Root/
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 | page->flags |= 1 << PG_slab; |
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 | * Lock slab, remove from the partial list and put the object into the |
1494 | * per cpu freelist. |
1495 | * |
1496 | * Returns a list of objects or NULL if it fails. |
1497 | * |
1498 | * Must hold list_lock. |
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 | do { |
1514 | freelist = page->freelist; |
1515 | counters = page->counters; |
1516 | new.counters = counters; |
1517 | if (mode) |
1518 | new.inuse = page->objects; |
1519 | |
1520 | VM_BUG_ON(new.frozen); |
1521 | new.frozen = 1; |
1522 | |
1523 | } while (!__cmpxchg_double_slab(s, page, |
1524 | freelist, counters, |
1525 | NULL, new.counters, |
1526 | "lock and freeze")); |
1527 | |
1528 | remove_partial(n, page); |
1529 | return freelist; |
1530 | } |
1531 | |
1532 | static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); |
1533 | |
1534 | /* |
1535 | * Try to allocate a partial slab from a specific node. |
1536 | */ |
1537 | static void *get_partial_node(struct kmem_cache *s, |
1538 | struct kmem_cache_node *n, struct kmem_cache_cpu *c) |
1539 | { |
1540 | struct page *page, *page2; |
1541 | void *object = NULL; |
1542 | |
1543 | /* |
1544 | * Racy check. If we mistakenly see no partial slabs then we |
1545 | * just allocate an empty slab. If we mistakenly try to get a |
1546 | * partial slab and there is none available then get_partials() |
1547 | * will return NULL. |
1548 | */ |
1549 | if (!n || !n->nr_partial) |
1550 | return NULL; |
1551 | |
1552 | spin_lock(&n->list_lock); |
1553 | list_for_each_entry_safe(page, page2, &n->partial, lru) { |
1554 | void *t = acquire_slab(s, n, page, object == NULL); |
1555 | int available; |
1556 | |
1557 | if (!t) |
1558 | break; |
1559 | |
1560 | if (!object) { |
1561 | c->page = page; |
1562 | c->node = page_to_nid(page); |
1563 | stat(s, ALLOC_FROM_PARTIAL); |
1564 | object = t; |
1565 | available = page->objects - page->inuse; |
1566 | } else { |
1567 | page->freelist = t; |
1568 | available = put_cpu_partial(s, page, 0); |
1569 | stat(s, CPU_PARTIAL_NODE); |
1570 | } |
1571 | if (kmem_cache_debug(s) || available > s->cpu_partial / 2) |
1572 | break; |
1573 | |
1574 | } |
1575 | spin_unlock(&n->list_lock); |
1576 | return object; |
1577 | } |
1578 | |
1579 | /* |
1580 | * Get a page from somewhere. Search in increasing NUMA distances. |
1581 | */ |
1582 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags, |
1583 | struct kmem_cache_cpu *c) |
1584 | { |
1585 | #ifdef CONFIG_NUMA |
1586 | struct zonelist *zonelist; |
1587 | struct zoneref *z; |
1588 | struct zone *zone; |
1589 | enum zone_type high_zoneidx = gfp_zone(flags); |
1590 | void *object; |
1591 | unsigned int cpuset_mems_cookie; |
1592 | |
1593 | /* |
1594 | * The defrag ratio allows a configuration of the tradeoffs between |
1595 | * inter node defragmentation and node local allocations. A lower |
1596 | * defrag_ratio increases the tendency to do local allocations |
1597 | * instead of attempting to obtain partial slabs from other nodes. |
1598 | * |
1599 | * If the defrag_ratio is set to 0 then kmalloc() always |
1600 | * returns node local objects. If the ratio is higher then kmalloc() |
1601 | * may return off node objects because partial slabs are obtained |
1602 | * from other nodes and filled up. |
1603 | * |
1604 | * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes |
1605 | * defrag_ratio = 1000) then every (well almost) allocation will |
1606 | * first attempt to defrag slab caches on other nodes. This means |
1607 | * scanning over all nodes to look for partial slabs which may be |
1608 | * expensive if we do it every time we are trying to find a slab |
1609 | * with available objects. |
1610 | */ |
1611 | if (!s->remote_node_defrag_ratio || |
1612 | get_cycles() % 1024 > s->remote_node_defrag_ratio) |
1613 | return NULL; |
1614 | |
1615 | do { |
1616 | cpuset_mems_cookie = get_mems_allowed(); |
1617 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); |
1618 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
1619 | struct kmem_cache_node *n; |
1620 | |
1621 | n = get_node(s, zone_to_nid(zone)); |
1622 | |
1623 | if (n && cpuset_zone_allowed_hardwall(zone, flags) && |
1624 | n->nr_partial > s->min_partial) { |
1625 | object = get_partial_node(s, n, c); |
1626 | if (object) { |
1627 | /* |
1628 | * Return the object even if |
1629 | * put_mems_allowed indicated that |
1630 | * the cpuset mems_allowed was |
1631 | * updated in parallel. It's a |
1632 | * harmless race between the alloc |
1633 | * and the cpuset update. |
1634 | */ |
1635 | put_mems_allowed(cpuset_mems_cookie); |
1636 | return object; |
1637 | } |
1638 | } |
1639 | } |
1640 | } while (!put_mems_allowed(cpuset_mems_cookie)); |
1641 | #endif |
1642 | return NULL; |
1643 | } |
1644 | |
1645 | /* |
1646 | * Get a partial page, lock it and return it. |
1647 | */ |
1648 | static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, |
1649 | struct kmem_cache_cpu *c) |
1650 | { |
1651 | void *object; |
1652 | int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; |
1653 | |
1654 | object = get_partial_node(s, get_node(s, searchnode), c); |
1655 | if (object || node != NUMA_NO_NODE) |
1656 | return object; |
1657 | |
1658 | return get_any_partial(s, flags, c); |
1659 | } |
1660 | |
1661 | #ifdef CONFIG_PREEMPT |
1662 | /* |
1663 | * Calculate the next globally unique transaction for disambiguiation |
1664 | * during cmpxchg. The transactions start with the cpu number and are then |
1665 | * incremented by CONFIG_NR_CPUS. |
1666 | */ |
1667 | #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) |
1668 | #else |
1669 | /* |
1670 | * No preemption supported therefore also no need to check for |
1671 | * different cpus. |
1672 | */ |
1673 | #define TID_STEP 1 |
1674 | #endif |
1675 | |
1676 | static inline unsigned long next_tid(unsigned long tid) |
1677 | { |
1678 | return tid + TID_STEP; |
1679 | } |
1680 | |
1681 | static inline unsigned int tid_to_cpu(unsigned long tid) |
1682 | { |
1683 | return tid % TID_STEP; |
1684 | } |
1685 | |
1686 | static inline unsigned long tid_to_event(unsigned long tid) |
1687 | { |
1688 | return tid / TID_STEP; |
1689 | } |
1690 | |
1691 | static inline unsigned int init_tid(int cpu) |
1692 | { |
1693 | return cpu; |
1694 | } |
1695 | |
1696 | static inline void note_cmpxchg_failure(const char *n, |
1697 | const struct kmem_cache *s, unsigned long tid) |
1698 | { |
1699 | #ifdef SLUB_DEBUG_CMPXCHG |
1700 | unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); |
1701 | |
1702 | printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name); |
1703 | |
1704 | #ifdef CONFIG_PREEMPT |
1705 | if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) |
1706 | printk("due to cpu change %d -> %d\n", |
1707 | tid_to_cpu(tid), tid_to_cpu(actual_tid)); |
1708 | else |
1709 | #endif |
1710 | if (tid_to_event(tid) != tid_to_event(actual_tid)) |
1711 | printk("due to cpu running other code. Event %ld->%ld\n", |
1712 | tid_to_event(tid), tid_to_event(actual_tid)); |
1713 | else |
1714 | printk("for unknown reason: actual=%lx was=%lx target=%lx\n", |
1715 | actual_tid, tid, next_tid(tid)); |
1716 | #endif |
1717 | stat(s, CMPXCHG_DOUBLE_CPU_FAIL); |
1718 | } |
1719 | |
1720 | void init_kmem_cache_cpus(struct kmem_cache *s) |
1721 | { |
1722 | int cpu; |
1723 | |
1724 | for_each_possible_cpu(cpu) |
1725 | per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); |
1726 | } |
1727 | |
1728 | /* |
1729 | * Remove the cpu slab |
1730 | */ |
1731 | static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
1732 | { |
1733 | enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; |
1734 | struct page *page = c->page; |
1735 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); |
1736 | int lock = 0; |
1737 | enum slab_modes l = M_NONE, m = M_NONE; |
1738 | void *freelist; |
1739 | void *nextfree; |
1740 | int tail = DEACTIVATE_TO_HEAD; |
1741 | struct page new; |
1742 | struct page old; |
1743 | |
1744 | if (page->freelist) { |
1745 | stat(s, DEACTIVATE_REMOTE_FREES); |
1746 | tail = DEACTIVATE_TO_TAIL; |
1747 | } |
1748 | |
1749 | c->tid = next_tid(c->tid); |
1750 | c->page = NULL; |
1751 | freelist = c->freelist; |
1752 | c->freelist = NULL; |
1753 | |
1754 | /* |
1755 | * Stage one: Free all available per cpu objects back |
1756 | * to the page freelist while it is still frozen. Leave the |
1757 | * last one. |
1758 | * |
1759 | * There is no need to take the list->lock because the page |
1760 | * is still frozen. |
1761 | */ |
1762 | while (freelist && (nextfree = get_freepointer(s, freelist))) { |
1763 | void *prior; |
1764 | unsigned long counters; |
1765 | |
1766 | do { |
1767 | prior = page->freelist; |
1768 | counters = page->counters; |
1769 | set_freepointer(s, freelist, prior); |
1770 | new.counters = counters; |
1771 | new.inuse--; |
1772 | VM_BUG_ON(!new.frozen); |
1773 | |
1774 | } while (!__cmpxchg_double_slab(s, page, |
1775 | prior, counters, |
1776 | freelist, new.counters, |
1777 | "drain percpu freelist")); |
1778 | |
1779 | freelist = nextfree; |
1780 | } |
1781 | |
1782 | /* |
1783 | * Stage two: Ensure that the page is unfrozen while the |
1784 | * list presence reflects the actual number of objects |
1785 | * during unfreeze. |
1786 | * |
1787 | * We setup the list membership and then perform a cmpxchg |
1788 | * with the count. If there is a mismatch then the page |
1789 | * is not unfrozen but the page is on the wrong list. |
1790 | * |
1791 | * Then we restart the process which may have to remove |
1792 | * the page from the list that we just put it on again |
1793 | * because the number of objects in the slab may have |
1794 | * changed. |
1795 | */ |
1796 | redo: |
1797 | |
1798 | old.freelist = page->freelist; |
1799 | old.counters = page->counters; |
1800 | VM_BUG_ON(!old.frozen); |
1801 | |
1802 | /* Determine target state of the slab */ |
1803 | new.counters = old.counters; |
1804 | if (freelist) { |
1805 | new.inuse--; |
1806 | set_freepointer(s, freelist, old.freelist); |
1807 | new.freelist = freelist; |
1808 | } else |
1809 | new.freelist = old.freelist; |
1810 | |
1811 | new.frozen = 0; |
1812 | |
1813 | if (!new.inuse && n->nr_partial > s->min_partial) |
1814 | m = M_FREE; |
1815 | else if (new.freelist) { |
1816 | m = M_PARTIAL; |
1817 | if (!lock) { |
1818 | lock = 1; |
1819 | /* |
1820 | * Taking the spinlock removes the possiblity |
1821 | * that acquire_slab() will see a slab page that |
1822 | * is frozen |
1823 | */ |
1824 | spin_lock(&n->list_lock); |
1825 | } |
1826 | } else { |
1827 | m = M_FULL; |
1828 | if (kmem_cache_debug(s) && !lock) { |
1829 | lock = 1; |
1830 | /* |
1831 | * This also ensures that the scanning of full |
1832 | * slabs from diagnostic functions will not see |
1833 | * any frozen slabs. |
1834 | */ |
1835 | spin_lock(&n->list_lock); |
1836 | } |
1837 | } |
1838 | |
1839 | if (l != m) { |
1840 | |
1841 | if (l == M_PARTIAL) |
1842 | |
1843 | remove_partial(n, page); |
1844 | |
1845 | else if (l == M_FULL) |
1846 | |
1847 | remove_full(s, page); |
1848 | |
1849 | if (m == M_PARTIAL) { |
1850 | |
1851 | add_partial(n, page, tail); |
1852 | stat(s, tail); |
1853 | |
1854 | } else if (m == M_FULL) { |
1855 | |
1856 | stat(s, DEACTIVATE_FULL); |
1857 | add_full(s, n, page); |
1858 | |
1859 | } |
1860 | } |
1861 | |
1862 | l = m; |
1863 | if (!__cmpxchg_double_slab(s, page, |
1864 | old.freelist, old.counters, |
1865 | new.freelist, new.counters, |
1866 | "unfreezing slab")) |
1867 | goto redo; |
1868 | |
1869 | if (lock) |
1870 | spin_unlock(&n->list_lock); |
1871 | |
1872 | if (m == M_FREE) { |
1873 | stat(s, DEACTIVATE_EMPTY); |
1874 | discard_slab(s, page); |
1875 | stat(s, FREE_SLAB); |
1876 | } |
1877 | } |
1878 | |
1879 | /* Unfreeze all the cpu partial slabs */ |
1880 | static void unfreeze_partials(struct kmem_cache *s) |
1881 | { |
1882 | struct kmem_cache_node *n = NULL; |
1883 | struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab); |
1884 | struct page *page, *discard_page = NULL; |
1885 | |
1886 | while ((page = c->partial)) { |
1887 | enum slab_modes { M_PARTIAL, M_FREE }; |
1888 | enum slab_modes l, m; |
1889 | struct page new; |
1890 | struct page old; |
1891 | |
1892 | c->partial = page->next; |
1893 | l = M_FREE; |
1894 | |
1895 | do { |
1896 | |
1897 | old.freelist = page->freelist; |
1898 | old.counters = page->counters; |
1899 | VM_BUG_ON(!old.frozen); |
1900 | |
1901 | new.counters = old.counters; |
1902 | new.freelist = old.freelist; |
1903 | |
1904 | new.frozen = 0; |
1905 | |
1906 | if (!new.inuse && (!n || n->nr_partial > s->min_partial)) |
1907 | m = M_FREE; |
1908 | else { |
1909 | struct kmem_cache_node *n2 = get_node(s, |
1910 | page_to_nid(page)); |
1911 | |
1912 | m = M_PARTIAL; |
1913 | if (n != n2) { |
1914 | if (n) |
1915 | spin_unlock(&n->list_lock); |
1916 | |
1917 | n = n2; |
1918 | spin_lock(&n->list_lock); |
1919 | } |
1920 | } |
1921 | |
1922 | if (l != m) { |
1923 | if (l == M_PARTIAL) { |
1924 | remove_partial(n, page); |
1925 | stat(s, FREE_REMOVE_PARTIAL); |
1926 | } else { |
1927 | add_partial(n, page, |
1928 | DEACTIVATE_TO_TAIL); |
1929 | stat(s, FREE_ADD_PARTIAL); |
1930 | } |
1931 | |
1932 | l = m; |
1933 | } |
1934 | |
1935 | } while (!cmpxchg_double_slab(s, page, |
1936 | old.freelist, old.counters, |
1937 | new.freelist, new.counters, |
1938 | "unfreezing slab")); |
1939 | |
1940 | if (m == M_FREE) { |
1941 | page->next = discard_page; |
1942 | discard_page = page; |
1943 | } |
1944 | } |
1945 | |
1946 | if (n) |
1947 | spin_unlock(&n->list_lock); |
1948 | |
1949 | while (discard_page) { |
1950 | page = discard_page; |
1951 | discard_page = discard_page->next; |
1952 | |
1953 | stat(s, DEACTIVATE_EMPTY); |
1954 | discard_slab(s, page); |
1955 | stat(s, FREE_SLAB); |
1956 | } |
1957 | } |
1958 | |
1959 | /* |
1960 | * Put a page that was just frozen (in __slab_free) into a partial page |
1961 | * slot if available. This is done without interrupts disabled and without |
1962 | * preemption disabled. The cmpxchg is racy and may put the partial page |
1963 | * onto a random cpus partial slot. |
1964 | * |
1965 | * If we did not find a slot then simply move all the partials to the |
1966 | * per node partial list. |
1967 | */ |
1968 | int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) |
1969 | { |
1970 | struct page *oldpage; |
1971 | int pages; |
1972 | int pobjects; |
1973 | |
1974 | do { |
1975 | pages = 0; |
1976 | pobjects = 0; |
1977 | oldpage = this_cpu_read(s->cpu_slab->partial); |
1978 | |
1979 | if (oldpage) { |
1980 | pobjects = oldpage->pobjects; |
1981 | pages = oldpage->pages; |
1982 | if (drain && pobjects > s->cpu_partial) { |
1983 | unsigned long flags; |
1984 | /* |
1985 | * partial array is full. Move the existing |
1986 | * set to the per node partial list. |
1987 | */ |
1988 | local_irq_save(flags); |
1989 | unfreeze_partials(s); |
1990 | local_irq_restore(flags); |
1991 | pobjects = 0; |
1992 | pages = 0; |
1993 | stat(s, CPU_PARTIAL_DRAIN); |
1994 | } |
1995 | } |
1996 | |
1997 | pages++; |
1998 | pobjects += page->objects - page->inuse; |
1999 | |
2000 | page->pages = pages; |
2001 | page->pobjects = pobjects; |
2002 | page->next = oldpage; |
2003 | |
2004 | } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage); |
2005 | return pobjects; |
2006 | } |
2007 | |
2008 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
2009 | { |
2010 | stat(s, CPUSLAB_FLUSH); |
2011 | deactivate_slab(s, c); |
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 kmem_cache_cpu *c, int node) |
2056 | { |
2057 | #ifdef CONFIG_NUMA |
2058 | if (node != NUMA_NO_NODE && c->node != 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 *object; |
2131 | struct kmem_cache_cpu *c; |
2132 | struct page *page = new_slab(s, flags, node); |
2133 | |
2134 | if (page) { |
2135 | c = __this_cpu_ptr(s->cpu_slab); |
2136 | if (c->page) |
2137 | flush_slab(s, c); |
2138 | |
2139 | /* |
2140 | * No other reference to the page yet so we can |
2141 | * muck around with it freely without cmpxchg |
2142 | */ |
2143 | object = page->freelist; |
2144 | page->freelist = NULL; |
2145 | |
2146 | stat(s, ALLOC_SLAB); |
2147 | c->node = page_to_nid(page); |
2148 | c->page = page; |
2149 | *pc = c; |
2150 | } else |
2151 | object = NULL; |
2152 | |
2153 | return object; |
2154 | } |
2155 | |
2156 | /* |
2157 | * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist |
2158 | * or deactivate the page. |
2159 | * |
2160 | * The page is still frozen if the return value is not NULL. |
2161 | * |
2162 | * If this function returns NULL then the page has been unfrozen. |
2163 | */ |
2164 | static inline void *get_freelist(struct kmem_cache *s, struct page *page) |
2165 | { |
2166 | struct page new; |
2167 | unsigned long counters; |
2168 | void *freelist; |
2169 | |
2170 | do { |
2171 | freelist = page->freelist; |
2172 | counters = page->counters; |
2173 | new.counters = counters; |
2174 | VM_BUG_ON(!new.frozen); |
2175 | |
2176 | new.inuse = page->objects; |
2177 | new.frozen = freelist != NULL; |
2178 | |
2179 | } while (!cmpxchg_double_slab(s, page, |
2180 | freelist, counters, |
2181 | NULL, new.counters, |
2182 | "get_freelist")); |
2183 | |
2184 | return freelist; |
2185 | } |
2186 | |
2187 | /* |
2188 | * Slow path. The lockless freelist is empty or we need to perform |
2189 | * debugging duties. |
2190 | * |
2191 | * Processing is still very fast if new objects have been freed to the |
2192 | * regular freelist. In that case we simply take over the regular freelist |
2193 | * as the lockless freelist and zap the regular freelist. |
2194 | * |
2195 | * If that is not working then we fall back to the partial lists. We take the |
2196 | * first element of the freelist as the object to allocate now and move the |
2197 | * rest of the freelist to the lockless freelist. |
2198 | * |
2199 | * And if we were unable to get a new slab from the partial slab lists then |
2200 | * we need to allocate a new slab. This is the slowest path since it involves |
2201 | * a call to the page allocator and the setup of a new slab. |
2202 | */ |
2203 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
2204 | unsigned long addr, struct kmem_cache_cpu *c) |
2205 | { |
2206 | void **object; |
2207 | unsigned long flags; |
2208 | |
2209 | local_irq_save(flags); |
2210 | #ifdef CONFIG_PREEMPT |
2211 | /* |
2212 | * We may have been preempted and rescheduled on a different |
2213 | * cpu before disabling interrupts. Need to reload cpu area |
2214 | * pointer. |
2215 | */ |
2216 | c = this_cpu_ptr(s->cpu_slab); |
2217 | #endif |
2218 | |
2219 | if (!c->page) |
2220 | goto new_slab; |
2221 | redo: |
2222 | if (unlikely(!node_match(c, node))) { |
2223 | stat(s, ALLOC_NODE_MISMATCH); |
2224 | deactivate_slab(s, c); |
2225 | goto new_slab; |
2226 | } |
2227 | |
2228 | /* must check again c->freelist in case of cpu migration or IRQ */ |
2229 | object = c->freelist; |
2230 | if (object) |
2231 | goto load_freelist; |
2232 | |
2233 | stat(s, ALLOC_SLOWPATH); |
2234 | |
2235 | object = get_freelist(s, c->page); |
2236 | |
2237 | if (!object) { |
2238 | c->page = NULL; |
2239 | stat(s, DEACTIVATE_BYPASS); |
2240 | goto new_slab; |
2241 | } |
2242 | |
2243 | stat(s, ALLOC_REFILL); |
2244 | |
2245 | load_freelist: |
2246 | c->freelist = get_freepointer(s, object); |
2247 | c->tid = next_tid(c->tid); |
2248 | local_irq_restore(flags); |
2249 | return object; |
2250 | |
2251 | new_slab: |
2252 | |
2253 | if (c->partial) { |
2254 | c->page = c->partial; |
2255 | c->partial = c->page->next; |
2256 | c->node = page_to_nid(c->page); |
2257 | stat(s, CPU_PARTIAL_ALLOC); |
2258 | c->freelist = NULL; |
2259 | goto redo; |
2260 | } |
2261 | |
2262 | /* Then do expensive stuff like retrieving pages from the partial lists */ |
2263 | object = get_partial(s, gfpflags, node, c); |
2264 | |
2265 | if (unlikely(!object)) { |
2266 | |
2267 | object = new_slab_objects(s, gfpflags, node, &c); |
2268 | |
2269 | if (unlikely(!object)) { |
2270 | if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) |
2271 | slab_out_of_memory(s, gfpflags, node); |
2272 | |
2273 | local_irq_restore(flags); |
2274 | return NULL; |
2275 | } |
2276 | } |
2277 | |
2278 | if (likely(!kmem_cache_debug(s))) |
2279 | goto load_freelist; |
2280 | |
2281 | /* Only entered in the debug case */ |
2282 | if (!alloc_debug_processing(s, c->page, object, addr)) |
2283 | goto new_slab; /* Slab failed checks. Next slab needed */ |
2284 | |
2285 | c->freelist = get_freepointer(s, object); |
2286 | deactivate_slab(s, c); |
2287 | c->node = NUMA_NO_NODE; |
2288 | local_irq_restore(flags); |
2289 | return object; |
2290 | } |
2291 | |
2292 | /* |
2293 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) |
2294 | * have the fastpath folded into their functions. So no function call |
2295 | * overhead for requests that can be satisfied on the fastpath. |
2296 | * |
2297 | * The fastpath works by first checking if the lockless freelist can be used. |
2298 | * If not then __slab_alloc is called for slow processing. |
2299 | * |
2300 | * Otherwise we can simply pick the next object from the lockless free list. |
2301 | */ |
2302 | static __always_inline void *slab_alloc(struct kmem_cache *s, |
2303 | gfp_t gfpflags, int node, unsigned long addr) |
2304 | { |
2305 | void **object; |
2306 | struct kmem_cache_cpu *c; |
2307 | unsigned long tid; |
2308 | |
2309 | if (slab_pre_alloc_hook(s, gfpflags)) |
2310 | return NULL; |
2311 | |
2312 | redo: |
2313 | |
2314 | /* |
2315 | * Must read kmem_cache cpu data via this cpu ptr. Preemption is |
2316 | * enabled. We may switch back and forth between cpus while |
2317 | * reading from one cpu area. That does not matter as long |
2318 | * as we end up on the original cpu again when doing the cmpxchg. |
2319 | */ |
2320 | c = __this_cpu_ptr(s->cpu_slab); |
2321 | |
2322 | /* |
2323 | * The transaction ids are globally unique per cpu and per operation on |
2324 | * a per cpu queue. Thus they can be guarantee that the cmpxchg_double |
2325 | * occurs on the right processor and that there was no operation on the |
2326 | * linked list in between. |
2327 | */ |
2328 | tid = c->tid; |
2329 | barrier(); |
2330 | |
2331 | object = c->freelist; |
2332 | if (unlikely(!object || !node_match(c, node))) |
2333 | |
2334 | object = __slab_alloc(s, gfpflags, node, addr, c); |
2335 | |
2336 | else { |
2337 | void *next_object = get_freepointer_safe(s, object); |
2338 | |
2339 | /* |
2340 | * The cmpxchg will only match if there was no additional |
2341 | * operation and if we are on the right processor. |
2342 | * |
2343 | * The cmpxchg does the following atomically (without lock semantics!) |
2344 | * 1. Relocate first pointer to the current per cpu area. |
2345 | * 2. Verify that tid and freelist have not been changed |
2346 | * 3. If they were not changed replace tid and freelist |
2347 | * |
2348 | * Since this is without lock semantics the protection is only against |
2349 | * code executing on this cpu *not* from access by other cpus. |
2350 | */ |
2351 | if (unlikely(!this_cpu_cmpxchg_double( |
2352 | s->cpu_slab->freelist, s->cpu_slab->tid, |
2353 | object, tid, |
2354 | next_object, next_tid(tid)))) { |
2355 | |
2356 | note_cmpxchg_failure("slab_alloc", s, tid); |
2357 | goto redo; |
2358 | } |
2359 | prefetch_freepointer(s, next_object); |
2360 | stat(s, ALLOC_FASTPATH); |
2361 | } |
2362 | |
2363 | if (unlikely(gfpflags & __GFP_ZERO) && object) |
2364 | memset(object, 0, s->objsize); |
2365 | |
2366 | slab_post_alloc_hook(s, gfpflags, object); |
2367 | |
2368 | return object; |
2369 | } |
2370 | |
2371 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) |
2372 | { |
2373 | void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); |
2374 | |
2375 | trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); |
2376 | |
2377 | return ret; |
2378 | } |
2379 | EXPORT_SYMBOL(kmem_cache_alloc); |
2380 | |
2381 | #ifdef CONFIG_TRACING |
2382 | void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) |
2383 | { |
2384 | void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); |
2385 | trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); |
2386 | return ret; |
2387 | } |
2388 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
2389 | |
2390 | void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) |
2391 | { |
2392 | void *ret = kmalloc_order(size, flags, order); |
2393 | trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags); |
2394 | return ret; |
2395 | } |
2396 | EXPORT_SYMBOL(kmalloc_order_trace); |
2397 | #endif |
2398 | |
2399 | #ifdef CONFIG_NUMA |
2400 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) |
2401 | { |
2402 | void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); |
2403 | |
2404 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
2405 | s->objsize, s->size, gfpflags, node); |
2406 | |
2407 | return ret; |
2408 | } |
2409 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
2410 | |
2411 | #ifdef CONFIG_TRACING |
2412 | void *kmem_cache_alloc_node_trace(struct kmem_cache *s, |
2413 | gfp_t gfpflags, |
2414 | int node, size_t size) |
2415 | { |
2416 | void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); |
2417 | |
2418 | trace_kmalloc_node(_RET_IP_, ret, |
2419 | size, s->size, gfpflags, node); |
2420 | return ret; |
2421 | } |
2422 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
2423 | #endif |
2424 | #endif |
2425 | |
2426 | /* |
2427 | * Slow patch handling. This may still be called frequently since objects |
2428 | * have a longer lifetime than the cpu slabs in most processing loads. |
2429 | * |
2430 | * So we still attempt to reduce cache line usage. Just take the slab |
2431 | * lock and free the item. If there is no additional partial page |
2432 | * handling required then we can return immediately. |
2433 | */ |
2434 | static void __slab_free(struct kmem_cache *s, struct page *page, |
2435 | void *x, unsigned long addr) |
2436 | { |
2437 | void *prior; |
2438 | void **object = (void *)x; |
2439 | int was_frozen; |
2440 | int inuse; |
2441 | struct page new; |
2442 | unsigned long counters; |
2443 | struct kmem_cache_node *n = NULL; |
2444 | unsigned long uninitialized_var(flags); |
2445 | |
2446 | stat(s, FREE_SLOWPATH); |
2447 | |
2448 | if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr)) |
2449 | return; |
2450 | |
2451 | do { |
2452 | prior = page->freelist; |
2453 | counters = page->counters; |
2454 | set_freepointer(s, object, prior); |
2455 | new.counters = counters; |
2456 | was_frozen = new.frozen; |
2457 | new.inuse--; |
2458 | if ((!new.inuse || !prior) && !was_frozen && !n) { |
2459 | |
2460 | if (!kmem_cache_debug(s) && !prior) |
2461 | |
2462 | /* |
2463 | * Slab was on no list before and will be partially empty |
2464 | * We can defer the list move and instead freeze it. |
2465 | */ |
2466 | new.frozen = 1; |
2467 | |
2468 | else { /* Needs to be taken off a list */ |
2469 | |
2470 | n = get_node(s, page_to_nid(page)); |
2471 | /* |
2472 | * Speculatively acquire the list_lock. |
2473 | * If the cmpxchg does not succeed then we may |
2474 | * drop the list_lock without any processing. |
2475 | * |
2476 | * Otherwise the list_lock will synchronize with |
2477 | * other processors updating the list of slabs. |
2478 | */ |
2479 | spin_lock_irqsave(&n->list_lock, flags); |
2480 | |
2481 | } |
2482 | } |
2483 | inuse = new.inuse; |
2484 | |
2485 | } while (!cmpxchg_double_slab(s, page, |
2486 | prior, counters, |
2487 | object, new.counters, |
2488 | "__slab_free")); |
2489 | |
2490 | if (likely(!n)) { |
2491 | |
2492 | /* |
2493 | * If we just froze the page then put it onto the |
2494 | * per cpu partial list. |
2495 | */ |
2496 | if (new.frozen && !was_frozen) { |
2497 | put_cpu_partial(s, page, 1); |
2498 | stat(s, CPU_PARTIAL_FREE); |
2499 | } |
2500 | /* |
2501 | * The list lock was not taken therefore no list |
2502 | * activity can be necessary. |
2503 | */ |
2504 | if (was_frozen) |
2505 | stat(s, FREE_FROZEN); |
2506 | return; |
2507 | } |
2508 | |
2509 | /* |
2510 | * was_frozen may have been set after we acquired the list_lock in |
2511 | * an earlier loop. So we need to check it here again. |
2512 | */ |
2513 | if (was_frozen) |
2514 | stat(s, FREE_FROZEN); |
2515 | else { |
2516 | if (unlikely(!inuse && n->nr_partial > s->min_partial)) |
2517 | goto slab_empty; |
2518 | |
2519 | /* |
2520 | * Objects left in the slab. If it was not on the partial list before |
2521 | * then add it. |
2522 | */ |
2523 | if (unlikely(!prior)) { |
2524 | remove_full(s, page); |
2525 | add_partial(n, page, DEACTIVATE_TO_TAIL); |
2526 | stat(s, FREE_ADD_PARTIAL); |
2527 | } |
2528 | } |
2529 | spin_unlock_irqrestore(&n->list_lock, flags); |
2530 | return; |
2531 | |
2532 | slab_empty: |
2533 | if (prior) { |
2534 | /* |
2535 | * Slab on the partial list. |
2536 | */ |
2537 | remove_partial(n, page); |
2538 | stat(s, FREE_REMOVE_PARTIAL); |
2539 | } else |
2540 | /* Slab must be on the full list */ |
2541 | remove_full(s, page); |
2542 | |
2543 | spin_unlock_irqrestore(&n->list_lock, flags); |
2544 | stat(s, FREE_SLAB); |
2545 | discard_slab(s, page); |
2546 | } |
2547 | |
2548 | /* |
2549 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that |
2550 | * can perform fastpath freeing without additional function calls. |
2551 | * |
2552 | * The fastpath is only possible if we are freeing to the current cpu slab |
2553 | * of this processor. This typically the case if we have just allocated |
2554 | * the item before. |
2555 | * |
2556 | * If fastpath is not possible then fall back to __slab_free where we deal |
2557 | * with all sorts of special processing. |
2558 | */ |
2559 | static __always_inline void slab_free(struct kmem_cache *s, |
2560 | struct page *page, void *x, unsigned long addr) |
2561 | { |
2562 | void **object = (void *)x; |
2563 | struct kmem_cache_cpu *c; |
2564 | unsigned long tid; |
2565 | |
2566 | slab_free_hook(s, x); |
2567 | |
2568 | redo: |
2569 | /* |
2570 | * Determine the currently cpus per cpu slab. |
2571 | * The cpu may change afterward. However that does not matter since |
2572 | * data is retrieved via this pointer. If we are on the same cpu |
2573 | * during the cmpxchg then the free will succedd. |
2574 | */ |
2575 | c = __this_cpu_ptr(s->cpu_slab); |
2576 | |
2577 | tid = c->tid; |
2578 | barrier(); |
2579 | |
2580 | if (likely(page == c->page)) { |
2581 | set_freepointer(s, object, c->freelist); |
2582 | |
2583 | if (unlikely(!this_cpu_cmpxchg_double( |
2584 | s->cpu_slab->freelist, s->cpu_slab->tid, |
2585 | c->freelist, tid, |
2586 | object, next_tid(tid)))) { |
2587 | |
2588 | note_cmpxchg_failure("slab_free", s, tid); |
2589 | goto redo; |
2590 | } |
2591 | stat(s, FREE_FASTPATH); |
2592 | } else |
2593 | __slab_free(s, page, x, addr); |
2594 | |
2595 | } |
2596 | |
2597 | void kmem_cache_free(struct kmem_cache *s, void *x) |
2598 | { |
2599 | struct page *page; |
2600 | |
2601 | page = virt_to_head_page(x); |
2602 | |
2603 | slab_free(s, page, x, _RET_IP_); |
2604 | |
2605 | trace_kmem_cache_free(_RET_IP_, x); |
2606 | } |
2607 | EXPORT_SYMBOL(kmem_cache_free); |
2608 | |
2609 | /* |
2610 | * Object placement in a slab is made very easy because we always start at |
2611 | * offset 0. If we tune the size of the object to the alignment then we can |
2612 | * get the required alignment by putting one properly sized object after |
2613 | * another. |
2614 | * |
2615 | * Notice that the allocation order determines the sizes of the per cpu |
2616 | * caches. Each processor has always one slab available for allocations. |
2617 | * Increasing the allocation order reduces the number of times that slabs |
2618 | * must be moved on and off the partial lists and is therefore a factor in |
2619 | * locking overhead. |
2620 | */ |
2621 | |
2622 | /* |
2623 | * Mininum / Maximum order of slab pages. This influences locking overhead |
2624 | * and slab fragmentation. A higher order reduces the number of partial slabs |
2625 | * and increases the number of allocations possible without having to |
2626 | * take the list_lock. |
2627 | */ |
2628 | static int slub_min_order; |
2629 | static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; |
2630 | static int slub_min_objects; |
2631 | |
2632 | /* |
2633 | * Merge control. If this is set then no merging of slab caches will occur. |
2634 | * (Could be removed. This was introduced to pacify the merge skeptics.) |
2635 | */ |
2636 | static int slub_nomerge; |
2637 | |
2638 | /* |
2639 | * Calculate the order of allocation given an slab object size. |
2640 | * |
2641 | * The order of allocation has significant impact on performance and other |
2642 | * system components. Generally order 0 allocations should be preferred since |
2643 | * order 0 does not cause fragmentation in the page allocator. Larger objects |
2644 | * be problematic to put into order 0 slabs because there may be too much |
2645 | * unused space left. We go to a higher order if more than 1/16th of the slab |
2646 | * would be wasted. |
2647 | * |
2648 | * In order to reach satisfactory performance we must ensure that a minimum |
2649 | * number of objects is in one slab. Otherwise we may generate too much |
2650 | * activity on the partial lists which requires taking the list_lock. This is |
2651 | * less a concern for large slabs though which are rarely used. |
2652 | * |
2653 | * slub_max_order specifies the order where we begin to stop considering the |
2654 | * number of objects in a slab as critical. If we reach slub_max_order then |
2655 | * we try to keep the page order as low as possible. So we accept more waste |
2656 | * of space in favor of a small page order. |
2657 | * |
2658 | * Higher order allocations also allow the placement of more objects in a |
2659 | * slab and thereby reduce object handling overhead. If the user has |
2660 | * requested a higher mininum order then we start with that one instead of |
2661 | * the smallest order which will fit the object. |
2662 | */ |
2663 | static inline int slab_order(int size, int min_objects, |
2664 | int max_order, int fract_leftover, int reserved) |
2665 | { |
2666 | int order; |
2667 | int rem; |
2668 | int min_order = slub_min_order; |
2669 | |
2670 | if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) |
2671 | return get_order(size * MAX_OBJS_PER_PAGE) - 1; |
2672 | |
2673 | for (order = max(min_order, |
2674 | fls(min_objects * size - 1) - PAGE_SHIFT); |
2675 | order <= max_order; order++) { |
2676 | |
2677 | unsigned long slab_size = PAGE_SIZE << order; |
2678 | |
2679 | if (slab_size < min_objects * size + reserved) |
2680 | continue; |
2681 | |
2682 | rem = (slab_size - reserved) % size; |
2683 | |
2684 | if (rem <= slab_size / fract_leftover) |
2685 | break; |
2686 | |
2687 | } |
2688 | |
2689 | return order; |
2690 | } |
2691 | |
2692 | static inline int calculate_order(int size, int reserved) |
2693 | { |
2694 | int order; |
2695 | int min_objects; |
2696 | int fraction; |
2697 | int max_objects; |
2698 | |
2699 | /* |
2700 | * Attempt to find best configuration for a slab. This |
2701 | * works by first attempting to generate a layout with |
2702 | * the best configuration and backing off gradually. |
2703 | * |
2704 | * First we reduce the acceptable waste in a slab. Then |
2705 | * we reduce the minimum objects required in a slab. |
2706 | */ |
2707 | min_objects = slub_min_objects; |
2708 | if (!min_objects) |
2709 | min_objects = 4 * (fls(nr_cpu_ids) + 1); |
2710 | max_objects = order_objects(slub_max_order, size, reserved); |
2711 | min_objects = min(min_objects, max_objects); |
2712 | |
2713 | while (min_objects > 1) { |
2714 | fraction = 16; |
2715 | while (fraction >= 4) { |
2716 | order = slab_order(size, min_objects, |
2717 | slub_max_order, fraction, reserved); |
2718 | if (order <= slub_max_order) |
2719 | return order; |
2720 | fraction /= 2; |
2721 | } |
2722 | min_objects--; |
2723 | } |
2724 | |
2725 | /* |
2726 | * We were unable to place multiple objects in a slab. Now |
2727 | * lets see if we can place a single object there. |
2728 | */ |
2729 | order = slab_order(size, 1, slub_max_order, 1, reserved); |
2730 | if (order <= slub_max_order) |
2731 | return order; |
2732 | |
2733 | /* |
2734 | * Doh this slab cannot be placed using slub_max_order. |
2735 | */ |
2736 | order = slab_order(size, 1, MAX_ORDER, 1, reserved); |
2737 | if (order < MAX_ORDER) |
2738 | return order; |
2739 | return -ENOSYS; |
2740 | } |
2741 | |
2742 | /* |
2743 | * Figure out what the alignment of the objects will be. |
2744 | */ |
2745 | static unsigned long calculate_alignment(unsigned long flags, |
2746 | unsigned long align, unsigned long size) |
2747 | { |
2748 | /* |
2749 | * If the user wants hardware cache aligned objects then follow that |
2750 | * suggestion if the object is sufficiently large. |
2751 | * |
2752 | * The hardware cache alignment cannot override the specified |
2753 | * alignment though. If that is greater then use it. |
2754 | */ |
2755 | if (flags & SLAB_HWCACHE_ALIGN) { |
2756 | unsigned long ralign = cache_line_size(); |
2757 | while (size <= ralign / 2) |
2758 | ralign /= 2; |
2759 | align = max(align, ralign); |
2760 | } |
2761 | |
2762 | if (align < ARCH_SLAB_MINALIGN) |
2763 | align = ARCH_SLAB_MINALIGN; |
2764 | |
2765 | return ALIGN(align, sizeof(void *)); |
2766 | } |
2767 | |
2768 | static void |
2769 | init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) |
2770 | { |
2771 | n->nr_partial = 0; |
2772 | spin_lock_init(&n->list_lock); |
2773 | INIT_LIST_HEAD(&n->partial); |
2774 | #ifdef CONFIG_SLUB_DEBUG |
2775 | atomic_long_set(&n->nr_slabs, 0); |
2776 | atomic_long_set(&n->total_objects, 0); |
2777 | INIT_LIST_HEAD(&n->full); |
2778 | #endif |
2779 | } |
2780 | |
2781 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
2782 | { |
2783 | BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < |
2784 | SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu)); |
2785 | |
2786 | /* |
2787 | * Must align to double word boundary for the double cmpxchg |
2788 | * instructions to work; see __pcpu_double_call_return_bool(). |
2789 | */ |
2790 | s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), |
2791 | 2 * sizeof(void *)); |
2792 | |
2793 | if (!s->cpu_slab) |
2794 | return 0; |
2795 | |
2796 | init_kmem_cache_cpus(s); |
2797 | |
2798 | return 1; |
2799 | } |
2800 | |
2801 | static struct kmem_cache *kmem_cache_node; |
2802 | |
2803 | /* |
2804 | * No kmalloc_node yet so do it by hand. We know that this is the first |
2805 | * slab on the node for this slabcache. There are no concurrent accesses |
2806 | * possible. |
2807 | * |
2808 | * Note that this function only works on the kmalloc_node_cache |
2809 | * when allocating for the kmalloc_node_cache. This is used for bootstrapping |
2810 | * memory on a fresh node that has no slab structures yet. |
2811 | */ |
2812 | static void early_kmem_cache_node_alloc(int node) |
2813 | { |
2814 | struct page *page; |
2815 | struct kmem_cache_node *n; |
2816 | |
2817 | BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); |
2818 | |
2819 | page = new_slab(kmem_cache_node, GFP_NOWAIT, node); |
2820 | |
2821 | BUG_ON(!page); |
2822 | if (page_to_nid(page) != node) { |
2823 | printk(KERN_ERR "SLUB: Unable to allocate memory from " |
2824 | "node %d\n", node); |
2825 | printk(KERN_ERR "SLUB: Allocating a useless per node structure " |
2826 | "in order to be able to continue\n"); |
2827 | } |
2828 | |
2829 | n = page->freelist; |
2830 | BUG_ON(!n); |
2831 | page->freelist = get_freepointer(kmem_cache_node, n); |
2832 | page->inuse = 1; |
2833 | page->frozen = 0; |
2834 | kmem_cache_node->node[node] = n; |
2835 | #ifdef CONFIG_SLUB_DEBUG |
2836 | init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); |
2837 | init_tracking(kmem_cache_node, n); |
2838 | #endif |
2839 | init_kmem_cache_node(n, kmem_cache_node); |
2840 | inc_slabs_node(kmem_cache_node, node, page->objects); |
2841 | |
2842 | add_partial(n, page, DEACTIVATE_TO_HEAD); |
2843 | } |
2844 | |
2845 | static void free_kmem_cache_nodes(struct kmem_cache *s) |
2846 | { |
2847 | int node; |
2848 | |
2849 | for_each_node_state(node, N_NORMAL_MEMORY) { |
2850 | struct kmem_cache_node *n = s->node[node]; |
2851 | |
2852 | if (n) |
2853 | kmem_cache_free(kmem_cache_node, n); |
2854 | |
2855 | s->node[node] = NULL; |
2856 | } |
2857 | } |
2858 | |
2859 | static int init_kmem_cache_nodes(struct kmem_cache *s) |
2860 | { |
2861 | int node; |
2862 | |
2863 | for_each_node_state(node, N_NORMAL_MEMORY) { |
2864 | struct kmem_cache_node *n; |
2865 | |
2866 | if (slab_state == DOWN) { |
2867 | early_kmem_cache_node_alloc(node); |
2868 | continue; |
2869 | } |
2870 | n = kmem_cache_alloc_node(kmem_cache_node, |
2871 | GFP_KERNEL, node); |
2872 | |
2873 | if (!n) { |
2874 | free_kmem_cache_nodes(s); |
2875 | return 0; |
2876 | } |
2877 | |
2878 | s->node[node] = n; |
2879 | init_kmem_cache_node(n, s); |
2880 | } |
2881 | return 1; |
2882 | } |
2883 | |
2884 | static void set_min_partial(struct kmem_cache *s, unsigned long min) |
2885 | { |
2886 | if (min < MIN_PARTIAL) |
2887 | min = MIN_PARTIAL; |
2888 | else if (min > MAX_PARTIAL) |
2889 | min = MAX_PARTIAL; |
2890 | s->min_partial = min; |
2891 | } |
2892 | |
2893 | /* |
2894 | * calculate_sizes() determines the order and the distribution of data within |
2895 | * a slab object. |
2896 | */ |
2897 | static int calculate_sizes(struct kmem_cache *s, int forced_order) |
2898 | { |
2899 | unsigned long flags = s->flags; |
2900 | unsigned long size = s->objsize; |
2901 | unsigned long align = s->align; |
2902 | int order; |
2903 | |
2904 | /* |
2905 | * Round up object size to the next word boundary. We can only |
2906 | * place the free pointer at word boundaries and this determines |
2907 | * the possible location of the free pointer. |
2908 | */ |
2909 | size = ALIGN(size, sizeof(void *)); |
2910 | |
2911 | #ifdef CONFIG_SLUB_DEBUG |
2912 | /* |
2913 | * Determine if we can poison the object itself. If the user of |
2914 | * the slab may touch the object after free or before allocation |
2915 | * then we should never poison the object itself. |
2916 | */ |
2917 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && |
2918 | !s->ctor) |
2919 | s->flags |= __OBJECT_POISON; |
2920 | else |
2921 | s->flags &= ~__OBJECT_POISON; |
2922 | |
2923 | |
2924 | /* |
2925 | * If we are Redzoning then check if there is some space between the |
2926 | * end of the object and the free pointer. If not then add an |
2927 | * additional word to have some bytes to store Redzone information. |
2928 | */ |
2929 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) |
2930 | size += sizeof(void *); |
2931 | #endif |
2932 | |
2933 | /* |
2934 | * With that we have determined the number of bytes in actual use |
2935 | * by the object. This is the potential offset to the free pointer. |
2936 | */ |
2937 | s->inuse = size; |
2938 | |
2939 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || |
2940 | s->ctor)) { |
2941 | /* |
2942 | * Relocate free pointer after the object if it is not |
2943 | * permitted to overwrite the first word of the object on |
2944 | * kmem_cache_free. |
2945 | * |
2946 | * This is the case if we do RCU, have a constructor or |
2947 | * destructor or are poisoning the objects. |
2948 | */ |
2949 | s->offset = size; |
2950 | size += sizeof(void *); |
2951 | } |
2952 | |
2953 | #ifdef CONFIG_SLUB_DEBUG |
2954 | if (flags & SLAB_STORE_USER) |
2955 | /* |
2956 | * Need to store information about allocs and frees after |
2957 | * the object. |
2958 | */ |
2959 | size += 2 * sizeof(struct track); |
2960 | |
2961 | if (flags & SLAB_RED_ZONE) |
2962 | /* |
2963 | * Add some empty padding so that we can catch |
2964 | * overwrites from earlier objects rather than let |
2965 | * tracking information or the free pointer be |
2966 | * corrupted if a user writes before the start |
2967 | * of the object. |
2968 | */ |
2969 | size += sizeof(void *); |
2970 | #endif |
2971 | |
2972 | /* |
2973 | * Determine the alignment based on various parameters that the |
2974 | * user specified and the dynamic determination of cache line size |
2975 | * on bootup. |
2976 | */ |
2977 | align = calculate_alignment(flags, align, s->objsize); |
2978 | s->align = align; |
2979 | |
2980 | /* |
2981 | * SLUB stores one object immediately after another beginning from |
2982 | * offset 0. In order to align the objects we have to simply size |
2983 | * each object to conform to the alignment. |
2984 | */ |
2985 | size = ALIGN(size, align); |
2986 | s->size = size; |
2987 | if (forced_order >= 0) |
2988 | order = forced_order; |
2989 | else |
2990 | order = calculate_order(size, s->reserved); |
2991 | |
2992 | if (order < 0) |
2993 | return 0; |
2994 | |
2995 | s->allocflags = 0; |
2996 | if (order) |
2997 | s->allocflags |= __GFP_COMP; |
2998 | |
2999 | if (s->flags & SLAB_CACHE_DMA) |
3000 | s->allocflags |= SLUB_DMA; |
3001 | |
3002 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
3003 | s->allocflags |= __GFP_RECLAIMABLE; |
3004 | |
3005 | /* |
3006 | * Determine the number of objects per slab |
3007 | */ |
3008 | s->oo = oo_make(order, size, s->reserved); |
3009 | s->min = oo_make(get_order(size), size, s->reserved); |
3010 | if (oo_objects(s->oo) > oo_objects(s->max)) |
3011 | s->max = s->oo; |
3012 | |
3013 | return !!oo_objects(s->oo); |
3014 | |
3015 | } |
3016 | |
3017 | static int kmem_cache_open(struct kmem_cache *s, |
3018 | const char *name, size_t size, |
3019 | size_t align, unsigned long flags, |
3020 | void (*ctor)(void *)) |
3021 | { |
3022 | memset(s, 0, kmem_size); |
3023 | s->name = name; |
3024 | s->ctor = ctor; |
3025 | s->objsize = size; |
3026 | s->align = align; |
3027 | s->flags = kmem_cache_flags(size, flags, name, ctor); |
3028 | s->reserved = 0; |
3029 | |
3030 | if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) |
3031 | s->reserved = sizeof(struct rcu_head); |
3032 | |
3033 | if (!calculate_sizes(s, -1)) |
3034 | goto error; |
3035 | if (disable_higher_order_debug) { |
3036 | /* |
3037 | * Disable debugging flags that store metadata if the min slab |
3038 | * order increased. |
3039 | */ |
3040 | if (get_order(s->size) > get_order(s->objsize)) { |
3041 | s->flags &= ~DEBUG_METADATA_FLAGS; |
3042 | s->offset = 0; |
3043 | if (!calculate_sizes(s, -1)) |
3044 | goto error; |
3045 | } |
3046 | } |
3047 | |
3048 | #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ |
3049 | defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) |
3050 | if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0) |
3051 | /* Enable fast mode */ |
3052 | s->flags |= __CMPXCHG_DOUBLE; |
3053 | #endif |
3054 | |
3055 | /* |
3056 | * The larger the object size is, the more pages we want on the partial |
3057 | * list to avoid pounding the page allocator excessively. |
3058 | */ |
3059 | set_min_partial(s, ilog2(s->size) / 2); |
3060 | |
3061 | /* |
3062 | * cpu_partial determined the maximum number of objects kept in the |
3063 | * per cpu partial lists of a processor. |
3064 | * |
3065 | * Per cpu partial lists mainly contain slabs that just have one |
3066 | * object freed. If they are used for allocation then they can be |
3067 | * filled up again with minimal effort. The slab will never hit the |
3068 | * per node partial lists and therefore no locking will be required. |
3069 | * |
3070 | * This setting also determines |
3071 | * |
3072 | * A) The number of objects from per cpu partial slabs dumped to the |
3073 | * per node list when we reach the limit. |
3074 | * B) The number of objects in cpu partial slabs to extract from the |
3075 | * per node list when we run out of per cpu objects. We only fetch 50% |
3076 | * to keep some capacity around for frees. |
3077 | */ |
3078 | if (kmem_cache_debug(s)) |
3079 | s->cpu_partial = 0; |
3080 | else if (s->size >= PAGE_SIZE) |
3081 | s->cpu_partial = 2; |
3082 | else if (s->size >= 1024) |
3083 | s->cpu_partial = 6; |
3084 | else if (s->size >= 256) |
3085 | s->cpu_partial = 13; |
3086 | else |
3087 | s->cpu_partial = 30; |
3088 | |
3089 | s->refcount = 1; |
3090 | #ifdef CONFIG_NUMA |
3091 | s->remote_node_defrag_ratio = 1000; |
3092 | #endif |
3093 | if (!init_kmem_cache_nodes(s)) |
3094 | goto error; |
3095 | |
3096 | if (alloc_kmem_cache_cpus(s)) |
3097 | return 1; |
3098 | |
3099 | free_kmem_cache_nodes(s); |
3100 | error: |
3101 | if (flags & SLAB_PANIC) |
3102 | panic("Cannot create slab %s size=%lu realsize=%u " |
3103 | "order=%u offset=%u flags=%lx\n", |
3104 | s->name, (unsigned long)size, s->size, oo_order(s->oo), |
3105 | s->offset, flags); |
3106 | return 0; |
3107 | } |
3108 | |
3109 | /* |
3110 | * Determine the size of a slab object |
3111 | */ |
3112 | unsigned int kmem_cache_size(struct kmem_cache *s) |
3113 | { |
3114 | return s->objsize; |
3115 | } |
3116 | EXPORT_SYMBOL(kmem_cache_size); |
3117 | |
3118 | static void list_slab_objects(struct kmem_cache *s, struct page *page, |
3119 | const char *text) |
3120 | { |
3121 | #ifdef CONFIG_SLUB_DEBUG |
3122 | void *addr = page_address(page); |
3123 | void *p; |
3124 | unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * |
3125 | sizeof(long), GFP_ATOMIC); |
3126 | if (!map) |
3127 | return; |
3128 | slab_err(s, page, "%s", text); |
3129 | slab_lock(page); |
3130 | |
3131 | get_map(s, page, map); |
3132 | for_each_object(p, s, addr, page->objects) { |
3133 | |
3134 | if (!test_bit(slab_index(p, s, addr), map)) { |
3135 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", |
3136 | p, p - addr); |
3137 | print_tracking(s, p); |
3138 | } |
3139 | } |
3140 | slab_unlock(page); |
3141 | kfree(map); |
3142 | #endif |
3143 | } |
3144 | |
3145 | /* |
3146 | * Attempt to free all partial slabs on a node. |
3147 | * This is called from kmem_cache_close(). We must be the last thread |
3148 | * using the cache and therefore we do not need to lock anymore. |
3149 | */ |
3150 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) |
3151 | { |
3152 | struct page *page, *h; |
3153 | |
3154 | list_for_each_entry_safe(page, h, &n->partial, lru) { |
3155 | if (!page->inuse) { |
3156 | remove_partial(n, page); |
3157 | discard_slab(s, page); |
3158 | } else { |
3159 | list_slab_objects(s, page, |
3160 | "Objects remaining on kmem_cache_close()"); |
3161 | } |
3162 | } |
3163 | } |
3164 | |
3165 | /* |
3166 | * Release all resources used by a slab cache. |
3167 | */ |
3168 | static inline int kmem_cache_close(struct kmem_cache *s) |
3169 | { |
3170 | int node; |
3171 | |
3172 | flush_all(s); |
3173 | free_percpu(s->cpu_slab); |
3174 | /* Attempt to free all objects */ |
3175 | for_each_node_state(node, N_NORMAL_MEMORY) { |
3176 | struct kmem_cache_node *n = get_node(s, node); |
3177 | |
3178 | free_partial(s, n); |
3179 | if (n->nr_partial || slabs_node(s, node)) |
3180 | return 1; |
3181 | } |
3182 | free_kmem_cache_nodes(s); |
3183 | return 0; |
3184 | } |
3185 | |
3186 | /* |
3187 | * Close a cache and release the kmem_cache structure |
3188 | * (must be used for caches created using kmem_cache_create) |
3189 | */ |
3190 | void kmem_cache_destroy(struct kmem_cache *s) |
3191 | { |
3192 | down_write(&slub_lock); |
3193 | s->refcount--; |
3194 | if (!s->refcount) { |
3195 | list_del(&s->list); |
3196 | up_write(&slub_lock); |
3197 | if (kmem_cache_close(s)) { |
3198 | printk(KERN_ERR "SLUB %s: %s called for cache that " |
3199 | "still has objects.\n", s->name, __func__); |
3200 | dump_stack(); |
3201 | } |
3202 | if (s->flags & SLAB_DESTROY_BY_RCU) |
3203 | rcu_barrier(); |
3204 | sysfs_slab_remove(s); |
3205 | } else |
3206 | up_write(&slub_lock); |
3207 | } |
3208 | EXPORT_SYMBOL(kmem_cache_destroy); |
3209 | |
3210 | /******************************************************************** |
3211 | * Kmalloc subsystem |
3212 | *******************************************************************/ |
3213 | |
3214 | struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT]; |
3215 | EXPORT_SYMBOL(kmalloc_caches); |
3216 | |
3217 | static struct kmem_cache *kmem_cache; |
3218 | |
3219 | #ifdef CONFIG_ZONE_DMA |
3220 | static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT]; |
3221 | #endif |
3222 | |
3223 | static int __init setup_slub_min_order(char *str) |
3224 | { |
3225 | get_option(&str, &slub_min_order); |
3226 | |
3227 | return 1; |
3228 | } |
3229 | |
3230 | __setup("slub_min_order=", setup_slub_min_order); |
3231 | |
3232 | static int __init setup_slub_max_order(char *str) |
3233 | { |
3234 | get_option(&str, &slub_max_order); |
3235 | slub_max_order = min(slub_max_order, MAX_ORDER - 1); |
3236 | |
3237 | return 1; |
3238 | } |
3239 | |
3240 | __setup("slub_max_order=", setup_slub_max_order); |
3241 | |
3242 | static int __init setup_slub_min_objects(char *str) |
3243 | { |
3244 | get_option(&str, &slub_min_objects); |
3245 | |
3246 | return 1; |
3247 | } |
3248 | |
3249 | __setup("slub_min_objects=", setup_slub_min_objects); |
3250 | |
3251 | static int __init setup_slub_nomerge(char *str) |
3252 | { |
3253 | slub_nomerge = 1; |
3254 | return 1; |
3255 | } |
3256 | |
3257 | __setup("slub_nomerge", setup_slub_nomerge); |
3258 | |
3259 | static struct kmem_cache *__init create_kmalloc_cache(const char *name, |
3260 | int size, unsigned int flags) |
3261 | { |
3262 | struct kmem_cache *s; |
3263 | |
3264 | s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); |
3265 | |
3266 | /* |
3267 | * This function is called with IRQs disabled during early-boot on |
3268 | * single CPU so there's no need to take slub_lock here. |
3269 | */ |
3270 | if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN, |
3271 | flags, NULL)) |
3272 | goto panic; |
3273 | |
3274 | list_add(&s->list, &slab_caches); |
3275 | return s; |
3276 | |
3277 | panic: |
3278 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); |
3279 | return NULL; |
3280 | } |
3281 | |
3282 | /* |
3283 | * Conversion table for small slabs sizes / 8 to the index in the |
3284 | * kmalloc array. This is necessary for slabs < 192 since we have non power |
3285 | * of two cache sizes there. The size of larger slabs can be determined using |
3286 | * fls. |
3287 | */ |
3288 | static s8 size_index[24] = { |
3289 | 3, /* 8 */ |
3290 | 4, /* 16 */ |
3291 | 5, /* 24 */ |
3292 | 5, /* 32 */ |
3293 | 6, /* 40 */ |
3294 | 6, /* 48 */ |
3295 | 6, /* 56 */ |
3296 | 6, /* 64 */ |
3297 | 1, /* 72 */ |
3298 | 1, /* 80 */ |
3299 | 1, /* 88 */ |
3300 | 1, /* 96 */ |
3301 | 7, /* 104 */ |
3302 | 7, /* 112 */ |
3303 | 7, /* 120 */ |
3304 | 7, /* 128 */ |
3305 | 2, /* 136 */ |
3306 | 2, /* 144 */ |
3307 | 2, /* 152 */ |
3308 | 2, /* 160 */ |
3309 | 2, /* 168 */ |
3310 | 2, /* 176 */ |
3311 | 2, /* 184 */ |
3312 | 2 /* 192 */ |
3313 | }; |
3314 | |
3315 | static inline int size_index_elem(size_t bytes) |
3316 | { |
3317 | return (bytes - 1) / 8; |
3318 | } |
3319 | |
3320 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) |
3321 | { |
3322 | int index; |
3323 | |
3324 | if (size <= 192) { |
3325 | if (!size) |
3326 | return ZERO_SIZE_PTR; |
3327 | |
3328 | index = size_index[size_index_elem(size)]; |
3329 | } else |
3330 | index = fls(size - 1); |
3331 | |
3332 | #ifdef CONFIG_ZONE_DMA |
3333 | if (unlikely((flags & SLUB_DMA))) |
3334 | return kmalloc_dma_caches[index]; |
3335 | |
3336 | #endif |
3337 | return kmalloc_caches[index]; |
3338 | } |
3339 | |
3340 | void *__kmalloc(size_t size, gfp_t flags) |
3341 | { |
3342 | struct kmem_cache *s; |
3343 | void *ret; |
3344 | |
3345 | if (unlikely(size > SLUB_MAX_SIZE)) |
3346 | return kmalloc_large(size, flags); |
3347 | |
3348 | s = get_slab(size, flags); |
3349 | |
3350 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
3351 | return s; |
3352 | |
3353 | ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); |
3354 | |
3355 | trace_kmalloc(_RET_IP_, ret, size, s->size, flags); |
3356 | |
3357 | return ret; |
3358 | } |
3359 | EXPORT_SYMBOL(__kmalloc); |
3360 | |
3361 | #ifdef CONFIG_NUMA |
3362 | static void *kmalloc_large_node(size_t size, gfp_t flags, int node) |
3363 | { |
3364 | struct page *page; |
3365 | void *ptr = NULL; |
3366 | |
3367 | flags |= __GFP_COMP | __GFP_NOTRACK; |
3368 | page = alloc_pages_node(node, flags, get_order(size)); |
3369 | if (page) |
3370 | ptr = page_address(page); |
3371 | |
3372 | kmemleak_alloc(ptr, size, 1, flags); |
3373 | return ptr; |
3374 | } |
3375 | |
3376 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3377 | { |
3378 | struct kmem_cache *s; |
3379 | void *ret; |
3380 | |
3381 | if (unlikely(size > SLUB_MAX_SIZE)) { |
3382 | ret = kmalloc_large_node(size, flags, node); |
3383 | |
3384 | trace_kmalloc_node(_RET_IP_, ret, |
3385 | size, PAGE_SIZE << get_order(size), |
3386 | flags, node); |
3387 | |
3388 | return ret; |
3389 | } |
3390 | |
3391 | s = get_slab(size, flags); |
3392 | |
3393 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
3394 | return s; |
3395 | |
3396 | ret = slab_alloc(s, flags, node, _RET_IP_); |
3397 | |
3398 | trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); |
3399 | |
3400 | return ret; |
3401 | } |
3402 | EXPORT_SYMBOL(__kmalloc_node); |
3403 | #endif |
3404 | |
3405 | size_t ksize(const void *object) |
3406 | { |
3407 | struct page *page; |
3408 | |
3409 | if (unlikely(object == ZERO_SIZE_PTR)) |
3410 | return 0; |
3411 | |
3412 | page = virt_to_head_page(object); |
3413 | |
3414 | if (unlikely(!PageSlab(page))) { |
3415 | WARN_ON(!PageCompound(page)); |
3416 | return PAGE_SIZE << compound_order(page); |
3417 | } |
3418 | |
3419 | return slab_ksize(page->slab); |
3420 | } |
3421 | EXPORT_SYMBOL(ksize); |
3422 | |
3423 | #ifdef CONFIG_SLUB_DEBUG |
3424 | bool verify_mem_not_deleted(const void *x) |
3425 | { |
3426 | struct page *page; |
3427 | void *object = (void *)x; |
3428 | unsigned long flags; |
3429 | bool rv; |
3430 | |
3431 | if (unlikely(ZERO_OR_NULL_PTR(x))) |
3432 | return false; |
3433 | |
3434 | local_irq_save(flags); |
3435 | |
3436 | page = virt_to_head_page(x); |
3437 | if (unlikely(!PageSlab(page))) { |
3438 | /* maybe it was from stack? */ |
3439 | rv = true; |
3440 | goto out_unlock; |
3441 | } |
3442 | |
3443 | slab_lock(page); |
3444 | if (on_freelist(page->slab, page, object)) { |
3445 | object_err(page->slab, page, object, "Object is on free-list"); |
3446 | rv = false; |
3447 | } else { |
3448 | rv = true; |
3449 | } |
3450 | slab_unlock(page); |
3451 | |
3452 | out_unlock: |
3453 | local_irq_restore(flags); |
3454 | return rv; |
3455 | } |
3456 | EXPORT_SYMBOL(verify_mem_not_deleted); |
3457 | #endif |
3458 | |
3459 | void kfree(const void *x) |
3460 | { |
3461 | struct page *page; |
3462 | void *object = (void *)x; |
3463 | |
3464 | trace_kfree(_RET_IP_, x); |
3465 | |
3466 | if (unlikely(ZERO_OR_NULL_PTR(x))) |
3467 | return; |
3468 | |
3469 | page = virt_to_head_page(x); |
3470 | if (unlikely(!PageSlab(page))) { |
3471 | BUG_ON(!PageCompound(page)); |
3472 | kmemleak_free(x); |
3473 | put_page(page); |
3474 | return; |
3475 | } |
3476 | slab_free(page->slab, page, object, _RET_IP_); |
3477 | } |
3478 | EXPORT_SYMBOL(kfree); |
3479 | |
3480 | /* |
3481 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts |
3482 | * the remaining slabs by the number of items in use. The slabs with the |
3483 | * most items in use come first. New allocations will then fill those up |
3484 | * and thus they can be removed from the partial lists. |
3485 | * |
3486 | * The slabs with the least items are placed last. This results in them |
3487 | * being allocated from last increasing the chance that the last objects |
3488 | * are freed in them. |
3489 | */ |
3490 | int kmem_cache_shrink(struct kmem_cache *s) |
3491 | { |
3492 | int node; |
3493 | int i; |
3494 | struct kmem_cache_node *n; |
3495 | struct page *page; |
3496 | struct page *t; |
3497 | int objects = oo_objects(s->max); |
3498 | struct list_head *slabs_by_inuse = |
3499 | kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); |
3500 | unsigned long flags; |
3501 | |
3502 | if (!slabs_by_inuse) |
3503 | return -ENOMEM; |
3504 | |
3505 | flush_all(s); |
3506 | for_each_node_state(node, N_NORMAL_MEMORY) { |
3507 | n = get_node(s, node); |
3508 | |
3509 | if (!n->nr_partial) |
3510 | continue; |
3511 | |
3512 | for (i = 0; i < objects; i++) |
3513 | INIT_LIST_HEAD(slabs_by_inuse + i); |
3514 | |
3515 | spin_lock_irqsave(&n->list_lock, flags); |
3516 | |
3517 | /* |
3518 | * Build lists indexed by the items in use in each slab. |
3519 | * |
3520 | * Note that concurrent frees may occur while we hold the |
3521 | * list_lock. page->inuse here is the upper limit. |
3522 | */ |
3523 | list_for_each_entry_safe(page, t, &n->partial, lru) { |
3524 | list_move(&page->lru, slabs_by_inuse + page->inuse); |
3525 | if (!page->inuse) |
3526 | n->nr_partial--; |
3527 | } |
3528 | |
3529 | /* |
3530 | * Rebuild the partial list with the slabs filled up most |
3531 | * first and the least used slabs at the end. |
3532 | */ |
3533 | for (i = objects - 1; i > 0; i--) |
3534 | list_splice(slabs_by_inuse + i, n->partial.prev); |
3535 | |
3536 | spin_unlock_irqrestore(&n->list_lock, flags); |
3537 | |
3538 | /* Release empty slabs */ |
3539 | list_for_each_entry_safe(page, t, slabs_by_inuse, lru) |
3540 | discard_slab(s, page); |
3541 | } |
3542 | |
3543 | kfree(slabs_by_inuse); |
3544 | return 0; |
3545 | } |
3546 | EXPORT_SYMBOL(kmem_cache_shrink); |
3547 | |
3548 | #if defined(CONFIG_MEMORY_HOTPLUG) |
3549 | static int slab_mem_going_offline_callback(void *arg) |
3550 | { |
3551 | struct kmem_cache *s; |
3552 | |
3553 | down_read(&slub_lock); |
3554 | list_for_each_entry(s, &slab_caches, list) |
3555 | kmem_cache_shrink(s); |
3556 | up_read(&slub_lock); |
3557 | |
3558 | return 0; |
3559 | } |
3560 | |
3561 | static void slab_mem_offline_callback(void *arg) |
3562 | { |
3563 | struct kmem_cache_node *n; |
3564 | struct kmem_cache *s; |
3565 | struct memory_notify *marg = arg; |
3566 | int offline_node; |
3567 | |
3568 | offline_node = marg->status_change_nid; |
3569 | |
3570 | /* |
3571 | * If the node still has available memory. we need kmem_cache_node |
3572 | * for it yet. |
3573 | */ |
3574 | if (offline_node < 0) |
3575 | return; |
3576 | |
3577 | down_read(&slub_lock); |
3578 | list_for_each_entry(s, &slab_caches, list) { |
3579 | n = get_node(s, offline_node); |
3580 | if (n) { |
3581 | /* |
3582 | * if n->nr_slabs > 0, slabs still exist on the node |
3583 | * that is going down. We were unable to free them, |
3584 | * and offline_pages() function shouldn't call this |
3585 | * callback. So, we must fail. |
3586 | */ |
3587 | BUG_ON(slabs_node(s, offline_node)); |
3588 | |
3589 | s->node[offline_node] = NULL; |
3590 | kmem_cache_free(kmem_cache_node, n); |
3591 | } |
3592 | } |
3593 | up_read(&slub_lock); |
3594 | } |
3595 | |
3596 | static int slab_mem_going_online_callback(void *arg) |
3597 | { |
3598 | struct kmem_cache_node *n; |
3599 | struct kmem_cache *s; |
3600 | struct memory_notify *marg = arg; |
3601 | int nid = marg->status_change_nid; |
3602 | int ret = 0; |
3603 | |
3604 | /* |
3605 | * If the node's memory is already available, then kmem_cache_node is |
3606 | * already created. Nothing to do. |
3607 | */ |
3608 | if (nid < 0) |
3609 | return 0; |
3610 | |
3611 | /* |
3612 | * We are bringing a node online. No memory is available yet. We must |
3613 | * allocate a kmem_cache_node structure in order to bring the node |
3614 | * online. |
3615 | */ |
3616 | down_read(&slub_lock); |
3617 | list_for_each_entry(s, &slab_caches, list) { |
3618 | /* |
3619 | * XXX: kmem_cache_alloc_node will fallback to other nodes |
3620 | * since memory is not yet available from the node that |
3621 | * is brought up. |
3622 | */ |
3623 | n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); |
3624 | if (!n) { |
3625 | ret = -ENOMEM; |
3626 | goto out; |
3627 | } |
3628 | init_kmem_cache_node(n, s); |
3629 | s->node[nid] = n; |
3630 | } |
3631 | out: |
3632 | up_read(&slub_lock); |
3633 | return ret; |
3634 | } |
3635 | |
3636 | static int slab_memory_callback(struct notifier_block *self, |
3637 | unsigned long action, void *arg) |
3638 | { |
3639 | int ret = 0; |
3640 | |
3641 | switch (action) { |
3642 | case MEM_GOING_ONLINE: |
3643 | ret = slab_mem_going_online_callback(arg); |
3644 | break; |
3645 | case MEM_GOING_OFFLINE: |
3646 | ret = slab_mem_going_offline_callback(arg); |
3647 | break; |
3648 | case MEM_OFFLINE: |
3649 | case MEM_CANCEL_ONLINE: |
3650 | slab_mem_offline_callback(arg); |
3651 | break; |
3652 | case MEM_ONLINE: |
3653 | case MEM_CANCEL_OFFLINE: |
3654 | break; |
3655 | } |
3656 | if (ret) |
3657 | ret = notifier_from_errno(ret); |
3658 | else |
3659 | ret = NOTIFY_OK; |
3660 | return ret; |
3661 | } |
3662 | |
3663 | #endif /* CONFIG_MEMORY_HOTPLUG */ |
3664 | |
3665 | /******************************************************************** |
3666 | * Basic setup of slabs |
3667 | *******************************************************************/ |
3668 | |
3669 | /* |
3670 | * Used for early kmem_cache structures that were allocated using |
3671 | * the page allocator |
3672 | */ |
3673 | |
3674 | static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s) |
3675 | { |
3676 | int node; |
3677 | |
3678 | list_add(&s->list, &slab_caches); |
3679 | s->refcount = -1; |
3680 | |
3681 | for_each_node_state(node, N_NORMAL_MEMORY) { |
3682 | struct kmem_cache_node *n = get_node(s, node); |
3683 | struct page *p; |
3684 | |
3685 | if (n) { |
3686 | list_for_each_entry(p, &n->partial, lru) |
3687 | p->slab = s; |
3688 | |
3689 | #ifdef CONFIG_SLUB_DEBUG |
3690 | list_for_each_entry(p, &n->full, lru) |
3691 | p->slab = s; |
3692 | #endif |
3693 | } |
3694 | } |
3695 | } |
3696 | |
3697 | void __init kmem_cache_init(void) |
3698 | { |
3699 | int i; |
3700 | int caches = 0; |
3701 | struct kmem_cache *temp_kmem_cache; |
3702 | int order; |
3703 | struct kmem_cache *temp_kmem_cache_node; |
3704 | unsigned long kmalloc_size; |
3705 | |
3706 | if (debug_guardpage_minorder()) |
3707 | slub_max_order = 0; |
3708 | |
3709 | kmem_size = offsetof(struct kmem_cache, node) + |
3710 | nr_node_ids * sizeof(struct kmem_cache_node *); |
3711 | |
3712 | /* Allocate two kmem_caches from the page allocator */ |
3713 | kmalloc_size = ALIGN(kmem_size, cache_line_size()); |
3714 | order = get_order(2 * kmalloc_size); |
3715 | kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order); |
3716 | |
3717 | /* |
3718 | * Must first have the slab cache available for the allocations of the |
3719 | * struct kmem_cache_node's. There is special bootstrap code in |
3720 | * kmem_cache_open for slab_state == DOWN. |
3721 | */ |
3722 | kmem_cache_node = (void *)kmem_cache + kmalloc_size; |
3723 | |
3724 | kmem_cache_open(kmem_cache_node, "kmem_cache_node", |
3725 | sizeof(struct kmem_cache_node), |
3726 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); |
3727 | |
3728 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); |
3729 | |
3730 | /* Able to allocate the per node structures */ |
3731 | slab_state = PARTIAL; |
3732 | |
3733 | temp_kmem_cache = kmem_cache; |
3734 | kmem_cache_open(kmem_cache, "kmem_cache", kmem_size, |
3735 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); |
3736 | kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); |
3737 | memcpy(kmem_cache, temp_kmem_cache, kmem_size); |
3738 | |
3739 | /* |
3740 | * Allocate kmem_cache_node properly from the kmem_cache slab. |
3741 | * kmem_cache_node is separately allocated so no need to |
3742 | * update any list pointers. |
3743 | */ |
3744 | temp_kmem_cache_node = kmem_cache_node; |
3745 | |
3746 | kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); |
3747 | memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size); |
3748 | |
3749 | kmem_cache_bootstrap_fixup(kmem_cache_node); |
3750 | |
3751 | caches++; |
3752 | kmem_cache_bootstrap_fixup(kmem_cache); |
3753 | caches++; |
3754 | /* Free temporary boot structure */ |
3755 | free_pages((unsigned long)temp_kmem_cache, order); |
3756 | |
3757 | /* Now we can use the kmem_cache to allocate kmalloc slabs */ |
3758 | |
3759 | /* |
3760 | * Patch up the size_index table if we have strange large alignment |
3761 | * requirements for the kmalloc array. This is only the case for |
3762 | * MIPS it seems. The standard arches will not generate any code here. |
3763 | * |
3764 | * Largest permitted alignment is 256 bytes due to the way we |
3765 | * handle the index determination for the smaller caches. |
3766 | * |
3767 | * Make sure that nothing crazy happens if someone starts tinkering |
3768 | * around with ARCH_KMALLOC_MINALIGN |
3769 | */ |
3770 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || |
3771 | (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); |
3772 | |
3773 | for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { |
3774 | int elem = size_index_elem(i); |
3775 | if (elem >= ARRAY_SIZE(size_index)) |
3776 | break; |
3777 | size_index[elem] = KMALLOC_SHIFT_LOW; |
3778 | } |
3779 | |
3780 | if (KMALLOC_MIN_SIZE == 64) { |
3781 | /* |
3782 | * The 96 byte size cache is not used if the alignment |
3783 | * is 64 byte. |
3784 | */ |
3785 | for (i = 64 + 8; i <= 96; i += 8) |
3786 | size_index[size_index_elem(i)] = 7; |
3787 | } else if (KMALLOC_MIN_SIZE == 128) { |
3788 | /* |
3789 | * The 192 byte sized cache is not used if the alignment |
3790 | * is 128 byte. Redirect kmalloc to use the 256 byte cache |
3791 | * instead. |
3792 | */ |
3793 | for (i = 128 + 8; i <= 192; i += 8) |
3794 | size_index[size_index_elem(i)] = 8; |
3795 | } |
3796 | |
3797 | /* Caches that are not of the two-to-the-power-of size */ |
3798 | if (KMALLOC_MIN_SIZE <= 32) { |
3799 | kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0); |
3800 | caches++; |
3801 | } |
3802 | |
3803 | if (KMALLOC_MIN_SIZE <= 64) { |
3804 | kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0); |
3805 | caches++; |
3806 | } |
3807 | |
3808 | for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { |
3809 | kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0); |
3810 | caches++; |
3811 | } |
3812 | |
3813 | slab_state = UP; |
3814 | |
3815 | /* Provide the correct kmalloc names now that the caches are up */ |
3816 | if (KMALLOC_MIN_SIZE <= 32) { |
3817 | kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT); |
3818 | BUG_ON(!kmalloc_caches[1]->name); |
3819 | } |
3820 | |
3821 | if (KMALLOC_MIN_SIZE <= 64) { |
3822 | kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT); |
3823 | BUG_ON(!kmalloc_caches[2]->name); |
3824 | } |
3825 | |
3826 | for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { |
3827 | char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); |
3828 | |
3829 | BUG_ON(!s); |
3830 | kmalloc_caches[i]->name = s; |
3831 | } |
3832 | |
3833 | #ifdef CONFIG_SMP |
3834 | register_cpu_notifier(&slab_notifier); |
3835 | #endif |
3836 | |
3837 | #ifdef CONFIG_ZONE_DMA |
3838 | for (i = 0; i < SLUB_PAGE_SHIFT; i++) { |
3839 | struct kmem_cache *s = kmalloc_caches[i]; |
3840 | |
3841 | if (s && s->size) { |
3842 | char *name = kasprintf(GFP_NOWAIT, |
3843 | "dma-kmalloc-%d", s->objsize); |
3844 | |
3845 | BUG_ON(!name); |
3846 | kmalloc_dma_caches[i] = create_kmalloc_cache(name, |
3847 | s->objsize, SLAB_CACHE_DMA); |
3848 | } |
3849 | } |
3850 | #endif |
3851 | printk(KERN_INFO |
3852 | "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," |
3853 | " CPUs=%d, Nodes=%d\n", |
3854 | caches, cache_line_size(), |
3855 | slub_min_order, slub_max_order, slub_min_objects, |
3856 | nr_cpu_ids, nr_node_ids); |
3857 | } |
3858 | |
3859 | void __init kmem_cache_init_late(void) |
3860 | { |
3861 | } |
3862 | |
3863 | /* |
3864 | * Find a mergeable slab cache |
3865 | */ |
3866 | static int slab_unmergeable(struct kmem_cache *s) |
3867 | { |
3868 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) |
3869 | return 1; |
3870 | |
3871 | if (s->ctor) |
3872 | return 1; |
3873 | |
3874 | /* |
3875 | * We may have set a slab to be unmergeable during bootstrap. |
3876 | */ |
3877 | if (s->refcount < 0) |
3878 | return 1; |
3879 | |
3880 | return 0; |
3881 | } |
3882 | |
3883 | static struct kmem_cache *find_mergeable(size_t size, |
3884 | size_t align, unsigned long flags, const char *name, |
3885 | void (*ctor)(void *)) |
3886 | { |
3887 | struct kmem_cache *s; |
3888 | |
3889 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) |
3890 | return NULL; |
3891 | |
3892 | if (ctor) |
3893 | return NULL; |
3894 | |
3895 | size = ALIGN(size, sizeof(void *)); |
3896 | align = calculate_alignment(flags, align, size); |
3897 | size = ALIGN(size, align); |
3898 | flags = kmem_cache_flags(size, flags, name, NULL); |
3899 | |
3900 | list_for_each_entry(s, &slab_caches, list) { |
3901 | if (slab_unmergeable(s)) |
3902 | continue; |
3903 | |
3904 | if (size > s->size) |
3905 | continue; |
3906 | |
3907 | if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) |
3908 | continue; |
3909 | /* |
3910 | * Check if alignment is compatible. |
3911 | * Courtesy of Adrian Drzewiecki |
3912 | */ |
3913 | if ((s->size & ~(align - 1)) != s->size) |
3914 | continue; |
3915 | |
3916 | if (s->size - size >= sizeof(void *)) |
3917 | continue; |
3918 | |
3919 | return s; |
3920 | } |
3921 | return NULL; |
3922 | } |
3923 | |
3924 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, |
3925 | size_t align, unsigned long flags, void (*ctor)(void *)) |
3926 | { |
3927 | struct kmem_cache *s; |
3928 | char *n; |
3929 | |
3930 | if (WARN_ON(!name)) |
3931 | return NULL; |
3932 | |
3933 | down_write(&slub_lock); |
3934 | s = find_mergeable(size, align, flags, name, ctor); |
3935 | if (s) { |
3936 | s->refcount++; |
3937 | /* |
3938 | * Adjust the object sizes so that we clear |
3939 | * the complete object on kzalloc. |
3940 | */ |
3941 | s->objsize = max(s->objsize, (int)size); |
3942 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); |
3943 | |
3944 | if (sysfs_slab_alias(s, name)) { |
3945 | s->refcount--; |
3946 | goto err; |
3947 | } |
3948 | up_write(&slub_lock); |
3949 | return s; |
3950 | } |
3951 | |
3952 | n = kstrdup(name, GFP_KERNEL); |
3953 | if (!n) |
3954 | goto err; |
3955 | |
3956 | s = kmalloc(kmem_size, GFP_KERNEL); |
3957 | if (s) { |
3958 | if (kmem_cache_open(s, n, |
3959 | size, align, flags, ctor)) { |
3960 | list_add(&s->list, &slab_caches); |
3961 | up_write(&slub_lock); |
3962 | if (sysfs_slab_add(s)) { |
3963 | down_write(&slub_lock); |
3964 | list_del(&s->list); |
3965 | kfree(n); |
3966 | kfree(s); |
3967 | goto err; |
3968 | } |
3969 | return s; |
3970 | } |
3971 | kfree(n); |
3972 | kfree(s); |
3973 | } |
3974 | err: |
3975 | up_write(&slub_lock); |
3976 | |
3977 | if (flags & SLAB_PANIC) |
3978 | panic("Cannot create slabcache %s\n", name); |
3979 | else |
3980 | s = NULL; |
3981 | return s; |
3982 | } |
3983 | EXPORT_SYMBOL(kmem_cache_create); |
3984 | |
3985 | #ifdef CONFIG_SMP |
3986 | /* |
3987 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
3988 | * necessary. |
3989 | */ |
3990 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, |
3991 | unsigned long action, void *hcpu) |
3992 | { |
3993 | long cpu = (long)hcpu; |
3994 | struct kmem_cache *s; |
3995 | unsigned long flags; |
3996 | |
3997 | switch (action) { |
3998 | case CPU_UP_CANCELED: |
3999 | case CPU_UP_CANCELED_FROZEN: |
4000 | case CPU_DEAD: |
4001 | case CPU_DEAD_FROZEN: |
4002 | down_read(&slub_lock); |
4003 | list_for_each_entry(s, &slab_caches, list) { |
4004 | local_irq_save(flags); |
4005 | __flush_cpu_slab(s, cpu); |
4006 | local_irq_restore(flags); |
4007 | } |
4008 | up_read(&slub_lock); |
4009 | break; |
4010 | default: |
4011 | break; |
4012 | } |
4013 | return NOTIFY_OK; |
4014 | } |
4015 | |
4016 | static struct notifier_block __cpuinitdata slab_notifier = { |
4017 | .notifier_call = slab_cpuup_callback |
4018 | }; |
4019 | |
4020 | #endif |
4021 | |
4022 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) |
4023 | { |
4024 | struct kmem_cache *s; |
4025 | void *ret; |
4026 | |
4027 | if (unlikely(size > SLUB_MAX_SIZE)) |
4028 | return kmalloc_large(size, gfpflags); |
4029 | |
4030 | s = get_slab(size, gfpflags); |
4031 | |
4032 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
4033 | return s; |
4034 | |
4035 | ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); |
4036 | |
4037 | /* Honor the call site pointer we received. */ |
4038 | trace_kmalloc(caller, ret, size, s->size, gfpflags); |
4039 | |
4040 | return ret; |
4041 | } |
4042 | |
4043 | #ifdef CONFIG_NUMA |
4044 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, |
4045 | int node, unsigned long caller) |
4046 | { |
4047 | struct kmem_cache *s; |
4048 | void *ret; |
4049 | |
4050 | if (unlikely(size > SLUB_MAX_SIZE)) { |
4051 | ret = kmalloc_large_node(size, gfpflags, node); |
4052 | |
4053 | trace_kmalloc_node(caller, ret, |
4054 | size, PAGE_SIZE << get_order(size), |
4055 | gfpflags, node); |
4056 | |
4057 | return ret; |
4058 | } |
4059 | |
4060 | s = get_slab(size, gfpflags); |
4061 | |
4062 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
4063 | return s; |
4064 | |
4065 | ret = slab_alloc(s, gfpflags, node, caller); |
4066 | |
4067 | /* Honor the call site pointer we received. */ |
4068 | trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); |
4069 | |
4070 | return ret; |
4071 | } |
4072 | #endif |
4073 | |
4074 | #ifdef CONFIG_SYSFS |
4075 | static int count_inuse(struct page *page) |
4076 | { |
4077 | return page->inuse; |
4078 | } |
4079 | |
4080 | static int count_total(struct page *page) |
4081 | { |
4082 | return page->objects; |
4083 | } |
4084 | #endif |
4085 | |
4086 | #ifdef CONFIG_SLUB_DEBUG |
4087 | static int validate_slab(struct kmem_cache *s, struct page *page, |
4088 | unsigned long *map) |
4089 | { |
4090 | void *p; |
4091 | void *addr = page_address(page); |
4092 | |
4093 | if (!check_slab(s, page) || |
4094 | !on_freelist(s, page, NULL)) |
4095 | return 0; |
4096 | |
4097 | /* Now we know that a valid freelist exists */ |
4098 | bitmap_zero(map, page->objects); |
4099 | |
4100 | get_map(s, page, map); |
4101 | for_each_object(p, s, addr, page->objects) { |
4102 | if (test_bit(slab_index(p, s, addr), map)) |
4103 | if (!check_object(s, page, p, SLUB_RED_INACTIVE)) |
4104 | return 0; |
4105 | } |
4106 | |
4107 | for_each_object(p, s, addr, page->objects) |
4108 | if (!test_bit(slab_index(p, s, addr), map)) |
4109 | if (!check_object(s, page, p, SLUB_RED_ACTIVE)) |
4110 | return 0; |
4111 | return 1; |
4112 | } |
4113 | |
4114 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, |
4115 | unsigned long *map) |
4116 | { |
4117 | slab_lock(page); |
4118 | validate_slab(s, page, map); |
4119 | slab_unlock(page); |
4120 | } |
4121 | |
4122 | static int validate_slab_node(struct kmem_cache *s, |
4123 | struct kmem_cache_node *n, unsigned long *map) |
4124 | { |
4125 | unsigned long count = 0; |
4126 | struct page *page; |
4127 | unsigned long flags; |
4128 | |
4129 | spin_lock_irqsave(&n->list_lock, flags); |
4130 | |
4131 | list_for_each_entry(page, &n->partial, lru) { |
4132 | validate_slab_slab(s, page, map); |
4133 | count++; |
4134 | } |
4135 | if (count != n->nr_partial) |
4136 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " |
4137 | "counter=%ld\n", s->name, count, n->nr_partial); |
4138 | |
4139 | if (!(s->flags & SLAB_STORE_USER)) |
4140 | goto out; |
4141 | |
4142 | list_for_each_entry(page, &n->full, lru) { |
4143 | validate_slab_slab(s, page, map); |
4144 | count++; |
4145 | } |
4146 | if (count != atomic_long_read(&n->nr_slabs)) |
4147 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " |
4148 | "counter=%ld\n", s->name, count, |
4149 | atomic_long_read(&n->nr_slabs)); |
4150 | |
4151 | out: |
4152 | spin_unlock_irqrestore(&n->list_lock, flags); |
4153 | return count; |
4154 | } |
4155 | |
4156 | static long validate_slab_cache(struct kmem_cache *s) |
4157 | { |
4158 | int node; |
4159 | unsigned long count = 0; |
4160 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * |
4161 | sizeof(unsigned long), GFP_KERNEL); |
4162 | |
4163 | if (!map) |
4164 | return -ENOMEM; |
4165 | |
4166 | flush_all(s); |
4167 | for_each_node_state(node, N_NORMAL_MEMORY) { |
4168 | struct kmem_cache_node *n = get_node(s, node); |
4169 | |
4170 | count += validate_slab_node(s, n, map); |
4171 | } |
4172 | kfree(map); |
4173 | return count; |
4174 | } |
4175 | /* |
4176 | * Generate lists of code addresses where slabcache objects are allocated |
4177 | * and freed. |
4178 | */ |
4179 | |
4180 | struct location { |
4181 | unsigned long count; |
4182 | unsigned long addr; |
4183 | long long sum_time; |
4184 | long min_time; |
4185 | long max_time; |
4186 | long min_pid; |
4187 | long max_pid; |
4188 | DECLARE_BITMAP(cpus, NR_CPUS); |
4189 | nodemask_t nodes; |
4190 | }; |
4191 | |
4192 | struct loc_track { |
4193 | unsigned long max; |
4194 | unsigned long count; |
4195 | struct location *loc; |
4196 | }; |
4197 | |
4198 | static void free_loc_track(struct loc_track *t) |
4199 | { |
4200 | if (t->max) |
4201 | free_pages((unsigned long)t->loc, |
4202 | get_order(sizeof(struct location) * t->max)); |
4203 | } |
4204 | |
4205 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) |
4206 | { |
4207 | struct location *l; |
4208 | int order; |
4209 | |
4210 | order = get_order(sizeof(struct location) * max); |
4211 | |
4212 | l = (void *)__get_free_pages(flags, order); |
4213 | if (!l) |
4214 | return 0; |
4215 | |
4216 | if (t->count) { |
4217 | memcpy(l, t->loc, sizeof(struct location) * t->count); |
4218 | free_loc_track(t); |
4219 | } |
4220 | t->max = max; |
4221 | t->loc = l; |
4222 | return 1; |
4223 | } |
4224 | |
4225 | static int add_location(struct loc_track *t, struct kmem_cache *s, |
4226 | const struct track *track) |
4227 | { |
4228 | long start, end, pos; |
4229 | struct location *l; |
4230 | unsigned long caddr; |
4231 | unsigned long age = jiffies - track->when; |
4232 | |
4233 | start = -1; |
4234 | end = t->count; |
4235 | |
4236 | for ( ; ; ) { |
4237 | pos = start + (end - start + 1) / 2; |
4238 | |
4239 | /* |
4240 | * There is nothing at "end". If we end up there |
4241 | * we need to add something to before end. |
4242 | */ |
4243 | if (pos == end) |
4244 | break; |
4245 | |
4246 | caddr = t->loc[pos].addr; |
4247 | if (track->addr == caddr) { |
4248 | |
4249 | l = &t->loc[pos]; |
4250 | l->count++; |
4251 | if (track->when) { |
4252 | l->sum_time += age; |
4253 | if (age < l->min_time) |
4254 | l->min_time = age; |
4255 | if (age > l->max_time) |
4256 | l->max_time = age; |
4257 | |
4258 | if (track->pid < l->min_pid) |
4259 | l->min_pid = track->pid; |
4260 | if (track->pid > l->max_pid) |
4261 | l->max_pid = track->pid; |
4262 | |
4263 | cpumask_set_cpu(track->cpu, |
4264 | to_cpumask(l->cpus)); |
4265 | } |
4266 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
4267 | return 1; |
4268 | } |
4269 | |
4270 | if (track->addr < caddr) |
4271 | end = pos; |
4272 | else |
4273 | start = pos; |
4274 | } |
4275 | |
4276 | /* |
4277 | * Not found. Insert new tracking element. |
4278 | */ |
4279 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) |
4280 | return 0; |
4281 | |
4282 | l = t->loc + pos; |
4283 | if (pos < t->count) |
4284 | memmove(l + 1, l, |
4285 | (t->count - pos) * sizeof(struct location)); |
4286 | t->count++; |
4287 | l->count = 1; |
4288 | l->addr = track->addr; |
4289 | l->sum_time = age; |
4290 | l->min_time = age; |
4291 | l->max_time = age; |
4292 | l->min_pid = track->pid; |
4293 | l->max_pid = track->pid; |
4294 | cpumask_clear(to_cpumask(l->cpus)); |
4295 | cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); |
4296 | nodes_clear(l->nodes); |
4297 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
4298 | return 1; |
4299 | } |
4300 | |
4301 | static void process_slab(struct loc_track *t, struct kmem_cache *s, |
4302 | struct page *page, enum track_item alloc, |
4303 | unsigned long *map) |
4304 | { |
4305 | void *addr = page_address(page); |
4306 | void *p; |
4307 | |
4308 | bitmap_zero(map, page->objects); |
4309 | get_map(s, page, map); |
4310 | |
4311 | for_each_object(p, s, addr, page->objects) |
4312 | if (!test_bit(slab_index(p, s, addr), map)) |
4313 | add_location(t, s, get_track(s, p, alloc)); |
4314 | } |
4315 | |
4316 | static int list_locations(struct kmem_cache *s, char *buf, |
4317 | enum track_item alloc) |
4318 | { |
4319 | int len = 0; |
4320 | unsigned long i; |
4321 | struct loc_track t = { 0, 0, NULL }; |
4322 | int node; |
4323 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * |
4324 | sizeof(unsigned long), GFP_KERNEL); |
4325 | |
4326 | if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), |
4327 | GFP_TEMPORARY)) { |
4328 | kfree(map); |
4329 | return sprintf(buf, "Out of memory\n"); |
4330 | } |
4331 | /* Push back cpu slabs */ |
4332 | flush_all(s); |
4333 | |
4334 | for_each_node_state(node, N_NORMAL_MEMORY) { |
4335 | struct kmem_cache_node *n = get_node(s, node); |
4336 | unsigned long flags; |
4337 | struct page *page; |
4338 | |
4339 | if (!atomic_long_read(&n->nr_slabs)) |
4340 | continue; |
4341 | |
4342 | spin_lock_irqsave(&n->list_lock, flags); |
4343 | list_for_each_entry(page, &n->partial, lru) |
4344 | process_slab(&t, s, page, alloc, map); |
4345 | list_for_each_entry(page, &n->full, lru) |
4346 | process_slab(&t, s, page, alloc, map); |
4347 | spin_unlock_irqrestore(&n->list_lock, flags); |
4348 | } |
4349 | |
4350 | for (i = 0; i < t.count; i++) { |
4351 | struct location *l = &t.loc[i]; |
4352 | |
4353 | if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) |
4354 | break; |
4355 | len += sprintf(buf + len, "%7ld ", l->count); |
4356 | |
4357 | if (l->addr) |
4358 | len += sprintf(buf + len, "%pS", (void *)l->addr); |
4359 | else |
4360 | len += sprintf(buf + len, "<not-available>"); |
4361 | |
4362 | if (l->sum_time != l->min_time) { |
4363 | len += sprintf(buf + len, " age=%ld/%ld/%ld", |
4364 | l->min_time, |
4365 | (long)div_u64(l->sum_time, l->count), |
4366 | l->max_time); |
4367 | } else |
4368 | len += sprintf(buf + len, " age=%ld", |
4369 | l->min_time); |
4370 | |
4371 | if (l->min_pid != l->max_pid) |
4372 | len += sprintf(buf + len, " pid=%ld-%ld", |
4373 | l->min_pid, l->max_pid); |
4374 | else |
4375 | len += sprintf(buf + len, " pid=%ld", |
4376 | l->min_pid); |
4377 | |
4378 | if (num_online_cpus() > 1 && |
4379 | !cpumask_empty(to_cpumask(l->cpus)) && |
4380 | len < PAGE_SIZE - 60) { |
4381 | len += sprintf(buf + len, " cpus="); |
4382 | len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, |
4383 | to_cpumask(l->cpus)); |
4384 | } |
4385 | |
4386 | if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && |
4387 | len < PAGE_SIZE - 60) { |
4388 | len += sprintf(buf + len, " nodes="); |
4389 | len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, |
4390 | l->nodes); |
4391 | } |
4392 | |
4393 | len += sprintf(buf + len, "\n"); |
4394 | } |
4395 | |
4396 | free_loc_track(&t); |
4397 | kfree(map); |
4398 | if (!t.count) |
4399 | len += sprintf(buf, "No data\n"); |
4400 | return len; |
4401 | } |
4402 | #endif |
4403 | |
4404 | #ifdef SLUB_RESILIENCY_TEST |
4405 | static void resiliency_test(void) |
4406 | { |
4407 | u8 *p; |
4408 | |
4409 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10); |
4410 | |
4411 | printk(KERN_ERR "SLUB resiliency testing\n"); |
4412 | printk(KERN_ERR "-----------------------\n"); |
4413 | printk(KERN_ERR "A. Corruption after allocation\n"); |
4414 | |
4415 | p = kzalloc(16, GFP_KERNEL); |
4416 | p[16] = 0x12; |
4417 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" |
4418 | " 0x12->0x%p\n\n", p + 16); |
4419 | |
4420 | validate_slab_cache(kmalloc_caches[4]); |
4421 | |
4422 | /* Hmmm... The next two are dangerous */ |
4423 | p = kzalloc(32, GFP_KERNEL); |
4424 | p[32 + sizeof(void *)] = 0x34; |
4425 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" |
4426 | " 0x34 -> -0x%p\n", p); |
4427 | printk(KERN_ERR |
4428 | "If allocated object is overwritten then not detectable\n\n"); |
4429 | |
4430 | validate_slab_cache(kmalloc_caches[5]); |
4431 | p = kzalloc(64, GFP_KERNEL); |
4432 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); |
4433 | *p = 0x56; |
4434 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", |
4435 | p); |
4436 | printk(KERN_ERR |
4437 | "If allocated object is overwritten then not detectable\n\n"); |
4438 | validate_slab_cache(kmalloc_caches[6]); |
4439 | |
4440 | printk(KERN_ERR "\nB. Corruption after free\n"); |
4441 | p = kzalloc(128, GFP_KERNEL); |
4442 | kfree(p); |
4443 | *p = 0x78; |
4444 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); |
4445 | validate_slab_cache(kmalloc_caches[7]); |
4446 | |
4447 | p = kzalloc(256, GFP_KERNEL); |
4448 | kfree(p); |
4449 | p[50] = 0x9a; |
4450 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", |
4451 | p); |
4452 | validate_slab_cache(kmalloc_caches[8]); |
4453 | |
4454 | p = kzalloc(512, GFP_KERNEL); |
4455 | kfree(p); |
4456 | p[512] = 0xab; |
4457 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); |
4458 | validate_slab_cache(kmalloc_caches[9]); |
4459 | } |
4460 | #else |
4461 | #ifdef CONFIG_SYSFS |
4462 | static void resiliency_test(void) {}; |
4463 | #endif |
4464 | #endif |
4465 | |
4466 | #ifdef CONFIG_SYSFS |
4467 | enum slab_stat_type { |
4468 | SL_ALL, /* All slabs */ |
4469 | SL_PARTIAL, /* Only partially allocated slabs */ |
4470 | SL_CPU, /* Only slabs used for cpu caches */ |
4471 | SL_OBJECTS, /* Determine allocated objects not slabs */ |
4472 | SL_TOTAL /* Determine object capacity not slabs */ |
4473 | }; |
4474 | |
4475 | #define SO_ALL (1 << SL_ALL) |
4476 | #define SO_PARTIAL (1 << SL_PARTIAL) |
4477 | #define SO_CPU (1 << SL_CPU) |
4478 | #define SO_OBJECTS (1 << SL_OBJECTS) |
4479 | #define SO_TOTAL (1 << SL_TOTAL) |
4480 | |
4481 | static ssize_t show_slab_objects(struct kmem_cache *s, |
4482 | char *buf, unsigned long flags) |
4483 | { |
4484 | unsigned long total = 0; |
4485 | int node; |
4486 | int x; |
4487 | unsigned long *nodes; |
4488 | unsigned long *per_cpu; |
4489 | |
4490 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); |
4491 | if (!nodes) |
4492 | return -ENOMEM; |
4493 | per_cpu = nodes + nr_node_ids; |
4494 | |
4495 | if (flags & SO_CPU) { |
4496 | int cpu; |
4497 | |
4498 | for_each_possible_cpu(cpu) { |
4499 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
4500 | int node = ACCESS_ONCE(c->node); |
4501 | struct page *page; |
4502 | |
4503 | if (node < 0) |
4504 | continue; |
4505 | page = ACCESS_ONCE(c->page); |
4506 | if (page) { |
4507 | if (flags & SO_TOTAL) |
4508 | x = page->objects; |
4509 | else if (flags & SO_OBJECTS) |
4510 | x = page->inuse; |
4511 | else |
4512 | x = 1; |
4513 | |
4514 | total += x; |
4515 | nodes[node] += x; |
4516 | } |
4517 | page = c->partial; |
4518 | |
4519 | if (page) { |
4520 | x = page->pobjects; |
4521 | total += x; |
4522 | nodes[node] += x; |
4523 | } |
4524 | per_cpu[node]++; |
4525 | } |
4526 | } |
4527 | |
4528 | lock_memory_hotplug(); |
4529 | #ifdef CONFIG_SLUB_DEBUG |
4530 | if (flags & SO_ALL) { |
4531 | for_each_node_state(node, N_NORMAL_MEMORY) { |
4532 | struct kmem_cache_node *n = get_node(s, node); |
4533 | |
4534 | if (flags & SO_TOTAL) |
4535 | x = atomic_long_read(&n->total_objects); |
4536 | else if (flags & SO_OBJECTS) |
4537 | x = atomic_long_read(&n->total_objects) - |
4538 | count_partial(n, count_free); |
4539 | |
4540 | else |
4541 | x = atomic_long_read(&n->nr_slabs); |
4542 | total += x; |
4543 | nodes[node] += x; |
4544 | } |
4545 | |
4546 | } else |
4547 | #endif |
4548 | if (flags & SO_PARTIAL) { |
4549 | for_each_node_state(node, N_NORMAL_MEMORY) { |
4550 | struct kmem_cache_node *n = get_node(s, node); |
4551 | |
4552 | if (flags & SO_TOTAL) |
4553 | x = count_partial(n, count_total); |
4554 | else if (flags & SO_OBJECTS) |
4555 | x = count_partial(n, count_inuse); |
4556 | else |
4557 | x = n->nr_partial; |
4558 | total += x; |
4559 | nodes[node] += x; |
4560 | } |
4561 | } |
4562 | x = sprintf(buf, "%lu", total); |
4563 | #ifdef CONFIG_NUMA |
4564 | for_each_node_state(node, N_NORMAL_MEMORY) |
4565 | if (nodes[node]) |
4566 | x += sprintf(buf + x, " N%d=%lu", |
4567 | node, nodes[node]); |
4568 | #endif |
4569 | unlock_memory_hotplug(); |
4570 | kfree(nodes); |
4571 | return x + sprintf(buf + x, "\n"); |
4572 | } |
4573 | |
4574 | #ifdef CONFIG_SLUB_DEBUG |
4575 | static int any_slab_objects(struct kmem_cache *s) |
4576 | { |
4577 | int node; |
4578 | |
4579 | for_each_online_node(node) { |
4580 | struct kmem_cache_node *n = get_node(s, node); |
4581 | |
4582 | if (!n) |
4583 | continue; |
4584 | |
4585 | if (atomic_long_read(&n->total_objects)) |
4586 | return 1; |
4587 | } |
4588 | return 0; |
4589 | } |
4590 | #endif |
4591 | |
4592 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) |
4593 | #define to_slab(n) container_of(n, struct kmem_cache, kobj) |
4594 | |
4595 | struct slab_attribute { |
4596 | struct attribute attr; |
4597 | ssize_t (*show)(struct kmem_cache *s, char *buf); |
4598 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); |
4599 | }; |
4600 | |
4601 | #define SLAB_ATTR_RO(_name) \ |
4602 | static struct slab_attribute _name##_attr = \ |
4603 | __ATTR(_name, 0400, _name##_show, NULL) |
4604 | |
4605 | #define SLAB_ATTR(_name) \ |
4606 | static struct slab_attribute _name##_attr = \ |
4607 | __ATTR(_name, 0600, _name##_show, _name##_store) |
4608 | |
4609 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
4610 | { |
4611 | return sprintf(buf, "%d\n", s->size); |
4612 | } |
4613 | SLAB_ATTR_RO(slab_size); |
4614 | |
4615 | static ssize_t align_show(struct kmem_cache *s, char *buf) |
4616 | { |
4617 | return sprintf(buf, "%d\n", s->align); |
4618 | } |
4619 | SLAB_ATTR_RO(align); |
4620 | |
4621 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) |
4622 | { |
4623 | return sprintf(buf, "%d\n", s->objsize); |
4624 | } |
4625 | SLAB_ATTR_RO(object_size); |
4626 | |
4627 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) |
4628 | { |
4629 | return sprintf(buf, "%d\n", oo_objects(s->oo)); |
4630 | } |
4631 | SLAB_ATTR_RO(objs_per_slab); |
4632 | |
4633 | static ssize_t order_store(struct kmem_cache *s, |
4634 | const char *buf, size_t length) |
4635 | { |
4636 | unsigned long order; |
4637 | int err; |
4638 | |
4639 | err = strict_strtoul(buf, 10, &order); |
4640 | if (err) |
4641 | return err; |
4642 | |
4643 | if (order > slub_max_order || order < slub_min_order) |
4644 | return -EINVAL; |
4645 | |
4646 | calculate_sizes(s, order); |
4647 | return length; |
4648 | } |
4649 | |
4650 | static ssize_t order_show(struct kmem_cache *s, char *buf) |
4651 | { |
4652 | return sprintf(buf, "%d\n", oo_order(s->oo)); |
4653 | } |
4654 | SLAB_ATTR(order); |
4655 | |
4656 | static ssize_t min_partial_show(struct kmem_cache *s, char *buf) |
4657 | { |
4658 | return sprintf(buf, "%lu\n", s->min_partial); |
4659 | } |
4660 | |
4661 | static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, |
4662 | size_t length) |
4663 | { |
4664 | unsigned long min; |
4665 | int err; |
4666 | |
4667 | err = strict_strtoul(buf, 10, &min); |
4668 | if (err) |
4669 | return err; |
4670 | |
4671 | set_min_partial(s, min); |
4672 | return length; |
4673 | } |
4674 | SLAB_ATTR(min_partial); |
4675 | |
4676 | static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) |
4677 | { |
4678 | return sprintf(buf, "%u\n", s->cpu_partial); |
4679 | } |
4680 | |
4681 | static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, |
4682 | size_t length) |
4683 | { |
4684 | unsigned long objects; |
4685 | int err; |
4686 | |
4687 | err = strict_strtoul(buf, 10, &objects); |
4688 | if (err) |
4689 | return err; |
4690 | if (objects && kmem_cache_debug(s)) |
4691 | return -EINVAL; |
4692 | |
4693 | s->cpu_partial = objects; |
4694 | flush_all(s); |
4695 | return length; |
4696 | } |
4697 | SLAB_ATTR(cpu_partial); |
4698 | |
4699 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) |
4700 | { |
4701 | if (!s->ctor) |
4702 | return 0; |
4703 | return sprintf(buf, "%pS\n", s->ctor); |
4704 | } |
4705 | SLAB_ATTR_RO(ctor); |
4706 | |
4707 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) |
4708 | { |
4709 | return sprintf(buf, "%d\n", s->refcount - 1); |
4710 | } |
4711 | SLAB_ATTR_RO(aliases); |
4712 | |
4713 | static ssize_t partial_show(struct kmem_cache *s, char *buf) |
4714 | { |
4715 | return show_slab_objects(s, buf, SO_PARTIAL); |
4716 | } |
4717 | SLAB_ATTR_RO(partial); |
4718 | |
4719 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) |
4720 | { |
4721 | return show_slab_objects(s, buf, SO_CPU); |
4722 | } |
4723 | SLAB_ATTR_RO(cpu_slabs); |
4724 | |
4725 | static ssize_t objects_show(struct kmem_cache *s, char *buf) |
4726 | { |
4727 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); |
4728 | } |
4729 | SLAB_ATTR_RO(objects); |
4730 | |
4731 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) |
4732 | { |
4733 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); |
4734 | } |
4735 | SLAB_ATTR_RO(objects_partial); |
4736 | |
4737 | static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) |
4738 | { |
4739 | int objects = 0; |
4740 | int pages = 0; |
4741 | int cpu; |
4742 | int len; |
4743 | |
4744 | for_each_online_cpu(cpu) { |
4745 | struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial; |
4746 | |
4747 | if (page) { |
4748 | pages += page->pages; |
4749 | objects += page->pobjects; |
4750 | } |
4751 | } |
4752 | |
4753 | len = sprintf(buf, "%d(%d)", objects, pages); |
4754 | |
4755 | #ifdef CONFIG_SMP |
4756 | for_each_online_cpu(cpu) { |
4757 | struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial; |
4758 | |
4759 | if (page && len < PAGE_SIZE - 20) |
4760 | len += sprintf(buf + len, " C%d=%d(%d)", cpu, |
4761 | page->pobjects, page->pages); |
4762 | } |
4763 | #endif |
4764 | return len + sprintf(buf + len, "\n"); |
4765 | } |
4766 | SLAB_ATTR_RO(slabs_cpu_partial); |
4767 | |
4768 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) |
4769 | { |
4770 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); |
4771 | } |
4772 | |
4773 | static ssize_t reclaim_account_store(struct kmem_cache *s, |
4774 | const char *buf, size_t length) |
4775 | { |
4776 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; |
4777 | if (buf[0] == '1') |
4778 | s->flags |= SLAB_RECLAIM_ACCOUNT; |
4779 | return length; |
4780 | } |
4781 | SLAB_ATTR(reclaim_account); |
4782 | |
4783 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) |
4784 | { |
4785 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); |
4786 | } |
4787 | SLAB_ATTR_RO(hwcache_align); |
4788 | |
4789 | #ifdef CONFIG_ZONE_DMA |
4790 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) |
4791 | { |
4792 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); |
4793 | } |
4794 | SLAB_ATTR_RO(cache_dma); |
4795 | #endif |
4796 | |
4797 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) |
4798 | { |
4799 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); |
4800 | } |
4801 | SLAB_ATTR_RO(destroy_by_rcu); |
4802 | |
4803 | static ssize_t reserved_show(struct kmem_cache *s, char *buf) |
4804 | { |
4805 | return sprintf(buf, "%d\n", s->reserved); |
4806 | } |
4807 | SLAB_ATTR_RO(reserved); |
4808 | |
4809 | #ifdef CONFIG_SLUB_DEBUG |
4810 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) |
4811 | { |
4812 | return show_slab_objects(s, buf, SO_ALL); |
4813 | } |
4814 | SLAB_ATTR_RO(slabs); |
4815 | |
4816 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) |
4817 | { |
4818 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); |
4819 | } |
4820 | SLAB_ATTR_RO(total_objects); |
4821 | |
4822 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) |
4823 | { |
4824 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); |
4825 | } |
4826 | |
4827 | static ssize_t sanity_checks_store(struct kmem_cache *s, |
4828 | const char *buf, size_t length) |
4829 | { |
4830 | s->flags &= ~SLAB_DEBUG_FREE; |
4831 | if (buf[0] == '1') { |
4832 | s->flags &= ~__CMPXCHG_DOUBLE; |
4833 | s->flags |= SLAB_DEBUG_FREE; |
4834 | } |
4835 | return length; |
4836 | } |
4837 | SLAB_ATTR(sanity_checks); |
4838 | |
4839 | static ssize_t trace_show(struct kmem_cache *s, char *buf) |
4840 | { |
4841 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); |
4842 | } |
4843 | |
4844 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, |
4845 | size_t length) |
4846 | { |
4847 | s->flags &= ~SLAB_TRACE; |
4848 | if (buf[0] == '1') { |
4849 | s->flags &= ~__CMPXCHG_DOUBLE; |
4850 | s->flags |= SLAB_TRACE; |
4851 | } |
4852 | return length; |
4853 | } |
4854 | SLAB_ATTR(trace); |
4855 | |
4856 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) |
4857 | { |
4858 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); |
4859 | } |
4860 | |
4861 | static ssize_t red_zone_store(struct kmem_cache *s, |
4862 | const char *buf, size_t length) |
4863 | { |
4864 | if (any_slab_objects(s)) |
4865 | return -EBUSY; |
4866 | |
4867 | s->flags &= ~SLAB_RED_ZONE; |
4868 | if (buf[0] == '1') { |
4869 | s->flags &= ~__CMPXCHG_DOUBLE; |
4870 | s->flags |= SLAB_RED_ZONE; |
4871 | } |
4872 | calculate_sizes(s, -1); |
4873 | return length; |
4874 | } |
4875 | SLAB_ATTR(red_zone); |
4876 | |
4877 | static ssize_t poison_show(struct kmem_cache *s, char *buf) |
4878 | { |
4879 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); |
4880 | } |
4881 | |
4882 | static ssize_t poison_store(struct kmem_cache *s, |
4883 | const char *buf, size_t length) |
4884 | { |
4885 | if (any_slab_objects(s)) |
4886 | return -EBUSY; |
4887 | |
4888 | s->flags &= ~SLAB_POISON; |
4889 | if (buf[0] == '1') { |
4890 | s->flags &= ~__CMPXCHG_DOUBLE; |
4891 | s->flags |= SLAB_POISON; |
4892 | } |
4893 | calculate_sizes(s, -1); |
4894 | return length; |
4895 | } |
4896 | SLAB_ATTR(poison); |
4897 | |
4898 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) |
4899 | { |
4900 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); |
4901 | } |
4902 | |
4903 | static ssize_t store_user_store(struct kmem_cache *s, |
4904 | const char *buf, size_t length) |
4905 | { |
4906 | if (any_slab_objects(s)) |
4907 | return -EBUSY; |
4908 | |
4909 | s->flags &= ~SLAB_STORE_USER; |
4910 | if (buf[0] == '1') { |
4911 | s->flags &= ~__CMPXCHG_DOUBLE; |
4912 | s->flags |= SLAB_STORE_USER; |
4913 | } |
4914 | calculate_sizes(s, -1); |
4915 | return length; |
4916 | } |
4917 | SLAB_ATTR(store_user); |
4918 | |
4919 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
4920 | { |
4921 | return 0; |
4922 | } |
4923 | |
4924 | static ssize_t validate_store(struct kmem_cache *s, |
4925 | const char *buf, size_t length) |
4926 | { |
4927 | int ret = -EINVAL; |
4928 | |
4929 | if (buf[0] == '1') { |
4930 | ret = validate_slab_cache(s); |
4931 | if (ret >= 0) |
4932 | ret = length; |
4933 | } |
4934 | return ret; |
4935 | } |
4936 | SLAB_ATTR(validate); |
4937 | |
4938 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) |
4939 | { |
4940 | if (!(s->flags & SLAB_STORE_USER)) |
4941 | return -ENOSYS; |
4942 | return list_locations(s, buf, TRACK_ALLOC); |
4943 | } |
4944 | SLAB_ATTR_RO(alloc_calls); |
4945 | |
4946 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) |
4947 | { |
4948 | if (!(s->flags & SLAB_STORE_USER)) |
4949 | return -ENOSYS; |
4950 | return list_locations(s, buf, TRACK_FREE); |
4951 | } |
4952 | SLAB_ATTR_RO(free_calls); |
4953 | #endif /* CONFIG_SLUB_DEBUG */ |
4954 | |
4955 | #ifdef CONFIG_FAILSLAB |
4956 | static ssize_t failslab_show(struct kmem_cache *s, char *buf) |
4957 | { |
4958 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); |
4959 | } |
4960 | |
4961 | static ssize_t failslab_store(struct kmem_cache *s, const char *buf, |
4962 | size_t length) |
4963 | { |
4964 | s->flags &= ~SLAB_FAILSLAB; |
4965 | if (buf[0] == '1') |
4966 | s->flags |= SLAB_FAILSLAB; |
4967 | return length; |
4968 | } |
4969 | SLAB_ATTR(failslab); |
4970 | #endif |
4971 | |
4972 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
4973 | { |
4974 | return 0; |
4975 | } |
4976 | |
4977 | static ssize_t shrink_store(struct kmem_cache *s, |
4978 | const char *buf, size_t length) |
4979 | { |
4980 | if (buf[0] == '1') { |
4981 | int rc = kmem_cache_shrink(s); |
4982 | |
4983 | if (rc) |
4984 | return rc; |
4985 | } else |
4986 | return -EINVAL; |
4987 | return length; |
4988 | } |
4989 | SLAB_ATTR(shrink); |
4990 | |
4991 | #ifdef CONFIG_NUMA |
4992 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) |
4993 | { |
4994 | return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); |
4995 | } |
4996 | |
4997 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, |
4998 | const char *buf, size_t length) |
4999 | { |
5000 | unsigned long ratio; |
5001 | int err; |
5002 | |
5003 | err = strict_strtoul(buf, 10, &ratio); |
5004 | if (err) |
5005 | return err; |
5006 | |
5007 | if (ratio <= 100) |
5008 | s->remote_node_defrag_ratio = ratio * 10; |
5009 | |
5010 | return length; |
5011 | } |
5012 | SLAB_ATTR(remote_node_defrag_ratio); |
5013 | #endif |
5014 | |
5015 | #ifdef CONFIG_SLUB_STATS |
5016 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) |
5017 | { |
5018 | unsigned long sum = 0; |
5019 | int cpu; |
5020 | int len; |
5021 | int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); |
5022 | |
5023 | if (!data) |
5024 | return -ENOMEM; |
5025 | |
5026 | for_each_online_cpu(cpu) { |
5027 | unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; |
5028 | |
5029 | data[cpu] = x; |
5030 | sum += x; |
5031 | } |
5032 | |
5033 | len = sprintf(buf, "%lu", sum); |
5034 | |
5035 | #ifdef CONFIG_SMP |
5036 | for_each_online_cpu(cpu) { |
5037 | if (data[cpu] && len < PAGE_SIZE - 20) |
5038 | len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); |
5039 | } |
5040 | #endif |
5041 | kfree(data); |
5042 | return len + sprintf(buf + len, "\n"); |
5043 | } |
5044 | |
5045 | static void clear_stat(struct kmem_cache *s, enum stat_item si) |
5046 | { |
5047 | int cpu; |
5048 | |
5049 | for_each_online_cpu(cpu) |
5050 | per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; |
5051 | } |
5052 | |
5053 | #define STAT_ATTR(si, text) \ |
5054 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ |
5055 | { \ |
5056 | return show_stat(s, buf, si); \ |
5057 | } \ |
5058 | static ssize_t text##_store(struct kmem_cache *s, \ |
5059 | const char *buf, size_t length) \ |
5060 | { \ |
5061 | if (buf[0] != '0') \ |
5062 | return -EINVAL; \ |
5063 | clear_stat(s, si); \ |
5064 | return length; \ |
5065 | } \ |
5066 | SLAB_ATTR(text); \ |
5067 | |
5068 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); |
5069 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); |
5070 | STAT_ATTR(FREE_FASTPATH, free_fastpath); |
5071 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); |
5072 | STAT_ATTR(FREE_FROZEN, free_frozen); |
5073 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); |
5074 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); |
5075 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); |
5076 | STAT_ATTR(ALLOC_SLAB, alloc_slab); |
5077 | STAT_ATTR(ALLOC_REFILL, alloc_refill); |
5078 | STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); |
5079 | STAT_ATTR(FREE_SLAB, free_slab); |
5080 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); |
5081 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); |
5082 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); |
5083 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); |
5084 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); |
5085 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); |
5086 | STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); |
5087 | STAT_ATTR(ORDER_FALLBACK, order_fallback); |
5088 | STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); |
5089 | STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); |
5090 | STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); |
5091 | STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); |
5092 | STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); |
5093 | STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); |
5094 | #endif |
5095 | |
5096 | static struct attribute *slab_attrs[] = { |
5097 | &slab_size_attr.attr, |
5098 | &object_size_attr.attr, |
5099 | &objs_per_slab_attr.attr, |
5100 | &order_attr.attr, |
5101 | &min_partial_attr.attr, |
5102 | &cpu_partial_attr.attr, |
5103 | &objects_attr.attr, |
5104 | &objects_partial_attr.attr, |
5105 | &partial_attr.attr, |
5106 | &cpu_slabs_attr.attr, |
5107 | &ctor_attr.attr, |
5108 | &aliases_attr.attr, |
5109 | &align_attr.attr, |
5110 | &hwcache_align_attr.attr, |
5111 | &reclaim_account_attr.attr, |
5112 | &destroy_by_rcu_attr.attr, |
5113 | &shrink_attr.attr, |
5114 | &reserved_attr.attr, |
5115 | &slabs_cpu_partial_attr.attr, |
5116 | #ifdef CONFIG_SLUB_DEBUG |
5117 | &total_objects_attr.attr, |
5118 | &slabs_attr.attr, |
5119 | &sanity_checks_attr.attr, |
5120 | &trace_attr.attr, |
5121 | &red_zone_attr.attr, |
5122 | &poison_attr.attr, |
5123 | &store_user_attr.attr, |
5124 | &validate_attr.attr, |
5125 | &alloc_calls_attr.attr, |
5126 | &free_calls_attr.attr, |
5127 | #endif |
5128 | #ifdef CONFIG_ZONE_DMA |
5129 | &cache_dma_attr.attr, |
5130 | #endif |
5131 | #ifdef CONFIG_NUMA |
5132 | &remote_node_defrag_ratio_attr.attr, |
5133 | #endif |
5134 | #ifdef CONFIG_SLUB_STATS |
5135 | &alloc_fastpath_attr.attr, |
5136 | &alloc_slowpath_attr.attr, |
5137 | &free_fastpath_attr.attr, |
5138 | &free_slowpath_attr.attr, |
5139 | &free_frozen_attr.attr, |
5140 | &free_add_partial_attr.attr, |
5141 | &free_remove_partial_attr.attr, |
5142 | &alloc_from_partial_attr.attr, |
5143 | &alloc_slab_attr.attr, |
5144 | &alloc_refill_attr.attr, |
5145 | &alloc_node_mismatch_attr.attr, |
5146 | &free_slab_attr.attr, |
5147 | &cpuslab_flush_attr.attr, |
5148 | &deactivate_full_attr.attr, |
5149 | &deactivate_empty_attr.attr, |
5150 | &deactivate_to_head_attr.attr, |
5151 | &deactivate_to_tail_attr.attr, |
5152 | &deactivate_remote_frees_attr.attr, |
5153 | &deactivate_bypass_attr.attr, |
5154 | &order_fallback_attr.attr, |
5155 | &cmpxchg_double_fail_attr.attr, |
5156 | &cmpxchg_double_cpu_fail_attr.attr, |
5157 | &cpu_partial_alloc_attr.attr, |
5158 | &cpu_partial_free_attr.attr, |
5159 | &cpu_partial_node_attr.attr, |
5160 | &cpu_partial_drain_attr.attr, |
5161 | #endif |
5162 | #ifdef CONFIG_FAILSLAB |
5163 | &failslab_attr.attr, |
5164 | #endif |
5165 | |
5166 | NULL |
5167 | }; |
5168 | |
5169 | static struct attribute_group slab_attr_group = { |
5170 | .attrs = slab_attrs, |
5171 | }; |
5172 | |
5173 | static ssize_t slab_attr_show(struct kobject *kobj, |
5174 | struct attribute *attr, |
5175 | char *buf) |
5176 | { |
5177 | struct slab_attribute *attribute; |
5178 | struct kmem_cache *s; |
5179 | int err; |
5180 | |
5181 | attribute = to_slab_attr(attr); |
5182 | s = to_slab(kobj); |
5183 | |
5184 | if (!attribute->show) |
5185 | return -EIO; |
5186 | |
5187 | err = attribute->show(s, buf); |
5188 | |
5189 | return err; |
5190 | } |
5191 | |
5192 | static ssize_t slab_attr_store(struct kobject *kobj, |
5193 | struct attribute *attr, |
5194 | const char *buf, size_t len) |
5195 | { |
5196 | struct slab_attribute *attribute; |
5197 | struct kmem_cache *s; |
5198 | int err; |
5199 | |
5200 | attribute = to_slab_attr(attr); |
5201 | s = to_slab(kobj); |
5202 | |
5203 | if (!attribute->store) |
5204 | return -EIO; |
5205 | |
5206 | err = attribute->store(s, buf, len); |
5207 | |
5208 | return err; |
5209 | } |
5210 | |
5211 | static void kmem_cache_release(struct kobject *kobj) |
5212 | { |
5213 | struct kmem_cache *s = to_slab(kobj); |
5214 | |
5215 | kfree(s->name); |
5216 | kfree(s); |
5217 | } |
5218 | |
5219 | static const struct sysfs_ops slab_sysfs_ops = { |
5220 | .show = slab_attr_show, |
5221 | .store = slab_attr_store, |
5222 | }; |
5223 | |
5224 | static struct kobj_type slab_ktype = { |
5225 | .sysfs_ops = &slab_sysfs_ops, |
5226 | .release = kmem_cache_release |
5227 | }; |
5228 | |
5229 | static int uevent_filter(struct kset *kset, struct kobject *kobj) |
5230 | { |
5231 | struct kobj_type *ktype = get_ktype(kobj); |
5232 | |
5233 | if (ktype == &slab_ktype) |
5234 | return 1; |
5235 | return 0; |
5236 | } |
5237 | |
5238 | static const struct kset_uevent_ops slab_uevent_ops = { |
5239 | .filter = uevent_filter, |
5240 | }; |
5241 | |
5242 | static struct kset *slab_kset; |
5243 | |
5244 | #define ID_STR_LENGTH 64 |
5245 | |
5246 | /* Create a unique string id for a slab cache: |
5247 | * |
5248 | * Format :[flags-]size |
5249 | */ |
5250 | static char *create_unique_id(struct kmem_cache *s) |
5251 | { |
5252 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); |
5253 | char *p = name; |
5254 | |
5255 | BUG_ON(!name); |
5256 | |
5257 | *p++ = ':'; |
5258 | /* |
5259 | * First flags affecting slabcache operations. We will only |
5260 | * get here for aliasable slabs so we do not need to support |
5261 | * too many flags. The flags here must cover all flags that |
5262 | * are matched during merging to guarantee that the id is |
5263 | * unique. |
5264 | */ |
5265 | if (s->flags & SLAB_CACHE_DMA) |
5266 | *p++ = 'd'; |
5267 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
5268 | *p++ = 'a'; |
5269 | if (s->flags & SLAB_DEBUG_FREE) |
5270 | *p++ = 'F'; |
5271 | if (!(s->flags & SLAB_NOTRACK)) |
5272 | *p++ = 't'; |
5273 | if (p != name + 1) |
5274 | *p++ = '-'; |
5275 | p += sprintf(p, "%07d", s->size); |
5276 | BUG_ON(p > name + ID_STR_LENGTH - 1); |
5277 | return name; |
5278 | } |
5279 | |
5280 | static int sysfs_slab_add(struct kmem_cache *s) |
5281 | { |
5282 | int err; |
5283 | const char *name; |
5284 | int unmergeable; |
5285 | |
5286 | if (slab_state < SYSFS) |
5287 | /* Defer until later */ |
5288 | return 0; |
5289 | |
5290 | unmergeable = slab_unmergeable(s); |
5291 | if (unmergeable) { |
5292 | /* |
5293 | * Slabcache can never be merged so we can use the name proper. |
5294 | * This is typically the case for debug situations. In that |
5295 | * case we can catch duplicate names easily. |
5296 | */ |
5297 | sysfs_remove_link(&slab_kset->kobj, s->name); |
5298 | name = s->name; |
5299 | } else { |
5300 | /* |
5301 | * Create a unique name for the slab as a target |
5302 | * for the symlinks. |
5303 | */ |
5304 | name = create_unique_id(s); |
5305 | } |
5306 | |
5307 | s->kobj.kset = slab_kset; |
5308 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); |
5309 | if (err) { |
5310 | kobject_put(&s->kobj); |
5311 | return err; |
5312 | } |
5313 | |
5314 | err = sysfs_create_group(&s->kobj, &slab_attr_group); |
5315 | if (err) { |
5316 | kobject_del(&s->kobj); |
5317 | kobject_put(&s->kobj); |
5318 | return err; |
5319 | } |
5320 | kobject_uevent(&s->kobj, KOBJ_ADD); |
5321 | if (!unmergeable) { |
5322 | /* Setup first alias */ |
5323 | sysfs_slab_alias(s, s->name); |
5324 | kfree(name); |
5325 | } |
5326 | return 0; |
5327 | } |
5328 | |
5329 | static void sysfs_slab_remove(struct kmem_cache *s) |
5330 | { |
5331 | if (slab_state < SYSFS) |
5332 | /* |
5333 | * Sysfs has not been setup yet so no need to remove the |
5334 | * cache from sysfs. |
5335 | */ |
5336 | return; |
5337 | |
5338 | kobject_uevent(&s->kobj, KOBJ_REMOVE); |
5339 | kobject_del(&s->kobj); |
5340 | kobject_put(&s->kobj); |
5341 | } |
5342 | |
5343 | /* |
5344 | * Need to buffer aliases during bootup until sysfs becomes |
5345 | * available lest we lose that information. |
5346 | */ |
5347 | struct saved_alias { |
5348 | struct kmem_cache *s; |
5349 | const char *name; |
5350 | struct saved_alias *next; |
5351 | }; |
5352 | |
5353 | static struct saved_alias *alias_list; |
5354 | |
5355 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) |
5356 | { |
5357 | struct saved_alias *al; |
5358 | |
5359 | if (slab_state == SYSFS) { |
5360 | /* |
5361 | * If we have a leftover link then remove it. |
5362 | */ |
5363 | sysfs_remove_link(&slab_kset->kobj, name); |
5364 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); |
5365 | } |
5366 | |
5367 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); |
5368 | if (!al) |
5369 | return -ENOMEM; |
5370 | |
5371 | al->s = s; |
5372 | al->name = name; |
5373 | al->next = alias_list; |
5374 | alias_list = al; |
5375 | return 0; |
5376 | } |
5377 | |
5378 | static int __init slab_sysfs_init(void) |
5379 | { |
5380 | struct kmem_cache *s; |
5381 | int err; |
5382 | |
5383 | down_write(&slub_lock); |
5384 | |
5385 | slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); |
5386 | if (!slab_kset) { |
5387 | up_write(&slub_lock); |
5388 | printk(KERN_ERR "Cannot register slab subsystem.\n"); |
5389 | return -ENOSYS; |
5390 | } |
5391 | |
5392 | slab_state = SYSFS; |
5393 | |
5394 | list_for_each_entry(s, &slab_caches, list) { |
5395 | err = sysfs_slab_add(s); |
5396 | if (err) |
5397 | printk(KERN_ERR "SLUB: Unable to add boot slab %s" |
5398 | " to sysfs\n", s->name); |
5399 | } |
5400 | |
5401 | while (alias_list) { |
5402 | struct saved_alias *al = alias_list; |
5403 | |
5404 | alias_list = alias_list->next; |
5405 | err = sysfs_slab_alias(al->s, al->name); |
5406 | if (err) |
5407 | printk(KERN_ERR "SLUB: Unable to add boot slab alias" |
5408 | " %s to sysfs\n", s->name); |
5409 | kfree(al); |
5410 | } |
5411 | |
5412 | up_write(&slub_lock); |
5413 | resiliency_test(); |
5414 | return 0; |
5415 | } |
5416 | |
5417 | __initcall(slab_sysfs_init); |
5418 | #endif /* CONFIG_SYSFS */ |
5419 | |
5420 | /* |
5421 | * The /proc/slabinfo ABI |
5422 | */ |
5423 | #ifdef CONFIG_SLABINFO |
5424 | static void print_slabinfo_header(struct seq_file *m) |
5425 | { |
5426 | seq_puts(m, "slabinfo - version: 2.1\n"); |
5427 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
5428 | "<objperslab> <pagesperslab>"); |
5429 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); |
5430 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); |
5431 | seq_putc(m, '\n'); |
5432 | } |
5433 | |
5434 | static void *s_start(struct seq_file *m, loff_t *pos) |
5435 | { |
5436 | loff_t n = *pos; |
5437 | |
5438 | down_read(&slub_lock); |
5439 | if (!n) |
5440 | print_slabinfo_header(m); |
5441 | |
5442 | return seq_list_start(&slab_caches, *pos); |
5443 | } |
5444 | |
5445 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) |
5446 | { |
5447 | return seq_list_next(p, &slab_caches, pos); |
5448 | } |
5449 | |
5450 | static void s_stop(struct seq_file *m, void *p) |
5451 | { |
5452 | up_read(&slub_lock); |
5453 | } |
5454 | |
5455 | static int s_show(struct seq_file *m, void *p) |
5456 | { |
5457 | unsigned long nr_partials = 0; |
5458 | unsigned long nr_slabs = 0; |
5459 | unsigned long nr_inuse = 0; |
5460 | unsigned long nr_objs = 0; |
5461 | unsigned long nr_free = 0; |
5462 | struct kmem_cache *s; |
5463 | int node; |
5464 | |
5465 | s = list_entry(p, struct kmem_cache, list); |
5466 | |
5467 | for_each_online_node(node) { |
5468 | struct kmem_cache_node *n = get_node(s, node); |
5469 | |
5470 | if (!n) |
5471 | continue; |
5472 | |
5473 | nr_partials += n->nr_partial; |
5474 | nr_slabs += atomic_long_read(&n->nr_slabs); |
5475 | nr_objs += atomic_long_read(&n->total_objects); |
5476 | nr_free += count_partial(n, count_free); |
5477 | } |
5478 | |
5479 | nr_inuse = nr_objs - nr_free; |
5480 | |
5481 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, |
5482 | nr_objs, s->size, oo_objects(s->oo), |
5483 | (1 << oo_order(s->oo))); |
5484 | seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); |
5485 | seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, |
5486 | 0UL); |
5487 | seq_putc(m, '\n'); |
5488 | return 0; |
5489 | } |
5490 | |
5491 | static const struct seq_operations slabinfo_op = { |
5492 | .start = s_start, |
5493 | .next = s_next, |
5494 | .stop = s_stop, |
5495 | .show = s_show, |
5496 | }; |
5497 | |
5498 | static int slabinfo_open(struct inode *inode, struct file *file) |
5499 | { |
5500 | return seq_open(file, &slabinfo_op); |
5501 | } |
5502 | |
5503 | static const struct file_operations proc_slabinfo_operations = { |
5504 | .open = slabinfo_open, |
5505 | .read = seq_read, |
5506 | .llseek = seq_lseek, |
5507 | .release = seq_release, |
5508 | }; |
5509 | |
5510 | static int __init slab_proc_init(void) |
5511 | { |
5512 | proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations); |
5513 | return 0; |
5514 | } |
5515 | module_init(slab_proc_init); |
5516 | #endif /* CONFIG_SLABINFO */ |
5517 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9