Root/
Source at commit be977234bfb4a6dca8a39e7c52165e4cd536ad71 created 12 years 9 months ago. By Lars-Peter Clausen, jz4740: Fix compile error | |
---|---|
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 and only |
6 | * uses a centralized lock to manage a pool of partial slabs. |
7 | * |
8 | * (C) 2007 SGI, Christoph Lameter |
9 | */ |
10 | |
11 | #include <linux/mm.h> |
12 | #include <linux/swap.h> /* struct reclaim_state */ |
13 | #include <linux/module.h> |
14 | #include <linux/bit_spinlock.h> |
15 | #include <linux/interrupt.h> |
16 | #include <linux/bitops.h> |
17 | #include <linux/slab.h> |
18 | #include <linux/proc_fs.h> |
19 | #include <linux/seq_file.h> |
20 | #include <linux/kmemcheck.h> |
21 | #include <linux/cpu.h> |
22 | #include <linux/cpuset.h> |
23 | #include <linux/mempolicy.h> |
24 | #include <linux/ctype.h> |
25 | #include <linux/debugobjects.h> |
26 | #include <linux/kallsyms.h> |
27 | #include <linux/memory.h> |
28 | #include <linux/math64.h> |
29 | #include <linux/fault-inject.h> |
30 | |
31 | #include <trace/events/kmem.h> |
32 | |
33 | /* |
34 | * Lock order: |
35 | * 1. slab_lock(page) |
36 | * 2. slab->list_lock |
37 | * |
38 | * The slab_lock protects operations on the object of a particular |
39 | * slab and its metadata in the page struct. If the slab lock |
40 | * has been taken then no allocations nor frees can be performed |
41 | * on the objects in the slab nor can the slab be added or removed |
42 | * from the partial or full lists since this would mean modifying |
43 | * the page_struct of the slab. |
44 | * |
45 | * The list_lock protects the partial and full list on each node and |
46 | * the partial slab counter. If taken then no new slabs may be added or |
47 | * removed from the lists nor make the number of partial slabs be modified. |
48 | * (Note that the total number of slabs is an atomic value that may be |
49 | * modified without taking the list lock). |
50 | * |
51 | * The list_lock is a centralized lock and thus we avoid taking it as |
52 | * much as possible. As long as SLUB does not have to handle partial |
53 | * slabs, operations can continue without any centralized lock. F.e. |
54 | * allocating a long series of objects that fill up slabs does not require |
55 | * the list lock. |
56 | * |
57 | * The lock order is sometimes inverted when we are trying to get a slab |
58 | * off a list. We take the list_lock and then look for a page on the list |
59 | * to use. While we do that objects in the slabs may be freed. We can |
60 | * only operate on the slab if we have also taken the slab_lock. So we use |
61 | * a slab_trylock() on the slab. If trylock was successful then no frees |
62 | * can occur anymore and we can use the slab for allocations etc. If the |
63 | * slab_trylock() does not succeed then frees are in progress in the slab and |
64 | * we must stay away from it for a while since we may cause a bouncing |
65 | * cacheline if we try to acquire the lock. So go onto the next slab. |
66 | * If all pages are busy then we may allocate a new slab instead of reusing |
67 | * a partial slab. A new slab has no one operating on it and thus there is |
68 | * no danger of cacheline contention. |
69 | * |
70 | * Interrupts are disabled during allocation and deallocation in order to |
71 | * make the slab allocator safe to use in the context of an irq. In addition |
72 | * interrupts are disabled to ensure that the processor does not change |
73 | * while handling per_cpu slabs, due to kernel preemption. |
74 | * |
75 | * SLUB assigns one slab for allocation to each processor. |
76 | * Allocations only occur from these slabs called cpu slabs. |
77 | * |
78 | * Slabs with free elements are kept on a partial list and during regular |
79 | * operations no list for full slabs is used. If an object in a full slab is |
80 | * freed then the slab will show up again on the partial lists. |
81 | * We track full slabs for debugging purposes though because otherwise we |
82 | * cannot scan all objects. |
83 | * |
84 | * Slabs are freed when they become empty. Teardown and setup is |
85 | * minimal so we rely on the page allocators per cpu caches for |
86 | * fast frees and allocs. |
87 | * |
88 | * Overloading of page flags that are otherwise used for LRU management. |
89 | * |
90 | * PageActive The slab is frozen and exempt from list processing. |
91 | * This means that the slab is dedicated to a purpose |
92 | * such as satisfying allocations for a specific |
93 | * processor. Objects may be freed in the slab while |
94 | * it is frozen but slab_free will then skip the usual |
95 | * list operations. It is up to the processor holding |
96 | * the slab to integrate the slab into the slab lists |
97 | * when the slab is no longer needed. |
98 | * |
99 | * One use of this flag is to mark slabs that are |
100 | * used for allocations. Then such a slab becomes a cpu |
101 | * slab. The cpu slab may be equipped with an additional |
102 | * freelist that allows lockless access to |
103 | * free objects in addition to the regular freelist |
104 | * that requires the slab lock. |
105 | * |
106 | * PageError Slab requires special handling due to debug |
107 | * options set. This moves slab handling out of |
108 | * the fast path and disables lockless freelists. |
109 | */ |
110 | |
111 | #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ |
112 | SLAB_TRACE | SLAB_DEBUG_FREE) |
113 | |
114 | static inline int kmem_cache_debug(struct kmem_cache *s) |
115 | { |
116 | #ifdef CONFIG_SLUB_DEBUG |
117 | return unlikely(s->flags & SLAB_DEBUG_FLAGS); |
118 | #else |
119 | return 0; |
120 | #endif |
121 | } |
122 | |
123 | /* |
124 | * Issues still to be resolved: |
125 | * |
126 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. |
127 | * |
128 | * - Variable sizing of the per node arrays |
129 | */ |
130 | |
131 | /* Enable to test recovery from slab corruption on boot */ |
132 | #undef SLUB_RESILIENCY_TEST |
133 | |
134 | /* |
135 | * Mininum number of partial slabs. These will be left on the partial |
136 | * lists even if they are empty. kmem_cache_shrink may reclaim them. |
137 | */ |
138 | #define MIN_PARTIAL 5 |
139 | |
140 | /* |
141 | * Maximum number of desirable partial slabs. |
142 | * The existence of more partial slabs makes kmem_cache_shrink |
143 | * sort the partial list by the number of objects in the. |
144 | */ |
145 | #define MAX_PARTIAL 10 |
146 | |
147 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ |
148 | SLAB_POISON | SLAB_STORE_USER) |
149 | |
150 | /* |
151 | * Debugging flags that require metadata to be stored in the slab. These get |
152 | * disabled when slub_debug=O is used and a cache's min order increases with |
153 | * metadata. |
154 | */ |
155 | #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) |
156 | |
157 | /* |
158 | * Set of flags that will prevent slab merging |
159 | */ |
160 | #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ |
161 | SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ |
162 | SLAB_FAILSLAB) |
163 | |
164 | #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ |
165 | SLAB_CACHE_DMA | SLAB_NOTRACK) |
166 | |
167 | #define OO_SHIFT 16 |
168 | #define OO_MASK ((1 << OO_SHIFT) - 1) |
169 | #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ |
170 | |
171 | /* Internal SLUB flags */ |
172 | #define __OBJECT_POISON 0x80000000UL /* Poison object */ |
173 | |
174 | static int kmem_size = sizeof(struct kmem_cache); |
175 | |
176 | #ifdef CONFIG_SMP |
177 | static struct notifier_block slab_notifier; |
178 | #endif |
179 | |
180 | static enum { |
181 | DOWN, /* No slab functionality available */ |
182 | PARTIAL, /* Kmem_cache_node works */ |
183 | UP, /* Everything works but does not show up in sysfs */ |
184 | SYSFS /* Sysfs up */ |
185 | } slab_state = DOWN; |
186 | |
187 | /* A list of all slab caches on the system */ |
188 | static DECLARE_RWSEM(slub_lock); |
189 | static LIST_HEAD(slab_caches); |
190 | |
191 | /* |
192 | * Tracking user of a slab. |
193 | */ |
194 | struct track { |
195 | unsigned long addr; /* Called from address */ |
196 | int cpu; /* Was running on cpu */ |
197 | int pid; /* Pid context */ |
198 | unsigned long when; /* When did the operation occur */ |
199 | }; |
200 | |
201 | enum track_item { TRACK_ALLOC, TRACK_FREE }; |
202 | |
203 | #ifdef CONFIG_SYSFS |
204 | static int sysfs_slab_add(struct kmem_cache *); |
205 | static int sysfs_slab_alias(struct kmem_cache *, const char *); |
206 | static void sysfs_slab_remove(struct kmem_cache *); |
207 | |
208 | #else |
209 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } |
210 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) |
211 | { return 0; } |
212 | static inline void sysfs_slab_remove(struct kmem_cache *s) |
213 | { |
214 | kfree(s->name); |
215 | kfree(s); |
216 | } |
217 | |
218 | #endif |
219 | |
220 | static inline void stat(const struct kmem_cache *s, enum stat_item si) |
221 | { |
222 | #ifdef CONFIG_SLUB_STATS |
223 | __this_cpu_inc(s->cpu_slab->stat[si]); |
224 | #endif |
225 | } |
226 | |
227 | /******************************************************************** |
228 | * Core slab cache functions |
229 | *******************************************************************/ |
230 | |
231 | int slab_is_available(void) |
232 | { |
233 | return slab_state >= UP; |
234 | } |
235 | |
236 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) |
237 | { |
238 | return s->node[node]; |
239 | } |
240 | |
241 | /* Verify that a pointer has an address that is valid within a slab page */ |
242 | static inline int check_valid_pointer(struct kmem_cache *s, |
243 | struct page *page, const void *object) |
244 | { |
245 | void *base; |
246 | |
247 | if (!object) |
248 | return 1; |
249 | |
250 | base = page_address(page); |
251 | if (object < base || object >= base + page->objects * s->size || |
252 | (object - base) % s->size) { |
253 | return 0; |
254 | } |
255 | |
256 | return 1; |
257 | } |
258 | |
259 | static inline void *get_freepointer(struct kmem_cache *s, void *object) |
260 | { |
261 | return *(void **)(object + s->offset); |
262 | } |
263 | |
264 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) |
265 | { |
266 | *(void **)(object + s->offset) = fp; |
267 | } |
268 | |
269 | /* Loop over all objects in a slab */ |
270 | #define for_each_object(__p, __s, __addr, __objects) \ |
271 | for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ |
272 | __p += (__s)->size) |
273 | |
274 | /* Scan freelist */ |
275 | #define for_each_free_object(__p, __s, __free) \ |
276 | for (__p = (__free); __p; __p = get_freepointer((__s), __p)) |
277 | |
278 | /* Determine object index from a given position */ |
279 | static inline int slab_index(void *p, struct kmem_cache *s, void *addr) |
280 | { |
281 | return (p - addr) / s->size; |
282 | } |
283 | |
284 | static inline size_t slab_ksize(const struct kmem_cache *s) |
285 | { |
286 | #ifdef CONFIG_SLUB_DEBUG |
287 | /* |
288 | * Debugging requires use of the padding between object |
289 | * and whatever may come after it. |
290 | */ |
291 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) |
292 | return s->objsize; |
293 | |
294 | #endif |
295 | /* |
296 | * If we have the need to store the freelist pointer |
297 | * back there or track user information then we can |
298 | * only use the space before that information. |
299 | */ |
300 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) |
301 | return s->inuse; |
302 | /* |
303 | * Else we can use all the padding etc for the allocation |
304 | */ |
305 | return s->size; |
306 | } |
307 | |
308 | static inline int order_objects(int order, unsigned long size, int reserved) |
309 | { |
310 | return ((PAGE_SIZE << order) - reserved) / size; |
311 | } |
312 | |
313 | static inline struct kmem_cache_order_objects oo_make(int order, |
314 | unsigned long size, int reserved) |
315 | { |
316 | struct kmem_cache_order_objects x = { |
317 | (order << OO_SHIFT) + order_objects(order, size, reserved) |
318 | }; |
319 | |
320 | return x; |
321 | } |
322 | |
323 | static inline int oo_order(struct kmem_cache_order_objects x) |
324 | { |
325 | return x.x >> OO_SHIFT; |
326 | } |
327 | |
328 | static inline int oo_objects(struct kmem_cache_order_objects x) |
329 | { |
330 | return x.x & OO_MASK; |
331 | } |
332 | |
333 | #ifdef CONFIG_SLUB_DEBUG |
334 | /* |
335 | * Debug settings: |
336 | */ |
337 | #ifdef CONFIG_SLUB_DEBUG_ON |
338 | static int slub_debug = DEBUG_DEFAULT_FLAGS; |
339 | #else |
340 | static int slub_debug; |
341 | #endif |
342 | |
343 | static char *slub_debug_slabs; |
344 | static int disable_higher_order_debug; |
345 | |
346 | /* |
347 | * Object debugging |
348 | */ |
349 | static void print_section(char *text, u8 *addr, unsigned int length) |
350 | { |
351 | int i, offset; |
352 | int newline = 1; |
353 | char ascii[17]; |
354 | |
355 | ascii[16] = 0; |
356 | |
357 | for (i = 0; i < length; i++) { |
358 | if (newline) { |
359 | printk(KERN_ERR "%8s 0x%p: ", text, addr + i); |
360 | newline = 0; |
361 | } |
362 | printk(KERN_CONT " %02x", addr[i]); |
363 | offset = i % 16; |
364 | ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; |
365 | if (offset == 15) { |
366 | printk(KERN_CONT " %s\n", ascii); |
367 | newline = 1; |
368 | } |
369 | } |
370 | if (!newline) { |
371 | i %= 16; |
372 | while (i < 16) { |
373 | printk(KERN_CONT " "); |
374 | ascii[i] = ' '; |
375 | i++; |
376 | } |
377 | printk(KERN_CONT " %s\n", ascii); |
378 | } |
379 | } |
380 | |
381 | static struct track *get_track(struct kmem_cache *s, void *object, |
382 | enum track_item alloc) |
383 | { |
384 | struct track *p; |
385 | |
386 | if (s->offset) |
387 | p = object + s->offset + sizeof(void *); |
388 | else |
389 | p = object + s->inuse; |
390 | |
391 | return p + alloc; |
392 | } |
393 | |
394 | static void set_track(struct kmem_cache *s, void *object, |
395 | enum track_item alloc, unsigned long addr) |
396 | { |
397 | struct track *p = get_track(s, object, alloc); |
398 | |
399 | if (addr) { |
400 | p->addr = addr; |
401 | p->cpu = smp_processor_id(); |
402 | p->pid = current->pid; |
403 | p->when = jiffies; |
404 | } else |
405 | memset(p, 0, sizeof(struct track)); |
406 | } |
407 | |
408 | static void init_tracking(struct kmem_cache *s, void *object) |
409 | { |
410 | if (!(s->flags & SLAB_STORE_USER)) |
411 | return; |
412 | |
413 | set_track(s, object, TRACK_FREE, 0UL); |
414 | set_track(s, object, TRACK_ALLOC, 0UL); |
415 | } |
416 | |
417 | static void print_track(const char *s, struct track *t) |
418 | { |
419 | if (!t->addr) |
420 | return; |
421 | |
422 | printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", |
423 | s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); |
424 | } |
425 | |
426 | static void print_tracking(struct kmem_cache *s, void *object) |
427 | { |
428 | if (!(s->flags & SLAB_STORE_USER)) |
429 | return; |
430 | |
431 | print_track("Allocated", get_track(s, object, TRACK_ALLOC)); |
432 | print_track("Freed", get_track(s, object, TRACK_FREE)); |
433 | } |
434 | |
435 | static void print_page_info(struct page *page) |
436 | { |
437 | printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", |
438 | page, page->objects, page->inuse, page->freelist, page->flags); |
439 | |
440 | } |
441 | |
442 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) |
443 | { |
444 | va_list args; |
445 | char buf[100]; |
446 | |
447 | va_start(args, fmt); |
448 | vsnprintf(buf, sizeof(buf), fmt, args); |
449 | va_end(args); |
450 | printk(KERN_ERR "========================================" |
451 | "=====================================\n"); |
452 | printk(KERN_ERR "BUG %s: %s\n", s->name, buf); |
453 | printk(KERN_ERR "----------------------------------------" |
454 | "-------------------------------------\n\n"); |
455 | } |
456 | |
457 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) |
458 | { |
459 | va_list args; |
460 | char buf[100]; |
461 | |
462 | va_start(args, fmt); |
463 | vsnprintf(buf, sizeof(buf), fmt, args); |
464 | va_end(args); |
465 | printk(KERN_ERR "FIX %s: %s\n", s->name, buf); |
466 | } |
467 | |
468 | static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) |
469 | { |
470 | unsigned int off; /* Offset of last byte */ |
471 | u8 *addr = page_address(page); |
472 | |
473 | print_tracking(s, p); |
474 | |
475 | print_page_info(page); |
476 | |
477 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", |
478 | p, p - addr, get_freepointer(s, p)); |
479 | |
480 | if (p > addr + 16) |
481 | print_section("Bytes b4", p - 16, 16); |
482 | |
483 | print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); |
484 | |
485 | if (s->flags & SLAB_RED_ZONE) |
486 | print_section("Redzone", p + s->objsize, |
487 | s->inuse - s->objsize); |
488 | |
489 | if (s->offset) |
490 | off = s->offset + sizeof(void *); |
491 | else |
492 | off = s->inuse; |
493 | |
494 | if (s->flags & SLAB_STORE_USER) |
495 | off += 2 * sizeof(struct track); |
496 | |
497 | if (off != s->size) |
498 | /* Beginning of the filler is the free pointer */ |
499 | print_section("Padding", p + off, s->size - off); |
500 | |
501 | dump_stack(); |
502 | } |
503 | |
504 | static void object_err(struct kmem_cache *s, struct page *page, |
505 | u8 *object, char *reason) |
506 | { |
507 | slab_bug(s, "%s", reason); |
508 | print_trailer(s, page, object); |
509 | } |
510 | |
511 | static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) |
512 | { |
513 | va_list args; |
514 | char buf[100]; |
515 | |
516 | va_start(args, fmt); |
517 | vsnprintf(buf, sizeof(buf), fmt, args); |
518 | va_end(args); |
519 | slab_bug(s, "%s", buf); |
520 | print_page_info(page); |
521 | dump_stack(); |
522 | } |
523 | |
524 | static void init_object(struct kmem_cache *s, void *object, u8 val) |
525 | { |
526 | u8 *p = object; |
527 | |
528 | if (s->flags & __OBJECT_POISON) { |
529 | memset(p, POISON_FREE, s->objsize - 1); |
530 | p[s->objsize - 1] = POISON_END; |
531 | } |
532 | |
533 | if (s->flags & SLAB_RED_ZONE) |
534 | memset(p + s->objsize, val, s->inuse - s->objsize); |
535 | } |
536 | |
537 | static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) |
538 | { |
539 | while (bytes) { |
540 | if (*start != (u8)value) |
541 | return start; |
542 | start++; |
543 | bytes--; |
544 | } |
545 | return NULL; |
546 | } |
547 | |
548 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, |
549 | void *from, void *to) |
550 | { |
551 | slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); |
552 | memset(from, data, to - from); |
553 | } |
554 | |
555 | static int check_bytes_and_report(struct kmem_cache *s, struct page *page, |
556 | u8 *object, char *what, |
557 | u8 *start, unsigned int value, unsigned int bytes) |
558 | { |
559 | u8 *fault; |
560 | u8 *end; |
561 | |
562 | fault = check_bytes(start, value, bytes); |
563 | if (!fault) |
564 | return 1; |
565 | |
566 | end = start + bytes; |
567 | while (end > fault && end[-1] == value) |
568 | end--; |
569 | |
570 | slab_bug(s, "%s overwritten", what); |
571 | printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", |
572 | fault, end - 1, fault[0], value); |
573 | print_trailer(s, page, object); |
574 | |
575 | restore_bytes(s, what, value, fault, end); |
576 | return 0; |
577 | } |
578 | |
579 | /* |
580 | * Object layout: |
581 | * |
582 | * object address |
583 | * Bytes of the object to be managed. |
584 | * If the freepointer may overlay the object then the free |
585 | * pointer is the first word of the object. |
586 | * |
587 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is |
588 | * 0xa5 (POISON_END) |
589 | * |
590 | * object + s->objsize |
591 | * Padding to reach word boundary. This is also used for Redzoning. |
592 | * Padding is extended by another word if Redzoning is enabled and |
593 | * objsize == inuse. |
594 | * |
595 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with |
596 | * 0xcc (RED_ACTIVE) for objects in use. |
597 | * |
598 | * object + s->inuse |
599 | * Meta data starts here. |
600 | * |
601 | * A. Free pointer (if we cannot overwrite object on free) |
602 | * B. Tracking data for SLAB_STORE_USER |
603 | * C. Padding to reach required alignment boundary or at mininum |
604 | * one word if debugging is on to be able to detect writes |
605 | * before the word boundary. |
606 | * |
607 | * Padding is done using 0x5a (POISON_INUSE) |
608 | * |
609 | * object + s->size |
610 | * Nothing is used beyond s->size. |
611 | * |
612 | * If slabcaches are merged then the objsize and inuse boundaries are mostly |
613 | * ignored. And therefore no slab options that rely on these boundaries |
614 | * may be used with merged slabcaches. |
615 | */ |
616 | |
617 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) |
618 | { |
619 | unsigned long off = s->inuse; /* The end of info */ |
620 | |
621 | if (s->offset) |
622 | /* Freepointer is placed after the object. */ |
623 | off += sizeof(void *); |
624 | |
625 | if (s->flags & SLAB_STORE_USER) |
626 | /* We also have user information there */ |
627 | off += 2 * sizeof(struct track); |
628 | |
629 | if (s->size == off) |
630 | return 1; |
631 | |
632 | return check_bytes_and_report(s, page, p, "Object padding", |
633 | p + off, POISON_INUSE, s->size - off); |
634 | } |
635 | |
636 | /* Check the pad bytes at the end of a slab page */ |
637 | static int slab_pad_check(struct kmem_cache *s, struct page *page) |
638 | { |
639 | u8 *start; |
640 | u8 *fault; |
641 | u8 *end; |
642 | int length; |
643 | int remainder; |
644 | |
645 | if (!(s->flags & SLAB_POISON)) |
646 | return 1; |
647 | |
648 | start = page_address(page); |
649 | length = (PAGE_SIZE << compound_order(page)) - s->reserved; |
650 | end = start + length; |
651 | remainder = length % s->size; |
652 | if (!remainder) |
653 | return 1; |
654 | |
655 | fault = check_bytes(end - remainder, POISON_INUSE, remainder); |
656 | if (!fault) |
657 | return 1; |
658 | while (end > fault && end[-1] == POISON_INUSE) |
659 | end--; |
660 | |
661 | slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); |
662 | print_section("Padding", end - remainder, remainder); |
663 | |
664 | restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); |
665 | return 0; |
666 | } |
667 | |
668 | static int check_object(struct kmem_cache *s, struct page *page, |
669 | void *object, u8 val) |
670 | { |
671 | u8 *p = object; |
672 | u8 *endobject = object + s->objsize; |
673 | |
674 | if (s->flags & SLAB_RED_ZONE) { |
675 | if (!check_bytes_and_report(s, page, object, "Redzone", |
676 | endobject, val, s->inuse - s->objsize)) |
677 | return 0; |
678 | } else { |
679 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { |
680 | check_bytes_and_report(s, page, p, "Alignment padding", |
681 | endobject, POISON_INUSE, s->inuse - s->objsize); |
682 | } |
683 | } |
684 | |
685 | if (s->flags & SLAB_POISON) { |
686 | if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && |
687 | (!check_bytes_and_report(s, page, p, "Poison", p, |
688 | POISON_FREE, s->objsize - 1) || |
689 | !check_bytes_and_report(s, page, p, "Poison", |
690 | p + s->objsize - 1, POISON_END, 1))) |
691 | return 0; |
692 | /* |
693 | * check_pad_bytes cleans up on its own. |
694 | */ |
695 | check_pad_bytes(s, page, p); |
696 | } |
697 | |
698 | if (!s->offset && val == SLUB_RED_ACTIVE) |
699 | /* |
700 | * Object and freepointer overlap. Cannot check |
701 | * freepointer while object is allocated. |
702 | */ |
703 | return 1; |
704 | |
705 | /* Check free pointer validity */ |
706 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { |
707 | object_err(s, page, p, "Freepointer corrupt"); |
708 | /* |
709 | * No choice but to zap it and thus lose the remainder |
710 | * of the free objects in this slab. May cause |
711 | * another error because the object count is now wrong. |
712 | */ |
713 | set_freepointer(s, p, NULL); |
714 | return 0; |
715 | } |
716 | return 1; |
717 | } |
718 | |
719 | static int check_slab(struct kmem_cache *s, struct page *page) |
720 | { |
721 | int maxobj; |
722 | |
723 | VM_BUG_ON(!irqs_disabled()); |
724 | |
725 | if (!PageSlab(page)) { |
726 | slab_err(s, page, "Not a valid slab page"); |
727 | return 0; |
728 | } |
729 | |
730 | maxobj = order_objects(compound_order(page), s->size, s->reserved); |
731 | if (page->objects > maxobj) { |
732 | slab_err(s, page, "objects %u > max %u", |
733 | s->name, page->objects, maxobj); |
734 | return 0; |
735 | } |
736 | if (page->inuse > page->objects) { |
737 | slab_err(s, page, "inuse %u > max %u", |
738 | s->name, page->inuse, page->objects); |
739 | return 0; |
740 | } |
741 | /* Slab_pad_check fixes things up after itself */ |
742 | slab_pad_check(s, page); |
743 | return 1; |
744 | } |
745 | |
746 | /* |
747 | * Determine if a certain object on a page is on the freelist. Must hold the |
748 | * slab lock to guarantee that the chains are in a consistent state. |
749 | */ |
750 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) |
751 | { |
752 | int nr = 0; |
753 | void *fp = page->freelist; |
754 | void *object = NULL; |
755 | unsigned long max_objects; |
756 | |
757 | while (fp && nr <= page->objects) { |
758 | if (fp == search) |
759 | return 1; |
760 | if (!check_valid_pointer(s, page, fp)) { |
761 | if (object) { |
762 | object_err(s, page, object, |
763 | "Freechain corrupt"); |
764 | set_freepointer(s, object, NULL); |
765 | break; |
766 | } else { |
767 | slab_err(s, page, "Freepointer corrupt"); |
768 | page->freelist = NULL; |
769 | page->inuse = page->objects; |
770 | slab_fix(s, "Freelist cleared"); |
771 | return 0; |
772 | } |
773 | break; |
774 | } |
775 | object = fp; |
776 | fp = get_freepointer(s, object); |
777 | nr++; |
778 | } |
779 | |
780 | max_objects = order_objects(compound_order(page), s->size, s->reserved); |
781 | if (max_objects > MAX_OBJS_PER_PAGE) |
782 | max_objects = MAX_OBJS_PER_PAGE; |
783 | |
784 | if (page->objects != max_objects) { |
785 | slab_err(s, page, "Wrong number of objects. Found %d but " |
786 | "should be %d", page->objects, max_objects); |
787 | page->objects = max_objects; |
788 | slab_fix(s, "Number of objects adjusted."); |
789 | } |
790 | if (page->inuse != page->objects - nr) { |
791 | slab_err(s, page, "Wrong object count. Counter is %d but " |
792 | "counted were %d", page->inuse, page->objects - nr); |
793 | page->inuse = page->objects - nr; |
794 | slab_fix(s, "Object count adjusted."); |
795 | } |
796 | return search == NULL; |
797 | } |
798 | |
799 | static void trace(struct kmem_cache *s, struct page *page, void *object, |
800 | int alloc) |
801 | { |
802 | if (s->flags & SLAB_TRACE) { |
803 | printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", |
804 | s->name, |
805 | alloc ? "alloc" : "free", |
806 | object, page->inuse, |
807 | page->freelist); |
808 | |
809 | if (!alloc) |
810 | print_section("Object", (void *)object, s->objsize); |
811 | |
812 | dump_stack(); |
813 | } |
814 | } |
815 | |
816 | /* |
817 | * Hooks for other subsystems that check memory allocations. In a typical |
818 | * production configuration these hooks all should produce no code at all. |
819 | */ |
820 | static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) |
821 | { |
822 | flags &= gfp_allowed_mask; |
823 | lockdep_trace_alloc(flags); |
824 | might_sleep_if(flags & __GFP_WAIT); |
825 | |
826 | return should_failslab(s->objsize, flags, s->flags); |
827 | } |
828 | |
829 | static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object) |
830 | { |
831 | flags &= gfp_allowed_mask; |
832 | kmemcheck_slab_alloc(s, flags, object, slab_ksize(s)); |
833 | kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags); |
834 | } |
835 | |
836 | static inline void slab_free_hook(struct kmem_cache *s, void *x) |
837 | { |
838 | kmemleak_free_recursive(x, s->flags); |
839 | |
840 | /* |
841 | * Trouble is that we may no longer disable interupts in the fast path |
842 | * So in order to make the debug calls that expect irqs to be |
843 | * disabled we need to disable interrupts temporarily. |
844 | */ |
845 | #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP) |
846 | { |
847 | unsigned long flags; |
848 | |
849 | local_irq_save(flags); |
850 | kmemcheck_slab_free(s, x, s->objsize); |
851 | debug_check_no_locks_freed(x, s->objsize); |
852 | local_irq_restore(flags); |
853 | } |
854 | #endif |
855 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) |
856 | debug_check_no_obj_freed(x, s->objsize); |
857 | } |
858 | |
859 | /* |
860 | * Tracking of fully allocated slabs for debugging purposes. |
861 | */ |
862 | static void add_full(struct kmem_cache_node *n, struct page *page) |
863 | { |
864 | spin_lock(&n->list_lock); |
865 | list_add(&page->lru, &n->full); |
866 | spin_unlock(&n->list_lock); |
867 | } |
868 | |
869 | static void remove_full(struct kmem_cache *s, struct page *page) |
870 | { |
871 | struct kmem_cache_node *n; |
872 | |
873 | if (!(s->flags & SLAB_STORE_USER)) |
874 | return; |
875 | |
876 | n = get_node(s, page_to_nid(page)); |
877 | |
878 | spin_lock(&n->list_lock); |
879 | list_del(&page->lru); |
880 | spin_unlock(&n->list_lock); |
881 | } |
882 | |
883 | /* Tracking of the number of slabs for debugging purposes */ |
884 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) |
885 | { |
886 | struct kmem_cache_node *n = get_node(s, node); |
887 | |
888 | return atomic_long_read(&n->nr_slabs); |
889 | } |
890 | |
891 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
892 | { |
893 | return atomic_long_read(&n->nr_slabs); |
894 | } |
895 | |
896 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) |
897 | { |
898 | struct kmem_cache_node *n = get_node(s, node); |
899 | |
900 | /* |
901 | * May be called early in order to allocate a slab for the |
902 | * kmem_cache_node structure. Solve the chicken-egg |
903 | * dilemma by deferring the increment of the count during |
904 | * bootstrap (see early_kmem_cache_node_alloc). |
905 | */ |
906 | if (n) { |
907 | atomic_long_inc(&n->nr_slabs); |
908 | atomic_long_add(objects, &n->total_objects); |
909 | } |
910 | } |
911 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) |
912 | { |
913 | struct kmem_cache_node *n = get_node(s, node); |
914 | |
915 | atomic_long_dec(&n->nr_slabs); |
916 | atomic_long_sub(objects, &n->total_objects); |
917 | } |
918 | |
919 | /* Object debug checks for alloc/free paths */ |
920 | static void setup_object_debug(struct kmem_cache *s, struct page *page, |
921 | void *object) |
922 | { |
923 | if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) |
924 | return; |
925 | |
926 | init_object(s, object, SLUB_RED_INACTIVE); |
927 | init_tracking(s, object); |
928 | } |
929 | |
930 | static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page, |
931 | void *object, unsigned long addr) |
932 | { |
933 | if (!check_slab(s, page)) |
934 | goto bad; |
935 | |
936 | if (!on_freelist(s, page, object)) { |
937 | object_err(s, page, object, "Object already allocated"); |
938 | goto bad; |
939 | } |
940 | |
941 | if (!check_valid_pointer(s, page, object)) { |
942 | object_err(s, page, object, "Freelist Pointer check fails"); |
943 | goto bad; |
944 | } |
945 | |
946 | if (!check_object(s, page, object, SLUB_RED_INACTIVE)) |
947 | goto bad; |
948 | |
949 | /* Success perform special debug activities for allocs */ |
950 | if (s->flags & SLAB_STORE_USER) |
951 | set_track(s, object, TRACK_ALLOC, addr); |
952 | trace(s, page, object, 1); |
953 | init_object(s, object, SLUB_RED_ACTIVE); |
954 | return 1; |
955 | |
956 | bad: |
957 | if (PageSlab(page)) { |
958 | /* |
959 | * If this is a slab page then lets do the best we can |
960 | * to avoid issues in the future. Marking all objects |
961 | * as used avoids touching the remaining objects. |
962 | */ |
963 | slab_fix(s, "Marking all objects used"); |
964 | page->inuse = page->objects; |
965 | page->freelist = NULL; |
966 | } |
967 | return 0; |
968 | } |
969 | |
970 | static noinline int free_debug_processing(struct kmem_cache *s, |
971 | struct page *page, void *object, unsigned long addr) |
972 | { |
973 | if (!check_slab(s, page)) |
974 | goto fail; |
975 | |
976 | if (!check_valid_pointer(s, page, object)) { |
977 | slab_err(s, page, "Invalid object pointer 0x%p", object); |
978 | goto fail; |
979 | } |
980 | |
981 | if (on_freelist(s, page, object)) { |
982 | object_err(s, page, object, "Object already free"); |
983 | goto fail; |
984 | } |
985 | |
986 | if (!check_object(s, page, object, SLUB_RED_ACTIVE)) |
987 | return 0; |
988 | |
989 | if (unlikely(s != page->slab)) { |
990 | if (!PageSlab(page)) { |
991 | slab_err(s, page, "Attempt to free object(0x%p) " |
992 | "outside of slab", object); |
993 | } else if (!page->slab) { |
994 | printk(KERN_ERR |
995 | "SLUB <none>: no slab for object 0x%p.\n", |
996 | object); |
997 | dump_stack(); |
998 | } else |
999 | object_err(s, page, object, |
1000 | "page slab pointer corrupt."); |
1001 | goto fail; |
1002 | } |
1003 | |
1004 | /* Special debug activities for freeing objects */ |
1005 | if (!PageSlubFrozen(page) && !page->freelist) |
1006 | remove_full(s, page); |
1007 | if (s->flags & SLAB_STORE_USER) |
1008 | set_track(s, object, TRACK_FREE, addr); |
1009 | trace(s, page, object, 0); |
1010 | init_object(s, object, SLUB_RED_INACTIVE); |
1011 | return 1; |
1012 | |
1013 | fail: |
1014 | slab_fix(s, "Object at 0x%p not freed", object); |
1015 | return 0; |
1016 | } |
1017 | |
1018 | static int __init setup_slub_debug(char *str) |
1019 | { |
1020 | slub_debug = DEBUG_DEFAULT_FLAGS; |
1021 | if (*str++ != '=' || !*str) |
1022 | /* |
1023 | * No options specified. Switch on full debugging. |
1024 | */ |
1025 | goto out; |
1026 | |
1027 | if (*str == ',') |
1028 | /* |
1029 | * No options but restriction on slabs. This means full |
1030 | * debugging for slabs matching a pattern. |
1031 | */ |
1032 | goto check_slabs; |
1033 | |
1034 | if (tolower(*str) == 'o') { |
1035 | /* |
1036 | * Avoid enabling debugging on caches if its minimum order |
1037 | * would increase as a result. |
1038 | */ |
1039 | disable_higher_order_debug = 1; |
1040 | goto out; |
1041 | } |
1042 | |
1043 | slub_debug = 0; |
1044 | if (*str == '-') |
1045 | /* |
1046 | * Switch off all debugging measures. |
1047 | */ |
1048 | goto out; |
1049 | |
1050 | /* |
1051 | * Determine which debug features should be switched on |
1052 | */ |
1053 | for (; *str && *str != ','; str++) { |
1054 | switch (tolower(*str)) { |
1055 | case 'f': |
1056 | slub_debug |= SLAB_DEBUG_FREE; |
1057 | break; |
1058 | case 'z': |
1059 | slub_debug |= SLAB_RED_ZONE; |
1060 | break; |
1061 | case 'p': |
1062 | slub_debug |= SLAB_POISON; |
1063 | break; |
1064 | case 'u': |
1065 | slub_debug |= SLAB_STORE_USER; |
1066 | break; |
1067 | case 't': |
1068 | slub_debug |= SLAB_TRACE; |
1069 | break; |
1070 | case 'a': |
1071 | slub_debug |= SLAB_FAILSLAB; |
1072 | break; |
1073 | default: |
1074 | printk(KERN_ERR "slub_debug option '%c' " |
1075 | "unknown. skipped\n", *str); |
1076 | } |
1077 | } |
1078 | |
1079 | check_slabs: |
1080 | if (*str == ',') |
1081 | slub_debug_slabs = str + 1; |
1082 | out: |
1083 | return 1; |
1084 | } |
1085 | |
1086 | __setup("slub_debug", setup_slub_debug); |
1087 | |
1088 | static unsigned long kmem_cache_flags(unsigned long objsize, |
1089 | unsigned long flags, const char *name, |
1090 | void (*ctor)(void *)) |
1091 | { |
1092 | /* |
1093 | * Enable debugging if selected on the kernel commandline. |
1094 | */ |
1095 | if (slub_debug && (!slub_debug_slabs || |
1096 | !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) |
1097 | flags |= slub_debug; |
1098 | |
1099 | return flags; |
1100 | } |
1101 | #else |
1102 | static inline void setup_object_debug(struct kmem_cache *s, |
1103 | struct page *page, void *object) {} |
1104 | |
1105 | static inline int alloc_debug_processing(struct kmem_cache *s, |
1106 | struct page *page, void *object, unsigned long addr) { return 0; } |
1107 | |
1108 | static inline int free_debug_processing(struct kmem_cache *s, |
1109 | struct page *page, void *object, unsigned long addr) { return 0; } |
1110 | |
1111 | static inline int slab_pad_check(struct kmem_cache *s, struct page *page) |
1112 | { return 1; } |
1113 | static inline int check_object(struct kmem_cache *s, struct page *page, |
1114 | void *object, u8 val) { return 1; } |
1115 | static inline void add_full(struct kmem_cache_node *n, struct page *page) {} |
1116 | static inline unsigned long kmem_cache_flags(unsigned long objsize, |
1117 | unsigned long flags, const char *name, |
1118 | void (*ctor)(void *)) |
1119 | { |
1120 | return flags; |
1121 | } |
1122 | #define slub_debug 0 |
1123 | |
1124 | #define disable_higher_order_debug 0 |
1125 | |
1126 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) |
1127 | { return 0; } |
1128 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
1129 | { return 0; } |
1130 | static inline void inc_slabs_node(struct kmem_cache *s, int node, |
1131 | int objects) {} |
1132 | static inline void dec_slabs_node(struct kmem_cache *s, int node, |
1133 | int objects) {} |
1134 | |
1135 | static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) |
1136 | { return 0; } |
1137 | |
1138 | static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, |
1139 | void *object) {} |
1140 | |
1141 | static inline void slab_free_hook(struct kmem_cache *s, void *x) {} |
1142 | |
1143 | #endif /* CONFIG_SLUB_DEBUG */ |
1144 | |
1145 | /* |
1146 | * Slab allocation and freeing |
1147 | */ |
1148 | static inline struct page *alloc_slab_page(gfp_t flags, int node, |
1149 | struct kmem_cache_order_objects oo) |
1150 | { |
1151 | int order = oo_order(oo); |
1152 | |
1153 | flags |= __GFP_NOTRACK; |
1154 | |
1155 | if (node == NUMA_NO_NODE) |
1156 | return alloc_pages(flags, order); |
1157 | else |
1158 | return alloc_pages_exact_node(node, flags, order); |
1159 | } |
1160 | |
1161 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) |
1162 | { |
1163 | struct page *page; |
1164 | struct kmem_cache_order_objects oo = s->oo; |
1165 | gfp_t alloc_gfp; |
1166 | |
1167 | flags |= s->allocflags; |
1168 | |
1169 | /* |
1170 | * Let the initial higher-order allocation fail under memory pressure |
1171 | * so we fall-back to the minimum order allocation. |
1172 | */ |
1173 | alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; |
1174 | |
1175 | page = alloc_slab_page(alloc_gfp, node, oo); |
1176 | if (unlikely(!page)) { |
1177 | oo = s->min; |
1178 | /* |
1179 | * Allocation may have failed due to fragmentation. |
1180 | * Try a lower order alloc if possible |
1181 | */ |
1182 | page = alloc_slab_page(flags, node, oo); |
1183 | if (!page) |
1184 | return NULL; |
1185 | |
1186 | stat(s, ORDER_FALLBACK); |
1187 | } |
1188 | |
1189 | if (kmemcheck_enabled |
1190 | && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { |
1191 | int pages = 1 << oo_order(oo); |
1192 | |
1193 | kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); |
1194 | |
1195 | /* |
1196 | * Objects from caches that have a constructor don't get |
1197 | * cleared when they're allocated, so we need to do it here. |
1198 | */ |
1199 | if (s->ctor) |
1200 | kmemcheck_mark_uninitialized_pages(page, pages); |
1201 | else |
1202 | kmemcheck_mark_unallocated_pages(page, pages); |
1203 | } |
1204 | |
1205 | page->objects = oo_objects(oo); |
1206 | mod_zone_page_state(page_zone(page), |
1207 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? |
1208 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, |
1209 | 1 << oo_order(oo)); |
1210 | |
1211 | return page; |
1212 | } |
1213 | |
1214 | static void setup_object(struct kmem_cache *s, struct page *page, |
1215 | void *object) |
1216 | { |
1217 | setup_object_debug(s, page, object); |
1218 | if (unlikely(s->ctor)) |
1219 | s->ctor(object); |
1220 | } |
1221 | |
1222 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) |
1223 | { |
1224 | struct page *page; |
1225 | void *start; |
1226 | void *last; |
1227 | void *p; |
1228 | |
1229 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
1230 | |
1231 | page = allocate_slab(s, |
1232 | flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); |
1233 | if (!page) |
1234 | goto out; |
1235 | |
1236 | inc_slabs_node(s, page_to_nid(page), page->objects); |
1237 | page->slab = s; |
1238 | page->flags |= 1 << PG_slab; |
1239 | |
1240 | start = page_address(page); |
1241 | |
1242 | if (unlikely(s->flags & SLAB_POISON)) |
1243 | memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); |
1244 | |
1245 | last = start; |
1246 | for_each_object(p, s, start, page->objects) { |
1247 | setup_object(s, page, last); |
1248 | set_freepointer(s, last, p); |
1249 | last = p; |
1250 | } |
1251 | setup_object(s, page, last); |
1252 | set_freepointer(s, last, NULL); |
1253 | |
1254 | page->freelist = start; |
1255 | page->inuse = 0; |
1256 | out: |
1257 | return page; |
1258 | } |
1259 | |
1260 | static void __free_slab(struct kmem_cache *s, struct page *page) |
1261 | { |
1262 | int order = compound_order(page); |
1263 | int pages = 1 << order; |
1264 | |
1265 | if (kmem_cache_debug(s)) { |
1266 | void *p; |
1267 | |
1268 | slab_pad_check(s, page); |
1269 | for_each_object(p, s, page_address(page), |
1270 | page->objects) |
1271 | check_object(s, page, p, SLUB_RED_INACTIVE); |
1272 | } |
1273 | |
1274 | kmemcheck_free_shadow(page, compound_order(page)); |
1275 | |
1276 | mod_zone_page_state(page_zone(page), |
1277 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? |
1278 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, |
1279 | -pages); |
1280 | |
1281 | __ClearPageSlab(page); |
1282 | reset_page_mapcount(page); |
1283 | if (current->reclaim_state) |
1284 | current->reclaim_state->reclaimed_slab += pages; |
1285 | __free_pages(page, order); |
1286 | } |
1287 | |
1288 | #define need_reserve_slab_rcu \ |
1289 | (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) |
1290 | |
1291 | static void rcu_free_slab(struct rcu_head *h) |
1292 | { |
1293 | struct page *page; |
1294 | |
1295 | if (need_reserve_slab_rcu) |
1296 | page = virt_to_head_page(h); |
1297 | else |
1298 | page = container_of((struct list_head *)h, struct page, lru); |
1299 | |
1300 | __free_slab(page->slab, page); |
1301 | } |
1302 | |
1303 | static void free_slab(struct kmem_cache *s, struct page *page) |
1304 | { |
1305 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { |
1306 | struct rcu_head *head; |
1307 | |
1308 | if (need_reserve_slab_rcu) { |
1309 | int order = compound_order(page); |
1310 | int offset = (PAGE_SIZE << order) - s->reserved; |
1311 | |
1312 | VM_BUG_ON(s->reserved != sizeof(*head)); |
1313 | head = page_address(page) + offset; |
1314 | } else { |
1315 | /* |
1316 | * RCU free overloads the RCU head over the LRU |
1317 | */ |
1318 | head = (void *)&page->lru; |
1319 | } |
1320 | |
1321 | call_rcu(head, rcu_free_slab); |
1322 | } else |
1323 | __free_slab(s, page); |
1324 | } |
1325 | |
1326 | static void discard_slab(struct kmem_cache *s, struct page *page) |
1327 | { |
1328 | dec_slabs_node(s, page_to_nid(page), page->objects); |
1329 | free_slab(s, page); |
1330 | } |
1331 | |
1332 | /* |
1333 | * Per slab locking using the pagelock |
1334 | */ |
1335 | static __always_inline void slab_lock(struct page *page) |
1336 | { |
1337 | bit_spin_lock(PG_locked, &page->flags); |
1338 | } |
1339 | |
1340 | static __always_inline void slab_unlock(struct page *page) |
1341 | { |
1342 | __bit_spin_unlock(PG_locked, &page->flags); |
1343 | } |
1344 | |
1345 | static __always_inline int slab_trylock(struct page *page) |
1346 | { |
1347 | int rc = 1; |
1348 | |
1349 | rc = bit_spin_trylock(PG_locked, &page->flags); |
1350 | return rc; |
1351 | } |
1352 | |
1353 | /* |
1354 | * Management of partially allocated slabs |
1355 | */ |
1356 | static void add_partial(struct kmem_cache_node *n, |
1357 | struct page *page, int tail) |
1358 | { |
1359 | spin_lock(&n->list_lock); |
1360 | n->nr_partial++; |
1361 | if (tail) |
1362 | list_add_tail(&page->lru, &n->partial); |
1363 | else |
1364 | list_add(&page->lru, &n->partial); |
1365 | spin_unlock(&n->list_lock); |
1366 | } |
1367 | |
1368 | static inline void __remove_partial(struct kmem_cache_node *n, |
1369 | struct page *page) |
1370 | { |
1371 | list_del(&page->lru); |
1372 | n->nr_partial--; |
1373 | } |
1374 | |
1375 | static void remove_partial(struct kmem_cache *s, struct page *page) |
1376 | { |
1377 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); |
1378 | |
1379 | spin_lock(&n->list_lock); |
1380 | __remove_partial(n, page); |
1381 | spin_unlock(&n->list_lock); |
1382 | } |
1383 | |
1384 | /* |
1385 | * Lock slab and remove from the partial list. |
1386 | * |
1387 | * Must hold list_lock. |
1388 | */ |
1389 | static inline int lock_and_freeze_slab(struct kmem_cache_node *n, |
1390 | struct page *page) |
1391 | { |
1392 | if (slab_trylock(page)) { |
1393 | __remove_partial(n, page); |
1394 | __SetPageSlubFrozen(page); |
1395 | return 1; |
1396 | } |
1397 | return 0; |
1398 | } |
1399 | |
1400 | /* |
1401 | * Try to allocate a partial slab from a specific node. |
1402 | */ |
1403 | static struct page *get_partial_node(struct kmem_cache_node *n) |
1404 | { |
1405 | struct page *page; |
1406 | |
1407 | /* |
1408 | * Racy check. If we mistakenly see no partial slabs then we |
1409 | * just allocate an empty slab. If we mistakenly try to get a |
1410 | * partial slab and there is none available then get_partials() |
1411 | * will return NULL. |
1412 | */ |
1413 | if (!n || !n->nr_partial) |
1414 | return NULL; |
1415 | |
1416 | spin_lock(&n->list_lock); |
1417 | list_for_each_entry(page, &n->partial, lru) |
1418 | if (lock_and_freeze_slab(n, page)) |
1419 | goto out; |
1420 | page = NULL; |
1421 | out: |
1422 | spin_unlock(&n->list_lock); |
1423 | return page; |
1424 | } |
1425 | |
1426 | /* |
1427 | * Get a page from somewhere. Search in increasing NUMA distances. |
1428 | */ |
1429 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) |
1430 | { |
1431 | #ifdef CONFIG_NUMA |
1432 | struct zonelist *zonelist; |
1433 | struct zoneref *z; |
1434 | struct zone *zone; |
1435 | enum zone_type high_zoneidx = gfp_zone(flags); |
1436 | struct page *page; |
1437 | |
1438 | /* |
1439 | * The defrag ratio allows a configuration of the tradeoffs between |
1440 | * inter node defragmentation and node local allocations. A lower |
1441 | * defrag_ratio increases the tendency to do local allocations |
1442 | * instead of attempting to obtain partial slabs from other nodes. |
1443 | * |
1444 | * If the defrag_ratio is set to 0 then kmalloc() always |
1445 | * returns node local objects. If the ratio is higher then kmalloc() |
1446 | * may return off node objects because partial slabs are obtained |
1447 | * from other nodes and filled up. |
1448 | * |
1449 | * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes |
1450 | * defrag_ratio = 1000) then every (well almost) allocation will |
1451 | * first attempt to defrag slab caches on other nodes. This means |
1452 | * scanning over all nodes to look for partial slabs which may be |
1453 | * expensive if we do it every time we are trying to find a slab |
1454 | * with available objects. |
1455 | */ |
1456 | if (!s->remote_node_defrag_ratio || |
1457 | get_cycles() % 1024 > s->remote_node_defrag_ratio) |
1458 | return NULL; |
1459 | |
1460 | get_mems_allowed(); |
1461 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); |
1462 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
1463 | struct kmem_cache_node *n; |
1464 | |
1465 | n = get_node(s, zone_to_nid(zone)); |
1466 | |
1467 | if (n && cpuset_zone_allowed_hardwall(zone, flags) && |
1468 | n->nr_partial > s->min_partial) { |
1469 | page = get_partial_node(n); |
1470 | if (page) { |
1471 | put_mems_allowed(); |
1472 | return page; |
1473 | } |
1474 | } |
1475 | } |
1476 | put_mems_allowed(); |
1477 | #endif |
1478 | return NULL; |
1479 | } |
1480 | |
1481 | /* |
1482 | * Get a partial page, lock it and return it. |
1483 | */ |
1484 | static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) |
1485 | { |
1486 | struct page *page; |
1487 | int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; |
1488 | |
1489 | page = get_partial_node(get_node(s, searchnode)); |
1490 | if (page || node != -1) |
1491 | return page; |
1492 | |
1493 | return get_any_partial(s, flags); |
1494 | } |
1495 | |
1496 | /* |
1497 | * Move a page back to the lists. |
1498 | * |
1499 | * Must be called with the slab lock held. |
1500 | * |
1501 | * On exit the slab lock will have been dropped. |
1502 | */ |
1503 | static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) |
1504 | __releases(bitlock) |
1505 | { |
1506 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); |
1507 | |
1508 | __ClearPageSlubFrozen(page); |
1509 | if (page->inuse) { |
1510 | |
1511 | if (page->freelist) { |
1512 | add_partial(n, page, tail); |
1513 | stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); |
1514 | } else { |
1515 | stat(s, DEACTIVATE_FULL); |
1516 | if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER)) |
1517 | add_full(n, page); |
1518 | } |
1519 | slab_unlock(page); |
1520 | } else { |
1521 | stat(s, DEACTIVATE_EMPTY); |
1522 | if (n->nr_partial < s->min_partial) { |
1523 | /* |
1524 | * Adding an empty slab to the partial slabs in order |
1525 | * to avoid page allocator overhead. This slab needs |
1526 | * to come after the other slabs with objects in |
1527 | * so that the others get filled first. That way the |
1528 | * size of the partial list stays small. |
1529 | * |
1530 | * kmem_cache_shrink can reclaim any empty slabs from |
1531 | * the partial list. |
1532 | */ |
1533 | add_partial(n, page, 1); |
1534 | slab_unlock(page); |
1535 | } else { |
1536 | slab_unlock(page); |
1537 | stat(s, FREE_SLAB); |
1538 | discard_slab(s, page); |
1539 | } |
1540 | } |
1541 | } |
1542 | |
1543 | #ifdef CONFIG_CMPXCHG_LOCAL |
1544 | #ifdef CONFIG_PREEMPT |
1545 | /* |
1546 | * Calculate the next globally unique transaction for disambiguiation |
1547 | * during cmpxchg. The transactions start with the cpu number and are then |
1548 | * incremented by CONFIG_NR_CPUS. |
1549 | */ |
1550 | #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) |
1551 | #else |
1552 | /* |
1553 | * No preemption supported therefore also no need to check for |
1554 | * different cpus. |
1555 | */ |
1556 | #define TID_STEP 1 |
1557 | #endif |
1558 | |
1559 | static inline unsigned long next_tid(unsigned long tid) |
1560 | { |
1561 | return tid + TID_STEP; |
1562 | } |
1563 | |
1564 | static inline unsigned int tid_to_cpu(unsigned long tid) |
1565 | { |
1566 | return tid % TID_STEP; |
1567 | } |
1568 | |
1569 | static inline unsigned long tid_to_event(unsigned long tid) |
1570 | { |
1571 | return tid / TID_STEP; |
1572 | } |
1573 | |
1574 | static inline unsigned int init_tid(int cpu) |
1575 | { |
1576 | return cpu; |
1577 | } |
1578 | |
1579 | static inline void note_cmpxchg_failure(const char *n, |
1580 | const struct kmem_cache *s, unsigned long tid) |
1581 | { |
1582 | #ifdef SLUB_DEBUG_CMPXCHG |
1583 | unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); |
1584 | |
1585 | printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name); |
1586 | |
1587 | #ifdef CONFIG_PREEMPT |
1588 | if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) |
1589 | printk("due to cpu change %d -> %d\n", |
1590 | tid_to_cpu(tid), tid_to_cpu(actual_tid)); |
1591 | else |
1592 | #endif |
1593 | if (tid_to_event(tid) != tid_to_event(actual_tid)) |
1594 | printk("due to cpu running other code. Event %ld->%ld\n", |
1595 | tid_to_event(tid), tid_to_event(actual_tid)); |
1596 | else |
1597 | printk("for unknown reason: actual=%lx was=%lx target=%lx\n", |
1598 | actual_tid, tid, next_tid(tid)); |
1599 | #endif |
1600 | stat(s, CMPXCHG_DOUBLE_CPU_FAIL); |
1601 | } |
1602 | |
1603 | #endif |
1604 | |
1605 | void init_kmem_cache_cpus(struct kmem_cache *s) |
1606 | { |
1607 | #ifdef CONFIG_CMPXCHG_LOCAL |
1608 | int cpu; |
1609 | |
1610 | for_each_possible_cpu(cpu) |
1611 | per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); |
1612 | #endif |
1613 | |
1614 | } |
1615 | /* |
1616 | * Remove the cpu slab |
1617 | */ |
1618 | static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
1619 | __releases(bitlock) |
1620 | { |
1621 | struct page *page = c->page; |
1622 | int tail = 1; |
1623 | |
1624 | if (page->freelist) |
1625 | stat(s, DEACTIVATE_REMOTE_FREES); |
1626 | /* |
1627 | * Merge cpu freelist into slab freelist. Typically we get here |
1628 | * because both freelists are empty. So this is unlikely |
1629 | * to occur. |
1630 | */ |
1631 | while (unlikely(c->freelist)) { |
1632 | void **object; |
1633 | |
1634 | tail = 0; /* Hot objects. Put the slab first */ |
1635 | |
1636 | /* Retrieve object from cpu_freelist */ |
1637 | object = c->freelist; |
1638 | c->freelist = get_freepointer(s, c->freelist); |
1639 | |
1640 | /* And put onto the regular freelist */ |
1641 | set_freepointer(s, object, page->freelist); |
1642 | page->freelist = object; |
1643 | page->inuse--; |
1644 | } |
1645 | c->page = NULL; |
1646 | #ifdef CONFIG_CMPXCHG_LOCAL |
1647 | c->tid = next_tid(c->tid); |
1648 | #endif |
1649 | unfreeze_slab(s, page, tail); |
1650 | } |
1651 | |
1652 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
1653 | { |
1654 | stat(s, CPUSLAB_FLUSH); |
1655 | slab_lock(c->page); |
1656 | deactivate_slab(s, c); |
1657 | } |
1658 | |
1659 | /* |
1660 | * Flush cpu slab. |
1661 | * |
1662 | * Called from IPI handler with interrupts disabled. |
1663 | */ |
1664 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) |
1665 | { |
1666 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
1667 | |
1668 | if (likely(c && c->page)) |
1669 | flush_slab(s, c); |
1670 | } |
1671 | |
1672 | static void flush_cpu_slab(void *d) |
1673 | { |
1674 | struct kmem_cache *s = d; |
1675 | |
1676 | __flush_cpu_slab(s, smp_processor_id()); |
1677 | } |
1678 | |
1679 | static void flush_all(struct kmem_cache *s) |
1680 | { |
1681 | on_each_cpu(flush_cpu_slab, s, 1); |
1682 | } |
1683 | |
1684 | /* |
1685 | * Check if the objects in a per cpu structure fit numa |
1686 | * locality expectations. |
1687 | */ |
1688 | static inline int node_match(struct kmem_cache_cpu *c, int node) |
1689 | { |
1690 | #ifdef CONFIG_NUMA |
1691 | if (node != NUMA_NO_NODE && c->node != node) |
1692 | return 0; |
1693 | #endif |
1694 | return 1; |
1695 | } |
1696 | |
1697 | static int count_free(struct page *page) |
1698 | { |
1699 | return page->objects - page->inuse; |
1700 | } |
1701 | |
1702 | static unsigned long count_partial(struct kmem_cache_node *n, |
1703 | int (*get_count)(struct page *)) |
1704 | { |
1705 | unsigned long flags; |
1706 | unsigned long x = 0; |
1707 | struct page *page; |
1708 | |
1709 | spin_lock_irqsave(&n->list_lock, flags); |
1710 | list_for_each_entry(page, &n->partial, lru) |
1711 | x += get_count(page); |
1712 | spin_unlock_irqrestore(&n->list_lock, flags); |
1713 | return x; |
1714 | } |
1715 | |
1716 | static inline unsigned long node_nr_objs(struct kmem_cache_node *n) |
1717 | { |
1718 | #ifdef CONFIG_SLUB_DEBUG |
1719 | return atomic_long_read(&n->total_objects); |
1720 | #else |
1721 | return 0; |
1722 | #endif |
1723 | } |
1724 | |
1725 | static noinline void |
1726 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) |
1727 | { |
1728 | int node; |
1729 | |
1730 | printk(KERN_WARNING |
1731 | "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", |
1732 | nid, gfpflags); |
1733 | printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " |
1734 | "default order: %d, min order: %d\n", s->name, s->objsize, |
1735 | s->size, oo_order(s->oo), oo_order(s->min)); |
1736 | |
1737 | if (oo_order(s->min) > get_order(s->objsize)) |
1738 | printk(KERN_WARNING " %s debugging increased min order, use " |
1739 | "slub_debug=O to disable.\n", s->name); |
1740 | |
1741 | for_each_online_node(node) { |
1742 | struct kmem_cache_node *n = get_node(s, node); |
1743 | unsigned long nr_slabs; |
1744 | unsigned long nr_objs; |
1745 | unsigned long nr_free; |
1746 | |
1747 | if (!n) |
1748 | continue; |
1749 | |
1750 | nr_free = count_partial(n, count_free); |
1751 | nr_slabs = node_nr_slabs(n); |
1752 | nr_objs = node_nr_objs(n); |
1753 | |
1754 | printk(KERN_WARNING |
1755 | " node %d: slabs: %ld, objs: %ld, free: %ld\n", |
1756 | node, nr_slabs, nr_objs, nr_free); |
1757 | } |
1758 | } |
1759 | |
1760 | /* |
1761 | * Slow path. The lockless freelist is empty or we need to perform |
1762 | * debugging duties. |
1763 | * |
1764 | * Interrupts are disabled. |
1765 | * |
1766 | * Processing is still very fast if new objects have been freed to the |
1767 | * regular freelist. In that case we simply take over the regular freelist |
1768 | * as the lockless freelist and zap the regular freelist. |
1769 | * |
1770 | * If that is not working then we fall back to the partial lists. We take the |
1771 | * first element of the freelist as the object to allocate now and move the |
1772 | * rest of the freelist to the lockless freelist. |
1773 | * |
1774 | * And if we were unable to get a new slab from the partial slab lists then |
1775 | * we need to allocate a new slab. This is the slowest path since it involves |
1776 | * a call to the page allocator and the setup of a new slab. |
1777 | */ |
1778 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
1779 | unsigned long addr, struct kmem_cache_cpu *c) |
1780 | { |
1781 | void **object; |
1782 | struct page *new; |
1783 | #ifdef CONFIG_CMPXCHG_LOCAL |
1784 | unsigned long flags; |
1785 | |
1786 | local_irq_save(flags); |
1787 | #ifdef CONFIG_PREEMPT |
1788 | /* |
1789 | * We may have been preempted and rescheduled on a different |
1790 | * cpu before disabling interrupts. Need to reload cpu area |
1791 | * pointer. |
1792 | */ |
1793 | c = this_cpu_ptr(s->cpu_slab); |
1794 | #endif |
1795 | #endif |
1796 | |
1797 | /* We handle __GFP_ZERO in the caller */ |
1798 | gfpflags &= ~__GFP_ZERO; |
1799 | |
1800 | if (!c->page) |
1801 | goto new_slab; |
1802 | |
1803 | slab_lock(c->page); |
1804 | if (unlikely(!node_match(c, node))) |
1805 | goto another_slab; |
1806 | |
1807 | stat(s, ALLOC_REFILL); |
1808 | |
1809 | load_freelist: |
1810 | object = c->page->freelist; |
1811 | if (unlikely(!object)) |
1812 | goto another_slab; |
1813 | if (kmem_cache_debug(s)) |
1814 | goto debug; |
1815 | |
1816 | c->freelist = get_freepointer(s, object); |
1817 | c->page->inuse = c->page->objects; |
1818 | c->page->freelist = NULL; |
1819 | c->node = page_to_nid(c->page); |
1820 | unlock_out: |
1821 | slab_unlock(c->page); |
1822 | #ifdef CONFIG_CMPXCHG_LOCAL |
1823 | c->tid = next_tid(c->tid); |
1824 | local_irq_restore(flags); |
1825 | #endif |
1826 | stat(s, ALLOC_SLOWPATH); |
1827 | return object; |
1828 | |
1829 | another_slab: |
1830 | deactivate_slab(s, c); |
1831 | |
1832 | new_slab: |
1833 | new = get_partial(s, gfpflags, node); |
1834 | if (new) { |
1835 | c->page = new; |
1836 | stat(s, ALLOC_FROM_PARTIAL); |
1837 | goto load_freelist; |
1838 | } |
1839 | |
1840 | gfpflags &= gfp_allowed_mask; |
1841 | if (gfpflags & __GFP_WAIT) |
1842 | local_irq_enable(); |
1843 | |
1844 | new = new_slab(s, gfpflags, node); |
1845 | |
1846 | if (gfpflags & __GFP_WAIT) |
1847 | local_irq_disable(); |
1848 | |
1849 | if (new) { |
1850 | c = __this_cpu_ptr(s->cpu_slab); |
1851 | stat(s, ALLOC_SLAB); |
1852 | if (c->page) |
1853 | flush_slab(s, c); |
1854 | slab_lock(new); |
1855 | __SetPageSlubFrozen(new); |
1856 | c->page = new; |
1857 | goto load_freelist; |
1858 | } |
1859 | if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) |
1860 | slab_out_of_memory(s, gfpflags, node); |
1861 | #ifdef CONFIG_CMPXCHG_LOCAL |
1862 | local_irq_restore(flags); |
1863 | #endif |
1864 | return NULL; |
1865 | debug: |
1866 | if (!alloc_debug_processing(s, c->page, object, addr)) |
1867 | goto another_slab; |
1868 | |
1869 | c->page->inuse++; |
1870 | c->page->freelist = get_freepointer(s, object); |
1871 | c->node = NUMA_NO_NODE; |
1872 | goto unlock_out; |
1873 | } |
1874 | |
1875 | /* |
1876 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) |
1877 | * have the fastpath folded into their functions. So no function call |
1878 | * overhead for requests that can be satisfied on the fastpath. |
1879 | * |
1880 | * The fastpath works by first checking if the lockless freelist can be used. |
1881 | * If not then __slab_alloc is called for slow processing. |
1882 | * |
1883 | * Otherwise we can simply pick the next object from the lockless free list. |
1884 | */ |
1885 | static __always_inline void *slab_alloc(struct kmem_cache *s, |
1886 | gfp_t gfpflags, int node, unsigned long addr) |
1887 | { |
1888 | void **object; |
1889 | struct kmem_cache_cpu *c; |
1890 | #ifdef CONFIG_CMPXCHG_LOCAL |
1891 | unsigned long tid; |
1892 | #else |
1893 | unsigned long flags; |
1894 | #endif |
1895 | |
1896 | if (slab_pre_alloc_hook(s, gfpflags)) |
1897 | return NULL; |
1898 | |
1899 | #ifndef CONFIG_CMPXCHG_LOCAL |
1900 | local_irq_save(flags); |
1901 | #else |
1902 | redo: |
1903 | #endif |
1904 | |
1905 | /* |
1906 | * Must read kmem_cache cpu data via this cpu ptr. Preemption is |
1907 | * enabled. We may switch back and forth between cpus while |
1908 | * reading from one cpu area. That does not matter as long |
1909 | * as we end up on the original cpu again when doing the cmpxchg. |
1910 | */ |
1911 | c = __this_cpu_ptr(s->cpu_slab); |
1912 | |
1913 | #ifdef CONFIG_CMPXCHG_LOCAL |
1914 | /* |
1915 | * The transaction ids are globally unique per cpu and per operation on |
1916 | * a per cpu queue. Thus they can be guarantee that the cmpxchg_double |
1917 | * occurs on the right processor and that there was no operation on the |
1918 | * linked list in between. |
1919 | */ |
1920 | tid = c->tid; |
1921 | barrier(); |
1922 | #endif |
1923 | |
1924 | object = c->freelist; |
1925 | if (unlikely(!object || !node_match(c, node))) |
1926 | |
1927 | object = __slab_alloc(s, gfpflags, node, addr, c); |
1928 | |
1929 | else { |
1930 | #ifdef CONFIG_CMPXCHG_LOCAL |
1931 | /* |
1932 | * The cmpxchg will only match if there was no additional |
1933 | * operation and if we are on the right processor. |
1934 | * |
1935 | * The cmpxchg does the following atomically (without lock semantics!) |
1936 | * 1. Relocate first pointer to the current per cpu area. |
1937 | * 2. Verify that tid and freelist have not been changed |
1938 | * 3. If they were not changed replace tid and freelist |
1939 | * |
1940 | * Since this is without lock semantics the protection is only against |
1941 | * code executing on this cpu *not* from access by other cpus. |
1942 | */ |
1943 | if (unlikely(!irqsafe_cpu_cmpxchg_double( |
1944 | s->cpu_slab->freelist, s->cpu_slab->tid, |
1945 | object, tid, |
1946 | get_freepointer(s, object), next_tid(tid)))) { |
1947 | |
1948 | note_cmpxchg_failure("slab_alloc", s, tid); |
1949 | goto redo; |
1950 | } |
1951 | #else |
1952 | c->freelist = get_freepointer(s, object); |
1953 | #endif |
1954 | stat(s, ALLOC_FASTPATH); |
1955 | } |
1956 | |
1957 | #ifndef CONFIG_CMPXCHG_LOCAL |
1958 | local_irq_restore(flags); |
1959 | #endif |
1960 | |
1961 | if (unlikely(gfpflags & __GFP_ZERO) && object) |
1962 | memset(object, 0, s->objsize); |
1963 | |
1964 | slab_post_alloc_hook(s, gfpflags, object); |
1965 | |
1966 | return object; |
1967 | } |
1968 | |
1969 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) |
1970 | { |
1971 | void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); |
1972 | |
1973 | trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); |
1974 | |
1975 | return ret; |
1976 | } |
1977 | EXPORT_SYMBOL(kmem_cache_alloc); |
1978 | |
1979 | #ifdef CONFIG_TRACING |
1980 | void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) |
1981 | { |
1982 | void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); |
1983 | trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); |
1984 | return ret; |
1985 | } |
1986 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
1987 | |
1988 | void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) |
1989 | { |
1990 | void *ret = kmalloc_order(size, flags, order); |
1991 | trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags); |
1992 | return ret; |
1993 | } |
1994 | EXPORT_SYMBOL(kmalloc_order_trace); |
1995 | #endif |
1996 | |
1997 | #ifdef CONFIG_NUMA |
1998 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) |
1999 | { |
2000 | void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); |
2001 | |
2002 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
2003 | s->objsize, s->size, gfpflags, node); |
2004 | |
2005 | return ret; |
2006 | } |
2007 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
2008 | |
2009 | #ifdef CONFIG_TRACING |
2010 | void *kmem_cache_alloc_node_trace(struct kmem_cache *s, |
2011 | gfp_t gfpflags, |
2012 | int node, size_t size) |
2013 | { |
2014 | void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); |
2015 | |
2016 | trace_kmalloc_node(_RET_IP_, ret, |
2017 | size, s->size, gfpflags, node); |
2018 | return ret; |
2019 | } |
2020 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
2021 | #endif |
2022 | #endif |
2023 | |
2024 | /* |
2025 | * Slow patch handling. This may still be called frequently since objects |
2026 | * have a longer lifetime than the cpu slabs in most processing loads. |
2027 | * |
2028 | * So we still attempt to reduce cache line usage. Just take the slab |
2029 | * lock and free the item. If there is no additional partial page |
2030 | * handling required then we can return immediately. |
2031 | */ |
2032 | static void __slab_free(struct kmem_cache *s, struct page *page, |
2033 | void *x, unsigned long addr) |
2034 | { |
2035 | void *prior; |
2036 | void **object = (void *)x; |
2037 | #ifdef CONFIG_CMPXCHG_LOCAL |
2038 | unsigned long flags; |
2039 | |
2040 | local_irq_save(flags); |
2041 | #endif |
2042 | slab_lock(page); |
2043 | stat(s, FREE_SLOWPATH); |
2044 | |
2045 | if (kmem_cache_debug(s)) |
2046 | goto debug; |
2047 | |
2048 | checks_ok: |
2049 | prior = page->freelist; |
2050 | set_freepointer(s, object, prior); |
2051 | page->freelist = object; |
2052 | page->inuse--; |
2053 | |
2054 | if (unlikely(PageSlubFrozen(page))) { |
2055 | stat(s, FREE_FROZEN); |
2056 | goto out_unlock; |
2057 | } |
2058 | |
2059 | if (unlikely(!page->inuse)) |
2060 | goto slab_empty; |
2061 | |
2062 | /* |
2063 | * Objects left in the slab. If it was not on the partial list before |
2064 | * then add it. |
2065 | */ |
2066 | if (unlikely(!prior)) { |
2067 | add_partial(get_node(s, page_to_nid(page)), page, 1); |
2068 | stat(s, FREE_ADD_PARTIAL); |
2069 | } |
2070 | |
2071 | out_unlock: |
2072 | slab_unlock(page); |
2073 | #ifdef CONFIG_CMPXCHG_LOCAL |
2074 | local_irq_restore(flags); |
2075 | #endif |
2076 | return; |
2077 | |
2078 | slab_empty: |
2079 | if (prior) { |
2080 | /* |
2081 | * Slab still on the partial list. |
2082 | */ |
2083 | remove_partial(s, page); |
2084 | stat(s, FREE_REMOVE_PARTIAL); |
2085 | } |
2086 | slab_unlock(page); |
2087 | #ifdef CONFIG_CMPXCHG_LOCAL |
2088 | local_irq_restore(flags); |
2089 | #endif |
2090 | stat(s, FREE_SLAB); |
2091 | discard_slab(s, page); |
2092 | return; |
2093 | |
2094 | debug: |
2095 | if (!free_debug_processing(s, page, x, addr)) |
2096 | goto out_unlock; |
2097 | goto checks_ok; |
2098 | } |
2099 | |
2100 | /* |
2101 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that |
2102 | * can perform fastpath freeing without additional function calls. |
2103 | * |
2104 | * The fastpath is only possible if we are freeing to the current cpu slab |
2105 | * of this processor. This typically the case if we have just allocated |
2106 | * the item before. |
2107 | * |
2108 | * If fastpath is not possible then fall back to __slab_free where we deal |
2109 | * with all sorts of special processing. |
2110 | */ |
2111 | static __always_inline void slab_free(struct kmem_cache *s, |
2112 | struct page *page, void *x, unsigned long addr) |
2113 | { |
2114 | void **object = (void *)x; |
2115 | struct kmem_cache_cpu *c; |
2116 | #ifdef CONFIG_CMPXCHG_LOCAL |
2117 | unsigned long tid; |
2118 | #else |
2119 | unsigned long flags; |
2120 | #endif |
2121 | |
2122 | slab_free_hook(s, x); |
2123 | |
2124 | #ifndef CONFIG_CMPXCHG_LOCAL |
2125 | local_irq_save(flags); |
2126 | |
2127 | #else |
2128 | redo: |
2129 | #endif |
2130 | |
2131 | /* |
2132 | * Determine the currently cpus per cpu slab. |
2133 | * The cpu may change afterward. However that does not matter since |
2134 | * data is retrieved via this pointer. If we are on the same cpu |
2135 | * during the cmpxchg then the free will succedd. |
2136 | */ |
2137 | c = __this_cpu_ptr(s->cpu_slab); |
2138 | |
2139 | #ifdef CONFIG_CMPXCHG_LOCAL |
2140 | tid = c->tid; |
2141 | barrier(); |
2142 | #endif |
2143 | |
2144 | if (likely(page == c->page && c->node != NUMA_NO_NODE)) { |
2145 | set_freepointer(s, object, c->freelist); |
2146 | |
2147 | #ifdef CONFIG_CMPXCHG_LOCAL |
2148 | if (unlikely(!irqsafe_cpu_cmpxchg_double( |
2149 | s->cpu_slab->freelist, s->cpu_slab->tid, |
2150 | c->freelist, tid, |
2151 | object, next_tid(tid)))) { |
2152 | |
2153 | note_cmpxchg_failure("slab_free", s, tid); |
2154 | goto redo; |
2155 | } |
2156 | #else |
2157 | c->freelist = object; |
2158 | #endif |
2159 | stat(s, FREE_FASTPATH); |
2160 | } else |
2161 | __slab_free(s, page, x, addr); |
2162 | |
2163 | #ifndef CONFIG_CMPXCHG_LOCAL |
2164 | local_irq_restore(flags); |
2165 | #endif |
2166 | } |
2167 | |
2168 | void kmem_cache_free(struct kmem_cache *s, void *x) |
2169 | { |
2170 | struct page *page; |
2171 | |
2172 | page = virt_to_head_page(x); |
2173 | |
2174 | slab_free(s, page, x, _RET_IP_); |
2175 | |
2176 | trace_kmem_cache_free(_RET_IP_, x); |
2177 | } |
2178 | EXPORT_SYMBOL(kmem_cache_free); |
2179 | |
2180 | /* |
2181 | * Object placement in a slab is made very easy because we always start at |
2182 | * offset 0. If we tune the size of the object to the alignment then we can |
2183 | * get the required alignment by putting one properly sized object after |
2184 | * another. |
2185 | * |
2186 | * Notice that the allocation order determines the sizes of the per cpu |
2187 | * caches. Each processor has always one slab available for allocations. |
2188 | * Increasing the allocation order reduces the number of times that slabs |
2189 | * must be moved on and off the partial lists and is therefore a factor in |
2190 | * locking overhead. |
2191 | */ |
2192 | |
2193 | /* |
2194 | * Mininum / Maximum order of slab pages. This influences locking overhead |
2195 | * and slab fragmentation. A higher order reduces the number of partial slabs |
2196 | * and increases the number of allocations possible without having to |
2197 | * take the list_lock. |
2198 | */ |
2199 | static int slub_min_order; |
2200 | static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; |
2201 | static int slub_min_objects; |
2202 | |
2203 | /* |
2204 | * Merge control. If this is set then no merging of slab caches will occur. |
2205 | * (Could be removed. This was introduced to pacify the merge skeptics.) |
2206 | */ |
2207 | static int slub_nomerge; |
2208 | |
2209 | /* |
2210 | * Calculate the order of allocation given an slab object size. |
2211 | * |
2212 | * The order of allocation has significant impact on performance and other |
2213 | * system components. Generally order 0 allocations should be preferred since |
2214 | * order 0 does not cause fragmentation in the page allocator. Larger objects |
2215 | * be problematic to put into order 0 slabs because there may be too much |
2216 | * unused space left. We go to a higher order if more than 1/16th of the slab |
2217 | * would be wasted. |
2218 | * |
2219 | * In order to reach satisfactory performance we must ensure that a minimum |
2220 | * number of objects is in one slab. Otherwise we may generate too much |
2221 | * activity on the partial lists which requires taking the list_lock. This is |
2222 | * less a concern for large slabs though which are rarely used. |
2223 | * |
2224 | * slub_max_order specifies the order where we begin to stop considering the |
2225 | * number of objects in a slab as critical. If we reach slub_max_order then |
2226 | * we try to keep the page order as low as possible. So we accept more waste |
2227 | * of space in favor of a small page order. |
2228 | * |
2229 | * Higher order allocations also allow the placement of more objects in a |
2230 | * slab and thereby reduce object handling overhead. If the user has |
2231 | * requested a higher mininum order then we start with that one instead of |
2232 | * the smallest order which will fit the object. |
2233 | */ |
2234 | static inline int slab_order(int size, int min_objects, |
2235 | int max_order, int fract_leftover, int reserved) |
2236 | { |
2237 | int order; |
2238 | int rem; |
2239 | int min_order = slub_min_order; |
2240 | |
2241 | if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) |
2242 | return get_order(size * MAX_OBJS_PER_PAGE) - 1; |
2243 | |
2244 | for (order = max(min_order, |
2245 | fls(min_objects * size - 1) - PAGE_SHIFT); |
2246 | order <= max_order; order++) { |
2247 | |
2248 | unsigned long slab_size = PAGE_SIZE << order; |
2249 | |
2250 | if (slab_size < min_objects * size + reserved) |
2251 | continue; |
2252 | |
2253 | rem = (slab_size - reserved) % size; |
2254 | |
2255 | if (rem <= slab_size / fract_leftover) |
2256 | break; |
2257 | |
2258 | } |
2259 | |
2260 | return order; |
2261 | } |
2262 | |
2263 | static inline int calculate_order(int size, int reserved) |
2264 | { |
2265 | int order; |
2266 | int min_objects; |
2267 | int fraction; |
2268 | int max_objects; |
2269 | |
2270 | /* |
2271 | * Attempt to find best configuration for a slab. This |
2272 | * works by first attempting to generate a layout with |
2273 | * the best configuration and backing off gradually. |
2274 | * |
2275 | * First we reduce the acceptable waste in a slab. Then |
2276 | * we reduce the minimum objects required in a slab. |
2277 | */ |
2278 | min_objects = slub_min_objects; |
2279 | if (!min_objects) |
2280 | min_objects = 4 * (fls(nr_cpu_ids) + 1); |
2281 | max_objects = order_objects(slub_max_order, size, reserved); |
2282 | min_objects = min(min_objects, max_objects); |
2283 | |
2284 | while (min_objects > 1) { |
2285 | fraction = 16; |
2286 | while (fraction >= 4) { |
2287 | order = slab_order(size, min_objects, |
2288 | slub_max_order, fraction, reserved); |
2289 | if (order <= slub_max_order) |
2290 | return order; |
2291 | fraction /= 2; |
2292 | } |
2293 | min_objects--; |
2294 | } |
2295 | |
2296 | /* |
2297 | * We were unable to place multiple objects in a slab. Now |
2298 | * lets see if we can place a single object there. |
2299 | */ |
2300 | order = slab_order(size, 1, slub_max_order, 1, reserved); |
2301 | if (order <= slub_max_order) |
2302 | return order; |
2303 | |
2304 | /* |
2305 | * Doh this slab cannot be placed using slub_max_order. |
2306 | */ |
2307 | order = slab_order(size, 1, MAX_ORDER, 1, reserved); |
2308 | if (order < MAX_ORDER) |
2309 | return order; |
2310 | return -ENOSYS; |
2311 | } |
2312 | |
2313 | /* |
2314 | * Figure out what the alignment of the objects will be. |
2315 | */ |
2316 | static unsigned long calculate_alignment(unsigned long flags, |
2317 | unsigned long align, unsigned long size) |
2318 | { |
2319 | /* |
2320 | * If the user wants hardware cache aligned objects then follow that |
2321 | * suggestion if the object is sufficiently large. |
2322 | * |
2323 | * The hardware cache alignment cannot override the specified |
2324 | * alignment though. If that is greater then use it. |
2325 | */ |
2326 | if (flags & SLAB_HWCACHE_ALIGN) { |
2327 | unsigned long ralign = cache_line_size(); |
2328 | while (size <= ralign / 2) |
2329 | ralign /= 2; |
2330 | align = max(align, ralign); |
2331 | } |
2332 | |
2333 | if (align < ARCH_SLAB_MINALIGN) |
2334 | align = ARCH_SLAB_MINALIGN; |
2335 | |
2336 | return ALIGN(align, sizeof(void *)); |
2337 | } |
2338 | |
2339 | static void |
2340 | init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) |
2341 | { |
2342 | n->nr_partial = 0; |
2343 | spin_lock_init(&n->list_lock); |
2344 | INIT_LIST_HEAD(&n->partial); |
2345 | #ifdef CONFIG_SLUB_DEBUG |
2346 | atomic_long_set(&n->nr_slabs, 0); |
2347 | atomic_long_set(&n->total_objects, 0); |
2348 | INIT_LIST_HEAD(&n->full); |
2349 | #endif |
2350 | } |
2351 | |
2352 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
2353 | { |
2354 | BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < |
2355 | SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu)); |
2356 | |
2357 | #ifdef CONFIG_CMPXCHG_LOCAL |
2358 | /* |
2359 | * Must align to double word boundary for the double cmpxchg instructions |
2360 | * to work. |
2361 | */ |
2362 | s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *)); |
2363 | #else |
2364 | /* Regular alignment is sufficient */ |
2365 | s->cpu_slab = alloc_percpu(struct kmem_cache_cpu); |
2366 | #endif |
2367 | |
2368 | if (!s->cpu_slab) |
2369 | return 0; |
2370 | |
2371 | init_kmem_cache_cpus(s); |
2372 | |
2373 | return 1; |
2374 | } |
2375 | |
2376 | static struct kmem_cache *kmem_cache_node; |
2377 | |
2378 | /* |
2379 | * No kmalloc_node yet so do it by hand. We know that this is the first |
2380 | * slab on the node for this slabcache. There are no concurrent accesses |
2381 | * possible. |
2382 | * |
2383 | * Note that this function only works on the kmalloc_node_cache |
2384 | * when allocating for the kmalloc_node_cache. This is used for bootstrapping |
2385 | * memory on a fresh node that has no slab structures yet. |
2386 | */ |
2387 | static void early_kmem_cache_node_alloc(int node) |
2388 | { |
2389 | struct page *page; |
2390 | struct kmem_cache_node *n; |
2391 | unsigned long flags; |
2392 | |
2393 | BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); |
2394 | |
2395 | page = new_slab(kmem_cache_node, GFP_NOWAIT, node); |
2396 | |
2397 | BUG_ON(!page); |
2398 | if (page_to_nid(page) != node) { |
2399 | printk(KERN_ERR "SLUB: Unable to allocate memory from " |
2400 | "node %d\n", node); |
2401 | printk(KERN_ERR "SLUB: Allocating a useless per node structure " |
2402 | "in order to be able to continue\n"); |
2403 | } |
2404 | |
2405 | n = page->freelist; |
2406 | BUG_ON(!n); |
2407 | page->freelist = get_freepointer(kmem_cache_node, n); |
2408 | page->inuse++; |
2409 | kmem_cache_node->node[node] = n; |
2410 | #ifdef CONFIG_SLUB_DEBUG |
2411 | init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); |
2412 | init_tracking(kmem_cache_node, n); |
2413 | #endif |
2414 | init_kmem_cache_node(n, kmem_cache_node); |
2415 | inc_slabs_node(kmem_cache_node, node, page->objects); |
2416 | |
2417 | /* |
2418 | * lockdep requires consistent irq usage for each lock |
2419 | * so even though there cannot be a race this early in |
2420 | * the boot sequence, we still disable irqs. |
2421 | */ |
2422 | local_irq_save(flags); |
2423 | add_partial(n, page, 0); |
2424 | local_irq_restore(flags); |
2425 | } |
2426 | |
2427 | static void free_kmem_cache_nodes(struct kmem_cache *s) |
2428 | { |
2429 | int node; |
2430 | |
2431 | for_each_node_state(node, N_NORMAL_MEMORY) { |
2432 | struct kmem_cache_node *n = s->node[node]; |
2433 | |
2434 | if (n) |
2435 | kmem_cache_free(kmem_cache_node, n); |
2436 | |
2437 | s->node[node] = NULL; |
2438 | } |
2439 | } |
2440 | |
2441 | static int init_kmem_cache_nodes(struct kmem_cache *s) |
2442 | { |
2443 | int node; |
2444 | |
2445 | for_each_node_state(node, N_NORMAL_MEMORY) { |
2446 | struct kmem_cache_node *n; |
2447 | |
2448 | if (slab_state == DOWN) { |
2449 | early_kmem_cache_node_alloc(node); |
2450 | continue; |
2451 | } |
2452 | n = kmem_cache_alloc_node(kmem_cache_node, |
2453 | GFP_KERNEL, node); |
2454 | |
2455 | if (!n) { |
2456 | free_kmem_cache_nodes(s); |
2457 | return 0; |
2458 | } |
2459 | |
2460 | s->node[node] = n; |
2461 | init_kmem_cache_node(n, s); |
2462 | } |
2463 | return 1; |
2464 | } |
2465 | |
2466 | static void set_min_partial(struct kmem_cache *s, unsigned long min) |
2467 | { |
2468 | if (min < MIN_PARTIAL) |
2469 | min = MIN_PARTIAL; |
2470 | else if (min > MAX_PARTIAL) |
2471 | min = MAX_PARTIAL; |
2472 | s->min_partial = min; |
2473 | } |
2474 | |
2475 | /* |
2476 | * calculate_sizes() determines the order and the distribution of data within |
2477 | * a slab object. |
2478 | */ |
2479 | static int calculate_sizes(struct kmem_cache *s, int forced_order) |
2480 | { |
2481 | unsigned long flags = s->flags; |
2482 | unsigned long size = s->objsize; |
2483 | unsigned long align = s->align; |
2484 | int order; |
2485 | |
2486 | /* |
2487 | * Round up object size to the next word boundary. We can only |
2488 | * place the free pointer at word boundaries and this determines |
2489 | * the possible location of the free pointer. |
2490 | */ |
2491 | size = ALIGN(size, sizeof(void *)); |
2492 | |
2493 | #ifdef CONFIG_SLUB_DEBUG |
2494 | /* |
2495 | * Determine if we can poison the object itself. If the user of |
2496 | * the slab may touch the object after free or before allocation |
2497 | * then we should never poison the object itself. |
2498 | */ |
2499 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && |
2500 | !s->ctor) |
2501 | s->flags |= __OBJECT_POISON; |
2502 | else |
2503 | s->flags &= ~__OBJECT_POISON; |
2504 | |
2505 | |
2506 | /* |
2507 | * If we are Redzoning then check if there is some space between the |
2508 | * end of the object and the free pointer. If not then add an |
2509 | * additional word to have some bytes to store Redzone information. |
2510 | */ |
2511 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) |
2512 | size += sizeof(void *); |
2513 | #endif |
2514 | |
2515 | /* |
2516 | * With that we have determined the number of bytes in actual use |
2517 | * by the object. This is the potential offset to the free pointer. |
2518 | */ |
2519 | s->inuse = size; |
2520 | |
2521 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || |
2522 | s->ctor)) { |
2523 | /* |
2524 | * Relocate free pointer after the object if it is not |
2525 | * permitted to overwrite the first word of the object on |
2526 | * kmem_cache_free. |
2527 | * |
2528 | * This is the case if we do RCU, have a constructor or |
2529 | * destructor or are poisoning the objects. |
2530 | */ |
2531 | s->offset = size; |
2532 | size += sizeof(void *); |
2533 | } |
2534 | |
2535 | #ifdef CONFIG_SLUB_DEBUG |
2536 | if (flags & SLAB_STORE_USER) |
2537 | /* |
2538 | * Need to store information about allocs and frees after |
2539 | * the object. |
2540 | */ |
2541 | size += 2 * sizeof(struct track); |
2542 | |
2543 | if (flags & SLAB_RED_ZONE) |
2544 | /* |
2545 | * Add some empty padding so that we can catch |
2546 | * overwrites from earlier objects rather than let |
2547 | * tracking information or the free pointer be |
2548 | * corrupted if a user writes before the start |
2549 | * of the object. |
2550 | */ |
2551 | size += sizeof(void *); |
2552 | #endif |
2553 | |
2554 | /* |
2555 | * Determine the alignment based on various parameters that the |
2556 | * user specified and the dynamic determination of cache line size |
2557 | * on bootup. |
2558 | */ |
2559 | align = calculate_alignment(flags, align, s->objsize); |
2560 | s->align = align; |
2561 | |
2562 | /* |
2563 | * SLUB stores one object immediately after another beginning from |
2564 | * offset 0. In order to align the objects we have to simply size |
2565 | * each object to conform to the alignment. |
2566 | */ |
2567 | size = ALIGN(size, align); |
2568 | s->size = size; |
2569 | if (forced_order >= 0) |
2570 | order = forced_order; |
2571 | else |
2572 | order = calculate_order(size, s->reserved); |
2573 | |
2574 | if (order < 0) |
2575 | return 0; |
2576 | |
2577 | s->allocflags = 0; |
2578 | if (order) |
2579 | s->allocflags |= __GFP_COMP; |
2580 | |
2581 | if (s->flags & SLAB_CACHE_DMA) |
2582 | s->allocflags |= SLUB_DMA; |
2583 | |
2584 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
2585 | s->allocflags |= __GFP_RECLAIMABLE; |
2586 | |
2587 | /* |
2588 | * Determine the number of objects per slab |
2589 | */ |
2590 | s->oo = oo_make(order, size, s->reserved); |
2591 | s->min = oo_make(get_order(size), size, s->reserved); |
2592 | if (oo_objects(s->oo) > oo_objects(s->max)) |
2593 | s->max = s->oo; |
2594 | |
2595 | return !!oo_objects(s->oo); |
2596 | |
2597 | } |
2598 | |
2599 | static int kmem_cache_open(struct kmem_cache *s, |
2600 | const char *name, size_t size, |
2601 | size_t align, unsigned long flags, |
2602 | void (*ctor)(void *)) |
2603 | { |
2604 | memset(s, 0, kmem_size); |
2605 | s->name = name; |
2606 | s->ctor = ctor; |
2607 | s->objsize = size; |
2608 | s->align = align; |
2609 | s->flags = kmem_cache_flags(size, flags, name, ctor); |
2610 | s->reserved = 0; |
2611 | |
2612 | if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) |
2613 | s->reserved = sizeof(struct rcu_head); |
2614 | |
2615 | if (!calculate_sizes(s, -1)) |
2616 | goto error; |
2617 | if (disable_higher_order_debug) { |
2618 | /* |
2619 | * Disable debugging flags that store metadata if the min slab |
2620 | * order increased. |
2621 | */ |
2622 | if (get_order(s->size) > get_order(s->objsize)) { |
2623 | s->flags &= ~DEBUG_METADATA_FLAGS; |
2624 | s->offset = 0; |
2625 | if (!calculate_sizes(s, -1)) |
2626 | goto error; |
2627 | } |
2628 | } |
2629 | |
2630 | /* |
2631 | * The larger the object size is, the more pages we want on the partial |
2632 | * list to avoid pounding the page allocator excessively. |
2633 | */ |
2634 | set_min_partial(s, ilog2(s->size)); |
2635 | s->refcount = 1; |
2636 | #ifdef CONFIG_NUMA |
2637 | s->remote_node_defrag_ratio = 1000; |
2638 | #endif |
2639 | if (!init_kmem_cache_nodes(s)) |
2640 | goto error; |
2641 | |
2642 | if (alloc_kmem_cache_cpus(s)) |
2643 | return 1; |
2644 | |
2645 | free_kmem_cache_nodes(s); |
2646 | error: |
2647 | if (flags & SLAB_PANIC) |
2648 | panic("Cannot create slab %s size=%lu realsize=%u " |
2649 | "order=%u offset=%u flags=%lx\n", |
2650 | s->name, (unsigned long)size, s->size, oo_order(s->oo), |
2651 | s->offset, flags); |
2652 | return 0; |
2653 | } |
2654 | |
2655 | /* |
2656 | * Determine the size of a slab object |
2657 | */ |
2658 | unsigned int kmem_cache_size(struct kmem_cache *s) |
2659 | { |
2660 | return s->objsize; |
2661 | } |
2662 | EXPORT_SYMBOL(kmem_cache_size); |
2663 | |
2664 | static void list_slab_objects(struct kmem_cache *s, struct page *page, |
2665 | const char *text) |
2666 | { |
2667 | #ifdef CONFIG_SLUB_DEBUG |
2668 | void *addr = page_address(page); |
2669 | void *p; |
2670 | unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * |
2671 | sizeof(long), GFP_ATOMIC); |
2672 | if (!map) |
2673 | return; |
2674 | slab_err(s, page, "%s", text); |
2675 | slab_lock(page); |
2676 | for_each_free_object(p, s, page->freelist) |
2677 | set_bit(slab_index(p, s, addr), map); |
2678 | |
2679 | for_each_object(p, s, addr, page->objects) { |
2680 | |
2681 | if (!test_bit(slab_index(p, s, addr), map)) { |
2682 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", |
2683 | p, p - addr); |
2684 | print_tracking(s, p); |
2685 | } |
2686 | } |
2687 | slab_unlock(page); |
2688 | kfree(map); |
2689 | #endif |
2690 | } |
2691 | |
2692 | /* |
2693 | * Attempt to free all partial slabs on a node. |
2694 | */ |
2695 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) |
2696 | { |
2697 | unsigned long flags; |
2698 | struct page *page, *h; |
2699 | |
2700 | spin_lock_irqsave(&n->list_lock, flags); |
2701 | list_for_each_entry_safe(page, h, &n->partial, lru) { |
2702 | if (!page->inuse) { |
2703 | __remove_partial(n, page); |
2704 | discard_slab(s, page); |
2705 | } else { |
2706 | list_slab_objects(s, page, |
2707 | "Objects remaining on kmem_cache_close()"); |
2708 | } |
2709 | } |
2710 | spin_unlock_irqrestore(&n->list_lock, flags); |
2711 | } |
2712 | |
2713 | /* |
2714 | * Release all resources used by a slab cache. |
2715 | */ |
2716 | static inline int kmem_cache_close(struct kmem_cache *s) |
2717 | { |
2718 | int node; |
2719 | |
2720 | flush_all(s); |
2721 | free_percpu(s->cpu_slab); |
2722 | /* Attempt to free all objects */ |
2723 | for_each_node_state(node, N_NORMAL_MEMORY) { |
2724 | struct kmem_cache_node *n = get_node(s, node); |
2725 | |
2726 | free_partial(s, n); |
2727 | if (n->nr_partial || slabs_node(s, node)) |
2728 | return 1; |
2729 | } |
2730 | free_kmem_cache_nodes(s); |
2731 | return 0; |
2732 | } |
2733 | |
2734 | /* |
2735 | * Close a cache and release the kmem_cache structure |
2736 | * (must be used for caches created using kmem_cache_create) |
2737 | */ |
2738 | void kmem_cache_destroy(struct kmem_cache *s) |
2739 | { |
2740 | down_write(&slub_lock); |
2741 | s->refcount--; |
2742 | if (!s->refcount) { |
2743 | list_del(&s->list); |
2744 | if (kmem_cache_close(s)) { |
2745 | printk(KERN_ERR "SLUB %s: %s called for cache that " |
2746 | "still has objects.\n", s->name, __func__); |
2747 | dump_stack(); |
2748 | } |
2749 | if (s->flags & SLAB_DESTROY_BY_RCU) |
2750 | rcu_barrier(); |
2751 | sysfs_slab_remove(s); |
2752 | } |
2753 | up_write(&slub_lock); |
2754 | } |
2755 | EXPORT_SYMBOL(kmem_cache_destroy); |
2756 | |
2757 | /******************************************************************** |
2758 | * Kmalloc subsystem |
2759 | *******************************************************************/ |
2760 | |
2761 | struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT]; |
2762 | EXPORT_SYMBOL(kmalloc_caches); |
2763 | |
2764 | static struct kmem_cache *kmem_cache; |
2765 | |
2766 | #ifdef CONFIG_ZONE_DMA |
2767 | static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT]; |
2768 | #endif |
2769 | |
2770 | static int __init setup_slub_min_order(char *str) |
2771 | { |
2772 | get_option(&str, &slub_min_order); |
2773 | |
2774 | return 1; |
2775 | } |
2776 | |
2777 | __setup("slub_min_order=", setup_slub_min_order); |
2778 | |
2779 | static int __init setup_slub_max_order(char *str) |
2780 | { |
2781 | get_option(&str, &slub_max_order); |
2782 | slub_max_order = min(slub_max_order, MAX_ORDER - 1); |
2783 | |
2784 | return 1; |
2785 | } |
2786 | |
2787 | __setup("slub_max_order=", setup_slub_max_order); |
2788 | |
2789 | static int __init setup_slub_min_objects(char *str) |
2790 | { |
2791 | get_option(&str, &slub_min_objects); |
2792 | |
2793 | return 1; |
2794 | } |
2795 | |
2796 | __setup("slub_min_objects=", setup_slub_min_objects); |
2797 | |
2798 | static int __init setup_slub_nomerge(char *str) |
2799 | { |
2800 | slub_nomerge = 1; |
2801 | return 1; |
2802 | } |
2803 | |
2804 | __setup("slub_nomerge", setup_slub_nomerge); |
2805 | |
2806 | static struct kmem_cache *__init create_kmalloc_cache(const char *name, |
2807 | int size, unsigned int flags) |
2808 | { |
2809 | struct kmem_cache *s; |
2810 | |
2811 | s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); |
2812 | |
2813 | /* |
2814 | * This function is called with IRQs disabled during early-boot on |
2815 | * single CPU so there's no need to take slub_lock here. |
2816 | */ |
2817 | if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN, |
2818 | flags, NULL)) |
2819 | goto panic; |
2820 | |
2821 | list_add(&s->list, &slab_caches); |
2822 | return s; |
2823 | |
2824 | panic: |
2825 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); |
2826 | return NULL; |
2827 | } |
2828 | |
2829 | /* |
2830 | * Conversion table for small slabs sizes / 8 to the index in the |
2831 | * kmalloc array. This is necessary for slabs < 192 since we have non power |
2832 | * of two cache sizes there. The size of larger slabs can be determined using |
2833 | * fls. |
2834 | */ |
2835 | static s8 size_index[24] = { |
2836 | 3, /* 8 */ |
2837 | 4, /* 16 */ |
2838 | 5, /* 24 */ |
2839 | 5, /* 32 */ |
2840 | 6, /* 40 */ |
2841 | 6, /* 48 */ |
2842 | 6, /* 56 */ |
2843 | 6, /* 64 */ |
2844 | 1, /* 72 */ |
2845 | 1, /* 80 */ |
2846 | 1, /* 88 */ |
2847 | 1, /* 96 */ |
2848 | 7, /* 104 */ |
2849 | 7, /* 112 */ |
2850 | 7, /* 120 */ |
2851 | 7, /* 128 */ |
2852 | 2, /* 136 */ |
2853 | 2, /* 144 */ |
2854 | 2, /* 152 */ |
2855 | 2, /* 160 */ |
2856 | 2, /* 168 */ |
2857 | 2, /* 176 */ |
2858 | 2, /* 184 */ |
2859 | 2 /* 192 */ |
2860 | }; |
2861 | |
2862 | static inline int size_index_elem(size_t bytes) |
2863 | { |
2864 | return (bytes - 1) / 8; |
2865 | } |
2866 | |
2867 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) |
2868 | { |
2869 | int index; |
2870 | |
2871 | if (size <= 192) { |
2872 | if (!size) |
2873 | return ZERO_SIZE_PTR; |
2874 | |
2875 | index = size_index[size_index_elem(size)]; |
2876 | } else |
2877 | index = fls(size - 1); |
2878 | |
2879 | #ifdef CONFIG_ZONE_DMA |
2880 | if (unlikely((flags & SLUB_DMA))) |
2881 | return kmalloc_dma_caches[index]; |
2882 | |
2883 | #endif |
2884 | return kmalloc_caches[index]; |
2885 | } |
2886 | |
2887 | void *__kmalloc(size_t size, gfp_t flags) |
2888 | { |
2889 | struct kmem_cache *s; |
2890 | void *ret; |
2891 | |
2892 | if (unlikely(size > SLUB_MAX_SIZE)) |
2893 | return kmalloc_large(size, flags); |
2894 | |
2895 | s = get_slab(size, flags); |
2896 | |
2897 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
2898 | return s; |
2899 | |
2900 | ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); |
2901 | |
2902 | trace_kmalloc(_RET_IP_, ret, size, s->size, flags); |
2903 | |
2904 | return ret; |
2905 | } |
2906 | EXPORT_SYMBOL(__kmalloc); |
2907 | |
2908 | #ifdef CONFIG_NUMA |
2909 | static void *kmalloc_large_node(size_t size, gfp_t flags, int node) |
2910 | { |
2911 | struct page *page; |
2912 | void *ptr = NULL; |
2913 | |
2914 | flags |= __GFP_COMP | __GFP_NOTRACK; |
2915 | page = alloc_pages_node(node, flags, get_order(size)); |
2916 | if (page) |
2917 | ptr = page_address(page); |
2918 | |
2919 | kmemleak_alloc(ptr, size, 1, flags); |
2920 | return ptr; |
2921 | } |
2922 | |
2923 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
2924 | { |
2925 | struct kmem_cache *s; |
2926 | void *ret; |
2927 | |
2928 | if (unlikely(size > SLUB_MAX_SIZE)) { |
2929 | ret = kmalloc_large_node(size, flags, node); |
2930 | |
2931 | trace_kmalloc_node(_RET_IP_, ret, |
2932 | size, PAGE_SIZE << get_order(size), |
2933 | flags, node); |
2934 | |
2935 | return ret; |
2936 | } |
2937 | |
2938 | s = get_slab(size, flags); |
2939 | |
2940 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
2941 | return s; |
2942 | |
2943 | ret = slab_alloc(s, flags, node, _RET_IP_); |
2944 | |
2945 | trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); |
2946 | |
2947 | return ret; |
2948 | } |
2949 | EXPORT_SYMBOL(__kmalloc_node); |
2950 | #endif |
2951 | |
2952 | size_t ksize(const void *object) |
2953 | { |
2954 | struct page *page; |
2955 | |
2956 | if (unlikely(object == ZERO_SIZE_PTR)) |
2957 | return 0; |
2958 | |
2959 | page = virt_to_head_page(object); |
2960 | |
2961 | if (unlikely(!PageSlab(page))) { |
2962 | WARN_ON(!PageCompound(page)); |
2963 | return PAGE_SIZE << compound_order(page); |
2964 | } |
2965 | |
2966 | return slab_ksize(page->slab); |
2967 | } |
2968 | EXPORT_SYMBOL(ksize); |
2969 | |
2970 | void kfree(const void *x) |
2971 | { |
2972 | struct page *page; |
2973 | void *object = (void *)x; |
2974 | |
2975 | trace_kfree(_RET_IP_, x); |
2976 | |
2977 | if (unlikely(ZERO_OR_NULL_PTR(x))) |
2978 | return; |
2979 | |
2980 | page = virt_to_head_page(x); |
2981 | if (unlikely(!PageSlab(page))) { |
2982 | BUG_ON(!PageCompound(page)); |
2983 | kmemleak_free(x); |
2984 | put_page(page); |
2985 | return; |
2986 | } |
2987 | slab_free(page->slab, page, object, _RET_IP_); |
2988 | } |
2989 | EXPORT_SYMBOL(kfree); |
2990 | |
2991 | /* |
2992 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts |
2993 | * the remaining slabs by the number of items in use. The slabs with the |
2994 | * most items in use come first. New allocations will then fill those up |
2995 | * and thus they can be removed from the partial lists. |
2996 | * |
2997 | * The slabs with the least items are placed last. This results in them |
2998 | * being allocated from last increasing the chance that the last objects |
2999 | * are freed in them. |
3000 | */ |
3001 | int kmem_cache_shrink(struct kmem_cache *s) |
3002 | { |
3003 | int node; |
3004 | int i; |
3005 | struct kmem_cache_node *n; |
3006 | struct page *page; |
3007 | struct page *t; |
3008 | int objects = oo_objects(s->max); |
3009 | struct list_head *slabs_by_inuse = |
3010 | kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); |
3011 | unsigned long flags; |
3012 | |
3013 | if (!slabs_by_inuse) |
3014 | return -ENOMEM; |
3015 | |
3016 | flush_all(s); |
3017 | for_each_node_state(node, N_NORMAL_MEMORY) { |
3018 | n = get_node(s, node); |
3019 | |
3020 | if (!n->nr_partial) |
3021 | continue; |
3022 | |
3023 | for (i = 0; i < objects; i++) |
3024 | INIT_LIST_HEAD(slabs_by_inuse + i); |
3025 | |
3026 | spin_lock_irqsave(&n->list_lock, flags); |
3027 | |
3028 | /* |
3029 | * Build lists indexed by the items in use in each slab. |
3030 | * |
3031 | * Note that concurrent frees may occur while we hold the |
3032 | * list_lock. page->inuse here is the upper limit. |
3033 | */ |
3034 | list_for_each_entry_safe(page, t, &n->partial, lru) { |
3035 | if (!page->inuse && slab_trylock(page)) { |
3036 | /* |
3037 | * Must hold slab lock here because slab_free |
3038 | * may have freed the last object and be |
3039 | * waiting to release the slab. |
3040 | */ |
3041 | __remove_partial(n, page); |
3042 | slab_unlock(page); |
3043 | discard_slab(s, page); |
3044 | } else { |
3045 | list_move(&page->lru, |
3046 | slabs_by_inuse + page->inuse); |
3047 | } |
3048 | } |
3049 | |
3050 | /* |
3051 | * Rebuild the partial list with the slabs filled up most |
3052 | * first and the least used slabs at the end. |
3053 | */ |
3054 | for (i = objects - 1; i >= 0; i--) |
3055 | list_splice(slabs_by_inuse + i, n->partial.prev); |
3056 | |
3057 | spin_unlock_irqrestore(&n->list_lock, flags); |
3058 | } |
3059 | |
3060 | kfree(slabs_by_inuse); |
3061 | return 0; |
3062 | } |
3063 | EXPORT_SYMBOL(kmem_cache_shrink); |
3064 | |
3065 | #if defined(CONFIG_MEMORY_HOTPLUG) |
3066 | static int slab_mem_going_offline_callback(void *arg) |
3067 | { |
3068 | struct kmem_cache *s; |
3069 | |
3070 | down_read(&slub_lock); |
3071 | list_for_each_entry(s, &slab_caches, list) |
3072 | kmem_cache_shrink(s); |
3073 | up_read(&slub_lock); |
3074 | |
3075 | return 0; |
3076 | } |
3077 | |
3078 | static void slab_mem_offline_callback(void *arg) |
3079 | { |
3080 | struct kmem_cache_node *n; |
3081 | struct kmem_cache *s; |
3082 | struct memory_notify *marg = arg; |
3083 | int offline_node; |
3084 | |
3085 | offline_node = marg->status_change_nid; |
3086 | |
3087 | /* |
3088 | * If the node still has available memory. we need kmem_cache_node |
3089 | * for it yet. |
3090 | */ |
3091 | if (offline_node < 0) |
3092 | return; |
3093 | |
3094 | down_read(&slub_lock); |
3095 | list_for_each_entry(s, &slab_caches, list) { |
3096 | n = get_node(s, offline_node); |
3097 | if (n) { |
3098 | /* |
3099 | * if n->nr_slabs > 0, slabs still exist on the node |
3100 | * that is going down. We were unable to free them, |
3101 | * and offline_pages() function shouldn't call this |
3102 | * callback. So, we must fail. |
3103 | */ |
3104 | BUG_ON(slabs_node(s, offline_node)); |
3105 | |
3106 | s->node[offline_node] = NULL; |
3107 | kmem_cache_free(kmem_cache_node, n); |
3108 | } |
3109 | } |
3110 | up_read(&slub_lock); |
3111 | } |
3112 | |
3113 | static int slab_mem_going_online_callback(void *arg) |
3114 | { |
3115 | struct kmem_cache_node *n; |
3116 | struct kmem_cache *s; |
3117 | struct memory_notify *marg = arg; |
3118 | int nid = marg->status_change_nid; |
3119 | int ret = 0; |
3120 | |
3121 | /* |
3122 | * If the node's memory is already available, then kmem_cache_node is |
3123 | * already created. Nothing to do. |
3124 | */ |
3125 | if (nid < 0) |
3126 | return 0; |
3127 | |
3128 | /* |
3129 | * We are bringing a node online. No memory is available yet. We must |
3130 | * allocate a kmem_cache_node structure in order to bring the node |
3131 | * online. |
3132 | */ |
3133 | down_read(&slub_lock); |
3134 | list_for_each_entry(s, &slab_caches, list) { |
3135 | /* |
3136 | * XXX: kmem_cache_alloc_node will fallback to other nodes |
3137 | * since memory is not yet available from the node that |
3138 | * is brought up. |
3139 | */ |
3140 | n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); |
3141 | if (!n) { |
3142 | ret = -ENOMEM; |
3143 | goto out; |
3144 | } |
3145 | init_kmem_cache_node(n, s); |
3146 | s->node[nid] = n; |
3147 | } |
3148 | out: |
3149 | up_read(&slub_lock); |
3150 | return ret; |
3151 | } |
3152 | |
3153 | static int slab_memory_callback(struct notifier_block *self, |
3154 | unsigned long action, void *arg) |
3155 | { |
3156 | int ret = 0; |
3157 | |
3158 | switch (action) { |
3159 | case MEM_GOING_ONLINE: |
3160 | ret = slab_mem_going_online_callback(arg); |
3161 | break; |
3162 | case MEM_GOING_OFFLINE: |
3163 | ret = slab_mem_going_offline_callback(arg); |
3164 | break; |
3165 | case MEM_OFFLINE: |
3166 | case MEM_CANCEL_ONLINE: |
3167 | slab_mem_offline_callback(arg); |
3168 | break; |
3169 | case MEM_ONLINE: |
3170 | case MEM_CANCEL_OFFLINE: |
3171 | break; |
3172 | } |
3173 | if (ret) |
3174 | ret = notifier_from_errno(ret); |
3175 | else |
3176 | ret = NOTIFY_OK; |
3177 | return ret; |
3178 | } |
3179 | |
3180 | #endif /* CONFIG_MEMORY_HOTPLUG */ |
3181 | |
3182 | /******************************************************************** |
3183 | * Basic setup of slabs |
3184 | *******************************************************************/ |
3185 | |
3186 | /* |
3187 | * Used for early kmem_cache structures that were allocated using |
3188 | * the page allocator |
3189 | */ |
3190 | |
3191 | static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s) |
3192 | { |
3193 | int node; |
3194 | |
3195 | list_add(&s->list, &slab_caches); |
3196 | s->refcount = -1; |
3197 | |
3198 | for_each_node_state(node, N_NORMAL_MEMORY) { |
3199 | struct kmem_cache_node *n = get_node(s, node); |
3200 | struct page *p; |
3201 | |
3202 | if (n) { |
3203 | list_for_each_entry(p, &n->partial, lru) |
3204 | p->slab = s; |
3205 | |
3206 | #ifdef CONFIG_SLAB_DEBUG |
3207 | list_for_each_entry(p, &n->full, lru) |
3208 | p->slab = s; |
3209 | #endif |
3210 | } |
3211 | } |
3212 | } |
3213 | |
3214 | void __init kmem_cache_init(void) |
3215 | { |
3216 | int i; |
3217 | int caches = 0; |
3218 | struct kmem_cache *temp_kmem_cache; |
3219 | int order; |
3220 | struct kmem_cache *temp_kmem_cache_node; |
3221 | unsigned long kmalloc_size; |
3222 | |
3223 | kmem_size = offsetof(struct kmem_cache, node) + |
3224 | nr_node_ids * sizeof(struct kmem_cache_node *); |
3225 | |
3226 | /* Allocate two kmem_caches from the page allocator */ |
3227 | kmalloc_size = ALIGN(kmem_size, cache_line_size()); |
3228 | order = get_order(2 * kmalloc_size); |
3229 | kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order); |
3230 | |
3231 | /* |
3232 | * Must first have the slab cache available for the allocations of the |
3233 | * struct kmem_cache_node's. There is special bootstrap code in |
3234 | * kmem_cache_open for slab_state == DOWN. |
3235 | */ |
3236 | kmem_cache_node = (void *)kmem_cache + kmalloc_size; |
3237 | |
3238 | kmem_cache_open(kmem_cache_node, "kmem_cache_node", |
3239 | sizeof(struct kmem_cache_node), |
3240 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); |
3241 | |
3242 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); |
3243 | |
3244 | /* Able to allocate the per node structures */ |
3245 | slab_state = PARTIAL; |
3246 | |
3247 | temp_kmem_cache = kmem_cache; |
3248 | kmem_cache_open(kmem_cache, "kmem_cache", kmem_size, |
3249 | 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); |
3250 | kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); |
3251 | memcpy(kmem_cache, temp_kmem_cache, kmem_size); |
3252 | |
3253 | /* |
3254 | * Allocate kmem_cache_node properly from the kmem_cache slab. |
3255 | * kmem_cache_node is separately allocated so no need to |
3256 | * update any list pointers. |
3257 | */ |
3258 | temp_kmem_cache_node = kmem_cache_node; |
3259 | |
3260 | kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); |
3261 | memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size); |
3262 | |
3263 | kmem_cache_bootstrap_fixup(kmem_cache_node); |
3264 | |
3265 | caches++; |
3266 | kmem_cache_bootstrap_fixup(kmem_cache); |
3267 | caches++; |
3268 | /* Free temporary boot structure */ |
3269 | free_pages((unsigned long)temp_kmem_cache, order); |
3270 | |
3271 | /* Now we can use the kmem_cache to allocate kmalloc slabs */ |
3272 | |
3273 | /* |
3274 | * Patch up the size_index table if we have strange large alignment |
3275 | * requirements for the kmalloc array. This is only the case for |
3276 | * MIPS it seems. The standard arches will not generate any code here. |
3277 | * |
3278 | * Largest permitted alignment is 256 bytes due to the way we |
3279 | * handle the index determination for the smaller caches. |
3280 | * |
3281 | * Make sure that nothing crazy happens if someone starts tinkering |
3282 | * around with ARCH_KMALLOC_MINALIGN |
3283 | */ |
3284 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || |
3285 | (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); |
3286 | |
3287 | for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { |
3288 | int elem = size_index_elem(i); |
3289 | if (elem >= ARRAY_SIZE(size_index)) |
3290 | break; |
3291 | size_index[elem] = KMALLOC_SHIFT_LOW; |
3292 | } |
3293 | |
3294 | if (KMALLOC_MIN_SIZE == 64) { |
3295 | /* |
3296 | * The 96 byte size cache is not used if the alignment |
3297 | * is 64 byte. |
3298 | */ |
3299 | for (i = 64 + 8; i <= 96; i += 8) |
3300 | size_index[size_index_elem(i)] = 7; |
3301 | } else if (KMALLOC_MIN_SIZE == 128) { |
3302 | /* |
3303 | * The 192 byte sized cache is not used if the alignment |
3304 | * is 128 byte. Redirect kmalloc to use the 256 byte cache |
3305 | * instead. |
3306 | */ |
3307 | for (i = 128 + 8; i <= 192; i += 8) |
3308 | size_index[size_index_elem(i)] = 8; |
3309 | } |
3310 | |
3311 | /* Caches that are not of the two-to-the-power-of size */ |
3312 | if (KMALLOC_MIN_SIZE <= 32) { |
3313 | kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0); |
3314 | caches++; |
3315 | } |
3316 | |
3317 | if (KMALLOC_MIN_SIZE <= 64) { |
3318 | kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0); |
3319 | caches++; |
3320 | } |
3321 | |
3322 | for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { |
3323 | kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0); |
3324 | caches++; |
3325 | } |
3326 | |
3327 | slab_state = UP; |
3328 | |
3329 | /* Provide the correct kmalloc names now that the caches are up */ |
3330 | if (KMALLOC_MIN_SIZE <= 32) { |
3331 | kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT); |
3332 | BUG_ON(!kmalloc_caches[1]->name); |
3333 | } |
3334 | |
3335 | if (KMALLOC_MIN_SIZE <= 64) { |
3336 | kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT); |
3337 | BUG_ON(!kmalloc_caches[2]->name); |
3338 | } |
3339 | |
3340 | for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { |
3341 | char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); |
3342 | |
3343 | BUG_ON(!s); |
3344 | kmalloc_caches[i]->name = s; |
3345 | } |
3346 | |
3347 | #ifdef CONFIG_SMP |
3348 | register_cpu_notifier(&slab_notifier); |
3349 | #endif |
3350 | |
3351 | #ifdef CONFIG_ZONE_DMA |
3352 | for (i = 0; i < SLUB_PAGE_SHIFT; i++) { |
3353 | struct kmem_cache *s = kmalloc_caches[i]; |
3354 | |
3355 | if (s && s->size) { |
3356 | char *name = kasprintf(GFP_NOWAIT, |
3357 | "dma-kmalloc-%d", s->objsize); |
3358 | |
3359 | BUG_ON(!name); |
3360 | kmalloc_dma_caches[i] = create_kmalloc_cache(name, |
3361 | s->objsize, SLAB_CACHE_DMA); |
3362 | } |
3363 | } |
3364 | #endif |
3365 | printk(KERN_INFO |
3366 | "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," |
3367 | " CPUs=%d, Nodes=%d\n", |
3368 | caches, cache_line_size(), |
3369 | slub_min_order, slub_max_order, slub_min_objects, |
3370 | nr_cpu_ids, nr_node_ids); |
3371 | } |
3372 | |
3373 | void __init kmem_cache_init_late(void) |
3374 | { |
3375 | } |
3376 | |
3377 | /* |
3378 | * Find a mergeable slab cache |
3379 | */ |
3380 | static int slab_unmergeable(struct kmem_cache *s) |
3381 | { |
3382 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) |
3383 | return 1; |
3384 | |
3385 | if (s->ctor) |
3386 | return 1; |
3387 | |
3388 | /* |
3389 | * We may have set a slab to be unmergeable during bootstrap. |
3390 | */ |
3391 | if (s->refcount < 0) |
3392 | return 1; |
3393 | |
3394 | return 0; |
3395 | } |
3396 | |
3397 | static struct kmem_cache *find_mergeable(size_t size, |
3398 | size_t align, unsigned long flags, const char *name, |
3399 | void (*ctor)(void *)) |
3400 | { |
3401 | struct kmem_cache *s; |
3402 | |
3403 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) |
3404 | return NULL; |
3405 | |
3406 | if (ctor) |
3407 | return NULL; |
3408 | |
3409 | size = ALIGN(size, sizeof(void *)); |
3410 | align = calculate_alignment(flags, align, size); |
3411 | size = ALIGN(size, align); |
3412 | flags = kmem_cache_flags(size, flags, name, NULL); |
3413 | |
3414 | list_for_each_entry(s, &slab_caches, list) { |
3415 | if (slab_unmergeable(s)) |
3416 | continue; |
3417 | |
3418 | if (size > s->size) |
3419 | continue; |
3420 | |
3421 | if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) |
3422 | continue; |
3423 | /* |
3424 | * Check if alignment is compatible. |
3425 | * Courtesy of Adrian Drzewiecki |
3426 | */ |
3427 | if ((s->size & ~(align - 1)) != s->size) |
3428 | continue; |
3429 | |
3430 | if (s->size - size >= sizeof(void *)) |
3431 | continue; |
3432 | |
3433 | return s; |
3434 | } |
3435 | return NULL; |
3436 | } |
3437 | |
3438 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, |
3439 | size_t align, unsigned long flags, void (*ctor)(void *)) |
3440 | { |
3441 | struct kmem_cache *s; |
3442 | char *n; |
3443 | |
3444 | if (WARN_ON(!name)) |
3445 | return NULL; |
3446 | |
3447 | down_write(&slub_lock); |
3448 | s = find_mergeable(size, align, flags, name, ctor); |
3449 | if (s) { |
3450 | s->refcount++; |
3451 | /* |
3452 | * Adjust the object sizes so that we clear |
3453 | * the complete object on kzalloc. |
3454 | */ |
3455 | s->objsize = max(s->objsize, (int)size); |
3456 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); |
3457 | |
3458 | if (sysfs_slab_alias(s, name)) { |
3459 | s->refcount--; |
3460 | goto err; |
3461 | } |
3462 | up_write(&slub_lock); |
3463 | return s; |
3464 | } |
3465 | |
3466 | n = kstrdup(name, GFP_KERNEL); |
3467 | if (!n) |
3468 | goto err; |
3469 | |
3470 | s = kmalloc(kmem_size, GFP_KERNEL); |
3471 | if (s) { |
3472 | if (kmem_cache_open(s, n, |
3473 | size, align, flags, ctor)) { |
3474 | list_add(&s->list, &slab_caches); |
3475 | if (sysfs_slab_add(s)) { |
3476 | list_del(&s->list); |
3477 | kfree(n); |
3478 | kfree(s); |
3479 | goto err; |
3480 | } |
3481 | up_write(&slub_lock); |
3482 | return s; |
3483 | } |
3484 | kfree(n); |
3485 | kfree(s); |
3486 | } |
3487 | err: |
3488 | up_write(&slub_lock); |
3489 | |
3490 | if (flags & SLAB_PANIC) |
3491 | panic("Cannot create slabcache %s\n", name); |
3492 | else |
3493 | s = NULL; |
3494 | return s; |
3495 | } |
3496 | EXPORT_SYMBOL(kmem_cache_create); |
3497 | |
3498 | #ifdef CONFIG_SMP |
3499 | /* |
3500 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
3501 | * necessary. |
3502 | */ |
3503 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, |
3504 | unsigned long action, void *hcpu) |
3505 | { |
3506 | long cpu = (long)hcpu; |
3507 | struct kmem_cache *s; |
3508 | unsigned long flags; |
3509 | |
3510 | switch (action) { |
3511 | case CPU_UP_CANCELED: |
3512 | case CPU_UP_CANCELED_FROZEN: |
3513 | case CPU_DEAD: |
3514 | case CPU_DEAD_FROZEN: |
3515 | down_read(&slub_lock); |
3516 | list_for_each_entry(s, &slab_caches, list) { |
3517 | local_irq_save(flags); |
3518 | __flush_cpu_slab(s, cpu); |
3519 | local_irq_restore(flags); |
3520 | } |
3521 | up_read(&slub_lock); |
3522 | break; |
3523 | default: |
3524 | break; |
3525 | } |
3526 | return NOTIFY_OK; |
3527 | } |
3528 | |
3529 | static struct notifier_block __cpuinitdata slab_notifier = { |
3530 | .notifier_call = slab_cpuup_callback |
3531 | }; |
3532 | |
3533 | #endif |
3534 | |
3535 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) |
3536 | { |
3537 | struct kmem_cache *s; |
3538 | void *ret; |
3539 | |
3540 | if (unlikely(size > SLUB_MAX_SIZE)) |
3541 | return kmalloc_large(size, gfpflags); |
3542 | |
3543 | s = get_slab(size, gfpflags); |
3544 | |
3545 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
3546 | return s; |
3547 | |
3548 | ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); |
3549 | |
3550 | /* Honor the call site pointer we received. */ |
3551 | trace_kmalloc(caller, ret, size, s->size, gfpflags); |
3552 | |
3553 | return ret; |
3554 | } |
3555 | |
3556 | #ifdef CONFIG_NUMA |
3557 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, |
3558 | int node, unsigned long caller) |
3559 | { |
3560 | struct kmem_cache *s; |
3561 | void *ret; |
3562 | |
3563 | if (unlikely(size > SLUB_MAX_SIZE)) { |
3564 | ret = kmalloc_large_node(size, gfpflags, node); |
3565 | |
3566 | trace_kmalloc_node(caller, ret, |
3567 | size, PAGE_SIZE << get_order(size), |
3568 | gfpflags, node); |
3569 | |
3570 | return ret; |
3571 | } |
3572 | |
3573 | s = get_slab(size, gfpflags); |
3574 | |
3575 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
3576 | return s; |
3577 | |
3578 | ret = slab_alloc(s, gfpflags, node, caller); |
3579 | |
3580 | /* Honor the call site pointer we received. */ |
3581 | trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); |
3582 | |
3583 | return ret; |
3584 | } |
3585 | #endif |
3586 | |
3587 | #ifdef CONFIG_SYSFS |
3588 | static int count_inuse(struct page *page) |
3589 | { |
3590 | return page->inuse; |
3591 | } |
3592 | |
3593 | static int count_total(struct page *page) |
3594 | { |
3595 | return page->objects; |
3596 | } |
3597 | #endif |
3598 | |
3599 | #ifdef CONFIG_SLUB_DEBUG |
3600 | static int validate_slab(struct kmem_cache *s, struct page *page, |
3601 | unsigned long *map) |
3602 | { |
3603 | void *p; |
3604 | void *addr = page_address(page); |
3605 | |
3606 | if (!check_slab(s, page) || |
3607 | !on_freelist(s, page, NULL)) |
3608 | return 0; |
3609 | |
3610 | /* Now we know that a valid freelist exists */ |
3611 | bitmap_zero(map, page->objects); |
3612 | |
3613 | for_each_free_object(p, s, page->freelist) { |
3614 | set_bit(slab_index(p, s, addr), map); |
3615 | if (!check_object(s, page, p, SLUB_RED_INACTIVE)) |
3616 | return 0; |
3617 | } |
3618 | |
3619 | for_each_object(p, s, addr, page->objects) |
3620 | if (!test_bit(slab_index(p, s, addr), map)) |
3621 | if (!check_object(s, page, p, SLUB_RED_ACTIVE)) |
3622 | return 0; |
3623 | return 1; |
3624 | } |
3625 | |
3626 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, |
3627 | unsigned long *map) |
3628 | { |
3629 | if (slab_trylock(page)) { |
3630 | validate_slab(s, page, map); |
3631 | slab_unlock(page); |
3632 | } else |
3633 | printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", |
3634 | s->name, page); |
3635 | } |
3636 | |
3637 | static int validate_slab_node(struct kmem_cache *s, |
3638 | struct kmem_cache_node *n, unsigned long *map) |
3639 | { |
3640 | unsigned long count = 0; |
3641 | struct page *page; |
3642 | unsigned long flags; |
3643 | |
3644 | spin_lock_irqsave(&n->list_lock, flags); |
3645 | |
3646 | list_for_each_entry(page, &n->partial, lru) { |
3647 | validate_slab_slab(s, page, map); |
3648 | count++; |
3649 | } |
3650 | if (count != n->nr_partial) |
3651 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " |
3652 | "counter=%ld\n", s->name, count, n->nr_partial); |
3653 | |
3654 | if (!(s->flags & SLAB_STORE_USER)) |
3655 | goto out; |
3656 | |
3657 | list_for_each_entry(page, &n->full, lru) { |
3658 | validate_slab_slab(s, page, map); |
3659 | count++; |
3660 | } |
3661 | if (count != atomic_long_read(&n->nr_slabs)) |
3662 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " |
3663 | "counter=%ld\n", s->name, count, |
3664 | atomic_long_read(&n->nr_slabs)); |
3665 | |
3666 | out: |
3667 | spin_unlock_irqrestore(&n->list_lock, flags); |
3668 | return count; |
3669 | } |
3670 | |
3671 | static long validate_slab_cache(struct kmem_cache *s) |
3672 | { |
3673 | int node; |
3674 | unsigned long count = 0; |
3675 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * |
3676 | sizeof(unsigned long), GFP_KERNEL); |
3677 | |
3678 | if (!map) |
3679 | return -ENOMEM; |
3680 | |
3681 | flush_all(s); |
3682 | for_each_node_state(node, N_NORMAL_MEMORY) { |
3683 | struct kmem_cache_node *n = get_node(s, node); |
3684 | |
3685 | count += validate_slab_node(s, n, map); |
3686 | } |
3687 | kfree(map); |
3688 | return count; |
3689 | } |
3690 | /* |
3691 | * Generate lists of code addresses where slabcache objects are allocated |
3692 | * and freed. |
3693 | */ |
3694 | |
3695 | struct location { |
3696 | unsigned long count; |
3697 | unsigned long addr; |
3698 | long long sum_time; |
3699 | long min_time; |
3700 | long max_time; |
3701 | long min_pid; |
3702 | long max_pid; |
3703 | DECLARE_BITMAP(cpus, NR_CPUS); |
3704 | nodemask_t nodes; |
3705 | }; |
3706 | |
3707 | struct loc_track { |
3708 | unsigned long max; |
3709 | unsigned long count; |
3710 | struct location *loc; |
3711 | }; |
3712 | |
3713 | static void free_loc_track(struct loc_track *t) |
3714 | { |
3715 | if (t->max) |
3716 | free_pages((unsigned long)t->loc, |
3717 | get_order(sizeof(struct location) * t->max)); |
3718 | } |
3719 | |
3720 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) |
3721 | { |
3722 | struct location *l; |
3723 | int order; |
3724 | |
3725 | order = get_order(sizeof(struct location) * max); |
3726 | |
3727 | l = (void *)__get_free_pages(flags, order); |
3728 | if (!l) |
3729 | return 0; |
3730 | |
3731 | if (t->count) { |
3732 | memcpy(l, t->loc, sizeof(struct location) * t->count); |
3733 | free_loc_track(t); |
3734 | } |
3735 | t->max = max; |
3736 | t->loc = l; |
3737 | return 1; |
3738 | } |
3739 | |
3740 | static int add_location(struct loc_track *t, struct kmem_cache *s, |
3741 | const struct track *track) |
3742 | { |
3743 | long start, end, pos; |
3744 | struct location *l; |
3745 | unsigned long caddr; |
3746 | unsigned long age = jiffies - track->when; |
3747 | |
3748 | start = -1; |
3749 | end = t->count; |
3750 | |
3751 | for ( ; ; ) { |
3752 | pos = start + (end - start + 1) / 2; |
3753 | |
3754 | /* |
3755 | * There is nothing at "end". If we end up there |
3756 | * we need to add something to before end. |
3757 | */ |
3758 | if (pos == end) |
3759 | break; |
3760 | |
3761 | caddr = t->loc[pos].addr; |
3762 | if (track->addr == caddr) { |
3763 | |
3764 | l = &t->loc[pos]; |
3765 | l->count++; |
3766 | if (track->when) { |
3767 | l->sum_time += age; |
3768 | if (age < l->min_time) |
3769 | l->min_time = age; |
3770 | if (age > l->max_time) |
3771 | l->max_time = age; |
3772 | |
3773 | if (track->pid < l->min_pid) |
3774 | l->min_pid = track->pid; |
3775 | if (track->pid > l->max_pid) |
3776 | l->max_pid = track->pid; |
3777 | |
3778 | cpumask_set_cpu(track->cpu, |
3779 | to_cpumask(l->cpus)); |
3780 | } |
3781 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
3782 | return 1; |
3783 | } |
3784 | |
3785 | if (track->addr < caddr) |
3786 | end = pos; |
3787 | else |
3788 | start = pos; |
3789 | } |
3790 | |
3791 | /* |
3792 | * Not found. Insert new tracking element. |
3793 | */ |
3794 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) |
3795 | return 0; |
3796 | |
3797 | l = t->loc + pos; |
3798 | if (pos < t->count) |
3799 | memmove(l + 1, l, |
3800 | (t->count - pos) * sizeof(struct location)); |
3801 | t->count++; |
3802 | l->count = 1; |
3803 | l->addr = track->addr; |
3804 | l->sum_time = age; |
3805 | l->min_time = age; |
3806 | l->max_time = age; |
3807 | l->min_pid = track->pid; |
3808 | l->max_pid = track->pid; |
3809 | cpumask_clear(to_cpumask(l->cpus)); |
3810 | cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); |
3811 | nodes_clear(l->nodes); |
3812 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
3813 | return 1; |
3814 | } |
3815 | |
3816 | static void process_slab(struct loc_track *t, struct kmem_cache *s, |
3817 | struct page *page, enum track_item alloc, |
3818 | unsigned long *map) |
3819 | { |
3820 | void *addr = page_address(page); |
3821 | void *p; |
3822 | |
3823 | bitmap_zero(map, page->objects); |
3824 | for_each_free_object(p, s, page->freelist) |
3825 | set_bit(slab_index(p, s, addr), map); |
3826 | |
3827 | for_each_object(p, s, addr, page->objects) |
3828 | if (!test_bit(slab_index(p, s, addr), map)) |
3829 | add_location(t, s, get_track(s, p, alloc)); |
3830 | } |
3831 | |
3832 | static int list_locations(struct kmem_cache *s, char *buf, |
3833 | enum track_item alloc) |
3834 | { |
3835 | int len = 0; |
3836 | unsigned long i; |
3837 | struct loc_track t = { 0, 0, NULL }; |
3838 | int node; |
3839 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * |
3840 | sizeof(unsigned long), GFP_KERNEL); |
3841 | |
3842 | if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), |
3843 | GFP_TEMPORARY)) { |
3844 | kfree(map); |
3845 | return sprintf(buf, "Out of memory\n"); |
3846 | } |
3847 | /* Push back cpu slabs */ |
3848 | flush_all(s); |
3849 | |
3850 | for_each_node_state(node, N_NORMAL_MEMORY) { |
3851 | struct kmem_cache_node *n = get_node(s, node); |
3852 | unsigned long flags; |
3853 | struct page *page; |
3854 | |
3855 | if (!atomic_long_read(&n->nr_slabs)) |
3856 | continue; |
3857 | |
3858 | spin_lock_irqsave(&n->list_lock, flags); |
3859 | list_for_each_entry(page, &n->partial, lru) |
3860 | process_slab(&t, s, page, alloc, map); |
3861 | list_for_each_entry(page, &n->full, lru) |
3862 | process_slab(&t, s, page, alloc, map); |
3863 | spin_unlock_irqrestore(&n->list_lock, flags); |
3864 | } |
3865 | |
3866 | for (i = 0; i < t.count; i++) { |
3867 | struct location *l = &t.loc[i]; |
3868 | |
3869 | if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) |
3870 | break; |
3871 | len += sprintf(buf + len, "%7ld ", l->count); |
3872 | |
3873 | if (l->addr) |
3874 | len += sprintf(buf + len, "%pS", (void *)l->addr); |
3875 | else |
3876 | len += sprintf(buf + len, "<not-available>"); |
3877 | |
3878 | if (l->sum_time != l->min_time) { |
3879 | len += sprintf(buf + len, " age=%ld/%ld/%ld", |
3880 | l->min_time, |
3881 | (long)div_u64(l->sum_time, l->count), |
3882 | l->max_time); |
3883 | } else |
3884 | len += sprintf(buf + len, " age=%ld", |
3885 | l->min_time); |
3886 | |
3887 | if (l->min_pid != l->max_pid) |
3888 | len += sprintf(buf + len, " pid=%ld-%ld", |
3889 | l->min_pid, l->max_pid); |
3890 | else |
3891 | len += sprintf(buf + len, " pid=%ld", |
3892 | l->min_pid); |
3893 | |
3894 | if (num_online_cpus() > 1 && |
3895 | !cpumask_empty(to_cpumask(l->cpus)) && |
3896 | len < PAGE_SIZE - 60) { |
3897 | len += sprintf(buf + len, " cpus="); |
3898 | len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, |
3899 | to_cpumask(l->cpus)); |
3900 | } |
3901 | |
3902 | if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && |
3903 | len < PAGE_SIZE - 60) { |
3904 | len += sprintf(buf + len, " nodes="); |
3905 | len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, |
3906 | l->nodes); |
3907 | } |
3908 | |
3909 | len += sprintf(buf + len, "\n"); |
3910 | } |
3911 | |
3912 | free_loc_track(&t); |
3913 | kfree(map); |
3914 | if (!t.count) |
3915 | len += sprintf(buf, "No data\n"); |
3916 | return len; |
3917 | } |
3918 | #endif |
3919 | |
3920 | #ifdef SLUB_RESILIENCY_TEST |
3921 | static void resiliency_test(void) |
3922 | { |
3923 | u8 *p; |
3924 | |
3925 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10); |
3926 | |
3927 | printk(KERN_ERR "SLUB resiliency testing\n"); |
3928 | printk(KERN_ERR "-----------------------\n"); |
3929 | printk(KERN_ERR "A. Corruption after allocation\n"); |
3930 | |
3931 | p = kzalloc(16, GFP_KERNEL); |
3932 | p[16] = 0x12; |
3933 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" |
3934 | " 0x12->0x%p\n\n", p + 16); |
3935 | |
3936 | validate_slab_cache(kmalloc_caches[4]); |
3937 | |
3938 | /* Hmmm... The next two are dangerous */ |
3939 | p = kzalloc(32, GFP_KERNEL); |
3940 | p[32 + sizeof(void *)] = 0x34; |
3941 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" |
3942 | " 0x34 -> -0x%p\n", p); |
3943 | printk(KERN_ERR |
3944 | "If allocated object is overwritten then not detectable\n\n"); |
3945 | |
3946 | validate_slab_cache(kmalloc_caches[5]); |
3947 | p = kzalloc(64, GFP_KERNEL); |
3948 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); |
3949 | *p = 0x56; |
3950 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", |
3951 | p); |
3952 | printk(KERN_ERR |
3953 | "If allocated object is overwritten then not detectable\n\n"); |
3954 | validate_slab_cache(kmalloc_caches[6]); |
3955 | |
3956 | printk(KERN_ERR "\nB. Corruption after free\n"); |
3957 | p = kzalloc(128, GFP_KERNEL); |
3958 | kfree(p); |
3959 | *p = 0x78; |
3960 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); |
3961 | validate_slab_cache(kmalloc_caches[7]); |
3962 | |
3963 | p = kzalloc(256, GFP_KERNEL); |
3964 | kfree(p); |
3965 | p[50] = 0x9a; |
3966 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", |
3967 | p); |
3968 | validate_slab_cache(kmalloc_caches[8]); |
3969 | |
3970 | p = kzalloc(512, GFP_KERNEL); |
3971 | kfree(p); |
3972 | p[512] = 0xab; |
3973 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); |
3974 | validate_slab_cache(kmalloc_caches[9]); |
3975 | } |
3976 | #else |
3977 | #ifdef CONFIG_SYSFS |
3978 | static void resiliency_test(void) {}; |
3979 | #endif |
3980 | #endif |
3981 | |
3982 | #ifdef CONFIG_SYSFS |
3983 | enum slab_stat_type { |
3984 | SL_ALL, /* All slabs */ |
3985 | SL_PARTIAL, /* Only partially allocated slabs */ |
3986 | SL_CPU, /* Only slabs used for cpu caches */ |
3987 | SL_OBJECTS, /* Determine allocated objects not slabs */ |
3988 | SL_TOTAL /* Determine object capacity not slabs */ |
3989 | }; |
3990 | |
3991 | #define SO_ALL (1 << SL_ALL) |
3992 | #define SO_PARTIAL (1 << SL_PARTIAL) |
3993 | #define SO_CPU (1 << SL_CPU) |
3994 | #define SO_OBJECTS (1 << SL_OBJECTS) |
3995 | #define SO_TOTAL (1 << SL_TOTAL) |
3996 | |
3997 | static ssize_t show_slab_objects(struct kmem_cache *s, |
3998 | char *buf, unsigned long flags) |
3999 | { |
4000 | unsigned long total = 0; |
4001 | int node; |
4002 | int x; |
4003 | unsigned long *nodes; |
4004 | unsigned long *per_cpu; |
4005 | |
4006 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); |
4007 | if (!nodes) |
4008 | return -ENOMEM; |
4009 | per_cpu = nodes + nr_node_ids; |
4010 | |
4011 | if (flags & SO_CPU) { |
4012 | int cpu; |
4013 | |
4014 | for_each_possible_cpu(cpu) { |
4015 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
4016 | |
4017 | if (!c || c->node < 0) |
4018 | continue; |
4019 | |
4020 | if (c->page) { |
4021 | if (flags & SO_TOTAL) |
4022 | x = c->page->objects; |
4023 | else if (flags & SO_OBJECTS) |
4024 | x = c->page->inuse; |
4025 | else |
4026 | x = 1; |
4027 | |
4028 | total += x; |
4029 | nodes[c->node] += x; |
4030 | } |
4031 | per_cpu[c->node]++; |
4032 | } |
4033 | } |
4034 | |
4035 | lock_memory_hotplug(); |
4036 | #ifdef CONFIG_SLUB_DEBUG |
4037 | if (flags & SO_ALL) { |
4038 | for_each_node_state(node, N_NORMAL_MEMORY) { |
4039 | struct kmem_cache_node *n = get_node(s, node); |
4040 | |
4041 | if (flags & SO_TOTAL) |
4042 | x = atomic_long_read(&n->total_objects); |
4043 | else if (flags & SO_OBJECTS) |
4044 | x = atomic_long_read(&n->total_objects) - |
4045 | count_partial(n, count_free); |
4046 | |
4047 | else |
4048 | x = atomic_long_read(&n->nr_slabs); |
4049 | total += x; |
4050 | nodes[node] += x; |
4051 | } |
4052 | |
4053 | } else |
4054 | #endif |
4055 | if (flags & SO_PARTIAL) { |
4056 | for_each_node_state(node, N_NORMAL_MEMORY) { |
4057 | struct kmem_cache_node *n = get_node(s, node); |
4058 | |
4059 | if (flags & SO_TOTAL) |
4060 | x = count_partial(n, count_total); |
4061 | else if (flags & SO_OBJECTS) |
4062 | x = count_partial(n, count_inuse); |
4063 | else |
4064 | x = n->nr_partial; |
4065 | total += x; |
4066 | nodes[node] += x; |
4067 | } |
4068 | } |
4069 | x = sprintf(buf, "%lu", total); |
4070 | #ifdef CONFIG_NUMA |
4071 | for_each_node_state(node, N_NORMAL_MEMORY) |
4072 | if (nodes[node]) |
4073 | x += sprintf(buf + x, " N%d=%lu", |
4074 | node, nodes[node]); |
4075 | #endif |
4076 | unlock_memory_hotplug(); |
4077 | kfree(nodes); |
4078 | return x + sprintf(buf + x, "\n"); |
4079 | } |
4080 | |
4081 | #ifdef CONFIG_SLUB_DEBUG |
4082 | static int any_slab_objects(struct kmem_cache *s) |
4083 | { |
4084 | int node; |
4085 | |
4086 | for_each_online_node(node) { |
4087 | struct kmem_cache_node *n = get_node(s, node); |
4088 | |
4089 | if (!n) |
4090 | continue; |
4091 | |
4092 | if (atomic_long_read(&n->total_objects)) |
4093 | return 1; |
4094 | } |
4095 | return 0; |
4096 | } |
4097 | #endif |
4098 | |
4099 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) |
4100 | #define to_slab(n) container_of(n, struct kmem_cache, kobj); |
4101 | |
4102 | struct slab_attribute { |
4103 | struct attribute attr; |
4104 | ssize_t (*show)(struct kmem_cache *s, char *buf); |
4105 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); |
4106 | }; |
4107 | |
4108 | #define SLAB_ATTR_RO(_name) \ |
4109 | static struct slab_attribute _name##_attr = __ATTR_RO(_name) |
4110 | |
4111 | #define SLAB_ATTR(_name) \ |
4112 | static struct slab_attribute _name##_attr = \ |
4113 | __ATTR(_name, 0644, _name##_show, _name##_store) |
4114 | |
4115 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
4116 | { |
4117 | return sprintf(buf, "%d\n", s->size); |
4118 | } |
4119 | SLAB_ATTR_RO(slab_size); |
4120 | |
4121 | static ssize_t align_show(struct kmem_cache *s, char *buf) |
4122 | { |
4123 | return sprintf(buf, "%d\n", s->align); |
4124 | } |
4125 | SLAB_ATTR_RO(align); |
4126 | |
4127 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) |
4128 | { |
4129 | return sprintf(buf, "%d\n", s->objsize); |
4130 | } |
4131 | SLAB_ATTR_RO(object_size); |
4132 | |
4133 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) |
4134 | { |
4135 | return sprintf(buf, "%d\n", oo_objects(s->oo)); |
4136 | } |
4137 | SLAB_ATTR_RO(objs_per_slab); |
4138 | |
4139 | static ssize_t order_store(struct kmem_cache *s, |
4140 | const char *buf, size_t length) |
4141 | { |
4142 | unsigned long order; |
4143 | int err; |
4144 | |
4145 | err = strict_strtoul(buf, 10, &order); |
4146 | if (err) |
4147 | return err; |
4148 | |
4149 | if (order > slub_max_order || order < slub_min_order) |
4150 | return -EINVAL; |
4151 | |
4152 | calculate_sizes(s, order); |
4153 | return length; |
4154 | } |
4155 | |
4156 | static ssize_t order_show(struct kmem_cache *s, char *buf) |
4157 | { |
4158 | return sprintf(buf, "%d\n", oo_order(s->oo)); |
4159 | } |
4160 | SLAB_ATTR(order); |
4161 | |
4162 | static ssize_t min_partial_show(struct kmem_cache *s, char *buf) |
4163 | { |
4164 | return sprintf(buf, "%lu\n", s->min_partial); |
4165 | } |
4166 | |
4167 | static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, |
4168 | size_t length) |
4169 | { |
4170 | unsigned long min; |
4171 | int err; |
4172 | |
4173 | err = strict_strtoul(buf, 10, &min); |
4174 | if (err) |
4175 | return err; |
4176 | |
4177 | set_min_partial(s, min); |
4178 | return length; |
4179 | } |
4180 | SLAB_ATTR(min_partial); |
4181 | |
4182 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) |
4183 | { |
4184 | if (!s->ctor) |
4185 | return 0; |
4186 | return sprintf(buf, "%pS\n", s->ctor); |
4187 | } |
4188 | SLAB_ATTR_RO(ctor); |
4189 | |
4190 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) |
4191 | { |
4192 | return sprintf(buf, "%d\n", s->refcount - 1); |
4193 | } |
4194 | SLAB_ATTR_RO(aliases); |
4195 | |
4196 | static ssize_t partial_show(struct kmem_cache *s, char *buf) |
4197 | { |
4198 | return show_slab_objects(s, buf, SO_PARTIAL); |
4199 | } |
4200 | SLAB_ATTR_RO(partial); |
4201 | |
4202 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) |
4203 | { |
4204 | return show_slab_objects(s, buf, SO_CPU); |
4205 | } |
4206 | SLAB_ATTR_RO(cpu_slabs); |
4207 | |
4208 | static ssize_t objects_show(struct kmem_cache *s, char *buf) |
4209 | { |
4210 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); |
4211 | } |
4212 | SLAB_ATTR_RO(objects); |
4213 | |
4214 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) |
4215 | { |
4216 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); |
4217 | } |
4218 | SLAB_ATTR_RO(objects_partial); |
4219 | |
4220 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) |
4221 | { |
4222 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); |
4223 | } |
4224 | |
4225 | static ssize_t reclaim_account_store(struct kmem_cache *s, |
4226 | const char *buf, size_t length) |
4227 | { |
4228 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; |
4229 | if (buf[0] == '1') |
4230 | s->flags |= SLAB_RECLAIM_ACCOUNT; |
4231 | return length; |
4232 | } |
4233 | SLAB_ATTR(reclaim_account); |
4234 | |
4235 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) |
4236 | { |
4237 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); |
4238 | } |
4239 | SLAB_ATTR_RO(hwcache_align); |
4240 | |
4241 | #ifdef CONFIG_ZONE_DMA |
4242 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) |
4243 | { |
4244 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); |
4245 | } |
4246 | SLAB_ATTR_RO(cache_dma); |
4247 | #endif |
4248 | |
4249 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) |
4250 | { |
4251 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); |
4252 | } |
4253 | SLAB_ATTR_RO(destroy_by_rcu); |
4254 | |
4255 | static ssize_t reserved_show(struct kmem_cache *s, char *buf) |
4256 | { |
4257 | return sprintf(buf, "%d\n", s->reserved); |
4258 | } |
4259 | SLAB_ATTR_RO(reserved); |
4260 | |
4261 | #ifdef CONFIG_SLUB_DEBUG |
4262 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) |
4263 | { |
4264 | return show_slab_objects(s, buf, SO_ALL); |
4265 | } |
4266 | SLAB_ATTR_RO(slabs); |
4267 | |
4268 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) |
4269 | { |
4270 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); |
4271 | } |
4272 | SLAB_ATTR_RO(total_objects); |
4273 | |
4274 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) |
4275 | { |
4276 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); |
4277 | } |
4278 | |
4279 | static ssize_t sanity_checks_store(struct kmem_cache *s, |
4280 | const char *buf, size_t length) |
4281 | { |
4282 | s->flags &= ~SLAB_DEBUG_FREE; |
4283 | if (buf[0] == '1') |
4284 | s->flags |= SLAB_DEBUG_FREE; |
4285 | return length; |
4286 | } |
4287 | SLAB_ATTR(sanity_checks); |
4288 | |
4289 | static ssize_t trace_show(struct kmem_cache *s, char *buf) |
4290 | { |
4291 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); |
4292 | } |
4293 | |
4294 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, |
4295 | size_t length) |
4296 | { |
4297 | s->flags &= ~SLAB_TRACE; |
4298 | if (buf[0] == '1') |
4299 | s->flags |= SLAB_TRACE; |
4300 | return length; |
4301 | } |
4302 | SLAB_ATTR(trace); |
4303 | |
4304 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) |
4305 | { |
4306 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); |
4307 | } |
4308 | |
4309 | static ssize_t red_zone_store(struct kmem_cache *s, |
4310 | const char *buf, size_t length) |
4311 | { |
4312 | if (any_slab_objects(s)) |
4313 | return -EBUSY; |
4314 | |
4315 | s->flags &= ~SLAB_RED_ZONE; |
4316 | if (buf[0] == '1') |
4317 | s->flags |= SLAB_RED_ZONE; |
4318 | calculate_sizes(s, -1); |
4319 | return length; |
4320 | } |
4321 | SLAB_ATTR(red_zone); |
4322 | |
4323 | static ssize_t poison_show(struct kmem_cache *s, char *buf) |
4324 | { |
4325 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); |
4326 | } |
4327 | |
4328 | static ssize_t poison_store(struct kmem_cache *s, |
4329 | const char *buf, size_t length) |
4330 | { |
4331 | if (any_slab_objects(s)) |
4332 | return -EBUSY; |
4333 | |
4334 | s->flags &= ~SLAB_POISON; |
4335 | if (buf[0] == '1') |
4336 | s->flags |= SLAB_POISON; |
4337 | calculate_sizes(s, -1); |
4338 | return length; |
4339 | } |
4340 | SLAB_ATTR(poison); |
4341 | |
4342 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) |
4343 | { |
4344 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); |
4345 | } |
4346 | |
4347 | static ssize_t store_user_store(struct kmem_cache *s, |
4348 | const char *buf, size_t length) |
4349 | { |
4350 | if (any_slab_objects(s)) |
4351 | return -EBUSY; |
4352 | |
4353 | s->flags &= ~SLAB_STORE_USER; |
4354 | if (buf[0] == '1') |
4355 | s->flags |= SLAB_STORE_USER; |
4356 | calculate_sizes(s, -1); |
4357 | return length; |
4358 | } |
4359 | SLAB_ATTR(store_user); |
4360 | |
4361 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
4362 | { |
4363 | return 0; |
4364 | } |
4365 | |
4366 | static ssize_t validate_store(struct kmem_cache *s, |
4367 | const char *buf, size_t length) |
4368 | { |
4369 | int ret = -EINVAL; |
4370 | |
4371 | if (buf[0] == '1') { |
4372 | ret = validate_slab_cache(s); |
4373 | if (ret >= 0) |
4374 | ret = length; |
4375 | } |
4376 | return ret; |
4377 | } |
4378 | SLAB_ATTR(validate); |
4379 | |
4380 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) |
4381 | { |
4382 | if (!(s->flags & SLAB_STORE_USER)) |
4383 | return -ENOSYS; |
4384 | return list_locations(s, buf, TRACK_ALLOC); |
4385 | } |
4386 | SLAB_ATTR_RO(alloc_calls); |
4387 | |
4388 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) |
4389 | { |
4390 | if (!(s->flags & SLAB_STORE_USER)) |
4391 | return -ENOSYS; |
4392 | return list_locations(s, buf, TRACK_FREE); |
4393 | } |
4394 | SLAB_ATTR_RO(free_calls); |
4395 | #endif /* CONFIG_SLUB_DEBUG */ |
4396 | |
4397 | #ifdef CONFIG_FAILSLAB |
4398 | static ssize_t failslab_show(struct kmem_cache *s, char *buf) |
4399 | { |
4400 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); |
4401 | } |
4402 | |
4403 | static ssize_t failslab_store(struct kmem_cache *s, const char *buf, |
4404 | size_t length) |
4405 | { |
4406 | s->flags &= ~SLAB_FAILSLAB; |
4407 | if (buf[0] == '1') |
4408 | s->flags |= SLAB_FAILSLAB; |
4409 | return length; |
4410 | } |
4411 | SLAB_ATTR(failslab); |
4412 | #endif |
4413 | |
4414 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
4415 | { |
4416 | return 0; |
4417 | } |
4418 | |
4419 | static ssize_t shrink_store(struct kmem_cache *s, |
4420 | const char *buf, size_t length) |
4421 | { |
4422 | if (buf[0] == '1') { |
4423 | int rc = kmem_cache_shrink(s); |
4424 | |
4425 | if (rc) |
4426 | return rc; |
4427 | } else |
4428 | return -EINVAL; |
4429 | return length; |
4430 | } |
4431 | SLAB_ATTR(shrink); |
4432 | |
4433 | #ifdef CONFIG_NUMA |
4434 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) |
4435 | { |
4436 | return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); |
4437 | } |
4438 | |
4439 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, |
4440 | const char *buf, size_t length) |
4441 | { |
4442 | unsigned long ratio; |
4443 | int err; |
4444 | |
4445 | err = strict_strtoul(buf, 10, &ratio); |
4446 | if (err) |
4447 | return err; |
4448 | |
4449 | if (ratio <= 100) |
4450 | s->remote_node_defrag_ratio = ratio * 10; |
4451 | |
4452 | return length; |
4453 | } |
4454 | SLAB_ATTR(remote_node_defrag_ratio); |
4455 | #endif |
4456 | |
4457 | #ifdef CONFIG_SLUB_STATS |
4458 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) |
4459 | { |
4460 | unsigned long sum = 0; |
4461 | int cpu; |
4462 | int len; |
4463 | int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); |
4464 | |
4465 | if (!data) |
4466 | return -ENOMEM; |
4467 | |
4468 | for_each_online_cpu(cpu) { |
4469 | unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; |
4470 | |
4471 | data[cpu] = x; |
4472 | sum += x; |
4473 | } |
4474 | |
4475 | len = sprintf(buf, "%lu", sum); |
4476 | |
4477 | #ifdef CONFIG_SMP |
4478 | for_each_online_cpu(cpu) { |
4479 | if (data[cpu] && len < PAGE_SIZE - 20) |
4480 | len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); |
4481 | } |
4482 | #endif |
4483 | kfree(data); |
4484 | return len + sprintf(buf + len, "\n"); |
4485 | } |
4486 | |
4487 | static void clear_stat(struct kmem_cache *s, enum stat_item si) |
4488 | { |
4489 | int cpu; |
4490 | |
4491 | for_each_online_cpu(cpu) |
4492 | per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; |
4493 | } |
4494 | |
4495 | #define STAT_ATTR(si, text) \ |
4496 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ |
4497 | { \ |
4498 | return show_stat(s, buf, si); \ |
4499 | } \ |
4500 | static ssize_t text##_store(struct kmem_cache *s, \ |
4501 | const char *buf, size_t length) \ |
4502 | { \ |
4503 | if (buf[0] != '0') \ |
4504 | return -EINVAL; \ |
4505 | clear_stat(s, si); \ |
4506 | return length; \ |
4507 | } \ |
4508 | SLAB_ATTR(text); \ |
4509 | |
4510 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); |
4511 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); |
4512 | STAT_ATTR(FREE_FASTPATH, free_fastpath); |
4513 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); |
4514 | STAT_ATTR(FREE_FROZEN, free_frozen); |
4515 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); |
4516 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); |
4517 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); |
4518 | STAT_ATTR(ALLOC_SLAB, alloc_slab); |
4519 | STAT_ATTR(ALLOC_REFILL, alloc_refill); |
4520 | STAT_ATTR(FREE_SLAB, free_slab); |
4521 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); |
4522 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); |
4523 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); |
4524 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); |
4525 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); |
4526 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); |
4527 | STAT_ATTR(ORDER_FALLBACK, order_fallback); |
4528 | #endif |
4529 | |
4530 | static struct attribute *slab_attrs[] = { |
4531 | &slab_size_attr.attr, |
4532 | &object_size_attr.attr, |
4533 | &objs_per_slab_attr.attr, |
4534 | &order_attr.attr, |
4535 | &min_partial_attr.attr, |
4536 | &objects_attr.attr, |
4537 | &objects_partial_attr.attr, |
4538 | &partial_attr.attr, |
4539 | &cpu_slabs_attr.attr, |
4540 | &ctor_attr.attr, |
4541 | &aliases_attr.attr, |
4542 | &align_attr.attr, |
4543 | &hwcache_align_attr.attr, |
4544 | &reclaim_account_attr.attr, |
4545 | &destroy_by_rcu_attr.attr, |
4546 | &shrink_attr.attr, |
4547 | &reserved_attr.attr, |
4548 | #ifdef CONFIG_SLUB_DEBUG |
4549 | &total_objects_attr.attr, |
4550 | &slabs_attr.attr, |
4551 | &sanity_checks_attr.attr, |
4552 | &trace_attr.attr, |
4553 | &red_zone_attr.attr, |
4554 | &poison_attr.attr, |
4555 | &store_user_attr.attr, |
4556 | &validate_attr.attr, |
4557 | &alloc_calls_attr.attr, |
4558 | &free_calls_attr.attr, |
4559 | #endif |
4560 | #ifdef CONFIG_ZONE_DMA |
4561 | &cache_dma_attr.attr, |
4562 | #endif |
4563 | #ifdef CONFIG_NUMA |
4564 | &remote_node_defrag_ratio_attr.attr, |
4565 | #endif |
4566 | #ifdef CONFIG_SLUB_STATS |
4567 | &alloc_fastpath_attr.attr, |
4568 | &alloc_slowpath_attr.attr, |
4569 | &free_fastpath_attr.attr, |
4570 | &free_slowpath_attr.attr, |
4571 | &free_frozen_attr.attr, |
4572 | &free_add_partial_attr.attr, |
4573 | &free_remove_partial_attr.attr, |
4574 | &alloc_from_partial_attr.attr, |
4575 | &alloc_slab_attr.attr, |
4576 | &alloc_refill_attr.attr, |
4577 | &free_slab_attr.attr, |
4578 | &cpuslab_flush_attr.attr, |
4579 | &deactivate_full_attr.attr, |
4580 | &deactivate_empty_attr.attr, |
4581 | &deactivate_to_head_attr.attr, |
4582 | &deactivate_to_tail_attr.attr, |
4583 | &deactivate_remote_frees_attr.attr, |
4584 | &order_fallback_attr.attr, |
4585 | #endif |
4586 | #ifdef CONFIG_FAILSLAB |
4587 | &failslab_attr.attr, |
4588 | #endif |
4589 | |
4590 | NULL |
4591 | }; |
4592 | |
4593 | static struct attribute_group slab_attr_group = { |
4594 | .attrs = slab_attrs, |
4595 | }; |
4596 | |
4597 | static ssize_t slab_attr_show(struct kobject *kobj, |
4598 | struct attribute *attr, |
4599 | char *buf) |
4600 | { |
4601 | struct slab_attribute *attribute; |
4602 | struct kmem_cache *s; |
4603 | int err; |
4604 | |
4605 | attribute = to_slab_attr(attr); |
4606 | s = to_slab(kobj); |
4607 | |
4608 | if (!attribute->show) |
4609 | return -EIO; |
4610 | |
4611 | err = attribute->show(s, buf); |
4612 | |
4613 | return err; |
4614 | } |
4615 | |
4616 | static ssize_t slab_attr_store(struct kobject *kobj, |
4617 | struct attribute *attr, |
4618 | const char *buf, size_t len) |
4619 | { |
4620 | struct slab_attribute *attribute; |
4621 | struct kmem_cache *s; |
4622 | int err; |
4623 | |
4624 | attribute = to_slab_attr(attr); |
4625 | s = to_slab(kobj); |
4626 | |
4627 | if (!attribute->store) |
4628 | return -EIO; |
4629 | |
4630 | err = attribute->store(s, buf, len); |
4631 | |
4632 | return err; |
4633 | } |
4634 | |
4635 | static void kmem_cache_release(struct kobject *kobj) |
4636 | { |
4637 | struct kmem_cache *s = to_slab(kobj); |
4638 | |
4639 | kfree(s->name); |
4640 | kfree(s); |
4641 | } |
4642 | |
4643 | static const struct sysfs_ops slab_sysfs_ops = { |
4644 | .show = slab_attr_show, |
4645 | .store = slab_attr_store, |
4646 | }; |
4647 | |
4648 | static struct kobj_type slab_ktype = { |
4649 | .sysfs_ops = &slab_sysfs_ops, |
4650 | .release = kmem_cache_release |
4651 | }; |
4652 | |
4653 | static int uevent_filter(struct kset *kset, struct kobject *kobj) |
4654 | { |
4655 | struct kobj_type *ktype = get_ktype(kobj); |
4656 | |
4657 | if (ktype == &slab_ktype) |
4658 | return 1; |
4659 | return 0; |
4660 | } |
4661 | |
4662 | static const struct kset_uevent_ops slab_uevent_ops = { |
4663 | .filter = uevent_filter, |
4664 | }; |
4665 | |
4666 | static struct kset *slab_kset; |
4667 | |
4668 | #define ID_STR_LENGTH 64 |
4669 | |
4670 | /* Create a unique string id for a slab cache: |
4671 | * |
4672 | * Format :[flags-]size |
4673 | */ |
4674 | static char *create_unique_id(struct kmem_cache *s) |
4675 | { |
4676 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); |
4677 | char *p = name; |
4678 | |
4679 | BUG_ON(!name); |
4680 | |
4681 | *p++ = ':'; |
4682 | /* |
4683 | * First flags affecting slabcache operations. We will only |
4684 | * get here for aliasable slabs so we do not need to support |
4685 | * too many flags. The flags here must cover all flags that |
4686 | * are matched during merging to guarantee that the id is |
4687 | * unique. |
4688 | */ |
4689 | if (s->flags & SLAB_CACHE_DMA) |
4690 | *p++ = 'd'; |
4691 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
4692 | *p++ = 'a'; |
4693 | if (s->flags & SLAB_DEBUG_FREE) |
4694 | *p++ = 'F'; |
4695 | if (!(s->flags & SLAB_NOTRACK)) |
4696 | *p++ = 't'; |
4697 | if (p != name + 1) |
4698 | *p++ = '-'; |
4699 | p += sprintf(p, "%07d", s->size); |
4700 | BUG_ON(p > name + ID_STR_LENGTH - 1); |
4701 | return name; |
4702 | } |
4703 | |
4704 | static int sysfs_slab_add(struct kmem_cache *s) |
4705 | { |
4706 | int err; |
4707 | const char *name; |
4708 | int unmergeable; |
4709 | |
4710 | if (slab_state < SYSFS) |
4711 | /* Defer until later */ |
4712 | return 0; |
4713 | |
4714 | unmergeable = slab_unmergeable(s); |
4715 | if (unmergeable) { |
4716 | /* |
4717 | * Slabcache can never be merged so we can use the name proper. |
4718 | * This is typically the case for debug situations. In that |
4719 | * case we can catch duplicate names easily. |
4720 | */ |
4721 | sysfs_remove_link(&slab_kset->kobj, s->name); |
4722 | name = s->name; |
4723 | } else { |
4724 | /* |
4725 | * Create a unique name for the slab as a target |
4726 | * for the symlinks. |
4727 | */ |
4728 | name = create_unique_id(s); |
4729 | } |
4730 | |
4731 | s->kobj.kset = slab_kset; |
4732 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); |
4733 | if (err) { |
4734 | kobject_put(&s->kobj); |
4735 | return err; |
4736 | } |
4737 | |
4738 | err = sysfs_create_group(&s->kobj, &slab_attr_group); |
4739 | if (err) { |
4740 | kobject_del(&s->kobj); |
4741 | kobject_put(&s->kobj); |
4742 | return err; |
4743 | } |
4744 | kobject_uevent(&s->kobj, KOBJ_ADD); |
4745 | if (!unmergeable) { |
4746 | /* Setup first alias */ |
4747 | sysfs_slab_alias(s, s->name); |
4748 | kfree(name); |
4749 | } |
4750 | return 0; |
4751 | } |
4752 | |
4753 | static void sysfs_slab_remove(struct kmem_cache *s) |
4754 | { |
4755 | if (slab_state < SYSFS) |
4756 | /* |
4757 | * Sysfs has not been setup yet so no need to remove the |
4758 | * cache from sysfs. |
4759 | */ |
4760 | return; |
4761 | |
4762 | kobject_uevent(&s->kobj, KOBJ_REMOVE); |
4763 | kobject_del(&s->kobj); |
4764 | kobject_put(&s->kobj); |
4765 | } |
4766 | |
4767 | /* |
4768 | * Need to buffer aliases during bootup until sysfs becomes |
4769 | * available lest we lose that information. |
4770 | */ |
4771 | struct saved_alias { |
4772 | struct kmem_cache *s; |
4773 | const char *name; |
4774 | struct saved_alias *next; |
4775 | }; |
4776 | |
4777 | static struct saved_alias *alias_list; |
4778 | |
4779 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) |
4780 | { |
4781 | struct saved_alias *al; |
4782 | |
4783 | if (slab_state == SYSFS) { |
4784 | /* |
4785 | * If we have a leftover link then remove it. |
4786 | */ |
4787 | sysfs_remove_link(&slab_kset->kobj, name); |
4788 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); |
4789 | } |
4790 | |
4791 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); |
4792 | if (!al) |
4793 | return -ENOMEM; |
4794 | |
4795 | al->s = s; |
4796 | al->name = name; |
4797 | al->next = alias_list; |
4798 | alias_list = al; |
4799 | return 0; |
4800 | } |
4801 | |
4802 | static int __init slab_sysfs_init(void) |
4803 | { |
4804 | struct kmem_cache *s; |
4805 | int err; |
4806 | |
4807 | down_write(&slub_lock); |
4808 | |
4809 | slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); |
4810 | if (!slab_kset) { |
4811 | up_write(&slub_lock); |
4812 | printk(KERN_ERR "Cannot register slab subsystem.\n"); |
4813 | return -ENOSYS; |
4814 | } |
4815 | |
4816 | slab_state = SYSFS; |
4817 | |
4818 | list_for_each_entry(s, &slab_caches, list) { |
4819 | err = sysfs_slab_add(s); |
4820 | if (err) |
4821 | printk(KERN_ERR "SLUB: Unable to add boot slab %s" |
4822 | " to sysfs\n", s->name); |
4823 | } |
4824 | |
4825 | while (alias_list) { |
4826 | struct saved_alias *al = alias_list; |
4827 | |
4828 | alias_list = alias_list->next; |
4829 | err = sysfs_slab_alias(al->s, al->name); |
4830 | if (err) |
4831 | printk(KERN_ERR "SLUB: Unable to add boot slab alias" |
4832 | " %s to sysfs\n", s->name); |
4833 | kfree(al); |
4834 | } |
4835 | |
4836 | up_write(&slub_lock); |
4837 | resiliency_test(); |
4838 | return 0; |
4839 | } |
4840 | |
4841 | __initcall(slab_sysfs_init); |
4842 | #endif /* CONFIG_SYSFS */ |
4843 | |
4844 | /* |
4845 | * The /proc/slabinfo ABI |
4846 | */ |
4847 | #ifdef CONFIG_SLABINFO |
4848 | static void print_slabinfo_header(struct seq_file *m) |
4849 | { |
4850 | seq_puts(m, "slabinfo - version: 2.1\n"); |
4851 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4852 | "<objperslab> <pagesperslab>"); |
4853 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); |
4854 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); |
4855 | seq_putc(m, '\n'); |
4856 | } |
4857 | |
4858 | static void *s_start(struct seq_file *m, loff_t *pos) |
4859 | { |
4860 | loff_t n = *pos; |
4861 | |
4862 | down_read(&slub_lock); |
4863 | if (!n) |
4864 | print_slabinfo_header(m); |
4865 | |
4866 | return seq_list_start(&slab_caches, *pos); |
4867 | } |
4868 | |
4869 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) |
4870 | { |
4871 | return seq_list_next(p, &slab_caches, pos); |
4872 | } |
4873 | |
4874 | static void s_stop(struct seq_file *m, void *p) |
4875 | { |
4876 | up_read(&slub_lock); |
4877 | } |
4878 | |
4879 | static int s_show(struct seq_file *m, void *p) |
4880 | { |
4881 | unsigned long nr_partials = 0; |
4882 | unsigned long nr_slabs = 0; |
4883 | unsigned long nr_inuse = 0; |
4884 | unsigned long nr_objs = 0; |
4885 | unsigned long nr_free = 0; |
4886 | struct kmem_cache *s; |
4887 | int node; |
4888 | |
4889 | s = list_entry(p, struct kmem_cache, list); |
4890 | |
4891 | for_each_online_node(node) { |
4892 | struct kmem_cache_node *n = get_node(s, node); |
4893 | |
4894 | if (!n) |
4895 | continue; |
4896 | |
4897 | nr_partials += n->nr_partial; |
4898 | nr_slabs += atomic_long_read(&n->nr_slabs); |
4899 | nr_objs += atomic_long_read(&n->total_objects); |
4900 | nr_free += count_partial(n, count_free); |
4901 | } |
4902 | |
4903 | nr_inuse = nr_objs - nr_free; |
4904 | |
4905 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, |
4906 | nr_objs, s->size, oo_objects(s->oo), |
4907 | (1 << oo_order(s->oo))); |
4908 | seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); |
4909 | seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, |
4910 | 0UL); |
4911 | seq_putc(m, '\n'); |
4912 | return 0; |
4913 | } |
4914 | |
4915 | static const struct seq_operations slabinfo_op = { |
4916 | .start = s_start, |
4917 | .next = s_next, |
4918 | .stop = s_stop, |
4919 | .show = s_show, |
4920 | }; |
4921 | |
4922 | static int slabinfo_open(struct inode *inode, struct file *file) |
4923 | { |
4924 | return seq_open(file, &slabinfo_op); |
4925 | } |
4926 | |
4927 | static const struct file_operations proc_slabinfo_operations = { |
4928 | .open = slabinfo_open, |
4929 | .read = seq_read, |
4930 | .llseek = seq_lseek, |
4931 | .release = seq_release, |
4932 | }; |
4933 | |
4934 | static int __init slab_proc_init(void) |
4935 | { |
4936 | proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations); |
4937 | return 0; |
4938 | } |
4939 | module_init(slab_proc_init); |
4940 | #endif /* CONFIG_SLABINFO */ |
4941 |
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