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