->flags &= ~SLAB_STORE_USER;
3986 if (buf[0] == '1')
3987 s->flags |= SLAB_STORE_USER;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003988 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003989 return length;
3990}
3991SLAB_ATTR(store_user);
3992
Christoph Lameter53e15af2007-05-06 14:49:43 -07003993static ssize_t validate_show(struct kmem_cache *s, char *buf)
3994{
3995 return 0;
3996}
3997
3998static ssize_t validate_store(struct kmem_cache *s,
3999 const char *buf, size_t length)
4000{
Christoph Lameter434e2452007-07-17 04:03:30 -07004001 int ret = -EINVAL;
4002
4003 if (buf[0] == '1') {
4004 ret = validate_slab_cache(s);
4005 if (ret >= 0)
4006 ret = length;
4007 }
4008 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004009}
4010SLAB_ATTR(validate);
4011
Christoph Lameter2086d262007-05-06 14:49:46 -07004012static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4013{
4014 return 0;
4015}
4016
4017static ssize_t shrink_store(struct kmem_cache *s,
4018 const char *buf, size_t length)
4019{
4020 if (buf[0] == '1') {
4021 int rc = kmem_cache_shrink(s);
4022
4023 if (rc)
4024 return rc;
4025 } else
4026 return -EINVAL;
4027 return length;
4028}
4029SLAB_ATTR(shrink);
4030
Christoph Lameter88a420e2007-05-06 14:49:45 -07004031static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4032{
4033 if (!(s->flags & SLAB_STORE_USER))
4034 return -ENOSYS;
4035 return list_locations(s, buf, TRACK_ALLOC);
4036}
4037SLAB_ATTR_RO(alloc_calls);
4038
4039static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4040{
4041 if (!(s->flags & SLAB_STORE_USER))
4042 return -ENOSYS;
4043 return list_locations(s, buf, TRACK_FREE);
4044}
4045SLAB_ATTR_RO(free_calls);
4046
Christoph Lameter81819f02007-05-06 14:49:36 -07004047#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004048static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004049{
Christoph Lameter98246012008-01-07 23:20:26 -08004050 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004051}
4052
Christoph Lameter98246012008-01-07 23:20:26 -08004053static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004054 const char *buf, size_t length)
4055{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004056 unsigned long ratio;
4057 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004058
Christoph Lameter0121c6192008-04-29 16:11:12 -07004059 err = strict_strtoul(buf, 10, &ratio);
4060 if (err)
4061 return err;
4062
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004063 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004064 s->remote_node_defrag_ratio = ratio * 10;
4065
Christoph Lameter81819f02007-05-06 14:49:36 -07004066 return length;
4067}
Christoph Lameter98246012008-01-07 23:20:26 -08004068SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004069#endif
4070
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004071#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004072static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4073{
4074 unsigned long sum = 0;
4075 int cpu;
4076 int len;
4077 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4078
4079 if (!data)
4080 return -ENOMEM;
4081
4082 for_each_online_cpu(cpu) {
4083 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4084
4085 data[cpu] = x;
4086 sum += x;
4087 }
4088
4089 len = sprintf(buf, "%lu", sum);
4090
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004091#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004092 for_each_online_cpu(cpu) {
4093 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004094 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004095 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004096#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004097 kfree(data);
4098 return len + sprintf(buf + len, "\n");
4099}
4100
4101#define STAT_ATTR(si, text) \
4102static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4103{ \
4104 return show_stat(s, buf, si); \
4105} \
4106SLAB_ATTR_RO(text); \
4107
4108STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4109STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4110STAT_ATTR(FREE_FASTPATH, free_fastpath);
4111STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4112STAT_ATTR(FREE_FROZEN, free_frozen);
4113STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4114STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4115STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4116STAT_ATTR(ALLOC_SLAB, alloc_slab);
4117STAT_ATTR(ALLOC_REFILL, alloc_refill);
4118STAT_ATTR(FREE_SLAB, free_slab);
4119STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4120STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4121STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4122STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4123STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4124STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter65c33762008-04-14 19:11:40 +03004125STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004126#endif
4127
Pekka Enberg06428782008-01-07 23:20:27 -08004128static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004129 &slab_size_attr.attr,
4130 &object_size_attr.attr,
4131 &objs_per_slab_attr.attr,
4132 &order_attr.attr,
4133 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004134 &objects_partial_attr.attr,
4135 &total_objects_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004136 &slabs_attr.attr,
4137 &partial_attr.attr,
4138 &cpu_slabs_attr.attr,
4139 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004140 &aliases_attr.attr,
4141 &align_attr.attr,
4142 &sanity_checks_attr.attr,
4143 &trace_attr.attr,
4144 &hwcache_align_attr.attr,
4145 &reclaim_account_attr.attr,
4146 &destroy_by_rcu_attr.attr,
4147 &red_zone_attr.attr,
4148 &poison_attr.attr,
4149 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004150 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07004151 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004152 &alloc_calls_attr.attr,
4153 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004154#ifdef CONFIG_ZONE_DMA
4155 &cache_dma_attr.attr,
4156#endif
4157#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004158 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004159#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004160#ifdef CONFIG_SLUB_STATS
4161 &alloc_fastpath_attr.attr,
4162 &alloc_slowpath_attr.attr,
4163 &free_fastpath_attr.attr,
4164 &free_slowpath_attr.attr,
4165 &free_frozen_attr.attr,
4166 &free_add_partial_attr.attr,
4167 &free_remove_partial_attr.attr,
4168 &alloc_from_partial_attr.attr,
4169 &alloc_slab_attr.attr,
4170 &alloc_refill_attr.attr,
4171 &free_slab_attr.attr,
4172 &cpuslab_flush_attr.attr,
4173 &deactivate_full_attr.attr,
4174 &deactivate_empty_attr.attr,
4175 &deactivate_to_head_attr.attr,
4176 &deactivate_to_tail_attr.attr,
4177 &deactivate_remote_frees_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004178 &order_fallback_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004179#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004180 NULL
4181};
4182
4183static struct attribute_group slab_attr_group = {
4184 .attrs = slab_attrs,
4185};
4186
4187static ssize_t slab_attr_show(struct kobject *kobj,
4188 struct attribute *attr,
4189 char *buf)
4190{
4191 struct slab_attribute *attribute;
4192 struct kmem_cache *s;
4193 int err;
4194
4195 attribute = to_slab_attr(attr);
4196 s = to_slab(kobj);
4197
4198 if (!attribute->show)
4199 return -EIO;
4200
4201 err = attribute->show(s, buf);
4202
4203 return err;
4204}
4205
4206static ssize_t slab_attr_store(struct kobject *kobj,
4207 struct attribute *attr,
4208 const char *buf, size_t len)
4209{
4210 struct slab_attribute *attribute;
4211 struct kmem_cache *s;
4212 int err;
4213
4214 attribute = to_slab_attr(attr);
4215 s = to_slab(kobj);
4216
4217 if (!attribute->store)
4218 return -EIO;
4219
4220 err = attribute->store(s, buf, len);
4221
4222 return err;
4223}
4224
Christoph Lameter151c6022008-01-07 22:29:05 -08004225static void kmem_cache_release(struct kobject *kobj)
4226{
4227 struct kmem_cache *s = to_slab(kobj);
4228
4229 kfree(s);
4230}
4231
Christoph Lameter81819f02007-05-06 14:49:36 -07004232static struct sysfs_ops slab_sysfs_ops = {
4233 .show = slab_attr_show,
4234 .store = slab_attr_store,
4235};
4236
4237static struct kobj_type slab_ktype = {
4238 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter151c6022008-01-07 22:29:05 -08004239 .release = kmem_cache_release
Christoph Lameter81819f02007-05-06 14:49:36 -07004240};
4241
4242static int uevent_filter(struct kset *kset, struct kobject *kobj)
4243{
4244 struct kobj_type *ktype = get_ktype(kobj);
4245
4246 if (ktype == &slab_ktype)
4247 return 1;
4248 return 0;
4249}
4250
4251static struct kset_uevent_ops slab_uevent_ops = {
4252 .filter = uevent_filter,
4253};
4254
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004255static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07004256
4257#define ID_STR_LENGTH 64
4258
4259/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08004260 *
4261 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07004262 */
4263static char *create_unique_id(struct kmem_cache *s)
4264{
4265 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4266 char *p = name;
4267
4268 BUG_ON(!name);
4269
4270 *p++ = ':';
4271 /*
4272 * First flags affecting slabcache operations. We will only
4273 * get here for aliasable slabs so we do not need to support
4274 * too many flags. The flags here must cover all flags that
4275 * are matched during merging to guarantee that the id is
4276 * unique.
4277 */
4278 if (s->flags & SLAB_CACHE_DMA)
4279 *p++ = 'd';
4280 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4281 *p++ = 'a';
4282 if (s->flags & SLAB_DEBUG_FREE)
4283 *p++ = 'F';
4284 if (p != name + 1)
4285 *p++ = '-';
4286 p += sprintf(p, "%07d", s->size);
4287 BUG_ON(p > name + ID_STR_LENGTH - 1);
4288 return name;
4289}
4290
4291static int sysfs_slab_add(struct kmem_cache *s)
4292{
4293 int err;
4294 const char *name;
4295 int unmergeable;
4296
4297 if (slab_state < SYSFS)
4298 /* Defer until later */
4299 return 0;
4300
4301 unmergeable = slab_unmergeable(s);
4302 if (unmergeable) {
4303 /*
4304 * Slabcache can never be merged so we can use the name proper.
4305 * This is typically the case for debug situations. In that
4306 * case we can catch duplicate names easily.
4307 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004308 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004309 name = s->name;
4310 } else {
4311 /*
4312 * Create a unique name for the slab as a target
4313 * for the symlinks.
4314 */
4315 name = create_unique_id(s);
4316 }
4317
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004318 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004319 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4320 if (err) {
4321 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004322 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07004323 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004324
4325 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4326 if (err)
4327 return err;
4328 kobject_uevent(&s->kobj, KOBJ_ADD);
4329 if (!unmergeable) {
4330 /* Setup first alias */
4331 sysfs_slab_alias(s, s->name);
4332 kfree(name);
4333 }
4334 return 0;
4335}
4336
4337static void sysfs_slab_remove(struct kmem_cache *s)
4338{
4339 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4340 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08004341 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07004342}
4343
4344/*
4345 * Need to buffer aliases during bootup until sysfs becomes
4346 * available lest we loose that information.
4347 */
4348struct saved_alias {
4349 struct kmem_cache *s;
4350 const char *name;
4351 struct saved_alias *next;
4352};
4353
Adrian Bunk5af328a2007-07-17 04:03:27 -07004354static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07004355
4356static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4357{
4358 struct saved_alias *al;
4359
4360 if (slab_state == SYSFS) {
4361 /*
4362 * If we have a leftover link then remove it.
4363 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004364 sysfs_remove_link(&slab_kset->kobj, name);
4365 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004366 }
4367
4368 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4369 if (!al)
4370 return -ENOMEM;
4371
4372 al->s = s;
4373 al->name = name;
4374 al->next = alias_list;
4375 alias_list = al;
4376 return 0;
4377}
4378
4379static int __init slab_sysfs_init(void)
4380{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004381 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004382 int err;
4383
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08004384 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06004385 if (!slab_kset) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004386 printk(KERN_ERR "Cannot register slab subsystem.\n");
4387 return -ENOSYS;
4388 }
4389
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004390 slab_state = SYSFS;
4391
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004392 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004393 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004394 if (err)
4395 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4396 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07004397 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004398
4399 while (alias_list) {
4400 struct saved_alias *al = alias_list;
4401
4402 alias_list = alias_list->next;
4403 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07004404 if (err)
4405 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4406 " %s to sysfs\n", s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07004407 kfree(al);
4408 }
4409
4410 resiliency_test();
4411 return 0;
4412}
4413
4414__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07004415#endif
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004416
4417/*
4418 * The /proc/slabinfo ABI
4419 */
Linus Torvalds158a9622008-01-02 13:04:48 -08004420#ifdef CONFIG_SLABINFO
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004421static void print_slabinfo_header(struct seq_file *m)
4422{
4423 seq_puts(m, "slabinfo - version: 2.1\n");
4424 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4425 "<objperslab> <pagesperslab>");
4426 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4427 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4428 seq_putc(m, '\n');
4429}
4430
4431static void *s_start(struct seq_file *m, loff_t *pos)
4432{
4433 loff_t n = *pos;
4434
4435 down_read(&slub_lock);
4436 if (!n)
4437 print_slabinfo_header(m);
4438
4439 return seq_list_start(&slab_caches, *pos);
4440}
4441
4442static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4443{
4444 return seq_list_next(p, &slab_caches, pos);
4445}
4446
4447static void s_stop(struct seq_file *m, void *p)
4448{
4449 up_read(&slub_lock);
4450}
4451
4452static int s_show(struct seq_file *m, void *p)
4453{
4454 unsigned long nr_partials = 0;
4455 unsigned long nr_slabs = 0;
4456 unsigned long nr_inuse = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004457 unsigned long nr_objs = 0;
4458 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004459 struct kmem_cache *s;
4460 int node;
4461
4462 s = list_entry(p, struct kmem_cache, list);
4463
4464 for_each_online_node(node) {
4465 struct kmem_cache_node *n = get_node(s, node);
4466
4467 if (!n)
4468 continue;
4469
4470 nr_partials += n->nr_partial;
4471 nr_slabs += atomic_long_read(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03004472 nr_objs += atomic_long_read(&n->total_objects);
4473 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004474 }
4475
Christoph Lameter205ab992008-04-14 19:11:40 +03004476 nr_inuse = nr_objs - nr_free;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004477
4478 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
Christoph Lameter834f3d12008-04-14 19:11:31 +03004479 nr_objs, s->size, oo_objects(s->oo),
4480 (1 << oo_order(s->oo)));
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004481 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4482 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4483 0UL);
4484 seq_putc(m, '\n');
4485 return 0;
4486}
4487
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004488static const struct seq_operations slabinfo_op = {
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01004489 .start = s_start,
4490 .next = s_next,
4491 .stop = s_stop,
4492 .show = s_show,
4493};
4494
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004495static int slabinfo_open(struct inode *inode, struct file *file)
4496{
4497 return seq_open(file, &slabinfo_op);
4498}
4499
4500static const struct file_operations proc_slabinfo_operations = {
4501 .open = slabinfo_open,
4502 .read = seq_read,
4503 .llseek = seq_lseek,
4504 .release = seq_release,
4505};
4506
4507static int __init slab_proc_init(void)
4508{
4509 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4510 return 0;
4511}
4512module_init(slab_proc_init);
Linus Torvalds158a9622008-01-02 13:04:48 -08004513#endif /* CONFIG_SLABINFO */