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 operations |
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> /* mm_account_reclaimed_pages() */ |
15 | #include <linux/module.h> |
16 | #include <linux/bit_spinlock.h> |
17 | #include <linux/interrupt.h> |
18 | #include <linux/swab.h> |
19 | #include <linux/bitops.h> |
20 | #include <linux/slab.h> |
21 | #include "slab.h" |
22 | #include <linux/proc_fs.h> |
23 | #include <linux/seq_file.h> |
24 | #include <linux/kasan.h> |
25 | #include <linux/kmsan.h> |
26 | #include <linux/cpu.h> |
27 | #include <linux/cpuset.h> |
28 | #include <linux/mempolicy.h> |
29 | #include <linux/ctype.h> |
30 | #include <linux/stackdepot.h> |
31 | #include <linux/debugobjects.h> |
32 | #include <linux/kallsyms.h> |
33 | #include <linux/kfence.h> |
34 | #include <linux/memory.h> |
35 | #include <linux/math64.h> |
36 | #include <linux/fault-inject.h> |
37 | #include <linux/stacktrace.h> |
38 | #include <linux/prefetch.h> |
39 | #include <linux/memcontrol.h> |
40 | #include <linux/random.h> |
41 | #include <kunit/test.h> |
42 | #include <kunit/test-bug.h> |
43 | #include <linux/sort.h> |
44 | |
45 | #include <linux/debugfs.h> |
46 | #include <trace/events/kmem.h> |
47 | |
48 | #include "internal.h" |
49 | |
50 | /* |
51 | * Lock order: |
52 | * 1. slab_mutex (Global Mutex) |
53 | * 2. node->list_lock (Spinlock) |
54 | * 3. kmem_cache->cpu_slab->lock (Local lock) |
55 | * 4. slab_lock(slab) (Only on some arches) |
56 | * 5. object_map_lock (Only for debugging) |
57 | * |
58 | * slab_mutex |
59 | * |
60 | * The role of the slab_mutex is to protect the list of all the slabs |
61 | * and to synchronize major metadata changes to slab cache structures. |
62 | * Also synchronizes memory hotplug callbacks. |
63 | * |
64 | * slab_lock |
65 | * |
66 | * The slab_lock is a wrapper around the page lock, thus it is a bit |
67 | * spinlock. |
68 | * |
69 | * The slab_lock is only used on arches that do not have the ability |
70 | * to do a cmpxchg_double. It only protects: |
71 | * |
72 | * A. slab->freelist -> List of free objects in a slab |
73 | * B. slab->inuse -> Number of objects in use |
74 | * C. slab->objects -> Number of objects in slab |
75 | * D. slab->frozen -> frozen state |
76 | * |
77 | * Frozen slabs |
78 | * |
79 | * If a slab is frozen then it is exempt from list management. It is not |
80 | * on any list except per cpu partial list. The processor that froze the |
81 | * slab is the one who can perform list operations on the slab. Other |
82 | * processors may put objects onto the freelist but the processor that |
83 | * froze the slab is the only one that can retrieve the objects from the |
84 | * slab's freelist. |
85 | * |
86 | * list_lock |
87 | * |
88 | * The list_lock protects the partial and full list on each node and |
89 | * the partial slab counter. If taken then no new slabs may be added or |
90 | * removed from the lists nor make the number of partial slabs be modified. |
91 | * (Note that the total number of slabs is an atomic value that may be |
92 | * modified without taking the list lock). |
93 | * |
94 | * The list_lock is a centralized lock and thus we avoid taking it as |
95 | * much as possible. As long as SLUB does not have to handle partial |
96 | * slabs, operations can continue without any centralized lock. F.e. |
97 | * allocating a long series of objects that fill up slabs does not require |
98 | * the list lock. |
99 | * |
100 | * For debug caches, all allocations are forced to go through a list_lock |
101 | * protected region to serialize against concurrent validation. |
102 | * |
103 | * cpu_slab->lock local lock |
104 | * |
105 | * This locks protect slowpath manipulation of all kmem_cache_cpu fields |
106 | * except the stat counters. This is a percpu structure manipulated only by |
107 | * the local cpu, so the lock protects against being preempted or interrupted |
108 | * by an irq. Fast path operations rely on lockless operations instead. |
109 | * |
110 | * On PREEMPT_RT, the local lock neither disables interrupts nor preemption |
111 | * which means the lockless fastpath cannot be used as it might interfere with |
112 | * an in-progress slow path operations. In this case the local lock is always |
113 | * taken but it still utilizes the freelist for the common operations. |
114 | * |
115 | * lockless fastpaths |
116 | * |
117 | * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free()) |
118 | * are fully lockless when satisfied from the percpu slab (and when |
119 | * cmpxchg_double is possible to use, otherwise slab_lock is taken). |
120 | * They also don't disable preemption or migration or irqs. They rely on |
121 | * the transaction id (tid) field to detect being preempted or moved to |
122 | * another cpu. |
123 | * |
124 | * irq, preemption, migration considerations |
125 | * |
126 | * Interrupts are disabled as part of list_lock or local_lock operations, or |
127 | * around the slab_lock operation, in order to make the slab allocator safe |
128 | * to use in the context of an irq. |
129 | * |
130 | * In addition, preemption (or migration on PREEMPT_RT) is disabled in the |
131 | * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the |
132 | * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer |
133 | * doesn't have to be revalidated in each section protected by the local lock. |
134 | * |
135 | * SLUB assigns one slab for allocation to each processor. |
136 | * Allocations only occur from these slabs called cpu slabs. |
137 | * |
138 | * Slabs with free elements are kept on a partial list and during regular |
139 | * operations no list for full slabs is used. If an object in a full slab is |
140 | * freed then the slab will show up again on the partial lists. |
141 | * We track full slabs for debugging purposes though because otherwise we |
142 | * cannot scan all objects. |
143 | * |
144 | * Slabs are freed when they become empty. Teardown and setup is |
145 | * minimal so we rely on the page allocators per cpu caches for |
146 | * fast frees and allocs. |
147 | * |
148 | * slab->frozen The slab is frozen and exempt from list processing. |
149 | * This means that the slab is dedicated to a purpose |
150 | * such as satisfying allocations for a specific |
151 | * processor. Objects may be freed in the slab while |
152 | * it is frozen but slab_free will then skip the usual |
153 | * list operations. It is up to the processor holding |
154 | * the slab to integrate the slab into the slab lists |
155 | * when the slab is no longer needed. |
156 | * |
157 | * One use of this flag is to mark slabs that are |
158 | * used for allocations. Then such a slab becomes a cpu |
159 | * slab. The cpu slab may be equipped with an additional |
160 | * freelist that allows lockless access to |
161 | * free objects in addition to the regular freelist |
162 | * that requires the slab lock. |
163 | * |
164 | * SLAB_DEBUG_FLAGS Slab requires special handling due to debug |
165 | * options set. This moves slab handling out of |
166 | * the fast path and disables lockless freelists. |
167 | */ |
168 | |
169 | /* |
170 | * We could simply use migrate_disable()/enable() but as long as it's a |
171 | * function call even on !PREEMPT_RT, use inline preempt_disable() there. |
172 | */ |
173 | #ifndef CONFIG_PREEMPT_RT |
174 | #define slub_get_cpu_ptr(var) get_cpu_ptr(var) |
175 | #define slub_put_cpu_ptr(var) put_cpu_ptr(var) |
176 | #define USE_LOCKLESS_FAST_PATH() (true) |
177 | #else |
178 | #define slub_get_cpu_ptr(var) \ |
179 | ({ \ |
180 | migrate_disable(); \ |
181 | this_cpu_ptr(var); \ |
182 | }) |
183 | #define slub_put_cpu_ptr(var) \ |
184 | do { \ |
185 | (void)(var); \ |
186 | migrate_enable(); \ |
187 | } while (0) |
188 | #define USE_LOCKLESS_FAST_PATH() (false) |
189 | #endif |
190 | |
191 | #ifndef CONFIG_SLUB_TINY |
192 | #define __fastpath_inline __always_inline |
193 | #else |
194 | #define __fastpath_inline |
195 | #endif |
196 | |
197 | #ifdef CONFIG_SLUB_DEBUG |
198 | #ifdef CONFIG_SLUB_DEBUG_ON |
199 | DEFINE_STATIC_KEY_TRUE(slub_debug_enabled); |
200 | #else |
201 | DEFINE_STATIC_KEY_FALSE(slub_debug_enabled); |
202 | #endif |
203 | #endif /* CONFIG_SLUB_DEBUG */ |
204 | |
205 | /* Structure holding parameters for get_partial() call chain */ |
206 | struct partial_context { |
207 | struct slab **slab; |
208 | gfp_t flags; |
209 | unsigned int orig_size; |
210 | }; |
211 | |
212 | static inline bool kmem_cache_debug(struct kmem_cache *s) |
213 | { |
214 | return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS); |
215 | } |
216 | |
217 | static inline bool slub_debug_orig_size(struct kmem_cache *s) |
218 | { |
219 | return (kmem_cache_debug_flags(s, SLAB_STORE_USER) && |
220 | (s->flags & SLAB_KMALLOC)); |
221 | } |
222 | |
223 | void *fixup_red_left(struct kmem_cache *s, void *p) |
224 | { |
225 | if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) |
226 | p += s->red_left_pad; |
227 | |
228 | return p; |
229 | } |
230 | |
231 | static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) |
232 | { |
233 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
234 | return !kmem_cache_debug(s); |
235 | #else |
236 | return false; |
237 | #endif |
238 | } |
239 | |
240 | /* |
241 | * Issues still to be resolved: |
242 | * |
243 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. |
244 | * |
245 | * - Variable sizing of the per node arrays |
246 | */ |
247 | |
248 | /* Enable to log cmpxchg failures */ |
249 | #undef SLUB_DEBUG_CMPXCHG |
250 | |
251 | #ifndef CONFIG_SLUB_TINY |
252 | /* |
253 | * Minimum number of partial slabs. These will be left on the partial |
254 | * lists even if they are empty. kmem_cache_shrink may reclaim them. |
255 | */ |
256 | #define MIN_PARTIAL 5 |
257 | |
258 | /* |
259 | * Maximum number of desirable partial slabs. |
260 | * The existence of more partial slabs makes kmem_cache_shrink |
261 | * sort the partial list by the number of objects in use. |
262 | */ |
263 | #define MAX_PARTIAL 10 |
264 | #else |
265 | #define MIN_PARTIAL 0 |
266 | #define MAX_PARTIAL 0 |
267 | #endif |
268 | |
269 | #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \ |
270 | SLAB_POISON | SLAB_STORE_USER) |
271 | |
272 | /* |
273 | * These debug flags cannot use CMPXCHG because there might be consistency |
274 | * issues when checking or reading debug information |
275 | */ |
276 | #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \ |
277 | SLAB_TRACE) |
278 | |
279 | |
280 | /* |
281 | * Debugging flags that require metadata to be stored in the slab. These get |
282 | * disabled when slub_debug=O is used and a cache's min order increases with |
283 | * metadata. |
284 | */ |
285 | #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) |
286 | |
287 | #define OO_SHIFT 16 |
288 | #define OO_MASK ((1 << OO_SHIFT) - 1) |
289 | #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */ |
290 | |
291 | /* Internal SLUB flags */ |
292 | /* Poison object */ |
293 | #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U) |
294 | /* Use cmpxchg_double */ |
295 | |
296 | #ifdef system_has_freelist_aba |
297 | #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U) |
298 | #else |
299 | #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0U) |
300 | #endif |
301 | |
302 | /* |
303 | * Tracking user of a slab. |
304 | */ |
305 | #define TRACK_ADDRS_COUNT 16 |
306 | struct track { |
307 | unsigned long addr; /* Called from address */ |
308 | #ifdef CONFIG_STACKDEPOT |
309 | depot_stack_handle_t handle; |
310 | #endif |
311 | int cpu; /* Was running on cpu */ |
312 | int pid; /* Pid context */ |
313 | unsigned long when; /* When did the operation occur */ |
314 | }; |
315 | |
316 | enum track_item { TRACK_ALLOC, TRACK_FREE }; |
317 | |
318 | #ifdef SLAB_SUPPORTS_SYSFS |
319 | static int sysfs_slab_add(struct kmem_cache *); |
320 | static int sysfs_slab_alias(struct kmem_cache *, const char *); |
321 | #else |
322 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } |
323 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) |
324 | { return 0; } |
325 | #endif |
326 | |
327 | #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG) |
328 | static void debugfs_slab_add(struct kmem_cache *); |
329 | #else |
330 | static inline void debugfs_slab_add(struct kmem_cache *s) { } |
331 | #endif |
332 | |
333 | static inline void stat(const struct kmem_cache *s, enum stat_item si) |
334 | { |
335 | #ifdef CONFIG_SLUB_STATS |
336 | /* |
337 | * The rmw is racy on a preemptible kernel but this is acceptable, so |
338 | * avoid this_cpu_add()'s irq-disable overhead. |
339 | */ |
340 | raw_cpu_inc(s->cpu_slab->stat[si]); |
341 | #endif |
342 | } |
343 | |
344 | /* |
345 | * Tracks for which NUMA nodes we have kmem_cache_nodes allocated. |
346 | * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily |
347 | * differ during memory hotplug/hotremove operations. |
348 | * Protected by slab_mutex. |
349 | */ |
350 | static nodemask_t slab_nodes; |
351 | |
352 | #ifndef CONFIG_SLUB_TINY |
353 | /* |
354 | * Workqueue used for flush_cpu_slab(). |
355 | */ |
356 | static struct workqueue_struct *flushwq; |
357 | #endif |
358 | |
359 | /******************************************************************** |
360 | * Core slab cache functions |
361 | *******************************************************************/ |
362 | |
363 | /* |
364 | * freeptr_t represents a SLUB freelist pointer, which might be encoded |
365 | * and not dereferenceable if CONFIG_SLAB_FREELIST_HARDENED is enabled. |
366 | */ |
367 | typedef struct { unsigned long v; } freeptr_t; |
368 | |
369 | /* |
370 | * Returns freelist pointer (ptr). With hardening, this is obfuscated |
371 | * with an XOR of the address where the pointer is held and a per-cache |
372 | * random number. |
373 | */ |
374 | static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s, |
375 | void *ptr, unsigned long ptr_addr) |
376 | { |
377 | unsigned long encoded; |
378 | |
379 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
380 | encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr); |
381 | #else |
382 | encoded = (unsigned long)ptr; |
383 | #endif |
384 | return (freeptr_t){.v = encoded}; |
385 | } |
386 | |
387 | static inline void *freelist_ptr_decode(const struct kmem_cache *s, |
388 | freeptr_t ptr, unsigned long ptr_addr) |
389 | { |
390 | void *decoded; |
391 | |
392 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
393 | decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr)); |
394 | #else |
395 | decoded = (void *)ptr.v; |
396 | #endif |
397 | return decoded; |
398 | } |
399 | |
400 | static inline void *get_freepointer(struct kmem_cache *s, void *object) |
401 | { |
402 | unsigned long ptr_addr; |
403 | freeptr_t p; |
404 | |
405 | object = kasan_reset_tag(addr: object); |
406 | ptr_addr = (unsigned long)object + s->offset; |
407 | p = *(freeptr_t *)(ptr_addr); |
408 | return freelist_ptr_decode(s, ptr: p, ptr_addr); |
409 | } |
410 | |
411 | #ifndef CONFIG_SLUB_TINY |
412 | static void prefetch_freepointer(const struct kmem_cache *s, void *object) |
413 | { |
414 | prefetchw(object + s->offset); |
415 | } |
416 | #endif |
417 | |
418 | /* |
419 | * When running under KMSAN, get_freepointer_safe() may return an uninitialized |
420 | * pointer value in the case the current thread loses the race for the next |
421 | * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in |
422 | * slab_alloc_node() will fail, so the uninitialized value won't be used, but |
423 | * KMSAN will still check all arguments of cmpxchg because of imperfect |
424 | * handling of inline assembly. |
425 | * To work around this problem, we apply __no_kmsan_checks to ensure that |
426 | * get_freepointer_safe() returns initialized memory. |
427 | */ |
428 | __no_kmsan_checks |
429 | static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) |
430 | { |
431 | unsigned long freepointer_addr; |
432 | freeptr_t p; |
433 | |
434 | if (!debug_pagealloc_enabled_static()) |
435 | return get_freepointer(s, object); |
436 | |
437 | object = kasan_reset_tag(addr: object); |
438 | freepointer_addr = (unsigned long)object + s->offset; |
439 | copy_from_kernel_nofault(dst: &p, src: (freeptr_t *)freepointer_addr, size: sizeof(p)); |
440 | return freelist_ptr_decode(s, ptr: p, ptr_addr: freepointer_addr); |
441 | } |
442 | |
443 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) |
444 | { |
445 | unsigned long freeptr_addr = (unsigned long)object + s->offset; |
446 | |
447 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
448 | BUG_ON(object == fp); /* naive detection of double free or corruption */ |
449 | #endif |
450 | |
451 | freeptr_addr = (unsigned long)kasan_reset_tag(addr: (void *)freeptr_addr); |
452 | *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, ptr: fp, ptr_addr: freeptr_addr); |
453 | } |
454 | |
455 | /* Loop over all objects in a slab */ |
456 | #define for_each_object(__p, __s, __addr, __objects) \ |
457 | for (__p = fixup_red_left(__s, __addr); \ |
458 | __p < (__addr) + (__objects) * (__s)->size; \ |
459 | __p += (__s)->size) |
460 | |
461 | static inline unsigned int order_objects(unsigned int order, unsigned int size) |
462 | { |
463 | return ((unsigned int)PAGE_SIZE << order) / size; |
464 | } |
465 | |
466 | static inline struct kmem_cache_order_objects oo_make(unsigned int order, |
467 | unsigned int size) |
468 | { |
469 | struct kmem_cache_order_objects x = { |
470 | (order << OO_SHIFT) + order_objects(order, size) |
471 | }; |
472 | |
473 | return x; |
474 | } |
475 | |
476 | static inline unsigned int oo_order(struct kmem_cache_order_objects x) |
477 | { |
478 | return x.x >> OO_SHIFT; |
479 | } |
480 | |
481 | static inline unsigned int oo_objects(struct kmem_cache_order_objects x) |
482 | { |
483 | return x.x & OO_MASK; |
484 | } |
485 | |
486 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
487 | static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects) |
488 | { |
489 | unsigned int nr_slabs; |
490 | |
491 | s->cpu_partial = nr_objects; |
492 | |
493 | /* |
494 | * We take the number of objects but actually limit the number of |
495 | * slabs on the per cpu partial list, in order to limit excessive |
496 | * growth of the list. For simplicity we assume that the slabs will |
497 | * be half-full. |
498 | */ |
499 | nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo)); |
500 | s->cpu_partial_slabs = nr_slabs; |
501 | } |
502 | #else |
503 | static inline void |
504 | slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects) |
505 | { |
506 | } |
507 | #endif /* CONFIG_SLUB_CPU_PARTIAL */ |
508 | |
509 | /* |
510 | * Per slab locking using the pagelock |
511 | */ |
512 | static __always_inline void slab_lock(struct slab *slab) |
513 | { |
514 | struct page *page = slab_page(slab); |
515 | |
516 | VM_BUG_ON_PAGE(PageTail(page), page); |
517 | bit_spin_lock(bitnum: PG_locked, addr: &page->flags); |
518 | } |
519 | |
520 | static __always_inline void slab_unlock(struct slab *slab) |
521 | { |
522 | struct page *page = slab_page(slab); |
523 | |
524 | VM_BUG_ON_PAGE(PageTail(page), page); |
525 | __bit_spin_unlock(bitnum: PG_locked, addr: &page->flags); |
526 | } |
527 | |
528 | static inline bool |
529 | __update_freelist_fast(struct slab *slab, |
530 | void *freelist_old, unsigned long counters_old, |
531 | void *freelist_new, unsigned long counters_new) |
532 | { |
533 | #ifdef system_has_freelist_aba |
534 | freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old }; |
535 | freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new }; |
536 | |
537 | return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full); |
538 | #else |
539 | return false; |
540 | #endif |
541 | } |
542 | |
543 | static inline bool |
544 | __update_freelist_slow(struct slab *slab, |
545 | void *freelist_old, unsigned long counters_old, |
546 | void *freelist_new, unsigned long counters_new) |
547 | { |
548 | bool ret = false; |
549 | |
550 | slab_lock(slab); |
551 | if (slab->freelist == freelist_old && |
552 | slab->counters == counters_old) { |
553 | slab->freelist = freelist_new; |
554 | slab->counters = counters_new; |
555 | ret = true; |
556 | } |
557 | slab_unlock(slab); |
558 | |
559 | return ret; |
560 | } |
561 | |
562 | /* |
563 | * Interrupts must be disabled (for the fallback code to work right), typically |
564 | * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is |
565 | * part of bit_spin_lock(), is sufficient because the policy is not to allow any |
566 | * allocation/ free operation in hardirq context. Therefore nothing can |
567 | * interrupt the operation. |
568 | */ |
569 | static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab, |
570 | void *freelist_old, unsigned long counters_old, |
571 | void *freelist_new, unsigned long counters_new, |
572 | const char *n) |
573 | { |
574 | bool ret; |
575 | |
576 | if (USE_LOCKLESS_FAST_PATH()) |
577 | lockdep_assert_irqs_disabled(); |
578 | |
579 | if (s->flags & __CMPXCHG_DOUBLE) { |
580 | ret = __update_freelist_fast(slab, freelist_old, counters_old, |
581 | freelist_new, counters_new); |
582 | } else { |
583 | ret = __update_freelist_slow(slab, freelist_old, counters_old, |
584 | freelist_new, counters_new); |
585 | } |
586 | if (likely(ret)) |
587 | return true; |
588 | |
589 | cpu_relax(); |
590 | stat(s, si: CMPXCHG_DOUBLE_FAIL); |
591 | |
592 | #ifdef SLUB_DEBUG_CMPXCHG |
593 | pr_info("%s %s: cmpxchg double redo " , n, s->name); |
594 | #endif |
595 | |
596 | return false; |
597 | } |
598 | |
599 | static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab, |
600 | void *freelist_old, unsigned long counters_old, |
601 | void *freelist_new, unsigned long counters_new, |
602 | const char *n) |
603 | { |
604 | bool ret; |
605 | |
606 | if (s->flags & __CMPXCHG_DOUBLE) { |
607 | ret = __update_freelist_fast(slab, freelist_old, counters_old, |
608 | freelist_new, counters_new); |
609 | } else { |
610 | unsigned long flags; |
611 | |
612 | local_irq_save(flags); |
613 | ret = __update_freelist_slow(slab, freelist_old, counters_old, |
614 | freelist_new, counters_new); |
615 | local_irq_restore(flags); |
616 | } |
617 | if (likely(ret)) |
618 | return true; |
619 | |
620 | cpu_relax(); |
621 | stat(s, si: CMPXCHG_DOUBLE_FAIL); |
622 | |
623 | #ifdef SLUB_DEBUG_CMPXCHG |
624 | pr_info("%s %s: cmpxchg double redo " , n, s->name); |
625 | #endif |
626 | |
627 | return false; |
628 | } |
629 | |
630 | #ifdef CONFIG_SLUB_DEBUG |
631 | static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)]; |
632 | static DEFINE_SPINLOCK(object_map_lock); |
633 | |
634 | static void __fill_map(unsigned long *obj_map, struct kmem_cache *s, |
635 | struct slab *slab) |
636 | { |
637 | void *addr = slab_address(slab); |
638 | void *p; |
639 | |
640 | bitmap_zero(obj_map, slab->objects); |
641 | |
642 | for (p = slab->freelist; p; p = get_freepointer(s, p)) |
643 | set_bit(__obj_to_index(s, addr, p), obj_map); |
644 | } |
645 | |
646 | #if IS_ENABLED(CONFIG_KUNIT) |
647 | static bool slab_add_kunit_errors(void) |
648 | { |
649 | struct kunit_resource *resource; |
650 | |
651 | if (!kunit_get_current_test()) |
652 | return false; |
653 | |
654 | resource = kunit_find_named_resource(current->kunit_test, "slab_errors" ); |
655 | if (!resource) |
656 | return false; |
657 | |
658 | (*(int *)resource->data)++; |
659 | kunit_put_resource(resource); |
660 | return true; |
661 | } |
662 | #else |
663 | static inline bool slab_add_kunit_errors(void) { return false; } |
664 | #endif |
665 | |
666 | static inline unsigned int size_from_object(struct kmem_cache *s) |
667 | { |
668 | if (s->flags & SLAB_RED_ZONE) |
669 | return s->size - s->red_left_pad; |
670 | |
671 | return s->size; |
672 | } |
673 | |
674 | static inline void *restore_red_left(struct kmem_cache *s, void *p) |
675 | { |
676 | if (s->flags & SLAB_RED_ZONE) |
677 | p -= s->red_left_pad; |
678 | |
679 | return p; |
680 | } |
681 | |
682 | /* |
683 | * Debug settings: |
684 | */ |
685 | #if defined(CONFIG_SLUB_DEBUG_ON) |
686 | static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS; |
687 | #else |
688 | static slab_flags_t slub_debug; |
689 | #endif |
690 | |
691 | static char *slub_debug_string; |
692 | static int disable_higher_order_debug; |
693 | |
694 | /* |
695 | * slub is about to manipulate internal object metadata. This memory lies |
696 | * outside the range of the allocated object, so accessing it would normally |
697 | * be reported by kasan as a bounds error. metadata_access_enable() is used |
698 | * to tell kasan that these accesses are OK. |
699 | */ |
700 | static inline void metadata_access_enable(void) |
701 | { |
702 | kasan_disable_current(); |
703 | } |
704 | |
705 | static inline void metadata_access_disable(void) |
706 | { |
707 | kasan_enable_current(); |
708 | } |
709 | |
710 | /* |
711 | * Object debugging |
712 | */ |
713 | |
714 | /* Verify that a pointer has an address that is valid within a slab page */ |
715 | static inline int check_valid_pointer(struct kmem_cache *s, |
716 | struct slab *slab, void *object) |
717 | { |
718 | void *base; |
719 | |
720 | if (!object) |
721 | return 1; |
722 | |
723 | base = slab_address(slab); |
724 | object = kasan_reset_tag(object); |
725 | object = restore_red_left(s, object); |
726 | if (object < base || object >= base + slab->objects * s->size || |
727 | (object - base) % s->size) { |
728 | return 0; |
729 | } |
730 | |
731 | return 1; |
732 | } |
733 | |
734 | static void print_section(char *level, char *text, u8 *addr, |
735 | unsigned int length) |
736 | { |
737 | metadata_access_enable(); |
738 | print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, |
739 | 16, 1, kasan_reset_tag((void *)addr), length, 1); |
740 | metadata_access_disable(); |
741 | } |
742 | |
743 | /* |
744 | * See comment in calculate_sizes(). |
745 | */ |
746 | static inline bool freeptr_outside_object(struct kmem_cache *s) |
747 | { |
748 | return s->offset >= s->inuse; |
749 | } |
750 | |
751 | /* |
752 | * Return offset of the end of info block which is inuse + free pointer if |
753 | * not overlapping with object. |
754 | */ |
755 | static inline unsigned int get_info_end(struct kmem_cache *s) |
756 | { |
757 | if (freeptr_outside_object(s)) |
758 | return s->inuse + sizeof(void *); |
759 | else |
760 | return s->inuse; |
761 | } |
762 | |
763 | static struct track *get_track(struct kmem_cache *s, void *object, |
764 | enum track_item alloc) |
765 | { |
766 | struct track *p; |
767 | |
768 | p = object + get_info_end(s); |
769 | |
770 | return kasan_reset_tag(p + alloc); |
771 | } |
772 | |
773 | #ifdef CONFIG_STACKDEPOT |
774 | static noinline depot_stack_handle_t set_track_prepare(void) |
775 | { |
776 | depot_stack_handle_t handle; |
777 | unsigned long entries[TRACK_ADDRS_COUNT]; |
778 | unsigned int nr_entries; |
779 | |
780 | nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3); |
781 | handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT); |
782 | |
783 | return handle; |
784 | } |
785 | #else |
786 | static inline depot_stack_handle_t set_track_prepare(void) |
787 | { |
788 | return 0; |
789 | } |
790 | #endif |
791 | |
792 | static void set_track_update(struct kmem_cache *s, void *object, |
793 | enum track_item alloc, unsigned long addr, |
794 | depot_stack_handle_t handle) |
795 | { |
796 | struct track *p = get_track(s, object, alloc); |
797 | |
798 | #ifdef CONFIG_STACKDEPOT |
799 | p->handle = handle; |
800 | #endif |
801 | p->addr = addr; |
802 | p->cpu = smp_processor_id(); |
803 | p->pid = current->pid; |
804 | p->when = jiffies; |
805 | } |
806 | |
807 | static __always_inline void set_track(struct kmem_cache *s, void *object, |
808 | enum track_item alloc, unsigned long addr) |
809 | { |
810 | depot_stack_handle_t handle = set_track_prepare(); |
811 | |
812 | set_track_update(s, object, alloc, addr, handle); |
813 | } |
814 | |
815 | static void init_tracking(struct kmem_cache *s, void *object) |
816 | { |
817 | struct track *p; |
818 | |
819 | if (!(s->flags & SLAB_STORE_USER)) |
820 | return; |
821 | |
822 | p = get_track(s, object, TRACK_ALLOC); |
823 | memset(p, 0, 2*sizeof(struct track)); |
824 | } |
825 | |
826 | static void print_track(const char *s, struct track *t, unsigned long pr_time) |
827 | { |
828 | depot_stack_handle_t handle __maybe_unused; |
829 | |
830 | if (!t->addr) |
831 | return; |
832 | |
833 | pr_err("%s in %pS age=%lu cpu=%u pid=%d\n" , |
834 | s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid); |
835 | #ifdef CONFIG_STACKDEPOT |
836 | handle = READ_ONCE(t->handle); |
837 | if (handle) |
838 | stack_depot_print(handle); |
839 | else |
840 | pr_err("object allocation/free stack trace missing\n" ); |
841 | #endif |
842 | } |
843 | |
844 | void print_tracking(struct kmem_cache *s, void *object) |
845 | { |
846 | unsigned long pr_time = jiffies; |
847 | if (!(s->flags & SLAB_STORE_USER)) |
848 | return; |
849 | |
850 | print_track("Allocated" , get_track(s, object, TRACK_ALLOC), pr_time); |
851 | print_track("Freed" , get_track(s, object, TRACK_FREE), pr_time); |
852 | } |
853 | |
854 | static void print_slab_info(const struct slab *slab) |
855 | { |
856 | struct folio *folio = (struct folio *)slab_folio(slab); |
857 | |
858 | pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n" , |
859 | slab, slab->objects, slab->inuse, slab->freelist, |
860 | folio_flags(folio, 0)); |
861 | } |
862 | |
863 | /* |
864 | * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API |
865 | * family will round up the real request size to these fixed ones, so |
866 | * there could be an extra area than what is requested. Save the original |
867 | * request size in the meta data area, for better debug and sanity check. |
868 | */ |
869 | static inline void set_orig_size(struct kmem_cache *s, |
870 | void *object, unsigned int orig_size) |
871 | { |
872 | void *p = kasan_reset_tag(object); |
873 | |
874 | if (!slub_debug_orig_size(s)) |
875 | return; |
876 | |
877 | #ifdef CONFIG_KASAN_GENERIC |
878 | /* |
879 | * KASAN could save its free meta data in object's data area at |
880 | * offset 0, if the size is larger than 'orig_size', it will |
881 | * overlap the data redzone in [orig_size+1, object_size], and |
882 | * the check should be skipped. |
883 | */ |
884 | if (kasan_metadata_size(s, true) > orig_size) |
885 | orig_size = s->object_size; |
886 | #endif |
887 | |
888 | p += get_info_end(s); |
889 | p += sizeof(struct track) * 2; |
890 | |
891 | *(unsigned int *)p = orig_size; |
892 | } |
893 | |
894 | static inline unsigned int get_orig_size(struct kmem_cache *s, void *object) |
895 | { |
896 | void *p = kasan_reset_tag(object); |
897 | |
898 | if (!slub_debug_orig_size(s)) |
899 | return s->object_size; |
900 | |
901 | p += get_info_end(s); |
902 | p += sizeof(struct track) * 2; |
903 | |
904 | return *(unsigned int *)p; |
905 | } |
906 | |
907 | void skip_orig_size_check(struct kmem_cache *s, const void *object) |
908 | { |
909 | set_orig_size(s, (void *)object, s->object_size); |
910 | } |
911 | |
912 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) |
913 | { |
914 | struct va_format vaf; |
915 | va_list args; |
916 | |
917 | va_start(args, fmt); |
918 | vaf.fmt = fmt; |
919 | vaf.va = &args; |
920 | pr_err("=============================================================================\n" ); |
921 | pr_err("BUG %s (%s): %pV\n" , s->name, print_tainted(), &vaf); |
922 | pr_err("-----------------------------------------------------------------------------\n\n" ); |
923 | va_end(args); |
924 | } |
925 | |
926 | __printf(2, 3) |
927 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) |
928 | { |
929 | struct va_format vaf; |
930 | va_list args; |
931 | |
932 | if (slab_add_kunit_errors()) |
933 | return; |
934 | |
935 | va_start(args, fmt); |
936 | vaf.fmt = fmt; |
937 | vaf.va = &args; |
938 | pr_err("FIX %s: %pV\n" , s->name, &vaf); |
939 | va_end(args); |
940 | } |
941 | |
942 | static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p) |
943 | { |
944 | unsigned int off; /* Offset of last byte */ |
945 | u8 *addr = slab_address(slab); |
946 | |
947 | print_tracking(s, p); |
948 | |
949 | print_slab_info(slab); |
950 | |
951 | pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n" , |
952 | p, p - addr, get_freepointer(s, p)); |
953 | |
954 | if (s->flags & SLAB_RED_ZONE) |
955 | print_section(KERN_ERR, "Redzone " , p - s->red_left_pad, |
956 | s->red_left_pad); |
957 | else if (p > addr + 16) |
958 | print_section(KERN_ERR, "Bytes b4 " , p - 16, 16); |
959 | |
960 | print_section(KERN_ERR, "Object " , p, |
961 | min_t(unsigned int, s->object_size, PAGE_SIZE)); |
962 | if (s->flags & SLAB_RED_ZONE) |
963 | print_section(KERN_ERR, "Redzone " , p + s->object_size, |
964 | s->inuse - s->object_size); |
965 | |
966 | off = get_info_end(s); |
967 | |
968 | if (s->flags & SLAB_STORE_USER) |
969 | off += 2 * sizeof(struct track); |
970 | |
971 | if (slub_debug_orig_size(s)) |
972 | off += sizeof(unsigned int); |
973 | |
974 | off += kasan_metadata_size(s, false); |
975 | |
976 | if (off != size_from_object(s)) |
977 | /* Beginning of the filler is the free pointer */ |
978 | print_section(KERN_ERR, "Padding " , p + off, |
979 | size_from_object(s) - off); |
980 | |
981 | dump_stack(); |
982 | } |
983 | |
984 | static void object_err(struct kmem_cache *s, struct slab *slab, |
985 | u8 *object, char *reason) |
986 | { |
987 | if (slab_add_kunit_errors()) |
988 | return; |
989 | |
990 | slab_bug(s, "%s" , reason); |
991 | print_trailer(s, slab, object); |
992 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
993 | } |
994 | |
995 | static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab, |
996 | void **freelist, void *nextfree) |
997 | { |
998 | if ((s->flags & SLAB_CONSISTENCY_CHECKS) && |
999 | !check_valid_pointer(s, slab, nextfree) && freelist) { |
1000 | object_err(s, slab, *freelist, "Freechain corrupt" ); |
1001 | *freelist = NULL; |
1002 | slab_fix(s, "Isolate corrupted freechain" ); |
1003 | return true; |
1004 | } |
1005 | |
1006 | return false; |
1007 | } |
1008 | |
1009 | static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab, |
1010 | const char *fmt, ...) |
1011 | { |
1012 | va_list args; |
1013 | char buf[100]; |
1014 | |
1015 | if (slab_add_kunit_errors()) |
1016 | return; |
1017 | |
1018 | va_start(args, fmt); |
1019 | vsnprintf(buf, sizeof(buf), fmt, args); |
1020 | va_end(args); |
1021 | slab_bug(s, "%s" , buf); |
1022 | print_slab_info(slab); |
1023 | dump_stack(); |
1024 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1025 | } |
1026 | |
1027 | static void init_object(struct kmem_cache *s, void *object, u8 val) |
1028 | { |
1029 | u8 *p = kasan_reset_tag(object); |
1030 | unsigned int poison_size = s->object_size; |
1031 | |
1032 | if (s->flags & SLAB_RED_ZONE) { |
1033 | memset(p - s->red_left_pad, val, s->red_left_pad); |
1034 | |
1035 | if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) { |
1036 | /* |
1037 | * Redzone the extra allocated space by kmalloc than |
1038 | * requested, and the poison size will be limited to |
1039 | * the original request size accordingly. |
1040 | */ |
1041 | poison_size = get_orig_size(s, object); |
1042 | } |
1043 | } |
1044 | |
1045 | if (s->flags & __OBJECT_POISON) { |
1046 | memset(p, POISON_FREE, poison_size - 1); |
1047 | p[poison_size - 1] = POISON_END; |
1048 | } |
1049 | |
1050 | if (s->flags & SLAB_RED_ZONE) |
1051 | memset(p + poison_size, val, s->inuse - poison_size); |
1052 | } |
1053 | |
1054 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, |
1055 | void *from, void *to) |
1056 | { |
1057 | slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x" , message, from, to - 1, data); |
1058 | memset(from, data, to - from); |
1059 | } |
1060 | |
1061 | static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab, |
1062 | u8 *object, char *what, |
1063 | u8 *start, unsigned int value, unsigned int bytes) |
1064 | { |
1065 | u8 *fault; |
1066 | u8 *end; |
1067 | u8 *addr = slab_address(slab); |
1068 | |
1069 | metadata_access_enable(); |
1070 | fault = memchr_inv(kasan_reset_tag(start), value, bytes); |
1071 | metadata_access_disable(); |
1072 | if (!fault) |
1073 | return 1; |
1074 | |
1075 | end = start + bytes; |
1076 | while (end > fault && end[-1] == value) |
1077 | end--; |
1078 | |
1079 | if (slab_add_kunit_errors()) |
1080 | goto skip_bug_print; |
1081 | |
1082 | slab_bug(s, "%s overwritten" , what); |
1083 | pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n" , |
1084 | fault, end - 1, fault - addr, |
1085 | fault[0], value); |
1086 | print_trailer(s, slab, object); |
1087 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1088 | |
1089 | skip_bug_print: |
1090 | restore_bytes(s, what, value, fault, end); |
1091 | return 0; |
1092 | } |
1093 | |
1094 | /* |
1095 | * Object layout: |
1096 | * |
1097 | * object address |
1098 | * Bytes of the object to be managed. |
1099 | * If the freepointer may overlay the object then the free |
1100 | * pointer is at the middle of the object. |
1101 | * |
1102 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is |
1103 | * 0xa5 (POISON_END) |
1104 | * |
1105 | * object + s->object_size |
1106 | * Padding to reach word boundary. This is also used for Redzoning. |
1107 | * Padding is extended by another word if Redzoning is enabled and |
1108 | * object_size == inuse. |
1109 | * |
1110 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with |
1111 | * 0xcc (RED_ACTIVE) for objects in use. |
1112 | * |
1113 | * object + s->inuse |
1114 | * Meta data starts here. |
1115 | * |
1116 | * A. Free pointer (if we cannot overwrite object on free) |
1117 | * B. Tracking data for SLAB_STORE_USER |
1118 | * C. Original request size for kmalloc object (SLAB_STORE_USER enabled) |
1119 | * D. Padding to reach required alignment boundary or at minimum |
1120 | * one word if debugging is on to be able to detect writes |
1121 | * before the word boundary. |
1122 | * |
1123 | * Padding is done using 0x5a (POISON_INUSE) |
1124 | * |
1125 | * object + s->size |
1126 | * Nothing is used beyond s->size. |
1127 | * |
1128 | * If slabcaches are merged then the object_size and inuse boundaries are mostly |
1129 | * ignored. And therefore no slab options that rely on these boundaries |
1130 | * may be used with merged slabcaches. |
1131 | */ |
1132 | |
1133 | static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p) |
1134 | { |
1135 | unsigned long off = get_info_end(s); /* The end of info */ |
1136 | |
1137 | if (s->flags & SLAB_STORE_USER) { |
1138 | /* We also have user information there */ |
1139 | off += 2 * sizeof(struct track); |
1140 | |
1141 | if (s->flags & SLAB_KMALLOC) |
1142 | off += sizeof(unsigned int); |
1143 | } |
1144 | |
1145 | off += kasan_metadata_size(s, false); |
1146 | |
1147 | if (size_from_object(s) == off) |
1148 | return 1; |
1149 | |
1150 | return check_bytes_and_report(s, slab, p, "Object padding" , |
1151 | p + off, POISON_INUSE, size_from_object(s) - off); |
1152 | } |
1153 | |
1154 | /* Check the pad bytes at the end of a slab page */ |
1155 | static void slab_pad_check(struct kmem_cache *s, struct slab *slab) |
1156 | { |
1157 | u8 *start; |
1158 | u8 *fault; |
1159 | u8 *end; |
1160 | u8 *pad; |
1161 | int length; |
1162 | int remainder; |
1163 | |
1164 | if (!(s->flags & SLAB_POISON)) |
1165 | return; |
1166 | |
1167 | start = slab_address(slab); |
1168 | length = slab_size(slab); |
1169 | end = start + length; |
1170 | remainder = length % s->size; |
1171 | if (!remainder) |
1172 | return; |
1173 | |
1174 | pad = end - remainder; |
1175 | metadata_access_enable(); |
1176 | fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder); |
1177 | metadata_access_disable(); |
1178 | if (!fault) |
1179 | return; |
1180 | while (end > fault && end[-1] == POISON_INUSE) |
1181 | end--; |
1182 | |
1183 | slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu" , |
1184 | fault, end - 1, fault - start); |
1185 | print_section(KERN_ERR, "Padding " , pad, remainder); |
1186 | |
1187 | restore_bytes(s, "slab padding" , POISON_INUSE, fault, end); |
1188 | } |
1189 | |
1190 | static int check_object(struct kmem_cache *s, struct slab *slab, |
1191 | void *object, u8 val) |
1192 | { |
1193 | u8 *p = object; |
1194 | u8 *endobject = object + s->object_size; |
1195 | unsigned int orig_size; |
1196 | |
1197 | if (s->flags & SLAB_RED_ZONE) { |
1198 | if (!check_bytes_and_report(s, slab, object, "Left Redzone" , |
1199 | object - s->red_left_pad, val, s->red_left_pad)) |
1200 | return 0; |
1201 | |
1202 | if (!check_bytes_and_report(s, slab, object, "Right Redzone" , |
1203 | endobject, val, s->inuse - s->object_size)) |
1204 | return 0; |
1205 | |
1206 | if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) { |
1207 | orig_size = get_orig_size(s, object); |
1208 | |
1209 | if (s->object_size > orig_size && |
1210 | !check_bytes_and_report(s, slab, object, |
1211 | "kmalloc Redzone" , p + orig_size, |
1212 | val, s->object_size - orig_size)) { |
1213 | return 0; |
1214 | } |
1215 | } |
1216 | } else { |
1217 | if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { |
1218 | check_bytes_and_report(s, slab, p, "Alignment padding" , |
1219 | endobject, POISON_INUSE, |
1220 | s->inuse - s->object_size); |
1221 | } |
1222 | } |
1223 | |
1224 | if (s->flags & SLAB_POISON) { |
1225 | if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && |
1226 | (!check_bytes_and_report(s, slab, p, "Poison" , p, |
1227 | POISON_FREE, s->object_size - 1) || |
1228 | !check_bytes_and_report(s, slab, p, "End Poison" , |
1229 | p + s->object_size - 1, POISON_END, 1))) |
1230 | return 0; |
1231 | /* |
1232 | * check_pad_bytes cleans up on its own. |
1233 | */ |
1234 | check_pad_bytes(s, slab, p); |
1235 | } |
1236 | |
1237 | if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE) |
1238 | /* |
1239 | * Object and freepointer overlap. Cannot check |
1240 | * freepointer while object is allocated. |
1241 | */ |
1242 | return 1; |
1243 | |
1244 | /* Check free pointer validity */ |
1245 | if (!check_valid_pointer(s, slab, get_freepointer(s, p))) { |
1246 | object_err(s, slab, p, "Freepointer corrupt" ); |
1247 | /* |
1248 | * No choice but to zap it and thus lose the remainder |
1249 | * of the free objects in this slab. May cause |
1250 | * another error because the object count is now wrong. |
1251 | */ |
1252 | set_freepointer(s, p, NULL); |
1253 | return 0; |
1254 | } |
1255 | return 1; |
1256 | } |
1257 | |
1258 | static int check_slab(struct kmem_cache *s, struct slab *slab) |
1259 | { |
1260 | int maxobj; |
1261 | |
1262 | if (!folio_test_slab(slab_folio(slab))) { |
1263 | slab_err(s, slab, "Not a valid slab page" ); |
1264 | return 0; |
1265 | } |
1266 | |
1267 | maxobj = order_objects(slab_order(slab), s->size); |
1268 | if (slab->objects > maxobj) { |
1269 | slab_err(s, slab, "objects %u > max %u" , |
1270 | slab->objects, maxobj); |
1271 | return 0; |
1272 | } |
1273 | if (slab->inuse > slab->objects) { |
1274 | slab_err(s, slab, "inuse %u > max %u" , |
1275 | slab->inuse, slab->objects); |
1276 | return 0; |
1277 | } |
1278 | /* Slab_pad_check fixes things up after itself */ |
1279 | slab_pad_check(s, slab); |
1280 | return 1; |
1281 | } |
1282 | |
1283 | /* |
1284 | * Determine if a certain object in a slab is on the freelist. Must hold the |
1285 | * slab lock to guarantee that the chains are in a consistent state. |
1286 | */ |
1287 | static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search) |
1288 | { |
1289 | int nr = 0; |
1290 | void *fp; |
1291 | void *object = NULL; |
1292 | int max_objects; |
1293 | |
1294 | fp = slab->freelist; |
1295 | while (fp && nr <= slab->objects) { |
1296 | if (fp == search) |
1297 | return 1; |
1298 | if (!check_valid_pointer(s, slab, fp)) { |
1299 | if (object) { |
1300 | object_err(s, slab, object, |
1301 | "Freechain corrupt" ); |
1302 | set_freepointer(s, object, NULL); |
1303 | } else { |
1304 | slab_err(s, slab, "Freepointer corrupt" ); |
1305 | slab->freelist = NULL; |
1306 | slab->inuse = slab->objects; |
1307 | slab_fix(s, "Freelist cleared" ); |
1308 | return 0; |
1309 | } |
1310 | break; |
1311 | } |
1312 | object = fp; |
1313 | fp = get_freepointer(s, object); |
1314 | nr++; |
1315 | } |
1316 | |
1317 | max_objects = order_objects(slab_order(slab), s->size); |
1318 | if (max_objects > MAX_OBJS_PER_PAGE) |
1319 | max_objects = MAX_OBJS_PER_PAGE; |
1320 | |
1321 | if (slab->objects != max_objects) { |
1322 | slab_err(s, slab, "Wrong number of objects. Found %d but should be %d" , |
1323 | slab->objects, max_objects); |
1324 | slab->objects = max_objects; |
1325 | slab_fix(s, "Number of objects adjusted" ); |
1326 | } |
1327 | if (slab->inuse != slab->objects - nr) { |
1328 | slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d" , |
1329 | slab->inuse, slab->objects - nr); |
1330 | slab->inuse = slab->objects - nr; |
1331 | slab_fix(s, "Object count adjusted" ); |
1332 | } |
1333 | return search == NULL; |
1334 | } |
1335 | |
1336 | static void trace(struct kmem_cache *s, struct slab *slab, void *object, |
1337 | int alloc) |
1338 | { |
1339 | if (s->flags & SLAB_TRACE) { |
1340 | pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n" , |
1341 | s->name, |
1342 | alloc ? "alloc" : "free" , |
1343 | object, slab->inuse, |
1344 | slab->freelist); |
1345 | |
1346 | if (!alloc) |
1347 | print_section(KERN_INFO, "Object " , (void *)object, |
1348 | s->object_size); |
1349 | |
1350 | dump_stack(); |
1351 | } |
1352 | } |
1353 | |
1354 | /* |
1355 | * Tracking of fully allocated slabs for debugging purposes. |
1356 | */ |
1357 | static void add_full(struct kmem_cache *s, |
1358 | struct kmem_cache_node *n, struct slab *slab) |
1359 | { |
1360 | if (!(s->flags & SLAB_STORE_USER)) |
1361 | return; |
1362 | |
1363 | lockdep_assert_held(&n->list_lock); |
1364 | list_add(&slab->slab_list, &n->full); |
1365 | } |
1366 | |
1367 | static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab) |
1368 | { |
1369 | if (!(s->flags & SLAB_STORE_USER)) |
1370 | return; |
1371 | |
1372 | lockdep_assert_held(&n->list_lock); |
1373 | list_del(&slab->slab_list); |
1374 | } |
1375 | |
1376 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
1377 | { |
1378 | return atomic_long_read(&n->nr_slabs); |
1379 | } |
1380 | |
1381 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) |
1382 | { |
1383 | struct kmem_cache_node *n = get_node(s, node); |
1384 | |
1385 | /* |
1386 | * May be called early in order to allocate a slab for the |
1387 | * kmem_cache_node structure. Solve the chicken-egg |
1388 | * dilemma by deferring the increment of the count during |
1389 | * bootstrap (see early_kmem_cache_node_alloc). |
1390 | */ |
1391 | if (likely(n)) { |
1392 | atomic_long_inc(&n->nr_slabs); |
1393 | atomic_long_add(objects, &n->total_objects); |
1394 | } |
1395 | } |
1396 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) |
1397 | { |
1398 | struct kmem_cache_node *n = get_node(s, node); |
1399 | |
1400 | atomic_long_dec(&n->nr_slabs); |
1401 | atomic_long_sub(objects, &n->total_objects); |
1402 | } |
1403 | |
1404 | /* Object debug checks for alloc/free paths */ |
1405 | static void setup_object_debug(struct kmem_cache *s, void *object) |
1406 | { |
1407 | if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)) |
1408 | return; |
1409 | |
1410 | init_object(s, object, SLUB_RED_INACTIVE); |
1411 | init_tracking(s, object); |
1412 | } |
1413 | |
1414 | static |
1415 | void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) |
1416 | { |
1417 | if (!kmem_cache_debug_flags(s, SLAB_POISON)) |
1418 | return; |
1419 | |
1420 | metadata_access_enable(); |
1421 | memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab)); |
1422 | metadata_access_disable(); |
1423 | } |
1424 | |
1425 | static inline int alloc_consistency_checks(struct kmem_cache *s, |
1426 | struct slab *slab, void *object) |
1427 | { |
1428 | if (!check_slab(s, slab)) |
1429 | return 0; |
1430 | |
1431 | if (!check_valid_pointer(s, slab, object)) { |
1432 | object_err(s, slab, object, "Freelist Pointer check fails" ); |
1433 | return 0; |
1434 | } |
1435 | |
1436 | if (!check_object(s, slab, object, SLUB_RED_INACTIVE)) |
1437 | return 0; |
1438 | |
1439 | return 1; |
1440 | } |
1441 | |
1442 | static noinline bool alloc_debug_processing(struct kmem_cache *s, |
1443 | struct slab *slab, void *object, int orig_size) |
1444 | { |
1445 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
1446 | if (!alloc_consistency_checks(s, slab, object)) |
1447 | goto bad; |
1448 | } |
1449 | |
1450 | /* Success. Perform special debug activities for allocs */ |
1451 | trace(s, slab, object, 1); |
1452 | set_orig_size(s, object, orig_size); |
1453 | init_object(s, object, SLUB_RED_ACTIVE); |
1454 | return true; |
1455 | |
1456 | bad: |
1457 | if (folio_test_slab(slab_folio(slab))) { |
1458 | /* |
1459 | * If this is a slab page then lets do the best we can |
1460 | * to avoid issues in the future. Marking all objects |
1461 | * as used avoids touching the remaining objects. |
1462 | */ |
1463 | slab_fix(s, "Marking all objects used" ); |
1464 | slab->inuse = slab->objects; |
1465 | slab->freelist = NULL; |
1466 | } |
1467 | return false; |
1468 | } |
1469 | |
1470 | static inline int free_consistency_checks(struct kmem_cache *s, |
1471 | struct slab *slab, void *object, unsigned long addr) |
1472 | { |
1473 | if (!check_valid_pointer(s, slab, object)) { |
1474 | slab_err(s, slab, "Invalid object pointer 0x%p" , object); |
1475 | return 0; |
1476 | } |
1477 | |
1478 | if (on_freelist(s, slab, object)) { |
1479 | object_err(s, slab, object, "Object already free" ); |
1480 | return 0; |
1481 | } |
1482 | |
1483 | if (!check_object(s, slab, object, SLUB_RED_ACTIVE)) |
1484 | return 0; |
1485 | |
1486 | if (unlikely(s != slab->slab_cache)) { |
1487 | if (!folio_test_slab(slab_folio(slab))) { |
1488 | slab_err(s, slab, "Attempt to free object(0x%p) outside of slab" , |
1489 | object); |
1490 | } else if (!slab->slab_cache) { |
1491 | pr_err("SLUB <none>: no slab for object 0x%p.\n" , |
1492 | object); |
1493 | dump_stack(); |
1494 | } else |
1495 | object_err(s, slab, object, |
1496 | "page slab pointer corrupt." ); |
1497 | return 0; |
1498 | } |
1499 | return 1; |
1500 | } |
1501 | |
1502 | /* |
1503 | * Parse a block of slub_debug options. Blocks are delimited by ';' |
1504 | * |
1505 | * @str: start of block |
1506 | * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified |
1507 | * @slabs: return start of list of slabs, or NULL when there's no list |
1508 | * @init: assume this is initial parsing and not per-kmem-create parsing |
1509 | * |
1510 | * returns the start of next block if there's any, or NULL |
1511 | */ |
1512 | static char * |
1513 | parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init) |
1514 | { |
1515 | bool higher_order_disable = false; |
1516 | |
1517 | /* Skip any completely empty blocks */ |
1518 | while (*str && *str == ';') |
1519 | str++; |
1520 | |
1521 | if (*str == ',') { |
1522 | /* |
1523 | * No options but restriction on slabs. This means full |
1524 | * debugging for slabs matching a pattern. |
1525 | */ |
1526 | *flags = DEBUG_DEFAULT_FLAGS; |
1527 | goto check_slabs; |
1528 | } |
1529 | *flags = 0; |
1530 | |
1531 | /* Determine which debug features should be switched on */ |
1532 | for (; *str && *str != ',' && *str != ';'; str++) { |
1533 | switch (tolower(*str)) { |
1534 | case '-': |
1535 | *flags = 0; |
1536 | break; |
1537 | case 'f': |
1538 | *flags |= SLAB_CONSISTENCY_CHECKS; |
1539 | break; |
1540 | case 'z': |
1541 | *flags |= SLAB_RED_ZONE; |
1542 | break; |
1543 | case 'p': |
1544 | *flags |= SLAB_POISON; |
1545 | break; |
1546 | case 'u': |
1547 | *flags |= SLAB_STORE_USER; |
1548 | break; |
1549 | case 't': |
1550 | *flags |= SLAB_TRACE; |
1551 | break; |
1552 | case 'a': |
1553 | *flags |= SLAB_FAILSLAB; |
1554 | break; |
1555 | case 'o': |
1556 | /* |
1557 | * Avoid enabling debugging on caches if its minimum |
1558 | * order would increase as a result. |
1559 | */ |
1560 | higher_order_disable = true; |
1561 | break; |
1562 | default: |
1563 | if (init) |
1564 | pr_err("slub_debug option '%c' unknown. skipped\n" , *str); |
1565 | } |
1566 | } |
1567 | check_slabs: |
1568 | if (*str == ',') |
1569 | *slabs = ++str; |
1570 | else |
1571 | *slabs = NULL; |
1572 | |
1573 | /* Skip over the slab list */ |
1574 | while (*str && *str != ';') |
1575 | str++; |
1576 | |
1577 | /* Skip any completely empty blocks */ |
1578 | while (*str && *str == ';') |
1579 | str++; |
1580 | |
1581 | if (init && higher_order_disable) |
1582 | disable_higher_order_debug = 1; |
1583 | |
1584 | if (*str) |
1585 | return str; |
1586 | else |
1587 | return NULL; |
1588 | } |
1589 | |
1590 | static int __init setup_slub_debug(char *str) |
1591 | { |
1592 | slab_flags_t flags; |
1593 | slab_flags_t global_flags; |
1594 | char *saved_str; |
1595 | char *slab_list; |
1596 | bool global_slub_debug_changed = false; |
1597 | bool slab_list_specified = false; |
1598 | |
1599 | global_flags = DEBUG_DEFAULT_FLAGS; |
1600 | if (*str++ != '=' || !*str) |
1601 | /* |
1602 | * No options specified. Switch on full debugging. |
1603 | */ |
1604 | goto out; |
1605 | |
1606 | saved_str = str; |
1607 | while (str) { |
1608 | str = parse_slub_debug_flags(str, &flags, &slab_list, true); |
1609 | |
1610 | if (!slab_list) { |
1611 | global_flags = flags; |
1612 | global_slub_debug_changed = true; |
1613 | } else { |
1614 | slab_list_specified = true; |
1615 | if (flags & SLAB_STORE_USER) |
1616 | stack_depot_request_early_init(); |
1617 | } |
1618 | } |
1619 | |
1620 | /* |
1621 | * For backwards compatibility, a single list of flags with list of |
1622 | * slabs means debugging is only changed for those slabs, so the global |
1623 | * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending |
1624 | * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as |
1625 | * long as there is no option specifying flags without a slab list. |
1626 | */ |
1627 | if (slab_list_specified) { |
1628 | if (!global_slub_debug_changed) |
1629 | global_flags = slub_debug; |
1630 | slub_debug_string = saved_str; |
1631 | } |
1632 | out: |
1633 | slub_debug = global_flags; |
1634 | if (slub_debug & SLAB_STORE_USER) |
1635 | stack_depot_request_early_init(); |
1636 | if (slub_debug != 0 || slub_debug_string) |
1637 | static_branch_enable(&slub_debug_enabled); |
1638 | else |
1639 | static_branch_disable(&slub_debug_enabled); |
1640 | if ((static_branch_unlikely(&init_on_alloc) || |
1641 | static_branch_unlikely(&init_on_free)) && |
1642 | (slub_debug & SLAB_POISON)) |
1643 | pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n" ); |
1644 | return 1; |
1645 | } |
1646 | |
1647 | __setup("slub_debug" , setup_slub_debug); |
1648 | |
1649 | /* |
1650 | * kmem_cache_flags - apply debugging options to the cache |
1651 | * @object_size: the size of an object without meta data |
1652 | * @flags: flags to set |
1653 | * @name: name of the cache |
1654 | * |
1655 | * Debug option(s) are applied to @flags. In addition to the debug |
1656 | * option(s), if a slab name (or multiple) is specified i.e. |
1657 | * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ... |
1658 | * then only the select slabs will receive the debug option(s). |
1659 | */ |
1660 | slab_flags_t kmem_cache_flags(unsigned int object_size, |
1661 | slab_flags_t flags, const char *name) |
1662 | { |
1663 | char *iter; |
1664 | size_t len; |
1665 | char *next_block; |
1666 | slab_flags_t block_flags; |
1667 | slab_flags_t slub_debug_local = slub_debug; |
1668 | |
1669 | if (flags & SLAB_NO_USER_FLAGS) |
1670 | return flags; |
1671 | |
1672 | /* |
1673 | * If the slab cache is for debugging (e.g. kmemleak) then |
1674 | * don't store user (stack trace) information by default, |
1675 | * but let the user enable it via the command line below. |
1676 | */ |
1677 | if (flags & SLAB_NOLEAKTRACE) |
1678 | slub_debug_local &= ~SLAB_STORE_USER; |
1679 | |
1680 | len = strlen(name); |
1681 | next_block = slub_debug_string; |
1682 | /* Go through all blocks of debug options, see if any matches our slab's name */ |
1683 | while (next_block) { |
1684 | next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false); |
1685 | if (!iter) |
1686 | continue; |
1687 | /* Found a block that has a slab list, search it */ |
1688 | while (*iter) { |
1689 | char *end, *glob; |
1690 | size_t cmplen; |
1691 | |
1692 | end = strchrnul(iter, ','); |
1693 | if (next_block && next_block < end) |
1694 | end = next_block - 1; |
1695 | |
1696 | glob = strnchr(iter, end - iter, '*'); |
1697 | if (glob) |
1698 | cmplen = glob - iter; |
1699 | else |
1700 | cmplen = max_t(size_t, len, (end - iter)); |
1701 | |
1702 | if (!strncmp(name, iter, cmplen)) { |
1703 | flags |= block_flags; |
1704 | return flags; |
1705 | } |
1706 | |
1707 | if (!*end || *end == ';') |
1708 | break; |
1709 | iter = end + 1; |
1710 | } |
1711 | } |
1712 | |
1713 | return flags | slub_debug_local; |
1714 | } |
1715 | #else /* !CONFIG_SLUB_DEBUG */ |
1716 | static inline void setup_object_debug(struct kmem_cache *s, void *object) {} |
1717 | static inline |
1718 | void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {} |
1719 | |
1720 | static inline bool alloc_debug_processing(struct kmem_cache *s, |
1721 | struct slab *slab, void *object, int orig_size) { return true; } |
1722 | |
1723 | static inline bool free_debug_processing(struct kmem_cache *s, |
1724 | struct slab *slab, void *head, void *tail, int *bulk_cnt, |
1725 | unsigned long addr, depot_stack_handle_t handle) { return true; } |
1726 | |
1727 | static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {} |
1728 | static inline int check_object(struct kmem_cache *s, struct slab *slab, |
1729 | void *object, u8 val) { return 1; } |
1730 | static inline depot_stack_handle_t set_track_prepare(void) { return 0; } |
1731 | static inline void set_track(struct kmem_cache *s, void *object, |
1732 | enum track_item alloc, unsigned long addr) {} |
1733 | static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, |
1734 | struct slab *slab) {} |
1735 | static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, |
1736 | struct slab *slab) {} |
1737 | slab_flags_t kmem_cache_flags(unsigned int object_size, |
1738 | slab_flags_t flags, const char *name) |
1739 | { |
1740 | return flags; |
1741 | } |
1742 | #define slub_debug 0 |
1743 | |
1744 | #define disable_higher_order_debug 0 |
1745 | |
1746 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
1747 | { return 0; } |
1748 | static inline void inc_slabs_node(struct kmem_cache *s, int node, |
1749 | int objects) {} |
1750 | static inline void dec_slabs_node(struct kmem_cache *s, int node, |
1751 | int objects) {} |
1752 | |
1753 | #ifndef CONFIG_SLUB_TINY |
1754 | static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab, |
1755 | void **freelist, void *nextfree) |
1756 | { |
1757 | return false; |
1758 | } |
1759 | #endif |
1760 | #endif /* CONFIG_SLUB_DEBUG */ |
1761 | |
1762 | /* |
1763 | * Hooks for other subsystems that check memory allocations. In a typical |
1764 | * production configuration these hooks all should produce no code at all. |
1765 | */ |
1766 | static __always_inline bool slab_free_hook(struct kmem_cache *s, |
1767 | void *x, bool init) |
1768 | { |
1769 | kmemleak_free_recursive(ptr: x, flags: s->flags); |
1770 | kmsan_slab_free(s, object: x); |
1771 | |
1772 | debug_check_no_locks_freed(from: x, len: s->object_size); |
1773 | |
1774 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) |
1775 | debug_check_no_obj_freed(address: x, size: s->object_size); |
1776 | |
1777 | /* Use KCSAN to help debug racy use-after-free. */ |
1778 | if (!(s->flags & SLAB_TYPESAFE_BY_RCU)) |
1779 | __kcsan_check_access(ptr: x, size: s->object_size, |
1780 | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT); |
1781 | |
1782 | /* |
1783 | * As memory initialization might be integrated into KASAN, |
1784 | * kasan_slab_free and initialization memset's must be |
1785 | * kept together to avoid discrepancies in behavior. |
1786 | * |
1787 | * The initialization memset's clear the object and the metadata, |
1788 | * but don't touch the SLAB redzone. |
1789 | */ |
1790 | if (init) { |
1791 | int rsize; |
1792 | |
1793 | if (!kasan_has_integrated_init()) |
1794 | memset(kasan_reset_tag(x), 0, s->object_size); |
1795 | rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0; |
1796 | memset((char *)kasan_reset_tag(x) + s->inuse, 0, |
1797 | s->size - s->inuse - rsize); |
1798 | } |
1799 | /* KASAN might put x into memory quarantine, delaying its reuse. */ |
1800 | return kasan_slab_free(s, object: x, init); |
1801 | } |
1802 | |
1803 | static inline bool slab_free_freelist_hook(struct kmem_cache *s, |
1804 | void **head, void **tail, |
1805 | int *cnt) |
1806 | { |
1807 | |
1808 | void *object; |
1809 | void *next = *head; |
1810 | void *old_tail = *tail ? *tail : *head; |
1811 | |
1812 | if (is_kfence_address(addr: next)) { |
1813 | slab_free_hook(s, x: next, init: false); |
1814 | return true; |
1815 | } |
1816 | |
1817 | /* Head and tail of the reconstructed freelist */ |
1818 | *head = NULL; |
1819 | *tail = NULL; |
1820 | |
1821 | do { |
1822 | object = next; |
1823 | next = get_freepointer(s, object); |
1824 | |
1825 | /* If object's reuse doesn't have to be delayed */ |
1826 | if (!slab_free_hook(s, x: object, init: slab_want_init_on_free(c: s))) { |
1827 | /* Move object to the new freelist */ |
1828 | set_freepointer(s, object, fp: *head); |
1829 | *head = object; |
1830 | if (!*tail) |
1831 | *tail = object; |
1832 | } else { |
1833 | /* |
1834 | * Adjust the reconstructed freelist depth |
1835 | * accordingly if object's reuse is delayed. |
1836 | */ |
1837 | --(*cnt); |
1838 | } |
1839 | } while (object != old_tail); |
1840 | |
1841 | if (*head == *tail) |
1842 | *tail = NULL; |
1843 | |
1844 | return *head != NULL; |
1845 | } |
1846 | |
1847 | static void *setup_object(struct kmem_cache *s, void *object) |
1848 | { |
1849 | setup_object_debug(s, object); |
1850 | object = kasan_init_slab_obj(cache: s, object); |
1851 | if (unlikely(s->ctor)) { |
1852 | kasan_unpoison_object_data(cache: s, object); |
1853 | s->ctor(object); |
1854 | kasan_poison_object_data(cache: s, object); |
1855 | } |
1856 | return object; |
1857 | } |
1858 | |
1859 | /* |
1860 | * Slab allocation and freeing |
1861 | */ |
1862 | static inline struct slab *alloc_slab_page(gfp_t flags, int node, |
1863 | struct kmem_cache_order_objects oo) |
1864 | { |
1865 | struct folio *folio; |
1866 | struct slab *slab; |
1867 | unsigned int order = oo_order(x: oo); |
1868 | |
1869 | if (node == NUMA_NO_NODE) |
1870 | folio = (struct folio *)alloc_pages(gfp: flags, order); |
1871 | else |
1872 | folio = (struct folio *)__alloc_pages_node(nid: node, gfp_mask: flags, order); |
1873 | |
1874 | if (!folio) |
1875 | return NULL; |
1876 | |
1877 | slab = folio_slab(folio); |
1878 | __folio_set_slab(folio); |
1879 | /* Make the flag visible before any changes to folio->mapping */ |
1880 | smp_wmb(); |
1881 | if (folio_is_pfmemalloc(folio)) |
1882 | slab_set_pfmemalloc(slab); |
1883 | |
1884 | return slab; |
1885 | } |
1886 | |
1887 | #ifdef CONFIG_SLAB_FREELIST_RANDOM |
1888 | /* Pre-initialize the random sequence cache */ |
1889 | static int init_cache_random_seq(struct kmem_cache *s) |
1890 | { |
1891 | unsigned int count = oo_objects(s->oo); |
1892 | int err; |
1893 | |
1894 | /* Bailout if already initialised */ |
1895 | if (s->random_seq) |
1896 | return 0; |
1897 | |
1898 | err = cache_random_seq_create(s, count, GFP_KERNEL); |
1899 | if (err) { |
1900 | pr_err("SLUB: Unable to initialize free list for %s\n" , |
1901 | s->name); |
1902 | return err; |
1903 | } |
1904 | |
1905 | /* Transform to an offset on the set of pages */ |
1906 | if (s->random_seq) { |
1907 | unsigned int i; |
1908 | |
1909 | for (i = 0; i < count; i++) |
1910 | s->random_seq[i] *= s->size; |
1911 | } |
1912 | return 0; |
1913 | } |
1914 | |
1915 | /* Initialize each random sequence freelist per cache */ |
1916 | static void __init init_freelist_randomization(void) |
1917 | { |
1918 | struct kmem_cache *s; |
1919 | |
1920 | mutex_lock(&slab_mutex); |
1921 | |
1922 | list_for_each_entry(s, &slab_caches, list) |
1923 | init_cache_random_seq(s); |
1924 | |
1925 | mutex_unlock(&slab_mutex); |
1926 | } |
1927 | |
1928 | /* Get the next entry on the pre-computed freelist randomized */ |
1929 | static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab, |
1930 | unsigned long *pos, void *start, |
1931 | unsigned long page_limit, |
1932 | unsigned long freelist_count) |
1933 | { |
1934 | unsigned int idx; |
1935 | |
1936 | /* |
1937 | * If the target page allocation failed, the number of objects on the |
1938 | * page might be smaller than the usual size defined by the cache. |
1939 | */ |
1940 | do { |
1941 | idx = s->random_seq[*pos]; |
1942 | *pos += 1; |
1943 | if (*pos >= freelist_count) |
1944 | *pos = 0; |
1945 | } while (unlikely(idx >= page_limit)); |
1946 | |
1947 | return (char *)start + idx; |
1948 | } |
1949 | |
1950 | /* Shuffle the single linked freelist based on a random pre-computed sequence */ |
1951 | static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab) |
1952 | { |
1953 | void *start; |
1954 | void *cur; |
1955 | void *next; |
1956 | unsigned long idx, pos, page_limit, freelist_count; |
1957 | |
1958 | if (slab->objects < 2 || !s->random_seq) |
1959 | return false; |
1960 | |
1961 | freelist_count = oo_objects(s->oo); |
1962 | pos = get_random_u32_below(freelist_count); |
1963 | |
1964 | page_limit = slab->objects * s->size; |
1965 | start = fixup_red_left(s, slab_address(slab)); |
1966 | |
1967 | /* First entry is used as the base of the freelist */ |
1968 | cur = next_freelist_entry(s, slab, &pos, start, page_limit, |
1969 | freelist_count); |
1970 | cur = setup_object(s, cur); |
1971 | slab->freelist = cur; |
1972 | |
1973 | for (idx = 1; idx < slab->objects; idx++) { |
1974 | next = next_freelist_entry(s, slab, &pos, start, page_limit, |
1975 | freelist_count); |
1976 | next = setup_object(s, next); |
1977 | set_freepointer(s, cur, next); |
1978 | cur = next; |
1979 | } |
1980 | set_freepointer(s, cur, NULL); |
1981 | |
1982 | return true; |
1983 | } |
1984 | #else |
1985 | static inline int init_cache_random_seq(struct kmem_cache *s) |
1986 | { |
1987 | return 0; |
1988 | } |
1989 | static inline void init_freelist_randomization(void) { } |
1990 | static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab) |
1991 | { |
1992 | return false; |
1993 | } |
1994 | #endif /* CONFIG_SLAB_FREELIST_RANDOM */ |
1995 | |
1996 | static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) |
1997 | { |
1998 | struct slab *slab; |
1999 | struct kmem_cache_order_objects oo = s->oo; |
2000 | gfp_t alloc_gfp; |
2001 | void *start, *p, *next; |
2002 | int idx; |
2003 | bool shuffle; |
2004 | |
2005 | flags &= gfp_allowed_mask; |
2006 | |
2007 | flags |= s->allocflags; |
2008 | |
2009 | /* |
2010 | * Let the initial higher-order allocation fail under memory pressure |
2011 | * so we fall-back to the minimum order allocation. |
2012 | */ |
2013 | alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; |
2014 | if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(x: oo) > oo_order(x: s->min)) |
2015 | alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM; |
2016 | |
2017 | slab = alloc_slab_page(flags: alloc_gfp, node, oo); |
2018 | if (unlikely(!slab)) { |
2019 | oo = s->min; |
2020 | alloc_gfp = flags; |
2021 | /* |
2022 | * Allocation may have failed due to fragmentation. |
2023 | * Try a lower order alloc if possible |
2024 | */ |
2025 | slab = alloc_slab_page(flags: alloc_gfp, node, oo); |
2026 | if (unlikely(!slab)) |
2027 | return NULL; |
2028 | stat(s, si: ORDER_FALLBACK); |
2029 | } |
2030 | |
2031 | slab->objects = oo_objects(x: oo); |
2032 | slab->inuse = 0; |
2033 | slab->frozen = 0; |
2034 | |
2035 | account_slab(slab, order: oo_order(x: oo), s, gfp: flags); |
2036 | |
2037 | slab->slab_cache = s; |
2038 | |
2039 | kasan_poison_slab(slab); |
2040 | |
2041 | start = slab_address(slab); |
2042 | |
2043 | setup_slab_debug(s, slab, addr: start); |
2044 | |
2045 | shuffle = shuffle_freelist(s, slab); |
2046 | |
2047 | if (!shuffle) { |
2048 | start = fixup_red_left(s, p: start); |
2049 | start = setup_object(s, object: start); |
2050 | slab->freelist = start; |
2051 | for (idx = 0, p = start; idx < slab->objects - 1; idx++) { |
2052 | next = p + s->size; |
2053 | next = setup_object(s, object: next); |
2054 | set_freepointer(s, object: p, fp: next); |
2055 | p = next; |
2056 | } |
2057 | set_freepointer(s, object: p, NULL); |
2058 | } |
2059 | |
2060 | return slab; |
2061 | } |
2062 | |
2063 | static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node) |
2064 | { |
2065 | if (unlikely(flags & GFP_SLAB_BUG_MASK)) |
2066 | flags = kmalloc_fix_flags(flags); |
2067 | |
2068 | WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO)); |
2069 | |
2070 | return allocate_slab(s, |
2071 | flags: flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); |
2072 | } |
2073 | |
2074 | static void __free_slab(struct kmem_cache *s, struct slab *slab) |
2075 | { |
2076 | struct folio *folio = slab_folio(slab); |
2077 | int order = folio_order(folio); |
2078 | int pages = 1 << order; |
2079 | |
2080 | __slab_clear_pfmemalloc(slab); |
2081 | folio->mapping = NULL; |
2082 | /* Make the mapping reset visible before clearing the flag */ |
2083 | smp_wmb(); |
2084 | __folio_clear_slab(folio); |
2085 | mm_account_reclaimed_pages(pages); |
2086 | unaccount_slab(slab, order, s); |
2087 | __free_pages(page: &folio->page, order); |
2088 | } |
2089 | |
2090 | static void rcu_free_slab(struct rcu_head *h) |
2091 | { |
2092 | struct slab *slab = container_of(h, struct slab, rcu_head); |
2093 | |
2094 | __free_slab(s: slab->slab_cache, slab); |
2095 | } |
2096 | |
2097 | static void free_slab(struct kmem_cache *s, struct slab *slab) |
2098 | { |
2099 | if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) { |
2100 | void *p; |
2101 | |
2102 | slab_pad_check(s, slab); |
2103 | for_each_object(p, s, slab_address(slab), slab->objects) |
2104 | check_object(s, slab, object: p, SLUB_RED_INACTIVE); |
2105 | } |
2106 | |
2107 | if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) |
2108 | call_rcu(head: &slab->rcu_head, func: rcu_free_slab); |
2109 | else |
2110 | __free_slab(s, slab); |
2111 | } |
2112 | |
2113 | static void discard_slab(struct kmem_cache *s, struct slab *slab) |
2114 | { |
2115 | dec_slabs_node(s, node: slab_nid(slab), objects: slab->objects); |
2116 | free_slab(s, slab); |
2117 | } |
2118 | |
2119 | /* |
2120 | * Management of partially allocated slabs. |
2121 | */ |
2122 | static inline void |
2123 | __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail) |
2124 | { |
2125 | n->nr_partial++; |
2126 | if (tail == DEACTIVATE_TO_TAIL) |
2127 | list_add_tail(new: &slab->slab_list, head: &n->partial); |
2128 | else |
2129 | list_add(new: &slab->slab_list, head: &n->partial); |
2130 | } |
2131 | |
2132 | static inline void add_partial(struct kmem_cache_node *n, |
2133 | struct slab *slab, int tail) |
2134 | { |
2135 | lockdep_assert_held(&n->list_lock); |
2136 | __add_partial(n, slab, tail); |
2137 | } |
2138 | |
2139 | static inline void remove_partial(struct kmem_cache_node *n, |
2140 | struct slab *slab) |
2141 | { |
2142 | lockdep_assert_held(&n->list_lock); |
2143 | list_del(entry: &slab->slab_list); |
2144 | n->nr_partial--; |
2145 | } |
2146 | |
2147 | /* |
2148 | * Called only for kmem_cache_debug() caches instead of acquire_slab(), with a |
2149 | * slab from the n->partial list. Remove only a single object from the slab, do |
2150 | * the alloc_debug_processing() checks and leave the slab on the list, or move |
2151 | * it to full list if it was the last free object. |
2152 | */ |
2153 | static void *alloc_single_from_partial(struct kmem_cache *s, |
2154 | struct kmem_cache_node *n, struct slab *slab, int orig_size) |
2155 | { |
2156 | void *object; |
2157 | |
2158 | lockdep_assert_held(&n->list_lock); |
2159 | |
2160 | object = slab->freelist; |
2161 | slab->freelist = get_freepointer(s, object); |
2162 | slab->inuse++; |
2163 | |
2164 | if (!alloc_debug_processing(s, slab, object, orig_size)) { |
2165 | remove_partial(n, slab); |
2166 | return NULL; |
2167 | } |
2168 | |
2169 | if (slab->inuse == slab->objects) { |
2170 | remove_partial(n, slab); |
2171 | add_full(s, n, slab); |
2172 | } |
2173 | |
2174 | return object; |
2175 | } |
2176 | |
2177 | /* |
2178 | * Called only for kmem_cache_debug() caches to allocate from a freshly |
2179 | * allocated slab. Allocate a single object instead of whole freelist |
2180 | * and put the slab to the partial (or full) list. |
2181 | */ |
2182 | static void *alloc_single_from_new_slab(struct kmem_cache *s, |
2183 | struct slab *slab, int orig_size) |
2184 | { |
2185 | int nid = slab_nid(slab); |
2186 | struct kmem_cache_node *n = get_node(s, node: nid); |
2187 | unsigned long flags; |
2188 | void *object; |
2189 | |
2190 | |
2191 | object = slab->freelist; |
2192 | slab->freelist = get_freepointer(s, object); |
2193 | slab->inuse = 1; |
2194 | |
2195 | if (!alloc_debug_processing(s, slab, object, orig_size)) |
2196 | /* |
2197 | * It's not really expected that this would fail on a |
2198 | * freshly allocated slab, but a concurrent memory |
2199 | * corruption in theory could cause that. |
2200 | */ |
2201 | return NULL; |
2202 | |
2203 | spin_lock_irqsave(&n->list_lock, flags); |
2204 | |
2205 | if (slab->inuse == slab->objects) |
2206 | add_full(s, n, slab); |
2207 | else |
2208 | add_partial(n, slab, tail: DEACTIVATE_TO_HEAD); |
2209 | |
2210 | inc_slabs_node(s, node: nid, objects: slab->objects); |
2211 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
2212 | |
2213 | return object; |
2214 | } |
2215 | |
2216 | /* |
2217 | * Remove slab from the partial list, freeze it and |
2218 | * return the pointer to the freelist. |
2219 | * |
2220 | * Returns a list of objects or NULL if it fails. |
2221 | */ |
2222 | static inline void *acquire_slab(struct kmem_cache *s, |
2223 | struct kmem_cache_node *n, struct slab *slab, |
2224 | int mode) |
2225 | { |
2226 | void *freelist; |
2227 | unsigned long counters; |
2228 | struct slab new; |
2229 | |
2230 | lockdep_assert_held(&n->list_lock); |
2231 | |
2232 | /* |
2233 | * Zap the freelist and set the frozen bit. |
2234 | * The old freelist is the list of objects for the |
2235 | * per cpu allocation list. |
2236 | */ |
2237 | freelist = slab->freelist; |
2238 | counters = slab->counters; |
2239 | new.counters = counters; |
2240 | if (mode) { |
2241 | new.inuse = slab->objects; |
2242 | new.freelist = NULL; |
2243 | } else { |
2244 | new.freelist = freelist; |
2245 | } |
2246 | |
2247 | VM_BUG_ON(new.frozen); |
2248 | new.frozen = 1; |
2249 | |
2250 | if (!__slab_update_freelist(s, slab, |
2251 | freelist_old: freelist, counters_old: counters, |
2252 | freelist_new: new.freelist, counters_new: new.counters, |
2253 | n: "acquire_slab" )) |
2254 | return NULL; |
2255 | |
2256 | remove_partial(n, slab); |
2257 | WARN_ON(!freelist); |
2258 | return freelist; |
2259 | } |
2260 | |
2261 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
2262 | static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain); |
2263 | #else |
2264 | static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab, |
2265 | int drain) { } |
2266 | #endif |
2267 | static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags); |
2268 | |
2269 | /* |
2270 | * Try to allocate a partial slab from a specific node. |
2271 | */ |
2272 | static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, |
2273 | struct partial_context *pc) |
2274 | { |
2275 | struct slab *slab, *slab2; |
2276 | void *object = NULL; |
2277 | unsigned long flags; |
2278 | unsigned int partial_slabs = 0; |
2279 | |
2280 | /* |
2281 | * Racy check. If we mistakenly see no partial slabs then we |
2282 | * just allocate an empty slab. If we mistakenly try to get a |
2283 | * partial slab and there is none available then get_partial() |
2284 | * will return NULL. |
2285 | */ |
2286 | if (!n || !n->nr_partial) |
2287 | return NULL; |
2288 | |
2289 | spin_lock_irqsave(&n->list_lock, flags); |
2290 | list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) { |
2291 | void *t; |
2292 | |
2293 | if (!pfmemalloc_match(slab, gfpflags: pc->flags)) |
2294 | continue; |
2295 | |
2296 | if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) { |
2297 | object = alloc_single_from_partial(s, n, slab, |
2298 | orig_size: pc->orig_size); |
2299 | if (object) |
2300 | break; |
2301 | continue; |
2302 | } |
2303 | |
2304 | t = acquire_slab(s, n, slab, mode: object == NULL); |
2305 | if (!t) |
2306 | break; |
2307 | |
2308 | if (!object) { |
2309 | *pc->slab = slab; |
2310 | stat(s, si: ALLOC_FROM_PARTIAL); |
2311 | object = t; |
2312 | } else { |
2313 | put_cpu_partial(s, slab, drain: 0); |
2314 | stat(s, si: CPU_PARTIAL_NODE); |
2315 | partial_slabs++; |
2316 | } |
2317 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
2318 | if (!kmem_cache_has_cpu_partial(s) |
2319 | || partial_slabs > s->cpu_partial_slabs / 2) |
2320 | break; |
2321 | #else |
2322 | break; |
2323 | #endif |
2324 | |
2325 | } |
2326 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
2327 | return object; |
2328 | } |
2329 | |
2330 | /* |
2331 | * Get a slab from somewhere. Search in increasing NUMA distances. |
2332 | */ |
2333 | static void *get_any_partial(struct kmem_cache *s, struct partial_context *pc) |
2334 | { |
2335 | #ifdef CONFIG_NUMA |
2336 | struct zonelist *zonelist; |
2337 | struct zoneref *z; |
2338 | struct zone *zone; |
2339 | enum zone_type highest_zoneidx = gfp_zone(flags: pc->flags); |
2340 | void *object; |
2341 | unsigned int cpuset_mems_cookie; |
2342 | |
2343 | /* |
2344 | * The defrag ratio allows a configuration of the tradeoffs between |
2345 | * inter node defragmentation and node local allocations. A lower |
2346 | * defrag_ratio increases the tendency to do local allocations |
2347 | * instead of attempting to obtain partial slabs from other nodes. |
2348 | * |
2349 | * If the defrag_ratio is set to 0 then kmalloc() always |
2350 | * returns node local objects. If the ratio is higher then kmalloc() |
2351 | * may return off node objects because partial slabs are obtained |
2352 | * from other nodes and filled up. |
2353 | * |
2354 | * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 |
2355 | * (which makes defrag_ratio = 1000) then every (well almost) |
2356 | * allocation will first attempt to defrag slab caches on other nodes. |
2357 | * This means scanning over all nodes to look for partial slabs which |
2358 | * may be expensive if we do it every time we are trying to find a slab |
2359 | * with available objects. |
2360 | */ |
2361 | if (!s->remote_node_defrag_ratio || |
2362 | get_cycles() % 1024 > s->remote_node_defrag_ratio) |
2363 | return NULL; |
2364 | |
2365 | do { |
2366 | cpuset_mems_cookie = read_mems_allowed_begin(); |
2367 | zonelist = node_zonelist(nid: mempolicy_slab_node(), flags: pc->flags); |
2368 | for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) { |
2369 | struct kmem_cache_node *n; |
2370 | |
2371 | n = get_node(s, node: zone_to_nid(zone)); |
2372 | |
2373 | if (n && cpuset_zone_allowed(z: zone, gfp_mask: pc->flags) && |
2374 | n->nr_partial > s->min_partial) { |
2375 | object = get_partial_node(s, n, pc); |
2376 | if (object) { |
2377 | /* |
2378 | * Don't check read_mems_allowed_retry() |
2379 | * here - if mems_allowed was updated in |
2380 | * parallel, that was a harmless race |
2381 | * between allocation and the cpuset |
2382 | * update |
2383 | */ |
2384 | return object; |
2385 | } |
2386 | } |
2387 | } |
2388 | } while (read_mems_allowed_retry(seq: cpuset_mems_cookie)); |
2389 | #endif /* CONFIG_NUMA */ |
2390 | return NULL; |
2391 | } |
2392 | |
2393 | /* |
2394 | * Get a partial slab, lock it and return it. |
2395 | */ |
2396 | static void *get_partial(struct kmem_cache *s, int node, struct partial_context *pc) |
2397 | { |
2398 | void *object; |
2399 | int searchnode = node; |
2400 | |
2401 | if (node == NUMA_NO_NODE) |
2402 | searchnode = numa_mem_id(); |
2403 | |
2404 | object = get_partial_node(s, n: get_node(s, node: searchnode), pc); |
2405 | if (object || node != NUMA_NO_NODE) |
2406 | return object; |
2407 | |
2408 | return get_any_partial(s, pc); |
2409 | } |
2410 | |
2411 | #ifndef CONFIG_SLUB_TINY |
2412 | |
2413 | #ifdef CONFIG_PREEMPTION |
2414 | /* |
2415 | * Calculate the next globally unique transaction for disambiguation |
2416 | * during cmpxchg. The transactions start with the cpu number and are then |
2417 | * incremented by CONFIG_NR_CPUS. |
2418 | */ |
2419 | #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) |
2420 | #else |
2421 | /* |
2422 | * No preemption supported therefore also no need to check for |
2423 | * different cpus. |
2424 | */ |
2425 | #define TID_STEP 1 |
2426 | #endif /* CONFIG_PREEMPTION */ |
2427 | |
2428 | static inline unsigned long next_tid(unsigned long tid) |
2429 | { |
2430 | return tid + TID_STEP; |
2431 | } |
2432 | |
2433 | #ifdef SLUB_DEBUG_CMPXCHG |
2434 | static inline unsigned int tid_to_cpu(unsigned long tid) |
2435 | { |
2436 | return tid % TID_STEP; |
2437 | } |
2438 | |
2439 | static inline unsigned long tid_to_event(unsigned long tid) |
2440 | { |
2441 | return tid / TID_STEP; |
2442 | } |
2443 | #endif |
2444 | |
2445 | static inline unsigned int init_tid(int cpu) |
2446 | { |
2447 | return cpu; |
2448 | } |
2449 | |
2450 | static inline void note_cmpxchg_failure(const char *n, |
2451 | const struct kmem_cache *s, unsigned long tid) |
2452 | { |
2453 | #ifdef SLUB_DEBUG_CMPXCHG |
2454 | unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); |
2455 | |
2456 | pr_info("%s %s: cmpxchg redo " , n, s->name); |
2457 | |
2458 | #ifdef CONFIG_PREEMPTION |
2459 | if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) |
2460 | pr_warn("due to cpu change %d -> %d\n" , |
2461 | tid_to_cpu(tid), tid_to_cpu(actual_tid)); |
2462 | else |
2463 | #endif |
2464 | if (tid_to_event(tid) != tid_to_event(actual_tid)) |
2465 | pr_warn("due to cpu running other code. Event %ld->%ld\n" , |
2466 | tid_to_event(tid), tid_to_event(actual_tid)); |
2467 | else |
2468 | pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n" , |
2469 | actual_tid, tid, next_tid(tid)); |
2470 | #endif |
2471 | stat(s, CMPXCHG_DOUBLE_CPU_FAIL); |
2472 | } |
2473 | |
2474 | static void init_kmem_cache_cpus(struct kmem_cache *s) |
2475 | { |
2476 | int cpu; |
2477 | struct kmem_cache_cpu *c; |
2478 | |
2479 | for_each_possible_cpu(cpu) { |
2480 | c = per_cpu_ptr(s->cpu_slab, cpu); |
2481 | local_lock_init(&c->lock); |
2482 | c->tid = init_tid(cpu); |
2483 | } |
2484 | } |
2485 | |
2486 | /* |
2487 | * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist, |
2488 | * unfreezes the slabs and puts it on the proper list. |
2489 | * Assumes the slab has been already safely taken away from kmem_cache_cpu |
2490 | * by the caller. |
2491 | */ |
2492 | static void deactivate_slab(struct kmem_cache *s, struct slab *slab, |
2493 | void *freelist) |
2494 | { |
2495 | enum slab_modes { M_NONE, M_PARTIAL, M_FREE, M_FULL_NOLIST }; |
2496 | struct kmem_cache_node *n = get_node(s, slab_nid(slab)); |
2497 | int free_delta = 0; |
2498 | enum slab_modes mode = M_NONE; |
2499 | void *nextfree, *freelist_iter, *freelist_tail; |
2500 | int tail = DEACTIVATE_TO_HEAD; |
2501 | unsigned long flags = 0; |
2502 | struct slab new; |
2503 | struct slab old; |
2504 | |
2505 | if (slab->freelist) { |
2506 | stat(s, DEACTIVATE_REMOTE_FREES); |
2507 | tail = DEACTIVATE_TO_TAIL; |
2508 | } |
2509 | |
2510 | /* |
2511 | * Stage one: Count the objects on cpu's freelist as free_delta and |
2512 | * remember the last object in freelist_tail for later splicing. |
2513 | */ |
2514 | freelist_tail = NULL; |
2515 | freelist_iter = freelist; |
2516 | while (freelist_iter) { |
2517 | nextfree = get_freepointer(s, freelist_iter); |
2518 | |
2519 | /* |
2520 | * If 'nextfree' is invalid, it is possible that the object at |
2521 | * 'freelist_iter' is already corrupted. So isolate all objects |
2522 | * starting at 'freelist_iter' by skipping them. |
2523 | */ |
2524 | if (freelist_corrupted(s, slab, &freelist_iter, nextfree)) |
2525 | break; |
2526 | |
2527 | freelist_tail = freelist_iter; |
2528 | free_delta++; |
2529 | |
2530 | freelist_iter = nextfree; |
2531 | } |
2532 | |
2533 | /* |
2534 | * Stage two: Unfreeze the slab while splicing the per-cpu |
2535 | * freelist to the head of slab's freelist. |
2536 | * |
2537 | * Ensure that the slab is unfrozen while the list presence |
2538 | * reflects the actual number of objects during unfreeze. |
2539 | * |
2540 | * We first perform cmpxchg holding lock and insert to list |
2541 | * when it succeed. If there is mismatch then the slab is not |
2542 | * unfrozen and number of objects in the slab may have changed. |
2543 | * Then release lock and retry cmpxchg again. |
2544 | */ |
2545 | redo: |
2546 | |
2547 | old.freelist = READ_ONCE(slab->freelist); |
2548 | old.counters = READ_ONCE(slab->counters); |
2549 | VM_BUG_ON(!old.frozen); |
2550 | |
2551 | /* Determine target state of the slab */ |
2552 | new.counters = old.counters; |
2553 | if (freelist_tail) { |
2554 | new.inuse -= free_delta; |
2555 | set_freepointer(s, freelist_tail, old.freelist); |
2556 | new.freelist = freelist; |
2557 | } else |
2558 | new.freelist = old.freelist; |
2559 | |
2560 | new.frozen = 0; |
2561 | |
2562 | if (!new.inuse && n->nr_partial >= s->min_partial) { |
2563 | mode = M_FREE; |
2564 | } else if (new.freelist) { |
2565 | mode = M_PARTIAL; |
2566 | /* |
2567 | * Taking the spinlock removes the possibility that |
2568 | * acquire_slab() will see a slab that is frozen |
2569 | */ |
2570 | spin_lock_irqsave(&n->list_lock, flags); |
2571 | } else { |
2572 | mode = M_FULL_NOLIST; |
2573 | } |
2574 | |
2575 | |
2576 | if (!slab_update_freelist(s, slab, |
2577 | old.freelist, old.counters, |
2578 | new.freelist, new.counters, |
2579 | "unfreezing slab" )) { |
2580 | if (mode == M_PARTIAL) |
2581 | spin_unlock_irqrestore(&n->list_lock, flags); |
2582 | goto redo; |
2583 | } |
2584 | |
2585 | |
2586 | if (mode == M_PARTIAL) { |
2587 | add_partial(n, slab, tail); |
2588 | spin_unlock_irqrestore(&n->list_lock, flags); |
2589 | stat(s, tail); |
2590 | } else if (mode == M_FREE) { |
2591 | stat(s, DEACTIVATE_EMPTY); |
2592 | discard_slab(s, slab); |
2593 | stat(s, FREE_SLAB); |
2594 | } else if (mode == M_FULL_NOLIST) { |
2595 | stat(s, DEACTIVATE_FULL); |
2596 | } |
2597 | } |
2598 | |
2599 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
2600 | static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab) |
2601 | { |
2602 | struct kmem_cache_node *n = NULL, *n2 = NULL; |
2603 | struct slab *slab, *slab_to_discard = NULL; |
2604 | unsigned long flags = 0; |
2605 | |
2606 | while (partial_slab) { |
2607 | struct slab new; |
2608 | struct slab old; |
2609 | |
2610 | slab = partial_slab; |
2611 | partial_slab = slab->next; |
2612 | |
2613 | n2 = get_node(s, slab_nid(slab)); |
2614 | if (n != n2) { |
2615 | if (n) |
2616 | spin_unlock_irqrestore(&n->list_lock, flags); |
2617 | |
2618 | n = n2; |
2619 | spin_lock_irqsave(&n->list_lock, flags); |
2620 | } |
2621 | |
2622 | do { |
2623 | |
2624 | old.freelist = slab->freelist; |
2625 | old.counters = slab->counters; |
2626 | VM_BUG_ON(!old.frozen); |
2627 | |
2628 | new.counters = old.counters; |
2629 | new.freelist = old.freelist; |
2630 | |
2631 | new.frozen = 0; |
2632 | |
2633 | } while (!__slab_update_freelist(s, slab, |
2634 | old.freelist, old.counters, |
2635 | new.freelist, new.counters, |
2636 | "unfreezing slab" )); |
2637 | |
2638 | if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) { |
2639 | slab->next = slab_to_discard; |
2640 | slab_to_discard = slab; |
2641 | } else { |
2642 | add_partial(n, slab, DEACTIVATE_TO_TAIL); |
2643 | stat(s, FREE_ADD_PARTIAL); |
2644 | } |
2645 | } |
2646 | |
2647 | if (n) |
2648 | spin_unlock_irqrestore(&n->list_lock, flags); |
2649 | |
2650 | while (slab_to_discard) { |
2651 | slab = slab_to_discard; |
2652 | slab_to_discard = slab_to_discard->next; |
2653 | |
2654 | stat(s, DEACTIVATE_EMPTY); |
2655 | discard_slab(s, slab); |
2656 | stat(s, FREE_SLAB); |
2657 | } |
2658 | } |
2659 | |
2660 | /* |
2661 | * Unfreeze all the cpu partial slabs. |
2662 | */ |
2663 | static void unfreeze_partials(struct kmem_cache *s) |
2664 | { |
2665 | struct slab *partial_slab; |
2666 | unsigned long flags; |
2667 | |
2668 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
2669 | partial_slab = this_cpu_read(s->cpu_slab->partial); |
2670 | this_cpu_write(s->cpu_slab->partial, NULL); |
2671 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
2672 | |
2673 | if (partial_slab) |
2674 | __unfreeze_partials(s, partial_slab); |
2675 | } |
2676 | |
2677 | static void unfreeze_partials_cpu(struct kmem_cache *s, |
2678 | struct kmem_cache_cpu *c) |
2679 | { |
2680 | struct slab *partial_slab; |
2681 | |
2682 | partial_slab = slub_percpu_partial(c); |
2683 | c->partial = NULL; |
2684 | |
2685 | if (partial_slab) |
2686 | __unfreeze_partials(s, partial_slab); |
2687 | } |
2688 | |
2689 | /* |
2690 | * Put a slab that was just frozen (in __slab_free|get_partial_node) into a |
2691 | * partial slab slot if available. |
2692 | * |
2693 | * If we did not find a slot then simply move all the partials to the |
2694 | * per node partial list. |
2695 | */ |
2696 | static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain) |
2697 | { |
2698 | struct slab *oldslab; |
2699 | struct slab *slab_to_unfreeze = NULL; |
2700 | unsigned long flags; |
2701 | int slabs = 0; |
2702 | |
2703 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
2704 | |
2705 | oldslab = this_cpu_read(s->cpu_slab->partial); |
2706 | |
2707 | if (oldslab) { |
2708 | if (drain && oldslab->slabs >= s->cpu_partial_slabs) { |
2709 | /* |
2710 | * Partial array is full. Move the existing set to the |
2711 | * per node partial list. Postpone the actual unfreezing |
2712 | * outside of the critical section. |
2713 | */ |
2714 | slab_to_unfreeze = oldslab; |
2715 | oldslab = NULL; |
2716 | } else { |
2717 | slabs = oldslab->slabs; |
2718 | } |
2719 | } |
2720 | |
2721 | slabs++; |
2722 | |
2723 | slab->slabs = slabs; |
2724 | slab->next = oldslab; |
2725 | |
2726 | this_cpu_write(s->cpu_slab->partial, slab); |
2727 | |
2728 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
2729 | |
2730 | if (slab_to_unfreeze) { |
2731 | __unfreeze_partials(s, slab_to_unfreeze); |
2732 | stat(s, CPU_PARTIAL_DRAIN); |
2733 | } |
2734 | } |
2735 | |
2736 | #else /* CONFIG_SLUB_CPU_PARTIAL */ |
2737 | |
2738 | static inline void unfreeze_partials(struct kmem_cache *s) { } |
2739 | static inline void unfreeze_partials_cpu(struct kmem_cache *s, |
2740 | struct kmem_cache_cpu *c) { } |
2741 | |
2742 | #endif /* CONFIG_SLUB_CPU_PARTIAL */ |
2743 | |
2744 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
2745 | { |
2746 | unsigned long flags; |
2747 | struct slab *slab; |
2748 | void *freelist; |
2749 | |
2750 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
2751 | |
2752 | slab = c->slab; |
2753 | freelist = c->freelist; |
2754 | |
2755 | c->slab = NULL; |
2756 | c->freelist = NULL; |
2757 | c->tid = next_tid(c->tid); |
2758 | |
2759 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
2760 | |
2761 | if (slab) { |
2762 | deactivate_slab(s, slab, freelist); |
2763 | stat(s, CPUSLAB_FLUSH); |
2764 | } |
2765 | } |
2766 | |
2767 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) |
2768 | { |
2769 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
2770 | void *freelist = c->freelist; |
2771 | struct slab *slab = c->slab; |
2772 | |
2773 | c->slab = NULL; |
2774 | c->freelist = NULL; |
2775 | c->tid = next_tid(c->tid); |
2776 | |
2777 | if (slab) { |
2778 | deactivate_slab(s, slab, freelist); |
2779 | stat(s, CPUSLAB_FLUSH); |
2780 | } |
2781 | |
2782 | unfreeze_partials_cpu(s, c); |
2783 | } |
2784 | |
2785 | struct slub_flush_work { |
2786 | struct work_struct work; |
2787 | struct kmem_cache *s; |
2788 | bool skip; |
2789 | }; |
2790 | |
2791 | /* |
2792 | * Flush cpu slab. |
2793 | * |
2794 | * Called from CPU work handler with migration disabled. |
2795 | */ |
2796 | static void flush_cpu_slab(struct work_struct *w) |
2797 | { |
2798 | struct kmem_cache *s; |
2799 | struct kmem_cache_cpu *c; |
2800 | struct slub_flush_work *sfw; |
2801 | |
2802 | sfw = container_of(w, struct slub_flush_work, work); |
2803 | |
2804 | s = sfw->s; |
2805 | c = this_cpu_ptr(s->cpu_slab); |
2806 | |
2807 | if (c->slab) |
2808 | flush_slab(s, c); |
2809 | |
2810 | unfreeze_partials(s); |
2811 | } |
2812 | |
2813 | static bool has_cpu_slab(int cpu, struct kmem_cache *s) |
2814 | { |
2815 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
2816 | |
2817 | return c->slab || slub_percpu_partial(c); |
2818 | } |
2819 | |
2820 | static DEFINE_MUTEX(flush_lock); |
2821 | static DEFINE_PER_CPU(struct slub_flush_work, slub_flush); |
2822 | |
2823 | static void flush_all_cpus_locked(struct kmem_cache *s) |
2824 | { |
2825 | struct slub_flush_work *sfw; |
2826 | unsigned int cpu; |
2827 | |
2828 | lockdep_assert_cpus_held(); |
2829 | mutex_lock(&flush_lock); |
2830 | |
2831 | for_each_online_cpu(cpu) { |
2832 | sfw = &per_cpu(slub_flush, cpu); |
2833 | if (!has_cpu_slab(cpu, s)) { |
2834 | sfw->skip = true; |
2835 | continue; |
2836 | } |
2837 | INIT_WORK(&sfw->work, flush_cpu_slab); |
2838 | sfw->skip = false; |
2839 | sfw->s = s; |
2840 | queue_work_on(cpu, flushwq, &sfw->work); |
2841 | } |
2842 | |
2843 | for_each_online_cpu(cpu) { |
2844 | sfw = &per_cpu(slub_flush, cpu); |
2845 | if (sfw->skip) |
2846 | continue; |
2847 | flush_work(&sfw->work); |
2848 | } |
2849 | |
2850 | mutex_unlock(&flush_lock); |
2851 | } |
2852 | |
2853 | static void flush_all(struct kmem_cache *s) |
2854 | { |
2855 | cpus_read_lock(); |
2856 | flush_all_cpus_locked(s); |
2857 | cpus_read_unlock(); |
2858 | } |
2859 | |
2860 | /* |
2861 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
2862 | * necessary. |
2863 | */ |
2864 | static int slub_cpu_dead(unsigned int cpu) |
2865 | { |
2866 | struct kmem_cache *s; |
2867 | |
2868 | mutex_lock(&slab_mutex); |
2869 | list_for_each_entry(s, &slab_caches, list) |
2870 | __flush_cpu_slab(s, cpu); |
2871 | mutex_unlock(&slab_mutex); |
2872 | return 0; |
2873 | } |
2874 | |
2875 | #else /* CONFIG_SLUB_TINY */ |
2876 | static inline void flush_all_cpus_locked(struct kmem_cache *s) { } |
2877 | static inline void flush_all(struct kmem_cache *s) { } |
2878 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { } |
2879 | static inline int slub_cpu_dead(unsigned int cpu) { return 0; } |
2880 | #endif /* CONFIG_SLUB_TINY */ |
2881 | |
2882 | /* |
2883 | * Check if the objects in a per cpu structure fit numa |
2884 | * locality expectations. |
2885 | */ |
2886 | static inline int node_match(struct slab *slab, int node) |
2887 | { |
2888 | #ifdef CONFIG_NUMA |
2889 | if (node != NUMA_NO_NODE && slab_nid(slab) != node) |
2890 | return 0; |
2891 | #endif |
2892 | return 1; |
2893 | } |
2894 | |
2895 | #ifdef CONFIG_SLUB_DEBUG |
2896 | static int count_free(struct slab *slab) |
2897 | { |
2898 | return slab->objects - slab->inuse; |
2899 | } |
2900 | |
2901 | static inline unsigned long node_nr_objs(struct kmem_cache_node *n) |
2902 | { |
2903 | return atomic_long_read(&n->total_objects); |
2904 | } |
2905 | |
2906 | /* Supports checking bulk free of a constructed freelist */ |
2907 | static inline bool free_debug_processing(struct kmem_cache *s, |
2908 | struct slab *slab, void *head, void *tail, int *bulk_cnt, |
2909 | unsigned long addr, depot_stack_handle_t handle) |
2910 | { |
2911 | bool checks_ok = false; |
2912 | void *object = head; |
2913 | int cnt = 0; |
2914 | |
2915 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
2916 | if (!check_slab(s, slab)) |
2917 | goto out; |
2918 | } |
2919 | |
2920 | if (slab->inuse < *bulk_cnt) { |
2921 | slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n" , |
2922 | slab->inuse, *bulk_cnt); |
2923 | goto out; |
2924 | } |
2925 | |
2926 | next_object: |
2927 | |
2928 | if (++cnt > *bulk_cnt) |
2929 | goto out_cnt; |
2930 | |
2931 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
2932 | if (!free_consistency_checks(s, slab, object, addr)) |
2933 | goto out; |
2934 | } |
2935 | |
2936 | if (s->flags & SLAB_STORE_USER) |
2937 | set_track_update(s, object, TRACK_FREE, addr, handle); |
2938 | trace(s, slab, object, 0); |
2939 | /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ |
2940 | init_object(s, object, SLUB_RED_INACTIVE); |
2941 | |
2942 | /* Reached end of constructed freelist yet? */ |
2943 | if (object != tail) { |
2944 | object = get_freepointer(s, object); |
2945 | goto next_object; |
2946 | } |
2947 | checks_ok = true; |
2948 | |
2949 | out_cnt: |
2950 | if (cnt != *bulk_cnt) { |
2951 | slab_err(s, slab, "Bulk free expected %d objects but found %d\n" , |
2952 | *bulk_cnt, cnt); |
2953 | *bulk_cnt = cnt; |
2954 | } |
2955 | |
2956 | out: |
2957 | |
2958 | if (!checks_ok) |
2959 | slab_fix(s, "Object at 0x%p not freed" , object); |
2960 | |
2961 | return checks_ok; |
2962 | } |
2963 | #endif /* CONFIG_SLUB_DEBUG */ |
2964 | |
2965 | #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS) |
2966 | static unsigned long count_partial(struct kmem_cache_node *n, |
2967 | int (*get_count)(struct slab *)) |
2968 | { |
2969 | unsigned long flags; |
2970 | unsigned long x = 0; |
2971 | struct slab *slab; |
2972 | |
2973 | spin_lock_irqsave(&n->list_lock, flags); |
2974 | list_for_each_entry(slab, &n->partial, slab_list) |
2975 | x += get_count(slab); |
2976 | spin_unlock_irqrestore(&n->list_lock, flags); |
2977 | return x; |
2978 | } |
2979 | #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */ |
2980 | |
2981 | #ifdef CONFIG_SLUB_DEBUG |
2982 | static noinline void |
2983 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) |
2984 | { |
2985 | static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, |
2986 | DEFAULT_RATELIMIT_BURST); |
2987 | int node; |
2988 | struct kmem_cache_node *n; |
2989 | |
2990 | if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) |
2991 | return; |
2992 | |
2993 | pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n" , |
2994 | nid, gfpflags, &gfpflags); |
2995 | pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n" , |
2996 | s->name, s->object_size, s->size, oo_order(s->oo), |
2997 | oo_order(s->min)); |
2998 | |
2999 | if (oo_order(s->min) > get_order(s->object_size)) |
3000 | pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n" , |
3001 | s->name); |
3002 | |
3003 | for_each_kmem_cache_node(s, node, n) { |
3004 | unsigned long nr_slabs; |
3005 | unsigned long nr_objs; |
3006 | unsigned long nr_free; |
3007 | |
3008 | nr_free = count_partial(n, count_free); |
3009 | nr_slabs = node_nr_slabs(n); |
3010 | nr_objs = node_nr_objs(n); |
3011 | |
3012 | pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n" , |
3013 | node, nr_slabs, nr_objs, nr_free); |
3014 | } |
3015 | } |
3016 | #else /* CONFIG_SLUB_DEBUG */ |
3017 | static inline void |
3018 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { } |
3019 | #endif |
3020 | |
3021 | static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags) |
3022 | { |
3023 | if (unlikely(slab_test_pfmemalloc(slab))) |
3024 | return gfp_pfmemalloc_allowed(gfp_mask: gfpflags); |
3025 | |
3026 | return true; |
3027 | } |
3028 | |
3029 | #ifndef CONFIG_SLUB_TINY |
3030 | static inline bool |
3031 | __update_cpu_freelist_fast(struct kmem_cache *s, |
3032 | void *freelist_old, void *freelist_new, |
3033 | unsigned long tid) |
3034 | { |
3035 | freelist_aba_t old = { .freelist = freelist_old, .counter = tid }; |
3036 | freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) }; |
3037 | |
3038 | return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full, |
3039 | &old.full, new.full); |
3040 | } |
3041 | |
3042 | /* |
3043 | * Check the slab->freelist and either transfer the freelist to the |
3044 | * per cpu freelist or deactivate the slab. |
3045 | * |
3046 | * The slab is still frozen if the return value is not NULL. |
3047 | * |
3048 | * If this function returns NULL then the slab has been unfrozen. |
3049 | */ |
3050 | static inline void *get_freelist(struct kmem_cache *s, struct slab *slab) |
3051 | { |
3052 | struct slab new; |
3053 | unsigned long counters; |
3054 | void *freelist; |
3055 | |
3056 | lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock)); |
3057 | |
3058 | do { |
3059 | freelist = slab->freelist; |
3060 | counters = slab->counters; |
3061 | |
3062 | new.counters = counters; |
3063 | VM_BUG_ON(!new.frozen); |
3064 | |
3065 | new.inuse = slab->objects; |
3066 | new.frozen = freelist != NULL; |
3067 | |
3068 | } while (!__slab_update_freelist(s, slab, |
3069 | freelist, counters, |
3070 | NULL, new.counters, |
3071 | "get_freelist" )); |
3072 | |
3073 | return freelist; |
3074 | } |
3075 | |
3076 | /* |
3077 | * Slow path. The lockless freelist is empty or we need to perform |
3078 | * debugging duties. |
3079 | * |
3080 | * Processing is still very fast if new objects have been freed to the |
3081 | * regular freelist. In that case we simply take over the regular freelist |
3082 | * as the lockless freelist and zap the regular freelist. |
3083 | * |
3084 | * If that is not working then we fall back to the partial lists. We take the |
3085 | * first element of the freelist as the object to allocate now and move the |
3086 | * rest of the freelist to the lockless freelist. |
3087 | * |
3088 | * And if we were unable to get a new slab from the partial slab lists then |
3089 | * we need to allocate a new slab. This is the slowest path since it involves |
3090 | * a call to the page allocator and the setup of a new slab. |
3091 | * |
3092 | * Version of __slab_alloc to use when we know that preemption is |
3093 | * already disabled (which is the case for bulk allocation). |
3094 | */ |
3095 | static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
3096 | unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size) |
3097 | { |
3098 | void *freelist; |
3099 | struct slab *slab; |
3100 | unsigned long flags; |
3101 | struct partial_context pc; |
3102 | |
3103 | stat(s, ALLOC_SLOWPATH); |
3104 | |
3105 | reread_slab: |
3106 | |
3107 | slab = READ_ONCE(c->slab); |
3108 | if (!slab) { |
3109 | /* |
3110 | * if the node is not online or has no normal memory, just |
3111 | * ignore the node constraint |
3112 | */ |
3113 | if (unlikely(node != NUMA_NO_NODE && |
3114 | !node_isset(node, slab_nodes))) |
3115 | node = NUMA_NO_NODE; |
3116 | goto new_slab; |
3117 | } |
3118 | redo: |
3119 | |
3120 | if (unlikely(!node_match(slab, node))) { |
3121 | /* |
3122 | * same as above but node_match() being false already |
3123 | * implies node != NUMA_NO_NODE |
3124 | */ |
3125 | if (!node_isset(node, slab_nodes)) { |
3126 | node = NUMA_NO_NODE; |
3127 | } else { |
3128 | stat(s, ALLOC_NODE_MISMATCH); |
3129 | goto deactivate_slab; |
3130 | } |
3131 | } |
3132 | |
3133 | /* |
3134 | * By rights, we should be searching for a slab page that was |
3135 | * PFMEMALLOC but right now, we are losing the pfmemalloc |
3136 | * information when the page leaves the per-cpu allocator |
3137 | */ |
3138 | if (unlikely(!pfmemalloc_match(slab, gfpflags))) |
3139 | goto deactivate_slab; |
3140 | |
3141 | /* must check again c->slab in case we got preempted and it changed */ |
3142 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3143 | if (unlikely(slab != c->slab)) { |
3144 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3145 | goto reread_slab; |
3146 | } |
3147 | freelist = c->freelist; |
3148 | if (freelist) |
3149 | goto load_freelist; |
3150 | |
3151 | freelist = get_freelist(s, slab); |
3152 | |
3153 | if (!freelist) { |
3154 | c->slab = NULL; |
3155 | c->tid = next_tid(c->tid); |
3156 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3157 | stat(s, DEACTIVATE_BYPASS); |
3158 | goto new_slab; |
3159 | } |
3160 | |
3161 | stat(s, ALLOC_REFILL); |
3162 | |
3163 | load_freelist: |
3164 | |
3165 | lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock)); |
3166 | |
3167 | /* |
3168 | * freelist is pointing to the list of objects to be used. |
3169 | * slab is pointing to the slab from which the objects are obtained. |
3170 | * That slab must be frozen for per cpu allocations to work. |
3171 | */ |
3172 | VM_BUG_ON(!c->slab->frozen); |
3173 | c->freelist = get_freepointer(s, freelist); |
3174 | c->tid = next_tid(c->tid); |
3175 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3176 | return freelist; |
3177 | |
3178 | deactivate_slab: |
3179 | |
3180 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3181 | if (slab != c->slab) { |
3182 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3183 | goto reread_slab; |
3184 | } |
3185 | freelist = c->freelist; |
3186 | c->slab = NULL; |
3187 | c->freelist = NULL; |
3188 | c->tid = next_tid(c->tid); |
3189 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3190 | deactivate_slab(s, slab, freelist); |
3191 | |
3192 | new_slab: |
3193 | |
3194 | if (slub_percpu_partial(c)) { |
3195 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3196 | if (unlikely(c->slab)) { |
3197 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3198 | goto reread_slab; |
3199 | } |
3200 | if (unlikely(!slub_percpu_partial(c))) { |
3201 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3202 | /* we were preempted and partial list got empty */ |
3203 | goto new_objects; |
3204 | } |
3205 | |
3206 | slab = c->slab = slub_percpu_partial(c); |
3207 | slub_set_percpu_partial(c, slab); |
3208 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3209 | stat(s, CPU_PARTIAL_ALLOC); |
3210 | goto redo; |
3211 | } |
3212 | |
3213 | new_objects: |
3214 | |
3215 | pc.flags = gfpflags; |
3216 | pc.slab = &slab; |
3217 | pc.orig_size = orig_size; |
3218 | freelist = get_partial(s, node, &pc); |
3219 | if (freelist) |
3220 | goto check_new_slab; |
3221 | |
3222 | slub_put_cpu_ptr(s->cpu_slab); |
3223 | slab = new_slab(s, gfpflags, node); |
3224 | c = slub_get_cpu_ptr(s->cpu_slab); |
3225 | |
3226 | if (unlikely(!slab)) { |
3227 | slab_out_of_memory(s, gfpflags, node); |
3228 | return NULL; |
3229 | } |
3230 | |
3231 | stat(s, ALLOC_SLAB); |
3232 | |
3233 | if (kmem_cache_debug(s)) { |
3234 | freelist = alloc_single_from_new_slab(s, slab, orig_size); |
3235 | |
3236 | if (unlikely(!freelist)) |
3237 | goto new_objects; |
3238 | |
3239 | if (s->flags & SLAB_STORE_USER) |
3240 | set_track(s, freelist, TRACK_ALLOC, addr); |
3241 | |
3242 | return freelist; |
3243 | } |
3244 | |
3245 | /* |
3246 | * No other reference to the slab yet so we can |
3247 | * muck around with it freely without cmpxchg |
3248 | */ |
3249 | freelist = slab->freelist; |
3250 | slab->freelist = NULL; |
3251 | slab->inuse = slab->objects; |
3252 | slab->frozen = 1; |
3253 | |
3254 | inc_slabs_node(s, slab_nid(slab), slab->objects); |
3255 | |
3256 | check_new_slab: |
3257 | |
3258 | if (kmem_cache_debug(s)) { |
3259 | /* |
3260 | * For debug caches here we had to go through |
3261 | * alloc_single_from_partial() so just store the tracking info |
3262 | * and return the object |
3263 | */ |
3264 | if (s->flags & SLAB_STORE_USER) |
3265 | set_track(s, freelist, TRACK_ALLOC, addr); |
3266 | |
3267 | return freelist; |
3268 | } |
3269 | |
3270 | if (unlikely(!pfmemalloc_match(slab, gfpflags))) { |
3271 | /* |
3272 | * For !pfmemalloc_match() case we don't load freelist so that |
3273 | * we don't make further mismatched allocations easier. |
3274 | */ |
3275 | deactivate_slab(s, slab, get_freepointer(s, freelist)); |
3276 | return freelist; |
3277 | } |
3278 | |
3279 | retry_load_slab: |
3280 | |
3281 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3282 | if (unlikely(c->slab)) { |
3283 | void *flush_freelist = c->freelist; |
3284 | struct slab *flush_slab = c->slab; |
3285 | |
3286 | c->slab = NULL; |
3287 | c->freelist = NULL; |
3288 | c->tid = next_tid(c->tid); |
3289 | |
3290 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3291 | |
3292 | deactivate_slab(s, flush_slab, flush_freelist); |
3293 | |
3294 | stat(s, CPUSLAB_FLUSH); |
3295 | |
3296 | goto retry_load_slab; |
3297 | } |
3298 | c->slab = slab; |
3299 | |
3300 | goto load_freelist; |
3301 | } |
3302 | |
3303 | /* |
3304 | * A wrapper for ___slab_alloc() for contexts where preemption is not yet |
3305 | * disabled. Compensates for possible cpu changes by refetching the per cpu area |
3306 | * pointer. |
3307 | */ |
3308 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
3309 | unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size) |
3310 | { |
3311 | void *p; |
3312 | |
3313 | #ifdef CONFIG_PREEMPT_COUNT |
3314 | /* |
3315 | * We may have been preempted and rescheduled on a different |
3316 | * cpu before disabling preemption. Need to reload cpu area |
3317 | * pointer. |
3318 | */ |
3319 | c = slub_get_cpu_ptr(s->cpu_slab); |
3320 | #endif |
3321 | |
3322 | p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size); |
3323 | #ifdef CONFIG_PREEMPT_COUNT |
3324 | slub_put_cpu_ptr(s->cpu_slab); |
3325 | #endif |
3326 | return p; |
3327 | } |
3328 | |
3329 | static __always_inline void *__slab_alloc_node(struct kmem_cache *s, |
3330 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
3331 | { |
3332 | struct kmem_cache_cpu *c; |
3333 | struct slab *slab; |
3334 | unsigned long tid; |
3335 | void *object; |
3336 | |
3337 | redo: |
3338 | /* |
3339 | * Must read kmem_cache cpu data via this cpu ptr. Preemption is |
3340 | * enabled. We may switch back and forth between cpus while |
3341 | * reading from one cpu area. That does not matter as long |
3342 | * as we end up on the original cpu again when doing the cmpxchg. |
3343 | * |
3344 | * We must guarantee that tid and kmem_cache_cpu are retrieved on the |
3345 | * same cpu. We read first the kmem_cache_cpu pointer and use it to read |
3346 | * the tid. If we are preempted and switched to another cpu between the |
3347 | * two reads, it's OK as the two are still associated with the same cpu |
3348 | * and cmpxchg later will validate the cpu. |
3349 | */ |
3350 | c = raw_cpu_ptr(s->cpu_slab); |
3351 | tid = READ_ONCE(c->tid); |
3352 | |
3353 | /* |
3354 | * Irqless object alloc/free algorithm used here depends on sequence |
3355 | * of fetching cpu_slab's data. tid should be fetched before anything |
3356 | * on c to guarantee that object and slab associated with previous tid |
3357 | * won't be used with current tid. If we fetch tid first, object and |
3358 | * slab could be one associated with next tid and our alloc/free |
3359 | * request will be failed. In this case, we will retry. So, no problem. |
3360 | */ |
3361 | barrier(); |
3362 | |
3363 | /* |
3364 | * The transaction ids are globally unique per cpu and per operation on |
3365 | * a per cpu queue. Thus they can be guarantee that the cmpxchg_double |
3366 | * occurs on the right processor and that there was no operation on the |
3367 | * linked list in between. |
3368 | */ |
3369 | |
3370 | object = c->freelist; |
3371 | slab = c->slab; |
3372 | |
3373 | if (!USE_LOCKLESS_FAST_PATH() || |
3374 | unlikely(!object || !slab || !node_match(slab, node))) { |
3375 | object = __slab_alloc(s, gfpflags, node, addr, c, orig_size); |
3376 | } else { |
3377 | void *next_object = get_freepointer_safe(s, object); |
3378 | |
3379 | /* |
3380 | * The cmpxchg will only match if there was no additional |
3381 | * operation and if we are on the right processor. |
3382 | * |
3383 | * The cmpxchg does the following atomically (without lock |
3384 | * semantics!) |
3385 | * 1. Relocate first pointer to the current per cpu area. |
3386 | * 2. Verify that tid and freelist have not been changed |
3387 | * 3. If they were not changed replace tid and freelist |
3388 | * |
3389 | * Since this is without lock semantics the protection is only |
3390 | * against code executing on this cpu *not* from access by |
3391 | * other cpus. |
3392 | */ |
3393 | if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) { |
3394 | note_cmpxchg_failure("slab_alloc" , s, tid); |
3395 | goto redo; |
3396 | } |
3397 | prefetch_freepointer(s, next_object); |
3398 | stat(s, ALLOC_FASTPATH); |
3399 | } |
3400 | |
3401 | return object; |
3402 | } |
3403 | #else /* CONFIG_SLUB_TINY */ |
3404 | static void *__slab_alloc_node(struct kmem_cache *s, |
3405 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
3406 | { |
3407 | struct partial_context pc; |
3408 | struct slab *slab; |
3409 | void *object; |
3410 | |
3411 | pc.flags = gfpflags; |
3412 | pc.slab = &slab; |
3413 | pc.orig_size = orig_size; |
3414 | object = get_partial(s, node, pc: &pc); |
3415 | |
3416 | if (object) |
3417 | return object; |
3418 | |
3419 | slab = new_slab(s, flags: gfpflags, node); |
3420 | if (unlikely(!slab)) { |
3421 | slab_out_of_memory(s, gfpflags, nid: node); |
3422 | return NULL; |
3423 | } |
3424 | |
3425 | object = alloc_single_from_new_slab(s, slab, orig_size); |
3426 | |
3427 | return object; |
3428 | } |
3429 | #endif /* CONFIG_SLUB_TINY */ |
3430 | |
3431 | /* |
3432 | * If the object has been wiped upon free, make sure it's fully initialized by |
3433 | * zeroing out freelist pointer. |
3434 | */ |
3435 | static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s, |
3436 | void *obj) |
3437 | { |
3438 | if (unlikely(slab_want_init_on_free(s)) && obj) |
3439 | memset((void *)((char *)kasan_reset_tag(obj) + s->offset), |
3440 | 0, sizeof(void *)); |
3441 | } |
3442 | |
3443 | /* |
3444 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) |
3445 | * have the fastpath folded into their functions. So no function call |
3446 | * overhead for requests that can be satisfied on the fastpath. |
3447 | * |
3448 | * The fastpath works by first checking if the lockless freelist can be used. |
3449 | * If not then __slab_alloc is called for slow processing. |
3450 | * |
3451 | * Otherwise we can simply pick the next object from the lockless free list. |
3452 | */ |
3453 | static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru, |
3454 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
3455 | { |
3456 | void *object; |
3457 | struct obj_cgroup *objcg = NULL; |
3458 | bool init = false; |
3459 | |
3460 | s = slab_pre_alloc_hook(s, lru, objcgp: &objcg, size: 1, flags: gfpflags); |
3461 | if (!s) |
3462 | return NULL; |
3463 | |
3464 | object = kfence_alloc(s, size: orig_size, flags: gfpflags); |
3465 | if (unlikely(object)) |
3466 | goto out; |
3467 | |
3468 | object = __slab_alloc_node(s, gfpflags, node, addr, orig_size); |
3469 | |
3470 | maybe_wipe_obj_freeptr(s, obj: object); |
3471 | init = slab_want_init_on_alloc(flags: gfpflags, c: s); |
3472 | |
3473 | out: |
3474 | /* |
3475 | * When init equals 'true', like for kzalloc() family, only |
3476 | * @orig_size bytes might be zeroed instead of s->object_size |
3477 | */ |
3478 | slab_post_alloc_hook(s, objcg, flags: gfpflags, size: 1, p: &object, init, orig_size); |
3479 | |
3480 | return object; |
3481 | } |
3482 | |
3483 | static __fastpath_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru, |
3484 | gfp_t gfpflags, unsigned long addr, size_t orig_size) |
3485 | { |
3486 | return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size); |
3487 | } |
3488 | |
3489 | static __fastpath_inline |
3490 | void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru, |
3491 | gfp_t gfpflags) |
3492 | { |
3493 | void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, orig_size: s->object_size); |
3494 | |
3495 | trace_kmem_cache_alloc(_RET_IP_, ptr: ret, s, gfp_flags: gfpflags, NUMA_NO_NODE); |
3496 | |
3497 | return ret; |
3498 | } |
3499 | |
3500 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) |
3501 | { |
3502 | return __kmem_cache_alloc_lru(s, NULL, gfpflags); |
3503 | } |
3504 | EXPORT_SYMBOL(kmem_cache_alloc); |
3505 | |
3506 | void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru, |
3507 | gfp_t gfpflags) |
3508 | { |
3509 | return __kmem_cache_alloc_lru(s, lru, gfpflags); |
3510 | } |
3511 | EXPORT_SYMBOL(kmem_cache_alloc_lru); |
3512 | |
3513 | void *__kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, |
3514 | int node, size_t orig_size, |
3515 | unsigned long caller) |
3516 | { |
3517 | return slab_alloc_node(s, NULL, gfpflags, node, |
3518 | addr: caller, orig_size); |
3519 | } |
3520 | |
3521 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) |
3522 | { |
3523 | void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, orig_size: s->object_size); |
3524 | |
3525 | trace_kmem_cache_alloc(_RET_IP_, ptr: ret, s, gfp_flags: gfpflags, node); |
3526 | |
3527 | return ret; |
3528 | } |
3529 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3530 | |
3531 | static noinline void free_to_partial_list( |
3532 | struct kmem_cache *s, struct slab *slab, |
3533 | void *head, void *tail, int bulk_cnt, |
3534 | unsigned long addr) |
3535 | { |
3536 | struct kmem_cache_node *n = get_node(s, node: slab_nid(slab)); |
3537 | struct slab *slab_free = NULL; |
3538 | int cnt = bulk_cnt; |
3539 | unsigned long flags; |
3540 | depot_stack_handle_t handle = 0; |
3541 | |
3542 | if (s->flags & SLAB_STORE_USER) |
3543 | handle = set_track_prepare(); |
3544 | |
3545 | spin_lock_irqsave(&n->list_lock, flags); |
3546 | |
3547 | if (free_debug_processing(s, slab, head, tail, bulk_cnt: &cnt, addr, handle)) { |
3548 | void *prior = slab->freelist; |
3549 | |
3550 | /* Perform the actual freeing while we still hold the locks */ |
3551 | slab->inuse -= cnt; |
3552 | set_freepointer(s, object: tail, fp: prior); |
3553 | slab->freelist = head; |
3554 | |
3555 | /* |
3556 | * If the slab is empty, and node's partial list is full, |
3557 | * it should be discarded anyway no matter it's on full or |
3558 | * partial list. |
3559 | */ |
3560 | if (slab->inuse == 0 && n->nr_partial >= s->min_partial) |
3561 | slab_free = slab; |
3562 | |
3563 | if (!prior) { |
3564 | /* was on full list */ |
3565 | remove_full(s, n, slab); |
3566 | if (!slab_free) { |
3567 | add_partial(n, slab, tail: DEACTIVATE_TO_TAIL); |
3568 | stat(s, si: FREE_ADD_PARTIAL); |
3569 | } |
3570 | } else if (slab_free) { |
3571 | remove_partial(n, slab); |
3572 | stat(s, si: FREE_REMOVE_PARTIAL); |
3573 | } |
3574 | } |
3575 | |
3576 | if (slab_free) { |
3577 | /* |
3578 | * Update the counters while still holding n->list_lock to |
3579 | * prevent spurious validation warnings |
3580 | */ |
3581 | dec_slabs_node(s, node: slab_nid(slab: slab_free), objects: slab_free->objects); |
3582 | } |
3583 | |
3584 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
3585 | |
3586 | if (slab_free) { |
3587 | stat(s, si: FREE_SLAB); |
3588 | free_slab(s, slab: slab_free); |
3589 | } |
3590 | } |
3591 | |
3592 | /* |
3593 | * Slow path handling. This may still be called frequently since objects |
3594 | * have a longer lifetime than the cpu slabs in most processing loads. |
3595 | * |
3596 | * So we still attempt to reduce cache line usage. Just take the slab |
3597 | * lock and free the item. If there is no additional partial slab |
3598 | * handling required then we can return immediately. |
3599 | */ |
3600 | static void __slab_free(struct kmem_cache *s, struct slab *slab, |
3601 | void *head, void *tail, int cnt, |
3602 | unsigned long addr) |
3603 | |
3604 | { |
3605 | void *prior; |
3606 | int was_frozen; |
3607 | struct slab new; |
3608 | unsigned long counters; |
3609 | struct kmem_cache_node *n = NULL; |
3610 | unsigned long flags; |
3611 | |
3612 | stat(s, si: FREE_SLOWPATH); |
3613 | |
3614 | if (kfence_free(addr: head)) |
3615 | return; |
3616 | |
3617 | if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) { |
3618 | free_to_partial_list(s, slab, head, tail, bulk_cnt: cnt, addr); |
3619 | return; |
3620 | } |
3621 | |
3622 | do { |
3623 | if (unlikely(n)) { |
3624 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
3625 | n = NULL; |
3626 | } |
3627 | prior = slab->freelist; |
3628 | counters = slab->counters; |
3629 | set_freepointer(s, object: tail, fp: prior); |
3630 | new.counters = counters; |
3631 | was_frozen = new.frozen; |
3632 | new.inuse -= cnt; |
3633 | if ((!new.inuse || !prior) && !was_frozen) { |
3634 | |
3635 | if (kmem_cache_has_cpu_partial(s) && !prior) { |
3636 | |
3637 | /* |
3638 | * Slab was on no list before and will be |
3639 | * partially empty |
3640 | * We can defer the list move and instead |
3641 | * freeze it. |
3642 | */ |
3643 | new.frozen = 1; |
3644 | |
3645 | } else { /* Needs to be taken off a list */ |
3646 | |
3647 | n = get_node(s, node: slab_nid(slab)); |
3648 | /* |
3649 | * Speculatively acquire the list_lock. |
3650 | * If the cmpxchg does not succeed then we may |
3651 | * drop the list_lock without any processing. |
3652 | * |
3653 | * Otherwise the list_lock will synchronize with |
3654 | * other processors updating the list of slabs. |
3655 | */ |
3656 | spin_lock_irqsave(&n->list_lock, flags); |
3657 | |
3658 | } |
3659 | } |
3660 | |
3661 | } while (!slab_update_freelist(s, slab, |
3662 | freelist_old: prior, counters_old: counters, |
3663 | freelist_new: head, counters_new: new.counters, |
3664 | n: "__slab_free" )); |
3665 | |
3666 | if (likely(!n)) { |
3667 | |
3668 | if (likely(was_frozen)) { |
3669 | /* |
3670 | * The list lock was not taken therefore no list |
3671 | * activity can be necessary. |
3672 | */ |
3673 | stat(s, si: FREE_FROZEN); |
3674 | } else if (new.frozen) { |
3675 | /* |
3676 | * If we just froze the slab then put it onto the |
3677 | * per cpu partial list. |
3678 | */ |
3679 | put_cpu_partial(s, slab, drain: 1); |
3680 | stat(s, si: CPU_PARTIAL_FREE); |
3681 | } |
3682 | |
3683 | return; |
3684 | } |
3685 | |
3686 | if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) |
3687 | goto slab_empty; |
3688 | |
3689 | /* |
3690 | * Objects left in the slab. If it was not on the partial list before |
3691 | * then add it. |
3692 | */ |
3693 | if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { |
3694 | remove_full(s, n, slab); |
3695 | add_partial(n, slab, tail: DEACTIVATE_TO_TAIL); |
3696 | stat(s, si: FREE_ADD_PARTIAL); |
3697 | } |
3698 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
3699 | return; |
3700 | |
3701 | slab_empty: |
3702 | if (prior) { |
3703 | /* |
3704 | * Slab on the partial list. |
3705 | */ |
3706 | remove_partial(n, slab); |
3707 | stat(s, si: FREE_REMOVE_PARTIAL); |
3708 | } else { |
3709 | /* Slab must be on the full list */ |
3710 | remove_full(s, n, slab); |
3711 | } |
3712 | |
3713 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
3714 | stat(s, si: FREE_SLAB); |
3715 | discard_slab(s, slab); |
3716 | } |
3717 | |
3718 | #ifndef CONFIG_SLUB_TINY |
3719 | /* |
3720 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that |
3721 | * can perform fastpath freeing without additional function calls. |
3722 | * |
3723 | * The fastpath is only possible if we are freeing to the current cpu slab |
3724 | * of this processor. This typically the case if we have just allocated |
3725 | * the item before. |
3726 | * |
3727 | * If fastpath is not possible then fall back to __slab_free where we deal |
3728 | * with all sorts of special processing. |
3729 | * |
3730 | * Bulk free of a freelist with several objects (all pointing to the |
3731 | * same slab) possible by specifying head and tail ptr, plus objects |
3732 | * count (cnt). Bulk free indicated by tail pointer being set. |
3733 | */ |
3734 | static __always_inline void do_slab_free(struct kmem_cache *s, |
3735 | struct slab *slab, void *head, void *tail, |
3736 | int cnt, unsigned long addr) |
3737 | { |
3738 | void *tail_obj = tail ? : head; |
3739 | struct kmem_cache_cpu *c; |
3740 | unsigned long tid; |
3741 | void **freelist; |
3742 | |
3743 | redo: |
3744 | /* |
3745 | * Determine the currently cpus per cpu slab. |
3746 | * The cpu may change afterward. However that does not matter since |
3747 | * data is retrieved via this pointer. If we are on the same cpu |
3748 | * during the cmpxchg then the free will succeed. |
3749 | */ |
3750 | c = raw_cpu_ptr(s->cpu_slab); |
3751 | tid = READ_ONCE(c->tid); |
3752 | |
3753 | /* Same with comment on barrier() in slab_alloc_node() */ |
3754 | barrier(); |
3755 | |
3756 | if (unlikely(slab != c->slab)) { |
3757 | __slab_free(s, slab, head, tail_obj, cnt, addr); |
3758 | return; |
3759 | } |
3760 | |
3761 | if (USE_LOCKLESS_FAST_PATH()) { |
3762 | freelist = READ_ONCE(c->freelist); |
3763 | |
3764 | set_freepointer(s, tail_obj, freelist); |
3765 | |
3766 | if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) { |
3767 | note_cmpxchg_failure("slab_free" , s, tid); |
3768 | goto redo; |
3769 | } |
3770 | } else { |
3771 | /* Update the free list under the local lock */ |
3772 | local_lock(&s->cpu_slab->lock); |
3773 | c = this_cpu_ptr(s->cpu_slab); |
3774 | if (unlikely(slab != c->slab)) { |
3775 | local_unlock(&s->cpu_slab->lock); |
3776 | goto redo; |
3777 | } |
3778 | tid = c->tid; |
3779 | freelist = c->freelist; |
3780 | |
3781 | set_freepointer(s, tail_obj, freelist); |
3782 | c->freelist = head; |
3783 | c->tid = next_tid(tid); |
3784 | |
3785 | local_unlock(&s->cpu_slab->lock); |
3786 | } |
3787 | stat(s, FREE_FASTPATH); |
3788 | } |
3789 | #else /* CONFIG_SLUB_TINY */ |
3790 | static void do_slab_free(struct kmem_cache *s, |
3791 | struct slab *slab, void *head, void *tail, |
3792 | int cnt, unsigned long addr) |
3793 | { |
3794 | void *tail_obj = tail ? : head; |
3795 | |
3796 | __slab_free(s, slab, head, tail: tail_obj, cnt, addr); |
3797 | } |
3798 | #endif /* CONFIG_SLUB_TINY */ |
3799 | |
3800 | static __fastpath_inline void slab_free(struct kmem_cache *s, struct slab *slab, |
3801 | void *head, void *tail, void **p, int cnt, |
3802 | unsigned long addr) |
3803 | { |
3804 | memcg_slab_free_hook(s, slab, p, objects: cnt); |
3805 | /* |
3806 | * With KASAN enabled slab_free_freelist_hook modifies the freelist |
3807 | * to remove objects, whose reuse must be delayed. |
3808 | */ |
3809 | if (slab_free_freelist_hook(s, head: &head, tail: &tail, cnt: &cnt)) |
3810 | do_slab_free(s, slab, head, tail, cnt, addr); |
3811 | } |
3812 | |
3813 | #ifdef CONFIG_KASAN_GENERIC |
3814 | void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) |
3815 | { |
3816 | do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr); |
3817 | } |
3818 | #endif |
3819 | |
3820 | void __kmem_cache_free(struct kmem_cache *s, void *x, unsigned long caller) |
3821 | { |
3822 | slab_free(s, slab: virt_to_slab(addr: x), head: x, NULL, p: &x, cnt: 1, addr: caller); |
3823 | } |
3824 | |
3825 | void kmem_cache_free(struct kmem_cache *s, void *x) |
3826 | { |
3827 | s = cache_from_obj(s, x); |
3828 | if (!s) |
3829 | return; |
3830 | trace_kmem_cache_free(_RET_IP_, ptr: x, s); |
3831 | slab_free(s, slab: virt_to_slab(addr: x), head: x, NULL, p: &x, cnt: 1, _RET_IP_); |
3832 | } |
3833 | EXPORT_SYMBOL(kmem_cache_free); |
3834 | |
3835 | struct detached_freelist { |
3836 | struct slab *slab; |
3837 | void *tail; |
3838 | void *freelist; |
3839 | int cnt; |
3840 | struct kmem_cache *s; |
3841 | }; |
3842 | |
3843 | /* |
3844 | * This function progressively scans the array with free objects (with |
3845 | * a limited look ahead) and extract objects belonging to the same |
3846 | * slab. It builds a detached freelist directly within the given |
3847 | * slab/objects. This can happen without any need for |
3848 | * synchronization, because the objects are owned by running process. |
3849 | * The freelist is build up as a single linked list in the objects. |
3850 | * The idea is, that this detached freelist can then be bulk |
3851 | * transferred to the real freelist(s), but only requiring a single |
3852 | * synchronization primitive. Look ahead in the array is limited due |
3853 | * to performance reasons. |
3854 | */ |
3855 | static inline |
3856 | int build_detached_freelist(struct kmem_cache *s, size_t size, |
3857 | void **p, struct detached_freelist *df) |
3858 | { |
3859 | int lookahead = 3; |
3860 | void *object; |
3861 | struct folio *folio; |
3862 | size_t same; |
3863 | |
3864 | object = p[--size]; |
3865 | folio = virt_to_folio(x: object); |
3866 | if (!s) { |
3867 | /* Handle kalloc'ed objects */ |
3868 | if (unlikely(!folio_test_slab(folio))) { |
3869 | free_large_kmalloc(folio, object); |
3870 | df->slab = NULL; |
3871 | return size; |
3872 | } |
3873 | /* Derive kmem_cache from object */ |
3874 | df->slab = folio_slab(folio); |
3875 | df->s = df->slab->slab_cache; |
3876 | } else { |
3877 | df->slab = folio_slab(folio); |
3878 | df->s = cache_from_obj(s, x: object); /* Support for memcg */ |
3879 | } |
3880 | |
3881 | /* Start new detached freelist */ |
3882 | df->tail = object; |
3883 | df->freelist = object; |
3884 | df->cnt = 1; |
3885 | |
3886 | if (is_kfence_address(addr: object)) |
3887 | return size; |
3888 | |
3889 | set_freepointer(s: df->s, object, NULL); |
3890 | |
3891 | same = size; |
3892 | while (size) { |
3893 | object = p[--size]; |
3894 | /* df->slab is always set at this point */ |
3895 | if (df->slab == virt_to_slab(addr: object)) { |
3896 | /* Opportunity build freelist */ |
3897 | set_freepointer(s: df->s, object, fp: df->freelist); |
3898 | df->freelist = object; |
3899 | df->cnt++; |
3900 | same--; |
3901 | if (size != same) |
3902 | swap(p[size], p[same]); |
3903 | continue; |
3904 | } |
3905 | |
3906 | /* Limit look ahead search */ |
3907 | if (!--lookahead) |
3908 | break; |
3909 | } |
3910 | |
3911 | return same; |
3912 | } |
3913 | |
3914 | /* Note that interrupts must be enabled when calling this function. */ |
3915 | void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) |
3916 | { |
3917 | if (!size) |
3918 | return; |
3919 | |
3920 | do { |
3921 | struct detached_freelist df; |
3922 | |
3923 | size = build_detached_freelist(s, size, p, df: &df); |
3924 | if (!df.slab) |
3925 | continue; |
3926 | |
3927 | slab_free(s: df.s, slab: df.slab, head: df.freelist, tail: df.tail, p: &p[size], cnt: df.cnt, |
3928 | _RET_IP_); |
3929 | } while (likely(size)); |
3930 | } |
3931 | EXPORT_SYMBOL(kmem_cache_free_bulk); |
3932 | |
3933 | #ifndef CONFIG_SLUB_TINY |
3934 | static inline int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, |
3935 | size_t size, void **p, struct obj_cgroup *objcg) |
3936 | { |
3937 | struct kmem_cache_cpu *c; |
3938 | unsigned long irqflags; |
3939 | int i; |
3940 | |
3941 | /* |
3942 | * Drain objects in the per cpu slab, while disabling local |
3943 | * IRQs, which protects against PREEMPT and interrupts |
3944 | * handlers invoking normal fastpath. |
3945 | */ |
3946 | c = slub_get_cpu_ptr(s->cpu_slab); |
3947 | local_lock_irqsave(&s->cpu_slab->lock, irqflags); |
3948 | |
3949 | for (i = 0; i < size; i++) { |
3950 | void *object = kfence_alloc(s, s->object_size, flags); |
3951 | |
3952 | if (unlikely(object)) { |
3953 | p[i] = object; |
3954 | continue; |
3955 | } |
3956 | |
3957 | object = c->freelist; |
3958 | if (unlikely(!object)) { |
3959 | /* |
3960 | * We may have removed an object from c->freelist using |
3961 | * the fastpath in the previous iteration; in that case, |
3962 | * c->tid has not been bumped yet. |
3963 | * Since ___slab_alloc() may reenable interrupts while |
3964 | * allocating memory, we should bump c->tid now. |
3965 | */ |
3966 | c->tid = next_tid(c->tid); |
3967 | |
3968 | local_unlock_irqrestore(&s->cpu_slab->lock, irqflags); |
3969 | |
3970 | /* |
3971 | * Invoking slow path likely have side-effect |
3972 | * of re-populating per CPU c->freelist |
3973 | */ |
3974 | p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, |
3975 | _RET_IP_, c, s->object_size); |
3976 | if (unlikely(!p[i])) |
3977 | goto error; |
3978 | |
3979 | c = this_cpu_ptr(s->cpu_slab); |
3980 | maybe_wipe_obj_freeptr(s, p[i]); |
3981 | |
3982 | local_lock_irqsave(&s->cpu_slab->lock, irqflags); |
3983 | |
3984 | continue; /* goto for-loop */ |
3985 | } |
3986 | c->freelist = get_freepointer(s, object); |
3987 | p[i] = object; |
3988 | maybe_wipe_obj_freeptr(s, p[i]); |
3989 | } |
3990 | c->tid = next_tid(c->tid); |
3991 | local_unlock_irqrestore(&s->cpu_slab->lock, irqflags); |
3992 | slub_put_cpu_ptr(s->cpu_slab); |
3993 | |
3994 | return i; |
3995 | |
3996 | error: |
3997 | slub_put_cpu_ptr(s->cpu_slab); |
3998 | slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size); |
3999 | kmem_cache_free_bulk(s, i, p); |
4000 | return 0; |
4001 | |
4002 | } |
4003 | #else /* CONFIG_SLUB_TINY */ |
4004 | static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, |
4005 | size_t size, void **p, struct obj_cgroup *objcg) |
4006 | { |
4007 | int i; |
4008 | |
4009 | for (i = 0; i < size; i++) { |
4010 | void *object = kfence_alloc(s, size: s->object_size, flags); |
4011 | |
4012 | if (unlikely(object)) { |
4013 | p[i] = object; |
4014 | continue; |
4015 | } |
4016 | |
4017 | p[i] = __slab_alloc_node(s, gfpflags: flags, NUMA_NO_NODE, |
4018 | _RET_IP_, orig_size: s->object_size); |
4019 | if (unlikely(!p[i])) |
4020 | goto error; |
4021 | |
4022 | maybe_wipe_obj_freeptr(s, obj: p[i]); |
4023 | } |
4024 | |
4025 | return i; |
4026 | |
4027 | error: |
4028 | slab_post_alloc_hook(s, objcg, flags, size: i, p, init: false, orig_size: s->object_size); |
4029 | kmem_cache_free_bulk(s, i, p); |
4030 | return 0; |
4031 | } |
4032 | #endif /* CONFIG_SLUB_TINY */ |
4033 | |
4034 | /* Note that interrupts must be enabled when calling this function. */ |
4035 | int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, |
4036 | void **p) |
4037 | { |
4038 | int i; |
4039 | struct obj_cgroup *objcg = NULL; |
4040 | |
4041 | if (!size) |
4042 | return 0; |
4043 | |
4044 | /* memcg and kmem_cache debug support */ |
4045 | s = slab_pre_alloc_hook(s, NULL, objcgp: &objcg, size, flags); |
4046 | if (unlikely(!s)) |
4047 | return 0; |
4048 | |
4049 | i = __kmem_cache_alloc_bulk(s, flags, size, p, objcg); |
4050 | |
4051 | /* |
4052 | * memcg and kmem_cache debug support and memory initialization. |
4053 | * Done outside of the IRQ disabled fastpath loop. |
4054 | */ |
4055 | if (i != 0) |
4056 | slab_post_alloc_hook(s, objcg, flags, size, p, |
4057 | init: slab_want_init_on_alloc(flags, c: s), orig_size: s->object_size); |
4058 | return i; |
4059 | } |
4060 | EXPORT_SYMBOL(kmem_cache_alloc_bulk); |
4061 | |
4062 | |
4063 | /* |
4064 | * Object placement in a slab is made very easy because we always start at |
4065 | * offset 0. If we tune the size of the object to the alignment then we can |
4066 | * get the required alignment by putting one properly sized object after |
4067 | * another. |
4068 | * |
4069 | * Notice that the allocation order determines the sizes of the per cpu |
4070 | * caches. Each processor has always one slab available for allocations. |
4071 | * Increasing the allocation order reduces the number of times that slabs |
4072 | * must be moved on and off the partial lists and is therefore a factor in |
4073 | * locking overhead. |
4074 | */ |
4075 | |
4076 | /* |
4077 | * Minimum / Maximum order of slab pages. This influences locking overhead |
4078 | * and slab fragmentation. A higher order reduces the number of partial slabs |
4079 | * and increases the number of allocations possible without having to |
4080 | * take the list_lock. |
4081 | */ |
4082 | static unsigned int slub_min_order; |
4083 | static unsigned int slub_max_order = |
4084 | IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER; |
4085 | static unsigned int slub_min_objects; |
4086 | |
4087 | /* |
4088 | * Calculate the order of allocation given an slab object size. |
4089 | * |
4090 | * The order of allocation has significant impact on performance and other |
4091 | * system components. Generally order 0 allocations should be preferred since |
4092 | * order 0 does not cause fragmentation in the page allocator. Larger objects |
4093 | * be problematic to put into order 0 slabs because there may be too much |
4094 | * unused space left. We go to a higher order if more than 1/16th of the slab |
4095 | * would be wasted. |
4096 | * |
4097 | * In order to reach satisfactory performance we must ensure that a minimum |
4098 | * number of objects is in one slab. Otherwise we may generate too much |
4099 | * activity on the partial lists which requires taking the list_lock. This is |
4100 | * less a concern for large slabs though which are rarely used. |
4101 | * |
4102 | * slub_max_order specifies the order where we begin to stop considering the |
4103 | * number of objects in a slab as critical. If we reach slub_max_order then |
4104 | * we try to keep the page order as low as possible. So we accept more waste |
4105 | * of space in favor of a small page order. |
4106 | * |
4107 | * Higher order allocations also allow the placement of more objects in a |
4108 | * slab and thereby reduce object handling overhead. If the user has |
4109 | * requested a higher minimum order then we start with that one instead of |
4110 | * the smallest order which will fit the object. |
4111 | */ |
4112 | static inline unsigned int calc_slab_order(unsigned int size, |
4113 | unsigned int min_order, unsigned int max_order, |
4114 | unsigned int fract_leftover) |
4115 | { |
4116 | unsigned int order; |
4117 | |
4118 | for (order = min_order; order <= max_order; order++) { |
4119 | |
4120 | unsigned int slab_size = (unsigned int)PAGE_SIZE << order; |
4121 | unsigned int rem; |
4122 | |
4123 | rem = slab_size % size; |
4124 | |
4125 | if (rem <= slab_size / fract_leftover) |
4126 | break; |
4127 | } |
4128 | |
4129 | return order; |
4130 | } |
4131 | |
4132 | static inline int calculate_order(unsigned int size) |
4133 | { |
4134 | unsigned int order; |
4135 | unsigned int min_objects; |
4136 | unsigned int max_objects; |
4137 | unsigned int min_order; |
4138 | |
4139 | min_objects = slub_min_objects; |
4140 | if (!min_objects) { |
4141 | /* |
4142 | * Some architectures will only update present cpus when |
4143 | * onlining them, so don't trust the number if it's just 1. But |
4144 | * we also don't want to use nr_cpu_ids always, as on some other |
4145 | * architectures, there can be many possible cpus, but never |
4146 | * onlined. Here we compromise between trying to avoid too high |
4147 | * order on systems that appear larger than they are, and too |
4148 | * low order on systems that appear smaller than they are. |
4149 | */ |
4150 | unsigned int nr_cpus = num_present_cpus(); |
4151 | if (nr_cpus <= 1) |
4152 | nr_cpus = nr_cpu_ids; |
4153 | min_objects = 4 * (fls(x: nr_cpus) + 1); |
4154 | } |
4155 | /* min_objects can't be 0 because get_order(0) is undefined */ |
4156 | max_objects = max(order_objects(slub_max_order, size), 1U); |
4157 | min_objects = min(min_objects, max_objects); |
4158 | |
4159 | min_order = max_t(unsigned int, slub_min_order, |
4160 | get_order(min_objects * size)); |
4161 | if (order_objects(order: min_order, size) > MAX_OBJS_PER_PAGE) |
4162 | return get_order(size: size * MAX_OBJS_PER_PAGE) - 1; |
4163 | |
4164 | /* |
4165 | * Attempt to find best configuration for a slab. This works by first |
4166 | * attempting to generate a layout with the best possible configuration |
4167 | * and backing off gradually. |
4168 | * |
4169 | * We start with accepting at most 1/16 waste and try to find the |
4170 | * smallest order from min_objects-derived/slub_min_order up to |
4171 | * slub_max_order that will satisfy the constraint. Note that increasing |
4172 | * the order can only result in same or less fractional waste, not more. |
4173 | * |
4174 | * If that fails, we increase the acceptable fraction of waste and try |
4175 | * again. The last iteration with fraction of 1/2 would effectively |
4176 | * accept any waste and give us the order determined by min_objects, as |
4177 | * long as at least single object fits within slub_max_order. |
4178 | */ |
4179 | for (unsigned int fraction = 16; fraction > 1; fraction /= 2) { |
4180 | order = calc_slab_order(size, min_order, max_order: slub_max_order, |
4181 | fract_leftover: fraction); |
4182 | if (order <= slub_max_order) |
4183 | return order; |
4184 | } |
4185 | |
4186 | /* |
4187 | * Doh this slab cannot be placed using slub_max_order. |
4188 | */ |
4189 | order = get_order(size); |
4190 | if (order <= MAX_ORDER) |
4191 | return order; |
4192 | return -ENOSYS; |
4193 | } |
4194 | |
4195 | static void |
4196 | init_kmem_cache_node(struct kmem_cache_node *n) |
4197 | { |
4198 | n->nr_partial = 0; |
4199 | spin_lock_init(&n->list_lock); |
4200 | INIT_LIST_HEAD(list: &n->partial); |
4201 | #ifdef CONFIG_SLUB_DEBUG |
4202 | atomic_long_set(&n->nr_slabs, 0); |
4203 | atomic_long_set(&n->total_objects, 0); |
4204 | INIT_LIST_HEAD(&n->full); |
4205 | #endif |
4206 | } |
4207 | |
4208 | #ifndef CONFIG_SLUB_TINY |
4209 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
4210 | { |
4211 | BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < |
4212 | NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH * |
4213 | sizeof(struct kmem_cache_cpu)); |
4214 | |
4215 | /* |
4216 | * Must align to double word boundary for the double cmpxchg |
4217 | * instructions to work; see __pcpu_double_call_return_bool(). |
4218 | */ |
4219 | s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), |
4220 | 2 * sizeof(void *)); |
4221 | |
4222 | if (!s->cpu_slab) |
4223 | return 0; |
4224 | |
4225 | init_kmem_cache_cpus(s); |
4226 | |
4227 | return 1; |
4228 | } |
4229 | #else |
4230 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
4231 | { |
4232 | return 1; |
4233 | } |
4234 | #endif /* CONFIG_SLUB_TINY */ |
4235 | |
4236 | static struct kmem_cache *kmem_cache_node; |
4237 | |
4238 | /* |
4239 | * No kmalloc_node yet so do it by hand. We know that this is the first |
4240 | * slab on the node for this slabcache. There are no concurrent accesses |
4241 | * possible. |
4242 | * |
4243 | * Note that this function only works on the kmem_cache_node |
4244 | * when allocating for the kmem_cache_node. This is used for bootstrapping |
4245 | * memory on a fresh node that has no slab structures yet. |
4246 | */ |
4247 | static void early_kmem_cache_node_alloc(int node) |
4248 | { |
4249 | struct slab *slab; |
4250 | struct kmem_cache_node *n; |
4251 | |
4252 | BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); |
4253 | |
4254 | slab = new_slab(s: kmem_cache_node, GFP_NOWAIT, node); |
4255 | |
4256 | BUG_ON(!slab); |
4257 | inc_slabs_node(s: kmem_cache_node, node: slab_nid(slab), objects: slab->objects); |
4258 | if (slab_nid(slab) != node) { |
4259 | pr_err("SLUB: Unable to allocate memory from node %d\n" , node); |
4260 | pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n" ); |
4261 | } |
4262 | |
4263 | n = slab->freelist; |
4264 | BUG_ON(!n); |
4265 | #ifdef CONFIG_SLUB_DEBUG |
4266 | init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); |
4267 | init_tracking(kmem_cache_node, n); |
4268 | #endif |
4269 | n = kasan_slab_alloc(s: kmem_cache_node, object: n, GFP_KERNEL, init: false); |
4270 | slab->freelist = get_freepointer(s: kmem_cache_node, object: n); |
4271 | slab->inuse = 1; |
4272 | kmem_cache_node->node[node] = n; |
4273 | init_kmem_cache_node(n); |
4274 | inc_slabs_node(s: kmem_cache_node, node, objects: slab->objects); |
4275 | |
4276 | /* |
4277 | * No locks need to be taken here as it has just been |
4278 | * initialized and there is no concurrent access. |
4279 | */ |
4280 | __add_partial(n, slab, tail: DEACTIVATE_TO_HEAD); |
4281 | } |
4282 | |
4283 | static void free_kmem_cache_nodes(struct kmem_cache *s) |
4284 | { |
4285 | int node; |
4286 | struct kmem_cache_node *n; |
4287 | |
4288 | for_each_kmem_cache_node(s, node, n) { |
4289 | s->node[node] = NULL; |
4290 | kmem_cache_free(kmem_cache_node, n); |
4291 | } |
4292 | } |
4293 | |
4294 | void __kmem_cache_release(struct kmem_cache *s) |
4295 | { |
4296 | cache_random_seq_destroy(cachep: s); |
4297 | #ifndef CONFIG_SLUB_TINY |
4298 | free_percpu(s->cpu_slab); |
4299 | #endif |
4300 | free_kmem_cache_nodes(s); |
4301 | } |
4302 | |
4303 | static int init_kmem_cache_nodes(struct kmem_cache *s) |
4304 | { |
4305 | int node; |
4306 | |
4307 | for_each_node_mask(node, slab_nodes) { |
4308 | struct kmem_cache_node *n; |
4309 | |
4310 | if (slab_state == DOWN) { |
4311 | early_kmem_cache_node_alloc(node); |
4312 | continue; |
4313 | } |
4314 | n = kmem_cache_alloc_node(kmem_cache_node, |
4315 | GFP_KERNEL, node); |
4316 | |
4317 | if (!n) { |
4318 | free_kmem_cache_nodes(s); |
4319 | return 0; |
4320 | } |
4321 | |
4322 | init_kmem_cache_node(n); |
4323 | s->node[node] = n; |
4324 | } |
4325 | return 1; |
4326 | } |
4327 | |
4328 | static void set_cpu_partial(struct kmem_cache *s) |
4329 | { |
4330 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
4331 | unsigned int nr_objects; |
4332 | |
4333 | /* |
4334 | * cpu_partial determined the maximum number of objects kept in the |
4335 | * per cpu partial lists of a processor. |
4336 | * |
4337 | * Per cpu partial lists mainly contain slabs that just have one |
4338 | * object freed. If they are used for allocation then they can be |
4339 | * filled up again with minimal effort. The slab will never hit the |
4340 | * per node partial lists and therefore no locking will be required. |
4341 | * |
4342 | * For backwards compatibility reasons, this is determined as number |
4343 | * of objects, even though we now limit maximum number of pages, see |
4344 | * slub_set_cpu_partial() |
4345 | */ |
4346 | if (!kmem_cache_has_cpu_partial(s)) |
4347 | nr_objects = 0; |
4348 | else if (s->size >= PAGE_SIZE) |
4349 | nr_objects = 6; |
4350 | else if (s->size >= 1024) |
4351 | nr_objects = 24; |
4352 | else if (s->size >= 256) |
4353 | nr_objects = 52; |
4354 | else |
4355 | nr_objects = 120; |
4356 | |
4357 | slub_set_cpu_partial(s, nr_objects); |
4358 | #endif |
4359 | } |
4360 | |
4361 | /* |
4362 | * calculate_sizes() determines the order and the distribution of data within |
4363 | * a slab object. |
4364 | */ |
4365 | static int calculate_sizes(struct kmem_cache *s) |
4366 | { |
4367 | slab_flags_t flags = s->flags; |
4368 | unsigned int size = s->object_size; |
4369 | unsigned int order; |
4370 | |
4371 | /* |
4372 | * Round up object size to the next word boundary. We can only |
4373 | * place the free pointer at word boundaries and this determines |
4374 | * the possible location of the free pointer. |
4375 | */ |
4376 | size = ALIGN(size, sizeof(void *)); |
4377 | |
4378 | #ifdef CONFIG_SLUB_DEBUG |
4379 | /* |
4380 | * Determine if we can poison the object itself. If the user of |
4381 | * the slab may touch the object after free or before allocation |
4382 | * then we should never poison the object itself. |
4383 | */ |
4384 | if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) && |
4385 | !s->ctor) |
4386 | s->flags |= __OBJECT_POISON; |
4387 | else |
4388 | s->flags &= ~__OBJECT_POISON; |
4389 | |
4390 | |
4391 | /* |
4392 | * If we are Redzoning then check if there is some space between the |
4393 | * end of the object and the free pointer. If not then add an |
4394 | * additional word to have some bytes to store Redzone information. |
4395 | */ |
4396 | if ((flags & SLAB_RED_ZONE) && size == s->object_size) |
4397 | size += sizeof(void *); |
4398 | #endif |
4399 | |
4400 | /* |
4401 | * With that we have determined the number of bytes in actual use |
4402 | * by the object and redzoning. |
4403 | */ |
4404 | s->inuse = size; |
4405 | |
4406 | if (slub_debug_orig_size(s) || |
4407 | (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) || |
4408 | ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) || |
4409 | s->ctor) { |
4410 | /* |
4411 | * Relocate free pointer after the object if it is not |
4412 | * permitted to overwrite the first word of the object on |
4413 | * kmem_cache_free. |
4414 | * |
4415 | * This is the case if we do RCU, have a constructor or |
4416 | * destructor, are poisoning the objects, or are |
4417 | * redzoning an object smaller than sizeof(void *). |
4418 | * |
4419 | * The assumption that s->offset >= s->inuse means free |
4420 | * pointer is outside of the object is used in the |
4421 | * freeptr_outside_object() function. If that is no |
4422 | * longer true, the function needs to be modified. |
4423 | */ |
4424 | s->offset = size; |
4425 | size += sizeof(void *); |
4426 | } else { |
4427 | /* |
4428 | * Store freelist pointer near middle of object to keep |
4429 | * it away from the edges of the object to avoid small |
4430 | * sized over/underflows from neighboring allocations. |
4431 | */ |
4432 | s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *)); |
4433 | } |
4434 | |
4435 | #ifdef CONFIG_SLUB_DEBUG |
4436 | if (flags & SLAB_STORE_USER) { |
4437 | /* |
4438 | * Need to store information about allocs and frees after |
4439 | * the object. |
4440 | */ |
4441 | size += 2 * sizeof(struct track); |
4442 | |
4443 | /* Save the original kmalloc request size */ |
4444 | if (flags & SLAB_KMALLOC) |
4445 | size += sizeof(unsigned int); |
4446 | } |
4447 | #endif |
4448 | |
4449 | kasan_cache_create(cache: s, size: &size, flags: &s->flags); |
4450 | #ifdef CONFIG_SLUB_DEBUG |
4451 | if (flags & SLAB_RED_ZONE) { |
4452 | /* |
4453 | * Add some empty padding so that we can catch |
4454 | * overwrites from earlier objects rather than let |
4455 | * tracking information or the free pointer be |
4456 | * corrupted if a user writes before the start |
4457 | * of the object. |
4458 | */ |
4459 | size += sizeof(void *); |
4460 | |
4461 | s->red_left_pad = sizeof(void *); |
4462 | s->red_left_pad = ALIGN(s->red_left_pad, s->align); |
4463 | size += s->red_left_pad; |
4464 | } |
4465 | #endif |
4466 | |
4467 | /* |
4468 | * SLUB stores one object immediately after another beginning from |
4469 | * offset 0. In order to align the objects we have to simply size |
4470 | * each object to conform to the alignment. |
4471 | */ |
4472 | size = ALIGN(size, s->align); |
4473 | s->size = size; |
4474 | s->reciprocal_size = reciprocal_value(d: size); |
4475 | order = calculate_order(size); |
4476 | |
4477 | if ((int)order < 0) |
4478 | return 0; |
4479 | |
4480 | s->allocflags = 0; |
4481 | if (order) |
4482 | s->allocflags |= __GFP_COMP; |
4483 | |
4484 | if (s->flags & SLAB_CACHE_DMA) |
4485 | s->allocflags |= GFP_DMA; |
4486 | |
4487 | if (s->flags & SLAB_CACHE_DMA32) |
4488 | s->allocflags |= GFP_DMA32; |
4489 | |
4490 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
4491 | s->allocflags |= __GFP_RECLAIMABLE; |
4492 | |
4493 | /* |
4494 | * Determine the number of objects per slab |
4495 | */ |
4496 | s->oo = oo_make(order, size); |
4497 | s->min = oo_make(order: get_order(size), size); |
4498 | |
4499 | return !!oo_objects(x: s->oo); |
4500 | } |
4501 | |
4502 | static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags) |
4503 | { |
4504 | s->flags = kmem_cache_flags(object_size: s->size, flags, name: s->name); |
4505 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
4506 | s->random = get_random_long(); |
4507 | #endif |
4508 | |
4509 | if (!calculate_sizes(s)) |
4510 | goto error; |
4511 | if (disable_higher_order_debug) { |
4512 | /* |
4513 | * Disable debugging flags that store metadata if the min slab |
4514 | * order increased. |
4515 | */ |
4516 | if (get_order(size: s->size) > get_order(size: s->object_size)) { |
4517 | s->flags &= ~DEBUG_METADATA_FLAGS; |
4518 | s->offset = 0; |
4519 | if (!calculate_sizes(s)) |
4520 | goto error; |
4521 | } |
4522 | } |
4523 | |
4524 | #ifdef system_has_freelist_aba |
4525 | if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) { |
4526 | /* Enable fast mode */ |
4527 | s->flags |= __CMPXCHG_DOUBLE; |
4528 | } |
4529 | #endif |
4530 | |
4531 | /* |
4532 | * The larger the object size is, the more slabs we want on the partial |
4533 | * list to avoid pounding the page allocator excessively. |
4534 | */ |
4535 | s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2); |
4536 | s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial); |
4537 | |
4538 | set_cpu_partial(s); |
4539 | |
4540 | #ifdef CONFIG_NUMA |
4541 | s->remote_node_defrag_ratio = 1000; |
4542 | #endif |
4543 | |
4544 | /* Initialize the pre-computed randomized freelist if slab is up */ |
4545 | if (slab_state >= UP) { |
4546 | if (init_cache_random_seq(s)) |
4547 | goto error; |
4548 | } |
4549 | |
4550 | if (!init_kmem_cache_nodes(s)) |
4551 | goto error; |
4552 | |
4553 | if (alloc_kmem_cache_cpus(s)) |
4554 | return 0; |
4555 | |
4556 | error: |
4557 | __kmem_cache_release(s); |
4558 | return -EINVAL; |
4559 | } |
4560 | |
4561 | static void list_slab_objects(struct kmem_cache *s, struct slab *slab, |
4562 | const char *text) |
4563 | { |
4564 | #ifdef CONFIG_SLUB_DEBUG |
4565 | void *addr = slab_address(slab); |
4566 | void *p; |
4567 | |
4568 | slab_err(s, slab, text, s->name); |
4569 | |
4570 | spin_lock(&object_map_lock); |
4571 | __fill_map(object_map, s, slab); |
4572 | |
4573 | for_each_object(p, s, addr, slab->objects) { |
4574 | |
4575 | if (!test_bit(__obj_to_index(s, addr, p), object_map)) { |
4576 | pr_err("Object 0x%p @offset=%tu\n" , p, p - addr); |
4577 | print_tracking(s, p); |
4578 | } |
4579 | } |
4580 | spin_unlock(&object_map_lock); |
4581 | #endif |
4582 | } |
4583 | |
4584 | /* |
4585 | * Attempt to free all partial slabs on a node. |
4586 | * This is called from __kmem_cache_shutdown(). We must take list_lock |
4587 | * because sysfs file might still access partial list after the shutdowning. |
4588 | */ |
4589 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) |
4590 | { |
4591 | LIST_HEAD(discard); |
4592 | struct slab *slab, *h; |
4593 | |
4594 | BUG_ON(irqs_disabled()); |
4595 | spin_lock_irq(lock: &n->list_lock); |
4596 | list_for_each_entry_safe(slab, h, &n->partial, slab_list) { |
4597 | if (!slab->inuse) { |
4598 | remove_partial(n, slab); |
4599 | list_add(new: &slab->slab_list, head: &discard); |
4600 | } else { |
4601 | list_slab_objects(s, slab, |
4602 | text: "Objects remaining in %s on __kmem_cache_shutdown()" ); |
4603 | } |
4604 | } |
4605 | spin_unlock_irq(lock: &n->list_lock); |
4606 | |
4607 | list_for_each_entry_safe(slab, h, &discard, slab_list) |
4608 | discard_slab(s, slab); |
4609 | } |
4610 | |
4611 | bool __kmem_cache_empty(struct kmem_cache *s) |
4612 | { |
4613 | int node; |
4614 | struct kmem_cache_node *n; |
4615 | |
4616 | for_each_kmem_cache_node(s, node, n) |
4617 | if (n->nr_partial || node_nr_slabs(n)) |
4618 | return false; |
4619 | return true; |
4620 | } |
4621 | |
4622 | /* |
4623 | * Release all resources used by a slab cache. |
4624 | */ |
4625 | int __kmem_cache_shutdown(struct kmem_cache *s) |
4626 | { |
4627 | int node; |
4628 | struct kmem_cache_node *n; |
4629 | |
4630 | flush_all_cpus_locked(s); |
4631 | /* Attempt to free all objects */ |
4632 | for_each_kmem_cache_node(s, node, n) { |
4633 | free_partial(s, n); |
4634 | if (n->nr_partial || node_nr_slabs(n)) |
4635 | return 1; |
4636 | } |
4637 | return 0; |
4638 | } |
4639 | |
4640 | #ifdef CONFIG_PRINTK |
4641 | void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab) |
4642 | { |
4643 | void *base; |
4644 | int __maybe_unused i; |
4645 | unsigned int objnr; |
4646 | void *objp; |
4647 | void *objp0; |
4648 | struct kmem_cache *s = slab->slab_cache; |
4649 | struct track __maybe_unused *trackp; |
4650 | |
4651 | kpp->kp_ptr = object; |
4652 | kpp->kp_slab = slab; |
4653 | kpp->kp_slab_cache = s; |
4654 | base = slab_address(slab); |
4655 | objp0 = kasan_reset_tag(addr: object); |
4656 | #ifdef CONFIG_SLUB_DEBUG |
4657 | objp = restore_red_left(s, objp0); |
4658 | #else |
4659 | objp = objp0; |
4660 | #endif |
4661 | objnr = obj_to_index(cache: s, slab, obj: objp); |
4662 | kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp); |
4663 | objp = base + s->size * objnr; |
4664 | kpp->kp_objp = objp; |
4665 | if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size |
4666 | || (objp - base) % s->size) || |
4667 | !(s->flags & SLAB_STORE_USER)) |
4668 | return; |
4669 | #ifdef CONFIG_SLUB_DEBUG |
4670 | objp = fixup_red_left(s, objp); |
4671 | trackp = get_track(s, objp, TRACK_ALLOC); |
4672 | kpp->kp_ret = (void *)trackp->addr; |
4673 | #ifdef CONFIG_STACKDEPOT |
4674 | { |
4675 | depot_stack_handle_t handle; |
4676 | unsigned long *entries; |
4677 | unsigned int nr_entries; |
4678 | |
4679 | handle = READ_ONCE(trackp->handle); |
4680 | if (handle) { |
4681 | nr_entries = stack_depot_fetch(handle, &entries); |
4682 | for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++) |
4683 | kpp->kp_stack[i] = (void *)entries[i]; |
4684 | } |
4685 | |
4686 | trackp = get_track(s, objp, TRACK_FREE); |
4687 | handle = READ_ONCE(trackp->handle); |
4688 | if (handle) { |
4689 | nr_entries = stack_depot_fetch(handle, &entries); |
4690 | for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++) |
4691 | kpp->kp_free_stack[i] = (void *)entries[i]; |
4692 | } |
4693 | } |
4694 | #endif |
4695 | #endif |
4696 | } |
4697 | #endif |
4698 | |
4699 | /******************************************************************** |
4700 | * Kmalloc subsystem |
4701 | *******************************************************************/ |
4702 | |
4703 | static int __init setup_slub_min_order(char *str) |
4704 | { |
4705 | get_option(str: &str, pint: (int *)&slub_min_order); |
4706 | |
4707 | if (slub_min_order > slub_max_order) |
4708 | slub_max_order = slub_min_order; |
4709 | |
4710 | return 1; |
4711 | } |
4712 | |
4713 | __setup("slub_min_order=" , setup_slub_min_order); |
4714 | |
4715 | static int __init setup_slub_max_order(char *str) |
4716 | { |
4717 | get_option(str: &str, pint: (int *)&slub_max_order); |
4718 | slub_max_order = min_t(unsigned int, slub_max_order, MAX_ORDER); |
4719 | |
4720 | if (slub_min_order > slub_max_order) |
4721 | slub_min_order = slub_max_order; |
4722 | |
4723 | return 1; |
4724 | } |
4725 | |
4726 | __setup("slub_max_order=" , setup_slub_max_order); |
4727 | |
4728 | static int __init setup_slub_min_objects(char *str) |
4729 | { |
4730 | get_option(str: &str, pint: (int *)&slub_min_objects); |
4731 | |
4732 | return 1; |
4733 | } |
4734 | |
4735 | __setup("slub_min_objects=" , setup_slub_min_objects); |
4736 | |
4737 | #ifdef CONFIG_HARDENED_USERCOPY |
4738 | /* |
4739 | * Rejects incorrectly sized objects and objects that are to be copied |
4740 | * to/from userspace but do not fall entirely within the containing slab |
4741 | * cache's usercopy region. |
4742 | * |
4743 | * Returns NULL if check passes, otherwise const char * to name of cache |
4744 | * to indicate an error. |
4745 | */ |
4746 | void __check_heap_object(const void *ptr, unsigned long n, |
4747 | const struct slab *slab, bool to_user) |
4748 | { |
4749 | struct kmem_cache *s; |
4750 | unsigned int offset; |
4751 | bool is_kfence = is_kfence_address(addr: ptr); |
4752 | |
4753 | ptr = kasan_reset_tag(addr: ptr); |
4754 | |
4755 | /* Find object and usable object size. */ |
4756 | s = slab->slab_cache; |
4757 | |
4758 | /* Reject impossible pointers. */ |
4759 | if (ptr < slab_address(slab)) |
4760 | usercopy_abort(name: "SLUB object not in SLUB page?!" , NULL, |
4761 | to_user, offset: 0, len: n); |
4762 | |
4763 | /* Find offset within object. */ |
4764 | if (is_kfence) |
4765 | offset = ptr - kfence_object_start(addr: ptr); |
4766 | else |
4767 | offset = (ptr - slab_address(slab)) % s->size; |
4768 | |
4769 | /* Adjust for redzone and reject if within the redzone. */ |
4770 | if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) { |
4771 | if (offset < s->red_left_pad) |
4772 | usercopy_abort(name: "SLUB object in left red zone" , |
4773 | detail: s->name, to_user, offset, len: n); |
4774 | offset -= s->red_left_pad; |
4775 | } |
4776 | |
4777 | /* Allow address range falling entirely within usercopy region. */ |
4778 | if (offset >= s->useroffset && |
4779 | offset - s->useroffset <= s->usersize && |
4780 | n <= s->useroffset - offset + s->usersize) |
4781 | return; |
4782 | |
4783 | usercopy_abort(name: "SLUB object" , detail: s->name, to_user, offset, len: n); |
4784 | } |
4785 | #endif /* CONFIG_HARDENED_USERCOPY */ |
4786 | |
4787 | #define SHRINK_PROMOTE_MAX 32 |
4788 | |
4789 | /* |
4790 | * kmem_cache_shrink discards empty slabs and promotes the slabs filled |
4791 | * up most to the head of the partial lists. New allocations will then |
4792 | * fill those up and thus they can be removed from the partial lists. |
4793 | * |
4794 | * The slabs with the least items are placed last. This results in them |
4795 | * being allocated from last increasing the chance that the last objects |
4796 | * are freed in them. |
4797 | */ |
4798 | static int __kmem_cache_do_shrink(struct kmem_cache *s) |
4799 | { |
4800 | int node; |
4801 | int i; |
4802 | struct kmem_cache_node *n; |
4803 | struct slab *slab; |
4804 | struct slab *t; |
4805 | struct list_head discard; |
4806 | struct list_head promote[SHRINK_PROMOTE_MAX]; |
4807 | unsigned long flags; |
4808 | int ret = 0; |
4809 | |
4810 | for_each_kmem_cache_node(s, node, n) { |
4811 | INIT_LIST_HEAD(list: &discard); |
4812 | for (i = 0; i < SHRINK_PROMOTE_MAX; i++) |
4813 | INIT_LIST_HEAD(list: promote + i); |
4814 | |
4815 | spin_lock_irqsave(&n->list_lock, flags); |
4816 | |
4817 | /* |
4818 | * Build lists of slabs to discard or promote. |
4819 | * |
4820 | * Note that concurrent frees may occur while we hold the |
4821 | * list_lock. slab->inuse here is the upper limit. |
4822 | */ |
4823 | list_for_each_entry_safe(slab, t, &n->partial, slab_list) { |
4824 | int free = slab->objects - slab->inuse; |
4825 | |
4826 | /* Do not reread slab->inuse */ |
4827 | barrier(); |
4828 | |
4829 | /* We do not keep full slabs on the list */ |
4830 | BUG_ON(free <= 0); |
4831 | |
4832 | if (free == slab->objects) { |
4833 | list_move(list: &slab->slab_list, head: &discard); |
4834 | n->nr_partial--; |
4835 | dec_slabs_node(s, node, objects: slab->objects); |
4836 | } else if (free <= SHRINK_PROMOTE_MAX) |
4837 | list_move(list: &slab->slab_list, head: promote + free - 1); |
4838 | } |
4839 | |
4840 | /* |
4841 | * Promote the slabs filled up most to the head of the |
4842 | * partial list. |
4843 | */ |
4844 | for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) |
4845 | list_splice(list: promote + i, head: &n->partial); |
4846 | |
4847 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
4848 | |
4849 | /* Release empty slabs */ |
4850 | list_for_each_entry_safe(slab, t, &discard, slab_list) |
4851 | free_slab(s, slab); |
4852 | |
4853 | if (node_nr_slabs(n)) |
4854 | ret = 1; |
4855 | } |
4856 | |
4857 | return ret; |
4858 | } |
4859 | |
4860 | int __kmem_cache_shrink(struct kmem_cache *s) |
4861 | { |
4862 | flush_all(s); |
4863 | return __kmem_cache_do_shrink(s); |
4864 | } |
4865 | |
4866 | static int slab_mem_going_offline_callback(void *arg) |
4867 | { |
4868 | struct kmem_cache *s; |
4869 | |
4870 | mutex_lock(&slab_mutex); |
4871 | list_for_each_entry(s, &slab_caches, list) { |
4872 | flush_all_cpus_locked(s); |
4873 | __kmem_cache_do_shrink(s); |
4874 | } |
4875 | mutex_unlock(lock: &slab_mutex); |
4876 | |
4877 | return 0; |
4878 | } |
4879 | |
4880 | static void slab_mem_offline_callback(void *arg) |
4881 | { |
4882 | struct memory_notify *marg = arg; |
4883 | int offline_node; |
4884 | |
4885 | offline_node = marg->status_change_nid_normal; |
4886 | |
4887 | /* |
4888 | * If the node still has available memory. we need kmem_cache_node |
4889 | * for it yet. |
4890 | */ |
4891 | if (offline_node < 0) |
4892 | return; |
4893 | |
4894 | mutex_lock(&slab_mutex); |
4895 | node_clear(offline_node, slab_nodes); |
4896 | /* |
4897 | * We no longer free kmem_cache_node structures here, as it would be |
4898 | * racy with all get_node() users, and infeasible to protect them with |
4899 | * slab_mutex. |
4900 | */ |
4901 | mutex_unlock(lock: &slab_mutex); |
4902 | } |
4903 | |
4904 | static int slab_mem_going_online_callback(void *arg) |
4905 | { |
4906 | struct kmem_cache_node *n; |
4907 | struct kmem_cache *s; |
4908 | struct memory_notify *marg = arg; |
4909 | int nid = marg->status_change_nid_normal; |
4910 | int ret = 0; |
4911 | |
4912 | /* |
4913 | * If the node's memory is already available, then kmem_cache_node is |
4914 | * already created. Nothing to do. |
4915 | */ |
4916 | if (nid < 0) |
4917 | return 0; |
4918 | |
4919 | /* |
4920 | * We are bringing a node online. No memory is available yet. We must |
4921 | * allocate a kmem_cache_node structure in order to bring the node |
4922 | * online. |
4923 | */ |
4924 | mutex_lock(&slab_mutex); |
4925 | list_for_each_entry(s, &slab_caches, list) { |
4926 | /* |
4927 | * The structure may already exist if the node was previously |
4928 | * onlined and offlined. |
4929 | */ |
4930 | if (get_node(s, node: nid)) |
4931 | continue; |
4932 | /* |
4933 | * XXX: kmem_cache_alloc_node will fallback to other nodes |
4934 | * since memory is not yet available from the node that |
4935 | * is brought up. |
4936 | */ |
4937 | n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); |
4938 | if (!n) { |
4939 | ret = -ENOMEM; |
4940 | goto out; |
4941 | } |
4942 | init_kmem_cache_node(n); |
4943 | s->node[nid] = n; |
4944 | } |
4945 | /* |
4946 | * Any cache created after this point will also have kmem_cache_node |
4947 | * initialized for the new node. |
4948 | */ |
4949 | node_set(nid, slab_nodes); |
4950 | out: |
4951 | mutex_unlock(lock: &slab_mutex); |
4952 | return ret; |
4953 | } |
4954 | |
4955 | static int slab_memory_callback(struct notifier_block *self, |
4956 | unsigned long action, void *arg) |
4957 | { |
4958 | int ret = 0; |
4959 | |
4960 | switch (action) { |
4961 | case MEM_GOING_ONLINE: |
4962 | ret = slab_mem_going_online_callback(arg); |
4963 | break; |
4964 | case MEM_GOING_OFFLINE: |
4965 | ret = slab_mem_going_offline_callback(arg); |
4966 | break; |
4967 | case MEM_OFFLINE: |
4968 | case MEM_CANCEL_ONLINE: |
4969 | slab_mem_offline_callback(arg); |
4970 | break; |
4971 | case MEM_ONLINE: |
4972 | case MEM_CANCEL_OFFLINE: |
4973 | break; |
4974 | } |
4975 | if (ret) |
4976 | ret = notifier_from_errno(err: ret); |
4977 | else |
4978 | ret = NOTIFY_OK; |
4979 | return ret; |
4980 | } |
4981 | |
4982 | /******************************************************************** |
4983 | * Basic setup of slabs |
4984 | *******************************************************************/ |
4985 | |
4986 | /* |
4987 | * Used for early kmem_cache structures that were allocated using |
4988 | * the page allocator. Allocate them properly then fix up the pointers |
4989 | * that may be pointing to the wrong kmem_cache structure. |
4990 | */ |
4991 | |
4992 | static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) |
4993 | { |
4994 | int node; |
4995 | struct kmem_cache *s = kmem_cache_zalloc(k: kmem_cache, GFP_NOWAIT); |
4996 | struct kmem_cache_node *n; |
4997 | |
4998 | memcpy(s, static_cache, kmem_cache->object_size); |
4999 | |
5000 | /* |
5001 | * This runs very early, and only the boot processor is supposed to be |
5002 | * up. Even if it weren't true, IRQs are not up so we couldn't fire |
5003 | * IPIs around. |
5004 | */ |
5005 | __flush_cpu_slab(s, smp_processor_id()); |
5006 | for_each_kmem_cache_node(s, node, n) { |
5007 | struct slab *p; |
5008 | |
5009 | list_for_each_entry(p, &n->partial, slab_list) |
5010 | p->slab_cache = s; |
5011 | |
5012 | #ifdef CONFIG_SLUB_DEBUG |
5013 | list_for_each_entry(p, &n->full, slab_list) |
5014 | p->slab_cache = s; |
5015 | #endif |
5016 | } |
5017 | list_add(new: &s->list, head: &slab_caches); |
5018 | return s; |
5019 | } |
5020 | |
5021 | void __init kmem_cache_init(void) |
5022 | { |
5023 | static __initdata struct kmem_cache boot_kmem_cache, |
5024 | boot_kmem_cache_node; |
5025 | int node; |
5026 | |
5027 | if (debug_guardpage_minorder()) |
5028 | slub_max_order = 0; |
5029 | |
5030 | /* Print slub debugging pointers without hashing */ |
5031 | if (__slub_debug_enabled()) |
5032 | no_hash_pointers_enable(NULL); |
5033 | |
5034 | kmem_cache_node = &boot_kmem_cache_node; |
5035 | kmem_cache = &boot_kmem_cache; |
5036 | |
5037 | /* |
5038 | * Initialize the nodemask for which we will allocate per node |
5039 | * structures. Here we don't need taking slab_mutex yet. |
5040 | */ |
5041 | for_each_node_state(node, N_NORMAL_MEMORY) |
5042 | node_set(node, slab_nodes); |
5043 | |
5044 | create_boot_cache(kmem_cache_node, name: "kmem_cache_node" , |
5045 | size: sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, useroffset: 0, usersize: 0); |
5046 | |
5047 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); |
5048 | |
5049 | /* Able to allocate the per node structures */ |
5050 | slab_state = PARTIAL; |
5051 | |
5052 | create_boot_cache(kmem_cache, name: "kmem_cache" , |
5053 | offsetof(struct kmem_cache, node) + |
5054 | nr_node_ids * sizeof(struct kmem_cache_node *), |
5055 | SLAB_HWCACHE_ALIGN, useroffset: 0, usersize: 0); |
5056 | |
5057 | kmem_cache = bootstrap(static_cache: &boot_kmem_cache); |
5058 | kmem_cache_node = bootstrap(static_cache: &boot_kmem_cache_node); |
5059 | |
5060 | /* Now we can use the kmem_cache to allocate kmalloc slabs */ |
5061 | setup_kmalloc_cache_index_table(); |
5062 | create_kmalloc_caches(0); |
5063 | |
5064 | /* Setup random freelists for each cache */ |
5065 | init_freelist_randomization(); |
5066 | |
5067 | cpuhp_setup_state_nocalls(state: CPUHP_SLUB_DEAD, name: "slub:dead" , NULL, |
5068 | teardown: slub_cpu_dead); |
5069 | |
5070 | pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n" , |
5071 | cache_line_size(), |
5072 | slub_min_order, slub_max_order, slub_min_objects, |
5073 | nr_cpu_ids, nr_node_ids); |
5074 | } |
5075 | |
5076 | void __init kmem_cache_init_late(void) |
5077 | { |
5078 | #ifndef CONFIG_SLUB_TINY |
5079 | flushwq = alloc_workqueue("slub_flushwq" , WQ_MEM_RECLAIM, 0); |
5080 | WARN_ON(!flushwq); |
5081 | #endif |
5082 | } |
5083 | |
5084 | struct kmem_cache * |
5085 | __kmem_cache_alias(const char *name, unsigned int size, unsigned int align, |
5086 | slab_flags_t flags, void (*ctor)(void *)) |
5087 | { |
5088 | struct kmem_cache *s; |
5089 | |
5090 | s = find_mergeable(size, align, flags, name, ctor); |
5091 | if (s) { |
5092 | if (sysfs_slab_alias(s, p: name)) |
5093 | return NULL; |
5094 | |
5095 | s->refcount++; |
5096 | |
5097 | /* |
5098 | * Adjust the object sizes so that we clear |
5099 | * the complete object on kzalloc. |
5100 | */ |
5101 | s->object_size = max(s->object_size, size); |
5102 | s->inuse = max(s->inuse, ALIGN(size, sizeof(void *))); |
5103 | } |
5104 | |
5105 | return s; |
5106 | } |
5107 | |
5108 | int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags) |
5109 | { |
5110 | int err; |
5111 | |
5112 | err = kmem_cache_open(s, flags); |
5113 | if (err) |
5114 | return err; |
5115 | |
5116 | /* Mutex is not taken during early boot */ |
5117 | if (slab_state <= UP) |
5118 | return 0; |
5119 | |
5120 | err = sysfs_slab_add(s); |
5121 | if (err) { |
5122 | __kmem_cache_release(s); |
5123 | return err; |
5124 | } |
5125 | |
5126 | if (s->flags & SLAB_STORE_USER) |
5127 | debugfs_slab_add(s); |
5128 | |
5129 | return 0; |
5130 | } |
5131 | |
5132 | #ifdef SLAB_SUPPORTS_SYSFS |
5133 | static int count_inuse(struct slab *slab) |
5134 | { |
5135 | return slab->inuse; |
5136 | } |
5137 | |
5138 | static int count_total(struct slab *slab) |
5139 | { |
5140 | return slab->objects; |
5141 | } |
5142 | #endif |
5143 | |
5144 | #ifdef CONFIG_SLUB_DEBUG |
5145 | static void validate_slab(struct kmem_cache *s, struct slab *slab, |
5146 | unsigned long *obj_map) |
5147 | { |
5148 | void *p; |
5149 | void *addr = slab_address(slab); |
5150 | |
5151 | if (!check_slab(s, slab) || !on_freelist(s, slab, NULL)) |
5152 | return; |
5153 | |
5154 | /* Now we know that a valid freelist exists */ |
5155 | __fill_map(obj_map, s, slab); |
5156 | for_each_object(p, s, addr, slab->objects) { |
5157 | u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ? |
5158 | SLUB_RED_INACTIVE : SLUB_RED_ACTIVE; |
5159 | |
5160 | if (!check_object(s, slab, p, val)) |
5161 | break; |
5162 | } |
5163 | } |
5164 | |
5165 | static int validate_slab_node(struct kmem_cache *s, |
5166 | struct kmem_cache_node *n, unsigned long *obj_map) |
5167 | { |
5168 | unsigned long count = 0; |
5169 | struct slab *slab; |
5170 | unsigned long flags; |
5171 | |
5172 | spin_lock_irqsave(&n->list_lock, flags); |
5173 | |
5174 | list_for_each_entry(slab, &n->partial, slab_list) { |
5175 | validate_slab(s, slab, obj_map); |
5176 | count++; |
5177 | } |
5178 | if (count != n->nr_partial) { |
5179 | pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n" , |
5180 | s->name, count, n->nr_partial); |
5181 | slab_add_kunit_errors(); |
5182 | } |
5183 | |
5184 | if (!(s->flags & SLAB_STORE_USER)) |
5185 | goto out; |
5186 | |
5187 | list_for_each_entry(slab, &n->full, slab_list) { |
5188 | validate_slab(s, slab, obj_map); |
5189 | count++; |
5190 | } |
5191 | if (count != node_nr_slabs(n)) { |
5192 | pr_err("SLUB: %s %ld slabs counted but counter=%ld\n" , |
5193 | s->name, count, node_nr_slabs(n)); |
5194 | slab_add_kunit_errors(); |
5195 | } |
5196 | |
5197 | out: |
5198 | spin_unlock_irqrestore(&n->list_lock, flags); |
5199 | return count; |
5200 | } |
5201 | |
5202 | long validate_slab_cache(struct kmem_cache *s) |
5203 | { |
5204 | int node; |
5205 | unsigned long count = 0; |
5206 | struct kmem_cache_node *n; |
5207 | unsigned long *obj_map; |
5208 | |
5209 | obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL); |
5210 | if (!obj_map) |
5211 | return -ENOMEM; |
5212 | |
5213 | flush_all(s); |
5214 | for_each_kmem_cache_node(s, node, n) |
5215 | count += validate_slab_node(s, n, obj_map); |
5216 | |
5217 | bitmap_free(obj_map); |
5218 | |
5219 | return count; |
5220 | } |
5221 | EXPORT_SYMBOL(validate_slab_cache); |
5222 | |
5223 | #ifdef CONFIG_DEBUG_FS |
5224 | /* |
5225 | * Generate lists of code addresses where slabcache objects are allocated |
5226 | * and freed. |
5227 | */ |
5228 | |
5229 | struct location { |
5230 | depot_stack_handle_t handle; |
5231 | unsigned long count; |
5232 | unsigned long addr; |
5233 | unsigned long waste; |
5234 | long long sum_time; |
5235 | long min_time; |
5236 | long max_time; |
5237 | long min_pid; |
5238 | long max_pid; |
5239 | DECLARE_BITMAP(cpus, NR_CPUS); |
5240 | nodemask_t nodes; |
5241 | }; |
5242 | |
5243 | struct loc_track { |
5244 | unsigned long max; |
5245 | unsigned long count; |
5246 | struct location *loc; |
5247 | loff_t idx; |
5248 | }; |
5249 | |
5250 | static struct dentry *slab_debugfs_root; |
5251 | |
5252 | static void free_loc_track(struct loc_track *t) |
5253 | { |
5254 | if (t->max) |
5255 | free_pages((unsigned long)t->loc, |
5256 | get_order(sizeof(struct location) * t->max)); |
5257 | } |
5258 | |
5259 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) |
5260 | { |
5261 | struct location *l; |
5262 | int order; |
5263 | |
5264 | order = get_order(sizeof(struct location) * max); |
5265 | |
5266 | l = (void *)__get_free_pages(flags, order); |
5267 | if (!l) |
5268 | return 0; |
5269 | |
5270 | if (t->count) { |
5271 | memcpy(l, t->loc, sizeof(struct location) * t->count); |
5272 | free_loc_track(t); |
5273 | } |
5274 | t->max = max; |
5275 | t->loc = l; |
5276 | return 1; |
5277 | } |
5278 | |
5279 | static int add_location(struct loc_track *t, struct kmem_cache *s, |
5280 | const struct track *track, |
5281 | unsigned int orig_size) |
5282 | { |
5283 | long start, end, pos; |
5284 | struct location *l; |
5285 | unsigned long caddr, chandle, cwaste; |
5286 | unsigned long age = jiffies - track->when; |
5287 | depot_stack_handle_t handle = 0; |
5288 | unsigned int waste = s->object_size - orig_size; |
5289 | |
5290 | #ifdef CONFIG_STACKDEPOT |
5291 | handle = READ_ONCE(track->handle); |
5292 | #endif |
5293 | start = -1; |
5294 | end = t->count; |
5295 | |
5296 | for ( ; ; ) { |
5297 | pos = start + (end - start + 1) / 2; |
5298 | |
5299 | /* |
5300 | * There is nothing at "end". If we end up there |
5301 | * we need to add something to before end. |
5302 | */ |
5303 | if (pos == end) |
5304 | break; |
5305 | |
5306 | l = &t->loc[pos]; |
5307 | caddr = l->addr; |
5308 | chandle = l->handle; |
5309 | cwaste = l->waste; |
5310 | if ((track->addr == caddr) && (handle == chandle) && |
5311 | (waste == cwaste)) { |
5312 | |
5313 | l->count++; |
5314 | if (track->when) { |
5315 | l->sum_time += age; |
5316 | if (age < l->min_time) |
5317 | l->min_time = age; |
5318 | if (age > l->max_time) |
5319 | l->max_time = age; |
5320 | |
5321 | if (track->pid < l->min_pid) |
5322 | l->min_pid = track->pid; |
5323 | if (track->pid > l->max_pid) |
5324 | l->max_pid = track->pid; |
5325 | |
5326 | cpumask_set_cpu(track->cpu, |
5327 | to_cpumask(l->cpus)); |
5328 | } |
5329 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
5330 | return 1; |
5331 | } |
5332 | |
5333 | if (track->addr < caddr) |
5334 | end = pos; |
5335 | else if (track->addr == caddr && handle < chandle) |
5336 | end = pos; |
5337 | else if (track->addr == caddr && handle == chandle && |
5338 | waste < cwaste) |
5339 | end = pos; |
5340 | else |
5341 | start = pos; |
5342 | } |
5343 | |
5344 | /* |
5345 | * Not found. Insert new tracking element. |
5346 | */ |
5347 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) |
5348 | return 0; |
5349 | |
5350 | l = t->loc + pos; |
5351 | if (pos < t->count) |
5352 | memmove(l + 1, l, |
5353 | (t->count - pos) * sizeof(struct location)); |
5354 | t->count++; |
5355 | l->count = 1; |
5356 | l->addr = track->addr; |
5357 | l->sum_time = age; |
5358 | l->min_time = age; |
5359 | l->max_time = age; |
5360 | l->min_pid = track->pid; |
5361 | l->max_pid = track->pid; |
5362 | l->handle = handle; |
5363 | l->waste = waste; |
5364 | cpumask_clear(to_cpumask(l->cpus)); |
5365 | cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); |
5366 | nodes_clear(l->nodes); |
5367 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
5368 | return 1; |
5369 | } |
5370 | |
5371 | static void process_slab(struct loc_track *t, struct kmem_cache *s, |
5372 | struct slab *slab, enum track_item alloc, |
5373 | unsigned long *obj_map) |
5374 | { |
5375 | void *addr = slab_address(slab); |
5376 | bool is_alloc = (alloc == TRACK_ALLOC); |
5377 | void *p; |
5378 | |
5379 | __fill_map(obj_map, s, slab); |
5380 | |
5381 | for_each_object(p, s, addr, slab->objects) |
5382 | if (!test_bit(__obj_to_index(s, addr, p), obj_map)) |
5383 | add_location(t, s, get_track(s, p, alloc), |
5384 | is_alloc ? get_orig_size(s, p) : |
5385 | s->object_size); |
5386 | } |
5387 | #endif /* CONFIG_DEBUG_FS */ |
5388 | #endif /* CONFIG_SLUB_DEBUG */ |
5389 | |
5390 | #ifdef SLAB_SUPPORTS_SYSFS |
5391 | enum slab_stat_type { |
5392 | SL_ALL, /* All slabs */ |
5393 | SL_PARTIAL, /* Only partially allocated slabs */ |
5394 | SL_CPU, /* Only slabs used for cpu caches */ |
5395 | SL_OBJECTS, /* Determine allocated objects not slabs */ |
5396 | SL_TOTAL /* Determine object capacity not slabs */ |
5397 | }; |
5398 | |
5399 | #define SO_ALL (1 << SL_ALL) |
5400 | #define SO_PARTIAL (1 << SL_PARTIAL) |
5401 | #define SO_CPU (1 << SL_CPU) |
5402 | #define SO_OBJECTS (1 << SL_OBJECTS) |
5403 | #define SO_TOTAL (1 << SL_TOTAL) |
5404 | |
5405 | static ssize_t show_slab_objects(struct kmem_cache *s, |
5406 | char *buf, unsigned long flags) |
5407 | { |
5408 | unsigned long total = 0; |
5409 | int node; |
5410 | int x; |
5411 | unsigned long *nodes; |
5412 | int len = 0; |
5413 | |
5414 | nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL); |
5415 | if (!nodes) |
5416 | return -ENOMEM; |
5417 | |
5418 | if (flags & SO_CPU) { |
5419 | int cpu; |
5420 | |
5421 | for_each_possible_cpu(cpu) { |
5422 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, |
5423 | cpu); |
5424 | int node; |
5425 | struct slab *slab; |
5426 | |
5427 | slab = READ_ONCE(c->slab); |
5428 | if (!slab) |
5429 | continue; |
5430 | |
5431 | node = slab_nid(slab); |
5432 | if (flags & SO_TOTAL) |
5433 | x = slab->objects; |
5434 | else if (flags & SO_OBJECTS) |
5435 | x = slab->inuse; |
5436 | else |
5437 | x = 1; |
5438 | |
5439 | total += x; |
5440 | nodes[node] += x; |
5441 | |
5442 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
5443 | slab = slub_percpu_partial_read_once(c); |
5444 | if (slab) { |
5445 | node = slab_nid(slab); |
5446 | if (flags & SO_TOTAL) |
5447 | WARN_ON_ONCE(1); |
5448 | else if (flags & SO_OBJECTS) |
5449 | WARN_ON_ONCE(1); |
5450 | else |
5451 | x = slab->slabs; |
5452 | total += x; |
5453 | nodes[node] += x; |
5454 | } |
5455 | #endif |
5456 | } |
5457 | } |
5458 | |
5459 | /* |
5460 | * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex" |
5461 | * already held which will conflict with an existing lock order: |
5462 | * |
5463 | * mem_hotplug_lock->slab_mutex->kernfs_mutex |
5464 | * |
5465 | * We don't really need mem_hotplug_lock (to hold off |
5466 | * slab_mem_going_offline_callback) here because slab's memory hot |
5467 | * unplug code doesn't destroy the kmem_cache->node[] data. |
5468 | */ |
5469 | |
5470 | #ifdef CONFIG_SLUB_DEBUG |
5471 | if (flags & SO_ALL) { |
5472 | struct kmem_cache_node *n; |
5473 | |
5474 | for_each_kmem_cache_node(s, node, n) { |
5475 | |
5476 | if (flags & SO_TOTAL) |
5477 | x = node_nr_objs(n); |
5478 | else if (flags & SO_OBJECTS) |
5479 | x = node_nr_objs(n) - count_partial(n, count_free); |
5480 | else |
5481 | x = node_nr_slabs(n); |
5482 | total += x; |
5483 | nodes[node] += x; |
5484 | } |
5485 | |
5486 | } else |
5487 | #endif |
5488 | if (flags & SO_PARTIAL) { |
5489 | struct kmem_cache_node *n; |
5490 | |
5491 | for_each_kmem_cache_node(s, node, n) { |
5492 | if (flags & SO_TOTAL) |
5493 | x = count_partial(n, count_total); |
5494 | else if (flags & SO_OBJECTS) |
5495 | x = count_partial(n, count_inuse); |
5496 | else |
5497 | x = n->nr_partial; |
5498 | total += x; |
5499 | nodes[node] += x; |
5500 | } |
5501 | } |
5502 | |
5503 | len += sysfs_emit_at(buf, len, "%lu" , total); |
5504 | #ifdef CONFIG_NUMA |
5505 | for (node = 0; node < nr_node_ids; node++) { |
5506 | if (nodes[node]) |
5507 | len += sysfs_emit_at(buf, len, " N%d=%lu" , |
5508 | node, nodes[node]); |
5509 | } |
5510 | #endif |
5511 | len += sysfs_emit_at(buf, len, "\n" ); |
5512 | kfree(nodes); |
5513 | |
5514 | return len; |
5515 | } |
5516 | |
5517 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) |
5518 | #define to_slab(n) container_of(n, struct kmem_cache, kobj) |
5519 | |
5520 | struct slab_attribute { |
5521 | struct attribute attr; |
5522 | ssize_t (*show)(struct kmem_cache *s, char *buf); |
5523 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); |
5524 | }; |
5525 | |
5526 | #define SLAB_ATTR_RO(_name) \ |
5527 | static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400) |
5528 | |
5529 | #define SLAB_ATTR(_name) \ |
5530 | static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600) |
5531 | |
5532 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
5533 | { |
5534 | return sysfs_emit(buf, "%u\n" , s->size); |
5535 | } |
5536 | SLAB_ATTR_RO(slab_size); |
5537 | |
5538 | static ssize_t align_show(struct kmem_cache *s, char *buf) |
5539 | { |
5540 | return sysfs_emit(buf, "%u\n" , s->align); |
5541 | } |
5542 | SLAB_ATTR_RO(align); |
5543 | |
5544 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) |
5545 | { |
5546 | return sysfs_emit(buf, "%u\n" , s->object_size); |
5547 | } |
5548 | SLAB_ATTR_RO(object_size); |
5549 | |
5550 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) |
5551 | { |
5552 | return sysfs_emit(buf, "%u\n" , oo_objects(s->oo)); |
5553 | } |
5554 | SLAB_ATTR_RO(objs_per_slab); |
5555 | |
5556 | static ssize_t order_show(struct kmem_cache *s, char *buf) |
5557 | { |
5558 | return sysfs_emit(buf, "%u\n" , oo_order(s->oo)); |
5559 | } |
5560 | SLAB_ATTR_RO(order); |
5561 | |
5562 | static ssize_t min_partial_show(struct kmem_cache *s, char *buf) |
5563 | { |
5564 | return sysfs_emit(buf, "%lu\n" , s->min_partial); |
5565 | } |
5566 | |
5567 | static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, |
5568 | size_t length) |
5569 | { |
5570 | unsigned long min; |
5571 | int err; |
5572 | |
5573 | err = kstrtoul(buf, 10, &min); |
5574 | if (err) |
5575 | return err; |
5576 | |
5577 | s->min_partial = min; |
5578 | return length; |
5579 | } |
5580 | SLAB_ATTR(min_partial); |
5581 | |
5582 | static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) |
5583 | { |
5584 | unsigned int nr_partial = 0; |
5585 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
5586 | nr_partial = s->cpu_partial; |
5587 | #endif |
5588 | |
5589 | return sysfs_emit(buf, "%u\n" , nr_partial); |
5590 | } |
5591 | |
5592 | static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, |
5593 | size_t length) |
5594 | { |
5595 | unsigned int objects; |
5596 | int err; |
5597 | |
5598 | err = kstrtouint(buf, 10, &objects); |
5599 | if (err) |
5600 | return err; |
5601 | if (objects && !kmem_cache_has_cpu_partial(s)) |
5602 | return -EINVAL; |
5603 | |
5604 | slub_set_cpu_partial(s, objects); |
5605 | flush_all(s); |
5606 | return length; |
5607 | } |
5608 | SLAB_ATTR(cpu_partial); |
5609 | |
5610 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) |
5611 | { |
5612 | if (!s->ctor) |
5613 | return 0; |
5614 | return sysfs_emit(buf, "%pS\n" , s->ctor); |
5615 | } |
5616 | SLAB_ATTR_RO(ctor); |
5617 | |
5618 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) |
5619 | { |
5620 | return sysfs_emit(buf, "%d\n" , s->refcount < 0 ? 0 : s->refcount - 1); |
5621 | } |
5622 | SLAB_ATTR_RO(aliases); |
5623 | |
5624 | static ssize_t partial_show(struct kmem_cache *s, char *buf) |
5625 | { |
5626 | return show_slab_objects(s, buf, SO_PARTIAL); |
5627 | } |
5628 | SLAB_ATTR_RO(partial); |
5629 | |
5630 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) |
5631 | { |
5632 | return show_slab_objects(s, buf, SO_CPU); |
5633 | } |
5634 | SLAB_ATTR_RO(cpu_slabs); |
5635 | |
5636 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) |
5637 | { |
5638 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); |
5639 | } |
5640 | SLAB_ATTR_RO(objects_partial); |
5641 | |
5642 | static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) |
5643 | { |
5644 | int objects = 0; |
5645 | int slabs = 0; |
5646 | int cpu __maybe_unused; |
5647 | int len = 0; |
5648 | |
5649 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
5650 | for_each_online_cpu(cpu) { |
5651 | struct slab *slab; |
5652 | |
5653 | slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); |
5654 | |
5655 | if (slab) |
5656 | slabs += slab->slabs; |
5657 | } |
5658 | #endif |
5659 | |
5660 | /* Approximate half-full slabs, see slub_set_cpu_partial() */ |
5661 | objects = (slabs * oo_objects(s->oo)) / 2; |
5662 | len += sysfs_emit_at(buf, len, "%d(%d)" , objects, slabs); |
5663 | |
5664 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
5665 | for_each_online_cpu(cpu) { |
5666 | struct slab *slab; |
5667 | |
5668 | slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); |
5669 | if (slab) { |
5670 | slabs = READ_ONCE(slab->slabs); |
5671 | objects = (slabs * oo_objects(s->oo)) / 2; |
5672 | len += sysfs_emit_at(buf, len, " C%d=%d(%d)" , |
5673 | cpu, objects, slabs); |
5674 | } |
5675 | } |
5676 | #endif |
5677 | len += sysfs_emit_at(buf, len, "\n" ); |
5678 | |
5679 | return len; |
5680 | } |
5681 | SLAB_ATTR_RO(slabs_cpu_partial); |
5682 | |
5683 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) |
5684 | { |
5685 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_RECLAIM_ACCOUNT)); |
5686 | } |
5687 | SLAB_ATTR_RO(reclaim_account); |
5688 | |
5689 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) |
5690 | { |
5691 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_HWCACHE_ALIGN)); |
5692 | } |
5693 | SLAB_ATTR_RO(hwcache_align); |
5694 | |
5695 | #ifdef CONFIG_ZONE_DMA |
5696 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) |
5697 | { |
5698 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_CACHE_DMA)); |
5699 | } |
5700 | SLAB_ATTR_RO(cache_dma); |
5701 | #endif |
5702 | |
5703 | #ifdef CONFIG_HARDENED_USERCOPY |
5704 | static ssize_t usersize_show(struct kmem_cache *s, char *buf) |
5705 | { |
5706 | return sysfs_emit(buf, "%u\n" , s->usersize); |
5707 | } |
5708 | SLAB_ATTR_RO(usersize); |
5709 | #endif |
5710 | |
5711 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) |
5712 | { |
5713 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_TYPESAFE_BY_RCU)); |
5714 | } |
5715 | SLAB_ATTR_RO(destroy_by_rcu); |
5716 | |
5717 | #ifdef CONFIG_SLUB_DEBUG |
5718 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) |
5719 | { |
5720 | return show_slab_objects(s, buf, SO_ALL); |
5721 | } |
5722 | SLAB_ATTR_RO(slabs); |
5723 | |
5724 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) |
5725 | { |
5726 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); |
5727 | } |
5728 | SLAB_ATTR_RO(total_objects); |
5729 | |
5730 | static ssize_t objects_show(struct kmem_cache *s, char *buf) |
5731 | { |
5732 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); |
5733 | } |
5734 | SLAB_ATTR_RO(objects); |
5735 | |
5736 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) |
5737 | { |
5738 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_CONSISTENCY_CHECKS)); |
5739 | } |
5740 | SLAB_ATTR_RO(sanity_checks); |
5741 | |
5742 | static ssize_t trace_show(struct kmem_cache *s, char *buf) |
5743 | { |
5744 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_TRACE)); |
5745 | } |
5746 | SLAB_ATTR_RO(trace); |
5747 | |
5748 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) |
5749 | { |
5750 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_RED_ZONE)); |
5751 | } |
5752 | |
5753 | SLAB_ATTR_RO(red_zone); |
5754 | |
5755 | static ssize_t poison_show(struct kmem_cache *s, char *buf) |
5756 | { |
5757 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_POISON)); |
5758 | } |
5759 | |
5760 | SLAB_ATTR_RO(poison); |
5761 | |
5762 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) |
5763 | { |
5764 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_STORE_USER)); |
5765 | } |
5766 | |
5767 | SLAB_ATTR_RO(store_user); |
5768 | |
5769 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
5770 | { |
5771 | return 0; |
5772 | } |
5773 | |
5774 | static ssize_t validate_store(struct kmem_cache *s, |
5775 | const char *buf, size_t length) |
5776 | { |
5777 | int ret = -EINVAL; |
5778 | |
5779 | if (buf[0] == '1' && kmem_cache_debug(s)) { |
5780 | ret = validate_slab_cache(s); |
5781 | if (ret >= 0) |
5782 | ret = length; |
5783 | } |
5784 | return ret; |
5785 | } |
5786 | SLAB_ATTR(validate); |
5787 | |
5788 | #endif /* CONFIG_SLUB_DEBUG */ |
5789 | |
5790 | #ifdef CONFIG_FAILSLAB |
5791 | static ssize_t failslab_show(struct kmem_cache *s, char *buf) |
5792 | { |
5793 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_FAILSLAB)); |
5794 | } |
5795 | |
5796 | static ssize_t failslab_store(struct kmem_cache *s, const char *buf, |
5797 | size_t length) |
5798 | { |
5799 | if (s->refcount > 1) |
5800 | return -EINVAL; |
5801 | |
5802 | if (buf[0] == '1') |
5803 | WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB); |
5804 | else |
5805 | WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB); |
5806 | |
5807 | return length; |
5808 | } |
5809 | SLAB_ATTR(failslab); |
5810 | #endif |
5811 | |
5812 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
5813 | { |
5814 | return 0; |
5815 | } |
5816 | |
5817 | static ssize_t shrink_store(struct kmem_cache *s, |
5818 | const char *buf, size_t length) |
5819 | { |
5820 | if (buf[0] == '1') |
5821 | kmem_cache_shrink(s); |
5822 | else |
5823 | return -EINVAL; |
5824 | return length; |
5825 | } |
5826 | SLAB_ATTR(shrink); |
5827 | |
5828 | #ifdef CONFIG_NUMA |
5829 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) |
5830 | { |
5831 | return sysfs_emit(buf, "%u\n" , s->remote_node_defrag_ratio / 10); |
5832 | } |
5833 | |
5834 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, |
5835 | const char *buf, size_t length) |
5836 | { |
5837 | unsigned int ratio; |
5838 | int err; |
5839 | |
5840 | err = kstrtouint(buf, 10, &ratio); |
5841 | if (err) |
5842 | return err; |
5843 | if (ratio > 100) |
5844 | return -ERANGE; |
5845 | |
5846 | s->remote_node_defrag_ratio = ratio * 10; |
5847 | |
5848 | return length; |
5849 | } |
5850 | SLAB_ATTR(remote_node_defrag_ratio); |
5851 | #endif |
5852 | |
5853 | #ifdef CONFIG_SLUB_STATS |
5854 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) |
5855 | { |
5856 | unsigned long sum = 0; |
5857 | int cpu; |
5858 | int len = 0; |
5859 | int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL); |
5860 | |
5861 | if (!data) |
5862 | return -ENOMEM; |
5863 | |
5864 | for_each_online_cpu(cpu) { |
5865 | unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; |
5866 | |
5867 | data[cpu] = x; |
5868 | sum += x; |
5869 | } |
5870 | |
5871 | len += sysfs_emit_at(buf, len, "%lu" , sum); |
5872 | |
5873 | #ifdef CONFIG_SMP |
5874 | for_each_online_cpu(cpu) { |
5875 | if (data[cpu]) |
5876 | len += sysfs_emit_at(buf, len, " C%d=%u" , |
5877 | cpu, data[cpu]); |
5878 | } |
5879 | #endif |
5880 | kfree(data); |
5881 | len += sysfs_emit_at(buf, len, "\n" ); |
5882 | |
5883 | return len; |
5884 | } |
5885 | |
5886 | static void clear_stat(struct kmem_cache *s, enum stat_item si) |
5887 | { |
5888 | int cpu; |
5889 | |
5890 | for_each_online_cpu(cpu) |
5891 | per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; |
5892 | } |
5893 | |
5894 | #define STAT_ATTR(si, text) \ |
5895 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ |
5896 | { \ |
5897 | return show_stat(s, buf, si); \ |
5898 | } \ |
5899 | static ssize_t text##_store(struct kmem_cache *s, \ |
5900 | const char *buf, size_t length) \ |
5901 | { \ |
5902 | if (buf[0] != '0') \ |
5903 | return -EINVAL; \ |
5904 | clear_stat(s, si); \ |
5905 | return length; \ |
5906 | } \ |
5907 | SLAB_ATTR(text); \ |
5908 | |
5909 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); |
5910 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); |
5911 | STAT_ATTR(FREE_FASTPATH, free_fastpath); |
5912 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); |
5913 | STAT_ATTR(FREE_FROZEN, free_frozen); |
5914 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); |
5915 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); |
5916 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); |
5917 | STAT_ATTR(ALLOC_SLAB, alloc_slab); |
5918 | STAT_ATTR(ALLOC_REFILL, alloc_refill); |
5919 | STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); |
5920 | STAT_ATTR(FREE_SLAB, free_slab); |
5921 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); |
5922 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); |
5923 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); |
5924 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); |
5925 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); |
5926 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); |
5927 | STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); |
5928 | STAT_ATTR(ORDER_FALLBACK, order_fallback); |
5929 | STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); |
5930 | STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); |
5931 | STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); |
5932 | STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); |
5933 | STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); |
5934 | STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); |
5935 | #endif /* CONFIG_SLUB_STATS */ |
5936 | |
5937 | #ifdef CONFIG_KFENCE |
5938 | static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf) |
5939 | { |
5940 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_SKIP_KFENCE)); |
5941 | } |
5942 | |
5943 | static ssize_t skip_kfence_store(struct kmem_cache *s, |
5944 | const char *buf, size_t length) |
5945 | { |
5946 | int ret = length; |
5947 | |
5948 | if (buf[0] == '0') |
5949 | s->flags &= ~SLAB_SKIP_KFENCE; |
5950 | else if (buf[0] == '1') |
5951 | s->flags |= SLAB_SKIP_KFENCE; |
5952 | else |
5953 | ret = -EINVAL; |
5954 | |
5955 | return ret; |
5956 | } |
5957 | SLAB_ATTR(skip_kfence); |
5958 | #endif |
5959 | |
5960 | static struct attribute *slab_attrs[] = { |
5961 | &slab_size_attr.attr, |
5962 | &object_size_attr.attr, |
5963 | &objs_per_slab_attr.attr, |
5964 | &order_attr.attr, |
5965 | &min_partial_attr.attr, |
5966 | &cpu_partial_attr.attr, |
5967 | &objects_partial_attr.attr, |
5968 | &partial_attr.attr, |
5969 | &cpu_slabs_attr.attr, |
5970 | &ctor_attr.attr, |
5971 | &aliases_attr.attr, |
5972 | &align_attr.attr, |
5973 | &hwcache_align_attr.attr, |
5974 | &reclaim_account_attr.attr, |
5975 | &destroy_by_rcu_attr.attr, |
5976 | &shrink_attr.attr, |
5977 | &slabs_cpu_partial_attr.attr, |
5978 | #ifdef CONFIG_SLUB_DEBUG |
5979 | &total_objects_attr.attr, |
5980 | &objects_attr.attr, |
5981 | &slabs_attr.attr, |
5982 | &sanity_checks_attr.attr, |
5983 | &trace_attr.attr, |
5984 | &red_zone_attr.attr, |
5985 | &poison_attr.attr, |
5986 | &store_user_attr.attr, |
5987 | &validate_attr.attr, |
5988 | #endif |
5989 | #ifdef CONFIG_ZONE_DMA |
5990 | &cache_dma_attr.attr, |
5991 | #endif |
5992 | #ifdef CONFIG_NUMA |
5993 | &remote_node_defrag_ratio_attr.attr, |
5994 | #endif |
5995 | #ifdef CONFIG_SLUB_STATS |
5996 | &alloc_fastpath_attr.attr, |
5997 | &alloc_slowpath_attr.attr, |
5998 | &free_fastpath_attr.attr, |
5999 | &free_slowpath_attr.attr, |
6000 | &free_frozen_attr.attr, |
6001 | &free_add_partial_attr.attr, |
6002 | &free_remove_partial_attr.attr, |
6003 | &alloc_from_partial_attr.attr, |
6004 | &alloc_slab_attr.attr, |
6005 | &alloc_refill_attr.attr, |
6006 | &alloc_node_mismatch_attr.attr, |
6007 | &free_slab_attr.attr, |
6008 | &cpuslab_flush_attr.attr, |
6009 | &deactivate_full_attr.attr, |
6010 | &deactivate_empty_attr.attr, |
6011 | &deactivate_to_head_attr.attr, |
6012 | &deactivate_to_tail_attr.attr, |
6013 | &deactivate_remote_frees_attr.attr, |
6014 | &deactivate_bypass_attr.attr, |
6015 | &order_fallback_attr.attr, |
6016 | &cmpxchg_double_fail_attr.attr, |
6017 | &cmpxchg_double_cpu_fail_attr.attr, |
6018 | &cpu_partial_alloc_attr.attr, |
6019 | &cpu_partial_free_attr.attr, |
6020 | &cpu_partial_node_attr.attr, |
6021 | &cpu_partial_drain_attr.attr, |
6022 | #endif |
6023 | #ifdef CONFIG_FAILSLAB |
6024 | &failslab_attr.attr, |
6025 | #endif |
6026 | #ifdef CONFIG_HARDENED_USERCOPY |
6027 | &usersize_attr.attr, |
6028 | #endif |
6029 | #ifdef CONFIG_KFENCE |
6030 | &skip_kfence_attr.attr, |
6031 | #endif |
6032 | |
6033 | NULL |
6034 | }; |
6035 | |
6036 | static const struct attribute_group slab_attr_group = { |
6037 | .attrs = slab_attrs, |
6038 | }; |
6039 | |
6040 | static ssize_t slab_attr_show(struct kobject *kobj, |
6041 | struct attribute *attr, |
6042 | char *buf) |
6043 | { |
6044 | struct slab_attribute *attribute; |
6045 | struct kmem_cache *s; |
6046 | |
6047 | attribute = to_slab_attr(attr); |
6048 | s = to_slab(kobj); |
6049 | |
6050 | if (!attribute->show) |
6051 | return -EIO; |
6052 | |
6053 | return attribute->show(s, buf); |
6054 | } |
6055 | |
6056 | static ssize_t slab_attr_store(struct kobject *kobj, |
6057 | struct attribute *attr, |
6058 | const char *buf, size_t len) |
6059 | { |
6060 | struct slab_attribute *attribute; |
6061 | struct kmem_cache *s; |
6062 | |
6063 | attribute = to_slab_attr(attr); |
6064 | s = to_slab(kobj); |
6065 | |
6066 | if (!attribute->store) |
6067 | return -EIO; |
6068 | |
6069 | return attribute->store(s, buf, len); |
6070 | } |
6071 | |
6072 | static void kmem_cache_release(struct kobject *k) |
6073 | { |
6074 | slab_kmem_cache_release(to_slab(k)); |
6075 | } |
6076 | |
6077 | static const struct sysfs_ops slab_sysfs_ops = { |
6078 | .show = slab_attr_show, |
6079 | .store = slab_attr_store, |
6080 | }; |
6081 | |
6082 | static const struct kobj_type slab_ktype = { |
6083 | .sysfs_ops = &slab_sysfs_ops, |
6084 | .release = kmem_cache_release, |
6085 | }; |
6086 | |
6087 | static struct kset *slab_kset; |
6088 | |
6089 | static inline struct kset *cache_kset(struct kmem_cache *s) |
6090 | { |
6091 | return slab_kset; |
6092 | } |
6093 | |
6094 | #define ID_STR_LENGTH 32 |
6095 | |
6096 | /* Create a unique string id for a slab cache: |
6097 | * |
6098 | * Format :[flags-]size |
6099 | */ |
6100 | static char *create_unique_id(struct kmem_cache *s) |
6101 | { |
6102 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); |
6103 | char *p = name; |
6104 | |
6105 | if (!name) |
6106 | return ERR_PTR(-ENOMEM); |
6107 | |
6108 | *p++ = ':'; |
6109 | /* |
6110 | * First flags affecting slabcache operations. We will only |
6111 | * get here for aliasable slabs so we do not need to support |
6112 | * too many flags. The flags here must cover all flags that |
6113 | * are matched during merging to guarantee that the id is |
6114 | * unique. |
6115 | */ |
6116 | if (s->flags & SLAB_CACHE_DMA) |
6117 | *p++ = 'd'; |
6118 | if (s->flags & SLAB_CACHE_DMA32) |
6119 | *p++ = 'D'; |
6120 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
6121 | *p++ = 'a'; |
6122 | if (s->flags & SLAB_CONSISTENCY_CHECKS) |
6123 | *p++ = 'F'; |
6124 | if (s->flags & SLAB_ACCOUNT) |
6125 | *p++ = 'A'; |
6126 | if (p != name + 1) |
6127 | *p++ = '-'; |
6128 | p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u" , s->size); |
6129 | |
6130 | if (WARN_ON(p > name + ID_STR_LENGTH - 1)) { |
6131 | kfree(name); |
6132 | return ERR_PTR(-EINVAL); |
6133 | } |
6134 | kmsan_unpoison_memory(name, p - name); |
6135 | return name; |
6136 | } |
6137 | |
6138 | static int sysfs_slab_add(struct kmem_cache *s) |
6139 | { |
6140 | int err; |
6141 | const char *name; |
6142 | struct kset *kset = cache_kset(s); |
6143 | int unmergeable = slab_unmergeable(s); |
6144 | |
6145 | if (!unmergeable && disable_higher_order_debug && |
6146 | (slub_debug & DEBUG_METADATA_FLAGS)) |
6147 | unmergeable = 1; |
6148 | |
6149 | if (unmergeable) { |
6150 | /* |
6151 | * Slabcache can never be merged so we can use the name proper. |
6152 | * This is typically the case for debug situations. In that |
6153 | * case we can catch duplicate names easily. |
6154 | */ |
6155 | sysfs_remove_link(&slab_kset->kobj, s->name); |
6156 | name = s->name; |
6157 | } else { |
6158 | /* |
6159 | * Create a unique name for the slab as a target |
6160 | * for the symlinks. |
6161 | */ |
6162 | name = create_unique_id(s); |
6163 | if (IS_ERR(name)) |
6164 | return PTR_ERR(name); |
6165 | } |
6166 | |
6167 | s->kobj.kset = kset; |
6168 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s" , name); |
6169 | if (err) |
6170 | goto out; |
6171 | |
6172 | err = sysfs_create_group(&s->kobj, &slab_attr_group); |
6173 | if (err) |
6174 | goto out_del_kobj; |
6175 | |
6176 | if (!unmergeable) { |
6177 | /* Setup first alias */ |
6178 | sysfs_slab_alias(s, s->name); |
6179 | } |
6180 | out: |
6181 | if (!unmergeable) |
6182 | kfree(name); |
6183 | return err; |
6184 | out_del_kobj: |
6185 | kobject_del(&s->kobj); |
6186 | goto out; |
6187 | } |
6188 | |
6189 | void sysfs_slab_unlink(struct kmem_cache *s) |
6190 | { |
6191 | if (slab_state >= FULL) |
6192 | kobject_del(&s->kobj); |
6193 | } |
6194 | |
6195 | void sysfs_slab_release(struct kmem_cache *s) |
6196 | { |
6197 | if (slab_state >= FULL) |
6198 | kobject_put(&s->kobj); |
6199 | } |
6200 | |
6201 | /* |
6202 | * Need to buffer aliases during bootup until sysfs becomes |
6203 | * available lest we lose that information. |
6204 | */ |
6205 | struct saved_alias { |
6206 | struct kmem_cache *s; |
6207 | const char *name; |
6208 | struct saved_alias *next; |
6209 | }; |
6210 | |
6211 | static struct saved_alias *alias_list; |
6212 | |
6213 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) |
6214 | { |
6215 | struct saved_alias *al; |
6216 | |
6217 | if (slab_state == FULL) { |
6218 | /* |
6219 | * If we have a leftover link then remove it. |
6220 | */ |
6221 | sysfs_remove_link(&slab_kset->kobj, name); |
6222 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); |
6223 | } |
6224 | |
6225 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); |
6226 | if (!al) |
6227 | return -ENOMEM; |
6228 | |
6229 | al->s = s; |
6230 | al->name = name; |
6231 | al->next = alias_list; |
6232 | alias_list = al; |
6233 | kmsan_unpoison_memory(al, sizeof(*al)); |
6234 | return 0; |
6235 | } |
6236 | |
6237 | static int __init slab_sysfs_init(void) |
6238 | { |
6239 | struct kmem_cache *s; |
6240 | int err; |
6241 | |
6242 | mutex_lock(&slab_mutex); |
6243 | |
6244 | slab_kset = kset_create_and_add("slab" , NULL, kernel_kobj); |
6245 | if (!slab_kset) { |
6246 | mutex_unlock(&slab_mutex); |
6247 | pr_err("Cannot register slab subsystem.\n" ); |
6248 | return -ENOMEM; |
6249 | } |
6250 | |
6251 | slab_state = FULL; |
6252 | |
6253 | list_for_each_entry(s, &slab_caches, list) { |
6254 | err = sysfs_slab_add(s); |
6255 | if (err) |
6256 | pr_err("SLUB: Unable to add boot slab %s to sysfs\n" , |
6257 | s->name); |
6258 | } |
6259 | |
6260 | while (alias_list) { |
6261 | struct saved_alias *al = alias_list; |
6262 | |
6263 | alias_list = alias_list->next; |
6264 | err = sysfs_slab_alias(al->s, al->name); |
6265 | if (err) |
6266 | pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n" , |
6267 | al->name); |
6268 | kfree(al); |
6269 | } |
6270 | |
6271 | mutex_unlock(&slab_mutex); |
6272 | return 0; |
6273 | } |
6274 | late_initcall(slab_sysfs_init); |
6275 | #endif /* SLAB_SUPPORTS_SYSFS */ |
6276 | |
6277 | #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS) |
6278 | static int slab_debugfs_show(struct seq_file *seq, void *v) |
6279 | { |
6280 | struct loc_track *t = seq->private; |
6281 | struct location *l; |
6282 | unsigned long idx; |
6283 | |
6284 | idx = (unsigned long) t->idx; |
6285 | if (idx < t->count) { |
6286 | l = &t->loc[idx]; |
6287 | |
6288 | seq_printf(seq, "%7ld " , l->count); |
6289 | |
6290 | if (l->addr) |
6291 | seq_printf(seq, "%pS" , (void *)l->addr); |
6292 | else |
6293 | seq_puts(seq, "<not-available>" ); |
6294 | |
6295 | if (l->waste) |
6296 | seq_printf(seq, " waste=%lu/%lu" , |
6297 | l->count * l->waste, l->waste); |
6298 | |
6299 | if (l->sum_time != l->min_time) { |
6300 | seq_printf(seq, " age=%ld/%llu/%ld" , |
6301 | l->min_time, div_u64(l->sum_time, l->count), |
6302 | l->max_time); |
6303 | } else |
6304 | seq_printf(seq, " age=%ld" , l->min_time); |
6305 | |
6306 | if (l->min_pid != l->max_pid) |
6307 | seq_printf(seq, " pid=%ld-%ld" , l->min_pid, l->max_pid); |
6308 | else |
6309 | seq_printf(seq, " pid=%ld" , |
6310 | l->min_pid); |
6311 | |
6312 | if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus))) |
6313 | seq_printf(seq, " cpus=%*pbl" , |
6314 | cpumask_pr_args(to_cpumask(l->cpus))); |
6315 | |
6316 | if (nr_online_nodes > 1 && !nodes_empty(l->nodes)) |
6317 | seq_printf(seq, " nodes=%*pbl" , |
6318 | nodemask_pr_args(&l->nodes)); |
6319 | |
6320 | #ifdef CONFIG_STACKDEPOT |
6321 | { |
6322 | depot_stack_handle_t handle; |
6323 | unsigned long *entries; |
6324 | unsigned int nr_entries, j; |
6325 | |
6326 | handle = READ_ONCE(l->handle); |
6327 | if (handle) { |
6328 | nr_entries = stack_depot_fetch(handle, &entries); |
6329 | seq_puts(seq, "\n" ); |
6330 | for (j = 0; j < nr_entries; j++) |
6331 | seq_printf(seq, " %pS\n" , (void *)entries[j]); |
6332 | } |
6333 | } |
6334 | #endif |
6335 | seq_puts(seq, "\n" ); |
6336 | } |
6337 | |
6338 | if (!idx && !t->count) |
6339 | seq_puts(seq, "No data\n" ); |
6340 | |
6341 | return 0; |
6342 | } |
6343 | |
6344 | static void slab_debugfs_stop(struct seq_file *seq, void *v) |
6345 | { |
6346 | } |
6347 | |
6348 | static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos) |
6349 | { |
6350 | struct loc_track *t = seq->private; |
6351 | |
6352 | t->idx = ++(*ppos); |
6353 | if (*ppos <= t->count) |
6354 | return ppos; |
6355 | |
6356 | return NULL; |
6357 | } |
6358 | |
6359 | static int cmp_loc_by_count(const void *a, const void *b, const void *data) |
6360 | { |
6361 | struct location *loc1 = (struct location *)a; |
6362 | struct location *loc2 = (struct location *)b; |
6363 | |
6364 | if (loc1->count > loc2->count) |
6365 | return -1; |
6366 | else |
6367 | return 1; |
6368 | } |
6369 | |
6370 | static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos) |
6371 | { |
6372 | struct loc_track *t = seq->private; |
6373 | |
6374 | t->idx = *ppos; |
6375 | return ppos; |
6376 | } |
6377 | |
6378 | static const struct seq_operations slab_debugfs_sops = { |
6379 | .start = slab_debugfs_start, |
6380 | .next = slab_debugfs_next, |
6381 | .stop = slab_debugfs_stop, |
6382 | .show = slab_debugfs_show, |
6383 | }; |
6384 | |
6385 | static int slab_debug_trace_open(struct inode *inode, struct file *filep) |
6386 | { |
6387 | |
6388 | struct kmem_cache_node *n; |
6389 | enum track_item alloc; |
6390 | int node; |
6391 | struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops, |
6392 | sizeof(struct loc_track)); |
6393 | struct kmem_cache *s = file_inode(filep)->i_private; |
6394 | unsigned long *obj_map; |
6395 | |
6396 | if (!t) |
6397 | return -ENOMEM; |
6398 | |
6399 | obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL); |
6400 | if (!obj_map) { |
6401 | seq_release_private(inode, filep); |
6402 | return -ENOMEM; |
6403 | } |
6404 | |
6405 | if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces" ) == 0) |
6406 | alloc = TRACK_ALLOC; |
6407 | else |
6408 | alloc = TRACK_FREE; |
6409 | |
6410 | if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) { |
6411 | bitmap_free(obj_map); |
6412 | seq_release_private(inode, filep); |
6413 | return -ENOMEM; |
6414 | } |
6415 | |
6416 | for_each_kmem_cache_node(s, node, n) { |
6417 | unsigned long flags; |
6418 | struct slab *slab; |
6419 | |
6420 | if (!node_nr_slabs(n)) |
6421 | continue; |
6422 | |
6423 | spin_lock_irqsave(&n->list_lock, flags); |
6424 | list_for_each_entry(slab, &n->partial, slab_list) |
6425 | process_slab(t, s, slab, alloc, obj_map); |
6426 | list_for_each_entry(slab, &n->full, slab_list) |
6427 | process_slab(t, s, slab, alloc, obj_map); |
6428 | spin_unlock_irqrestore(&n->list_lock, flags); |
6429 | } |
6430 | |
6431 | /* Sort locations by count */ |
6432 | sort_r(t->loc, t->count, sizeof(struct location), |
6433 | cmp_loc_by_count, NULL, NULL); |
6434 | |
6435 | bitmap_free(obj_map); |
6436 | return 0; |
6437 | } |
6438 | |
6439 | static int slab_debug_trace_release(struct inode *inode, struct file *file) |
6440 | { |
6441 | struct seq_file *seq = file->private_data; |
6442 | struct loc_track *t = seq->private; |
6443 | |
6444 | free_loc_track(t); |
6445 | return seq_release_private(inode, file); |
6446 | } |
6447 | |
6448 | static const struct file_operations slab_debugfs_fops = { |
6449 | .open = slab_debug_trace_open, |
6450 | .read = seq_read, |
6451 | .llseek = seq_lseek, |
6452 | .release = slab_debug_trace_release, |
6453 | }; |
6454 | |
6455 | static void debugfs_slab_add(struct kmem_cache *s) |
6456 | { |
6457 | struct dentry *slab_cache_dir; |
6458 | |
6459 | if (unlikely(!slab_debugfs_root)) |
6460 | return; |
6461 | |
6462 | slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root); |
6463 | |
6464 | debugfs_create_file("alloc_traces" , 0400, |
6465 | slab_cache_dir, s, &slab_debugfs_fops); |
6466 | |
6467 | debugfs_create_file("free_traces" , 0400, |
6468 | slab_cache_dir, s, &slab_debugfs_fops); |
6469 | } |
6470 | |
6471 | void debugfs_slab_release(struct kmem_cache *s) |
6472 | { |
6473 | debugfs_lookup_and_remove(s->name, slab_debugfs_root); |
6474 | } |
6475 | |
6476 | static int __init slab_debugfs_init(void) |
6477 | { |
6478 | struct kmem_cache *s; |
6479 | |
6480 | slab_debugfs_root = debugfs_create_dir("slab" , NULL); |
6481 | |
6482 | list_for_each_entry(s, &slab_caches, list) |
6483 | if (s->flags & SLAB_STORE_USER) |
6484 | debugfs_slab_add(s); |
6485 | |
6486 | return 0; |
6487 | |
6488 | } |
6489 | __initcall(slab_debugfs_init); |
6490 | #endif |
6491 | /* |
6492 | * The /proc/slabinfo ABI |
6493 | */ |
6494 | #ifdef CONFIG_SLUB_DEBUG |
6495 | void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) |
6496 | { |
6497 | unsigned long nr_slabs = 0; |
6498 | unsigned long nr_objs = 0; |
6499 | unsigned long nr_free = 0; |
6500 | int node; |
6501 | struct kmem_cache_node *n; |
6502 | |
6503 | for_each_kmem_cache_node(s, node, n) { |
6504 | nr_slabs += node_nr_slabs(n); |
6505 | nr_objs += node_nr_objs(n); |
6506 | nr_free += count_partial(n, count_free); |
6507 | } |
6508 | |
6509 | sinfo->active_objs = nr_objs - nr_free; |
6510 | sinfo->num_objs = nr_objs; |
6511 | sinfo->active_slabs = nr_slabs; |
6512 | sinfo->num_slabs = nr_slabs; |
6513 | sinfo->objects_per_slab = oo_objects(s->oo); |
6514 | sinfo->cache_order = oo_order(s->oo); |
6515 | } |
6516 | |
6517 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) |
6518 | { |
6519 | } |
6520 | |
6521 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
6522 | size_t count, loff_t *ppos) |
6523 | { |
6524 | return -EIO; |
6525 | } |
6526 | #endif /* CONFIG_SLUB_DEBUG */ |
6527 | |