]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0 | |
2 | /* | |
3 | * SLUB: A slab allocator that limits cache line use instead of queuing | |
4 | * objects in per cpu and per node lists. | |
5 | * | |
6 | * The allocator synchronizes using per slab locks or atomic operatios | |
7 | * and only uses a centralized lock to manage a pool of partial slabs. | |
8 | * | |
9 | * (C) 2007 SGI, Christoph Lameter | |
10 | * (C) 2011 Linux Foundation, Christoph Lameter | |
11 | */ | |
12 | ||
13 | #include <linux/mm.h> | |
14 | #include <linux/swap.h> /* struct reclaim_state */ | |
15 | #include <linux/module.h> | |
16 | #include <linux/bit_spinlock.h> | |
17 | #include <linux/interrupt.h> | |
18 | #include <linux/bitops.h> | |
19 | #include <linux/slab.h> | |
20 | #include "slab.h" | |
21 | #include <linux/proc_fs.h> | |
22 | #include <linux/notifier.h> | |
23 | #include <linux/seq_file.h> | |
24 | #include <linux/kasan.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 | #include <linux/random.h> | |
38 | ||
39 | #include <trace/events/kmem.h> | |
40 | ||
41 | #include "internal.h" | |
42 | ||
43 | /* | |
44 | * Lock order: | |
45 | * 1. slab_mutex (Global Mutex) | |
46 | * 2. node->list_lock | |
47 | * 3. slab_lock(page) (Only on some arches and for debugging) | |
48 | * | |
49 | * slab_mutex | |
50 | * | |
51 | * The role of the slab_mutex is to protect the list of all the slabs | |
52 | * and to synchronize major metadata changes to slab cache structures. | |
53 | * | |
54 | * The slab_lock is only used for debugging and on arches that do not | |
55 | * have the ability to do a cmpxchg_double. It only protects the second | |
56 | * double word in the page struct. Meaning | |
57 | * A. page->freelist -> List of object free in a page | |
58 | * B. page->counters -> Counters of objects | |
59 | * C. page->frozen -> frozen state | |
60 | * | |
61 | * If a slab is frozen then it is exempt from list management. It is not | |
62 | * on any list. The processor that froze the slab is the one who can | |
63 | * perform list operations on the page. Other processors may put objects | |
64 | * onto the freelist but the processor that froze the slab is the only | |
65 | * one that can retrieve the objects from the page's freelist. | |
66 | * | |
67 | * The list_lock protects the partial and full list on each node and | |
68 | * the partial slab counter. If taken then no new slabs may be added or | |
69 | * removed from the lists nor make the number of partial slabs be modified. | |
70 | * (Note that the total number of slabs is an atomic value that may be | |
71 | * modified without taking the list lock). | |
72 | * | |
73 | * The list_lock is a centralized lock and thus we avoid taking it as | |
74 | * much as possible. As long as SLUB does not have to handle partial | |
75 | * slabs, operations can continue without any centralized lock. F.e. | |
76 | * allocating a long series of objects that fill up slabs does not require | |
77 | * the list lock. | |
78 | * Interrupts are disabled during allocation and deallocation in order to | |
79 | * make the slab allocator safe to use in the context of an irq. In addition | |
80 | * interrupts are disabled to ensure that the processor does not change | |
81 | * while handling per_cpu slabs, due to kernel preemption. | |
82 | * | |
83 | * SLUB assigns one slab for allocation to each processor. | |
84 | * Allocations only occur from these slabs called cpu slabs. | |
85 | * | |
86 | * Slabs with free elements are kept on a partial list and during regular | |
87 | * operations no list for full slabs is used. If an object in a full slab is | |
88 | * freed then the slab will show up again on the partial lists. | |
89 | * We track full slabs for debugging purposes though because otherwise we | |
90 | * cannot scan all objects. | |
91 | * | |
92 | * Slabs are freed when they become empty. Teardown and setup is | |
93 | * minimal so we rely on the page allocators per cpu caches for | |
94 | * fast frees and allocs. | |
95 | * | |
96 | * Overloading of page flags that are otherwise used for LRU management. | |
97 | * | |
98 | * PageActive The slab is frozen and exempt from list processing. | |
99 | * This means that the slab is dedicated to a purpose | |
100 | * such as satisfying allocations for a specific | |
101 | * processor. Objects may be freed in the slab while | |
102 | * it is frozen but slab_free will then skip the usual | |
103 | * list operations. It is up to the processor holding | |
104 | * the slab to integrate the slab into the slab lists | |
105 | * when the slab is no longer needed. | |
106 | * | |
107 | * One use of this flag is to mark slabs that are | |
108 | * used for allocations. Then such a slab becomes a cpu | |
109 | * slab. The cpu slab may be equipped with an additional | |
110 | * freelist that allows lockless access to | |
111 | * free objects in addition to the regular freelist | |
112 | * that requires the slab lock. | |
113 | * | |
114 | * PageError Slab requires special handling due to debug | |
115 | * options set. This moves slab handling out of | |
116 | * the fast path and disables lockless freelists. | |
117 | */ | |
118 | ||
119 | static inline int kmem_cache_debug(struct kmem_cache *s) | |
120 | { | |
121 | #ifdef CONFIG_SLUB_DEBUG | |
122 | return unlikely(s->flags & SLAB_DEBUG_FLAGS); | |
123 | #else | |
124 | return 0; | |
125 | #endif | |
126 | } | |
127 | ||
128 | void *fixup_red_left(struct kmem_cache *s, void *p) | |
129 | { | |
130 | if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) | |
131 | p += s->red_left_pad; | |
132 | ||
133 | return p; | |
134 | } | |
135 | ||
136 | static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) | |
137 | { | |
138 | #ifdef CONFIG_SLUB_CPU_PARTIAL | |
139 | return !kmem_cache_debug(s); | |
140 | #else | |
141 | return false; | |
142 | #endif | |
143 | } | |
144 | ||
145 | /* | |
146 | * Issues still to be resolved: | |
147 | * | |
148 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. | |
149 | * | |
150 | * - Variable sizing of the per node arrays | |
151 | */ | |
152 | ||
153 | /* Enable to test recovery from slab corruption on boot */ | |
154 | #undef SLUB_RESILIENCY_TEST | |
155 | ||
156 | /* Enable to log cmpxchg failures */ | |
157 | #undef SLUB_DEBUG_CMPXCHG | |
158 | ||
159 | /* | |
160 | * Mininum number of partial slabs. These will be left on the partial | |
161 | * lists even if they are empty. kmem_cache_shrink may reclaim them. | |
162 | */ | |
163 | #define MIN_PARTIAL 5 | |
164 | ||
165 | /* | |
166 | * Maximum number of desirable partial slabs. | |
167 | * The existence of more partial slabs makes kmem_cache_shrink | |
168 | * sort the partial list by the number of objects in use. | |
169 | */ | |
170 | #define MAX_PARTIAL 10 | |
171 | ||
172 | #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \ | |
173 | SLAB_POISON | SLAB_STORE_USER) | |
174 | ||
175 | /* | |
176 | * These debug flags cannot use CMPXCHG because there might be consistency | |
177 | * issues when checking or reading debug information | |
178 | */ | |
179 | #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \ | |
180 | SLAB_TRACE) | |
181 | ||
182 | ||
183 | /* | |
184 | * Debugging flags that require metadata to be stored in the slab. These get | |
185 | * disabled when slub_debug=O is used and a cache's min order increases with | |
186 | * metadata. | |
187 | */ | |
188 | #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) | |
189 | ||
190 | #define OO_SHIFT 16 | |
191 | #define OO_MASK ((1 << OO_SHIFT) - 1) | |
192 | #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */ | |
193 | ||
194 | /* Internal SLUB flags */ | |
195 | /* Poison object */ | |
196 | #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U) | |
197 | /* Use cmpxchg_double */ | |
198 | #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U) | |
199 | ||
200 | /* | |
201 | * Tracking user of a slab. | |
202 | */ | |
203 | #define TRACK_ADDRS_COUNT 16 | |
204 | struct track { | |
205 | unsigned long addr; /* Called from address */ | |
206 | #ifdef CONFIG_STACKTRACE | |
207 | unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ | |
208 | #endif | |
209 | int cpu; /* Was running on cpu */ | |
210 | int pid; /* Pid context */ | |
211 | unsigned long when; /* When did the operation occur */ | |
212 | }; | |
213 | ||
214 | enum track_item { TRACK_ALLOC, TRACK_FREE }; | |
215 | ||
216 | #ifdef CONFIG_SYSFS | |
217 | static int sysfs_slab_add(struct kmem_cache *); | |
218 | static int sysfs_slab_alias(struct kmem_cache *, const char *); | |
219 | static void memcg_propagate_slab_attrs(struct kmem_cache *s); | |
220 | static void sysfs_slab_remove(struct kmem_cache *s); | |
221 | #else | |
222 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } | |
223 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) | |
224 | { return 0; } | |
225 | static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { } | |
226 | static inline void sysfs_slab_remove(struct kmem_cache *s) { } | |
227 | #endif | |
228 | ||
229 | static inline void stat(const struct kmem_cache *s, enum stat_item si) | |
230 | { | |
231 | #ifdef CONFIG_SLUB_STATS | |
232 | /* | |
233 | * The rmw is racy on a preemptible kernel but this is acceptable, so | |
234 | * avoid this_cpu_add()'s irq-disable overhead. | |
235 | */ | |
236 | raw_cpu_inc(s->cpu_slab->stat[si]); | |
237 | #endif | |
238 | } | |
239 | ||
240 | /******************************************************************** | |
241 | * Core slab cache functions | |
242 | *******************************************************************/ | |
243 | ||
244 | /* | |
245 | * Returns freelist pointer (ptr). With hardening, this is obfuscated | |
246 | * with an XOR of the address where the pointer is held and a per-cache | |
247 | * random number. | |
248 | */ | |
249 | static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr, | |
250 | unsigned long ptr_addr) | |
251 | { | |
252 | #ifdef CONFIG_SLAB_FREELIST_HARDENED | |
253 | return (void *)((unsigned long)ptr ^ s->random ^ ptr_addr); | |
254 | #else | |
255 | return ptr; | |
256 | #endif | |
257 | } | |
258 | ||
259 | /* Returns the freelist pointer recorded at location ptr_addr. */ | |
260 | static inline void *freelist_dereference(const struct kmem_cache *s, | |
261 | void *ptr_addr) | |
262 | { | |
263 | return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr), | |
264 | (unsigned long)ptr_addr); | |
265 | } | |
266 | ||
267 | static inline void *get_freepointer(struct kmem_cache *s, void *object) | |
268 | { | |
269 | return freelist_dereference(s, object + s->offset); | |
270 | } | |
271 | ||
272 | static void prefetch_freepointer(const struct kmem_cache *s, void *object) | |
273 | { | |
274 | if (object) | |
275 | prefetch(freelist_dereference(s, object + s->offset)); | |
276 | } | |
277 | ||
278 | static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) | |
279 | { | |
280 | unsigned long freepointer_addr; | |
281 | void *p; | |
282 | ||
283 | if (!debug_pagealloc_enabled()) | |
284 | return get_freepointer(s, object); | |
285 | ||
286 | freepointer_addr = (unsigned long)object + s->offset; | |
287 | probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p)); | |
288 | return freelist_ptr(s, p, freepointer_addr); | |
289 | } | |
290 | ||
291 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
292 | { | |
293 | unsigned long freeptr_addr = (unsigned long)object + s->offset; | |
294 | ||
295 | #ifdef CONFIG_SLAB_FREELIST_HARDENED | |
296 | BUG_ON(object == fp); /* naive detection of double free or corruption */ | |
297 | #endif | |
298 | ||
299 | *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr); | |
300 | } | |
301 | ||
302 | /* Loop over all objects in a slab */ | |
303 | #define for_each_object(__p, __s, __addr, __objects) \ | |
304 | for (__p = fixup_red_left(__s, __addr); \ | |
305 | __p < (__addr) + (__objects) * (__s)->size; \ | |
306 | __p += (__s)->size) | |
307 | ||
308 | #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \ | |
309 | for (__p = fixup_red_left(__s, __addr), __idx = 1; \ | |
310 | __idx <= __objects; \ | |
311 | __p += (__s)->size, __idx++) | |
312 | ||
313 | /* Determine object index from a given position */ | |
314 | static inline int slab_index(void *p, struct kmem_cache *s, void *addr) | |
315 | { | |
316 | return (p - addr) / s->size; | |
317 | } | |
318 | ||
319 | static inline int order_objects(int order, unsigned long size, int reserved) | |
320 | { | |
321 | return ((PAGE_SIZE << order) - reserved) / size; | |
322 | } | |
323 | ||
324 | static inline struct kmem_cache_order_objects oo_make(int order, | |
325 | unsigned long size, int reserved) | |
326 | { | |
327 | struct kmem_cache_order_objects x = { | |
328 | (order << OO_SHIFT) + order_objects(order, size, reserved) | |
329 | }; | |
330 | ||
331 | return x; | |
332 | } | |
333 | ||
334 | static inline int oo_order(struct kmem_cache_order_objects x) | |
335 | { | |
336 | return x.x >> OO_SHIFT; | |
337 | } | |
338 | ||
339 | static inline int oo_objects(struct kmem_cache_order_objects x) | |
340 | { | |
341 | return x.x & OO_MASK; | |
342 | } | |
343 | ||
344 | /* | |
345 | * Per slab locking using the pagelock | |
346 | */ | |
347 | static __always_inline void slab_lock(struct page *page) | |
348 | { | |
349 | VM_BUG_ON_PAGE(PageTail(page), page); | |
350 | bit_spin_lock(PG_locked, &page->flags); | |
351 | } | |
352 | ||
353 | static __always_inline void slab_unlock(struct page *page) | |
354 | { | |
355 | VM_BUG_ON_PAGE(PageTail(page), page); | |
356 | __bit_spin_unlock(PG_locked, &page->flags); | |
357 | } | |
358 | ||
359 | static inline void set_page_slub_counters(struct page *page, unsigned long counters_new) | |
360 | { | |
361 | struct page tmp; | |
362 | tmp.counters = counters_new; | |
363 | /* | |
364 | * page->counters can cover frozen/inuse/objects as well | |
365 | * as page->_refcount. If we assign to ->counters directly | |
366 | * we run the risk of losing updates to page->_refcount, so | |
367 | * be careful and only assign to the fields we need. | |
368 | */ | |
369 | page->frozen = tmp.frozen; | |
370 | page->inuse = tmp.inuse; | |
371 | page->objects = tmp.objects; | |
372 | } | |
373 | ||
374 | /* Interrupts must be disabled (for the fallback code to work right) */ | |
375 | static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, | |
376 | void *freelist_old, unsigned long counters_old, | |
377 | void *freelist_new, unsigned long counters_new, | |
378 | const char *n) | |
379 | { | |
380 | VM_BUG_ON(!irqs_disabled()); | |
381 | #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ | |
382 | defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) | |
383 | if (s->flags & __CMPXCHG_DOUBLE) { | |
384 | if (cmpxchg_double(&page->freelist, &page->counters, | |
385 | freelist_old, counters_old, | |
386 | freelist_new, counters_new)) | |
387 | return true; | |
388 | } else | |
389 | #endif | |
390 | { | |
391 | slab_lock(page); | |
392 | if (page->freelist == freelist_old && | |
393 | page->counters == counters_old) { | |
394 | page->freelist = freelist_new; | |
395 | set_page_slub_counters(page, counters_new); | |
396 | slab_unlock(page); | |
397 | return true; | |
398 | } | |
399 | slab_unlock(page); | |
400 | } | |
401 | ||
402 | cpu_relax(); | |
403 | stat(s, CMPXCHG_DOUBLE_FAIL); | |
404 | ||
405 | #ifdef SLUB_DEBUG_CMPXCHG | |
406 | pr_info("%s %s: cmpxchg double redo ", n, s->name); | |
407 | #endif | |
408 | ||
409 | return false; | |
410 | } | |
411 | ||
412 | static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page, | |
413 | void *freelist_old, unsigned long counters_old, | |
414 | void *freelist_new, unsigned long counters_new, | |
415 | const char *n) | |
416 | { | |
417 | #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ | |
418 | defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) | |
419 | if (s->flags & __CMPXCHG_DOUBLE) { | |
420 | if (cmpxchg_double(&page->freelist, &page->counters, | |
421 | freelist_old, counters_old, | |
422 | freelist_new, counters_new)) | |
423 | return true; | |
424 | } else | |
425 | #endif | |
426 | { | |
427 | unsigned long flags; | |
428 | ||
429 | local_irq_save(flags); | |
430 | slab_lock(page); | |
431 | if (page->freelist == freelist_old && | |
432 | page->counters == counters_old) { | |
433 | page->freelist = freelist_new; | |
434 | set_page_slub_counters(page, counters_new); | |
435 | slab_unlock(page); | |
436 | local_irq_restore(flags); | |
437 | return true; | |
438 | } | |
439 | slab_unlock(page); | |
440 | local_irq_restore(flags); | |
441 | } | |
442 | ||
443 | cpu_relax(); | |
444 | stat(s, CMPXCHG_DOUBLE_FAIL); | |
445 | ||
446 | #ifdef SLUB_DEBUG_CMPXCHG | |
447 | pr_info("%s %s: cmpxchg double redo ", n, s->name); | |
448 | #endif | |
449 | ||
450 | return false; | |
451 | } | |
452 | ||
453 | #ifdef CONFIG_SLUB_DEBUG | |
454 | /* | |
455 | * Determine a map of object in use on a page. | |
456 | * | |
457 | * Node listlock must be held to guarantee that the page does | |
458 | * not vanish from under us. | |
459 | */ | |
460 | static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) | |
461 | { | |
462 | void *p; | |
463 | void *addr = page_address(page); | |
464 | ||
465 | for (p = page->freelist; p; p = get_freepointer(s, p)) | |
466 | set_bit(slab_index(p, s, addr), map); | |
467 | } | |
468 | ||
469 | static inline int size_from_object(struct kmem_cache *s) | |
470 | { | |
471 | if (s->flags & SLAB_RED_ZONE) | |
472 | return s->size - s->red_left_pad; | |
473 | ||
474 | return s->size; | |
475 | } | |
476 | ||
477 | static inline void *restore_red_left(struct kmem_cache *s, void *p) | |
478 | { | |
479 | if (s->flags & SLAB_RED_ZONE) | |
480 | p -= s->red_left_pad; | |
481 | ||
482 | return p; | |
483 | } | |
484 | ||
485 | /* | |
486 | * Debug settings: | |
487 | */ | |
488 | #if defined(CONFIG_SLUB_DEBUG_ON) | |
489 | static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS; | |
490 | #else | |
491 | static slab_flags_t slub_debug; | |
492 | #endif | |
493 | ||
494 | static char *slub_debug_slabs; | |
495 | static int disable_higher_order_debug; | |
496 | ||
497 | /* | |
498 | * slub is about to manipulate internal object metadata. This memory lies | |
499 | * outside the range of the allocated object, so accessing it would normally | |
500 | * be reported by kasan as a bounds error. metadata_access_enable() is used | |
501 | * to tell kasan that these accesses are OK. | |
502 | */ | |
503 | static inline void metadata_access_enable(void) | |
504 | { | |
505 | kasan_disable_current(); | |
506 | } | |
507 | ||
508 | static inline void metadata_access_disable(void) | |
509 | { | |
510 | kasan_enable_current(); | |
511 | } | |
512 | ||
513 | /* | |
514 | * Object debugging | |
515 | */ | |
516 | ||
517 | /* Verify that a pointer has an address that is valid within a slab page */ | |
518 | static inline int check_valid_pointer(struct kmem_cache *s, | |
519 | struct page *page, void *object) | |
520 | { | |
521 | void *base; | |
522 | ||
523 | if (!object) | |
524 | return 1; | |
525 | ||
526 | base = page_address(page); | |
527 | object = restore_red_left(s, object); | |
528 | if (object < base || object >= base + page->objects * s->size || | |
529 | (object - base) % s->size) { | |
530 | return 0; | |
531 | } | |
532 | ||
533 | return 1; | |
534 | } | |
535 | ||
536 | static void print_section(char *level, char *text, u8 *addr, | |
537 | unsigned int length) | |
538 | { | |
539 | metadata_access_enable(); | |
540 | print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr, | |
541 | length, 1); | |
542 | metadata_access_disable(); | |
543 | } | |
544 | ||
545 | static struct track *get_track(struct kmem_cache *s, void *object, | |
546 | enum track_item alloc) | |
547 | { | |
548 | struct track *p; | |
549 | ||
550 | if (s->offset) | |
551 | p = object + s->offset + sizeof(void *); | |
552 | else | |
553 | p = object + s->inuse; | |
554 | ||
555 | return p + alloc; | |
556 | } | |
557 | ||
558 | static void set_track(struct kmem_cache *s, void *object, | |
559 | enum track_item alloc, unsigned long addr) | |
560 | { | |
561 | struct track *p = get_track(s, object, alloc); | |
562 | ||
563 | if (addr) { | |
564 | #ifdef CONFIG_STACKTRACE | |
565 | struct stack_trace trace; | |
566 | int i; | |
567 | ||
568 | trace.nr_entries = 0; | |
569 | trace.max_entries = TRACK_ADDRS_COUNT; | |
570 | trace.entries = p->addrs; | |
571 | trace.skip = 3; | |
572 | metadata_access_enable(); | |
573 | save_stack_trace(&trace); | |
574 | metadata_access_disable(); | |
575 | ||
576 | /* See rant in lockdep.c */ | |
577 | if (trace.nr_entries != 0 && | |
578 | trace.entries[trace.nr_entries - 1] == ULONG_MAX) | |
579 | trace.nr_entries--; | |
580 | ||
581 | for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++) | |
582 | p->addrs[i] = 0; | |
583 | #endif | |
584 | p->addr = addr; | |
585 | p->cpu = smp_processor_id(); | |
586 | p->pid = current->pid; | |
587 | p->when = jiffies; | |
588 | } else | |
589 | memset(p, 0, sizeof(struct track)); | |
590 | } | |
591 | ||
592 | static void init_tracking(struct kmem_cache *s, void *object) | |
593 | { | |
594 | if (!(s->flags & SLAB_STORE_USER)) | |
595 | return; | |
596 | ||
597 | set_track(s, object, TRACK_FREE, 0UL); | |
598 | set_track(s, object, TRACK_ALLOC, 0UL); | |
599 | } | |
600 | ||
601 | static void print_track(const char *s, struct track *t) | |
602 | { | |
603 | if (!t->addr) | |
604 | return; | |
605 | ||
606 | pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n", | |
607 | s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); | |
608 | #ifdef CONFIG_STACKTRACE | |
609 | { | |
610 | int i; | |
611 | for (i = 0; i < TRACK_ADDRS_COUNT; i++) | |
612 | if (t->addrs[i]) | |
613 | pr_err("\t%pS\n", (void *)t->addrs[i]); | |
614 | else | |
615 | break; | |
616 | } | |
617 | #endif | |
618 | } | |
619 | ||
620 | static void print_tracking(struct kmem_cache *s, void *object) | |
621 | { | |
622 | if (!(s->flags & SLAB_STORE_USER)) | |
623 | return; | |
624 | ||
625 | print_track("Allocated", get_track(s, object, TRACK_ALLOC)); | |
626 | print_track("Freed", get_track(s, object, TRACK_FREE)); | |
627 | } | |
628 | ||
629 | static void print_page_info(struct page *page) | |
630 | { | |
631 | pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", | |
632 | page, page->objects, page->inuse, page->freelist, page->flags); | |
633 | ||
634 | } | |
635 | ||
636 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) | |
637 | { | |
638 | struct va_format vaf; | |
639 | va_list args; | |
640 | ||
641 | va_start(args, fmt); | |
642 | vaf.fmt = fmt; | |
643 | vaf.va = &args; | |
644 | pr_err("=============================================================================\n"); | |
645 | pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf); | |
646 | pr_err("-----------------------------------------------------------------------------\n\n"); | |
647 | ||
648 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); | |
649 | va_end(args); | |
650 | } | |
651 | ||
652 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) | |
653 | { | |
654 | struct va_format vaf; | |
655 | va_list args; | |
656 | ||
657 | va_start(args, fmt); | |
658 | vaf.fmt = fmt; | |
659 | vaf.va = &args; | |
660 | pr_err("FIX %s: %pV\n", s->name, &vaf); | |
661 | va_end(args); | |
662 | } | |
663 | ||
664 | static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) | |
665 | { | |
666 | unsigned int off; /* Offset of last byte */ | |
667 | u8 *addr = page_address(page); | |
668 | ||
669 | print_tracking(s, p); | |
670 | ||
671 | print_page_info(page); | |
672 | ||
673 | pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", | |
674 | p, p - addr, get_freepointer(s, p)); | |
675 | ||
676 | if (s->flags & SLAB_RED_ZONE) | |
677 | print_section(KERN_ERR, "Redzone ", p - s->red_left_pad, | |
678 | s->red_left_pad); | |
679 | else if (p > addr + 16) | |
680 | print_section(KERN_ERR, "Bytes b4 ", p - 16, 16); | |
681 | ||
682 | print_section(KERN_ERR, "Object ", p, | |
683 | min_t(unsigned long, s->object_size, PAGE_SIZE)); | |
684 | if (s->flags & SLAB_RED_ZONE) | |
685 | print_section(KERN_ERR, "Redzone ", p + s->object_size, | |
686 | s->inuse - s->object_size); | |
687 | ||
688 | if (s->offset) | |
689 | off = s->offset + sizeof(void *); | |
690 | else | |
691 | off = s->inuse; | |
692 | ||
693 | if (s->flags & SLAB_STORE_USER) | |
694 | off += 2 * sizeof(struct track); | |
695 | ||
696 | off += kasan_metadata_size(s); | |
697 | ||
698 | if (off != size_from_object(s)) | |
699 | /* Beginning of the filler is the free pointer */ | |
700 | print_section(KERN_ERR, "Padding ", p + off, | |
701 | size_from_object(s) - off); | |
702 | ||
703 | dump_stack(); | |
704 | } | |
705 | ||
706 | void object_err(struct kmem_cache *s, struct page *page, | |
707 | u8 *object, char *reason) | |
708 | { | |
709 | slab_bug(s, "%s", reason); | |
710 | print_trailer(s, page, object); | |
711 | } | |
712 | ||
713 | static void slab_err(struct kmem_cache *s, struct page *page, | |
714 | const char *fmt, ...) | |
715 | { | |
716 | va_list args; | |
717 | char buf[100]; | |
718 | ||
719 | va_start(args, fmt); | |
720 | vsnprintf(buf, sizeof(buf), fmt, args); | |
721 | va_end(args); | |
722 | slab_bug(s, "%s", buf); | |
723 | print_page_info(page); | |
724 | dump_stack(); | |
725 | } | |
726 | ||
727 | static void init_object(struct kmem_cache *s, void *object, u8 val) | |
728 | { | |
729 | u8 *p = object; | |
730 | ||
731 | if (s->flags & SLAB_RED_ZONE) | |
732 | memset(p - s->red_left_pad, val, s->red_left_pad); | |
733 | ||
734 | if (s->flags & __OBJECT_POISON) { | |
735 | memset(p, POISON_FREE, s->object_size - 1); | |
736 | p[s->object_size - 1] = POISON_END; | |
737 | } | |
738 | ||
739 | if (s->flags & SLAB_RED_ZONE) | |
740 | memset(p + s->object_size, val, s->inuse - s->object_size); | |
741 | } | |
742 | ||
743 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
744 | void *from, void *to) | |
745 | { | |
746 | slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); | |
747 | memset(from, data, to - from); | |
748 | } | |
749 | ||
750 | static int check_bytes_and_report(struct kmem_cache *s, struct page *page, | |
751 | u8 *object, char *what, | |
752 | u8 *start, unsigned int value, unsigned int bytes) | |
753 | { | |
754 | u8 *fault; | |
755 | u8 *end; | |
756 | ||
757 | metadata_access_enable(); | |
758 | fault = memchr_inv(start, value, bytes); | |
759 | metadata_access_disable(); | |
760 | if (!fault) | |
761 | return 1; | |
762 | ||
763 | end = start + bytes; | |
764 | while (end > fault && end[-1] == value) | |
765 | end--; | |
766 | ||
767 | slab_bug(s, "%s overwritten", what); | |
768 | pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", | |
769 | fault, end - 1, fault[0], value); | |
770 | print_trailer(s, page, object); | |
771 | ||
772 | restore_bytes(s, what, value, fault, end); | |
773 | return 0; | |
774 | } | |
775 | ||
776 | /* | |
777 | * Object layout: | |
778 | * | |
779 | * object address | |
780 | * Bytes of the object to be managed. | |
781 | * If the freepointer may overlay the object then the free | |
782 | * pointer is the first word of the object. | |
783 | * | |
784 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is | |
785 | * 0xa5 (POISON_END) | |
786 | * | |
787 | * object + s->object_size | |
788 | * Padding to reach word boundary. This is also used for Redzoning. | |
789 | * Padding is extended by another word if Redzoning is enabled and | |
790 | * object_size == inuse. | |
791 | * | |
792 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with | |
793 | * 0xcc (RED_ACTIVE) for objects in use. | |
794 | * | |
795 | * object + s->inuse | |
796 | * Meta data starts here. | |
797 | * | |
798 | * A. Free pointer (if we cannot overwrite object on free) | |
799 | * B. Tracking data for SLAB_STORE_USER | |
800 | * C. Padding to reach required alignment boundary or at mininum | |
801 | * one word if debugging is on to be able to detect writes | |
802 | * before the word boundary. | |
803 | * | |
804 | * Padding is done using 0x5a (POISON_INUSE) | |
805 | * | |
806 | * object + s->size | |
807 | * Nothing is used beyond s->size. | |
808 | * | |
809 | * If slabcaches are merged then the object_size and inuse boundaries are mostly | |
810 | * ignored. And therefore no slab options that rely on these boundaries | |
811 | * may be used with merged slabcaches. | |
812 | */ | |
813 | ||
814 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) | |
815 | { | |
816 | unsigned long off = s->inuse; /* The end of info */ | |
817 | ||
818 | if (s->offset) | |
819 | /* Freepointer is placed after the object. */ | |
820 | off += sizeof(void *); | |
821 | ||
822 | if (s->flags & SLAB_STORE_USER) | |
823 | /* We also have user information there */ | |
824 | off += 2 * sizeof(struct track); | |
825 | ||
826 | off += kasan_metadata_size(s); | |
827 | ||
828 | if (size_from_object(s) == off) | |
829 | return 1; | |
830 | ||
831 | return check_bytes_and_report(s, page, p, "Object padding", | |
832 | p + off, POISON_INUSE, size_from_object(s) - off); | |
833 | } | |
834 | ||
835 | /* Check the pad bytes at the end of a slab page */ | |
836 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
837 | { | |
838 | u8 *start; | |
839 | u8 *fault; | |
840 | u8 *end; | |
841 | int length; | |
842 | int remainder; | |
843 | ||
844 | if (!(s->flags & SLAB_POISON)) | |
845 | return 1; | |
846 | ||
847 | start = page_address(page); | |
848 | length = (PAGE_SIZE << compound_order(page)) - s->reserved; | |
849 | end = start + length; | |
850 | remainder = length % s->size; | |
851 | if (!remainder) | |
852 | return 1; | |
853 | ||
854 | metadata_access_enable(); | |
855 | fault = memchr_inv(end - remainder, POISON_INUSE, remainder); | |
856 | metadata_access_disable(); | |
857 | if (!fault) | |
858 | return 1; | |
859 | while (end > fault && end[-1] == POISON_INUSE) | |
860 | end--; | |
861 | ||
862 | slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); | |
863 | print_section(KERN_ERR, "Padding ", end - remainder, remainder); | |
864 | ||
865 | restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); | |
866 | return 0; | |
867 | } | |
868 | ||
869 | static int check_object(struct kmem_cache *s, struct page *page, | |
870 | void *object, u8 val) | |
871 | { | |
872 | u8 *p = object; | |
873 | u8 *endobject = object + s->object_size; | |
874 | ||
875 | if (s->flags & SLAB_RED_ZONE) { | |
876 | if (!check_bytes_and_report(s, page, object, "Redzone", | |
877 | object - s->red_left_pad, val, s->red_left_pad)) | |
878 | return 0; | |
879 | ||
880 | if (!check_bytes_and_report(s, page, object, "Redzone", | |
881 | endobject, val, s->inuse - s->object_size)) | |
882 | return 0; | |
883 | } else { | |
884 | if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { | |
885 | check_bytes_and_report(s, page, p, "Alignment padding", | |
886 | endobject, POISON_INUSE, | |
887 | s->inuse - s->object_size); | |
888 | } | |
889 | } | |
890 | ||
891 | if (s->flags & SLAB_POISON) { | |
892 | if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && | |
893 | (!check_bytes_and_report(s, page, p, "Poison", p, | |
894 | POISON_FREE, s->object_size - 1) || | |
895 | !check_bytes_and_report(s, page, p, "Poison", | |
896 | p + s->object_size - 1, POISON_END, 1))) | |
897 | return 0; | |
898 | /* | |
899 | * check_pad_bytes cleans up on its own. | |
900 | */ | |
901 | check_pad_bytes(s, page, p); | |
902 | } | |
903 | ||
904 | if (!s->offset && val == SLUB_RED_ACTIVE) | |
905 | /* | |
906 | * Object and freepointer overlap. Cannot check | |
907 | * freepointer while object is allocated. | |
908 | */ | |
909 | return 1; | |
910 | ||
911 | /* Check free pointer validity */ | |
912 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
913 | object_err(s, page, p, "Freepointer corrupt"); | |
914 | /* | |
915 | * No choice but to zap it and thus lose the remainder | |
916 | * of the free objects in this slab. May cause | |
917 | * another error because the object count is now wrong. | |
918 | */ | |
919 | set_freepointer(s, p, NULL); | |
920 | return 0; | |
921 | } | |
922 | return 1; | |
923 | } | |
924 | ||
925 | static int check_slab(struct kmem_cache *s, struct page *page) | |
926 | { | |
927 | int maxobj; | |
928 | ||
929 | VM_BUG_ON(!irqs_disabled()); | |
930 | ||
931 | if (!PageSlab(page)) { | |
932 | slab_err(s, page, "Not a valid slab page"); | |
933 | return 0; | |
934 | } | |
935 | ||
936 | maxobj = order_objects(compound_order(page), s->size, s->reserved); | |
937 | if (page->objects > maxobj) { | |
938 | slab_err(s, page, "objects %u > max %u", | |
939 | page->objects, maxobj); | |
940 | return 0; | |
941 | } | |
942 | if (page->inuse > page->objects) { | |
943 | slab_err(s, page, "inuse %u > max %u", | |
944 | page->inuse, page->objects); | |
945 | return 0; | |
946 | } | |
947 | /* Slab_pad_check fixes things up after itself */ | |
948 | slab_pad_check(s, page); | |
949 | return 1; | |
950 | } | |
951 | ||
952 | /* | |
953 | * Determine if a certain object on a page is on the freelist. Must hold the | |
954 | * slab lock to guarantee that the chains are in a consistent state. | |
955 | */ | |
956 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
957 | { | |
958 | int nr = 0; | |
959 | void *fp; | |
960 | void *object = NULL; | |
961 | int max_objects; | |
962 | ||
963 | fp = page->freelist; | |
964 | while (fp && nr <= page->objects) { | |
965 | if (fp == search) | |
966 | return 1; | |
967 | if (!check_valid_pointer(s, page, fp)) { | |
968 | if (object) { | |
969 | object_err(s, page, object, | |
970 | "Freechain corrupt"); | |
971 | set_freepointer(s, object, NULL); | |
972 | } else { | |
973 | slab_err(s, page, "Freepointer corrupt"); | |
974 | page->freelist = NULL; | |
975 | page->inuse = page->objects; | |
976 | slab_fix(s, "Freelist cleared"); | |
977 | return 0; | |
978 | } | |
979 | break; | |
980 | } | |
981 | object = fp; | |
982 | fp = get_freepointer(s, object); | |
983 | nr++; | |
984 | } | |
985 | ||
986 | max_objects = order_objects(compound_order(page), s->size, s->reserved); | |
987 | if (max_objects > MAX_OBJS_PER_PAGE) | |
988 | max_objects = MAX_OBJS_PER_PAGE; | |
989 | ||
990 | if (page->objects != max_objects) { | |
991 | slab_err(s, page, "Wrong number of objects. Found %d but should be %d", | |
992 | page->objects, max_objects); | |
993 | page->objects = max_objects; | |
994 | slab_fix(s, "Number of objects adjusted."); | |
995 | } | |
996 | if (page->inuse != page->objects - nr) { | |
997 | slab_err(s, page, "Wrong object count. Counter is %d but counted were %d", | |
998 | page->inuse, page->objects - nr); | |
999 | page->inuse = page->objects - nr; | |
1000 | slab_fix(s, "Object count adjusted."); | |
1001 | } | |
1002 | return search == NULL; | |
1003 | } | |
1004 | ||
1005 | static void trace(struct kmem_cache *s, struct page *page, void *object, | |
1006 | int alloc) | |
1007 | { | |
1008 | if (s->flags & SLAB_TRACE) { | |
1009 | pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n", | |
1010 | s->name, | |
1011 | alloc ? "alloc" : "free", | |
1012 | object, page->inuse, | |
1013 | page->freelist); | |
1014 | ||
1015 | if (!alloc) | |
1016 | print_section(KERN_INFO, "Object ", (void *)object, | |
1017 | s->object_size); | |
1018 | ||
1019 | dump_stack(); | |
1020 | } | |
1021 | } | |
1022 | ||
1023 | /* | |
1024 | * Tracking of fully allocated slabs for debugging purposes. | |
1025 | */ | |
1026 | static void add_full(struct kmem_cache *s, | |
1027 | struct kmem_cache_node *n, struct page *page) | |
1028 | { | |
1029 | if (!(s->flags & SLAB_STORE_USER)) | |
1030 | return; | |
1031 | ||
1032 | lockdep_assert_held(&n->list_lock); | |
1033 | list_add(&page->lru, &n->full); | |
1034 | } | |
1035 | ||
1036 | static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page) | |
1037 | { | |
1038 | if (!(s->flags & SLAB_STORE_USER)) | |
1039 | return; | |
1040 | ||
1041 | lockdep_assert_held(&n->list_lock); | |
1042 | list_del(&page->lru); | |
1043 | } | |
1044 | ||
1045 | /* Tracking of the number of slabs for debugging purposes */ | |
1046 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
1047 | { | |
1048 | struct kmem_cache_node *n = get_node(s, node); | |
1049 | ||
1050 | return atomic_long_read(&n->nr_slabs); | |
1051 | } | |
1052 | ||
1053 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) | |
1054 | { | |
1055 | return atomic_long_read(&n->nr_slabs); | |
1056 | } | |
1057 | ||
1058 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) | |
1059 | { | |
1060 | struct kmem_cache_node *n = get_node(s, node); | |
1061 | ||
1062 | /* | |
1063 | * May be called early in order to allocate a slab for the | |
1064 | * kmem_cache_node structure. Solve the chicken-egg | |
1065 | * dilemma by deferring the increment of the count during | |
1066 | * bootstrap (see early_kmem_cache_node_alloc). | |
1067 | */ | |
1068 | if (likely(n)) { | |
1069 | atomic_long_inc(&n->nr_slabs); | |
1070 | atomic_long_add(objects, &n->total_objects); | |
1071 | } | |
1072 | } | |
1073 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) | |
1074 | { | |
1075 | struct kmem_cache_node *n = get_node(s, node); | |
1076 | ||
1077 | atomic_long_dec(&n->nr_slabs); | |
1078 | atomic_long_sub(objects, &n->total_objects); | |
1079 | } | |
1080 | ||
1081 | /* Object debug checks for alloc/free paths */ | |
1082 | static void setup_object_debug(struct kmem_cache *s, struct page *page, | |
1083 | void *object) | |
1084 | { | |
1085 | if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) | |
1086 | return; | |
1087 | ||
1088 | init_object(s, object, SLUB_RED_INACTIVE); | |
1089 | init_tracking(s, object); | |
1090 | } | |
1091 | ||
1092 | static inline int alloc_consistency_checks(struct kmem_cache *s, | |
1093 | struct page *page, | |
1094 | void *object, unsigned long addr) | |
1095 | { | |
1096 | if (!check_slab(s, page)) | |
1097 | return 0; | |
1098 | ||
1099 | if (!check_valid_pointer(s, page, object)) { | |
1100 | object_err(s, page, object, "Freelist Pointer check fails"); | |
1101 | return 0; | |
1102 | } | |
1103 | ||
1104 | if (!check_object(s, page, object, SLUB_RED_INACTIVE)) | |
1105 | return 0; | |
1106 | ||
1107 | return 1; | |
1108 | } | |
1109 | ||
1110 | static noinline int alloc_debug_processing(struct kmem_cache *s, | |
1111 | struct page *page, | |
1112 | void *object, unsigned long addr) | |
1113 | { | |
1114 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { | |
1115 | if (!alloc_consistency_checks(s, page, object, addr)) | |
1116 | goto bad; | |
1117 | } | |
1118 | ||
1119 | /* Success perform special debug activities for allocs */ | |
1120 | if (s->flags & SLAB_STORE_USER) | |
1121 | set_track(s, object, TRACK_ALLOC, addr); | |
1122 | trace(s, page, object, 1); | |
1123 | init_object(s, object, SLUB_RED_ACTIVE); | |
1124 | return 1; | |
1125 | ||
1126 | bad: | |
1127 | if (PageSlab(page)) { | |
1128 | /* | |
1129 | * If this is a slab page then lets do the best we can | |
1130 | * to avoid issues in the future. Marking all objects | |
1131 | * as used avoids touching the remaining objects. | |
1132 | */ | |
1133 | slab_fix(s, "Marking all objects used"); | |
1134 | page->inuse = page->objects; | |
1135 | page->freelist = NULL; | |
1136 | } | |
1137 | return 0; | |
1138 | } | |
1139 | ||
1140 | static inline int free_consistency_checks(struct kmem_cache *s, | |
1141 | struct page *page, void *object, unsigned long addr) | |
1142 | { | |
1143 | if (!check_valid_pointer(s, page, object)) { | |
1144 | slab_err(s, page, "Invalid object pointer 0x%p", object); | |
1145 | return 0; | |
1146 | } | |
1147 | ||
1148 | if (on_freelist(s, page, object)) { | |
1149 | object_err(s, page, object, "Object already free"); | |
1150 | return 0; | |
1151 | } | |
1152 | ||
1153 | if (!check_object(s, page, object, SLUB_RED_ACTIVE)) | |
1154 | return 0; | |
1155 | ||
1156 | if (unlikely(s != page->slab_cache)) { | |
1157 | if (!PageSlab(page)) { | |
1158 | slab_err(s, page, "Attempt to free object(0x%p) outside of slab", | |
1159 | object); | |
1160 | } else if (!page->slab_cache) { | |
1161 | pr_err("SLUB <none>: no slab for object 0x%p.\n", | |
1162 | object); | |
1163 | dump_stack(); | |
1164 | } else | |
1165 | object_err(s, page, object, | |
1166 | "page slab pointer corrupt."); | |
1167 | return 0; | |
1168 | } | |
1169 | return 1; | |
1170 | } | |
1171 | ||
1172 | /* Supports checking bulk free of a constructed freelist */ | |
1173 | static noinline int free_debug_processing( | |
1174 | struct kmem_cache *s, struct page *page, | |
1175 | void *head, void *tail, int bulk_cnt, | |
1176 | unsigned long addr) | |
1177 | { | |
1178 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1179 | void *object = head; | |
1180 | int cnt = 0; | |
1181 | unsigned long uninitialized_var(flags); | |
1182 | int ret = 0; | |
1183 | ||
1184 | spin_lock_irqsave(&n->list_lock, flags); | |
1185 | slab_lock(page); | |
1186 | ||
1187 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { | |
1188 | if (!check_slab(s, page)) | |
1189 | goto out; | |
1190 | } | |
1191 | ||
1192 | next_object: | |
1193 | cnt++; | |
1194 | ||
1195 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { | |
1196 | if (!free_consistency_checks(s, page, object, addr)) | |
1197 | goto out; | |
1198 | } | |
1199 | ||
1200 | if (s->flags & SLAB_STORE_USER) | |
1201 | set_track(s, object, TRACK_FREE, addr); | |
1202 | trace(s, page, object, 0); | |
1203 | /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ | |
1204 | init_object(s, object, SLUB_RED_INACTIVE); | |
1205 | ||
1206 | /* Reached end of constructed freelist yet? */ | |
1207 | if (object != tail) { | |
1208 | object = get_freepointer(s, object); | |
1209 | goto next_object; | |
1210 | } | |
1211 | ret = 1; | |
1212 | ||
1213 | out: | |
1214 | if (cnt != bulk_cnt) | |
1215 | slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n", | |
1216 | bulk_cnt, cnt); | |
1217 | ||
1218 | slab_unlock(page); | |
1219 | spin_unlock_irqrestore(&n->list_lock, flags); | |
1220 | if (!ret) | |
1221 | slab_fix(s, "Object at 0x%p not freed", object); | |
1222 | return ret; | |
1223 | } | |
1224 | ||
1225 | static int __init setup_slub_debug(char *str) | |
1226 | { | |
1227 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
1228 | if (*str++ != '=' || !*str) | |
1229 | /* | |
1230 | * No options specified. Switch on full debugging. | |
1231 | */ | |
1232 | goto out; | |
1233 | ||
1234 | if (*str == ',') | |
1235 | /* | |
1236 | * No options but restriction on slabs. This means full | |
1237 | * debugging for slabs matching a pattern. | |
1238 | */ | |
1239 | goto check_slabs; | |
1240 | ||
1241 | slub_debug = 0; | |
1242 | if (*str == '-') | |
1243 | /* | |
1244 | * Switch off all debugging measures. | |
1245 | */ | |
1246 | goto out; | |
1247 | ||
1248 | /* | |
1249 | * Determine which debug features should be switched on | |
1250 | */ | |
1251 | for (; *str && *str != ','; str++) { | |
1252 | switch (tolower(*str)) { | |
1253 | case 'f': | |
1254 | slub_debug |= SLAB_CONSISTENCY_CHECKS; | |
1255 | break; | |
1256 | case 'z': | |
1257 | slub_debug |= SLAB_RED_ZONE; | |
1258 | break; | |
1259 | case 'p': | |
1260 | slub_debug |= SLAB_POISON; | |
1261 | break; | |
1262 | case 'u': | |
1263 | slub_debug |= SLAB_STORE_USER; | |
1264 | break; | |
1265 | case 't': | |
1266 | slub_debug |= SLAB_TRACE; | |
1267 | break; | |
1268 | case 'a': | |
1269 | slub_debug |= SLAB_FAILSLAB; | |
1270 | break; | |
1271 | case 'o': | |
1272 | /* | |
1273 | * Avoid enabling debugging on caches if its minimum | |
1274 | * order would increase as a result. | |
1275 | */ | |
1276 | disable_higher_order_debug = 1; | |
1277 | break; | |
1278 | default: | |
1279 | pr_err("slub_debug option '%c' unknown. skipped\n", | |
1280 | *str); | |
1281 | } | |
1282 | } | |
1283 | ||
1284 | check_slabs: | |
1285 | if (*str == ',') | |
1286 | slub_debug_slabs = str + 1; | |
1287 | out: | |
1288 | return 1; | |
1289 | } | |
1290 | ||
1291 | __setup("slub_debug", setup_slub_debug); | |
1292 | ||
1293 | slab_flags_t kmem_cache_flags(unsigned long object_size, | |
1294 | slab_flags_t flags, const char *name, | |
1295 | void (*ctor)(void *)) | |
1296 | { | |
1297 | /* | |
1298 | * Enable debugging if selected on the kernel commandline. | |
1299 | */ | |
1300 | if (slub_debug && (!slub_debug_slabs || (name && | |
1301 | !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))) | |
1302 | flags |= slub_debug; | |
1303 | ||
1304 | return flags; | |
1305 | } | |
1306 | #else /* !CONFIG_SLUB_DEBUG */ | |
1307 | static inline void setup_object_debug(struct kmem_cache *s, | |
1308 | struct page *page, void *object) {} | |
1309 | ||
1310 | static inline int alloc_debug_processing(struct kmem_cache *s, | |
1311 | struct page *page, void *object, unsigned long addr) { return 0; } | |
1312 | ||
1313 | static inline int free_debug_processing( | |
1314 | struct kmem_cache *s, struct page *page, | |
1315 | void *head, void *tail, int bulk_cnt, | |
1316 | unsigned long addr) { return 0; } | |
1317 | ||
1318 | static inline int slab_pad_check(struct kmem_cache *s, struct page *page) | |
1319 | { return 1; } | |
1320 | static inline int check_object(struct kmem_cache *s, struct page *page, | |
1321 | void *object, u8 val) { return 1; } | |
1322 | static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, | |
1323 | struct page *page) {} | |
1324 | static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, | |
1325 | struct page *page) {} | |
1326 | slab_flags_t kmem_cache_flags(unsigned long object_size, | |
1327 | slab_flags_t flags, const char *name, | |
1328 | void (*ctor)(void *)) | |
1329 | { | |
1330 | return flags; | |
1331 | } | |
1332 | #define slub_debug 0 | |
1333 | ||
1334 | #define disable_higher_order_debug 0 | |
1335 | ||
1336 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
1337 | { return 0; } | |
1338 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) | |
1339 | { return 0; } | |
1340 | static inline void inc_slabs_node(struct kmem_cache *s, int node, | |
1341 | int objects) {} | |
1342 | static inline void dec_slabs_node(struct kmem_cache *s, int node, | |
1343 | int objects) {} | |
1344 | ||
1345 | #endif /* CONFIG_SLUB_DEBUG */ | |
1346 | ||
1347 | /* | |
1348 | * Hooks for other subsystems that check memory allocations. In a typical | |
1349 | * production configuration these hooks all should produce no code at all. | |
1350 | */ | |
1351 | static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags) | |
1352 | { | |
1353 | kmemleak_alloc(ptr, size, 1, flags); | |
1354 | kasan_kmalloc_large(ptr, size, flags); | |
1355 | } | |
1356 | ||
1357 | static inline void kfree_hook(const void *x) | |
1358 | { | |
1359 | kmemleak_free(x); | |
1360 | kasan_kfree_large(x); | |
1361 | } | |
1362 | ||
1363 | static inline void *slab_free_hook(struct kmem_cache *s, void *x) | |
1364 | { | |
1365 | void *freeptr; | |
1366 | ||
1367 | kmemleak_free_recursive(x, s->flags); | |
1368 | ||
1369 | /* | |
1370 | * Trouble is that we may no longer disable interrupts in the fast path | |
1371 | * So in order to make the debug calls that expect irqs to be | |
1372 | * disabled we need to disable interrupts temporarily. | |
1373 | */ | |
1374 | #ifdef CONFIG_LOCKDEP | |
1375 | { | |
1376 | unsigned long flags; | |
1377 | ||
1378 | local_irq_save(flags); | |
1379 | debug_check_no_locks_freed(x, s->object_size); | |
1380 | local_irq_restore(flags); | |
1381 | } | |
1382 | #endif | |
1383 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) | |
1384 | debug_check_no_obj_freed(x, s->object_size); | |
1385 | ||
1386 | freeptr = get_freepointer(s, x); | |
1387 | /* | |
1388 | * kasan_slab_free() may put x into memory quarantine, delaying its | |
1389 | * reuse. In this case the object's freelist pointer is changed. | |
1390 | */ | |
1391 | kasan_slab_free(s, x); | |
1392 | return freeptr; | |
1393 | } | |
1394 | ||
1395 | static inline void slab_free_freelist_hook(struct kmem_cache *s, | |
1396 | void *head, void *tail) | |
1397 | { | |
1398 | /* | |
1399 | * Compiler cannot detect this function can be removed if slab_free_hook() | |
1400 | * evaluates to nothing. Thus, catch all relevant config debug options here. | |
1401 | */ | |
1402 | #if defined(CONFIG_LOCKDEP) || \ | |
1403 | defined(CONFIG_DEBUG_KMEMLEAK) || \ | |
1404 | defined(CONFIG_DEBUG_OBJECTS_FREE) || \ | |
1405 | defined(CONFIG_KASAN) | |
1406 | ||
1407 | void *object = head; | |
1408 | void *tail_obj = tail ? : head; | |
1409 | void *freeptr; | |
1410 | ||
1411 | do { | |
1412 | freeptr = slab_free_hook(s, object); | |
1413 | } while ((object != tail_obj) && (object = freeptr)); | |
1414 | #endif | |
1415 | } | |
1416 | ||
1417 | static void setup_object(struct kmem_cache *s, struct page *page, | |
1418 | void *object) | |
1419 | { | |
1420 | setup_object_debug(s, page, object); | |
1421 | kasan_init_slab_obj(s, object); | |
1422 | if (unlikely(s->ctor)) { | |
1423 | kasan_unpoison_object_data(s, object); | |
1424 | s->ctor(object); | |
1425 | kasan_poison_object_data(s, object); | |
1426 | } | |
1427 | } | |
1428 | ||
1429 | /* | |
1430 | * Slab allocation and freeing | |
1431 | */ | |
1432 | static inline struct page *alloc_slab_page(struct kmem_cache *s, | |
1433 | gfp_t flags, int node, struct kmem_cache_order_objects oo) | |
1434 | { | |
1435 | struct page *page; | |
1436 | int order = oo_order(oo); | |
1437 | ||
1438 | if (node == NUMA_NO_NODE) | |
1439 | page = alloc_pages(flags, order); | |
1440 | else | |
1441 | page = __alloc_pages_node(node, flags, order); | |
1442 | ||
1443 | if (page && memcg_charge_slab(page, flags, order, s)) { | |
1444 | __free_pages(page, order); | |
1445 | page = NULL; | |
1446 | } | |
1447 | ||
1448 | return page; | |
1449 | } | |
1450 | ||
1451 | #ifdef CONFIG_SLAB_FREELIST_RANDOM | |
1452 | /* Pre-initialize the random sequence cache */ | |
1453 | static int init_cache_random_seq(struct kmem_cache *s) | |
1454 | { | |
1455 | int err; | |
1456 | unsigned long i, count = oo_objects(s->oo); | |
1457 | ||
1458 | /* Bailout if already initialised */ | |
1459 | if (s->random_seq) | |
1460 | return 0; | |
1461 | ||
1462 | err = cache_random_seq_create(s, count, GFP_KERNEL); | |
1463 | if (err) { | |
1464 | pr_err("SLUB: Unable to initialize free list for %s\n", | |
1465 | s->name); | |
1466 | return err; | |
1467 | } | |
1468 | ||
1469 | /* Transform to an offset on the set of pages */ | |
1470 | if (s->random_seq) { | |
1471 | for (i = 0; i < count; i++) | |
1472 | s->random_seq[i] *= s->size; | |
1473 | } | |
1474 | return 0; | |
1475 | } | |
1476 | ||
1477 | /* Initialize each random sequence freelist per cache */ | |
1478 | static void __init init_freelist_randomization(void) | |
1479 | { | |
1480 | struct kmem_cache *s; | |
1481 | ||
1482 | mutex_lock(&slab_mutex); | |
1483 | ||
1484 | list_for_each_entry(s, &slab_caches, list) | |
1485 | init_cache_random_seq(s); | |
1486 | ||
1487 | mutex_unlock(&slab_mutex); | |
1488 | } | |
1489 | ||
1490 | /* Get the next entry on the pre-computed freelist randomized */ | |
1491 | static void *next_freelist_entry(struct kmem_cache *s, struct page *page, | |
1492 | unsigned long *pos, void *start, | |
1493 | unsigned long page_limit, | |
1494 | unsigned long freelist_count) | |
1495 | { | |
1496 | unsigned int idx; | |
1497 | ||
1498 | /* | |
1499 | * If the target page allocation failed, the number of objects on the | |
1500 | * page might be smaller than the usual size defined by the cache. | |
1501 | */ | |
1502 | do { | |
1503 | idx = s->random_seq[*pos]; | |
1504 | *pos += 1; | |
1505 | if (*pos >= freelist_count) | |
1506 | *pos = 0; | |
1507 | } while (unlikely(idx >= page_limit)); | |
1508 | ||
1509 | return (char *)start + idx; | |
1510 | } | |
1511 | ||
1512 | /* Shuffle the single linked freelist based on a random pre-computed sequence */ | |
1513 | static bool shuffle_freelist(struct kmem_cache *s, struct page *page) | |
1514 | { | |
1515 | void *start; | |
1516 | void *cur; | |
1517 | void *next; | |
1518 | unsigned long idx, pos, page_limit, freelist_count; | |
1519 | ||
1520 | if (page->objects < 2 || !s->random_seq) | |
1521 | return false; | |
1522 | ||
1523 | freelist_count = oo_objects(s->oo); | |
1524 | pos = get_random_int() % freelist_count; | |
1525 | ||
1526 | page_limit = page->objects * s->size; | |
1527 | start = fixup_red_left(s, page_address(page)); | |
1528 | ||
1529 | /* First entry is used as the base of the freelist */ | |
1530 | cur = next_freelist_entry(s, page, &pos, start, page_limit, | |
1531 | freelist_count); | |
1532 | page->freelist = cur; | |
1533 | ||
1534 | for (idx = 1; idx < page->objects; idx++) { | |
1535 | setup_object(s, page, cur); | |
1536 | next = next_freelist_entry(s, page, &pos, start, page_limit, | |
1537 | freelist_count); | |
1538 | set_freepointer(s, cur, next); | |
1539 | cur = next; | |
1540 | } | |
1541 | setup_object(s, page, cur); | |
1542 | set_freepointer(s, cur, NULL); | |
1543 | ||
1544 | return true; | |
1545 | } | |
1546 | #else | |
1547 | static inline int init_cache_random_seq(struct kmem_cache *s) | |
1548 | { | |
1549 | return 0; | |
1550 | } | |
1551 | static inline void init_freelist_randomization(void) { } | |
1552 | static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page) | |
1553 | { | |
1554 | return false; | |
1555 | } | |
1556 | #endif /* CONFIG_SLAB_FREELIST_RANDOM */ | |
1557 | ||
1558 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1559 | { | |
1560 | struct page *page; | |
1561 | struct kmem_cache_order_objects oo = s->oo; | |
1562 | gfp_t alloc_gfp; | |
1563 | void *start, *p; | |
1564 | int idx, order; | |
1565 | bool shuffle; | |
1566 | ||
1567 | flags &= gfp_allowed_mask; | |
1568 | ||
1569 | if (gfpflags_allow_blocking(flags)) | |
1570 | local_irq_enable(); | |
1571 | ||
1572 | flags |= s->allocflags; | |
1573 | ||
1574 | /* | |
1575 | * Let the initial higher-order allocation fail under memory pressure | |
1576 | * so we fall-back to the minimum order allocation. | |
1577 | */ | |
1578 | alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; | |
1579 | if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min)) | |
1580 | alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL); | |
1581 | ||
1582 | page = alloc_slab_page(s, alloc_gfp, node, oo); | |
1583 | if (unlikely(!page)) { | |
1584 | oo = s->min; | |
1585 | alloc_gfp = flags; | |
1586 | /* | |
1587 | * Allocation may have failed due to fragmentation. | |
1588 | * Try a lower order alloc if possible | |
1589 | */ | |
1590 | page = alloc_slab_page(s, alloc_gfp, node, oo); | |
1591 | if (unlikely(!page)) | |
1592 | goto out; | |
1593 | stat(s, ORDER_FALLBACK); | |
1594 | } | |
1595 | ||
1596 | page->objects = oo_objects(oo); | |
1597 | ||
1598 | order = compound_order(page); | |
1599 | page->slab_cache = s; | |
1600 | __SetPageSlab(page); | |
1601 | if (page_is_pfmemalloc(page)) | |
1602 | SetPageSlabPfmemalloc(page); | |
1603 | ||
1604 | start = page_address(page); | |
1605 | ||
1606 | if (unlikely(s->flags & SLAB_POISON)) | |
1607 | memset(start, POISON_INUSE, PAGE_SIZE << order); | |
1608 | ||
1609 | kasan_poison_slab(page); | |
1610 | ||
1611 | shuffle = shuffle_freelist(s, page); | |
1612 | ||
1613 | if (!shuffle) { | |
1614 | for_each_object_idx(p, idx, s, start, page->objects) { | |
1615 | setup_object(s, page, p); | |
1616 | if (likely(idx < page->objects)) | |
1617 | set_freepointer(s, p, p + s->size); | |
1618 | else | |
1619 | set_freepointer(s, p, NULL); | |
1620 | } | |
1621 | page->freelist = fixup_red_left(s, start); | |
1622 | } | |
1623 | ||
1624 | page->inuse = page->objects; | |
1625 | page->frozen = 1; | |
1626 | ||
1627 | out: | |
1628 | if (gfpflags_allow_blocking(flags)) | |
1629 | local_irq_disable(); | |
1630 | if (!page) | |
1631 | return NULL; | |
1632 | ||
1633 | mod_lruvec_page_state(page, | |
1634 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1635 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1636 | 1 << oo_order(oo)); | |
1637 | ||
1638 | inc_slabs_node(s, page_to_nid(page), page->objects); | |
1639 | ||
1640 | return page; | |
1641 | } | |
1642 | ||
1643 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1644 | { | |
1645 | if (unlikely(flags & GFP_SLAB_BUG_MASK)) { | |
1646 | gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK; | |
1647 | flags &= ~GFP_SLAB_BUG_MASK; | |
1648 | pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n", | |
1649 | invalid_mask, &invalid_mask, flags, &flags); | |
1650 | dump_stack(); | |
1651 | } | |
1652 | ||
1653 | return allocate_slab(s, | |
1654 | flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); | |
1655 | } | |
1656 | ||
1657 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
1658 | { | |
1659 | int order = compound_order(page); | |
1660 | int pages = 1 << order; | |
1661 | ||
1662 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { | |
1663 | void *p; | |
1664 | ||
1665 | slab_pad_check(s, page); | |
1666 | for_each_object(p, s, page_address(page), | |
1667 | page->objects) | |
1668 | check_object(s, page, p, SLUB_RED_INACTIVE); | |
1669 | } | |
1670 | ||
1671 | mod_lruvec_page_state(page, | |
1672 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1673 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1674 | -pages); | |
1675 | ||
1676 | __ClearPageSlabPfmemalloc(page); | |
1677 | __ClearPageSlab(page); | |
1678 | ||
1679 | page_mapcount_reset(page); | |
1680 | if (current->reclaim_state) | |
1681 | current->reclaim_state->reclaimed_slab += pages; | |
1682 | memcg_uncharge_slab(page, order, s); | |
1683 | __free_pages(page, order); | |
1684 | } | |
1685 | ||
1686 | #define need_reserve_slab_rcu \ | |
1687 | (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) | |
1688 | ||
1689 | static void rcu_free_slab(struct rcu_head *h) | |
1690 | { | |
1691 | struct page *page; | |
1692 | ||
1693 | if (need_reserve_slab_rcu) | |
1694 | page = virt_to_head_page(h); | |
1695 | else | |
1696 | page = container_of((struct list_head *)h, struct page, lru); | |
1697 | ||
1698 | __free_slab(page->slab_cache, page); | |
1699 | } | |
1700 | ||
1701 | static void free_slab(struct kmem_cache *s, struct page *page) | |
1702 | { | |
1703 | if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) { | |
1704 | struct rcu_head *head; | |
1705 | ||
1706 | if (need_reserve_slab_rcu) { | |
1707 | int order = compound_order(page); | |
1708 | int offset = (PAGE_SIZE << order) - s->reserved; | |
1709 | ||
1710 | VM_BUG_ON(s->reserved != sizeof(*head)); | |
1711 | head = page_address(page) + offset; | |
1712 | } else { | |
1713 | head = &page->rcu_head; | |
1714 | } | |
1715 | ||
1716 | call_rcu(head, rcu_free_slab); | |
1717 | } else | |
1718 | __free_slab(s, page); | |
1719 | } | |
1720 | ||
1721 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
1722 | { | |
1723 | dec_slabs_node(s, page_to_nid(page), page->objects); | |
1724 | free_slab(s, page); | |
1725 | } | |
1726 | ||
1727 | /* | |
1728 | * Management of partially allocated slabs. | |
1729 | */ | |
1730 | static inline void | |
1731 | __add_partial(struct kmem_cache_node *n, struct page *page, int tail) | |
1732 | { | |
1733 | n->nr_partial++; | |
1734 | if (tail == DEACTIVATE_TO_TAIL) | |
1735 | list_add_tail(&page->lru, &n->partial); | |
1736 | else | |
1737 | list_add(&page->lru, &n->partial); | |
1738 | } | |
1739 | ||
1740 | static inline void add_partial(struct kmem_cache_node *n, | |
1741 | struct page *page, int tail) | |
1742 | { | |
1743 | lockdep_assert_held(&n->list_lock); | |
1744 | __add_partial(n, page, tail); | |
1745 | } | |
1746 | ||
1747 | static inline void remove_partial(struct kmem_cache_node *n, | |
1748 | struct page *page) | |
1749 | { | |
1750 | lockdep_assert_held(&n->list_lock); | |
1751 | list_del(&page->lru); | |
1752 | n->nr_partial--; | |
1753 | } | |
1754 | ||
1755 | /* | |
1756 | * Remove slab from the partial list, freeze it and | |
1757 | * return the pointer to the freelist. | |
1758 | * | |
1759 | * Returns a list of objects or NULL if it fails. | |
1760 | */ | |
1761 | static inline void *acquire_slab(struct kmem_cache *s, | |
1762 | struct kmem_cache_node *n, struct page *page, | |
1763 | int mode, int *objects) | |
1764 | { | |
1765 | void *freelist; | |
1766 | unsigned long counters; | |
1767 | struct page new; | |
1768 | ||
1769 | lockdep_assert_held(&n->list_lock); | |
1770 | ||
1771 | /* | |
1772 | * Zap the freelist and set the frozen bit. | |
1773 | * The old freelist is the list of objects for the | |
1774 | * per cpu allocation list. | |
1775 | */ | |
1776 | freelist = page->freelist; | |
1777 | counters = page->counters; | |
1778 | new.counters = counters; | |
1779 | *objects = new.objects - new.inuse; | |
1780 | if (mode) { | |
1781 | new.inuse = page->objects; | |
1782 | new.freelist = NULL; | |
1783 | } else { | |
1784 | new.freelist = freelist; | |
1785 | } | |
1786 | ||
1787 | VM_BUG_ON(new.frozen); | |
1788 | new.frozen = 1; | |
1789 | ||
1790 | if (!__cmpxchg_double_slab(s, page, | |
1791 | freelist, counters, | |
1792 | new.freelist, new.counters, | |
1793 | "acquire_slab")) | |
1794 | return NULL; | |
1795 | ||
1796 | remove_partial(n, page); | |
1797 | WARN_ON(!freelist); | |
1798 | return freelist; | |
1799 | } | |
1800 | ||
1801 | static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); | |
1802 | static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags); | |
1803 | ||
1804 | /* | |
1805 | * Try to allocate a partial slab from a specific node. | |
1806 | */ | |
1807 | static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, | |
1808 | struct kmem_cache_cpu *c, gfp_t flags) | |
1809 | { | |
1810 | struct page *page, *page2; | |
1811 | void *object = NULL; | |
1812 | int available = 0; | |
1813 | int objects; | |
1814 | ||
1815 | /* | |
1816 | * Racy check. If we mistakenly see no partial slabs then we | |
1817 | * just allocate an empty slab. If we mistakenly try to get a | |
1818 | * partial slab and there is none available then get_partials() | |
1819 | * will return NULL. | |
1820 | */ | |
1821 | if (!n || !n->nr_partial) | |
1822 | return NULL; | |
1823 | ||
1824 | spin_lock(&n->list_lock); | |
1825 | list_for_each_entry_safe(page, page2, &n->partial, lru) { | |
1826 | void *t; | |
1827 | ||
1828 | if (!pfmemalloc_match(page, flags)) | |
1829 | continue; | |
1830 | ||
1831 | t = acquire_slab(s, n, page, object == NULL, &objects); | |
1832 | if (!t) | |
1833 | break; | |
1834 | ||
1835 | available += objects; | |
1836 | if (!object) { | |
1837 | c->page = page; | |
1838 | stat(s, ALLOC_FROM_PARTIAL); | |
1839 | object = t; | |
1840 | } else { | |
1841 | put_cpu_partial(s, page, 0); | |
1842 | stat(s, CPU_PARTIAL_NODE); | |
1843 | } | |
1844 | if (!kmem_cache_has_cpu_partial(s) | |
1845 | || available > slub_cpu_partial(s) / 2) | |
1846 | break; | |
1847 | ||
1848 | } | |
1849 | spin_unlock(&n->list_lock); | |
1850 | return object; | |
1851 | } | |
1852 | ||
1853 | /* | |
1854 | * Get a page from somewhere. Search in increasing NUMA distances. | |
1855 | */ | |
1856 | static void *get_any_partial(struct kmem_cache *s, gfp_t flags, | |
1857 | struct kmem_cache_cpu *c) | |
1858 | { | |
1859 | #ifdef CONFIG_NUMA | |
1860 | struct zonelist *zonelist; | |
1861 | struct zoneref *z; | |
1862 | struct zone *zone; | |
1863 | enum zone_type high_zoneidx = gfp_zone(flags); | |
1864 | void *object; | |
1865 | unsigned int cpuset_mems_cookie; | |
1866 | ||
1867 | /* | |
1868 | * The defrag ratio allows a configuration of the tradeoffs between | |
1869 | * inter node defragmentation and node local allocations. A lower | |
1870 | * defrag_ratio increases the tendency to do local allocations | |
1871 | * instead of attempting to obtain partial slabs from other nodes. | |
1872 | * | |
1873 | * If the defrag_ratio is set to 0 then kmalloc() always | |
1874 | * returns node local objects. If the ratio is higher then kmalloc() | |
1875 | * may return off node objects because partial slabs are obtained | |
1876 | * from other nodes and filled up. | |
1877 | * | |
1878 | * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 | |
1879 | * (which makes defrag_ratio = 1000) then every (well almost) | |
1880 | * allocation will first attempt to defrag slab caches on other nodes. | |
1881 | * This means scanning over all nodes to look for partial slabs which | |
1882 | * may be expensive if we do it every time we are trying to find a slab | |
1883 | * with available objects. | |
1884 | */ | |
1885 | if (!s->remote_node_defrag_ratio || | |
1886 | get_cycles() % 1024 > s->remote_node_defrag_ratio) | |
1887 | return NULL; | |
1888 | ||
1889 | do { | |
1890 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1891 | zonelist = node_zonelist(mempolicy_slab_node(), flags); | |
1892 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { | |
1893 | struct kmem_cache_node *n; | |
1894 | ||
1895 | n = get_node(s, zone_to_nid(zone)); | |
1896 | ||
1897 | if (n && cpuset_zone_allowed(zone, flags) && | |
1898 | n->nr_partial > s->min_partial) { | |
1899 | object = get_partial_node(s, n, c, flags); | |
1900 | if (object) { | |
1901 | /* | |
1902 | * Don't check read_mems_allowed_retry() | |
1903 | * here - if mems_allowed was updated in | |
1904 | * parallel, that was a harmless race | |
1905 | * between allocation and the cpuset | |
1906 | * update | |
1907 | */ | |
1908 | return object; | |
1909 | } | |
1910 | } | |
1911 | } | |
1912 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); | |
1913 | #endif | |
1914 | return NULL; | |
1915 | } | |
1916 | ||
1917 | /* | |
1918 | * Get a partial page, lock it and return it. | |
1919 | */ | |
1920 | static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, | |
1921 | struct kmem_cache_cpu *c) | |
1922 | { | |
1923 | void *object; | |
1924 | int searchnode = node; | |
1925 | ||
1926 | if (node == NUMA_NO_NODE) | |
1927 | searchnode = numa_mem_id(); | |
1928 | else if (!node_present_pages(node)) | |
1929 | searchnode = node_to_mem_node(node); | |
1930 | ||
1931 | object = get_partial_node(s, get_node(s, searchnode), c, flags); | |
1932 | if (object || node != NUMA_NO_NODE) | |
1933 | return object; | |
1934 | ||
1935 | return get_any_partial(s, flags, c); | |
1936 | } | |
1937 | ||
1938 | #ifdef CONFIG_PREEMPT | |
1939 | /* | |
1940 | * Calculate the next globally unique transaction for disambiguiation | |
1941 | * during cmpxchg. The transactions start with the cpu number and are then | |
1942 | * incremented by CONFIG_NR_CPUS. | |
1943 | */ | |
1944 | #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) | |
1945 | #else | |
1946 | /* | |
1947 | * No preemption supported therefore also no need to check for | |
1948 | * different cpus. | |
1949 | */ | |
1950 | #define TID_STEP 1 | |
1951 | #endif | |
1952 | ||
1953 | static inline unsigned long next_tid(unsigned long tid) | |
1954 | { | |
1955 | return tid + TID_STEP; | |
1956 | } | |
1957 | ||
1958 | static inline unsigned int tid_to_cpu(unsigned long tid) | |
1959 | { | |
1960 | return tid % TID_STEP; | |
1961 | } | |
1962 | ||
1963 | static inline unsigned long tid_to_event(unsigned long tid) | |
1964 | { | |
1965 | return tid / TID_STEP; | |
1966 | } | |
1967 | ||
1968 | static inline unsigned int init_tid(int cpu) | |
1969 | { | |
1970 | return cpu; | |
1971 | } | |
1972 | ||
1973 | static inline void note_cmpxchg_failure(const char *n, | |
1974 | const struct kmem_cache *s, unsigned long tid) | |
1975 | { | |
1976 | #ifdef SLUB_DEBUG_CMPXCHG | |
1977 | unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); | |
1978 | ||
1979 | pr_info("%s %s: cmpxchg redo ", n, s->name); | |
1980 | ||
1981 | #ifdef CONFIG_PREEMPT | |
1982 | if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) | |
1983 | pr_warn("due to cpu change %d -> %d\n", | |
1984 | tid_to_cpu(tid), tid_to_cpu(actual_tid)); | |
1985 | else | |
1986 | #endif | |
1987 | if (tid_to_event(tid) != tid_to_event(actual_tid)) | |
1988 | pr_warn("due to cpu running other code. Event %ld->%ld\n", | |
1989 | tid_to_event(tid), tid_to_event(actual_tid)); | |
1990 | else | |
1991 | pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n", | |
1992 | actual_tid, tid, next_tid(tid)); | |
1993 | #endif | |
1994 | stat(s, CMPXCHG_DOUBLE_CPU_FAIL); | |
1995 | } | |
1996 | ||
1997 | static void init_kmem_cache_cpus(struct kmem_cache *s) | |
1998 | { | |
1999 | int cpu; | |
2000 | ||
2001 | for_each_possible_cpu(cpu) | |
2002 | per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); | |
2003 | } | |
2004 | ||
2005 | /* | |
2006 | * Remove the cpu slab | |
2007 | */ | |
2008 | static void deactivate_slab(struct kmem_cache *s, struct page *page, | |
2009 | void *freelist, struct kmem_cache_cpu *c) | |
2010 | { | |
2011 | enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; | |
2012 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
2013 | int lock = 0; | |
2014 | enum slab_modes l = M_NONE, m = M_NONE; | |
2015 | void *nextfree; | |
2016 | int tail = DEACTIVATE_TO_HEAD; | |
2017 | struct page new; | |
2018 | struct page old; | |
2019 | ||
2020 | if (page->freelist) { | |
2021 | stat(s, DEACTIVATE_REMOTE_FREES); | |
2022 | tail = DEACTIVATE_TO_TAIL; | |
2023 | } | |
2024 | ||
2025 | /* | |
2026 | * Stage one: Free all available per cpu objects back | |
2027 | * to the page freelist while it is still frozen. Leave the | |
2028 | * last one. | |
2029 | * | |
2030 | * There is no need to take the list->lock because the page | |
2031 | * is still frozen. | |
2032 | */ | |
2033 | while (freelist && (nextfree = get_freepointer(s, freelist))) { | |
2034 | void *prior; | |
2035 | unsigned long counters; | |
2036 | ||
2037 | do { | |
2038 | prior = page->freelist; | |
2039 | counters = page->counters; | |
2040 | set_freepointer(s, freelist, prior); | |
2041 | new.counters = counters; | |
2042 | new.inuse--; | |
2043 | VM_BUG_ON(!new.frozen); | |
2044 | ||
2045 | } while (!__cmpxchg_double_slab(s, page, | |
2046 | prior, counters, | |
2047 | freelist, new.counters, | |
2048 | "drain percpu freelist")); | |
2049 | ||
2050 | freelist = nextfree; | |
2051 | } | |
2052 | ||
2053 | /* | |
2054 | * Stage two: Ensure that the page is unfrozen while the | |
2055 | * list presence reflects the actual number of objects | |
2056 | * during unfreeze. | |
2057 | * | |
2058 | * We setup the list membership and then perform a cmpxchg | |
2059 | * with the count. If there is a mismatch then the page | |
2060 | * is not unfrozen but the page is on the wrong list. | |
2061 | * | |
2062 | * Then we restart the process which may have to remove | |
2063 | * the page from the list that we just put it on again | |
2064 | * because the number of objects in the slab may have | |
2065 | * changed. | |
2066 | */ | |
2067 | redo: | |
2068 | ||
2069 | old.freelist = page->freelist; | |
2070 | old.counters = page->counters; | |
2071 | VM_BUG_ON(!old.frozen); | |
2072 | ||
2073 | /* Determine target state of the slab */ | |
2074 | new.counters = old.counters; | |
2075 | if (freelist) { | |
2076 | new.inuse--; | |
2077 | set_freepointer(s, freelist, old.freelist); | |
2078 | new.freelist = freelist; | |
2079 | } else | |
2080 | new.freelist = old.freelist; | |
2081 | ||
2082 | new.frozen = 0; | |
2083 | ||
2084 | if (!new.inuse && n->nr_partial >= s->min_partial) | |
2085 | m = M_FREE; | |
2086 | else if (new.freelist) { | |
2087 | m = M_PARTIAL; | |
2088 | if (!lock) { | |
2089 | lock = 1; | |
2090 | /* | |
2091 | * Taking the spinlock removes the possiblity | |
2092 | * that acquire_slab() will see a slab page that | |
2093 | * is frozen | |
2094 | */ | |
2095 | spin_lock(&n->list_lock); | |
2096 | } | |
2097 | } else { | |
2098 | m = M_FULL; | |
2099 | if (kmem_cache_debug(s) && !lock) { | |
2100 | lock = 1; | |
2101 | /* | |
2102 | * This also ensures that the scanning of full | |
2103 | * slabs from diagnostic functions will not see | |
2104 | * any frozen slabs. | |
2105 | */ | |
2106 | spin_lock(&n->list_lock); | |
2107 | } | |
2108 | } | |
2109 | ||
2110 | if (l != m) { | |
2111 | ||
2112 | if (l == M_PARTIAL) | |
2113 | ||
2114 | remove_partial(n, page); | |
2115 | ||
2116 | else if (l == M_FULL) | |
2117 | ||
2118 | remove_full(s, n, page); | |
2119 | ||
2120 | if (m == M_PARTIAL) { | |
2121 | ||
2122 | add_partial(n, page, tail); | |
2123 | stat(s, tail); | |
2124 | ||
2125 | } else if (m == M_FULL) { | |
2126 | ||
2127 | stat(s, DEACTIVATE_FULL); | |
2128 | add_full(s, n, page); | |
2129 | ||
2130 | } | |
2131 | } | |
2132 | ||
2133 | l = m; | |
2134 | if (!__cmpxchg_double_slab(s, page, | |
2135 | old.freelist, old.counters, | |
2136 | new.freelist, new.counters, | |
2137 | "unfreezing slab")) | |
2138 | goto redo; | |
2139 | ||
2140 | if (lock) | |
2141 | spin_unlock(&n->list_lock); | |
2142 | ||
2143 | if (m == M_FREE) { | |
2144 | stat(s, DEACTIVATE_EMPTY); | |
2145 | discard_slab(s, page); | |
2146 | stat(s, FREE_SLAB); | |
2147 | } | |
2148 | ||
2149 | c->page = NULL; | |
2150 | c->freelist = NULL; | |
2151 | } | |
2152 | ||
2153 | /* | |
2154 | * Unfreeze all the cpu partial slabs. | |
2155 | * | |
2156 | * This function must be called with interrupts disabled | |
2157 | * for the cpu using c (or some other guarantee must be there | |
2158 | * to guarantee no concurrent accesses). | |
2159 | */ | |
2160 | static void unfreeze_partials(struct kmem_cache *s, | |
2161 | struct kmem_cache_cpu *c) | |
2162 | { | |
2163 | #ifdef CONFIG_SLUB_CPU_PARTIAL | |
2164 | struct kmem_cache_node *n = NULL, *n2 = NULL; | |
2165 | struct page *page, *discard_page = NULL; | |
2166 | ||
2167 | while ((page = c->partial)) { | |
2168 | struct page new; | |
2169 | struct page old; | |
2170 | ||
2171 | c->partial = page->next; | |
2172 | ||
2173 | n2 = get_node(s, page_to_nid(page)); | |
2174 | if (n != n2) { | |
2175 | if (n) | |
2176 | spin_unlock(&n->list_lock); | |
2177 | ||
2178 | n = n2; | |
2179 | spin_lock(&n->list_lock); | |
2180 | } | |
2181 | ||
2182 | do { | |
2183 | ||
2184 | old.freelist = page->freelist; | |
2185 | old.counters = page->counters; | |
2186 | VM_BUG_ON(!old.frozen); | |
2187 | ||
2188 | new.counters = old.counters; | |
2189 | new.freelist = old.freelist; | |
2190 | ||
2191 | new.frozen = 0; | |
2192 | ||
2193 | } while (!__cmpxchg_double_slab(s, page, | |
2194 | old.freelist, old.counters, | |
2195 | new.freelist, new.counters, | |
2196 | "unfreezing slab")); | |
2197 | ||
2198 | if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) { | |
2199 | page->next = discard_page; | |
2200 | discard_page = page; | |
2201 | } else { | |
2202 | add_partial(n, page, DEACTIVATE_TO_TAIL); | |
2203 | stat(s, FREE_ADD_PARTIAL); | |
2204 | } | |
2205 | } | |
2206 | ||
2207 | if (n) | |
2208 | spin_unlock(&n->list_lock); | |
2209 | ||
2210 | while (discard_page) { | |
2211 | page = discard_page; | |
2212 | discard_page = discard_page->next; | |
2213 | ||
2214 | stat(s, DEACTIVATE_EMPTY); | |
2215 | discard_slab(s, page); | |
2216 | stat(s, FREE_SLAB); | |
2217 | } | |
2218 | #endif | |
2219 | } | |
2220 | ||
2221 | /* | |
2222 | * Put a page that was just frozen (in __slab_free) into a partial page | |
2223 | * slot if available. This is done without interrupts disabled and without | |
2224 | * preemption disabled. The cmpxchg is racy and may put the partial page | |
2225 | * onto a random cpus partial slot. | |
2226 | * | |
2227 | * If we did not find a slot then simply move all the partials to the | |
2228 | * per node partial list. | |
2229 | */ | |
2230 | static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) | |
2231 | { | |
2232 | #ifdef CONFIG_SLUB_CPU_PARTIAL | |
2233 | struct page *oldpage; | |
2234 | int pages; | |
2235 | int pobjects; | |
2236 | ||
2237 | preempt_disable(); | |
2238 | do { | |
2239 | pages = 0; | |
2240 | pobjects = 0; | |
2241 | oldpage = this_cpu_read(s->cpu_slab->partial); | |
2242 | ||
2243 | if (oldpage) { | |
2244 | pobjects = oldpage->pobjects; | |
2245 | pages = oldpage->pages; | |
2246 | if (drain && pobjects > s->cpu_partial) { | |
2247 | unsigned long flags; | |
2248 | /* | |
2249 | * partial array is full. Move the existing | |
2250 | * set to the per node partial list. | |
2251 | */ | |
2252 | local_irq_save(flags); | |
2253 | unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); | |
2254 | local_irq_restore(flags); | |
2255 | oldpage = NULL; | |
2256 | pobjects = 0; | |
2257 | pages = 0; | |
2258 | stat(s, CPU_PARTIAL_DRAIN); | |
2259 | } | |
2260 | } | |
2261 | ||
2262 | pages++; | |
2263 | pobjects += page->objects - page->inuse; | |
2264 | ||
2265 | page->pages = pages; | |
2266 | page->pobjects = pobjects; | |
2267 | page->next = oldpage; | |
2268 | ||
2269 | } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) | |
2270 | != oldpage); | |
2271 | if (unlikely(!s->cpu_partial)) { | |
2272 | unsigned long flags; | |
2273 | ||
2274 | local_irq_save(flags); | |
2275 | unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); | |
2276 | local_irq_restore(flags); | |
2277 | } | |
2278 | preempt_enable(); | |
2279 | #endif | |
2280 | } | |
2281 | ||
2282 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) | |
2283 | { | |
2284 | stat(s, CPUSLAB_FLUSH); | |
2285 | deactivate_slab(s, c->page, c->freelist, c); | |
2286 | ||
2287 | c->tid = next_tid(c->tid); | |
2288 | } | |
2289 | ||
2290 | /* | |
2291 | * Flush cpu slab. | |
2292 | * | |
2293 | * Called from IPI handler with interrupts disabled. | |
2294 | */ | |
2295 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) | |
2296 | { | |
2297 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); | |
2298 | ||
2299 | if (likely(c)) { | |
2300 | if (c->page) | |
2301 | flush_slab(s, c); | |
2302 | ||
2303 | unfreeze_partials(s, c); | |
2304 | } | |
2305 | } | |
2306 | ||
2307 | static void flush_cpu_slab(void *d) | |
2308 | { | |
2309 | struct kmem_cache *s = d; | |
2310 | ||
2311 | __flush_cpu_slab(s, smp_processor_id()); | |
2312 | } | |
2313 | ||
2314 | static bool has_cpu_slab(int cpu, void *info) | |
2315 | { | |
2316 | struct kmem_cache *s = info; | |
2317 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); | |
2318 | ||
2319 | return c->page || slub_percpu_partial(c); | |
2320 | } | |
2321 | ||
2322 | static void flush_all(struct kmem_cache *s) | |
2323 | { | |
2324 | on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC); | |
2325 | } | |
2326 | ||
2327 | /* | |
2328 | * Use the cpu notifier to insure that the cpu slabs are flushed when | |
2329 | * necessary. | |
2330 | */ | |
2331 | static int slub_cpu_dead(unsigned int cpu) | |
2332 | { | |
2333 | struct kmem_cache *s; | |
2334 | unsigned long flags; | |
2335 | ||
2336 | mutex_lock(&slab_mutex); | |
2337 | list_for_each_entry(s, &slab_caches, list) { | |
2338 | local_irq_save(flags); | |
2339 | __flush_cpu_slab(s, cpu); | |
2340 | local_irq_restore(flags); | |
2341 | } | |
2342 | mutex_unlock(&slab_mutex); | |
2343 | return 0; | |
2344 | } | |
2345 | ||
2346 | /* | |
2347 | * Check if the objects in a per cpu structure fit numa | |
2348 | * locality expectations. | |
2349 | */ | |
2350 | static inline int node_match(struct page *page, int node) | |
2351 | { | |
2352 | #ifdef CONFIG_NUMA | |
2353 | if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node)) | |
2354 | return 0; | |
2355 | #endif | |
2356 | return 1; | |
2357 | } | |
2358 | ||
2359 | #ifdef CONFIG_SLUB_DEBUG | |
2360 | static int count_free(struct page *page) | |
2361 | { | |
2362 | return page->objects - page->inuse; | |
2363 | } | |
2364 | ||
2365 | static inline unsigned long node_nr_objs(struct kmem_cache_node *n) | |
2366 | { | |
2367 | return atomic_long_read(&n->total_objects); | |
2368 | } | |
2369 | #endif /* CONFIG_SLUB_DEBUG */ | |
2370 | ||
2371 | #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS) | |
2372 | static unsigned long count_partial(struct kmem_cache_node *n, | |
2373 | int (*get_count)(struct page *)) | |
2374 | { | |
2375 | unsigned long flags; | |
2376 | unsigned long x = 0; | |
2377 | struct page *page; | |
2378 | ||
2379 | spin_lock_irqsave(&n->list_lock, flags); | |
2380 | list_for_each_entry(page, &n->partial, lru) | |
2381 | x += get_count(page); | |
2382 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2383 | return x; | |
2384 | } | |
2385 | #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */ | |
2386 | ||
2387 | static noinline void | |
2388 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) | |
2389 | { | |
2390 | #ifdef CONFIG_SLUB_DEBUG | |
2391 | static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, | |
2392 | DEFAULT_RATELIMIT_BURST); | |
2393 | int node; | |
2394 | struct kmem_cache_node *n; | |
2395 | ||
2396 | if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) | |
2397 | return; | |
2398 | ||
2399 | pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", | |
2400 | nid, gfpflags, &gfpflags); | |
2401 | pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n", | |
2402 | s->name, s->object_size, s->size, oo_order(s->oo), | |
2403 | oo_order(s->min)); | |
2404 | ||
2405 | if (oo_order(s->min) > get_order(s->object_size)) | |
2406 | pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n", | |
2407 | s->name); | |
2408 | ||
2409 | for_each_kmem_cache_node(s, node, n) { | |
2410 | unsigned long nr_slabs; | |
2411 | unsigned long nr_objs; | |
2412 | unsigned long nr_free; | |
2413 | ||
2414 | nr_free = count_partial(n, count_free); | |
2415 | nr_slabs = node_nr_slabs(n); | |
2416 | nr_objs = node_nr_objs(n); | |
2417 | ||
2418 | pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n", | |
2419 | node, nr_slabs, nr_objs, nr_free); | |
2420 | } | |
2421 | #endif | |
2422 | } | |
2423 | ||
2424 | static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, | |
2425 | int node, struct kmem_cache_cpu **pc) | |
2426 | { | |
2427 | void *freelist; | |
2428 | struct kmem_cache_cpu *c = *pc; | |
2429 | struct page *page; | |
2430 | ||
2431 | freelist = get_partial(s, flags, node, c); | |
2432 | ||
2433 | if (freelist) | |
2434 | return freelist; | |
2435 | ||
2436 | page = new_slab(s, flags, node); | |
2437 | if (page) { | |
2438 | c = raw_cpu_ptr(s->cpu_slab); | |
2439 | if (c->page) | |
2440 | flush_slab(s, c); | |
2441 | ||
2442 | /* | |
2443 | * No other reference to the page yet so we can | |
2444 | * muck around with it freely without cmpxchg | |
2445 | */ | |
2446 | freelist = page->freelist; | |
2447 | page->freelist = NULL; | |
2448 | ||
2449 | stat(s, ALLOC_SLAB); | |
2450 | c->page = page; | |
2451 | *pc = c; | |
2452 | } else | |
2453 | freelist = NULL; | |
2454 | ||
2455 | return freelist; | |
2456 | } | |
2457 | ||
2458 | static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags) | |
2459 | { | |
2460 | if (unlikely(PageSlabPfmemalloc(page))) | |
2461 | return gfp_pfmemalloc_allowed(gfpflags); | |
2462 | ||
2463 | return true; | |
2464 | } | |
2465 | ||
2466 | /* | |
2467 | * Check the page->freelist of a page and either transfer the freelist to the | |
2468 | * per cpu freelist or deactivate the page. | |
2469 | * | |
2470 | * The page is still frozen if the return value is not NULL. | |
2471 | * | |
2472 | * If this function returns NULL then the page has been unfrozen. | |
2473 | * | |
2474 | * This function must be called with interrupt disabled. | |
2475 | */ | |
2476 | static inline void *get_freelist(struct kmem_cache *s, struct page *page) | |
2477 | { | |
2478 | struct page new; | |
2479 | unsigned long counters; | |
2480 | void *freelist; | |
2481 | ||
2482 | do { | |
2483 | freelist = page->freelist; | |
2484 | counters = page->counters; | |
2485 | ||
2486 | new.counters = counters; | |
2487 | VM_BUG_ON(!new.frozen); | |
2488 | ||
2489 | new.inuse = page->objects; | |
2490 | new.frozen = freelist != NULL; | |
2491 | ||
2492 | } while (!__cmpxchg_double_slab(s, page, | |
2493 | freelist, counters, | |
2494 | NULL, new.counters, | |
2495 | "get_freelist")); | |
2496 | ||
2497 | return freelist; | |
2498 | } | |
2499 | ||
2500 | /* | |
2501 | * Slow path. The lockless freelist is empty or we need to perform | |
2502 | * debugging duties. | |
2503 | * | |
2504 | * Processing is still very fast if new objects have been freed to the | |
2505 | * regular freelist. In that case we simply take over the regular freelist | |
2506 | * as the lockless freelist and zap the regular freelist. | |
2507 | * | |
2508 | * If that is not working then we fall back to the partial lists. We take the | |
2509 | * first element of the freelist as the object to allocate now and move the | |
2510 | * rest of the freelist to the lockless freelist. | |
2511 | * | |
2512 | * And if we were unable to get a new slab from the partial slab lists then | |
2513 | * we need to allocate a new slab. This is the slowest path since it involves | |
2514 | * a call to the page allocator and the setup of a new slab. | |
2515 | * | |
2516 | * Version of __slab_alloc to use when we know that interrupts are | |
2517 | * already disabled (which is the case for bulk allocation). | |
2518 | */ | |
2519 | static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, | |
2520 | unsigned long addr, struct kmem_cache_cpu *c) | |
2521 | { | |
2522 | void *freelist; | |
2523 | struct page *page; | |
2524 | ||
2525 | page = c->page; | |
2526 | if (!page) | |
2527 | goto new_slab; | |
2528 | redo: | |
2529 | ||
2530 | if (unlikely(!node_match(page, node))) { | |
2531 | int searchnode = node; | |
2532 | ||
2533 | if (node != NUMA_NO_NODE && !node_present_pages(node)) | |
2534 | searchnode = node_to_mem_node(node); | |
2535 | ||
2536 | if (unlikely(!node_match(page, searchnode))) { | |
2537 | stat(s, ALLOC_NODE_MISMATCH); | |
2538 | deactivate_slab(s, page, c->freelist, c); | |
2539 | goto new_slab; | |
2540 | } | |
2541 | } | |
2542 | ||
2543 | /* | |
2544 | * By rights, we should be searching for a slab page that was | |
2545 | * PFMEMALLOC but right now, we are losing the pfmemalloc | |
2546 | * information when the page leaves the per-cpu allocator | |
2547 | */ | |
2548 | if (unlikely(!pfmemalloc_match(page, gfpflags))) { | |
2549 | deactivate_slab(s, page, c->freelist, c); | |
2550 | goto new_slab; | |
2551 | } | |
2552 | ||
2553 | /* must check again c->freelist in case of cpu migration or IRQ */ | |
2554 | freelist = c->freelist; | |
2555 | if (freelist) | |
2556 | goto load_freelist; | |
2557 | ||
2558 | freelist = get_freelist(s, page); | |
2559 | ||
2560 | if (!freelist) { | |
2561 | c->page = NULL; | |
2562 | stat(s, DEACTIVATE_BYPASS); | |
2563 | goto new_slab; | |
2564 | } | |
2565 | ||
2566 | stat(s, ALLOC_REFILL); | |
2567 | ||
2568 | load_freelist: | |
2569 | /* | |
2570 | * freelist is pointing to the list of objects to be used. | |
2571 | * page is pointing to the page from which the objects are obtained. | |
2572 | * That page must be frozen for per cpu allocations to work. | |
2573 | */ | |
2574 | VM_BUG_ON(!c->page->frozen); | |
2575 | c->freelist = get_freepointer(s, freelist); | |
2576 | c->tid = next_tid(c->tid); | |
2577 | return freelist; | |
2578 | ||
2579 | new_slab: | |
2580 | ||
2581 | if (slub_percpu_partial(c)) { | |
2582 | page = c->page = slub_percpu_partial(c); | |
2583 | slub_set_percpu_partial(c, page); | |
2584 | stat(s, CPU_PARTIAL_ALLOC); | |
2585 | goto redo; | |
2586 | } | |
2587 | ||
2588 | freelist = new_slab_objects(s, gfpflags, node, &c); | |
2589 | ||
2590 | if (unlikely(!freelist)) { | |
2591 | slab_out_of_memory(s, gfpflags, node); | |
2592 | return NULL; | |
2593 | } | |
2594 | ||
2595 | page = c->page; | |
2596 | if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags))) | |
2597 | goto load_freelist; | |
2598 | ||
2599 | /* Only entered in the debug case */ | |
2600 | if (kmem_cache_debug(s) && | |
2601 | !alloc_debug_processing(s, page, freelist, addr)) | |
2602 | goto new_slab; /* Slab failed checks. Next slab needed */ | |
2603 | ||
2604 | deactivate_slab(s, page, get_freepointer(s, freelist), c); | |
2605 | return freelist; | |
2606 | } | |
2607 | ||
2608 | /* | |
2609 | * Another one that disabled interrupt and compensates for possible | |
2610 | * cpu changes by refetching the per cpu area pointer. | |
2611 | */ | |
2612 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, | |
2613 | unsigned long addr, struct kmem_cache_cpu *c) | |
2614 | { | |
2615 | void *p; | |
2616 | unsigned long flags; | |
2617 | ||
2618 | local_irq_save(flags); | |
2619 | #ifdef CONFIG_PREEMPT | |
2620 | /* | |
2621 | * We may have been preempted and rescheduled on a different | |
2622 | * cpu before disabling interrupts. Need to reload cpu area | |
2623 | * pointer. | |
2624 | */ | |
2625 | c = this_cpu_ptr(s->cpu_slab); | |
2626 | #endif | |
2627 | ||
2628 | p = ___slab_alloc(s, gfpflags, node, addr, c); | |
2629 | local_irq_restore(flags); | |
2630 | return p; | |
2631 | } | |
2632 | ||
2633 | /* | |
2634 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) | |
2635 | * have the fastpath folded into their functions. So no function call | |
2636 | * overhead for requests that can be satisfied on the fastpath. | |
2637 | * | |
2638 | * The fastpath works by first checking if the lockless freelist can be used. | |
2639 | * If not then __slab_alloc is called for slow processing. | |
2640 | * | |
2641 | * Otherwise we can simply pick the next object from the lockless free list. | |
2642 | */ | |
2643 | static __always_inline void *slab_alloc_node(struct kmem_cache *s, | |
2644 | gfp_t gfpflags, int node, unsigned long addr) | |
2645 | { | |
2646 | void *object; | |
2647 | struct kmem_cache_cpu *c; | |
2648 | struct page *page; | |
2649 | unsigned long tid; | |
2650 | ||
2651 | s = slab_pre_alloc_hook(s, gfpflags); | |
2652 | if (!s) | |
2653 | return NULL; | |
2654 | redo: | |
2655 | /* | |
2656 | * Must read kmem_cache cpu data via this cpu ptr. Preemption is | |
2657 | * enabled. We may switch back and forth between cpus while | |
2658 | * reading from one cpu area. That does not matter as long | |
2659 | * as we end up on the original cpu again when doing the cmpxchg. | |
2660 | * | |
2661 | * We should guarantee that tid and kmem_cache are retrieved on | |
2662 | * the same cpu. It could be different if CONFIG_PREEMPT so we need | |
2663 | * to check if it is matched or not. | |
2664 | */ | |
2665 | do { | |
2666 | tid = this_cpu_read(s->cpu_slab->tid); | |
2667 | c = raw_cpu_ptr(s->cpu_slab); | |
2668 | } while (IS_ENABLED(CONFIG_PREEMPT) && | |
2669 | unlikely(tid != READ_ONCE(c->tid))); | |
2670 | ||
2671 | /* | |
2672 | * Irqless object alloc/free algorithm used here depends on sequence | |
2673 | * of fetching cpu_slab's data. tid should be fetched before anything | |
2674 | * on c to guarantee that object and page associated with previous tid | |
2675 | * won't be used with current tid. If we fetch tid first, object and | |
2676 | * page could be one associated with next tid and our alloc/free | |
2677 | * request will be failed. In this case, we will retry. So, no problem. | |
2678 | */ | |
2679 | barrier(); | |
2680 | ||
2681 | /* | |
2682 | * The transaction ids are globally unique per cpu and per operation on | |
2683 | * a per cpu queue. Thus they can be guarantee that the cmpxchg_double | |
2684 | * occurs on the right processor and that there was no operation on the | |
2685 | * linked list in between. | |
2686 | */ | |
2687 | ||
2688 | object = c->freelist; | |
2689 | page = c->page; | |
2690 | if (unlikely(!object || !node_match(page, node))) { | |
2691 | object = __slab_alloc(s, gfpflags, node, addr, c); | |
2692 | stat(s, ALLOC_SLOWPATH); | |
2693 | } else { | |
2694 | void *next_object = get_freepointer_safe(s, object); | |
2695 | ||
2696 | /* | |
2697 | * The cmpxchg will only match if there was no additional | |
2698 | * operation and if we are on the right processor. | |
2699 | * | |
2700 | * The cmpxchg does the following atomically (without lock | |
2701 | * semantics!) | |
2702 | * 1. Relocate first pointer to the current per cpu area. | |
2703 | * 2. Verify that tid and freelist have not been changed | |
2704 | * 3. If they were not changed replace tid and freelist | |
2705 | * | |
2706 | * Since this is without lock semantics the protection is only | |
2707 | * against code executing on this cpu *not* from access by | |
2708 | * other cpus. | |
2709 | */ | |
2710 | if (unlikely(!this_cpu_cmpxchg_double( | |
2711 | s->cpu_slab->freelist, s->cpu_slab->tid, | |
2712 | object, tid, | |
2713 | next_object, next_tid(tid)))) { | |
2714 | ||
2715 | note_cmpxchg_failure("slab_alloc", s, tid); | |
2716 | goto redo; | |
2717 | } | |
2718 | prefetch_freepointer(s, next_object); | |
2719 | stat(s, ALLOC_FASTPATH); | |
2720 | } | |
2721 | ||
2722 | if (unlikely(gfpflags & __GFP_ZERO) && object) | |
2723 | memset(object, 0, s->object_size); | |
2724 | ||
2725 | slab_post_alloc_hook(s, gfpflags, 1, &object); | |
2726 | ||
2727 | return object; | |
2728 | } | |
2729 | ||
2730 | static __always_inline void *slab_alloc(struct kmem_cache *s, | |
2731 | gfp_t gfpflags, unsigned long addr) | |
2732 | { | |
2733 | return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr); | |
2734 | } | |
2735 | ||
2736 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
2737 | { | |
2738 | void *ret = slab_alloc(s, gfpflags, _RET_IP_); | |
2739 | ||
2740 | trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, | |
2741 | s->size, gfpflags); | |
2742 | ||
2743 | return ret; | |
2744 | } | |
2745 | EXPORT_SYMBOL(kmem_cache_alloc); | |
2746 | ||
2747 | #ifdef CONFIG_TRACING | |
2748 | void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) | |
2749 | { | |
2750 | void *ret = slab_alloc(s, gfpflags, _RET_IP_); | |
2751 | trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); | |
2752 | kasan_kmalloc(s, ret, size, gfpflags); | |
2753 | return ret; | |
2754 | } | |
2755 | EXPORT_SYMBOL(kmem_cache_alloc_trace); | |
2756 | #endif | |
2757 | ||
2758 | #ifdef CONFIG_NUMA | |
2759 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
2760 | { | |
2761 | void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); | |
2762 | ||
2763 | trace_kmem_cache_alloc_node(_RET_IP_, ret, | |
2764 | s->object_size, s->size, gfpflags, node); | |
2765 | ||
2766 | return ret; | |
2767 | } | |
2768 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
2769 | ||
2770 | #ifdef CONFIG_TRACING | |
2771 | void *kmem_cache_alloc_node_trace(struct kmem_cache *s, | |
2772 | gfp_t gfpflags, | |
2773 | int node, size_t size) | |
2774 | { | |
2775 | void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); | |
2776 | ||
2777 | trace_kmalloc_node(_RET_IP_, ret, | |
2778 | size, s->size, gfpflags, node); | |
2779 | ||
2780 | kasan_kmalloc(s, ret, size, gfpflags); | |
2781 | return ret; | |
2782 | } | |
2783 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); | |
2784 | #endif | |
2785 | #endif | |
2786 | ||
2787 | /* | |
2788 | * Slow path handling. This may still be called frequently since objects | |
2789 | * have a longer lifetime than the cpu slabs in most processing loads. | |
2790 | * | |
2791 | * So we still attempt to reduce cache line usage. Just take the slab | |
2792 | * lock and free the item. If there is no additional partial page | |
2793 | * handling required then we can return immediately. | |
2794 | */ | |
2795 | static void __slab_free(struct kmem_cache *s, struct page *page, | |
2796 | void *head, void *tail, int cnt, | |
2797 | unsigned long addr) | |
2798 | ||
2799 | { | |
2800 | void *prior; | |
2801 | int was_frozen; | |
2802 | struct page new; | |
2803 | unsigned long counters; | |
2804 | struct kmem_cache_node *n = NULL; | |
2805 | unsigned long uninitialized_var(flags); | |
2806 | ||
2807 | stat(s, FREE_SLOWPATH); | |
2808 | ||
2809 | if (kmem_cache_debug(s) && | |
2810 | !free_debug_processing(s, page, head, tail, cnt, addr)) | |
2811 | return; | |
2812 | ||
2813 | do { | |
2814 | if (unlikely(n)) { | |
2815 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2816 | n = NULL; | |
2817 | } | |
2818 | prior = page->freelist; | |
2819 | counters = page->counters; | |
2820 | set_freepointer(s, tail, prior); | |
2821 | new.counters = counters; | |
2822 | was_frozen = new.frozen; | |
2823 | new.inuse -= cnt; | |
2824 | if ((!new.inuse || !prior) && !was_frozen) { | |
2825 | ||
2826 | if (kmem_cache_has_cpu_partial(s) && !prior) { | |
2827 | ||
2828 | /* | |
2829 | * Slab was on no list before and will be | |
2830 | * partially empty | |
2831 | * We can defer the list move and instead | |
2832 | * freeze it. | |
2833 | */ | |
2834 | new.frozen = 1; | |
2835 | ||
2836 | } else { /* Needs to be taken off a list */ | |
2837 | ||
2838 | n = get_node(s, page_to_nid(page)); | |
2839 | /* | |
2840 | * Speculatively acquire the list_lock. | |
2841 | * If the cmpxchg does not succeed then we may | |
2842 | * drop the list_lock without any processing. | |
2843 | * | |
2844 | * Otherwise the list_lock will synchronize with | |
2845 | * other processors updating the list of slabs. | |
2846 | */ | |
2847 | spin_lock_irqsave(&n->list_lock, flags); | |
2848 | ||
2849 | } | |
2850 | } | |
2851 | ||
2852 | } while (!cmpxchg_double_slab(s, page, | |
2853 | prior, counters, | |
2854 | head, new.counters, | |
2855 | "__slab_free")); | |
2856 | ||
2857 | if (likely(!n)) { | |
2858 | ||
2859 | /* | |
2860 | * If we just froze the page then put it onto the | |
2861 | * per cpu partial list. | |
2862 | */ | |
2863 | if (new.frozen && !was_frozen) { | |
2864 | put_cpu_partial(s, page, 1); | |
2865 | stat(s, CPU_PARTIAL_FREE); | |
2866 | } | |
2867 | /* | |
2868 | * The list lock was not taken therefore no list | |
2869 | * activity can be necessary. | |
2870 | */ | |
2871 | if (was_frozen) | |
2872 | stat(s, FREE_FROZEN); | |
2873 | return; | |
2874 | } | |
2875 | ||
2876 | if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) | |
2877 | goto slab_empty; | |
2878 | ||
2879 | /* | |
2880 | * Objects left in the slab. If it was not on the partial list before | |
2881 | * then add it. | |
2882 | */ | |
2883 | if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { | |
2884 | if (kmem_cache_debug(s)) | |
2885 | remove_full(s, n, page); | |
2886 | add_partial(n, page, DEACTIVATE_TO_TAIL); | |
2887 | stat(s, FREE_ADD_PARTIAL); | |
2888 | } | |
2889 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2890 | return; | |
2891 | ||
2892 | slab_empty: | |
2893 | if (prior) { | |
2894 | /* | |
2895 | * Slab on the partial list. | |
2896 | */ | |
2897 | remove_partial(n, page); | |
2898 | stat(s, FREE_REMOVE_PARTIAL); | |
2899 | } else { | |
2900 | /* Slab must be on the full list */ | |
2901 | remove_full(s, n, page); | |
2902 | } | |
2903 | ||
2904 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2905 | stat(s, FREE_SLAB); | |
2906 | discard_slab(s, page); | |
2907 | } | |
2908 | ||
2909 | /* | |
2910 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that | |
2911 | * can perform fastpath freeing without additional function calls. | |
2912 | * | |
2913 | * The fastpath is only possible if we are freeing to the current cpu slab | |
2914 | * of this processor. This typically the case if we have just allocated | |
2915 | * the item before. | |
2916 | * | |
2917 | * If fastpath is not possible then fall back to __slab_free where we deal | |
2918 | * with all sorts of special processing. | |
2919 | * | |
2920 | * Bulk free of a freelist with several objects (all pointing to the | |
2921 | * same page) possible by specifying head and tail ptr, plus objects | |
2922 | * count (cnt). Bulk free indicated by tail pointer being set. | |
2923 | */ | |
2924 | static __always_inline void do_slab_free(struct kmem_cache *s, | |
2925 | struct page *page, void *head, void *tail, | |
2926 | int cnt, unsigned long addr) | |
2927 | { | |
2928 | void *tail_obj = tail ? : head; | |
2929 | struct kmem_cache_cpu *c; | |
2930 | unsigned long tid; | |
2931 | redo: | |
2932 | /* | |
2933 | * Determine the currently cpus per cpu slab. | |
2934 | * The cpu may change afterward. However that does not matter since | |
2935 | * data is retrieved via this pointer. If we are on the same cpu | |
2936 | * during the cmpxchg then the free will succeed. | |
2937 | */ | |
2938 | do { | |
2939 | tid = this_cpu_read(s->cpu_slab->tid); | |
2940 | c = raw_cpu_ptr(s->cpu_slab); | |
2941 | } while (IS_ENABLED(CONFIG_PREEMPT) && | |
2942 | unlikely(tid != READ_ONCE(c->tid))); | |
2943 | ||
2944 | /* Same with comment on barrier() in slab_alloc_node() */ | |
2945 | barrier(); | |
2946 | ||
2947 | if (likely(page == c->page)) { | |
2948 | set_freepointer(s, tail_obj, c->freelist); | |
2949 | ||
2950 | if (unlikely(!this_cpu_cmpxchg_double( | |
2951 | s->cpu_slab->freelist, s->cpu_slab->tid, | |
2952 | c->freelist, tid, | |
2953 | head, next_tid(tid)))) { | |
2954 | ||
2955 | note_cmpxchg_failure("slab_free", s, tid); | |
2956 | goto redo; | |
2957 | } | |
2958 | stat(s, FREE_FASTPATH); | |
2959 | } else | |
2960 | __slab_free(s, page, head, tail_obj, cnt, addr); | |
2961 | ||
2962 | } | |
2963 | ||
2964 | static __always_inline void slab_free(struct kmem_cache *s, struct page *page, | |
2965 | void *head, void *tail, int cnt, | |
2966 | unsigned long addr) | |
2967 | { | |
2968 | slab_free_freelist_hook(s, head, tail); | |
2969 | /* | |
2970 | * slab_free_freelist_hook() could have put the items into quarantine. | |
2971 | * If so, no need to free them. | |
2972 | */ | |
2973 | if (s->flags & SLAB_KASAN && !(s->flags & SLAB_TYPESAFE_BY_RCU)) | |
2974 | return; | |
2975 | do_slab_free(s, page, head, tail, cnt, addr); | |
2976 | } | |
2977 | ||
2978 | #ifdef CONFIG_KASAN | |
2979 | void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) | |
2980 | { | |
2981 | do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr); | |
2982 | } | |
2983 | #endif | |
2984 | ||
2985 | void kmem_cache_free(struct kmem_cache *s, void *x) | |
2986 | { | |
2987 | s = cache_from_obj(s, x); | |
2988 | if (!s) | |
2989 | return; | |
2990 | slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_); | |
2991 | trace_kmem_cache_free(_RET_IP_, x); | |
2992 | } | |
2993 | EXPORT_SYMBOL(kmem_cache_free); | |
2994 | ||
2995 | struct detached_freelist { | |
2996 | struct page *page; | |
2997 | void *tail; | |
2998 | void *freelist; | |
2999 | int cnt; | |
3000 | struct kmem_cache *s; | |
3001 | }; | |
3002 | ||
3003 | /* | |
3004 | * This function progressively scans the array with free objects (with | |
3005 | * a limited look ahead) and extract objects belonging to the same | |
3006 | * page. It builds a detached freelist directly within the given | |
3007 | * page/objects. This can happen without any need for | |
3008 | * synchronization, because the objects are owned by running process. | |
3009 | * The freelist is build up as a single linked list in the objects. | |
3010 | * The idea is, that this detached freelist can then be bulk | |
3011 | * transferred to the real freelist(s), but only requiring a single | |
3012 | * synchronization primitive. Look ahead in the array is limited due | |
3013 | * to performance reasons. | |
3014 | */ | |
3015 | static inline | |
3016 | int build_detached_freelist(struct kmem_cache *s, size_t size, | |
3017 | void **p, struct detached_freelist *df) | |
3018 | { | |
3019 | size_t first_skipped_index = 0; | |
3020 | int lookahead = 3; | |
3021 | void *object; | |
3022 | struct page *page; | |
3023 | ||
3024 | /* Always re-init detached_freelist */ | |
3025 | df->page = NULL; | |
3026 | ||
3027 | do { | |
3028 | object = p[--size]; | |
3029 | /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */ | |
3030 | } while (!object && size); | |
3031 | ||
3032 | if (!object) | |
3033 | return 0; | |
3034 | ||
3035 | page = virt_to_head_page(object); | |
3036 | if (!s) { | |
3037 | /* Handle kalloc'ed objects */ | |
3038 | if (unlikely(!PageSlab(page))) { | |
3039 | BUG_ON(!PageCompound(page)); | |
3040 | kfree_hook(object); | |
3041 | __free_pages(page, compound_order(page)); | |
3042 | p[size] = NULL; /* mark object processed */ | |
3043 | return size; | |
3044 | } | |
3045 | /* Derive kmem_cache from object */ | |
3046 | df->s = page->slab_cache; | |
3047 | } else { | |
3048 | df->s = cache_from_obj(s, object); /* Support for memcg */ | |
3049 | } | |
3050 | ||
3051 | /* Start new detached freelist */ | |
3052 | df->page = page; | |
3053 | set_freepointer(df->s, object, NULL); | |
3054 | df->tail = object; | |
3055 | df->freelist = object; | |
3056 | p[size] = NULL; /* mark object processed */ | |
3057 | df->cnt = 1; | |
3058 | ||
3059 | while (size) { | |
3060 | object = p[--size]; | |
3061 | if (!object) | |
3062 | continue; /* Skip processed objects */ | |
3063 | ||
3064 | /* df->page is always set at this point */ | |
3065 | if (df->page == virt_to_head_page(object)) { | |
3066 | /* Opportunity build freelist */ | |
3067 | set_freepointer(df->s, object, df->freelist); | |
3068 | df->freelist = object; | |
3069 | df->cnt++; | |
3070 | p[size] = NULL; /* mark object processed */ | |
3071 | ||
3072 | continue; | |
3073 | } | |
3074 | ||
3075 | /* Limit look ahead search */ | |
3076 | if (!--lookahead) | |
3077 | break; | |
3078 | ||
3079 | if (!first_skipped_index) | |
3080 | first_skipped_index = size + 1; | |
3081 | } | |
3082 | ||
3083 | return first_skipped_index; | |
3084 | } | |
3085 | ||
3086 | /* Note that interrupts must be enabled when calling this function. */ | |
3087 | void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) | |
3088 | { | |
3089 | if (WARN_ON(!size)) | |
3090 | return; | |
3091 | ||
3092 | do { | |
3093 | struct detached_freelist df; | |
3094 | ||
3095 | size = build_detached_freelist(s, size, p, &df); | |
3096 | if (!df.page) | |
3097 | continue; | |
3098 | ||
3099 | slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_); | |
3100 | } while (likely(size)); | |
3101 | } | |
3102 | EXPORT_SYMBOL(kmem_cache_free_bulk); | |
3103 | ||
3104 | /* Note that interrupts must be enabled when calling this function. */ | |
3105 | int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, | |
3106 | void **p) | |
3107 | { | |
3108 | struct kmem_cache_cpu *c; | |
3109 | int i; | |
3110 | ||
3111 | /* memcg and kmem_cache debug support */ | |
3112 | s = slab_pre_alloc_hook(s, flags); | |
3113 | if (unlikely(!s)) | |
3114 | return false; | |
3115 | /* | |
3116 | * Drain objects in the per cpu slab, while disabling local | |
3117 | * IRQs, which protects against PREEMPT and interrupts | |
3118 | * handlers invoking normal fastpath. | |
3119 | */ | |
3120 | local_irq_disable(); | |
3121 | c = this_cpu_ptr(s->cpu_slab); | |
3122 | ||
3123 | for (i = 0; i < size; i++) { | |
3124 | void *object = c->freelist; | |
3125 | ||
3126 | if (unlikely(!object)) { | |
3127 | /* | |
3128 | * Invoking slow path likely have side-effect | |
3129 | * of re-populating per CPU c->freelist | |
3130 | */ | |
3131 | p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, | |
3132 | _RET_IP_, c); | |
3133 | if (unlikely(!p[i])) | |
3134 | goto error; | |
3135 | ||
3136 | c = this_cpu_ptr(s->cpu_slab); | |
3137 | continue; /* goto for-loop */ | |
3138 | } | |
3139 | c->freelist = get_freepointer(s, object); | |
3140 | p[i] = object; | |
3141 | } | |
3142 | c->tid = next_tid(c->tid); | |
3143 | local_irq_enable(); | |
3144 | ||
3145 | /* Clear memory outside IRQ disabled fastpath loop */ | |
3146 | if (unlikely(flags & __GFP_ZERO)) { | |
3147 | int j; | |
3148 | ||
3149 | for (j = 0; j < i; j++) | |
3150 | memset(p[j], 0, s->object_size); | |
3151 | } | |
3152 | ||
3153 | /* memcg and kmem_cache debug support */ | |
3154 | slab_post_alloc_hook(s, flags, size, p); | |
3155 | return i; | |
3156 | error: | |
3157 | local_irq_enable(); | |
3158 | slab_post_alloc_hook(s, flags, i, p); | |
3159 | __kmem_cache_free_bulk(s, i, p); | |
3160 | return 0; | |
3161 | } | |
3162 | EXPORT_SYMBOL(kmem_cache_alloc_bulk); | |
3163 | ||
3164 | ||
3165 | /* | |
3166 | * Object placement in a slab is made very easy because we always start at | |
3167 | * offset 0. If we tune the size of the object to the alignment then we can | |
3168 | * get the required alignment by putting one properly sized object after | |
3169 | * another. | |
3170 | * | |
3171 | * Notice that the allocation order determines the sizes of the per cpu | |
3172 | * caches. Each processor has always one slab available for allocations. | |
3173 | * Increasing the allocation order reduces the number of times that slabs | |
3174 | * must be moved on and off the partial lists and is therefore a factor in | |
3175 | * locking overhead. | |
3176 | */ | |
3177 | ||
3178 | /* | |
3179 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
3180 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
3181 | * and increases the number of allocations possible without having to | |
3182 | * take the list_lock. | |
3183 | */ | |
3184 | static int slub_min_order; | |
3185 | static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; | |
3186 | static int slub_min_objects; | |
3187 | ||
3188 | /* | |
3189 | * Calculate the order of allocation given an slab object size. | |
3190 | * | |
3191 | * The order of allocation has significant impact on performance and other | |
3192 | * system components. Generally order 0 allocations should be preferred since | |
3193 | * order 0 does not cause fragmentation in the page allocator. Larger objects | |
3194 | * be problematic to put into order 0 slabs because there may be too much | |
3195 | * unused space left. We go to a higher order if more than 1/16th of the slab | |
3196 | * would be wasted. | |
3197 | * | |
3198 | * In order to reach satisfactory performance we must ensure that a minimum | |
3199 | * number of objects is in one slab. Otherwise we may generate too much | |
3200 | * activity on the partial lists which requires taking the list_lock. This is | |
3201 | * less a concern for large slabs though which are rarely used. | |
3202 | * | |
3203 | * slub_max_order specifies the order where we begin to stop considering the | |
3204 | * number of objects in a slab as critical. If we reach slub_max_order then | |
3205 | * we try to keep the page order as low as possible. So we accept more waste | |
3206 | * of space in favor of a small page order. | |
3207 | * | |
3208 | * Higher order allocations also allow the placement of more objects in a | |
3209 | * slab and thereby reduce object handling overhead. If the user has | |
3210 | * requested a higher mininum order then we start with that one instead of | |
3211 | * the smallest order which will fit the object. | |
3212 | */ | |
3213 | static inline int slab_order(int size, int min_objects, | |
3214 | int max_order, int fract_leftover, int reserved) | |
3215 | { | |
3216 | int order; | |
3217 | int rem; | |
3218 | int min_order = slub_min_order; | |
3219 | ||
3220 | if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) | |
3221 | return get_order(size * MAX_OBJS_PER_PAGE) - 1; | |
3222 | ||
3223 | for (order = max(min_order, get_order(min_objects * size + reserved)); | |
3224 | order <= max_order; order++) { | |
3225 | ||
3226 | unsigned long slab_size = PAGE_SIZE << order; | |
3227 | ||
3228 | rem = (slab_size - reserved) % size; | |
3229 | ||
3230 | if (rem <= slab_size / fract_leftover) | |
3231 | break; | |
3232 | } | |
3233 | ||
3234 | return order; | |
3235 | } | |
3236 | ||
3237 | static inline int calculate_order(int size, int reserved) | |
3238 | { | |
3239 | int order; | |
3240 | int min_objects; | |
3241 | int fraction; | |
3242 | int max_objects; | |
3243 | ||
3244 | /* | |
3245 | * Attempt to find best configuration for a slab. This | |
3246 | * works by first attempting to generate a layout with | |
3247 | * the best configuration and backing off gradually. | |
3248 | * | |
3249 | * First we increase the acceptable waste in a slab. Then | |
3250 | * we reduce the minimum objects required in a slab. | |
3251 | */ | |
3252 | min_objects = slub_min_objects; | |
3253 | if (!min_objects) | |
3254 | min_objects = 4 * (fls(nr_cpu_ids) + 1); | |
3255 | max_objects = order_objects(slub_max_order, size, reserved); | |
3256 | min_objects = min(min_objects, max_objects); | |
3257 | ||
3258 | while (min_objects > 1) { | |
3259 | fraction = 16; | |
3260 | while (fraction >= 4) { | |
3261 | order = slab_order(size, min_objects, | |
3262 | slub_max_order, fraction, reserved); | |
3263 | if (order <= slub_max_order) | |
3264 | return order; | |
3265 | fraction /= 2; | |
3266 | } | |
3267 | min_objects--; | |
3268 | } | |
3269 | ||
3270 | /* | |
3271 | * We were unable to place multiple objects in a slab. Now | |
3272 | * lets see if we can place a single object there. | |
3273 | */ | |
3274 | order = slab_order(size, 1, slub_max_order, 1, reserved); | |
3275 | if (order <= slub_max_order) | |
3276 | return order; | |
3277 | ||
3278 | /* | |
3279 | * Doh this slab cannot be placed using slub_max_order. | |
3280 | */ | |
3281 | order = slab_order(size, 1, MAX_ORDER, 1, reserved); | |
3282 | if (order < MAX_ORDER) | |
3283 | return order; | |
3284 | return -ENOSYS; | |
3285 | } | |
3286 | ||
3287 | static void | |
3288 | init_kmem_cache_node(struct kmem_cache_node *n) | |
3289 | { | |
3290 | n->nr_partial = 0; | |
3291 | spin_lock_init(&n->list_lock); | |
3292 | INIT_LIST_HEAD(&n->partial); | |
3293 | #ifdef CONFIG_SLUB_DEBUG | |
3294 | atomic_long_set(&n->nr_slabs, 0); | |
3295 | atomic_long_set(&n->total_objects, 0); | |
3296 | INIT_LIST_HEAD(&n->full); | |
3297 | #endif | |
3298 | } | |
3299 | ||
3300 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) | |
3301 | { | |
3302 | BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < | |
3303 | KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu)); | |
3304 | ||
3305 | /* | |
3306 | * Must align to double word boundary for the double cmpxchg | |
3307 | * instructions to work; see __pcpu_double_call_return_bool(). | |
3308 | */ | |
3309 | s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), | |
3310 | 2 * sizeof(void *)); | |
3311 | ||
3312 | if (!s->cpu_slab) | |
3313 | return 0; | |
3314 | ||
3315 | init_kmem_cache_cpus(s); | |
3316 | ||
3317 | return 1; | |
3318 | } | |
3319 | ||
3320 | static struct kmem_cache *kmem_cache_node; | |
3321 | ||
3322 | /* | |
3323 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
3324 | * slab on the node for this slabcache. There are no concurrent accesses | |
3325 | * possible. | |
3326 | * | |
3327 | * Note that this function only works on the kmem_cache_node | |
3328 | * when allocating for the kmem_cache_node. This is used for bootstrapping | |
3329 | * memory on a fresh node that has no slab structures yet. | |
3330 | */ | |
3331 | static void early_kmem_cache_node_alloc(int node) | |
3332 | { | |
3333 | struct page *page; | |
3334 | struct kmem_cache_node *n; | |
3335 | ||
3336 | BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); | |
3337 | ||
3338 | page = new_slab(kmem_cache_node, GFP_NOWAIT, node); | |
3339 | ||
3340 | BUG_ON(!page); | |
3341 | if (page_to_nid(page) != node) { | |
3342 | pr_err("SLUB: Unable to allocate memory from node %d\n", node); | |
3343 | pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n"); | |
3344 | } | |
3345 | ||
3346 | n = page->freelist; | |
3347 | BUG_ON(!n); | |
3348 | page->freelist = get_freepointer(kmem_cache_node, n); | |
3349 | page->inuse = 1; | |
3350 | page->frozen = 0; | |
3351 | kmem_cache_node->node[node] = n; | |
3352 | #ifdef CONFIG_SLUB_DEBUG | |
3353 | init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); | |
3354 | init_tracking(kmem_cache_node, n); | |
3355 | #endif | |
3356 | kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node), | |
3357 | GFP_KERNEL); | |
3358 | init_kmem_cache_node(n); | |
3359 | inc_slabs_node(kmem_cache_node, node, page->objects); | |
3360 | ||
3361 | /* | |
3362 | * No locks need to be taken here as it has just been | |
3363 | * initialized and there is no concurrent access. | |
3364 | */ | |
3365 | __add_partial(n, page, DEACTIVATE_TO_HEAD); | |
3366 | } | |
3367 | ||
3368 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
3369 | { | |
3370 | int node; | |
3371 | struct kmem_cache_node *n; | |
3372 | ||
3373 | for_each_kmem_cache_node(s, node, n) { | |
3374 | s->node[node] = NULL; | |
3375 | kmem_cache_free(kmem_cache_node, n); | |
3376 | } | |
3377 | } | |
3378 | ||
3379 | void __kmem_cache_release(struct kmem_cache *s) | |
3380 | { | |
3381 | cache_random_seq_destroy(s); | |
3382 | free_percpu(s->cpu_slab); | |
3383 | free_kmem_cache_nodes(s); | |
3384 | } | |
3385 | ||
3386 | static int init_kmem_cache_nodes(struct kmem_cache *s) | |
3387 | { | |
3388 | int node; | |
3389 | ||
3390 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3391 | struct kmem_cache_node *n; | |
3392 | ||
3393 | if (slab_state == DOWN) { | |
3394 | early_kmem_cache_node_alloc(node); | |
3395 | continue; | |
3396 | } | |
3397 | n = kmem_cache_alloc_node(kmem_cache_node, | |
3398 | GFP_KERNEL, node); | |
3399 | ||
3400 | if (!n) { | |
3401 | free_kmem_cache_nodes(s); | |
3402 | return 0; | |
3403 | } | |
3404 | ||
3405 | init_kmem_cache_node(n); | |
3406 | s->node[node] = n; | |
3407 | } | |
3408 | return 1; | |
3409 | } | |
3410 | ||
3411 | static void set_min_partial(struct kmem_cache *s, unsigned long min) | |
3412 | { | |
3413 | if (min < MIN_PARTIAL) | |
3414 | min = MIN_PARTIAL; | |
3415 | else if (min > MAX_PARTIAL) | |
3416 | min = MAX_PARTIAL; | |
3417 | s->min_partial = min; | |
3418 | } | |
3419 | ||
3420 | static void set_cpu_partial(struct kmem_cache *s) | |
3421 | { | |
3422 | #ifdef CONFIG_SLUB_CPU_PARTIAL | |
3423 | /* | |
3424 | * cpu_partial determined the maximum number of objects kept in the | |
3425 | * per cpu partial lists of a processor. | |
3426 | * | |
3427 | * Per cpu partial lists mainly contain slabs that just have one | |
3428 | * object freed. If they are used for allocation then they can be | |
3429 | * filled up again with minimal effort. The slab will never hit the | |
3430 | * per node partial lists and therefore no locking will be required. | |
3431 | * | |
3432 | * This setting also determines | |
3433 | * | |
3434 | * A) The number of objects from per cpu partial slabs dumped to the | |
3435 | * per node list when we reach the limit. | |
3436 | * B) The number of objects in cpu partial slabs to extract from the | |
3437 | * per node list when we run out of per cpu objects. We only fetch | |
3438 | * 50% to keep some capacity around for frees. | |
3439 | */ | |
3440 | if (!kmem_cache_has_cpu_partial(s)) | |
3441 | s->cpu_partial = 0; | |
3442 | else if (s->size >= PAGE_SIZE) | |
3443 | s->cpu_partial = 2; | |
3444 | else if (s->size >= 1024) | |
3445 | s->cpu_partial = 6; | |
3446 | else if (s->size >= 256) | |
3447 | s->cpu_partial = 13; | |
3448 | else | |
3449 | s->cpu_partial = 30; | |
3450 | #endif | |
3451 | } | |
3452 | ||
3453 | /* | |
3454 | * calculate_sizes() determines the order and the distribution of data within | |
3455 | * a slab object. | |
3456 | */ | |
3457 | static int calculate_sizes(struct kmem_cache *s, int forced_order) | |
3458 | { | |
3459 | slab_flags_t flags = s->flags; | |
3460 | size_t size = s->object_size; | |
3461 | int order; | |
3462 | ||
3463 | /* | |
3464 | * Round up object size to the next word boundary. We can only | |
3465 | * place the free pointer at word boundaries and this determines | |
3466 | * the possible location of the free pointer. | |
3467 | */ | |
3468 | size = ALIGN(size, sizeof(void *)); | |
3469 | ||
3470 | #ifdef CONFIG_SLUB_DEBUG | |
3471 | /* | |
3472 | * Determine if we can poison the object itself. If the user of | |
3473 | * the slab may touch the object after free or before allocation | |
3474 | * then we should never poison the object itself. | |
3475 | */ | |
3476 | if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) && | |
3477 | !s->ctor) | |
3478 | s->flags |= __OBJECT_POISON; | |
3479 | else | |
3480 | s->flags &= ~__OBJECT_POISON; | |
3481 | ||
3482 | ||
3483 | /* | |
3484 | * If we are Redzoning then check if there is some space between the | |
3485 | * end of the object and the free pointer. If not then add an | |
3486 | * additional word to have some bytes to store Redzone information. | |
3487 | */ | |
3488 | if ((flags & SLAB_RED_ZONE) && size == s->object_size) | |
3489 | size += sizeof(void *); | |
3490 | #endif | |
3491 | ||
3492 | /* | |
3493 | * With that we have determined the number of bytes in actual use | |
3494 | * by the object. This is the potential offset to the free pointer. | |
3495 | */ | |
3496 | s->inuse = size; | |
3497 | ||
3498 | if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) || | |
3499 | s->ctor)) { | |
3500 | /* | |
3501 | * Relocate free pointer after the object if it is not | |
3502 | * permitted to overwrite the first word of the object on | |
3503 | * kmem_cache_free. | |
3504 | * | |
3505 | * This is the case if we do RCU, have a constructor or | |
3506 | * destructor or are poisoning the objects. | |
3507 | */ | |
3508 | s->offset = size; | |
3509 | size += sizeof(void *); | |
3510 | } | |
3511 | ||
3512 | #ifdef CONFIG_SLUB_DEBUG | |
3513 | if (flags & SLAB_STORE_USER) | |
3514 | /* | |
3515 | * Need to store information about allocs and frees after | |
3516 | * the object. | |
3517 | */ | |
3518 | size += 2 * sizeof(struct track); | |
3519 | #endif | |
3520 | ||
3521 | kasan_cache_create(s, &size, &s->flags); | |
3522 | #ifdef CONFIG_SLUB_DEBUG | |
3523 | if (flags & SLAB_RED_ZONE) { | |
3524 | /* | |
3525 | * Add some empty padding so that we can catch | |
3526 | * overwrites from earlier objects rather than let | |
3527 | * tracking information or the free pointer be | |
3528 | * corrupted if a user writes before the start | |
3529 | * of the object. | |
3530 | */ | |
3531 | size += sizeof(void *); | |
3532 | ||
3533 | s->red_left_pad = sizeof(void *); | |
3534 | s->red_left_pad = ALIGN(s->red_left_pad, s->align); | |
3535 | size += s->red_left_pad; | |
3536 | } | |
3537 | #endif | |
3538 | ||
3539 | /* | |
3540 | * SLUB stores one object immediately after another beginning from | |
3541 | * offset 0. In order to align the objects we have to simply size | |
3542 | * each object to conform to the alignment. | |
3543 | */ | |
3544 | size = ALIGN(size, s->align); | |
3545 | s->size = size; | |
3546 | if (forced_order >= 0) | |
3547 | order = forced_order; | |
3548 | else | |
3549 | order = calculate_order(size, s->reserved); | |
3550 | ||
3551 | if (order < 0) | |
3552 | return 0; | |
3553 | ||
3554 | s->allocflags = 0; | |
3555 | if (order) | |
3556 | s->allocflags |= __GFP_COMP; | |
3557 | ||
3558 | if (s->flags & SLAB_CACHE_DMA) | |
3559 | s->allocflags |= GFP_DMA; | |
3560 | ||
3561 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
3562 | s->allocflags |= __GFP_RECLAIMABLE; | |
3563 | ||
3564 | /* | |
3565 | * Determine the number of objects per slab | |
3566 | */ | |
3567 | s->oo = oo_make(order, size, s->reserved); | |
3568 | s->min = oo_make(get_order(size), size, s->reserved); | |
3569 | if (oo_objects(s->oo) > oo_objects(s->max)) | |
3570 | s->max = s->oo; | |
3571 | ||
3572 | return !!oo_objects(s->oo); | |
3573 | } | |
3574 | ||
3575 | static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags) | |
3576 | { | |
3577 | s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor); | |
3578 | s->reserved = 0; | |
3579 | #ifdef CONFIG_SLAB_FREELIST_HARDENED | |
3580 | s->random = get_random_long(); | |
3581 | #endif | |
3582 | ||
3583 | if (need_reserve_slab_rcu && (s->flags & SLAB_TYPESAFE_BY_RCU)) | |
3584 | s->reserved = sizeof(struct rcu_head); | |
3585 | ||
3586 | if (!calculate_sizes(s, -1)) | |
3587 | goto error; | |
3588 | if (disable_higher_order_debug) { | |
3589 | /* | |
3590 | * Disable debugging flags that store metadata if the min slab | |
3591 | * order increased. | |
3592 | */ | |
3593 | if (get_order(s->size) > get_order(s->object_size)) { | |
3594 | s->flags &= ~DEBUG_METADATA_FLAGS; | |
3595 | s->offset = 0; | |
3596 | if (!calculate_sizes(s, -1)) | |
3597 | goto error; | |
3598 | } | |
3599 | } | |
3600 | ||
3601 | #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ | |
3602 | defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) | |
3603 | if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0) | |
3604 | /* Enable fast mode */ | |
3605 | s->flags |= __CMPXCHG_DOUBLE; | |
3606 | #endif | |
3607 | ||
3608 | /* | |
3609 | * The larger the object size is, the more pages we want on the partial | |
3610 | * list to avoid pounding the page allocator excessively. | |
3611 | */ | |
3612 | set_min_partial(s, ilog2(s->size) / 2); | |
3613 | ||
3614 | set_cpu_partial(s); | |
3615 | ||
3616 | #ifdef CONFIG_NUMA | |
3617 | s->remote_node_defrag_ratio = 1000; | |
3618 | #endif | |
3619 | ||
3620 | /* Initialize the pre-computed randomized freelist if slab is up */ | |
3621 | if (slab_state >= UP) { | |
3622 | if (init_cache_random_seq(s)) | |
3623 | goto error; | |
3624 | } | |
3625 | ||
3626 | if (!init_kmem_cache_nodes(s)) | |
3627 | goto error; | |
3628 | ||
3629 | if (alloc_kmem_cache_cpus(s)) | |
3630 | return 0; | |
3631 | ||
3632 | free_kmem_cache_nodes(s); | |
3633 | error: | |
3634 | if (flags & SLAB_PANIC) | |
3635 | panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n", | |
3636 | s->name, (unsigned long)s->size, s->size, | |
3637 | oo_order(s->oo), s->offset, (unsigned long)flags); | |
3638 | return -EINVAL; | |
3639 | } | |
3640 | ||
3641 | static void list_slab_objects(struct kmem_cache *s, struct page *page, | |
3642 | const char *text) | |
3643 | { | |
3644 | #ifdef CONFIG_SLUB_DEBUG | |
3645 | void *addr = page_address(page); | |
3646 | void *p; | |
3647 | unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * | |
3648 | sizeof(long), GFP_ATOMIC); | |
3649 | if (!map) | |
3650 | return; | |
3651 | slab_err(s, page, text, s->name); | |
3652 | slab_lock(page); | |
3653 | ||
3654 | get_map(s, page, map); | |
3655 | for_each_object(p, s, addr, page->objects) { | |
3656 | ||
3657 | if (!test_bit(slab_index(p, s, addr), map)) { | |
3658 | pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr); | |
3659 | print_tracking(s, p); | |
3660 | } | |
3661 | } | |
3662 | slab_unlock(page); | |
3663 | kfree(map); | |
3664 | #endif | |
3665 | } | |
3666 | ||
3667 | /* | |
3668 | * Attempt to free all partial slabs on a node. | |
3669 | * This is called from __kmem_cache_shutdown(). We must take list_lock | |
3670 | * because sysfs file might still access partial list after the shutdowning. | |
3671 | */ | |
3672 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) | |
3673 | { | |
3674 | LIST_HEAD(discard); | |
3675 | struct page *page, *h; | |
3676 | ||
3677 | BUG_ON(irqs_disabled()); | |
3678 | spin_lock_irq(&n->list_lock); | |
3679 | list_for_each_entry_safe(page, h, &n->partial, lru) { | |
3680 | if (!page->inuse) { | |
3681 | remove_partial(n, page); | |
3682 | list_add(&page->lru, &discard); | |
3683 | } else { | |
3684 | list_slab_objects(s, page, | |
3685 | "Objects remaining in %s on __kmem_cache_shutdown()"); | |
3686 | } | |
3687 | } | |
3688 | spin_unlock_irq(&n->list_lock); | |
3689 | ||
3690 | list_for_each_entry_safe(page, h, &discard, lru) | |
3691 | discard_slab(s, page); | |
3692 | } | |
3693 | ||
3694 | /* | |
3695 | * Release all resources used by a slab cache. | |
3696 | */ | |
3697 | int __kmem_cache_shutdown(struct kmem_cache *s) | |
3698 | { | |
3699 | int node; | |
3700 | struct kmem_cache_node *n; | |
3701 | ||
3702 | flush_all(s); | |
3703 | /* Attempt to free all objects */ | |
3704 | for_each_kmem_cache_node(s, node, n) { | |
3705 | free_partial(s, n); | |
3706 | if (n->nr_partial || slabs_node(s, node)) | |
3707 | return 1; | |
3708 | } | |
3709 | sysfs_slab_remove(s); | |
3710 | return 0; | |
3711 | } | |
3712 | ||
3713 | /******************************************************************** | |
3714 | * Kmalloc subsystem | |
3715 | *******************************************************************/ | |
3716 | ||
3717 | static int __init setup_slub_min_order(char *str) | |
3718 | { | |
3719 | get_option(&str, &slub_min_order); | |
3720 | ||
3721 | return 1; | |
3722 | } | |
3723 | ||
3724 | __setup("slub_min_order=", setup_slub_min_order); | |
3725 | ||
3726 | static int __init setup_slub_max_order(char *str) | |
3727 | { | |
3728 | get_option(&str, &slub_max_order); | |
3729 | slub_max_order = min(slub_max_order, MAX_ORDER - 1); | |
3730 | ||
3731 | return 1; | |
3732 | } | |
3733 | ||
3734 | __setup("slub_max_order=", setup_slub_max_order); | |
3735 | ||
3736 | static int __init setup_slub_min_objects(char *str) | |
3737 | { | |
3738 | get_option(&str, &slub_min_objects); | |
3739 | ||
3740 | return 1; | |
3741 | } | |
3742 | ||
3743 | __setup("slub_min_objects=", setup_slub_min_objects); | |
3744 | ||
3745 | void *__kmalloc(size_t size, gfp_t flags) | |
3746 | { | |
3747 | struct kmem_cache *s; | |
3748 | void *ret; | |
3749 | ||
3750 | if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) | |
3751 | return kmalloc_large(size, flags); | |
3752 | ||
3753 | s = kmalloc_slab(size, flags); | |
3754 | ||
3755 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3756 | return s; | |
3757 | ||
3758 | ret = slab_alloc(s, flags, _RET_IP_); | |
3759 | ||
3760 | trace_kmalloc(_RET_IP_, ret, size, s->size, flags); | |
3761 | ||
3762 | kasan_kmalloc(s, ret, size, flags); | |
3763 | ||
3764 | return ret; | |
3765 | } | |
3766 | EXPORT_SYMBOL(__kmalloc); | |
3767 | ||
3768 | #ifdef CONFIG_NUMA | |
3769 | static void *kmalloc_large_node(size_t size, gfp_t flags, int node) | |
3770 | { | |
3771 | struct page *page; | |
3772 | void *ptr = NULL; | |
3773 | ||
3774 | flags |= __GFP_COMP; | |
3775 | page = alloc_pages_node(node, flags, get_order(size)); | |
3776 | if (page) | |
3777 | ptr = page_address(page); | |
3778 | ||
3779 | kmalloc_large_node_hook(ptr, size, flags); | |
3780 | return ptr; | |
3781 | } | |
3782 | ||
3783 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3784 | { | |
3785 | struct kmem_cache *s; | |
3786 | void *ret; | |
3787 | ||
3788 | if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { | |
3789 | ret = kmalloc_large_node(size, flags, node); | |
3790 | ||
3791 | trace_kmalloc_node(_RET_IP_, ret, | |
3792 | size, PAGE_SIZE << get_order(size), | |
3793 | flags, node); | |
3794 | ||
3795 | return ret; | |
3796 | } | |
3797 | ||
3798 | s = kmalloc_slab(size, flags); | |
3799 | ||
3800 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3801 | return s; | |
3802 | ||
3803 | ret = slab_alloc_node(s, flags, node, _RET_IP_); | |
3804 | ||
3805 | trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); | |
3806 | ||
3807 | kasan_kmalloc(s, ret, size, flags); | |
3808 | ||
3809 | return ret; | |
3810 | } | |
3811 | EXPORT_SYMBOL(__kmalloc_node); | |
3812 | #endif | |
3813 | ||
3814 | #ifdef CONFIG_HARDENED_USERCOPY | |
3815 | /* | |
3816 | * Rejects objects that are incorrectly sized. | |
3817 | * | |
3818 | * Returns NULL if check passes, otherwise const char * to name of cache | |
3819 | * to indicate an error. | |
3820 | */ | |
3821 | const char *__check_heap_object(const void *ptr, unsigned long n, | |
3822 | struct page *page) | |
3823 | { | |
3824 | struct kmem_cache *s; | |
3825 | unsigned long offset; | |
3826 | size_t object_size; | |
3827 | ||
3828 | /* Find object and usable object size. */ | |
3829 | s = page->slab_cache; | |
3830 | object_size = slab_ksize(s); | |
3831 | ||
3832 | /* Reject impossible pointers. */ | |
3833 | if (ptr < page_address(page)) | |
3834 | return s->name; | |
3835 | ||
3836 | /* Find offset within object. */ | |
3837 | offset = (ptr - page_address(page)) % s->size; | |
3838 | ||
3839 | /* Adjust for redzone and reject if within the redzone. */ | |
3840 | if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) { | |
3841 | if (offset < s->red_left_pad) | |
3842 | return s->name; | |
3843 | offset -= s->red_left_pad; | |
3844 | } | |
3845 | ||
3846 | /* Allow address range falling entirely within object size. */ | |
3847 | if (offset <= object_size && n <= object_size - offset) | |
3848 | return NULL; | |
3849 | ||
3850 | return s->name; | |
3851 | } | |
3852 | #endif /* CONFIG_HARDENED_USERCOPY */ | |
3853 | ||
3854 | static size_t __ksize(const void *object) | |
3855 | { | |
3856 | struct page *page; | |
3857 | ||
3858 | if (unlikely(object == ZERO_SIZE_PTR)) | |
3859 | return 0; | |
3860 | ||
3861 | page = virt_to_head_page(object); | |
3862 | ||
3863 | if (unlikely(!PageSlab(page))) { | |
3864 | WARN_ON(!PageCompound(page)); | |
3865 | return PAGE_SIZE << compound_order(page); | |
3866 | } | |
3867 | ||
3868 | return slab_ksize(page->slab_cache); | |
3869 | } | |
3870 | ||
3871 | size_t ksize(const void *object) | |
3872 | { | |
3873 | size_t size = __ksize(object); | |
3874 | /* We assume that ksize callers could use whole allocated area, | |
3875 | * so we need to unpoison this area. | |
3876 | */ | |
3877 | kasan_unpoison_shadow(object, size); | |
3878 | return size; | |
3879 | } | |
3880 | EXPORT_SYMBOL(ksize); | |
3881 | ||
3882 | void kfree(const void *x) | |
3883 | { | |
3884 | struct page *page; | |
3885 | void *object = (void *)x; | |
3886 | ||
3887 | trace_kfree(_RET_IP_, x); | |
3888 | ||
3889 | if (unlikely(ZERO_OR_NULL_PTR(x))) | |
3890 | return; | |
3891 | ||
3892 | page = virt_to_head_page(x); | |
3893 | if (unlikely(!PageSlab(page))) { | |
3894 | BUG_ON(!PageCompound(page)); | |
3895 | kfree_hook(x); | |
3896 | __free_pages(page, compound_order(page)); | |
3897 | return; | |
3898 | } | |
3899 | slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_); | |
3900 | } | |
3901 | EXPORT_SYMBOL(kfree); | |
3902 | ||
3903 | #define SHRINK_PROMOTE_MAX 32 | |
3904 | ||
3905 | /* | |
3906 | * kmem_cache_shrink discards empty slabs and promotes the slabs filled | |
3907 | * up most to the head of the partial lists. New allocations will then | |
3908 | * fill those up and thus they can be removed from the partial lists. | |
3909 | * | |
3910 | * The slabs with the least items are placed last. This results in them | |
3911 | * being allocated from last increasing the chance that the last objects | |
3912 | * are freed in them. | |
3913 | */ | |
3914 | int __kmem_cache_shrink(struct kmem_cache *s) | |
3915 | { | |
3916 | int node; | |
3917 | int i; | |
3918 | struct kmem_cache_node *n; | |
3919 | struct page *page; | |
3920 | struct page *t; | |
3921 | struct list_head discard; | |
3922 | struct list_head promote[SHRINK_PROMOTE_MAX]; | |
3923 | unsigned long flags; | |
3924 | int ret = 0; | |
3925 | ||
3926 | flush_all(s); | |
3927 | for_each_kmem_cache_node(s, node, n) { | |
3928 | INIT_LIST_HEAD(&discard); | |
3929 | for (i = 0; i < SHRINK_PROMOTE_MAX; i++) | |
3930 | INIT_LIST_HEAD(promote + i); | |
3931 | ||
3932 | spin_lock_irqsave(&n->list_lock, flags); | |
3933 | ||
3934 | /* | |
3935 | * Build lists of slabs to discard or promote. | |
3936 | * | |
3937 | * Note that concurrent frees may occur while we hold the | |
3938 | * list_lock. page->inuse here is the upper limit. | |
3939 | */ | |
3940 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
3941 | int free = page->objects - page->inuse; | |
3942 | ||
3943 | /* Do not reread page->inuse */ | |
3944 | barrier(); | |
3945 | ||
3946 | /* We do not keep full slabs on the list */ | |
3947 | BUG_ON(free <= 0); | |
3948 | ||
3949 | if (free == page->objects) { | |
3950 | list_move(&page->lru, &discard); | |
3951 | n->nr_partial--; | |
3952 | } else if (free <= SHRINK_PROMOTE_MAX) | |
3953 | list_move(&page->lru, promote + free - 1); | |
3954 | } | |
3955 | ||
3956 | /* | |
3957 | * Promote the slabs filled up most to the head of the | |
3958 | * partial list. | |
3959 | */ | |
3960 | for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) | |
3961 | list_splice(promote + i, &n->partial); | |
3962 | ||
3963 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3964 | ||
3965 | /* Release empty slabs */ | |
3966 | list_for_each_entry_safe(page, t, &discard, lru) | |
3967 | discard_slab(s, page); | |
3968 | ||
3969 | if (slabs_node(s, node)) | |
3970 | ret = 1; | |
3971 | } | |
3972 | ||
3973 | return ret; | |
3974 | } | |
3975 | ||
3976 | #ifdef CONFIG_MEMCG | |
3977 | static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s) | |
3978 | { | |
3979 | /* | |
3980 | * Called with all the locks held after a sched RCU grace period. | |
3981 | * Even if @s becomes empty after shrinking, we can't know that @s | |
3982 | * doesn't have allocations already in-flight and thus can't | |
3983 | * destroy @s until the associated memcg is released. | |
3984 | * | |
3985 | * However, let's remove the sysfs files for empty caches here. | |
3986 | * Each cache has a lot of interface files which aren't | |
3987 | * particularly useful for empty draining caches; otherwise, we can | |
3988 | * easily end up with millions of unnecessary sysfs files on | |
3989 | * systems which have a lot of memory and transient cgroups. | |
3990 | */ | |
3991 | if (!__kmem_cache_shrink(s)) | |
3992 | sysfs_slab_remove(s); | |
3993 | } | |
3994 | ||
3995 | void __kmemcg_cache_deactivate(struct kmem_cache *s) | |
3996 | { | |
3997 | /* | |
3998 | * Disable empty slabs caching. Used to avoid pinning offline | |
3999 | * memory cgroups by kmem pages that can be freed. | |
4000 | */ | |
4001 | slub_set_cpu_partial(s, 0); | |
4002 | s->min_partial = 0; | |
4003 | ||
4004 | /* | |
4005 | * s->cpu_partial is checked locklessly (see put_cpu_partial), so | |
4006 | * we have to make sure the change is visible before shrinking. | |
4007 | */ | |
4008 | slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu); | |
4009 | } | |
4010 | #endif | |
4011 | ||
4012 | static int slab_mem_going_offline_callback(void *arg) | |
4013 | { | |
4014 | struct kmem_cache *s; | |
4015 | ||
4016 | mutex_lock(&slab_mutex); | |
4017 | list_for_each_entry(s, &slab_caches, list) | |
4018 | __kmem_cache_shrink(s); | |
4019 | mutex_unlock(&slab_mutex); | |
4020 | ||
4021 | return 0; | |
4022 | } | |
4023 | ||
4024 | static void slab_mem_offline_callback(void *arg) | |
4025 | { | |
4026 | struct kmem_cache_node *n; | |
4027 | struct kmem_cache *s; | |
4028 | struct memory_notify *marg = arg; | |
4029 | int offline_node; | |
4030 | ||
4031 | offline_node = marg->status_change_nid_normal; | |
4032 | ||
4033 | /* | |
4034 | * If the node still has available memory. we need kmem_cache_node | |
4035 | * for it yet. | |
4036 | */ | |
4037 | if (offline_node < 0) | |
4038 | return; | |
4039 | ||
4040 | mutex_lock(&slab_mutex); | |
4041 | list_for_each_entry(s, &slab_caches, list) { | |
4042 | n = get_node(s, offline_node); | |
4043 | if (n) { | |
4044 | /* | |
4045 | * if n->nr_slabs > 0, slabs still exist on the node | |
4046 | * that is going down. We were unable to free them, | |
4047 | * and offline_pages() function shouldn't call this | |
4048 | * callback. So, we must fail. | |
4049 | */ | |
4050 | BUG_ON(slabs_node(s, offline_node)); | |
4051 | ||
4052 | s->node[offline_node] = NULL; | |
4053 | kmem_cache_free(kmem_cache_node, n); | |
4054 | } | |
4055 | } | |
4056 | mutex_unlock(&slab_mutex); | |
4057 | } | |
4058 | ||
4059 | static int slab_mem_going_online_callback(void *arg) | |
4060 | { | |
4061 | struct kmem_cache_node *n; | |
4062 | struct kmem_cache *s; | |
4063 | struct memory_notify *marg = arg; | |
4064 | int nid = marg->status_change_nid_normal; | |
4065 | int ret = 0; | |
4066 | ||
4067 | /* | |
4068 | * If the node's memory is already available, then kmem_cache_node is | |
4069 | * already created. Nothing to do. | |
4070 | */ | |
4071 | if (nid < 0) | |
4072 | return 0; | |
4073 | ||
4074 | /* | |
4075 | * We are bringing a node online. No memory is available yet. We must | |
4076 | * allocate a kmem_cache_node structure in order to bring the node | |
4077 | * online. | |
4078 | */ | |
4079 | mutex_lock(&slab_mutex); | |
4080 | list_for_each_entry(s, &slab_caches, list) { | |
4081 | /* | |
4082 | * XXX: kmem_cache_alloc_node will fallback to other nodes | |
4083 | * since memory is not yet available from the node that | |
4084 | * is brought up. | |
4085 | */ | |
4086 | n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); | |
4087 | if (!n) { | |
4088 | ret = -ENOMEM; | |
4089 | goto out; | |
4090 | } | |
4091 | init_kmem_cache_node(n); | |
4092 | s->node[nid] = n; | |
4093 | } | |
4094 | out: | |
4095 | mutex_unlock(&slab_mutex); | |
4096 | return ret; | |
4097 | } | |
4098 | ||
4099 | static int slab_memory_callback(struct notifier_block *self, | |
4100 | unsigned long action, void *arg) | |
4101 | { | |
4102 | int ret = 0; | |
4103 | ||
4104 | switch (action) { | |
4105 | case MEM_GOING_ONLINE: | |
4106 | ret = slab_mem_going_online_callback(arg); | |
4107 | break; | |
4108 | case MEM_GOING_OFFLINE: | |
4109 | ret = slab_mem_going_offline_callback(arg); | |
4110 | break; | |
4111 | case MEM_OFFLINE: | |
4112 | case MEM_CANCEL_ONLINE: | |
4113 | slab_mem_offline_callback(arg); | |
4114 | break; | |
4115 | case MEM_ONLINE: | |
4116 | case MEM_CANCEL_OFFLINE: | |
4117 | break; | |
4118 | } | |
4119 | if (ret) | |
4120 | ret = notifier_from_errno(ret); | |
4121 | else | |
4122 | ret = NOTIFY_OK; | |
4123 | return ret; | |
4124 | } | |
4125 | ||
4126 | static struct notifier_block slab_memory_callback_nb = { | |
4127 | .notifier_call = slab_memory_callback, | |
4128 | .priority = SLAB_CALLBACK_PRI, | |
4129 | }; | |
4130 | ||
4131 | /******************************************************************** | |
4132 | * Basic setup of slabs | |
4133 | *******************************************************************/ | |
4134 | ||
4135 | /* | |
4136 | * Used for early kmem_cache structures that were allocated using | |
4137 | * the page allocator. Allocate them properly then fix up the pointers | |
4138 | * that may be pointing to the wrong kmem_cache structure. | |
4139 | */ | |
4140 | ||
4141 | static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) | |
4142 | { | |
4143 | int node; | |
4144 | struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); | |
4145 | struct kmem_cache_node *n; | |
4146 | ||
4147 | memcpy(s, static_cache, kmem_cache->object_size); | |
4148 | ||
4149 | /* | |
4150 | * This runs very early, and only the boot processor is supposed to be | |
4151 | * up. Even if it weren't true, IRQs are not up so we couldn't fire | |
4152 | * IPIs around. | |
4153 | */ | |
4154 | __flush_cpu_slab(s, smp_processor_id()); | |
4155 | for_each_kmem_cache_node(s, node, n) { | |
4156 | struct page *p; | |
4157 | ||
4158 | list_for_each_entry(p, &n->partial, lru) | |
4159 | p->slab_cache = s; | |
4160 | ||
4161 | #ifdef CONFIG_SLUB_DEBUG | |
4162 | list_for_each_entry(p, &n->full, lru) | |
4163 | p->slab_cache = s; | |
4164 | #endif | |
4165 | } | |
4166 | slab_init_memcg_params(s); | |
4167 | list_add(&s->list, &slab_caches); | |
4168 | memcg_link_cache(s); | |
4169 | return s; | |
4170 | } | |
4171 | ||
4172 | void __init kmem_cache_init(void) | |
4173 | { | |
4174 | static __initdata struct kmem_cache boot_kmem_cache, | |
4175 | boot_kmem_cache_node; | |
4176 | ||
4177 | if (debug_guardpage_minorder()) | |
4178 | slub_max_order = 0; | |
4179 | ||
4180 | kmem_cache_node = &boot_kmem_cache_node; | |
4181 | kmem_cache = &boot_kmem_cache; | |
4182 | ||
4183 | create_boot_cache(kmem_cache_node, "kmem_cache_node", | |
4184 | sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN); | |
4185 | ||
4186 | register_hotmemory_notifier(&slab_memory_callback_nb); | |
4187 | ||
4188 | /* Able to allocate the per node structures */ | |
4189 | slab_state = PARTIAL; | |
4190 | ||
4191 | create_boot_cache(kmem_cache, "kmem_cache", | |
4192 | offsetof(struct kmem_cache, node) + | |
4193 | nr_node_ids * sizeof(struct kmem_cache_node *), | |
4194 | SLAB_HWCACHE_ALIGN); | |
4195 | ||
4196 | kmem_cache = bootstrap(&boot_kmem_cache); | |
4197 | ||
4198 | /* | |
4199 | * Allocate kmem_cache_node properly from the kmem_cache slab. | |
4200 | * kmem_cache_node is separately allocated so no need to | |
4201 | * update any list pointers. | |
4202 | */ | |
4203 | kmem_cache_node = bootstrap(&boot_kmem_cache_node); | |
4204 | ||
4205 | /* Now we can use the kmem_cache to allocate kmalloc slabs */ | |
4206 | setup_kmalloc_cache_index_table(); | |
4207 | create_kmalloc_caches(0); | |
4208 | ||
4209 | /* Setup random freelists for each cache */ | |
4210 | init_freelist_randomization(); | |
4211 | ||
4212 | cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL, | |
4213 | slub_cpu_dead); | |
4214 | ||
4215 | pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%u, Nodes=%d\n", | |
4216 | cache_line_size(), | |
4217 | slub_min_order, slub_max_order, slub_min_objects, | |
4218 | nr_cpu_ids, nr_node_ids); | |
4219 | } | |
4220 | ||
4221 | void __init kmem_cache_init_late(void) | |
4222 | { | |
4223 | } | |
4224 | ||
4225 | struct kmem_cache * | |
4226 | __kmem_cache_alias(const char *name, size_t size, size_t align, | |
4227 | slab_flags_t flags, void (*ctor)(void *)) | |
4228 | { | |
4229 | struct kmem_cache *s, *c; | |
4230 | ||
4231 | s = find_mergeable(size, align, flags, name, ctor); | |
4232 | if (s) { | |
4233 | s->refcount++; | |
4234 | ||
4235 | /* | |
4236 | * Adjust the object sizes so that we clear | |
4237 | * the complete object on kzalloc. | |
4238 | */ | |
4239 | s->object_size = max(s->object_size, (int)size); | |
4240 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); | |
4241 | ||
4242 | for_each_memcg_cache(c, s) { | |
4243 | c->object_size = s->object_size; | |
4244 | c->inuse = max_t(int, c->inuse, | |
4245 | ALIGN(size, sizeof(void *))); | |
4246 | } | |
4247 | ||
4248 | if (sysfs_slab_alias(s, name)) { | |
4249 | s->refcount--; | |
4250 | s = NULL; | |
4251 | } | |
4252 | } | |
4253 | ||
4254 | return s; | |
4255 | } | |
4256 | ||
4257 | int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags) | |
4258 | { | |
4259 | int err; | |
4260 | ||
4261 | err = kmem_cache_open(s, flags); | |
4262 | if (err) | |
4263 | return err; | |
4264 | ||
4265 | /* Mutex is not taken during early boot */ | |
4266 | if (slab_state <= UP) | |
4267 | return 0; | |
4268 | ||
4269 | memcg_propagate_slab_attrs(s); | |
4270 | err = sysfs_slab_add(s); | |
4271 | if (err) | |
4272 | __kmem_cache_release(s); | |
4273 | ||
4274 | return err; | |
4275 | } | |
4276 | ||
4277 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) | |
4278 | { | |
4279 | struct kmem_cache *s; | |
4280 | void *ret; | |
4281 | ||
4282 | if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) | |
4283 | return kmalloc_large(size, gfpflags); | |
4284 | ||
4285 | s = kmalloc_slab(size, gfpflags); | |
4286 | ||
4287 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
4288 | return s; | |
4289 | ||
4290 | ret = slab_alloc(s, gfpflags, caller); | |
4291 | ||
4292 | /* Honor the call site pointer we received. */ | |
4293 | trace_kmalloc(caller, ret, size, s->size, gfpflags); | |
4294 | ||
4295 | return ret; | |
4296 | } | |
4297 | ||
4298 | #ifdef CONFIG_NUMA | |
4299 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
4300 | int node, unsigned long caller) | |
4301 | { | |
4302 | struct kmem_cache *s; | |
4303 | void *ret; | |
4304 | ||
4305 | if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { | |
4306 | ret = kmalloc_large_node(size, gfpflags, node); | |
4307 | ||
4308 | trace_kmalloc_node(caller, ret, | |
4309 | size, PAGE_SIZE << get_order(size), | |
4310 | gfpflags, node); | |
4311 | ||
4312 | return ret; | |
4313 | } | |
4314 | ||
4315 | s = kmalloc_slab(size, gfpflags); | |
4316 | ||
4317 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
4318 | return s; | |
4319 | ||
4320 | ret = slab_alloc_node(s, gfpflags, node, caller); | |
4321 | ||
4322 | /* Honor the call site pointer we received. */ | |
4323 | trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); | |
4324 | ||
4325 | return ret; | |
4326 | } | |
4327 | #endif | |
4328 | ||
4329 | #ifdef CONFIG_SYSFS | |
4330 | static int count_inuse(struct page *page) | |
4331 | { | |
4332 | return page->inuse; | |
4333 | } | |
4334 | ||
4335 | static int count_total(struct page *page) | |
4336 | { | |
4337 | return page->objects; | |
4338 | } | |
4339 | #endif | |
4340 | ||
4341 | #ifdef CONFIG_SLUB_DEBUG | |
4342 | static int validate_slab(struct kmem_cache *s, struct page *page, | |
4343 | unsigned long *map) | |
4344 | { | |
4345 | void *p; | |
4346 | void *addr = page_address(page); | |
4347 | ||
4348 | if (!check_slab(s, page) || | |
4349 | !on_freelist(s, page, NULL)) | |
4350 | return 0; | |
4351 | ||
4352 | /* Now we know that a valid freelist exists */ | |
4353 | bitmap_zero(map, page->objects); | |
4354 | ||
4355 | get_map(s, page, map); | |
4356 | for_each_object(p, s, addr, page->objects) { | |
4357 | if (test_bit(slab_index(p, s, addr), map)) | |
4358 | if (!check_object(s, page, p, SLUB_RED_INACTIVE)) | |
4359 | return 0; | |
4360 | } | |
4361 | ||
4362 | for_each_object(p, s, addr, page->objects) | |
4363 | if (!test_bit(slab_index(p, s, addr), map)) | |
4364 | if (!check_object(s, page, p, SLUB_RED_ACTIVE)) | |
4365 | return 0; | |
4366 | return 1; | |
4367 | } | |
4368 | ||
4369 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, | |
4370 | unsigned long *map) | |
4371 | { | |
4372 | slab_lock(page); | |
4373 | validate_slab(s, page, map); | |
4374 | slab_unlock(page); | |
4375 | } | |
4376 | ||
4377 | static int validate_slab_node(struct kmem_cache *s, | |
4378 | struct kmem_cache_node *n, unsigned long *map) | |
4379 | { | |
4380 | unsigned long count = 0; | |
4381 | struct page *page; | |
4382 | unsigned long flags; | |
4383 | ||
4384 | spin_lock_irqsave(&n->list_lock, flags); | |
4385 | ||
4386 | list_for_each_entry(page, &n->partial, lru) { | |
4387 | validate_slab_slab(s, page, map); | |
4388 | count++; | |
4389 | } | |
4390 | if (count != n->nr_partial) | |
4391 | pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n", | |
4392 | s->name, count, n->nr_partial); | |
4393 | ||
4394 | if (!(s->flags & SLAB_STORE_USER)) | |
4395 | goto out; | |
4396 | ||
4397 | list_for_each_entry(page, &n->full, lru) { | |
4398 | validate_slab_slab(s, page, map); | |
4399 | count++; | |
4400 | } | |
4401 | if (count != atomic_long_read(&n->nr_slabs)) | |
4402 | pr_err("SLUB: %s %ld slabs counted but counter=%ld\n", | |
4403 | s->name, count, atomic_long_read(&n->nr_slabs)); | |
4404 | ||
4405 | out: | |
4406 | spin_unlock_irqrestore(&n->list_lock, flags); | |
4407 | return count; | |
4408 | } | |
4409 | ||
4410 | static long validate_slab_cache(struct kmem_cache *s) | |
4411 | { | |
4412 | int node; | |
4413 | unsigned long count = 0; | |
4414 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * | |
4415 | sizeof(unsigned long), GFP_KERNEL); | |
4416 | struct kmem_cache_node *n; | |
4417 | ||
4418 | if (!map) | |
4419 | return -ENOMEM; | |
4420 | ||
4421 | flush_all(s); | |
4422 | for_each_kmem_cache_node(s, node, n) | |
4423 | count += validate_slab_node(s, n, map); | |
4424 | kfree(map); | |
4425 | return count; | |
4426 | } | |
4427 | /* | |
4428 | * Generate lists of code addresses where slabcache objects are allocated | |
4429 | * and freed. | |
4430 | */ | |
4431 | ||
4432 | struct location { | |
4433 | unsigned long count; | |
4434 | unsigned long addr; | |
4435 | long long sum_time; | |
4436 | long min_time; | |
4437 | long max_time; | |
4438 | long min_pid; | |
4439 | long max_pid; | |
4440 | DECLARE_BITMAP(cpus, NR_CPUS); | |
4441 | nodemask_t nodes; | |
4442 | }; | |
4443 | ||
4444 | struct loc_track { | |
4445 | unsigned long max; | |
4446 | unsigned long count; | |
4447 | struct location *loc; | |
4448 | }; | |
4449 | ||
4450 | static void free_loc_track(struct loc_track *t) | |
4451 | { | |
4452 | if (t->max) | |
4453 | free_pages((unsigned long)t->loc, | |
4454 | get_order(sizeof(struct location) * t->max)); | |
4455 | } | |
4456 | ||
4457 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) | |
4458 | { | |
4459 | struct location *l; | |
4460 | int order; | |
4461 | ||
4462 | order = get_order(sizeof(struct location) * max); | |
4463 | ||
4464 | l = (void *)__get_free_pages(flags, order); | |
4465 | if (!l) | |
4466 | return 0; | |
4467 | ||
4468 | if (t->count) { | |
4469 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
4470 | free_loc_track(t); | |
4471 | } | |
4472 | t->max = max; | |
4473 | t->loc = l; | |
4474 | return 1; | |
4475 | } | |
4476 | ||
4477 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
4478 | const struct track *track) | |
4479 | { | |
4480 | long start, end, pos; | |
4481 | struct location *l; | |
4482 | unsigned long caddr; | |
4483 | unsigned long age = jiffies - track->when; | |
4484 | ||
4485 | start = -1; | |
4486 | end = t->count; | |
4487 | ||
4488 | for ( ; ; ) { | |
4489 | pos = start + (end - start + 1) / 2; | |
4490 | ||
4491 | /* | |
4492 | * There is nothing at "end". If we end up there | |
4493 | * we need to add something to before end. | |
4494 | */ | |
4495 | if (pos == end) | |
4496 | break; | |
4497 | ||
4498 | caddr = t->loc[pos].addr; | |
4499 | if (track->addr == caddr) { | |
4500 | ||
4501 | l = &t->loc[pos]; | |
4502 | l->count++; | |
4503 | if (track->when) { | |
4504 | l->sum_time += age; | |
4505 | if (age < l->min_time) | |
4506 | l->min_time = age; | |
4507 | if (age > l->max_time) | |
4508 | l->max_time = age; | |
4509 | ||
4510 | if (track->pid < l->min_pid) | |
4511 | l->min_pid = track->pid; | |
4512 | if (track->pid > l->max_pid) | |
4513 | l->max_pid = track->pid; | |
4514 | ||
4515 | cpumask_set_cpu(track->cpu, | |
4516 | to_cpumask(l->cpus)); | |
4517 | } | |
4518 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
4519 | return 1; | |
4520 | } | |
4521 | ||
4522 | if (track->addr < caddr) | |
4523 | end = pos; | |
4524 | else | |
4525 | start = pos; | |
4526 | } | |
4527 | ||
4528 | /* | |
4529 | * Not found. Insert new tracking element. | |
4530 | */ | |
4531 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) | |
4532 | return 0; | |
4533 | ||
4534 | l = t->loc + pos; | |
4535 | if (pos < t->count) | |
4536 | memmove(l + 1, l, | |
4537 | (t->count - pos) * sizeof(struct location)); | |
4538 | t->count++; | |
4539 | l->count = 1; | |
4540 | l->addr = track->addr; | |
4541 | l->sum_time = age; | |
4542 | l->min_time = age; | |
4543 | l->max_time = age; | |
4544 | l->min_pid = track->pid; | |
4545 | l->max_pid = track->pid; | |
4546 | cpumask_clear(to_cpumask(l->cpus)); | |
4547 | cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); | |
4548 | nodes_clear(l->nodes); | |
4549 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
4550 | return 1; | |
4551 | } | |
4552 | ||
4553 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
4554 | struct page *page, enum track_item alloc, | |
4555 | unsigned long *map) | |
4556 | { | |
4557 | void *addr = page_address(page); | |
4558 | void *p; | |
4559 | ||
4560 | bitmap_zero(map, page->objects); | |
4561 | get_map(s, page, map); | |
4562 | ||
4563 | for_each_object(p, s, addr, page->objects) | |
4564 | if (!test_bit(slab_index(p, s, addr), map)) | |
4565 | add_location(t, s, get_track(s, p, alloc)); | |
4566 | } | |
4567 | ||
4568 | static int list_locations(struct kmem_cache *s, char *buf, | |
4569 | enum track_item alloc) | |
4570 | { | |
4571 | int len = 0; | |
4572 | unsigned long i; | |
4573 | struct loc_track t = { 0, 0, NULL }; | |
4574 | int node; | |
4575 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * | |
4576 | sizeof(unsigned long), GFP_KERNEL); | |
4577 | struct kmem_cache_node *n; | |
4578 | ||
4579 | if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), | |
4580 | GFP_KERNEL)) { | |
4581 | kfree(map); | |
4582 | return sprintf(buf, "Out of memory\n"); | |
4583 | } | |
4584 | /* Push back cpu slabs */ | |
4585 | flush_all(s); | |
4586 | ||
4587 | for_each_kmem_cache_node(s, node, n) { | |
4588 | unsigned long flags; | |
4589 | struct page *page; | |
4590 | ||
4591 | if (!atomic_long_read(&n->nr_slabs)) | |
4592 | continue; | |
4593 | ||
4594 | spin_lock_irqsave(&n->list_lock, flags); | |
4595 | list_for_each_entry(page, &n->partial, lru) | |
4596 | process_slab(&t, s, page, alloc, map); | |
4597 | list_for_each_entry(page, &n->full, lru) | |
4598 | process_slab(&t, s, page, alloc, map); | |
4599 | spin_unlock_irqrestore(&n->list_lock, flags); | |
4600 | } | |
4601 | ||
4602 | for (i = 0; i < t.count; i++) { | |
4603 | struct location *l = &t.loc[i]; | |
4604 | ||
4605 | if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) | |
4606 | break; | |
4607 | len += sprintf(buf + len, "%7ld ", l->count); | |
4608 | ||
4609 | if (l->addr) | |
4610 | len += sprintf(buf + len, "%pS", (void *)l->addr); | |
4611 | else | |
4612 | len += sprintf(buf + len, "<not-available>"); | |
4613 | ||
4614 | if (l->sum_time != l->min_time) { | |
4615 | len += sprintf(buf + len, " age=%ld/%ld/%ld", | |
4616 | l->min_time, | |
4617 | (long)div_u64(l->sum_time, l->count), | |
4618 | l->max_time); | |
4619 | } else | |
4620 | len += sprintf(buf + len, " age=%ld", | |
4621 | l->min_time); | |
4622 | ||
4623 | if (l->min_pid != l->max_pid) | |
4624 | len += sprintf(buf + len, " pid=%ld-%ld", | |
4625 | l->min_pid, l->max_pid); | |
4626 | else | |
4627 | len += sprintf(buf + len, " pid=%ld", | |
4628 | l->min_pid); | |
4629 | ||
4630 | if (num_online_cpus() > 1 && | |
4631 | !cpumask_empty(to_cpumask(l->cpus)) && | |
4632 | len < PAGE_SIZE - 60) | |
4633 | len += scnprintf(buf + len, PAGE_SIZE - len - 50, | |
4634 | " cpus=%*pbl", | |
4635 | cpumask_pr_args(to_cpumask(l->cpus))); | |
4636 | ||
4637 | if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && | |
4638 | len < PAGE_SIZE - 60) | |
4639 | len += scnprintf(buf + len, PAGE_SIZE - len - 50, | |
4640 | " nodes=%*pbl", | |
4641 | nodemask_pr_args(&l->nodes)); | |
4642 | ||
4643 | len += sprintf(buf + len, "\n"); | |
4644 | } | |
4645 | ||
4646 | free_loc_track(&t); | |
4647 | kfree(map); | |
4648 | if (!t.count) | |
4649 | len += sprintf(buf, "No data\n"); | |
4650 | return len; | |
4651 | } | |
4652 | #endif | |
4653 | ||
4654 | #ifdef SLUB_RESILIENCY_TEST | |
4655 | static void __init resiliency_test(void) | |
4656 | { | |
4657 | u8 *p; | |
4658 | ||
4659 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10); | |
4660 | ||
4661 | pr_err("SLUB resiliency testing\n"); | |
4662 | pr_err("-----------------------\n"); | |
4663 | pr_err("A. Corruption after allocation\n"); | |
4664 | ||
4665 | p = kzalloc(16, GFP_KERNEL); | |
4666 | p[16] = 0x12; | |
4667 | pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n", | |
4668 | p + 16); | |
4669 | ||
4670 | validate_slab_cache(kmalloc_caches[4]); | |
4671 | ||
4672 | /* Hmmm... The next two are dangerous */ | |
4673 | p = kzalloc(32, GFP_KERNEL); | |
4674 | p[32 + sizeof(void *)] = 0x34; | |
4675 | pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n", | |
4676 | p); | |
4677 | pr_err("If allocated object is overwritten then not detectable\n\n"); | |
4678 | ||
4679 | validate_slab_cache(kmalloc_caches[5]); | |
4680 | p = kzalloc(64, GFP_KERNEL); | |
4681 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
4682 | *p = 0x56; | |
4683 | pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
4684 | p); | |
4685 | pr_err("If allocated object is overwritten then not detectable\n\n"); | |
4686 | validate_slab_cache(kmalloc_caches[6]); | |
4687 | ||
4688 | pr_err("\nB. Corruption after free\n"); | |
4689 | p = kzalloc(128, GFP_KERNEL); | |
4690 | kfree(p); | |
4691 | *p = 0x78; | |
4692 | pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
4693 | validate_slab_cache(kmalloc_caches[7]); | |
4694 | ||
4695 | p = kzalloc(256, GFP_KERNEL); | |
4696 | kfree(p); | |
4697 | p[50] = 0x9a; | |
4698 | pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); | |
4699 | validate_slab_cache(kmalloc_caches[8]); | |
4700 | ||
4701 | p = kzalloc(512, GFP_KERNEL); | |
4702 | kfree(p); | |
4703 | p[512] = 0xab; | |
4704 | pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
4705 | validate_slab_cache(kmalloc_caches[9]); | |
4706 | } | |
4707 | #else | |
4708 | #ifdef CONFIG_SYSFS | |
4709 | static void resiliency_test(void) {}; | |
4710 | #endif | |
4711 | #endif | |
4712 | ||
4713 | #ifdef CONFIG_SYSFS | |
4714 | enum slab_stat_type { | |
4715 | SL_ALL, /* All slabs */ | |
4716 | SL_PARTIAL, /* Only partially allocated slabs */ | |
4717 | SL_CPU, /* Only slabs used for cpu caches */ | |
4718 | SL_OBJECTS, /* Determine allocated objects not slabs */ | |
4719 | SL_TOTAL /* Determine object capacity not slabs */ | |
4720 | }; | |
4721 | ||
4722 | #define SO_ALL (1 << SL_ALL) | |
4723 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
4724 | #define SO_CPU (1 << SL_CPU) | |
4725 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
4726 | #define SO_TOTAL (1 << SL_TOTAL) | |
4727 | ||
4728 | #ifdef CONFIG_MEMCG | |
4729 | static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON); | |
4730 | ||
4731 | static int __init setup_slub_memcg_sysfs(char *str) | |
4732 | { | |
4733 | int v; | |
4734 | ||
4735 | if (get_option(&str, &v) > 0) | |
4736 | memcg_sysfs_enabled = v; | |
4737 | ||
4738 | return 1; | |
4739 | } | |
4740 | ||
4741 | __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs); | |
4742 | #endif | |
4743 | ||
4744 | static ssize_t show_slab_objects(struct kmem_cache *s, | |
4745 | char *buf, unsigned long flags) | |
4746 | { | |
4747 | unsigned long total = 0; | |
4748 | int node; | |
4749 | int x; | |
4750 | unsigned long *nodes; | |
4751 | ||
4752 | nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
4753 | if (!nodes) | |
4754 | return -ENOMEM; | |
4755 | ||
4756 | if (flags & SO_CPU) { | |
4757 | int cpu; | |
4758 | ||
4759 | for_each_possible_cpu(cpu) { | |
4760 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, | |
4761 | cpu); | |
4762 | int node; | |
4763 | struct page *page; | |
4764 | ||
4765 | page = READ_ONCE(c->page); | |
4766 | if (!page) | |
4767 | continue; | |
4768 | ||
4769 | node = page_to_nid(page); | |
4770 | if (flags & SO_TOTAL) | |
4771 | x = page->objects; | |
4772 | else if (flags & SO_OBJECTS) | |
4773 | x = page->inuse; | |
4774 | else | |
4775 | x = 1; | |
4776 | ||
4777 | total += x; | |
4778 | nodes[node] += x; | |
4779 | ||
4780 | page = slub_percpu_partial_read_once(c); | |
4781 | if (page) { | |
4782 | node = page_to_nid(page); | |
4783 | if (flags & SO_TOTAL) | |
4784 | WARN_ON_ONCE(1); | |
4785 | else if (flags & SO_OBJECTS) | |
4786 | WARN_ON_ONCE(1); | |
4787 | else | |
4788 | x = page->pages; | |
4789 | total += x; | |
4790 | nodes[node] += x; | |
4791 | } | |
4792 | } | |
4793 | } | |
4794 | ||
4795 | get_online_mems(); | |
4796 | #ifdef CONFIG_SLUB_DEBUG | |
4797 | if (flags & SO_ALL) { | |
4798 | struct kmem_cache_node *n; | |
4799 | ||
4800 | for_each_kmem_cache_node(s, node, n) { | |
4801 | ||
4802 | if (flags & SO_TOTAL) | |
4803 | x = atomic_long_read(&n->total_objects); | |
4804 | else if (flags & SO_OBJECTS) | |
4805 | x = atomic_long_read(&n->total_objects) - | |
4806 | count_partial(n, count_free); | |
4807 | else | |
4808 | x = atomic_long_read(&n->nr_slabs); | |
4809 | total += x; | |
4810 | nodes[node] += x; | |
4811 | } | |
4812 | ||
4813 | } else | |
4814 | #endif | |
4815 | if (flags & SO_PARTIAL) { | |
4816 | struct kmem_cache_node *n; | |
4817 | ||
4818 | for_each_kmem_cache_node(s, node, n) { | |
4819 | if (flags & SO_TOTAL) | |
4820 | x = count_partial(n, count_total); | |
4821 | else if (flags & SO_OBJECTS) | |
4822 | x = count_partial(n, count_inuse); | |
4823 | else | |
4824 | x = n->nr_partial; | |
4825 | total += x; | |
4826 | nodes[node] += x; | |
4827 | } | |
4828 | } | |
4829 | x = sprintf(buf, "%lu", total); | |
4830 | #ifdef CONFIG_NUMA | |
4831 | for (node = 0; node < nr_node_ids; node++) | |
4832 | if (nodes[node]) | |
4833 | x += sprintf(buf + x, " N%d=%lu", | |
4834 | node, nodes[node]); | |
4835 | #endif | |
4836 | put_online_mems(); | |
4837 | kfree(nodes); | |
4838 | return x + sprintf(buf + x, "\n"); | |
4839 | } | |
4840 | ||
4841 | #ifdef CONFIG_SLUB_DEBUG | |
4842 | static int any_slab_objects(struct kmem_cache *s) | |
4843 | { | |
4844 | int node; | |
4845 | struct kmem_cache_node *n; | |
4846 | ||
4847 | for_each_kmem_cache_node(s, node, n) | |
4848 | if (atomic_long_read(&n->total_objects)) | |
4849 | return 1; | |
4850 | ||
4851 | return 0; | |
4852 | } | |
4853 | #endif | |
4854 | ||
4855 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
4856 | #define to_slab(n) container_of(n, struct kmem_cache, kobj) | |
4857 | ||
4858 | struct slab_attribute { | |
4859 | struct attribute attr; | |
4860 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
4861 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
4862 | }; | |
4863 | ||
4864 | #define SLAB_ATTR_RO(_name) \ | |
4865 | static struct slab_attribute _name##_attr = \ | |
4866 | __ATTR(_name, 0400, _name##_show, NULL) | |
4867 | ||
4868 | #define SLAB_ATTR(_name) \ | |
4869 | static struct slab_attribute _name##_attr = \ | |
4870 | __ATTR(_name, 0600, _name##_show, _name##_store) | |
4871 | ||
4872 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) | |
4873 | { | |
4874 | return sprintf(buf, "%d\n", s->size); | |
4875 | } | |
4876 | SLAB_ATTR_RO(slab_size); | |
4877 | ||
4878 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
4879 | { | |
4880 | return sprintf(buf, "%d\n", s->align); | |
4881 | } | |
4882 | SLAB_ATTR_RO(align); | |
4883 | ||
4884 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
4885 | { | |
4886 | return sprintf(buf, "%d\n", s->object_size); | |
4887 | } | |
4888 | SLAB_ATTR_RO(object_size); | |
4889 | ||
4890 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
4891 | { | |
4892 | return sprintf(buf, "%d\n", oo_objects(s->oo)); | |
4893 | } | |
4894 | SLAB_ATTR_RO(objs_per_slab); | |
4895 | ||
4896 | static ssize_t order_store(struct kmem_cache *s, | |
4897 | const char *buf, size_t length) | |
4898 | { | |
4899 | unsigned long order; | |
4900 | int err; | |
4901 | ||
4902 | err = kstrtoul(buf, 10, &order); | |
4903 | if (err) | |
4904 | return err; | |
4905 | ||
4906 | if (order > slub_max_order || order < slub_min_order) | |
4907 | return -EINVAL; | |
4908 | ||
4909 | calculate_sizes(s, order); | |
4910 | return length; | |
4911 | } | |
4912 | ||
4913 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
4914 | { | |
4915 | return sprintf(buf, "%d\n", oo_order(s->oo)); | |
4916 | } | |
4917 | SLAB_ATTR(order); | |
4918 | ||
4919 | static ssize_t min_partial_show(struct kmem_cache *s, char *buf) | |
4920 | { | |
4921 | return sprintf(buf, "%lu\n", s->min_partial); | |
4922 | } | |
4923 | ||
4924 | static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, | |
4925 | size_t length) | |
4926 | { | |
4927 | unsigned long min; | |
4928 | int err; | |
4929 | ||
4930 | err = kstrtoul(buf, 10, &min); | |
4931 | if (err) | |
4932 | return err; | |
4933 | ||
4934 | set_min_partial(s, min); | |
4935 | return length; | |
4936 | } | |
4937 | SLAB_ATTR(min_partial); | |
4938 | ||
4939 | static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) | |
4940 | { | |
4941 | return sprintf(buf, "%u\n", slub_cpu_partial(s)); | |
4942 | } | |
4943 | ||
4944 | static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, | |
4945 | size_t length) | |
4946 | { | |
4947 | unsigned long objects; | |
4948 | int err; | |
4949 | ||
4950 | err = kstrtoul(buf, 10, &objects); | |
4951 | if (err) | |
4952 | return err; | |
4953 | if (objects && !kmem_cache_has_cpu_partial(s)) | |
4954 | return -EINVAL; | |
4955 | ||
4956 | slub_set_cpu_partial(s, objects); | |
4957 | flush_all(s); | |
4958 | return length; | |
4959 | } | |
4960 | SLAB_ATTR(cpu_partial); | |
4961 | ||
4962 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
4963 | { | |
4964 | if (!s->ctor) | |
4965 | return 0; | |
4966 | return sprintf(buf, "%pS\n", s->ctor); | |
4967 | } | |
4968 | SLAB_ATTR_RO(ctor); | |
4969 | ||
4970 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) | |
4971 | { | |
4972 | return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1); | |
4973 | } | |
4974 | SLAB_ATTR_RO(aliases); | |
4975 | ||
4976 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
4977 | { | |
4978 | return show_slab_objects(s, buf, SO_PARTIAL); | |
4979 | } | |
4980 | SLAB_ATTR_RO(partial); | |
4981 | ||
4982 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
4983 | { | |
4984 | return show_slab_objects(s, buf, SO_CPU); | |
4985 | } | |
4986 | SLAB_ATTR_RO(cpu_slabs); | |
4987 | ||
4988 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
4989 | { | |
4990 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); | |
4991 | } | |
4992 | SLAB_ATTR_RO(objects); | |
4993 | ||
4994 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) | |
4995 | { | |
4996 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); | |
4997 | } | |
4998 | SLAB_ATTR_RO(objects_partial); | |
4999 | ||
5000 | static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) | |
5001 | { | |
5002 | int objects = 0; | |
5003 | int pages = 0; | |
5004 | int cpu; | |
5005 | int len; | |
5006 | ||
5007 | for_each_online_cpu(cpu) { | |
5008 | struct page *page; | |
5009 | ||
5010 | page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); | |
5011 | ||
5012 | if (page) { | |
5013 | pages += page->pages; | |
5014 | objects += page->pobjects; | |
5015 | } | |
5016 | } | |
5017 | ||
5018 | len = sprintf(buf, "%d(%d)", objects, pages); | |
5019 | ||
5020 | #ifdef CONFIG_SMP | |
5021 | for_each_online_cpu(cpu) { | |
5022 | struct page *page; | |
5023 | ||
5024 | page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); | |
5025 | ||
5026 | if (page && len < PAGE_SIZE - 20) | |
5027 | len += sprintf(buf + len, " C%d=%d(%d)", cpu, | |
5028 | page->pobjects, page->pages); | |
5029 | } | |
5030 | #endif | |
5031 | return len + sprintf(buf + len, "\n"); | |
5032 | } | |
5033 | SLAB_ATTR_RO(slabs_cpu_partial); | |
5034 | ||
5035 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
5036 | { | |
5037 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
5038 | } | |
5039 | ||
5040 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
5041 | const char *buf, size_t length) | |
5042 | { | |
5043 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
5044 | if (buf[0] == '1') | |
5045 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
5046 | return length; | |
5047 | } | |
5048 | SLAB_ATTR(reclaim_account); | |
5049 | ||
5050 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
5051 | { | |
5052 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); | |
5053 | } | |
5054 | SLAB_ATTR_RO(hwcache_align); | |
5055 | ||
5056 | #ifdef CONFIG_ZONE_DMA | |
5057 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
5058 | { | |
5059 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
5060 | } | |
5061 | SLAB_ATTR_RO(cache_dma); | |
5062 | #endif | |
5063 | ||
5064 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
5065 | { | |
5066 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU)); | |
5067 | } | |
5068 | SLAB_ATTR_RO(destroy_by_rcu); | |
5069 | ||
5070 | static ssize_t reserved_show(struct kmem_cache *s, char *buf) | |
5071 | { | |
5072 | return sprintf(buf, "%d\n", s->reserved); | |
5073 | } | |
5074 | SLAB_ATTR_RO(reserved); | |
5075 | ||
5076 | #ifdef CONFIG_SLUB_DEBUG | |
5077 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
5078 | { | |
5079 | return show_slab_objects(s, buf, SO_ALL); | |
5080 | } | |
5081 | SLAB_ATTR_RO(slabs); | |
5082 | ||
5083 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) | |
5084 | { | |
5085 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); | |
5086 | } | |
5087 | SLAB_ATTR_RO(total_objects); | |
5088 | ||
5089 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
5090 | { | |
5091 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS)); | |
5092 | } | |
5093 | ||
5094 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
5095 | const char *buf, size_t length) | |
5096 | { | |
5097 | s->flags &= ~SLAB_CONSISTENCY_CHECKS; | |
5098 | if (buf[0] == '1') { | |
5099 | s->flags &= ~__CMPXCHG_DOUBLE; | |
5100 | s->flags |= SLAB_CONSISTENCY_CHECKS; | |
5101 | } | |
5102 | return length; | |
5103 | } | |
5104 | SLAB_ATTR(sanity_checks); | |
5105 | ||
5106 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
5107 | { | |
5108 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
5109 | } | |
5110 | ||
5111 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
5112 | size_t length) | |
5113 | { | |
5114 | /* | |
5115 | * Tracing a merged cache is going to give confusing results | |
5116 | * as well as cause other issues like converting a mergeable | |
5117 | * cache into an umergeable one. | |
5118 | */ | |
5119 | if (s->refcount > 1) | |
5120 | return -EINVAL; | |
5121 | ||
5122 | s->flags &= ~SLAB_TRACE; | |
5123 | if (buf[0] == '1') { | |
5124 | s->flags &= ~__CMPXCHG_DOUBLE; | |
5125 | s->flags |= SLAB_TRACE; | |
5126 | } | |
5127 | return length; | |
5128 | } | |
5129 | SLAB_ATTR(trace); | |
5130 | ||
5131 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
5132 | { | |
5133 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
5134 | } | |
5135 | ||
5136 | static ssize_t red_zone_store(struct kmem_cache *s, | |
5137 | const char *buf, size_t length) | |
5138 | { | |
5139 | if (any_slab_objects(s)) | |
5140 | return -EBUSY; | |
5141 | ||
5142 | s->flags &= ~SLAB_RED_ZONE; | |
5143 | if (buf[0] == '1') { | |
5144 | s->flags |= SLAB_RED_ZONE; | |
5145 | } | |
5146 | calculate_sizes(s, -1); | |
5147 | return length; | |
5148 | } | |
5149 | SLAB_ATTR(red_zone); | |
5150 | ||
5151 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
5152 | { | |
5153 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
5154 | } | |
5155 | ||
5156 | static ssize_t poison_store(struct kmem_cache *s, | |
5157 | const char *buf, size_t length) | |
5158 | { | |
5159 | if (any_slab_objects(s)) | |
5160 | return -EBUSY; | |
5161 | ||
5162 | s->flags &= ~SLAB_POISON; | |
5163 | if (buf[0] == '1') { | |
5164 | s->flags |= SLAB_POISON; | |
5165 | } | |
5166 | calculate_sizes(s, -1); | |
5167 | return length; | |
5168 | } | |
5169 | SLAB_ATTR(poison); | |
5170 | ||
5171 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
5172 | { | |
5173 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
5174 | } | |
5175 | ||
5176 | static ssize_t store_user_store(struct kmem_cache *s, | |
5177 | const char *buf, size_t length) | |
5178 | { | |
5179 | if (any_slab_objects(s)) | |
5180 | return -EBUSY; | |
5181 | ||
5182 | s->flags &= ~SLAB_STORE_USER; | |
5183 | if (buf[0] == '1') { | |
5184 | s->flags &= ~__CMPXCHG_DOUBLE; | |
5185 | s->flags |= SLAB_STORE_USER; | |
5186 | } | |
5187 | calculate_sizes(s, -1); | |
5188 | return length; | |
5189 | } | |
5190 | SLAB_ATTR(store_user); | |
5191 | ||
5192 | static ssize_t validate_show(struct kmem_cache *s, char *buf) | |
5193 | { | |
5194 | return 0; | |
5195 | } | |
5196 | ||
5197 | static ssize_t validate_store(struct kmem_cache *s, | |
5198 | const char *buf, size_t length) | |
5199 | { | |
5200 | int ret = -EINVAL; | |
5201 | ||
5202 | if (buf[0] == '1') { | |
5203 | ret = validate_slab_cache(s); | |
5204 | if (ret >= 0) | |
5205 | ret = length; | |
5206 | } | |
5207 | return ret; | |
5208 | } | |
5209 | SLAB_ATTR(validate); | |
5210 | ||
5211 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) | |
5212 | { | |
5213 | if (!(s->flags & SLAB_STORE_USER)) | |
5214 | return -ENOSYS; | |
5215 | return list_locations(s, buf, TRACK_ALLOC); | |
5216 | } | |
5217 | SLAB_ATTR_RO(alloc_calls); | |
5218 | ||
5219 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
5220 | { | |
5221 | if (!(s->flags & SLAB_STORE_USER)) | |
5222 | return -ENOSYS; | |
5223 | return list_locations(s, buf, TRACK_FREE); | |
5224 | } | |
5225 | SLAB_ATTR_RO(free_calls); | |
5226 | #endif /* CONFIG_SLUB_DEBUG */ | |
5227 | ||
5228 | #ifdef CONFIG_FAILSLAB | |
5229 | static ssize_t failslab_show(struct kmem_cache *s, char *buf) | |
5230 | { | |
5231 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); | |
5232 | } | |
5233 | ||
5234 | static ssize_t failslab_store(struct kmem_cache *s, const char *buf, | |
5235 | size_t length) | |
5236 | { | |
5237 | if (s->refcount > 1) | |
5238 | return -EINVAL; | |
5239 | ||
5240 | s->flags &= ~SLAB_FAILSLAB; | |
5241 | if (buf[0] == '1') | |
5242 | s->flags |= SLAB_FAILSLAB; | |
5243 | return length; | |
5244 | } | |
5245 | SLAB_ATTR(failslab); | |
5246 | #endif | |
5247 | ||
5248 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) | |
5249 | { | |
5250 | return 0; | |
5251 | } | |
5252 | ||
5253 | static ssize_t shrink_store(struct kmem_cache *s, | |
5254 | const char *buf, size_t length) | |
5255 | { | |
5256 | if (buf[0] == '1') | |
5257 | kmem_cache_shrink(s); | |
5258 | else | |
5259 | return -EINVAL; | |
5260 | return length; | |
5261 | } | |
5262 | SLAB_ATTR(shrink); | |
5263 | ||
5264 | #ifdef CONFIG_NUMA | |
5265 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) | |
5266 | { | |
5267 | return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); | |
5268 | } | |
5269 | ||
5270 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, | |
5271 | const char *buf, size_t length) | |
5272 | { | |
5273 | unsigned long ratio; | |
5274 | int err; | |
5275 | ||
5276 | err = kstrtoul(buf, 10, &ratio); | |
5277 | if (err) | |
5278 | return err; | |
5279 | ||
5280 | if (ratio <= 100) | |
5281 | s->remote_node_defrag_ratio = ratio * 10; | |
5282 | ||
5283 | return length; | |
5284 | } | |
5285 | SLAB_ATTR(remote_node_defrag_ratio); | |
5286 | #endif | |
5287 | ||
5288 | #ifdef CONFIG_SLUB_STATS | |
5289 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) | |
5290 | { | |
5291 | unsigned long sum = 0; | |
5292 | int cpu; | |
5293 | int len; | |
5294 | int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); | |
5295 | ||
5296 | if (!data) | |
5297 | return -ENOMEM; | |
5298 | ||
5299 | for_each_online_cpu(cpu) { | |
5300 | unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; | |
5301 | ||
5302 | data[cpu] = x; | |
5303 | sum += x; | |
5304 | } | |
5305 | ||
5306 | len = sprintf(buf, "%lu", sum); | |
5307 | ||
5308 | #ifdef CONFIG_SMP | |
5309 | for_each_online_cpu(cpu) { | |
5310 | if (data[cpu] && len < PAGE_SIZE - 20) | |
5311 | len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); | |
5312 | } | |
5313 | #endif | |
5314 | kfree(data); | |
5315 | return len + sprintf(buf + len, "\n"); | |
5316 | } | |
5317 | ||
5318 | static void clear_stat(struct kmem_cache *s, enum stat_item si) | |
5319 | { | |
5320 | int cpu; | |
5321 | ||
5322 | for_each_online_cpu(cpu) | |
5323 | per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; | |
5324 | } | |
5325 | ||
5326 | #define STAT_ATTR(si, text) \ | |
5327 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ | |
5328 | { \ | |
5329 | return show_stat(s, buf, si); \ | |
5330 | } \ | |
5331 | static ssize_t text##_store(struct kmem_cache *s, \ | |
5332 | const char *buf, size_t length) \ | |
5333 | { \ | |
5334 | if (buf[0] != '0') \ | |
5335 | return -EINVAL; \ | |
5336 | clear_stat(s, si); \ | |
5337 | return length; \ | |
5338 | } \ | |
5339 | SLAB_ATTR(text); \ | |
5340 | ||
5341 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); | |
5342 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); | |
5343 | STAT_ATTR(FREE_FASTPATH, free_fastpath); | |
5344 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); | |
5345 | STAT_ATTR(FREE_FROZEN, free_frozen); | |
5346 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); | |
5347 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); | |
5348 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); | |
5349 | STAT_ATTR(ALLOC_SLAB, alloc_slab); | |
5350 | STAT_ATTR(ALLOC_REFILL, alloc_refill); | |
5351 | STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); | |
5352 | STAT_ATTR(FREE_SLAB, free_slab); | |
5353 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); | |
5354 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); | |
5355 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); | |
5356 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); | |
5357 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); | |
5358 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); | |
5359 | STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); | |
5360 | STAT_ATTR(ORDER_FALLBACK, order_fallback); | |
5361 | STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); | |
5362 | STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); | |
5363 | STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); | |
5364 | STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); | |
5365 | STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); | |
5366 | STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); | |
5367 | #endif | |
5368 | ||
5369 | static struct attribute *slab_attrs[] = { | |
5370 | &slab_size_attr.attr, | |
5371 | &object_size_attr.attr, | |
5372 | &objs_per_slab_attr.attr, | |
5373 | &order_attr.attr, | |
5374 | &min_partial_attr.attr, | |
5375 | &cpu_partial_attr.attr, | |
5376 | &objects_attr.attr, | |
5377 | &objects_partial_attr.attr, | |
5378 | &partial_attr.attr, | |
5379 | &cpu_slabs_attr.attr, | |
5380 | &ctor_attr.attr, | |
5381 | &aliases_attr.attr, | |
5382 | &align_attr.attr, | |
5383 | &hwcache_align_attr.attr, | |
5384 | &reclaim_account_attr.attr, | |
5385 | &destroy_by_rcu_attr.attr, | |
5386 | &shrink_attr.attr, | |
5387 | &reserved_attr.attr, | |
5388 | &slabs_cpu_partial_attr.attr, | |
5389 | #ifdef CONFIG_SLUB_DEBUG | |
5390 | &total_objects_attr.attr, | |
5391 | &slabs_attr.attr, | |
5392 | &sanity_checks_attr.attr, | |
5393 | &trace_attr.attr, | |
5394 | &red_zone_attr.attr, | |
5395 | &poison_attr.attr, | |
5396 | &store_user_attr.attr, | |
5397 | &validate_attr.attr, | |
5398 | &alloc_calls_attr.attr, | |
5399 | &free_calls_attr.attr, | |
5400 | #endif | |
5401 | #ifdef CONFIG_ZONE_DMA | |
5402 | &cache_dma_attr.attr, | |
5403 | #endif | |
5404 | #ifdef CONFIG_NUMA | |
5405 | &remote_node_defrag_ratio_attr.attr, | |
5406 | #endif | |
5407 | #ifdef CONFIG_SLUB_STATS | |
5408 | &alloc_fastpath_attr.attr, | |
5409 | &alloc_slowpath_attr.attr, | |
5410 | &free_fastpath_attr.attr, | |
5411 | &free_slowpath_attr.attr, | |
5412 | &free_frozen_attr.attr, | |
5413 | &free_add_partial_attr.attr, | |
5414 | &free_remove_partial_attr.attr, | |
5415 | &alloc_from_partial_attr.attr, | |
5416 | &alloc_slab_attr.attr, | |
5417 | &alloc_refill_attr.attr, | |
5418 | &alloc_node_mismatch_attr.attr, | |
5419 | &free_slab_attr.attr, | |
5420 | &cpuslab_flush_attr.attr, | |
5421 | &deactivate_full_attr.attr, | |
5422 | &deactivate_empty_attr.attr, | |
5423 | &deactivate_to_head_attr.attr, | |
5424 | &deactivate_to_tail_attr.attr, | |
5425 | &deactivate_remote_frees_attr.attr, | |
5426 | &deactivate_bypass_attr.attr, | |
5427 | &order_fallback_attr.attr, | |
5428 | &cmpxchg_double_fail_attr.attr, | |
5429 | &cmpxchg_double_cpu_fail_attr.attr, | |
5430 | &cpu_partial_alloc_attr.attr, | |
5431 | &cpu_partial_free_attr.attr, | |
5432 | &cpu_partial_node_attr.attr, | |
5433 | &cpu_partial_drain_attr.attr, | |
5434 | #endif | |
5435 | #ifdef CONFIG_FAILSLAB | |
5436 | &failslab_attr.attr, | |
5437 | #endif | |
5438 | ||
5439 | NULL | |
5440 | }; | |
5441 | ||
5442 | static const struct attribute_group slab_attr_group = { | |
5443 | .attrs = slab_attrs, | |
5444 | }; | |
5445 | ||
5446 | static ssize_t slab_attr_show(struct kobject *kobj, | |
5447 | struct attribute *attr, | |
5448 | char *buf) | |
5449 | { | |
5450 | struct slab_attribute *attribute; | |
5451 | struct kmem_cache *s; | |
5452 | int err; | |
5453 | ||
5454 | attribute = to_slab_attr(attr); | |
5455 | s = to_slab(kobj); | |
5456 | ||
5457 | if (!attribute->show) | |
5458 | return -EIO; | |
5459 | ||
5460 | err = attribute->show(s, buf); | |
5461 | ||
5462 | return err; | |
5463 | } | |
5464 | ||
5465 | static ssize_t slab_attr_store(struct kobject *kobj, | |
5466 | struct attribute *attr, | |
5467 | const char *buf, size_t len) | |
5468 | { | |
5469 | struct slab_attribute *attribute; | |
5470 | struct kmem_cache *s; | |
5471 | int err; | |
5472 | ||
5473 | attribute = to_slab_attr(attr); | |
5474 | s = to_slab(kobj); | |
5475 | ||
5476 | if (!attribute->store) | |
5477 | return -EIO; | |
5478 | ||
5479 | err = attribute->store(s, buf, len); | |
5480 | #ifdef CONFIG_MEMCG | |
5481 | if (slab_state >= FULL && err >= 0 && is_root_cache(s)) { | |
5482 | struct kmem_cache *c; | |
5483 | ||
5484 | mutex_lock(&slab_mutex); | |
5485 | if (s->max_attr_size < len) | |
5486 | s->max_attr_size = len; | |
5487 | ||
5488 | /* | |
5489 | * This is a best effort propagation, so this function's return | |
5490 | * value will be determined by the parent cache only. This is | |
5491 | * basically because not all attributes will have a well | |
5492 | * defined semantics for rollbacks - most of the actions will | |
5493 | * have permanent effects. | |
5494 | * | |
5495 | * Returning the error value of any of the children that fail | |
5496 | * is not 100 % defined, in the sense that users seeing the | |
5497 | * error code won't be able to know anything about the state of | |
5498 | * the cache. | |
5499 | * | |
5500 | * Only returning the error code for the parent cache at least | |
5501 | * has well defined semantics. The cache being written to | |
5502 | * directly either failed or succeeded, in which case we loop | |
5503 | * through the descendants with best-effort propagation. | |
5504 | */ | |
5505 | for_each_memcg_cache(c, s) | |
5506 | attribute->store(c, buf, len); | |
5507 | mutex_unlock(&slab_mutex); | |
5508 | } | |
5509 | #endif | |
5510 | return err; | |
5511 | } | |
5512 | ||
5513 | static void memcg_propagate_slab_attrs(struct kmem_cache *s) | |
5514 | { | |
5515 | #ifdef CONFIG_MEMCG | |
5516 | int i; | |
5517 | char *buffer = NULL; | |
5518 | struct kmem_cache *root_cache; | |
5519 | ||
5520 | if (is_root_cache(s)) | |
5521 | return; | |
5522 | ||
5523 | root_cache = s->memcg_params.root_cache; | |
5524 | ||
5525 | /* | |
5526 | * This mean this cache had no attribute written. Therefore, no point | |
5527 | * in copying default values around | |
5528 | */ | |
5529 | if (!root_cache->max_attr_size) | |
5530 | return; | |
5531 | ||
5532 | for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) { | |
5533 | char mbuf[64]; | |
5534 | char *buf; | |
5535 | struct slab_attribute *attr = to_slab_attr(slab_attrs[i]); | |
5536 | ssize_t len; | |
5537 | ||
5538 | if (!attr || !attr->store || !attr->show) | |
5539 | continue; | |
5540 | ||
5541 | /* | |
5542 | * It is really bad that we have to allocate here, so we will | |
5543 | * do it only as a fallback. If we actually allocate, though, | |
5544 | * we can just use the allocated buffer until the end. | |
5545 | * | |
5546 | * Most of the slub attributes will tend to be very small in | |
5547 | * size, but sysfs allows buffers up to a page, so they can | |
5548 | * theoretically happen. | |
5549 | */ | |
5550 | if (buffer) | |
5551 | buf = buffer; | |
5552 | else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf)) | |
5553 | buf = mbuf; | |
5554 | else { | |
5555 | buffer = (char *) get_zeroed_page(GFP_KERNEL); | |
5556 | if (WARN_ON(!buffer)) | |
5557 | continue; | |
5558 | buf = buffer; | |
5559 | } | |
5560 | ||
5561 | len = attr->show(root_cache, buf); | |
5562 | if (len > 0) | |
5563 | attr->store(s, buf, len); | |
5564 | } | |
5565 | ||
5566 | if (buffer) | |
5567 | free_page((unsigned long)buffer); | |
5568 | #endif | |
5569 | } | |
5570 | ||
5571 | static void kmem_cache_release(struct kobject *k) | |
5572 | { | |
5573 | slab_kmem_cache_release(to_slab(k)); | |
5574 | } | |
5575 | ||
5576 | static const struct sysfs_ops slab_sysfs_ops = { | |
5577 | .show = slab_attr_show, | |
5578 | .store = slab_attr_store, | |
5579 | }; | |
5580 | ||
5581 | static struct kobj_type slab_ktype = { | |
5582 | .sysfs_ops = &slab_sysfs_ops, | |
5583 | .release = kmem_cache_release, | |
5584 | }; | |
5585 | ||
5586 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
5587 | { | |
5588 | struct kobj_type *ktype = get_ktype(kobj); | |
5589 | ||
5590 | if (ktype == &slab_ktype) | |
5591 | return 1; | |
5592 | return 0; | |
5593 | } | |
5594 | ||
5595 | static const struct kset_uevent_ops slab_uevent_ops = { | |
5596 | .filter = uevent_filter, | |
5597 | }; | |
5598 | ||
5599 | static struct kset *slab_kset; | |
5600 | ||
5601 | static inline struct kset *cache_kset(struct kmem_cache *s) | |
5602 | { | |
5603 | #ifdef CONFIG_MEMCG | |
5604 | if (!is_root_cache(s)) | |
5605 | return s->memcg_params.root_cache->memcg_kset; | |
5606 | #endif | |
5607 | return slab_kset; | |
5608 | } | |
5609 | ||
5610 | #define ID_STR_LENGTH 64 | |
5611 | ||
5612 | /* Create a unique string id for a slab cache: | |
5613 | * | |
5614 | * Format :[flags-]size | |
5615 | */ | |
5616 | static char *create_unique_id(struct kmem_cache *s) | |
5617 | { | |
5618 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
5619 | char *p = name; | |
5620 | ||
5621 | BUG_ON(!name); | |
5622 | ||
5623 | *p++ = ':'; | |
5624 | /* | |
5625 | * First flags affecting slabcache operations. We will only | |
5626 | * get here for aliasable slabs so we do not need to support | |
5627 | * too many flags. The flags here must cover all flags that | |
5628 | * are matched during merging to guarantee that the id is | |
5629 | * unique. | |
5630 | */ | |
5631 | if (s->flags & SLAB_CACHE_DMA) | |
5632 | *p++ = 'd'; | |
5633 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
5634 | *p++ = 'a'; | |
5635 | if (s->flags & SLAB_CONSISTENCY_CHECKS) | |
5636 | *p++ = 'F'; | |
5637 | if (s->flags & SLAB_ACCOUNT) | |
5638 | *p++ = 'A'; | |
5639 | if (p != name + 1) | |
5640 | *p++ = '-'; | |
5641 | p += sprintf(p, "%07d", s->size); | |
5642 | ||
5643 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
5644 | return name; | |
5645 | } | |
5646 | ||
5647 | static void sysfs_slab_remove_workfn(struct work_struct *work) | |
5648 | { | |
5649 | struct kmem_cache *s = | |
5650 | container_of(work, struct kmem_cache, kobj_remove_work); | |
5651 | ||
5652 | if (!s->kobj.state_in_sysfs) | |
5653 | /* | |
5654 | * For a memcg cache, this may be called during | |
5655 | * deactivation and again on shutdown. Remove only once. | |
5656 | * A cache is never shut down before deactivation is | |
5657 | * complete, so no need to worry about synchronization. | |
5658 | */ | |
5659 | goto out; | |
5660 | ||
5661 | #ifdef CONFIG_MEMCG | |
5662 | kset_unregister(s->memcg_kset); | |
5663 | #endif | |
5664 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
5665 | kobject_del(&s->kobj); | |
5666 | out: | |
5667 | kobject_put(&s->kobj); | |
5668 | } | |
5669 | ||
5670 | static int sysfs_slab_add(struct kmem_cache *s) | |
5671 | { | |
5672 | int err; | |
5673 | const char *name; | |
5674 | struct kset *kset = cache_kset(s); | |
5675 | int unmergeable = slab_unmergeable(s); | |
5676 | ||
5677 | INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn); | |
5678 | ||
5679 | if (!kset) { | |
5680 | kobject_init(&s->kobj, &slab_ktype); | |
5681 | return 0; | |
5682 | } | |
5683 | ||
5684 | if (!unmergeable && disable_higher_order_debug && | |
5685 | (slub_debug & DEBUG_METADATA_FLAGS)) | |
5686 | unmergeable = 1; | |
5687 | ||
5688 | if (unmergeable) { | |
5689 | /* | |
5690 | * Slabcache can never be merged so we can use the name proper. | |
5691 | * This is typically the case for debug situations. In that | |
5692 | * case we can catch duplicate names easily. | |
5693 | */ | |
5694 | sysfs_remove_link(&slab_kset->kobj, s->name); | |
5695 | name = s->name; | |
5696 | } else { | |
5697 | /* | |
5698 | * Create a unique name for the slab as a target | |
5699 | * for the symlinks. | |
5700 | */ | |
5701 | name = create_unique_id(s); | |
5702 | } | |
5703 | ||
5704 | s->kobj.kset = kset; | |
5705 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name); | |
5706 | if (err) | |
5707 | goto out; | |
5708 | ||
5709 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
5710 | if (err) | |
5711 | goto out_del_kobj; | |
5712 | ||
5713 | #ifdef CONFIG_MEMCG | |
5714 | if (is_root_cache(s) && memcg_sysfs_enabled) { | |
5715 | s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj); | |
5716 | if (!s->memcg_kset) { | |
5717 | err = -ENOMEM; | |
5718 | goto out_del_kobj; | |
5719 | } | |
5720 | } | |
5721 | #endif | |
5722 | ||
5723 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
5724 | if (!unmergeable) { | |
5725 | /* Setup first alias */ | |
5726 | sysfs_slab_alias(s, s->name); | |
5727 | } | |
5728 | out: | |
5729 | if (!unmergeable) | |
5730 | kfree(name); | |
5731 | return err; | |
5732 | out_del_kobj: | |
5733 | kobject_del(&s->kobj); | |
5734 | goto out; | |
5735 | } | |
5736 | ||
5737 | static void sysfs_slab_remove(struct kmem_cache *s) | |
5738 | { | |
5739 | if (slab_state < FULL) | |
5740 | /* | |
5741 | * Sysfs has not been setup yet so no need to remove the | |
5742 | * cache from sysfs. | |
5743 | */ | |
5744 | return; | |
5745 | ||
5746 | kobject_get(&s->kobj); | |
5747 | schedule_work(&s->kobj_remove_work); | |
5748 | } | |
5749 | ||
5750 | void sysfs_slab_release(struct kmem_cache *s) | |
5751 | { | |
5752 | if (slab_state >= FULL) | |
5753 | kobject_put(&s->kobj); | |
5754 | } | |
5755 | ||
5756 | /* | |
5757 | * Need to buffer aliases during bootup until sysfs becomes | |
5758 | * available lest we lose that information. | |
5759 | */ | |
5760 | struct saved_alias { | |
5761 | struct kmem_cache *s; | |
5762 | const char *name; | |
5763 | struct saved_alias *next; | |
5764 | }; | |
5765 | ||
5766 | static struct saved_alias *alias_list; | |
5767 | ||
5768 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
5769 | { | |
5770 | struct saved_alias *al; | |
5771 | ||
5772 | if (slab_state == FULL) { | |
5773 | /* | |
5774 | * If we have a leftover link then remove it. | |
5775 | */ | |
5776 | sysfs_remove_link(&slab_kset->kobj, name); | |
5777 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); | |
5778 | } | |
5779 | ||
5780 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
5781 | if (!al) | |
5782 | return -ENOMEM; | |
5783 | ||
5784 | al->s = s; | |
5785 | al->name = name; | |
5786 | al->next = alias_list; | |
5787 | alias_list = al; | |
5788 | return 0; | |
5789 | } | |
5790 | ||
5791 | static int __init slab_sysfs_init(void) | |
5792 | { | |
5793 | struct kmem_cache *s; | |
5794 | int err; | |
5795 | ||
5796 | mutex_lock(&slab_mutex); | |
5797 | ||
5798 | slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); | |
5799 | if (!slab_kset) { | |
5800 | mutex_unlock(&slab_mutex); | |
5801 | pr_err("Cannot register slab subsystem.\n"); | |
5802 | return -ENOSYS; | |
5803 | } | |
5804 | ||
5805 | slab_state = FULL; | |
5806 | ||
5807 | list_for_each_entry(s, &slab_caches, list) { | |
5808 | err = sysfs_slab_add(s); | |
5809 | if (err) | |
5810 | pr_err("SLUB: Unable to add boot slab %s to sysfs\n", | |
5811 | s->name); | |
5812 | } | |
5813 | ||
5814 | while (alias_list) { | |
5815 | struct saved_alias *al = alias_list; | |
5816 | ||
5817 | alias_list = alias_list->next; | |
5818 | err = sysfs_slab_alias(al->s, al->name); | |
5819 | if (err) | |
5820 | pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n", | |
5821 | al->name); | |
5822 | kfree(al); | |
5823 | } | |
5824 | ||
5825 | mutex_unlock(&slab_mutex); | |
5826 | resiliency_test(); | |
5827 | return 0; | |
5828 | } | |
5829 | ||
5830 | __initcall(slab_sysfs_init); | |
5831 | #endif /* CONFIG_SYSFS */ | |
5832 | ||
5833 | /* | |
5834 | * The /proc/slabinfo ABI | |
5835 | */ | |
5836 | #ifdef CONFIG_SLUB_DEBUG | |
5837 | void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) | |
5838 | { | |
5839 | unsigned long nr_slabs = 0; | |
5840 | unsigned long nr_objs = 0; | |
5841 | unsigned long nr_free = 0; | |
5842 | int node; | |
5843 | struct kmem_cache_node *n; | |
5844 | ||
5845 | for_each_kmem_cache_node(s, node, n) { | |
5846 | nr_slabs += node_nr_slabs(n); | |
5847 | nr_objs += node_nr_objs(n); | |
5848 | nr_free += count_partial(n, count_free); | |
5849 | } | |
5850 | ||
5851 | sinfo->active_objs = nr_objs - nr_free; | |
5852 | sinfo->num_objs = nr_objs; | |
5853 | sinfo->active_slabs = nr_slabs; | |
5854 | sinfo->num_slabs = nr_slabs; | |
5855 | sinfo->objects_per_slab = oo_objects(s->oo); | |
5856 | sinfo->cache_order = oo_order(s->oo); | |
5857 | } | |
5858 | ||
5859 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) | |
5860 | { | |
5861 | } | |
5862 | ||
5863 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, | |
5864 | size_t count, loff_t *ppos) | |
5865 | { | |
5866 | return -EIO; | |
5867 | } | |
5868 | #endif /* CONFIG_SLUB_DEBUG */ |