]>
Commit | Line | Data |
---|---|---|
1 | /* | |
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 or atomic operatios | |
6 | * and only uses a centralized lock to manage a pool of partial slabs. | |
7 | * | |
8 | * (C) 2007 SGI, Christoph Lameter | |
9 | * (C) 2011 Linux Foundation, Christoph Lameter | |
10 | */ | |
11 | ||
12 | #include <linux/mm.h> | |
13 | #include <linux/swap.h> /* struct reclaim_state */ | |
14 | #include <linux/module.h> | |
15 | #include <linux/bit_spinlock.h> | |
16 | #include <linux/interrupt.h> | |
17 | #include <linux/bitops.h> | |
18 | #include <linux/slab.h> | |
19 | #include <linux/proc_fs.h> | |
20 | #include <linux/seq_file.h> | |
21 | #include <linux/kmemcheck.h> | |
22 | #include <linux/cpu.h> | |
23 | #include <linux/cpuset.h> | |
24 | #include <linux/mempolicy.h> | |
25 | #include <linux/ctype.h> | |
26 | #include <linux/debugobjects.h> | |
27 | #include <linux/kallsyms.h> | |
28 | #include <linux/memory.h> | |
29 | #include <linux/math64.h> | |
30 | #include <linux/fault-inject.h> | |
31 | #include <linux/stacktrace.h> | |
32 | #include <linux/prefetch.h> | |
33 | ||
34 | #include <trace/events/kmem.h> | |
35 | ||
36 | /* | |
37 | * Lock order: | |
38 | * 1. slub_lock (Global Semaphore) | |
39 | * 2. node->list_lock | |
40 | * 3. slab_lock(page) (Only on some arches and for debugging) | |
41 | * | |
42 | * slub_lock | |
43 | * | |
44 | * The role of the slub_lock is to protect the list of all the slabs | |
45 | * and to synchronize major metadata changes to slab cache structures. | |
46 | * | |
47 | * The slab_lock is only used for debugging and on arches that do not | |
48 | * have the ability to do a cmpxchg_double. It only protects the second | |
49 | * double word in the page struct. Meaning | |
50 | * A. page->freelist -> List of object free in a page | |
51 | * B. page->counters -> Counters of objects | |
52 | * C. page->frozen -> frozen state | |
53 | * | |
54 | * If a slab is frozen then it is exempt from list management. It is not | |
55 | * on any list. The processor that froze the slab is the one who can | |
56 | * perform list operations on the page. Other processors may put objects | |
57 | * onto the freelist but the processor that froze the slab is the only | |
58 | * one that can retrieve the objects from the page's freelist. | |
59 | * | |
60 | * The list_lock protects the partial and full list on each node and | |
61 | * the partial slab counter. If taken then no new slabs may be added or | |
62 | * removed from the lists nor make the number of partial slabs be modified. | |
63 | * (Note that the total number of slabs is an atomic value that may be | |
64 | * modified without taking the list lock). | |
65 | * | |
66 | * The list_lock is a centralized lock and thus we avoid taking it as | |
67 | * much as possible. As long as SLUB does not have to handle partial | |
68 | * slabs, operations can continue without any centralized lock. F.e. | |
69 | * allocating a long series of objects that fill up slabs does not require | |
70 | * the list lock. | |
71 | * Interrupts are disabled during allocation and deallocation in order to | |
72 | * make the slab allocator safe to use in the context of an irq. In addition | |
73 | * interrupts are disabled to ensure that the processor does not change | |
74 | * while handling per_cpu slabs, due to kernel preemption. | |
75 | * | |
76 | * SLUB assigns one slab for allocation to each processor. | |
77 | * Allocations only occur from these slabs called cpu slabs. | |
78 | * | |
79 | * Slabs with free elements are kept on a partial list and during regular | |
80 | * operations no list for full slabs is used. If an object in a full slab is | |
81 | * freed then the slab will show up again on the partial lists. | |
82 | * We track full slabs for debugging purposes though because otherwise we | |
83 | * cannot scan all objects. | |
84 | * | |
85 | * Slabs are freed when they become empty. Teardown and setup is | |
86 | * minimal so we rely on the page allocators per cpu caches for | |
87 | * fast frees and allocs. | |
88 | * | |
89 | * Overloading of page flags that are otherwise used for LRU management. | |
90 | * | |
91 | * PageActive The slab is frozen and exempt from list processing. | |
92 | * This means that the slab is dedicated to a purpose | |
93 | * such as satisfying allocations for a specific | |
94 | * processor. Objects may be freed in the slab while | |
95 | * it is frozen but slab_free will then skip the usual | |
96 | * list operations. It is up to the processor holding | |
97 | * the slab to integrate the slab into the slab lists | |
98 | * when the slab is no longer needed. | |
99 | * | |
100 | * One use of this flag is to mark slabs that are | |
101 | * used for allocations. Then such a slab becomes a cpu | |
102 | * slab. The cpu slab may be equipped with an additional | |
103 | * freelist that allows lockless access to | |
104 | * free objects in addition to the regular freelist | |
105 | * that requires the slab lock. | |
106 | * | |
107 | * PageError Slab requires special handling due to debug | |
108 | * options set. This moves slab handling out of | |
109 | * the fast path and disables lockless freelists. | |
110 | */ | |
111 | ||
112 | #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ | |
113 | SLAB_TRACE | SLAB_DEBUG_FREE) | |
114 | ||
115 | static inline int kmem_cache_debug(struct kmem_cache *s) | |
116 | { | |
117 | #ifdef CONFIG_SLUB_DEBUG | |
118 | return unlikely(s->flags & SLAB_DEBUG_FLAGS); | |
119 | #else | |
120 | return 0; | |
121 | #endif | |
122 | } | |
123 | ||
124 | /* | |
125 | * Issues still to be resolved: | |
126 | * | |
127 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. | |
128 | * | |
129 | * - Variable sizing of the per node arrays | |
130 | */ | |
131 | ||
132 | /* Enable to test recovery from slab corruption on boot */ | |
133 | #undef SLUB_RESILIENCY_TEST | |
134 | ||
135 | /* Enable to log cmpxchg failures */ | |
136 | #undef SLUB_DEBUG_CMPXCHG | |
137 | ||
138 | /* | |
139 | * Mininum number of partial slabs. These will be left on the partial | |
140 | * lists even if they are empty. kmem_cache_shrink may reclaim them. | |
141 | */ | |
142 | #define MIN_PARTIAL 5 | |
143 | ||
144 | /* | |
145 | * Maximum number of desirable partial slabs. | |
146 | * The existence of more partial slabs makes kmem_cache_shrink | |
147 | * sort the partial list by the number of objects in the. | |
148 | */ | |
149 | #define MAX_PARTIAL 10 | |
150 | ||
151 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ | |
152 | SLAB_POISON | SLAB_STORE_USER) | |
153 | ||
154 | /* | |
155 | * Debugging flags that require metadata to be stored in the slab. These get | |
156 | * disabled when slub_debug=O is used and a cache's min order increases with | |
157 | * metadata. | |
158 | */ | |
159 | #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) | |
160 | ||
161 | /* | |
162 | * Set of flags that will prevent slab merging | |
163 | */ | |
164 | #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ | |
165 | SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ | |
166 | SLAB_FAILSLAB) | |
167 | ||
168 | #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ | |
169 | SLAB_CACHE_DMA | SLAB_NOTRACK) | |
170 | ||
171 | #define OO_SHIFT 16 | |
172 | #define OO_MASK ((1 << OO_SHIFT) - 1) | |
173 | #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */ | |
174 | ||
175 | /* Internal SLUB flags */ | |
176 | #define __OBJECT_POISON 0x80000000UL /* Poison object */ | |
177 | #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */ | |
178 | ||
179 | static int kmem_size = sizeof(struct kmem_cache); | |
180 | ||
181 | #ifdef CONFIG_SMP | |
182 | static struct notifier_block slab_notifier; | |
183 | #endif | |
184 | ||
185 | static enum { | |
186 | DOWN, /* No slab functionality available */ | |
187 | PARTIAL, /* Kmem_cache_node works */ | |
188 | UP, /* Everything works but does not show up in sysfs */ | |
189 | SYSFS /* Sysfs up */ | |
190 | } slab_state = DOWN; | |
191 | ||
192 | /* A list of all slab caches on the system */ | |
193 | static DECLARE_RWSEM(slub_lock); | |
194 | static LIST_HEAD(slab_caches); | |
195 | ||
196 | /* | |
197 | * Tracking user of a slab. | |
198 | */ | |
199 | #define TRACK_ADDRS_COUNT 16 | |
200 | struct track { | |
201 | unsigned long addr; /* Called from address */ | |
202 | #ifdef CONFIG_STACKTRACE | |
203 | unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ | |
204 | #endif | |
205 | int cpu; /* Was running on cpu */ | |
206 | int pid; /* Pid context */ | |
207 | unsigned long when; /* When did the operation occur */ | |
208 | }; | |
209 | ||
210 | enum track_item { TRACK_ALLOC, TRACK_FREE }; | |
211 | ||
212 | #ifdef CONFIG_SYSFS | |
213 | static int sysfs_slab_add(struct kmem_cache *); | |
214 | static int sysfs_slab_alias(struct kmem_cache *, const char *); | |
215 | static void sysfs_slab_remove(struct kmem_cache *); | |
216 | ||
217 | #else | |
218 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } | |
219 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) | |
220 | { return 0; } | |
221 | static inline void sysfs_slab_remove(struct kmem_cache *s) | |
222 | { | |
223 | kfree(s->name); | |
224 | kfree(s); | |
225 | } | |
226 | ||
227 | #endif | |
228 | ||
229 | static inline void stat(const struct kmem_cache *s, enum stat_item si) | |
230 | { | |
231 | #ifdef CONFIG_SLUB_STATS | |
232 | __this_cpu_inc(s->cpu_slab->stat[si]); | |
233 | #endif | |
234 | } | |
235 | ||
236 | /******************************************************************** | |
237 | * Core slab cache functions | |
238 | *******************************************************************/ | |
239 | ||
240 | int slab_is_available(void) | |
241 | { | |
242 | return slab_state >= UP; | |
243 | } | |
244 | ||
245 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) | |
246 | { | |
247 | return s->node[node]; | |
248 | } | |
249 | ||
250 | /* Verify that a pointer has an address that is valid within a slab page */ | |
251 | static inline int check_valid_pointer(struct kmem_cache *s, | |
252 | struct page *page, const void *object) | |
253 | { | |
254 | void *base; | |
255 | ||
256 | if (!object) | |
257 | return 1; | |
258 | ||
259 | base = page_address(page); | |
260 | if (object < base || object >= base + page->objects * s->size || | |
261 | (object - base) % s->size) { | |
262 | return 0; | |
263 | } | |
264 | ||
265 | return 1; | |
266 | } | |
267 | ||
268 | static inline void *get_freepointer(struct kmem_cache *s, void *object) | |
269 | { | |
270 | return *(void **)(object + s->offset); | |
271 | } | |
272 | ||
273 | static void prefetch_freepointer(const struct kmem_cache *s, void *object) | |
274 | { | |
275 | prefetch(object + s->offset); | |
276 | } | |
277 | ||
278 | static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) | |
279 | { | |
280 | void *p; | |
281 | ||
282 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
283 | probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p)); | |
284 | #else | |
285 | p = get_freepointer(s, object); | |
286 | #endif | |
287 | return p; | |
288 | } | |
289 | ||
290 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
291 | { | |
292 | *(void **)(object + s->offset) = fp; | |
293 | } | |
294 | ||
295 | /* Loop over all objects in a slab */ | |
296 | #define for_each_object(__p, __s, __addr, __objects) \ | |
297 | for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ | |
298 | __p += (__s)->size) | |
299 | ||
300 | /* Determine object index from a given position */ | |
301 | static inline int slab_index(void *p, struct kmem_cache *s, void *addr) | |
302 | { | |
303 | return (p - addr) / s->size; | |
304 | } | |
305 | ||
306 | static inline size_t slab_ksize(const struct kmem_cache *s) | |
307 | { | |
308 | #ifdef CONFIG_SLUB_DEBUG | |
309 | /* | |
310 | * Debugging requires use of the padding between object | |
311 | * and whatever may come after it. | |
312 | */ | |
313 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) | |
314 | return s->objsize; | |
315 | ||
316 | #endif | |
317 | /* | |
318 | * If we have the need to store the freelist pointer | |
319 | * back there or track user information then we can | |
320 | * only use the space before that information. | |
321 | */ | |
322 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) | |
323 | return s->inuse; | |
324 | /* | |
325 | * Else we can use all the padding etc for the allocation | |
326 | */ | |
327 | return s->size; | |
328 | } | |
329 | ||
330 | static inline int order_objects(int order, unsigned long size, int reserved) | |
331 | { | |
332 | return ((PAGE_SIZE << order) - reserved) / size; | |
333 | } | |
334 | ||
335 | static inline struct kmem_cache_order_objects oo_make(int order, | |
336 | unsigned long size, int reserved) | |
337 | { | |
338 | struct kmem_cache_order_objects x = { | |
339 | (order << OO_SHIFT) + order_objects(order, size, reserved) | |
340 | }; | |
341 | ||
342 | return x; | |
343 | } | |
344 | ||
345 | static inline int oo_order(struct kmem_cache_order_objects x) | |
346 | { | |
347 | return x.x >> OO_SHIFT; | |
348 | } | |
349 | ||
350 | static inline int oo_objects(struct kmem_cache_order_objects x) | |
351 | { | |
352 | return x.x & OO_MASK; | |
353 | } | |
354 | ||
355 | /* | |
356 | * Per slab locking using the pagelock | |
357 | */ | |
358 | static __always_inline void slab_lock(struct page *page) | |
359 | { | |
360 | bit_spin_lock(PG_locked, &page->flags); | |
361 | } | |
362 | ||
363 | static __always_inline void slab_unlock(struct page *page) | |
364 | { | |
365 | __bit_spin_unlock(PG_locked, &page->flags); | |
366 | } | |
367 | ||
368 | /* Interrupts must be disabled (for the fallback code to work right) */ | |
369 | static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, | |
370 | void *freelist_old, unsigned long counters_old, | |
371 | void *freelist_new, unsigned long counters_new, | |
372 | const char *n) | |
373 | { | |
374 | VM_BUG_ON(!irqs_disabled()); | |
375 | #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ | |
376 | defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) | |
377 | if (s->flags & __CMPXCHG_DOUBLE) { | |
378 | if (cmpxchg_double(&page->freelist, &page->counters, | |
379 | freelist_old, counters_old, | |
380 | freelist_new, counters_new)) | |
381 | return 1; | |
382 | } else | |
383 | #endif | |
384 | { | |
385 | slab_lock(page); | |
386 | if (page->freelist == freelist_old && page->counters == counters_old) { | |
387 | page->freelist = freelist_new; | |
388 | page->counters = counters_new; | |
389 | slab_unlock(page); | |
390 | return 1; | |
391 | } | |
392 | slab_unlock(page); | |
393 | } | |
394 | ||
395 | cpu_relax(); | |
396 | stat(s, CMPXCHG_DOUBLE_FAIL); | |
397 | ||
398 | #ifdef SLUB_DEBUG_CMPXCHG | |
399 | printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name); | |
400 | #endif | |
401 | ||
402 | return 0; | |
403 | } | |
404 | ||
405 | static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page, | |
406 | void *freelist_old, unsigned long counters_old, | |
407 | void *freelist_new, unsigned long counters_new, | |
408 | const char *n) | |
409 | { | |
410 | #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ | |
411 | defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) | |
412 | if (s->flags & __CMPXCHG_DOUBLE) { | |
413 | if (cmpxchg_double(&page->freelist, &page->counters, | |
414 | freelist_old, counters_old, | |
415 | freelist_new, counters_new)) | |
416 | return 1; | |
417 | } else | |
418 | #endif | |
419 | { | |
420 | unsigned long flags; | |
421 | ||
422 | local_irq_save(flags); | |
423 | slab_lock(page); | |
424 | if (page->freelist == freelist_old && page->counters == counters_old) { | |
425 | page->freelist = freelist_new; | |
426 | page->counters = counters_new; | |
427 | slab_unlock(page); | |
428 | local_irq_restore(flags); | |
429 | return 1; | |
430 | } | |
431 | slab_unlock(page); | |
432 | local_irq_restore(flags); | |
433 | } | |
434 | ||
435 | cpu_relax(); | |
436 | stat(s, CMPXCHG_DOUBLE_FAIL); | |
437 | ||
438 | #ifdef SLUB_DEBUG_CMPXCHG | |
439 | printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name); | |
440 | #endif | |
441 | ||
442 | return 0; | |
443 | } | |
444 | ||
445 | #ifdef CONFIG_SLUB_DEBUG | |
446 | /* | |
447 | * Determine a map of object in use on a page. | |
448 | * | |
449 | * Node listlock must be held to guarantee that the page does | |
450 | * not vanish from under us. | |
451 | */ | |
452 | static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) | |
453 | { | |
454 | void *p; | |
455 | void *addr = page_address(page); | |
456 | ||
457 | for (p = page->freelist; p; p = get_freepointer(s, p)) | |
458 | set_bit(slab_index(p, s, addr), map); | |
459 | } | |
460 | ||
461 | /* | |
462 | * Debug settings: | |
463 | */ | |
464 | #ifdef CONFIG_SLUB_DEBUG_ON | |
465 | static int slub_debug = DEBUG_DEFAULT_FLAGS; | |
466 | #else | |
467 | static int slub_debug; | |
468 | #endif | |
469 | ||
470 | static char *slub_debug_slabs; | |
471 | static int disable_higher_order_debug; | |
472 | ||
473 | /* | |
474 | * Object debugging | |
475 | */ | |
476 | static void print_section(char *text, u8 *addr, unsigned int length) | |
477 | { | |
478 | print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr, | |
479 | length, 1); | |
480 | } | |
481 | ||
482 | static struct track *get_track(struct kmem_cache *s, void *object, | |
483 | enum track_item alloc) | |
484 | { | |
485 | struct track *p; | |
486 | ||
487 | if (s->offset) | |
488 | p = object + s->offset + sizeof(void *); | |
489 | else | |
490 | p = object + s->inuse; | |
491 | ||
492 | return p + alloc; | |
493 | } | |
494 | ||
495 | static void set_track(struct kmem_cache *s, void *object, | |
496 | enum track_item alloc, unsigned long addr) | |
497 | { | |
498 | struct track *p = get_track(s, object, alloc); | |
499 | ||
500 | if (addr) { | |
501 | #ifdef CONFIG_STACKTRACE | |
502 | struct stack_trace trace; | |
503 | int i; | |
504 | ||
505 | trace.nr_entries = 0; | |
506 | trace.max_entries = TRACK_ADDRS_COUNT; | |
507 | trace.entries = p->addrs; | |
508 | trace.skip = 3; | |
509 | save_stack_trace(&trace); | |
510 | ||
511 | /* See rant in lockdep.c */ | |
512 | if (trace.nr_entries != 0 && | |
513 | trace.entries[trace.nr_entries - 1] == ULONG_MAX) | |
514 | trace.nr_entries--; | |
515 | ||
516 | for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++) | |
517 | p->addrs[i] = 0; | |
518 | #endif | |
519 | p->addr = addr; | |
520 | p->cpu = smp_processor_id(); | |
521 | p->pid = current->pid; | |
522 | p->when = jiffies; | |
523 | } else | |
524 | memset(p, 0, sizeof(struct track)); | |
525 | } | |
526 | ||
527 | static void init_tracking(struct kmem_cache *s, void *object) | |
528 | { | |
529 | if (!(s->flags & SLAB_STORE_USER)) | |
530 | return; | |
531 | ||
532 | set_track(s, object, TRACK_FREE, 0UL); | |
533 | set_track(s, object, TRACK_ALLOC, 0UL); | |
534 | } | |
535 | ||
536 | static void print_track(const char *s, struct track *t) | |
537 | { | |
538 | if (!t->addr) | |
539 | return; | |
540 | ||
541 | printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", | |
542 | s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); | |
543 | #ifdef CONFIG_STACKTRACE | |
544 | { | |
545 | int i; | |
546 | for (i = 0; i < TRACK_ADDRS_COUNT; i++) | |
547 | if (t->addrs[i]) | |
548 | printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]); | |
549 | else | |
550 | break; | |
551 | } | |
552 | #endif | |
553 | } | |
554 | ||
555 | static void print_tracking(struct kmem_cache *s, void *object) | |
556 | { | |
557 | if (!(s->flags & SLAB_STORE_USER)) | |
558 | return; | |
559 | ||
560 | print_track("Allocated", get_track(s, object, TRACK_ALLOC)); | |
561 | print_track("Freed", get_track(s, object, TRACK_FREE)); | |
562 | } | |
563 | ||
564 | static void print_page_info(struct page *page) | |
565 | { | |
566 | printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", | |
567 | page, page->objects, page->inuse, page->freelist, page->flags); | |
568 | ||
569 | } | |
570 | ||
571 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) | |
572 | { | |
573 | va_list args; | |
574 | char buf[100]; | |
575 | ||
576 | va_start(args, fmt); | |
577 | vsnprintf(buf, sizeof(buf), fmt, args); | |
578 | va_end(args); | |
579 | printk(KERN_ERR "========================================" | |
580 | "=====================================\n"); | |
581 | printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf); | |
582 | printk(KERN_ERR "----------------------------------------" | |
583 | "-------------------------------------\n\n"); | |
584 | } | |
585 | ||
586 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) | |
587 | { | |
588 | va_list args; | |
589 | char buf[100]; | |
590 | ||
591 | va_start(args, fmt); | |
592 | vsnprintf(buf, sizeof(buf), fmt, args); | |
593 | va_end(args); | |
594 | printk(KERN_ERR "FIX %s: %s\n", s->name, buf); | |
595 | } | |
596 | ||
597 | static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) | |
598 | { | |
599 | unsigned int off; /* Offset of last byte */ | |
600 | u8 *addr = page_address(page); | |
601 | ||
602 | print_tracking(s, p); | |
603 | ||
604 | print_page_info(page); | |
605 | ||
606 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", | |
607 | p, p - addr, get_freepointer(s, p)); | |
608 | ||
609 | if (p > addr + 16) | |
610 | print_section("Bytes b4 ", p - 16, 16); | |
611 | ||
612 | print_section("Object ", p, min_t(unsigned long, s->objsize, | |
613 | PAGE_SIZE)); | |
614 | if (s->flags & SLAB_RED_ZONE) | |
615 | print_section("Redzone ", p + s->objsize, | |
616 | s->inuse - s->objsize); | |
617 | ||
618 | if (s->offset) | |
619 | off = s->offset + sizeof(void *); | |
620 | else | |
621 | off = s->inuse; | |
622 | ||
623 | if (s->flags & SLAB_STORE_USER) | |
624 | off += 2 * sizeof(struct track); | |
625 | ||
626 | if (off != s->size) | |
627 | /* Beginning of the filler is the free pointer */ | |
628 | print_section("Padding ", p + off, s->size - off); | |
629 | ||
630 | dump_stack(); | |
631 | } | |
632 | ||
633 | static void object_err(struct kmem_cache *s, struct page *page, | |
634 | u8 *object, char *reason) | |
635 | { | |
636 | slab_bug(s, "%s", reason); | |
637 | print_trailer(s, page, object); | |
638 | } | |
639 | ||
640 | static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) | |
641 | { | |
642 | va_list args; | |
643 | char buf[100]; | |
644 | ||
645 | va_start(args, fmt); | |
646 | vsnprintf(buf, sizeof(buf), fmt, args); | |
647 | va_end(args); | |
648 | slab_bug(s, "%s", buf); | |
649 | print_page_info(page); | |
650 | dump_stack(); | |
651 | } | |
652 | ||
653 | static void init_object(struct kmem_cache *s, void *object, u8 val) | |
654 | { | |
655 | u8 *p = object; | |
656 | ||
657 | if (s->flags & __OBJECT_POISON) { | |
658 | memset(p, POISON_FREE, s->objsize - 1); | |
659 | p[s->objsize - 1] = POISON_END; | |
660 | } | |
661 | ||
662 | if (s->flags & SLAB_RED_ZONE) | |
663 | memset(p + s->objsize, val, s->inuse - s->objsize); | |
664 | } | |
665 | ||
666 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
667 | void *from, void *to) | |
668 | { | |
669 | slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); | |
670 | memset(from, data, to - from); | |
671 | } | |
672 | ||
673 | static int check_bytes_and_report(struct kmem_cache *s, struct page *page, | |
674 | u8 *object, char *what, | |
675 | u8 *start, unsigned int value, unsigned int bytes) | |
676 | { | |
677 | u8 *fault; | |
678 | u8 *end; | |
679 | ||
680 | fault = memchr_inv(start, value, bytes); | |
681 | if (!fault) | |
682 | return 1; | |
683 | ||
684 | end = start + bytes; | |
685 | while (end > fault && end[-1] == value) | |
686 | end--; | |
687 | ||
688 | slab_bug(s, "%s overwritten", what); | |
689 | printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", | |
690 | fault, end - 1, fault[0], value); | |
691 | print_trailer(s, page, object); | |
692 | ||
693 | restore_bytes(s, what, value, fault, end); | |
694 | return 0; | |
695 | } | |
696 | ||
697 | /* | |
698 | * Object layout: | |
699 | * | |
700 | * object address | |
701 | * Bytes of the object to be managed. | |
702 | * If the freepointer may overlay the object then the free | |
703 | * pointer is the first word of the object. | |
704 | * | |
705 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is | |
706 | * 0xa5 (POISON_END) | |
707 | * | |
708 | * object + s->objsize | |
709 | * Padding to reach word boundary. This is also used for Redzoning. | |
710 | * Padding is extended by another word if Redzoning is enabled and | |
711 | * objsize == inuse. | |
712 | * | |
713 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with | |
714 | * 0xcc (RED_ACTIVE) for objects in use. | |
715 | * | |
716 | * object + s->inuse | |
717 | * Meta data starts here. | |
718 | * | |
719 | * A. Free pointer (if we cannot overwrite object on free) | |
720 | * B. Tracking data for SLAB_STORE_USER | |
721 | * C. Padding to reach required alignment boundary or at mininum | |
722 | * one word if debugging is on to be able to detect writes | |
723 | * before the word boundary. | |
724 | * | |
725 | * Padding is done using 0x5a (POISON_INUSE) | |
726 | * | |
727 | * object + s->size | |
728 | * Nothing is used beyond s->size. | |
729 | * | |
730 | * If slabcaches are merged then the objsize and inuse boundaries are mostly | |
731 | * ignored. And therefore no slab options that rely on these boundaries | |
732 | * may be used with merged slabcaches. | |
733 | */ | |
734 | ||
735 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) | |
736 | { | |
737 | unsigned long off = s->inuse; /* The end of info */ | |
738 | ||
739 | if (s->offset) | |
740 | /* Freepointer is placed after the object. */ | |
741 | off += sizeof(void *); | |
742 | ||
743 | if (s->flags & SLAB_STORE_USER) | |
744 | /* We also have user information there */ | |
745 | off += 2 * sizeof(struct track); | |
746 | ||
747 | if (s->size == off) | |
748 | return 1; | |
749 | ||
750 | return check_bytes_and_report(s, page, p, "Object padding", | |
751 | p + off, POISON_INUSE, s->size - off); | |
752 | } | |
753 | ||
754 | /* Check the pad bytes at the end of a slab page */ | |
755 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
756 | { | |
757 | u8 *start; | |
758 | u8 *fault; | |
759 | u8 *end; | |
760 | int length; | |
761 | int remainder; | |
762 | ||
763 | if (!(s->flags & SLAB_POISON)) | |
764 | return 1; | |
765 | ||
766 | start = page_address(page); | |
767 | length = (PAGE_SIZE << compound_order(page)) - s->reserved; | |
768 | end = start + length; | |
769 | remainder = length % s->size; | |
770 | if (!remainder) | |
771 | return 1; | |
772 | ||
773 | fault = memchr_inv(end - remainder, POISON_INUSE, remainder); | |
774 | if (!fault) | |
775 | return 1; | |
776 | while (end > fault && end[-1] == POISON_INUSE) | |
777 | end--; | |
778 | ||
779 | slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); | |
780 | print_section("Padding ", end - remainder, remainder); | |
781 | ||
782 | restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); | |
783 | return 0; | |
784 | } | |
785 | ||
786 | static int check_object(struct kmem_cache *s, struct page *page, | |
787 | void *object, u8 val) | |
788 | { | |
789 | u8 *p = object; | |
790 | u8 *endobject = object + s->objsize; | |
791 | ||
792 | if (s->flags & SLAB_RED_ZONE) { | |
793 | if (!check_bytes_and_report(s, page, object, "Redzone", | |
794 | endobject, val, s->inuse - s->objsize)) | |
795 | return 0; | |
796 | } else { | |
797 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { | |
798 | check_bytes_and_report(s, page, p, "Alignment padding", | |
799 | endobject, POISON_INUSE, s->inuse - s->objsize); | |
800 | } | |
801 | } | |
802 | ||
803 | if (s->flags & SLAB_POISON) { | |
804 | if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && | |
805 | (!check_bytes_and_report(s, page, p, "Poison", p, | |
806 | POISON_FREE, s->objsize - 1) || | |
807 | !check_bytes_and_report(s, page, p, "Poison", | |
808 | p + s->objsize - 1, POISON_END, 1))) | |
809 | return 0; | |
810 | /* | |
811 | * check_pad_bytes cleans up on its own. | |
812 | */ | |
813 | check_pad_bytes(s, page, p); | |
814 | } | |
815 | ||
816 | if (!s->offset && val == SLUB_RED_ACTIVE) | |
817 | /* | |
818 | * Object and freepointer overlap. Cannot check | |
819 | * freepointer while object is allocated. | |
820 | */ | |
821 | return 1; | |
822 | ||
823 | /* Check free pointer validity */ | |
824 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
825 | object_err(s, page, p, "Freepointer corrupt"); | |
826 | /* | |
827 | * No choice but to zap it and thus lose the remainder | |
828 | * of the free objects in this slab. May cause | |
829 | * another error because the object count is now wrong. | |
830 | */ | |
831 | set_freepointer(s, p, NULL); | |
832 | return 0; | |
833 | } | |
834 | return 1; | |
835 | } | |
836 | ||
837 | static int check_slab(struct kmem_cache *s, struct page *page) | |
838 | { | |
839 | int maxobj; | |
840 | ||
841 | VM_BUG_ON(!irqs_disabled()); | |
842 | ||
843 | if (!PageSlab(page)) { | |
844 | slab_err(s, page, "Not a valid slab page"); | |
845 | return 0; | |
846 | } | |
847 | ||
848 | maxobj = order_objects(compound_order(page), s->size, s->reserved); | |
849 | if (page->objects > maxobj) { | |
850 | slab_err(s, page, "objects %u > max %u", | |
851 | s->name, page->objects, maxobj); | |
852 | return 0; | |
853 | } | |
854 | if (page->inuse > page->objects) { | |
855 | slab_err(s, page, "inuse %u > max %u", | |
856 | s->name, page->inuse, page->objects); | |
857 | return 0; | |
858 | } | |
859 | /* Slab_pad_check fixes things up after itself */ | |
860 | slab_pad_check(s, page); | |
861 | return 1; | |
862 | } | |
863 | ||
864 | /* | |
865 | * Determine if a certain object on a page is on the freelist. Must hold the | |
866 | * slab lock to guarantee that the chains are in a consistent state. | |
867 | */ | |
868 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
869 | { | |
870 | int nr = 0; | |
871 | void *fp; | |
872 | void *object = NULL; | |
873 | unsigned long max_objects; | |
874 | ||
875 | fp = page->freelist; | |
876 | while (fp && nr <= page->objects) { | |
877 | if (fp == search) | |
878 | return 1; | |
879 | if (!check_valid_pointer(s, page, fp)) { | |
880 | if (object) { | |
881 | object_err(s, page, object, | |
882 | "Freechain corrupt"); | |
883 | set_freepointer(s, object, NULL); | |
884 | break; | |
885 | } else { | |
886 | slab_err(s, page, "Freepointer corrupt"); | |
887 | page->freelist = NULL; | |
888 | page->inuse = page->objects; | |
889 | slab_fix(s, "Freelist cleared"); | |
890 | return 0; | |
891 | } | |
892 | break; | |
893 | } | |
894 | object = fp; | |
895 | fp = get_freepointer(s, object); | |
896 | nr++; | |
897 | } | |
898 | ||
899 | max_objects = order_objects(compound_order(page), s->size, s->reserved); | |
900 | if (max_objects > MAX_OBJS_PER_PAGE) | |
901 | max_objects = MAX_OBJS_PER_PAGE; | |
902 | ||
903 | if (page->objects != max_objects) { | |
904 | slab_err(s, page, "Wrong number of objects. Found %d but " | |
905 | "should be %d", page->objects, max_objects); | |
906 | page->objects = max_objects; | |
907 | slab_fix(s, "Number of objects adjusted."); | |
908 | } | |
909 | if (page->inuse != page->objects - nr) { | |
910 | slab_err(s, page, "Wrong object count. Counter is %d but " | |
911 | "counted were %d", page->inuse, page->objects - nr); | |
912 | page->inuse = page->objects - nr; | |
913 | slab_fix(s, "Object count adjusted."); | |
914 | } | |
915 | return search == NULL; | |
916 | } | |
917 | ||
918 | static void trace(struct kmem_cache *s, struct page *page, void *object, | |
919 | int alloc) | |
920 | { | |
921 | if (s->flags & SLAB_TRACE) { | |
922 | printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", | |
923 | s->name, | |
924 | alloc ? "alloc" : "free", | |
925 | object, page->inuse, | |
926 | page->freelist); | |
927 | ||
928 | if (!alloc) | |
929 | print_section("Object ", (void *)object, s->objsize); | |
930 | ||
931 | dump_stack(); | |
932 | } | |
933 | } | |
934 | ||
935 | /* | |
936 | * Hooks for other subsystems that check memory allocations. In a typical | |
937 | * production configuration these hooks all should produce no code at all. | |
938 | */ | |
939 | static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) | |
940 | { | |
941 | flags &= gfp_allowed_mask; | |
942 | lockdep_trace_alloc(flags); | |
943 | might_sleep_if(flags & __GFP_WAIT); | |
944 | ||
945 | return should_failslab(s->objsize, flags, s->flags); | |
946 | } | |
947 | ||
948 | static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object) | |
949 | { | |
950 | flags &= gfp_allowed_mask; | |
951 | kmemcheck_slab_alloc(s, flags, object, slab_ksize(s)); | |
952 | kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags); | |
953 | } | |
954 | ||
955 | static inline void slab_free_hook(struct kmem_cache *s, void *x) | |
956 | { | |
957 | kmemleak_free_recursive(x, s->flags); | |
958 | ||
959 | /* | |
960 | * Trouble is that we may no longer disable interupts in the fast path | |
961 | * So in order to make the debug calls that expect irqs to be | |
962 | * disabled we need to disable interrupts temporarily. | |
963 | */ | |
964 | #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP) | |
965 | { | |
966 | unsigned long flags; | |
967 | ||
968 | local_irq_save(flags); | |
969 | kmemcheck_slab_free(s, x, s->objsize); | |
970 | debug_check_no_locks_freed(x, s->objsize); | |
971 | local_irq_restore(flags); | |
972 | } | |
973 | #endif | |
974 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) | |
975 | debug_check_no_obj_freed(x, s->objsize); | |
976 | } | |
977 | ||
978 | /* | |
979 | * Tracking of fully allocated slabs for debugging purposes. | |
980 | * | |
981 | * list_lock must be held. | |
982 | */ | |
983 | static void add_full(struct kmem_cache *s, | |
984 | struct kmem_cache_node *n, struct page *page) | |
985 | { | |
986 | if (!(s->flags & SLAB_STORE_USER)) | |
987 | return; | |
988 | ||
989 | list_add(&page->lru, &n->full); | |
990 | } | |
991 | ||
992 | /* | |
993 | * list_lock must be held. | |
994 | */ | |
995 | static void remove_full(struct kmem_cache *s, struct page *page) | |
996 | { | |
997 | if (!(s->flags & SLAB_STORE_USER)) | |
998 | return; | |
999 | ||
1000 | list_del(&page->lru); | |
1001 | } | |
1002 | ||
1003 | /* Tracking of the number of slabs for debugging purposes */ | |
1004 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
1005 | { | |
1006 | struct kmem_cache_node *n = get_node(s, node); | |
1007 | ||
1008 | return atomic_long_read(&n->nr_slabs); | |
1009 | } | |
1010 | ||
1011 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) | |
1012 | { | |
1013 | return atomic_long_read(&n->nr_slabs); | |
1014 | } | |
1015 | ||
1016 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) | |
1017 | { | |
1018 | struct kmem_cache_node *n = get_node(s, node); | |
1019 | ||
1020 | /* | |
1021 | * May be called early in order to allocate a slab for the | |
1022 | * kmem_cache_node structure. Solve the chicken-egg | |
1023 | * dilemma by deferring the increment of the count during | |
1024 | * bootstrap (see early_kmem_cache_node_alloc). | |
1025 | */ | |
1026 | if (n) { | |
1027 | atomic_long_inc(&n->nr_slabs); | |
1028 | atomic_long_add(objects, &n->total_objects); | |
1029 | } | |
1030 | } | |
1031 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) | |
1032 | { | |
1033 | struct kmem_cache_node *n = get_node(s, node); | |
1034 | ||
1035 | atomic_long_dec(&n->nr_slabs); | |
1036 | atomic_long_sub(objects, &n->total_objects); | |
1037 | } | |
1038 | ||
1039 | /* Object debug checks for alloc/free paths */ | |
1040 | static void setup_object_debug(struct kmem_cache *s, struct page *page, | |
1041 | void *object) | |
1042 | { | |
1043 | if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) | |
1044 | return; | |
1045 | ||
1046 | init_object(s, object, SLUB_RED_INACTIVE); | |
1047 | init_tracking(s, object); | |
1048 | } | |
1049 | ||
1050 | static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page, | |
1051 | void *object, unsigned long addr) | |
1052 | { | |
1053 | if (!check_slab(s, page)) | |
1054 | goto bad; | |
1055 | ||
1056 | if (!check_valid_pointer(s, page, object)) { | |
1057 | object_err(s, page, object, "Freelist Pointer check fails"); | |
1058 | goto bad; | |
1059 | } | |
1060 | ||
1061 | if (!check_object(s, page, object, SLUB_RED_INACTIVE)) | |
1062 | goto bad; | |
1063 | ||
1064 | /* Success perform special debug activities for allocs */ | |
1065 | if (s->flags & SLAB_STORE_USER) | |
1066 | set_track(s, object, TRACK_ALLOC, addr); | |
1067 | trace(s, page, object, 1); | |
1068 | init_object(s, object, SLUB_RED_ACTIVE); | |
1069 | return 1; | |
1070 | ||
1071 | bad: | |
1072 | if (PageSlab(page)) { | |
1073 | /* | |
1074 | * If this is a slab page then lets do the best we can | |
1075 | * to avoid issues in the future. Marking all objects | |
1076 | * as used avoids touching the remaining objects. | |
1077 | */ | |
1078 | slab_fix(s, "Marking all objects used"); | |
1079 | page->inuse = page->objects; | |
1080 | page->freelist = NULL; | |
1081 | } | |
1082 | return 0; | |
1083 | } | |
1084 | ||
1085 | static noinline int free_debug_processing(struct kmem_cache *s, | |
1086 | struct page *page, void *object, unsigned long addr) | |
1087 | { | |
1088 | unsigned long flags; | |
1089 | int rc = 0; | |
1090 | ||
1091 | local_irq_save(flags); | |
1092 | slab_lock(page); | |
1093 | ||
1094 | if (!check_slab(s, page)) | |
1095 | goto fail; | |
1096 | ||
1097 | if (!check_valid_pointer(s, page, object)) { | |
1098 | slab_err(s, page, "Invalid object pointer 0x%p", object); | |
1099 | goto fail; | |
1100 | } | |
1101 | ||
1102 | if (on_freelist(s, page, object)) { | |
1103 | object_err(s, page, object, "Object already free"); | |
1104 | goto fail; | |
1105 | } | |
1106 | ||
1107 | if (!check_object(s, page, object, SLUB_RED_ACTIVE)) | |
1108 | goto out; | |
1109 | ||
1110 | if (unlikely(s != page->slab)) { | |
1111 | if (!PageSlab(page)) { | |
1112 | slab_err(s, page, "Attempt to free object(0x%p) " | |
1113 | "outside of slab", object); | |
1114 | } else if (!page->slab) { | |
1115 | printk(KERN_ERR | |
1116 | "SLUB <none>: no slab for object 0x%p.\n", | |
1117 | object); | |
1118 | dump_stack(); | |
1119 | } else | |
1120 | object_err(s, page, object, | |
1121 | "page slab pointer corrupt."); | |
1122 | goto fail; | |
1123 | } | |
1124 | ||
1125 | if (s->flags & SLAB_STORE_USER) | |
1126 | set_track(s, object, TRACK_FREE, addr); | |
1127 | trace(s, page, object, 0); | |
1128 | init_object(s, object, SLUB_RED_INACTIVE); | |
1129 | rc = 1; | |
1130 | out: | |
1131 | slab_unlock(page); | |
1132 | local_irq_restore(flags); | |
1133 | return rc; | |
1134 | ||
1135 | fail: | |
1136 | slab_fix(s, "Object at 0x%p not freed", object); | |
1137 | goto out; | |
1138 | } | |
1139 | ||
1140 | static int __init setup_slub_debug(char *str) | |
1141 | { | |
1142 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
1143 | if (*str++ != '=' || !*str) | |
1144 | /* | |
1145 | * No options specified. Switch on full debugging. | |
1146 | */ | |
1147 | goto out; | |
1148 | ||
1149 | if (*str == ',') | |
1150 | /* | |
1151 | * No options but restriction on slabs. This means full | |
1152 | * debugging for slabs matching a pattern. | |
1153 | */ | |
1154 | goto check_slabs; | |
1155 | ||
1156 | if (tolower(*str) == 'o') { | |
1157 | /* | |
1158 | * Avoid enabling debugging on caches if its minimum order | |
1159 | * would increase as a result. | |
1160 | */ | |
1161 | disable_higher_order_debug = 1; | |
1162 | goto out; | |
1163 | } | |
1164 | ||
1165 | slub_debug = 0; | |
1166 | if (*str == '-') | |
1167 | /* | |
1168 | * Switch off all debugging measures. | |
1169 | */ | |
1170 | goto out; | |
1171 | ||
1172 | /* | |
1173 | * Determine which debug features should be switched on | |
1174 | */ | |
1175 | for (; *str && *str != ','; str++) { | |
1176 | switch (tolower(*str)) { | |
1177 | case 'f': | |
1178 | slub_debug |= SLAB_DEBUG_FREE; | |
1179 | break; | |
1180 | case 'z': | |
1181 | slub_debug |= SLAB_RED_ZONE; | |
1182 | break; | |
1183 | case 'p': | |
1184 | slub_debug |= SLAB_POISON; | |
1185 | break; | |
1186 | case 'u': | |
1187 | slub_debug |= SLAB_STORE_USER; | |
1188 | break; | |
1189 | case 't': | |
1190 | slub_debug |= SLAB_TRACE; | |
1191 | break; | |
1192 | case 'a': | |
1193 | slub_debug |= SLAB_FAILSLAB; | |
1194 | break; | |
1195 | default: | |
1196 | printk(KERN_ERR "slub_debug option '%c' " | |
1197 | "unknown. skipped\n", *str); | |
1198 | } | |
1199 | } | |
1200 | ||
1201 | check_slabs: | |
1202 | if (*str == ',') | |
1203 | slub_debug_slabs = str + 1; | |
1204 | out: | |
1205 | return 1; | |
1206 | } | |
1207 | ||
1208 | __setup("slub_debug", setup_slub_debug); | |
1209 | ||
1210 | static unsigned long kmem_cache_flags(unsigned long objsize, | |
1211 | unsigned long flags, const char *name, | |
1212 | void (*ctor)(void *)) | |
1213 | { | |
1214 | /* | |
1215 | * Enable debugging if selected on the kernel commandline. | |
1216 | */ | |
1217 | if (slub_debug && (!slub_debug_slabs || | |
1218 | !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) | |
1219 | flags |= slub_debug; | |
1220 | ||
1221 | return flags; | |
1222 | } | |
1223 | #else | |
1224 | static inline void setup_object_debug(struct kmem_cache *s, | |
1225 | struct page *page, void *object) {} | |
1226 | ||
1227 | static inline int alloc_debug_processing(struct kmem_cache *s, | |
1228 | struct page *page, void *object, unsigned long addr) { return 0; } | |
1229 | ||
1230 | static inline int free_debug_processing(struct kmem_cache *s, | |
1231 | struct page *page, void *object, unsigned long addr) { return 0; } | |
1232 | ||
1233 | static inline int slab_pad_check(struct kmem_cache *s, struct page *page) | |
1234 | { return 1; } | |
1235 | static inline int check_object(struct kmem_cache *s, struct page *page, | |
1236 | void *object, u8 val) { return 1; } | |
1237 | static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, | |
1238 | struct page *page) {} | |
1239 | static inline void remove_full(struct kmem_cache *s, struct page *page) {} | |
1240 | static inline unsigned long kmem_cache_flags(unsigned long objsize, | |
1241 | unsigned long flags, const char *name, | |
1242 | void (*ctor)(void *)) | |
1243 | { | |
1244 | return flags; | |
1245 | } | |
1246 | #define slub_debug 0 | |
1247 | ||
1248 | #define disable_higher_order_debug 0 | |
1249 | ||
1250 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
1251 | { return 0; } | |
1252 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) | |
1253 | { return 0; } | |
1254 | static inline void inc_slabs_node(struct kmem_cache *s, int node, | |
1255 | int objects) {} | |
1256 | static inline void dec_slabs_node(struct kmem_cache *s, int node, | |
1257 | int objects) {} | |
1258 | ||
1259 | static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) | |
1260 | { return 0; } | |
1261 | ||
1262 | static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, | |
1263 | void *object) {} | |
1264 | ||
1265 | static inline void slab_free_hook(struct kmem_cache *s, void *x) {} | |
1266 | ||
1267 | #endif /* CONFIG_SLUB_DEBUG */ | |
1268 | ||
1269 | /* | |
1270 | * Slab allocation and freeing | |
1271 | */ | |
1272 | static inline struct page *alloc_slab_page(gfp_t flags, int node, | |
1273 | struct kmem_cache_order_objects oo) | |
1274 | { | |
1275 | int order = oo_order(oo); | |
1276 | ||
1277 | flags |= __GFP_NOTRACK; | |
1278 | ||
1279 | if (node == NUMA_NO_NODE) | |
1280 | return alloc_pages(flags, order); | |
1281 | else | |
1282 | return alloc_pages_exact_node(node, flags, order); | |
1283 | } | |
1284 | ||
1285 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1286 | { | |
1287 | struct page *page; | |
1288 | struct kmem_cache_order_objects oo = s->oo; | |
1289 | gfp_t alloc_gfp; | |
1290 | ||
1291 | flags &= gfp_allowed_mask; | |
1292 | ||
1293 | if (flags & __GFP_WAIT) | |
1294 | local_irq_enable(); | |
1295 | ||
1296 | flags |= s->allocflags; | |
1297 | ||
1298 | /* | |
1299 | * Let the initial higher-order allocation fail under memory pressure | |
1300 | * so we fall-back to the minimum order allocation. | |
1301 | */ | |
1302 | alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; | |
1303 | ||
1304 | page = alloc_slab_page(alloc_gfp, node, oo); | |
1305 | if (unlikely(!page)) { | |
1306 | oo = s->min; | |
1307 | /* | |
1308 | * Allocation may have failed due to fragmentation. | |
1309 | * Try a lower order alloc if possible | |
1310 | */ | |
1311 | page = alloc_slab_page(flags, node, oo); | |
1312 | ||
1313 | if (page) | |
1314 | stat(s, ORDER_FALLBACK); | |
1315 | } | |
1316 | ||
1317 | if (flags & __GFP_WAIT) | |
1318 | local_irq_disable(); | |
1319 | ||
1320 | if (!page) | |
1321 | return NULL; | |
1322 | ||
1323 | if (kmemcheck_enabled | |
1324 | && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { | |
1325 | int pages = 1 << oo_order(oo); | |
1326 | ||
1327 | kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); | |
1328 | ||
1329 | /* | |
1330 | * Objects from caches that have a constructor don't get | |
1331 | * cleared when they're allocated, so we need to do it here. | |
1332 | */ | |
1333 | if (s->ctor) | |
1334 | kmemcheck_mark_uninitialized_pages(page, pages); | |
1335 | else | |
1336 | kmemcheck_mark_unallocated_pages(page, pages); | |
1337 | } | |
1338 | ||
1339 | page->objects = oo_objects(oo); | |
1340 | mod_zone_page_state(page_zone(page), | |
1341 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1342 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1343 | 1 << oo_order(oo)); | |
1344 | ||
1345 | return page; | |
1346 | } | |
1347 | ||
1348 | static void setup_object(struct kmem_cache *s, struct page *page, | |
1349 | void *object) | |
1350 | { | |
1351 | setup_object_debug(s, page, object); | |
1352 | if (unlikely(s->ctor)) | |
1353 | s->ctor(object); | |
1354 | } | |
1355 | ||
1356 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1357 | { | |
1358 | struct page *page; | |
1359 | void *start; | |
1360 | void *last; | |
1361 | void *p; | |
1362 | ||
1363 | BUG_ON(flags & GFP_SLAB_BUG_MASK); | |
1364 | ||
1365 | page = allocate_slab(s, | |
1366 | flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); | |
1367 | if (!page) | |
1368 | goto out; | |
1369 | ||
1370 | inc_slabs_node(s, page_to_nid(page), page->objects); | |
1371 | page->slab = s; | |
1372 | __SetPageSlab(page); | |
1373 | ||
1374 | start = page_address(page); | |
1375 | ||
1376 | if (unlikely(s->flags & SLAB_POISON)) | |
1377 | memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); | |
1378 | ||
1379 | last = start; | |
1380 | for_each_object(p, s, start, page->objects) { | |
1381 | setup_object(s, page, last); | |
1382 | set_freepointer(s, last, p); | |
1383 | last = p; | |
1384 | } | |
1385 | setup_object(s, page, last); | |
1386 | set_freepointer(s, last, NULL); | |
1387 | ||
1388 | page->freelist = start; | |
1389 | page->inuse = page->objects; | |
1390 | page->frozen = 1; | |
1391 | out: | |
1392 | return page; | |
1393 | } | |
1394 | ||
1395 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
1396 | { | |
1397 | int order = compound_order(page); | |
1398 | int pages = 1 << order; | |
1399 | ||
1400 | if (kmem_cache_debug(s)) { | |
1401 | void *p; | |
1402 | ||
1403 | slab_pad_check(s, page); | |
1404 | for_each_object(p, s, page_address(page), | |
1405 | page->objects) | |
1406 | check_object(s, page, p, SLUB_RED_INACTIVE); | |
1407 | } | |
1408 | ||
1409 | kmemcheck_free_shadow(page, compound_order(page)); | |
1410 | ||
1411 | mod_zone_page_state(page_zone(page), | |
1412 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1413 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1414 | -pages); | |
1415 | ||
1416 | __ClearPageSlab(page); | |
1417 | reset_page_mapcount(page); | |
1418 | if (current->reclaim_state) | |
1419 | current->reclaim_state->reclaimed_slab += pages; | |
1420 | __free_pages(page, order); | |
1421 | } | |
1422 | ||
1423 | #define need_reserve_slab_rcu \ | |
1424 | (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) | |
1425 | ||
1426 | static void rcu_free_slab(struct rcu_head *h) | |
1427 | { | |
1428 | struct page *page; | |
1429 | ||
1430 | if (need_reserve_slab_rcu) | |
1431 | page = virt_to_head_page(h); | |
1432 | else | |
1433 | page = container_of((struct list_head *)h, struct page, lru); | |
1434 | ||
1435 | __free_slab(page->slab, page); | |
1436 | } | |
1437 | ||
1438 | static void free_slab(struct kmem_cache *s, struct page *page) | |
1439 | { | |
1440 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { | |
1441 | struct rcu_head *head; | |
1442 | ||
1443 | if (need_reserve_slab_rcu) { | |
1444 | int order = compound_order(page); | |
1445 | int offset = (PAGE_SIZE << order) - s->reserved; | |
1446 | ||
1447 | VM_BUG_ON(s->reserved != sizeof(*head)); | |
1448 | head = page_address(page) + offset; | |
1449 | } else { | |
1450 | /* | |
1451 | * RCU free overloads the RCU head over the LRU | |
1452 | */ | |
1453 | head = (void *)&page->lru; | |
1454 | } | |
1455 | ||
1456 | call_rcu(head, rcu_free_slab); | |
1457 | } else | |
1458 | __free_slab(s, page); | |
1459 | } | |
1460 | ||
1461 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
1462 | { | |
1463 | dec_slabs_node(s, page_to_nid(page), page->objects); | |
1464 | free_slab(s, page); | |
1465 | } | |
1466 | ||
1467 | /* | |
1468 | * Management of partially allocated slabs. | |
1469 | * | |
1470 | * list_lock must be held. | |
1471 | */ | |
1472 | static inline void add_partial(struct kmem_cache_node *n, | |
1473 | struct page *page, int tail) | |
1474 | { | |
1475 | n->nr_partial++; | |
1476 | if (tail == DEACTIVATE_TO_TAIL) | |
1477 | list_add_tail(&page->lru, &n->partial); | |
1478 | else | |
1479 | list_add(&page->lru, &n->partial); | |
1480 | } | |
1481 | ||
1482 | /* | |
1483 | * list_lock must be held. | |
1484 | */ | |
1485 | static inline void remove_partial(struct kmem_cache_node *n, | |
1486 | struct page *page) | |
1487 | { | |
1488 | list_del(&page->lru); | |
1489 | n->nr_partial--; | |
1490 | } | |
1491 | ||
1492 | /* | |
1493 | * Remove slab from the partial list, freeze it and | |
1494 | * return the pointer to the freelist. | |
1495 | * | |
1496 | * Returns a list of objects or NULL if it fails. | |
1497 | * | |
1498 | * Must hold list_lock since we modify the partial list. | |
1499 | */ | |
1500 | static inline void *acquire_slab(struct kmem_cache *s, | |
1501 | struct kmem_cache_node *n, struct page *page, | |
1502 | int mode) | |
1503 | { | |
1504 | void *freelist; | |
1505 | unsigned long counters; | |
1506 | struct page new; | |
1507 | ||
1508 | /* | |
1509 | * Zap the freelist and set the frozen bit. | |
1510 | * The old freelist is the list of objects for the | |
1511 | * per cpu allocation list. | |
1512 | */ | |
1513 | freelist = page->freelist; | |
1514 | counters = page->counters; | |
1515 | new.counters = counters; | |
1516 | if (mode) { | |
1517 | new.inuse = page->objects; | |
1518 | new.freelist = NULL; | |
1519 | } else { | |
1520 | new.freelist = freelist; | |
1521 | } | |
1522 | ||
1523 | VM_BUG_ON(new.frozen); | |
1524 | new.frozen = 1; | |
1525 | ||
1526 | if (!__cmpxchg_double_slab(s, page, | |
1527 | freelist, counters, | |
1528 | new.freelist, new.counters, | |
1529 | "acquire_slab")) | |
1530 | return NULL; | |
1531 | ||
1532 | remove_partial(n, page); | |
1533 | WARN_ON(!freelist); | |
1534 | return freelist; | |
1535 | } | |
1536 | ||
1537 | static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); | |
1538 | ||
1539 | /* | |
1540 | * Try to allocate a partial slab from a specific node. | |
1541 | */ | |
1542 | static void *get_partial_node(struct kmem_cache *s, | |
1543 | struct kmem_cache_node *n, struct kmem_cache_cpu *c) | |
1544 | { | |
1545 | struct page *page, *page2; | |
1546 | void *object = NULL; | |
1547 | ||
1548 | /* | |
1549 | * Racy check. If we mistakenly see no partial slabs then we | |
1550 | * just allocate an empty slab. If we mistakenly try to get a | |
1551 | * partial slab and there is none available then get_partials() | |
1552 | * will return NULL. | |
1553 | */ | |
1554 | if (!n || !n->nr_partial) | |
1555 | return NULL; | |
1556 | ||
1557 | spin_lock(&n->list_lock); | |
1558 | list_for_each_entry_safe(page, page2, &n->partial, lru) { | |
1559 | void *t = acquire_slab(s, n, page, object == NULL); | |
1560 | int available; | |
1561 | ||
1562 | if (!t) | |
1563 | break; | |
1564 | ||
1565 | if (!object) { | |
1566 | c->page = page; | |
1567 | stat(s, ALLOC_FROM_PARTIAL); | |
1568 | object = t; | |
1569 | available = page->objects - page->inuse; | |
1570 | } else { | |
1571 | available = put_cpu_partial(s, page, 0); | |
1572 | stat(s, CPU_PARTIAL_NODE); | |
1573 | } | |
1574 | if (kmem_cache_debug(s) || available > s->cpu_partial / 2) | |
1575 | break; | |
1576 | ||
1577 | } | |
1578 | spin_unlock(&n->list_lock); | |
1579 | return object; | |
1580 | } | |
1581 | ||
1582 | /* | |
1583 | * Get a page from somewhere. Search in increasing NUMA distances. | |
1584 | */ | |
1585 | static void *get_any_partial(struct kmem_cache *s, gfp_t flags, | |
1586 | struct kmem_cache_cpu *c) | |
1587 | { | |
1588 | #ifdef CONFIG_NUMA | |
1589 | struct zonelist *zonelist; | |
1590 | struct zoneref *z; | |
1591 | struct zone *zone; | |
1592 | enum zone_type high_zoneidx = gfp_zone(flags); | |
1593 | void *object; | |
1594 | unsigned int cpuset_mems_cookie; | |
1595 | ||
1596 | /* | |
1597 | * The defrag ratio allows a configuration of the tradeoffs between | |
1598 | * inter node defragmentation and node local allocations. A lower | |
1599 | * defrag_ratio increases the tendency to do local allocations | |
1600 | * instead of attempting to obtain partial slabs from other nodes. | |
1601 | * | |
1602 | * If the defrag_ratio is set to 0 then kmalloc() always | |
1603 | * returns node local objects. If the ratio is higher then kmalloc() | |
1604 | * may return off node objects because partial slabs are obtained | |
1605 | * from other nodes and filled up. | |
1606 | * | |
1607 | * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes | |
1608 | * defrag_ratio = 1000) then every (well almost) allocation will | |
1609 | * first attempt to defrag slab caches on other nodes. This means | |
1610 | * scanning over all nodes to look for partial slabs which may be | |
1611 | * expensive if we do it every time we are trying to find a slab | |
1612 | * with available objects. | |
1613 | */ | |
1614 | if (!s->remote_node_defrag_ratio || | |
1615 | get_cycles() % 1024 > s->remote_node_defrag_ratio) | |
1616 | return NULL; | |
1617 | ||
1618 | do { | |
1619 | cpuset_mems_cookie = get_mems_allowed(); | |
1620 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); | |
1621 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { | |
1622 | struct kmem_cache_node *n; | |
1623 | ||
1624 | n = get_node(s, zone_to_nid(zone)); | |
1625 | ||
1626 | if (n && cpuset_zone_allowed_hardwall(zone, flags) && | |
1627 | n->nr_partial > s->min_partial) { | |
1628 | object = get_partial_node(s, n, c); | |
1629 | if (object) { | |
1630 | /* | |
1631 | * Return the object even if | |
1632 | * put_mems_allowed indicated that | |
1633 | * the cpuset mems_allowed was | |
1634 | * updated in parallel. It's a | |
1635 | * harmless race between the alloc | |
1636 | * and the cpuset update. | |
1637 | */ | |
1638 | put_mems_allowed(cpuset_mems_cookie); | |
1639 | return object; | |
1640 | } | |
1641 | } | |
1642 | } | |
1643 | } while (!put_mems_allowed(cpuset_mems_cookie)); | |
1644 | #endif | |
1645 | return NULL; | |
1646 | } | |
1647 | ||
1648 | /* | |
1649 | * Get a partial page, lock it and return it. | |
1650 | */ | |
1651 | static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, | |
1652 | struct kmem_cache_cpu *c) | |
1653 | { | |
1654 | void *object; | |
1655 | int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; | |
1656 | ||
1657 | object = get_partial_node(s, get_node(s, searchnode), c); | |
1658 | if (object || node != NUMA_NO_NODE) | |
1659 | return object; | |
1660 | ||
1661 | return get_any_partial(s, flags, c); | |
1662 | } | |
1663 | ||
1664 | #ifdef CONFIG_PREEMPT | |
1665 | /* | |
1666 | * Calculate the next globally unique transaction for disambiguiation | |
1667 | * during cmpxchg. The transactions start with the cpu number and are then | |
1668 | * incremented by CONFIG_NR_CPUS. | |
1669 | */ | |
1670 | #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) | |
1671 | #else | |
1672 | /* | |
1673 | * No preemption supported therefore also no need to check for | |
1674 | * different cpus. | |
1675 | */ | |
1676 | #define TID_STEP 1 | |
1677 | #endif | |
1678 | ||
1679 | static inline unsigned long next_tid(unsigned long tid) | |
1680 | { | |
1681 | return tid + TID_STEP; | |
1682 | } | |
1683 | ||
1684 | static inline unsigned int tid_to_cpu(unsigned long tid) | |
1685 | { | |
1686 | return tid % TID_STEP; | |
1687 | } | |
1688 | ||
1689 | static inline unsigned long tid_to_event(unsigned long tid) | |
1690 | { | |
1691 | return tid / TID_STEP; | |
1692 | } | |
1693 | ||
1694 | static inline unsigned int init_tid(int cpu) | |
1695 | { | |
1696 | return cpu; | |
1697 | } | |
1698 | ||
1699 | static inline void note_cmpxchg_failure(const char *n, | |
1700 | const struct kmem_cache *s, unsigned long tid) | |
1701 | { | |
1702 | #ifdef SLUB_DEBUG_CMPXCHG | |
1703 | unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); | |
1704 | ||
1705 | printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name); | |
1706 | ||
1707 | #ifdef CONFIG_PREEMPT | |
1708 | if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) | |
1709 | printk("due to cpu change %d -> %d\n", | |
1710 | tid_to_cpu(tid), tid_to_cpu(actual_tid)); | |
1711 | else | |
1712 | #endif | |
1713 | if (tid_to_event(tid) != tid_to_event(actual_tid)) | |
1714 | printk("due to cpu running other code. Event %ld->%ld\n", | |
1715 | tid_to_event(tid), tid_to_event(actual_tid)); | |
1716 | else | |
1717 | printk("for unknown reason: actual=%lx was=%lx target=%lx\n", | |
1718 | actual_tid, tid, next_tid(tid)); | |
1719 | #endif | |
1720 | stat(s, CMPXCHG_DOUBLE_CPU_FAIL); | |
1721 | } | |
1722 | ||
1723 | void init_kmem_cache_cpus(struct kmem_cache *s) | |
1724 | { | |
1725 | int cpu; | |
1726 | ||
1727 | for_each_possible_cpu(cpu) | |
1728 | per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); | |
1729 | } | |
1730 | ||
1731 | /* | |
1732 | * Remove the cpu slab | |
1733 | */ | |
1734 | static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist) | |
1735 | { | |
1736 | enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; | |
1737 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1738 | int lock = 0; | |
1739 | enum slab_modes l = M_NONE, m = M_NONE; | |
1740 | void *nextfree; | |
1741 | int tail = DEACTIVATE_TO_HEAD; | |
1742 | struct page new; | |
1743 | struct page old; | |
1744 | ||
1745 | if (page->freelist) { | |
1746 | stat(s, DEACTIVATE_REMOTE_FREES); | |
1747 | tail = DEACTIVATE_TO_TAIL; | |
1748 | } | |
1749 | ||
1750 | /* | |
1751 | * Stage one: Free all available per cpu objects back | |
1752 | * to the page freelist while it is still frozen. Leave the | |
1753 | * last one. | |
1754 | * | |
1755 | * There is no need to take the list->lock because the page | |
1756 | * is still frozen. | |
1757 | */ | |
1758 | while (freelist && (nextfree = get_freepointer(s, freelist))) { | |
1759 | void *prior; | |
1760 | unsigned long counters; | |
1761 | ||
1762 | do { | |
1763 | prior = page->freelist; | |
1764 | counters = page->counters; | |
1765 | set_freepointer(s, freelist, prior); | |
1766 | new.counters = counters; | |
1767 | new.inuse--; | |
1768 | VM_BUG_ON(!new.frozen); | |
1769 | ||
1770 | } while (!__cmpxchg_double_slab(s, page, | |
1771 | prior, counters, | |
1772 | freelist, new.counters, | |
1773 | "drain percpu freelist")); | |
1774 | ||
1775 | freelist = nextfree; | |
1776 | } | |
1777 | ||
1778 | /* | |
1779 | * Stage two: Ensure that the page is unfrozen while the | |
1780 | * list presence reflects the actual number of objects | |
1781 | * during unfreeze. | |
1782 | * | |
1783 | * We setup the list membership and then perform a cmpxchg | |
1784 | * with the count. If there is a mismatch then the page | |
1785 | * is not unfrozen but the page is on the wrong list. | |
1786 | * | |
1787 | * Then we restart the process which may have to remove | |
1788 | * the page from the list that we just put it on again | |
1789 | * because the number of objects in the slab may have | |
1790 | * changed. | |
1791 | */ | |
1792 | redo: | |
1793 | ||
1794 | old.freelist = page->freelist; | |
1795 | old.counters = page->counters; | |
1796 | VM_BUG_ON(!old.frozen); | |
1797 | ||
1798 | /* Determine target state of the slab */ | |
1799 | new.counters = old.counters; | |
1800 | if (freelist) { | |
1801 | new.inuse--; | |
1802 | set_freepointer(s, freelist, old.freelist); | |
1803 | new.freelist = freelist; | |
1804 | } else | |
1805 | new.freelist = old.freelist; | |
1806 | ||
1807 | new.frozen = 0; | |
1808 | ||
1809 | if (!new.inuse && n->nr_partial > s->min_partial) | |
1810 | m = M_FREE; | |
1811 | else if (new.freelist) { | |
1812 | m = M_PARTIAL; | |
1813 | if (!lock) { | |
1814 | lock = 1; | |
1815 | /* | |
1816 | * Taking the spinlock removes the possiblity | |
1817 | * that acquire_slab() will see a slab page that | |
1818 | * is frozen | |
1819 | */ | |
1820 | spin_lock(&n->list_lock); | |
1821 | } | |
1822 | } else { | |
1823 | m = M_FULL; | |
1824 | if (kmem_cache_debug(s) && !lock) { | |
1825 | lock = 1; | |
1826 | /* | |
1827 | * This also ensures that the scanning of full | |
1828 | * slabs from diagnostic functions will not see | |
1829 | * any frozen slabs. | |
1830 | */ | |
1831 | spin_lock(&n->list_lock); | |
1832 | } | |
1833 | } | |
1834 | ||
1835 | if (l != m) { | |
1836 | ||
1837 | if (l == M_PARTIAL) | |
1838 | ||
1839 | remove_partial(n, page); | |
1840 | ||
1841 | else if (l == M_FULL) | |
1842 | ||
1843 | remove_full(s, page); | |
1844 | ||
1845 | if (m == M_PARTIAL) { | |
1846 | ||
1847 | add_partial(n, page, tail); | |
1848 | stat(s, tail); | |
1849 | ||
1850 | } else if (m == M_FULL) { | |
1851 | ||
1852 | stat(s, DEACTIVATE_FULL); | |
1853 | add_full(s, n, page); | |
1854 | ||
1855 | } | |
1856 | } | |
1857 | ||
1858 | l = m; | |
1859 | if (!__cmpxchg_double_slab(s, page, | |
1860 | old.freelist, old.counters, | |
1861 | new.freelist, new.counters, | |
1862 | "unfreezing slab")) | |
1863 | goto redo; | |
1864 | ||
1865 | if (lock) | |
1866 | spin_unlock(&n->list_lock); | |
1867 | ||
1868 | if (m == M_FREE) { | |
1869 | stat(s, DEACTIVATE_EMPTY); | |
1870 | discard_slab(s, page); | |
1871 | stat(s, FREE_SLAB); | |
1872 | } | |
1873 | } | |
1874 | ||
1875 | /* Unfreeze all the cpu partial slabs */ | |
1876 | static void unfreeze_partials(struct kmem_cache *s) | |
1877 | { | |
1878 | struct kmem_cache_node *n = NULL; | |
1879 | struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab); | |
1880 | struct page *page, *discard_page = NULL; | |
1881 | ||
1882 | while ((page = c->partial)) { | |
1883 | enum slab_modes { M_PARTIAL, M_FREE }; | |
1884 | enum slab_modes l, m; | |
1885 | struct page new; | |
1886 | struct page old; | |
1887 | ||
1888 | c->partial = page->next; | |
1889 | l = M_FREE; | |
1890 | ||
1891 | do { | |
1892 | ||
1893 | old.freelist = page->freelist; | |
1894 | old.counters = page->counters; | |
1895 | VM_BUG_ON(!old.frozen); | |
1896 | ||
1897 | new.counters = old.counters; | |
1898 | new.freelist = old.freelist; | |
1899 | ||
1900 | new.frozen = 0; | |
1901 | ||
1902 | if (!new.inuse && (!n || n->nr_partial > s->min_partial)) | |
1903 | m = M_FREE; | |
1904 | else { | |
1905 | struct kmem_cache_node *n2 = get_node(s, | |
1906 | page_to_nid(page)); | |
1907 | ||
1908 | m = M_PARTIAL; | |
1909 | if (n != n2) { | |
1910 | if (n) | |
1911 | spin_unlock(&n->list_lock); | |
1912 | ||
1913 | n = n2; | |
1914 | spin_lock(&n->list_lock); | |
1915 | } | |
1916 | } | |
1917 | ||
1918 | if (l != m) { | |
1919 | if (l == M_PARTIAL) { | |
1920 | remove_partial(n, page); | |
1921 | stat(s, FREE_REMOVE_PARTIAL); | |
1922 | } else { | |
1923 | add_partial(n, page, | |
1924 | DEACTIVATE_TO_TAIL); | |
1925 | stat(s, FREE_ADD_PARTIAL); | |
1926 | } | |
1927 | ||
1928 | l = m; | |
1929 | } | |
1930 | ||
1931 | } while (!cmpxchg_double_slab(s, page, | |
1932 | old.freelist, old.counters, | |
1933 | new.freelist, new.counters, | |
1934 | "unfreezing slab")); | |
1935 | ||
1936 | if (m == M_FREE) { | |
1937 | page->next = discard_page; | |
1938 | discard_page = page; | |
1939 | } | |
1940 | } | |
1941 | ||
1942 | if (n) | |
1943 | spin_unlock(&n->list_lock); | |
1944 | ||
1945 | while (discard_page) { | |
1946 | page = discard_page; | |
1947 | discard_page = discard_page->next; | |
1948 | ||
1949 | stat(s, DEACTIVATE_EMPTY); | |
1950 | discard_slab(s, page); | |
1951 | stat(s, FREE_SLAB); | |
1952 | } | |
1953 | } | |
1954 | ||
1955 | /* | |
1956 | * Put a page that was just frozen (in __slab_free) into a partial page | |
1957 | * slot if available. This is done without interrupts disabled and without | |
1958 | * preemption disabled. The cmpxchg is racy and may put the partial page | |
1959 | * onto a random cpus partial slot. | |
1960 | * | |
1961 | * If we did not find a slot then simply move all the partials to the | |
1962 | * per node partial list. | |
1963 | */ | |
1964 | int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) | |
1965 | { | |
1966 | struct page *oldpage; | |
1967 | int pages; | |
1968 | int pobjects; | |
1969 | ||
1970 | do { | |
1971 | pages = 0; | |
1972 | pobjects = 0; | |
1973 | oldpage = this_cpu_read(s->cpu_slab->partial); | |
1974 | ||
1975 | if (oldpage) { | |
1976 | pobjects = oldpage->pobjects; | |
1977 | pages = oldpage->pages; | |
1978 | if (drain && pobjects > s->cpu_partial) { | |
1979 | unsigned long flags; | |
1980 | /* | |
1981 | * partial array is full. Move the existing | |
1982 | * set to the per node partial list. | |
1983 | */ | |
1984 | local_irq_save(flags); | |
1985 | unfreeze_partials(s); | |
1986 | local_irq_restore(flags); | |
1987 | pobjects = 0; | |
1988 | pages = 0; | |
1989 | stat(s, CPU_PARTIAL_DRAIN); | |
1990 | } | |
1991 | } | |
1992 | ||
1993 | pages++; | |
1994 | pobjects += page->objects - page->inuse; | |
1995 | ||
1996 | page->pages = pages; | |
1997 | page->pobjects = pobjects; | |
1998 | page->next = oldpage; | |
1999 | ||
2000 | } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage); | |
2001 | return pobjects; | |
2002 | } | |
2003 | ||
2004 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) | |
2005 | { | |
2006 | stat(s, CPUSLAB_FLUSH); | |
2007 | deactivate_slab(s, c->page, c->freelist); | |
2008 | ||
2009 | c->tid = next_tid(c->tid); | |
2010 | c->page = NULL; | |
2011 | c->freelist = NULL; | |
2012 | } | |
2013 | ||
2014 | /* | |
2015 | * Flush cpu slab. | |
2016 | * | |
2017 | * Called from IPI handler with interrupts disabled. | |
2018 | */ | |
2019 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) | |
2020 | { | |
2021 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); | |
2022 | ||
2023 | if (likely(c)) { | |
2024 | if (c->page) | |
2025 | flush_slab(s, c); | |
2026 | ||
2027 | unfreeze_partials(s); | |
2028 | } | |
2029 | } | |
2030 | ||
2031 | static void flush_cpu_slab(void *d) | |
2032 | { | |
2033 | struct kmem_cache *s = d; | |
2034 | ||
2035 | __flush_cpu_slab(s, smp_processor_id()); | |
2036 | } | |
2037 | ||
2038 | static bool has_cpu_slab(int cpu, void *info) | |
2039 | { | |
2040 | struct kmem_cache *s = info; | |
2041 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); | |
2042 | ||
2043 | return c->page || c->partial; | |
2044 | } | |
2045 | ||
2046 | static void flush_all(struct kmem_cache *s) | |
2047 | { | |
2048 | on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC); | |
2049 | } | |
2050 | ||
2051 | /* | |
2052 | * Check if the objects in a per cpu structure fit numa | |
2053 | * locality expectations. | |
2054 | */ | |
2055 | static inline int node_match(struct page *page, int node) | |
2056 | { | |
2057 | #ifdef CONFIG_NUMA | |
2058 | if (node != NUMA_NO_NODE && page_to_nid(page) != node) | |
2059 | return 0; | |
2060 | #endif | |
2061 | return 1; | |
2062 | } | |
2063 | ||
2064 | static int count_free(struct page *page) | |
2065 | { | |
2066 | return page->objects - page->inuse; | |
2067 | } | |
2068 | ||
2069 | static unsigned long count_partial(struct kmem_cache_node *n, | |
2070 | int (*get_count)(struct page *)) | |
2071 | { | |
2072 | unsigned long flags; | |
2073 | unsigned long x = 0; | |
2074 | struct page *page; | |
2075 | ||
2076 | spin_lock_irqsave(&n->list_lock, flags); | |
2077 | list_for_each_entry(page, &n->partial, lru) | |
2078 | x += get_count(page); | |
2079 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2080 | return x; | |
2081 | } | |
2082 | ||
2083 | static inline unsigned long node_nr_objs(struct kmem_cache_node *n) | |
2084 | { | |
2085 | #ifdef CONFIG_SLUB_DEBUG | |
2086 | return atomic_long_read(&n->total_objects); | |
2087 | #else | |
2088 | return 0; | |
2089 | #endif | |
2090 | } | |
2091 | ||
2092 | static noinline void | |
2093 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) | |
2094 | { | |
2095 | int node; | |
2096 | ||
2097 | printk(KERN_WARNING | |
2098 | "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
2099 | nid, gfpflags); | |
2100 | printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " | |
2101 | "default order: %d, min order: %d\n", s->name, s->objsize, | |
2102 | s->size, oo_order(s->oo), oo_order(s->min)); | |
2103 | ||
2104 | if (oo_order(s->min) > get_order(s->objsize)) | |
2105 | printk(KERN_WARNING " %s debugging increased min order, use " | |
2106 | "slub_debug=O to disable.\n", s->name); | |
2107 | ||
2108 | for_each_online_node(node) { | |
2109 | struct kmem_cache_node *n = get_node(s, node); | |
2110 | unsigned long nr_slabs; | |
2111 | unsigned long nr_objs; | |
2112 | unsigned long nr_free; | |
2113 | ||
2114 | if (!n) | |
2115 | continue; | |
2116 | ||
2117 | nr_free = count_partial(n, count_free); | |
2118 | nr_slabs = node_nr_slabs(n); | |
2119 | nr_objs = node_nr_objs(n); | |
2120 | ||
2121 | printk(KERN_WARNING | |
2122 | " node %d: slabs: %ld, objs: %ld, free: %ld\n", | |
2123 | node, nr_slabs, nr_objs, nr_free); | |
2124 | } | |
2125 | } | |
2126 | ||
2127 | static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, | |
2128 | int node, struct kmem_cache_cpu **pc) | |
2129 | { | |
2130 | void *freelist; | |
2131 | struct kmem_cache_cpu *c = *pc; | |
2132 | struct page *page; | |
2133 | ||
2134 | freelist = get_partial(s, flags, node, c); | |
2135 | ||
2136 | if (freelist) | |
2137 | return freelist; | |
2138 | ||
2139 | page = new_slab(s, flags, node); | |
2140 | if (page) { | |
2141 | c = __this_cpu_ptr(s->cpu_slab); | |
2142 | if (c->page) | |
2143 | flush_slab(s, c); | |
2144 | ||
2145 | /* | |
2146 | * No other reference to the page yet so we can | |
2147 | * muck around with it freely without cmpxchg | |
2148 | */ | |
2149 | freelist = page->freelist; | |
2150 | page->freelist = NULL; | |
2151 | ||
2152 | stat(s, ALLOC_SLAB); | |
2153 | c->page = page; | |
2154 | *pc = c; | |
2155 | } else | |
2156 | freelist = NULL; | |
2157 | ||
2158 | return freelist; | |
2159 | } | |
2160 | ||
2161 | /* | |
2162 | * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist | |
2163 | * or deactivate the page. | |
2164 | * | |
2165 | * The page is still frozen if the return value is not NULL. | |
2166 | * | |
2167 | * If this function returns NULL then the page has been unfrozen. | |
2168 | */ | |
2169 | static inline void *get_freelist(struct kmem_cache *s, struct page *page) | |
2170 | { | |
2171 | struct page new; | |
2172 | unsigned long counters; | |
2173 | void *freelist; | |
2174 | ||
2175 | do { | |
2176 | freelist = page->freelist; | |
2177 | counters = page->counters; | |
2178 | ||
2179 | new.counters = counters; | |
2180 | VM_BUG_ON(!new.frozen); | |
2181 | ||
2182 | new.inuse = page->objects; | |
2183 | new.frozen = freelist != NULL; | |
2184 | ||
2185 | } while (!cmpxchg_double_slab(s, page, | |
2186 | freelist, counters, | |
2187 | NULL, new.counters, | |
2188 | "get_freelist")); | |
2189 | ||
2190 | return freelist; | |
2191 | } | |
2192 | ||
2193 | /* | |
2194 | * Slow path. The lockless freelist is empty or we need to perform | |
2195 | * debugging duties. | |
2196 | * | |
2197 | * Processing is still very fast if new objects have been freed to the | |
2198 | * regular freelist. In that case we simply take over the regular freelist | |
2199 | * as the lockless freelist and zap the regular freelist. | |
2200 | * | |
2201 | * If that is not working then we fall back to the partial lists. We take the | |
2202 | * first element of the freelist as the object to allocate now and move the | |
2203 | * rest of the freelist to the lockless freelist. | |
2204 | * | |
2205 | * And if we were unable to get a new slab from the partial slab lists then | |
2206 | * we need to allocate a new slab. This is the slowest path since it involves | |
2207 | * a call to the page allocator and the setup of a new slab. | |
2208 | */ | |
2209 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, | |
2210 | unsigned long addr, struct kmem_cache_cpu *c) | |
2211 | { | |
2212 | void *freelist; | |
2213 | struct page *page; | |
2214 | unsigned long flags; | |
2215 | ||
2216 | local_irq_save(flags); | |
2217 | #ifdef CONFIG_PREEMPT | |
2218 | /* | |
2219 | * We may have been preempted and rescheduled on a different | |
2220 | * cpu before disabling interrupts. Need to reload cpu area | |
2221 | * pointer. | |
2222 | */ | |
2223 | c = this_cpu_ptr(s->cpu_slab); | |
2224 | #endif | |
2225 | ||
2226 | page = c->page; | |
2227 | if (!page) | |
2228 | goto new_slab; | |
2229 | redo: | |
2230 | ||
2231 | if (unlikely(!node_match(page, node))) { | |
2232 | stat(s, ALLOC_NODE_MISMATCH); | |
2233 | deactivate_slab(s, page, c->freelist); | |
2234 | c->page = NULL; | |
2235 | c->freelist = NULL; | |
2236 | goto new_slab; | |
2237 | } | |
2238 | ||
2239 | /* must check again c->freelist in case of cpu migration or IRQ */ | |
2240 | freelist = c->freelist; | |
2241 | if (freelist) | |
2242 | goto load_freelist; | |
2243 | ||
2244 | stat(s, ALLOC_SLOWPATH); | |
2245 | ||
2246 | freelist = get_freelist(s, page); | |
2247 | ||
2248 | if (!freelist) { | |
2249 | c->page = NULL; | |
2250 | stat(s, DEACTIVATE_BYPASS); | |
2251 | goto new_slab; | |
2252 | } | |
2253 | ||
2254 | stat(s, ALLOC_REFILL); | |
2255 | ||
2256 | load_freelist: | |
2257 | /* | |
2258 | * freelist is pointing to the list of objects to be used. | |
2259 | * page is pointing to the page from which the objects are obtained. | |
2260 | * That page must be frozen for per cpu allocations to work. | |
2261 | */ | |
2262 | VM_BUG_ON(!c->page->frozen); | |
2263 | c->freelist = get_freepointer(s, freelist); | |
2264 | c->tid = next_tid(c->tid); | |
2265 | local_irq_restore(flags); | |
2266 | return freelist; | |
2267 | ||
2268 | new_slab: | |
2269 | ||
2270 | if (c->partial) { | |
2271 | page = c->page = c->partial; | |
2272 | c->partial = page->next; | |
2273 | stat(s, CPU_PARTIAL_ALLOC); | |
2274 | c->freelist = NULL; | |
2275 | goto redo; | |
2276 | } | |
2277 | ||
2278 | freelist = new_slab_objects(s, gfpflags, node, &c); | |
2279 | ||
2280 | if (unlikely(!freelist)) { | |
2281 | if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) | |
2282 | slab_out_of_memory(s, gfpflags, node); | |
2283 | ||
2284 | local_irq_restore(flags); | |
2285 | return NULL; | |
2286 | } | |
2287 | ||
2288 | page = c->page; | |
2289 | if (likely(!kmem_cache_debug(s))) | |
2290 | goto load_freelist; | |
2291 | ||
2292 | /* Only entered in the debug case */ | |
2293 | if (!alloc_debug_processing(s, page, freelist, addr)) | |
2294 | goto new_slab; /* Slab failed checks. Next slab needed */ | |
2295 | ||
2296 | deactivate_slab(s, page, get_freepointer(s, freelist)); | |
2297 | c->page = NULL; | |
2298 | c->freelist = NULL; | |
2299 | local_irq_restore(flags); | |
2300 | return freelist; | |
2301 | } | |
2302 | ||
2303 | /* | |
2304 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) | |
2305 | * have the fastpath folded into their functions. So no function call | |
2306 | * overhead for requests that can be satisfied on the fastpath. | |
2307 | * | |
2308 | * The fastpath works by first checking if the lockless freelist can be used. | |
2309 | * If not then __slab_alloc is called for slow processing. | |
2310 | * | |
2311 | * Otherwise we can simply pick the next object from the lockless free list. | |
2312 | */ | |
2313 | static __always_inline void *slab_alloc(struct kmem_cache *s, | |
2314 | gfp_t gfpflags, int node, unsigned long addr) | |
2315 | { | |
2316 | void **object; | |
2317 | struct kmem_cache_cpu *c; | |
2318 | struct page *page; | |
2319 | unsigned long tid; | |
2320 | ||
2321 | if (slab_pre_alloc_hook(s, gfpflags)) | |
2322 | return NULL; | |
2323 | ||
2324 | redo: | |
2325 | ||
2326 | /* | |
2327 | * Must read kmem_cache cpu data via this cpu ptr. Preemption is | |
2328 | * enabled. We may switch back and forth between cpus while | |
2329 | * reading from one cpu area. That does not matter as long | |
2330 | * as we end up on the original cpu again when doing the cmpxchg. | |
2331 | */ | |
2332 | c = __this_cpu_ptr(s->cpu_slab); | |
2333 | ||
2334 | /* | |
2335 | * The transaction ids are globally unique per cpu and per operation on | |
2336 | * a per cpu queue. Thus they can be guarantee that the cmpxchg_double | |
2337 | * occurs on the right processor and that there was no operation on the | |
2338 | * linked list in between. | |
2339 | */ | |
2340 | tid = c->tid; | |
2341 | barrier(); | |
2342 | ||
2343 | object = c->freelist; | |
2344 | page = c->page; | |
2345 | if (unlikely(!object || !node_match(page, node))) | |
2346 | ||
2347 | object = __slab_alloc(s, gfpflags, node, addr, c); | |
2348 | ||
2349 | else { | |
2350 | void *next_object = get_freepointer_safe(s, object); | |
2351 | ||
2352 | /* | |
2353 | * The cmpxchg will only match if there was no additional | |
2354 | * operation and if we are on the right processor. | |
2355 | * | |
2356 | * The cmpxchg does the following atomically (without lock semantics!) | |
2357 | * 1. Relocate first pointer to the current per cpu area. | |
2358 | * 2. Verify that tid and freelist have not been changed | |
2359 | * 3. If they were not changed replace tid and freelist | |
2360 | * | |
2361 | * Since this is without lock semantics the protection is only against | |
2362 | * code executing on this cpu *not* from access by other cpus. | |
2363 | */ | |
2364 | if (unlikely(!this_cpu_cmpxchg_double( | |
2365 | s->cpu_slab->freelist, s->cpu_slab->tid, | |
2366 | object, tid, | |
2367 | next_object, next_tid(tid)))) { | |
2368 | ||
2369 | note_cmpxchg_failure("slab_alloc", s, tid); | |
2370 | goto redo; | |
2371 | } | |
2372 | prefetch_freepointer(s, next_object); | |
2373 | stat(s, ALLOC_FASTPATH); | |
2374 | } | |
2375 | ||
2376 | if (unlikely(gfpflags & __GFP_ZERO) && object) | |
2377 | memset(object, 0, s->objsize); | |
2378 | ||
2379 | slab_post_alloc_hook(s, gfpflags, object); | |
2380 | ||
2381 | return object; | |
2382 | } | |
2383 | ||
2384 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
2385 | { | |
2386 | void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); | |
2387 | ||
2388 | trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); | |
2389 | ||
2390 | return ret; | |
2391 | } | |
2392 | EXPORT_SYMBOL(kmem_cache_alloc); | |
2393 | ||
2394 | #ifdef CONFIG_TRACING | |
2395 | void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) | |
2396 | { | |
2397 | void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); | |
2398 | trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); | |
2399 | return ret; | |
2400 | } | |
2401 | EXPORT_SYMBOL(kmem_cache_alloc_trace); | |
2402 | ||
2403 | void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) | |
2404 | { | |
2405 | void *ret = kmalloc_order(size, flags, order); | |
2406 | trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags); | |
2407 | return ret; | |
2408 | } | |
2409 | EXPORT_SYMBOL(kmalloc_order_trace); | |
2410 | #endif | |
2411 | ||
2412 | #ifdef CONFIG_NUMA | |
2413 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
2414 | { | |
2415 | void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); | |
2416 | ||
2417 | trace_kmem_cache_alloc_node(_RET_IP_, ret, | |
2418 | s->objsize, s->size, gfpflags, node); | |
2419 | ||
2420 | return ret; | |
2421 | } | |
2422 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
2423 | ||
2424 | #ifdef CONFIG_TRACING | |
2425 | void *kmem_cache_alloc_node_trace(struct kmem_cache *s, | |
2426 | gfp_t gfpflags, | |
2427 | int node, size_t size) | |
2428 | { | |
2429 | void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); | |
2430 | ||
2431 | trace_kmalloc_node(_RET_IP_, ret, | |
2432 | size, s->size, gfpflags, node); | |
2433 | return ret; | |
2434 | } | |
2435 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); | |
2436 | #endif | |
2437 | #endif | |
2438 | ||
2439 | /* | |
2440 | * Slow patch handling. This may still be called frequently since objects | |
2441 | * have a longer lifetime than the cpu slabs in most processing loads. | |
2442 | * | |
2443 | * So we still attempt to reduce cache line usage. Just take the slab | |
2444 | * lock and free the item. If there is no additional partial page | |
2445 | * handling required then we can return immediately. | |
2446 | */ | |
2447 | static void __slab_free(struct kmem_cache *s, struct page *page, | |
2448 | void *x, unsigned long addr) | |
2449 | { | |
2450 | void *prior; | |
2451 | void **object = (void *)x; | |
2452 | int was_frozen; | |
2453 | int inuse; | |
2454 | struct page new; | |
2455 | unsigned long counters; | |
2456 | struct kmem_cache_node *n = NULL; | |
2457 | unsigned long uninitialized_var(flags); | |
2458 | ||
2459 | stat(s, FREE_SLOWPATH); | |
2460 | ||
2461 | if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr)) | |
2462 | return; | |
2463 | ||
2464 | do { | |
2465 | prior = page->freelist; | |
2466 | counters = page->counters; | |
2467 | set_freepointer(s, object, prior); | |
2468 | new.counters = counters; | |
2469 | was_frozen = new.frozen; | |
2470 | new.inuse--; | |
2471 | if ((!new.inuse || !prior) && !was_frozen && !n) { | |
2472 | ||
2473 | if (!kmem_cache_debug(s) && !prior) | |
2474 | ||
2475 | /* | |
2476 | * Slab was on no list before and will be partially empty | |
2477 | * We can defer the list move and instead freeze it. | |
2478 | */ | |
2479 | new.frozen = 1; | |
2480 | ||
2481 | else { /* Needs to be taken off a list */ | |
2482 | ||
2483 | n = get_node(s, page_to_nid(page)); | |
2484 | /* | |
2485 | * Speculatively acquire the list_lock. | |
2486 | * If the cmpxchg does not succeed then we may | |
2487 | * drop the list_lock without any processing. | |
2488 | * | |
2489 | * Otherwise the list_lock will synchronize with | |
2490 | * other processors updating the list of slabs. | |
2491 | */ | |
2492 | spin_lock_irqsave(&n->list_lock, flags); | |
2493 | ||
2494 | } | |
2495 | } | |
2496 | inuse = new.inuse; | |
2497 | ||
2498 | } while (!cmpxchg_double_slab(s, page, | |
2499 | prior, counters, | |
2500 | object, new.counters, | |
2501 | "__slab_free")); | |
2502 | ||
2503 | if (likely(!n)) { | |
2504 | ||
2505 | /* | |
2506 | * If we just froze the page then put it onto the | |
2507 | * per cpu partial list. | |
2508 | */ | |
2509 | if (new.frozen && !was_frozen) { | |
2510 | put_cpu_partial(s, page, 1); | |
2511 | stat(s, CPU_PARTIAL_FREE); | |
2512 | } | |
2513 | /* | |
2514 | * The list lock was not taken therefore no list | |
2515 | * activity can be necessary. | |
2516 | */ | |
2517 | if (was_frozen) | |
2518 | stat(s, FREE_FROZEN); | |
2519 | return; | |
2520 | } | |
2521 | ||
2522 | /* | |
2523 | * was_frozen may have been set after we acquired the list_lock in | |
2524 | * an earlier loop. So we need to check it here again. | |
2525 | */ | |
2526 | if (was_frozen) | |
2527 | stat(s, FREE_FROZEN); | |
2528 | else { | |
2529 | if (unlikely(!inuse && n->nr_partial > s->min_partial)) | |
2530 | goto slab_empty; | |
2531 | ||
2532 | /* | |
2533 | * Objects left in the slab. If it was not on the partial list before | |
2534 | * then add it. | |
2535 | */ | |
2536 | if (unlikely(!prior)) { | |
2537 | remove_full(s, page); | |
2538 | add_partial(n, page, DEACTIVATE_TO_TAIL); | |
2539 | stat(s, FREE_ADD_PARTIAL); | |
2540 | } | |
2541 | } | |
2542 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2543 | return; | |
2544 | ||
2545 | slab_empty: | |
2546 | if (prior) { | |
2547 | /* | |
2548 | * Slab on the partial list. | |
2549 | */ | |
2550 | remove_partial(n, page); | |
2551 | stat(s, FREE_REMOVE_PARTIAL); | |
2552 | } else | |
2553 | /* Slab must be on the full list */ | |
2554 | remove_full(s, page); | |
2555 | ||
2556 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2557 | stat(s, FREE_SLAB); | |
2558 | discard_slab(s, page); | |
2559 | } | |
2560 | ||
2561 | /* | |
2562 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that | |
2563 | * can perform fastpath freeing without additional function calls. | |
2564 | * | |
2565 | * The fastpath is only possible if we are freeing to the current cpu slab | |
2566 | * of this processor. This typically the case if we have just allocated | |
2567 | * the item before. | |
2568 | * | |
2569 | * If fastpath is not possible then fall back to __slab_free where we deal | |
2570 | * with all sorts of special processing. | |
2571 | */ | |
2572 | static __always_inline void slab_free(struct kmem_cache *s, | |
2573 | struct page *page, void *x, unsigned long addr) | |
2574 | { | |
2575 | void **object = (void *)x; | |
2576 | struct kmem_cache_cpu *c; | |
2577 | unsigned long tid; | |
2578 | ||
2579 | slab_free_hook(s, x); | |
2580 | ||
2581 | redo: | |
2582 | /* | |
2583 | * Determine the currently cpus per cpu slab. | |
2584 | * The cpu may change afterward. However that does not matter since | |
2585 | * data is retrieved via this pointer. If we are on the same cpu | |
2586 | * during the cmpxchg then the free will succedd. | |
2587 | */ | |
2588 | c = __this_cpu_ptr(s->cpu_slab); | |
2589 | ||
2590 | tid = c->tid; | |
2591 | barrier(); | |
2592 | ||
2593 | if (likely(page == c->page)) { | |
2594 | set_freepointer(s, object, c->freelist); | |
2595 | ||
2596 | if (unlikely(!this_cpu_cmpxchg_double( | |
2597 | s->cpu_slab->freelist, s->cpu_slab->tid, | |
2598 | c->freelist, tid, | |
2599 | object, next_tid(tid)))) { | |
2600 | ||
2601 | note_cmpxchg_failure("slab_free", s, tid); | |
2602 | goto redo; | |
2603 | } | |
2604 | stat(s, FREE_FASTPATH); | |
2605 | } else | |
2606 | __slab_free(s, page, x, addr); | |
2607 | ||
2608 | } | |
2609 | ||
2610 | void kmem_cache_free(struct kmem_cache *s, void *x) | |
2611 | { | |
2612 | struct page *page; | |
2613 | ||
2614 | page = virt_to_head_page(x); | |
2615 | ||
2616 | slab_free(s, page, x, _RET_IP_); | |
2617 | ||
2618 | trace_kmem_cache_free(_RET_IP_, x); | |
2619 | } | |
2620 | EXPORT_SYMBOL(kmem_cache_free); | |
2621 | ||
2622 | /* | |
2623 | * Object placement in a slab is made very easy because we always start at | |
2624 | * offset 0. If we tune the size of the object to the alignment then we can | |
2625 | * get the required alignment by putting one properly sized object after | |
2626 | * another. | |
2627 | * | |
2628 | * Notice that the allocation order determines the sizes of the per cpu | |
2629 | * caches. Each processor has always one slab available for allocations. | |
2630 | * Increasing the allocation order reduces the number of times that slabs | |
2631 | * must be moved on and off the partial lists and is therefore a factor in | |
2632 | * locking overhead. | |
2633 | */ | |
2634 | ||
2635 | /* | |
2636 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
2637 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
2638 | * and increases the number of allocations possible without having to | |
2639 | * take the list_lock. | |
2640 | */ | |
2641 | static int slub_min_order; | |
2642 | static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; | |
2643 | static int slub_min_objects; | |
2644 | ||
2645 | /* | |
2646 | * Merge control. If this is set then no merging of slab caches will occur. | |
2647 | * (Could be removed. This was introduced to pacify the merge skeptics.) | |
2648 | */ | |
2649 | static int slub_nomerge; | |
2650 | ||
2651 | /* | |
2652 | * Calculate the order of allocation given an slab object size. | |
2653 | * | |
2654 | * The order of allocation has significant impact on performance and other | |
2655 | * system components. Generally order 0 allocations should be preferred since | |
2656 | * order 0 does not cause fragmentation in the page allocator. Larger objects | |
2657 | * be problematic to put into order 0 slabs because there may be too much | |
2658 | * unused space left. We go to a higher order if more than 1/16th of the slab | |
2659 | * would be wasted. | |
2660 | * | |
2661 | * In order to reach satisfactory performance we must ensure that a minimum | |
2662 | * number of objects is in one slab. Otherwise we may generate too much | |
2663 | * activity on the partial lists which requires taking the list_lock. This is | |
2664 | * less a concern for large slabs though which are rarely used. | |
2665 | * | |
2666 | * slub_max_order specifies the order where we begin to stop considering the | |
2667 | * number of objects in a slab as critical. If we reach slub_max_order then | |
2668 | * we try to keep the page order as low as possible. So we accept more waste | |
2669 | * of space in favor of a small page order. | |
2670 | * | |
2671 | * Higher order allocations also allow the placement of more objects in a | |
2672 | * slab and thereby reduce object handling overhead. If the user has | |
2673 | * requested a higher mininum order then we start with that one instead of | |
2674 | * the smallest order which will fit the object. | |
2675 | */ | |
2676 | static inline int slab_order(int size, int min_objects, | |
2677 | int max_order, int fract_leftover, int reserved) | |
2678 | { | |
2679 | int order; | |
2680 | int rem; | |
2681 | int min_order = slub_min_order; | |
2682 | ||
2683 | if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) | |
2684 | return get_order(size * MAX_OBJS_PER_PAGE) - 1; | |
2685 | ||
2686 | for (order = max(min_order, | |
2687 | fls(min_objects * size - 1) - PAGE_SHIFT); | |
2688 | order <= max_order; order++) { | |
2689 | ||
2690 | unsigned long slab_size = PAGE_SIZE << order; | |
2691 | ||
2692 | if (slab_size < min_objects * size + reserved) | |
2693 | continue; | |
2694 | ||
2695 | rem = (slab_size - reserved) % size; | |
2696 | ||
2697 | if (rem <= slab_size / fract_leftover) | |
2698 | break; | |
2699 | ||
2700 | } | |
2701 | ||
2702 | return order; | |
2703 | } | |
2704 | ||
2705 | static inline int calculate_order(int size, int reserved) | |
2706 | { | |
2707 | int order; | |
2708 | int min_objects; | |
2709 | int fraction; | |
2710 | int max_objects; | |
2711 | ||
2712 | /* | |
2713 | * Attempt to find best configuration for a slab. This | |
2714 | * works by first attempting to generate a layout with | |
2715 | * the best configuration and backing off gradually. | |
2716 | * | |
2717 | * First we reduce the acceptable waste in a slab. Then | |
2718 | * we reduce the minimum objects required in a slab. | |
2719 | */ | |
2720 | min_objects = slub_min_objects; | |
2721 | if (!min_objects) | |
2722 | min_objects = 4 * (fls(nr_cpu_ids) + 1); | |
2723 | max_objects = order_objects(slub_max_order, size, reserved); | |
2724 | min_objects = min(min_objects, max_objects); | |
2725 | ||
2726 | while (min_objects > 1) { | |
2727 | fraction = 16; | |
2728 | while (fraction >= 4) { | |
2729 | order = slab_order(size, min_objects, | |
2730 | slub_max_order, fraction, reserved); | |
2731 | if (order <= slub_max_order) | |
2732 | return order; | |
2733 | fraction /= 2; | |
2734 | } | |
2735 | min_objects--; | |
2736 | } | |
2737 | ||
2738 | /* | |
2739 | * We were unable to place multiple objects in a slab. Now | |
2740 | * lets see if we can place a single object there. | |
2741 | */ | |
2742 | order = slab_order(size, 1, slub_max_order, 1, reserved); | |
2743 | if (order <= slub_max_order) | |
2744 | return order; | |
2745 | ||
2746 | /* | |
2747 | * Doh this slab cannot be placed using slub_max_order. | |
2748 | */ | |
2749 | order = slab_order(size, 1, MAX_ORDER, 1, reserved); | |
2750 | if (order < MAX_ORDER) | |
2751 | return order; | |
2752 | return -ENOSYS; | |
2753 | } | |
2754 | ||
2755 | /* | |
2756 | * Figure out what the alignment of the objects will be. | |
2757 | */ | |
2758 | static unsigned long calculate_alignment(unsigned long flags, | |
2759 | unsigned long align, unsigned long size) | |
2760 | { | |
2761 | /* | |
2762 | * If the user wants hardware cache aligned objects then follow that | |
2763 | * suggestion if the object is sufficiently large. | |
2764 | * | |
2765 | * The hardware cache alignment cannot override the specified | |
2766 | * alignment though. If that is greater then use it. | |
2767 | */ | |
2768 | if (flags & SLAB_HWCACHE_ALIGN) { | |
2769 | unsigned long ralign = cache_line_size(); | |
2770 | while (size <= ralign / 2) | |
2771 | ralign /= 2; | |
2772 | align = max(align, ralign); | |
2773 | } | |
2774 | ||
2775 | if (align < ARCH_SLAB_MINALIGN) | |
2776 | align = ARCH_SLAB_MINALIGN; | |
2777 | ||
2778 | return ALIGN(align, sizeof(void *)); | |
2779 | } | |
2780 | ||
2781 | static void | |
2782 | init_kmem_cache_node(struct kmem_cache_node *n) | |
2783 | { | |
2784 | n->nr_partial = 0; | |
2785 | spin_lock_init(&n->list_lock); | |
2786 | INIT_LIST_HEAD(&n->partial); | |
2787 | #ifdef CONFIG_SLUB_DEBUG | |
2788 | atomic_long_set(&n->nr_slabs, 0); | |
2789 | atomic_long_set(&n->total_objects, 0); | |
2790 | INIT_LIST_HEAD(&n->full); | |
2791 | #endif | |
2792 | } | |
2793 | ||
2794 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) | |
2795 | { | |
2796 | BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < | |
2797 | SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu)); | |
2798 | ||
2799 | /* | |
2800 | * Must align to double word boundary for the double cmpxchg | |
2801 | * instructions to work; see __pcpu_double_call_return_bool(). | |
2802 | */ | |
2803 | s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), | |
2804 | 2 * sizeof(void *)); | |
2805 | ||
2806 | if (!s->cpu_slab) | |
2807 | return 0; | |
2808 | ||
2809 | init_kmem_cache_cpus(s); | |
2810 | ||
2811 | return 1; | |
2812 | } | |
2813 | ||
2814 | static struct kmem_cache *kmem_cache_node; | |
2815 | ||
2816 | /* | |
2817 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
2818 | * slab on the node for this slabcache. There are no concurrent accesses | |
2819 | * possible. | |
2820 | * | |
2821 | * Note that this function only works on the kmalloc_node_cache | |
2822 | * when allocating for the kmalloc_node_cache. This is used for bootstrapping | |
2823 | * memory on a fresh node that has no slab structures yet. | |
2824 | */ | |
2825 | static void early_kmem_cache_node_alloc(int node) | |
2826 | { | |
2827 | struct page *page; | |
2828 | struct kmem_cache_node *n; | |
2829 | ||
2830 | BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); | |
2831 | ||
2832 | page = new_slab(kmem_cache_node, GFP_NOWAIT, node); | |
2833 | ||
2834 | BUG_ON(!page); | |
2835 | if (page_to_nid(page) != node) { | |
2836 | printk(KERN_ERR "SLUB: Unable to allocate memory from " | |
2837 | "node %d\n", node); | |
2838 | printk(KERN_ERR "SLUB: Allocating a useless per node structure " | |
2839 | "in order to be able to continue\n"); | |
2840 | } | |
2841 | ||
2842 | n = page->freelist; | |
2843 | BUG_ON(!n); | |
2844 | page->freelist = get_freepointer(kmem_cache_node, n); | |
2845 | page->inuse = 1; | |
2846 | page->frozen = 0; | |
2847 | kmem_cache_node->node[node] = n; | |
2848 | #ifdef CONFIG_SLUB_DEBUG | |
2849 | init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); | |
2850 | init_tracking(kmem_cache_node, n); | |
2851 | #endif | |
2852 | init_kmem_cache_node(n); | |
2853 | inc_slabs_node(kmem_cache_node, node, page->objects); | |
2854 | ||
2855 | add_partial(n, page, DEACTIVATE_TO_HEAD); | |
2856 | } | |
2857 | ||
2858 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
2859 | { | |
2860 | int node; | |
2861 | ||
2862 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2863 | struct kmem_cache_node *n = s->node[node]; | |
2864 | ||
2865 | if (n) | |
2866 | kmem_cache_free(kmem_cache_node, n); | |
2867 | ||
2868 | s->node[node] = NULL; | |
2869 | } | |
2870 | } | |
2871 | ||
2872 | static int init_kmem_cache_nodes(struct kmem_cache *s) | |
2873 | { | |
2874 | int node; | |
2875 | ||
2876 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2877 | struct kmem_cache_node *n; | |
2878 | ||
2879 | if (slab_state == DOWN) { | |
2880 | early_kmem_cache_node_alloc(node); | |
2881 | continue; | |
2882 | } | |
2883 | n = kmem_cache_alloc_node(kmem_cache_node, | |
2884 | GFP_KERNEL, node); | |
2885 | ||
2886 | if (!n) { | |
2887 | free_kmem_cache_nodes(s); | |
2888 | return 0; | |
2889 | } | |
2890 | ||
2891 | s->node[node] = n; | |
2892 | init_kmem_cache_node(n); | |
2893 | } | |
2894 | return 1; | |
2895 | } | |
2896 | ||
2897 | static void set_min_partial(struct kmem_cache *s, unsigned long min) | |
2898 | { | |
2899 | if (min < MIN_PARTIAL) | |
2900 | min = MIN_PARTIAL; | |
2901 | else if (min > MAX_PARTIAL) | |
2902 | min = MAX_PARTIAL; | |
2903 | s->min_partial = min; | |
2904 | } | |
2905 | ||
2906 | /* | |
2907 | * calculate_sizes() determines the order and the distribution of data within | |
2908 | * a slab object. | |
2909 | */ | |
2910 | static int calculate_sizes(struct kmem_cache *s, int forced_order) | |
2911 | { | |
2912 | unsigned long flags = s->flags; | |
2913 | unsigned long size = s->objsize; | |
2914 | unsigned long align = s->align; | |
2915 | int order; | |
2916 | ||
2917 | /* | |
2918 | * Round up object size to the next word boundary. We can only | |
2919 | * place the free pointer at word boundaries and this determines | |
2920 | * the possible location of the free pointer. | |
2921 | */ | |
2922 | size = ALIGN(size, sizeof(void *)); | |
2923 | ||
2924 | #ifdef CONFIG_SLUB_DEBUG | |
2925 | /* | |
2926 | * Determine if we can poison the object itself. If the user of | |
2927 | * the slab may touch the object after free or before allocation | |
2928 | * then we should never poison the object itself. | |
2929 | */ | |
2930 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && | |
2931 | !s->ctor) | |
2932 | s->flags |= __OBJECT_POISON; | |
2933 | else | |
2934 | s->flags &= ~__OBJECT_POISON; | |
2935 | ||
2936 | ||
2937 | /* | |
2938 | * If we are Redzoning then check if there is some space between the | |
2939 | * end of the object and the free pointer. If not then add an | |
2940 | * additional word to have some bytes to store Redzone information. | |
2941 | */ | |
2942 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) | |
2943 | size += sizeof(void *); | |
2944 | #endif | |
2945 | ||
2946 | /* | |
2947 | * With that we have determined the number of bytes in actual use | |
2948 | * by the object. This is the potential offset to the free pointer. | |
2949 | */ | |
2950 | s->inuse = size; | |
2951 | ||
2952 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || | |
2953 | s->ctor)) { | |
2954 | /* | |
2955 | * Relocate free pointer after the object if it is not | |
2956 | * permitted to overwrite the first word of the object on | |
2957 | * kmem_cache_free. | |
2958 | * | |
2959 | * This is the case if we do RCU, have a constructor or | |
2960 | * destructor or are poisoning the objects. | |
2961 | */ | |
2962 | s->offset = size; | |
2963 | size += sizeof(void *); | |
2964 | } | |
2965 | ||
2966 | #ifdef CONFIG_SLUB_DEBUG | |
2967 | if (flags & SLAB_STORE_USER) | |
2968 | /* | |
2969 | * Need to store information about allocs and frees after | |
2970 | * the object. | |
2971 | */ | |
2972 | size += 2 * sizeof(struct track); | |
2973 | ||
2974 | if (flags & SLAB_RED_ZONE) | |
2975 | /* | |
2976 | * Add some empty padding so that we can catch | |
2977 | * overwrites from earlier objects rather than let | |
2978 | * tracking information or the free pointer be | |
2979 | * corrupted if a user writes before the start | |
2980 | * of the object. | |
2981 | */ | |
2982 | size += sizeof(void *); | |
2983 | #endif | |
2984 | ||
2985 | /* | |
2986 | * Determine the alignment based on various parameters that the | |
2987 | * user specified and the dynamic determination of cache line size | |
2988 | * on bootup. | |
2989 | */ | |
2990 | align = calculate_alignment(flags, align, s->objsize); | |
2991 | s->align = align; | |
2992 | ||
2993 | /* | |
2994 | * SLUB stores one object immediately after another beginning from | |
2995 | * offset 0. In order to align the objects we have to simply size | |
2996 | * each object to conform to the alignment. | |
2997 | */ | |
2998 | size = ALIGN(size, align); | |
2999 | s->size = size; | |
3000 | if (forced_order >= 0) | |
3001 | order = forced_order; | |
3002 | else | |
3003 | order = calculate_order(size, s->reserved); | |
3004 | ||
3005 | if (order < 0) | |
3006 | return 0; | |
3007 | ||
3008 | s->allocflags = 0; | |
3009 | if (order) | |
3010 | s->allocflags |= __GFP_COMP; | |
3011 | ||
3012 | if (s->flags & SLAB_CACHE_DMA) | |
3013 | s->allocflags |= SLUB_DMA; | |
3014 | ||
3015 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
3016 | s->allocflags |= __GFP_RECLAIMABLE; | |
3017 | ||
3018 | /* | |
3019 | * Determine the number of objects per slab | |
3020 | */ | |
3021 | s->oo = oo_make(order, size, s->reserved); | |
3022 | s->min = oo_make(get_order(size), size, s->reserved); | |
3023 | if (oo_objects(s->oo) > oo_objects(s->max)) | |
3024 | s->max = s->oo; | |
3025 | ||
3026 | return !!oo_objects(s->oo); | |
3027 | ||
3028 | } | |
3029 | ||
3030 | static int kmem_cache_open(struct kmem_cache *s, | |
3031 | const char *name, size_t size, | |
3032 | size_t align, unsigned long flags, | |
3033 | void (*ctor)(void *)) | |
3034 | { | |
3035 | memset(s, 0, kmem_size); | |
3036 | s->name = name; | |
3037 | s->ctor = ctor; | |
3038 | s->objsize = size; | |
3039 | s->align = align; | |
3040 | s->flags = kmem_cache_flags(size, flags, name, ctor); | |
3041 | s->reserved = 0; | |
3042 | ||
3043 | if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) | |
3044 | s->reserved = sizeof(struct rcu_head); | |
3045 | ||
3046 | if (!calculate_sizes(s, -1)) | |
3047 | goto error; | |
3048 | if (disable_higher_order_debug) { | |
3049 | /* | |
3050 | * Disable debugging flags that store metadata if the min slab | |
3051 | * order increased. | |
3052 | */ | |
3053 | if (get_order(s->size) > get_order(s->objsize)) { | |
3054 | s->flags &= ~DEBUG_METADATA_FLAGS; | |
3055 | s->offset = 0; | |
3056 | if (!calculate_sizes(s, -1)) | |
3057 | goto error; | |
3058 | } | |
3059 | } | |
3060 | ||
3061 | #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ | |
3062 | defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) | |
3063 | if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0) | |
3064 | /* Enable fast mode */ | |
3065 | s->flags |= __CMPXCHG_DOUBLE; | |
3066 | #endif | |
3067 | ||
3068 | /* | |
3069 | * The larger the object size is, the more pages we want on the partial | |
3070 | * list to avoid pounding the page allocator excessively. | |
3071 | */ | |
3072 | set_min_partial(s, ilog2(s->size) / 2); | |
3073 | ||
3074 | /* | |
3075 | * cpu_partial determined the maximum number of objects kept in the | |
3076 | * per cpu partial lists of a processor. | |
3077 | * | |
3078 | * Per cpu partial lists mainly contain slabs that just have one | |
3079 | * object freed. If they are used for allocation then they can be | |
3080 | * filled up again with minimal effort. The slab will never hit the | |
3081 | * per node partial lists and therefore no locking will be required. | |
3082 | * | |
3083 | * This setting also determines | |
3084 | * | |
3085 | * A) The number of objects from per cpu partial slabs dumped to the | |
3086 | * per node list when we reach the limit. | |
3087 | * B) The number of objects in cpu partial slabs to extract from the | |
3088 | * per node list when we run out of per cpu objects. We only fetch 50% | |
3089 | * to keep some capacity around for frees. | |
3090 | */ | |
3091 | if (kmem_cache_debug(s)) | |
3092 | s->cpu_partial = 0; | |
3093 | else if (s->size >= PAGE_SIZE) | |
3094 | s->cpu_partial = 2; | |
3095 | else if (s->size >= 1024) | |
3096 | s->cpu_partial = 6; | |
3097 | else if (s->size >= 256) | |
3098 | s->cpu_partial = 13; | |
3099 | else | |
3100 | s->cpu_partial = 30; | |
3101 | ||
3102 | s->refcount = 1; | |
3103 | #ifdef CONFIG_NUMA | |
3104 | s->remote_node_defrag_ratio = 1000; | |
3105 | #endif | |
3106 | if (!init_kmem_cache_nodes(s)) | |
3107 | goto error; | |
3108 | ||
3109 | if (alloc_kmem_cache_cpus(s)) | |
3110 | return 1; | |
3111 | ||
3112 | free_kmem_cache_nodes(s); | |
3113 | error: | |
3114 | if (flags & SLAB_PANIC) | |
3115 | panic("Cannot create slab %s size=%lu realsize=%u " | |
3116 | "order=%u offset=%u flags=%lx\n", | |
3117 | s->name, (unsigned long)size, s->size, oo_order(s->oo), | |
3118 | s->offset, flags); | |
3119 | return 0; | |
3120 | } | |
3121 | ||
3122 | /* | |
3123 | * Determine the size of a slab object | |
3124 | */ | |
3125 | unsigned int kmem_cache_size(struct kmem_cache *s) | |
3126 | { | |
3127 | return s->objsize; | |
3128 | } | |
3129 | EXPORT_SYMBOL(kmem_cache_size); | |
3130 | ||
3131 | static void list_slab_objects(struct kmem_cache *s, struct page *page, | |
3132 | const char *text) | |
3133 | { | |
3134 | #ifdef CONFIG_SLUB_DEBUG | |
3135 | void *addr = page_address(page); | |
3136 | void *p; | |
3137 | unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * | |
3138 | sizeof(long), GFP_ATOMIC); | |
3139 | if (!map) | |
3140 | return; | |
3141 | slab_err(s, page, "%s", text); | |
3142 | slab_lock(page); | |
3143 | ||
3144 | get_map(s, page, map); | |
3145 | for_each_object(p, s, addr, page->objects) { | |
3146 | ||
3147 | if (!test_bit(slab_index(p, s, addr), map)) { | |
3148 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", | |
3149 | p, p - addr); | |
3150 | print_tracking(s, p); | |
3151 | } | |
3152 | } | |
3153 | slab_unlock(page); | |
3154 | kfree(map); | |
3155 | #endif | |
3156 | } | |
3157 | ||
3158 | /* | |
3159 | * Attempt to free all partial slabs on a node. | |
3160 | * This is called from kmem_cache_close(). We must be the last thread | |
3161 | * using the cache and therefore we do not need to lock anymore. | |
3162 | */ | |
3163 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) | |
3164 | { | |
3165 | struct page *page, *h; | |
3166 | ||
3167 | list_for_each_entry_safe(page, h, &n->partial, lru) { | |
3168 | if (!page->inuse) { | |
3169 | remove_partial(n, page); | |
3170 | discard_slab(s, page); | |
3171 | } else { | |
3172 | list_slab_objects(s, page, | |
3173 | "Objects remaining on kmem_cache_close()"); | |
3174 | } | |
3175 | } | |
3176 | } | |
3177 | ||
3178 | /* | |
3179 | * Release all resources used by a slab cache. | |
3180 | */ | |
3181 | static inline int kmem_cache_close(struct kmem_cache *s) | |
3182 | { | |
3183 | int node; | |
3184 | ||
3185 | flush_all(s); | |
3186 | free_percpu(s->cpu_slab); | |
3187 | /* Attempt to free all objects */ | |
3188 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3189 | struct kmem_cache_node *n = get_node(s, node); | |
3190 | ||
3191 | free_partial(s, n); | |
3192 | if (n->nr_partial || slabs_node(s, node)) | |
3193 | return 1; | |
3194 | } | |
3195 | free_kmem_cache_nodes(s); | |
3196 | return 0; | |
3197 | } | |
3198 | ||
3199 | /* | |
3200 | * Close a cache and release the kmem_cache structure | |
3201 | * (must be used for caches created using kmem_cache_create) | |
3202 | */ | |
3203 | void kmem_cache_destroy(struct kmem_cache *s) | |
3204 | { | |
3205 | down_write(&slub_lock); | |
3206 | s->refcount--; | |
3207 | if (!s->refcount) { | |
3208 | list_del(&s->list); | |
3209 | up_write(&slub_lock); | |
3210 | if (kmem_cache_close(s)) { | |
3211 | printk(KERN_ERR "SLUB %s: %s called for cache that " | |
3212 | "still has objects.\n", s->name, __func__); | |
3213 | dump_stack(); | |
3214 | } | |
3215 | if (s->flags & SLAB_DESTROY_BY_RCU) | |
3216 | rcu_barrier(); | |
3217 | sysfs_slab_remove(s); | |
3218 | } else | |
3219 | up_write(&slub_lock); | |
3220 | } | |
3221 | EXPORT_SYMBOL(kmem_cache_destroy); | |
3222 | ||
3223 | /******************************************************************** | |
3224 | * Kmalloc subsystem | |
3225 | *******************************************************************/ | |
3226 | ||
3227 | struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT]; | |
3228 | EXPORT_SYMBOL(kmalloc_caches); | |
3229 | ||
3230 | static struct kmem_cache *kmem_cache; | |
3231 | ||
3232 | #ifdef CONFIG_ZONE_DMA | |
3233 | static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT]; | |
3234 | #endif | |
3235 | ||
3236 | static int __init setup_slub_min_order(char *str) | |
3237 | { | |
3238 | get_option(&str, &slub_min_order); | |
3239 | ||
3240 | return 1; | |
3241 | } | |
3242 | ||
3243 | __setup("slub_min_order=", setup_slub_min_order); | |
3244 | ||
3245 | static int __init setup_slub_max_order(char *str) | |
3246 | { | |
3247 | get_option(&str, &slub_max_order); | |
3248 | slub_max_order = min(slub_max_order, MAX_ORDER - 1); | |
3249 | ||
3250 | return 1; | |
3251 | } | |
3252 | ||
3253 | __setup("slub_max_order=", setup_slub_max_order); | |
3254 | ||
3255 | static int __init setup_slub_min_objects(char *str) | |
3256 | { | |
3257 | get_option(&str, &slub_min_objects); | |
3258 | ||
3259 | return 1; | |
3260 | } | |
3261 | ||
3262 | __setup("slub_min_objects=", setup_slub_min_objects); | |
3263 | ||
3264 | static int __init setup_slub_nomerge(char *str) | |
3265 | { | |
3266 | slub_nomerge = 1; | |
3267 | return 1; | |
3268 | } | |
3269 | ||
3270 | __setup("slub_nomerge", setup_slub_nomerge); | |
3271 | ||
3272 | static struct kmem_cache *__init create_kmalloc_cache(const char *name, | |
3273 | int size, unsigned int flags) | |
3274 | { | |
3275 | struct kmem_cache *s; | |
3276 | ||
3277 | s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); | |
3278 | ||
3279 | /* | |
3280 | * This function is called with IRQs disabled during early-boot on | |
3281 | * single CPU so there's no need to take slub_lock here. | |
3282 | */ | |
3283 | if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN, | |
3284 | flags, NULL)) | |
3285 | goto panic; | |
3286 | ||
3287 | list_add(&s->list, &slab_caches); | |
3288 | return s; | |
3289 | ||
3290 | panic: | |
3291 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); | |
3292 | return NULL; | |
3293 | } | |
3294 | ||
3295 | /* | |
3296 | * Conversion table for small slabs sizes / 8 to the index in the | |
3297 | * kmalloc array. This is necessary for slabs < 192 since we have non power | |
3298 | * of two cache sizes there. The size of larger slabs can be determined using | |
3299 | * fls. | |
3300 | */ | |
3301 | static s8 size_index[24] = { | |
3302 | 3, /* 8 */ | |
3303 | 4, /* 16 */ | |
3304 | 5, /* 24 */ | |
3305 | 5, /* 32 */ | |
3306 | 6, /* 40 */ | |
3307 | 6, /* 48 */ | |
3308 | 6, /* 56 */ | |
3309 | 6, /* 64 */ | |
3310 | 1, /* 72 */ | |
3311 | 1, /* 80 */ | |
3312 | 1, /* 88 */ | |
3313 | 1, /* 96 */ | |
3314 | 7, /* 104 */ | |
3315 | 7, /* 112 */ | |
3316 | 7, /* 120 */ | |
3317 | 7, /* 128 */ | |
3318 | 2, /* 136 */ | |
3319 | 2, /* 144 */ | |
3320 | 2, /* 152 */ | |
3321 | 2, /* 160 */ | |
3322 | 2, /* 168 */ | |
3323 | 2, /* 176 */ | |
3324 | 2, /* 184 */ | |
3325 | 2 /* 192 */ | |
3326 | }; | |
3327 | ||
3328 | static inline int size_index_elem(size_t bytes) | |
3329 | { | |
3330 | return (bytes - 1) / 8; | |
3331 | } | |
3332 | ||
3333 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) | |
3334 | { | |
3335 | int index; | |
3336 | ||
3337 | if (size <= 192) { | |
3338 | if (!size) | |
3339 | return ZERO_SIZE_PTR; | |
3340 | ||
3341 | index = size_index[size_index_elem(size)]; | |
3342 | } else | |
3343 | index = fls(size - 1); | |
3344 | ||
3345 | #ifdef CONFIG_ZONE_DMA | |
3346 | if (unlikely((flags & SLUB_DMA))) | |
3347 | return kmalloc_dma_caches[index]; | |
3348 | ||
3349 | #endif | |
3350 | return kmalloc_caches[index]; | |
3351 | } | |
3352 | ||
3353 | void *__kmalloc(size_t size, gfp_t flags) | |
3354 | { | |
3355 | struct kmem_cache *s; | |
3356 | void *ret; | |
3357 | ||
3358 | if (unlikely(size > SLUB_MAX_SIZE)) | |
3359 | return kmalloc_large(size, flags); | |
3360 | ||
3361 | s = get_slab(size, flags); | |
3362 | ||
3363 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3364 | return s; | |
3365 | ||
3366 | ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); | |
3367 | ||
3368 | trace_kmalloc(_RET_IP_, ret, size, s->size, flags); | |
3369 | ||
3370 | return ret; | |
3371 | } | |
3372 | EXPORT_SYMBOL(__kmalloc); | |
3373 | ||
3374 | #ifdef CONFIG_NUMA | |
3375 | static void *kmalloc_large_node(size_t size, gfp_t flags, int node) | |
3376 | { | |
3377 | struct page *page; | |
3378 | void *ptr = NULL; | |
3379 | ||
3380 | flags |= __GFP_COMP | __GFP_NOTRACK; | |
3381 | page = alloc_pages_node(node, flags, get_order(size)); | |
3382 | if (page) | |
3383 | ptr = page_address(page); | |
3384 | ||
3385 | kmemleak_alloc(ptr, size, 1, flags); | |
3386 | return ptr; | |
3387 | } | |
3388 | ||
3389 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3390 | { | |
3391 | struct kmem_cache *s; | |
3392 | void *ret; | |
3393 | ||
3394 | if (unlikely(size > SLUB_MAX_SIZE)) { | |
3395 | ret = kmalloc_large_node(size, flags, node); | |
3396 | ||
3397 | trace_kmalloc_node(_RET_IP_, ret, | |
3398 | size, PAGE_SIZE << get_order(size), | |
3399 | flags, node); | |
3400 | ||
3401 | return ret; | |
3402 | } | |
3403 | ||
3404 | s = get_slab(size, flags); | |
3405 | ||
3406 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3407 | return s; | |
3408 | ||
3409 | ret = slab_alloc(s, flags, node, _RET_IP_); | |
3410 | ||
3411 | trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); | |
3412 | ||
3413 | return ret; | |
3414 | } | |
3415 | EXPORT_SYMBOL(__kmalloc_node); | |
3416 | #endif | |
3417 | ||
3418 | size_t ksize(const void *object) | |
3419 | { | |
3420 | struct page *page; | |
3421 | ||
3422 | if (unlikely(object == ZERO_SIZE_PTR)) | |
3423 | return 0; | |
3424 | ||
3425 | page = virt_to_head_page(object); | |
3426 | ||
3427 | if (unlikely(!PageSlab(page))) { | |
3428 | WARN_ON(!PageCompound(page)); | |
3429 | return PAGE_SIZE << compound_order(page); | |
3430 | } | |
3431 | ||
3432 | return slab_ksize(page->slab); | |
3433 | } | |
3434 | EXPORT_SYMBOL(ksize); | |
3435 | ||
3436 | #ifdef CONFIG_SLUB_DEBUG | |
3437 | bool verify_mem_not_deleted(const void *x) | |
3438 | { | |
3439 | struct page *page; | |
3440 | void *object = (void *)x; | |
3441 | unsigned long flags; | |
3442 | bool rv; | |
3443 | ||
3444 | if (unlikely(ZERO_OR_NULL_PTR(x))) | |
3445 | return false; | |
3446 | ||
3447 | local_irq_save(flags); | |
3448 | ||
3449 | page = virt_to_head_page(x); | |
3450 | if (unlikely(!PageSlab(page))) { | |
3451 | /* maybe it was from stack? */ | |
3452 | rv = true; | |
3453 | goto out_unlock; | |
3454 | } | |
3455 | ||
3456 | slab_lock(page); | |
3457 | if (on_freelist(page->slab, page, object)) { | |
3458 | object_err(page->slab, page, object, "Object is on free-list"); | |
3459 | rv = false; | |
3460 | } else { | |
3461 | rv = true; | |
3462 | } | |
3463 | slab_unlock(page); | |
3464 | ||
3465 | out_unlock: | |
3466 | local_irq_restore(flags); | |
3467 | return rv; | |
3468 | } | |
3469 | EXPORT_SYMBOL(verify_mem_not_deleted); | |
3470 | #endif | |
3471 | ||
3472 | void kfree(const void *x) | |
3473 | { | |
3474 | struct page *page; | |
3475 | void *object = (void *)x; | |
3476 | ||
3477 | trace_kfree(_RET_IP_, x); | |
3478 | ||
3479 | if (unlikely(ZERO_OR_NULL_PTR(x))) | |
3480 | return; | |
3481 | ||
3482 | page = virt_to_head_page(x); | |
3483 | if (unlikely(!PageSlab(page))) { | |
3484 | BUG_ON(!PageCompound(page)); | |
3485 | kmemleak_free(x); | |
3486 | put_page(page); | |
3487 | return; | |
3488 | } | |
3489 | slab_free(page->slab, page, object, _RET_IP_); | |
3490 | } | |
3491 | EXPORT_SYMBOL(kfree); | |
3492 | ||
3493 | /* | |
3494 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts | |
3495 | * the remaining slabs by the number of items in use. The slabs with the | |
3496 | * most items in use come first. New allocations will then fill those up | |
3497 | * and thus they can be removed from the partial lists. | |
3498 | * | |
3499 | * The slabs with the least items are placed last. This results in them | |
3500 | * being allocated from last increasing the chance that the last objects | |
3501 | * are freed in them. | |
3502 | */ | |
3503 | int kmem_cache_shrink(struct kmem_cache *s) | |
3504 | { | |
3505 | int node; | |
3506 | int i; | |
3507 | struct kmem_cache_node *n; | |
3508 | struct page *page; | |
3509 | struct page *t; | |
3510 | int objects = oo_objects(s->max); | |
3511 | struct list_head *slabs_by_inuse = | |
3512 | kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); | |
3513 | unsigned long flags; | |
3514 | ||
3515 | if (!slabs_by_inuse) | |
3516 | return -ENOMEM; | |
3517 | ||
3518 | flush_all(s); | |
3519 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3520 | n = get_node(s, node); | |
3521 | ||
3522 | if (!n->nr_partial) | |
3523 | continue; | |
3524 | ||
3525 | for (i = 0; i < objects; i++) | |
3526 | INIT_LIST_HEAD(slabs_by_inuse + i); | |
3527 | ||
3528 | spin_lock_irqsave(&n->list_lock, flags); | |
3529 | ||
3530 | /* | |
3531 | * Build lists indexed by the items in use in each slab. | |
3532 | * | |
3533 | * Note that concurrent frees may occur while we hold the | |
3534 | * list_lock. page->inuse here is the upper limit. | |
3535 | */ | |
3536 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
3537 | list_move(&page->lru, slabs_by_inuse + page->inuse); | |
3538 | if (!page->inuse) | |
3539 | n->nr_partial--; | |
3540 | } | |
3541 | ||
3542 | /* | |
3543 | * Rebuild the partial list with the slabs filled up most | |
3544 | * first and the least used slabs at the end. | |
3545 | */ | |
3546 | for (i = objects - 1; i > 0; i--) | |
3547 | list_splice(slabs_by_inuse + i, n->partial.prev); | |
3548 | ||
3549 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3550 | ||
3551 | /* Release empty slabs */ | |
3552 | list_for_each_entry_safe(page, t, slabs_by_inuse, lru) | |
3553 | discard_slab(s, page); | |
3554 | } | |
3555 | ||
3556 | kfree(slabs_by_inuse); | |
3557 | return 0; | |
3558 | } | |
3559 | EXPORT_SYMBOL(kmem_cache_shrink); | |
3560 | ||
3561 | #if defined(CONFIG_MEMORY_HOTPLUG) | |
3562 | static int slab_mem_going_offline_callback(void *arg) | |
3563 | { | |
3564 | struct kmem_cache *s; | |
3565 | ||
3566 | down_read(&slub_lock); | |
3567 | list_for_each_entry(s, &slab_caches, list) | |
3568 | kmem_cache_shrink(s); | |
3569 | up_read(&slub_lock); | |
3570 | ||
3571 | return 0; | |
3572 | } | |
3573 | ||
3574 | static void slab_mem_offline_callback(void *arg) | |
3575 | { | |
3576 | struct kmem_cache_node *n; | |
3577 | struct kmem_cache *s; | |
3578 | struct memory_notify *marg = arg; | |
3579 | int offline_node; | |
3580 | ||
3581 | offline_node = marg->status_change_nid; | |
3582 | ||
3583 | /* | |
3584 | * If the node still has available memory. we need kmem_cache_node | |
3585 | * for it yet. | |
3586 | */ | |
3587 | if (offline_node < 0) | |
3588 | return; | |
3589 | ||
3590 | down_read(&slub_lock); | |
3591 | list_for_each_entry(s, &slab_caches, list) { | |
3592 | n = get_node(s, offline_node); | |
3593 | if (n) { | |
3594 | /* | |
3595 | * if n->nr_slabs > 0, slabs still exist on the node | |
3596 | * that is going down. We were unable to free them, | |
3597 | * and offline_pages() function shouldn't call this | |
3598 | * callback. So, we must fail. | |
3599 | */ | |
3600 | BUG_ON(slabs_node(s, offline_node)); | |
3601 | ||
3602 | s->node[offline_node] = NULL; | |
3603 | kmem_cache_free(kmem_cache_node, n); | |
3604 | } | |
3605 | } | |
3606 | up_read(&slub_lock); | |
3607 | } | |
3608 | ||
3609 | static int slab_mem_going_online_callback(void *arg) | |
3610 | { | |
3611 | struct kmem_cache_node *n; | |
3612 | struct kmem_cache *s; | |
3613 | struct memory_notify *marg = arg; | |
3614 | int nid = marg->status_change_nid; | |
3615 | int ret = 0; | |
3616 | ||
3617 | /* | |
3618 | * If the node's memory is already available, then kmem_cache_node is | |
3619 | * already created. Nothing to do. | |
3620 | */ | |
3621 | if (nid < 0) | |
3622 | return 0; | |
3623 | ||
3624 | /* | |
3625 | * We are bringing a node online. No memory is available yet. We must | |
3626 | * allocate a kmem_cache_node structure in order to bring the node | |
3627 | * online. | |
3628 | */ | |
3629 | down_read(&slub_lock); | |
3630 | list_for_each_entry(s, &slab_caches, list) { | |
3631 | /* | |
3632 | * XXX: kmem_cache_alloc_node will fallback to other nodes | |
3633 | * since memory is not yet available from the node that | |
3634 | * is brought up. | |
3635 | */ | |
3636 | n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); | |
3637 | if (!n) { | |
3638 | ret = -ENOMEM; | |
3639 | goto out; | |
3640 | } | |
3641 | init_kmem_cache_node(n); | |
3642 | s->node[nid] = n; | |
3643 | } | |
3644 | out: | |
3645 | up_read(&slub_lock); | |
3646 | return ret; | |
3647 | } | |
3648 | ||
3649 | static int slab_memory_callback(struct notifier_block *self, | |
3650 | unsigned long action, void *arg) | |
3651 | { | |
3652 | int ret = 0; | |
3653 | ||
3654 | switch (action) { | |
3655 | case MEM_GOING_ONLINE: | |
3656 | ret = slab_mem_going_online_callback(arg); | |
3657 | break; | |
3658 | case MEM_GOING_OFFLINE: | |
3659 | ret = slab_mem_going_offline_callback(arg); | |
3660 | break; | |
3661 | case MEM_OFFLINE: | |
3662 | case MEM_CANCEL_ONLINE: | |
3663 | slab_mem_offline_callback(arg); | |
3664 | break; | |
3665 | case MEM_ONLINE: | |
3666 | case MEM_CANCEL_OFFLINE: | |
3667 | break; | |
3668 | } | |
3669 | if (ret) | |
3670 | ret = notifier_from_errno(ret); | |
3671 | else | |
3672 | ret = NOTIFY_OK; | |
3673 | return ret; | |
3674 | } | |
3675 | ||
3676 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
3677 | ||
3678 | /******************************************************************** | |
3679 | * Basic setup of slabs | |
3680 | *******************************************************************/ | |
3681 | ||
3682 | /* | |
3683 | * Used for early kmem_cache structures that were allocated using | |
3684 | * the page allocator | |
3685 | */ | |
3686 | ||
3687 | static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s) | |
3688 | { | |
3689 | int node; | |
3690 | ||
3691 | list_add(&s->list, &slab_caches); | |
3692 | s->refcount = -1; | |
3693 | ||
3694 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3695 | struct kmem_cache_node *n = get_node(s, node); | |
3696 | struct page *p; | |
3697 | ||
3698 | if (n) { | |
3699 | list_for_each_entry(p, &n->partial, lru) | |
3700 | p->slab = s; | |
3701 | ||
3702 | #ifdef CONFIG_SLUB_DEBUG | |
3703 | list_for_each_entry(p, &n->full, lru) | |
3704 | p->slab = s; | |
3705 | #endif | |
3706 | } | |
3707 | } | |
3708 | } | |
3709 | ||
3710 | void __init kmem_cache_init(void) | |
3711 | { | |
3712 | int i; | |
3713 | int caches = 0; | |
3714 | struct kmem_cache *temp_kmem_cache; | |
3715 | int order; | |
3716 | struct kmem_cache *temp_kmem_cache_node; | |
3717 | unsigned long kmalloc_size; | |
3718 | ||
3719 | if (debug_guardpage_minorder()) | |
3720 | slub_max_order = 0; | |
3721 | ||
3722 | kmem_size = offsetof(struct kmem_cache, node) + | |
3723 | nr_node_ids * sizeof(struct kmem_cache_node *); | |
3724 | ||
3725 | /* Allocate two kmem_caches from the page allocator */ | |
3726 | kmalloc_size = ALIGN(kmem_size, cache_line_size()); | |
3727 | order = get_order(2 * kmalloc_size); | |
3728 | kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order); | |
3729 | ||
3730 | /* | |
3731 | * Must first have the slab cache available for the allocations of the | |
3732 | * struct kmem_cache_node's. There is special bootstrap code in | |
3733 | * kmem_cache_open for slab_state == DOWN. | |
3734 | */ | |
3735 | kmem_cache_node = (void *)kmem_cache + kmalloc_size; | |
3736 | ||
3737 | kmem_cache_open(kmem_cache_node, "kmem_cache_node", | |
3738 | sizeof(struct kmem_cache_node), | |
3739 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); | |
3740 | ||
3741 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
3742 | ||
3743 | /* Able to allocate the per node structures */ | |
3744 | slab_state = PARTIAL; | |
3745 | ||
3746 | temp_kmem_cache = kmem_cache; | |
3747 | kmem_cache_open(kmem_cache, "kmem_cache", kmem_size, | |
3748 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); | |
3749 | kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); | |
3750 | memcpy(kmem_cache, temp_kmem_cache, kmem_size); | |
3751 | ||
3752 | /* | |
3753 | * Allocate kmem_cache_node properly from the kmem_cache slab. | |
3754 | * kmem_cache_node is separately allocated so no need to | |
3755 | * update any list pointers. | |
3756 | */ | |
3757 | temp_kmem_cache_node = kmem_cache_node; | |
3758 | ||
3759 | kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); | |
3760 | memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size); | |
3761 | ||
3762 | kmem_cache_bootstrap_fixup(kmem_cache_node); | |
3763 | ||
3764 | caches++; | |
3765 | kmem_cache_bootstrap_fixup(kmem_cache); | |
3766 | caches++; | |
3767 | /* Free temporary boot structure */ | |
3768 | free_pages((unsigned long)temp_kmem_cache, order); | |
3769 | ||
3770 | /* Now we can use the kmem_cache to allocate kmalloc slabs */ | |
3771 | ||
3772 | /* | |
3773 | * Patch up the size_index table if we have strange large alignment | |
3774 | * requirements for the kmalloc array. This is only the case for | |
3775 | * MIPS it seems. The standard arches will not generate any code here. | |
3776 | * | |
3777 | * Largest permitted alignment is 256 bytes due to the way we | |
3778 | * handle the index determination for the smaller caches. | |
3779 | * | |
3780 | * Make sure that nothing crazy happens if someone starts tinkering | |
3781 | * around with ARCH_KMALLOC_MINALIGN | |
3782 | */ | |
3783 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || | |
3784 | (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); | |
3785 | ||
3786 | for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { | |
3787 | int elem = size_index_elem(i); | |
3788 | if (elem >= ARRAY_SIZE(size_index)) | |
3789 | break; | |
3790 | size_index[elem] = KMALLOC_SHIFT_LOW; | |
3791 | } | |
3792 | ||
3793 | if (KMALLOC_MIN_SIZE == 64) { | |
3794 | /* | |
3795 | * The 96 byte size cache is not used if the alignment | |
3796 | * is 64 byte. | |
3797 | */ | |
3798 | for (i = 64 + 8; i <= 96; i += 8) | |
3799 | size_index[size_index_elem(i)] = 7; | |
3800 | } else if (KMALLOC_MIN_SIZE == 128) { | |
3801 | /* | |
3802 | * The 192 byte sized cache is not used if the alignment | |
3803 | * is 128 byte. Redirect kmalloc to use the 256 byte cache | |
3804 | * instead. | |
3805 | */ | |
3806 | for (i = 128 + 8; i <= 192; i += 8) | |
3807 | size_index[size_index_elem(i)] = 8; | |
3808 | } | |
3809 | ||
3810 | /* Caches that are not of the two-to-the-power-of size */ | |
3811 | if (KMALLOC_MIN_SIZE <= 32) { | |
3812 | kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0); | |
3813 | caches++; | |
3814 | } | |
3815 | ||
3816 | if (KMALLOC_MIN_SIZE <= 64) { | |
3817 | kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0); | |
3818 | caches++; | |
3819 | } | |
3820 | ||
3821 | for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { | |
3822 | kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0); | |
3823 | caches++; | |
3824 | } | |
3825 | ||
3826 | slab_state = UP; | |
3827 | ||
3828 | /* Provide the correct kmalloc names now that the caches are up */ | |
3829 | if (KMALLOC_MIN_SIZE <= 32) { | |
3830 | kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT); | |
3831 | BUG_ON(!kmalloc_caches[1]->name); | |
3832 | } | |
3833 | ||
3834 | if (KMALLOC_MIN_SIZE <= 64) { | |
3835 | kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT); | |
3836 | BUG_ON(!kmalloc_caches[2]->name); | |
3837 | } | |
3838 | ||
3839 | for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { | |
3840 | char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); | |
3841 | ||
3842 | BUG_ON(!s); | |
3843 | kmalloc_caches[i]->name = s; | |
3844 | } | |
3845 | ||
3846 | #ifdef CONFIG_SMP | |
3847 | register_cpu_notifier(&slab_notifier); | |
3848 | #endif | |
3849 | ||
3850 | #ifdef CONFIG_ZONE_DMA | |
3851 | for (i = 0; i < SLUB_PAGE_SHIFT; i++) { | |
3852 | struct kmem_cache *s = kmalloc_caches[i]; | |
3853 | ||
3854 | if (s && s->size) { | |
3855 | char *name = kasprintf(GFP_NOWAIT, | |
3856 | "dma-kmalloc-%d", s->objsize); | |
3857 | ||
3858 | BUG_ON(!name); | |
3859 | kmalloc_dma_caches[i] = create_kmalloc_cache(name, | |
3860 | s->objsize, SLAB_CACHE_DMA); | |
3861 | } | |
3862 | } | |
3863 | #endif | |
3864 | printk(KERN_INFO | |
3865 | "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," | |
3866 | " CPUs=%d, Nodes=%d\n", | |
3867 | caches, cache_line_size(), | |
3868 | slub_min_order, slub_max_order, slub_min_objects, | |
3869 | nr_cpu_ids, nr_node_ids); | |
3870 | } | |
3871 | ||
3872 | void __init kmem_cache_init_late(void) | |
3873 | { | |
3874 | } | |
3875 | ||
3876 | /* | |
3877 | * Find a mergeable slab cache | |
3878 | */ | |
3879 | static int slab_unmergeable(struct kmem_cache *s) | |
3880 | { | |
3881 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) | |
3882 | return 1; | |
3883 | ||
3884 | if (s->ctor) | |
3885 | return 1; | |
3886 | ||
3887 | /* | |
3888 | * We may have set a slab to be unmergeable during bootstrap. | |
3889 | */ | |
3890 | if (s->refcount < 0) | |
3891 | return 1; | |
3892 | ||
3893 | return 0; | |
3894 | } | |
3895 | ||
3896 | static struct kmem_cache *find_mergeable(size_t size, | |
3897 | size_t align, unsigned long flags, const char *name, | |
3898 | void (*ctor)(void *)) | |
3899 | { | |
3900 | struct kmem_cache *s; | |
3901 | ||
3902 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) | |
3903 | return NULL; | |
3904 | ||
3905 | if (ctor) | |
3906 | return NULL; | |
3907 | ||
3908 | size = ALIGN(size, sizeof(void *)); | |
3909 | align = calculate_alignment(flags, align, size); | |
3910 | size = ALIGN(size, align); | |
3911 | flags = kmem_cache_flags(size, flags, name, NULL); | |
3912 | ||
3913 | list_for_each_entry(s, &slab_caches, list) { | |
3914 | if (slab_unmergeable(s)) | |
3915 | continue; | |
3916 | ||
3917 | if (size > s->size) | |
3918 | continue; | |
3919 | ||
3920 | if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) | |
3921 | continue; | |
3922 | /* | |
3923 | * Check if alignment is compatible. | |
3924 | * Courtesy of Adrian Drzewiecki | |
3925 | */ | |
3926 | if ((s->size & ~(align - 1)) != s->size) | |
3927 | continue; | |
3928 | ||
3929 | if (s->size - size >= sizeof(void *)) | |
3930 | continue; | |
3931 | ||
3932 | return s; | |
3933 | } | |
3934 | return NULL; | |
3935 | } | |
3936 | ||
3937 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, | |
3938 | size_t align, unsigned long flags, void (*ctor)(void *)) | |
3939 | { | |
3940 | struct kmem_cache *s; | |
3941 | char *n; | |
3942 | ||
3943 | if (WARN_ON(!name)) | |
3944 | return NULL; | |
3945 | ||
3946 | down_write(&slub_lock); | |
3947 | s = find_mergeable(size, align, flags, name, ctor); | |
3948 | if (s) { | |
3949 | s->refcount++; | |
3950 | /* | |
3951 | * Adjust the object sizes so that we clear | |
3952 | * the complete object on kzalloc. | |
3953 | */ | |
3954 | s->objsize = max(s->objsize, (int)size); | |
3955 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); | |
3956 | ||
3957 | if (sysfs_slab_alias(s, name)) { | |
3958 | s->refcount--; | |
3959 | goto err; | |
3960 | } | |
3961 | up_write(&slub_lock); | |
3962 | return s; | |
3963 | } | |
3964 | ||
3965 | n = kstrdup(name, GFP_KERNEL); | |
3966 | if (!n) | |
3967 | goto err; | |
3968 | ||
3969 | s = kmalloc(kmem_size, GFP_KERNEL); | |
3970 | if (s) { | |
3971 | if (kmem_cache_open(s, n, | |
3972 | size, align, flags, ctor)) { | |
3973 | list_add(&s->list, &slab_caches); | |
3974 | up_write(&slub_lock); | |
3975 | if (sysfs_slab_add(s)) { | |
3976 | down_write(&slub_lock); | |
3977 | list_del(&s->list); | |
3978 | kfree(n); | |
3979 | kfree(s); | |
3980 | goto err; | |
3981 | } | |
3982 | return s; | |
3983 | } | |
3984 | kfree(s); | |
3985 | } | |
3986 | kfree(n); | |
3987 | err: | |
3988 | up_write(&slub_lock); | |
3989 | ||
3990 | if (flags & SLAB_PANIC) | |
3991 | panic("Cannot create slabcache %s\n", name); | |
3992 | else | |
3993 | s = NULL; | |
3994 | return s; | |
3995 | } | |
3996 | EXPORT_SYMBOL(kmem_cache_create); | |
3997 | ||
3998 | #ifdef CONFIG_SMP | |
3999 | /* | |
4000 | * Use the cpu notifier to insure that the cpu slabs are flushed when | |
4001 | * necessary. | |
4002 | */ | |
4003 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, | |
4004 | unsigned long action, void *hcpu) | |
4005 | { | |
4006 | long cpu = (long)hcpu; | |
4007 | struct kmem_cache *s; | |
4008 | unsigned long flags; | |
4009 | ||
4010 | switch (action) { | |
4011 | case CPU_UP_CANCELED: | |
4012 | case CPU_UP_CANCELED_FROZEN: | |
4013 | case CPU_DEAD: | |
4014 | case CPU_DEAD_FROZEN: | |
4015 | down_read(&slub_lock); | |
4016 | list_for_each_entry(s, &slab_caches, list) { | |
4017 | local_irq_save(flags); | |
4018 | __flush_cpu_slab(s, cpu); | |
4019 | local_irq_restore(flags); | |
4020 | } | |
4021 | up_read(&slub_lock); | |
4022 | break; | |
4023 | default: | |
4024 | break; | |
4025 | } | |
4026 | return NOTIFY_OK; | |
4027 | } | |
4028 | ||
4029 | static struct notifier_block __cpuinitdata slab_notifier = { | |
4030 | .notifier_call = slab_cpuup_callback | |
4031 | }; | |
4032 | ||
4033 | #endif | |
4034 | ||
4035 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) | |
4036 | { | |
4037 | struct kmem_cache *s; | |
4038 | void *ret; | |
4039 | ||
4040 | if (unlikely(size > SLUB_MAX_SIZE)) | |
4041 | return kmalloc_large(size, gfpflags); | |
4042 | ||
4043 | s = get_slab(size, gfpflags); | |
4044 | ||
4045 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
4046 | return s; | |
4047 | ||
4048 | ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); | |
4049 | ||
4050 | /* Honor the call site pointer we received. */ | |
4051 | trace_kmalloc(caller, ret, size, s->size, gfpflags); | |
4052 | ||
4053 | return ret; | |
4054 | } | |
4055 | ||
4056 | #ifdef CONFIG_NUMA | |
4057 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
4058 | int node, unsigned long caller) | |
4059 | { | |
4060 | struct kmem_cache *s; | |
4061 | void *ret; | |
4062 | ||
4063 | if (unlikely(size > SLUB_MAX_SIZE)) { | |
4064 | ret = kmalloc_large_node(size, gfpflags, node); | |
4065 | ||
4066 | trace_kmalloc_node(caller, ret, | |
4067 | size, PAGE_SIZE << get_order(size), | |
4068 | gfpflags, node); | |
4069 | ||
4070 | return ret; | |
4071 | } | |
4072 | ||
4073 | s = get_slab(size, gfpflags); | |
4074 | ||
4075 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
4076 | return s; | |
4077 | ||
4078 | ret = slab_alloc(s, gfpflags, node, caller); | |
4079 | ||
4080 | /* Honor the call site pointer we received. */ | |
4081 | trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); | |
4082 | ||
4083 | return ret; | |
4084 | } | |
4085 | #endif | |
4086 | ||
4087 | #ifdef CONFIG_SYSFS | |
4088 | static int count_inuse(struct page *page) | |
4089 | { | |
4090 | return page->inuse; | |
4091 | } | |
4092 | ||
4093 | static int count_total(struct page *page) | |
4094 | { | |
4095 | return page->objects; | |
4096 | } | |
4097 | #endif | |
4098 | ||
4099 | #ifdef CONFIG_SLUB_DEBUG | |
4100 | static int validate_slab(struct kmem_cache *s, struct page *page, | |
4101 | unsigned long *map) | |
4102 | { | |
4103 | void *p; | |
4104 | void *addr = page_address(page); | |
4105 | ||
4106 | if (!check_slab(s, page) || | |
4107 | !on_freelist(s, page, NULL)) | |
4108 | return 0; | |
4109 | ||
4110 | /* Now we know that a valid freelist exists */ | |
4111 | bitmap_zero(map, page->objects); | |
4112 | ||
4113 | get_map(s, page, map); | |
4114 | for_each_object(p, s, addr, page->objects) { | |
4115 | if (test_bit(slab_index(p, s, addr), map)) | |
4116 | if (!check_object(s, page, p, SLUB_RED_INACTIVE)) | |
4117 | return 0; | |
4118 | } | |
4119 | ||
4120 | for_each_object(p, s, addr, page->objects) | |
4121 | if (!test_bit(slab_index(p, s, addr), map)) | |
4122 | if (!check_object(s, page, p, SLUB_RED_ACTIVE)) | |
4123 | return 0; | |
4124 | return 1; | |
4125 | } | |
4126 | ||
4127 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, | |
4128 | unsigned long *map) | |
4129 | { | |
4130 | slab_lock(page); | |
4131 | validate_slab(s, page, map); | |
4132 | slab_unlock(page); | |
4133 | } | |
4134 | ||
4135 | static int validate_slab_node(struct kmem_cache *s, | |
4136 | struct kmem_cache_node *n, unsigned long *map) | |
4137 | { | |
4138 | unsigned long count = 0; | |
4139 | struct page *page; | |
4140 | unsigned long flags; | |
4141 | ||
4142 | spin_lock_irqsave(&n->list_lock, flags); | |
4143 | ||
4144 | list_for_each_entry(page, &n->partial, lru) { | |
4145 | validate_slab_slab(s, page, map); | |
4146 | count++; | |
4147 | } | |
4148 | if (count != n->nr_partial) | |
4149 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " | |
4150 | "counter=%ld\n", s->name, count, n->nr_partial); | |
4151 | ||
4152 | if (!(s->flags & SLAB_STORE_USER)) | |
4153 | goto out; | |
4154 | ||
4155 | list_for_each_entry(page, &n->full, lru) { | |
4156 | validate_slab_slab(s, page, map); | |
4157 | count++; | |
4158 | } | |
4159 | if (count != atomic_long_read(&n->nr_slabs)) | |
4160 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " | |
4161 | "counter=%ld\n", s->name, count, | |
4162 | atomic_long_read(&n->nr_slabs)); | |
4163 | ||
4164 | out: | |
4165 | spin_unlock_irqrestore(&n->list_lock, flags); | |
4166 | return count; | |
4167 | } | |
4168 | ||
4169 | static long validate_slab_cache(struct kmem_cache *s) | |
4170 | { | |
4171 | int node; | |
4172 | unsigned long count = 0; | |
4173 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * | |
4174 | sizeof(unsigned long), GFP_KERNEL); | |
4175 | ||
4176 | if (!map) | |
4177 | return -ENOMEM; | |
4178 | ||
4179 | flush_all(s); | |
4180 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
4181 | struct kmem_cache_node *n = get_node(s, node); | |
4182 | ||
4183 | count += validate_slab_node(s, n, map); | |
4184 | } | |
4185 | kfree(map); | |
4186 | return count; | |
4187 | } | |
4188 | /* | |
4189 | * Generate lists of code addresses where slabcache objects are allocated | |
4190 | * and freed. | |
4191 | */ | |
4192 | ||
4193 | struct location { | |
4194 | unsigned long count; | |
4195 | unsigned long addr; | |
4196 | long long sum_time; | |
4197 | long min_time; | |
4198 | long max_time; | |
4199 | long min_pid; | |
4200 | long max_pid; | |
4201 | DECLARE_BITMAP(cpus, NR_CPUS); | |
4202 | nodemask_t nodes; | |
4203 | }; | |
4204 | ||
4205 | struct loc_track { | |
4206 | unsigned long max; | |
4207 | unsigned long count; | |
4208 | struct location *loc; | |
4209 | }; | |
4210 | ||
4211 | static void free_loc_track(struct loc_track *t) | |
4212 | { | |
4213 | if (t->max) | |
4214 | free_pages((unsigned long)t->loc, | |
4215 | get_order(sizeof(struct location) * t->max)); | |
4216 | } | |
4217 | ||
4218 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) | |
4219 | { | |
4220 | struct location *l; | |
4221 | int order; | |
4222 | ||
4223 | order = get_order(sizeof(struct location) * max); | |
4224 | ||
4225 | l = (void *)__get_free_pages(flags, order); | |
4226 | if (!l) | |
4227 | return 0; | |
4228 | ||
4229 | if (t->count) { | |
4230 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
4231 | free_loc_track(t); | |
4232 | } | |
4233 | t->max = max; | |
4234 | t->loc = l; | |
4235 | return 1; | |
4236 | } | |
4237 | ||
4238 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
4239 | const struct track *track) | |
4240 | { | |
4241 | long start, end, pos; | |
4242 | struct location *l; | |
4243 | unsigned long caddr; | |
4244 | unsigned long age = jiffies - track->when; | |
4245 | ||
4246 | start = -1; | |
4247 | end = t->count; | |
4248 | ||
4249 | for ( ; ; ) { | |
4250 | pos = start + (end - start + 1) / 2; | |
4251 | ||
4252 | /* | |
4253 | * There is nothing at "end". If we end up there | |
4254 | * we need to add something to before end. | |
4255 | */ | |
4256 | if (pos == end) | |
4257 | break; | |
4258 | ||
4259 | caddr = t->loc[pos].addr; | |
4260 | if (track->addr == caddr) { | |
4261 | ||
4262 | l = &t->loc[pos]; | |
4263 | l->count++; | |
4264 | if (track->when) { | |
4265 | l->sum_time += age; | |
4266 | if (age < l->min_time) | |
4267 | l->min_time = age; | |
4268 | if (age > l->max_time) | |
4269 | l->max_time = age; | |
4270 | ||
4271 | if (track->pid < l->min_pid) | |
4272 | l->min_pid = track->pid; | |
4273 | if (track->pid > l->max_pid) | |
4274 | l->max_pid = track->pid; | |
4275 | ||
4276 | cpumask_set_cpu(track->cpu, | |
4277 | to_cpumask(l->cpus)); | |
4278 | } | |
4279 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
4280 | return 1; | |
4281 | } | |
4282 | ||
4283 | if (track->addr < caddr) | |
4284 | end = pos; | |
4285 | else | |
4286 | start = pos; | |
4287 | } | |
4288 | ||
4289 | /* | |
4290 | * Not found. Insert new tracking element. | |
4291 | */ | |
4292 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) | |
4293 | return 0; | |
4294 | ||
4295 | l = t->loc + pos; | |
4296 | if (pos < t->count) | |
4297 | memmove(l + 1, l, | |
4298 | (t->count - pos) * sizeof(struct location)); | |
4299 | t->count++; | |
4300 | l->count = 1; | |
4301 | l->addr = track->addr; | |
4302 | l->sum_time = age; | |
4303 | l->min_time = age; | |
4304 | l->max_time = age; | |
4305 | l->min_pid = track->pid; | |
4306 | l->max_pid = track->pid; | |
4307 | cpumask_clear(to_cpumask(l->cpus)); | |
4308 | cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); | |
4309 | nodes_clear(l->nodes); | |
4310 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
4311 | return 1; | |
4312 | } | |
4313 | ||
4314 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
4315 | struct page *page, enum track_item alloc, | |
4316 | unsigned long *map) | |
4317 | { | |
4318 | void *addr = page_address(page); | |
4319 | void *p; | |
4320 | ||
4321 | bitmap_zero(map, page->objects); | |
4322 | get_map(s, page, map); | |
4323 | ||
4324 | for_each_object(p, s, addr, page->objects) | |
4325 | if (!test_bit(slab_index(p, s, addr), map)) | |
4326 | add_location(t, s, get_track(s, p, alloc)); | |
4327 | } | |
4328 | ||
4329 | static int list_locations(struct kmem_cache *s, char *buf, | |
4330 | enum track_item alloc) | |
4331 | { | |
4332 | int len = 0; | |
4333 | unsigned long i; | |
4334 | struct loc_track t = { 0, 0, NULL }; | |
4335 | int node; | |
4336 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * | |
4337 | sizeof(unsigned long), GFP_KERNEL); | |
4338 | ||
4339 | if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), | |
4340 | GFP_TEMPORARY)) { | |
4341 | kfree(map); | |
4342 | return sprintf(buf, "Out of memory\n"); | |
4343 | } | |
4344 | /* Push back cpu slabs */ | |
4345 | flush_all(s); | |
4346 | ||
4347 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
4348 | struct kmem_cache_node *n = get_node(s, node); | |
4349 | unsigned long flags; | |
4350 | struct page *page; | |
4351 | ||
4352 | if (!atomic_long_read(&n->nr_slabs)) | |
4353 | continue; | |
4354 | ||
4355 | spin_lock_irqsave(&n->list_lock, flags); | |
4356 | list_for_each_entry(page, &n->partial, lru) | |
4357 | process_slab(&t, s, page, alloc, map); | |
4358 | list_for_each_entry(page, &n->full, lru) | |
4359 | process_slab(&t, s, page, alloc, map); | |
4360 | spin_unlock_irqrestore(&n->list_lock, flags); | |
4361 | } | |
4362 | ||
4363 | for (i = 0; i < t.count; i++) { | |
4364 | struct location *l = &t.loc[i]; | |
4365 | ||
4366 | if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) | |
4367 | break; | |
4368 | len += sprintf(buf + len, "%7ld ", l->count); | |
4369 | ||
4370 | if (l->addr) | |
4371 | len += sprintf(buf + len, "%pS", (void *)l->addr); | |
4372 | else | |
4373 | len += sprintf(buf + len, "<not-available>"); | |
4374 | ||
4375 | if (l->sum_time != l->min_time) { | |
4376 | len += sprintf(buf + len, " age=%ld/%ld/%ld", | |
4377 | l->min_time, | |
4378 | (long)div_u64(l->sum_time, l->count), | |
4379 | l->max_time); | |
4380 | } else | |
4381 | len += sprintf(buf + len, " age=%ld", | |
4382 | l->min_time); | |
4383 | ||
4384 | if (l->min_pid != l->max_pid) | |
4385 | len += sprintf(buf + len, " pid=%ld-%ld", | |
4386 | l->min_pid, l->max_pid); | |
4387 | else | |
4388 | len += sprintf(buf + len, " pid=%ld", | |
4389 | l->min_pid); | |
4390 | ||
4391 | if (num_online_cpus() > 1 && | |
4392 | !cpumask_empty(to_cpumask(l->cpus)) && | |
4393 | len < PAGE_SIZE - 60) { | |
4394 | len += sprintf(buf + len, " cpus="); | |
4395 | len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
4396 | to_cpumask(l->cpus)); | |
4397 | } | |
4398 | ||
4399 | if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && | |
4400 | len < PAGE_SIZE - 60) { | |
4401 | len += sprintf(buf + len, " nodes="); | |
4402 | len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
4403 | l->nodes); | |
4404 | } | |
4405 | ||
4406 | len += sprintf(buf + len, "\n"); | |
4407 | } | |
4408 | ||
4409 | free_loc_track(&t); | |
4410 | kfree(map); | |
4411 | if (!t.count) | |
4412 | len += sprintf(buf, "No data\n"); | |
4413 | return len; | |
4414 | } | |
4415 | #endif | |
4416 | ||
4417 | #ifdef SLUB_RESILIENCY_TEST | |
4418 | static void resiliency_test(void) | |
4419 | { | |
4420 | u8 *p; | |
4421 | ||
4422 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10); | |
4423 | ||
4424 | printk(KERN_ERR "SLUB resiliency testing\n"); | |
4425 | printk(KERN_ERR "-----------------------\n"); | |
4426 | printk(KERN_ERR "A. Corruption after allocation\n"); | |
4427 | ||
4428 | p = kzalloc(16, GFP_KERNEL); | |
4429 | p[16] = 0x12; | |
4430 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" | |
4431 | " 0x12->0x%p\n\n", p + 16); | |
4432 | ||
4433 | validate_slab_cache(kmalloc_caches[4]); | |
4434 | ||
4435 | /* Hmmm... The next two are dangerous */ | |
4436 | p = kzalloc(32, GFP_KERNEL); | |
4437 | p[32 + sizeof(void *)] = 0x34; | |
4438 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" | |
4439 | " 0x34 -> -0x%p\n", p); | |
4440 | printk(KERN_ERR | |
4441 | "If allocated object is overwritten then not detectable\n\n"); | |
4442 | ||
4443 | validate_slab_cache(kmalloc_caches[5]); | |
4444 | p = kzalloc(64, GFP_KERNEL); | |
4445 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
4446 | *p = 0x56; | |
4447 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
4448 | p); | |
4449 | printk(KERN_ERR | |
4450 | "If allocated object is overwritten then not detectable\n\n"); | |
4451 | validate_slab_cache(kmalloc_caches[6]); | |
4452 | ||
4453 | printk(KERN_ERR "\nB. Corruption after free\n"); | |
4454 | p = kzalloc(128, GFP_KERNEL); | |
4455 | kfree(p); | |
4456 | *p = 0x78; | |
4457 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
4458 | validate_slab_cache(kmalloc_caches[7]); | |
4459 | ||
4460 | p = kzalloc(256, GFP_KERNEL); | |
4461 | kfree(p); | |
4462 | p[50] = 0x9a; | |
4463 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", | |
4464 | p); | |
4465 | validate_slab_cache(kmalloc_caches[8]); | |
4466 | ||
4467 | p = kzalloc(512, GFP_KERNEL); | |
4468 | kfree(p); | |
4469 | p[512] = 0xab; | |
4470 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
4471 | validate_slab_cache(kmalloc_caches[9]); | |
4472 | } | |
4473 | #else | |
4474 | #ifdef CONFIG_SYSFS | |
4475 | static void resiliency_test(void) {}; | |
4476 | #endif | |
4477 | #endif | |
4478 | ||
4479 | #ifdef CONFIG_SYSFS | |
4480 | enum slab_stat_type { | |
4481 | SL_ALL, /* All slabs */ | |
4482 | SL_PARTIAL, /* Only partially allocated slabs */ | |
4483 | SL_CPU, /* Only slabs used for cpu caches */ | |
4484 | SL_OBJECTS, /* Determine allocated objects not slabs */ | |
4485 | SL_TOTAL /* Determine object capacity not slabs */ | |
4486 | }; | |
4487 | ||
4488 | #define SO_ALL (1 << SL_ALL) | |
4489 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
4490 | #define SO_CPU (1 << SL_CPU) | |
4491 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
4492 | #define SO_TOTAL (1 << SL_TOTAL) | |
4493 | ||
4494 | static ssize_t show_slab_objects(struct kmem_cache *s, | |
4495 | char *buf, unsigned long flags) | |
4496 | { | |
4497 | unsigned long total = 0; | |
4498 | int node; | |
4499 | int x; | |
4500 | unsigned long *nodes; | |
4501 | unsigned long *per_cpu; | |
4502 | ||
4503 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
4504 | if (!nodes) | |
4505 | return -ENOMEM; | |
4506 | per_cpu = nodes + nr_node_ids; | |
4507 | ||
4508 | if (flags & SO_CPU) { | |
4509 | int cpu; | |
4510 | ||
4511 | for_each_possible_cpu(cpu) { | |
4512 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); | |
4513 | int node; | |
4514 | struct page *page; | |
4515 | ||
4516 | page = ACCESS_ONCE(c->page); | |
4517 | if (!page) | |
4518 | continue; | |
4519 | ||
4520 | node = page_to_nid(page); | |
4521 | if (flags & SO_TOTAL) | |
4522 | x = page->objects; | |
4523 | else if (flags & SO_OBJECTS) | |
4524 | x = page->inuse; | |
4525 | else | |
4526 | x = 1; | |
4527 | ||
4528 | total += x; | |
4529 | nodes[node] += x; | |
4530 | ||
4531 | page = ACCESS_ONCE(c->partial); | |
4532 | if (page) { | |
4533 | x = page->pobjects; | |
4534 | total += x; | |
4535 | nodes[node] += x; | |
4536 | } | |
4537 | ||
4538 | per_cpu[node]++; | |
4539 | } | |
4540 | } | |
4541 | ||
4542 | lock_memory_hotplug(); | |
4543 | #ifdef CONFIG_SLUB_DEBUG | |
4544 | if (flags & SO_ALL) { | |
4545 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
4546 | struct kmem_cache_node *n = get_node(s, node); | |
4547 | ||
4548 | if (flags & SO_TOTAL) | |
4549 | x = atomic_long_read(&n->total_objects); | |
4550 | else if (flags & SO_OBJECTS) | |
4551 | x = atomic_long_read(&n->total_objects) - | |
4552 | count_partial(n, count_free); | |
4553 | ||
4554 | else | |
4555 | x = atomic_long_read(&n->nr_slabs); | |
4556 | total += x; | |
4557 | nodes[node] += x; | |
4558 | } | |
4559 | ||
4560 | } else | |
4561 | #endif | |
4562 | if (flags & SO_PARTIAL) { | |
4563 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
4564 | struct kmem_cache_node *n = get_node(s, node); | |
4565 | ||
4566 | if (flags & SO_TOTAL) | |
4567 | x = count_partial(n, count_total); | |
4568 | else if (flags & SO_OBJECTS) | |
4569 | x = count_partial(n, count_inuse); | |
4570 | else | |
4571 | x = n->nr_partial; | |
4572 | total += x; | |
4573 | nodes[node] += x; | |
4574 | } | |
4575 | } | |
4576 | x = sprintf(buf, "%lu", total); | |
4577 | #ifdef CONFIG_NUMA | |
4578 | for_each_node_state(node, N_NORMAL_MEMORY) | |
4579 | if (nodes[node]) | |
4580 | x += sprintf(buf + x, " N%d=%lu", | |
4581 | node, nodes[node]); | |
4582 | #endif | |
4583 | unlock_memory_hotplug(); | |
4584 | kfree(nodes); | |
4585 | return x + sprintf(buf + x, "\n"); | |
4586 | } | |
4587 | ||
4588 | #ifdef CONFIG_SLUB_DEBUG | |
4589 | static int any_slab_objects(struct kmem_cache *s) | |
4590 | { | |
4591 | int node; | |
4592 | ||
4593 | for_each_online_node(node) { | |
4594 | struct kmem_cache_node *n = get_node(s, node); | |
4595 | ||
4596 | if (!n) | |
4597 | continue; | |
4598 | ||
4599 | if (atomic_long_read(&n->total_objects)) | |
4600 | return 1; | |
4601 | } | |
4602 | return 0; | |
4603 | } | |
4604 | #endif | |
4605 | ||
4606 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
4607 | #define to_slab(n) container_of(n, struct kmem_cache, kobj) | |
4608 | ||
4609 | struct slab_attribute { | |
4610 | struct attribute attr; | |
4611 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
4612 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
4613 | }; | |
4614 | ||
4615 | #define SLAB_ATTR_RO(_name) \ | |
4616 | static struct slab_attribute _name##_attr = \ | |
4617 | __ATTR(_name, 0400, _name##_show, NULL) | |
4618 | ||
4619 | #define SLAB_ATTR(_name) \ | |
4620 | static struct slab_attribute _name##_attr = \ | |
4621 | __ATTR(_name, 0600, _name##_show, _name##_store) | |
4622 | ||
4623 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) | |
4624 | { | |
4625 | return sprintf(buf, "%d\n", s->size); | |
4626 | } | |
4627 | SLAB_ATTR_RO(slab_size); | |
4628 | ||
4629 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
4630 | { | |
4631 | return sprintf(buf, "%d\n", s->align); | |
4632 | } | |
4633 | SLAB_ATTR_RO(align); | |
4634 | ||
4635 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
4636 | { | |
4637 | return sprintf(buf, "%d\n", s->objsize); | |
4638 | } | |
4639 | SLAB_ATTR_RO(object_size); | |
4640 | ||
4641 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
4642 | { | |
4643 | return sprintf(buf, "%d\n", oo_objects(s->oo)); | |
4644 | } | |
4645 | SLAB_ATTR_RO(objs_per_slab); | |
4646 | ||
4647 | static ssize_t order_store(struct kmem_cache *s, | |
4648 | const char *buf, size_t length) | |
4649 | { | |
4650 | unsigned long order; | |
4651 | int err; | |
4652 | ||
4653 | err = strict_strtoul(buf, 10, &order); | |
4654 | if (err) | |
4655 | return err; | |
4656 | ||
4657 | if (order > slub_max_order || order < slub_min_order) | |
4658 | return -EINVAL; | |
4659 | ||
4660 | calculate_sizes(s, order); | |
4661 | return length; | |
4662 | } | |
4663 | ||
4664 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
4665 | { | |
4666 | return sprintf(buf, "%d\n", oo_order(s->oo)); | |
4667 | } | |
4668 | SLAB_ATTR(order); | |
4669 | ||
4670 | static ssize_t min_partial_show(struct kmem_cache *s, char *buf) | |
4671 | { | |
4672 | return sprintf(buf, "%lu\n", s->min_partial); | |
4673 | } | |
4674 | ||
4675 | static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, | |
4676 | size_t length) | |
4677 | { | |
4678 | unsigned long min; | |
4679 | int err; | |
4680 | ||
4681 | err = strict_strtoul(buf, 10, &min); | |
4682 | if (err) | |
4683 | return err; | |
4684 | ||
4685 | set_min_partial(s, min); | |
4686 | return length; | |
4687 | } | |
4688 | SLAB_ATTR(min_partial); | |
4689 | ||
4690 | static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) | |
4691 | { | |
4692 | return sprintf(buf, "%u\n", s->cpu_partial); | |
4693 | } | |
4694 | ||
4695 | static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, | |
4696 | size_t length) | |
4697 | { | |
4698 | unsigned long objects; | |
4699 | int err; | |
4700 | ||
4701 | err = strict_strtoul(buf, 10, &objects); | |
4702 | if (err) | |
4703 | return err; | |
4704 | if (objects && kmem_cache_debug(s)) | |
4705 | return -EINVAL; | |
4706 | ||
4707 | s->cpu_partial = objects; | |
4708 | flush_all(s); | |
4709 | return length; | |
4710 | } | |
4711 | SLAB_ATTR(cpu_partial); | |
4712 | ||
4713 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
4714 | { | |
4715 | if (!s->ctor) | |
4716 | return 0; | |
4717 | return sprintf(buf, "%pS\n", s->ctor); | |
4718 | } | |
4719 | SLAB_ATTR_RO(ctor); | |
4720 | ||
4721 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) | |
4722 | { | |
4723 | return sprintf(buf, "%d\n", s->refcount - 1); | |
4724 | } | |
4725 | SLAB_ATTR_RO(aliases); | |
4726 | ||
4727 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
4728 | { | |
4729 | return show_slab_objects(s, buf, SO_PARTIAL); | |
4730 | } | |
4731 | SLAB_ATTR_RO(partial); | |
4732 | ||
4733 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
4734 | { | |
4735 | return show_slab_objects(s, buf, SO_CPU); | |
4736 | } | |
4737 | SLAB_ATTR_RO(cpu_slabs); | |
4738 | ||
4739 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
4740 | { | |
4741 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); | |
4742 | } | |
4743 | SLAB_ATTR_RO(objects); | |
4744 | ||
4745 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) | |
4746 | { | |
4747 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); | |
4748 | } | |
4749 | SLAB_ATTR_RO(objects_partial); | |
4750 | ||
4751 | static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) | |
4752 | { | |
4753 | int objects = 0; | |
4754 | int pages = 0; | |
4755 | int cpu; | |
4756 | int len; | |
4757 | ||
4758 | for_each_online_cpu(cpu) { | |
4759 | struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial; | |
4760 | ||
4761 | if (page) { | |
4762 | pages += page->pages; | |
4763 | objects += page->pobjects; | |
4764 | } | |
4765 | } | |
4766 | ||
4767 | len = sprintf(buf, "%d(%d)", objects, pages); | |
4768 | ||
4769 | #ifdef CONFIG_SMP | |
4770 | for_each_online_cpu(cpu) { | |
4771 | struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial; | |
4772 | ||
4773 | if (page && len < PAGE_SIZE - 20) | |
4774 | len += sprintf(buf + len, " C%d=%d(%d)", cpu, | |
4775 | page->pobjects, page->pages); | |
4776 | } | |
4777 | #endif | |
4778 | return len + sprintf(buf + len, "\n"); | |
4779 | } | |
4780 | SLAB_ATTR_RO(slabs_cpu_partial); | |
4781 | ||
4782 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
4783 | { | |
4784 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
4785 | } | |
4786 | ||
4787 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
4788 | const char *buf, size_t length) | |
4789 | { | |
4790 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
4791 | if (buf[0] == '1') | |
4792 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
4793 | return length; | |
4794 | } | |
4795 | SLAB_ATTR(reclaim_account); | |
4796 | ||
4797 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
4798 | { | |
4799 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); | |
4800 | } | |
4801 | SLAB_ATTR_RO(hwcache_align); | |
4802 | ||
4803 | #ifdef CONFIG_ZONE_DMA | |
4804 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
4805 | { | |
4806 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
4807 | } | |
4808 | SLAB_ATTR_RO(cache_dma); | |
4809 | #endif | |
4810 | ||
4811 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
4812 | { | |
4813 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); | |
4814 | } | |
4815 | SLAB_ATTR_RO(destroy_by_rcu); | |
4816 | ||
4817 | static ssize_t reserved_show(struct kmem_cache *s, char *buf) | |
4818 | { | |
4819 | return sprintf(buf, "%d\n", s->reserved); | |
4820 | } | |
4821 | SLAB_ATTR_RO(reserved); | |
4822 | ||
4823 | #ifdef CONFIG_SLUB_DEBUG | |
4824 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
4825 | { | |
4826 | return show_slab_objects(s, buf, SO_ALL); | |
4827 | } | |
4828 | SLAB_ATTR_RO(slabs); | |
4829 | ||
4830 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) | |
4831 | { | |
4832 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); | |
4833 | } | |
4834 | SLAB_ATTR_RO(total_objects); | |
4835 | ||
4836 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
4837 | { | |
4838 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); | |
4839 | } | |
4840 | ||
4841 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
4842 | const char *buf, size_t length) | |
4843 | { | |
4844 | s->flags &= ~SLAB_DEBUG_FREE; | |
4845 | if (buf[0] == '1') { | |
4846 | s->flags &= ~__CMPXCHG_DOUBLE; | |
4847 | s->flags |= SLAB_DEBUG_FREE; | |
4848 | } | |
4849 | return length; | |
4850 | } | |
4851 | SLAB_ATTR(sanity_checks); | |
4852 | ||
4853 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
4854 | { | |
4855 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
4856 | } | |
4857 | ||
4858 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
4859 | size_t length) | |
4860 | { | |
4861 | s->flags &= ~SLAB_TRACE; | |
4862 | if (buf[0] == '1') { | |
4863 | s->flags &= ~__CMPXCHG_DOUBLE; | |
4864 | s->flags |= SLAB_TRACE; | |
4865 | } | |
4866 | return length; | |
4867 | } | |
4868 | SLAB_ATTR(trace); | |
4869 | ||
4870 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
4871 | { | |
4872 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
4873 | } | |
4874 | ||
4875 | static ssize_t red_zone_store(struct kmem_cache *s, | |
4876 | const char *buf, size_t length) | |
4877 | { | |
4878 | if (any_slab_objects(s)) | |
4879 | return -EBUSY; | |
4880 | ||
4881 | s->flags &= ~SLAB_RED_ZONE; | |
4882 | if (buf[0] == '1') { | |
4883 | s->flags &= ~__CMPXCHG_DOUBLE; | |
4884 | s->flags |= SLAB_RED_ZONE; | |
4885 | } | |
4886 | calculate_sizes(s, -1); | |
4887 | return length; | |
4888 | } | |
4889 | SLAB_ATTR(red_zone); | |
4890 | ||
4891 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
4892 | { | |
4893 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
4894 | } | |
4895 | ||
4896 | static ssize_t poison_store(struct kmem_cache *s, | |
4897 | const char *buf, size_t length) | |
4898 | { | |
4899 | if (any_slab_objects(s)) | |
4900 | return -EBUSY; | |
4901 | ||
4902 | s->flags &= ~SLAB_POISON; | |
4903 | if (buf[0] == '1') { | |
4904 | s->flags &= ~__CMPXCHG_DOUBLE; | |
4905 | s->flags |= SLAB_POISON; | |
4906 | } | |
4907 | calculate_sizes(s, -1); | |
4908 | return length; | |
4909 | } | |
4910 | SLAB_ATTR(poison); | |
4911 | ||
4912 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
4913 | { | |
4914 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
4915 | } | |
4916 | ||
4917 | static ssize_t store_user_store(struct kmem_cache *s, | |
4918 | const char *buf, size_t length) | |
4919 | { | |
4920 | if (any_slab_objects(s)) | |
4921 | return -EBUSY; | |
4922 | ||
4923 | s->flags &= ~SLAB_STORE_USER; | |
4924 | if (buf[0] == '1') { | |
4925 | s->flags &= ~__CMPXCHG_DOUBLE; | |
4926 | s->flags |= SLAB_STORE_USER; | |
4927 | } | |
4928 | calculate_sizes(s, -1); | |
4929 | return length; | |
4930 | } | |
4931 | SLAB_ATTR(store_user); | |
4932 | ||
4933 | static ssize_t validate_show(struct kmem_cache *s, char *buf) | |
4934 | { | |
4935 | return 0; | |
4936 | } | |
4937 | ||
4938 | static ssize_t validate_store(struct kmem_cache *s, | |
4939 | const char *buf, size_t length) | |
4940 | { | |
4941 | int ret = -EINVAL; | |
4942 | ||
4943 | if (buf[0] == '1') { | |
4944 | ret = validate_slab_cache(s); | |
4945 | if (ret >= 0) | |
4946 | ret = length; | |
4947 | } | |
4948 | return ret; | |
4949 | } | |
4950 | SLAB_ATTR(validate); | |
4951 | ||
4952 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) | |
4953 | { | |
4954 | if (!(s->flags & SLAB_STORE_USER)) | |
4955 | return -ENOSYS; | |
4956 | return list_locations(s, buf, TRACK_ALLOC); | |
4957 | } | |
4958 | SLAB_ATTR_RO(alloc_calls); | |
4959 | ||
4960 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
4961 | { | |
4962 | if (!(s->flags & SLAB_STORE_USER)) | |
4963 | return -ENOSYS; | |
4964 | return list_locations(s, buf, TRACK_FREE); | |
4965 | } | |
4966 | SLAB_ATTR_RO(free_calls); | |
4967 | #endif /* CONFIG_SLUB_DEBUG */ | |
4968 | ||
4969 | #ifdef CONFIG_FAILSLAB | |
4970 | static ssize_t failslab_show(struct kmem_cache *s, char *buf) | |
4971 | { | |
4972 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); | |
4973 | } | |
4974 | ||
4975 | static ssize_t failslab_store(struct kmem_cache *s, const char *buf, | |
4976 | size_t length) | |
4977 | { | |
4978 | s->flags &= ~SLAB_FAILSLAB; | |
4979 | if (buf[0] == '1') | |
4980 | s->flags |= SLAB_FAILSLAB; | |
4981 | return length; | |
4982 | } | |
4983 | SLAB_ATTR(failslab); | |
4984 | #endif | |
4985 | ||
4986 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) | |
4987 | { | |
4988 | return 0; | |
4989 | } | |
4990 | ||
4991 | static ssize_t shrink_store(struct kmem_cache *s, | |
4992 | const char *buf, size_t length) | |
4993 | { | |
4994 | if (buf[0] == '1') { | |
4995 | int rc = kmem_cache_shrink(s); | |
4996 | ||
4997 | if (rc) | |
4998 | return rc; | |
4999 | } else | |
5000 | return -EINVAL; | |
5001 | return length; | |
5002 | } | |
5003 | SLAB_ATTR(shrink); | |
5004 | ||
5005 | #ifdef CONFIG_NUMA | |
5006 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) | |
5007 | { | |
5008 | return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); | |
5009 | } | |
5010 | ||
5011 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, | |
5012 | const char *buf, size_t length) | |
5013 | { | |
5014 | unsigned long ratio; | |
5015 | int err; | |
5016 | ||
5017 | err = strict_strtoul(buf, 10, &ratio); | |
5018 | if (err) | |
5019 | return err; | |
5020 | ||
5021 | if (ratio <= 100) | |
5022 | s->remote_node_defrag_ratio = ratio * 10; | |
5023 | ||
5024 | return length; | |
5025 | } | |
5026 | SLAB_ATTR(remote_node_defrag_ratio); | |
5027 | #endif | |
5028 | ||
5029 | #ifdef CONFIG_SLUB_STATS | |
5030 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) | |
5031 | { | |
5032 | unsigned long sum = 0; | |
5033 | int cpu; | |
5034 | int len; | |
5035 | int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); | |
5036 | ||
5037 | if (!data) | |
5038 | return -ENOMEM; | |
5039 | ||
5040 | for_each_online_cpu(cpu) { | |
5041 | unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; | |
5042 | ||
5043 | data[cpu] = x; | |
5044 | sum += x; | |
5045 | } | |
5046 | ||
5047 | len = sprintf(buf, "%lu", sum); | |
5048 | ||
5049 | #ifdef CONFIG_SMP | |
5050 | for_each_online_cpu(cpu) { | |
5051 | if (data[cpu] && len < PAGE_SIZE - 20) | |
5052 | len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); | |
5053 | } | |
5054 | #endif | |
5055 | kfree(data); | |
5056 | return len + sprintf(buf + len, "\n"); | |
5057 | } | |
5058 | ||
5059 | static void clear_stat(struct kmem_cache *s, enum stat_item si) | |
5060 | { | |
5061 | int cpu; | |
5062 | ||
5063 | for_each_online_cpu(cpu) | |
5064 | per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; | |
5065 | } | |
5066 | ||
5067 | #define STAT_ATTR(si, text) \ | |
5068 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ | |
5069 | { \ | |
5070 | return show_stat(s, buf, si); \ | |
5071 | } \ | |
5072 | static ssize_t text##_store(struct kmem_cache *s, \ | |
5073 | const char *buf, size_t length) \ | |
5074 | { \ | |
5075 | if (buf[0] != '0') \ | |
5076 | return -EINVAL; \ | |
5077 | clear_stat(s, si); \ | |
5078 | return length; \ | |
5079 | } \ | |
5080 | SLAB_ATTR(text); \ | |
5081 | ||
5082 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); | |
5083 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); | |
5084 | STAT_ATTR(FREE_FASTPATH, free_fastpath); | |
5085 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); | |
5086 | STAT_ATTR(FREE_FROZEN, free_frozen); | |
5087 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); | |
5088 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); | |
5089 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); | |
5090 | STAT_ATTR(ALLOC_SLAB, alloc_slab); | |
5091 | STAT_ATTR(ALLOC_REFILL, alloc_refill); | |
5092 | STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); | |
5093 | STAT_ATTR(FREE_SLAB, free_slab); | |
5094 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); | |
5095 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); | |
5096 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); | |
5097 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); | |
5098 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); | |
5099 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); | |
5100 | STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); | |
5101 | STAT_ATTR(ORDER_FALLBACK, order_fallback); | |
5102 | STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); | |
5103 | STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); | |
5104 | STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); | |
5105 | STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); | |
5106 | STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); | |
5107 | STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); | |
5108 | #endif | |
5109 | ||
5110 | static struct attribute *slab_attrs[] = { | |
5111 | &slab_size_attr.attr, | |
5112 | &object_size_attr.attr, | |
5113 | &objs_per_slab_attr.attr, | |
5114 | &order_attr.attr, | |
5115 | &min_partial_attr.attr, | |
5116 | &cpu_partial_attr.attr, | |
5117 | &objects_attr.attr, | |
5118 | &objects_partial_attr.attr, | |
5119 | &partial_attr.attr, | |
5120 | &cpu_slabs_attr.attr, | |
5121 | &ctor_attr.attr, | |
5122 | &aliases_attr.attr, | |
5123 | &align_attr.attr, | |
5124 | &hwcache_align_attr.attr, | |
5125 | &reclaim_account_attr.attr, | |
5126 | &destroy_by_rcu_attr.attr, | |
5127 | &shrink_attr.attr, | |
5128 | &reserved_attr.attr, | |
5129 | &slabs_cpu_partial_attr.attr, | |
5130 | #ifdef CONFIG_SLUB_DEBUG | |
5131 | &total_objects_attr.attr, | |
5132 | &slabs_attr.attr, | |
5133 | &sanity_checks_attr.attr, | |
5134 | &trace_attr.attr, | |
5135 | &red_zone_attr.attr, | |
5136 | &poison_attr.attr, | |
5137 | &store_user_attr.attr, | |
5138 | &validate_attr.attr, | |
5139 | &alloc_calls_attr.attr, | |
5140 | &free_calls_attr.attr, | |
5141 | #endif | |
5142 | #ifdef CONFIG_ZONE_DMA | |
5143 | &cache_dma_attr.attr, | |
5144 | #endif | |
5145 | #ifdef CONFIG_NUMA | |
5146 | &remote_node_defrag_ratio_attr.attr, | |
5147 | #endif | |
5148 | #ifdef CONFIG_SLUB_STATS | |
5149 | &alloc_fastpath_attr.attr, | |
5150 | &alloc_slowpath_attr.attr, | |
5151 | &free_fastpath_attr.attr, | |
5152 | &free_slowpath_attr.attr, | |
5153 | &free_frozen_attr.attr, | |
5154 | &free_add_partial_attr.attr, | |
5155 | &free_remove_partial_attr.attr, | |
5156 | &alloc_from_partial_attr.attr, | |
5157 | &alloc_slab_attr.attr, | |
5158 | &alloc_refill_attr.attr, | |
5159 | &alloc_node_mismatch_attr.attr, | |
5160 | &free_slab_attr.attr, | |
5161 | &cpuslab_flush_attr.attr, | |
5162 | &deactivate_full_attr.attr, | |
5163 | &deactivate_empty_attr.attr, | |
5164 | &deactivate_to_head_attr.attr, | |
5165 | &deactivate_to_tail_attr.attr, | |
5166 | &deactivate_remote_frees_attr.attr, | |
5167 | &deactivate_bypass_attr.attr, | |
5168 | &order_fallback_attr.attr, | |
5169 | &cmpxchg_double_fail_attr.attr, | |
5170 | &cmpxchg_double_cpu_fail_attr.attr, | |
5171 | &cpu_partial_alloc_attr.attr, | |
5172 | &cpu_partial_free_attr.attr, | |
5173 | &cpu_partial_node_attr.attr, | |
5174 | &cpu_partial_drain_attr.attr, | |
5175 | #endif | |
5176 | #ifdef CONFIG_FAILSLAB | |
5177 | &failslab_attr.attr, | |
5178 | #endif | |
5179 | ||
5180 | NULL | |
5181 | }; | |
5182 | ||
5183 | static struct attribute_group slab_attr_group = { | |
5184 | .attrs = slab_attrs, | |
5185 | }; | |
5186 | ||
5187 | static ssize_t slab_attr_show(struct kobject *kobj, | |
5188 | struct attribute *attr, | |
5189 | char *buf) | |
5190 | { | |
5191 | struct slab_attribute *attribute; | |
5192 | struct kmem_cache *s; | |
5193 | int err; | |
5194 | ||
5195 | attribute = to_slab_attr(attr); | |
5196 | s = to_slab(kobj); | |
5197 | ||
5198 | if (!attribute->show) | |
5199 | return -EIO; | |
5200 | ||
5201 | err = attribute->show(s, buf); | |
5202 | ||
5203 | return err; | |
5204 | } | |
5205 | ||
5206 | static ssize_t slab_attr_store(struct kobject *kobj, | |
5207 | struct attribute *attr, | |
5208 | const char *buf, size_t len) | |
5209 | { | |
5210 | struct slab_attribute *attribute; | |
5211 | struct kmem_cache *s; | |
5212 | int err; | |
5213 | ||
5214 | attribute = to_slab_attr(attr); | |
5215 | s = to_slab(kobj); | |
5216 | ||
5217 | if (!attribute->store) | |
5218 | return -EIO; | |
5219 | ||
5220 | err = attribute->store(s, buf, len); | |
5221 | ||
5222 | return err; | |
5223 | } | |
5224 | ||
5225 | static void kmem_cache_release(struct kobject *kobj) | |
5226 | { | |
5227 | struct kmem_cache *s = to_slab(kobj); | |
5228 | ||
5229 | kfree(s->name); | |
5230 | kfree(s); | |
5231 | } | |
5232 | ||
5233 | static const struct sysfs_ops slab_sysfs_ops = { | |
5234 | .show = slab_attr_show, | |
5235 | .store = slab_attr_store, | |
5236 | }; | |
5237 | ||
5238 | static struct kobj_type slab_ktype = { | |
5239 | .sysfs_ops = &slab_sysfs_ops, | |
5240 | .release = kmem_cache_release | |
5241 | }; | |
5242 | ||
5243 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
5244 | { | |
5245 | struct kobj_type *ktype = get_ktype(kobj); | |
5246 | ||
5247 | if (ktype == &slab_ktype) | |
5248 | return 1; | |
5249 | return 0; | |
5250 | } | |
5251 | ||
5252 | static const struct kset_uevent_ops slab_uevent_ops = { | |
5253 | .filter = uevent_filter, | |
5254 | }; | |
5255 | ||
5256 | static struct kset *slab_kset; | |
5257 | ||
5258 | #define ID_STR_LENGTH 64 | |
5259 | ||
5260 | /* Create a unique string id for a slab cache: | |
5261 | * | |
5262 | * Format :[flags-]size | |
5263 | */ | |
5264 | static char *create_unique_id(struct kmem_cache *s) | |
5265 | { | |
5266 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
5267 | char *p = name; | |
5268 | ||
5269 | BUG_ON(!name); | |
5270 | ||
5271 | *p++ = ':'; | |
5272 | /* | |
5273 | * First flags affecting slabcache operations. We will only | |
5274 | * get here for aliasable slabs so we do not need to support | |
5275 | * too many flags. The flags here must cover all flags that | |
5276 | * are matched during merging to guarantee that the id is | |
5277 | * unique. | |
5278 | */ | |
5279 | if (s->flags & SLAB_CACHE_DMA) | |
5280 | *p++ = 'd'; | |
5281 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
5282 | *p++ = 'a'; | |
5283 | if (s->flags & SLAB_DEBUG_FREE) | |
5284 | *p++ = 'F'; | |
5285 | if (!(s->flags & SLAB_NOTRACK)) | |
5286 | *p++ = 't'; | |
5287 | if (p != name + 1) | |
5288 | *p++ = '-'; | |
5289 | p += sprintf(p, "%07d", s->size); | |
5290 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
5291 | return name; | |
5292 | } | |
5293 | ||
5294 | static int sysfs_slab_add(struct kmem_cache *s) | |
5295 | { | |
5296 | int err; | |
5297 | const char *name; | |
5298 | int unmergeable; | |
5299 | ||
5300 | if (slab_state < SYSFS) | |
5301 | /* Defer until later */ | |
5302 | return 0; | |
5303 | ||
5304 | unmergeable = slab_unmergeable(s); | |
5305 | if (unmergeable) { | |
5306 | /* | |
5307 | * Slabcache can never be merged so we can use the name proper. | |
5308 | * This is typically the case for debug situations. In that | |
5309 | * case we can catch duplicate names easily. | |
5310 | */ | |
5311 | sysfs_remove_link(&slab_kset->kobj, s->name); | |
5312 | name = s->name; | |
5313 | } else { | |
5314 | /* | |
5315 | * Create a unique name for the slab as a target | |
5316 | * for the symlinks. | |
5317 | */ | |
5318 | name = create_unique_id(s); | |
5319 | } | |
5320 | ||
5321 | s->kobj.kset = slab_kset; | |
5322 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); | |
5323 | if (err) { | |
5324 | kobject_put(&s->kobj); | |
5325 | return err; | |
5326 | } | |
5327 | ||
5328 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
5329 | if (err) { | |
5330 | kobject_del(&s->kobj); | |
5331 | kobject_put(&s->kobj); | |
5332 | return err; | |
5333 | } | |
5334 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
5335 | if (!unmergeable) { | |
5336 | /* Setup first alias */ | |
5337 | sysfs_slab_alias(s, s->name); | |
5338 | kfree(name); | |
5339 | } | |
5340 | return 0; | |
5341 | } | |
5342 | ||
5343 | static void sysfs_slab_remove(struct kmem_cache *s) | |
5344 | { | |
5345 | if (slab_state < SYSFS) | |
5346 | /* | |
5347 | * Sysfs has not been setup yet so no need to remove the | |
5348 | * cache from sysfs. | |
5349 | */ | |
5350 | return; | |
5351 | ||
5352 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
5353 | kobject_del(&s->kobj); | |
5354 | kobject_put(&s->kobj); | |
5355 | } | |
5356 | ||
5357 | /* | |
5358 | * Need to buffer aliases during bootup until sysfs becomes | |
5359 | * available lest we lose that information. | |
5360 | */ | |
5361 | struct saved_alias { | |
5362 | struct kmem_cache *s; | |
5363 | const char *name; | |
5364 | struct saved_alias *next; | |
5365 | }; | |
5366 | ||
5367 | static struct saved_alias *alias_list; | |
5368 | ||
5369 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
5370 | { | |
5371 | struct saved_alias *al; | |
5372 | ||
5373 | if (slab_state == SYSFS) { | |
5374 | /* | |
5375 | * If we have a leftover link then remove it. | |
5376 | */ | |
5377 | sysfs_remove_link(&slab_kset->kobj, name); | |
5378 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); | |
5379 | } | |
5380 | ||
5381 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
5382 | if (!al) | |
5383 | return -ENOMEM; | |
5384 | ||
5385 | al->s = s; | |
5386 | al->name = name; | |
5387 | al->next = alias_list; | |
5388 | alias_list = al; | |
5389 | return 0; | |
5390 | } | |
5391 | ||
5392 | static int __init slab_sysfs_init(void) | |
5393 | { | |
5394 | struct kmem_cache *s; | |
5395 | int err; | |
5396 | ||
5397 | down_write(&slub_lock); | |
5398 | ||
5399 | slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); | |
5400 | if (!slab_kset) { | |
5401 | up_write(&slub_lock); | |
5402 | printk(KERN_ERR "Cannot register slab subsystem.\n"); | |
5403 | return -ENOSYS; | |
5404 | } | |
5405 | ||
5406 | slab_state = SYSFS; | |
5407 | ||
5408 | list_for_each_entry(s, &slab_caches, list) { | |
5409 | err = sysfs_slab_add(s); | |
5410 | if (err) | |
5411 | printk(KERN_ERR "SLUB: Unable to add boot slab %s" | |
5412 | " to sysfs\n", s->name); | |
5413 | } | |
5414 | ||
5415 | while (alias_list) { | |
5416 | struct saved_alias *al = alias_list; | |
5417 | ||
5418 | alias_list = alias_list->next; | |
5419 | err = sysfs_slab_alias(al->s, al->name); | |
5420 | if (err) | |
5421 | printk(KERN_ERR "SLUB: Unable to add boot slab alias" | |
5422 | " %s to sysfs\n", s->name); | |
5423 | kfree(al); | |
5424 | } | |
5425 | ||
5426 | up_write(&slub_lock); | |
5427 | resiliency_test(); | |
5428 | return 0; | |
5429 | } | |
5430 | ||
5431 | __initcall(slab_sysfs_init); | |
5432 | #endif /* CONFIG_SYSFS */ | |
5433 | ||
5434 | /* | |
5435 | * The /proc/slabinfo ABI | |
5436 | */ | |
5437 | #ifdef CONFIG_SLABINFO | |
5438 | static void print_slabinfo_header(struct seq_file *m) | |
5439 | { | |
5440 | seq_puts(m, "slabinfo - version: 2.1\n"); | |
5441 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " | |
5442 | "<objperslab> <pagesperslab>"); | |
5443 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
5444 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
5445 | seq_putc(m, '\n'); | |
5446 | } | |
5447 | ||
5448 | static void *s_start(struct seq_file *m, loff_t *pos) | |
5449 | { | |
5450 | loff_t n = *pos; | |
5451 | ||
5452 | down_read(&slub_lock); | |
5453 | if (!n) | |
5454 | print_slabinfo_header(m); | |
5455 | ||
5456 | return seq_list_start(&slab_caches, *pos); | |
5457 | } | |
5458 | ||
5459 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
5460 | { | |
5461 | return seq_list_next(p, &slab_caches, pos); | |
5462 | } | |
5463 | ||
5464 | static void s_stop(struct seq_file *m, void *p) | |
5465 | { | |
5466 | up_read(&slub_lock); | |
5467 | } | |
5468 | ||
5469 | static int s_show(struct seq_file *m, void *p) | |
5470 | { | |
5471 | unsigned long nr_partials = 0; | |
5472 | unsigned long nr_slabs = 0; | |
5473 | unsigned long nr_inuse = 0; | |
5474 | unsigned long nr_objs = 0; | |
5475 | unsigned long nr_free = 0; | |
5476 | struct kmem_cache *s; | |
5477 | int node; | |
5478 | ||
5479 | s = list_entry(p, struct kmem_cache, list); | |
5480 | ||
5481 | for_each_online_node(node) { | |
5482 | struct kmem_cache_node *n = get_node(s, node); | |
5483 | ||
5484 | if (!n) | |
5485 | continue; | |
5486 | ||
5487 | nr_partials += n->nr_partial; | |
5488 | nr_slabs += atomic_long_read(&n->nr_slabs); | |
5489 | nr_objs += atomic_long_read(&n->total_objects); | |
5490 | nr_free += count_partial(n, count_free); | |
5491 | } | |
5492 | ||
5493 | nr_inuse = nr_objs - nr_free; | |
5494 | ||
5495 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, | |
5496 | nr_objs, s->size, oo_objects(s->oo), | |
5497 | (1 << oo_order(s->oo))); | |
5498 | seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); | |
5499 | seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, | |
5500 | 0UL); | |
5501 | seq_putc(m, '\n'); | |
5502 | return 0; | |
5503 | } | |
5504 | ||
5505 | static const struct seq_operations slabinfo_op = { | |
5506 | .start = s_start, | |
5507 | .next = s_next, | |
5508 | .stop = s_stop, | |
5509 | .show = s_show, | |
5510 | }; | |
5511 | ||
5512 | static int slabinfo_open(struct inode *inode, struct file *file) | |
5513 | { | |
5514 | return seq_open(file, &slabinfo_op); | |
5515 | } | |
5516 | ||
5517 | static const struct file_operations proc_slabinfo_operations = { | |
5518 | .open = slabinfo_open, | |
5519 | .read = seq_read, | |
5520 | .llseek = seq_lseek, | |
5521 | .release = seq_release, | |
5522 | }; | |
5523 | ||
5524 | static int __init slab_proc_init(void) | |
5525 | { | |
5526 | proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations); | |
5527 | return 0; | |
5528 | } | |
5529 | module_init(slab_proc_init); | |
5530 | #endif /* CONFIG_SLABINFO */ |