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