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